release: v0.1.6

Since v0.1.5:
- fix(messages): Claude models no longer 400 with 'Invalid signature in
  thinking block' — thinking is replayed only for the active tool loop
  and only when genuinely signed (cross-backend/cross-model histories
  are stripped)
- fix(codex): ChatGPT Codex no longer 400s with 'System messages are
  not allowed' — system prompts ride the instructions field and
  encrypted reasoning is requested for stateless replay
- docs: AGENTS.md records both wire contracts

.github/workflows/warm-cache.yml

@@ -0,0 +1,85 @@
+name: Warm build cache
+
+# Release builds run on TAG refs, and GitHub Actions cache isolation only
+# lets a run restore caches created on its OWN ref or the DEFAULT branch.
+# No workflow ran on `main`, so every release compiled the whole workspace
+# cold on all five targets (~50 min wall clock, gated by Windows). This
+# workflow builds the same `release-dist` profile on `main` so tag builds
+# restore a warm default-branch cache.
+#
+# Triggers: dependency-affecting pushes to main (each release's version-bump
+# commit warms the cache for the NEXT release), a weekly refresh so the
+# cache never hits GitHub's 7-day unused-eviction, and manual dispatch.
+#
+# The setup steps mirror release.yml's build job (toolchain, target,
+# dotslash/protoc, rust-cache key) — keep them in lockstep, or the cache
+# key won't match and releases go back to cold builds.
+
+on:
+  push:
+    branches: ["main"]
+    paths:
+      - "Cargo.lock"
+      - "Cargo.toml"
+      - "rust-toolchain.toml"
+      - ".github/workflows/warm-cache.yml"
+  schedule:
+    - cron: "17 5 * * 1"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+env:
+  CARGO_TERM_COLOR: always
+
+concurrency:
+  group: warm-cache
+  cancel-in-progress: true
+
+jobs:
+  warm:
+    name: warm (${{ matrix.target }})
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - target: aarch64-apple-darwin
+            os: macos-14
+          - target: x86_64-apple-darwin
+            os: macos-14
+          - target: x86_64-unknown-linux-gnu
+            os: ubuntu-24.04
+          - target: aarch64-unknown-linux-gnu
+            os: ubuntu-24.04-arm
+          - target: x86_64-pc-windows-msvc
+            os: windows-2022
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install toolchain (rust-toolchain.toml)
+        run: rustup show
+
+      - name: Add build target
+        run: rustup target add ${{ matrix.target }}
+
+      - name: Install dotslash (protoc launcher)
+        run: cargo install dotslash --locked
+
+      - name: Install protoc (Windows PATH fallback)
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $url = "https://github.com/protocolbuffers/protobuf/releases/download/v29.3/protoc-29.3-win64.zip"
+          Invoke-WebRequest -Uri $url -OutFile protoc.zip
+          Expand-Archive protoc.zip -DestinationPath "$env:USERPROFILE\protoc"
+          Add-Content $env:GITHUB_PATH "$env:USERPROFILE\protoc\bin"
+
+      # Same key as release.yml so tag builds restore this cache verbatim.
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: ${{ matrix.target }}
+
+      - name: Build kigi (release-dist)
+        run: cargo build --profile release-dist -p kigi-bin --locked --target ${{ matrix.target }}

AGENTS.md

@@ -18,9 +18,12 @@ import) or any `KIMI_*` env var.
 
 - **Zero egress**: outbound connections are limited to
   `auth.kimi.com`, `api.kimi.com`, `api.moonshot.cn`, `api.moonshot.ai`,
-  GitHub Releases domains, and user-configured MCP servers. No telemetry,
-  no analytics, ever. `crates/codegen/kigi-env` is the single home of
-  first-party endpoints.
+  GitHub Releases domains, user-configured MCP servers, the endpoints of
+  provider platforms the user has credentialed, and `models.dev` (model
+  metadata refresh — reached ONLY when an enabled platform's `/models` wire
+  lacks metadata, `wire_serves_metadata=false`; Kimi/Moonshot never trigger
+  it; `KIGI_MODELS_DEV_URL=0` disables). No telemetry, no analytics, ever.
+  `crates/codegen/kigi-env` is the single home of first-party endpoints.
 - **Toolchain**: Rust 1.97.0 (rust-toolchain.toml), edition 2024.
 - **Gates** (all must stay green):
   `cargo check --workspace --all-targets`,
@@ -29,9 +32,19 @@ import) or any `KIMI_*` env var.
 - **Observability is local**: `kigi-log` (unified session log, `--debug`
   firehose, subsystem file logs, opt-in instrumentation) writes under
   `~/.kigi` only. Its zero-network property is a contract.
+- **Atomic file replace goes through `util::fs::replace_file`** (tmp+rename
+  commit step; async callers wrap in `spawn_blocking`). Never inline a bare
+  `fs::rename` replace: Windows `MoveFileExW(REPLACE_EXISTING)` fails with a
+  sharing violation while AV/indexer/cloud-sync holds the destination open —
+  the "persists on macOS, silently doesn't on Windows" class (a /model
+  switch that never stuck). Plain rename stays correct only for true moves
+  whose destination doesn't pre-exist (worktree-pool markers, corrupt-file
+  backups). Write failures must at least `warn!` — never `let _ =`.
 - The root `Cargo.toml` is hand-maintained (upstream's generator is not in
   this repo). Members sorted; versions inherited from
-  `workspace.package.version` (0.1.0).
+  `workspace.package.version` — the single source of truth for the release
+  version (`kigi_version::VERSION` derives from it; the release workflow
+  gates the `v*` tag against it).
 
 ## Layout
 
@@ -76,8 +89,9 @@ node as one ordinary goal — the agentic loop lives INSIDE the node; the
 edges stay deterministic Rust. The harness appends a terminal
 `gn-final` verification node depending on every planner node.
 
-- Feature flag `KIGI_GRAPH=1` (default off); availability additionally
-  requires the goal harness (`BuiltinGate::Graph`).
+- Enabled by default (`KIGI_GRAPH=0` is the off-switch; the G0 gray
+  release is over); availability additionally requires the goal harness
+  (`BuiltinGate::Graph`).
 - Key modules (kigi-shell): `session/graph_tracker.rs` (pure state
   machine; reuses `GoalStatus`/`GoalPhase`/`GoalPauseReason`),
   `session/graph_plan.rs` (planner-JSON contract + validation + fnv id
@@ -150,6 +164,248 @@ edges stay deterministic Rust. The harness appends a terminal
   the replan cap; `{"ops": []}` is a respected free no-op; failures
   degrade.
 
+## Provider registry & API-key auth (post-0.1.3 expansion)
+
+- The platform registry is compiled-in spec rows in `kigi-models`
+  (`PlatformSpec`; adding a platform = enum variant + `ALL` entry + `spec()`
+  arm + row; registry tests enforce completeness/uniqueness/row shape).
+- API-key resolution precedence, per platform: platform env var(s) >
+  `auth.json` scope named by the platform id (`moonshot-cn`, …) >
+  legacy `[platforms.<id>]` in config.toml (read-only fallback).
+- The TUI login picker persists pasted keys to `auth.json` (platform-id
+  scope, `api_key` mode) — never to config.toml. The keyring holds ONLY the
+  OAuth session scope; platform keys are file-only.
+- Auth method ids over ACP equal the platform ids; interactive picker rows
+  are built generically from advertised methods (`AuthMethodKind::
+  ApiKeyPlatform`), so new registry rows appear in the picker with no TUI
+  changes.
+- Every OAuth picker row carries ITS OWN method id (`PendingMenuItem::Login
+  { method_id }` → `Action::LoginWith`); an unknown id fails closed. The
+  id-less `Action::Login` (auto-login, 401 re-auth) resolves the first
+  interactive method. `/login` opens the picker (`Action::OpenLoginPicker`),
+  never a flow directly; mid-session its last row is Cancel, not Quit.
+- `_meta.connected` on an advertised method is DISPLAY state (green badge on
+  the picker): stamped at `initialize()` from stored credentials
+  (`connected_method_ids` + `stamp_connected_meta`), kept fresh TUI-side
+  after in-session logins (`auth_in_flight_method` → `AuthComplete`). It is
+  never an authorization input.
+- Refreshable-OAuth providers beyond Kimi Code use a GENERIC path, NOT Kimi's
+  bespoke wire. A `uses_oauth` platform carrying `oauth: Some(&OAuthConfig)`
+  (client id / auth host / start+token paths / `token_host` / `scope` /
+  `scope_key` / optional extra device field / `flow` / `token_body`) drives a
+  scope-keyed `AuthManager::new_oauth_provider` +
+  `refresh::GenericDeviceRefresher` (selected by `build_refresher` via
+  `oauth_config_for_scope_key`; the refresher dispatches the refresh body by
+  `token_body`: form → `auth::oauth_device`, JSON → `auth::oauth_pkce`,
+  `GithubCopilotExchange` → `auth::github_copilot` copilot-token re-mint). Kimi
+  Code keeps `oauth: None` and its bespoke path unchanged. The interactive
+  login is dispatched by `OAuthConfig.flow` (in `run_oauth_provider_flow`):
+  - `OAuthFlow::DeviceCode` → `auth::oauth_device` (RFC-8628 device-code, plain
+    kigi UA, no X-Msh headers). Provider: `xai-grok` (`scope_key oauth/xai`,
+    base `api.x.ai/v1`, form token body, same wire as the API-key `xai` row).
+  - `OAuthFlow::PkceLocalhost { redirect_port, redirect_path }` →
+    `auth::oauth_pkce` (authorization-code + PKCE S256,
+    `127.0.0.1:redirect_port{redirect_path}` loopback with STRICT `state`
+    validation + manual-paste fallback, authorize host possibly ≠ token host).
+    The `token_body` selects the login dialect: `Json` = claude (`state ==
+    verifier`, JSON exchange carrying `state`), `Form` = codex (fresh-random
+    `state`, FORM exchange without `state` via `exchange_code_form`).
+    `OAuthConfig.authorize_extra` appends provider-only authorize params (empty
+    for all but codex). Providers:
+    - `claude-pro-max` (`scope_key oauth/claude-pro-max`, port 53692
+      `/callback`, JSON body, base `api.anthropic.com/v1`, Anthropic Messages +
+      listing wire reached with an OAuth `sk-ant-oat…` Bearer). Its Messages
+      requests take the OAuth adaptation — `anthropic-beta claude-code-…,oauth-…`
+      + `claude-cli` UA + `x-app cli` + the required "You are Claude Code…"
+      system prefix — gated on `SamplerConfig.anthropic_oauth` (claude-pro-max
+      only), so API-key `anthropic`/`minimax` Messages requests stay
+      byte-identical. Its `/v1/models` listing rides the same Bearer +
+      oauth-beta headers. THINKING REPLAY (`prune_replayed_thinking`,
+      all Messages requests): Anthropic validates every replayed
+      `thinking` block (signature model-bound, non-empty required), so
+      only the final assistant message's signed thinking is replayed and
+      only while its tool loop is open (request ends on the tool
+      results); everything else — unsigned cross-backend history, `tco_*`
+      Responses blobs, stale-model blocks — is stripped, or the request
+      400s "Invalid `signature` in `thinking` block".
+    - `openai-codex` (ChatGPT Plus/Pro, `scope_key oauth/openai-codex`, port
+      1455 `/auth/callback`, FORM body, authorize+token host `auth.openai.com`,
+      client `app_EMoam…`, scope `openid profile email offline_access`, the 3
+      authorize-extra params `id_token_add_organizations`/
+      `codex_cli_simplified_flow`/`originator=codex_cli_rs`). Refresh is a plain
+      `refresh_token` FORM grant (the generic refresher's `Form` path →
+      `auth::oauth_device`). The minted `access_token` is a JWT; login FAILS
+      FAST unless it carries the `["https://api.openai.com/auth"]
+      ["chatgpt_account_id"]` claim (`chatgpt_account_id_from_jwt`) — that
+      account id is NOT persisted but re-derived STATELESSLY from the current
+      bearer at every request. INFERENCE reuses the EXISTING Responses wire
+      against base `chatgpt.com/backend-api/codex` (→ `{base}/responses`) with
+      a Codex-gated adaptation (`SamplerConfig.openai_codex` /
+      `PlatformId::sends_codex_responses_headers()`): headers
+      `chatgpt-account-id` (per-request from the JWT), `originator codex_cli_rs`,
+      `OpenAI-Beta responses=experimental`, a codex `User-Agent`; `store:false`
+      is the shared Responses default. BODY adaptation
+      (`adapt_body_for_codex_backend`, same gate): the backend 400s
+      `role:system` input ("System messages are not allowed") — system items
+      are hoisted into the top-level `instructions` field — and stateless
+      reasoning replay requires `include:["reasoning.encrypted_content"]`.
+      API-key `openai` Responses requests carry NONE of this
+      (byte-identical, pinned by a control wire test). `reasoning.effort`
+      carries the thinking level (incl. the codex-only `ultra`). NO
+      websocket, NO base_instructions.
+      CATALOG is HARDCODED (`PlatformId::hardcoded_catalog` →
+      `openai_codex_wire_models`, mapped through the SAME
+      `platform_wire_model_to_entry` output): exactly the 4 `visibility=list` &&
+      `supported_in_api=true` models (`gpt-5.6-sol/terra/luna`, `gpt-5.5`, ctx
+      272000, per-model efforts) — NO live `/models` fetch, NO codex-CLI /
+      `~/.codex` dependency; `gpt-5.3-codex-spark` (api=false) and
+      `gpt-5.4`/`gpt-5.4-mini`/`codex-auto-review` (hidden) are EXCLUDED.
+  - `OAuthFlow::GithubDeviceCopilot` → `auth::github_copilot` (TWO-STAGE).
+    Provider: `github-copilot` (`scope_key oauth/github-copilot`, base
+    `api.individual.githubcopilot.com`, ChatCompletions wire). Stage 1 is an
+    RFC-8628 device flow on `github.com` (client `Iv1.b507a08c87ecfe98`, scope
+    `read:user`) whose errors ride a `200` body (not `4xx`) — it mints the
+    DURABLE github token. Stage 2 (`GET api.github.com/copilot_internal/v2/token`
+    with `copilot_exchange` + editor headers) re-mints the SHORT-LIVED copilot
+    token. Persisted as `KimiAuth.key = copilot token`, `refresh_token = github
+    token`, `expires_at = copilot expiry`; the "refresh" is a copilot-token
+    RE-MINT (GET, not a `refresh_token` grant). Every `/models` listing AND
+    `/chat/completions` request carries the VS Code editor-identity headers
+    (`User-Agent GitHubCopilotChat/…`, `Editor-Version`, `Editor-Plugin-Version`,
+    `Copilot-Integration-Id`; `+X-GitHub-Api-Version` on `/models`, `+X-Initiator
+    user` on inference) — gated on `SamplerConfig.github_copilot` /
+    `PlatformId::sends_copilot_editor_headers()` so every other ChatCompletions
+    provider stays byte-identical. WIRE-COMPAT SCOPE: Kigi is one-wire-per-
+    platform, so the catalog is FILTERED (`parse_github_copilot_listing`) to the
+    openai-completions-served models — keep iff `model_picker_enabled` &&
+    `policy.state != "disabled"` && `tool_calls != false` AND the id is NOT a
+    `claude-(haiku|sonnet|opus)-[45]` (anthropic-messages) or `gpt-5/oswe/mai-`
+    (responses-only) model. Those excluded models need per-model wire routing
+    (deferred, documented debt), NOT included lest they fail at inference.
+    KNOWN LIMITATION: Kigi does NOT port Pi's per-model policy-acceptance step
+    (`POST {base}/models/{id}/policy {state:"enabled"}`). A kept model whose
+    Copilot policy is unconfigured can list yet `403` at inference until the user
+    enables it once in GitHub's UI — a deliberate omission (it mutates account
+    state and is unverifiable without a live Copilot account), not a silent gap.
+  These are INTERACTIVE login rows advertised right after `kimi-code`
+  (`AuthMethodKind::OAuthPlatform`, in `PlatformId::ALL` order: `xai-grok`,
+  `claude-pro-max`, `github-copilot`, `openai-codex`). The catalog fetch resolves each such platform's OWN session
+  token (`resolve_generic_oauth_tokens`, refreshed on expiry) and routes
+  `platform.oauth().is_some()` → `platform.base_url()` (kimi-code alone →
+  `proxy_url()`). Tokens/codes/verifiers are NEVER logged.
+- A stored subscription-OAuth session IS a catalog fetch source. Every
+  fetch-PLAN decision that cannot afford real token resolution — the startup
+  prefetch arming gate, `on_auth_changed`'s wipe guard
+  (`should_wipe_catalog_on_auth_change`), and cache-origin computation — uses
+  the sync presence probes `models_fetch::stored_oauth_platforms` /
+  `stored_oauth_token_stubs` (auth.json scope scan; names only, no bearers).
+  All three must see the SAME enabled-platform set the real fetch enables, or
+  a claude-pro-max-only user boots onto the bundled Kimi table with an empty
+  picker. Managed catalog entries stamp `meta.provider` (platform display
+  name) for the client's `/model` picker; the picker itself lists the fetched
+  catalog, which is by construction the credentialed providers' models.
+- INFERENCE-AUTH CHOKEPOINT (security): `auth::credential_authority::CredentialAuthority`
+  is the ONE authority answering "WHICH credential — if any — may ride this
+  request?". It holds the session's effective `EndpointsConfig` and
+  its primary `AuthManager` PRIVATELY and answers only
+  `(platform, base_url)` questions: `credential_class` → `CredentialClass::{Pooled,
+  Primary, None}` (the outer term of `auth_method::session_token_auth_gate`),
+  `manager_for` (the governing manager
+  for refresh / 401 recovery), `credential_for` (the request's `api_key`) and
+  `bearer_resolver_for` (the aux/summary/session resolver). Do NOT re-derive this
+  rule anywhere else — three rounds of leaks came from exactly that.
+  The rule: a subscription-OAuth platform rides ITS OWN pooled `AuthManager` (so
+  it keeps a live `bearer_resolver` and mid-session refresh despite a
+  non-first-party base URL) and ONLY at its own `platform.base_url()` — a
+  `[model."claude-pro-max/x"]` override keeps `info.id` but can point `base_url`
+  anywhere. `kimi-code` and a platform-less model (a bare slug / `[model.*]`
+  entry) ride the PRIMARY session, and ONLY at the session's own effective coding
+  endpoint: `EndpointsConfig::proxy_url()` (which prefers `[endpoints]
+  coding_api_base_url` from **config.toml** — what the managed-config sync writes
+  — over `KIGI_CODE_BASE_URL`), `models_base_url`, loopback, or the compiled
+  production endpoint. Never a blanket allow: BYOK is `has_own_credentials()`,
+  which probes `std::env::var` at call time, so a `[model.*]` block with an unset
+  `env_key` classifies `NotByok`. Every API-key registry platform rides NOTHING.
+  STRUCTURAL ENFORCEMENT: the authority is the only producer of
+  `SessionCredential`, an opaque type with no production constructor, and every
+  API that stamps a session bearer onto a request (`resolve_credentials`,
+  `resolve_aux_model_sampling_config`, `try_resolve_model_credentials`,
+  `resolve_chat_state_auth_type`) takes
+  `Option<&SessionCredential>` rather than `Option<&str>` — a new call site
+  cannot express the leak. `stamp_session_local_sampler_fields` likewise takes the
+  aux `bearer_resolver` explicitly instead of copying the session's and relying on
+  the caller to re-point it, and `sampler_turn::aux_bearer_resolver_for` is the
+  ONE definition of the aux/summary resolver rule (the session actor and
+  `MvpAgent::build_summary_client` both call it; a private second copy is how the
+  summary client stayed ungated after M3). NEVER HAND-CARRY A CREDENTIAL TO A
+  GUARD (C1): "may **a** session credential ride here" is `true` for
+  a subscription-OAuth platform at its own host — where the credential that may
+  ride is that platform's POOLED token, never the primary. Ask
+  `credential_for(platform, base_url)` and stamp what it returns, so the question
+  and the credential are the same object; where you cannot (a credential the
+  authority does not own), MATCH on `credential_class` rather than reach for a
+  boolean. There is deliberately no second, similarly-named predicate to pick
+  wrongly. The shared `sampling_config.api_key`
+  (the subagent baseline and the unresolved-model fallback) has exactly TWO
+  production writers, both guarded by the authority:
+  `MvpAgent::stamp_session_credential` (the `cached_token` / `kimi.com/oidc`
+  login handlers and the `new_session` / `load_session` seed), which asks
+  `credential_for`; and the `xai.api_key` handler in `acp_agent.rs`, which stamps
+  the house `KIGI_API_KEY` read from the environment — a credential the authority
+  does not own — only when `credential_class` is `Primary`, the class the house
+  key rides and the one every OAuth platform's own host is NOT.
+- MODEL→PLATFORM LOOKUP (security): `SamplingConfig::model` is the BARE routing
+  slug, and duplicate slugs across platforms are BY DESIGN — an API-key platform
+  and its subscription-OAuth twin list identical ids (`xai`/`xai-grok`,
+  `anthropic`/`claude-pro-max`, `openai`/`openai-codex`), with the API-key
+  platform FIRST in `PlatformId::ALL`. The auth layer therefore resolves the
+  platform from the catalog KEY the picker selected, held PER SESSION in
+  `SessionActor::selected_catalog_key` (seeded at spawn by
+  `agent::models::selected_catalog_key_for_spawn`, rewritten by `SetSessionModel`,
+  CLEARED by `OverrideModelName` when the rename makes it stale), via
+  `agent::models::entry_for_slug`/`platform_for_slug`.
+  NEVER `ModelsManager::current_model_id()`: that cell is process-global,
+  last-writer-wins across concurrent sessions, and Leader mode never writes it at
+  all (`agent/handlers/model_switch.rs`). The shared `MvpAgent::sampling_config`
+  is BUILT from that cell — but ONCE, at startup, and never rebuilt, while the
+  cell moves on every non-Leader switch. Its guards therefore read
+  `MvpAgent::sampling_config_platform`, the platform captured WITH the config by
+  the same `ModelsManager::sampling_config()` call, never a fresh lookup against
+  the live cell: once the two drift, re-resolving the config's bare slug falls
+  through to `resolve_catalog_key`'s `.rev()` scan, answers the API-key twin, and
+  a post-expiry `kigi login` silently leaves the EXPIRED bearer in the config
+  that seeds every subagent (H-a). REFUSE RATHER THAN GUESS (H-b):
+  when the per-session key does not name the slug, `platform_for_slug` returns
+  `None` for a slug that collides across platforms rather than trusting
+  `resolve_catalog_key`'s `.rev()` last match — which is the subscription-OAuth
+  twin, so the guess hands an API-key session the pooled bearer that REPLACES its
+  own key on the wire. `None` then routes purely by the ENDPOINT, which for an
+  OAuth host means no credential, no resolver and no adaptation.
+  Anything else (aux models, subagent
+  overrides) falls back to the picker's own `resolve_catalog_key`, and
+  `config::find_model_by_id`'s slug scan takes the LAST match so the two can
+  never disagree. Resolving the wrong twin costs the OAuth platform its live
+  `bearer_resolver` (unrecoverable 401 ~1h in), its Messages adaptation and its
+  Copilot/Codex identity headers — and hands the API-key twin's session a pooled
+  OAuth bearer stamped over the user's own key.
+- CATALOG VISIBILITY: `platform_wire_model_to_entry` stamps
+  `supported_in_api = platform != KimiCode`. `ModelInfo::visible_for_auth`
+  reads only the PRIMARY manager's auth mode, so gating the other OAuth
+  platforms on it would hide every model from a user who signed in with ONLY a
+  Claude Pro/Max, ChatGPT, Copilot, or Grok subscription. Only `kimi-code`
+  rides the primary session, so only it may be gated on it.
+- Model metadata (context window, thinking levels) comes from the provider
+  wire when served; metadata-poor listings are enriched from models.dev
+  (`kigi-models/src/enrichment.rs` — bundled raw snapshot regenerated by
+  `scripts/gen_enrichment_snapshot.py`, single Rust transform
+  `parse_api_json` for bundled + runtime refresh; 24h cache
+  `~/.kigi/models_dev_cache.json`). Wire values always win; enrichment
+  never invents model availability. Canonical reasoning efforts:
+  none/minimal/low/medium/high/xhigh/max/ultra (`max` split from `xhigh`
+  2026-07; `ultra` is codex-only, above `max`, surfaced only via a model's
+  server-declared effort menu; Kimi wire spells its top tier `max`, kimi_compat
+  renames).
+
 ## Milestones (PRD §8.3)
 
 - M0 (done): rename, deletions (voice/telemetry/announcements/marketplace/

Cargo.lock

@@ -5442,7 +5442,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-acp-lib"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "async-trait",
@@ -5456,7 +5456,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-agent"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -5486,7 +5486,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-agent-lifecycle"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "async-trait",
  "tokio",
@@ -5495,7 +5495,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-auth"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "async-trait",
  "http 1.4.2",
@@ -5508,7 +5508,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-bin"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "clap",
@@ -5543,7 +5543,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-chat-state"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "indexmap",
  "kigi-compaction",
@@ -5560,7 +5560,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-codebase-graph"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "ahash",
  "clap",
@@ -5596,7 +5596,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-compaction"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -5609,7 +5609,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-config"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "base64",
  "blake3",
@@ -5632,7 +5632,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-config-types"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "indexmap",
@@ -5646,7 +5646,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-crash-handler"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "backtrace",
  "libc",
@@ -5657,7 +5657,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-env"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "tracing",
  "url",
@@ -5665,7 +5665,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-fast-worktree"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "bytes",
@@ -5697,7 +5697,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-file-utils"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "aws-config",
@@ -5721,7 +5721,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-fsnotify"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "criterion",
  "dunce",
@@ -5742,7 +5742,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-gix-status"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "gix",
  "kigi-test-utils",
@@ -5752,7 +5752,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hooks"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "fastrand",
  "kigi-config",
@@ -5771,7 +5771,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hooks-plugins-types"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "serde",
  "serde_json",
@@ -5779,7 +5779,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-http"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "kigi-auth",
  "kigi-log",
@@ -5794,7 +5794,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hunk-tracker"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "chrono",
  "dunce",
@@ -5815,14 +5815,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-interjection-core"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "serde",
 ]
 
 [[package]]
 name = "kigi-log"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "chrono",
@@ -5840,7 +5840,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-markdown"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anstyle",
  "anstyle-lossy",
@@ -5864,14 +5864,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-markdown-core"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "pulldown-cmark",
 ]
 
 [[package]]
 name = "kigi-mcp"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "async-trait",
@@ -5908,7 +5908,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-memory"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "arc-swap",
@@ -5942,7 +5942,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-mermaid"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "fontdb",
  "image",
@@ -5960,16 +5960,17 @@ dependencies = [
 
 [[package]]
 name = "kigi-models"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "kigi-env",
  "serde",
  "serde_json",
+ "tracing",
 ]
 
 [[package]]
 name = "kigi-pager-minimal"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "chrono",
  "crossterm",
@@ -5986,7 +5987,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-pager-pty-harness"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "alacritty_terminal",
  "anyhow",
@@ -6011,7 +6012,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-pager-render"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "anstyle",
@@ -6063,7 +6064,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-paths"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "camino",
  "serde",
@@ -6073,7 +6074,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-prompt-queue"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "serde",
  "serde_json",
@@ -6081,7 +6082,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-proto-build"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "pbjson-build",
@@ -6092,7 +6093,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-ratatui-inline"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "ansi-width",
  "anstyle-parse 0.2.7",
@@ -6109,7 +6110,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-ratatui-textarea"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "arboard",
  "chrono",
@@ -6130,7 +6131,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sampler"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "async-openai",
  "async-stream",
@@ -6153,10 +6154,11 @@ dependencies = [
 
 [[package]]
 name = "kigi-sampling-types"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "assert_matches",
  "async-openai",
+ "base64",
  "indexmap",
  "kigi-compaction",
  "kigi-tools",
@@ -6169,7 +6171,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sandbox"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "chrono",
@@ -6190,7 +6192,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-secrets"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "regex",
  "serde_json",
@@ -6228,7 +6230,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-shell"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6365,7 +6367,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-shell-base"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "chrono",
@@ -6390,7 +6392,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sqlite-journal"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "libc",
  "rusqlite",
@@ -6401,7 +6403,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-subagent-resolution"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "kigi-sampling-types",
  "kigi-tool-types",
@@ -6416,7 +6418,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-system-power"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "windows-sys 0.59.0",
  "zbus",
@@ -6424,7 +6426,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-test-support"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6446,7 +6448,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-test-utils"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "runfiles",
  "tracing",
@@ -6455,11 +6457,11 @@ dependencies = [
 
 [[package]]
 name = "kigi-token-estimation"
-version = "0.1.3"
+version = "0.1.6"
 
 [[package]]
 name = "kigi-tool-protocol"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "kigi-tool-types",
  "serde",
@@ -6470,7 +6472,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tool-runtime"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -6488,7 +6490,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tool-types"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "minijinja",
  "schemars 1.2.1",
@@ -6498,7 +6500,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tools"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "arc-swap",
@@ -6575,7 +6577,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tools-api"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "kigi-proto-build",
  "kigi-tool-protocol",
@@ -6588,11 +6590,11 @@ dependencies = [
 
 [[package]]
 name = "kigi-tracing-macros"
-version = "0.1.3"
+version = "0.1.6"
 
 [[package]]
 name = "kigi-tty-utils"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "libc",
  "nix 0.30.1",
@@ -6602,7 +6604,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tui"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "ansi-to-tui",
@@ -6689,7 +6691,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-update"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "dunce",
@@ -6718,14 +6720,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-version"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "semver",
 ]
 
 [[package]]
 name = "kigi-workspace"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6804,7 +6806,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-workspace-types"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "base64",
  "chrono",
@@ -8838,7 +8840,7 @@ dependencies = [
 
 [[package]]
 name = "ptyctl"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "alacritty_terminal",
  "anyhow",
@@ -8856,7 +8858,7 @@ dependencies = [
 
 [[package]]
 name = "ptyctl-cli"
-version = "0.1.3"
+version = "0.1.6"
 dependencies = [
  "anyhow",
  "axum",

Cargo.toml

@@ -76,7 +76,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.1.3"
+version = "0.1.6"
 edition = "2024"
 license = "Apache-2.0"
 

NOTICE

@@ -19,3 +19,8 @@ Modifications relative to the upstream work (Apache License 2.0 §4(b) notice):
 This product includes code ported from openai/codex and sst/opencode; see
 crates/codegen/kigi-tools/THIRD_PARTY_NOTICES.md for the license terms and
 the per-file change notices.
+
+This product bundles model metadata derived from models.dev
+(https://github.com/sst/models.dev, MIT License) as
+crates/codegen/kigi-models/enrichment_snapshot.json, and may refresh it
+from https://models.dev/api.json at runtime.

README.md

@@ -2,17 +2,23 @@
 
 <h1>Kigi (<code>kigi</code>) 🌘</h1>
 
-**Kigi** is an unofficial Kimi Code CLI community build — a terminal-based
-AI coding agent re-targeted at the Kimi Code subscription API and the
-Moonshot open platform, built on the Apache-2.0 sources of
-[xai-org/grok-build](https://github.com/xai-org/grok-build).
+<h3>🕸️ The world's first CLI with built-in <em>Graph Engineering</em></h3>
 
-It runs as a full-screen TUI that understands your codebase, edits files,
-executes shell commands, searches the web, and manages long-running tasks —
-interactively, headlessly for scripting/CI, or embedded in editors via the
-Agent Client Protocol (ACP).
+<p><code>/graph</code> turns one objective into a dependency graph of
+autonomous, self-verifying agent loops — planned, parallelized,
+adversarially verified, and merged back, end to end.</p>
+
+**Kigi** is a coding agent that lives in your terminal. It reads the repo,
+writes the patch, runs the tests, and keeps going while you do something else.
+Full-screen, headless in CI with `-p`, or docked in your editor over ACP.
+
+**Already paying for Claude Pro/Max, ChatGPT Plus/Pro, GitHub Copilot, or
+Grok? Sign in and use it.** No API key, no second bill. Rather bring your own
+key? OpenAI, Anthropic, Google, DeepSeek, Groq, Moonshot and
+[two dozen more](#providers-and-api-keys) are wired in.
 
 [Installation](#installation) ·
+[Graph engineering](#graph-engineering) ·
 [Providers and API keys](#providers-and-api-keys) ·
 [Building from source](#building-from-source) ·
 [Coexistence with the official CLI](#coexistence-with-the-official-kimi-cli) ·
@@ -28,10 +34,6 @@ Agent Client Protocol (ACP).
 
 ## Installation
 
-Prebuilt single-file binaries for macOS (arm64/x86_64), Linux (arm64/x86_64),
-and Windows (x86_64) are published on
-[GitHub Releases](https://github.com/ZacharyZhang-NY/Kigi-CLI/releases):
-
 ```sh
 # macOS / Linux
 curl -fsSL https://raw.githubusercontent.com/ZacharyZhang-NY/Kigi-CLI/main/install.sh | bash
@@ -43,61 +45,115 @@ irm https://raw.githubusercontent.com/ZacharyZhang-NY/Kigi-CLI/main/install.ps1
 ```
 
 ```sh
-kigi --version   # kigi 0.1.1 … unofficial Kimi Code CLI community build
-kigi login       # sign in with your Kimi Code subscription (device-code flow)
-kigi             # start the TUI
+kigi login   # pick a provider, sign in
+kigi         # go
 ```
 
-The installer verifies every download against the release's `SHA256SUMS`,
-installs into `~/.kigi/bin/kigi` (`%USERPROFILE%\.kigi\bin\kigi.exe` on
-Windows), and prints the PATH line to add. Later releases arrive through the
-built-in self-updater (`kigi update`, gated by `KIGI_AUTO_UPDATE`), which
-pulls from the same GitHub Releases feed.
+Single file, no runtime. macOS and Linux on arm64/x86_64, Windows on x86_64,
+checksummed against the release's `SHA256SUMS`. `kigi update` handles upgrades.
+
+## Graph engineering
+
+Every other agent runs a loop: think, act, repeat — one thread, one thing at a
+time. `/graph` runs a dependency graph instead. Work that doesn't block other
+work happens at the same time, in separate worktrees, and nothing merges until
+something else has tried to tear it apart.
+
+```
+/graph <objective> [--budget <tokens>]   # decompose + run fully autonomously
+/graph status                            # node tree, budget, current work
+/graph show                              # box-drawing DAG view
+/graph pause | resume [--budget <n>]     # halt / continue (budget top-up)
+/graph clear                             # abandon the graph
+```
+
+One command runs the whole thing, start to finish:
+
+- A planner breaks your objective into a dependency DAG, then validates it.
+- Independent nodes fan out as parallel workers, each in its own git worktree.
+- Every node has to get past an adversarial verifier before it merges back.
+- Find something out of scope? Say `DISCOVERED:` and the graph replans —
+  append-only, so nothing already agreed on gets rewritten.
+- Between passes, a topology optimizer drops dependencies that were never real.
+- A final node re-checks the *whole* objective before the graph is allowed to
+  call itself done.
+
+State lives in `.kigi/graph.jsonl`, next to your code. Close the laptop, come
+back tomorrow, `/graph resume`. A teammate can pick it up from the same file.
+
+On by default. `KIGI_GRAPH=0` turns it off; `KIGI_GRAPH_CONCURRENCY` (default
+3) controls how many nodes run at once.
 
 ## Providers and API keys
 
-Kigi talks to a fixed three-platform registry:
+29 platforms ship compiled in: 5 you sign into, 24 you hand a key. Nothing is
+registered at runtime — if it's not in this list, it's not there.
 
-| Platform id   | Base URL                         | Auth                        |
-| ------------- | -------------------------------- | --------------------------- |
-| `kimi-code`   | `https://api.kimi.com/coding/v1` | Kimi Code subscription OAuth (`kigi login`) |
-| `moonshot-cn` | `https://api.moonshot.cn/v1`     | Moonshot open-platform API key |
-| `moonshot-ai` | `https://api.moonshot.ai/v1`     | Moonshot open-platform API key |
+**Sign in with a subscription you already pay for.** Run `kigi login` and pick.
+Each provider's token is stored under its own key, and one provider's
+credentials are never sent to another.
 
-Moonshot API keys come from the environment or `~/.kigi/config.toml`
-(environment wins; values are never logged):
+| Platform id      | Provider                  | Sign-in                                     |
+| ---------------- | ------------------------- | ------------------------------------------- |
+| `kimi-code`      | Kimi Code (original target)| Subscription OAuth (device code)           |
+| `claude-pro-max` | Claude Pro/Max            | Subscription OAuth (browser, PKCE)          |
+| `openai-codex`   | ChatGPT Plus/Pro (Codex)  | Subscription OAuth (browser, PKCE)          |
+| `github-copilot` | GitHub Copilot            | Subscription OAuth (device code)            |
+| `xai-grok`       | xAI Grok                  | Subscription OAuth (device code)            |
+
+You get whatever models your plan actually serves — the list is fetched at
+sign-in, not hardcoded. (ChatGPT/Codex is the exception: its backend publishes
+no model endpoint, so those four are compiled in.)
+
+**API-key providers.** Export the env var, or drop the key in
+`~/.kigi/config.toml`. Keys are never logged.
+
+| Provider                  | Platform id            | API key env                                     |
+| ------------------------- | ---------------------- | ----------------------------------------------- |
+| Moonshot (moonshot.cn)    | `moonshot-cn`          | `KIGI_MOONSHOT_CN_API_KEY` (or `KIGI_MOONSHOT_API_KEY`) |
+| Moonshot (moonshot.ai)    | `moonshot-ai`          | `KIGI_MOONSHOT_AI_API_KEY` (or `KIGI_MOONSHOT_API_KEY`) |
+| OpenAI                    | `openai`               | `OPENAI_API_KEY`                                |
+| Anthropic                 | `anthropic`            | `ANTHROPIC_API_KEY`                             |
+| DeepSeek                  | `deepseek`             | `DEEPSEEK_API_KEY`                              |
+| Groq                      | `groq`                 | `GROQ_API_KEY`                                  |
+| Mistral                   | `mistral`              | `MISTRAL_API_KEY`                               |
+| Fireworks AI              | `fireworks`            | `FIREWORKS_API_KEY`                             |
+| Google Gemini             | `google`               | `GEMINI_API_KEY`                                |
+| OpenRouter                | `openrouter`           | `OPENROUTER_API_KEY`                            |
+| Together AI               | `together`             | `TOGETHER_API_KEY`                              |
+| Cerebras                  | `cerebras`             | `CEREBRAS_API_KEY`                              |
+| NVIDIA NIM                | `nvidia`               | `NVIDIA_API_KEY`                                |
+| Vercel AI Gateway         | `vercel-ai-gateway`    | `AI_GATEWAY_API_KEY`                            |
+| xAI (Grok)                | `xai`                  | `XAI_API_KEY`                                   |
+| Qwen Token Plan           | `qwen-token-plan`      | `QWEN_TOKEN_PLAN_API_KEY`                       |
+| Qwen Token Plan (China)   | `qwen-token-plan-cn`   | `QWEN_TOKEN_PLAN_CN_API_KEY`                    |
+| Kimi For Coding           | `kimi-coding`          | `KIMI_API_KEY`                                  |
+| Z.AI                      | `zai`                  | `ZAI_API_KEY`                                   |
+| Z.AI Coding (China)       | `zai-coding-cn`        | `ZAI_CODING_CN_API_KEY`                         |
+| Xiaomi MiMo               | `xiaomi`               | `XIAOMI_API_KEY`                                |
+| Xiaomi Token Plan (China) | `xiaomi-token-plan-cn` | `XIAOMI_TOKEN_PLAN_CN_API_KEY`                  |
+| MiniMax                   | `minimax`              | `MINIMAX_API_KEY`                               |
+| MiniMax (China)           | `minimax-cn`           | `MINIMAX_CN_API_KEY`                            |
 
 ```sh
-export KIGI_MOONSHOT_API_KEY=sk-...     # applies to both open platforms
-export KIGI_MOONSHOT_CN_API_KEY=sk-...  # platform-scoped, beats the generic name
-export KIGI_MOONSHOT_AI_API_KEY=sk-...
+export OPENAI_API_KEY=sk-...
+export XAI_API_KEY=xai-...
 ```
 
 ```toml
 # ~/.kigi/config.toml
-[platforms.moonshot-cn]
+[platforms.openai]
 api_key = "sk-..."
 
-[platforms.moonshot-ai]
-api_key = "sk-..."
+[platforms.xai]
+api_key = "xai-..."
 ```
 
-On login and on startup Kigi syncs each configured platform's model list
-from `GET {base}/models` and shows the merged catalog in the model picker
-(catalog keys are `{platform_id}/{model_id}`). Models that advertise
-selectable thinking levels (e.g. K3's `low`/`high`/`max`) expose them in
-`/model` and `/effort`. If the sync fails, the last cached catalog is used;
-with no cache, a small built-in fallback list applies. Model selection
-resolves as `--model` CLI flag > `KIGI_DEFAULT_MODEL` > `[models] default`
-in config.toml > server-delivered list > built-in fallback.
+Model lists sync on startup. Pick one with `/model`, set its thinking level
+with `/effort`.
 
-`KIGI_CODE_BASE_URL` re-points the subscription platform (useful for
-testing); `KIGI_MOONSHOT_CN_BASE_URL` / `KIGI_MOONSHOT_AI_BASE_URL` are the
-equivalent dev/test overrides for the open platforms.
-
-The web `search`/`fetch` tools ride the Kimi Code subscription services and
-are present only on OAuth sessions — API-key-only sessions run without
-them, matching the official client.
+Web `search`/`fetch` need a Kimi Code subscription; API-key sessions run
+without them, same as the official client.
 
 ## Building from source
 
@@ -113,19 +169,15 @@ launcher at `bin/protoc`; install dotslash (`brew install dotslash` or
 
 ## Coexistence with the official Kimi CLI
 
-Kigi is not affiliated with Moonshot AI or xAI, and it coexists with the
-official `kimi` CLI on the same machine: independent binary name,
-independent config directory (`~/.kigi`), independent keyring credentials
-(service `kigi`), and a `KIGI_*` environment-variable namespace. Nothing
-the official client installs or stores is ever read at runtime or written.
-On first launch Kigi offers a **one-time, strictly read-only** import of
-your existing `~/.kimi` configuration (MCP servers, custom providers,
-default model) via `kigi import-kimi` — file contents and mtimes under
-`~/.kimi` are left untouched, verified by tests.
+Kigi started as an unofficial Kimi Code CLI — a community fork of
+[xai-org/grok-build](https://github.com/xai-org/grok-build), not affiliated
+with Moonshot AI or xAI. It keeps its own binary, its own `~/.kigi`, its own
+keyring entry, and its own `KIGI_*` env vars, and never touches what the
+official `kimi` CLI installed. `kigi import-kimi` copies your old config over
+once, read-only.
 
-Kigi is **zero-telemetry**: the only outbound connections are the
-inference/auth APIs you configure, GitHub Releases for updates, and MCP
-servers you add.
+**Zero telemetry.** It talks to the APIs you configured, GitHub Releases, and
+your own MCP servers. Nothing else.
 
 ## License
 

crates/codegen/kigi-chat-state/src/actor/state.rs

@@ -272,6 +272,7 @@ mod tests {
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             extra_headers: Default::default(),
             context_window: std::num::NonZeroU64::new(128_000).unwrap(),
             reasoning_effort: None,

crates/codegen/kigi-chat-state/src/actor/tests.rs

@@ -23,6 +23,7 @@ fn test_config_with_window(context_window: u64) -> SamplingConfig {
         temperature: None,
         top_p: None,
         api_backend: Default::default(),
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(context_window)
             .expect("test context_window must be non-zero"),
@@ -913,6 +914,7 @@ async fn update_sampling_config_is_queryable() {
         temperature: Some(0.5),
         top_p: None,
         api_backend: Default::default(),
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(200_000).unwrap(),
         reasoning_effort: None,
@@ -1298,6 +1300,7 @@ async fn build_request_uses_sampling_config() {
         temperature: Some(0.7),
         top_p: Some(0.9),
         api_backend: Default::default(),
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(128_000).unwrap(),
         reasoning_effort: None,
@@ -3401,6 +3404,7 @@ async fn sampling_config_survives_compaction_replacement() {
         temperature: Some(0.7),
         top_p: Some(0.95),
         api_backend: ApiBackend::Responses,
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(500_000).unwrap(),
         reasoning_effort: None,
@@ -3481,6 +3485,7 @@ async fn model_metadata_lost_after_compaction_then_recovered_on_next_turn() {
         temperature: Some(0.7),
         top_p: Some(0.95),
         api_backend: Default::default(),
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(500_000).unwrap(),
         reasoning_effort: None,
@@ -3569,6 +3574,7 @@ async fn context_window_downgrade_triggers_auto_compact() {
         temperature: Some(0.7),
         top_p: Some(0.95),
         api_backend: ApiBackend::Responses,
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(500_000).unwrap(),
         reasoning_effort: None,

crates/codegen/kigi-chat-state/src/commands.rs

@@ -370,6 +370,7 @@ mod tests {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 extra_headers: Default::default(),
                 context_window: std::num::NonZeroU64::new(128_000).unwrap(),
                 reasoning_effort: None,

crates/codegen/kigi-chat-state/src/types.rs

@@ -178,6 +178,7 @@ mod tests {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 extra_headers: Default::default(),
                 context_window: NonZeroU64::new(128_000).unwrap(),
                 reasoning_effort: None,
@@ -221,6 +222,7 @@ mod tests {
                 temperature: Some(0.7),
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 extra_headers: Default::default(),
                 context_window: NonZeroU64::new(128_000).unwrap(),
                 reasoning_effort: None,

crates/codegen/kigi-models/Cargo.toml

@@ -9,6 +9,7 @@ description = "Kimi platform registry, /models wire contract, capability derivat
 kigi-env = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
+tracing = { workspace = true }
 
 [lints]
 workspace = true

crates/codegen/kigi-models/src/enrichment.rs

@@ -0,0 +1,388 @@
+//! models.dev metadata enrichment (multi-provider expansion).
+//!
+//! Most provider `/models` listings return bare ids — no context window, no
+//! thinking levels. This module carries per-model metadata keyed by
+//! models.dev provider id, sourced from the bundled snapshot
+//! (`enrichment_snapshot.json`, regenerated by
+//! `scripts/gen_enrichment_snapshot.py`) or a runtime refresh fetched by the
+//! shell. It NEVER invents model availability: the live listing is the only
+//! source of which models exist — enrichment only fills metadata gaps, and
+//! wire-served values always win (see [`enrich_wire_model`]).
+
+use std::collections::BTreeMap;
+use std::sync::LazyLock;
+
+/// One model's enrichment metadata. All fields optional-by-default so the
+/// snapshot stays minimal.
+#[derive(Debug, Clone, Default, PartialEq, serde::Deserialize, serde::Serialize)]
+pub struct EnrichmentModel {
+    /// Max context window in tokens (`limit.context`).
+    #[serde(default, skip_serializing_if = "is_zero")]
+    pub context: u64,
+    /// Max output tokens (`limit.output`); fills a wire-unserved output cap
+    /// (Anthropic 400s when `max_tokens` exceeds the model's limit).
+    #[serde(default, skip_serializing_if = "is_zero")]
+    pub output: u64,
+    /// Model supports reasoning/thinking.
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    pub reasoning: bool,
+    /// Selectable effort levels (canonical tokens, e.g. ["low","high","max"]).
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub efforts: Vec<String>,
+    /// Accepts image input.
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    pub image_in: bool,
+    /// Supports tool calling (used later to filter non-agentic listings).
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    pub tool_call: bool,
+    /// Human display name.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub name: Option<String>,
+}
+
+fn is_zero(v: &u64) -> bool {
+    *v == 0
+}
+
+/// `models.dev provider id -> model id -> metadata`.
+pub type EnrichmentCatalog = BTreeMap<String, BTreeMap<String, EnrichmentModel>>;
+
+/// The bundled snapshot, embedded at compile time — RAW models.dev shape,
+/// filtered to the providers kigi references (pure-filter script:
+/// `scripts/gen_enrichment_snapshot.py`). OFFLINE FALLBACK for the runtime
+/// refresh; both parse through [`parse_api_json`] so there is exactly one
+/// field interpretation.
+pub const ENRICHMENT_SNAPSHOT_JSON: &str = include_str!("../enrichment_snapshot.json");
+
+static BUNDLED: LazyLock<EnrichmentCatalog> = LazyLock::new(|| {
+    // Baked-in JSON — a mismatch here is a developer error, not a runtime
+    // condition (same policy as default_models.json).
+    parse_api_json(ENRICHMENT_SNAPSHOT_JSON, None)
+        .expect("enrichment_snapshot.json: invalid JSON (regenerate via script)")
+});
+
+// ── Raw models.dev api.json shape (parse-only) ──────────────────────────────
+
+#[derive(serde::Deserialize)]
+struct RawProvider {
+    #[serde(default)]
+    models: BTreeMap<String, RawModel>,
+}
+
+#[derive(serde::Deserialize, Default)]
+#[serde(default)]
+struct RawModel {
+    name: Option<String>,
+    reasoning: bool,
+    reasoning_options: Vec<RawReasoningOption>,
+    limit: RawLimit,
+    modalities: RawModalities,
+    tool_call: bool,
+}
+
+#[derive(serde::Deserialize, Default)]
+#[serde(default)]
+struct RawReasoningOption {
+    r#type: String,
+    values: Vec<String>,
+}
+
+#[derive(serde::Deserialize, Default)]
+#[serde(default)]
+struct RawLimit {
+    context: u64,
+    output: u64,
+}
+
+#[derive(serde::Deserialize, Default)]
+#[serde(default)]
+struct RawModalities {
+    input: Vec<String>,
+}
+
+/// Parse a models.dev `api.json` document (full download or the bundled
+/// filtered snapshot) into the in-memory catalog. `keep`: restrict to these
+/// provider ids (runtime refresh filters the full 3MB download to the
+/// registry's providers); `None` keeps everything present.
+///
+/// Blast-radius confinement: providers are filtered on RAW keys first and
+/// typed-parsed individually — a schema drift in one of the ~70 providers
+/// kigi never keeps cannot fail the whole refresh, and a malformed KEPT
+/// provider is warn-skipped (its models fall to defaults) rather than
+/// killing the others. Only a document that isn't a JSON object errors
+/// (the caller falls back to cache/bundled — never a silently empty
+/// catalog). Keys starting with `_` (provenance stamps) are skipped.
+pub fn parse_api_json(
+    json: &str,
+    keep: Option<&std::collections::BTreeSet<&str>>,
+) -> Result<EnrichmentCatalog, serde_json::Error> {
+    let raw: BTreeMap<String, serde_json::Value> = serde_json::from_str(json)?;
+    let mut catalog = EnrichmentCatalog::new();
+    for (pid, value) in raw {
+        if pid.starts_with('_') {
+            continue;
+        }
+        if let Some(keep) = keep
+            && !keep.contains(pid.as_str())
+        {
+            continue;
+        }
+        let provider: RawProvider = match serde_json::from_value(value) {
+            Ok(p) => p,
+            Err(e) => {
+                tracing::warn!(provider = %pid, error = %e,
+                    "models.dev provider entry malformed; skipping (models fall to defaults)");
+                continue;
+            }
+        };
+        let models = provider
+            .models
+            .into_iter()
+            .map(|(mid, m)| {
+                let efforts = m
+                    .reasoning_options
+                    .into_iter()
+                    .find(|o| o.r#type == "effort" && !o.values.is_empty())
+                    .map(|o| o.values)
+                    .unwrap_or_default();
+                let meta = EnrichmentModel {
+                    context: m.limit.context,
+                    output: m.limit.output,
+                    reasoning: m.reasoning,
+                    efforts,
+                    image_in: m.modalities.input.iter().any(|s| s == "image"),
+                    tool_call: m.tool_call,
+                    name: m.name,
+                };
+                (mid, meta)
+            })
+            .collect();
+        catalog.insert(pid, models);
+    }
+    Ok(catalog)
+}
+
+/// The compiled-in enrichment catalog.
+pub fn bundled_enrichment() -> &'static EnrichmentCatalog {
+    &BUNDLED
+}
+
+/// Look up one model's metadata. `None` when the provider or model is
+/// unknown to the catalog (callers fall through to defaults).
+pub fn lookup<'a>(
+    catalog: &'a EnrichmentCatalog,
+    models_dev_id: &str,
+    model_id: &str,
+) -> Option<&'a EnrichmentModel> {
+    catalog.get(models_dev_id)?.get(model_id)
+}
+
+/// Fill metadata gaps on a live listing entry. WIRE WINS: a field the
+/// provider served is never overwritten — enrichment only supplies what the
+/// wire left absent/zero. Never changes the model id (availability stays
+/// wire-truth).
+pub fn enrich_wire_model(wire: &mut crate::WireModel, meta: &EnrichmentModel) {
+    if wire.context_length == 0 && meta.context > 0 {
+        wire.context_length = meta.context;
+    }
+    if wire.max_output_tokens == 0 && meta.output > 0 {
+        wire.max_output_tokens = meta.output;
+    }
+    if meta.reasoning {
+        wire.supports_reasoning = true;
+    }
+    if meta.image_in {
+        wire.supports_image_in = true;
+    }
+    if wire.display_name.is_none() {
+        wire.display_name = meta.name.clone();
+    }
+    if wire.think_efforts.is_none() && !meta.efforts.is_empty() {
+        wire.think_efforts = Some(crate::WireThinkEfforts {
+            support: true,
+            valid_efforts: meta.efforts.clone(),
+            // The provider's implicit default applies when the user doesn't
+            // pick a level; models.dev doesn't record one.
+            default_effort: None,
+        });
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The bundled snapshot parses and covers the providers this expansion
+    /// references; kimi entries cross-check against the live wire values the
+    /// registry already knows (guards against a corrupted regeneration).
+    #[test]
+    fn bundled_snapshot_parses_and_cross_checks_kimi() {
+        let catalog = bundled_enrichment();
+        assert!(
+            catalog.len() >= 25,
+            "snapshot lost providers: {}",
+            catalog.len()
+        );
+        let k3 = lookup(catalog, "kimi-for-coding", "k3").expect("k3 present");
+        assert_eq!(k3.context, 1_048_576, "k3 context must match the live wire");
+        assert_eq!(k3.efforts, ["low", "high", "max"]);
+        assert!(k3.reasoning);
+        let opus = lookup(catalog, "anthropic", "claude-opus-4-8").expect("opus present");
+        assert_eq!(opus.context, 1_000_000);
+        assert_eq!(opus.efforts, ["low", "medium", "high", "xhigh", "max"]);
+    }
+
+    /// Wire-served fields are never overwritten; absent fields are filled.
+    #[test]
+    fn enrich_fills_gaps_and_never_overwrites_wire() {
+        let meta = EnrichmentModel {
+            context: 400_000,
+            output: 64_000,
+            reasoning: true,
+            efforts: vec!["low".into(), "high".into()],
+            image_in: true,
+            name: Some("GPT Test".into()),
+            ..Default::default()
+        };
+        // Bare listing entry (OpenAI-style: id only).
+        let mut bare: crate::WireModel =
+            serde_json::from_value(serde_json::json!({ "id": "gpt-test" })).unwrap();
+        enrich_wire_model(&mut bare, &meta);
+        assert_eq!(bare.context_length, 400_000);
+        assert_eq!(bare.max_output_tokens, 64_000, "output cap filled");
+        assert!(bare.supports_reasoning);
+        assert!(bare.supports_image_in);
+        assert_eq!(bare.display_name.as_deref(), Some("GPT Test"));
+        let efforts = bare.think_efforts.expect("efforts filled");
+        assert!(efforts.support);
+        assert_eq!(efforts.valid_efforts, ["low", "high"]);
+        assert_eq!(efforts.default_effort, None);
+
+        // Wire-served entry: nothing may change.
+        let mut served: crate::WireModel = serde_json::from_value(serde_json::json!({
+            "id": "gpt-test",
+            "context_length": 123,
+            "max_output_tokens": 77,
+            "display_name": "Wire Name",
+            "think_efforts": { "support": true, "valid_efforts": ["max"] }
+        }))
+        .unwrap();
+        enrich_wire_model(&mut served, &meta);
+        assert_eq!(served.context_length, 123, "wire context wins");
+        assert_eq!(served.max_output_tokens, 77, "wire output cap wins");
+        assert_eq!(served.display_name.as_deref(), Some("Wire Name"));
+        assert_eq!(
+            served.think_efforts.unwrap().valid_efforts,
+            ["max"],
+            "wire efforts win"
+        );
+    }
+
+    /// The runtime refresh parses the FULL api.json (extra fields like cost/
+    /// env/doc present) and filters to the registry's provider ids.
+    #[test]
+    fn parse_api_json_filters_and_tolerates_unknown_fields() {
+        let full = serde_json::json!({
+            "openai": {
+                "id": "openai", "env": ["OPENAI_API_KEY"], "doc": "https://x",
+                "models": {
+                    "gpt-test": {
+                        "name": "GPT Test",
+                        "reasoning": true,
+                        "reasoning_options": [
+                            {"type": "effort", "values": ["low", "high", "max"]}
+                        ],
+                        "limit": {"context": 400000, "output": 128000},
+                        "modalities": {"input": ["text", "image"], "output": ["text"]},
+                        "tool_call": true,
+                        "cost": {"input": 1.25, "output": 10}
+                    }
+                }
+            },
+            "unwanted-provider": { "models": { "m": {} } }
+        })
+        .to_string();
+        let keep: std::collections::BTreeSet<&str> = ["openai"].into();
+        let catalog = parse_api_json(&full, Some(&keep)).unwrap();
+        assert!(!catalog.contains_key("unwanted-provider"));
+        let m = lookup(&catalog, "openai", "gpt-test").unwrap();
+        assert_eq!(m.context, 400_000);
+        assert_eq!(m.output, 128_000);
+        assert_eq!(m.efforts, ["low", "high", "max"]);
+        assert!(m.reasoning && m.image_in && m.tool_call);
+        assert_eq!(m.name.as_deref(), Some("GPT Test"));
+        // Malformed document errors — callers must fall back loudly, never
+        // proceed with a silently empty catalog.
+        assert!(parse_api_json("not json", None).is_err());
+    }
+
+    /// Every registry `models_dev_id` must be covered by the bundled
+    /// snapshot — a registry row added without updating the script's
+    /// TARGETS would silently diverge bundled-vs-refresh behavior.
+    #[test]
+    fn bundled_snapshot_covers_every_registry_models_dev_id() {
+        let catalog = bundled_enrichment();
+        for platform in crate::PlatformId::ALL {
+            if let Some(dev_id) = platform.models_dev_id() {
+                assert!(
+                    catalog.contains_key(dev_id),
+                    "{}: models_dev_id {dev_id:?} missing from the bundled \
+                     snapshot — add it to scripts/gen_enrichment_snapshot.py \
+                     TARGETS and regenerate",
+                    platform.as_str(),
+                );
+            }
+        }
+    }
+
+    /// Field-coverage guard against script/parser drift: for every field
+    /// `RawModel` reads, at least one bundled model must carry a non-default
+    /// value. A regeneration that dropped a MODEL_KEYS entry (e.g.
+    /// `modalities`) would zero that field across the whole snapshot and
+    /// fail here, instead of silently diverging from runtime refreshes.
+    #[test]
+    fn bundled_snapshot_carries_every_parsed_field() {
+        let all: Vec<&EnrichmentModel> = bundled_enrichment()
+            .values()
+            .flat_map(|models| models.values())
+            .collect();
+        assert!(all.iter().any(|m| m.context > 0), "no context anywhere");
+        assert!(all.iter().any(|m| m.output > 0), "no output anywhere");
+        assert!(all.iter().any(|m| m.reasoning), "no reasoning anywhere");
+        assert!(
+            all.iter().any(|m| !m.efforts.is_empty()),
+            "no efforts anywhere"
+        );
+        assert!(all.iter().any(|m| m.image_in), "no image_in anywhere");
+        assert!(all.iter().any(|m| m.tool_call), "no tool_call anywhere");
+        assert!(all.iter().any(|m| m.name.is_some()), "no names anywhere");
+    }
+
+    /// One malformed provider (schema drift on models.dev) must not kill
+    /// the whole refresh — kept siblings still parse; only a non-object
+    /// document errors.
+    #[test]
+    fn malformed_provider_is_skipped_not_fatal() {
+        let doc = serde_json::json!({
+            "good": { "models": { "m": { "limit": {"context": 7} } } },
+            "drifted": { "models": "this is not an object" },
+            "_meta": { "source": "stamp, must be ignored" }
+        })
+        .to_string();
+        let catalog = parse_api_json(&doc, None).expect("document parses");
+        assert_eq!(
+            lookup(&catalog, "good", "m").map(|m| m.context),
+            Some(7),
+            "sibling providers must survive one drifted provider"
+        );
+        assert!(!catalog.contains_key("drifted"));
+        assert!(!catalog.contains_key("_meta"));
+    }
+
+    #[test]
+    fn lookup_misses_are_none() {
+        let catalog = bundled_enrichment();
+        assert!(lookup(catalog, "no-such-provider", "x").is_none());
+        assert!(lookup(catalog, "openai", "no-such-model").is_none());
+    }
+}

crates/codegen/kigi-pager-pty-harness/src/content.rs

@@ -97,7 +97,7 @@ impl ContentController {
             // pins them to the mock so no PTY test can reach a live endpoint.
             ("KIGI_MOONSHOT_CN_BASE_URL".into(), self.url()),
             ("KIGI_MOONSHOT_AI_BASE_URL".into(), self.url()),
-            ("XAI_API_KEY".into(), "test-key-for-ci".into()),
+            ("KIGI_API_KEY".into(), "test-key-for-ci".into()),
             ("KIGI_TELEMETRY_ENABLED".into(), "false".into()),
             ("KIGI_FEEDBACK_ENABLED".into(), "false".into()),
             ("KIGI_TRACE_UPLOAD".into(), "false".into()),
@@ -287,7 +287,7 @@ mod tests {
         assert_eq!(get("KIGI_API_BASE_URL"), Some(content.url()));
         assert_eq!(get("KIGI_MOONSHOT_CN_BASE_URL"), Some(content.url()));
         assert_eq!(get("KIGI_MOONSHOT_AI_BASE_URL"), Some(content.url()));
-        assert_eq!(get("XAI_API_KEY").as_deref(), Some("test-key-for-ci"));
+        assert_eq!(get("KIGI_API_KEY").as_deref(), Some("test-key-for-ci"));
         assert_eq!(get("KIGI_TELEMETRY_ENABLED").as_deref(), Some("false"));
         assert_eq!(get("KIGI_FEEDBACK_ENABLED").as_deref(), Some("false"));
         assert_eq!(get("KIGI_TRACE_UPLOAD").as_deref(), Some("false"));

crates/codegen/kigi-pager-pty-harness/src/flows.rs

@@ -96,11 +96,12 @@ pub fn seed_fake_oauth(content: &ContentController, user: &str) {
     .expect("seed fake oauth auth.json");
 }
 
-/// [`ContentController::env_for_pager`] minus `XAI_API_KEY`, so the entry
-/// written by [`seed_fake_oauth`] is the active credential.
+/// [`ContentController::env_for_pager`] minus the house BYOK key
+/// (`KIGI_API_KEY`), so the entry written by [`seed_fake_oauth`] is the active
+/// credential.
 pub fn oauth_env_for_pager(content: &ContentController) -> Vec<(String, String)> {
     let mut env = content.env_for_pager();
-    env.retain(|(k, _)| k != "XAI_API_KEY");
+    env.retain(|(k, _)| k != "KIGI_API_KEY");
     env
 }
 

crates/codegen/kigi-sampler/src/actor/state.rs

@@ -89,6 +89,10 @@ mod tests {
             top_p: None,
             api_backend: ApiBackend::ChatCompletions,
             auth_scheme: Default::default(),
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
+            chat_compat: Default::default(),
             extra_headers: IndexMap::new(),
             context_window: 8192,
             force_http1: false,

crates/codegen/kigi-sampler/src/client.rs

@@ -41,6 +41,35 @@ pub use kigi_sampling_types::ApiBackend;
 const AGENT_PRODUCT: &str = "kigi";
 const ANTHROPIC_DEFAULT_MAX_TOKENS: u32 = 128_000;
 
+/// Prepend the required Claude-Code system block to a Messages request's
+/// `system` field (Claude Pro/Max OAuth path). Anthropic inspects the FIRST
+/// system block, so the prefix is inserted as a distinct leading `text` block
+/// while preserving any caller-supplied prompt (string or block form).
+/// Idempotent: a leading block already equal to the prefix is not re-added.
+fn prepend_claude_code_system_prefix(system: &mut Option<messages::SystemParam>) {
+    use messages::{SystemParam, TextBlock};
+    let text_block = |text: String| TextBlock {
+        r#type: "text".to_string(),
+        text,
+        cache_control: None,
+    };
+    let mut blocks = match system.take() {
+        None => Vec::new(),
+        Some(SystemParam::Text(text)) => vec![text_block(text)],
+        Some(SystemParam::Blocks(blocks)) => blocks,
+    };
+    let already_present = blocks
+        .first()
+        .is_some_and(|b| b.text == kigi_sampling_types::CLAUDE_CODE_SYSTEM_PREFIX);
+    if !already_present {
+        blocks.insert(
+            0,
+            text_block(kigi_sampling_types::CLAUDE_CODE_SYSTEM_PREFIX.to_string()),
+        );
+    }
+    *system = Some(SystemParam::Blocks(blocks));
+}
+
 /// Parse the `Retry-After` response header as delta-seconds.
 /// Our inference backends only emit integer seconds (never HTTP-date),
 /// so we only handle that form. HTTP-dates silently return `None` and
@@ -70,6 +99,9 @@ fn deserialize_response_event(data: &str) -> Result<rs::ResponseStreamEvent> {
         Err(first_err) => {
             // Try sanitizing: parse as Value, strip unknown tools, retry.
             if let Ok(mut value) = serde_json::from_str::<serde_json::Value>(data) {
+                // A `max` reasoning-effort echo is unrepresentable in the
+                // typed enum; drop it so the event parses.
+                kigi_sampling_types::normalize_effort_echo(&mut value);
                 // Strip tools that async_openai's rs::Tool can't deserialize
                 // (e.g., xAI-specific "x_search"). Instead of maintaining a
                 // hardcoded allowlist, try deserializing each tool entry —
@@ -267,8 +299,14 @@ struct ClientDefaults {
     top_p: Option<f32>,
     api_backend: ApiBackend,
     auth_scheme: AuthScheme,
+    chat_compat: kigi_sampling_types::ChatCompat,
     stream_tool_calls: bool,
     doom_loop_recovery: Option<kigi_sampling_types::DoomLoopRecoveryPolicy>,
+    /// Claude Pro/Max OAuth Messages adaptation (see [`SamplerConfig`]).
+    anthropic_oauth: bool,
+    /// ChatGPT/Codex Responses adaptation (see [`SamplerConfig`]). Gates the
+    /// per-request `chatgpt-account-id` header derived from the bearer JWT.
+    openai_codex: bool,
 }
 
 // =============================================================================
@@ -368,6 +406,14 @@ impl SamplingClient {
                         )
                     })?;
                     headers.insert(HeaderName::from_static("x-api-key"), header_value);
+                    if config.api_backend == kigi_sampling_types::ApiBackend::Messages {
+                        // The real Anthropic Messages wire rejects requests
+                        // without this; compatible endpoints ignore it.
+                        headers.insert(
+                            HeaderName::from_static("anthropic-version"),
+                            HeaderValue::from_static(kigi_sampling_types::ANTHROPIC_VERSION),
+                        );
+                    }
                 }
                 AuthScheme::Bearer => {
                     let bearer = format!("Bearer {}", api_key);
@@ -386,6 +432,74 @@ impl SamplingClient {
             }
         }
 
+        // Claude Pro/Max OAuth identity headers (claude-pro-max only). The
+        // OAuth `sk-ant-oat…` bearer is Claude-Code-scoped, so Anthropic
+        // rejects the Messages request without the oauth beta + claude-cli
+        // identity. Gated on `anthropic_oauth` so API-key anthropic/minimax
+        // requests carry none of this and stay byte-identical. `Accept` is set
+        // per-request (text/event-stream for streams), so it is NOT added here.
+        if config.anthropic_oauth {
+            headers.insert(
+                HeaderName::from_static("anthropic-version"),
+                HeaderValue::from_static(kigi_sampling_types::ANTHROPIC_VERSION),
+            );
+            headers.insert(
+                HeaderName::from_static("anthropic-beta"),
+                HeaderValue::from_static(kigi_sampling_types::ANTHROPIC_OAUTH_BETA),
+            );
+            headers.insert(
+                HeaderName::from_static("x-app"),
+                HeaderValue::from_static("cli"),
+            );
+            headers.insert(
+                HeaderName::from_static("anthropic-dangerous-direct-browser-access"),
+                HeaderValue::from_static("true"),
+            );
+        }
+
+        // GitHub Copilot editor-identity headers (github-copilot only). Copilot's
+        // proxy validates the VS Code editor identity, so the ChatCompletions
+        // request MUST carry it. `User-Agent` is set in the UA block below (it
+        // would otherwise be overwritten); here we add the other three editor
+        // headers plus `X-Initiator: user`. Gated on `github_copilot` so every
+        // other ChatCompletions provider (groq, …) stays byte-identical.
+        if config.github_copilot {
+            headers.insert(
+                HeaderName::from_static("editor-version"),
+                HeaderValue::from_static(kigi_sampling_types::COPILOT_EDITOR_VERSION),
+            );
+            headers.insert(
+                HeaderName::from_static("editor-plugin-version"),
+                HeaderValue::from_static(kigi_sampling_types::COPILOT_EDITOR_PLUGIN_VERSION),
+            );
+            headers.insert(
+                HeaderName::from_static("copilot-integration-id"),
+                HeaderValue::from_static(kigi_sampling_types::COPILOT_INTEGRATION_ID),
+            );
+            headers.insert(
+                HeaderName::from_static("x-initiator"),
+                HeaderValue::from_static(kigi_sampling_types::COPILOT_INITIATOR),
+            );
+        }
+
+        // ChatGPT/Codex Responses identity headers (openai-codex only). The
+        // Codex backend authorizes the OAuth bearer AND validates the Codex
+        // client identity. The static pieces (`originator`, `OpenAI-Beta`) ride
+        // every request; `chatgpt-account-id` is dynamic (derived per-request
+        // from the bearer JWT in `post()`), and `User-Agent` is set in the UA
+        // block below. Gated on `openai_codex` so API-key `openai` Responses
+        // requests stay byte-identical (`store: false` is the shared default).
+        if config.openai_codex {
+            headers.insert(
+                HeaderName::from_static("originator"),
+                HeaderValue::from_static(kigi_sampling_types::CODEX_ORIGINATOR),
+            );
+            headers.insert(
+                HeaderName::from_static("openai-beta"),
+                HeaderValue::from_static(kigi_sampling_types::CODEX_OPENAI_BETA),
+            );
+        }
+
         // Apply all extra headers verbatim. This is the single
         // injection point for proxy-auth headers and any other URL- or
         // environment-specific headers the session decides to set.
@@ -403,12 +517,23 @@ impl SamplingClient {
         // (PRD F3: auth is a plain bearer; kimi-cli sends only User-Agent
         // plus the OAuth device headers, src/kimi_cli/llm.py:317-323).
         {
-            let ua_string = match config.origin_client.as_ref() {
+            // Claude Pro/Max OAuth presents the claude-cli identity; GitHub
+            // Copilot presents the VS Code Copilot Chat identity; every other
+            // path keeps the kigi User-Agent.
+            let ua_string = if config.anthropic_oauth {
+                kigi_sampling_types::CLAUDE_CODE_USER_AGENT.to_string()
+            } else if config.github_copilot {
+                kigi_sampling_types::COPILOT_USER_AGENT.to_string()
+            } else if config.openai_codex {
+                kigi_sampling_types::CODEX_USER_AGENT.to_string()
+            } else {
+                match config.origin_client.as_ref() {
                     Some(origin) => user_agent_string_for(origin),
                     None => user_agent_string_for(&OriginClientInfo {
                         product: AGENT_PRODUCT.to_string(),
                         version: Some(agent_version()),
                     }),
+                }
             };
             if let Ok(v) = HeaderValue::from_str(&ua_string) {
                 headers.insert(USER_AGENT, v);
@@ -445,8 +570,11 @@ impl SamplingClient {
             top_p: config.top_p,
             api_backend: config.api_backend,
             auth_scheme: config.auth_scheme,
+            chat_compat: config.chat_compat,
             stream_tool_calls: config.stream_tool_calls,
             doom_loop_recovery: config.doom_loop_recovery,
+            anthropic_oauth: config.anthropic_oauth,
+            openai_codex: config.openai_codex,
         };
 
         Ok(Self {
@@ -486,6 +614,23 @@ impl SamplingClient {
                 }
             }
         }
+        // ChatGPT/Codex `chatgpt-account-id` (openai-codex only): derive it
+        // STATELESSLY from the bearer that will actually ride this request (the
+        // resolver-fresh one just set, or the construction-time bearer in
+        // `default_headers`) by decoding its JWT claim. A refreshed token still
+        // carries the claim, so there is no persisted account-id field.
+        // openai-codex-gated → API-key `openai` never gets this header.
+        // SECURITY: the bearer and account id are never logged here.
+        if self.defaults.openai_codex
+            && let Some(bearer) = headers
+                .get(AUTHORIZATION)
+                .and_then(|v| v.to_str().ok())
+                .and_then(|s| s.strip_prefix("Bearer "))
+            && let Some(account_id) = kigi_sampling_types::chatgpt_account_id_from_jwt(bearer)
+            && let Ok(v) = HeaderValue::from_str(&account_id)
+        {
+            headers.insert(HeaderName::from_static("chatgpt-account-id"), v);
+        }
         {
             let auth_prefix = headers
                 .get(AUTHORIZATION)
@@ -595,6 +740,9 @@ impl SamplingClient {
             || lower.contains("apikey")
             || lower.contains("token")
             || lower.contains("secret")
+            // `chatgpt-account-id` (Codex) and any other account identifier: a
+            // stable per-user id that must never reach a log.
+            || lower.contains("account-id")
     }
 
     /// Format a single header for error messages, redacting sensitive values.
@@ -764,7 +912,10 @@ impl SamplingClient {
             tracing::error!("Failed to serialize chat/completions request: {}", e);
             SamplingError::Serialization(e)
         })?;
-        crate::kimi_compat::adapt_chat_completions_body(&mut request_body);
+        crate::kimi_compat::adapt_chat_completions_body_for(
+            self.defaults.chat_compat,
+            &mut request_body,
+        );
 
         let http_request = self
             .post(self.endpoint("chat/completions"))
@@ -818,7 +969,10 @@ impl SamplingClient {
             tracing::error!("Failed to serialize chat/completions request: {}", e);
             SamplingError::Serialization(e)
         })?;
-        crate::kimi_compat::adapt_chat_completions_body(&mut request_body);
+        crate::kimi_compat::adapt_chat_completions_body_for(
+            self.defaults.chat_compat,
+            &mut request_body,
+        );
 
         let http_request = self
             .post(self.endpoint("chat/completions"))
@@ -1028,6 +1182,10 @@ impl SamplingClient {
         // it in post-serialize. This is the last surviving piece of the
         // old raw_output machinery.
         kigi_sampling_types::patch_reasoning_text_types(&mut request_body);
+        kigi_sampling_types::patch_reasoning_effort(&mut request_body, request.reasoning_effort);
+        if self.defaults.openai_codex {
+            kigi_sampling_types::adapt_body_for_codex_backend(&mut request_body);
+        }
         let http_request = self.post(self.endpoint("responses")).json(&request_body);
 
         let response = http_request.send().await.map_err(|e| {
@@ -1074,7 +1232,16 @@ impl SamplingClient {
             });
         }
 
-        let response_obj = serde_json::from_slice::<rs::Response>(&bytes).map_err(|e| {
+        let mut response_value =
+            serde_json::from_slice::<serde_json::Value>(&bytes).map_err(|e| {
+                let raw_body = String::from_utf8_lossy(&bytes);
+                tracing::error!(error = %e, raw_body = %raw_body, "Response body is not JSON");
+                SamplingError::Serialization(e)
+            })?;
+        // A `max` effort echo is unrepresentable in the typed enum — drop it
+        // rather than failing the whole response.
+        kigi_sampling_types::normalize_effort_echo(&mut response_value);
+        let response_obj = serde_json::from_value::<rs::Response>(response_value).map_err(|e| {
             let raw_body = String::from_utf8_lossy(&bytes);
             tracing::error!(
                 error = %e,
@@ -1156,6 +1323,10 @@ impl SamplingClient {
             }
         }
         kigi_sampling_types::patch_reasoning_text_types(&mut request_body);
+        kigi_sampling_types::patch_reasoning_effort(&mut request_body, request.reasoning_effort);
+        if self.defaults.openai_codex {
+            kigi_sampling_types::adapt_body_for_codex_backend(&mut request_body);
+        }
         // Fresh per attempt so signals never leak across retries; `None`
         // (check disabled) sends no header and does no peek work per event.
         let doom_loop = self
@@ -1314,6 +1485,15 @@ impl SamplingClient {
 
     /// Apply default configuration to a Messages API request.
     fn apply_message_defaults(&self, request: &mut MessagesRequestWrapper) -> Result<()> {
+        // Claude Pro/Max OAuth adaptation (claude-pro-max only): the OAuth
+        // token is Claude-Code-scoped, so the request MUST lead with the exact
+        // "You are Claude Code…" system block or Anthropic rejects it. Prepend
+        // it as a distinct first system block, preserving any caller prompt.
+        // Gated on `anthropic_oauth` so API-key anthropic/minimax are untouched.
+        if self.defaults.anthropic_oauth {
+            prepend_claude_code_system_prefix(&mut request.inner.system);
+        }
+
         // Apply model default if not specified
         if request.inner.model.is_empty() {
             request.inner.model = self.defaults.model.clone();
@@ -1687,6 +1867,7 @@ impl SamplingClient {
         let responses_request: rs::CreateResponse = (&request).into();
 
         let mut wrapper = CreateResponseWrapper::new(responses_request);
+        wrapper.reasoning_effort = request.reasoning_effort;
         wrapper.x_kigi_conv_id = x_kigi_conv_id;
         wrapper.x_kigi_req_id = x_kigi_req_id;
         wrapper.x_kigi_session_id = x_kigi_session_id;
@@ -1720,6 +1901,7 @@ impl SamplingClient {
         let responses_request: rs::CreateResponse = (&request).into();
 
         let mut wrapper = CreateResponseWrapper::new(responses_request);
+        wrapper.reasoning_effort = request.reasoning_effort;
         wrapper.x_kigi_conv_id = x_kigi_conv_id;
         wrapper.x_kigi_req_id = x_kigi_req_id;
         wrapper.x_kigi_session_id = x_kigi_session_id;
@@ -1856,6 +2038,10 @@ mod tests {
             top_p: None,
             api_backend: ApiBackend::ChatCompletions,
             auth_scheme: AuthScheme::Bearer,
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
+            chat_compat: Default::default(),
             extra_headers: IndexMap::new(),
             context_window: 8192,
             force_http1: false,
@@ -1874,6 +2060,337 @@ mod tests {
         }
     }
 
+    /// The real Anthropic Messages wire rejects requests without
+    /// `anthropic-version`; the XApiKey+Messages client must carry it in its
+    /// default headers, and Bearer/ChatCompletions clients must NOT.
+    #[test]
+    fn x_api_key_messages_client_sends_anthropic_version() {
+        let mut config = minimal_config();
+        config.auth_scheme = AuthScheme::XApiKey;
+        config.api_backend = ApiBackend::Messages;
+        let client = SamplingClient::new(config).expect("client builds");
+        assert_eq!(
+            client
+                .default_headers
+                .get("anthropic-version")
+                .and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::ANTHROPIC_VERSION)
+        );
+        assert!(client.default_headers.get("x-api-key").is_some());
+
+        let bearer = SamplingClient::new(minimal_config()).expect("client builds");
+        assert!(
+            bearer.default_headers.get("anthropic-version").is_none(),
+            "non-anthropic clients must not grow the header"
+        );
+        let mut x_api_chat = minimal_config();
+        x_api_chat.auth_scheme = AuthScheme::XApiKey;
+        let x_api_chat = SamplingClient::new(x_api_chat).expect("client builds");
+        assert!(
+            x_api_chat
+                .default_headers
+                .get("anthropic-version")
+                .is_none(),
+            "XApiKey without the Messages backend must not grow the header"
+        );
+    }
+
+    /// Claude Pro/Max OAuth Messages client (`anthropic_oauth = true`, Bearer,
+    /// Messages) carries the full OAuth identity: Bearer auth, anthropic-version,
+    /// the oauth `anthropic-beta`, `x-app: cli`, the claude-cli User-Agent, and
+    /// the direct-browser-access header.
+    #[test]
+    fn anthropic_oauth_messages_client_sends_oauth_identity_headers() {
+        let mut config = minimal_config();
+        config.api_key = Some("sk-ant-oat-secret".to_string());
+        config.auth_scheme = AuthScheme::Bearer;
+        config.api_backend = ApiBackend::Messages;
+        config.anthropic_oauth = true;
+        let client = SamplingClient::new(config).expect("client builds");
+        let h = &client.default_headers;
+        assert_eq!(
+            h.get(AUTHORIZATION).and_then(|v| v.to_str().ok()),
+            Some("Bearer sk-ant-oat-secret"),
+            "OAuth path rides Authorization: Bearer, never x-api-key"
+        );
+        assert!(
+            h.get("x-api-key").is_none(),
+            "OAuth path must not send x-api-key"
+        );
+        assert_eq!(
+            h.get("anthropic-version").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::ANTHROPIC_VERSION)
+        );
+        assert_eq!(
+            h.get("anthropic-beta").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::ANTHROPIC_OAUTH_BETA)
+        );
+        assert_eq!(h.get("x-app").and_then(|v| v.to_str().ok()), Some("cli"));
+        assert_eq!(
+            h.get(USER_AGENT).and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::CLAUDE_CODE_USER_AGENT)
+        );
+        assert_eq!(
+            h.get("anthropic-dangerous-direct-browser-access")
+                .and_then(|v| v.to_str().ok()),
+            Some("true")
+        );
+    }
+
+    /// REGRESSION: an API-key Anthropic Messages client (XApiKey, NOT oauth)
+    /// carries NONE of the OAuth identity — no anthropic-beta, no x-app, and
+    /// the kigi User-Agent — so the API-key path stays byte-identical.
+    #[test]
+    fn api_key_anthropic_messages_client_has_no_oauth_identity() {
+        let mut config = minimal_config();
+        config.auth_scheme = AuthScheme::XApiKey;
+        config.api_backend = ApiBackend::Messages;
+        // anthropic_oauth stays false.
+        let client = SamplingClient::new(config).expect("client builds");
+        let h = &client.default_headers;
+        assert!(
+            h.get("anthropic-beta").is_none(),
+            "API-key anthropic must NOT send the oauth beta"
+        );
+        assert!(
+            h.get("x-app").is_none(),
+            "API-key anthropic must NOT send x-app"
+        );
+        assert!(h.get("anthropic-dangerous-direct-browser-access").is_none());
+        assert!(
+            h.get(USER_AGENT)
+                .and_then(|v| v.to_str().ok())
+                .is_some_and(|ua| ua.starts_with("kigi/")),
+            "API-key anthropic keeps the kigi User-Agent"
+        );
+    }
+
+    /// GitHub Copilot ChatCompletions client (`github_copilot = true`) carries
+    /// the VS Code Copilot editor-identity headers + `X-Initiator: user` and
+    /// presents the Copilot User-Agent (overriding the kigi UA).
+    #[test]
+    fn github_copilot_client_sends_editor_identity_headers() {
+        let mut config = minimal_config();
+        config.github_copilot = true;
+        let client = SamplingClient::new(config).expect("client builds");
+        let h = &client.default_headers;
+        assert_eq!(
+            h.get(USER_AGENT).and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::COPILOT_USER_AGENT),
+            "Copilot presents the VS Code Copilot User-Agent"
+        );
+        assert_eq!(
+            h.get("editor-version").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::COPILOT_EDITOR_VERSION)
+        );
+        assert_eq!(
+            h.get("editor-plugin-version").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::COPILOT_EDITOR_PLUGIN_VERSION)
+        );
+        assert_eq!(
+            h.get("copilot-integration-id")
+                .and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::COPILOT_INTEGRATION_ID)
+        );
+        assert_eq!(
+            h.get("x-initiator").and_then(|v| v.to_str().ok()),
+            Some("user"),
+            "inference carries X-Initiator: user"
+        );
+    }
+
+    /// SECURITY REGRESSION: `chatgpt-account-id` (the Codex per-user account
+    /// identifier) must be redacted by the header loggers. It is not a token, so
+    /// the substring list has to name it explicitly — without this, one debug
+    /// request log would write the user's stable ChatGPT account id to disk.
+    #[test]
+    fn account_id_headers_are_redacted_from_logs() {
+        assert!(
+            SamplingClient::is_sensitive_header("chatgpt-account-id"),
+            "chatgpt-account-id must be treated as sensitive"
+        );
+        assert!(
+            SamplingClient::is_sensitive_header("ChatGPT-Account-Id"),
+            "redaction is case-insensitive"
+        );
+        let rendered = SamplingClient::format_header("chatgpt-account-id", "acct-abc-123");
+        assert!(
+            rendered.contains("[REDACTED]") && !rendered.contains("acct-abc-123"),
+            "the account id value must never appear verbatim, got {rendered:?}"
+        );
+        // Sanity: an ordinary header is still shown (redaction stays targeted).
+        assert!(!SamplingClient::is_sensitive_header("content-type"));
+    }
+
+    /// REGRESSION: a plain ChatCompletions client (github-copilot OFF, standing
+    /// in for groq) carries NONE of the Copilot editor headers and keeps the
+    /// kigi User-Agent — every other ChatCompletions provider stays untouched.
+    #[test]
+    fn plain_chat_completions_client_has_no_copilot_editor_headers() {
+        // github_copilot stays false (as it is for groq and every other
+        // ChatCompletions platform).
+        let client = SamplingClient::new(minimal_config()).expect("client builds");
+        let h = &client.default_headers;
+        assert!(
+            h.get("editor-version").is_none(),
+            "groq must NOT send Editor-Version"
+        );
+        assert!(
+            h.get("editor-plugin-version").is_none(),
+            "groq must NOT send Editor-Plugin-Version"
+        );
+        assert!(
+            h.get("copilot-integration-id").is_none(),
+            "groq must NOT send Copilot-Integration-Id"
+        );
+        assert!(
+            h.get("x-initiator").is_none(),
+            "groq must NOT send X-Initiator"
+        );
+        assert!(
+            h.get(USER_AGENT)
+                .and_then(|v| v.to_str().ok())
+                .is_some_and(|ua| ua.starts_with("kigi/")),
+            "groq keeps the kigi User-Agent"
+        );
+    }
+
+    // A JWT whose payload carries `["https://api.openai.com/auth"]
+    // ["chatgpt_account_id"] = "acct-test-42"` (alg=none; signature is cosmetic
+    // — the extractor only base64url-decodes the payload segment).
+    const CODEX_TEST_JWT: &str = "eyJhbGciOiJub25lIiwidHlwIjoiSldUIn0.eyJodHRwczovL2FwaS5vcGVuYWkuY29tL2F1dGgiOnsiY2hhdGdwdF9hY2NvdW50X2lkIjoiYWNjdC10ZXN0LTQyIn0sInN1YiI6InVzZXItMSJ9.sig";
+
+    /// openai-codex Responses client (`openai_codex = true`): the STATIC codex
+    /// identity headers (`originator`, `OpenAI-Beta`, codex `User-Agent`) ride
+    /// `default_headers`, and a built `post()` request derives `chatgpt-account-
+    /// id` from the bearer JWT. The Responses body default keeps `store: false`.
+    #[test]
+    fn openai_codex_client_sends_codex_identity_headers() {
+        let mut config = minimal_config();
+        config.api_backend = ApiBackend::Responses;
+        config.openai_codex = true;
+        config.api_key = Some(CODEX_TEST_JWT.to_string());
+        let client = SamplingClient::new(config).expect("client builds");
+        let h = &client.default_headers;
+        assert_eq!(
+            h.get("originator").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::CODEX_ORIGINATOR),
+            "codex must send originator: codex_cli_rs"
+        );
+        assert_eq!(
+            h.get("openai-beta").and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::CODEX_OPENAI_BETA),
+            "codex must opt into OpenAI-Beta: responses=experimental"
+        );
+        assert_eq!(
+            h.get(USER_AGENT).and_then(|v| v.to_str().ok()),
+            Some(kigi_sampling_types::CODEX_USER_AGENT),
+            "codex presents the codex User-Agent"
+        );
+        // The account id is derived PER REQUEST from the bearer JWT in post().
+        let req = client
+            .post("https://chatgpt.com/backend-api/codex/responses")
+            .build()
+            .expect("build request");
+        assert_eq!(
+            req.headers()
+                .get("chatgpt-account-id")
+                .and_then(|v| v.to_str().ok()),
+            Some("acct-test-42"),
+            "chatgpt-account-id must be decoded from the bearer JWT claim"
+        );
+    }
+
+    /// REGRESSION: an API-key `openai` Responses client (`openai_codex = false`)
+    /// carries NONE of the Codex identity headers — not the static ones and not
+    /// the per-request `chatgpt-account-id` — so its request stays byte-identical.
+    #[test]
+    fn api_key_openai_responses_client_has_no_codex_headers() {
+        let mut config = minimal_config();
+        config.api_backend = ApiBackend::Responses;
+        // openai_codex stays false (as it is for API-key openai). Even with a
+        // JWT-shaped key, no account-id header is derived.
+        config.api_key = Some(CODEX_TEST_JWT.to_string());
+        let client = SamplingClient::new(config).expect("client builds");
+        let h = &client.default_headers;
+        assert!(
+            h.get("originator").is_none(),
+            "openai must NOT send originator"
+        );
+        assert!(
+            h.get("openai-beta").is_none(),
+            "openai must NOT send the codex OpenAI-Beta"
+        );
+        assert!(
+            h.get(USER_AGENT)
+                .and_then(|v| v.to_str().ok())
+                .is_some_and(|ua| ua.starts_with("kigi/")),
+            "API-key openai keeps the kigi User-Agent"
+        );
+        let req = client
+            .post("https://api.openai.com/v1/responses")
+            .build()
+            .expect("build request");
+        assert!(
+            req.headers().get("chatgpt-account-id").is_none(),
+            "API-key openai must NOT send chatgpt-account-id (regression)"
+        );
+    }
+
+    /// The system-prompt prefix is prepended as a distinct leading `text`
+    /// block for each `system` shape (absent / string / blocks), preserving the
+    /// caller's prompt, and is idempotent (not stamped twice).
+    #[test]
+    fn claude_code_system_prefix_prepends_and_is_idempotent() {
+        use messages::{SystemParam, TextBlock};
+        let prefix = kigi_sampling_types::CLAUDE_CODE_SYSTEM_PREFIX;
+
+        // Absent system → a single prefix block.
+        let mut none = None;
+        prepend_claude_code_system_prefix(&mut none);
+        match none {
+            Some(SystemParam::Blocks(b)) => {
+                assert_eq!(b.len(), 1);
+                assert_eq!(b[0].text, prefix);
+            }
+            other => panic!("expected one prefix block, got {other:?}"),
+        }
+
+        // String system → [prefix, original].
+        let mut text = Some(SystemParam::Text("do the thing".into()));
+        prepend_claude_code_system_prefix(&mut text);
+        match text {
+            Some(SystemParam::Blocks(b)) => {
+                assert_eq!(b.len(), 2);
+                assert_eq!(b[0].text, prefix);
+                assert_eq!(b[1].text, "do the thing");
+            }
+            other => panic!("expected two blocks, got {other:?}"),
+        }
+
+        // Idempotent: a leading prefix block is not re-added.
+        let mut already = Some(SystemParam::Blocks(vec![
+            TextBlock {
+                r#type: "text".into(),
+                text: prefix.to_string(),
+                cache_control: None,
+            },
+            TextBlock {
+                r#type: "text".into(),
+                text: "tail".into(),
+                cache_control: None,
+            },
+        ]));
+        prepend_claude_code_system_prefix(&mut already);
+        match already {
+            Some(SystemParam::Blocks(b)) => {
+                assert_eq!(b.len(), 2, "prefix must not be duplicated");
+                assert_eq!(b[0].text, prefix);
+            }
+            other => panic!("expected unchanged blocks, got {other:?}"),
+        }
+    }
+
     /// Verify the serialized shape of StreamingChatRequest matches the
     /// expected wire format: all ChatCompletionRequest fields flattened at
     /// top level, plus `stream: true` and `stream_options.include_usage: true`.
@@ -2017,6 +2534,7 @@ mod tests {
             api_key: Some("bearer-key-abc123".to_string()),
             api_backend: ApiBackend::Messages,
             auth_scheme: AuthScheme::Bearer,
+            chat_compat: Default::default(),
             ..minimal_config()
         };
         let client = SamplingClient::new(cfg).expect("client should build");
@@ -2183,6 +2701,7 @@ mod tests {
             api_key: Some("stale-bearer".to_string()),
             api_backend: ApiBackend::Messages,
             auth_scheme: AuthScheme::Bearer,
+            chat_compat: Default::default(),
             bearer_resolver: Some(std::sync::Arc::new(StaticBearerResolver("fresh-bearer"))),
             ..minimal_config()
         };
@@ -2211,6 +2730,7 @@ mod tests {
             api_key: Some("stale-bearer".to_string()),
             api_backend: ApiBackend::Responses,
             auth_scheme: AuthScheme::Bearer,
+            chat_compat: Default::default(),
             bearer_resolver: Some(std::sync::Arc::new(StaticBearerResolver("fresh-bearer"))),
             ..minimal_config()
         };

crates/codegen/kigi-sampler/src/config.rs

@@ -33,7 +33,7 @@ pub enum AuthScheme {
 ///
 /// `SamplerConfig` is the single source of truth for sampler
 /// configuration. The shell builds it directly (see
-/// `agent::config::resolve_model_to_sampling_config` and
+/// `agent::config::sampling_config_for_model` and
 /// `session::acp_session::SessionActor::reconstruct_full_config`) by
 /// composing chat-state's `kigi_sampling_types::SamplingConfig`
 /// with `Credentials` (api key, client version).
@@ -55,6 +55,29 @@ pub struct SamplerConfig {
     pub api_backend: ApiBackend,
     #[serde(default)]
     pub auth_scheme: AuthScheme,
+    /// Claude Pro/Max OAuth adaptation (claude-pro-max only). When true the
+    /// Messages request carries the OAuth identity headers (`anthropic-beta`
+    /// oauth, `claude-cli` User-Agent, `x-app: cli`) and its system prompt is
+    /// prefixed with the required "You are Claude Code…" line. Gated so the
+    /// API-key `anthropic` + `minimax` Messages requests stay byte-identical.
+    #[serde(default)]
+    pub anthropic_oauth: bool,
+    /// GitHub Copilot ChatCompletions adaptation (github-copilot only). When
+    /// true the request carries the VS Code Copilot editor-identity headers
+    /// (User-Agent `GitHubCopilotChat/…`, `Editor-Version`,
+    /// `Editor-Plugin-Version`, `Copilot-Integration-Id`) plus `X-Initiator:
+    /// user`. Gated so every other ChatCompletions provider (groq, …) stays
+    /// byte-identical.
+    #[serde(default)]
+    pub github_copilot: bool,
+    /// ChatGPT/Codex Responses adaptation (openai-codex only). When true the
+    /// `/codex/responses` request carries the Codex identity headers
+    /// (`chatgpt-account-id` derived per-request from the bearer JWT,
+    /// `originator: codex_cli_rs`, `OpenAI-Beta: responses=experimental`, a codex
+    /// `User-Agent`). Gated so the API-key `openai` Responses requests stay
+    /// byte-identical (`store: false` is already the shared Responses default).
+    #[serde(default)]
+    pub openai_codex: bool,
     /// Extra request headers applied verbatim. The sampler never inspects
     /// the URL to derive headers; callers (the session) inject proxy auth
     /// and other access headers here before constructing the config.
@@ -70,6 +93,11 @@ pub struct SamplerConfig {
 
     // Reasoning effort
     pub reasoning_effort: Option<ReasoningEffort>,
+    /// ChatCompletions body-adaptation dialect (per-platform; BYOK/custom
+    /// endpoints default to the historical Kimi behavior; lenient default on
+    /// deserialize so persisted configs from before the field parse).
+    #[serde(default)]
+    pub chat_compat: kigi_sampling_types::ChatCompat,
 
     /// Client identity for the User-Agent header (`kigi/{version}` plus an
     /// optional origin product). The old xAI proxy's identity headers
@@ -135,9 +163,13 @@ impl Default for SamplerConfig {
             model: String::new(),
             max_completion_tokens: None,
             temperature: None,
+            chat_compat: kigi_sampling_types::ChatCompat::default(),
             top_p: None,
             api_backend: ApiBackend::default(),
             auth_scheme: AuthScheme::default(),
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
             extra_headers: IndexMap::new(),
             context_window: 0,
             force_http1: false,

crates/codegen/kigi-sampler/src/kimi_compat.rs

@@ -29,6 +29,89 @@ pub(crate) fn adapt_chat_completions_body(body: &mut Value) {
     adapt_tool_schemas(body);
 }
 
+/// Dialect-dispatched body adaptation. Kimi keeps the full historical
+/// pipeline (thinking + message hygiene + schema normalization — all built
+/// for the Kimi wire's strictness); DeepSeek differs ONLY in how thinking
+/// rides the body; Passthrough providers take OpenAI-style bodies verbatim
+/// (their `reasoning_effort` scalar is already the wire form).
+pub(crate) fn adapt_chat_completions_body_for(
+    compat: kigi_sampling_types::ChatCompat,
+    body: &mut Value,
+) {
+    match compat {
+        kigi_sampling_types::ChatCompat::Kimi => adapt_chat_completions_body(body),
+        kigi_sampling_types::ChatCompat::DeepSeek => {
+            adapt_thinking_deepseek(body);
+            strip_kigi_private_message_fields(body);
+        }
+        kigi_sampling_types::ChatCompat::Passthrough => {
+            strip_kigi_private_message_fields(body);
+        }
+        kigi_sampling_types::ChatCompat::StrictOpenAi => {
+            strip_kigi_private_message_fields(body);
+            strip_stream_options(body);
+        }
+    }
+}
+
+/// Mistral's strict Pydantic validator 422-rejects `stream_options`
+/// (`extra_forbidden` on `stream_options.include_usage`; its request model
+/// has no such field). kigi injects `stream_options.include_usage` on every
+/// streaming request for the other providers, so strip the whole object for
+/// Mistral. Streaming usage falls back to token estimation (as for any
+/// provider that omits streaming usage).
+fn strip_stream_options(body: &mut Value) {
+    if let Some(obj) = body.as_object_mut() {
+        obj.remove("stream_options");
+    }
+}
+
+/// Remove kigi-internal history artifacts from input messages before they
+/// reach a non-Kimi wire. `reasoning_content` is Kimi's replayed-thinking
+/// field (Kimi consumes it; DeepSeek documents it as prefix-mode-only and
+/// historically 400s on it; other providers don't know it) and `model_id`
+/// is kigi's private per-message provenance. Kimi's own pipeline handles
+/// these in `adapt_messages`.
+fn strip_kigi_private_message_fields(body: &mut Value) {
+    let Some(messages) = body.get_mut("messages").and_then(|m| m.as_array_mut()) else {
+        return;
+    };
+    for message in messages {
+        if let Some(obj) = message.as_object_mut() {
+            obj.remove("reasoning_content");
+            obj.remove("model_id");
+        }
+    }
+}
+
+/// DeepSeek spells the thinking control `thinking:{type, reasoning_effort}`
+/// (api-docs.deepseek.com create-chat-completion; the server maps
+/// low/medium→high and xhigh→max itself, so the canonical level passes
+/// through verbatim). `none` disables thinking; absent leaves the server
+/// default (enabled).
+fn adapt_thinking_deepseek(body: &mut Value) {
+    let Some(obj) = body.as_object_mut() else {
+        return;
+    };
+    let Some(effort) = obj.remove("reasoning_effort") else {
+        return;
+    };
+    let Some(level) = effort.as_str().map(str::to_owned) else {
+        return;
+    };
+    if level == "none" {
+        obj.insert(
+            "thinking".to_string(),
+            serde_json::json!({ "type": "disabled" }),
+        );
+    } else {
+        obj.insert(
+            "thinking".to_string(),
+            serde_json::json!({ "type": "enabled", "reasoning_effort": level }),
+        );
+    }
+}
+
 /// Map the OpenAI-style `reasoning_effort` knob onto Kimi's `thinking`
 /// request field and drop `reasoning_effort` from the wire.
 ///
@@ -280,6 +363,123 @@ mod tests {
     use super::*;
     use serde_json::json;
 
+    #[test]
+    fn deepseek_dialect_spells_thinking_reasoning_effort() {
+        use kigi_sampling_types::ChatCompat;
+        // Official docs: thinking:{type, reasoning_effort}; server maps
+        // low/medium→high, xhigh→max itself — levels pass through verbatim.
+        let mut body = json!({ "model": "deepseek-v4-pro", "reasoning_effort": "high" });
+        adapt_chat_completions_body_for(ChatCompat::DeepSeek, &mut body);
+        assert_eq!(body.get("reasoning_effort"), None);
+        assert_eq!(
+            body["thinking"],
+            json!({ "type": "enabled", "reasoning_effort": "high" })
+        );
+
+        let mut body = json!({ "model": "deepseek-v4-flash", "reasoning_effort": "max" });
+        adapt_chat_completions_body_for(ChatCompat::DeepSeek, &mut body);
+        assert_eq!(
+            body["thinking"],
+            json!({ "type": "enabled", "reasoning_effort": "max" })
+        );
+
+        // none disables; absent leaves the server default (no thinking key).
+        let mut body = json!({ "reasoning_effort": "none" });
+        adapt_chat_completions_body_for(ChatCompat::DeepSeek, &mut body);
+        assert_eq!(body["thinking"], json!({ "type": "disabled" }));
+        let mut body = json!({ "model": "deepseek-chat" });
+        adapt_chat_completions_body_for(ChatCompat::DeepSeek, &mut body);
+        assert_eq!(body.get("thinking"), None);
+
+        // DeepSeek does NOT get kimi's message/tool-schema rewrites (empty
+        // assistant tool-call content survives — DeepSeek's documented
+        // function-calling round-trip uses that shape), but kigi-private
+        // fields are stripped: replayed reasoning_content is prefix-mode-only
+        // on the DeepSeek wire (historically a 400 in input messages).
+        let mut body = json!({
+            "reasoning_effort": "high",
+            "messages": [
+                { "role": "assistant", "content": "", "tool_calls": [{}],
+                  "reasoning_content": "replayed thinking", "model_id": "kigi/x" }
+            ]
+        });
+        adapt_chat_completions_body_for(ChatCompat::DeepSeek, &mut body);
+        assert_eq!(body["messages"][0]["content"], json!(""));
+        assert_eq!(body["messages"][0].get("reasoning_content"), None);
+        assert_eq!(body["messages"][0].get("model_id"), None);
+    }
+
+    #[test]
+    fn strict_openai_dialect_strips_stream_options_and_private_fields() {
+        use kigi_sampling_types::ChatCompat;
+        // Mistral 422s on stream_options (extra_forbidden) and doesn't know
+        // kigi's private message fields; OpenAI-style reasoning_effort stays.
+        let mut body = json!({
+            "model": "mistral-medium-latest",
+            "reasoning_effort": "high",
+            "stream": true,
+            "stream_options": { "include_usage": true },
+            "messages": [
+                { "role": "assistant", "content": "hi",
+                  "reasoning_content": "internal", "model_id": "kigi/x" }
+            ]
+        });
+        adapt_chat_completions_body_for(ChatCompat::StrictOpenAi, &mut body);
+        assert_eq!(
+            body.get("stream_options"),
+            None,
+            "stream_options must be stripped"
+        );
+        assert_eq!(body["stream"], json!(true), "stream flag stays");
+        assert_eq!(
+            body["reasoning_effort"],
+            json!("high"),
+            "OpenAI-style effort passes through (Mistral accepts it natively)"
+        );
+        assert_eq!(body["messages"][0].get("reasoning_content"), None);
+        assert_eq!(body["messages"][0].get("model_id"), None);
+        assert_eq!(body["messages"][0]["content"], json!("hi"));
+    }
+
+    #[test]
+    fn passthrough_dialect_leaves_openai_body_verbatim() {
+        use kigi_sampling_types::ChatCompat;
+        // Verbatim EXCEPT kigi-private history artifacts, which no non-Kimi
+        // wire understands.
+        let mut body = json!({
+            "model": "gpt-oss",
+            "reasoning_effort": "high",
+            "messages": [
+                { "role": "user", "content": "hi" },
+                { "role": "assistant", "content": "yo",
+                  "reasoning_content": "internal", "model_id": "kigi/x" }
+            ]
+        });
+        adapt_chat_completions_body_for(ChatCompat::Passthrough, &mut body);
+        assert_eq!(
+            body,
+            json!({
+                "model": "gpt-oss",
+                "reasoning_effort": "high",
+                "messages": [
+                    { "role": "user", "content": "hi" },
+                    { "role": "assistant", "content": "yo" }
+                ]
+            }),
+            "reasoning_effort stays OpenAI-style; private fields are stripped"
+        );
+    }
+
+    #[test]
+    fn kimi_dialect_dispatch_matches_legacy_pipeline() {
+        use kigi_sampling_types::ChatCompat;
+        let mut via_dispatch = json!({ "model": "k3", "reasoning_effort": "max" });
+        adapt_chat_completions_body_for(ChatCompat::Kimi, &mut via_dispatch);
+        let mut via_legacy = json!({ "model": "k3", "reasoning_effort": "max" });
+        adapt_chat_completions_body(&mut via_legacy);
+        assert_eq!(via_dispatch, via_legacy, "Kimi dispatch = legacy pipeline");
+    }
+
     #[test]
     fn reasoning_effort_maps_to_kimi_thinking_field() {
         // Level rides along as thinking.effort (live wire: 200 with
@@ -292,8 +492,9 @@ mod tests {
             json!({ "type": "enabled", "effort": "high" })
         );
 
-        // Canonical `xhigh` is spelled `max` on the Kimi wire (the K3
-        // valid_efforts vocabulary is low/high/max).
+        // Legacy canonical `xhigh` (pre-Max configs/sessions) is spelled
+        // `max` on the Kimi wire (the K3 valid_efforts vocabulary is
+        // low/high/max — there is no `xhigh` there).
         let mut body = json!({ "model": "k3", "reasoning_effort": "xhigh" });
         adapt_chat_completions_body(&mut body);
         assert_eq!(
@@ -301,6 +502,15 @@ mod tests {
             json!({ "type": "enabled", "effort": "max" })
         );
 
+        // Canonical `max` (what the K3 menu token parses to since the
+        // ReasoningEffort::Max split) passes through unchanged.
+        let mut body = json!({ "model": "k3", "reasoning_effort": "max" });
+        adapt_chat_completions_body(&mut body);
+        assert_eq!(
+            body["thinking"],
+            json!({ "type": "enabled", "effort": "max" })
+        );
+
         // kimi.py:218: "off" (our ReasoningEffort::None) → disabled, and no
         // effort key (a disabled+effort combination would be contradictory).
         let mut body = json!({ "reasoning_effort": "none" });

crates/codegen/kigi-sampler/src/stream/chat_completions.rs

@@ -517,6 +517,85 @@ mod tests {
         }
     }
 
+    /// Deserialize a full chunk from a JSON `delta` so it flows through the
+    /// `#[serde(from = "RawChatChunkDelta")]` content-split path (Mistral
+    /// sends array `content`).
+    fn chunk_from_delta(delta: serde_json::Value) -> ChatCompletionChunk {
+        serde_json::from_value(serde_json::json!({
+            "id": "c",
+            "object": "chat.completion.chunk",
+            "created": 0,
+            "model": "mistral-medium-latest",
+            "choices": [{ "index": 0, "delta": delta, "finish_reason": null }],
+        }))
+        .expect("mistral chunk deserializes")
+    }
+
+    /// End-to-end: a Mistral reasoning stream (array `content` thinking
+    /// deltas → a transition delta carrying both a thinking and a text chunk
+    /// → plain-string answer deltas) is split by deserialization and consumed
+    /// into the SAME reasoning-sibling + assistant-answer result the
+    /// `reasoning_content` string path produces. Closes the array-path
+    /// integration gap.
+    #[tokio::test]
+    async fn mistral_reasoning_array_stream_splits_into_reasoning_and_answer() {
+        let chunks: Vec<Result<ChatCompletionChunk, SamplingError>> = vec![
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking",
+                  "thinking": [{ "type": "text", "text": "Let me think. " }] }
+            ]}))),
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking",
+                  "thinking": [{ "type": "text", "text": "It's 22." }] }
+            ]}))),
+            // Transition: one array with a closing thinking chunk AND the
+            // first answer text chunk.
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking", "thinking": [{ "type": "text", "text": " Done." }] },
+                { "type": "text", "text": "Answer: " }
+            ]}))),
+            // Answer phase: plain-string deltas (no longer arrays).
+            Ok(chunk_from_delta(serde_json::json!({ "content": "22." }))),
+            Ok(final_chunk(FinishReason::Stop)),
+        ];
+        let raw = stream::iter(chunks).boxed();
+        let events = collect(stream_chat_completions(
+            raw,
+            None,
+            rid(),
+            Duration::from_secs(60),
+        ))
+        .await;
+
+        // Channel tokens: thinking rode the Reasoning channel, answer the Text
+        // channel — never crossed.
+        let mut reasoning = String::new();
+        let mut answer = String::new();
+        for e in &events {
+            if let SamplingEvent::ChannelToken { channel, text, .. } = e {
+                match channel {
+                    SamplingChannel::Reasoning => reasoning.push_str(text),
+                    SamplingChannel::Text => answer.push_str(text),
+                }
+            }
+        }
+        assert_eq!(reasoning, "Let me think. It's 22. Done.");
+        assert_eq!(answer, "Answer: 22.");
+
+        // The accumulated final response carries the same split.
+        match events.last().unwrap() {
+            SamplingEvent::Completed { response, .. } => {
+                let r = response
+                    .reasoning_items()
+                    .next()
+                    .expect("array thinking became a reasoning sibling");
+                let rs::SummaryPart::SummaryText(t) = &r.summary[0];
+                assert_eq!(t.text, "Let me think. It's 22. Done.");
+            }
+            other => panic!("expected Completed, got {other:?}"),
+        }
+    }
+
     #[tokio::test]
     async fn tool_call_stream_emits_deltas_and_assembles_final_call() {
         // First chunk has id + name + part of arguments.

crates/codegen/kigi-sampler/tests/test_actor.rs

@@ -78,6 +78,10 @@ fn test_config(base_url: String, model: &str) -> SamplerConfig {
         top_p: None,
         api_backend: ApiBackend::ChatCompletions,
         auth_scheme: Default::default(),
+        anthropic_oauth: false,
+        github_copilot: false,
+        openai_codex: false,
+        chat_compat: Default::default(),
         extra_headers: IndexMap::new(),
         context_window: 128_000,
         force_http1: false,

crates/codegen/kigi-sampling-types/Cargo.toml

@@ -7,6 +7,7 @@ description = "Pure data types for the xAI sampling / chat-completion API layer"
 
 [dependencies]
 async-openai = { workspace = true }
+base64 = { workspace = true }
 indexmap = { workspace = true, features = ["serde"] }
 reqwest = { workspace = true }
 serde = { workspace = true, features = ["derive"] }

crates/codegen/kigi-sampling-types/src/conversation.rs

@@ -250,8 +250,13 @@ pub struct AssistantItem {
     /// `response.reasoning.effort` (Responses API). Stored beside
     /// `model_id`/`model_fingerprint` so per-response effort survives
     /// mid-session model/effort switches. `None` for synthetic items and
-    /// backends that don't echo it.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
+    /// backends that don't echo it. Lenient on read: an unknown token from a
+    /// newer kigi drops to `None` rather than failing the history line.
+    #[serde(
+        default,
+        skip_serializing_if = "Option::is_none",
+        deserialize_with = "crate::types::lenient_reasoning_effort_opt"
+    )]
     pub reasoning_effort: Option<crate::ReasoningEffort>,
 }
 
@@ -2153,7 +2158,10 @@ impl From<&ConversationRequest> for rs::CreateResponse {
             prompt_cache_key: None,
             prompt_cache_retention: None,
             reasoning: Some(rs::Reasoning {
-                effort: req.reasoning_effort.map(|e| e.to_responses_api()),
+                // Effort is written onto the serialized body by
+                // `patch_reasoning_effort` (the typed enum cannot spell
+                // `max`); the wrapper carries the canonical value.
+                effort: None,
                 summary: Some(rs::ReasoningSummary::Concise),
             }),
             safety_identifier: None,
@@ -3193,6 +3201,8 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
     flush_assistant(&mut pending_assistant, &mut messages);
     flush_tool_results(&mut pending_tool_results, &mut messages);
 
+    prune_replayed_thinking(&mut messages);
+
     // Attach cache_control: {type: "ephemeral"} to last system block
     if let Some(last) = system_blocks.last_mut() {
         last.cache_control = Some(CacheControl {
@@ -3282,6 +3292,74 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
     }
 }
 
+/// Strip replayed `thinking` blocks the Anthropic Messages API would
+/// reject — keep exactly the one it requires.
+///
+/// Anthropic validates EVERY `thinking` block in the request: the
+/// signature is bound to the emitting model and must be non-empty, so
+/// history from another backend (`encrypted_content: None` replays as
+/// `signature: ""`), a Responses-API `tco_*` blob (signature bytes with no
+/// text), or a block signed by a DIFFERENT model after a mid-session
+/// `/model` switch 400s the whole request with
+/// "messages.N.content.0: Invalid `signature` in `thinking` block".
+///
+/// The API only NEEDS thinking for the ACTIVE tool-use continuation: the
+/// final assistant message whose tool_use results follow must carry its
+/// signed thinking back verbatim. Prior turns' thinking is ignored even
+/// when valid (Pi/Claude Code replay exactly this way). So: keep the
+/// final assistant message's thinking when the loop is open and the block
+/// is genuinely signed (non-empty text AND signature — an open loop can
+/// never span a model switch, so that signature is always the current
+/// model's); strip every other thinking block. An assistant message left
+/// EMPTY by the strip (a thinking-only aborted turn) is removed — the API
+/// rejects empty content arrays.
+fn prune_replayed_thinking(messages: &mut Vec<crate::messages::Message>) {
+    use crate::messages::{ContentBlock, MessageContent, MessageRole};
+    let last_assistant = messages
+        .iter()
+        .rposition(|m| matches!(m.role, MessageRole::Assistant));
+    let active_tool_loop = last_assistant.is_some_and(|i| {
+        let has_tool_use = matches!(
+            &messages[i].content,
+            MessageContent::Blocks(blocks)
+                if blocks.iter().any(|b| matches!(b, ContentBlock::ToolUse { .. }))
+        );
+        // The loop is OPEN only while the request ends on the tool results:
+        // a later plain user turn closes it (the results answered, the model
+        // replied — its thinking is history the API ignores or rejects).
+        let continuation = &messages[i + 1..];
+        let ends_on_results = !continuation.is_empty()
+            && continuation.iter().all(|m| {
+                matches!(
+                    &m.content,
+                    MessageContent::Blocks(blocks)
+                        if blocks.iter().any(|b| matches!(b, ContentBlock::ToolResult { .. }))
+                )
+            });
+        has_tool_use && ends_on_results
+    });
+    let mut index = 0;
+    messages.retain_mut(|m| {
+        let i = index;
+        index += 1;
+        if !matches!(m.role, MessageRole::Assistant) {
+            return true;
+        }
+        let MessageContent::Blocks(blocks) = &mut m.content else {
+            return true;
+        };
+        let keep_thinking = active_tool_loop && Some(i) == last_assistant;
+        blocks.retain(|b| match b {
+            ContentBlock::Thinking {
+                thinking,
+                signature,
+            } => keep_thinking && !thinking.is_empty() && !signature.is_empty(),
+            _ => true,
+        });
+        !blocks.is_empty()
+    });
+}
+
 /// Convert a MessagesResponse to a single Assistant `ConversationItem`.
 ///
 /// Note: Anthropic `Thinking` blocks are dropped here because this `From`
@@ -5076,13 +5154,35 @@ mod tests {
         }
     }
 
+    #[test]
+    fn assistant_item_unknown_reasoning_effort_token_degrades_to_none() {
+        // History lines written by a NEWER kigi (grown effort vocabulary)
+        // must not fail this binary's history parse — the lenient
+        // deserializer must be wired on the AssistantItem field itself.
+        let json = serde_json::json!({
+            "role": "assistant",
+            "content": "hi",
+            "reasoning_effort": "hypermax"
+        });
+        let item: AssistantItem = serde_json::from_value(json).expect("history line must survive");
+        assert_eq!(item.reasoning_effort, None);
+        let json = serde_json::json!({
+            "role": "assistant",
+            "content": "hi",
+            "reasoning_effort": "max"
+        });
+        let item: AssistantItem = serde_json::from_value(json).unwrap();
+        assert_eq!(item.reasoning_effort, Some(crate::ReasoningEffort::Max));
+    }
+
     #[test]
     fn test_messages_request_wire_format_for_supported_variants() {
         for (variant, expected) in [
             (crate::ReasoningEffort::Low, "low"),
             (crate::ReasoningEffort::Medium, "medium"),
             (crate::ReasoningEffort::High, "high"),
-            (crate::ReasoningEffort::Xhigh, "max"),
+            (crate::ReasoningEffort::Xhigh, "xhigh"),
+            (crate::ReasoningEffort::Max, "max"),
         ] {
             let req = messages_test_request(Some(variant));
             let msgs = build_messages_request(&req);
@@ -5131,6 +5231,7 @@ mod tests {
             (crate::ReasoningEffort::Medium, "medium"),
             (crate::ReasoningEffort::High, "high"),
             (crate::ReasoningEffort::Xhigh, "xhigh"),
+            (crate::ReasoningEffort::Max, "max"),
         ] {
             let req = ConversationRequest::from_items(vec![ConversationItem::user("hi")])
                 .with_model("test");
@@ -5163,6 +5264,10 @@ mod tests {
 
     #[test]
     fn test_responses_request_carries_reasoning_effort_nested() {
+        // Wire contract since the Max split: the typed conversion leaves
+        // reasoning.effort UNSET (async-openai's enum cannot spell `max`);
+        // `patch_reasoning_effort` writes the canonical token onto the
+        // serialized body — every level, `xhigh` and `max` distinct.
         for (variant, expected) in [
             (crate::ReasoningEffort::None, "none"),
             (crate::ReasoningEffort::Minimal, "minimal"),
@@ -5170,6 +5275,7 @@ mod tests {
             (crate::ReasoningEffort::Medium, "medium"),
             (crate::ReasoningEffort::High, "high"),
             (crate::ReasoningEffort::Xhigh, "xhigh"),
+            (crate::ReasoningEffort::Max, "max"),
         ] {
             let req = ConversationRequest {
                 reasoning_effort: Some(variant),
@@ -5177,15 +5283,56 @@ mod tests {
                     .with_model("test")
             };
             let resp: crate::rs::CreateResponse = (&req).into();
-            let json = serde_json::to_value(&resp).unwrap();
+            let mut json = serde_json::to_value(&resp).unwrap();
+            assert_eq!(
+                json.pointer("/reasoning/effort"),
+                None,
+                "typed conversion must leave effort unset ({variant:?})"
+            );
+            crate::patch_reasoning_effort(&mut json, req.reasoning_effort);
             assert_eq!(
                 json.pointer("/reasoning/effort").and_then(|v| v.as_str()),
                 Some(expected),
-                "{variant:?} should serialize as reasoning.effort={expected:?}; got: {json:#}",
+                "{variant:?} should be patched as reasoning.effort={expected:?}; got: {json:#}",
+            );
+            assert_eq!(
+                json.pointer("/reasoning/summary").and_then(|v| v.as_str()),
+                Some("concise"),
+                "the patch must not clobber reasoning.summary"
             );
         }
     }
 
+    #[test]
+    fn normalize_effort_echo_drops_only_unrepresentable_tokens() {
+        // `max` echo (bare response shape) is dropped so typed parsing
+        // succeeds; known tokens pass through; stream-event envelopes are
+        // handled via /response/reasoning.
+        let mut bare = serde_json::json!({ "reasoning": { "effort": "max", "summary": "c" } });
+        crate::normalize_effort_echo(&mut bare);
+        assert_eq!(bare.pointer("/reasoning/effort"), None);
+        assert_eq!(
+            bare.pointer("/reasoning/summary").and_then(|v| v.as_str()),
+            Some("c")
+        );
+
+        let mut known = serde_json::json!({ "reasoning": { "effort": "high" } });
+        crate::normalize_effort_echo(&mut known);
+        assert_eq!(
+            known.pointer("/reasoning/effort").and_then(|v| v.as_str()),
+            Some("high"),
+            "representable echoes must pass through"
+        );
+
+        let mut event = serde_json::json!({ "response": { "reasoning": { "effort": "max" } } });
+        crate::normalize_effort_echo(&mut event);
+        assert_eq!(event.pointer("/response/reasoning/effort"), None);
+
+        let mut absent = serde_json::json!({ "output": [] });
+        crate::normalize_effort_echo(&mut absent);
+        assert_eq!(absent, serde_json::json!({ "output": [] }));
+    }
+
     #[test]
     fn test_responses_request_omits_effort_when_unset() {
         let req =
@@ -5255,6 +5402,160 @@ mod tests {
     /// messages while setting top-level `thinking: null` — the Messages API
     /// rejects this with a 400.  Verify that stripped reasoning produces a
     /// valid request with no thinking blocks in messages.
+    /// Collect `(message_index, thinking, signature)` for every thinking
+    /// block in a built Messages request.
+    fn thinking_blocks(json: &serde_json::Value) -> Vec<(usize, String, String)> {
+        let mut out = Vec::new();
+        for (i, m) in json["messages"].as_array().unwrap().iter().enumerate() {
+            if let Some(content) = m.get("content").and_then(|c| c.as_array()) {
+                for b in content {
+                    if b.get("type").and_then(|t| t.as_str()) == Some("thinking") {
+                        out.push((
+                            i,
+                            b["thinking"].as_str().unwrap_or_default().to_string(),
+                            b["signature"].as_str().unwrap_or_default().to_string(),
+                        ));
+                    }
+                }
+            }
+        }
+        out
+    }
+
+    fn reasoning(text: &str, encrypted: Option<&str>) -> ConversationItem {
+        ConversationItem::Reasoning(rs::ReasoningItem {
+            id: String::new(),
+            summary: if text.is_empty() {
+                vec![]
+            } else {
+                vec![rs::SummaryPart::SummaryText(rs::SummaryTextContent {
+                    text: text.to_string(),
+                })]
+            },
+            content: None,
+            encrypted_content: encrypted.map(str::to_owned),
+            status: None,
+        })
+    }
+
+    fn assistant_text(text: &str) -> ConversationItem {
+        ConversationItem::Assistant(AssistantItem {
+            content: text.into(),
+            tool_calls: vec![],
+            model_id: None,
+            model_fingerprint: None,
+            reasoning_effort: None,
+        })
+    }
+
+    /// Anthropic validates EVERY replayed `thinking` block: an unsigned one
+    /// (history synthesized by another backend replays as `signature: ""`),
+    /// a Responses `tco_*` blob (signature with no text), or a block signed
+    /// by a different model after a mid-session `/model` switch 400s the
+    /// whole request — "messages.N.content.0: Invalid `signature` in
+    /// `thinking` block". The API only NEEDS thinking for the active
+    /// tool-use continuation, so outside one the builder must replay NO
+    /// thinking blocks at all.
+    #[test]
+    fn messages_request_strips_thinking_outside_active_tool_loop() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::system("sys"),
+            ConversationItem::user("q1"),
+            // Cross-backend history: unsigned reasoning (the Windows repro —
+            // session started on another model, then switched to Claude).
+            reasoning("some thinking", None),
+            assistant_text("a1"),
+            ConversationItem::user("q2"),
+            // Responses-API blob: signature-shaped bytes, no text.
+            reasoning("", Some("tco_blob")),
+            assistant_text("a2"),
+            ConversationItem::user("q3"),
+            // Genuinely signed — but its tool loop (none) is closed, so the
+            // API ignores it when valid and 400s it after a model switch.
+            reasoning("signed thinking", Some("sig-real")),
+            assistant_text("a3"),
+            ConversationItem::user("q4"),
+        ]);
+        let json = serde_json::to_value(build_messages_request(&req)).unwrap();
+        assert_eq!(
+            thinking_blocks(&json),
+            vec![],
+            "no thinking block may be replayed outside an active tool loop:\n{json:#}"
+        );
+    }
+
+    /// The active tool-use continuation is the one place Anthropic REQUIRES
+    /// the signed thinking block back: the final assistant message issued
+    /// tool_use and its results follow. Exactly that block is kept; a
+    /// prior turn's signed thinking is still stripped.
+    #[test]
+    fn messages_request_keeps_signed_thinking_for_active_tool_loop() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q0"),
+            reasoning("old turn", Some("sig-old")),
+            assistant_text("a0"),
+            ConversationItem::user("q1"),
+            reasoning("current turn", Some("sig-current")),
+            ConversationItem::Assistant(AssistantItem {
+                content: "".into(),
+                tool_calls: vec![ToolCall {
+                    id: std::sync::Arc::from("tc1"),
+                    name: "read_file".to_string(),
+                    arguments: std::sync::Arc::from("{}"),
+                }],
+                model_id: None,
+                model_fingerprint: None,
+                reasoning_effort: None,
+            }),
+            ConversationItem::tool_result("tc1", "file contents"),
+        ]);
+        let json = serde_json::to_value(build_messages_request(&req)).unwrap();
+        let blocks = thinking_blocks(&json);
+        assert_eq!(
+            blocks.len(),
+            1,
+            "exactly the active loop's thinking survives:\n{json:#}"
+        );
+        let (msg_idx, thinking, signature) = &blocks[0];
+        assert_eq!(thinking, "current turn");
+        assert_eq!(signature, "sig-current");
+        // It sits at content.0 of the final assistant message.
+        let msg = &json["messages"].as_array().unwrap()[*msg_idx];
+        assert_eq!(msg["role"], "assistant");
+        assert_eq!(msg["content"][0]["type"], "thinking");
+        assert!(
+            msg["content"]
+                .as_array()
+                .unwrap()
+                .iter()
+                .any(|b| b["type"] == "tool_use"),
+            "the kept thinking belongs to the tool_use turn"
+        );
+    }
+
+    /// A thinking-only assistant turn (aborted before any text/tool output)
+    /// must not survive as an EMPTY assistant message after the strip —
+    /// Anthropic rejects empty content arrays.
+    #[test]
+    fn messages_request_drops_assistant_message_emptied_by_thinking_strip() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q"),
+            reasoning("aborted turn thinking", Some("sig")),
+            assistant_text(""),
+            ConversationItem::user("follow-up"),
+        ]);
+        let json = serde_json::to_value(build_messages_request(&req)).unwrap();
+        for m in json["messages"].as_array().unwrap() {
+            if let Some(content) = m.get("content").and_then(|c| c.as_array()) {
+                assert!(
+                    !content.is_empty(),
+                    "no message may ship an empty content array:\n{json:#}"
+                );
+            }
+        }
+        assert_eq!(thinking_blocks(&json), vec![]);
+    }
+
     #[test]
     fn test_btw_stripped_reasoning_produces_no_thinking_blocks() {
         // Simulate a conversation where the model responded with thinking.

crates/codegen/kigi-sampling-types/src/types.rs

@@ -487,7 +487,69 @@ pub enum FinishReason {
     FunctionCall,
 }
 
+/// Wire `content` field on a chat response/delta. OpenAI-compatible providers
+/// send a plain string, but reasoning providers (Mistral) send an ARRAY of
+/// typed chunks: `{"type":"text","text":".."}` for the answer and
+/// `{"type":"thinking","thinking":[{"type":"text","text":".."}],...}` for the
+/// chain of thought (verified against the mistralai/client-python SDK models).
+/// The chunk union is OPEN — unknown chunk types are ignored, never fatal.
+#[derive(Deserialize)]
+#[serde(untagged)]
+enum WireContent {
+    Text(String),
+    Chunks(Vec<Value>),
+}
+
+/// Read a chunk's `text` string field into `dst` when present (defensive: a
+/// reference/tool-reference/unknown chunk simply has no `text`).
+fn push_chunk_text(dst: &mut String, chunk: &Value) {
+    if let Some(t) = chunk.get("text").and_then(Value::as_str) {
+        dst.push_str(t);
+    }
+}
+
+fn non_empty(s: String) -> Option<String> {
+    if s.is_empty() { None } else { Some(s) }
+}
+
+impl WireContent {
+    /// Split into `(visible answer text, thinking text)`. A plain string is
+    /// the answer with no thinking; an array routes `text` chunks to the
+    /// answer and the nested `text` of `thinking` chunks to the reasoning
+    /// channel. Byte-identical for string content (the only shape non-Mistral
+    /// providers ever send).
+    fn split(self) -> (Option<String>, Option<String>) {
+        match self {
+            // Preserve string content verbatim (incl. empty) — byte-identical
+            // to the pre-change `content: Option<String>` for every provider
+            // that sends a string.
+            WireContent::Text(s) => (Some(s), None),
+            WireContent::Chunks(chunks) => {
+                let mut answer = String::new();
+                let mut thinking = String::new();
+                for c in &chunks {
+                    match c.get("type").and_then(Value::as_str) {
+                        Some("text") => push_chunk_text(&mut answer, c),
+                        Some("thinking") => {
+                            // `thinking` is a nested list of chunks.
+                            if let Some(inner) = c.get("thinking").and_then(Value::as_array) {
+                                for ic in inner {
+                                    push_chunk_text(&mut thinking, ic);
+                                }
+                            }
+                        }
+                        // reference / tool_reference / unknown → not text.
+                        _ => {}
+                    }
+                }
+                (non_empty(answer), non_empty(thinking))
+            }
+        }
+    }
+}
+
 #[derive(Debug, Serialize, Deserialize, Clone)]
+#[serde(from = "RawChatResponseMessage")]
 pub struct ChatResponseMessage {
     pub role: Role,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -502,6 +564,38 @@ pub struct ChatResponseMessage {
     pub citations: Option<Vec<String>>,
 }
 
+/// Deserialization mirror of [`ChatResponseMessage`] that accepts array
+/// `content` and folds thinking chunks into `reasoning_content`.
+#[derive(Deserialize)]
+struct RawChatResponseMessage {
+    role: Role,
+    #[serde(default)]
+    content: Option<WireContent>,
+    #[serde(default)]
+    reasoning_content: Option<String>,
+    #[serde(default)]
+    tool_calls: Vec<ToolCallResponse>,
+    #[serde(default)]
+    tool_call_id: Option<String>,
+    #[serde(default)]
+    citations: Option<Vec<String>>,
+}
+
+impl From<RawChatResponseMessage> for ChatResponseMessage {
+    fn from(r: RawChatResponseMessage) -> Self {
+        let (answer, thinking) = r.content.map(WireContent::split).unwrap_or((None, None));
+        ChatResponseMessage {
+            role: r.role,
+            content: answer,
+            // The wire's own `reasoning_content` wins; else thinking chunks.
+            reasoning_content: r.reasoning_content.or(thinking),
+            tool_calls: r.tool_calls,
+            tool_call_id: r.tool_call_id,
+            citations: r.citations,
+        }
+    }
+}
+
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct ToolCallResponse {
     pub id: String,
@@ -649,6 +743,7 @@ pub struct ToolCallFunctionDelta {
 }
 
 #[derive(Debug, Serialize, Deserialize, Clone, Default)]
+#[serde(from = "RawChatChunkDelta")]
 pub struct ChatChunkDelta {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub role: Option<Role>,
@@ -656,16 +751,43 @@ pub struct ChatChunkDelta {
     pub content: Option<String>,
     pub reasoning_content: Option<String>,
     /// Tool call deltas. Handles `null` in JSON as empty vec.
-    #[serde(
-        default,
-        skip_serializing_if = "Vec::is_empty",
-        deserialize_with = "deserialize_null_default"
-    )]
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
     pub tool_calls: Vec<ToolCallDelta>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_call_id: Option<String>,
 }
 
+/// Deserialization mirror of [`ChatChunkDelta`] that accepts array `content`
+/// (Mistral reasoning streams a thinking-chunk array during the thinking
+/// phase, then plain-string answer deltas) and folds thinking into
+/// `reasoning_content`.
+#[derive(Deserialize)]
+struct RawChatChunkDelta {
+    #[serde(default)]
+    role: Option<Role>,
+    #[serde(default)]
+    content: Option<WireContent>,
+    #[serde(default)]
+    reasoning_content: Option<String>,
+    #[serde(default, deserialize_with = "deserialize_null_default")]
+    tool_calls: Vec<ToolCallDelta>,
+    #[serde(default)]
+    tool_call_id: Option<String>,
+}
+
+impl From<RawChatChunkDelta> for ChatChunkDelta {
+    fn from(r: RawChatChunkDelta) -> Self {
+        let (answer, thinking) = r.content.map(WireContent::split).unwrap_or((None, None));
+        ChatChunkDelta {
+            role: r.role,
+            content: answer,
+            reasoning_content: r.reasoning_content.or(thinking),
+            tool_calls: r.tool_calls,
+            tool_call_id: r.tool_call_id,
+        }
+    }
+}
+
 /// Parameters to control realtime data.
 #[derive(Serialize, Deserialize, Clone, Debug)]
 pub struct SearchParameters {
@@ -787,22 +909,23 @@ pub enum ReasoningEffort {
     Medium,
     High,
     Xhigh,
+    /// Distinct top tier above `xhigh` (OpenAI Responses and Anthropic
+    /// Messages both accept `xhigh` AND `max` as separate levels in 2026;
+    /// the Kimi wire spells its top tier `max` with no `xhigh`).
+    Max,
+    /// Codex-only top tier above `max` (the ChatGPT Codex backend exposes an
+    /// `ultra` reasoning effort on its flagship models). Reachable ONLY via a
+    /// model's server-declared effort menu (openai-codex); no built-in fallback
+    /// menu offers it, so other providers never emit it.
+    Ultra,
 }
 
 impl ReasoningEffort {
-    pub fn to_responses_api(self) -> crate::rs::ReasoningEffort {
-        match self {
-            Self::None => crate::rs::ReasoningEffort::None,
-            Self::Minimal => crate::rs::ReasoningEffort::Minimal,
-            Self::Low => crate::rs::ReasoningEffort::Low,
-            Self::Medium => crate::rs::ReasoningEffort::Medium,
-            Self::High => crate::rs::ReasoningEffort::High,
-            Self::Xhigh => crate::rs::ReasoningEffort::Xhigh,
-        }
-    }
-
-    /// Inverse of [`to_responses_api`](Self::to_responses_api): the effort the
-    /// Responses API echoes back on `response.reasoning.effort`.
+    /// The canonical effort behind a typed Responses-API echo
+    /// (`response.reasoning.effort`). `max` echoes never reach the typed
+    /// enum — [`normalize_effort_echo`] drops them pre-parse (async-openai
+    /// has no such variant); the request direction writes the wire string
+    /// via [`patch_reasoning_effort`].
     pub fn from_responses_api(effort: crate::rs::ReasoningEffort) -> Self {
         match effort {
             crate::rs::ReasoningEffort::None => Self::None,
@@ -822,17 +945,25 @@ impl ReasoningEffort {
             Self::Medium => "medium",
             Self::High => "high",
             Self::Xhigh => "xhigh",
+            Self::Max => "max",
+            Self::Ultra => "ultra",
         }
     }
 
-    /// Anthropic Messages API `output_config.effort` string; `None` for unsupported variants.
+    /// Anthropic Messages API effort string; `None` for unsupported variants.
+    /// `xhigh` and `max` are distinct levels on the 2026 Messages API (both
+    /// appear in `GET /v1/models` `capabilities.effort`). `ultra` is codex-only
+    /// and never selected on an Anthropic model, but maps to its own string for
+    /// completeness (the Responses path writes effort via `as_str`, not this).
     pub fn to_messages_api(self) -> Option<&'static str> {
         match self {
             Self::None | Self::Minimal => None,
             Self::Low => Some("low"),
             Self::Medium => Some("medium"),
             Self::High => Some("high"),
-            Self::Xhigh => Some("max"),
+            Self::Xhigh => Some("xhigh"),
+            Self::Max => Some("max"),
+            Self::Ultra => Some("ultra"),
         }
     }
 }
@@ -853,19 +984,277 @@ impl std::str::FromStr for ReasoningEffort {
             "low" => Ok(Self::Low),
             "medium" => Ok(Self::Medium),
             "high" => Ok(Self::High),
-            "xhigh" | "max" => Ok(Self::Xhigh), // max is a CLI/UX alias of xhigh
+            "xhigh" => Ok(Self::Xhigh),
+            "max" => Ok(Self::Max),
+            "ultra" => Ok(Self::Ultra),
             _ => Err(format!(
-                "invalid reasoning effort: {s:?} (expected one of: none, minimal, low, medium, high, xhigh, max)"
+                "invalid reasoning effort: {s:?} (expected one of: none, minimal, low, medium, high, xhigh, max, ultra)"
             )),
         }
     }
 }
 
-/// Canonical wire parse only (`max` → `Xhigh`); remapped menu ids need a model catalog.
+/// Canonical wire parse; remapped menu ids need a model catalog.
 pub fn parse_canonical_effort_token(token: &str) -> Option<ReasoningEffort> {
     token.parse().ok()
 }
 
+/// Deserialize an optional effort LENIENTLY for persisted stores (session
+/// summaries, chat history): a token this binary doesn't know — written by a
+/// newer kigi after the effort vocabulary grew, exactly what happened when
+/// `max` split from `xhigh` — degrades to `None` with a warning instead of
+/// failing the whole record, so listings and resume survive version
+/// rollback. Non-string values still error (real corruption stays loud), and
+/// config-TOML parsing stays strict — its layer warn-skips explicitly.
+pub fn lenient_reasoning_effort_opt<'de, D>(d: D) -> Result<Option<ReasoningEffort>, D::Error>
+where
+    D: Deserializer<'de>,
+{
+    let raw = Option::<String>::deserialize(d)?;
+    Ok(raw.and_then(|s| match s.parse::<ReasoningEffort>() {
+        Ok(effort) => Some(effort),
+        Err(error) => {
+            tracing::warn!(token = %s, %error, "persisted reasoning_effort unknown; dropping");
+            None
+        }
+    }))
+}
+
+/// Write the canonical effort onto a serialized Responses request body
+/// (`body.reasoning.effort`). This is the ONLY place effort reaches the
+/// Responses wire: the typed `rs::ReasoningEffort` tops out at `xhigh`, so
+/// `max` must be written post-serialize. A `None` effort leaves the body
+/// untouched (the provider default applies).
+pub fn patch_reasoning_effort(body: &mut Value, effort: Option<ReasoningEffort>) {
+    let Some(effort) = effort else { return };
+    let Some(obj) = body.as_object_mut() else {
+        return;
+    };
+    let reasoning = obj
+        .entry("reasoning")
+        .or_insert_with(|| Value::Object(serde_json::Map::new()));
+    if let Some(reasoning) = reasoning.as_object_mut() {
+        reasoning.insert(
+            "effort".to_string(),
+            Value::String(effort.as_str().to_string()),
+        );
+    }
+}
+
+/// Adapt a serialized Responses request body to the ChatGPT/Codex backend
+/// contract (`chatgpt.com/backend-api/codex/responses`) — ported from the
+/// same reference as the identity headers (official Codex CLI + Pi's
+/// `api/openai-codex-responses.ts`):
+///
+/// 1. The backend rejects `role: system` input items outright
+///    (400 `{"detail":"System messages are not allowed"}`); its system
+///    channel is the top-level `instructions` field. Hoist every system
+///    input message there (order preserved, blank-line joined, appended to
+///    any existing instructions) and remove them from `input`.
+/// 2. `store` is always `false` on this backend, so reasoning continuity
+///    is stateless: `include: ["reasoning.encrypted_content"]` is required
+///    for the response to carry replayable encrypted reasoning.
+///
+/// openai-codex-GATED at the call sites — API-key `openai` Responses
+/// bodies stay byte-identical.
+pub fn adapt_body_for_codex_backend(body: &mut Value) {
+    // 1. Hoist system messages into `instructions`.
+    let mut hoisted: Vec<String> = Vec::new();
+    if let Some(input) = body.get_mut("input").and_then(|v| v.as_array_mut()) {
+        input.retain(|item| {
+            let is_system = item.get("role").and_then(|r| r.as_str()) == Some("system");
+            if is_system {
+                match item.get("content") {
+                    Some(Value::String(s)) => hoisted.push(s.clone()),
+                    Some(Value::Array(parts)) => {
+                        for p in parts {
+                            if let Some(t) = p.get("text").and_then(|t| t.as_str()) {
+                                hoisted.push(t.to_string());
+                            }
+                        }
+                    }
+                    _ => {}
+                }
+            }
+            !is_system
+        });
+    }
+    if !hoisted.is_empty() {
+        let mut instructions = body
+            .get("instructions")
+            .and_then(|v| v.as_str())
+            .map(str::to_owned)
+            .unwrap_or_default();
+        for part in hoisted {
+            if !instructions.is_empty() {
+                instructions.push_str("\n\n");
+            }
+            instructions.push_str(&part);
+        }
+        body["instructions"] = Value::String(instructions);
+    }
+
+    // 2. Request replayable encrypted reasoning.
+    let include = body
+        .as_object_mut()
+        .map(|obj| obj.entry("include").or_insert_with(|| Value::Array(vec![])));
+    if let Some(Value::Array(entries)) = include {
+        let key = Value::String("reasoning.encrypted_content".to_string());
+        if !entries.contains(&key) {
+            entries.push(key);
+        }
+    }
+}
+
+/// Neutralize a `reasoning.effort` echo the typed `rs` enum cannot parse
+/// (`max`): remove it so response deserialization succeeds. The turn's
+/// canonical effort lives in the session sampling config regardless; only
+/// the per-item echo metadata is dropped, recorded as not-echoed.
+/// (Ceiling: async-openai lacks a Max variant; delete this when it grows
+/// one.) Handles both bare response bodies (`/reasoning/effort`) and
+/// stream-event envelopes (`/response/reasoning/effort`).
+pub fn normalize_effort_echo(value: &mut Value) {
+    for path in ["/reasoning", "/response/reasoning"] {
+        if let Some(reasoning) = value.pointer_mut(path).and_then(|v| v.as_object_mut())
+            && let Some(effort) = reasoning.get("effort").and_then(|v| v.as_str())
+            && crate::rs::ReasoningEffort::deserialize(serde_json::Value::String(
+                effort.to_string(),
+            ))
+            .is_err()
+        {
+            tracing::debug!(
+                effort,
+                "reasoning.effort echo unrepresentable in the typed enum; dropping"
+            );
+            reasoning.remove("effort");
+        }
+    }
+}
+
+/// The `anthropic-version` header value kigi speaks on Anthropic-style
+/// wires (Messages inference and the /v1/models listing).
+pub const ANTHROPIC_VERSION: &str = "2023-06-01";
+
+/// `anthropic-beta` value for the Claude-Code OAuth (Pro/Max subscription)
+/// path — required on BOTH the Messages inference request and the `/v1/models`
+/// listing when the bearer is an OAuth `sk-ant-oat…` token. API-key Anthropic
+/// and MiniMax NEVER send this (their requests stay byte-identical).
+pub const ANTHROPIC_OAUTH_BETA: &str = "claude-code-20250219,oauth-2025-04-20";
+
+/// User-Agent kigi presents on the Claude-Code OAuth path (mirrors the
+/// official Claude Code CLI). OAuth-gated: unrelated to the default kigi UA.
+pub const CLAUDE_CODE_USER_AGENT: &str = "claude-cli/2.1.75";
+
+/// System-prompt prefix REQUIRED on the Claude-Code OAuth Messages path: the
+/// OAuth token is Claude-Code-scoped, so Anthropic rejects the request unless
+/// the system prompt's first block is exactly this line. OAuth-gated.
+pub const CLAUDE_CODE_SYSTEM_PREFIX: &str =
+    "You are Claude Code, Anthropic's official CLI for Claude.";
+
+// ── GitHub Copilot editor-identity headers ──────────────────────────────────
+// The VS Code Copilot Chat client identity. Copilot's proxy authorizes the
+// short-lived copilot token AND validates these editor headers, so they ride
+// the `copilot_internal/v2/token` exchange, the `/models` listing, and every
+// `/chat/completions` inference request. github-copilot-GATED: no other
+// platform sends them, so their requests stay byte-identical. Values are
+// non-secret wire constants (ported from Pi `api/github-copilot-headers.ts` +
+// `auth/oauth/github-copilot.ts`).
+
+/// `User-Agent` for the Copilot path (overrides the default kigi UA, OAuth-gated).
+pub const COPILOT_USER_AGENT: &str = "GitHubCopilotChat/0.35.0";
+/// `Editor-Version` — the host editor Copilot believes it is talking to.
+pub const COPILOT_EDITOR_VERSION: &str = "vscode/1.107.0";
+/// `Editor-Plugin-Version` — the Copilot Chat plugin build.
+pub const COPILOT_EDITOR_PLUGIN_VERSION: &str = "copilot-chat/0.35.0";
+/// `Copilot-Integration-Id` — the integration the token is scoped to.
+pub const COPILOT_INTEGRATION_ID: &str = "vscode-chat";
+/// `X-GitHub-Api-Version` — sent ONLY on the `/models` listing.
+pub const COPILOT_API_VERSION: &str = "2026-06-01";
+/// `X-Initiator` value — sent ONLY on inference (`user`, per the spec).
+pub const COPILOT_INITIATOR: &str = "user";
+
+// ── ChatGPT/Codex (openai-codex) OAuth-inference headers ─────────────────────
+// The ChatGPT Codex backend authorizes an OAuth bearer AND validates the Codex
+// client identity. These ride the `/codex/responses` inference request ONLY.
+// openai-codex-GATED: no other Responses provider (API-key `openai`) sends them,
+// so their requests stay byte-identical. Values are non-secret wire constants
+// (ported from the official Codex CLI + Pi `api/openai-codex-responses.ts`).
+
+/// `originator` header identifying the Codex CLI client (matches the authorize
+/// `originator` param).
+pub const CODEX_ORIGINATOR: &str = "codex_cli_rs";
+/// `OpenAI-Beta` opt-in the Codex Responses endpoint requires.
+pub const CODEX_OPENAI_BETA: &str = "responses=experimental";
+/// `User-Agent` presented on the Codex path (overrides the default kigi UA,
+/// openai-codex-gated). The Codex backend does not strictly validate the UA
+/// string (Pi ships its own and it works), so this is a stable best-effort
+/// identity, not a pinned build.
+pub const CODEX_USER_AGENT: &str = "codex_cli_rs/0.104.0";
+/// JWT payload claim namespace carrying the ChatGPT account id.
+const CODEX_JWT_AUTH_CLAIM: &str = "https://api.openai.com/auth";
+
+/// Extract the `chatgpt_account_id` from a Codex OAuth access token (a JWT):
+/// base64url-decode the payload segment and read
+/// `["https://api.openai.com/auth"]["chatgpt_account_id"]`. Returns `None` when
+/// the token is not a well-formed JWT or the claim is missing/empty.
+///
+/// Used BOTH at login (fail-fast: a token without the claim is useless) and at
+/// inference (the header is derived STATELESSLY from the current bearer, so a
+/// refreshed token — which still carries the claim — needs no persisted field).
+///
+/// SECURITY: the token, its payload, and the returned account id are NEVER
+/// logged by this function or its callers.
+pub fn chatgpt_account_id_from_jwt(token: &str) -> Option<String> {
+    use base64::Engine;
+    // A JWT is exactly three dot-separated segments; anything else is not a
+    // token we can read (fail closed rather than decode a lookalike).
+    let mut segments = token.split('.');
+    let (_header, payload_b64, _signature) = (segments.next()?, segments.next()?, segments.next()?);
+    if segments.next().is_some() {
+        return None;
+    }
+    // JWT payloads are base64url without padding; be tolerant of either.
+    let bytes = base64::engine::general_purpose::URL_SAFE_NO_PAD
+        .decode(payload_b64)
+        .or_else(|_| base64::engine::general_purpose::URL_SAFE.decode(payload_b64))
+        .ok()?;
+    let claims: serde_json::Value = serde_json::from_slice(&bytes).ok()?;
+    let account_id = claims
+        .get(CODEX_JWT_AUTH_CLAIM)?
+        .get("chatgpt_account_id")?
+        .as_str()?;
+    (!account_id.is_empty()).then(|| account_id.to_string())
+}
+
+/// ChatCompletions request-body adaptation dialect. Providers disagree on
+/// how thinking rides an OpenAI-compatible body: Kimi wants
+/// `thinking:{type,effort}`, DeepSeek wants
+/// `thinking:{type,reasoning_effort}`, most others take the OpenAI-style
+/// `reasoning_effort` scalar untouched.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum ChatCompat {
+    /// Kimi wire (`thinking:{type,effort}`, canonical xhigh spelled max).
+    /// The default: BYOK/custom ChatCompletions endpoints keep the
+    /// historical kigi behavior.
+    #[default]
+    Kimi,
+    /// DeepSeek wire (`thinking:{type,reasoning_effort}`, official docs:
+    /// low/medium map to high and xhigh to max server-side).
+    DeepSeek,
+    /// Leave the body as-is (OpenAI-style `reasoning_effort` passes through).
+    Passthrough,
+    /// Strict OpenAI-compatible validators (Mistral, Cerebras) reject any
+    /// out-of-schema request field with a 4xx (`additionalProperties:false`).
+    /// kigi injects `stream_options.include_usage` on every streaming
+    /// request, which such validators reject, so it is stripped (streaming
+    /// usage falls back to token estimation). `reasoning_effort` passes
+    /// through; private message fields are stripped like Passthrough.
+    /// (Serde alias `mistral` keeps sessions persisted before the rename.)
+    #[serde(alias = "mistral")]
+    StrictOpenAi,
+}
+
 pub const REASONING_EFFORT_META_KEY: &str = "reasoningEffort";
 pub const SUPPORTS_REASONING_EFFORT_META_KEY: &str = "supportsReasoningEffort";
 
@@ -1054,6 +1443,11 @@ pub struct SamplingConfig {
     pub max_completion_tokens: Option<u32>,
     pub temperature: Option<f32>,
     pub top_p: Option<f32>,
+    /// ChatCompletions body-adaptation dialect (per-platform; serde-default
+    /// Kimi keeps pre-field sessions and BYOK endpoints on the historical
+    /// behavior).
+    #[serde(default)]
+    pub chat_compat: ChatCompat,
     /// Which API backend to use for this model
     #[serde(default)]
     pub api_backend: ApiBackend,
@@ -1081,6 +1475,12 @@ pub struct CreateResponseWrapper {
     /// The inner Responses API request.
     pub inner: crate::rs::CreateResponse,
 
+    /// Canonical reasoning effort for this request. The typed
+    /// `inner.reasoning.effort` stays `None` (async-openai's enum cannot
+    /// spell `max`); [`patch_reasoning_effort`] writes this value onto the
+    /// serialized body just before send.
+    pub reasoning_effort: Option<ReasoningEffort>,
+
     /// Custom header: conversation ID for tracking.
     pub x_kigi_conv_id: Option<String>,
 
@@ -1107,6 +1507,7 @@ impl CreateResponseWrapper {
     pub fn new(inner: crate::rs::CreateResponse) -> Self {
         Self {
             inner,
+            reasoning_effort: None,
             x_kigi_conv_id: None,
             x_kigi_req_id: None,
             x_kigi_session_id: None,
@@ -1215,6 +1616,173 @@ mod tests {
     use super::*;
     use serde_json::json;
 
+    /// The Codex backend rejects `role: system` input outright
+    /// (400 `{"detail":"System messages are not allowed"}`) — its system
+    /// channel is the top-level `instructions` field, and stateless
+    /// (`store: false`) reasoning replay needs
+    /// `include: ["reasoning.encrypted_content"]`. The adapter must hoist
+    /// every system item (string or parts content, order preserved),
+    /// append to existing instructions, and leave the rest of the input
+    /// untouched.
+    #[test]
+    fn codex_adapter_hoists_system_messages_and_requests_encrypted_reasoning() {
+        let mut body = json!({
+            "model": "gpt-5.2-codex",
+            "instructions": "base",
+            "input": [
+                {"type": "message", "role": "system", "content": "sys head"},
+                {"type": "message", "role": "user", "content": "hello"},
+                {"type": "message", "role": "system", "content": [
+                    {"type": "input_text", "text": "memory reminder"}
+                ]},
+                {"type": "message", "role": "assistant", "content": "hi"}
+            ]
+        });
+        adapt_body_for_codex_backend(&mut body);
+
+        let input = body["input"].as_array().unwrap();
+        assert_eq!(input.len(), 2, "system items removed from input: {body:#}");
+        assert!(
+            input.iter().all(|i| i["role"] != "system"),
+            "no system role may remain: {body:#}"
+        );
+        assert_eq!(
+            body["instructions"], "base\n\nsys head\n\nmemory reminder",
+            "system content hoisted into instructions, order preserved"
+        );
+        assert_eq!(
+            body["include"],
+            json!(["reasoning.encrypted_content"]),
+            "stateless reasoning replay requires the include"
+        );
+
+        // Idempotent: a second pass changes nothing.
+        let before = body.clone();
+        adapt_body_for_codex_backend(&mut body);
+        assert_eq!(body, before);
+    }
+
+    /// No system items and no prior instructions: input untouched, no
+    /// empty-string instructions invented, include still requested.
+    #[test]
+    fn codex_adapter_without_system_messages_only_adds_include() {
+        let mut body = json!({
+            "model": "gpt-5.2-codex",
+            "input": [{"type": "message", "role": "user", "content": "q"}]
+        });
+        adapt_body_for_codex_backend(&mut body);
+        assert!(body.get("instructions").is_none(), "{body:#}");
+        assert_eq!(body["input"].as_array().unwrap().len(), 1);
+        assert_eq!(body["include"], json!(["reasoning.encrypted_content"]));
+    }
+
+    /// String content (the only shape non-Mistral providers send) stays the
+    /// answer verbatim with no thinking — byte-identical to the pre-change
+    /// deserialization.
+    /// The StrictOpenAi dialect kept the serde alias `mistral`, so sessions
+    /// persisted before the rename still deserialize.
+    #[test]
+    fn chat_compat_mistral_alias_deserializes_to_strict_openai() {
+        let v: ChatCompat = serde_json::from_str("\"mistral\"").unwrap();
+        assert_eq!(v, ChatCompat::StrictOpenAi);
+        // New value round-trips as strict_open_ai.
+        let v: ChatCompat = serde_json::from_str("\"strict_open_ai\"").unwrap();
+        assert_eq!(v, ChatCompat::StrictOpenAi);
+        assert_eq!(
+            serde_json::to_string(&ChatCompat::StrictOpenAi).unwrap(),
+            "\"strict_open_ai\""
+        );
+    }
+
+    #[test]
+    fn chunk_delta_string_content_unchanged() {
+        let delta: ChatChunkDelta =
+            serde_json::from_value(json!({ "content": "hello world" })).unwrap();
+        assert_eq!(delta.content.as_deref(), Some("hello world"));
+        assert_eq!(delta.reasoning_content, None);
+        // A DeepSeek-style wire reasoning_content still rides its own field.
+        let delta: ChatChunkDelta =
+            serde_json::from_value(json!({ "content": "hi", "reasoning_content": "because" }))
+                .unwrap();
+        assert_eq!(delta.content.as_deref(), Some("hi"));
+        assert_eq!(delta.reasoning_content.as_deref(), Some("because"));
+        // Null / absent content → None.
+        let delta: ChatChunkDelta = serde_json::from_value(json!({})).unwrap();
+        assert_eq!(delta.content, None);
+    }
+
+    /// Mistral reasoning: array content routes `text` chunks to the answer
+    /// and the nested `text` of `thinking` chunks to reasoning_content
+    /// (verified shapes from the mistralai/client-python SDK).
+    #[test]
+    fn chunk_delta_array_content_splits_thinking_and_answer() {
+        // Thinking phase: array with only a thinking chunk (nested list).
+        let delta: ChatChunkDelta = serde_json::from_value(json!({
+            "content": [
+                { "type": "thinking", "thinking": [{ "type": "text", "text": "step 1" }],
+                  "signature": null, "closed": false }
+            ]
+        }))
+        .unwrap();
+        assert_eq!(delta.content, None, "thinking phase has no visible answer");
+        assert_eq!(delta.reasoning_content.as_deref(), Some("step 1"));
+
+        // Answer phase: plain string delta.
+        let delta: ChatChunkDelta =
+            serde_json::from_value(json!({ "content": "the answer" })).unwrap();
+        assert_eq!(delta.content.as_deref(), Some("the answer"));
+        assert_eq!(delta.reasoning_content, None);
+
+        // Transition delta: one array carrying both a closing thinking chunk
+        // and the first text chunk → both channels populated.
+        let delta: ChatChunkDelta = serde_json::from_value(json!({
+            "content": [
+                { "type": "thinking", "thinking": [{ "type": "text", "text": "done" }] },
+                { "type": "text", "text": "Answer: 22" }
+            ]
+        }))
+        .unwrap();
+        assert_eq!(delta.content.as_deref(), Some("Answer: 22"));
+        assert_eq!(delta.reasoning_content.as_deref(), Some("done"));
+    }
+
+    /// The chunk union is OPEN: unknown chunk types and reference chunks are
+    /// ignored, never fatal (SDK's UnknownContentChunk fallback).
+    #[test]
+    fn content_array_tolerates_unknown_chunk_types() {
+        let delta: ChatChunkDelta = serde_json::from_value(json!({
+            "content": [
+                { "type": "reference", "reference_ids": [1, 2] },
+                { "type": "some_future_chunk", "payload": { "x": 1 } },
+                { "type": "text", "text": "visible" }
+            ]
+        }))
+        .unwrap();
+        assert_eq!(delta.content.as_deref(), Some("visible"));
+        assert_eq!(delta.reasoning_content, None);
+    }
+
+    /// Non-streaming ChatResponseMessage gets the same array handling.
+    #[test]
+    fn response_message_array_content_splits() {
+        let msg: ChatResponseMessage = serde_json::from_value(json!({
+            "role": "assistant",
+            "content": [
+                { "type": "thinking", "thinking": [{ "type": "text", "text": "reasoning" }] },
+                { "type": "text", "text": "final" }
+            ]
+        }))
+        .unwrap();
+        assert_eq!(msg.content.as_deref(), Some("final"));
+        assert_eq!(msg.reasoning_content.as_deref(), Some("reasoning"));
+        // String content still works.
+        let msg: ChatResponseMessage = serde_json::from_value(json!({
+            "role": "assistant", "content": "plain"
+        }))
+        .unwrap();
+        assert_eq!(msg.content.as_deref(), Some("plain"));
+    }
+
     #[test]
     fn reasoning_effort_serde_lowercase_round_trip() {
         for v in [
@@ -1224,6 +1792,7 @@ mod tests {
             ReasoningEffort::Medium,
             ReasoningEffort::High,
             ReasoningEffort::Xhigh,
+            ReasoningEffort::Max,
         ] {
             let json = serde_json::to_string(&v).unwrap();
             assert_eq!(json, format!("\"{}\"", v.as_str()), "serialize {v:?}");
@@ -1231,31 +1800,112 @@ mod tests {
             assert_eq!(back, v, "round-trip {v:?}");
         }
         assert!(serde_json::from_str::<ReasoningEffort>("\"BOGUS\"").is_err());
-        assert!(serde_json::from_str::<ReasoningEffort>("\"max\"").is_err());
     }
 
     #[test]
-    fn reasoning_effort_from_str_accepts_max_as_xhigh() {
+    fn reasoning_effort_from_str_max_and_xhigh_are_distinct() {
         assert_eq!(
             "max".parse::<ReasoningEffort>().unwrap(),
-            ReasoningEffort::Xhigh
+            ReasoningEffort::Max
         );
         assert_eq!(
             "MAX".parse::<ReasoningEffort>().unwrap(),
-            ReasoningEffort::Xhigh
+            ReasoningEffort::Max
         );
         assert_eq!(
             "xhigh".parse::<ReasoningEffort>().unwrap(),
             ReasoningEffort::Xhigh
         );
-        assert_eq!(ReasoningEffort::Xhigh.as_str(), "xhigh");
+        // Messages API: distinct wire tokens per level (2026 capabilities).
+        assert_eq!(ReasoningEffort::Xhigh.to_messages_api(), Some("xhigh"));
+        assert_eq!(ReasoningEffort::Max.to_messages_api(), Some("max"));
+    }
+
+    /// The codex-only `ultra` tier parses, serializes, and patches onto a
+    /// Responses body as `reasoning.effort = "ultra"` (the crux of surfacing a
+    /// codex model's full thinking menu). It is a DISTINCT level above `max`.
+    #[test]
+    fn reasoning_effort_ultra_is_a_distinct_codex_tier() {
+        assert_eq!(
+            "ultra".parse::<ReasoningEffort>().unwrap(),
+            ReasoningEffort::Ultra
+        );
+        assert_eq!(
+            "ULTRA".parse::<ReasoningEffort>().unwrap(),
+            ReasoningEffort::Ultra
+        );
+        assert_ne!(ReasoningEffort::Ultra, ReasoningEffort::Max);
+        assert_eq!(ReasoningEffort::Ultra.as_str(), "ultra");
+        let json = serde_json::to_string(&ReasoningEffort::Ultra).unwrap();
+        assert_eq!(json, "\"ultra\"");
+        assert_eq!(
+            serde_json::from_str::<ReasoningEffort>("\"ultra\"").unwrap(),
+            ReasoningEffort::Ultra
+        );
+        let mut body = serde_json::json!({ "model": "gpt-5.6-sol" });
+        patch_reasoning_effort(&mut body, Some(ReasoningEffort::Ultra));
+        assert_eq!(body["reasoning"]["effort"], "ultra");
+    }
+
+    /// The account id is decoded STATELESSLY from the bearer JWT payload's
+    /// `["https://api.openai.com/auth"]["chatgpt_account_id"]` claim; a token
+    /// without the claim (or not a JWT) yields `None` (login fails fast on it).
+    #[test]
+    fn chatgpt_account_id_extracted_from_jwt_claim() {
+        use base64::Engine;
+        let make_jwt = |payload: serde_json::Value| -> String {
+            let header =
+                base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(br#"{"alg":"none"}"#);
+            let body = base64::engine::general_purpose::URL_SAFE_NO_PAD
+                .encode(serde_json::to_vec(&payload).unwrap());
+            format!("{header}.{body}.sig")
+        };
+        let good = make_jwt(serde_json::json!({
+            "https://api.openai.com/auth": { "chatgpt_account_id": "acct-abc-123" },
+            "sub": "user-1"
+        }));
+        assert_eq!(
+            chatgpt_account_id_from_jwt(&good).as_deref(),
+            Some("acct-abc-123")
+        );
+        // Claim namespace present but no account id → None (fail-fast).
+        let no_account = make_jwt(serde_json::json!({
+            "https://api.openai.com/auth": { "user_id": "u" }
+        }));
+        assert_eq!(chatgpt_account_id_from_jwt(&no_account), None);
+        // Empty account id → None.
+        let empty = make_jwt(serde_json::json!({
+            "https://api.openai.com/auth": { "chatgpt_account_id": "" }
+        }));
+        assert_eq!(chatgpt_account_id_from_jwt(&empty), None);
+        // Not a JWT (no payload segment) → None.
+        assert_eq!(chatgpt_account_id_from_jwt("not-a-jwt"), None);
+        assert_eq!(chatgpt_account_id_from_jwt(""), None);
+        // A JWT is EXACTLY three segments: a lookalike with too few or too many
+        // is rejected outright rather than decoded (fail closed).
+        let payload = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(
+            serde_json::to_vec(&serde_json::json!({
+                "https://api.openai.com/auth": { "chatgpt_account_id": "acct-abc-123" }
+            }))
+            .unwrap(),
+        );
+        assert_eq!(
+            chatgpt_account_id_from_jwt(&format!("hdr.{payload}")),
+            None,
+            "two segments is not a JWT"
+        );
+        assert_eq!(
+            chatgpt_account_id_from_jwt(&format!("hdr.{payload}.sig.extra")),
+            None,
+            "four segments is not a JWT"
+        );
     }
 
     #[test]
     fn parse_canonical_effort_token_helper() {
         assert_eq!(
             parse_canonical_effort_token("max"),
-            Some(ReasoningEffort::Xhigh)
+            Some(ReasoningEffort::Max)
         );
         assert_eq!(
             parse_canonical_effort_token("high"),
@@ -1415,7 +2065,8 @@ mod tests {
         );
         let bad_type = as_map(serde_json::json!({"reasoningEffort": 3}));
         assert_eq!(parse_reasoning_effort_meta(Some(&bad_type)), None);
-        let unknown = as_map(serde_json::json!({"reasoningEffort": "ULTRA"}));
+        // `ultra` is now a real codex tier; a genuinely-unknown token still None.
+        let unknown = as_map(serde_json::json!({"reasoningEffort": "MEGA"}));
         assert_eq!(parse_reasoning_effort_meta(Some(&unknown)), None);
     }
 

crates/codegen/kigi-shell-base/src/util/fs.rs

@@ -0,0 +1,87 @@
+//! Filesystem primitives shared across the shell.
+
+use std::io;
+use std::path::Path;
+
+/// Replace `dest` with `tmp` — the commit step of every tmp+rename atomic
+/// write in the product. This is the ONE place that knows how to make that
+/// commit stick on Windows; call sites must never inline a bare
+/// `fs::rename` replace again.
+///
+/// Unix `rename(2)` replaces atomically and needs no help. Windows
+/// `MoveFileExW(REPLACE_EXISTING)` fails with a sharing violation while
+/// ANOTHER process (antivirus scanner, search indexer, cloud sync) holds
+/// `dest` open — the classic "persists on macOS, silently doesn't on
+/// Windows" failure (a model switch that never sticks, a stale models
+/// cache). On Windows a failed rename therefore deletes the destination
+/// first (the pattern `auth/storage.rs` shipped first) and retries with two
+/// short back-offs for scanners that hold the file for a few milliseconds.
+///
+/// On final failure the tmp file is removed (no litter) and the error is
+/// returned — callers decide severity, but MUST at least log it (errors
+/// never pass silently).
+pub fn replace_file(tmp: &Path, dest: &Path) -> io::Result<()> {
+    let result = replace_file_inner(tmp, dest);
+    if result.is_err() {
+        let _ = std::fs::remove_file(tmp);
+    }
+    result
+}
+
+#[cfg(not(windows))]
+fn replace_file_inner(tmp: &Path, dest: &Path) -> io::Result<()> {
+    std::fs::rename(tmp, dest)
+}
+
+#[cfg(windows)]
+fn replace_file_inner(tmp: &Path, dest: &Path) -> io::Result<()> {
+    let mut last = match std::fs::rename(tmp, dest) {
+        Ok(()) => return Ok(()),
+        Err(e) => e,
+    };
+    for backoff_ms in [0u64, 10, 50] {
+        if backoff_ms > 0 {
+            std::thread::sleep(std::time::Duration::from_millis(backoff_ms));
+        }
+        // Delete-first: marks an open-with-delete-sharing file for deletion
+        // and clears the way for a plain rename; harmless when absent.
+        let _ = std::fs::remove_file(dest);
+        match std::fs::rename(tmp, dest) {
+            Ok(()) => return Ok(()),
+            Err(e) => last = e,
+        }
+    }
+    Err(last)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The common contract on every platform: replace over an existing
+    /// destination, create a missing one, and error (cleaning the tmp)
+    /// when the tmp itself is missing.
+    #[test]
+    fn replace_file_commits_and_cleans_up() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let dest = dir.path().join("target.json");
+        let tmp = dir.path().join("target.json.tmp");
+
+        // Create-missing.
+        std::fs::write(&tmp, b"v1").unwrap();
+        replace_file(&tmp, &dest).expect("create");
+        assert_eq!(std::fs::read(&dest).unwrap(), b"v1");
+        assert!(!tmp.exists(), "tmp must be consumed");
+
+        // Replace-existing.
+        std::fs::write(&tmp, b"v2").unwrap();
+        replace_file(&tmp, &dest).expect("replace");
+        assert_eq!(std::fs::read(&dest).unwrap(), b"v2");
+        assert!(!tmp.exists());
+
+        // Missing tmp → error, dest untouched.
+        let err = replace_file(&tmp, &dest).expect_err("missing tmp must fail");
+        assert_eq!(err.kind(), io::ErrorKind::NotFound);
+        assert_eq!(std::fs::read(&dest).unwrap(), b"v2");
+    }
+}

crates/codegen/kigi-shell-base/src/util/mod.rs

@@ -1,4 +1,5 @@
 pub mod event_id;
+pub mod fs;
 pub mod kigi_home;
 pub mod secure_file;
 pub mod tips;
@@ -34,7 +35,17 @@ pub fn random_f64() -> f64 {
 pub fn probabilistic_sample(rate: f64) -> bool {
     random_f64() < rate
 }
-fn matches_trusted_base_url(candidate: &str, trusted_base: &str) -> bool {
+/// True when `candidate` is `trusted_base` or a path below it, comparing
+/// scheme, host and effective port exactly (so suffix attacks such as
+/// `api.kimi.com.evil.example` never match).
+///
+/// Public because the credential chokepoint
+/// ([`kigi_shell::auth::credential_authority`](../../../kigi_shell/auth/credential_authority/index.html))
+/// must compare a request URL against the SESSION's effective endpoints
+/// (`[endpoints] coding_api_base_url` from config.toml, `models_base_url`, a
+/// platform's own registry host) — none of which the env-var-only predicates
+/// below can see.
+pub fn matches_trusted_base_url(candidate: &str, trusted_base: &str) -> bool {
     let Ok(candidate) = reqwest::Url::parse(candidate) else {
         return false;
     };

crates/codegen/kigi-shell/build.rs

@@ -109,21 +109,43 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
         t = asset_triple
     );
 
+    // Transient CDN hiccups (502/503/timeouts) are retried with backoff: a
+    // single flaky response must not kill a ~50-minute release build (it
+    // did — the v0.1.5 tag build failed on one 502). Genuine failures
+    // (404, offline) still error out with the offline-build hint.
     let bytes: Vec<u8> = {
-        let resp = reqwest::blocking::get(&url).map_err(|e| {
-            format!(
-                "Failed to download ripgrep: {}\nSet KIGI_SHELL_BUNDLE_RG_PATH to a local rg for offline builds.",
-                e
-            )
-        })?;
-        if !resp.status().is_success() {
-            return Err(format!(
-                "HTTP {} downloading ripgrep. Set KIGI_SHELL_BUNDLE_RG_PATH for offline builds.",
-                resp.status()
-            )
-            .into());
+        let mut last_err = String::new();
+        let mut bytes = None;
+        for (attempt, backoff_secs) in [0u64, 2, 8].into_iter().enumerate() {
+            if backoff_secs > 0 {
+                std::thread::sleep(std::time::Duration::from_secs(backoff_secs));
             }
-        resp.bytes()?.to_vec()
+            match reqwest::blocking::get(&url) {
+                Ok(resp) if resp.status().is_success() => match resp.bytes() {
+                    Ok(b) => {
+                        bytes = Some(b.to_vec());
+                        break;
+                    }
+                    Err(e) => last_err = format!("reading ripgrep body: {e}"),
+                },
+                Ok(resp) => {
+                    let status = resp.status();
+                    last_err = format!("HTTP {status} downloading ripgrep");
+                    // Only server-side/transient statuses are worth retrying.
+                    if !(status.is_server_error() || status.as_u16() == 429) {
+                        break;
+                    }
+                }
+                Err(e) => last_err = format!("Failed to download ripgrep: {e}"),
+            }
+            println!(
+                "cargo:warning=ripgrep download attempt {} failed: {last_err}",
+                attempt + 1
+            );
+        }
+        bytes.ok_or_else(|| {
+            format!("{last_err}. Set KIGI_SHELL_BUNDLE_RG_PATH for offline builds.")
+        })?
     };
 
     let gz = flate2::read::GzDecoder::new(&bytes[..]);

crates/codegen/kigi-shell/src/active_sessions.rs

@@ -170,9 +170,7 @@ fn write_data_file_atomic(
     let json = serde_json::to_string_pretty(sessions)
         .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
     fs::write(tmp_path, json.as_bytes())?;
-    fs::rename(tmp_path, data_path).inspect_err(|_| {
-        let _ = fs::remove_file(tmp_path);
-    })
+    crate::util::fs::replace_file(tmp_path, data_path)
 }
 
 fn is_pid_alive(pid: u32) -> bool {

crates/codegen/kigi-shell/src/agent/app.rs

@@ -178,7 +178,13 @@ async fn prefetch_models(agent_config: &AgentConfig) -> Option<IndexMap<String,
 
     if auth.is_some() || endpoints.has_custom_endpoint() || platform_keys.any() {
         tokio::task::spawn_blocking(move || {
-            prefetch_models_blocking(&endpoints, auth.as_ref(), fetch_auth, &platform_keys)
+            prefetch_models_blocking(
+                &endpoints,
+                auth.as_ref(),
+                &Default::default(),
+                fetch_auth,
+                &platform_keys,
+            )
         })
         .await
         .ok()
@@ -621,6 +627,7 @@ pub async fn run_leader(
         crate::agent::models::prefetch_models_blocking(
             &endpoints_for_prefetch,
             auth_for_prefetch.as_ref(),
+            &Default::default(),
             fetch_auth_for_prefetch,
             &platform_keys_for_prefetch,
         )

crates/codegen/kigi-shell/src/agent/config_model_override_parse.rs

@@ -311,7 +311,7 @@ mod tests {
             w.kind == ModelOverrideWarningKind::DuplicateAlias
                 && w.field.as_deref() == Some("send_compactions_remaining")
         }));
-        let resolved = crate::agent::config::resolve_model_list(&cfg, None);
+        let resolved = crate::agent::config::resolve_model_list(&cfg, None, &Default::default());
         assert!(resolved.contains_key("kigi-4.5"));
     }
 

crates/codegen/kigi-shell/src/agent/enrichment_fetch.rs

@@ -0,0 +1,432 @@
+//! models.dev enrichment catalog loading (agent-side IO).
+//!
+//! Providers whose `/models` wire serves no context/thinking metadata get it
+//! from models.dev (see `kigi_models::enrichment`). This module owns the IO:
+//! a 24h-TTL disk cache under `~/.kigi`, a runtime refresh of
+//! `https://models.dev/api.json` (filtered to registry providers before
+//! caching), and the bundled-snapshot fallback. NO NETWORK unless some
+//! enabled platform actually needs enrichment (`wire_serves_metadata` false)
+//! — with only Kimi/Moonshot configured this module never leaves disk.
+
+use std::collections::BTreeSet;
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
+use kigi_models::enrichment::{EnrichmentCatalog, bundled_enrichment, parse_api_json};
+
+/// Override the refresh URL (e2e mock), or disable refresh entirely with
+/// `0`/`off` (bundled snapshot + existing cache only).
+pub(crate) const MODELS_DEV_URL_ENV: &str = "KIGI_MODELS_DEV_URL";
+const DEFAULT_MODELS_DEV_URL: &str = "https://models.dev/api.json";
+const CACHE_FILE: &str = "models_dev_cache.json";
+const CACHE_TTL: Duration = Duration::from_secs(24 * 60 * 60);
+
+#[derive(serde::Serialize, serde::Deserialize)]
+struct DiskCache {
+    /// Unix seconds of the successful fetch.
+    fetched_at: u64,
+    /// The kigi version that wrote the cache — a different binary (upgrade
+    /// OR downgrade) refetches rather than trusting old filtering rules.
+    #[serde(default)]
+    kigi_version: String,
+    /// The keep-set the catalog was filtered to. A registry change (new
+    /// provider row, models_dev_id rename) invalidates the cache instead of
+    /// silently serving a catalog missing the new provider for up to 24h.
+    #[serde(default)]
+    keep_set: Vec<String>,
+    /// Already filtered + transformed catalog.
+    catalog: EnrichmentCatalog,
+}
+
+fn current_keep_set() -> Vec<String> {
+    registry_models_dev_ids()
+        .into_iter()
+        .map(str::to_owned)
+        .collect()
+}
+
+/// Whether any of `platforms` needs enrichment at all.
+pub(crate) fn any_platform_needs_enrichment(platforms: &[kigi_models::PlatformId]) -> bool {
+    platforms.iter().any(|p| !p.wire_serves_metadata())
+}
+
+/// The registry's models.dev provider ids (the refresh filter).
+fn registry_models_dev_ids() -> BTreeSet<&'static str> {
+    kigi_models::PlatformId::ALL
+        .into_iter()
+        .filter_map(|p| p.models_dev_id())
+        .collect()
+}
+
+/// Cache dir override (tests re-home the cache away from the real
+/// `~/.kigi` — same pattern as `KIGI_MODELS_CACHE_DIR`).
+pub(crate) const MODELS_DEV_CACHE_DIR_ENV: &str = "KIGI_MODELS_DEV_CACHE_DIR";
+
+fn cache_path() -> std::path::PathBuf {
+    match std::env::var(MODELS_DEV_CACHE_DIR_ENV) {
+        Ok(dir) if !dir.trim().is_empty() => std::path::PathBuf::from(dir).join(CACHE_FILE),
+        _ => crate::util::kigi_home::kigi_home().join(CACHE_FILE),
+    }
+}
+
+fn refresh_url() -> Option<String> {
+    match std::env::var(MODELS_DEV_URL_ENV) {
+        Ok(v)
+            if matches!(
+                v.trim().to_ascii_lowercase().as_str(),
+                "0" | "off" | "false"
+            ) =>
+        {
+            None
+        }
+        Ok(v) if !v.trim().is_empty() => Some(v.trim().to_string()),
+        _ => Some(DEFAULT_MODELS_DEV_URL.to_string()),
+    }
+}
+
+fn now_unix() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0)
+}
+
+/// Load the enrichment catalog for a fetch pass over `enabled` platforms.
+///
+/// Fast path: nothing needs enrichment → empty catalog, zero IO (the merge
+/// branch is never taken for wire-served platforms, and NOT forcing the
+/// bundled parse keeps its cost/panic surface off the kimi/moonshot path).
+/// Otherwise: fresh valid disk cache → use it; else refresh over HTTP
+/// (filter + transform + best-effort cache write); on refresh failure fall
+/// back to a STALE cache, then the bundled snapshot — each step logged,
+/// never silent.
+pub(crate) fn load_enrichment_catalog(
+    enabled: &[kigi_models::PlatformId],
+) -> std::borrow::Cow<'static, EnrichmentCatalog> {
+    if !any_platform_needs_enrichment(enabled) {
+        return std::borrow::Cow::Owned(EnrichmentCatalog::new());
+    }
+    load_enrichment_catalog_at(&cache_path())
+}
+
+/// Path-injectable core (tests use a tempdir path directly — no env, no
+/// `kigi_home()` OnceLock interaction).
+fn load_enrichment_catalog_at(
+    path: &std::path::Path,
+) -> std::borrow::Cow<'static, EnrichmentCatalog> {
+    use std::borrow::Cow;
+
+    let cached: Option<DiskCache> =
+        std::fs::read_to_string(path)
+            .ok()
+            .and_then(|s| match serde_json::from_str(&s) {
+                Ok(c) => Some(c),
+                Err(e) => {
+                    tracing::warn!(path = %path.display(), error = %e,
+                    "models.dev cache unreadable; refetching");
+                    None
+                }
+            });
+    let now = now_unix();
+    let cache_is_fresh = cached.as_ref().is_some_and(|c| {
+        // A future fetched_at (clock jump backwards, corrupt stamp) is
+        // stale, not fresh-forever; a different binary or keep-set means the
+        // cache was filtered under other rules — refetch instead of serving
+        // a catalog that may miss newly-registered providers for 24h.
+        c.fetched_at <= now
+            && now - c.fetched_at < CACHE_TTL.as_secs()
+            && c.kigi_version == kigi_version::VERSION
+            && c.keep_set == current_keep_set()
+    });
+    if cache_is_fresh {
+        tracing::debug!("models.dev enrichment: fresh disk cache");
+        return Cow::Owned(cached.expect("cache_is_fresh implies Some").catalog);
+    }
+
+    match refresh_url() {
+        Some(url) => match fetch_and_filter(&url) {
+            Ok(catalog) => {
+                let cache = DiskCache {
+                    fetched_at: now_unix(),
+                    kigi_version: kigi_version::VERSION.to_string(),
+                    keep_set: current_keep_set(),
+                    catalog,
+                };
+                // Best-effort write: a read-only home must not fail the fetch.
+                match serde_json::to_string(&cache) {
+                    Ok(body) => {
+                        if let Err(e) = crate::util::config::atomic_write_string(path, &body) {
+                            tracing::warn!(path = %path.display(), error = %e,
+                                "models.dev cache write failed; continuing in-memory");
+                        }
+                    }
+                    Err(e) => {
+                        tracing::warn!(error = %e, "models.dev cache serialize failed")
+                    }
+                }
+                tracing::info!("models.dev enrichment refreshed");
+                Cow::Owned(cache.catalog)
+            }
+            Err(e) => {
+                if let Some(c) = cached {
+                    tracing::warn!(error = %e,
+                        "models.dev refresh failed; using STALE cache");
+                    Cow::Owned(c.catalog)
+                } else {
+                    tracing::warn!(error = %e,
+                        "models.dev refresh failed; using bundled snapshot");
+                    Cow::Borrowed(bundled_enrichment())
+                }
+            }
+        },
+        None => {
+            tracing::info!("models.dev refresh disabled; using cache/bundled");
+            match cached {
+                Some(c) => Cow::Owned(c.catalog),
+                None => Cow::Borrowed(bundled_enrichment()),
+            }
+        }
+    }
+}
+
+fn fetch_and_filter(url: &str) -> anyhow::Result<EnrichmentCatalog> {
+    let response = crate::http::shared_blocking_client().get(url).send()?;
+    let status = response.status();
+    anyhow::ensure!(status.is_success(), "GET {url}: HTTP {}", status.as_u16());
+    let body = response.text()?;
+    let keep = registry_models_dev_ids();
+    Ok(parse_api_json(&body, Some(&keep))?)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use kigi_test_support::EnvGuard;
+    use serial_test::serial;
+
+    /// Wire-served-only platform sets never trigger IO — and never force the
+    /// bundled parse (empty owned catalog; the merge branch is gated off).
+    /// Kimi/Moonshot users therefore keep a zero-egress, zero-cache fetch
+    /// path even now that enrichment-needing platforms (OpenAI) exist.
+    #[test]
+    fn wire_served_platforms_get_empty_catalog_without_io() {
+        let wire_served = [
+            kigi_models::PlatformId::KimiCode,
+            kigi_models::PlatformId::MoonshotCn,
+            kigi_models::PlatformId::MoonshotAi,
+        ];
+        assert!(!any_platform_needs_enrichment(&wire_served));
+        let catalog = load_enrichment_catalog(&wire_served);
+        assert!(catalog.is_empty());
+        // The full registry now DOES need enrichment (OpenAI is
+        // wire_serves_metadata=false) — the fast path must not hide that.
+        assert!(any_platform_needs_enrichment(&kigi_models::PlatformId::ALL));
+    }
+
+    fn cache_file_in(dir: &tempfile::TempDir) -> std::path::PathBuf {
+        dir.path().join(CACHE_FILE)
+    }
+
+    fn write_cache(path: &std::path::Path, fetched_at: u64, versioned: bool) {
+        let cache = DiskCache {
+            fetched_at,
+            kigi_version: if versioned {
+                kigi_version::VERSION.to_string()
+            } else {
+                "0.0.0-other".to_string()
+            },
+            keep_set: current_keep_set(),
+            catalog: EnrichmentCatalog::from([(
+                "moonshotai".to_string(),
+                std::collections::BTreeMap::from([(
+                    "from-cache".to_string(),
+                    kigi_models::enrichment::EnrichmentModel {
+                        context: 111,
+                        ..Default::default()
+                    },
+                )]),
+            )]),
+        };
+        std::fs::write(path, serde_json::to_string(&cache).unwrap()).unwrap();
+    }
+
+    /// Fresh valid cache short-circuits — no HTTP (no mock server mounted:
+    /// a fetch attempt would fail and fall to bundled, failing the assert).
+    #[test]
+    #[serial]
+    fn fresh_valid_cache_is_served_without_refresh() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        write_cache(&path, now_unix(), true);
+        let _url = EnvGuard::set(MODELS_DEV_URL_ENV, "http://127.0.0.1:1/api.json");
+        let catalog = load_enrichment_catalog_at(&path);
+        assert!(
+            kigi_models::enrichment::lookup(&catalog, "moonshotai", "from-cache").is_some(),
+            "fresh cache must be served"
+        );
+    }
+
+    /// Version/keep-set/future-stamp guards: each invalidates a fresh-aged
+    /// cache. Refresh is disabled here, so invalidation falls through to the
+    /// STALE cache (resource degradation, not data loss) — proving both the
+    /// guard firing and the fallback order.
+    #[test]
+    #[serial]
+    fn cache_guards_invalidate_and_fall_back_to_stale() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        let _url = EnvGuard::set(MODELS_DEV_URL_ENV, "0");
+        // Wrong binary version → not fresh → (refresh disabled) → stale used.
+        write_cache(&path, now_unix(), false);
+        let catalog = load_enrichment_catalog_at(&path);
+        assert!(
+            kigi_models::enrichment::lookup(&catalog, "moonshotai", "from-cache").is_some(),
+            "stale-fallback must still serve the cached data"
+        );
+        // Future fetched_at → same path (guard fired: debug-log absence is
+        // not observable here; the behavioral pin is refresh-disabled + the
+        // wiremock test below proving a fired guard refetches).
+        write_cache(&path, now_unix() + 10_000, true);
+        let catalog = load_enrichment_catalog_at(&path);
+        assert!(kigi_models::enrichment::lookup(&catalog, "moonshotai", "from-cache").is_some());
+    }
+
+    /// An invalidated cache (wrong version) REFETCHES when refresh is
+    /// enabled: wiremock expect(1) proves the HTTP call happened; the new
+    /// cache file carries the current version + keep-set and the fetched
+    /// content replaces the stale entry.
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn invalidated_cache_refetches_and_rewrites() {
+        let _ = tracing_subscriber::fmt()
+            .with_env_filter(tracing_subscriber::EnvFilter::from_default_env())
+            .with_test_writer()
+            .try_init();
+        let server = wiremock::MockServer::start().await;
+        wiremock::Mock::given(wiremock::matchers::method("GET"))
+            .and(wiremock::matchers::path("/api.json"))
+            .respond_with(
+                wiremock::ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                    "moonshotai": { "models": { "from-refresh": {
+                        "limit": {"context": 222}
+                    }}}
+                })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        write_cache(&path, now_unix(), false);
+        let _url = EnvGuard::set(MODELS_DEV_URL_ENV, format!("{}/api.json", server.uri()));
+        let path2 = path.clone();
+        let catalog =
+            tokio::task::spawn_blocking(move || load_enrichment_catalog_at(&path2).into_owned())
+                .await
+                .unwrap();
+        assert!(
+            kigi_models::enrichment::lookup(&catalog, "moonshotai", "from-refresh").is_some(),
+            "guard-invalidated cache must refetch"
+        );
+        let rewritten: DiskCache =
+            serde_json::from_str(&std::fs::read_to_string(path).unwrap()).unwrap();
+        assert_eq!(rewritten.kigi_version, kigi_version::VERSION);
+        assert_eq!(rewritten.keep_set, current_keep_set());
+        assert!(rewritten.catalog.contains_key("moonshotai"));
+    }
+
+    /// Corrupted cache file → refetch (not a crash, not trust-garbage).
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn corrupted_cache_refetches() {
+        let server = wiremock::MockServer::start().await;
+        wiremock::Mock::given(wiremock::matchers::method("GET"))
+            .and(wiremock::matchers::path("/api.json"))
+            .respond_with(
+                wiremock::ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "moonshotai": { "models": {} } })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        std::fs::write(&path, "not json {").unwrap();
+        let _url = EnvGuard::set(MODELS_DEV_URL_ENV, format!("{}/api.json", server.uri()));
+        let path2 = path.clone();
+        let catalog =
+            tokio::task::spawn_blocking(move || load_enrichment_catalog_at(&path2).into_owned())
+                .await
+                .unwrap();
+        assert!(catalog.contains_key("moonshotai"));
+    }
+
+    /// Refresh failure with NO cache → bundled snapshot fallback.
+    #[test]
+    #[serial]
+    fn refresh_failure_without_cache_falls_back_to_bundled() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        let _url = EnvGuard::set(MODELS_DEV_URL_ENV, "http://127.0.0.1:1/api.json");
+        let catalog = load_enrichment_catalog_at(&path);
+        assert!(
+            kigi_models::enrichment::lookup(&catalog, "kimi-for-coding", "k3").is_some(),
+            "bundled snapshot must back a total refresh failure"
+        );
+        assert!(!path.exists(), "failed refresh must not write a cache");
+    }
+
+    /// Kill switch through the FULL load path (not just refresh_url): no
+    /// cache + refresh disabled → bundled, no HTTP attempted (an attempt
+    /// against the sentinel URL would be a hang/refusal, not bundled data).
+    #[test]
+    #[serial]
+    fn kill_switch_full_path_serves_bundled() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = cache_file_in(&dir);
+        for token in ["0", "off", "FALSE", " Off "] {
+            let _url = EnvGuard::set(MODELS_DEV_URL_ENV, token);
+            assert!(refresh_url().is_none(), "token {token:?} must disable");
+            let catalog = load_enrichment_catalog_at(&path);
+            assert!(kigi_models::enrichment::lookup(&catalog, "kimi-for-coding", "k3").is_some());
+        }
+    }
+
+    /// Refresh path: mock server → transform runs and the keep-set filters
+    /// to registry providers (`moonshotai` is a real registry models_dev id;
+    /// unknown providers are dropped).
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn refresh_transforms_and_filters_to_registry_ids() {
+        let server = wiremock::MockServer::start().await;
+        wiremock::Mock::given(wiremock::matchers::method("GET"))
+            .and(wiremock::matchers::path("/api.json"))
+            .respond_with(
+                wiremock::ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                    "moonshotai": { "models": { "kimi-test": {
+                        "limit": {"context": 262144},
+                        "reasoning": true,
+                        "reasoning_options": [
+                            {"type": "effort", "values": ["low", "high"]}
+                        ]
+                    }}},
+                    "not-in-registry": { "models": { "m": {} } }
+                })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let url = format!("{}/api.json", server.uri());
+        let catalog = tokio::task::spawn_blocking(move || fetch_and_filter(&url))
+            .await
+            .unwrap()
+            .expect("fetch must succeed");
+        assert!(
+            !catalog.contains_key("not-in-registry"),
+            "keep-set must drop providers outside the registry"
+        );
+        let m = kigi_models::enrichment::lookup(&catalog, "moonshotai", "kimi-test")
+            .expect("registry provider survives the filter");
+        assert_eq!(m.context, 262_144);
+        assert_eq!(m.efforts, ["low", "high"]);
+    }
+}

crates/codegen/kigi-shell/src/agent/handlers/model_switch.rs

@@ -110,6 +110,29 @@ pub(crate) async fn apply(
             .models_manager
             .model_supports_reasoning_effort(model_id.0.as_ref())
         {
+            // Legacy migration: pre-split sessions persisted canonical
+            // `xhigh` for models whose live menu now spells the top tier
+            // `max` (K3). The wire is identical either way (kimi_compat
+            // renames xhigh→max), but the menu has no xhigh-valued row, so
+            // display/active-row would drift from the model vocabulary and
+            // the stale token would be re-persisted forever. Migrate once.
+            let eff = if eff == kigi_sampling_types::ReasoningEffort::Xhigh
+                && !agent
+                    .models_manager
+                    .model_offers_effort(model_id.0.as_ref(), eff)
+                && agent.models_manager.model_offers_effort(
+                    model_id.0.as_ref(),
+                    kigi_sampling_types::ReasoningEffort::Max,
+                ) {
+                tracing::info!(
+                    session_id = % session_id.0,
+                    "set_session_model: migrating legacy xhigh override to max \
+                     (model menu offers max, not xhigh)"
+                );
+                kigi_sampling_types::ReasoningEffort::Max
+            } else {
+                eff
+            };
             tracing::info!(
                 session_id = % session_id.0, effort = % eff,
                 "set_session_model: applying reasoning_effort override from meta"
@@ -169,9 +192,18 @@ pub(crate) async fn apply(
             model.map(|e| &e.info),
         )
     };
+    // H4: hand the session the catalog KEY the picker actually resolved. The
+    // slug in `model_sampling.model` cannot distinguish `xai/grok-*` from
+    // `xai-grok/grok-*` (duplicate ids across an API-key platform and its
+    // subscription-OAuth twin are by design), and the process-global
+    // `current_model_id()` below is not written at all in Leader mode.
+    let catalog_key =
+        crate::agent::models::resolve_catalog_key(&agent.models_manager.models(), &model_id)
+            .map(|k| k.0.to_string());
     let (tx, rx) = oneshot::channel();
     let _ = handle.cmd_tx.send(SessionCommand::SetSessionModel {
         sampling_config: model_sampling,
+        catalog_key,
         use_concise,
         apply_prompt_override,
         skip_prompt_rewrite: did_rebuild || model_unchanged,

crates/codegen/kigi-shell/src/agent/mod.rs

@@ -4,6 +4,7 @@ pub mod auth_method;
 pub mod chat_modes;
 pub mod config;
 pub mod config_model_override_parse;
+pub(crate) mod enrichment_fetch;
 mod ext_parsers;
 pub(crate) mod feedback_client;
 pub mod folder_trust;

crates/codegen/kigi-shell/src/agent/mvp_agent/acp_agent.rs

@@ -165,7 +165,7 @@ impl acp::Agent for MvpAgent {
                 &crate::util::kigi_home::kigi_home(),
             )
         {
-            unsafe { std::env::set_var("XAI_API_KEY", &api_key) };
+            unsafe { std::env::set_var("KIGI_API_KEY", &api_key) };
             tracing::info!("auth: loaded API key from auth.json (xai::api_key scope)");
             kigi_log::unified_log::info(
                 "auth: loaded API key from auth.json (xai::api_key scope)",
@@ -221,7 +221,24 @@ impl acp::Agent for MvpAgent {
             has_cached_token,
             login_label: None,
         });
-        let auth_methods = built.methods;
+        let mut auth_methods = built.methods;
+        // Connected badges for the client's login picker: probe stored
+        // credentials once (auth.json scopes + resolved platform keys) and
+        // stamp `_meta.connected` on every method that already has one.
+        {
+            let store = crate::auth::read_auth_json(
+                &crate::util::kigi_home::kigi_home().join("auth.json"),
+            )
+            .unwrap_or_default();
+            let keys = crate::agent::models::PlatformApiKeys::resolve_from_effective_config();
+            let connected = auth_method::connected_method_ids(
+                has_cached_token,
+                has_external_api_key,
+                |scope| store.contains_key(scope),
+                |p| keys.key_for(p).is_some(),
+            );
+            auth_method::stamp_connected_meta(&mut auth_methods, &connected);
+        }
         kigi_log::unified_log::info(
             "auth: initialize() built auth_methods for ACP response",
             None,
@@ -330,10 +347,28 @@ impl acp::Agent for MvpAgent {
         );
         match arguments.method_id.0.as_ref() {
             auth_method::XAI_API_KEY_METHOD_ID => {
+                // C1: the SECOND writer of the shared `sampling_config.api_key`.
+                // The house `KIGI_API_KEY` is the user's own key for the
+                // session's own endpoint, so the stamp requires the authority to
+                // classify that endpoint `CredentialClass::Primary` — NOT merely
+                // "takes some session credential", which is also true on a
+                // subscription-OAuth platform's host, where this key has no
+                // business (the shared config is the subagent baseline and the
+                // unresolved-model fallback). The key is still persisted to
+                // auth.json either way; only the stamp is guarded.
+                let takes_house_key = self.shared_config_takes_house_key();
                 let mut sampling_config = self.sampling_config.borrow_mut();
                 if sampling_config.api_key.is_none() {
                     if let Ok(api_key) = auth_method::read_xai_api_key_env() {
+                        if takes_house_key {
                             sampling_config.api_key = Some(api_key.clone());
+                        } else {
+                            tracing::debug!(
+                                model = sampling_config.model.as_str(),
+                                "auth: house api key withheld from the shared sampling config \
+                                 (its endpoint is not this session's own)"
+                            );
+                        }
                         if let Err(e) = crate::auth::store_api_key(
                             &crate::util::kigi_home::kigi_home(),
                             &api_key,
@@ -357,7 +392,7 @@ impl acp::Agent for MvpAgent {
                         return Err(
                             acp::Error::auth_required()
                                 .data(
-                                    "Set XAI_API_KEY or add api_key/env_key to config.toml.",
+                                    "Set KIGI_API_KEY or add api_key/env_key to config.toml.",
                                 ),
                         );
                     }
@@ -389,7 +424,7 @@ impl acp::Agent for MvpAgent {
                         ),
                     ),
                 );
-                let Some(auth) = self.auth_manager.current() else {
+                let Some(_auth) = self.auth_manager.current() else {
                     let message = if self.auth_manager.is_expired() {
                         "Session expired, re-authentication required"
                     } else {
@@ -408,9 +443,15 @@ impl acp::Agent for MvpAgent {
                         .await;
                 };
                 self.emit_settings_update_notification();
-                        {
-                    let mut sampling_config = self.sampling_config.borrow_mut();
-                    sampling_config.api_key = Some(auth.key);
+                // H2/C1: route the stamp through the ONE guard, which asks the
+                // authority which credential governs the shared config rather
+                // than being handed this one. That config may already point at
+                // a third-party model or at another provider's subscription
+                // host (it is the subagent baseline and the unresolved-model
+                // fallback), and `auth.key` authorizes only the session's own
+                // coding endpoint. The manager already holds this token, so the
+                // authority reads it back where it belongs.
+                if self.stamp_session_credential(true) {
                     tracing::debug!(
                         "auth: cached_token handler set api_key (SessionToken)"
                     );
@@ -484,9 +525,15 @@ impl acp::Agent for MvpAgent {
                         err.message = e.to_string();
                         err
                     })?;
-                {
-                    let mut sampling_config = self.sampling_config.borrow_mut();
-                    sampling_config.api_key = Some(auth.key.clone());
+                // C1: hot-swap FIRST, then let the authority read the fresh
+                // token back where it belongs. Nothing hand-carries `auth.key`
+                // to the shared config any more — the stamp is whatever
+                // credential governs that config's own model + endpoint, which
+                // for a session whose current model is another provider's
+                // subscription model is that provider's pooled token, and for a
+                // third-party host is nothing at all.
+                self.auth_manager.hot_swap(auth.clone());
+                if self.stamp_session_credential(true) {
                     tracing::debug!(
                         "auth: kimi.com/oidc handler set api_key (SessionToken)"
                     );
@@ -496,7 +543,6 @@ impl acp::Agent for MvpAgent {
                         None,
                     );
                 }
-                self.auth_manager.hot_swap(auth.clone());
                 self.emit_settings_update_notification();
                         self.set_auth_method(arguments.method_id.clone());
                 self.models_manager.on_auth_changed().await;
@@ -508,15 +554,22 @@ impl acp::Agent for MvpAgent {
                 );
                 Ok(self.auth_response_with_meta())
             }
-            auth_method::MOONSHOT_CN_METHOD_ID | auth_method::MOONSHOT_AI_METHOD_ID => {
-                let platform = auth_method::moonshot_platform_for_method_id(
+            _ => {
+                if let Some(platform) = auth_method::AuthMethodKind::from_id(
                     &arguments.method_id,
                 )
-                    .expect("match arm guarantees a moonshot method id");
-                self.authenticate_moonshot(platform, arguments.method_id.clone())
+                .oauth_platform()
+                {
+                    self.authenticate_oauth_platform(platform, arguments).await
+                } else if let Some(platform) =
+                    auth_method::platform_for_method_id(&arguments.method_id)
+                {
+                    self.authenticate_api_key_platform(
+                        platform,
+                        arguments.method_id.clone(),
+                    )
                     .await
-            }
-            _ => {
+                } else {
                     Err(
                         acp::Error::invalid_params()
                             .data(
@@ -526,6 +579,7 @@ impl acp::Agent for MvpAgent {
                 }
             }
         }
+    }
     async fn new_session(
         &self,
         arguments: acp::NewSessionRequest,

crates/codegen/kigi-shell/src/agent/mvp_agent/agent_ops.rs

@@ -25,20 +25,35 @@ impl MvpAgent {
         primary: &SamplingConfig,
     ) -> Result<(OaiCompatClient, String), acp::Error> {
         let slug = self.resolve_session_summary_model();
-        let session_key = self.auth_manager.current_or_expired().map(|a| a.key.clone());
         let models = self.models_manager.models();
+        // Resolve the aux token by the summary model's OWN platform AND
+        // endpoint: a grok (oauth-platform) summary model draws its pooled grok
+        // token or `None`, and an API-key registry platform draws NOTHING —
+        // NEVER the primary Kimi session token (which `resolve_credentials`
+        // would otherwise stamp onto an api.x.ai / api.deepseek.com request).
+        // The first-party subscription channel still gets the primary
+        // (byte-identical).
+        let authority = self.credential_authority();
+        let session_key = authority.credential_for_slug(&models, None, &slug);
         let endpoints = self.models_manager.endpoints();
         let alpha_test_key = self.cfg.borrow().endpoints.alpha_test_key.clone();
         let config = match crate::agent::config::resolve_aux_model_sampling_config(
             &slug,
             &models,
             &endpoints,
-            session_key.as_deref(),
+            session_key.as_ref(),
             alpha_test_key,
         ) {
             Some(mut cfg) => {
                 cfg.attribution_callback = primary.attribution_callback.clone();
-                cfg.bearer_resolver = primary.bearer_resolver.clone();
+                // The SESSION model's bearer_resolver must not ride to a summary
+                // model on a different provider — `SamplingClient::post`
+                // REPLACES the request's auth header from it, overwriting the
+                // summary model's own resolved key on ITS host. The chokepoint
+                // resolves the resolver from the SUMMARY model's platform +
+                // endpoint instead; the session's is never even read.
+                cfg.bearer_resolver =
+                    self.summary_bearer_resolver(&models, &slug, &cfg.base_url);
                 cfg.max_retries = primary.max_retries;
                 cfg
             }
@@ -52,6 +67,36 @@ impl MvpAgent {
         let client = OaiCompatClient::new(config).map_err(map_sampling_err_to_acp)?;
         Ok((client, model))
     }
+    /// The `bearer_resolver` the SESSION-SUMMARY client may carry — the SAME
+    /// rule the session actor's aux path applies
+    /// ([`crate::session::acp_session::sampler_turn::aux_bearer_resolver_for`]),
+    /// not a second copy of it.
+    ///
+    /// M3 completed: the aux path was gated on the session-token gate and this
+    /// one was not, so an api-key / house-key session whose
+    /// `[model.session-summary]` block carries its OWN `env_key` on the
+    /// session's own coding endpoint had that key REPLACED on the wire by the
+    /// primary bearer on every summary request. Named (rather than inlined
+    /// above) so the gate is reachable from a test — the ungated version stayed
+    /// green because the resolver is consumed by `OaiCompatClient::new`.
+    ///
+    /// Aux slugs are not the session's selection, so `current_key = None`
+    /// (H-b then refuses a collided slug rather than guessing its OAuth twin).
+    pub(super) fn summary_bearer_resolver(
+        &self,
+        models: &indexmap::IndexMap<String, ModelEntry>,
+        slug: &str,
+        base_url: &str,
+    ) -> Option<kigi_sampler::SharedBearerResolver> {
+        let auth_method = self.auth_method_id.load();
+        crate::session::acp_session::sampler_turn::aux_bearer_resolver_for(
+            &self.credential_authority(),
+            auth_method.as_deref(),
+            crate::agent::models::platform_for_slug(models, None, slug),
+            crate::agent::config::resolve_model_auth_facts(slug).byok,
+            base_url,
+        )
+    }
     /// `true` for session-based ACP auth methods.
     fn is_session_based_auth(&self) -> bool {
         self.auth_method_id
@@ -452,18 +497,19 @@ impl MvpAgent {
             )
             .await
     }
-    /// `authenticate(moonshot-cn / moonshot-ai)`: interactive open-platform
-    /// API-key login from the welcome picker.
+    /// `authenticate(<api-key platform id>)`: interactive API-key login from
+    /// the welcome picker for any non-OAuth registry platform.
     ///
     /// Reloads the platform keys from disk+env (the TUI persists the pasted
-    /// key to `[platforms.<id>]` in config.toml immediately before this call),
-    /// fails with an actionable error when none is configured, validates the
-    /// key against `GET {platform_base}/models`, then marks the session
-    /// authenticated exactly like an external API key: publish the method id
-    /// (NOT session-based — no token refresh), swap the freshly-stamped config
-    /// into the models manager, and trigger the model sync so the catalog
-    /// gains the platform's entries. The key itself is never logged.
-    pub(super) async fn authenticate_moonshot(
+    /// key to auth.json under the platform-id scope immediately before this
+    /// call), fails with an actionable error when none is configured,
+    /// validates the key against `GET {platform_base}/models`, then marks the
+    /// session authenticated exactly like an external API key: publish the
+    /// method id (NOT session-based — no token refresh), swap the
+    /// freshly-stamped config into the models manager, and trigger the model
+    /// sync so the catalog gains the platform's entries. The key itself is
+    /// never logged.
+    pub(super) async fn authenticate_api_key_platform(
         &self,
         platform: kigi_models::PlatformId,
         method_id: acp::AuthMethodId,
@@ -480,10 +526,11 @@ impl MvpAgent {
                     Some("platform_key_invalid_or_missing"),
                 );
             })?;
-        // Swap the on-disk config (now carrying the key) into the models
-        // manager so `apply_platform_credentials` stamps the platform's
-        // catalog entries; a parse failure keeps the last-known-good config
-        // (`on_auth_changed` below still re-resolves keys from disk itself).
+        // Rebuild the catalog from the on-disk config: the rebuild freshly
+        // resolves platform keys (env > auth.json > config), so the key just
+        // persisted to auth.json is stamped onto the platform's entries; a
+        // parse failure keeps the last-known-good config (`on_auth_changed`
+        // below still re-resolves keys from disk itself).
         match crate::config::load_effective_config()
             .map_err(|e| e.to_string())
             .and_then(|raw| crate::agent::config::Config::new_from_toml_cfg(&raw))
@@ -512,6 +559,78 @@ impl MvpAgent {
             .and_then(|v| v.as_object().cloned());
         Ok(AuthenticateResponse::new().meta(meta))
     }
+    /// `authenticate(<generic-oauth platform id>)`: interactive device-code
+    /// login for a `uses_oauth` platform carrying an `OAuthConfig` (xai-grok).
+    ///
+    /// Uses a per-provider [`AuthManager`] scoped to the platform's `scope_key`
+    /// (NOT the primary Kimi manager) so the minted session is persisted under
+    /// its own `auth.json` scope, then triggers a catalog re-sync so the
+    /// platform's models appear (their bearer is resolved per-provider at fetch
+    /// / sampling time). The tokens are never logged.
+    pub(super) async fn authenticate_oauth_platform(
+        &self,
+        platform: kigi_models::PlatformId,
+        arguments: acp::AuthenticateRequest,
+    ) -> Result<AuthenticateResponse, acp::Error> {
+        let method_id = arguments.method_id.clone();
+        let oauth = platform
+            .oauth()
+            .expect("oauth_platform() guarantees a device-code OAuthConfig");
+        let auth_meta = AuthRequestMeta::from_json(arguments.meta.as_ref());
+        tracing::info!(
+            method = method_id.0.as_ref(),
+            headless = auth_meta.headless,
+            reauth = auth_meta.reauth,
+            "auth: generic oauth device login",
+        );
+        // M4: the SAME home the pool reads from
+        // (`oauth_registry::pool_home()`), not `kigi_home()` directly —
+        // identical in production, but a login driven from a lib test would
+        // otherwise write into the developer's real `~/.kigi` while every
+        // inference-time lookup read the disposable test pool home.
+        let kigi_home = crate::auth::oauth_registry::pool_home();
+        let auth_manager =
+            std::sync::Arc::new(crate::auth::AuthManager::new_oauth_provider(&kigi_home, oauth));
+        auth_manager.configure_refresher();
+
+        let flow_result = if !auth_meta.headless {
+            let (url_tx, url_rx) = tokio::sync::oneshot::channel();
+            let (code_tx, code_rx) = tokio::sync::mpsc::channel(1);
+            *self.auth_code_tx.borrow_mut() = Some(code_tx);
+            *self.auth_url_rx.borrow_mut() = Some(url_rx);
+            let result = crate::auth::run_oauth_provider_flow(
+                &auth_manager,
+                oauth,
+                auth_meta.reauth,
+                Some(crate::auth::AuthChannels {
+                    url_tx: Some(url_tx),
+                    code_rx,
+                }),
+            )
+            .await;
+            *self.auth_code_tx.borrow_mut() = None;
+            *self.auth_url_rx.borrow_mut() = None;
+            result
+        } else {
+            crate::auth::run_oauth_provider_flow(&auth_manager, oauth, auth_meta.reauth, None).await
+        };
+
+        let (_auth, _did_auth) = flow_result.map_err(|e| {
+            emit_login_span(false, method_id.0.as_ref(), None, Some("login_flow_failed"));
+            let mut err = acp::Error::auth_required();
+            err.message = e.to_string();
+            err
+        })?;
+
+        // Do NOT stamp this token onto the shared sampling_config: it authorizes
+        // ONLY this platform's models (api.x.ai/v1), not the primary session.
+        // The catalog re-sync below resolves it per-provider.
+        self.set_auth_method(method_id.clone());
+        self.models_manager.on_auth_changed().await;
+        emit_login_span(true, method_id.0.as_ref(), None, None);
+        Ok(self.auth_response_with_meta())
+    }
+
     pub(crate) fn deployment_key(&self) -> Option<String> {
         self.cfg.borrow().endpoints.deployment_key.clone()
     }
@@ -591,21 +710,60 @@ impl MvpAgent {
         );
         Ok(entry.clone())
     }
+    /// This agent's credential chokepoint: the EFFECTIVE endpoints (so a
+    /// managed `[endpoints] coding_api_base_url` deployment keeps the session
+    /// bearer — H3) plus the primary manager, which the authority keeps private.
+    pub(crate) fn credential_authority(
+        &self,
+    ) -> crate::auth::credential_authority::CredentialAuthority {
+        crate::auth::credential_authority::CredentialAuthority::new(
+            self.models_manager.endpoints(),
+            Some(self.auth_manager.clone()),
+        )
+    }
+    /// The SESSION credential for `model`, resolved by the model's OWN platform
+    /// AND endpoint at the chokepoint — the single guard against the
+    /// api_key-channel token leak (C1).
+    ///
+    /// A subscription-OAuth model draws from ITS OWN pooled manager (`None`,
+    /// never the Kimi key, when that provider has no stored session). Every
+    /// API-key registry platform and every `[model.*]` block pointed at a
+    /// third-party host gets `None`: `resolve_credentials` would otherwise take
+    /// the `else if let Some(key) = session_key` arm and set
+    /// `api_key = <Kimi bearer>` with the THIRD-PARTY `base_url`, which
+    /// `SamplingClient` builds into `Authorization: Bearer …` — reachable with
+    /// ZERO configuration, since `default_models.json` bundles `moonshot-cn/*`
+    /// and `moonshot-ai/*` entries a Kimi-subscription user sees on first
+    /// launch / offline. The first-party subscription channel uses the primary,
+    /// and only under a session-based auth method — byte-identical.
+    ///
+    /// SECURITY: the resolved token is never logged.
+    fn session_token_for_model(
+        &self,
+        model: &ModelEntry,
+    ) -> Option<crate::auth::credential_authority::SessionCredential> {
+        let is_primary_channel = crate::auth::credential_authority::entry_platform(model)
+            .is_none_or(|platform| platform.oauth().is_none());
+        if is_primary_channel && !self.is_session_based_auth() {
+            return None;
+        }
+        self.credential_authority().credential_for_model(model)
+    }
     pub(crate) fn prepare_sampling_config_for_model(
         &self,
         model: &ModelEntry,
         origin_client: Option<crate::http::OriginClientInfo>,
     ) -> SamplingConfig {
-        let session = if self.is_session_based_auth() {
-            self.auth_manager.current_or_expired()
-        } else {
-            None
-        };
+        // Resolve the session token by the MODEL's platform, not the primary
+        // auth method: an oauth-platform model (xai-grok) uses its OWN
+        // pool-backed token (`None` — never the Kimi key — when the user has not
+        // logged into that provider), closing the api_key-channel leak where the
+        // primary Kimi session token was stamped onto a grok request. A
+        // first-party / Kimi model is unchanged. GUARANTEE: a grok model's
+        // api_key is its own grok token or `None`, never the primary Kimi key.
+        let session = self.session_token_for_model(model);
         let has_session_key = session.is_some();
-        let mut credentials = resolve_credentials(
-            model,
-            session.as_ref().map(|a| a.key.as_str()),
-        );
+        let mut credentials = resolve_credentials(model, session.as_ref());
         if !has_session_key && credentials.auth_type == kigi_chat_state::AuthType::ApiKey
             && !model.has_own_credentials() && self.is_session_based_auth()
         {
@@ -792,7 +950,8 @@ impl MvpAgent {
         models_manager: crate::agent::models::ModelsManager,
     ) -> Self {
         models_manager.set_gateway(gateway.clone());
-        let sampling_config = models_manager.sampling_config();
+        // H-a: the config AND the platform it was built from, from ONE call.
+        let baseline = models_manager.sampling_config();
         let storage_mode = cfg.storage_mode;
         let default_yolo_mode = cfg.default_yolo_mode;
         let default_auto_mode = cfg.default_auto_mode;
@@ -847,7 +1006,8 @@ impl MvpAgent {
             models_manager,
             cfg: RefCell::new(cfg.clone()),
             auth_method_id: crate::agent::auth_method::new_shared_auth_method_id(None),
-            sampling_config: RefCell::new(sampling_config),
+            sampling_config: RefCell::new(baseline.config),
+            sampling_config_platform: std::cell::Cell::new(baseline.platform),
             auth_manager,
             auth_code_tx: RefCell::new(None),
             auth_url_rx: RefCell::new(None),
@@ -1437,32 +1597,128 @@ impl MvpAgent {
         );
         (serde_json::json!({ "options" : config_options }), serde_json::json!(detail))
     }
+    /// The registry platform the SHARED `sampling_config` routes to: the one
+    /// captured WITH the config when it was built, never a fresh lookup.
+    ///
+    /// H-a: `ModelsManager::sampling_config` builds the shared config from the
+    /// catalog entry `current_model_id()` named AT THAT MOMENT, and the config
+    /// is never rebuilt. A guard that re-resolves `config.model` (the BARE
+    /// routing slug) against the LIVE cell therefore answers about a DIFFERENT
+    /// entry the moment the two drift — and they drift on every non-Leader model
+    /// switch (`handlers/model_switch.rs`) and on catalog reselection.
+    /// `entry_for_slug_resolution`'s `entry.info.model == slug` test then fails
+    /// and the resolution falls through to `resolve_catalog_key`'s `.rev()`
+    /// scan; `PlatformId::ALL` lists `kimi-code` first and its API-key twin
+    /// `kimi-coding` 19th, so the LAST match is the twin, which takes no session
+    /// credential. For a Kimi-subscription user who also has `KIMI_API_KEY` set,
+    /// a post-expiry `kigi login` then found no governing manager and — because
+    /// the stamp only overwrites on success — left the EXPIRED bearer in place:
+    /// every unresolved-model fallback and every subagent baseline turn 401'd
+    /// until restart.
+    fn shared_config_platform(&self) -> Option<kigi_models::PlatformId> {
+        self.sampling_config_platform.get()
+    }
+    /// H2/C1: stamp the shared `sampling_config`'s OWN governing session
+    /// credential — whatever the authority says that is for this config's model
+    /// and endpoint.
+    ///
+    /// The shared config is not inert: `resolve_sampling_config_for_model`
+    /// returns it verbatim whenever a model id fails to resolve, and
+    /// `SubagentSpawnContext` clones it as every subagent's baseline — so an
+    /// `api_key` stamped here reaches the wire against whatever `base_url` the
+    /// config carries. The login/seed sites used to stamp it unconditionally,
+    /// exactly the mistake `authenticate_oauth_platform` already documents ("Do
+    /// NOT stamp this token onto the shared sampling_config").
+    ///
+    /// C1: this function does NOT take a credential. The previous shape took
+    /// `key: String` — always the primary Kimi bearer — and guarded it with a
+    /// predicate asking whether **a** session credential may ride. For a
+    /// subscription-OAuth platform at its own host that is correctly `true`, but
+    /// the credential that may ride there is that platform's POOLED token: a
+    /// Claude Pro/Max user running `kigi login` stamped the Kimi subscription
+    /// bearer onto a config routed at `api.anthropic.com`. Asking
+    /// `credential_for` instead makes the question and the credential the same
+    /// object, so the pairing cannot be wrong — and there is no primary handle
+    /// here to hand-carry (M2).
+    ///
+    /// `overwrite = false` keeps the historical "only if missing" seeding
+    /// behaviour; the login handlers pass `true` because a fresh login must
+    /// replace a stale bearer, and they call this AFTER the manager holds the
+    /// new token so it is read back through the authority.
+    ///
+    /// SECURITY: the token is never logged.
+    pub(super) fn stamp_session_credential(&self, overwrite: bool) -> bool {
+        let (model, base_url) = {
+            let sampling_config = self.sampling_config.borrow();
+            if !overwrite && sampling_config.api_key.is_some() {
+                return false;
+            }
+            (
+                sampling_config.model.clone(),
+                sampling_config.base_url.clone(),
+            )
+        };
+        let platform = self.shared_config_platform();
+        let Some(credential) = self.credential_authority().credential_for(platform, &base_url)
+        else {
+            tracing::debug!(
+                model = model.as_str(),
+                "auth: no session credential governs the shared sampling config \
+                 (its model + endpoint take none, or the governing provider has no session)"
+            );
+            return false;
+        };
+        self.sampling_config.borrow_mut().api_key = Some(credential.expose().to_owned());
+        true
+    }
+    /// Whether the shared `sampling_config` may receive a credential that
+    /// authorizes only the SESSION's own first-party endpoint — the house
+    /// `KIGI_API_KEY` the `xai.api_key` login handler reads from the
+    /// environment, which the authority does not own and so cannot produce.
+    ///
+    /// The house key rides the [`CredentialClass::Primary`] channel and no
+    /// other. `Pooled` is deliberately excluded: a subscription-OAuth platform's
+    /// own host DOES take a session credential, but that credential is the
+    /// platform's pooled token, where the house key has no business (C1).
+    pub(super) fn shared_config_takes_house_key(&self) -> bool {
+        let base_url = self.sampling_config.borrow().base_url.clone();
+        matches!(
+            self.credential_authority()
+                .credential_class(self.shared_config_platform(), &base_url),
+            crate::auth::credential_authority::CredentialClass::Primary
+        )
+    }
     /// Seed the global sampling config with login auth when available.
     ///
-    /// Only sets the `api_key` if missing. Does NOT resolve `base_url` from
+    /// Only sets the `api_key` if missing, and only with the credential that
+    /// governs the config's own model + endpoint (see
+    /// [`Self::stamp_session_credential`]). Does NOT resolve `base_url` from
     /// `current_model_id` — that's deferred to session creation time to avoid
     /// cross-client contamination in leader mode (where `current_model_id` is
     /// shared mutable state).
     pub(super) fn seed_client_config_auth_if_available(&self) {
-        let mut sampling_config = self.sampling_config.borrow_mut();
-        if sampling_config.api_key.is_none() {
-            if let Some(auth) = self.auth_manager.current_or_expired() {
-                sampling_config.api_key = Some(auth.key);
+        if self.sampling_config.borrow().api_key.is_some() {
+            return;
+        }
+        if self.stamp_session_credential(false) {
             tracing::debug!("auth: seed_client_config set auth (SessionToken)");
             kigi_log::unified_log::debug(
                 "auth: seed_client_config set auth (SessionToken)",
                 None,
                 None,
             );
-            } else if !self
+            return;
+        }
+        // No credential was stamped. Only the total-absence case is worth
+        // warning about: a withheld-by-endpoint seed is the rule working.
+        if self.auth_manager.current_or_expired().is_none()
+            && !self
                 .models_manager
                 .models()
                 .values()
                 .any(|m| m.has_own_credentials())
         {
-                tracing::warn!(
-                    "No credentials found: no login token and no model api_key/env_key"
-                );
+            tracing::warn!("No credentials found: no login token and no model api_key/env_key");
             kigi_log::unified_log::warn(
                 "No credentials found: no login token and no model api_key/env_key",
                 None,
@@ -1470,7 +1726,6 @@ impl MvpAgent {
             );
         }
     }
-    }
     /// Allocate the next monotonic telemetry turn number for a session.
     ///
     /// Returns the current turn number and advances the counter. The counter is
@@ -2190,12 +2445,34 @@ impl MvpAgent {
         }
         let (mut handle, agent_system_prompt, session_thread) = {
             let _timer = crate::instrumentation_timer!("session.spawn_actor_call");
-            let session_key = self.auth_manager.current_or_expired().map(|a| a.key);
+            // Classify the credential's auth_type against the bearer this
+            // model's endpoint would ACTUALLY receive, not the raw primary:
+            // an API-key-platform model has no session credential at all, and
+            // reading the primary here reported one.
+            //
+            // H-a: the platform is resolved with the SAME catalog key the
+            // actor about to be spawned seeds itself with
+            // (`selected_catalog_key_for_spawn`, spawn.rs), so this
+            // classification and every per-turn decision that session makes
+            // agree by construction instead of re-resolving the bare slug.
+            let models_for_key = self.models_manager.models();
+            let session_key = self.credential_authority().credential_for(
+                crate::agent::models::platform_for_slug(
+                    &models_for_key,
+                    crate::agent::models::selected_catalog_key_for_spawn(
+                        &models_for_key,
+                        &session_model_id,
+                    )
+                    .as_deref(),
+                    sampling_config.model.as_str(),
+                ),
+                &sampling_config.base_url,
+            );
             let credentials = kigi_chat_state::Credentials {
                 api_key: sampling_config.api_key.clone(),
                 auth_type: crate::agent::config::resolve_chat_state_auth_type(
                     sampling_config.model.as_str(),
-                    session_key.as_deref(),
+                    session_key.as_ref(),
                     self.auth_type(),
                 ),
                 alpha_test_key: self.alpha_test_key(),

crates/codegen/kigi-shell/src/agent/mvp_agent/mod.rs

@@ -501,6 +501,19 @@ pub struct MvpAgent {
     /// only api_key is written here (same for all clients). Per-session base_url
     /// is resolved at session creation time in `new_session` / `load_session`.
     pub(crate) sampling_config: RefCell<SamplingConfig>,
+    /// The registry platform [`Self::sampling_config`] was BUILT from, captured
+    /// in the same `ModelsManager::sampling_config()` call that produced it.
+    ///
+    /// H-a: the shared config is built ONCE (`Self::with_models`) and never
+    /// rebuilt, but `ModelsManager::current_model_id()` moves on every
+    /// non-Leader model switch and on catalog reselection. Its guards
+    /// (`stamp_session_credential`, `shared_config_takes_house_key`)
+    /// therefore read THIS cell, never the live one: after a switch the live
+    /// cell names a different entry, and re-resolving the config's bare slug
+    /// against it fell through to `resolve_catalog_key`'s `.rev()` scan —
+    /// answering the API-key twin, which takes no session credential, so a
+    /// successful `kigi login` silently failed to replace the expired bearer.
+    pub(crate) sampling_config_platform: std::cell::Cell<Option<kigi_models::PlatformId>>,
     pub(crate) auth_manager: Arc<AuthManager>,
     pub(crate) models_manager: crate::agent::models::ModelsManager,
     /// Forwards pasted codes from `handle_auth_submit_code` to the auth flow.

crates/codegen/kigi-shell/src/agent/mvp_agent/tests.rs

@@ -1158,6 +1158,80 @@ fn build_agent_with_auth(auth: crate::auth::KimiAuth) -> MvpAgent {
     let cfg = AgentConfig::default();
     MvpAgent::new(gateway, &cfg, auth_manager, None).expect("valid test config")
 }
+/// Regression (token-leak, Facet B via the api_key channel): under a
+/// session-based (Kimi) primary auth method, `prepare_sampling_config_for_model`
+/// must resolve the session token by the MODEL's platform — so a grok
+/// (oauth-platform) model NEVER carries the primary Kimi session key as its
+/// `api_key`, while a first-party Kimi model still does. Relies on the
+/// process-global OAuth pool (there is no per-session snapshot).
+///
+/// Reverting the fix (resolving the session token from the primary regardless of
+/// the model's platform) fails the grok assertion below — it would stamp the
+/// live Kimi key on a request bound for api.x.ai.
+#[tokio::test]
+#[serial_test::serial]
+async fn prepare_sampling_config_never_stamps_kimi_key_on_grok_model() {
+    use crate::agent::auth_method::{
+        CACHED_TOKEN_AUTH_METHOD_ID, HOUSE_API_KEY_ENV_VAR, LEGACY_XAI_API_KEY_ENV_VAR,
+        XAI_API_KEY_ENV_VAR,
+    };
+    use crate::agent::config::{EndpointsConfig, ModelEntry};
+    use kigi_test_support::EnvGuard;
+
+    const KIMI_KEY: &str = "kimi-session-secret-DO-NOT-LEAK";
+
+    // No ambient BYOK env key: a grok model with no stored oauth session then
+    // resolves to no api_key at all, rather than a global-key fallback that could
+    // mask the leak under test.
+    let _house = EnvGuard::unset(HOUSE_API_KEY_ENV_VAR);
+    let _xai = EnvGuard::unset(XAI_API_KEY_ENV_VAR);
+    let _legacy = EnvGuard::unset(LEGACY_XAI_API_KEY_ENV_VAR);
+
+    // Primary: a live Kimi session token under a session-based auth method.
+    let agent = build_agent_with_auth(crate::auth::KimiAuth {
+        key: KIMI_KEY.to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    agent.set_auth_method(acp::AuthMethodId::new(CACHED_TOKEN_AUTH_METHOD_ID));
+
+    let endpoints = EndpointsConfig::default();
+
+    // First-party SUBSCRIPTION model (kimi-code): the primary session key IS its
+    // api_key — the byte-identical primary path, and proof the Kimi token is
+    // live (so it WOULD leak if mis-routed onto a grok request). This assertion
+    // also confirms the session-based primary path is active.
+    //
+    // This used to use `moonshot-cn/kimi-k2-0905-preview` and assert the SAME
+    // thing, which encoded the C1 defect: moonshot-cn is an API-key registry
+    // platform on `api.moonshot.cn`, NOT first-party, so "must carry the primary
+    // session key" was asserting the leak. `api_key_channel_leak_tests` now pins
+    // the opposite for every moonshot entry.
+    let mut kimi_model = ModelEntry::fallback("kimi-for-coding", &endpoints);
+    kimi_model.info.id = Some("kimi-code/kimi-for-coding".to_string());
+    kimi_model.info.base_url = kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url.to_string();
+    assert!(!kimi_model.has_own_credentials());
+    let kimi_cfg = agent.prepare_sampling_config_for_model(&kimi_model, None);
+    assert_eq!(
+        kimi_cfg.api_key.as_deref(),
+        Some(KIMI_KEY),
+        "the first-party subscription model must carry the primary session key \
+         (primary path unchanged)"
+    );
+
+    // xai-grok model (oauth platform): the session token resolves from its OWN
+    // pool-backed manager, INDEPENDENT of the Kimi primary — so its api_key can
+    // NEVER be the Kimi session key.
+    let mut grok_model = ModelEntry::fallback("grok-4-latest", &endpoints);
+    grok_model.info.id = Some("xai-grok/grok-4-latest".to_string());
+    assert!(!grok_model.has_own_credentials());
+    let grok_cfg = agent.prepare_sampling_config_for_model(&grok_model, None);
+    assert_ne!(
+        grok_cfg.api_key.as_deref(),
+        Some(KIMI_KEY),
+        "LEAK: a grok model must never carry the primary Kimi session key as api_key"
+    );
+}
 /// Regression: boot-time plugin discovery is deferred past ACP
 /// `initialize`, so the shared plugin registry starts empty.
 /// `resolve_mcp_servers` reads that snapshot to merge plugin-contributed
@@ -1214,6 +1288,10 @@ async fn ensure_plugin_registry_lazily_populates_snapshot() {
     );
 }
 mod subagent_spawn_context_tests;
+/// LEAK guard for the `api_key` channel (C1/C2), through the real
+/// `prepare_sampling_config_for_model` resolution path.
+mod api_key_channel_leak_tests;
+mod chokepoint_leak_tests;
 /// No load in flight and no session → the wait returns immediately
 /// (the caller then surfaces "unknown session id" exactly as before).
 #[tokio::test]
@@ -1904,10 +1982,12 @@ async fn cached_token_fallthrough_prefers_api_key_for_deployment_key() {
 #[serial_test::serial]
 async fn cached_token_fallthrough_falls_to_kigi_com_without_credentials() {
     use crate::agent::auth_method::{
-        KIMI_CODE_METHOD_ID, LEGACY_XAI_API_KEY_ENV_VAR, XAI_API_KEY_ENV_VAR,
+        HOUSE_API_KEY_ENV_VAR, KIMI_CODE_METHOD_ID, LEGACY_XAI_API_KEY_ENV_VAR,
+        XAI_API_KEY_ENV_VAR,
     };
     use kigi_test_support::EnvGuard;
     let _lockdown = EnvGuard::unset("KIGI_DISABLE_API_KEY_AUTH");
+    let _house = EnvGuard::unset(HOUSE_API_KEY_ENV_VAR);
     let _new = EnvGuard::unset(XAI_API_KEY_ENV_VAR);
     let _legacy = EnvGuard::unset(LEGACY_XAI_API_KEY_ENV_VAR);
     let agent = build_minimal_agent_for_tests();

crates/codegen/kigi-shell/src/agent/mvp_agent/tests/api_key_channel_leak_tests.rs

@@ -0,0 +1,314 @@
+//! LEAK GUARD (`api_key` channel) — C1/C2, driven through the REAL resolution
+//! path, `MvpAgent::prepare_sampling_config_for_model`.
+//!
+//! This is the channel the `bearer_resolver` guard does NOT close, and the one
+//! the first round of leak tests assumed away by hand-stamping a provider key
+//! into chat state. The chain: `session_token_for_model` used to fall through to
+//! `self.auth_manager.current_or_expired()` (the primary Kimi bearer) for every
+//! non-OAuth model, `resolve_credentials` then took its
+//! `else if let Some(key) = session_key` arm and set `api_key = <Kimi token>`
+//! with the THIRD-PARTY `base_url`, and `SamplingClient` builds
+//! `Authorization: Bearer <api_key>` straight into `default_headers` — which
+//! `post()` only overrides when a resolver exists, so `bearer_resolver: None`
+//! does not save it.
+//!
+//! Nothing here stamps a credential by hand: every assertion reads what the
+//! resolution path actually produced.
+
+use super::super::*;
+use crate::agent::auth_method::{
+    CACHED_TOKEN_AUTH_METHOD_ID, HOUSE_API_KEY_ENV_VAR, LEGACY_XAI_API_KEY_ENV_VAR,
+    XAI_API_KEY_ENV_VAR,
+};
+use crate::agent::config::{Config as AgentConfig, EndpointsConfig, EnvKeys, ModelEntry};
+use crate::auth::{AuthManager, AuthMode, KimiAuth, KimiCodeConfig};
+use kigi_test_support::EnvGuard;
+
+pub(super) const KIMI_TOKEN: &str = "kimi-subscription-token-DO-NOT-LEAK";
+
+/// Ambient BYOK env vars unset, so a model with no resolvable credential ends up
+/// with `api_key == None` rather than a global-key fallback that could mask the
+/// leak under test. Every test holding these must be `#[serial]`.
+pub(super) fn without_ambient_byok_env() -> [EnvGuard; 3] {
+    [
+        EnvGuard::unset(HOUSE_API_KEY_ENV_VAR),
+        EnvGuard::unset(XAI_API_KEY_ENV_VAR),
+        EnvGuard::unset(LEGACY_XAI_API_KEY_ENV_VAR),
+    ]
+}
+
+/// An `MvpAgent` on a session-based (`cached_token`) ACP method holding a live
+/// Kimi subscription bearer — the mainstream configuration in which the leak
+/// fires. `(tempdir, agent)`; the tempdir is the auth store and is returned so
+/// the caller keeps it alive.
+pub(super) fn kimi_session_agent() -> (tempfile::TempDir, MvpAgent) {
+    let dir = tempfile::tempdir().expect("tempdir");
+    let auth_manager = std::sync::Arc::new(AuthManager::new(dir.path(), KimiCodeConfig::default()));
+    auth_manager.hot_swap(KimiAuth {
+        key: KIMI_TOKEN.to_string(),
+        auth_mode: AuthMode::OAuth,
+        refresh_token: Some("rt".into()),
+        expires_at: Some(chrono::Utc::now() + chrono::Duration::hours(1)),
+        ..KimiAuth::test_default()
+    });
+    let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
+    let agent = MvpAgent::new(
+        GatewaySender::new(tx),
+        &AgentConfig::default(),
+        auth_manager,
+        None,
+    )
+    .expect("valid test config");
+    agent.set_auth_method(acp::AuthMethodId::new(CACHED_TOKEN_AUTH_METHOD_ID));
+    (dir, agent)
+}
+
+/// A catalog entry as `resolve_model_list` builds one for a fetched registry
+/// model: managed catalog key, platform base URL, no credential of its own.
+pub(super) fn platform_entry(catalog_key: &str, slug: &str, base_url: &str) -> ModelEntry {
+    let mut entry = ModelEntry::fallback(slug, &EndpointsConfig::default());
+    entry.info.id = Some(catalog_key.to_string());
+    entry.info.base_url = base_url.to_string();
+    entry
+}
+
+/// C1, the ZERO-CONFIGURATION repro. `default_models.json` bundles
+/// `moonshot-cn/*` and `moonshot-ai/*` entries with `api_key: None`, and
+/// `resolve_model_list` keeps the bundled defaults whenever no catalog fetch has
+/// succeeded — so on first launch / offline a Kimi-subscription user sees them
+/// in the picker with no configuration whatsoever. Selecting one used to send
+/// `Authorization: Bearer <Kimi OAuth token>` to `api.moonshot.cn`, which is NOT
+/// first-party.
+///
+/// Revert-to-red: make `CredentialAuthority::governing_manager`'s
+/// `Some(platform) => None` arm return `self.primary.clone()` and every `api_key`
+/// below becomes `Some(KIMI_TOKEN)`.
+#[tokio::test]
+#[serial_test::serial]
+async fn bundled_default_moonshot_models_never_carry_the_kimi_bearer() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    let bundled = crate::agent::config::default_model_entries(&EndpointsConfig::default());
+    let moonshot: Vec<_> = bundled
+        .iter()
+        .filter(|(key, _)| key.starts_with("moonshot-cn/") || key.starts_with("moonshot-ai/"))
+        .collect();
+    assert_eq!(
+        moonshot.len(),
+        4,
+        "default_models.json still bundles the four moonshot open-platform entries"
+    );
+
+    for (key, entry) in moonshot {
+        assert!(
+            !entry.has_own_credentials(),
+            "{key}: the bundled entry carries no credential of its own"
+        );
+        assert!(
+            !crate::util::is_effective_coding_endpoint_url(&entry.info().base_url),
+            "{key}: routes to a third-party host ({})",
+            entry.info().base_url
+        );
+        let cfg = agent.prepare_sampling_config_for_model(entry, None);
+        assert_ne!(
+            cfg.api_key.as_deref(),
+            Some(KIMI_TOKEN),
+            "LEAK: selecting the bundled {key} sent the Kimi subscription bearer to {}",
+            entry.info().base_url
+        );
+        assert_eq!(
+            cfg.base_url,
+            entry.info().base_url,
+            "{key}: still routes to its own host (the fix must not reroute traffic)"
+        );
+    }
+}
+
+/// C1 across the API-key registry platform shapes a fetched catalog produces.
+#[tokio::test]
+#[serial_test::serial]
+async fn api_key_platform_models_never_carry_the_kimi_bearer_as_api_key() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    for (catalog_key, slug, base_url) in [
+        ("deepseek/deepseek-chat", "deepseek-chat", "https://api.deepseek.com/v1"),
+        ("openai/gpt-5.2", "gpt-5.2", "https://api.openai.com/v1"),
+        ("anthropic/claude-opus-4-8", "claude-opus-4-8", "https://api.anthropic.com/v1"),
+        ("groq/llama-4", "llama-4", "https://api.groq.com/openai/v1"),
+        ("xai/grok-4.5", "grok-4.5", "https://api.x.ai/v1"),
+    ] {
+        let entry = platform_entry(catalog_key, slug, base_url);
+        let cfg = agent.prepare_sampling_config_for_model(&entry, None);
+        assert_ne!(
+            cfg.api_key.as_deref(),
+            Some(KIMI_TOKEN),
+            "LEAK: {catalog_key} carried the primary Kimi session bearer to {base_url}"
+        );
+    }
+}
+
+/// C2, the `[model.*]` repro. A `[model.gpt-4o]` block has `info.id == None`, so
+/// it has no platform at all — which used to be a blanket allow. BYOK is
+/// `has_own_credentials()`, which probes `std::env::var` AT CALL TIME, so an
+/// unset (or mistyped) `env_key` classifies the model NotByok and the Kimi
+/// bearer went to `api.openai.com` on BOTH channels.
+///
+/// Revert-to-red: make `CredentialAuthority::is_session_coding_endpoint` return
+/// `true` unconditionally and `api_key` here becomes `Some(KIMI_TOKEN)`.
+#[tokio::test]
+#[serial_test::serial]
+async fn config_model_with_an_unset_env_key_never_carries_the_kimi_bearer() {
+    let _env = without_ambient_byok_env();
+    let _typo = EnvGuard::unset("OPENAI_API_KEY_TYPO");
+    let (_dir, agent) = kimi_session_agent();
+
+    let mut entry = ModelEntry::fallback("gpt-4o", &EndpointsConfig::default());
+    entry.info.id = None; // a `[model.gpt-4o]` config block
+    entry.info.base_url = "https://api.openai.com/v1".to_string();
+    entry.env_key = Some(EnvKeys::single("OPENAI_API_KEY_TYPO"));
+    assert!(
+        !entry.has_own_credentials(),
+        "the env var is unset, so this classifies NotByok — the precondition of the defect"
+    );
+
+    let cfg = agent.prepare_sampling_config_for_model(&entry, None);
+    assert_ne!(
+        cfg.api_key.as_deref(),
+        Some(KIMI_TOKEN),
+        "LEAK: a [model.*] block with an unset env_key sent the Kimi bearer to api.openai.com"
+    );
+    assert_eq!(cfg.api_key, None, "no credential resolves — fail fast");
+}
+
+/// The first-party subscription channel must stay BYTE-IDENTICAL: `kimi-code/*`
+/// (and a `[model.*]` block on the session's own coding endpoint, including a
+/// `KIGI_CODE_BASE_URL` deployment / a local dev proxy) still carries the
+/// primary session key. This assertion is also what proves the Kimi token is
+/// live in the tests above — it WOULD leak if the guard were missing.
+#[tokio::test]
+#[serial_test::serial]
+async fn the_first_party_subscription_channel_still_carries_the_session_key() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    let kimi = platform_entry(
+        "kimi-code/kimi-for-coding",
+        "kimi-for-coding",
+        kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+    );
+    assert_eq!(
+        agent
+            .prepare_sampling_config_for_model(&kimi, None)
+            .api_key
+            .as_deref(),
+        Some(KIMI_TOKEN),
+        "the kimi-code subscription channel must be unchanged"
+    );
+
+    for base_url in [
+        kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+        "http://127.0.0.1:4141/v1",
+        "http://localhost:8080/v1",
+    ] {
+        let mut bare = ModelEntry::fallback("kigi-4.5", &EndpointsConfig::default());
+        bare.info.id = None;
+        bare.info.base_url = base_url.to_string();
+        assert_eq!(
+            agent
+                .prepare_sampling_config_for_model(&bare, None)
+                .api_key
+                .as_deref(),
+            Some(KIMI_TOKEN),
+            "{base_url}: a custom deployment / local proxy keeps the session key"
+        );
+    }
+}
+
+/// A subscription-OAuth model draws its `api_key` from ITS OWN pooled manager,
+/// never the Kimi primary — and never falls back to it when that provider has no
+/// stored session (the pool home is an empty TempDir under `cfg(test)`).
+#[tokio::test]
+#[serial_test::serial]
+async fn oauth_platform_models_never_carry_the_kimi_bearer_as_api_key() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    for (catalog_key, slug, base_url) in [
+        ("xai-grok/grok-4-latest", "grok-4-latest", "https://api.x.ai/v1"),
+        (
+            "claude-pro-max/claude-opus-4-8",
+            "claude-opus-4-8",
+            "https://api.anthropic.com/v1",
+        ),
+        ("github-copilot/gpt-4.1", "gpt-4.1", "https://api.githubcopilot.com"),
+        (
+            "openai-codex/gpt-5.5",
+            "gpt-5.5",
+            "https://chatgpt.com/backend-api/codex",
+        ),
+    ] {
+        let entry = platform_entry(catalog_key, slug, base_url);
+        let cfg = agent.prepare_sampling_config_for_model(&entry, None);
+        assert_ne!(
+            cfg.api_key.as_deref(),
+            Some(KIMI_TOKEN),
+            "LEAK: {catalog_key} carried the primary Kimi session bearer to {base_url}"
+        );
+    }
+}
+
+/// H5 at the api_key channel: `resolve_model_id` (the picker's own lookup) must
+/// hand `prepare_sampling_config_for_model` the entry the user SELECTED, even
+/// when an API-key platform and its subscription-OAuth twin list the same
+/// routing slug in `PlatformId::ALL` order. Selecting the OAuth twin by catalog
+/// key must not resolve the API-key twin — and vice versa.
+#[tokio::test]
+#[serial_test::serial]
+async fn dual_credential_slug_collision_resolves_the_selected_catalog_key() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    // API-key platform FIRST, exactly as `PlatformId::ALL` orders them.
+    for key in ["xai/grok-4.5", "xai-grok/grok-4.5"] {
+        agent.models_manager.insert_test_entry(
+            key,
+            platform_entry(key, "grok-4.5", "https://api.x.ai/v1"),
+        );
+    }
+
+    for key in ["xai/grok-4.5", "xai-grok/grok-4.5"] {
+        let resolved = agent
+            .resolve_model_id(&acp::ModelId::new(key))
+            .expect("both twins resolve");
+        assert_eq!(
+            resolved.info().id.as_deref(),
+            Some(key),
+            "selecting {key} must resolve THAT catalog entry, not its slug twin"
+        );
+    }
+
+    // And the bare slug resolves the same entry the picker's `resolve_catalog_key`
+    // does — one direction, one answer (the auth layer used to first-match).
+    let by_slug = agent
+        .resolve_model_id(&acp::ModelId::new("grok-4.5"))
+        .expect("the bare slug resolves");
+    let models = agent.models_manager.models();
+    let picker_key = crate::agent::models::resolve_catalog_key(
+        &models,
+        &acp::ModelId::new("grok-4.5"),
+    )
+    .expect("the picker resolves the bare slug");
+    assert_eq!(
+        by_slug.info().id.as_deref(),
+        Some(picker_key.0.as_ref()),
+        "the auth layer and the picker must resolve the SAME entry for one slug"
+    );
+    assert_eq!(
+        crate::agent::config::find_model_by_id(&models, "grok-4.5")
+            .and_then(|e| e.info().id.as_deref()),
+        Some(picker_key.0.as_ref()),
+        "find_model_by_id must agree with resolve_catalog_key by construction"
+    );
+}

crates/codegen/kigi-shell/src/agent/mvp_agent/tests/chokepoint_leak_tests.rs

@@ -0,0 +1,589 @@
+//! LEAK GUARD (the chokepoint's own call sites) — H1, H2 and the H3 regression.
+//!
+//! The companion of `api_key_channel_leak_tests`, which drives
+//! `MvpAgent::prepare_sampling_config_for_model`. These three pin the OTHER
+//! producers of an outgoing credential: `ModelsManager::sampling_config()` (the
+//! agent-wide baseline), the shared `MvpAgent::sampling_config` the login/seed
+//! paths stamp, and the SESSION's effective coding endpoint as configured from
+//! config.toml rather than the environment.
+//!
+//! Every assertion reads what the real resolution path produced; nothing is
+//! hand-stamped.
+
+use super::super::*;
+use super::api_key_channel_leak_tests::{
+    KIMI_TOKEN, kimi_session_agent, platform_entry, without_ambient_byok_env,
+};
+use crate::agent::auth_method::CACHED_TOKEN_AUTH_METHOD_ID;
+use crate::agent::config::{Config as AgentConfig, EndpointsConfig, ModelEntry};
+use crate::auth::credential_authority::CredentialClass;
+use crate::auth::{AuthManager, AuthMode, KimiAuth, KimiCodeConfig};
+use kigi_sampler::BearerResolver;
+use kigi_test_support::EnvGuard;
+
+/// Re-seed the shared config the way `MvpAgent::with_models` does: the config
+/// AND the platform it was BUILT from, from ONE `ModelsManager::sampling_config()`
+/// against whatever `current_model_id` names right now. A test can therefore
+/// never set one without the other — which is the whole point of H-a.
+fn rebuild_shared_config(agent: &MvpAgent) {
+    let baseline = agent.models_manager.sampling_config();
+    agent.sampling_config_platform.set(baseline.platform);
+    *agent.sampling_config.borrow_mut() = baseline.config;
+}
+
+/// H1 — `ModelsManager::sampling_config()`, the OTHER `api_key` producer, and a
+/// byte-for-byte repeat of the round-1 defect: it resolved the session bearer
+/// itself (`platform.oauth()`, else `auth_manager.current_or_expired()`), so
+/// every non-OAuth platform got the primary Kimi token.
+///
+/// This config is not incidental — it is the `MvpAgent` baseline
+/// (`Self::with_models`), which `resolve_sampling_config_for_model` returns
+/// verbatim for an unresolved model id and `SubagentSpawnContext` clones as
+/// every subagent's baseline, so its `api_key` reaches the wire against its own
+/// `base_url`. Zero-config repro: a Kimi subscription + a bundled
+/// `moonshot-cn/*` default.
+///
+/// Revert-to-red (L: this edit COMPILES — the previous wording named a
+/// `Option<String>` argument that the `Option<&SessionCredential>` signature
+/// rejects, so it could never have been run): in
+/// `ModelsManager::sampling_config`, ask the authority about the SESSION's
+/// endpoint instead of the current model's own —
+/// `.credential_for(None, &config.endpoints.proxy_url())` in place of
+/// `.credential_for_model(current_model)`. That is the round-1 defect's shape
+/// (the credential decided by something other than the model's own platform +
+/// endpoint) and every `assert_ne!` below sees `Some(KIMI_TOKEN)`.
+#[tokio::test]
+#[serial_test::serial]
+async fn models_manager_sampling_config_never_carries_the_kimi_bearer() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    for (catalog_key, slug, base_url) in [
+        (
+            "moonshot-cn/kimi-k2-turbo-preview",
+            "kimi-k2-turbo-preview",
+            "https://api.moonshot.cn/v1",
+        ),
+        (
+            "deepseek/deepseek-chat",
+            "deepseek-chat",
+            "https://api.deepseek.com/v1",
+        ),
+        ("openai/gpt-5.2", "gpt-5.2", "https://api.openai.com/v1"),
+    ] {
+        agent
+            .models_manager
+            .insert_test_entry(catalog_key, platform_entry(catalog_key, slug, base_url));
+        agent
+            .models_manager
+            .set_current_model_id(acp::ModelId::new(catalog_key));
+        let cfg = agent.models_manager.sampling_config().config;
+        assert_eq!(cfg.base_url, base_url, "{catalog_key}: routed to its own host");
+        assert_ne!(
+            cfg.api_key.as_deref(),
+            Some(KIMI_TOKEN),
+            "LEAK: ModelsManager::sampling_config sent the Kimi bearer to {base_url}"
+        );
+    }
+
+    // …and the first-party subscription channel is unchanged, which is what
+    // proves the Kimi bearer was reachable above.
+    let kimi_key = "kimi-code/kimi-for-coding";
+    agent.models_manager.insert_test_entry(
+        kimi_key,
+        platform_entry(
+            kimi_key,
+            "kimi-for-coding",
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+        ),
+    );
+    agent
+        .models_manager
+        .set_current_model_id(acp::ModelId::new(kimi_key));
+    assert_eq!(
+        agent
+            .models_manager
+            .sampling_config()
+            .config
+            .api_key
+            .as_deref(),
+        Some(KIMI_TOKEN),
+        "the kimi-code subscription channel must be byte-identical"
+    );
+}
+
+/// H2 — the SHARED `MvpAgent::sampling_config`. `seed_client_config_auth_if_available`
+/// (from `new_session` / `load_session`) and the `cached_token` / `kimi.com/oidc`
+/// login handlers all stamped `sampling_config.api_key = Some(<primary bearer>)`
+/// with NO platform or endpoint guard — while that very config may point at a
+/// third-party model, and is both the subagent baseline and the
+/// unresolved-model fallback. `agent_ops`' generic-OAuth login handler already
+/// documents the correct rule ("Do NOT stamp this token onto the shared
+/// sampling_config"); the Kimi handlers violated it.
+///
+/// Revert-to-red: make `stamp_session_credential` skip the authority entirely —
+/// `self.sampling_config.borrow_mut().api_key =
+/// self.auth_manager.current_or_expired().map(|a| a.key); return true;` — and
+/// the third-party rows below become `Some(KIMI_TOKEN)`.
+#[tokio::test]
+#[serial_test::serial]
+async fn shared_sampling_config_is_never_stamped_off_the_session_endpoint() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    for (catalog_key, slug, base_url) in [
+        (
+            "deepseek/deepseek-chat",
+            "deepseek-chat",
+            "https://api.deepseek.com/v1",
+        ),
+        (
+            "moonshot-cn/kimi-k2-turbo-preview",
+            "kimi-k2-turbo-preview",
+            "https://api.moonshot.cn/v1",
+        ),
+    ] {
+        agent
+            .models_manager
+            .insert_test_entry(catalog_key, platform_entry(catalog_key, slug, base_url));
+        agent
+            .models_manager
+            .set_current_model_id(acp::ModelId::new(catalog_key));
+        rebuild_shared_config(&agent);
+        {
+            let mut shared = agent.sampling_config.borrow_mut();
+            assert_eq!(shared.model, slug, "{catalog_key}: built from this entry");
+            assert_eq!(shared.base_url, base_url);
+            shared.api_key = None;
+        }
+        // The `new_session` / `load_session` seed…
+        agent.seed_client_config_auth_if_available();
+        assert_eq!(
+            agent.sampling_config.borrow().api_key, None,
+            "LEAK: seeding stamped the Kimi bearer onto a config routed at {base_url}"
+        );
+        // …and the login handlers, which overwrite rather than seed.
+        assert!(
+            !agent.stamp_session_credential(true),
+            "LEAK: a login handler stamped the Kimi bearer onto a config routed at {base_url}"
+        );
+        assert_eq!(agent.sampling_config.borrow().api_key, None);
+    }
+
+    // Byte-identical on the session's own endpoint: both paths still stamp.
+    let kimi_key = "kimi-code/kimi-for-coding";
+    agent.models_manager.insert_test_entry(
+        kimi_key,
+        platform_entry(
+            kimi_key,
+            "kimi-for-coding",
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+        ),
+    );
+    agent
+        .models_manager
+        .set_current_model_id(acp::ModelId::new(kimi_key));
+    rebuild_shared_config(&agent);
+    agent.sampling_config.borrow_mut().api_key = None;
+    agent.seed_client_config_auth_if_available();
+    assert_eq!(
+        agent.sampling_config.borrow().api_key.as_deref(),
+        Some(KIMI_TOKEN),
+        "the subscription endpoint must still be seeded (this is what makes the \
+         assertions above meaningful)"
+    );
+}
+
+/// C1 — THE CRITICAL. The login stamp used to pair the right QUESTION (*may
+/// **a** session credential ride here?*) with the wrong CREDENTIAL (always
+/// `auth_manager.current_or_expired().key`, the primary Kimi bearer). For a
+/// subscription-OAuth platform at its OWN host the answer is correctly "yes" —
+/// the credential that may ride there is that platform's POOLED token
+/// ([`CredentialClass::Pooled`]) — so a Claude Pro/Max user whose current model is
+/// `claude-pro-max/*` ran `kigi login` and the Kimi subscription bearer landed
+/// on a config routed at `api.anthropic.com`. From there it reaches the wire via
+/// `resolve_sampling_config_for_model`'s verbatim fallback (offline / stale
+/// catalog) and via `SubagentSpawnContext`'s baseline clone.
+///
+/// All FOUR subscription platforms, each at its own registry host, derived from
+/// the registry so the fixture cannot drift.
+///
+/// The pooled managers are empty here (`pool_home()` is a per-process temp path
+/// under `cfg(test)`), so the correct answer is `None` — and `None` is also what
+/// proves the primary is not being substituted, because the same agent DOES
+/// stamp `KIMI_TOKEN` on its own coding endpoint at the end of the test.
+///
+/// Revert-to-red (production, compiles): restore the old shape in
+/// `MvpAgent::stamp_session_credential` —
+/// ```ignore
+/// if self.credential_authority().credential_class(platform, &base_url)
+///     == CredentialClass::None
+/// {
+///     return false;
+/// }
+/// let Some(auth) = self.auth_manager.current_or_expired() else { return false };
+/// self.sampling_config.borrow_mut().api_key = Some(auth.key);
+/// true
+/// ```
+/// and every OAuth row below becomes `Some(KIMI_TOKEN)`.
+#[tokio::test]
+#[serial_test::serial]
+async fn oauth_platform_shared_config_never_receives_the_primary_on_login() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    for (platform_id, slug) in [
+        ("claude-pro-max", "claude-opus-4-8"),
+        ("openai-codex", "gpt-5.5-codex"),
+        ("github-copilot", "gpt-4.1"),
+        ("xai-grok", "grok-4.5"),
+    ] {
+        let platform = kigi_models::PlatformId::parse(platform_id).expect("known platform");
+        let base_url = platform.base_url();
+        let catalog_key = format!("{platform_id}/{slug}");
+        agent
+            .models_manager
+            .insert_test_entry(&catalog_key, platform_entry(&catalog_key, slug, &base_url));
+        agent
+            .models_manager
+            .set_current_model_id(acp::ModelId::new(catalog_key.clone()));
+        rebuild_shared_config(&agent);
+        {
+            let mut shared = agent.sampling_config.borrow_mut();
+            assert_eq!(shared.model, slug, "{catalog_key}: built from this entry");
+            assert_eq!(shared.base_url, base_url);
+            shared.api_key = None;
+        }
+
+        // Precondition: this endpoint DOES take a session credential — that is
+        // exactly why guarding a hand-carried primary with "may a session
+        // credential ride?" was the bug. The class names WHICH one: the
+        // platform's POOLED token, never the primary / house key.
+        assert_eq!(
+            agent
+                .credential_authority()
+                .credential_class(Some(platform), &base_url),
+            CredentialClass::Pooled,
+            "{catalog_key}: precondition — its own host takes its POOLED token, \
+             and never the primary / house credential"
+        );
+
+        // `kigi login` (cached_token and kimi.com/oidc both land here) …
+        assert!(
+            !agent.stamp_session_credential(true),
+            "LEAK: a Kimi login stamped a credential onto a config routed at {base_url}"
+        );
+        assert_eq!(
+            agent.sampling_config.borrow().api_key,
+            None,
+            "LEAK: {catalog_key} received a bearer that is not its own pooled token"
+        );
+        // … and the `new_session` / `load_session` seed.
+        agent.seed_client_config_auth_if_available();
+        assert_ne!(
+            agent.sampling_config.borrow().api_key.as_deref(),
+            Some(KIMI_TOKEN),
+            "LEAK: seeding sent the Kimi subscription bearer to {base_url}"
+        );
+    }
+
+    // The first-party channel is untouched — this is what makes every
+    // assertion above meaningful (the Kimi bearer IS live and IS stampable).
+    let kimi_key = "kimi-code/kimi-for-coding";
+    agent.models_manager.insert_test_entry(
+        kimi_key,
+        platform_entry(
+            kimi_key,
+            "kimi-for-coding",
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+        ),
+    );
+    agent
+        .models_manager
+        .set_current_model_id(acp::ModelId::new(kimi_key));
+    rebuild_shared_config(&agent);
+    agent.sampling_config.borrow_mut().api_key = None;
+    assert!(agent.stamp_session_credential(true));
+    assert_eq!(
+        agent.sampling_config.borrow().api_key.as_deref(),
+        Some(KIMI_TOKEN),
+        "the Kimi subscription channel must be byte-identical"
+    );
+}
+
+/// H-a (AVAILABILITY, HIGH) — the stamp guard must resolve the model the shared
+/// config was BUILT from, not the one `current_model_id()` names NOW.
+///
+/// The shared config is built ONCE (`MvpAgent::with_models`) and never rebuilt,
+/// while `current_model_id()` moves on every non-Leader model switch
+/// (`handlers/model_switch.rs`) and on catalog reselection. Once they drift, a
+/// guard re-resolving the config's BARE slug against the live cell fails
+/// `entry_for_slug_resolution`'s `entry.info.model == slug` test and falls
+/// through to `resolve_catalog_key`'s `.rev()` scan.
+///
+/// `kimi-code` (subscription, `uses_oauth`) and `kimi-coding` (API-key twin,
+/// SAME coding host, `uses_oauth: false`) list the same routing slug, and
+/// `PlatformId::ALL` puts `kimi-code` 1st and `kimi-coding` 19th — so that scan
+/// answers `kimi-coding`, which takes NO session credential and therefore has NO
+/// governing manager. Because `stamp_session_credential` only overwrites ON
+/// SUCCESS, a Kimi-subscription user who also has `KIMI_API_KEY` set kept the
+/// EXPIRED bearer in the shared config after a successful re-login: every
+/// unresolved-model fallback and every subagent baseline turn 401'd until
+/// restart.
+///
+/// The switch happens AFTER the config is built — the previous version of this
+/// test left both on the same model, so it could not catch this.
+///
+/// Revert-to-red (production, compiles): make `MvpAgent::shared_config_platform`
+/// re-resolve from the live cell instead of returning the captured value —
+/// ```ignore
+/// fn shared_config_platform(&self) -> Option<kigi_models::PlatformId> {
+///     let model = self.sampling_config.borrow().model.clone();
+///     let current_key = self.models_manager.current_model_id();
+///     crate::agent::models::platform_for_slug(
+///         &self.models_manager.models(),
+///         Some(current_key.0.as_ref()),
+///         &model,
+///     )
+/// }
+/// ```
+/// and the re-login assertion below sees the stale bearer.
+#[tokio::test]
+#[serial_test::serial]
+async fn relogin_restamps_the_shared_config_after_the_model_switched() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+
+    let coding_host = kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url;
+    let slug = "kimi-for-coding";
+    // Insertion order mirrors `PlatformId::ALL`: the subscription platform
+    // first, its API-key twin later — so the `.rev()` scan answers the twin.
+    for catalog_key in ["kimi-code/kimi-for-coding", "kimi-coding/kimi-for-coding"] {
+        agent
+            .models_manager
+            .insert_test_entry(catalog_key, platform_entry(catalog_key, slug, coding_host));
+    }
+    assert_eq!(
+        crate::agent::models::platform_for_slug(&agent.models_manager.models(), None, slug),
+        Some(kigi_models::PlatformId::KimiCoding),
+        "precondition: a `None`-keyed slug scan answers the API-key twin, which takes \
+         no session credential"
+    );
+
+    // Startup: the picker selected the SUBSCRIPTION entry, and the shared config
+    // was built from it (config + platform, one call).
+    agent
+        .models_manager
+        .set_current_model_id(acp::ModelId::new("kimi-code/kimi-for-coding"));
+    rebuild_shared_config(&agent);
+    assert_eq!(
+        agent.sampling_config.borrow().model,
+        slug,
+        "precondition: the shared config carries the BARE slug, which collides"
+    );
+
+    // The user switches model: `current_model_id` moves to the API-key twin
+    // while the shared config still represents the subscription entry.
+    agent
+        .models_manager
+        .set_current_model_id(acp::ModelId::new("kimi-coding/kimi-for-coding"));
+
+    // The session expires and `kigi login` re-mints it. The handlers overwrite
+    // (`stamp_session_credential(true)`) AFTER the manager holds the new token.
+    agent.sampling_config.borrow_mut().api_key = Some("expired-bearer".to_string());
+    assert!(
+        agent.stamp_session_credential(true),
+        "a successful re-login must restamp the shared config"
+    );
+    assert_eq!(
+        agent.sampling_config.borrow().api_key.as_deref(),
+        Some(KIMI_TOKEN),
+        "the re-login must REPLACE the expired bearer: resolving the API-key twin \
+         instead finds no governing manager, leaves the stale token in place, and \
+         401s every subagent baseline turn until restart"
+    );
+
+    // And the seed path (`new_session` / `load_session`) agrees.
+    agent.sampling_config.borrow_mut().api_key = None;
+    agent.seed_client_config_auth_if_available();
+    assert_eq!(
+        agent.sampling_config.borrow().api_key.as_deref(),
+        Some(KIMI_TOKEN),
+    );
+}
+
+/// H3 (REGRESSION) — a MANAGED deployment configures its coding endpoint with
+/// `[endpoints] coding_api_base_url` in **config.toml** (what the managed-config
+/// sync writes), NOT the `KIGI_CODE_BASE_URL` env var. The previous round's
+/// predicate knew only the env var, so `EndpointsConfig::proxy_url()`'s
+/// config-key branch was invisible: every model inheriting the managed endpoint
+/// classified third-party, lost its api_key AND its resolver, and 401'd on every
+/// turn.
+///
+/// NOTE: no env var is set anywhere in this test — that is the point.
+///
+/// Revert-to-red: drop the `proxy_url()` arm from
+/// `CredentialAuthority::is_session_coding_endpoint` and both `assert_eq!`s
+/// below become `None`.
+#[tokio::test]
+#[serial_test::serial]
+async fn managed_config_toml_coding_endpoint_keeps_the_session_bearer() {
+    let _env = without_ambient_byok_env();
+    let _no_env_override = EnvGuard::unset("KIGI_CODE_BASE_URL");
+    let managed = "https://proxy.acme.example/v1";
+
+    let dir = tempfile::tempdir().expect("tempdir");
+    let auth_manager = std::sync::Arc::new(AuthManager::new(dir.path(), KimiCodeConfig::default()));
+    auth_manager.hot_swap(KimiAuth {
+        key: KIMI_TOKEN.to_string(),
+        auth_mode: AuthMode::OAuth,
+        refresh_token: Some("rt".into()),
+        expires_at: Some(chrono::Utc::now() + chrono::Duration::hours(1)),
+        ..KimiAuth::test_default()
+    });
+    let cfg = AgentConfig {
+        endpoints: EndpointsConfig {
+            coding_api_base_url: Some(managed.to_string()),
+            ..EndpointsConfig::default()
+        },
+        ..AgentConfig::default()
+    };
+    let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
+    let agent = MvpAgent::new(GatewaySender::new(tx), &cfg, auth_manager, None)
+        .expect("valid test config");
+    agent.set_auth_method(acp::AuthMethodId::new(CACHED_TOKEN_AUTH_METHOD_ID));
+    assert_eq!(
+        agent.models_manager.endpoints().proxy_url(),
+        managed,
+        "precondition: the session's effective coding endpoint is the config.toml key"
+    );
+
+    // A `[model.*]` entry inheriting the managed endpoint …
+    let mut bare = ModelEntry::fallback("kigi-4.5", &cfg.endpoints);
+    bare.info.id = None;
+    bare.info.base_url = managed.to_string();
+    assert_eq!(
+        agent
+            .prepare_sampling_config_for_model(&bare, None)
+            .api_key
+            .as_deref(),
+        Some(KIMI_TOKEN),
+        "a managed deployment must still receive the session bearer"
+    );
+
+    // … and the kimi-code catalog entry, whose base_url IS `proxy_url()`.
+    let kimi = platform_entry("kimi-code/kimi-for-coding", "kimi-for-coding", managed);
+    assert_eq!(
+        agent
+            .prepare_sampling_config_for_model(&kimi, None)
+            .api_key
+            .as_deref(),
+        Some(KIMI_TOKEN),
+        "the managed kimi-code entry must still receive the session bearer"
+    );
+
+    // The spec's requirement was "rides AND still refreshes": an api_key alone
+    // freezes at login and 401s unrecoverably ~1h in. The managed endpoint must
+    // also keep a LIVE manager — the primary's, so mid-session refresh and 401
+    // recovery run against the credential that actually owns that host.
+    for platform in [None, Some(kigi_models::PlatformId::KimiCode)] {
+        let manager = agent
+            .credential_authority()
+            .manager_for(platform, managed)
+            .unwrap_or_else(|| {
+                panic!("{platform:?}: a managed deployment must keep a live manager")
+            });
+        assert!(
+            std::sync::Arc::ptr_eq(&manager, &agent.auth_manager),
+            "{platform:?}: and it must be the session's OWN primary manager"
+        );
+        let resolver = agent
+            .credential_authority()
+            .bearer_resolver_for(platform, managed)
+            .unwrap_or_else(|| panic!("{platform:?}: … exposed as a live bearer_resolver"));
+        assert_eq!(
+            resolver.current_bearer(),
+            Some(KIMI_TOKEN.to_string()),
+            "{platform:?}: the resolver reads the live primary session bearer"
+        );
+    }
+
+    // The guard still holds: a third-party host under the SAME managed config
+    // gets nothing — no key, and no resolver either.
+    let third_party = platform_entry(
+        "deepseek/deepseek-chat",
+        "deepseek-chat",
+        "https://api.deepseek.com/v1",
+    );
+    assert_eq!(
+        agent
+            .prepare_sampling_config_for_model(&third_party, None)
+            .api_key,
+        None,
+        "LEAK: a managed deployment must not widen the trust set to third parties"
+    );
+    assert!(
+        agent
+            .credential_authority()
+            .bearer_resolver_for(
+                Some(kigi_models::PlatformId::DeepSeek),
+                "https://api.deepseek.com/v1"
+            )
+            .is_none(),
+        "LEAK: nor hand it a live resolver over the primary"
+    );
+}
+
+/// M3 (COMPLETION) — the SUMMARY client's `bearer_resolver` must honour the
+/// session-token gate, exactly as the session actor's aux path does.
+///
+/// `build_summary_client` set `cfg.bearer_resolver =
+/// authority.bearer_resolver_for(platform_for_slug(…), &cfg.base_url)` with NO
+/// gate while `SessionActor::aux_bearer_resolver` had one. `SamplingClient::post`
+/// REPLACES the request's auth header from that resolver, so an api-key /
+/// house-key session whose `[model.session-summary]` block carries its OWN
+/// `env_key` on the session's own coding endpoint had that key overwritten by
+/// the primary bearer on every summary request. Both now go through ONE rule,
+/// `sampler_turn::aux_bearer_resolver_for`.
+///
+/// The summary slug is deliberately absent from the catalog and from any
+/// config: it classifies `NotByok` definitively, so the only variable left is
+/// the ACP auth method — which is the gate term under test.
+///
+/// Revert-to-red (production, compiles): in `MvpAgent::summary_bearer_resolver`,
+/// return `self.credential_authority().bearer_resolver_for(platform, base_url)`
+/// directly (the pre-fix shape) and the api-key row below resolves `KIMI_TOKEN`.
+#[tokio::test]
+#[serial_test::serial]
+async fn summary_client_resolver_honours_the_session_gate() {
+    let _env = without_ambient_byok_env();
+    let (_dir, agent) = kimi_session_agent();
+    let coding_host = kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url;
+    let slug = "kigi-summary-aux-not-in-any-catalog";
+    let models = agent.models_manager.models();
+
+    // A session-based method on the session's OWN endpoint: byte-identical, the
+    // summary model keeps a LIVE resolver over the primary.
+    agent.set_auth_method(acp::AuthMethodId::new(CACHED_TOKEN_AUTH_METHOD_ID));
+    let resolver = agent
+        .summary_bearer_resolver(&models, slug, coding_host)
+        .expect("the first-party subscription summary channel keeps its resolver");
+    assert_eq!(
+        resolver.current_bearer(),
+        Some(KIMI_TOKEN.to_string()),
+        "…and it reads the live primary session bearer"
+    );
+
+    // An API-KEY session: the gate is inactive, so the summary model's own key
+    // must survive to the wire instead of being replaced by the primary.
+    agent.set_auth_method(acp::AuthMethodId::new(
+        crate::agent::auth_method::XAI_API_KEY_METHOD_ID,
+    ));
+    assert!(
+        agent
+            .summary_bearer_resolver(&models, slug, coding_host)
+            .is_none(),
+        "LEAK: an api-key session's summary model had its own key replaced on the \
+         wire by the primary bearer"
+    );
+}

crates/codegen/kigi-shell/src/agent/server.rs

@@ -176,6 +176,7 @@ async fn handle_connection(ws: WebSocket, state: Arc<ServerState>, peer_addr: So
                         prefetch_models_blocking(
                             &agent_config.endpoints,
                             auth.as_ref(),
+                            &Default::default(),
                             fetch_auth,
                             &platform_keys,
                         )

crates/codegen/kigi-shell/src/agent/session_config.rs

@@ -59,6 +59,8 @@ fn effort_label(effort: ReasoningEffort) -> String {
         ReasoningEffort::Medium => "Medium",
         ReasoningEffort::High => "High",
         ReasoningEffort::Xhigh => "X-High",
+        ReasoningEffort::Max => "Max",
+        ReasoningEffort::Ultra => "Ultra",
     }
     .to_string()
 }

crates/codegen/kigi-shell/src/agent/subagent/mod.rs

@@ -880,6 +880,24 @@ async fn read_parent_sampling_config(
             let auth_scheme = crate::agent::config::try_resolve_model_credentials(&cfg.model, None)
                 .map(|r| r.auth_scheme)
                 .unwrap_or_default();
+            // Claude Pro/Max OAuth Messages adaptation inherits from the parent
+            // model's platform (claude-pro-max → true); every other platform,
+            // and BYOK, → false, so the API-key paths stay byte-identical.
+            let anthropic_oauth = kigi_models::parse_managed_model_key(ctx.model_id.0.as_ref())
+                .is_some_and(|(platform, _)| {
+                    platform.oauth().is_some()
+                        && platform.wire_api() == kigi_models::PlatformWireApi::Messages
+                });
+            // GitHub Copilot editor headers inherit from the parent model's
+            // platform (github-copilot → true); every other platform / BYOK →
+            // false, so the other ChatCompletions paths stay byte-identical.
+            let github_copilot = kigi_models::parse_managed_model_key(ctx.model_id.0.as_ref())
+                .is_some_and(|(platform, _)| platform.sends_copilot_editor_headers());
+            // ChatGPT/Codex headers inherit from the parent model's platform
+            // (openai-codex → true); every other platform / BYOK → false, so the
+            // API-key openai Responses path stays byte-identical.
+            let openai_codex = kigi_models::parse_managed_model_key(ctx.model_id.0.as_ref())
+                .is_some_and(|(platform, _)| platform.sends_codex_responses_headers());
             let inherited = kigi_sampler::SamplerConfig {
                 api_key: creds.api_key,
                 base_url: cfg.base_url,
@@ -889,6 +907,10 @@ async fn read_parent_sampling_config(
                 top_p: cfg.top_p,
                 api_backend: cfg.api_backend,
                 auth_scheme,
+                anthropic_oauth,
+                github_copilot,
+                openai_codex,
+                chat_compat: cfg.chat_compat,
                 extra_headers,
                 context_window: cfg.context_window.get(),
                 reasoning_effort: cfg.reasoning_effort,
@@ -981,9 +1003,21 @@ fn resolve_model_override_to_config(
     } else {
         acp::ModelId::new(entry.info().model.clone())
     };
-    let session_key = ctx.auth.as_ref().map(|a| a.key.as_str());
+    // Resolve the child's session token by the OVERRIDE model's OWN platform
+    // AND endpoint, not the parent's primary auth: a grok (oauth-platform)
+    // override draws its pooled grok token or `None`, and an API-key registry
+    // platform / a third-party `[model.*]` host draws NOTHING — NEVER the
+    // primary Kimi session token, which `resolve_credentials` would otherwise
+    // stamp onto the child's api.x.ai / api.moonshot.cn credentials. The
+    // first-party subscription channel still resolves to the primary
+    // (byte-identical).
+    let session_key = crate::auth::credential_authority::CredentialAuthority::new(
+        ctx.models_manager.endpoints(),
+        Some(ctx.auth_manager.clone()),
+    )
+    .credential_for_model(&entry);
     let has_session_key = session_key.is_some();
-    let mut credentials = resolve_credentials(&entry, session_key);
+    let mut credentials = resolve_credentials(&entry, session_key.as_ref());
     credentials.auth_type = subagent_auth_type(Some(&entry), &ctx.auth_method_id);
     let resolved_auth_type = credentials.auth_type;
     let config = sampling_config_for_model(&entry, credentials, ctx.alpha_test_key.clone());

crates/codegen/kigi-shell/src/agent/subagent/tests/mod.rs

@@ -3131,19 +3131,25 @@ fn fresh_tool_model_rejects_unavailable_exact_key_over_visible_slug_collision()
         "validation must inspect the unavailable exact-key entry selected by execution"
     );
 }
+/// Validation must inspect the SAME slug-collision entry execution selects.
+/// Both go through `find_model_by_id`, whose slug scan takes the LAST match —
+/// aligned with the picker's `resolve_catalog_key` so the auth layer and the
+/// picker can never resolve different platforms for one slug (the H5 collision).
+/// So a blocked LAST entry must be rejected even though an available earlier one
+/// shares the slug.
 #[test]
-fn fresh_tool_model_rejects_unavailable_first_slug_collision() {
+fn fresh_tool_model_rejects_unavailable_last_slug_collision() {
     let mut models = indexmap::IndexMap::new();
-    let mut unavailable_first = test_model_entry("shared-routing-slug");
-    unavailable_first.info.user_selectable = false;
-    models.insert("blocked-first".to_string(), unavailable_first);
-    models.insert("visible-second".to_string(), test_model_entry("shared-routing-slug"));
+    models.insert("visible-first".to_string(), test_model_entry("shared-routing-slug"));
+    let mut unavailable_last = test_model_entry("shared-routing-slug");
+    unavailable_last.info.user_selectable = false;
+    models.insert("blocked-last".to_string(), unavailable_last);
     assert_eq!(
         super::handle_request::task_model_override_error(Some("shared-routing-slug"),
         ModelOverrideProvenance::Tool, false, & models, false,).as_deref(),
         Some("Unknown Task.model slug 'shared-routing-slug'. Valid model slugs: \
-                 visible-second. Omit `model` to inherit the parent model."),
-        "validation must inspect the first routing-slug entry selected by execution"
+                 visible-first. Omit `model` to inherit the parent model."),
+        "validation must inspect the last routing-slug entry selected by execution"
     );
 }
 #[test]
@@ -3272,6 +3278,7 @@ fn test_sampling_config(model_slug: &str) -> kigi_sampling_types::SamplingConfig
         temperature: None,
         top_p: None,
         api_backend: Default::default(),
+        chat_compat: Default::default(),
         extra_headers: Default::default(),
         context_window: NonZeroU64::new(256_000).expect("non-zero context window"),
         reasoning_effort: None,

crates/codegen/kigi-shell/src/agent/subagent/tests/rest.rs

@@ -2467,6 +2467,111 @@ async fn resolve_subagent_config_override_unknown_model_falls_through_to_inherit
     assert_eq!(config.model, "kigi-4.5");
     assert_eq!(model_id.0.as_ref(), "kigi-4.5");
 }
+/// Build an `Arc<AuthManager>` (primary Kimi) holding `key` as its live bearer.
+/// The `TempDir` is returned so the caller keeps it alive.
+fn kimi_primary_with_token(key: &str) -> (tempfile::TempDir, std::sync::Arc<crate::auth::AuthManager>) {
+    let dir = tempfile::tempdir().unwrap();
+    let manager = std::sync::Arc::new(crate::auth::AuthManager::new(
+        dir.path(),
+        crate::auth::KimiCodeConfig::default(),
+    ));
+    manager.hot_swap(crate::auth::KimiAuth {
+        key: key.to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    (dir, manager)
+}
+/// LEAK 2 (subagent model-override): a grok (oauth-platform) override with a
+/// Kimi primary must NEVER receive the primary Kimi session token as its
+/// `api_key` — it draws grok's own pooled token (or `None`). Revert-to-red: the
+/// pre-fix code passed `ctx.auth` (Kimi) straight to `resolve_credentials`, so
+/// `config.api_key == "kimi-secret"` and this assertion fails.
+#[tokio::test]
+async fn subagent_override_grok_model_never_leaks_kimi_session_token() {
+    let (_kd, manager) = kimi_primary_with_token("kimi-secret");
+    let mut grok = test_model_entry("grok-4-latest");
+    grok.info.id = Some("xai-grok/grok-4-latest".to_string());
+    grok.info.base_url = "https://api.x.ai/v1".to_string();
+    let mut models = indexmap::IndexMap::new();
+    models.insert("grok".to_string(), grok);
+    let mut ctx = ctx_with_toggle(HashMap::new());
+    ctx.available_models = models;
+    ctx.auth = Some(crate::auth::KimiAuth {
+        key: "kimi-secret".to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    ctx.auth_manager = manager;
+    let (config, _model_id) =
+        resolve_model_override_to_config("grok", &ctx).expect("grok override resolves to a config");
+    assert_ne!(
+        config.api_key.as_deref(),
+        Some("kimi-secret"),
+        "a grok override must never receive the primary Kimi session token",
+    );
+}
+/// Byte-identical guard: an override on the SESSION's own first-party endpoint
+/// (the kimi-code subscription channel) still resolves to the primary session
+/// token — the primary path is unchanged.
+#[tokio::test]
+async fn subagent_override_first_party_model_still_gets_primary_token() {
+    let (_kd, manager) = kimi_primary_with_token("kimi-secret");
+    let mut entry = test_model_entry("kimi-for-coding");
+    entry.info.id = Some("kimi-code/kimi-for-coding".to_string());
+    entry.info.base_url = kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url.to_string();
+    let mut models = indexmap::IndexMap::new();
+    models.insert("kfc".to_string(), entry);
+    let mut ctx = ctx_with_toggle(HashMap::new());
+    ctx.available_models = models;
+    ctx.auth = Some(crate::auth::KimiAuth {
+        key: "kimi-secret".to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    ctx.auth_manager = manager;
+    let (config, _model_id) = resolve_model_override_to_config("kfc", &ctx)
+        .expect("first-party override resolves to a config");
+    assert_eq!(
+        config.api_key.as_deref(),
+        Some("kimi-secret"),
+        "a first-party override must still receive the primary session token",
+    );
+}
+/// LEAK guard (C1, subagent-override `api_key` channel): an API-key registry
+/// platform override must NOT receive the parent's primary Kimi session token —
+/// `resolve_credentials` would stamp it as the child's `api_key` on
+/// `api.moonshot.cn`. This test previously asserted the opposite
+/// (`subagent_override_non_oauth_model_still_gets_primary_token`), which encoded
+/// the defect.
+///
+/// Revert-to-red: make `CredentialAuthority::governing_manager`'s
+/// `Some(platform) => None` arm return `self.primary.clone()` and `api_key`
+/// becomes `Some("kimi-secret")`.
+#[tokio::test]
+async fn subagent_override_api_key_platform_never_gets_the_primary_token() {
+    let (_kd, manager) = kimi_primary_with_token("kimi-secret");
+    let mut entry = test_model_entry("kimi-k2-0905-preview");
+    entry.info.id = Some("moonshot-cn/kimi-k2".to_string());
+    entry.info.base_url = "https://api.moonshot.cn/v1".to_string();
+    let mut models = indexmap::IndexMap::new();
+    models.insert("k2".to_string(), entry);
+    let mut ctx = ctx_with_toggle(HashMap::new());
+    ctx.available_models = models;
+    ctx.auth = Some(crate::auth::KimiAuth {
+        key: "kimi-secret".to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    ctx.auth_manager = manager;
+    let (config, _model_id) = resolve_model_override_to_config("k2", &ctx)
+        .expect("an API-key-platform override still resolves to a config");
+    assert_ne!(
+        config.api_key.as_deref(),
+        Some("kimi-secret"),
+        "LEAK: an API-key-platform override must never receive the primary Kimi session token",
+    );
+}
 /// An unresolvable `AgentDefinition.model` pin (model absent from
 /// `available_models`) falls through to inherit the parent model.
 #[tokio::test]

crates/codegen/kigi-shell/src/auth/credential_authority.rs

@@ -0,0 +1,568 @@
+//! THE credential chokepoint.
+//!
+//! One authority answers, for every outgoing inference request, the only
+//! question that matters: **which credential — if any — may ride it?**
+//! ([`CredentialAuthority::credential_class`]). Before this module the answer was
+//! re-derived — differently — at every call site (`ModelsManager`, `MvpAgent`,
+//! `SessionActor`, the aux/summary/subagent paths), and three separate rounds
+//! of fixes each closed some sites and missed others.
+//!
+//! # How omission is structurally prevented
+//!
+//! 1. [`SessionCredential`] wraps the bearer and has **no production
+//!    constructor outside this module**. The only function in the crate that
+//!    can build one is [`CredentialAuthority::credential_for`], which *requires*
+//!    `(platform, base_url)` and holds the session's `EndpointsConfig` and
+//!    primary [`AuthManager`] privately.
+//! 2. Every API that stamps a session credential onto a request —
+//!    `resolve_credentials`, `resolve_aux_model_sampling_config`,
+//!    `try_resolve_model_credentials`,
+//!    `resolve_chat_state_auth_type` — takes `Option<&SessionCredential>`,
+//!    never `Option<&str>`. A new call site therefore *cannot compile* a leak:
+//!    there is no way to produce the value without going through the rule.
+//! 3. The authority owns the primary manager privately and exposes it only via
+//!    [`CredentialAuthority::manager_for`] /
+//!    [`CredentialAuthority::bearer_resolver_for`], which take the same
+//!    `(platform, base_url)` pair — so the `bearer_resolver` sink is funnelled
+//!    through the identical rule as the `api_key` sink.
+//! 4. A guard asks [`CredentialAuthority::credential_class`] and MATCHES on the
+//!    answer. There is no second, similarly-named boolean to pick by mistake:
+//!    the round-3 defect (C1) was `takes_session_credential` — *may **a**
+//!    session credential ride?* — paired with a hand-carried PRIMARY bearer,
+//!    and the two predicates that made that pairing expressible are gone.
+//!
+//! SECURITY: no token is ever logged, `Debug`-printed or `Display`ed here.
+
+use std::sync::Arc;
+
+use crate::agent::config::{EndpointsConfig, ModelEntry};
+use crate::auth::AuthManager;
+
+/// A session bearer this authority has cleared for one specific request
+/// endpoint.
+///
+/// Opaque by construction: the inner `String` is private, the type is not
+/// `Debug`/`Clone`-into-`String`, and the only production constructor is
+/// [`CredentialAuthority::credential_for`]. See the module docs for why that
+/// matters.
+pub(crate) struct SessionCredential(String);
+
+impl SessionCredential {
+    /// The raw bearer. SECURITY: callers stamp this straight onto a request —
+    /// never log it.
+    pub(crate) fn expose(&self) -> &str {
+        &self.0
+    }
+
+    /// Test-only forgery, so unit tests can exercise the *downstream*
+    /// credential plumbing (`resolve_credentials`' BYOK-vs-session precedence,
+    /// aux config shapes) without standing up an `AuthManager`. Deliberately
+    /// `#[cfg(test)]`: production code has no way to build one.
+    #[cfg(test)]
+    pub(crate) fn for_test(key: &str) -> Self {
+        Self(key.to_owned())
+    }
+}
+
+/// WHICH credential — if any — may ride a request routed to a given
+/// `(platform, base_url)` pair.
+///
+/// ONE question with three answers, replacing the two look-alike booleans
+/// `takes_session_credential` / `takes_primary_credential` (identical
+/// signatures, near-identical names, opposite answers on a subscription host).
+/// C1 was caused by asking the first and stamping the credential the second
+/// describes; with a single classifier a call site must MATCH on the answer, so
+/// that mistake is no longer expressible.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) enum CredentialClass {
+    /// `platform`'s OWN pooled subscription-OAuth token, at its own registry
+    /// host. NEVER the primary session bearer and never the house key.
+    Pooled,
+    /// The credential that authorizes the SESSION's own coding endpoint:
+    /// the primary (`kimi-code` / platform-less) bearer.
+    ///
+    /// Deliberately NOT split into a separate `HouseKey` variant: the house
+    /// `KIGI_API_KEY` is accepted by exactly this endpoint and no other, so it
+    /// rides precisely this class. A fourth variant would re-create the
+    /// two-similar-answers hazard this enum exists to remove.
+    Primary,
+    /// Nothing rides: every API-key registry platform, an OAuth platform
+    /// redirected off its own host, and any endpoint that is not the session's.
+    None,
+}
+
+impl CredentialClass {
+    /// Stable label for structured logs. SECURITY: names a channel, never a
+    /// token.
+    pub(crate) fn as_str(self) -> &'static str {
+        match self {
+            Self::Pooled => "pooled",
+            Self::Primary => "primary",
+            Self::None => "none",
+        }
+    }
+}
+
+/// The single authority over inference-time session credentials.
+///
+/// Construct one from the session's EFFECTIVE endpoints plus its primary
+/// (first-party / Kimi) manager, then ask it about a request. Cheap to build
+/// (a handful of `Option<String>` clones + an `Arc` clone).
+#[derive(Clone)]
+pub(crate) struct CredentialAuthority {
+    /// The session's effective `[endpoints]` — config.toml layered over env.
+    /// H3: `EndpointsConfig::proxy_url()` prefers `[endpoints]
+    /// coding_api_base_url` from **config.toml** and only then falls back to
+    /// `KIGI_CODE_BASE_URL`. A predicate that knows only the env var makes a
+    /// managed/enterprise deployment lose its session bearer entirely (401 on
+    /// every turn), so the endpoints are part of the authority's identity, not
+    /// an afterthought.
+    endpoints: EndpointsConfig,
+    /// The primary session manager. PRIVATE: nothing hands it back, so a path
+    /// holding a `CredentialAuthority` cannot reach `current_or_expired()`
+    /// without naming an endpoint.
+    primary: Option<Arc<AuthManager>>,
+}
+
+impl CredentialAuthority {
+    pub(crate) fn new(endpoints: EndpointsConfig, primary: Option<Arc<AuthManager>>) -> Self {
+        Self { endpoints, primary }
+    }
+
+    /// THE rule, stated once.
+    ///
+    /// - a subscription-OAuth platform (claude-pro-max, openai-codex,
+    ///   github-copilot, xai-grok) rides ITS OWN pooled manager — never the
+    ///   primary — and only to its own registry host (L10: a
+    ///   `[model."claude-pro-max/x"]` override keeps `info.id` but can point
+    ///   `base_url` anywhere, and used to ship the Claude OAuth bearer there);
+    /// - `kimi-code` — the one `uses_oauth` platform with no `OAuthConfig` —
+    ///   rides the PRIMARY session, and only at the session's own effective
+    ///   coding endpoint;
+    /// - every API-key registry platform (deepseek, openai, anthropic,
+    ///   moonshot-*, …) rides NOTHING: its credential is that platform's API
+    ///   key, already resolved into the catalog entry;
+    /// - a platform-less model (a bare slug or a `[model.*]` block) is decided
+    ///   purely by the ENDPOINT — BYOK detection probes `std::env::var` at call
+    ///   time, so an unset/mistyped `env_key` must not turn into "send the
+    ///   subscription bearer to `api.openai.com`".
+    pub(crate) fn credential_class(
+        &self,
+        platform: Option<kigi_models::PlatformId>,
+        base_url: &str,
+    ) -> CredentialClass {
+        match platform {
+            Some(platform) => match platform.oauth() {
+                Some(_) if self.endpoint_is_platform_host(platform, base_url) => {
+                    CredentialClass::Pooled
+                }
+                // An OAuth platform pointed at a host that is NOT its own.
+                Some(_) => CredentialClass::None,
+                None if platform.uses_oauth() && self.is_session_coding_endpoint(base_url) => {
+                    CredentialClass::Primary
+                }
+                // `kimi-code` off the session's endpoint, and every API-key
+                // registry platform.
+                None => CredentialClass::None,
+            },
+            None if self.is_session_coding_endpoint(base_url) => CredentialClass::Primary,
+            None => CredentialClass::None,
+        }
+    }
+
+    /// The manager behind [`Self::credential_class`]. Derived from the class, so
+    /// the rule is stated exactly once and the two can never disagree.
+    fn governing_manager(
+        &self,
+        platform: Option<kigi_models::PlatformId>,
+        base_url: &str,
+    ) -> Option<Arc<AuthManager>> {
+        match self.credential_class(platform, base_url) {
+            CredentialClass::Pooled => {
+                platform
+                    .and_then(kigi_models::PlatformId::oauth)
+                    .map(|oauth| {
+                        crate::auth::oauth_registry::global_manager_for(
+                            &crate::auth::oauth_registry::pool_home(),
+                            oauth,
+                        )
+                    })
+            }
+            CredentialClass::Primary => self.primary.clone(),
+            CredentialClass::None => None,
+        }
+    }
+
+    /// Whether `base_url` is the SESSION's own coding endpoint: the effective
+    /// `[endpoints] coding_api_base_url` from **config.toml** (what a managed /
+    /// enterprise deployment actually sets — H3), the `models_base_url`
+    /// custom-endpoint mode, the `KIGI_CODE_BASE_URL` env override, a loopback
+    /// dev proxy, or the compiled production endpoint.
+    ///
+    /// Deliberately NOT [`crate::util::is_first_party_url`], which is
+    /// production-only and would break every custom deployment.
+    fn is_session_coding_endpoint(&self, base_url: &str) -> bool {
+        if crate::util::is_effective_coding_endpoint_url(base_url) {
+            return true;
+        }
+        if crate::util::matches_trusted_base_url(base_url, &self.endpoints.proxy_url()) {
+            return true;
+        }
+        self.endpoints
+            .models_base_url
+            .as_deref()
+            .is_some_and(|models_base| crate::util::matches_trusted_base_url(base_url, models_base))
+    }
+
+    /// Whether `base_url` is `platform`'s own registry host — the guard that
+    /// keeps a subscription-OAuth bearer from riding a redirected `[model.*]`
+    /// override to a third party (L10).
+    fn endpoint_is_platform_host(&self, platform: kigi_models::PlatformId, base_url: &str) -> bool {
+        crate::util::matches_trusted_base_url(base_url, &platform.base_url())
+    }
+
+    /// The `AuthManager` that governs this request's bearer resolution,
+    /// mid-session refresh and 401 recovery — or `None` when no session
+    /// credential may ride (fail fast; never a silent fallback to the primary).
+    pub(crate) fn manager_for(
+        &self,
+        platform: Option<kigi_models::PlatformId>,
+        base_url: &str,
+    ) -> Option<Arc<AuthManager>> {
+        self.governing_manager(platform, base_url)
+    }
+
+    /// The session bearer to stamp as this request's `api_key`, or `None`.
+    ///
+    /// The ONLY production constructor of [`SessionCredential`].
+    pub(crate) fn credential_for(
+        &self,
+        platform: Option<kigi_models::PlatformId>,
+        base_url: &str,
+    ) -> Option<SessionCredential> {
+        self.governing_manager(platform, base_url)
+            .and_then(|am| am.current_or_expired())
+            .map(|auth| SessionCredential(auth.key))
+    }
+
+    /// A live sampler `bearer_resolver` over the governing manager, so the
+    /// request keeps mid-session refresh / 401 recovery against the credential
+    /// that actually belongs to its host.
+    pub(crate) fn bearer_resolver_for(
+        &self,
+        platform: Option<kigi_models::PlatformId>,
+        base_url: &str,
+    ) -> Option<kigi_sampler::SharedBearerResolver> {
+        self.manager_for(platform, base_url)
+            .map(crate::session::acp_session::sampler_turn::auth_manager_bearer_resolver)
+    }
+
+    /// [`Self::credential_for`] for a resolved catalog entry: derives the
+    /// platform and the base URL from the SAME entry, so the two can never be
+    /// mismatched by a call site.
+    pub(crate) fn credential_for_model(&self, entry: &ModelEntry) -> Option<SessionCredential> {
+        let info = entry.info();
+        self.credential_for(entry_platform(entry), &info.base_url)
+    }
+
+    /// [`Self::credential_for`] for the catalog model a routing slug resolves
+    /// to. `current_key` is the SESSION's own selected catalog key (see
+    /// [`crate::agent::models::entry_for_slug`]); pass `None` for aux /
+    /// override slugs, which are not the session's selection.
+    ///
+    /// M5: a slug that is NOT in the catalog resolves through the SAME endpoint
+    /// rule against the aux fallback endpoint
+    /// (`EndpointsConfig::resolve_inference_base_url`, which is exactly where
+    /// `resolve_aux_model_sampling_config`'s Tier-2 entry routes) instead of
+    /// being handed the primary unconditionally — the old "first-party by
+    /// construction" justification was false once `models_base_url` could point
+    /// anywhere.
+    ///
+    /// M6: the platform and the base URL come from ONE
+    /// [`crate::agent::models::entry_for_slug`] lookup, so they can no longer
+    /// disagree (the aux path used to resolve the platform with `current_key`
+    /// and the credential with a separate `find_model_by_id`).
+    pub(crate) fn credential_for_slug(
+        &self,
+        models: &indexmap::IndexMap<String, ModelEntry>,
+        current_key: Option<&str>,
+        slug: &str,
+    ) -> Option<SessionCredential> {
+        match crate::agent::models::entry_for_slug(models, current_key, slug) {
+            Some(entry) => self.credential_for_model(entry),
+            None => self.credential_for(None, &self.endpoints.resolve_inference_base_url()),
+        }
+    }
+}
+
+/// The registry platform a catalog entry belongs to (`info.id` is the managed
+/// key `{platform}/{model}`). `None` for a bare / `[model.*]` entry.
+pub(crate) fn entry_platform(entry: &ModelEntry) -> Option<kigi_models::PlatformId> {
+    entry
+        .info()
+        .id
+        .as_deref()
+        .and_then(kigi_models::parse_managed_model_key)
+        .map(|(platform, _)| platform)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::auth::{AuthMode, KimiAuth, KimiCodeConfig};
+
+    /// A primary holding a fixed in-memory bearer. The `TempDir` is returned so
+    /// the caller keeps it alive; the token is read from memory, so on-disk
+    /// contents are irrelevant.
+    fn primary(key: &str) -> (tempfile::TempDir, Arc<AuthManager>) {
+        let dir = tempfile::tempdir().unwrap();
+        let manager = Arc::new(AuthManager::new(dir.path(), KimiCodeConfig::default()));
+        manager.hot_swap(KimiAuth {
+            key: key.to_string(),
+            auth_mode: AuthMode::OAuth,
+            ..KimiAuth::test_default()
+        });
+        (dir, manager)
+    }
+
+    fn authority(endpoints: EndpointsConfig, primary: Arc<AuthManager>) -> CredentialAuthority {
+        CredentialAuthority::new(endpoints, Some(primary))
+    }
+
+    fn platform(id: &str) -> kigi_models::PlatformId {
+        kigi_models::PlatformId::parse(id).expect("known platform")
+    }
+
+    /// H3 (REGRESSION): the effective coding endpoint is
+    /// `EndpointsConfig::proxy_url()`, which prefers `[endpoints]
+    /// coding_api_base_url` from **config.toml** — the key the managed-config
+    /// sync writes. A predicate that knows only `KIGI_CODE_BASE_URL` classifies
+    /// such a deployment as third-party, withholds the api_key AND the
+    /// resolver, and 401s on every turn.
+    ///
+    /// Revert-to-red: drop the `proxy_url()` arm from
+    /// `is_session_coding_endpoint` (leaving only
+    /// `is_effective_coding_endpoint_url`) and every assertion here fails —
+    /// with NO env var set anywhere in the test.
+    #[test]
+    fn config_toml_coding_endpoint_still_rides_the_session_bearer() {
+        let (_d, kimi) = primary("kimi-tok");
+        let managed = "https://proxy.acme.com/v1";
+        let auth = authority(
+            EndpointsConfig {
+                coding_api_base_url: Some(managed.to_string()),
+                ..EndpointsConfig::default()
+            },
+            kimi.clone(),
+        );
+        assert_eq!(
+            auth.credential_class(None, managed),
+            CredentialClass::Primary,
+            "a [model.*] entry inheriting the managed coding endpoint takes the session bearer"
+        );
+        assert_eq!(
+            auth.credential_for(None, managed)
+                .map(|c| c.expose().to_owned()),
+            Some("kimi-tok".to_string()),
+            "the managed deployment must still receive the session bearer"
+        );
+        assert!(
+            auth.manager_for(None, managed).is_some(),
+            "and must keep a live manager, or it loses refresh and 401 recovery"
+        );
+        // kimi-code entries route to `proxy_url()` too (models_fetch's
+        // `platform_fetch_base`), so the platform arm must honour it as well.
+        assert_eq!(
+            auth.credential_for(Some(platform("kimi-code")), managed)
+                .map(|c| c.expose().to_owned()),
+            Some("kimi-tok".to_string()),
+        );
+        // A DIFFERENT authority (no managed key configured) must NOT trust it.
+        let default_auth = authority(EndpointsConfig::default(), kimi);
+        assert_eq!(
+            default_auth.credential_class(None, managed),
+            CredentialClass::None,
+            "the managed host is only trusted for the session that configured it"
+        );
+    }
+
+    /// The `models_base_url` custom-endpoint mode is equally invisible to the
+    /// env-var-only predicate.
+    #[test]
+    fn config_toml_models_base_url_still_rides_the_session_bearer() {
+        let (_d, kimi) = primary("kimi-tok");
+        let custom = "https://models.acme.internal/v1";
+        let auth = authority(
+            EndpointsConfig {
+                models_base_url: Some(custom.to_string()),
+                ..EndpointsConfig::default()
+            },
+            kimi,
+        );
+        assert_eq!(
+            auth.credential_for(None, custom)
+                .map(|c| c.expose().to_owned()),
+            Some("kimi-tok".to_string()),
+        );
+    }
+
+    /// The compiled production endpoint and loopback proxies are unchanged.
+    #[test]
+    fn production_and_loopback_endpoints_are_unchanged() {
+        let (_d, kimi) = primary("kimi-tok");
+        let auth = authority(EndpointsConfig::default(), kimi);
+        for url in [
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+            "http://127.0.0.1:8080/v1",
+            "http://localhost:3000/v1",
+            "http://[::1]:9000/v1",
+        ] {
+            assert_eq!(
+                auth.credential_class(None, url),
+                CredentialClass::Primary,
+                "{url}: the session's own endpoint is byte-identical"
+            );
+        }
+    }
+
+    /// LEAK guard: every API-key registry platform, and any platform-less model
+    /// on a third-party host, gets NO session credential and NO manager.
+    #[test]
+    fn third_party_endpoints_never_receive_the_primary() {
+        let (_d, kimi) = primary("kimi-tok");
+        let auth = authority(EndpointsConfig::default(), kimi);
+        for id in [
+            "deepseek",
+            "openai",
+            "anthropic",
+            "moonshot-cn",
+            "moonshot-ai",
+        ] {
+            let p = platform(id);
+            assert!(
+                auth.credential_for(Some(p), &p.base_url()).is_none(),
+                "LEAK: {id} is an API-key platform — no session bearer may ride there"
+            );
+            assert!(auth.manager_for(Some(p), &p.base_url()).is_none());
+        }
+        for url in [
+            "https://api.openai.com/v1",
+            "https://api.deepseek.com/v1",
+            "https://api.moonshot.cn/v1",
+            "",
+        ] {
+            assert!(
+                auth.credential_for(None, url).is_none(),
+                "LEAK: {url} is a third-party host"
+            );
+        }
+    }
+
+    /// L10: an OAuth platform whose `[model.*]` override redirects `base_url`
+    /// to a third-party host keeps `info.id` — and must NOT ship that
+    /// platform's pooled OAuth bearer there.
+    #[tokio::test]
+    async fn oauth_platform_redirected_to_a_third_party_host_gets_nothing() {
+        let (_d, kimi) = primary("kimi-tok");
+        let auth = authority(EndpointsConfig::default(), kimi);
+        for id in [
+            "claude-pro-max",
+            "openai-codex",
+            "github-copilot",
+            "xai-grok",
+        ] {
+            let p = platform(id);
+            assert!(
+                auth.manager_for(Some(p), &p.base_url()).is_some(),
+                "{id} keeps its own pooled manager on its own host"
+            );
+            assert!(
+                auth.manager_for(Some(p), "https://third.party/v1")
+                    .is_none(),
+                "LEAK: {id} redirected to a third-party host must ship no bearer"
+            );
+            assert_eq!(
+                auth.credential_class(Some(p), "https://third.party/v1"),
+                CredentialClass::None,
+                "LEAK: {id} redirected to a third-party host takes no session credential"
+            );
+        }
+    }
+
+    /// C1 — a subscription platform's own host DOES take a session credential,
+    /// but it is that platform's POOLED token, never the primary / house key.
+    /// Two look-alike booleans used to encode this, and picking the wrong one is
+    /// the whole defect; one classifier makes the distinction impossible to
+    /// mis-read.
+    #[test]
+    fn a_subscription_host_classifies_pooled_never_primary() {
+        let (_d, kimi) = primary("kimi-tok");
+        let auth = authority(EndpointsConfig::default(), kimi);
+        for id in [
+            "claude-pro-max",
+            "openai-codex",
+            "github-copilot",
+            "xai-grok",
+        ] {
+            let p = platform(id);
+            assert_eq!(
+                auth.credential_class(Some(p), &p.base_url()),
+                CredentialClass::Pooled,
+                "LEAK: {id}'s own host takes its POOLED token — never the primary / house key"
+            );
+        }
+        // Every API-key registry platform: nothing at all.
+        for id in ["deepseek", "openai", "anthropic", "moonshot-cn"] {
+            let p = platform(id);
+            assert_eq!(
+                auth.credential_class(Some(p), &p.base_url()),
+                CredentialClass::None
+            );
+        }
+        // The primary channel is unchanged: kimi-code and a platform-less model
+        // on the session's own endpoint, and nothing on a third-party host.
+        for url in [
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+            "http://127.0.0.1:8080/v1",
+        ] {
+            assert_eq!(auth.credential_class(None, url), CredentialClass::Primary);
+            assert_eq!(
+                auth.credential_class(Some(platform("kimi-code")), url),
+                CredentialClass::Primary
+            );
+        }
+        assert_eq!(
+            auth.credential_class(None, "https://api.openai.com/v1"),
+            CredentialClass::None
+        );
+    }
+
+    /// The four subscription-OAuth platforms draw from their OWN pooled
+    /// managers — never the primary Kimi one, even under a Kimi session.
+    #[tokio::test]
+    async fn oauth_platforms_never_resolve_the_primary() {
+        let (_d, kimi) = primary("kimi-tok");
+        let auth = authority(EndpointsConfig::default(), kimi.clone());
+        for id in [
+            "claude-pro-max",
+            "openai-codex",
+            "github-copilot",
+            "xai-grok",
+        ] {
+            let p = platform(id);
+            let resolved = auth
+                .manager_for(Some(p), &p.base_url())
+                .expect("pooled manager");
+            assert!(
+                !Arc::ptr_eq(&resolved, &kimi),
+                "{id} must NOT resolve the primary Kimi manager"
+            );
+            assert_ne!(
+                auth.credential_for(Some(p), &p.base_url())
+                    .map(|c| c.expose().to_owned()),
+                Some("kimi-tok".to_string()),
+                "{id} must never receive the primary Kimi bearer"
+            );
+        }
+    }
+}

crates/codegen/kigi-shell/src/auth/device_code.rs

@@ -13,15 +13,70 @@
 
 use std::sync::Arc;
 
-use crate::auth::kimi_oauth::{
-    DeviceAuthorization, DevicePollResult, poll_device_token, request_device_authorization,
-};
+use kigi_models::OAuthConfig;
+
+use crate::auth::kimi_oauth::{DeviceAuthorization, DevicePollResult};
 use crate::auth::{AuthChannels, AuthManager, AuthUrlInfo, AuthUrlMode, KimiAuth};
 
 /// Extra wait added to the poll interval when the server answers `slow_down`
 /// (OAuth-standard device-flow backpressure).
 const SLOW_DOWN_INCREMENT_SECS: u64 = 5;
 
+/// The wire behind a device-code login. The `Kimi` arm calls the bespoke Kimi
+/// Code wire (X-Msh headers, `/api/oauth/*`) verbatim — byte-identical to the
+/// pre-generalization path; the `Generic` arm drives a registry
+/// [`OAuthConfig`] provider (xai-grok) through [`crate::auth::oauth_device`].
+enum DeviceFlowBackend<'a> {
+    Kimi {
+        host: &'a str,
+    },
+    Generic(&'a OAuthConfig),
+    /// GitHub Copilot two-stage flow: the device authorization is the generic
+    /// one, but the token POLL reads GitHub's 200-body errors, and login
+    /// FINALIZES the durable github token into a copilot session token via the
+    /// Stage-2 exchange (see [`DeviceFlowBackend::finalize`]).
+    GithubCopilot(&'a OAuthConfig),
+}
+
+impl DeviceFlowBackend<'_> {
+    async fn request(&self) -> anyhow::Result<DeviceAuthorization> {
+        match self {
+            Self::Kimi { host } => {
+                crate::auth::kimi_oauth::request_device_authorization(host).await
+            }
+            Self::Generic(cfg) | Self::GithubCopilot(cfg) => {
+                crate::auth::oauth_device::request_device_authorization(cfg).await
+            }
+        }
+    }
+    async fn poll(&self, device_code: &str) -> anyhow::Result<DevicePollResult> {
+        match self {
+            Self::Kimi { host } => {
+                crate::auth::kimi_oauth::poll_device_token(host, device_code).await
+            }
+            Self::Generic(cfg) => {
+                crate::auth::oauth_device::poll_device_token(cfg, device_code).await
+            }
+            Self::GithubCopilot(cfg) => {
+                crate::auth::github_copilot::poll_github_device_token(cfg, device_code).await
+            }
+        }
+    }
+    /// Transform the device-grant credential before it is persisted. The Kimi
+    /// and generic flows persist the poll result verbatim; the GitHub Copilot
+    /// flow exchanges the durable github token (in `auth.key`) for the
+    /// short-lived copilot token, persisting BOTH (copilot as `key`, github as
+    /// `refresh_token`).
+    async fn finalize(&self, auth: KimiAuth) -> anyhow::Result<KimiAuth> {
+        match self {
+            Self::Kimi { .. } | Self::Generic(_) => Ok(auth),
+            Self::GithubCopilot(cfg) => {
+                crate::auth::github_copilot::exchange_copilot_token(cfg, &auth.key).await
+            }
+        }
+    }
+}
+
 /// Outcome of one full poll loop over a single device authorization.
 enum PollLoopOutcome {
     /// Access token issued.
@@ -40,11 +95,46 @@ pub async fn run_device_code_login_channels(
     host: &str,
     auth_manager: &Arc<AuthManager>,
     channels: &mut Option<AuthChannels>,
+) -> anyhow::Result<(KimiAuth, bool)> {
+    run_device_code_login_backend(DeviceFlowBackend::Kimi { host }, auth_manager, channels).await
+}
+
+/// Device-code login for a GENERIC [`OAuthConfig`] provider (xai-grok). Same
+/// TUI/CLI presentation as the Kimi login; only the wire differs.
+pub async fn run_device_code_login_generic(
+    oauth: &OAuthConfig,
+    auth_manager: &Arc<AuthManager>,
+    channels: &mut Option<AuthChannels>,
+) -> anyhow::Result<(KimiAuth, bool)> {
+    run_device_code_login_backend(DeviceFlowBackend::Generic(oauth), auth_manager, channels).await
+}
+
+/// GitHub Copilot two-stage login (github-copilot): the same device-flow
+/// presentation as the generic path, but the token poll reads GitHub's 200-body
+/// errors and the minted github token is finalized into a copilot session token
+/// before it is persisted (see [`DeviceFlowBackend::finalize`]).
+pub async fn run_device_code_login_github_copilot(
+    oauth: &OAuthConfig,
+    auth_manager: &Arc<AuthManager>,
+    channels: &mut Option<AuthChannels>,
+) -> anyhow::Result<(KimiAuth, bool)> {
+    run_device_code_login_backend(
+        DeviceFlowBackend::GithubCopilot(oauth),
+        auth_manager,
+        channels,
+    )
+    .await
+}
+
+async fn run_device_code_login_backend(
+    backend: DeviceFlowBackend<'_>,
+    auth_manager: &Arc<AuthManager>,
+    channels: &mut Option<AuthChannels>,
 ) -> anyhow::Result<(KimiAuth, bool)> {
     let interactive_tui = channels.is_some();
     let mut channels = channels.take();
     loop {
-        let device_auth = request_device_authorization(host).await?;
+        let device_auth = backend.request().await?;
         let display_uri = device_auth.verification_uri_complete.clone();
 
         if interactive_tui {
@@ -62,10 +152,14 @@ pub async fn run_device_code_login_channels(
             prompt_on_stderr(&device_auth).await;
         }
 
-        match complete_device_code_login(host, &device_auth).await? {
+        match complete_device_code_login(&backend, &device_auth).await? {
             PollLoopOutcome::Done(auth) => {
+                // Finalize before persisting: the GitHub Copilot flow exchanges
+                // the durable github token for the short-lived copilot token
+                // here; the Kimi / generic flows pass the credential through.
+                let auth = backend.finalize(*auth).await?;
                 let auth = auth_manager
-                    .update(*auth)
+                    .update(auth)
                     .await
                     .map_err(|e| anyhow::anyhow!("Failed to save credentials: {e}"))?;
                 return Ok((auth, true));
@@ -112,7 +206,7 @@ async fn prompt_on_stderr(device_auth: &DeviceAuthorization) {
 /// Poll the token endpoint until the user approves, the device code expires
 /// (→ [`PollLoopOutcome::Restart`]), or the wire fails.
 async fn complete_device_code_login(
-    host: &str,
+    backend: &DeviceFlowBackend<'_>,
     device_auth: &DeviceAuthorization,
 ) -> anyhow::Result<PollLoopOutcome> {
     let mut poll_interval = std::time::Duration::from_secs(device_auth.interval.max(1) as u64);
@@ -120,7 +214,7 @@ async fn complete_device_code_login(
         // Sleep first: an immediate poll on a fresh code only returns
         // authorization_pending (and risks slow_down).
         tokio::time::sleep(poll_interval).await;
-        match poll_device_token(host, &device_auth.device_code).await? {
+        match backend.poll(&device_auth.device_code).await? {
             DevicePollResult::Success(auth) => {
                 tracing::info!("auth: device login authorized");
                 return Ok(PollLoopOutcome::Done(auth));
@@ -148,8 +242,8 @@ async fn complete_device_code_login(
 /// Open `url` in the browser off-thread: `webbrowser::open` is synchronous and
 /// would stall the single-threaded TUI loop. Returns `true` on success so the
 /// caller can decide how to notify the user (eprintln on CLI, nothing on TUI
-/// where the URL is already rendered in the widget).
-async fn open_browser_detached(url: &str) -> bool {
+/// where the URL is already rendered in the widget). Shared with the PKCE flow.
+pub(super) async fn open_browser_detached(url: &str) -> bool {
     // Unit tests drive the full login flow against mock servers — their
     // fixture URLs must never reach a real browser.
     if cfg!(test) {
@@ -338,6 +432,100 @@ mod tests {
         );
     }
 
+    /// GitHub Copilot two-stage login e2e (mock wire): device authorization →
+    /// poll (pending → github token) → Stage-2 copilot-token exchange (Bearer
+    /// github token + editor headers → copilot token + expiry). The persisted
+    /// credential keys the COPILOT token, keeps the GITHUB token as
+    /// `refresh_token`, and carries the copilot expiry.
+    #[tokio::test]
+    async fn github_copilot_two_stage_login_persists_copilot_and_github() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        let cfg = OAuthConfig {
+            auth_host: host,
+            token_host: host,
+            copilot_exchange: Some((host, "/copilot_internal/v2/token")),
+            ..kigi_models::COPILOT_OAUTH_CONFIG
+        };
+        // Stage 1a: device authorization (github.com/login/device/code).
+        Mock::given(method("POST"))
+            .and(path("/login/device/code"))
+            .and(body_string_contains("client_id=Iv1.b507a08c87ecfe98"))
+            .and(body_string_contains("scope=read"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "device_code": "gh-dev-1",
+                "user_code": "WDJB-MJHT",
+                "verification_uri": "https://github.com/login/device",
+                "expires_in": 900,
+                "interval": 0,
+            })))
+            .mount(&server)
+            .await;
+        // Stage 1b: token poll — GitHub returns errors AND success in a 200 body.
+        Mock::given(method("POST"))
+            .and(path("/login/oauth/access_token"))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "error": "authorization_pending" })),
+            )
+            .up_to_n_times(1)
+            .mount(&server)
+            .await;
+        Mock::given(method("POST"))
+            .and(path("/login/oauth/access_token"))
+            .and(body_string_contains("device_code=gh-dev-1"))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "access_token": "gho_github_tok" })),
+            )
+            .mount(&server)
+            .await;
+        // Stage 2: copilot-token exchange (Bearer github token + editor headers).
+        let future = (chrono::Utc::now() + chrono::Duration::minutes(30)).timestamp();
+        Mock::given(method("GET"))
+            .and(path("/copilot_internal/v2/token"))
+            .and(wiremock::matchers::header(
+                "Authorization",
+                "Bearer gho_github_tok",
+            ))
+            .and(wiremock::matchers::header(
+                "Editor-Plugin-Version",
+                "copilot-chat/0.35.0",
+            ))
+            .respond_with(ResponseTemplate::new(200).set_body_json(
+                serde_json::json!({ "token": "copilot-session-tok", "expires_at": future }),
+            ))
+            .expect(1)
+            .mount(&server)
+            .await;
+
+        let dir = tempfile::tempdir().unwrap();
+        let mgr = auth_manager(&dir);
+        let mut channels = None;
+        let (auth, is_new) = run_device_code_login_github_copilot(&cfg, &mgr, &mut channels)
+            .await
+            .unwrap();
+        assert!(is_new);
+        assert_eq!(
+            auth.key, "copilot-session-tok",
+            "key = copilot session token"
+        );
+        assert_eq!(
+            auth.refresh_token.as_deref(),
+            Some("gho_github_tok"),
+            "the durable github token is persisted as refresh_token"
+        );
+        assert!(
+            auth.expires_at.is_some_and(|e| e > chrono::Utc::now()),
+            "the copilot expiry must be persisted"
+        );
+        assert_eq!(
+            mgr.current_or_expired().map(|a| a.key),
+            Some("copilot-session-tok".into()),
+            "login must land the copilot token in the manager cache"
+        );
+    }
+
     /// A 5xx from the token endpoint is a hard error (kimi-cli parity).
     #[tokio::test]
     async fn server_error_during_poll_fails_login() {

crates/codegen/kigi-shell/src/auth/flow.rs

@@ -67,6 +67,178 @@ pub async fn run_auth_flow(
     run_auth_flow_inner(auth_manager, kimi_code_config, reauth, false, channels).await
 }
 
+/// Login flow for a GENERIC OAuth provider (xai-grok device-code,
+/// claude-pro-max PKCE-localhost): use a valid cached session unless re-authing,
+/// otherwise dispatch by `oauth.flow` to the device-code or PKCE-localhost login
+/// (persisting under the provider's own scope via `auth_manager`). Unlike the
+/// Kimi flow this does not run the silent-refresh dance — the login's
+/// `AuthManager::update` persists a fresh token set directly.
+pub async fn run_oauth_provider_flow(
+    auth_manager: &Arc<AuthManager>,
+    oauth: &'static kigi_models::OAuthConfig,
+    reauth: bool,
+    channels: Option<AuthChannels>,
+) -> anyhow::Result<(KimiAuth, bool)> {
+    tracing::info!(
+        scope_key = oauth.scope_key,
+        reauth,
+        "auth: starting generic oauth login"
+    );
+    if reauth {
+        auth_manager.clear()?;
+    }
+    if !reauth && let Some(auth) = auth_manager.current() {
+        tracing::info!(
+            scope_key = oauth.scope_key,
+            "auth: using cached oauth session"
+        );
+        return Ok((auth, false));
+    }
+    let mut channels = channels;
+    match oauth.flow {
+        kigi_models::OAuthFlow::DeviceCode => {
+            crate::auth::device_code::run_device_code_login_generic(
+                oauth,
+                auth_manager,
+                &mut channels,
+            )
+            .await
+        }
+        kigi_models::OAuthFlow::PkceLocalhost {
+            redirect_port,
+            redirect_path,
+        } => {
+            run_pkce_localhost_login(
+                oauth,
+                redirect_port,
+                redirect_path,
+                auth_manager,
+                &mut channels,
+            )
+            .await
+        }
+        kigi_models::OAuthFlow::GithubDeviceCopilot => {
+            crate::auth::device_code::run_device_code_login_github_copilot(
+                oauth,
+                auth_manager,
+                &mut channels,
+            )
+            .await
+        }
+    }
+}
+
+/// PKCE-localhost login (claude-pro-max JSON, openai-codex FORM): generate PKCE,
+/// present the browser authorize URL (TUI channel or stderr), open the browser,
+/// then await the code from EITHER the `127.0.0.1:{redirect_port}{redirect_path}`
+/// loopback callback OR a manual paste (headless fallback). Exchange it at the
+/// token endpoint and persist.
+///
+/// The `token_body` selects the wire dialect: `Json` is the claude path
+/// (`state == verifier`, JSON exchange carrying `state`); `Form` is the codex
+/// path (fresh-random `state`, FORM exchange WITHOUT `state`, then a FAIL-FAST
+/// check that the minted JWT carries a `chatgpt_account_id` — a token without it
+/// is useless for inference, so the login bails rather than persisting it).
+///
+/// SECURITY: the verifier / code / tokens / JWT / account id are never logged;
+/// the loopback binds `127.0.0.1` only and validates `state` strictly.
+async fn run_pkce_localhost_login(
+    oauth: &'static kigi_models::OAuthConfig,
+    redirect_port: u16,
+    redirect_path: &'static str,
+    auth_manager: &Arc<AuthManager>,
+    channels: &mut Option<AuthChannels>,
+) -> anyhow::Result<(KimiAuth, bool)> {
+    use crate::auth::oauth_pkce;
+
+    // The token-body encoding also selects the PKCE `state` convention: the JSON
+    // dialect is Claude's/Pi's `state == verifier`, while form-encoded endpoints
+    // (the OAuth norm) get an INDEPENDENT random state. A new form provider
+    // inheriting the standard random state is correct by default.
+    let uses_form_exchange = matches!(oauth.token_body, kigi_models::OAuthTokenBody::Form);
+    let pkce = if uses_form_exchange {
+        oauth_pkce::generate_pkce_random_state()
+    } else {
+        oauth_pkce::generate_pkce()
+    };
+    let redirect = oauth_pkce::redirect_uri(redirect_port, redirect_path);
+    let authorize_url = oauth_pkce::build_authorize_url(oauth, &redirect, &pkce);
+
+    let mut chans = channels.take();
+    if let Some(tx) = chans.as_mut().and_then(|c| c.url_tx.take()) {
+        // TUI: push the URL BEFORE opening the browser (never block the UI on a
+        // slow/headless browser launch).
+        let _ = tx.send(AuthUrlInfo {
+            url: authorize_url.clone(),
+            mode: AuthUrlMode::Device,
+        });
+        crate::auth::device_code::open_browser_detached(&authorize_url).await;
+    } else {
+        eprintln!();
+        eprintln!("To sign in, open this URL in your browser:");
+        eprintln!();
+        eprintln!("  {authorize_url}");
+        eprintln!();
+        if !crate::auth::device_code::open_browser_detached(&authorize_url).await {
+            eprintln!("  (Could not open the browser automatically — open the URL above.)");
+            eprintln!();
+        }
+        eprintln!("Waiting for the sign-in to complete...");
+    }
+
+    let code = await_pkce_code(redirect_port, redirect_path, &pkce, chans.as_mut()).await?;
+    let auth = if uses_form_exchange {
+        oauth_pkce::exchange_code_form(oauth, &code, &pkce, &redirect).await?
+    } else {
+        oauth_pkce::exchange_code(oauth, &code, &pkce, &redirect).await?
+    };
+    // FAIL FAST: an access token that yields no chatgpt_account_id cannot
+    // authorize inference (it becomes the `chatgpt-account-id` header) — bail
+    // rather than persist a dead session. Gated on the EXPLICIT provider fact,
+    // never on the token-body encoding.
+    if oauth.requires_chatgpt_account_id
+        && kigi_sampling_types::chatgpt_account_id_from_jwt(&auth.key).is_none()
+    {
+        anyhow::bail!(
+            "ChatGPT login did not return a usable account id \
+             (the access token is missing the chatgpt_account_id claim)"
+        );
+    }
+    let auth = auth_manager
+        .update(auth)
+        .await
+        .map_err(|e| anyhow::anyhow!("Failed to save credentials: {e}"))?;
+    Ok((auth, true))
+}
+
+/// Await the authorization code: the loopback callback is primary; when a TUI
+/// channel is present, a pasted code (redirect URL / `code#state` / bare code)
+/// is accepted concurrently as a headless fallback. State is validated in both
+/// arms (strict on the loopback, mismatch-rejecting on the paste).
+async fn await_pkce_code(
+    redirect_port: u16,
+    redirect_path: &str,
+    pkce: &crate::auth::oauth_pkce::PkceCodes,
+    channels: Option<&mut AuthChannels>,
+) -> anyhow::Result<String> {
+    use crate::auth::oauth_pkce;
+    match channels {
+        Some(ch) => {
+            tokio::select! {
+                code = oauth_pkce::await_loopback_code(redirect_port, redirect_path, &pkce.state) => code,
+                pasted = ch.code_rx.recv() => {
+                    let pasted = pasted
+                        .ok_or_else(|| anyhow::anyhow!("auth code channel closed before a code arrived"))?;
+                    let params = oauth_pkce::parse_manual_paste(&pasted)?;
+                    oauth_pkce::validate_pasted_state(&params, &pkce.state)?;
+                    Ok(params.code)
+                }
+            }
+        }
+        None => oauth_pkce::await_loopback_code(redirect_port, redirect_path, &pkce.state).await,
+    }
+}
+
 async fn run_auth_flow_inner(
     auth_manager: &Arc<AuthManager>,
     _kimi_code_config: &KimiCodeConfig,

crates/codegen/kigi-shell/src/auth/github_copilot.rs

@@ -0,0 +1,485 @@
+//! GitHub Copilot two-stage OAuth wire (github-copilot), driven by a registry
+//! [`kigi_models::OAuthConfig`] whose `flow` is [`OAuthFlow::GithubDeviceCopilot`].
+//!
+//! Stage 1 — RFC-8628 device flow on `auth_host` (github.com). The device
+//! authorization POST is the generic one ([`super::oauth_device`]); the token
+//! POLL is Copilot-specific because GitHub returns its device errors in a `200`
+//! body (`{error: "authorization_pending"|"slow_down"|"expired_token"}`), not a
+//! `4xx`, and the success payload carries ONLY `access_token` (the DURABLE
+//! GitHub token — no refresh token, no expiry).
+//!
+//! Stage 2 — copilot-token exchange: `GET {copilot_exchange}` bearing the GitHub
+//! token + the editor headers re-mints the SHORT-LIVED copilot session token
+//! (`{token, expires_at}`). This runs at login ([`exchange_copilot_token`]) and
+//! on every "refresh" ([`remint_copilot_token`], dispatched by the generic
+//! refresher) — the github token is unchanged and re-persisted as the
+//! `refresh_token`; the copilot token becomes the `key`.
+//!
+//! SECURITY: the github token and the copilot token are NEVER logged (only
+//! non-secret events: poll succeeded, copilot token minted/re-minted).
+
+use chrono::{DateTime, Utc};
+use kigi_models::OAuthConfig;
+use serde::Deserialize;
+
+use super::kimi_oauth::{DevicePollResult, RefreshError};
+use super::model::{AuthMode, KimiAuth};
+
+/// RFC-8628 device grant type (shared with the generic device wire).
+const DEVICE_GRANT_TYPE: &str = "urn:ietf:params:oauth:grant-type:device_code";
+/// Copilot-exchange retry budget over 5xx / network blips (parity with the
+/// device/PKCE refresh wires); 401/403 fails fast (the github token is dead).
+const MAX_EXCHANGE_RETRIES: u32 = 3;
+const RETRYABLE_EXCHANGE_STATUSES: [u16; 5] = [429, 500, 502, 503, 504];
+
+/// The four VS Code Copilot editor-identity headers every Copilot request
+/// carries. Non-secret wire constants owned by `kigi_sampling_types` (the
+/// single source shared with the sampler's inference gate).
+fn editor_headers() -> [(&'static str, &'static str); 4] {
+    [
+        ("User-Agent", kigi_sampling_types::COPILOT_USER_AGENT),
+        (
+            "Editor-Version",
+            kigi_sampling_types::COPILOT_EDITOR_VERSION,
+        ),
+        (
+            "Editor-Plugin-Version",
+            kigi_sampling_types::COPILOT_EDITOR_PLUGIN_VERSION,
+        ),
+        (
+            "Copilot-Integration-Id",
+            kigi_sampling_types::COPILOT_INTEGRATION_ID,
+        ),
+    ]
+}
+
+/// GitHub's device-token poll response: EITHER `access_token` (the durable
+/// GitHub token) OR an `error` (in a `200` body). No refresh token / expiry.
+#[derive(Deserialize, Default)]
+struct GithubDeviceTokenResponse {
+    #[serde(default)]
+    access_token: Option<String>,
+    #[serde(default)]
+    error: Option<String>,
+}
+
+/// One poll of `POST {auth_host}{token_path}` (github.com/login/oauth/
+/// access_token) with the device grant. GitHub answers `200` for BOTH success
+/// and the pending/slow_down/expired errors, so the outcome is read from the
+/// body, not the status. On success the [`KimiAuth`] carries the GitHub token as
+/// `key` with NO refresh token / expiry — the caller finalizes it via the
+/// copilot exchange before persisting.
+pub(crate) async fn poll_github_device_token(
+    cfg: &OAuthConfig,
+    device_code: &str,
+) -> anyhow::Result<DevicePollResult> {
+    let url = format!("{}{}", cfg.auth_host.trim_end_matches('/'), cfg.token_path);
+    let resp = crate::http::shared_client()
+        .post(&url)
+        .header("Accept", "application/json")
+        .header("User-Agent", kigi_sampling_types::COPILOT_USER_AGENT)
+        .form(&[
+            ("client_id", cfg.client_id),
+            ("device_code", device_code),
+            ("grant_type", DEVICE_GRANT_TYPE),
+        ])
+        .send()
+        .await
+        .map_err(|e| anyhow::anyhow!("Token polling request failed: {e}"))?;
+
+    let status = resp.status();
+    if status.is_server_error() {
+        anyhow::bail!("Token polling server error: {status}");
+    }
+    let body = resp.bytes().await?;
+    let parsed: GithubDeviceTokenResponse = serde_json::from_slice(&body).unwrap_or_default();
+
+    if let Some(access) = parsed.access_token.filter(|t| !t.is_empty()) {
+        tracing::info!("auth: github device poll succeeded, github token issued (copilot)");
+        return Ok(DevicePollResult::Success(Box::new(github_token_auth(
+            access,
+        ))));
+    }
+    match parsed.error.as_deref() {
+        Some("expired_token") => {
+            tracing::info!("auth: github device code expired; restarting (copilot)");
+            Ok(DevicePollResult::Expired)
+        }
+        Some(error) => {
+            tracing::debug!(error, "auth: github device poll pending (copilot)");
+            Ok(DevicePollResult::Pending {
+                error: error.to_owned(),
+                description: None,
+            })
+        }
+        None => Ok(DevicePollResult::Pending {
+            error: "missing_access_token".to_owned(),
+            description: None,
+        }),
+    }
+}
+
+/// A transient [`KimiAuth`] holding ONLY the durable GitHub token (no refresh
+/// token / expiry) — the intermediate device-flow result, finalized by the
+/// copilot exchange before it is ever persisted.
+fn github_token_auth(github_token: String) -> KimiAuth {
+    KimiAuth {
+        key: github_token,
+        auth_mode: AuthMode::OAuth,
+        create_time: Utc::now(),
+        user_id: String::new(),
+        email: None,
+        refresh_token: None,
+        expires_at: None,
+        expires_in: None,
+        scope: None,
+        token_type: None,
+    }
+}
+
+/// Stage-2 copilot-token exchange response (`GET copilot_internal/v2/token`).
+/// `endpoints`/`proxy-ep` are ignored: Kigi resolves the base URL from the
+/// platform registry (the individual-subscription endpoint, or the
+/// `KIGI_COPILOT_BASE_URL` override).
+#[derive(Deserialize)]
+struct CopilotTokenResponse {
+    token: String,
+    /// Unix seconds when the copilot token expires (~30 min out).
+    expires_at: i64,
+}
+
+/// Materialize the persisted credential from a copilot-token exchange:
+/// `key` = the short-lived copilot token, `refresh_token` = the DURABLE github
+/// token (so every re-mint re-exchanges it), `expires_at` = the copilot expiry.
+fn copilot_auth(resp: CopilotTokenResponse, github_token: &str) -> anyhow::Result<KimiAuth> {
+    let now = Utc::now();
+    // FAIL-FAST: an uninterpretable expiry means we cannot schedule the re-mint,
+    // so reject it rather than silently falling back to a long default TTL — which
+    // would let the ~30-min copilot token 401 on the wire ~30 min later.
+    let expires_at = DateTime::from_timestamp(resp.expires_at, 0).ok_or_else(|| {
+        anyhow::anyhow!(
+            "copilot token has an out-of-range expires_at: {}",
+            resp.expires_at
+        )
+    })?;
+    // The manager's dynamic threshold (`max(300, expires_in × 0.5)`) drives the
+    // proactive re-mint; `expires_in` is the copilot token's remaining life.
+    let expires_in = (expires_at - now).num_seconds();
+    Ok(KimiAuth {
+        key: resp.token,
+        auth_mode: AuthMode::OAuth,
+        create_time: now,
+        user_id: String::new(),
+        email: None,
+        refresh_token: Some(github_token.to_owned()),
+        expires_at: Some(expires_at),
+        expires_in: Some(expires_in),
+        scope: None,
+        token_type: Some("bearer".to_owned()),
+    })
+}
+
+/// The `(host, path)` of the copilot-token exchange endpoint, or a fatal error
+/// when the config lacks it (a non-Copilot config reaching this wire is a bug).
+fn exchange_url(cfg: &OAuthConfig) -> anyhow::Result<String> {
+    let (host, path) = cfg.copilot_exchange.ok_or_else(|| {
+        anyhow::anyhow!("github-copilot config missing copilot_exchange endpoint")
+    })?;
+    Ok(format!("{}{path}", host.trim_end_matches('/')))
+}
+
+/// `GET {copilot_exchange}` bearing the github token + editor headers.
+async fn send_copilot_exchange(
+    cfg: &OAuthConfig,
+    github_token: &str,
+) -> anyhow::Result<reqwest::Response> {
+    let url = exchange_url(cfg)?;
+    let mut req = crate::http::shared_client()
+        .get(&url)
+        .header("Accept", "application/json")
+        .header("Authorization", format!("Bearer {github_token}"));
+    for (name, value) in editor_headers() {
+        req = req.header(name, value);
+    }
+    req.send()
+        .await
+        .map_err(|e| anyhow::anyhow!("copilot-token exchange request failed: {e}"))
+}
+
+/// Exchange the durable GitHub token for a copilot session token (login path).
+/// FAIL-FAST: a non-2xx response aborts login (never a silent fallback).
+pub(crate) async fn exchange_copilot_token(
+    cfg: &OAuthConfig,
+    github_token: &str,
+) -> anyhow::Result<KimiAuth> {
+    tracing::info!(
+        scope_key = cfg.scope_key,
+        "auth: exchanging github token for copilot token"
+    );
+    let resp = send_copilot_exchange(cfg, github_token).await?;
+    let status = resp.status();
+    if !status.is_success() {
+        let body = resp.text().await.unwrap_or_default();
+        tracing::warn!(%status, scope_key = cfg.scope_key, "auth: copilot-token exchange failed");
+        anyhow::bail!("Copilot token exchange failed (HTTP {status}): {body}");
+    }
+    let parsed: CopilotTokenResponse = resp
+        .json()
+        .await
+        .map_err(|e| anyhow::anyhow!("malformed copilot token payload: {e}"))?;
+    tracing::info!(scope_key = cfg.scope_key, "auth: copilot token minted");
+    copilot_auth(parsed, github_token)
+}
+
+/// Re-mint the copilot token from the durable GitHub token (refresh path,
+/// dispatched from the generic refresher). This is NOT a `refresh_token` grant:
+/// it re-runs the copilot exchange. `refresh_token` is the github token; the
+/// returned [`KimiAuth`] preserves it. Retries 5xx / network blips; 401/403
+/// (the github token is revoked) fails fast as [`RefreshError::Unauthorized`].
+pub(crate) async fn remint_copilot_token(
+    cfg: &OAuthConfig,
+    github_token: &str,
+) -> Result<KimiAuth, RefreshError> {
+    let mut last_error = String::from("no attempt made");
+    for attempt in 0..MAX_EXCHANGE_RETRIES {
+        if attempt > 0 {
+            let backoff = std::time::Duration::from_secs(1 << (attempt - 1));
+            tracing::warn!(
+                attempt,
+                backoff_secs = backoff.as_secs(),
+                "auth: retrying copilot-token re-mint"
+            );
+            tokio::time::sleep(backoff).await;
+        }
+        let resp = match send_copilot_exchange(cfg, github_token).await {
+            Ok(resp) => resp,
+            Err(e) => {
+                last_error = format!("{e}");
+                continue;
+            }
+        };
+        let status = resp.status().as_u16();
+        let bytes = resp.bytes().await.unwrap_or_default();
+        if status == 401 || status == 403 {
+            return Err(RefreshError::Unauthorized {
+                status,
+                description: "GitHub token rejected at copilot-token exchange.".to_owned(),
+            });
+        }
+        if status == 200 {
+            return match serde_json::from_slice::<CopilotTokenResponse>(&bytes) {
+                Ok(parsed) => match copilot_auth(parsed, github_token) {
+                    Ok(auth) => {
+                        tracing::info!(scope_key = cfg.scope_key, "auth: copilot token re-minted");
+                        Ok(auth)
+                    }
+                    Err(e) => Err(RefreshError::Fatal {
+                        status,
+                        description: format!("{e}"),
+                    }),
+                },
+                Err(e) => Err(RefreshError::Fatal {
+                    status,
+                    description: format!("malformed copilot token payload: {e}"),
+                }),
+            };
+        }
+        let description = format!("copilot-token exchange failed (HTTP {status}).");
+        if RETRYABLE_EXCHANGE_STATUSES.contains(&status) {
+            last_error = description;
+            continue;
+        }
+        return Err(RefreshError::Fatal {
+            status,
+            description,
+        });
+    }
+    Err(RefreshError::Exhausted { last_error })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use chrono::Duration;
+    use kigi_models::COPILOT_OAUTH_CONFIG;
+    use wiremock::matchers::{body_string_contains, header, method, path};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    /// A COPILOT_OAUTH_CONFIG pointed at a mock server for both stages.
+    fn mock_cfg(host: &'static str, exchange: &'static str) -> OAuthConfig {
+        OAuthConfig {
+            auth_host: host,
+            token_host: host,
+            copilot_exchange: Some((exchange, "/copilot_internal/v2/token")),
+            ..COPILOT_OAUTH_CONFIG
+        }
+    }
+
+    /// GitHub's device poll returns pending errors in a 200 body — mapped to
+    /// Pending (authorization_pending / slow_down) and Expired (expired_token),
+    /// never mis-read as a token.
+    #[tokio::test]
+    async fn github_device_poll_maps_200_body_errors() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/login/oauth/access_token"))
+            .and(body_string_contains("grant_type=urn"))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "error": "authorization_pending" })),
+            )
+            .mount(&server)
+            .await;
+        let result = poll_github_device_token(&mock_cfg(host, host), "dev-1")
+            .await
+            .unwrap();
+        assert!(
+            matches!(result, DevicePollResult::Pending { error, .. } if error == "authorization_pending")
+        );
+    }
+
+    #[tokio::test]
+    async fn github_device_poll_expired_restarts() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/login/oauth/access_token"))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "error": "expired_token" })),
+            )
+            .mount(&server)
+            .await;
+        let result = poll_github_device_token(&mock_cfg(host, host), "dev-1")
+            .await
+            .unwrap();
+        assert!(matches!(result, DevicePollResult::Expired));
+    }
+
+    /// A successful poll yields the DURABLE github token as `key` with NO
+    /// refresh token / expiry (the copilot exchange finalizes it next).
+    #[tokio::test]
+    async fn github_device_poll_success_is_bare_github_token() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/login/oauth/access_token"))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .set_body_json(serde_json::json!({ "access_token": "gho_github_tok" })),
+            )
+            .mount(&server)
+            .await;
+        let DevicePollResult::Success(auth) = poll_github_device_token(&mock_cfg(host, host), "d")
+            .await
+            .unwrap()
+        else {
+            panic!("expected success");
+        };
+        assert_eq!(auth.key, "gho_github_tok");
+        assert_eq!(
+            auth.refresh_token, None,
+            "github token is not a refresh grant"
+        );
+        assert_eq!(auth.expires_at, None, "the github token is long-lived");
+    }
+
+    /// The Stage-2 exchange rides the github Bearer + editor headers and maps
+    /// `{token, expires_at}` onto `key=copilot`, `refresh_token=github`, with a
+    /// future `expires_at`.
+    #[tokio::test]
+    async fn copilot_exchange_maps_token_and_persists_github_as_refresh() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        let future = (Utc::now() + Duration::minutes(30)).timestamp();
+        Mock::given(method("GET"))
+            .and(path("/copilot_internal/v2/token"))
+            .and(header("Authorization", "Bearer gho_github_tok"))
+            .and(header("Editor-Version", "vscode/1.107.0"))
+            .and(header("Copilot-Integration-Id", "vscode-chat"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(
+                serde_json::json!({ "token": "tid=abc;copilot-tok", "expires_at": future }),
+            ))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let auth = exchange_copilot_token(&mock_cfg(host, host), "gho_github_tok")
+            .await
+            .unwrap();
+        assert_eq!(auth.key, "tid=abc;copilot-tok", "key = copilot token");
+        assert_eq!(
+            auth.refresh_token.as_deref(),
+            Some("gho_github_tok"),
+            "the durable github token is persisted as refresh_token"
+        );
+        assert!(
+            auth.expires_at.is_some_and(|e| e > Utc::now()),
+            "copilot expiry must be in the future"
+        );
+    }
+
+    /// The copilot re-mint (refresh) re-exchanges the github token for a NEW
+    /// copilot token, keeping the github token as refresh_token.
+    #[tokio::test]
+    async fn copilot_remint_returns_new_copilot_token() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        let future = (Utc::now() + Duration::minutes(30)).timestamp();
+        Mock::given(method("GET"))
+            .and(path("/copilot_internal/v2/token"))
+            .and(header("Authorization", "Bearer gho_github_tok"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(
+                serde_json::json!({ "token": "copilot-tok-2", "expires_at": future }),
+            ))
+            .mount(&server)
+            .await;
+        let auth = remint_copilot_token(&mock_cfg(host, host), "gho_github_tok")
+            .await
+            .unwrap();
+        assert_eq!(auth.key, "copilot-tok-2");
+        assert_eq!(auth.refresh_token.as_deref(), Some("gho_github_tok"));
+    }
+
+    /// A 401 at the exchange (github token revoked) fails fast as Unauthorized
+    /// (drives the manager's permanent-failure / re-login path).
+    #[tokio::test]
+    async fn copilot_remint_401_is_unauthorized() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("GET"))
+            .and(path("/copilot_internal/v2/token"))
+            .respond_with(ResponseTemplate::new(401))
+            .mount(&server)
+            .await;
+        let err = remint_copilot_token(&mock_cfg(host, host), "dead-github-tok")
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(err, RefreshError::Unauthorized { status: 401, .. }),
+            "got {err:?}"
+        );
+    }
+
+    /// FAIL-FAST: an out-of-range `expires_at` is rejected rather than silently
+    /// degrading to a long default TTL (which would 401 on the wire ~30 min in).
+    #[tokio::test]
+    async fn copilot_exchange_rejects_out_of_range_expiry() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("GET"))
+            .and(path("/copilot_internal/v2/token"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(
+                serde_json::json!({ "token": "copilot-tok", "expires_at": i64::MAX }),
+            ))
+            .mount(&server)
+            .await;
+        let err = exchange_copilot_token(&mock_cfg(host, host), "gho_github_tok")
+            .await
+            .unwrap_err();
+        assert!(
+            format!("{err}").contains("out-of-range expires_at"),
+            "got {err}"
+        );
+    }
+}

crates/codegen/kigi-shell/src/auth/kimi_oauth.rs

@@ -130,8 +130,9 @@ fn with_device_headers(
 }
 
 /// Defend against control characters / non-https redirects from a
-/// compromised or mis-configured OAuth host.
-fn validate_verification_uri(uri: &str) -> anyhow::Result<()> {
+/// compromised or mis-configured OAuth host. Shared with the generic
+/// device-code wire ([`super::oauth_device`]).
+pub(crate) fn validate_verification_uri(uri: &str) -> anyhow::Result<()> {
     if uri.chars().any(|c| c.is_ascii_control()) {
         anyhow::bail!("Server returned invalid verification URI");
     }

crates/codegen/kigi-shell/src/auth/manager.rs

@@ -12,6 +12,7 @@ use tokio_util::sync::CancellationToken;
 
 #[path = "manager/lock.rs"]
 mod lock;
+pub(crate) use lock::try_lock_auth_file_nonblocking;
 #[path = "manager/sleep_gate.rs"]
 mod sleep_gate;
 
@@ -384,6 +385,59 @@ impl AuthManager {
         )
     }
 
+    /// Build a manager for a GENERIC device-code OAuth provider (xai-grok),
+    /// scoped to `oauth.scope_key`. Unlike [`Self::new`] this path is
+    /// file-store only (no keyring — that is gated to the default Kimi install)
+    /// and ignores the Kimi-specific `KIGI_AUTH` inline-credential env; it
+    /// otherwise shares the same multi-scope `auth.json` (honoring
+    /// `KIGI_AUTH_PATH`). The refresher is selected from the scope by
+    /// [`super::refresh::build_refresher`].
+    pub(crate) fn new_oauth_provider(
+        kigi_home: &Path,
+        oauth: &'static kigi_models::OAuthConfig,
+    ) -> Self {
+        let scope = oauth.scope_key.to_owned();
+        let path = std::env::var("KIGI_AUTH_PATH")
+            .map(PathBuf::from)
+            .unwrap_or_else(|_| kigi_home.join("auth.json"));
+
+        let (auth, disk_state) = match read_auth_json(&path) {
+            Ok(map) => {
+                let found = lookup_auth(&map, &scope);
+                let state = if found.is_some() {
+                    DiskAuthState::Ok
+                } else {
+                    DiskAuthState::EntryMissing
+                };
+                (found, state)
+            }
+            Err(e) => {
+                let state = if e.kind() == std::io::ErrorKind::NotFound {
+                    DiskAuthState::FileMissing
+                } else {
+                    DiskAuthState::Unreadable
+                };
+                (None, state)
+            }
+        };
+        kigi_log::unified_log::info(
+            "AuthManager::new_oauth_provider",
+            None,
+            Some(serde_json::json!({
+                "scope": &scope,
+                "found": auth.is_some(),
+                "is_expired": auth.as_ref().map(is_expired),
+            })),
+        );
+        Self::assemble(
+            auth,
+            path,
+            scope,
+            KimiCodeConfig::default(),
+            Some(disk_state),
+        )
+    }
+
     /// Single field-assembly point for [`Self::new`]'s two construction paths
     /// (inline `KIGI_AUTH` vs. on-disk `auth.json`), which differ only in the
     /// threaded fields. One literal means a newly added field can't be silently
@@ -788,6 +842,13 @@ impl AuthManager {
         &self.kimi_code_config
     }
 
+    /// The auth.json / keyring scope key this manager persists under
+    /// (`oauth/kimi-code` for Kimi, `oauth/xai` for xai-grok, …). Drives the
+    /// refresher selection in [`super::refresh::build_refresher`].
+    pub(crate) fn scope(&self) -> &str {
+        &self.scope
+    }
+
     /// Handle notified after every successful token refresh.
     ///
     /// Used by [`ModelsManager`] to trigger model catalog recovery

crates/codegen/kigi-shell/src/auth/mod.rs

@@ -1,13 +1,18 @@
 pub(crate) mod attribution;
 mod config;
+pub(crate) mod credential_authority;
 pub mod credential_provider;
 pub(crate) mod device;
 pub mod device_code;
 pub mod error;
 mod flow;
+pub(crate) mod github_copilot;
 pub(crate) mod kimi_oauth;
 pub(crate) mod manager;
 mod model;
+pub(crate) mod oauth_device;
+pub(crate) mod oauth_pkce;
+pub(crate) mod oauth_registry;
 pub(crate) mod recovery;
 pub(crate) mod refresh;
 mod storage;
@@ -17,7 +22,8 @@ pub(crate) use flow::try_ensure_session_noninteractive;
 pub use flow::{
     AuthChannels, AuthUrlInfo, AuthUrlMode, LogoutResult, ensure_authenticated,
     ensure_authenticated_or_noninteractive, perform_logout, run_auth_flow,
-    run_auth_flow_with_stderr_bridge, run_cli_login, run_cli_logout, try_ensure_fresh_auth,
+    run_auth_flow_with_stderr_bridge, run_cli_login, run_cli_logout, run_oauth_provider_flow,
+    try_ensure_fresh_auth,
 };
 mod meta;
 pub use device::device_headers;
@@ -27,5 +33,6 @@ pub use meta::AuthMeta;
 pub use model::{AuthMode, KimiAuth, lookup_auth};
 pub(crate) use model::{TOKEN_TTL, is_expired, token_suffix};
 pub use storage::{
-    clear_api_key, read_api_key, read_auth_json, read_token_by_scope, store_api_key,
+    clear_api_key, read_api_key, read_auth_json, read_platform_api_key, read_token_by_scope,
+    store_api_key, store_platform_api_key,
 };

crates/codegen/kigi-shell/src/auth/oauth_device.rs

@@ -0,0 +1,423 @@
+//! Generic RFC-8628 device-code OAuth wire, driven by a registry
+//! [`kigi_models::OAuthConfig`] (xai-grok today; Copilot/Claude later).
+//!
+//! Three `application/x-www-form-urlencoded` POSTs against `{auth_host}`:
+//!
+//! - `POST {device_path}` — form `client_id` + `scope` + the optional
+//!   `extra_device_field` (e.g. `referrer=kigi`)
+//! - `POST {token_path}` (poll) — form `client_id` + `device_code` +
+//!   `grant_type=urn:ietf:params:oauth:grant-type:device_code`
+//! - `POST {token_path}` (refresh) — form `client_id` +
+//!   `grant_type=refresh_token` + `refresh_token`, with the same exponential
+//!   backoff / status handling as the Kimi wire.
+//!
+//! Unlike [`super::kimi_oauth`] this sends NO X-Msh device headers — just the
+//! shared kigi `User-Agent` and `Accept: application/json`. Access/refresh
+//! tokens are NEVER logged (only non-secret events: requested, poll succeeded,
+//! refreshed).
+
+use kigi_models::OAuthConfig;
+use serde::Deserialize;
+
+use super::kimi_oauth::{
+    DeviceAuthorization, DevicePollResult, RefreshError, TokenResponse, validate_verification_uri,
+};
+
+const DEVICE_GRANT_TYPE: &str = "urn:ietf:params:oauth:grant-type:device_code";
+const REFRESH_GRANT_TYPE: &str = "refresh_token";
+/// Refresh retry budget over the retryable statuses / network blips.
+const MAX_REFRESH_RETRIES: u32 = 3;
+/// HTTP statuses worth retrying a refresh for (kimi-cli parity).
+const RETRYABLE_REFRESH_STATUSES: [u16; 5] = [429, 500, 502, 503, 504];
+
+#[derive(Deserialize)]
+struct DeviceAuthorizationResponse {
+    user_code: String,
+    device_code: String,
+    #[serde(default)]
+    verification_uri: Option<String>,
+    /// Optional here (the Kimi wire requires it): Pi's xAI response may omit
+    /// `verification_uri_complete` and carry only `verification_uri`.
+    #[serde(default)]
+    verification_uri_complete: Option<String>,
+    #[serde(default)]
+    expires_in: Option<i64>,
+    #[serde(default)]
+    interval: Option<i64>,
+}
+
+#[derive(Deserialize, Default)]
+struct OAuthErrorBody {
+    #[serde(default)]
+    error: Option<String>,
+    #[serde(default)]
+    error_description: Option<String>,
+}
+
+fn oauth_url(host: &str, path: &str) -> String {
+    format!("{}{path}", host.trim_end_matches('/'))
+}
+
+/// The device-authorization form fields: `client_id`, `scope`, and the
+/// optional non-standard `extra_device_field`.
+fn device_form(cfg: &OAuthConfig) -> Vec<(&'static str, &'static str)> {
+    let mut form = vec![("client_id", cfg.client_id), ("scope", cfg.scope)];
+    if let Some((name, value)) = cfg.extra_device_field {
+        form.push((name, value));
+    }
+    form
+}
+
+/// `POST {auth_host}{device_path}` — start a device login.
+pub(crate) async fn request_device_authorization(
+    cfg: &OAuthConfig,
+) -> anyhow::Result<DeviceAuthorization> {
+    let url = oauth_url(cfg.auth_host, cfg.device_path);
+    tracing::info!(url = %url, "auth: requesting device authorization (generic oauth)");
+    let resp = crate::http::shared_client()
+        .post(&url)
+        .header("Accept", "application/json")
+        .form(&device_form(cfg))
+        .send()
+        .await?;
+
+    let status = resp.status();
+    if !status.is_success() {
+        let body = resp.text().await.unwrap_or_default();
+        tracing::warn!(%status, "auth: device authorization failed (generic oauth)");
+        anyhow::bail!("Device authorization failed (HTTP {status}): {body}");
+    }
+    let parsed: DeviceAuthorizationResponse = resp.json().await?;
+
+    if !parsed
+        .user_code
+        .chars()
+        .all(|c| c.is_ascii_alphanumeric() || c == '-')
+    {
+        anyhow::bail!("Server returned invalid user_code format (expected [A-Z0-9-])");
+    }
+    // Pi forces the displayed URI to https; we require a valid https (or
+    // localhost) verification target, preferring the pre-filled complete form.
+    let verification_uri_complete = parsed
+        .verification_uri_complete
+        .clone()
+        .or_else(|| parsed.verification_uri.clone())
+        .ok_or_else(|| anyhow::anyhow!("Server returned no verification URI"))?;
+    validate_verification_uri(&verification_uri_complete)?;
+    if let Some(ref uri) = parsed.verification_uri {
+        validate_verification_uri(uri)?;
+    }
+
+    tracing::info!(
+        user_code = %parsed.user_code,
+        interval = parsed.interval.unwrap_or(5),
+        expires_in = ?parsed.expires_in,
+        "auth: device authorization issued (generic oauth)"
+    );
+    Ok(DeviceAuthorization {
+        user_code: parsed.user_code,
+        device_code: parsed.device_code,
+        verification_uri: parsed.verification_uri.filter(|u| !u.is_empty()),
+        verification_uri_complete,
+        expires_in: parsed.expires_in.filter(|&e| e > 0),
+        interval: parsed.interval.unwrap_or(5),
+    })
+}
+
+/// One poll of `POST {auth_host}{token_path}` with the device grant.
+pub(crate) async fn poll_device_token(
+    cfg: &OAuthConfig,
+    device_code: &str,
+) -> anyhow::Result<DevicePollResult> {
+    let url = oauth_url(cfg.auth_host, cfg.token_path);
+    let resp = crate::http::shared_client()
+        .post(&url)
+        .header("Accept", "application/json")
+        .form(&[
+            ("client_id", cfg.client_id),
+            ("device_code", device_code),
+            ("grant_type", DEVICE_GRANT_TYPE),
+        ])
+        .send()
+        .await
+        .map_err(|e| anyhow::anyhow!("Token polling request failed: {e}"))?;
+
+    let status = resp.status();
+    if status.is_server_error() {
+        anyhow::bail!("Token polling server error: {status}");
+    }
+    let body = resp.bytes().await?;
+    if status.is_success() {
+        if let Ok(tokens) = serde_json::from_slice::<TokenResponse>(&body) {
+            tracing::info!("auth: device poll succeeded, access token issued (generic oauth)");
+            return Ok(DevicePollResult::Success(Box::new(tokens.into_auth())));
+        }
+        tracing::warn!(
+            "auth: device poll returned 200 without access_token; continuing (generic oauth)"
+        );
+        return Ok(DevicePollResult::Pending {
+            error: "missing_access_token".to_owned(),
+            description: None,
+        });
+    }
+    let err: OAuthErrorBody = serde_json::from_slice(&body).unwrap_or_default();
+    let error = err.error.unwrap_or_else(|| "unknown_error".to_owned());
+    if error == "expired_token" {
+        tracing::info!(
+            "auth: device code expired; restarting device authorization (generic oauth)"
+        );
+        return Ok(DevicePollResult::Expired);
+    }
+    tracing::debug!(error = %error, "auth: device poll pending (generic oauth)");
+    Ok(DevicePollResult::Pending {
+        error,
+        description: err.error_description,
+    })
+}
+
+/// `POST {auth_host}{token_path}` with `grant_type=refresh_token`. Retries the
+/// retryable statuses / network errors with exponential backoff; 401/403
+/// returns immediately as [`RefreshError::Unauthorized`].
+pub(crate) async fn refresh_token(
+    cfg: &OAuthConfig,
+    refresh_token: &str,
+) -> Result<super::model::KimiAuth, RefreshError> {
+    let url = oauth_url(cfg.auth_host, cfg.token_path);
+    let mut last_error = String::from("no attempt made");
+    for attempt in 0..MAX_REFRESH_RETRIES {
+        if attempt > 0 {
+            let backoff = std::time::Duration::from_secs(1 << (attempt - 1));
+            tracing::warn!(
+                attempt,
+                backoff_secs = backoff.as_secs(),
+                last_error = %last_error,
+                "auth: retrying token refresh (generic oauth)"
+            );
+            tokio::time::sleep(backoff).await;
+        }
+        tracing::info!(attempt, "auth: token refresh attempt (generic oauth)");
+        let send_result = crate::http::shared_client()
+            .post(&url)
+            .header("Accept", "application/json")
+            .form(&[
+                ("client_id", cfg.client_id),
+                ("grant_type", REFRESH_GRANT_TYPE),
+                ("refresh_token", refresh_token),
+            ])
+            .send()
+            .await;
+
+        let resp = match send_result {
+            Ok(resp) => resp,
+            Err(e) => {
+                last_error = format!("network error: {e}");
+                continue;
+            }
+        };
+        let status = resp.status().as_u16();
+        let body = resp.bytes().await.unwrap_or_default();
+        if status == 401 || status == 403 {
+            let err: OAuthErrorBody = serde_json::from_slice(&body).unwrap_or_default();
+            return Err(RefreshError::Unauthorized {
+                status,
+                description: err
+                    .error_description
+                    .unwrap_or_else(|| "Token refresh unauthorized.".to_owned()),
+            });
+        }
+        if status == 200 {
+            return match serde_json::from_slice::<TokenResponse>(&body) {
+                Ok(tokens) => Ok(tokens.into_auth()),
+                Err(e) => Err(RefreshError::Fatal {
+                    status,
+                    description: format!("malformed token payload: {e}"),
+                }),
+            };
+        }
+        let err: OAuthErrorBody = serde_json::from_slice(&body).unwrap_or_default();
+        let description = err
+            .error_description
+            .unwrap_or_else(|| format!("Token refresh failed (HTTP {status})."));
+        if RETRYABLE_REFRESH_STATUSES.contains(&status) {
+            last_error = description;
+            continue;
+        }
+        return Err(RefreshError::Fatal {
+            status,
+            description,
+        });
+    }
+    Err(RefreshError::Exhausted { last_error })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use kigi_models::XAI_OAUTH_CONFIG;
+    use wiremock::matchers::{body_string_contains, method, path};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    /// An OAuthConfig pointed at a mock server (copies XAI's client_id/scope/
+    /// paths but overrides the host).
+    fn mock_cfg(host: &'static str) -> OAuthConfig {
+        OAuthConfig {
+            auth_host: host,
+            ..XAI_OAUTH_CONFIG
+        }
+    }
+
+    fn token_json(access: &str, refresh: &str) -> serde_json::Value {
+        serde_json::json!({
+            "access_token": access,
+            "refresh_token": refresh,
+            "expires_in": 3600,
+            "scope": "grok-cli:access",
+            "token_type": "bearer",
+        })
+    }
+
+    #[tokio::test]
+    async fn device_authorization_sends_client_scope_and_referrer() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/device/code"))
+            .and(body_string_contains(
+                "client_id=b1a00492-073a-47ea-816f-4c329264a828",
+            ))
+            .and(body_string_contains("scope=openid"))
+            .and(body_string_contains("referrer=kigi"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "user_code": "GROK-1234",
+                "device_code": "dev-xai-1",
+                "verification_uri": "https://x.ai/device",
+                "verification_uri_complete": "https://x.ai/device?user_code=GROK-1234",
+                "expires_in": 900,
+                "interval": 5,
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let auth = request_device_authorization(&mock_cfg(host)).await.unwrap();
+        assert_eq!(auth.user_code, "GROK-1234");
+        assert_eq!(auth.device_code, "dev-xai-1");
+        assert_eq!(
+            auth.verification_uri_complete,
+            "https://x.ai/device?user_code=GROK-1234"
+        );
+        assert_eq!(auth.expires_in, Some(900));
+    }
+
+    /// A response with only `verification_uri` (no `_complete`) still yields a
+    /// valid display URI.
+    #[tokio::test]
+    async fn device_authorization_falls_back_to_verification_uri() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/device/code"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "user_code": "GROK-9",
+                "device_code": "d",
+                "verification_uri": "https://x.ai/device",
+            })))
+            .mount(&server)
+            .await;
+        let auth = request_device_authorization(&mock_cfg(host)).await.unwrap();
+        assert_eq!(auth.verification_uri_complete, "https://x.ai/device");
+        assert_eq!(auth.interval, 5, "default interval");
+    }
+
+    #[tokio::test]
+    async fn poll_success_builds_auth_with_expiry() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .and(body_string_contains("grant_type=urn"))
+            .and(body_string_contains("device_code=dev-xai-1"))
+            .respond_with(
+                ResponseTemplate::new(200).set_body_json(token_json("grok-at", "grok-rt")),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let result = poll_device_token(&mock_cfg(host), "dev-xai-1")
+            .await
+            .unwrap();
+        let DevicePollResult::Success(auth) = result else {
+            panic!("expected success, got {result:?}");
+        };
+        assert_eq!(auth.key, "grok-at");
+        assert_eq!(auth.refresh_token.as_deref(), Some("grok-rt"));
+        assert_eq!(auth.expires_in, Some(3600));
+    }
+
+    #[tokio::test]
+    async fn poll_maps_authorization_pending_to_pending() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .respond_with(
+                ResponseTemplate::new(400)
+                    .set_body_json(serde_json::json!({ "error": "authorization_pending" })),
+            )
+            .mount(&server)
+            .await;
+        let result = poll_device_token(&mock_cfg(host), "dev-xai-1")
+            .await
+            .unwrap();
+        match result {
+            DevicePollResult::Pending { error, .. } => assert_eq!(error, "authorization_pending"),
+            other => panic!("expected pending, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn refresh_success_round_trip() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .and(body_string_contains("grant_type=refresh_token"))
+            .and(body_string_contains("refresh_token=grok-rt-old"))
+            .respond_with(
+                ResponseTemplate::new(200).set_body_json(token_json("grok-at-new", "grok-rt-new")),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let auth = refresh_token(&mock_cfg(host), "grok-rt-old").await.unwrap();
+        assert_eq!(auth.key, "grok-at-new");
+        assert_eq!(auth.refresh_token.as_deref(), Some("grok-rt-new"));
+    }
+
+    #[tokio::test]
+    async fn refresh_401_maps_to_unauthorized() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .respond_with(
+                ResponseTemplate::new(401)
+                    .set_body_json(serde_json::json!({ "error_description": "refresh revoked" })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let err = refresh_token(&mock_cfg(host), "grok-rt-dead")
+            .await
+            .unwrap_err();
+        match err {
+            RefreshError::Unauthorized {
+                status,
+                description,
+            } => {
+                assert_eq!(status, 401);
+                assert_eq!(description, "refresh revoked");
+            }
+            other => panic!("expected Unauthorized, got {other:?}"),
+        }
+    }
+}

crates/codegen/kigi-shell/src/auth/oauth_pkce.rs

@@ -0,0 +1,904 @@
+//! Generic authorization-code + PKCE (S256) OAuth wire with a `127.0.0.1`
+//! loopback callback, driven by a registry [`kigi_models::OAuthConfig`] whose
+//! `flow` is [`OAuthFlow::PkceLocalhost`] (claude-pro-max JSON + openai-codex
+//! FORM).
+//!
+//! Shape (Pi `earendil-works/pi` `auth/oauth/{anthropic,openai-codex}.ts`):
+//! - `verifier = base64url(32 random bytes)`; `challenge = base64url(SHA-256(
+//!   verifier))`. `state` is `verifier` for claude ([`generate_pkce`]) or a
+//!   fresh-random value for codex ([`generate_pkce_random_state`]).
+//! - Browser opens `{auth_host}{device_path}?client_id&response_type=code&
+//!   scope&redirect_uri&state&code_challenge&code_challenge_method=S256` plus any
+//!   `authorize_extra` params (codex only).
+//! - The code returns to a loopback listener on `127.0.0.1:{redirect_port}`
+//!   answering ONLY `{redirect_path}` (claude `/callback`, codex
+//!   `/auth/callback`), with STRICT `state` validation (a mismatch is rejected —
+//!   CSRF guard). A manual paste (redirect URL / `code#state` / bare code) is
+//!   accepted as a headless fallback.
+//! - Code → token exchange POSTs `{token_host}{token_path}` as JSON
+//!   ([`exchange_code`], claude) or FORM ([`exchange_code_form`], codex; NO
+//!   `state` field). Refresh: claude JSON here ([`refresh_token`], rotating);
+//!   codex takes the generic device refresher's FORM path.
+//!
+//! SECURITY: the verifier, authorization code, access token, and refresh token
+//! are NEVER logged (only non-secret events: authorize URL requested, callback
+//! received, token issued, token refreshed).
+
+use anyhow::Context;
+use base64::Engine;
+use kigi_models::{OAuthConfig, OAuthTokenBody};
+use serde::Deserialize;
+use sha2::{Digest, Sha256};
+
+use super::kimi_oauth::{RefreshError, TokenResponse};
+use super::model::KimiAuth;
+
+const CODE_GRANT_TYPE: &str = "authorization_code";
+const REFRESH_GRANT_TYPE: &str = "refresh_token";
+/// Refresh retry budget over the retryable statuses / network blips.
+const MAX_REFRESH_RETRIES: u32 = 3;
+/// HTTP statuses worth retrying a refresh for (parity with the device wire).
+const RETRYABLE_REFRESH_STATUSES: [u16; 5] = [429, 500, 502, 503, 504];
+
+/// PKCE secrets for one login attempt. Two `state` conventions ship:
+/// [`generate_pkce`] sets `state == verifier` (Pi's/Claude's convention), while
+/// [`generate_pkce_random_state`] mints an INDEPENDENT random state (the OAuth
+/// standard, used by ChatGPT/Codex). Either way the state is validated on the
+/// callback as the CSRF guard.
+#[derive(Debug, Clone)]
+pub(crate) struct PkceCodes {
+    /// `code_verifier` — the 43-char base64url secret, sent at token exchange.
+    pub verifier: String,
+    /// `code_challenge = base64url(SHA-256(verifier))`, sent at authorize.
+    pub challenge: String,
+    /// `state` — the verifier itself, or an independent random value depending
+    /// on the provider's dialect; validated on the callback (CSRF guard).
+    pub state: String,
+}
+
+/// Generate PKCE S256 codes: `verifier = base64url(32 random bytes)`,
+/// `challenge = base64url(SHA-256(verifier))`, `state = verifier`.
+pub(crate) fn generate_pkce() -> PkceCodes {
+    use rand::RngCore;
+    let mut raw = [0u8; 32];
+    rand::rng().fill_bytes(&mut raw);
+    let verifier = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(raw);
+    let digest = Sha256::digest(verifier.as_bytes());
+    let challenge = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(digest);
+    PkceCodes {
+        state: verifier.clone(),
+        verifier,
+        challenge,
+    }
+}
+
+/// Like [`generate_pkce`] but with an INDEPENDENT fresh-random `state` (16
+/// random bytes) instead of `state == verifier`. The ChatGPT/Codex flow uses a
+/// distinct state (the verifier never doubles as the CSRF token there), so the
+/// verifier stays out of the state carried on the loopback callback.
+pub(crate) fn generate_pkce_random_state() -> PkceCodes {
+    use rand::RngCore;
+    let mut raw = [0u8; 16];
+    rand::rng().fill_bytes(&mut raw);
+    let state = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(raw);
+    PkceCodes {
+        state,
+        ..generate_pkce()
+    }
+}
+
+/// The loopback redirect URI for a PKCE-localhost provider (claude `/callback`,
+/// codex `/auth/callback`).
+pub(crate) fn redirect_uri(redirect_port: u16, redirect_path: &str) -> String {
+    format!("http://localhost:{redirect_port}{redirect_path}")
+}
+
+/// Build the browser authorize URL:
+/// `{auth_host}{device_path}?client_id&response_type=code&scope&redirect_uri&
+/// state&code_challenge&code_challenge_method=S256` plus any config
+/// `authorize_extra` params (empty for every config but codex, so their URLs
+/// stay byte-identical).
+pub(crate) fn build_authorize_url(
+    cfg: &OAuthConfig,
+    redirect_uri: &str,
+    pkce: &PkceCodes,
+) -> String {
+    let base = format!("{}{}", cfg.auth_host.trim_end_matches('/'), cfg.device_path);
+    let mut serializer = url::form_urlencoded::Serializer::new(String::new());
+    serializer
+        .append_pair("client_id", cfg.client_id)
+        .append_pair("response_type", "code")
+        .append_pair("scope", cfg.scope)
+        .append_pair("redirect_uri", redirect_uri)
+        .append_pair("state", &pkce.state)
+        .append_pair("code_challenge", &pkce.challenge)
+        .append_pair("code_challenge_method", "S256");
+    for (key, value) in cfg.authorize_extra {
+        serializer.append_pair(key, value);
+    }
+    format!("{base}?{}", serializer.finish())
+}
+
+/// `code` + `state` extracted from a callback (loopback query OR manual paste).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(crate) struct CallbackParams {
+    pub code: String,
+    pub state: Option<String>,
+}
+
+/// Parse `code`/`state` from the raw query string of a `/callback?…` request
+/// (e.g. `code=abc&state=xyz`). An `error=` param surfaces as an `Err`.
+pub(crate) fn parse_callback_query(query: &str) -> anyhow::Result<CallbackParams> {
+    let mut code = None;
+    let mut state = None;
+    let mut error = None;
+    for (k, v) in url::form_urlencoded::parse(query.as_bytes()) {
+        match k.as_ref() {
+            "code" => code = Some(v.into_owned()),
+            "state" => state = Some(v.into_owned()),
+            "error" => error = Some(v.into_owned()),
+            _ => {}
+        }
+    }
+    if let Some(error) = error {
+        anyhow::bail!("Authorization server returned an error: {error}");
+    }
+    let code = code.context("callback missing authorization code")?;
+    if code.is_empty() {
+        anyhow::bail!("callback authorization code was empty");
+    }
+    Ok(CallbackParams { code, state })
+}
+
+/// Parse a MANUAL paste (headless fallback). Accepts, in order:
+/// - a full redirect URL (`http://localhost:…/callback?code=…&state=…`),
+/// - a `code#state` pair (Anthropic's console shows this form),
+/// - a bare `code` (state then unknown → `None`, caller validation applies).
+pub(crate) fn parse_manual_paste(input: &str) -> anyhow::Result<CallbackParams> {
+    let trimmed = input.trim();
+    if trimmed.is_empty() {
+        anyhow::bail!("empty paste");
+    }
+    // Full redirect URL.
+    if trimmed.starts_with("http://") || trimmed.starts_with("https://") {
+        let url = url::Url::parse(trimmed).context("pasted value is not a valid URL")?;
+        return parse_callback_query(url.query().unwrap_or_default());
+    }
+    // `code#state`.
+    if let Some((code, state)) = trimmed.split_once('#') {
+        if code.is_empty() {
+            anyhow::bail!("pasted code was empty");
+        }
+        return Ok(CallbackParams {
+            code: code.to_owned(),
+            state: (!state.is_empty()).then(|| state.to_owned()),
+        });
+    }
+    // Bare code.
+    Ok(CallbackParams {
+        code: trimmed.to_owned(),
+        state: None,
+    })
+}
+
+/// STRICT state validation (CSRF guard): the callback `state` MUST be present
+/// AND equal to the expected value. A mismatch (or absence, when a paste has no
+/// state) is rejected — the flow NEVER proceeds on an unverified callback.
+pub(crate) fn validate_state(params: &CallbackParams, expected_state: &str) -> anyhow::Result<()> {
+    match params.state.as_deref() {
+        Some(state) if state == expected_state => Ok(()),
+        Some(_) => anyhow::bail!("OAuth state mismatch — rejecting callback (CSRF guard)"),
+        None => anyhow::bail!("OAuth callback carried no state — rejecting (CSRF guard)"),
+    }
+}
+
+/// State validation for a MANUAL paste (headless fallback): a present state
+/// MUST match (mismatch rejected — CSRF guard), but an ABSENT state is allowed
+/// — a bare-code paste is user-initiated (not a network-reachable callback), so
+/// there is no state to check. The loopback path uses the stricter
+/// [`validate_state`] (an absent state there IS rejected).
+pub(crate) fn validate_pasted_state(
+    params: &CallbackParams,
+    expected_state: &str,
+) -> anyhow::Result<()> {
+    match params.state.as_deref() {
+        Some(state) if state == expected_state => Ok(()),
+        Some(_) => anyhow::bail!("OAuth state mismatch — rejecting pasted code (CSRF guard)"),
+        None => Ok(()),
+    }
+}
+
+/// The token-endpoint URL (`{token_host}{token_path}`).
+fn token_url(cfg: &OAuthConfig) -> String {
+    format!("{}{}", cfg.token_host.trim_end_matches('/'), cfg.token_path)
+}
+
+/// POST the token endpoint with a JSON body, honoring `cfg.token_body`. Claude
+/// is JSON; a `Form`-bodied config would be handled by the device wire, so the
+/// PKCE path asserts JSON (never silently mis-encodes).
+async fn post_token_json(
+    cfg: &OAuthConfig,
+    body: serde_json::Value,
+) -> reqwest::Result<reqwest::Response> {
+    debug_assert!(
+        matches!(cfg.token_body, OAuthTokenBody::Json),
+        "PKCE token exchange expects a JSON token body"
+    );
+    crate::http::shared_client()
+        .post(token_url(cfg))
+        .header("Accept", "application/json")
+        .json(&body)
+        .send()
+        .await
+}
+
+/// Exchange an authorization `code` for a token set (JSON body):
+/// `{grant_type:"authorization_code", code, state, client_id, redirect_uri,
+/// code_verifier}`. Returns the materialized [`KimiAuth`].
+pub(crate) async fn exchange_code(
+    cfg: &OAuthConfig,
+    code: &str,
+    pkce: &PkceCodes,
+    redirect_uri: &str,
+) -> anyhow::Result<KimiAuth> {
+    let body = serde_json::json!({
+        "grant_type": CODE_GRANT_TYPE,
+        "code": code,
+        "state": pkce.state,
+        "client_id": cfg.client_id,
+        "redirect_uri": redirect_uri,
+        "code_verifier": pkce.verifier,
+    });
+    tracing::info!(
+        scope_key = cfg.scope_key,
+        "auth: exchanging code for token (pkce)"
+    );
+    let resp = post_token_json(cfg, body)
+        .await
+        .context("token exchange request failed")?;
+    let status = resp.status();
+    if !status.is_success() {
+        let body = resp.text().await.unwrap_or_default();
+        tracing::warn!(%status, scope_key = cfg.scope_key, "auth: code exchange failed (pkce)");
+        anyhow::bail!("Token exchange failed (HTTP {status}): {body}");
+    }
+    let tokens: TokenResponse = resp.json().await.context("malformed token payload")?;
+    tracing::info!(
+        scope_key = cfg.scope_key,
+        "auth: pkce code exchange succeeded"
+    );
+    Ok(tokens.into_auth())
+}
+
+/// Exchange an authorization `code` for a token set with a FORM body (codex):
+/// `{grant_type=authorization_code, client_id, code, code_verifier,
+/// redirect_uri}`. Unlike [`exchange_code`], the `state` is NOT sent in the
+/// token body (the ChatGPT/Codex token endpoint does not expect it). Asserts a
+/// `Form` config so a JSON provider can never silently mis-encode.
+pub(crate) async fn exchange_code_form(
+    cfg: &OAuthConfig,
+    code: &str,
+    pkce: &PkceCodes,
+    redirect_uri: &str,
+) -> anyhow::Result<KimiAuth> {
+    debug_assert!(
+        matches!(cfg.token_body, OAuthTokenBody::Form),
+        "PKCE form exchange expects a Form token body"
+    );
+    tracing::info!(
+        scope_key = cfg.scope_key,
+        "auth: exchanging code for token (pkce form)"
+    );
+    let resp = crate::http::shared_client()
+        .post(token_url(cfg))
+        .header("Accept", "application/json")
+        .form(&[
+            ("grant_type", CODE_GRANT_TYPE),
+            ("client_id", cfg.client_id),
+            ("code", code),
+            ("code_verifier", pkce.verifier.as_str()),
+            ("redirect_uri", redirect_uri),
+        ])
+        .send()
+        .await
+        .context("token exchange request failed")?;
+    let status = resp.status();
+    if !status.is_success() {
+        let body = resp.text().await.unwrap_or_default();
+        tracing::warn!(%status, scope_key = cfg.scope_key, "auth: code exchange failed (pkce form)");
+        anyhow::bail!("Token exchange failed (HTTP {status}): {body}");
+    }
+    let tokens: TokenResponse = resp.json().await.context("malformed token payload")?;
+    tracing::info!(
+        scope_key = cfg.scope_key,
+        "auth: pkce form code exchange succeeded"
+    );
+    Ok(tokens.into_auth())
+}
+
+/// `POST {token_host}{token_path}` with `grant_type=refresh_token` (JSON body).
+/// Claude ROTATES the refresh token, so the caller MUST persist the returned
+/// one. Retries the retryable statuses / network errors with exponential
+/// backoff; 401/403 returns immediately as [`RefreshError::Unauthorized`].
+pub(crate) async fn refresh_token(
+    cfg: &OAuthConfig,
+    refresh_token: &str,
+) -> Result<KimiAuth, RefreshError> {
+    let mut last_error = String::from("no attempt made");
+    for attempt in 0..MAX_REFRESH_RETRIES {
+        if attempt > 0 {
+            let backoff = std::time::Duration::from_secs(1 << (attempt - 1));
+            tracing::warn!(
+                attempt,
+                backoff_secs = backoff.as_secs(),
+                last_error = %last_error,
+                "auth: retrying token refresh (pkce)"
+            );
+            tokio::time::sleep(backoff).await;
+        }
+        tracing::info!(
+            attempt,
+            scope_key = cfg.scope_key,
+            "auth: token refresh attempt (pkce)"
+        );
+        let body = serde_json::json!({
+            "grant_type": REFRESH_GRANT_TYPE,
+            "client_id": cfg.client_id,
+            "refresh_token": refresh_token,
+        });
+        let resp = match post_token_json(cfg, body).await {
+            Ok(resp) => resp,
+            Err(e) => {
+                last_error = format!("network error: {e}");
+                continue;
+            }
+        };
+        let status = resp.status().as_u16();
+        let bytes = resp.bytes().await.unwrap_or_default();
+        if status == 401 || status == 403 {
+            let err: OAuthErrorBody = serde_json::from_slice(&bytes).unwrap_or_default();
+            return Err(RefreshError::Unauthorized {
+                status,
+                description: err
+                    .error_description
+                    .unwrap_or_else(|| "Token refresh unauthorized.".to_owned()),
+            });
+        }
+        if status == 200 {
+            return match serde_json::from_slice::<TokenResponse>(&bytes) {
+                Ok(tokens) => Ok(tokens.into_auth()),
+                Err(e) => Err(RefreshError::Fatal {
+                    status,
+                    description: format!("malformed token payload: {e}"),
+                }),
+            };
+        }
+        let err: OAuthErrorBody = serde_json::from_slice(&bytes).unwrap_or_default();
+        let description = err
+            .error_description
+            .unwrap_or_else(|| format!("Token refresh failed (HTTP {status})."));
+        if RETRYABLE_REFRESH_STATUSES.contains(&status) {
+            last_error = description;
+            continue;
+        }
+        return Err(RefreshError::Fatal {
+            status,
+            description,
+        });
+    }
+    Err(RefreshError::Exhausted { last_error })
+}
+
+#[derive(Deserialize, Default)]
+struct OAuthErrorBody {
+    #[serde(default)]
+    error_description: Option<String>,
+}
+
+/// Bind a loopback HTTP listener on `127.0.0.1:{redirect_port}` and wait for a
+/// single `GET {redirect_path}?code=…&state=…`, validating `state` STRICTLY
+/// against `expected_state` (mismatch → rejected). Returns the authorization
+/// code.
+///
+/// The listener answers ONLY `redirect_path` (claude `/callback`, codex
+/// `/auth/callback`); any other path gets 404. It binds `127.0.0.1` (never
+/// `0.0.0.0`), so no non-loopback host can reach it.
+pub(crate) async fn await_loopback_code(
+    redirect_port: u16,
+    redirect_path: &str,
+    expected_state: &str,
+) -> anyhow::Result<String> {
+    let listener = tokio::net::TcpListener::bind(("127.0.0.1", redirect_port))
+        .await
+        .with_context(|| format!("could not bind loopback 127.0.0.1:{redirect_port}"))?;
+    tracing::info!(port = redirect_port, "auth: pkce loopback listener bound");
+    loop {
+        let (stream, _peer) = listener.accept().await.context("loopback accept failed")?;
+        match handle_loopback_conn(stream, redirect_path, expected_state).await {
+            LoopbackOutcome::Code(code) => return Ok(code),
+            LoopbackOutcome::Rejected(err) => return Err(err),
+            // Not the callback GET (favicon, health probe): keep listening.
+            LoopbackOutcome::Ignore => continue,
+        }
+    }
+}
+
+enum LoopbackOutcome {
+    Code(String),
+    Rejected(anyhow::Error),
+    Ignore,
+}
+
+/// Read the request line of one loopback connection, answer with a small HTML
+/// page, and classify the outcome. STRICT: a `/callback` with a bad/missing
+/// state is [`LoopbackOutcome::Rejected`] (the browser sees an error page).
+async fn handle_loopback_conn(
+    mut stream: tokio::net::TcpStream,
+    redirect_path: &str,
+    expected_state: &str,
+) -> LoopbackOutcome {
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+    // Read only enough for the request line — a GET has no body.
+    let mut buf = [0u8; 8192];
+    let n = match stream.read(&mut buf).await {
+        Ok(0) => return LoopbackOutcome::Ignore,
+        Ok(n) => n,
+        Err(_) => return LoopbackOutcome::Ignore,
+    };
+    let head = String::from_utf8_lossy(&buf[..n]);
+    let Some(request_line) = head.lines().next() else {
+        return LoopbackOutcome::Ignore;
+    };
+    // `GET {redirect_path}?code=…&state=… HTTP/1.1`
+    let mut parts = request_line.split_whitespace();
+    let (Some(method), Some(target)) = (parts.next(), parts.next()) else {
+        return LoopbackOutcome::Ignore;
+    };
+    if method != "GET" {
+        let _ = write_http(&mut stream, 405, "Method Not Allowed").await;
+        return LoopbackOutcome::Ignore;
+    }
+    let (path, query) = target.split_once('?').unwrap_or((target, ""));
+    if path != redirect_path {
+        let _ = write_http(&mut stream, 404, "Not Found").await;
+        return LoopbackOutcome::Ignore;
+    }
+
+    let result = parse_callback_query(query)
+        .and_then(|params| validate_state(&params, expected_state).map(|()| params.code));
+    match result {
+        Ok(code) => {
+            let _ = write_http(
+                &mut stream,
+                200,
+                "Signed in. You can close this window and return to kigi.",
+            )
+            .await;
+            let _ = stream.flush().await;
+            LoopbackOutcome::Code(code)
+        }
+        Err(e) => {
+            let _ = write_http(&mut stream, 400, "Login failed — return to kigi and retry.").await;
+            let _ = stream.flush().await;
+            LoopbackOutcome::Rejected(e)
+        }
+    }
+}
+
+/// Write a minimal HTTP/1.1 response with an HTML body.
+async fn write_http(
+    stream: &mut tokio::net::TcpStream,
+    status: u16,
+    message: &str,
+) -> std::io::Result<()> {
+    use tokio::io::AsyncWriteExt;
+    let reason = match status {
+        200 => "OK",
+        400 => "Bad Request",
+        404 => "Not Found",
+        405 => "Method Not Allowed",
+        _ => "Error",
+    };
+    let body = format!("<!doctype html><meta charset=utf-8><p>{message}</p>");
+    let response = format!(
+        "HTTP/1.1 {status} {reason}\r\nContent-Type: text/html; charset=utf-8\r\n\
+         Content-Length: {}\r\nConnection: close\r\n\r\n{body}",
+        body.len()
+    );
+    stream.write_all(response.as_bytes()).await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use kigi_models::CLAUDE_OAUTH_CONFIG;
+    use wiremock::matchers::{body_string_contains, method, path};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    /// A config pointed at a mock token host (copies Claude's client_id/scope/
+    /// paths but overrides the token host).
+    fn mock_cfg(token_host: &'static str) -> OAuthConfig {
+        OAuthConfig {
+            token_host,
+            ..CLAUDE_OAUTH_CONFIG
+        }
+    }
+
+    /// PKCE codes: verifier/challenge are non-empty base64url (no padding), the
+    /// challenge is the base64url SHA-256 of the verifier, and state == verifier.
+    #[test]
+    fn generate_pkce_produces_valid_s256_codes() {
+        let pkce = generate_pkce();
+        assert_eq!(pkce.state, pkce.verifier, "state must equal the verifier");
+        assert!(!pkce.verifier.is_empty() && !pkce.challenge.is_empty());
+        for s in [&pkce.verifier, &pkce.challenge] {
+            assert!(
+                s.chars()
+                    .all(|c| c.is_ascii_alphanumeric() || c == '-' || c == '_'),
+                "base64url (no pad) only: {s}"
+            );
+            assert!(!s.contains('='), "no padding: {s}");
+        }
+        // challenge == base64url(SHA-256(verifier)).
+        let expect = base64::engine::general_purpose::URL_SAFE_NO_PAD
+            .encode(Sha256::digest(pkce.verifier.as_bytes()));
+        assert_eq!(pkce.challenge, expect);
+        // Fresh entropy each call.
+        assert_ne!(pkce.verifier, generate_pkce().verifier);
+    }
+
+    /// The authorize URL carries the fixed params + the PKCE state and S256
+    /// challenge, and targets `claude.ai/oauth/authorize`.
+    #[test]
+    fn authorize_url_has_state_and_s256_challenge() {
+        let pkce = generate_pkce();
+        let redirect = redirect_uri(53692, "/callback");
+        let url = build_authorize_url(&CLAUDE_OAUTH_CONFIG, &redirect, &pkce);
+        let parsed = url::Url::parse(&url).expect("valid URL");
+        assert_eq!(parsed.host_str(), Some("claude.ai"));
+        assert_eq!(parsed.path(), "/oauth/authorize");
+        let q: std::collections::HashMap<_, _> = parsed.query_pairs().into_owned().collect();
+        assert_eq!(q.get("response_type").map(String::as_str), Some("code"));
+        assert_eq!(
+            q.get("code_challenge_method").map(String::as_str),
+            Some("S256")
+        );
+        assert_eq!(
+            q.get("state").map(String::as_str),
+            Some(pkce.state.as_str())
+        );
+        assert_eq!(
+            q.get("code_challenge").map(String::as_str),
+            Some(pkce.challenge.as_str())
+        );
+        assert_eq!(
+            q.get("client_id").map(String::as_str),
+            Some(CLAUDE_OAUTH_CONFIG.client_id)
+        );
+        assert_eq!(
+            q.get("redirect_uri").map(String::as_str),
+            Some(redirect.as_str())
+        );
+        // The verifier itself must NEVER appear in the browser URL.
+        assert!(
+            !url.contains("code_verifier"),
+            "the verifier must not ride the authorize URL"
+        );
+    }
+
+    /// STRICT state validation: an exact match passes; a mismatch or an absent
+    /// state is REJECTED (CSRF guard — the flow must never proceed).
+    #[test]
+    fn state_validation_is_strict() {
+        let ok = CallbackParams {
+            code: "c".into(),
+            state: Some("expected".into()),
+        };
+        assert!(validate_state(&ok, "expected").is_ok());
+        let mismatch = CallbackParams {
+            code: "c".into(),
+            state: Some("attacker".into()),
+        };
+        assert!(
+            validate_state(&mismatch, "expected").is_err(),
+            "a state mismatch MUST be rejected"
+        );
+        let missing = CallbackParams {
+            code: "c".into(),
+            state: None,
+        };
+        assert!(
+            validate_state(&missing, "expected").is_err(),
+            "an absent state MUST be rejected"
+        );
+    }
+
+    /// A loopback `/callback` with the WRONG state is rejected end-to-end (the
+    /// listener returns an error, never a code) — the CSRF guard on the wire.
+    #[tokio::test]
+    async fn loopback_rejects_state_mismatch() {
+        // Ephemeral port: bind, learn the port, then drive a client at it.
+        let probe = tokio::net::TcpListener::bind(("127.0.0.1", 0))
+            .await
+            .unwrap();
+        let port = probe.local_addr().unwrap().port();
+        drop(probe);
+
+        let server =
+            tokio::spawn(
+                async move { await_loopback_code(port, "/callback", "the-real-state").await },
+            );
+        // Give the listener a moment to bind.
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        // Attacker callback: valid code, WRONG state.
+        let _ = reqwest::get(format!(
+            "http://127.0.0.1:{port}/callback?code=stolen&state=wrong-state"
+        ))
+        .await;
+        let outcome = server.await.unwrap();
+        let err = outcome.expect_err("a state mismatch must be rejected, never yield a code");
+        assert!(err.to_string().contains("state mismatch"), "{err}");
+    }
+
+    /// A loopback `/callback` with the MATCHING state yields the code.
+    #[tokio::test]
+    async fn loopback_returns_code_on_valid_state() {
+        let probe = tokio::net::TcpListener::bind(("127.0.0.1", 0))
+            .await
+            .unwrap();
+        let port = probe.local_addr().unwrap().port();
+        drop(probe);
+
+        let server =
+            tokio::spawn(async move { await_loopback_code(port, "/callback", "good-state").await });
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        let _ = reqwest::get(format!(
+            "http://127.0.0.1:{port}/callback?code=auth-code-123&state=good-state"
+        ))
+        .await;
+        let code = server.await.unwrap().expect("valid state yields the code");
+        assert_eq!(code, "auth-code-123");
+    }
+
+    /// Manual-paste parsing: full redirect URL, `code#state`, and bare code.
+    #[test]
+    fn manual_paste_parses_all_three_forms() {
+        let from_url =
+            parse_manual_paste("http://localhost:53692/callback?code=abc123&state=st-9").unwrap();
+        assert_eq!(from_url.code, "abc123");
+        assert_eq!(from_url.state.as_deref(), Some("st-9"));
+
+        let from_hash = parse_manual_paste("abc123#st-9").unwrap();
+        assert_eq!(from_hash.code, "abc123");
+        assert_eq!(from_hash.state.as_deref(), Some("st-9"));
+
+        let bare = parse_manual_paste("  abc123  ").unwrap();
+        assert_eq!(bare.code, "abc123");
+        assert_eq!(bare.state, None);
+
+        assert!(parse_manual_paste("").is_err());
+        // A pasted redirect that carries an error param surfaces the error.
+        assert!(parse_manual_paste("http://localhost/callback?error=access_denied").is_err());
+    }
+
+    /// Code → token exchange: JSON body carries the grant + verifier, response
+    /// materializes a `KimiAuth` with the rotating refresh token.
+    #[tokio::test]
+    async fn exchange_code_posts_json_and_returns_auth() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/v1/oauth/token"))
+            .and(body_string_contains(
+                "\"grant_type\":\"authorization_code\"",
+            ))
+            .and(body_string_contains("\"code\":\"auth-code-xyz\""))
+            .and(body_string_contains("\"code_verifier\""))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "access_token": "sk-ant-oat-new",
+                "refresh_token": "sk-ant-ort-new",
+                "expires_in": 3600,
+                "token_type": "bearer",
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let cfg = mock_cfg(host);
+        let pkce = generate_pkce();
+        let auth = exchange_code(
+            &cfg,
+            "auth-code-xyz",
+            &pkce,
+            &redirect_uri(53692, "/callback"),
+        )
+        .await
+        .unwrap();
+        assert_eq!(auth.key, "sk-ant-oat-new");
+        assert_eq!(auth.refresh_token.as_deref(), Some("sk-ant-ort-new"));
+        assert_eq!(auth.expires_in, Some(3600));
+    }
+
+    /// Refresh rotates the refresh token (JSON body, refresh grant).
+    #[tokio::test]
+    async fn refresh_rotates_refresh_token() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/v1/oauth/token"))
+            .and(body_string_contains("\"grant_type\":\"refresh_token\""))
+            .and(body_string_contains("\"refresh_token\":\"ort-old\""))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "access_token": "oat-fresh",
+                "refresh_token": "ort-rotated",
+                "expires_in": 3600,
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let auth = refresh_token(&mock_cfg(host), "ort-old").await.unwrap();
+        assert_eq!(auth.key, "oat-fresh");
+        assert_eq!(
+            auth.refresh_token.as_deref(),
+            Some("ort-rotated"),
+            "the rotated refresh token must be adopted"
+        );
+    }
+
+    /// A 401 on refresh maps to Unauthorized (drives the permanent-failure path).
+    #[tokio::test]
+    async fn refresh_401_maps_to_unauthorized() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/v1/oauth/token"))
+            .respond_with(
+                ResponseTemplate::new(401)
+                    .set_body_json(serde_json::json!({ "error_description": "refresh revoked" })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let err = refresh_token(&mock_cfg(host), "ort-dead")
+            .await
+            .unwrap_err();
+        match err {
+            RefreshError::Unauthorized {
+                status,
+                description,
+            } => {
+                assert_eq!(status, 401);
+                assert_eq!(description, "refresh revoked");
+            }
+            other => panic!("expected Unauthorized, got {other:?}"),
+        }
+    }
+
+    // ── ChatGPT/Codex PKCE (openai-codex) ────────────────────────────────────
+
+    /// Codex PKCE uses an INDEPENDENT fresh-random state (NOT `state ==
+    /// verifier`) so the verifier never rides the callback.
+    #[test]
+    fn codex_pkce_state_is_independent_of_the_verifier() {
+        let pkce = generate_pkce_random_state();
+        assert_ne!(
+            pkce.state, pkce.verifier,
+            "codex state must be fresh-random, not the verifier"
+        );
+        assert!(!pkce.state.is_empty() && !pkce.verifier.is_empty());
+        // Challenge is still the S256 of the verifier.
+        let expect = base64::engine::general_purpose::URL_SAFE_NO_PAD
+            .encode(Sha256::digest(pkce.verifier.as_bytes()));
+        assert_eq!(pkce.challenge, expect);
+        assert_ne!(
+            generate_pkce_random_state().state,
+            pkce.state,
+            "fresh state"
+        );
+    }
+
+    /// The codex authorize URL carries the PKCE state + S256 challenge AND the
+    /// three codex-only extra params, and targets `auth.openai.com/oauth/
+    /// authorize` with the `/auth/callback` redirect. The verifier never rides it.
+    #[test]
+    fn codex_authorize_url_has_state_challenge_and_three_extra_params() {
+        use kigi_models::CODEX_OAUTH_CONFIG;
+        let pkce = generate_pkce_random_state();
+        let redirect = redirect_uri(1455, "/auth/callback");
+        assert_eq!(redirect, "http://localhost:1455/auth/callback");
+        let url = build_authorize_url(&CODEX_OAUTH_CONFIG, &redirect, &pkce);
+        let parsed = url::Url::parse(&url).expect("valid URL");
+        assert_eq!(parsed.host_str(), Some("auth.openai.com"));
+        assert_eq!(parsed.path(), "/oauth/authorize");
+        let q: std::collections::HashMap<_, _> = parsed.query_pairs().into_owned().collect();
+        assert_eq!(
+            q.get("state").map(String::as_str),
+            Some(pkce.state.as_str())
+        );
+        assert_eq!(
+            q.get("code_challenge").map(String::as_str),
+            Some(pkce.challenge.as_str())
+        );
+        assert_eq!(
+            q.get("code_challenge_method").map(String::as_str),
+            Some("S256")
+        );
+        // The three codex-only extra params.
+        assert_eq!(
+            q.get("id_token_add_organizations").map(String::as_str),
+            Some("true")
+        );
+        assert_eq!(
+            q.get("codex_cli_simplified_flow").map(String::as_str),
+            Some("true")
+        );
+        assert_eq!(
+            q.get("originator").map(String::as_str),
+            Some("codex_cli_rs")
+        );
+        assert!(
+            !url.contains("code_verifier"),
+            "the verifier must not ride the authorize URL"
+        );
+    }
+
+    /// The claude authorize URL is UNCHANGED (no extra params) — its empty
+    /// `authorize_extra` keeps it byte-identical.
+    #[test]
+    fn claude_authorize_url_carries_no_extra_params() {
+        let pkce = generate_pkce();
+        let url = build_authorize_url(
+            &CLAUDE_OAUTH_CONFIG,
+            &redirect_uri(53692, "/callback"),
+            &pkce,
+        );
+        assert!(!url.contains("id_token_add_organizations"));
+        assert!(!url.contains("codex_cli_simplified_flow"));
+        assert!(!url.contains("originator"));
+    }
+
+    fn codex_mock_cfg(token_host: &'static str) -> OAuthConfig {
+        OAuthConfig {
+            token_host,
+            ..kigi_models::CODEX_OAUTH_CONFIG
+        }
+    }
+
+    /// Codex code→token exchange posts a FORM body carrying the grant + code +
+    /// verifier + redirect_uri, and NOTABLY NO `state` field (the codex token
+    /// endpoint does not expect it). Response materializes a `KimiAuth`.
+    #[tokio::test]
+    async fn codex_exchange_code_posts_form_without_state() {
+        use wiremock::matchers::{body_string_contains, header, method, path};
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth/token"))
+            .and(header("content-type", "application/x-www-form-urlencoded"))
+            .and(body_string_contains("grant_type=authorization_code"))
+            .and(body_string_contains("code=codex-auth-code"))
+            .and(body_string_contains("code_verifier="))
+            .and(body_string_contains("redirect_uri="))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "access_token": "codex-access-jwt",
+                "refresh_token": "codex-refresh",
+                "expires_in": 3600,
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let cfg = codex_mock_cfg(host);
+        let pkce = generate_pkce_random_state();
+        let auth = exchange_code_form(
+            &cfg,
+            "codex-auth-code",
+            &pkce,
+            "http://localhost:1455/auth/callback",
+        )
+        .await
+        .unwrap();
+        assert_eq!(auth.key, "codex-access-jwt");
+        assert_eq!(auth.refresh_token.as_deref(), Some("codex-refresh"));
+        assert_eq!(auth.expires_in, Some(3600));
+    }
+}

crates/codegen/kigi-shell/src/auth/oauth_registry.rs

@@ -0,0 +1,193 @@
+//! Process-global per-provider OAuth `AuthManager` pool for INFERENCE-time auth.
+//!
+//! A session binds its primary (Kimi / first-party) [`AuthManager`] for the
+//! subscription path, but a `uses_oauth` platform that carries an
+//! [`kigi_models::OAuthConfig`] (xai-grok today) needs its OWN scope-keyed
+//! manager for every per-turn decision — bearer resolution, proactive /
+//! on-expiry refresh, and 401 recovery. Reusing the Kimi manager for a grok
+//! turn would transmit the Kimi subscription bearer to `api.x.ai` (a
+//! cross-provider leak, guaranteed 401) and, without proactive refresh, would
+//! 401 every turn once the ~1h grok token expired until a process restart.
+//!
+//! The pool is the SINGLE SOURCE OF TRUTH: one long-lived `AuthManager` per
+//! generic-oauth scope, each wired with the SAME lifecycle as the primary Kimi
+//! manager (`configure_refresher()` + `start_proactive_refresh()`) so the
+//! on-disk token stays fresh and a 401 recovers via the provider's own manager.
+//! Managers are built ON DEMAND from the on-disk token ([`global_manager_for`]),
+//! so a login landing AFTER a session spawned self-heals — no frozen per-session
+//! snapshot.
+//!
+//! ROUTING LIVES ELSEWHERE. This module is only the pool; the decision of which
+//! credential governs a request belongs to the single chokepoint,
+//! [`crate::auth::credential_authority::CredentialAuthority`]. Keeping the two
+//! apart is deliberate: three rounds of leaks came from routing rules being
+//! re-derived per call site.
+//!
+//! SECURITY: access/refresh tokens and resolved bearers are NEVER logged here.
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::{Arc, OnceLock};
+
+use parking_lot::Mutex;
+
+use crate::auth::AuthManager;
+
+/// Process-wide pool of live per-scope OAuth managers.
+///
+/// Auth is process-global (one user), so a single manager per scope is correct
+/// and lets the proactive-refresh task start exactly once per scope no matter
+/// how many sessions spawn. Keyed by the OAuth `scope_key` (`oauth/xai`, …).
+fn oauth_manager_pool() -> &'static Mutex<HashMap<&'static str, Arc<AuthManager>>> {
+    static POOL: OnceLock<Mutex<HashMap<&'static str, Arc<AuthManager>>>> = OnceLock::new();
+    POOL.get_or_init(|| Mutex::new(HashMap::new()))
+}
+
+/// The kigi home EVERY OAuth-provider construction in this crate resolves from
+/// — the pool, the catalog fetch's per-platform token resolution, the
+/// aux/summary token routing and the session's inference manager — so they can
+/// never read different homes.
+///
+/// Production: [`crate::util::kigi_home::kigi_home`]. LIB TESTS: a
+/// per-process path under the system temp dir that is deliberately **never
+/// created**. The pool is process-global and every manager it builds starts a
+/// never-cancelled proactive-refresh loop, so a unit test resolving the real
+/// `~/.kigi` would read the developer's stored OAuth tokens and, 60 s later,
+/// fire REAL refresh requests against them. Deliberately not a per-test opt-in
+/// that can be forgotten: `kigi_home()` is itself a `OnceLock` an earlier test
+/// has usually already resolved to the real home, so setting `KIGI_SHARE_DIR`
+/// in a test cannot pin it after the fact.
+///
+/// M4 — THE LIMIT, STATED: `cfg(test)` is set only for THIS crate's `--lib`
+/// tests. `crates/codegen/kigi-shell/tests/*.rs` link the library built WITHOUT
+/// it, so for an integration test this resolves the real home unless that test
+/// binary itself isolates one, which it must do through the two overrides the
+/// auth stack already honours and BEFORE anything resolves `kigi_home()`:
+/// `KIGI_SHARE_DIR` (read by `kigi_home()`, a `OnceLock`) or `KIGI_AUTH_PATH`
+/// (read by [`AuthManager::new_oauth_provider`], which pins the token file
+/// outright and so overrides this home entirely). 12 of the 28 integration
+/// binaries under `crates/codegen/kigi-shell/tests/` set `KIGI_SHARE_DIR`; the
+/// other 16 never reach an OAuth-platform inference path today, which is a
+/// property of those tests, not a guarantee of this function. No
+/// production-readable env override is added here on purpose: a knob that
+/// redirects where OAuth tokens are read from is not worth a test convenience.
+///
+/// M8: this used to be a `static OnceLock<TempDir>`. Statics are never dropped,
+/// so that leaked one temp directory per test binary — against the project's
+/// "tests are TempDir self-cleaning" discipline. Nothing is created here
+/// instead, and nothing in the lib-test suite creates it: a manager reads a
+/// missing `auth.json` as "no session", and the only two paths that WRITE one
+/// are a successful token refresh (which needs a stored refresh token that by
+/// construction does not exist here) and a completed device login
+/// ([`crate::agent::mvp_agent::MvpAgent::authenticate_oauth_platform`], which
+/// M4 repointed at this same home). Both require the network, so no lib test
+/// performs either. That is an observation about the suite, not an invariant of
+/// this function — [`tests::test_pool_home_is_disposable_and_never_the_real_home`]
+/// asserts the directory does not exist and is the tripwire if one ever does
+/// (the path is per-PROCESS under the system temp dir, so the blast radius of a
+/// future login-driving test is one disposable directory, never `~/.kigi`).
+pub(crate) fn pool_home() -> std::path::PathBuf {
+    #[cfg(test)]
+    {
+        std::env::temp_dir().join(format!("kigi-oauth-pool-test-{}", std::process::id()))
+    }
+    #[cfg(not(test))]
+    crate::util::kigi_home::kigi_home()
+}
+
+/// Get-or-create the process-global manager for `oauth`, wiring the same
+/// refresher + proactive-refresh lifecycle as the primary Kimi manager the
+/// FIRST time a scope is seen. The manager reads the on-disk token at
+/// construction (thereafter kept fresh by the proactive-refresh loop), so a
+/// grok login that lands after this scope was first built is adopted on the
+/// manager's own refresh tick — no session ever needs re-spawning.
+///
+/// MUST be called from within a Tokio runtime (the proactive-refresh loop
+/// spawns a task, mirroring the primary).
+pub(crate) fn global_manager_for(
+    kigi_home: &Path,
+    oauth: &'static kigi_models::OAuthConfig,
+) -> Arc<AuthManager> {
+    let mut pool = oauth_manager_pool().lock();
+    if let Some(existing) = pool.get(oauth.scope_key) {
+        return existing.clone();
+    }
+    let manager = Arc::new(AuthManager::new_oauth_provider(kigi_home, oauth));
+    manager.configure_refresher();
+    // Never-cancelled token = process-lifetime, matching the api-server /
+    // per-session eager-refresh sites that pass a fresh token.
+    manager.start_proactive_refresh(tokio_util::sync::CancellationToken::new());
+    pool.insert(oauth.scope_key, manager.clone());
+    manager
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn oauth_for(id: &str) -> &'static kigi_models::OAuthConfig {
+        kigi_models::PlatformId::parse(id)
+            .expect("known platform")
+            .oauth()
+            .expect("subscription-OAuth platform carries an OAuthConfig")
+    }
+
+    /// Each subscription-OAuth platform gets its OWN process-global pooled
+    /// manager, and no two share one. (Which credential governs a REQUEST is
+    /// not decided here — see
+    /// [`crate::auth::credential_authority::CredentialAuthority`].)
+    #[tokio::test]
+    async fn every_oauth_scope_gets_its_own_pooled_manager() {
+        let home = tempfile::tempdir().unwrap();
+        let ids = [
+            "xai-grok",
+            "claude-pro-max",
+            "github-copilot",
+            "openai-codex",
+        ];
+        let managers: Vec<_> = ids
+            .iter()
+            .map(|id| global_manager_for(home.path(), oauth_for(id)))
+            .collect();
+        for (i, a) in managers.iter().enumerate() {
+            assert!(
+                Arc::ptr_eq(a, &global_manager_for(home.path(), oauth_for(ids[i]))),
+                "{}: the pool must return the SAME manager for a scope",
+                ids[i]
+            );
+            for (j, b) in managers.iter().enumerate() {
+                if i != j {
+                    assert!(
+                        !Arc::ptr_eq(a, b),
+                        "{} and {} must not share a pooled manager",
+                        ids[i],
+                        ids[j]
+                    );
+                }
+            }
+        }
+    }
+
+    /// M8: the test pool home is a per-process path that is never created, so a
+    /// test binary leaves nothing behind (and never resolves the developer's
+    /// real `~/.kigi`, whose stored OAuth tokens the pool would otherwise read
+    /// and proactively refresh over the network).
+    ///
+    /// This is also the tripwire for the cleanup claim in [`pool_home`]: a
+    /// completed device login through `authenticate_oauth_platform` WOULD create
+    /// this directory, so if a lib test ever drives one, this assertion fires
+    /// and the cleanup has to be added rather than silently regressing the
+    /// "tests are TempDir self-cleaning" discipline.
+    #[test]
+    fn test_pool_home_is_disposable_and_never_the_real_home() {
+        let home = pool_home();
+        assert!(
+            home.starts_with(std::env::temp_dir()),
+            "the test pool home must live under the system temp dir, got {home:?}"
+        );
+        assert!(
+            !home.exists(),
+            "the test pool home must not be created — nothing to clean up"
+        );
+    }
+}

crates/codegen/kigi-shell/src/auth/refresh/generic_refresher.rs

@@ -0,0 +1,314 @@
+//! Generic OAuth token refresher for any [`kigi_models::OAuthConfig`] provider,
+//! driven through the [`TokenRefresher`] seam. The wire call is selected by the
+//! config's `token_body`: form-encoded `grant_type=refresh_token` (xai-grok,
+//! openai-codex), JSON (claude-pro-max), or the GitHub Copilot copilot-token
+//! RE-MINT. Providers with `requires_chatgpt_account_id` additionally fail fast
+//! when the refreshed token drops the claim.
+//!
+//! Structurally identical to [`super::kimi_refresher::KimiRefresher`] — same
+//! sibling-adoption + post-401 grace — but the wire call goes through
+//! [`crate::auth::oauth_device`] (no X-Msh headers) instead of the Kimi wire.
+//! Access/refresh tokens are NEVER logged.
+
+use std::sync::Arc;
+
+use kigi_models::{OAuthConfig, OAuthTokenBody};
+
+use crate::auth::error::RefreshTokenFailedReason;
+use crate::auth::kimi_oauth::RefreshError;
+use crate::auth::manager::RefreshReason;
+use crate::auth::{github_copilot, oauth_device, oauth_pkce};
+
+use super::{AuthSnapshot, RefreshOutcome, TokenRefresher};
+
+/// Grace period after a 401/403 before concluding the refresh token is dead:
+/// a concurrent instance may still be persisting its rotated token.
+const POST_UNAUTHORIZED_GRACE: std::time::Duration = std::time::Duration::from_secs(1);
+
+pub(crate) struct GenericDeviceRefresher {
+    auth: Arc<dyn AuthSnapshot>,
+    cfg: &'static OAuthConfig,
+}
+
+impl GenericDeviceRefresher {
+    pub(crate) fn new(auth: Arc<dyn AuthSnapshot>, cfg: &'static OAuthConfig) -> Self {
+        Self { auth, cfg }
+    }
+
+    /// Post-401 sibling check: wait a beat, re-read the persisted credential,
+    /// and adopt it when its refresh token differs from the rejected one.
+    async fn adopt_rotation_after_unauthorized(&self, tried_rt: &str) -> Option<RefreshOutcome> {
+        tokio::time::sleep(POST_UNAUTHORIZED_GRACE).await;
+        let latest = self.auth.read_disk_auth()?;
+        let latest_rt = latest.refresh_token.as_deref()?;
+        if latest_rt == tried_rt {
+            return None;
+        }
+        kigi_log::unified_log::info(
+            "auth.refresh.adopted_rotation_after_401",
+            None,
+            Some(serde_json::json!({
+                "scope_key": self.cfg.scope_key,
+                "adopted_rt_prefix": crate::auth::token_suffix(latest_rt),
+                "rejected_rt_prefix": crate::auth::token_suffix(tried_rt),
+            })),
+        );
+        Some(RefreshOutcome::success(latest))
+    }
+}
+
+#[async_trait::async_trait]
+impl TokenRefresher for GenericDeviceRefresher {
+    async fn refresh(&self, reason: RefreshReason) -> RefreshOutcome {
+        tracing::info!(
+            ?reason,
+            scope_key = self.cfg.scope_key,
+            "auth: generic refresh attempt"
+        );
+
+        let disk_auth = self.auth.read_disk_auth();
+
+        // Sibling short-circuit: a valid persisted token whose key differs from
+        // in-memory means another process refreshed already — adopt directly.
+        if let Some(ref d) = disk_auth
+            && !crate::auth::is_expired(d)
+            && self.auth.current().map(|a| a.key).as_deref() != Some(&d.key)
+        {
+            kigi_log::unified_log::info(
+                "auth.refresh.adopted_sibling_token",
+                None,
+                Some(serde_json::json!({
+                    "scope_key": self.cfg.scope_key,
+                    "disk_key_prefix": crate::auth::token_suffix(&d.key),
+                })),
+            );
+            return RefreshOutcome::success(d.clone());
+        }
+
+        let Some(auth) = super::resolve_refresh_credential(self.auth.as_ref(), disk_auth, reason)
+        else {
+            tracing::warn!(
+                ?reason,
+                "auth: no credential available for refresh (generic)"
+            );
+            return RefreshOutcome::transient("no token with refresh_token available");
+        };
+        let Some(refresh_token) = auth.refresh_token.clone() else {
+            tracing::warn!(
+                ?reason,
+                "auth: resolved credential has no refresh token (generic)"
+            );
+            return RefreshOutcome::transient("credential has no refresh token");
+        };
+
+        tracing::info!(
+            rt_prefix = crate::auth::token_suffix(&refresh_token),
+            expires_at = ?auth.expires_at,
+            "auth: sending refresh_token grant (generic oauth)"
+        );
+
+        // Refresh over the provider's token-body encoding: xai's endpoint is
+        // form-encoded (device wire); Claude's is JSON (PKCE wire); GitHub
+        // Copilot's "refresh" is a copilot-token RE-MINT — a `GET
+        // copilot_internal/v2/token` bearing the durable github token (the
+        // `refresh_token` field here), NOT a refresh_token grant. All three
+        // return the same `Result<KimiAuth, RefreshError>`.
+        let wire_result = match self.cfg.token_body {
+            OAuthTokenBody::Form => oauth_device::refresh_token(self.cfg, &refresh_token).await,
+            OAuthTokenBody::Json => oauth_pkce::refresh_token(self.cfg, &refresh_token).await,
+            OAuthTokenBody::GithubCopilotExchange => {
+                github_copilot::remint_copilot_token(self.cfg, &refresh_token).await
+            }
+        };
+        match wire_result {
+            Ok(new_auth)
+                if self.cfg.requires_chatgpt_account_id
+                    && kigi_sampling_types::chatgpt_account_id_from_jwt(&new_auth.key)
+                        .is_none() =>
+            {
+                // FAIL FAST (never silently): the refreshed token carries no
+                // `chatgpt_account_id`, so every inference request would go out
+                // WITHOUT the required `chatgpt-account-id` header and draw an
+                // opaque backend 4xx. Surface it as a permanent failure so the
+                // user is told to re-login.
+                tracing::warn!(
+                    scope_key = self.cfg.scope_key,
+                    "auth: refreshed token is missing the chatgpt_account_id claim"
+                );
+                RefreshOutcome::permanent(RefreshTokenFailedReason::Other, Some(refresh_token))
+            }
+            Ok(new_auth) => {
+                kigi_log::unified_log::info(
+                    "auth.refresh.token_rotated",
+                    None,
+                    Some(serde_json::json!({
+                        "scope_key": self.cfg.scope_key,
+                        "new_key_prefix": crate::auth::token_suffix(&new_auth.key),
+                        "expires_at": new_auth.expires_at.map(|e| e.to_rfc3339()),
+                    })),
+                );
+                RefreshOutcome::success(new_auth)
+            }
+            Err(RefreshError::Unauthorized {
+                status,
+                description,
+            }) => {
+                tracing::warn!(status, %description, "auth: refresh token rejected (generic)");
+                if let Some(adopted) = self.adopt_rotation_after_unauthorized(&refresh_token).await
+                {
+                    return adopted;
+                }
+                kigi_log::unified_log::warn(
+                    "auth.refresh.unauthorized",
+                    None,
+                    Some(serde_json::json!({
+                        "scope_key": self.cfg.scope_key,
+                        "status": status,
+                        "description": description,
+                        "rt_prefix": crate::auth::token_suffix(&refresh_token),
+                    })),
+                );
+                RefreshOutcome::permanent(
+                    RefreshTokenFailedReason::RefreshTokenRejected,
+                    Some(refresh_token),
+                )
+            }
+            Err(
+                e @ (RefreshError::Exhausted { .. }
+                | RefreshError::Fatal { .. }
+                | RefreshError::Local(_)),
+            ) => {
+                tracing::warn!(error = %e, "auth: refresh attempt failed (transient, generic)");
+                kigi_log::unified_log::warn(
+                    "auth.refresh.transient_wire_failure",
+                    None,
+                    Some(serde_json::json!({
+                        "scope_key": self.cfg.scope_key,
+                        "error": format!("{e}"),
+                    })),
+                );
+                RefreshOutcome::transient(format!("token refresh failed: {e}"))
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::auth::model::KimiAuth;
+    use chrono::{Duration, Utc};
+    use kigi_models::XAI_OAUTH_CONFIG;
+    use parking_lot::Mutex;
+    use wiremock::matchers::{body_string_contains, method, path};
+    use wiremock::{Mock, MockServer, ResponseTemplate};
+
+    struct FakeSnapshot {
+        current: Mutex<Option<KimiAuth>>,
+        disk: Mutex<Option<KimiAuth>>,
+    }
+    impl FakeSnapshot {
+        fn new(current: Option<KimiAuth>, disk: Option<KimiAuth>) -> Arc<Self> {
+            Arc::new(Self {
+                current: Mutex::new(current),
+                disk: Mutex::new(disk),
+            })
+        }
+    }
+    impl AuthSnapshot for FakeSnapshot {
+        fn current(&self) -> Option<KimiAuth> {
+            self.current
+                .lock()
+                .clone()
+                .filter(|a| !crate::auth::is_expired(a))
+        }
+        fn expired_auth(&self) -> Option<KimiAuth> {
+            self.current.lock().clone().filter(crate::auth::is_expired)
+        }
+        fn read_disk_auth(&self) -> Option<KimiAuth> {
+            self.disk.lock().clone()
+        }
+        fn is_expired(&self) -> bool {
+            self.current
+                .lock()
+                .as_ref()
+                .is_some_and(crate::auth::is_expired)
+        }
+    }
+
+    fn expired_session(key: &str, rt: &str) -> KimiAuth {
+        KimiAuth {
+            key: key.into(),
+            refresh_token: Some(rt.into()),
+            expires_at: Some(Utc::now() - Duration::hours(1)),
+            expires_in: Some(3600),
+            ..KimiAuth::test_default()
+        }
+    }
+
+    fn mock_cfg(host: &'static str) -> OAuthConfig {
+        OAuthConfig {
+            auth_host: host,
+            ..XAI_OAUTH_CONFIG
+        }
+    }
+
+    /// A successful refresh rotates the token via the generic wire.
+    #[tokio::test]
+    async fn refresh_success_returns_rotated_token() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .and(body_string_contains("refresh_token=grok-rt-old"))
+            .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "access_token": "grok-at-new",
+                "refresh_token": "grok-rt-new",
+                "expires_in": 3600,
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let stale = expired_session("grok-at-old", "grok-rt-old");
+        let snap = FakeSnapshot::new(Some(stale.clone()), Some(stale));
+        let cfg: &'static OAuthConfig = Box::leak(Box::new(mock_cfg(host)));
+        let refresher = GenericDeviceRefresher::new(snap, cfg);
+        let outcome = refresher.refresh(RefreshReason::PreRequest).await;
+        let RefreshOutcome::Success(new_auth) = outcome else {
+            panic!("expected success, got {outcome:?}");
+        };
+        assert_eq!(new_auth.key, "grok-at-new");
+        assert_eq!(new_auth.refresh_token.as_deref(), Some("grok-rt-new"));
+    }
+
+    /// A 401 on refresh (with no sibling rotation) tombstones the rejected
+    /// refresh token as a permanent failure.
+    #[tokio::test]
+    async fn unauthorized_is_permanent_failure() {
+        let server = MockServer::start().await;
+        let host: &'static str = Box::leak(server.uri().into_boxed_str());
+        Mock::given(method("POST"))
+            .and(path("/oauth2/token"))
+            .respond_with(
+                ResponseTemplate::new(401)
+                    .set_body_json(serde_json::json!({ "error_description": "revoked" })),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+        let stale = expired_session("grok-at-old", "grok-rt-dead");
+        let snap = FakeSnapshot::new(Some(stale.clone()), Some(stale));
+        let cfg: &'static OAuthConfig = Box::leak(Box::new(mock_cfg(host)));
+        let refresher = GenericDeviceRefresher::new(snap, cfg);
+        let outcome = refresher.refresh(RefreshReason::PreRequest).await;
+        let RefreshOutcome::PermanentFailure {
+            error,
+            rejected_refresh_token,
+        } = outcome
+        else {
+            panic!("expected permanent failure, got {outcome:?}");
+        };
+        assert_eq!(error.reason, RefreshTokenFailedReason::RefreshTokenRejected);
+        assert_eq!(rejected_refresh_token.as_deref(), Some("grok-rt-dead"));
+    }
+}

crates/codegen/kigi-shell/src/auth/refresh/mod.rs

@@ -1,3 +1,4 @@
+mod generic_refresher;
 mod kimi_refresher;
 
 use std::sync::Arc;
@@ -6,6 +7,7 @@ use crate::auth::manager::AuthManager;
 pub(crate) use crate::auth::manager::RefreshReason;
 use crate::auth::model::KimiAuth;
 
+pub(crate) use generic_refresher::GenericDeviceRefresher;
 pub(crate) use kimi_refresher::KimiRefresher;
 
 /// Read-only view of `AuthManager` for refreshers. Enforces the
@@ -120,8 +122,20 @@ pub(crate) trait TokenRefresher: Send + Sync {
     async fn refresh(&self, reason: RefreshReason) -> RefreshOutcome;
 }
 
-/// Build the production refresher against `kigi_env::oauth_host()`.
+/// Build the production refresher for this manager's scope. A scope that maps
+/// to a generic device-code [`kigi_models::OAuthConfig`] (xai-grok) gets the
+/// [`GenericDeviceRefresher`]; every other scope — Kimi Code, whose registry
+/// `oauth` field is `None` by design — gets the bespoke [`KimiRefresher`]
+/// against `kigi_env::oauth_host()`.
 pub(crate) fn build_refresher(auth_manager: Arc<AuthManager>) -> Arc<dyn TokenRefresher> {
+    match kigi_models::oauth_config_for_scope_key(auth_manager.scope()) {
+        Some(cfg) => {
+            let snapshot: Arc<dyn AuthSnapshot> = auth_manager;
+            Arc::new(GenericDeviceRefresher::new(snapshot, cfg))
+        }
+        None => {
             let snapshot: Arc<dyn AuthSnapshot> = auth_manager;
             Arc::new(KimiRefresher::new(snapshot, kigi_env::oauth_host()))
+        }
+    }
 }

crates/codegen/kigi-shell/src/auth/storage.rs

@@ -96,9 +96,9 @@ pub(crate) fn disable_mock_keyring_for_test() {
 fn keyring_entry() -> Result<&'static keyring::Entry, keyring::Error> {
     static ENTRY: std::sync::OnceLock<Result<keyring::Entry, keyring::Error>> =
         std::sync::OnceLock::new();
-    match ENTRY.get_or_init(|| {
-        keyring::Entry::new(KEYRING_SERVICE, crate::auth::config::KIMI_CODE_OAUTH_SCOPE)
-    }) {
+    match ENTRY
+        .get_or_init(|| keyring::Entry::new(KEYRING_SERVICE, crate::auth::KIMI_CODE_OAUTH_SCOPE))
+    {
         Ok(entry) => Ok(entry),
         // `keyring::Error` is not `Clone`; surface a stable equivalent.
         Err(e) => {
@@ -430,17 +430,12 @@ fn write_store_to(path: &Path, auth_store: &AuthStore) -> std::io::Result<()> {
     Ok(())
 }
 
-/// Atomic write: tmp + rename. Unix `rename(2)` replaces atomically;
-/// Windows `rename` requires removing the target first.
+/// Atomic write: tmp + Windows-safe replace (see `util::fs::replace_file`,
+/// which this site's inline delete-first pattern graduated into).
 fn write_auth_json_atomic(auth_file: &Path, auth_store: &AuthStore) -> std::io::Result<()> {
     let tmp = auth_file.with_extension(format!("json.{}.tmp", std::process::id()));
     write_store_to(&tmp, auth_store)?;
-    #[cfg(windows)]
-    {
-        let _ = std::fs::remove_file(auth_file);
-    }
-    std::fs::rename(&tmp, auth_file)?;
-    Ok(())
+    crate::util::fs::replace_file(&tmp, auth_file)
 }
 
 /// Non-atomic fallback: truncate and rewrite `auth.json` in place.
@@ -536,6 +531,71 @@ pub fn store_api_key(kigi_home: &Path, api_key: &str) -> std::io::Result<()> {
     write_auth_json(&path, &map)
 }
 
+/// Read an API-key platform's key from auth.json. The scope is the platform
+/// id itself (`anthropic`, `moonshot-cn`, …) — the stable per-provider
+/// auth.json key contract. `None` when absent or unreadable.
+pub fn read_platform_api_key(
+    kigi_home: &Path,
+    platform: kigi_models::PlatformId,
+) -> Option<String> {
+    let path = kigi_home.join("auth.json");
+    let map = read_auth_json(&path).ok()?;
+    map.get(platform.as_str()).map(|a| a.key.clone())
+}
+
+/// Store an API-key platform's key in auth.json under its platform-id scope.
+/// Same corrupt-recovery + atomic-write path as [`store_api_key`]; all other
+/// scopes (OAuth session, other platforms) are preserved.
+///
+/// SECURITY: the key must never be logged; errors carry only IO context.
+pub fn store_platform_api_key(
+    kigi_home: &Path,
+    platform: kigi_models::PlatformId,
+    api_key: &str,
+) -> std::io::Result<()> {
+    if platform.uses_oauth() {
+        // Real error, not debug_assert: an OAuth scope written here would
+        // shadow the session entry in release builds too.
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidInput,
+            format!(
+                "{} authenticates via OAuth and takes no API key",
+                platform.as_str()
+            ),
+        ));
+    }
+    let path = kigi_home.join("auth.json");
+    // Serialize with the manager's cross-process auth.json writers (token
+    // refresh holds the same flock): an unlocked read-modify-write here
+    // could write back a pre-refresh map and revert a rotated refresh
+    // token (token-family revocation → forced re-login). Bounded retry;
+    // a sustained holder fails loudly rather than racing.
+    let mut lock = None;
+    for _ in 0..20 {
+        lock = super::manager::try_lock_auth_file_nonblocking(&path);
+        if lock.is_some() {
+            break;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(25));
+    }
+    let Some(_lock) = lock else {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::WouldBlock,
+            "auth.json is locked by another kigi process; try again",
+        ));
+    };
+    let mut map = read_auth_json_or_empty_recovering_corrupt(&path)?;
+    map.insert(
+        platform.as_str().to_owned(),
+        KimiAuth {
+            key: api_key.to_owned(),
+            auth_mode: AuthMode::ApiKey,
+            ..Default::default()
+        },
+    );
+    write_auth_json(&path, &map)
+}
+
 /// Remove the `kigi::api_key` scope from auth.json.
 pub fn clear_api_key(kigi_home: &Path) -> std::io::Result<()> {
     let path = kigi_home.join("auth.json");
@@ -550,6 +610,80 @@ pub fn clear_api_key(kigi_home: &Path) -> std::io::Result<()> {
     Ok(())
 }
 
+#[cfg(test)]
+mod platform_key_tests {
+    use super::*;
+
+    /// Store/read round-trip under the platform-id scope; the OAuth session
+    /// scope and other platform scopes in the same file are preserved.
+    #[test]
+    fn platform_key_round_trip_preserves_other_scopes() {
+        let dir = tempfile::tempdir().unwrap();
+        let home = dir.path();
+        // Pre-existing OAuth session entry must survive platform-key writes.
+        let path = home.join("auth.json");
+        let mut map = AuthStore::new();
+        map.insert(
+            crate::auth::KIMI_CODE_OAUTH_SCOPE.to_owned(),
+            KimiAuth {
+                key: "oauth-token".to_owned(),
+                auth_mode: AuthMode::OAuth,
+                ..Default::default()
+            },
+        );
+        write_auth_json(&path, &map).unwrap();
+
+        store_platform_api_key(home, kigi_models::PlatformId::MoonshotCn, "sk-cn").unwrap();
+        store_platform_api_key(home, kigi_models::PlatformId::MoonshotAi, "sk-ai").unwrap();
+
+        assert_eq!(
+            read_platform_api_key(home, kigi_models::PlatformId::MoonshotCn).as_deref(),
+            Some("sk-cn")
+        );
+        assert_eq!(
+            read_platform_api_key(home, kigi_models::PlatformId::MoonshotAi).as_deref(),
+            Some("sk-ai")
+        );
+        let stored = read_auth_json(&path).unwrap();
+        assert_eq!(
+            stored
+                .get(crate::auth::KIMI_CODE_OAUTH_SCOPE)
+                .map(|a| a.key.as_str()),
+            Some("oauth-token"),
+            "platform-key writes must not clobber the OAuth session scope"
+        );
+        assert_eq!(
+            stored.get("moonshot-cn").map(|a| a.auth_mode.clone()),
+            Some(AuthMode::ApiKey),
+            "platform keys are stored as api_key mode under the platform id"
+        );
+    }
+
+    /// The OAuth platform takes no API key — a real error in release builds.
+    #[test]
+    fn storing_key_for_oauth_platform_is_invalid_input() {
+        let dir = tempfile::tempdir().unwrap();
+        let err = store_platform_api_key(dir.path(), kigi_models::PlatformId::KimiCode, "sk-x")
+            .expect_err("oauth platform must reject api keys");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+        assert!(
+            !dir.path().join("auth.json").exists(),
+            "rejected write must not create auth.json"
+        );
+    }
+
+    /// Missing file reads as None (not an error) — resolution treats absent
+    /// auth.json as "no stored key".
+    #[test]
+    fn reading_platform_key_without_auth_json_is_none() {
+        let dir = tempfile::tempdir().unwrap();
+        assert_eq!(
+            read_platform_api_key(dir.path(), kigi_models::PlatformId::MoonshotCn),
+            None
+        );
+    }
+}
+
 #[cfg(test)]
 mod write_fallback_tests {
     use super::*;

crates/codegen/kigi-shell/src/claude_import.rs

@@ -670,10 +670,7 @@ fn write_import_marker(config_path: &Path) -> anyhow::Result<()> {
         let _ = std::fs::remove_file(&tmp);
         return Err(e.into());
     }
-    if let Err(e) = std::fs::rename(&tmp, config_path) {
-        let _ = std::fs::remove_file(&tmp);
-        return Err(e.into());
-    }
+    crate::util::fs::replace_file(&tmp, config_path)?;
     Ok(())
 }
 
@@ -854,7 +851,7 @@ fn apply_items_to_config(config_path: &Path, items: &[ImportableItem]) -> anyhow
             std::fs::create_dir_all(parent)?;
         }
         std::fs::write(&tmp, &toml_str)?;
-        std::fs::rename(&tmp, config_path)?;
+        crate::util::fs::replace_file(&tmp, config_path)?;
         info!(
             path = %config_path.display(),
             count,
@@ -1204,7 +1201,7 @@ fn apply_hooks_to_dir(hooks_dir: &Path, items: &[ImportableItem]) -> anyhow::Res
         let json_str = serde_json::to_string_pretty(&root)?;
         let tmp = target.with_extension("json.tmp");
         std::fs::write(&tmp, &json_str)?;
-        std::fs::rename(&tmp, &target)?;
+        crate::util::fs::replace_file(&tmp, &target)?;
         info!(
             path = %target.display(),
             count,

crates/codegen/kigi-shell/src/claude_import_state.rs

@@ -90,7 +90,7 @@ pub fn save_import_state(state: &ImportState) -> std::io::Result<()> {
     // `claude_import_state.json.tmp` (the last extension is replaced).
     let tmp = path.with_extension("json.tmp");
     std::fs::write(&tmp, &json)?;
-    std::fs::rename(&tmp, &path)?;
+    crate::util::fs::replace_file(&tmp, &path)?;
     Ok(())
 }
 

crates/codegen/kigi-shell/src/cli_models.rs

@@ -35,7 +35,11 @@ impl AuthStatus {
                 .unwrap_or(&origin);
             return Self::LoggedIn(host.to_owned());
         }
-        let models = crate::agent::config::resolve_model_list(agent_config, None);
+        let models = crate::agent::config::resolve_model_list(
+            agent_config,
+            None,
+            &crate::agent::models::PlatformApiKeys::resolve(&agent_config.platforms),
+        );
         if crate::agent::auth_method::should_advertise_xai_api_key(models.values())
             && let Some(name) = models
                 .iter()
@@ -88,7 +92,9 @@ pub async fn list_models(
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::agent::auth_method::{LEGACY_XAI_API_KEY_ENV_VAR, XAI_API_KEY_ENV_VAR};
+    use crate::agent::auth_method::{
+        HOUSE_API_KEY_ENV_VAR, LEGACY_XAI_API_KEY_ENV_VAR, XAI_API_KEY_ENV_VAR,
+    };
     use crate::agent::config::Config;
     use crate::auth::{AuthMode, KimiAuth};
     use kigi_test_support::EnvGuard;
@@ -98,10 +104,11 @@ mod tests {
     ///
     /// Uses `KIGI_AUTH_PATH` (not `KIGI_SHARE_DIR`) so a OnceLock-cached real home
     /// with `auth.json` cannot leak into these tests.
-    fn isolate_auth_sources() -> (tempfile::TempDir, [EnvGuard; 7]) {
+    fn isolate_auth_sources() -> (tempfile::TempDir, [EnvGuard; 8]) {
         let dir = tempfile::tempdir().unwrap();
         let auth_path = dir.path().join("no-auth.json");
         let guards = [
+            EnvGuard::unset(HOUSE_API_KEY_ENV_VAR),
             EnvGuard::unset(XAI_API_KEY_ENV_VAR),
             EnvGuard::unset(LEGACY_XAI_API_KEY_ENV_VAR),
             EnvGuard::unset("KIGI_AUTH"),

crates/codegen/kigi-shell/src/extensions/auth.rs

@@ -56,18 +56,18 @@ fn handle_set_api_key(args: &acp::ExtRequest) -> ExtResult {
             crate::auth::clear_api_key(&kigi_home)
                 .map_err(|e| acp::Error::internal_error().data(e.to_string()))?;
             // SAFETY: ext_method is single-threaded per agent
-            unsafe { std::env::remove_var("XAI_API_KEY") };
+            unsafe { std::env::remove_var("KIGI_API_KEY") };
         } else {
             crate::auth::store_api_key(&kigi_home, k)
                 .map_err(|e| acp::Error::internal_error().data(e.to_string()))?;
             // SAFETY: ext_method is single-threaded per agent
-            unsafe { std::env::set_var("XAI_API_KEY", k) };
+            unsafe { std::env::set_var("KIGI_API_KEY", k) };
         }
     } else {
         crate::auth::clear_api_key(&kigi_home)
             .map_err(|e| acp::Error::internal_error().data(e.to_string()))?;
         // SAFETY: ext_method is single-threaded per agent
-        unsafe { std::env::remove_var("XAI_API_KEY") };
+        unsafe { std::env::remove_var("KIGI_API_KEY") };
     }
     ExtMethodResult::success(serde_json::json!({ "ok": true }))
         .to_ext_response()

crates/codegen/kigi-shell/src/kimi_import.rs

@@ -160,18 +160,22 @@ struct KimiProviderToml {
 }
 
 /// Built-in kigi platform a kimi provider duplicates, if any: provider type
-/// `kimi` is the Kimi Code subscription channel; the two Moonshot open
-/// platforms are recognized by their fixed production hosts (the same hosts
-/// `kigi_models::PlatformId::base_url` compiles in).
+/// `kimi` is the Kimi Code subscription channel; API-key platforms are
+/// recognized by their production hosts (the same hosts
+/// `kigi_models::PlatformId::base_url` compiles in — moonshot, openai, and
+/// every future registry row automatically).
 fn builtin_platform(provider: &KimiProviderToml) -> Option<PlatformId> {
     if provider.provider_type == "kimi" {
         return Some(PlatformId::KimiCode);
     }
-    match url_host(&provider.base_url) {
-        Some("api.moonshot.cn") => Some(PlatformId::MoonshotCn),
-        Some("api.moonshot.ai") => Some(PlatformId::MoonshotAi),
-        _ => None,
+    let host = url_host(&provider.base_url)?;
+    PlatformId::ALL.into_iter().find(|platform| {
+        if platform.uses_oauth() {
+            return false;
         }
+        let base = platform.base_url();
+        url_host(&base) == Some(host)
+    })
 }
 
 /// Host component of an http(s) URL. `None` for other schemes.
@@ -524,10 +528,7 @@ pub fn apply_at(plan: &KimiImportPlan, kigi_home: &Path) -> anyhow::Result<KimiA
             let _ = std::fs::remove_file(&tmp);
             return Err(e.into());
         }
-        if let Err(e) = std::fs::rename(&tmp, &config_path) {
-            let _ = std::fs::remove_file(&tmp);
-            return Err(e.into());
-        }
+        crate::util::fs::replace_file(&tmp, &config_path)?;
         info!(
             path = %config_path.display(),
             added = applied.total_added(),
@@ -607,7 +608,7 @@ model = "kimi-for-coding"
 max_context_size = 262144
 
 [models.my-openai]
-provider = "openrouter"
+provider = "customllm"
 model = "gpt-x"
 max_context_size = 128000
 
@@ -616,9 +617,9 @@ type = "kimi"
 base_url = "https://api.kimi.com/coding/v1"
 api_key = "sk-kimi-secret"
 
-[providers.openrouter]
+[providers.customllm]
 type = "openai_legacy"
-base_url = "https://openrouter.ai/api/v1"
+base_url = "https://llm.example.test/v1"
 api_key = "sk-or-secret"
 "#,
         )
@@ -707,7 +708,7 @@ api_key = "sk-or-secret"
         let m = &plan.custom_models[0];
         assert_eq!(m.alias, "my-openai");
         assert_eq!(m.model, "gpt-x");
-        assert_eq!(m.base_url, "https://openrouter.ai/api/v1");
+        assert_eq!(m.base_url, "https://llm.example.test/v1");
         assert_eq!(m.api_key.as_deref(), Some("sk-or-secret"));
         assert_eq!(m.context_window, Some(128_000));
 
@@ -846,7 +847,7 @@ api_key = "sk-ms"
         assert_eq!(model["model"].as_str().unwrap(), "gpt-x");
         assert_eq!(
             model["base_url"].as_str().unwrap(),
-            "https://openrouter.ai/api/v1"
+            "https://llm.example.test/v1"
         );
         assert_eq!(model["api_key"].as_str().unwrap(), "sk-or-secret");
         assert_eq!(model["context_window"].as_integer().unwrap(), 128_000);

crates/codegen/kigi-shell/src/sampling/error.rs

@@ -76,7 +76,7 @@ pub fn map_sampling_err_to_acp(err: SamplingError) -> acp::Error {
                     && crate::agent::auth_method::has_xai_api_key_env()
                 {
                     format!(
-                        "{message}\n\nYou have an API key set (XAI_API_KEY). \
+                        "{message}\n\nYou have an API key set (KIGI_API_KEY). \
                          Your cached OAuth session is being used instead. \
                          To use your API key, run `kigi logout` or type /logout in the TUI."
                     )
@@ -441,24 +441,32 @@ mod tests {
         );
     }
 
-    /// Helper: run a closure with XAI_API_KEY temporarily set (or cleared).
-    /// Cleans up even if the closure panics.
+    /// Helper: run a closure with the house BYOK key (KIGI_API_KEY) temporarily
+    /// set (or cleared). Clears every house-key env (KIGI_API_KEY plus the
+    /// back-compat XAI_API_KEY / KIGI_CODE_XAI_API_KEY) so the "no key" case is
+    /// hermetic. Cleans up even if the closure panics.
     fn with_api_key_env<F: FnOnce()>(key: Option<&str>, f: F) {
+        let prev_house = std::env::var("KIGI_API_KEY").ok();
         let prev = std::env::var("XAI_API_KEY").ok();
         let prev_legacy = std::env::var("KIGI_CODE_XAI_API_KEY").ok();
         // SAFETY: serial_test ensures no concurrent env mutation.
         unsafe {
+            std::env::remove_var("KIGI_API_KEY");
             std::env::remove_var("XAI_API_KEY");
             std::env::remove_var("KIGI_CODE_XAI_API_KEY");
             if let Some(k) = key {
-                std::env::set_var("XAI_API_KEY", k);
+                std::env::set_var("KIGI_API_KEY", k);
             }
         }
         let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(f));
         // Restore original state.
         unsafe {
+            std::env::remove_var("KIGI_API_KEY");
             std::env::remove_var("XAI_API_KEY");
             std::env::remove_var("KIGI_CODE_XAI_API_KEY");
+            if let Some(v) = prev_house {
+                std::env::set_var("KIGI_API_KEY", v);
+            }
             if let Some(v) = prev {
                 std::env::set_var("XAI_API_KEY", v);
             }

crates/codegen/kigi-shell/src/session/acp_session.rs

@@ -135,7 +135,7 @@ use prompt_build::*;
 mod session_mode;
 use session_mode::*;
 #[path = "acp_session_impl/sampler_turn.rs"]
-mod sampler_turn;
+pub(crate) mod sampler_turn;
 use sampler_turn::*;
 #[path = "acp_session_impl/tool_dispatch.rs"]
 mod tool_dispatch;
@@ -417,6 +417,21 @@ pub(crate) struct SessionActor {
     /// [`SessionActor::model_auth_facts`].
     pub(crate) model_auth_facts:
         std::cell::RefCell<Option<(String, crate::agent::config::ModelAuthFacts)>>,
+    /// The catalog KEY this session's model was selected by (`{platform}/{model}`
+    /// for a registry model), owned PER SESSION.
+    ///
+    /// H4: `SamplingConfig::model` is the bare routing slug, and duplicate slugs
+    /// across an API-key platform and its subscription-OAuth twin
+    /// (`xai`/`xai-grok`, `anthropic`/`claude-pro-max`, `openai`/`openai-codex`)
+    /// are BY DESIGN, so the slug alone cannot name the platform. This used to
+    /// be read from `ModelsManager::current_model_id()` — a single
+    /// PROCESS-GLOBAL cell that Leader mode never writes
+    /// (`agent/handlers/model_switch.rs`) and that is last-writer-wins across
+    /// concurrent sessions, so both collision directions resolved the wrong
+    /// platform: the subscription session lost its resolver (unrecoverable 401
+    /// ~1h in) and the API-key session got the pooled OAuth bearer stamped over
+    /// its own `sk-…` key. Written at spawn and on every `SetSessionModel`.
+    pub(crate) selected_catalog_key: std::cell::RefCell<Option<String>>,
     /// 401-attribution callback. Joined with the bearer the
     /// sampler sends on the wire to emit an `auth 401 attribution`
     /// event at each of the six `OaiCompatClient` 401 arms in
@@ -1132,7 +1147,7 @@ fn persist_chat_history_jsonl_sync(session_info: &SessionInfo, conversation: &[C
             buf.push(b'\n');
         }
         std::fs::File::create(&tmp_path)?.write_all(&buf)?;
-        std::fs::rename(&tmp_path, &final_path)?;
+        crate::util::fs::replace_file(&tmp_path, &final_path)?;
         Ok(())
     })();
     if let Err(e) = result {
@@ -1256,6 +1271,17 @@ mod rewind_synthetic_turn_tests;
 #[cfg(test)]
 #[path = "acp_session_tests/rewrite_zero_turn_prefix_tests.rs"]
 mod rewrite_zero_turn_prefix_tests;
+/// The same guard for the model→platform lookup (the dual-credential slug
+/// collision) and for the stamped aux/summary configs.
+#[cfg(test)]
+#[path = "acp_session_tests/session_bearer_leak_platform_tests.rs"]
+mod session_bearer_leak_platform_tests;
+/// LEAK guard: the primary Kimi subscription bearer must never ride a request
+/// to an API-key registry platform's host, while the subscription-OAuth
+/// platforms keep a live resolver from their OWN pooled manager.
+#[cfg(test)]
+#[path = "acp_session_tests/session_bearer_leak_tests.rs"]
+mod session_bearer_leak_tests;
 /// Pins the `SubagentFinished` usage-fold attribution gate.
 #[cfg(test)]
 #[path = "acp_session_tests/subagent_usage_fold_tests.rs"]

crates/codegen/kigi-shell/src/session/acp_session_impl/model_switch.rs

@@ -5,12 +5,18 @@ impl SessionActor {
     pub(super) async fn handle_set_session_model(
         &self,
         sampling_config: kigi_sampler::SamplerConfig,
+        catalog_key: Option<String>,
         use_concise: bool,
         apply_prompt_override: bool,
         skip_prompt_rewrite: bool,
         auto_compact_threshold_percent: u8,
     ) -> Result<acp::ModelId, acp::Error> {
         let model_id = acp::ModelId::new(sampling_config.model.clone());
+        // H4: record the picker's catalog KEY as this SESSION's own selection.
+        // `sampling_config.model` is the ambiguous bare slug; the key is what
+        // disambiguates an API-key platform from its subscription-OAuth twin,
+        // and it must never come from the process-global `current_model_id()`.
+        *self.selected_catalog_key.borrow_mut() = catalog_key;
         let new_context_window = self.compaction.context_window_override.unwrap_or_else(|| {
             std::num::NonZeroU64::new(sampling_config.context_window).unwrap_or_else(|| {
                 std::num::NonZeroU64::new(DEFAULT_CONTEXT_WINDOW)
@@ -53,22 +59,24 @@ impl SessionActor {
                 temperature: sampling_config.temperature,
                 top_p: sampling_config.top_p,
                 api_backend: sampling_config.api_backend.clone(),
+                chat_compat: sampling_config.chat_compat,
                 extra_headers: sampling_config.extra_headers.clone(),
                 context_window: new_context_window,
                 reasoning_effort: sampling_config.reasoning_effort,
                 stream_tool_calls: Some(sampling_config.stream_tool_calls),
             });
         let existing = self.chat_state_handle.get_credentials().await;
-        let session_key = self
-            .auth_manager
-            .as_ref()
-            .and_then(|am| am.current_or_expired().map(|a| a.key));
+        // Read the session bearer from the switched-to model's OWN manager: a
+        // grok model reads the xai-grok token (used only to classify the
+        // credential's auth_type here), never the Kimi one. Kimi / non-oauth
+        // models resolve to the primary — byte-identical.
+        let session_key = self.session_credential_for_model(&sampling_config.model);
         self.chat_state_handle
             .update_credentials(kigi_chat_state::Credentials {
                 api_key: sampling_config.api_key.clone(),
                 auth_type: crate::agent::config::resolve_chat_state_auth_type(
                     sampling_config.model.as_str(),
-                    session_key.as_deref(),
+                    session_key.as_ref(),
                     existing.auth_type,
                 ),
                 alpha_test_key: existing.alpha_test_key,

crates/codegen/kigi-shell/src/session/acp_session_impl/prompt_build.rs

@@ -624,10 +624,20 @@ impl SessionActor {
         let resolved_describe = self
             .resolve_aux_sampler_config(&self.image_description_model)
             .await;
+        // LEAK 1b: the aux bearer_resolver is decided at the chokepoint from the
+        // IMAGE-DESCRIBE model's own platform + endpoint and passed in
+        // explicitly, so an aux model on another provider can never inherit the
+        // session (Kimi) resolver and have its own key overwritten on the aux
+        // host. The `None` fallback yields the SESSION config verbatim, whose
+        // own resolver must stay as-is.
+        let describe_resolver = resolved_describe
+            .as_ref()
+            .map(|cfg| self.aux_bearer_resolver(&self.image_description_model, &cfg.base_url));
         let (describe_model, sampler_config) =
             crate::agent::config::finalize_image_describe_sampler_config(
                 resolved_describe,
                 &active_session_config,
+                describe_resolver.flatten(),
                 Some(self.max_retries),
             );
         let client = kigi_sampler::SamplingClient::new(sampler_config).map_err(|e| {

crates/codegen/kigi-shell/src/session/acp_session_impl/run_loop.rs

@@ -300,10 +300,10 @@ pub(super) async fn run_session(
                    SessionActor::maybe_start_running_task(session.clone(), completion_tx
                    .clone()). await; } SessionCommand::SessionMode { session_mode, responds_to }
                    => { session.handle_session_mode(session_mode). await; let _ = responds_to
-                   .send(()); } SessionCommand::SetSessionModel { sampling_config, use_concise,
-                   apply_prompt_override, skip_prompt_rewrite, auto_compact_threshold_percent,
-                   responds_to } => { let updated_model_id = session
-                   .handle_set_session_model(sampling_config, use_concise,
+                   .send(()); } SessionCommand::SetSessionModel { sampling_config, catalog_key,
+                   use_concise, apply_prompt_override, skip_prompt_rewrite,
+                   auto_compact_threshold_percent, responds_to } => { let updated_model_id =
+                   session.handle_set_session_model(sampling_config, catalog_key, use_concise,
                    apply_prompt_override, skip_prompt_rewrite, auto_compact_threshold_percent).
                    await; let _ = responds_to.send(updated_model_id); }
                    SessionCommand::RebuildAgentForDefinition { definition, responds_to } => {
@@ -319,11 +319,24 @@ pub(super) async fn run_session(
                    .signals_handle().set_primary_model(& model_name); cfg.model = model_name
                    .clone(); cfg.extra_headers.extend(extra_headers); if let Some(cw) =
                    context_window && session.compaction.context_window_override.is_none() { cfg
-                   .context_window = cw; } session.chat_state_handle
+                   .context_window = cw; } let override_base_url = cfg
+                   .base_url.clone(); session.chat_state_handle
                    .update_sampling_config(cfg); let existing = session.chat_state_handle
-                   .get_credentials(). await; if let Some(r) = crate
-                   ::agent::config::try_resolve_model_credentials(model_name.as_str(), existing
-                   .api_key.as_deref()) { session.chat_state_handle
+                   .get_credentials(). await;
+                   // H-c: the rename makes the session's own selected catalog
+                   // key stale unless it still names this model; a stale key is
+                   // exactly what the model→platform rule must not trust.
+                   session.retain_selected_catalog_key_for(& model_name);
+                   // The override model routes to the SAME endpoint the session
+                   // already had; ask the chokepoint whether that endpoint takes
+                   // a session credential rather than re-offering the key
+                   // already in chat state. The platform comes from the session's
+                   // OWN lookup so this and every later turn agree.
+                   let override_session_key = session.credential_authority()
+                   .credential_for(session.model_platform(model_name.as_str()), &
+                   override_base_url); if let Some(r) = crate
+                   ::agent::config::try_resolve_model_credentials(model_name.as_str(),
+                   override_session_key.as_ref()) { session.chat_state_handle
                    .update_credentials(kigi_chat_state::Credentials { api_key : r.api_key,
                    auth_type : r.auth_type, alpha_test_key : existing.alpha_test_key, }); } session.model_auth_facts
                    .replace(None); } } SessionCommand::GetCurrentModel { responds_to } => { let

crates/codegen/kigi-shell/src/session/acp_session_impl/sampler_turn.rs

@@ -29,37 +29,105 @@ pub(super) fn is_auth_tool_error(err: &kigi_tool_runtime::ToolError) -> bool {
 /// Gate inputs bundled with the composed decision so the 401-recovery log can
 /// report the components.
 #[derive(Clone, Copy)]
-struct SessionTokenAuthGate {
+pub(crate) struct SessionTokenAuthGate {
     is_session_based: bool,
     model_byok: crate::agent::auth_method::ModelByok,
     /// Whether the request targets a first-party host. Lets an `Unknown`
     /// BYOK status still refresh against the first-party cli-chat-proxy hosts without
     /// risking a session-token leak to a third-party BYOK endpoint.
     endpoint_is_first_party: bool,
+    /// WHICH credential this model's platform/endpoint pair accepts, per the
+    /// single credential chokepoint
+    /// ([`crate::auth::credential_authority::CredentialAuthority::credential_class`]).
+    /// `None` for every API-key registry platform, which keeps the primary Kimi
+    /// bearer off `api.deepseek.com` / `api.openai.com` / … .
+    credential_class: crate::auth::credential_authority::CredentialClass,
 }
 impl SessionTokenAuthGate {
     /// Single place `is_session_based` / `endpoint_is_first_party` are derived,
-    /// so all call sites assemble the gate identically.
-    fn new(
+    /// so all call sites assemble the gate identically. `model_platform` is the
+    /// registry platform the model routes to (`None` for a bare / `[model.*]`
+    /// entry) — it MUST be derived from the same lookup
+    /// ([`SessionActor::model_platform`]) that
+    /// [`SessionActor::auth_manager_for_endpoint`] uses, so the gate's verdict
+    /// and the manager actually wrapped as the bearer resolver can never
+    /// disagree. `authority` is that same chokepoint, so the gate cannot answer
+    /// the endpoint question differently from the manager routing.
+    pub(crate) fn new(
         auth_method_id: Option<&acp::AuthMethodId>,
         model_byok: crate::agent::auth_method::ModelByok,
         base_url: &str,
+        model_platform: Option<kigi_models::PlatformId>,
+        authority: &crate::auth::credential_authority::CredentialAuthority,
     ) -> Self {
         Self {
+            // L13: a model whose OWN credential is a pooled subscription-OAuth
+            // session is session-based BY ITSELF, whatever the primary ACP
+            // method is. A user logged in with an API-KEY platform (e.g.
+            // `deepseek`) who selects a `claude-pro-max/*` model still gets that
+            // platform's pooled bearer as the request's `api_key` — without this
+            // term the gate would be inactive, so the config would carry NO
+            // resolver: the token freezes at selection time and the session dies
+            // with an unrecoverable 401 once it expires (~1h). The outer
+            // `credential_class` conjunct keeps this confined to that
+            // platform's own host.
             is_session_based: auth_method_id
-                .is_some_and(crate::agent::auth_method::is_session_based_method),
+                .is_some_and(crate::agent::auth_method::is_session_based_method)
+                || model_platform.is_some_and(|p| p.oauth().is_some()),
             model_byok,
             endpoint_is_first_party: crate::util::is_first_party_url(base_url),
+            credential_class: authority.credential_class(model_platform, base_url),
         }
     }
-    fn active(self) -> bool {
+    pub(crate) fn active(self) -> bool {
         crate::agent::auth_method::session_token_auth_gate(
             self.is_session_based,
             self.model_byok,
             self.endpoint_is_first_party,
+            self.credential_class,
         )
     }
 }
+/// THE aux / summary `bearer_resolver` rule, stated ONCE.
+///
+/// `SamplingClient::post` REPLACES the request's auth header from the resolver,
+/// so an aux model on a different provider would have its own correctly-resolved
+/// key overwritten by the session bearer ON THE AUX HOST. An OAuth aux model
+/// gets a live resolver over ITS OWN pooled manager (keeping mid-session
+/// refresh); a first-party aux model gets the primary's, but ONLY when the
+/// session-token gate is active; everything else gets `None`, so the aux model's
+/// own key survives to the wire.
+///
+/// M3 — the FIRST-PARTY case honours the gate, which is what the old "copy
+/// `active_session_config.bearer_resolver`" shape did implicitly: that field is
+/// `None` whenever the gate is inactive. Without it, a BYOK / api-key session
+/// with a `[model.*]` aux entry carrying its OWN `env_key` on the session's own
+/// coding endpoint has that key REPLACED on the wire by the primary bearer on
+/// every image-describe / auto-mode-classifier / summary request. A
+/// subscription-OAuth aux model is deliberately NOT gated this way: its pooled
+/// token IS its credential, and withholding the resolver only costs it
+/// mid-session refresh (L13).
+///
+/// Shared by [`SessionActor::aux_bearer_resolver`] and
+/// `MvpAgent::summary_bearer_resolver`: the summary client is built by the
+/// AGENT, not the session actor, and its own private copy of this rule is
+/// exactly how it stayed ungated after M3 closed the session-actor side.
+pub(crate) fn aux_bearer_resolver_for(
+    authority: &crate::auth::credential_authority::CredentialAuthority,
+    auth_method_id: Option<&acp::AuthMethodId>,
+    platform: Option<kigi_models::PlatformId>,
+    model_byok: crate::agent::auth_method::ModelByok,
+    base_url: &str,
+) -> Option<kigi_sampler::SharedBearerResolver> {
+    let is_primary_channel = platform.is_none_or(|p| p.oauth().is_none());
+    if is_primary_channel
+        && !SessionTokenAuthGate::new(auth_method_id, model_byok, base_url, platform, authority)
+            .active()
+    {
+        return None;
+    }
+    authority.bearer_resolver_for(platform, base_url)
+}
 /// Run a tool call; on an auth-shaped failure, attempt recovery via
 /// `AuthManager` and one retry. When `shared_recovery` is `Some`, concurrent
 /// 401s in the same batch deduplicate via `OnceCell::get_or_init`.
@@ -103,6 +171,31 @@ where
         result
     }
 }
+/// Wraps an [`AuthManager`](crate::auth::AuthManager) as a sampler
+/// [`BearerResolver`](kigi_sampler::BearerResolver), resolving the live
+/// (current-or-expired) bearer at request time. Shared by
+/// [`SessionActor::reconstruct_full_config`] (the session model) and the
+/// aux-model bearer routing
+/// ([`CredentialAuthority::bearer_resolver_for`](crate::auth::credential_authority::CredentialAuthority::bearer_resolver_for))
+/// so both wrap ONE definition. SECURITY: the bearer is resolved per request
+/// and never logged.
+pub(crate) struct AuthManagerBearerResolver(pub(crate) std::sync::Arc<crate::auth::AuthManager>);
+impl std::fmt::Debug for AuthManagerBearerResolver {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AuthManagerBearerResolver").finish()
+    }
+}
+impl kigi_sampler::BearerResolver for AuthManagerBearerResolver {
+    fn current_bearer(&self) -> Option<String> {
+        self.0.current_or_expired().map(|a| a.key)
+    }
+}
+/// Wrap `am` as a shared sampler bearer resolver.
+pub(crate) fn auth_manager_bearer_resolver(
+    am: std::sync::Arc<crate::auth::AuthManager>,
+) -> kigi_sampler::SharedBearerResolver {
+    std::sync::Arc::new(AuthManagerBearerResolver(am))
+}
 impl SessionActor {
     pub(super) async fn prepare_tool_definitions_timed(&self) -> (Vec<ToolDefinition>, u64) {
         let mcp_wait_start = std::time::Instant::now();
@@ -146,8 +239,8 @@ impl SessionActor {
         let plan_active = self.plan_mode.lock().is_active();
         filter_cursor_tools_by_plan_mode(defs, plan_active)
     }
-    /// Memoized per-model [`ModelAuthFacts`](crate::agent::config::ModelAuthFacts),
-    /// keyed by `model_id`.
+    /// Memoized per-model [`ModelAuthFacts`](crate::agent::config::ModelAuthFacts)
+    /// for the SESSION's own model, keyed by `model_id`.
     ///
     /// A fresh `Unknown` (config currently unparseable) falls back to the last
     /// definite value for the same `model_id` rather than demoting a live session
@@ -156,6 +249,32 @@ impl SessionActor {
     /// `model_id`, keying on `model_id` alone is insufficient — each
     /// model/credential chokepoint must clear this memo (`replace(None)`).
     pub(super) fn model_auth_facts(&self, model_id: &str) -> crate::agent::config::ModelAuthFacts {
+        self.resolve_auth_facts(model_id, true)
+    }
+    /// [`Self::model_auth_facts`] for a model that is NOT the session's own — an
+    /// AUX / summary / image-describe slug.
+    ///
+    /// Identical resolution, but it NEVER WRITES the slot. The memo is a SINGLE
+    /// slot: when the aux path shared it, one classifier or image-describe call
+    /// evicted the session model's entry, and (a) the next
+    /// [`Self::reconstruct_full_config`] paid another `load_effective_config()`
+    /// + `resolve_model_list()` — the per-turn disk read M7/M9 removed — while
+    /// (b) a transient `Unknown` for the SESSION model then had no same-`model_id`
+    /// definite value to fall back to, so it degraded to `endpoint_is_first_party`
+    /// — `false` for every subscription-OAuth host, costing the session its
+    /// `bearer_resolver` and 401ing unrecoverably ~1h in (the failure L13
+    /// prevents). Reading a matching entry is still allowed: it can only hit when
+    /// the slot already names this same slug.
+    fn aux_model_auth_facts(&self, model_id: &str) -> crate::agent::config::ModelAuthFacts {
+        self.resolve_auth_facts(model_id, false)
+    }
+    /// Shared body of [`Self::model_auth_facts`] / [`Self::aux_model_auth_facts`].
+    /// `memoize` is the ONLY difference, so the two can never resolve differently.
+    fn resolve_auth_facts(
+        &self,
+        model_id: &str,
+        memoize: bool,
+    ) -> crate::agent::config::ModelAuthFacts {
         use crate::agent::auth_method::ModelByok;
         if let Some((cached_id, facts)) = self.model_auth_facts.borrow().as_ref()
             && cached_id == model_id
@@ -172,17 +291,187 @@ impl SessionActor {
             }
             return fresh;
         }
+        if memoize {
             *self.model_auth_facts.borrow_mut() = Some((model_id.to_string(), fresh));
+        }
         fresh
     }
-    /// Gate inputs for `model_id` routed to `base_url`. See
+    /// Gate inputs for the SESSION model `model_id` routed to `base_url`. See
     /// [`crate::agent::auth_method::session_token_auth_gate`] for the rationale
     /// (`base_url` keeps an `Unknown` BYOK status refreshable only
     /// against first-party xAI hosts).
     fn auth_gate(&self, model_id: &str, base_url: &str) -> SessionTokenAuthGate {
         let byok = self.model_auth_facts(model_id).byok;
         let auth_method = self.auth_method_id.load();
-        SessionTokenAuthGate::new(auth_method.as_deref(), byok, base_url)
+        SessionTokenAuthGate::new(
+            auth_method.as_deref(),
+            byok,
+            base_url,
+            self.model_platform(model_id),
+            &self.credential_authority(),
+        )
+    }
+    /// This session's credential chokepoint: its EFFECTIVE endpoints (so a
+    /// managed `[endpoints] coding_api_base_url` deployment keeps the session
+    /// bearer — H3) plus its primary manager, which the authority keeps
+    /// private. Every inference-auth question this actor asks goes through it.
+    pub(crate) fn credential_authority(
+        &self,
+    ) -> crate::auth::credential_authority::CredentialAuthority {
+        crate::auth::credential_authority::CredentialAuthority::new(
+            self.models_manager.endpoints(),
+            self.auth_manager.clone(),
+        )
+    }
+    /// The [`AuthManager`](crate::auth::AuthManager) that governs INFERENCE auth
+    /// for the routing slug `model` against the endpoint the request will
+    /// ACTUALLY be sent to, from the ONE chokepoint
+    /// ([`crate::auth::credential_authority::CredentialAuthority`]).
+    ///
+    /// A subscription-OAuth platform routes to ITS OWN scope-keyed pooled
+    /// manager; `kimi-code` and a platform-less model on the session's own
+    /// coding endpoint route to the primary; every API-key registry platform,
+    /// and any endpoint that is neither, routes to `None` — fail fast, never a
+    /// silent fallback to the primary. `None` also for a BYOK / test session
+    /// with no primary.
+    ///
+    /// Callers pass the LIVE sampling config's `base_url` so the manager, the
+    /// gate and the wire can never be resolved against three different endpoints
+    /// (an `OverrideModelName` session keeps its original `base_url` under a
+    /// routing name absent from the catalog). A `model`-only sibling that
+    /// re-derived the endpoint from the CATALOG instead used to exist beside
+    /// this; it had zero callers and was deleted rather than left as a second,
+    /// unexercised way to answer the same question (`lib.rs`'s
+    /// `#![allow(dead_code)]` means such a helper raises no warning).
+    pub(super) fn auth_manager_for_endpoint(
+        &self,
+        model: &str,
+        base_url: &str,
+    ) -> Option<std::sync::Arc<crate::auth::AuthManager>> {
+        self.credential_authority()
+            .manager_for(self.model_platform(model), base_url)
+    }
+    /// The SESSION credential (if any) that may ride a request for the routing
+    /// slug `model`. The only producer is the chokepoint.
+    pub(super) fn session_credential_for_model(
+        &self,
+        model: &str,
+    ) -> Option<crate::auth::credential_authority::SessionCredential> {
+        self.credential_authority()
+            .credential_for(self.model_platform(model), &self.model_base_url(model))
+    }
+    /// The base URL a routing slug actually resolves to in the live catalog.
+    /// Falls back to the session's own inference endpoint for an unlisted slug,
+    /// which is exactly where `resolve_aux_model_sampling_config`'s Tier-2
+    /// fallback entry routes — so the endpoint the rule is applied to is always
+    /// the endpoint the request is sent to.
+    fn model_base_url(&self, model: &str) -> String {
+        let models = self.models_manager.models();
+        match crate::agent::config::find_model_by_id(&models, model) {
+            Some(entry) => entry.info().base_url.clone(),
+            None => self.models_manager.endpoints().resolve_inference_base_url(),
+        }
+    }
+    /// This SESSION's own selected catalog key (H4) — never the process-global
+    /// `ModelsManager::current_model_id()`, which Leader mode never writes and
+    /// which is last-writer-wins across concurrent sessions.
+    pub(super) fn selected_catalog_key(&self) -> Option<String> {
+        self.selected_catalog_key.borrow().clone()
+    }
+    /// Keep the session's own selected catalog key consistent with an
+    /// `OverrideModelName` rename: KEEP it when it still names `model_name`
+    /// (same entry, new routing name), otherwise CLEAR it.
+    ///
+    /// H-c: `OverrideModelName` is the one command that rewrites
+    /// `SamplingConfig::model` without going through `SetSessionModel`, so it
+    /// used to leave the field naming a model the session is no longer on.
+    /// Clearing rather than re-resolving is deliberate: re-resolving would put
+    /// `resolve_catalog_key`'s `.rev()` guess INTO the field the whole rule
+    /// treats as the session's deliberate selection, and a cleared field
+    /// refuses a collided slug instead of guessing its OAuth twin (H-b).
+    pub(super) fn retain_selected_catalog_key_for(&self, model_name: &str) {
+        let models = self.models_manager.models();
+        let still_names_it = self.selected_catalog_key().is_some_and(|key| {
+            key == model_name
+                || models
+                    .get(key.as_str())
+                    .is_some_and(|entry| entry.info.model == model_name)
+        });
+        if !still_names_it {
+            *self.selected_catalog_key.borrow_mut() = None;
+        }
+    }
+    /// The registry platform the routing slug `model` belongs to, from the SAME
+    /// lookup [`Self::auth_manager_for_endpoint`] routes on. `None` for a bare /
+    /// `[model.*]` / unlisted model.
+    pub(super) fn model_platform(&self, model: &str) -> Option<kigi_models::PlatformId> {
+        let models = self.models_manager.models();
+        crate::agent::models::platform_for_slug(
+            &models,
+            self.selected_catalog_key().as_deref(),
+            model,
+        )
+    }
+    /// Whether `model` routes to the Claude Pro/Max OAuth-Messages platform
+    /// (claude-pro-max) — the gate for the sampler's OAuth Messages adaptation
+    /// (identity headers + "You are Claude Code" system prefix). A generic-OAuth
+    /// platform speaking the Messages wire; every other model (incl. xai-grok,
+    /// which is ChatCompletions) returns `false`, keeping the API-key Anthropic
+    /// / MiniMax Messages requests byte-identical.
+    fn model_is_anthropic_oauth(&self, model: &str) -> bool {
+        self.model_platform(model).is_some_and(|platform| {
+            platform.oauth().is_some()
+                && platform.wire_api() == kigi_models::PlatformWireApi::Messages
+        })
+    }
+    /// Whether `model` routes to the GitHub Copilot ChatCompletions platform
+    /// (github-copilot) — the gate for the sampler's editor-identity headers +
+    /// `X-Initiator`. Every other model returns `false`, keeping the other
+    /// ChatCompletions providers byte-identical.
+    fn model_is_github_copilot(&self, model: &str) -> bool {
+        self.model_platform(model)
+            .is_some_and(kigi_models::PlatformId::sends_copilot_editor_headers)
+    }
+    /// Whether `model` routes to the ChatGPT/Codex Responses platform
+    /// (openai-codex) — the gate for the sampler's Codex identity headers
+    /// (`chatgpt-account-id` + originator + OpenAI-Beta). Every other model
+    /// returns `false`, keeping the API-key `openai` Responses request
+    /// byte-identical.
+    fn model_is_openai_codex(&self, model: &str) -> bool {
+        self.model_platform(model)
+            .is_some_and(kigi_models::PlatformId::sends_codex_responses_headers)
+    }
+    /// The `bearer_resolver` an AUX / summary `SamplerConfig` may carry — the
+    /// shared [`aux_bearer_resolver_for`] rule applied to the AUX model's own
+    /// platform + endpoint.
+    ///
+    /// The aux config never inherits the session resolver: it is passed this
+    /// value explicitly (see
+    /// [`crate::agent::config::stamp_session_local_sampler_fields`]), so
+    /// "forgot to re-point" is not expressible.
+    ///
+    /// The BYOK status comes from [`Self::aux_model_auth_facts`], which does NOT
+    /// write the session model's single-slot memo.
+    pub(super) fn aux_bearer_resolver(
+        &self,
+        slug: &str,
+        base_url: &str,
+    ) -> Option<kigi_sampler::SharedBearerResolver> {
+        let auth_method = self.auth_method_id.load();
+        // An aux slug is NOT the session's selection, so it must not be resolved
+        // against `selected_catalog_key` — the same rule the aux `api_key` obeys
+        // (`credential_for_slug(.., None, ..)`). Keying an aux model on the
+        // SESSION's selection let a colliding same-vendor slug resolve the OAuth
+        // twin, whose pooled resolver would then overwrite the user's own key on
+        // the aux request.
+        let models = self.models_manager.models();
+        aux_bearer_resolver_for(
+            &self.credential_authority(),
+            auth_method.as_deref(),
+            crate::agent::models::platform_for_slug(&models, None, slug),
+            self.aux_model_auth_facts(slug).byok,
+            base_url,
+        )
     }
     /// Emit a unified-log breadcrumb whenever the session-token refresh gate is
     /// evaluated with an **`Unknown`** per-model BYOK status on a session-based
@@ -202,8 +491,8 @@ impl SessionActor {
         let ctx = serde_json::json!(
             { "site" : site, "model_byok" : gate.model_byok.as_str(), "is_session_based"
             : gate.is_session_based, "endpoint_is_first_party" : gate
-            .endpoint_is_first_party, "refresh_active" : refresh_active, "base_url" :
-            base_url, }
+            .endpoint_is_first_party, "credential_class" : gate.credential_class
+            .as_str(), "refresh_active" : refresh_active, "base_url" : base_url, }
         );
         let sid = Some(self.session_info.id.0.as_ref());
         if refresh_active {
@@ -237,18 +526,6 @@ impl SessionActor {
                 }
             }
         }
-        #[allow(clippy::items_after_statements)]
-        struct AuthManagerBearerResolver(std::sync::Arc<crate::auth::AuthManager>);
-        impl std::fmt::Debug for AuthManagerBearerResolver {
-            fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-                f.debug_struct("AuthManagerBearerResolver").finish()
-            }
-        }
-        impl kigi_sampler::BearerResolver for AuthManagerBearerResolver {
-            fn current_bearer(&self) -> Option<String> {
-                self.0.current_or_expired().map(|a| a.key)
-            }
-        }
         let cfg = self
             .chat_state_handle
             .get_sampling_config()
@@ -260,6 +537,7 @@ impl SessionActor {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 extra_headers: Default::default(),
                 context_window: std::num::NonZeroU64::new(256_000).unwrap(),
                 reasoning_effort: None,
@@ -268,10 +546,32 @@ impl SessionActor {
         let creds = self.chat_state_handle.get_credentials().await;
         let model_facts = self.model_auth_facts(cfg.model.as_str());
         let auth_method = self.auth_method_id.load();
-        let gate =
-            SessionTokenAuthGate::new(auth_method.as_deref(), model_facts.byok, &cfg.base_url);
+        let gate = SessionTokenAuthGate::new(
+            auth_method.as_deref(),
+            model_facts.byok,
+            &cfg.base_url,
+            self.model_platform(cfg.model.as_str()),
+            &self.credential_authority(),
+        );
         let use_bearer_resolver = gate.active();
         self.log_auth_gate_unknown("reconstruct_full_config", gate, &cfg.base_url);
+        // Resolve the bearer from the ACTIVE model's OWN manager: a grok model
+        // wraps the xai-grok manager, never the Kimi one (captured before
+        // `cfg.model` is moved into the struct below). `None` when the gate is
+        // inactive or the oauth provider has no manager (fail-fast, no Kimi
+        // fallback).
+        let inference_auth_manager = if use_bearer_resolver {
+            self.auth_manager_for_endpoint(&cfg.model, &cfg.base_url)
+        } else {
+            None
+        };
+        // Claude Pro/Max OAuth Messages adaptation for THIS turn's model
+        // (captured before `cfg.model` is moved into the struct below).
+        let anthropic_oauth = self.model_is_anthropic_oauth(&cfg.model);
+        // GitHub Copilot editor-identity headers for THIS turn's model.
+        let github_copilot = self.model_is_github_copilot(&cfg.model);
+        // ChatGPT/Codex identity headers for THIS turn's model.
+        let openai_codex = self.model_is_openai_codex(&cfg.model);
         let auth_scheme = model_facts.auth_scheme;
         let mut extra_headers = cfg.extra_headers;
         crate::agent::config::inject_url_derived_headers(
@@ -312,6 +612,10 @@ impl SessionActor {
             top_p: cfg.top_p,
             api_backend: cfg.api_backend,
             auth_scheme,
+            anthropic_oauth,
+            github_copilot,
+            openai_codex,
+            chat_compat: cfg.chat_compat,
             extra_headers,
             context_window: cfg.context_window.get(),
             reasoning_effort: cfg.reasoning_effort,
@@ -321,15 +625,7 @@ impl SessionActor {
             idle_timeout_secs: None,
             origin_client: self.origin_client.clone(),
             attribution_callback: self.attribution_callback.clone(),
-            bearer_resolver: if use_bearer_resolver {
-                self.auth_manager
-                    .as_ref()
-                    .map(|am| -> kigi_sampler::SharedBearerResolver {
-                        std::sync::Arc::new(AuthManagerBearerResolver(am.clone()))
-                    })
-            } else {
-                None
-            },
+            bearer_resolver: inference_auth_manager.map(auth_manager_bearer_resolver),
             supports_backend_search: self.supports_backend_search.get(),
             compactions_remaining: self.compactions_remaining.get(),
             compaction_at_tokens: self.compaction_at_tokens.get(),
@@ -459,17 +755,27 @@ impl SessionActor {
         slug: &str,
     ) -> Option<kigi_sampler::SamplerConfig> {
         let creds = self.chat_state_handle.get_credentials().await;
-        let session_key = self
-            .auth_manager
-            .as_ref()
-            .and_then(|am| am.current_or_expired().map(|a| a.key.clone()));
         let models = self.models_manager.models();
+        // Resolve the aux token by the aux model's OWN platform AND endpoint: a
+        // grok (oauth-platform) aux model draws its pooled grok token or `None`,
+        // and an API-key registry platform draws NOTHING — NEVER the primary
+        // Kimi session token (which `resolve_credentials` would otherwise stamp
+        // onto an api.x.ai / api.deepseek.com request). The first-party
+        // subscription channel still gets the primary (byte-identical).
+        // M6: ONE lookup. The platform AND the base URL the rule is applied to
+        // both come from `credential_for_slug`'s single resolution of `slug`
+        // against this catalog, so the platform and the endpoint can no longer
+        // disagree (they were previously resolved by two different lookups).
+        // Aux slugs are not the session's selection, so no `current_key`.
+        let session_key = self
+            .credential_authority()
+            .credential_for_slug(&models, None, slug);
         let endpoints = self.models_manager.endpoints();
         crate::agent::config::resolve_aux_model_sampling_config(
             slug,
             &models,
             &endpoints,
-            session_key.as_deref(),
+            session_key.as_ref(),
             creds.alpha_test_key.clone(),
         )
     }
@@ -484,9 +790,15 @@ impl SessionActor {
     ) -> Option<(kigi_sampler::SamplingClient, String)> {
         let active_session_config = self.reconstruct_full_config().await;
         let mut cfg = self.resolve_aux_sampler_config(slug).await?;
+        // LEAK 1b: the aux classifier must NOT inherit the SESSION model's
+        // (Kimi) bearer_resolver — the resolver is decided by the AUX model's
+        // own platform + endpoint at the chokepoint and passed in explicitly,
+        // so there is no "copy then remember to re-point" step to forget.
+        let aux_resolver = self.aux_bearer_resolver(slug, &cfg.base_url);
         crate::agent::config::stamp_session_local_sampler_fields(
             &mut cfg,
             &active_session_config,
+            aux_resolver,
             Some(self.max_retries),
         );
         let model = cfg.model.clone();
@@ -634,6 +946,7 @@ impl SessionActor {
                     session_id = % self.session_info.id.0, is_session_based = gate
                     .is_session_based, model_byok = gate.model_byok.as_str(),
                     endpoint_is_first_party = gate.endpoint_is_first_party,
+                    credential_class = gate.credential_class.as_str(),
                     "auth recovery: sampler 401 not refreshable (api-key auth) — surfacing 401",
                 );
                 kigi_log::unified_log::warn(
@@ -643,7 +956,8 @@ impl SessionActor {
                         { "kind" : error.kind.as_str(), "status_code" : error
                         .status_code, "is_session_based" : gate.is_session_based,
                         "model_byok" : gate.model_byok.as_str(),
-                        "endpoint_is_first_party" : gate.endpoint_is_first_party, }
+                        "endpoint_is_first_party" : gate.endpoint_is_first_party,
+                        "credential_class" : gate.credential_class.as_str(), }
                     )),
                 );
             }
@@ -659,7 +973,17 @@ impl SessionActor {
                 )),
             );
         }
-        if auth_recovery_eligible && let Some(ref am) = self.auth_manager {
+        // Recover via the ACTIVE model's OWN manager: a grok 401 recovers the
+        // xai-grok session via the xai-grok manager, never the Kimi one. For a
+        // Kimi / non-oauth model this resolves to the primary — byte-identical.
+        if auth_recovery_eligible {
+            let (recovery_model, recovery_base_url) = self
+                .chat_state_handle
+                .get_sampling_config()
+                .await
+                .map(|c| (c.model, c.base_url))
+                .unwrap_or_default();
+            if let Some(am) = self.auth_manager_for_endpoint(&recovery_model, &recovery_base_url) {
                 if am.try_recover_unauthorized().await {
                     tracing::info!(
                         session_id = % self.session_info.id.0,
@@ -683,6 +1007,7 @@ impl SessionActor {
                     None,
                 );
             }
+        }
         if matches!(error.kind, SamplingErrorKind::IdleTimeout) {
             self.signals_handle().record_idle_timeout();
         }
@@ -844,14 +1169,17 @@ impl SessionActor {
     }
     /// Proactively refresh the auth token if near expiry.
     pub(super) async fn refresh_token_if_expired(&self) {
-        if let Some(ref am) = self.auth_manager {
-            let creds = self.chat_state_handle.get_credentials().await;
         let (model_id, base_url) = self
             .chat_state_handle
             .get_sampling_config()
             .await
             .map(|c| (c.model, c.base_url))
             .unwrap_or_default();
+        // Refresh the ACTIVE model's OWN manager: a grok model refreshes the
+        // xai-grok token via the xai-grok manager, never the Kimi one. For a
+        // Kimi / non-oauth model this resolves to the primary — byte-identical.
+        if let Some(am) = self.auth_manager_for_endpoint(&model_id, &base_url) {
+            let creds = self.chat_state_handle.get_credentials().await;
             if self.auth_gate(&model_id, &base_url).active()
                 && let Ok(key) = am.get_valid_token().await
             {
@@ -882,6 +1210,25 @@ impl SessionActor {
             .map(|c| c.model)
             .unwrap_or_default();
         let Some(ref key) = current_key else { return };
+        // M7/M9: a registry-platform model's key normally comes from that
+        // platform's credential resolved into its catalog entry, so with the
+        // session gate now inactive for every API-key platform those turns all
+        // fell through to here and paid a `load_effective_config()` disk read
+        // PER TURN, then logged a permanently false "Model not found in
+        // config.toml [model.*]" warning.
+        //
+        // But a `[model."deepseek/deepseek-chat"]` override DOES keep the base
+        // entry's `info.id` (`ConfigModelOverride::apply`), so "has a platform"
+        // does NOT imply "has no `[model.*]` block" — skipping on the platform
+        // alone would freeze an on-disk key rotation for the whole session.
+        // Skip only when the catalog entry carries no own credential at all,
+        // which is exactly the "key came from the platform, not from config"
+        // case the disk read cannot improve on.
+        if self.model_platform(&current_model_id).is_some()
+            && !self.model_has_own_credential(&current_model_id)
+        {
+            return;
+        }
         let Some(new_key) = self.reload_api_key_from_config(&current_model_id) else {
             return;
         };
@@ -896,6 +1243,15 @@ impl SessionActor {
         creds.api_key = Some(new_key);
         self.chat_state_handle.update_credentials(creds);
     }
+    /// Whether the live catalog entry for `slug` carries its own credential —
+    /// an `api_key`/`env_key` from a `[model.*]` block, which a config edit can
+    /// rotate mid-session. A platform entry whose key came from the platform
+    /// credential has none.
+    fn model_has_own_credential(&self, slug: &str) -> bool {
+        let models = self.models_manager.models();
+        crate::agent::config::find_model_by_id(&models, slug)
+            .is_some_and(crate::agent::config::ModelEntry::has_own_credentials)
+    }
     fn reload_api_key_from_config(&self, current_model_id: &str) -> Option<String> {
         let raw_config = crate::config::load_effective_config()
             .map_err(|e| tracing::warn!(error = % e, "Failed to reload config"))
@@ -973,3 +1329,47 @@ impl SessionActor {
             .push_assistant_response(assistant_item);
     }
 }
+#[cfg(test)]
+mod bearer_resolver_tests {
+    use super::AuthManagerBearerResolver;
+    use kigi_sampler::BearerResolver;
+
+    /// LEAK 1b: the shared `AuthManagerBearerResolver` resolves the LIVE bearer
+    /// of the manager it wraps. So an aux bearer_resolver built over grok's OWN
+    /// (oauth) pooled manager yields grok's token (or `None`) — NEVER the Kimi
+    /// session token that a Kimi-manager resolver would. The
+    /// [`CredentialAuthority::bearer_resolver_for`](crate::auth::credential_authority::CredentialAuthority::bearer_resolver_for)
+    /// wraps exactly this grok manager for a grok aux model.
+    #[tokio::test]
+    async fn resolver_resolves_the_wrapped_manager_never_kimi() {
+        let dir = tempfile::tempdir().unwrap();
+        let kimi = std::sync::Arc::new(crate::auth::AuthManager::new(
+            dir.path(),
+            crate::auth::KimiCodeConfig::default(),
+        ));
+        kimi.hot_swap(crate::auth::KimiAuth {
+            key: "kimi-tok".to_string(),
+            auth_mode: crate::auth::AuthMode::OAuth,
+            ..crate::auth::KimiAuth::test_default()
+        });
+        // The Kimi-manager resolver yields the Kimi bearer.
+        assert_eq!(
+            AuthManagerBearerResolver(kimi.clone()).current_bearer(),
+            Some("kimi-tok".to_string()),
+        );
+        // The grok (oauth) pooled manager is distinct — its resolver never
+        // yields the Kimi bearer (grok's own token, or None).
+        let oauth = kigi_models::PlatformId::XaiGrok
+            .oauth()
+            .expect("xai-grok carries an OAuthConfig");
+        let grok = crate::auth::oauth_registry::global_manager_for(
+            &crate::auth::oauth_registry::pool_home(),
+            oauth,
+        );
+        assert_ne!(
+            AuthManagerBearerResolver(grok).current_bearer(),
+            Some("kimi-tok".to_string()),
+            "a grok aux bearer_resolver must never resolve the Kimi session token",
+        );
+    }
+}

crates/codegen/kigi-shell/src/session/acp_session_impl/spawn.rs

@@ -346,6 +346,7 @@ pub(crate) async fn spawn_session_actor(
         temperature: sampling_config.temperature,
         top_p: sampling_config.top_p,
         api_backend: sampling_config.api_backend.clone(),
+        chat_compat: sampling_config.chat_compat,
         extra_headers: sampling_config.extra_headers.clone(),
         context_window: context_window_override.unwrap_or(baseline_context_window),
         reasoning_effort: sampling_config.reasoning_effort,
@@ -979,10 +980,21 @@ pub(crate) async fn spawn_session_actor(
         }
     };
     let doom_loop_recovery = effective_config.resolve_doom_loop_recovery();
+    let session_model_id_for_actor = session_model_id.clone();
     let session = Arc::new_cyclic(|weak: &std::sync::Weak<SessionActor>| SessionActor {
         session_info: session_info.clone(),
         auth_method_id,
         model_auth_facts: std::cell::RefCell::new(None),
+        // H4: seed the session's OWN selected catalog key from the model it was
+        // spawned with, resolved through the picker's lookup. Never the
+        // process-global `current_model_id()`. H-c: the rule lives in
+        // `selected_catalog_key_for_spawn` so it is covered by a test.
+        selected_catalog_key: std::cell::RefCell::new(
+            crate::agent::models::selected_catalog_key_for_spawn(
+                &models_manager.models(),
+                &session_model_id_for_actor,
+            ),
+        ),
         attribution_callback,
         auth_manager,
         state,

crates/codegen/kigi-shell/src/session/acp_session_tests/auth_error_no_retry_tests.rs

@@ -485,23 +485,44 @@ async fn no_legacy_hint_for_oidc_auth() {
 #[test]
 fn session_token_auth_gate_truth_table() {
     use crate::agent::auth_method::{ModelByok, session_token_auth_gate as gate};
+    use crate::auth::credential_authority::CredentialClass;
     // Non-session methods never refresh, regardless of BYOK status or endpoint.
+    // `Pooled` (an OAuth platform's own pool) and `Primary` (kimi-code, or a
+    // bare / [model.*] model on the session's own endpoint) behave identically
+    // here: each names a credential that IS refreshable on that host.
     for fp in [false, true] {
-        assert!(!gate(false, ModelByok::NotByok, fp));
-        assert!(!gate(false, ModelByok::Byok, fp));
-        assert!(!gate(false, ModelByok::Unknown, fp));
-        // Session method: a definite classification ignores the endpoint —
-        // NotByok always refreshes (only ever routes to the session endpoint),
-        // a genuine per-model Byok never does.
-        assert!(gate(true, ModelByok::NotByok, fp));
-        assert!(!gate(true, ModelByok::Byok, fp));
+        for class in [CredentialClass::Pooled, CredentialClass::Primary] {
+            assert!(!gate(false, ModelByok::NotByok, fp, class));
+            assert!(!gate(false, ModelByok::Byok, fp, class));
+            assert!(!gate(false, ModelByok::Unknown, fp, class));
+            // Session method on an endpoint that DOES take a session
+            // credential: a definite classification ignores the endpoint —
+            // NotByok refreshes, a genuine per-model Byok never does.
+            assert!(gate(true, ModelByok::NotByok, fp, class));
+            assert!(!gate(true, ModelByok::Byok, fp, class));
+        }
+        // …and an API-key registry platform endpoint (`CredentialClass::None`)
+        // is refused on every arm, first-party flag included: the leak guard.
+        assert!(!gate(true, ModelByok::NotByok, fp, CredentialClass::None));
+        assert!(!gate(true, ModelByok::Byok, fp, CredentialClass::None));
+        assert!(!gate(true, ModelByok::Unknown, fp, CredentialClass::None));
     }
     // Session method + Unknown BYOK: refresh only against a first-party xAI
     // host, so a transiently-unclassifiable config can't demote a live session
     // (the stale-token 401 regression) yet the session token never leaks to a
     // third-party BYOK endpoint. This arm was unconditionally `false` pre-fix.
-    assert!(gate(true, ModelByok::Unknown, true));
-    assert!(!gate(true, ModelByok::Unknown, false));
+    assert!(gate(
+        true,
+        ModelByok::Unknown,
+        true,
+        CredentialClass::Primary
+    ));
+    assert!(!gate(
+        true,
+        ModelByok::Unknown,
+        false,
+        CredentialClass::Primary
+    ));
 }
 
 /// Pre-fix, the gate read `auth_type` and skipped recovery here, 401'ing every
@@ -847,7 +868,11 @@ async fn set_session_model_invalidates_byok_memo_for_same_model_id() {
                 temperature: None,
                 top_p: None,
                 api_backend: crate::sampling::ApiBackend::ChatCompletions,
+                chat_compat: Default::default(),
                 auth_scheme: Default::default(),
+                anthropic_oauth: false,
+                github_copilot: false,
+                openai_codex: false,
                 extra_headers: Default::default(),
                 context_window: 256_000,
                 force_http1: false,
@@ -865,7 +890,7 @@ async fn set_session_model_invalidates_byok_memo_for_same_model_id() {
                 header_injector: None,
             };
             let _ = actor
-                .handle_set_session_model(cfg, false, false, true, 85)
+                .handle_set_session_model(cfg, None, false, false, true, 85)
                 .await;
 
             assert!(

crates/codegen/kigi-shell/src/session/acp_session_tests/cancel_running_task_tests.rs

@@ -44,7 +44,11 @@ async fn persist_ack_waits_for_disk_flush_before_success() {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 auth_scheme: Default::default(),
+                anthropic_oauth: false,
+                github_copilot: false,
+                openai_codex: false,
                 extra_headers: Default::default(),
                 context_window: 100_000,
                 force_http1: false,
@@ -87,6 +91,7 @@ async fn persist_ack_waits_for_disk_flush_before_success() {
                     temperature: None,
                     top_p: None,
                     api_backend: Default::default(),
+                    chat_compat: Default::default(),
                     extra_headers: Default::default(),
                     context_window: std::num::NonZeroU64::new(100_000).unwrap(),
                     reasoning_effort: None,
@@ -104,6 +109,7 @@ async fn persist_ack_waits_for_disk_flush_before_success() {
                 session_info,
                 auth_method_id: test_auth_method_id("test-auth"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 attribution_callback: None,
                 auth_manager: None,
                 state: TokioMutex::new(State {
@@ -338,7 +344,11 @@ async fn first_turn_memory_injection_persists_to_chat_history() {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 auth_scheme: Default::default(),
+                anthropic_oauth: false,
+                github_copilot: false,
+                openai_codex: false,
                 context_window: 100_000,
                 force_http1: false,
                 max_retries: None,
@@ -381,6 +391,7 @@ async fn first_turn_memory_injection_persists_to_chat_history() {
                     temperature: None,
                     top_p: None,
                     api_backend: Default::default(),
+                    chat_compat: Default::default(),
                     extra_headers: Default::default(),
                     context_window: std::num::NonZeroU64::new(100_000).unwrap(),
                     reasoning_effort: None,
@@ -466,7 +477,11 @@ async fn first_turn_memory_injection_disabled_does_not_persist_to_chat_history()
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 auth_scheme: Default::default(),
+                anthropic_oauth: false,
+                github_copilot: false,
+                openai_codex: false,
                 context_window: 100_000,
                 force_http1: false,
                 max_retries: None,
@@ -513,6 +528,7 @@ async fn first_turn_memory_injection_disabled_does_not_persist_to_chat_history()
                     temperature: None,
                     top_p: None,
                     api_backend: Default::default(),
+                    chat_compat: Default::default(),
                     extra_headers: Default::default(),
                     context_window: std::num::NonZeroU64::new(100_000).unwrap(),
                     reasoning_effort: None,
@@ -547,6 +563,7 @@ async fn first_turn_memory_injection_disabled_does_not_persist_to_chat_history()
                 session_info: session_info.clone(),
                 auth_method_id: test_auth_method_id("test-auth"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 attribution_callback: None,
                 auth_manager: None,
                 state: TokioMutex::new(State {
@@ -810,6 +827,7 @@ async fn cancel_running_task_teardown_clears_running_and_pending_work() {
                 },
                 auth_method_id: test_auth_method_id("test-auth"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 attribution_callback: None,
                 auth_manager: None,
                 state,
@@ -1732,7 +1750,11 @@ async fn cancel_propagates_to_sampler_handle_so_no_further_emission() {
                 temperature: None,
                 top_p: None,
                 api_backend: kigi_sampler::ApiBackend::Responses,
+                chat_compat: Default::default(),
                 auth_scheme: Default::default(),
+                anthropic_oauth: false,
+                github_copilot: false,
+                openai_codex: false,
                 extra_headers: Default::default(),
                 context_window: 100_000,
                 force_http1: false,
@@ -1798,6 +1820,7 @@ async fn cancel_propagates_to_sampler_handle_so_no_further_emission() {
                 },
                 auth_method_id: test_auth_method_id("test-auth"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 attribution_callback: None,
                 auth_manager: None,
                 state,

crates/codegen/kigi-shell/src/session/acp_session_tests/idle_resume_tests.rs

@@ -104,6 +104,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                     temperature: None,
                     top_p: None,
                     api_backend: Default::default(),
+                    chat_compat: Default::default(),
                     extra_headers: Default::default(),
                     context_window: std::num::NonZeroU64::new(200_000).unwrap(),
                     reasoning_effort: None,
@@ -127,6 +128,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                 attribution_callback: None,
                 auth_method_id: test_auth_method_id("cached_token"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 auth_manager: {
                     let dir = tempfile::tempdir().unwrap();
                     let mgr = std::sync::Arc::new(crate::auth::AuthManager::new(

crates/codegen/kigi-shell/src/session/acp_session_tests/inline_auto_compact_flow_tests.rs

@@ -52,6 +52,7 @@ async fn create_test_actor(
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             extra_headers: Default::default(),
             context_window: std::num::NonZeroU64::new(context_window)
                 .expect("test context_window must be non-zero"),
@@ -71,6 +72,7 @@ async fn create_test_actor(
         rebuild_spec: crate::session::agent_rebuild::test_rebuild_spec_default(),
         auth_method_id: test_auth_method_id("test-auth"),
         model_auth_facts: std::cell::RefCell::new(None),
+        selected_catalog_key: std::cell::RefCell::new(None),
         attribution_callback: None,
         auth_manager: None,
         state,
@@ -486,6 +488,7 @@ async fn create_test_actor_with_memory(
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             extra_headers: Default::default(),
             context_window: std::num::NonZeroU64::new(context_window)
                 .expect("test context_window must be non-zero"),
@@ -509,6 +512,7 @@ async fn create_test_actor_with_memory(
         rebuild_spec: crate::session::agent_rebuild::test_rebuild_spec_default(),
         auth_method_id: test_auth_method_id("test-auth"),
         model_auth_facts: std::cell::RefCell::new(None),
+        selected_catalog_key: std::cell::RefCell::new(None),
         attribution_callback: None,
         auth_manager: None,
         state,
@@ -1240,6 +1244,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                     temperature: None,
                     top_p: None,
                     api_backend: Default::default(),
+                    chat_compat: Default::default(),
                     extra_headers: Default::default(),
                     context_window: std::num::NonZeroU64::new(200_000).unwrap(),
                     reasoning_effort: None,
@@ -1263,6 +1268,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                 rebuild_spec: crate::session::agent_rebuild::test_rebuild_spec_default(),
                 auth_method_id: test_auth_method_id("cached_token"),
                 model_auth_facts: std::cell::RefCell::new(None),
+                selected_catalog_key: std::cell::RefCell::new(None),
                 auth_manager: {
                     let dir = tempfile::tempdir().unwrap();
                     let mgr = std::sync::Arc::new(crate::auth::AuthManager::new(

crates/codegen/kigi-shell/src/session/acp_session_tests/memory_config_tests.rs

@@ -104,6 +104,7 @@ async fn create_test_actor_with_memory(
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             extra_headers: Default::default(),
             context_window: std::num::NonZeroU64::new(context_window)
                 .expect("test context_window must be non-zero"),
@@ -126,6 +127,7 @@ async fn create_test_actor_with_memory(
         },
         auth_method_id: test_auth_method_id("test-auth"),
         model_auth_facts: std::cell::RefCell::new(None),
+        selected_catalog_key: std::cell::RefCell::new(None),
         attribution_callback: None,
         auth_manager: None,
         state,

crates/codegen/kigi-shell/src/session/acp_session_tests/replay_buffer_send_update_tests.rs

@@ -78,6 +78,7 @@ pub(super) async fn make_replay_send_update_fixture() -> ReplaySendUpdateFixture
         },
         auth_method_id: test_auth_method_id("test-auth"),
         model_auth_facts: std::cell::RefCell::new(None),
+        selected_catalog_key: std::cell::RefCell::new(None),
         attribution_callback: None,
         auth_manager: None,
         state,

crates/codegen/kigi-shell/src/session/acp_session_tests/session_bearer_leak_platform_tests.rs

@@ -0,0 +1,990 @@
+//! LEAK GUARD, part 2: the model→platform lookup (H5) and the AUX resolver
+//! decision (H3/H4). Shares the fixtures in
+//! [`super::session_bearer_leak_tests`]; see that module's header for the chain
+//! and the storage-discipline contract.
+
+use super::session_bearer_leak_tests::{
+    KIMI_TOKEN, actor_on_managed_model, actor_with_catalog, managed_entry,
+};
+use super::*;
+use kigi_sampler::BearerResolver;
+use kigi_test_support::EnvGuard;
+use std::sync::Arc;
+
+/// The host BOTH halves of the `anthropic` / `claude-pro-max` collision route
+/// to, derived from the registry (as the sibling at
+/// [`oauth_platform_models_keep_a_live_resolver_from_their_own_pool`] does) so
+/// the fixture cannot drift, with the twin agreement asserted rather than
+/// assumed — the collision is only a collision because both platforms serve the
+/// same host.
+fn anthropic_collision_host() -> String {
+    let oauth_host = kigi_models::PlatformId::ClaudeProMax.base_url();
+    assert_eq!(
+        kigi_models::PlatformId::Anthropic.base_url(),
+        oauth_host,
+        "the API-key platform and its subscription-OAuth twin must serve the same host, \
+         or this fixture is not testing the dual-credential collision"
+    );
+    oauth_host
+}
+
+/// Ambient BYOK env unset. `resolve_model_auth_facts` probes `std::env::var` at
+/// call time, so a developer (or CI) holding `ANTHROPIC_API_KEY` flips the
+/// fixture to `Byok` and switches off the session-token gate for a reason that
+/// has nothing to do with the platform lookup under test. Every holder must be
+/// `#[serial]`.
+fn anthropic_collision_env_guard() -> [EnvGuard; 2] {
+    [
+        EnvGuard::unset("ANTHROPIC_API_KEY"),
+        EnvGuard::unset("KIGI_CODE_BASE_URL"),
+    ]
+}
+
+#[tokio::test(flavor = "current_thread")]
+async fn dual_credential_slug_collision_resolves_the_selected_oauth_platform() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            // (api-key twin, oauth twin, shared slug, host)
+            for (api_key_twin, oauth_twin, slug, base_url) in [
+                (
+                    "xai/grok-4.5",
+                    "xai-grok/grok-4.5",
+                    "grok-4.5",
+                    "https://api.x.ai/v1",
+                ),
+                (
+                    "anthropic/claude-opus-4-8",
+                    "claude-pro-max/claude-opus-4-8",
+                    "claude-opus-4-8",
+                    "https://api.anthropic.com/v1",
+                ),
+                (
+                    "openai/gpt-5.5-codex",
+                    "openai-codex/gpt-5.5-codex",
+                    "gpt-5.5-codex",
+                    "https://chatgpt.com/backend-api/codex",
+                ),
+            ] {
+                // API-key platform FIRST, exactly as `PlatformId::ALL` orders them.
+                let catalog = vec![
+                    managed_entry(api_key_twin, slug, base_url),
+                    managed_entry(oauth_twin, slug, base_url),
+                ];
+                let (_dir, actor, _rx) = actor_with_catalog(catalog, oauth_twin, "unused").await;
+                let cfg = actor.reconstruct_full_config().await;
+
+                let resolver = cfg.bearer_resolver.as_ref().unwrap_or_else(|| {
+                    panic!(
+                        "{oauth_twin}: selecting the OAuth twin must keep a LIVE bearer_resolver \
+                         (mid-session refresh); resolving {api_key_twin} instead drops it"
+                    )
+                });
+                assert_ne!(
+                    resolver.current_bearer(),
+                    Some(KIMI_TOKEN.to_string()),
+                    "{oauth_twin}: the resolver must read its OWN pool, never the Kimi primary"
+                );
+
+                let platform = kigi_models::parse_managed_model_key(oauth_twin)
+                    .expect("managed key")
+                    .0;
+                assert_eq!(
+                    cfg.anthropic_oauth,
+                    platform.wire_api() == kigi_models::PlatformWireApi::Messages,
+                    "{oauth_twin}: the Claude OAuth Messages adaptation must follow the \
+                     SELECTED platform"
+                );
+                assert_eq!(
+                    cfg.openai_codex,
+                    platform.sends_codex_responses_headers(),
+                    "{oauth_twin}: the Codex identity headers must follow the SELECTED platform"
+                );
+                assert_eq!(
+                    cfg.github_copilot,
+                    platform.sends_copilot_editor_headers(),
+                    "{oauth_twin}: the Copilot editor headers must follow the SELECTED platform"
+                );
+            }
+        })
+        .await;
+}
+
+/// The other half of H5: selecting the API-KEY twin of a colliding slug must
+/// still resolve the API-key platform — no bearer_resolver, no adaptations. The
+/// unified lookup must not simply prefer OAuth.
+#[tokio::test(flavor = "current_thread")]
+async fn dual_credential_slug_collision_resolves_the_selected_api_key_platform() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let catalog = vec![
+                managed_entry(
+                    "anthropic/claude-opus-4-8",
+                    "claude-opus-4-8",
+                    "https://api.anthropic.com/v1",
+                ),
+                managed_entry(
+                    "claude-pro-max/claude-opus-4-8",
+                    "claude-opus-4-8",
+                    "https://api.anthropic.com/v1",
+                ),
+            ];
+            let (_dir, actor, _rx) =
+                actor_with_catalog(catalog, "anthropic/claude-opus-4-8", "sk-ant-byok").await;
+            let cfg = actor.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_none(),
+                "selecting the API-key twin must get NO session bearer resolver"
+            );
+            assert!(
+                !cfg.anthropic_oauth,
+                "the API-key Anthropic Messages request must stay byte-identical"
+            );
+            assert_eq!(cfg.api_key.as_deref(), Some("sk-ant-byok"));
+        })
+        .await;
+}
+
+/// MANDATORY counterpart: the subscription-OAuth platforms have NON-first-party
+/// base URLs, so the fix must not disable their resolver. Each must still get a
+/// LIVE `bearer_resolver` — and it must read THAT platform's own pooled
+/// `AuthManager`, never the Kimi primary. The pooled managers are empty here (a
+/// TempDir pool home), which is what makes `current_bearer() == None` a proof
+/// that the Kimi bearer cannot be what they resolve.
+#[tokio::test(flavor = "current_thread")]
+async fn oauth_platform_models_keep_a_live_resolver_from_their_own_pool() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            // The base URL is the platform's OWN registry host, exactly as
+            // `models_fetch::platform_fetch_base` builds every fetched entry —
+            // derived here rather than hard-coded so the fixture cannot drift
+            // from the registry (L10 compares against precisely this).
+            for (catalog_key, slug) in [
+                ("claude-pro-max/claude-opus-4-8", "claude-opus-4-8"),
+                ("github-copilot/gpt-4.1", "gpt-4.1"),
+                ("xai-grok/grok-4-latest", "grok-4-latest"),
+                ("openai-codex/gpt-5.5", "gpt-5.5"),
+            ] {
+                let base_url = kigi_models::parse_managed_model_key(catalog_key)
+                    .expect("managed key")
+                    .0
+                    .base_url();
+                let base_url = base_url.as_str();
+                let (_dir, actor, _rx) =
+                    actor_on_managed_model(catalog_key, slug, base_url, "unused").await;
+                let cfg = actor.reconstruct_full_config().await;
+                let resolver = cfg.bearer_resolver.as_ref().unwrap_or_else(|| {
+                    panic!("{catalog_key}: must keep a live bearer_resolver for refresh")
+                });
+                assert_ne!(
+                    resolver.current_bearer(),
+                    Some(KIMI_TOKEN.to_string()),
+                    "{catalog_key}: the Kimi bearer must never be what it resolves"
+                );
+
+                // The resolver is LIVE over that platform's pooled manager: a
+                // token rotated inside the pool is observed by the
+                // already-built resolver (this is what mid-session refresh
+                // does). The pool is read here, never mutated.
+                let pooled = actor
+                    .credential_authority()
+                    .manager_for(
+                        kigi_models::parse_managed_model_key(catalog_key).map(|(p, _)| p),
+                        base_url,
+                    )
+                    .expect("an OAuth platform on its own host always resolves a manager");
+                assert!(
+                    !Arc::ptr_eq(
+                        &pooled,
+                        actor.auth_manager.as_ref().expect("primary is present")
+                    ),
+                    "{catalog_key}: must route to its OWN pooled manager, not the Kimi primary"
+                );
+                assert_eq!(
+                    resolver.current_bearer(),
+                    pooled.current_or_expired().map(|a| a.key),
+                    "{catalog_key}: the resolver must read THIS platform's pooled manager"
+                );
+            }
+        })
+        .await;
+}
+
+/// H3/H4 — the stamped AUX paths (image-describe, the auto-mode classifier and
+/// the session-summary client all funnel through
+/// `CredentialAuthority::bearer_resolver_for`). `SamplingClient::post` REPLACES
+/// the request's auth header from the resolver, so an API-key-platform aux model
+/// would have its own key overwritten by the Kimi bearer ON THE AUX HOST.
+///
+/// Revert-to-red: make `CredentialAuthority::governing_manager`'s
+/// `Some(platform) => None` arm return `self.primary.clone()` and the deepseek /
+/// openai / moonshot rows resolve `KIMI_TOKEN`.
+#[test]
+fn aux_bearer_resolver_clears_the_session_resolver_off_a_third_party_aux_host() {
+    let dir = tempfile::tempdir().expect("tempdir");
+    let primary = Arc::new(crate::auth::AuthManager::new(
+        dir.path(),
+        crate::auth::KimiCodeConfig::default(),
+    ));
+    primary.hot_swap(crate::auth::KimiAuth {
+        key: KIMI_TOKEN.to_string(),
+        auth_mode: crate::auth::AuthMode::OAuth,
+        ..crate::auth::KimiAuth::test_default()
+    });
+    let authority = crate::auth::credential_authority::CredentialAuthority::new(
+        crate::agent::config::EndpointsConfig::default(),
+        Some(primary),
+    );
+    let platform =
+        |key: &str| kigi_models::parse_managed_model_key(key).map(|(platform, _)| platform);
+
+    // Cleared: every API-key registry platform, and a `[model.*]` aux model
+    // pointed at a third-party host.
+    for (key, base_url) in [
+        ("deepseek/deepseek-chat", "https://api.deepseek.com/v1"),
+        ("openai/gpt-5-mini", "https://api.openai.com/v1"),
+        (
+            "moonshot-cn/kimi-k2-turbo-preview",
+            "https://api.moonshot.cn/v1",
+        ),
+    ] {
+        assert!(
+            authority
+                .bearer_resolver_for(platform(key), base_url)
+                .is_none(),
+            "LEAK: an aux model on {base_url} must not inherit the session bearer resolver"
+        );
+    }
+    assert!(
+        authority
+            .bearer_resolver_for(None, "https://api.openai.com/v1")
+            .is_none(),
+        "LEAK: a [model.*] aux model on a third-party host must not inherit it either"
+    );
+
+    // Kept (byte-identical): the first-party subscription channel and a
+    // platform-less aux model on the session's own endpoint.
+    for (key, base_url) in [
+        (
+            Some("kimi-code/kimi-for-coding"),
+            kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+        ),
+        (None, kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url),
+        (None, "http://127.0.0.1:4141/v1"),
+    ] {
+        let resolved = authority
+            .bearer_resolver_for(key.and_then(platform), base_url)
+            .unwrap_or_else(|| panic!("{key:?} @ {base_url} must keep the session resolver"));
+        assert_eq!(resolved.current_bearer(), Some(KIMI_TOKEN.to_string()));
+    }
+
+    // Re-pointed: an OAuth aux model on ITS OWN host gets a LIVE resolver over
+    // its own pool (empty here), never the Kimi primary. L10: the same model
+    // redirected to a third-party host gets NOTHING.
+    let rt = tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .expect("runtime for the pooled manager's refresh task");
+    rt.block_on(async {
+        for key in [
+            "xai-grok/grok-4-latest",
+            "claude-pro-max/claude-opus-4-8",
+            "github-copilot/gpt-4.1",
+            "openai-codex/gpt-5.5",
+        ] {
+            let p = platform(key).expect("managed key");
+            let resolved = authority
+                .bearer_resolver_for(Some(p), &p.base_url())
+                .unwrap_or_else(|| panic!("{key} must keep a live resolver from its own pool"));
+            assert_ne!(
+                resolved.current_bearer(),
+                Some(KIMI_TOKEN.to_string()),
+                "{key}: the aux resolver must never resolve the Kimi session bearer"
+            );
+            assert!(
+                authority
+                    .bearer_resolver_for(Some(p), "https://example.invalid/v1")
+                    .is_none(),
+                "LEAK ({key}): an OAuth aux model redirected off its own host gets nothing"
+            );
+        }
+    });
+}
+
+/// H4 — TWO CONCURRENT SESSIONS on a colliding slug. The model→platform lookup
+/// used to key on `ModelsManager::current_model_id()`, a single PROCESS-GLOBAL
+/// `RwLock<acp::ModelId>` written by whichever session switched last. With one
+/// session on `xai-grok/grok-4.5` and another on `xai/grok-4.5` — same routing
+/// slug, by design — the loser resolved the OTHER session's platform:
+/// the subscription session lost its live resolver (unrecoverable 401 ~1h in)
+/// and the API-key session got the pooled OAuth bearer stamped over its own
+/// `sk-…` key, which the provider rejects.
+///
+/// Here the global cell is deliberately set to the API-key twin for BOTH
+/// sessions (last writer wins, and it was the API-key one). Each session must
+/// still resolve ITS OWN selection.
+///
+/// Revert-to-red: replace `self.selected_catalog_key()` in
+/// `SessionActor::model_platform` with
+/// `Some(self.models_manager.current_model_id().0.as_ref())`. Under THIS
+/// fixture both of the subscription session's assertions fail. (L: the fixture
+/// is what makes that true — `managed_entry` carries `api_key: None` and
+/// `actor_with_catalog` pins `NotByok`, which is exactly what a FETCHED registry
+/// entry resolves to. A user who additionally sets `ANTHROPIC_API_KEY` /
+/// `[model.*] env_key` classifies `Byok`, the gate is inactive for that reason
+/// alone, and only the `anthropic_oauth` assertion would still catch the
+/// mis-resolution — hence the env guard below.)
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn concurrent_sessions_on_a_colliding_slug_each_resolve_their_own_platform() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let api_key_twin = "anthropic/claude-opus-4-8";
+            let oauth_twin = "claude-pro-max/claude-opus-4-8";
+            let slug = "claude-opus-4-8";
+            let host = &anthropic_collision_host();
+            let catalog = || {
+                vec![
+                    // API-key platform FIRST, exactly as `PlatformId::ALL` orders them.
+                    managed_entry(api_key_twin, slug, host),
+                    managed_entry(oauth_twin, slug, host),
+                ]
+            };
+
+            let (_d1, subscription, _r1) =
+                actor_with_catalog(catalog(), oauth_twin, "unused").await;
+            let (_d2, api_key, _r2) =
+                actor_with_catalog(catalog(), api_key_twin, "sk-ant-user").await;
+            // The other session switched last: the process-global cell now names
+            // the API-key twin for BOTH.
+            for actor in [&subscription, &api_key] {
+                actor
+                    .models_manager
+                    .set_current_model_id(acp::ModelId::new(api_key_twin.to_string()));
+            }
+
+            let sub_cfg = subscription.reconstruct_full_config().await;
+            let resolver = sub_cfg.bearer_resolver.as_ref().expect(
+                "the subscription session must keep its own live bearer_resolver even when \
+                 another session switched the process-global model last",
+            );
+            assert_ne!(
+                resolver.current_bearer(),
+                Some(KIMI_TOKEN.to_string()),
+                "it must read the claude-pro-max pool, never the Kimi primary"
+            );
+            assert!(
+                sub_cfg.anthropic_oauth,
+                "the Claude OAuth Messages adaptation must follow the SUBSCRIPTION session"
+            );
+
+            let api_cfg = api_key.reconstruct_full_config().await;
+            assert!(
+                api_cfg.bearer_resolver.is_none(),
+                "LEAK: the API-key session must get no session bearer_resolver"
+            );
+            assert!(
+                !api_cfg.anthropic_oauth,
+                "the API-key session must not get the OAuth Messages adaptation"
+            );
+            assert_eq!(
+                api_cfg.api_key.as_deref(),
+                Some("sk-ant-user"),
+                "the API-key session keeps its own provider key"
+            );
+        })
+        .await;
+}
+
+/// H4 in LEADER mode, where `agent/handlers/model_switch.rs` skips
+/// `set_current_model_id` ENTIRELY, so the process-global cell is frozen at the
+/// startup default for the whole process lifetime. `platform_for_slug` then
+/// fell through to the `.rev()` scan, which returns the LAST match — the OAuth
+/// twin — so a Leader-mode session on the API-KEY twin was handed the pooled
+/// OAuth bearer plus the Messages adaptation, and Anthropic rejects both.
+///
+/// Revert-to-red: same edit as above; with the global cell naming the startup
+/// default (not in this catalog) the `.rev()` fallback resolves
+/// `claude-pro-max/*` and both assertions fail.
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn leader_mode_session_resolves_its_own_platform_without_the_global_cell() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let slug = "claude-opus-4-8";
+            let host = &anthropic_collision_host();
+            let (_dir, actor, _rx) = actor_with_catalog(
+                vec![
+                    managed_entry("anthropic/claude-opus-4-8", slug, host),
+                    managed_entry("claude-pro-max/claude-opus-4-8", slug, host),
+                ],
+                "anthropic/claude-opus-4-8",
+                "sk-ant-user",
+            )
+            .await;
+            // Leader mode never writes the global cell: it still names the
+            // startup default, which is not in this catalog at all.
+            assert!(
+                !actor
+                    .models_manager
+                    .models()
+                    .contains_key(actor.models_manager.current_model_id().0.as_ref()),
+                "precondition: the process-global model id is stale (Leader mode)"
+            );
+
+            let cfg = actor.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_none(),
+                "LEAK: a Leader-mode API-key session must get no session bearer_resolver"
+            );
+            assert!(
+                !cfg.anthropic_oauth,
+                "a Leader-mode API-key session must not get the OAuth Messages adaptation"
+            );
+        })
+        .await;
+}
+
+/// L13 — a user whose ACP auth method is an API-KEY registry platform (e.g.
+/// `deepseek`) can still SELECT a subscription-OAuth model, and the chokepoint
+/// hands it that platform's pooled bearer as the request's `api_key`. The gate
+/// keys on the primary method, which is not session-based, so the config used
+/// to carry NO `bearer_resolver`: the pooled token froze at selection time and
+/// the session died with an unrecoverable 401 once it expired (~1h).
+///
+/// The model's own credential now makes the gate session-based, confined to
+/// that platform's own host by the gate's `credential_class` conjunct.
+#[tokio::test(flavor = "current_thread")]
+async fn oauth_model_under_an_api_key_auth_method_keeps_its_pooled_resolver() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let base_url = kigi_models::PlatformId::ClaudeProMax.base_url();
+            let (_dir, actor, _rx) = actor_with_catalog(
+                vec![managed_entry(
+                    "claude-pro-max/claude-opus-4-8",
+                    "claude-opus-4-8",
+                    &base_url,
+                )],
+                "claude-pro-max/claude-opus-4-8",
+                "unused",
+            )
+            .await;
+            // The PRIMARY ACP method is an API-key registry platform login.
+            actor
+                .auth_method_id
+                .store(Some(std::sync::Arc::new(acp::AuthMethodId::new(
+                    "deepseek",
+                ))));
+
+            let cfg = actor.reconstruct_full_config().await;
+            let resolver = cfg.bearer_resolver.as_ref().expect(
+                "a subscription-OAuth model keeps a live resolver whatever the primary \
+                 ACP auth method is, or it cannot refresh mid-session",
+            );
+            assert_ne!(
+                resolver.current_bearer(),
+                Some(KIMI_TOKEN.to_string()),
+                "and it reads the claude-pro-max pool, never the primary"
+            );
+
+            // An API-key-platform model under the same method stays resolver-free.
+            let (_d2, deepseek, _r2) = actor_with_catalog(
+                vec![managed_entry(
+                    "deepseek/deepseek-chat",
+                    "deepseek-chat",
+                    "https://api.deepseek.com/v1",
+                )],
+                "deepseek/deepseek-chat",
+                "sk-deepseek",
+            )
+            .await;
+            deepseek
+                .auth_method_id
+                .store(Some(std::sync::Arc::new(acp::AuthMethodId::new(
+                    "deepseek",
+                ))));
+            assert!(
+                deepseek
+                    .reconstruct_full_config()
+                    .await
+                    .bearer_resolver
+                    .is_none(),
+                "LEAK: an API-key-platform model must never get a session resolver"
+            );
+        })
+        .await;
+}
+
+/// H-b — a `None` or STALE per-session catalog key must REFUSE, not degrade to
+/// the subscription-OAuth twin.
+///
+/// `model_platform` falls through to `resolve_catalog_key`'s `.rev()` scan when
+/// the session's own key does not name the slug, and that scan returns the LAST
+/// match — the OAuth twin, because `PlatformId::ALL` orders every API-key
+/// platform first. Combined with the L13 disjunct (a model whose own credential
+/// is a pooled OAuth session is session-based BY ITSELF), an API-KEY session on
+/// `anthropic/claude-opus-4-8` with no per-session key got
+/// `is_session_based = true`, `credential_class = Pooled` (same
+/// host) and, at `NotByok`, an ACTIVE gate — so `manager_for` handed it the
+/// Claude POOLED manager, whose `bearer_resolver` REPLACES the user's own
+/// `sk-ant-…` on the wire, plus the OAuth Messages adaptation. Anthropic rejects
+/// both. This is exactly what H4 prevents, reached through the `None` path.
+///
+/// A key that is absent or names a different model is not evidence for either
+/// twin: resolve to NO platform, which the chokepoint then decides purely by the
+/// ENDPOINT (the OAuth host is not this session's coding endpoint ⇒ nothing
+/// rides).
+///
+/// Revert-to-red (production, compiles): delete the
+/// `platform.oauth().is_some() && !disambiguated && slug_collides_across_platforms(..)`
+/// refusal from `crate::agent::models::platform_for_slug` and every
+/// `bearer_resolver` / `anthropic_oauth` assertion below fails.
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn a_missing_or_stale_session_key_refuses_instead_of_guessing_the_oauth_twin() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let api_key_twin = "anthropic/claude-opus-4-8";
+            let slug = "claude-opus-4-8";
+            let host = anthropic_collision_host();
+            let catalog = || {
+                vec![
+                    // API-key platform FIRST, exactly as `PlatformId::ALL` orders them.
+                    managed_entry(api_key_twin, slug, &host),
+                    managed_entry("claude-pro-max/claude-opus-4-8", slug, &host),
+                ]
+            };
+
+            for (case, stale_key) in [
+                // No key at all: a session spawned on a model that left the
+                // catalog, or one an older build never seeded.
+                ("absent", None),
+                // Stale: an `OverrideModelName` rename, or a key naming a model
+                // this session is no longer on.
+                ("stale", Some("claude-pro-max/some-other-model".to_string())),
+            ] {
+                let (_dir, actor, _rx) =
+                    actor_with_catalog(catalog(), api_key_twin, "sk-ant-user").await;
+                *actor.selected_catalog_key.borrow_mut() = stale_key;
+
+                let cfg = actor.reconstruct_full_config().await;
+                assert!(
+                    cfg.bearer_resolver.is_none(),
+                    "{case}: LEAK — an unresolvable selection must get NO session bearer \
+                     resolver; the pooled OAuth bearer would REPLACE the user's own key"
+                );
+                assert!(
+                    !cfg.anthropic_oauth,
+                    "{case}: nor the Claude OAuth Messages adaptation"
+                );
+                assert_eq!(
+                    cfg.api_key.as_deref(),
+                    Some("sk-ant-user"),
+                    "{case}: the user's own provider key must survive untouched"
+                );
+                assert!(
+                    actor
+                        .credential_authority()
+                        .manager_for(
+                            crate::agent::models::platform_for_slug(
+                                &actor.models_manager.models(),
+                                actor.selected_catalog_key().as_deref(),
+                                slug,
+                            ),
+                            &host,
+                        )
+                        .is_none(),
+                    "{case}: and no manager either — refuse, never guess"
+                );
+            }
+
+            // …while a session that DID select the OAuth twin still gets its
+            // pooled resolver: the refusal is about the guess, not the platform.
+            let (_dir, selected, _rx) =
+                actor_with_catalog(catalog(), "claude-pro-max/claude-opus-4-8", "unused").await;
+            let cfg = selected.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_some() && cfg.anthropic_oauth,
+                "a DELIBERATE subscription selection keeps its pooled resolver and \
+                 adaptation (this is what makes the refusals above meaningful)"
+            );
+        })
+        .await;
+}
+
+/// H-c — coverage for the FIRST of the two production writers of
+/// `selected_catalog_key`: the spawn seed
+/// (`crate::agent::models::selected_catalog_key_for_spawn`, called from
+/// `spawn.rs`). Every other test in this module sets the field by hand, so a
+/// wrong seed was silent.
+///
+/// Both spawn shapes are covered: a FRESH session, spawned on the catalog key
+/// the picker resolved, and a RESUME/LOAD, which spawns with the RAW persisted
+/// `summary.current_model_id` — a BARE routing slug after any `SetSessionModel`,
+/// since `handle_set_session_model` persists `sampling_config.model`. This seed
+/// is where that slug becomes a key. The assertion is end-to-end: the seeded key
+/// is fed to the very function the auth layer keys on.
+#[test]
+fn spawn_seeds_the_session_key_the_auth_layer_keys_on() {
+    let slug = "claude-opus-4-8";
+    let host = anthropic_collision_host();
+    let api_key_twin = "anthropic/claude-opus-4-8";
+    let oauth_twin = "claude-pro-max/claude-opus-4-8";
+    let models: indexmap::IndexMap<String, crate::agent::config::ModelEntry> = [
+        managed_entry(api_key_twin, slug, &host),
+        managed_entry(oauth_twin, slug, &host),
+    ]
+    .into_iter()
+    .collect();
+
+    // FRESH: spawned on the catalog key. Idempotent, and it disambiguates.
+    for selected in [api_key_twin, oauth_twin] {
+        let seeded = crate::agent::models::selected_catalog_key_for_spawn(
+            &models,
+            &acp::ModelId::new(selected.to_string()),
+        );
+        assert_eq!(
+            seeded.as_deref(),
+            Some(selected),
+            "a fresh session must record the catalog key it was spawned with"
+        );
+        assert_eq!(
+            crate::agent::models::platform_for_slug(&models, seeded.as_deref(), slug),
+            kigi_models::parse_managed_model_key(selected).map(|(p, _)| p),
+            "…and that key must resolve THIS session's own platform for the bare slug"
+        );
+    }
+
+    // RESUME/LOAD: `acp_agent::load_session` spawns with the RAW persisted id,
+    // which after any model switch is the bare routing slug. THIS seed is what
+    // turns it into a key — the picker's `.rev()` answer, which is the resume
+    // default for a collided slug.
+    let resumed = crate::agent::models::selected_catalog_key_for_spawn(
+        &models,
+        &acp::ModelId::new(slug.to_string()),
+    );
+    assert_eq!(
+        resumed.as_deref(),
+        Some(oauth_twin),
+        "a bare persisted slug resolves through the picker's own lookup"
+    );
+
+    // A model that is no longer in the catalog seeds NOTHING, which (H-b) then
+    // refuses rather than guessing a twin.
+    assert_eq!(
+        crate::agent::models::selected_catalog_key_for_spawn(
+            &models,
+            &acp::ModelId::new("gone/model".to_string()),
+        ),
+        None,
+        "a model that left the catalog must not seed a key"
+    );
+}
+
+/// H-c — coverage for the SECOND production writer: `SetSessionModel`
+/// (`handle_set_session_model`), the picker's own path. The existing test
+/// through this handler passes `None`, so a handler that dropped the key on the
+/// floor stayed green.
+///
+/// End-to-end: after the switch the session's per-turn config must carry the
+/// SELECTED twin's pooled resolver and adaptation, even though the bare slug in
+/// the config is ambiguous.
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn set_session_model_records_the_key_the_next_turn_resolves_on() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let slug = "claude-opus-4-8";
+            let host = anthropic_collision_host();
+            let api_key_twin = "anthropic/claude-opus-4-8";
+            let oauth_twin = "claude-pro-max/claude-opus-4-8";
+            let (_dir, actor, _rx) = actor_with_catalog(
+                vec![
+                    managed_entry(api_key_twin, slug, &host),
+                    managed_entry(oauth_twin, slug, &host),
+                ],
+                api_key_twin,
+                "sk-ant-user",
+            )
+            .await;
+
+            // Switch to the SUBSCRIPTION twin, exactly as
+            // `agent/handlers/model_switch.rs` does: the ambiguous slug in the
+            // sampler config plus the catalog KEY the picker resolved.
+            let models = actor.models_manager.models();
+            let entry = models.get(oauth_twin).expect("catalog entry");
+            let sampler = crate::agent::config::sampling_config_for_model(
+                entry,
+                crate::agent::config::resolve_credentials(entry, None),
+                None,
+            );
+            actor
+                .handle_set_session_model(
+                    sampler,
+                    Some(oauth_twin.to_string()),
+                    false,
+                    false,
+                    true,
+                    85,
+                )
+                .await
+                .expect("model switch");
+
+            assert_eq!(
+                actor.selected_catalog_key().as_deref(),
+                Some(oauth_twin),
+                "SetSessionModel must record the picker's catalog key"
+            );
+            let cfg = actor.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_some(),
+                "the switched-to subscription model must keep a live pooled resolver"
+            );
+            assert!(
+                cfg.anthropic_oauth,
+                "…and the Claude OAuth Messages adaptation"
+            );
+
+            // And back: switching to the API-key twin must UNDO both.
+            let entry = models.get(api_key_twin).expect("catalog entry");
+            let sampler = crate::agent::config::sampling_config_for_model(
+                entry,
+                crate::agent::config::resolve_credentials(entry, None),
+                None,
+            );
+            actor
+                .handle_set_session_model(
+                    sampler,
+                    Some(api_key_twin.to_string()),
+                    false,
+                    false,
+                    true,
+                    85,
+                )
+                .await
+                .expect("model switch");
+            let cfg = actor.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_none() && !cfg.anthropic_oauth,
+                "LEAK: switching back to the API-key twin must drop the pooled resolver \
+                 and the OAuth adaptation"
+            );
+        })
+        .await;
+}
+
+/// H-c — `OverrideModelName` is the one command that rewrites
+/// `SamplingConfig::model` WITHOUT going through `SetSessionModel`, so it used
+/// to leave `selected_catalog_key` naming a model the session is no longer on.
+/// It must keep the field consistent: KEEP when the key still names the new
+/// routing name, CLEAR otherwise — never re-resolve, which would put the
+/// `.rev()` guess into the field the rule treats as a deliberate selection.
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn override_model_name_keeps_the_session_key_consistent() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let slug = "claude-opus-4-8";
+            let host = anthropic_collision_host();
+            let oauth_twin = "claude-pro-max/claude-opus-4-8";
+            let (_dir, actor, _rx) = actor_with_catalog(
+                vec![
+                    managed_entry("anthropic/claude-opus-4-8", slug, &host),
+                    managed_entry(oauth_twin, slug, &host),
+                ],
+                oauth_twin,
+                "unused",
+            )
+            .await;
+
+            // A rename to the SAME model's routing slug (or to its catalog key)
+            // keeps the selection.
+            for same in [slug, oauth_twin] {
+                actor.retain_selected_catalog_key_for(same);
+                assert_eq!(
+                    actor.selected_catalog_key().as_deref(),
+                    Some(oauth_twin),
+                    "{same}: still names the selected entry — keep it"
+                );
+            }
+
+            // A rename to a DIFFERENT name makes the key stale: clear it, so the
+            // collided slug refuses (H-b) instead of resolving the old model.
+            actor.retain_selected_catalog_key_for("some-harness-model-name");
+            assert_eq!(
+                actor.selected_catalog_key(),
+                None,
+                "a stale key must be cleared, not carried into the next turn's \
+                 platform lookup"
+            );
+        })
+        .await;
+}
+
+/// M3 — the FIRST-PARTY aux case must honour the session gate, which is what
+/// the old shape did implicitly.
+///
+/// `stamp_session_local_sampler_fields` used to copy
+/// `active_session_config.bearer_resolver`, and that field is `None` whenever
+/// the gate is inactive. Re-pointing the aux resolver at the chokepoint (the
+/// LEAK 1b fix) made it `Some(primary)` for the session's own coding endpoint
+/// REGARDLESS of the gate — so a BYOK / api-key session with a `[model.*]` aux
+/// entry carrying its own key on that endpoint had that key REPLACED by the
+/// primary bearer on every image-describe / auto-mode-classifier / summary
+/// request (`SamplingClient::post` overrides the auth header from the resolver).
+///
+/// Revert-to-red (production, compiles): delete the
+/// `if is_primary_channel && !SessionTokenAuthGate::new(…).active()` early
+/// return from `sampler_turn::aux_bearer_resolver_for` and the first two rows
+/// below resolve `KIMI_TOKEN`.
+#[tokio::test(flavor = "current_thread")]
+async fn first_party_aux_resolver_honours_the_session_gate() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let coding_host = kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url;
+            let aux_slug = "kigi-aux";
+            let mut info = crate::agent::config::ModelInfo::fallback(aux_slug);
+            info.id = None; // a `[model.kigi-aux]` block, not a registry entry
+            info.base_url = coding_host.to_string();
+            let aux_entry = crate::agent::config::ModelEntry {
+                info,
+                api_key: None,
+                env_key: None,
+                api_base_url: None,
+            };
+
+            // (case, ACP auth method, the aux model's own BYOK status, expected)
+            for (case, auth_method, byok, expect_resolver) in [
+                (
+                    "an API-key session: the aux model's own key must survive",
+                    "deepseek",
+                    crate::agent::auth_method::ModelByok::NotByok,
+                    false,
+                ),
+                (
+                    "a BYOK aux entry under a session method: its env_key wins",
+                    "cached_token",
+                    crate::agent::auth_method::ModelByok::Byok,
+                    false,
+                ),
+                (
+                    "the first-party subscription aux channel: byte-identical",
+                    "cached_token",
+                    crate::agent::auth_method::ModelByok::NotByok,
+                    true,
+                ),
+            ] {
+                let (_dir, actor, _rx) = actor_with_catalog(
+                    vec![(aux_slug.to_string(), aux_entry.clone())],
+                    aux_slug,
+                    "",
+                )
+                .await;
+                actor
+                    .auth_method_id
+                    .store(Some(Arc::new(acp::AuthMethodId::new(auth_method))));
+                actor.model_auth_facts.replace(Some((
+                    aux_slug.to_string(),
+                    crate::agent::config::ModelAuthFacts {
+                        byok,
+                        auth_scheme: Default::default(),
+                    },
+                )));
+
+                let resolved = actor.aux_bearer_resolver(aux_slug, coding_host);
+                assert_eq!(
+                    resolved.is_some(),
+                    expect_resolver,
+                    "{case}: aux resolver presence on the session's own endpoint"
+                );
+                if let Some(resolver) = resolved {
+                    assert_eq!(
+                        resolver.current_bearer(),
+                        Some(KIMI_TOKEN.to_string()),
+                        "{case}: and when it IS kept it is the primary's, live"
+                    );
+                }
+            }
+        })
+        .await;
+}
+
+/// M-aux (REGRESSION this remediation introduced) — an aux call must NOT evict
+/// the SESSION model's memoized auth facts.
+///
+/// `SessionActor::model_auth_facts` is a SINGLE slot. When `aux_bearer_resolver`
+/// began asking it about the AUX slug, a definite result overwrote the session
+/// model's entry, and:
+///   (a) the next `reconstruct_full_config` re-paid `load_effective_config()` +
+///       `resolve_model_list()` — the per-turn disk read M7/M9 removed — on top
+///       of the one the aux call itself paid; and
+///   (b) the memo's documented purpose (a transient `Unknown` falling back to
+///       the last DEFINITE value FOR THE SAME model_id) was defeated: with the
+///       aux slug in the slot, the session model's `Unknown` degrades to
+///       `endpoint_is_first_party`, which is `false` for every
+///       subscription-OAuth host — the session loses its `bearer_resolver` and
+///       401s unrecoverably ~1h in, the failure L13 exists to prevent.
+///
+/// Round 3's deleted `repoint_aux_bearer_resolver` never touched the memo.
+///
+/// Revert-to-red (production, compiles): make `SessionActor::aux_bearer_resolver`
+/// call `self.model_auth_facts(slug)` instead of `self.aux_model_auth_facts(slug)`
+/// — the slot then names the aux slug and both assertions below fail.
+#[tokio::test(flavor = "current_thread")]
+#[serial_test::serial]
+async fn an_aux_call_does_not_evict_the_session_models_auth_facts() {
+    let _env = anthropic_collision_env_guard();
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let session_slug = "claude-opus-4-8";
+            let host = anthropic_collision_host();
+            let oauth_twin = "claude-pro-max/claude-opus-4-8";
+            let (_dir, actor, _rx) = actor_with_catalog(
+                vec![managed_entry(oauth_twin, session_slug, &host)],
+                oauth_twin,
+                "unused",
+            )
+            .await;
+
+            // The session model's DEFINITE facts, as a turn would have memoized
+            // them.
+            actor.model_auth_facts.replace(Some((
+                session_slug.to_string(),
+                crate::agent::config::ModelAuthFacts {
+                    byok: crate::agent::auth_method::ModelByok::NotByok,
+                    auth_scheme: Default::default(),
+                },
+            )));
+
+            // An aux turn: the auto-mode classifier / image-describe slug, which
+            // is NOT the session's model.
+            let _ = actor.aux_bearer_resolver("kigi-aux-classifier", &host);
+
+            let memo = actor.model_auth_facts.borrow();
+            let (cached_id, facts) = memo
+                .as_ref()
+                .expect("the session model's memo must survive an aux call");
+            assert_eq!(
+                cached_id, session_slug,
+                "an aux call evicted the SESSION model's memo: the next turn re-reads \
+                 config from disk, and a transient Unknown loses its definite fallback"
+            );
+            assert_eq!(facts.byok, crate::agent::auth_method::ModelByok::NotByok);
+        })
+        .await;
+}

crates/codegen/kigi-shell/src/session/acp_session_tests/session_bearer_leak_tests.rs

@@ -0,0 +1,437 @@
+//! LEAK GUARD (bearer_resolver channel): the primary (Kimi) subscription bearer
+//! must never be stamped on a request to a host that does not own it.
+//!
+//! Chain the guard closes: a session-based ACP method (`cached_token` /
+//! `kimi-code` / any OAuth platform) + a selected API-key-platform model
+//! classifies `ModelByok::NotByok` (the model carries no `[model.*]` key), the
+//! pre-fix `session_token_auth_gate` returned `true` unconditionally on that
+//! arm, the manager lookup fell through to the primary Kimi manager for a
+//! non-OAuth platform, and `SamplingClient::post` then REPLACED the correctly
+//! resolved provider key with the Kimi bearer on the wire.
+//!
+//! The `api_key` half of the same defect (the config never even gets the
+//! provider key, because `resolve_credentials` stamps the session token) is
+//! pinned in `agent/mvp_agent/tests/api_key_channel_leak_tests.rs`, which drives
+//! the real `prepare_sampling_config_for_model` resolution path. These tests
+//! deliberately do NOT hand-stamp a provider key except where the assertion is
+//! about the resolver overwriting one that already resolved correctly.
+//!
+//! The counterpart contract these tests also pin: the four subscription-OAuth
+//! platforms have non-first-party base URLs but MUST keep a live
+//! `bearer_resolver` drawn from their OWN pooled `AuthManager`, or they lose
+//! mid-session token refresh.
+//!
+//! STORAGE DISCIPLINE (H6/M8): nothing here touches the developer's real
+//! `~/.kigi` and nothing hot-swaps the process-global OAuth pool. Under
+//! `cfg(test)` `oauth_registry::pool_home()` is a per-process temp path that is
+//! never created, so every pooled manager is empty — exactly what the
+//! assertions need (a live resolver that is provably NOT the Kimi one) — and
+//! the binary leaves nothing behind.
+
+use super::support::*;
+use super::*;
+use crate::agent::auth_method::ModelByok;
+use crate::agent::config::{ModelAuthFacts, ModelEntry, ModelInfo};
+use crate::auth::{AuthManager, AuthMode, KimiAuth, KimiCodeConfig};
+use kigi_sampler::BearerResolver;
+use std::sync::Arc;
+use tokio::sync::mpsc;
+
+/// The primary session bearer. Any occurrence of this string in an outgoing
+/// request to a third-party host is the defect.
+pub(super) const KIMI_TOKEN: &str = "kimi-subscription-token-DO-NOT-LEAK";
+
+/// `(tempdir, manager)` standing in for the session's primary Kimi
+/// `AuthManager`, holding a live (unexpired) OAuth session bearer.
+fn kimi_primary() -> (tempfile::TempDir, Arc<AuthManager>) {
+    let dir = tempfile::tempdir().expect("tempdir");
+    let am = Arc::new(AuthManager::new(dir.path(), KimiCodeConfig::default()));
+    am.hot_swap(KimiAuth {
+        key: KIMI_TOKEN.to_string(),
+        auth_mode: AuthMode::OAuth,
+        refresh_token: Some("rt".into()),
+        expires_at: Some(chrono::Utc::now() + chrono::Duration::hours(1)),
+        ..KimiAuth::test_default()
+    });
+    (dir, am)
+}
+
+/// One catalog entry: catalog key `catalog_key`, routing slug `slug`, routed at
+/// `base_url`, carrying no credential of its own (the shape every fetched
+/// registry model has).
+pub(super) fn managed_entry(catalog_key: &str, slug: &str, base_url: &str) -> (String, ModelEntry) {
+    let mut info = ModelInfo::fallback(slug);
+    info.id = Some(catalog_key.to_string());
+    info.base_url = base_url.to_string();
+    (
+        catalog_key.to_string(),
+        ModelEntry {
+            info,
+            api_key: None,
+            env_key: None,
+            api_base_url: None,
+        },
+    )
+}
+
+/// A `SessionActor` on a session-based ACP method with a live Kimi primary,
+/// whose live catalog holds `catalog` and whose SELECTED model is the catalog
+/// key `selected` (the picker's own notion of "current"). `wire_key` is the
+/// already-correctly-resolved provider credential sitting in chat state.
+///
+/// The per-model BYOK memo is pinned to `NotByok` on purpose: that is what a
+/// fetched registry model actually resolves to (`resolve_model_auth_facts` only
+/// ever sees `default_models.json` + `[model.*]`), and pinning it keeps the test
+/// independent of the developer's on-disk `~/.kigi/config.toml`.
+pub(super) async fn actor_with_catalog(
+    catalog: Vec<(String, ModelEntry)>,
+    selected: &str,
+    wire_key: &str,
+) -> (
+    tempfile::TempDir,
+    Arc<SessionActor>,
+    mpsc::UnboundedReceiver<PersistenceMsg>,
+) {
+    let (dir, am) = kimi_primary();
+    let (gateway_tx, _gateway_rx) = mpsc::unbounded_channel();
+    let (persistence_tx, persistence_rx) = mpsc::unbounded_channel();
+    let mut actor = create_test_actor(50_000, 200_000, 85, gateway_tx, persistence_tx).await;
+    actor.auth_manager = Some(am);
+    actor.auth_method_id = test_auth_method_id("cached_token");
+
+    let mut selected_entry = None;
+    for (key, entry) in catalog {
+        if key == selected {
+            selected_entry = Some(entry.clone());
+        }
+        actor.models_manager.insert_test_entry(key, entry);
+    }
+    let selected_entry = selected_entry.expect("the selected key must be in the catalog");
+    // H4: the SESSION owns its selection. The process-global
+    // `ModelsManager::current_model_id()` is deliberately left UNSET (it still
+    // names the startup default, which is not in this catalog) — exactly what
+    // Leader mode produces, since `agent/handlers/model_switch.rs` never calls
+    // `set_current_model_id` there, and what a second concurrent session on a
+    // colliding slug produces (last writer wins). Every assertion below
+    // therefore rides the per-session key, not the global cell.
+    *actor.selected_catalog_key.borrow_mut() = Some(selected.to_string());
+
+    let slug = selected_entry.info().model.clone();
+    actor
+        .chat_state_handle
+        .update_sampling_config(kigi_sampling_types::SamplingConfig {
+            base_url: selected_entry.info().base_url.clone(),
+            model: slug.clone(),
+            max_completion_tokens: None,
+            temperature: None,
+            top_p: None,
+            api_backend: Default::default(),
+            chat_compat: Default::default(),
+            extra_headers: Default::default(),
+            context_window: std::num::NonZeroU64::new(200_000).unwrap(),
+            reasoning_effort: None,
+            stream_tool_calls: None,
+        });
+    actor
+        .chat_state_handle
+        .update_credentials(kigi_chat_state::Credentials {
+            api_key: Some(wire_key.to_string()),
+            auth_type: kigi_chat_state::AuthType::SessionToken,
+            ..Default::default()
+        });
+    actor.model_auth_facts.replace(Some((
+        slug,
+        ModelAuthFacts {
+            byok: ModelByok::NotByok,
+            auth_scheme: Default::default(),
+        },
+    )));
+    (dir, Arc::new(actor), persistence_rx)
+}
+
+/// Single-entry convenience over [`actor_with_catalog`].
+pub(super) async fn actor_on_managed_model(
+    catalog_key: &str,
+    slug: &str,
+    base_url: &str,
+    wire_key: &str,
+) -> (
+    tempfile::TempDir,
+    Arc<SessionActor>,
+    mpsc::UnboundedReceiver<PersistenceMsg>,
+) {
+    actor_with_catalog(
+        vec![managed_entry(catalog_key, slug, base_url)],
+        catalog_key,
+        wire_key,
+    )
+    .await
+}
+
+/// THE leak test, at the wire. A `deepseek/deepseek-chat` turn on a session
+/// (`cached_token`) method with a live Kimi primary must send DeepSeek's own key
+/// — the Kimi subscription bearer must not appear anywhere in the request.
+///
+/// Revert-to-red: dropping the `credential_class` conjunct from
+/// `session_token_auth_gate` puts `Bearer <KIMI_TOKEN>` on this request.
+#[tokio::test(flavor = "multi_thread")]
+async fn deepseek_turn_under_a_kimi_session_sends_no_kimi_bearer_on_the_wire() {
+    let server = wiremock::MockServer::start().await;
+    wiremock::Mock::given(wiremock::matchers::method("POST"))
+        .and(wiremock::matchers::path("/chat/completions"))
+        .respond_with(
+            wiremock::ResponseTemplate::new(200).set_body_json(serde_json::json!({
+                "id": "cmpl-1",
+                "object": "chat.completion",
+                "created": 0,
+                "model": "deepseek-chat",
+                "choices": [{
+                    "index": 0,
+                    "message": { "role": "assistant", "content": "ok" },
+                    "finish_reason": "stop"
+                }]
+            })),
+        )
+        .mount(&server)
+        .await;
+    let uri = server.uri();
+
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let (_dir, actor, _rx) = actor_on_managed_model(
+                "deepseek/deepseek-chat",
+                "deepseek-chat",
+                &uri,
+                "sk-deepseek-provider-key",
+            )
+            .await;
+
+            let cfg = actor.reconstruct_full_config().await;
+            assert!(
+                cfg.bearer_resolver.is_none(),
+                "an API-key platform model must get NO session bearer resolver"
+            );
+            let client =
+                kigi_sampler::SamplingClient::new(cfg).expect("sampling client must construct");
+            let _ = client
+                .chat_completion(kigi_sampling_types::ChatCompletionRequest::new(
+                    "deepseek-chat",
+                    vec![kigi_sampling_types::ChatRequestMessage::user("hi")],
+                ))
+                .await;
+        })
+        .await;
+
+    let requests = server
+        .received_requests()
+        .await
+        .expect("wiremock records requests");
+    assert_eq!(requests.len(), 1, "exactly one inference request was sent");
+    let auth = requests[0]
+        .headers
+        .get("authorization")
+        .and_then(|v| v.to_str().ok())
+        .expect("the request must carry an Authorization header")
+        .to_string();
+    assert!(
+        !auth.contains(KIMI_TOKEN),
+        "the Kimi subscription bearer must never reach a third-party inference host"
+    );
+    assert_eq!(
+        auth, "Bearer sk-deepseek-provider-key",
+        "the correctly-resolved provider key must survive to the wire"
+    );
+}
+
+/// The same guard for every other API-key registry platform shape: OpenAI
+/// (Responses), Anthropic (x-api-key/Messages), Groq, Together and Z.AI CN — all
+/// classify `NotByok`, all route to a non-first-party host, none may receive a
+/// session bearer resolver.
+#[tokio::test(flavor = "current_thread")]
+async fn api_key_platform_models_get_no_session_bearer_resolver() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            for (catalog_key, slug, base_url) in [
+                ("openai/gpt-5.2", "gpt-5.2", "https://api.openai.com/v1"),
+                (
+                    "anthropic/claude-opus-4-8",
+                    "claude-opus-4-8",
+                    "https://api.anthropic.com/v1",
+                ),
+                ("groq/llama-4", "llama-4", "https://api.groq.com/openai/v1"),
+                ("together/qwen-3", "qwen-3", "https://api.together.xyz/v1"),
+                (
+                    "zai-coding-cn/glm-5",
+                    "glm-5",
+                    "https://open.bigmodel.cn/api/paas/v4",
+                ),
+            ] {
+                let (_dir, actor, _rx) =
+                    actor_on_managed_model(catalog_key, slug, base_url, "sk-provider-key").await;
+                let cfg = actor.reconstruct_full_config().await;
+                assert!(
+                    cfg.bearer_resolver.is_none(),
+                    "{catalog_key}: an API-key platform must get no session bearer resolver"
+                );
+                assert_eq!(
+                    cfg.api_key.as_deref(),
+                    Some("sk-provider-key"),
+                    "{catalog_key}: the provider key must stay on the config"
+                );
+            }
+        })
+        .await;
+}
+
+/// C2 at the resolver channel: a `[model.*]` entry has NO platform
+/// (`info.id == None`), which used to be a blanket allow. Pointed at a
+/// third-party host it must get no session resolver; pointed at the session's
+/// own coding endpoint (a config.toml `[endpoints] coding_api_base_url`
+/// deployment, a `KIGI_CODE_BASE_URL` override, or a local dev proxy) it must
+/// keep one — that is why the predicate is not `is_first_party_url`.
+///
+/// Revert-to-red: make `CredentialAuthority::is_session_coding_endpoint` return
+/// `true` unconditionally and a Kimi resolver lands on the openai.com config.
+#[tokio::test(flavor = "current_thread")]
+async fn config_model_entry_takes_a_session_resolver_only_on_its_own_endpoint() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            for base_url in ["https://api.openai.com/v1", "https://api.deepseek.com/v1"] {
+                let mut info = ModelInfo::fallback("gpt-4o");
+                info.id = None; // a `[model.gpt-4o]` block
+                info.base_url = base_url.to_string();
+                let entry = ModelEntry {
+                    info,
+                    api_key: None,
+                    // An env_key that is NOT set: `has_own_credentials()` probes
+                    // `std::env::var` at call time, so this classifies NotByok.
+                    env_key: None,
+                    api_base_url: None,
+                };
+                let (_dir, actor, _rx) =
+                    actor_with_catalog(vec![("gpt-4o".to_string(), entry)], "gpt-4o", "").await;
+                let cfg = actor.reconstruct_full_config().await;
+                assert!(
+                    cfg.bearer_resolver.is_none(),
+                    "LEAK: a [model.*] block at {base_url} must get no session bearer resolver"
+                );
+            }
+
+            for base_url in [
+                kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+                "http://127.0.0.1:4141/v1",
+            ] {
+                let mut info = ModelInfo::fallback("kigi-4.5");
+                info.id = None;
+                info.base_url = base_url.to_string();
+                let entry = ModelEntry {
+                    info,
+                    api_key: None,
+                    env_key: None,
+                    api_base_url: None,
+                };
+                let (_dir, actor, _rx) =
+                    actor_with_catalog(vec![("kigi-4.5".to_string(), entry)], "kigi-4.5", "").await;
+                let resolver = actor
+                    .reconstruct_full_config()
+                    .await
+                    .bearer_resolver
+                    .expect("the session's own endpoint keeps the session resolver");
+                assert_eq!(
+                    resolver.current_bearer(),
+                    Some(KIMI_TOKEN.to_string()),
+                    "{base_url}: a custom deployment / dev proxy is unchanged"
+                );
+            }
+        })
+        .await;
+}
+
+/// The Kimi / first-party subscription channel must be BYTE-IDENTICAL: the
+/// session model keeps its live bearer_resolver AND the pre-flight refresh
+/// still heals a stale buffered key. This is also what proves the Kimi bearer is
+/// live in every LEAK assertion in this module — it WOULD leak if the guard
+/// were missing.
+///
+/// (L12: the slug-collision commentary that used to sit here belongs to the
+/// collision tests in `session_bearer_leak_platform_tests`, which is where its
+/// revert-to-red actually reproduces; on this first-party test it never could.)
+#[tokio::test(flavor = "current_thread")]
+async fn kimi_first_party_model_still_rides_the_primary_session_bearer() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let (_dir, actor, _rx) = actor_on_managed_model(
+                "kimi-code/kimi-for-coding",
+                "kimi-for-coding",
+                kigi_env::PRODUCTION_ENDPOINTS.coding_api_base_url,
+                "stale-buffered-token",
+            )
+            .await;
+
+            let cfg = actor.reconstruct_full_config().await;
+            let resolver = cfg
+                .bearer_resolver
+                .as_ref()
+                .expect("the subscription model must keep the live session resolver");
+            assert_eq!(
+                resolver.current_bearer(),
+                Some(KIMI_TOKEN.to_string()),
+                "the first-party model resolves the primary session bearer"
+            );
+
+            actor.refresh_token_if_expired().await;
+            assert_eq!(
+                actor
+                    .chat_state_handle
+                    .get_credentials()
+                    .await
+                    .api_key
+                    .as_deref(),
+                Some(KIMI_TOKEN),
+                "the first-party pre-flight refresh must still heal the stale key"
+            );
+        })
+        .await;
+}
+
+/// The persistence half of the defect: `refresh_token_if_expired` used to write
+/// the Kimi session token into `chat_state` `creds.api_key` for ANY
+/// session-method turn, from where it propagated to subagents and aux configs.
+/// A deepseek turn must leave the provider key untouched.
+///
+/// M7 rides along: a registry-platform model must not fall into
+/// `reload_api_key_from_config` at all (a `load_effective_config()` disk read
+/// per turn plus a permanently false "not found in config.toml" warning), so
+/// the key is left exactly as resolved.
+#[tokio::test(flavor = "current_thread")]
+async fn preflight_refresh_never_writes_the_kimi_token_into_a_platform_credential() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let (_dir, actor, _rx) = actor_on_managed_model(
+                "deepseek/deepseek-chat",
+                "deepseek-chat",
+                "https://api.deepseek.com/v1",
+                "sk-deepseek-provider-key",
+            )
+            .await;
+
+            actor.refresh_token_if_expired().await;
+
+            assert_eq!(
+                actor
+                    .chat_state_handle
+                    .get_credentials()
+                    .await
+                    .api_key
+                    .as_deref(),
+                Some("sk-deepseek-provider-key"),
+                "the Kimi session token must never overwrite a platform credential"
+            );
+        })
+        .await;
+}

crates/codegen/kigi-shell/src/session/acp_session_tests/support.rs

@@ -168,6 +168,7 @@ pub(crate) async fn create_test_actor_ex(
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             extra_headers: Default::default(),
             context_window: std::num::NonZeroU64::new(context_window)
                 .expect("test context_window must be non-zero"),
@@ -187,6 +188,7 @@ pub(crate) async fn create_test_actor_ex(
         },
         auth_method_id: test_auth_method_id("test-auth"),
         model_auth_facts: std::cell::RefCell::new(None),
+        selected_catalog_key: std::cell::RefCell::new(None),
         attribution_callback: None,
         auth_manager: None,
         state,

crates/codegen/kigi-shell/src/session/commands.rs

@@ -145,6 +145,13 @@ pub enum SessionCommand {
     },
     SetSessionModel {
         sampling_config: kigi_sampler::SamplerConfig,
+        /// The catalog KEY the picker resolved (`{platform}/{model}` for a
+        /// registry model), which `sampling_config.model` — the bare routing
+        /// slug — cannot express when an API-key platform and its
+        /// subscription-OAuth twin list the same id. The session stores it as
+        /// its OWN selection instead of reading the process-global
+        /// `ModelsManager::current_model_id()` (H4).
+        catalog_key: Option<String>,
         use_concise: bool,
         /// When `false`, skip the system prompt rewrite (concise/default swap).
         /// Set to `false` for forked sessions so mid-session model switches

crates/codegen/kigi-shell/src/session/compaction.rs

@@ -2149,6 +2149,7 @@ mod inline_auto_compact_flow_tests {
                 temperature: None,
                 top_p: None,
                 api_backend: Default::default(),
+                chat_compat: Default::default(),
                 extra_headers: Default::default(),
                 context_window: std::num::NonZeroU64::new(context_window)
                     .expect("test context_window must be non-zero"),
@@ -2167,6 +2168,7 @@ mod inline_auto_compact_flow_tests {
             },
             auth_method_id: test_auth_method_id("test-auth"),
             model_auth_facts: std::cell::RefCell::new(None),
+            selected_catalog_key: std::cell::RefCell::new(None),
             attribution_callback: None,
             auth_manager: None,
             state,

crates/codegen/kigi-shell/src/session/goal_classifier.rs

@@ -594,10 +594,10 @@ async fn write_patch_file_atomic(path: &Path, body: &str) -> std::io::Result<()>
         .unwrap_or("goal-classifier.patch");
     let tmp = dir.join(format!(".{file_name}.{}.tmp", uuid::Uuid::now_v7()));
     tokio::fs::write(&tmp, body).await?;
-    if let Err(err) = tokio::fs::rename(&tmp, path).await {
-        let _ = tokio::fs::remove_file(&tmp).await;
-        return Err(err);
-    }
+    let dest = path.to_path_buf();
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &dest))
+        .await
+        .map_err(std::io::Error::other)??;
     Ok(())
 }
 

crates/codegen/kigi-shell/src/session/goal_tracker.rs

@@ -895,7 +895,8 @@ impl GoalTracker {
         };
         let dest = goal_dir.join(name);
         let _ = std::fs::create_dir_all(&goal_dir);
-        if std::fs::rename(&src, &dest).is_ok() || copy_no_follow(&src, &dest).is_ok() {
+        if crate::util::fs::replace_file(&src, &dest).is_ok() || copy_no_follow(&src, &dest).is_ok()
+        {
             append_skeptic_reports(&scratch_root, &dest);
             o.last_classifier_details_path = Some(dest.to_string_lossy().into_owned());
         }

crates/codegen/kigi-shell/src/session/graph_project.rs

@@ -120,7 +120,7 @@ pub fn project(dir: &Path, state: &GraphOrchestration) -> std::io::Result<()> {
         f.write_all(&buf)?;
         f.sync_all()?;
     }
-    std::fs::rename(&tmp, &target)
+    crate::util::fs::replace_file(&tmp, &target)
 }
 
 /// Load the projected graph, `Ok(None)` when absent. Malformed content

crates/codegen/kigi-shell/src/session/helpers/session_compact.rs

@@ -1588,7 +1588,11 @@ mod reasoning_compaction_regression_tests {
             temperature: Some(0.7),
             top_p: None,
             api_backend: ApiBackend::ChatCompletions,
+            chat_compat: Default::default(),
             auth_scheme: Default::default(),
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
             extra_headers: Default::default(),
             context_window: 256_000,
             force_http1: false,

crates/codegen/kigi-shell/src/session/persistence.rs

@@ -879,7 +879,14 @@ pub struct Summary {
     /// `None` for sessions created before this field existed.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub sandbox_profile: Option<String>,
-    #[serde(default, skip_serializing_if = "Option::is_none")]
+    /// Lenient on read: an unknown token written by a newer kigi (grown
+    /// effort vocabulary) drops to `None` rather than hiding the whole
+    /// session from listings / failing resume after a version rollback.
+    #[serde(
+        default,
+        skip_serializing_if = "Option::is_none",
+        deserialize_with = "kigi_sampling_types::lenient_reasoning_effort_opt"
+    )]
     pub reasoning_effort: Option<ReasoningEffort>,
 }
 
@@ -1002,6 +1009,22 @@ mod is_hidden_tests {
         }
     }
 
+    #[test]
+    fn summary_unknown_reasoning_effort_token_degrades_to_none() {
+        // A summary written by a NEWER kigi with a grown effort vocabulary
+        // must not vanish from listings or fail resume on this binary.
+        let mut s = summary_with_kind(None);
+        s.reasoning_effort = Some(ReasoningEffort::Max);
+        let json = serde_json::to_string(&s).unwrap();
+        let future = json.replace("\"max\"", "\"hypermax\"");
+        assert_ne!(json, future, "fixture must actually carry the token");
+        let back: Summary = serde_json::from_str(&future).expect("record must survive");
+        assert_eq!(back.reasoning_effort, None);
+        // Known tokens (including post-split "max") still round-trip.
+        let back: Summary = serde_json::from_str(&json).unwrap();
+        assert_eq!(back.reasoning_effort, Some(ReasoningEffort::Max));
+    }
+
     #[test]
     fn summary_round_trips_and_defaults_reasoning_effort() {
         let mut s = summary_with_kind(None);

crates/codegen/kigi-shell/src/session/prompt_history.rs

@@ -98,7 +98,7 @@ pub fn truncate_if_needed(cwd: &str) -> io::Result<()> {
         }
     }
 
-    std::fs::rename(temp_path, path)?;
+    crate::util::fs::replace_file(&temp_path, &path)?;
     Ok(())
 }
 

crates/codegen/kigi-shell/src/session/storage/jsonl/mod.rs

@@ -13,6 +13,16 @@ use std::fs::OpenOptions;
 use std::io::{self, Read};
 use std::path::{Path, PathBuf};
 use tokio::io::AsyncWriteExt;
+
+/// Commit `tmp` over `target` with the shared Windows-safe replace
+/// (`util::fs::replace_file`): a bare async rename silently lost session
+/// state — including the switched model — on Windows whenever AV/indexer
+/// held the destination open.
+async fn replace_file_async(tmp: PathBuf, target: PathBuf) -> io::Result<()> {
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &target))
+        .await
+        .unwrap_or_else(|e| Err(io::Error::other(e)))
+}
 /// How the adapter resolves the session directory on disk.
 ///
 /// - `FromRoot` (default): computes `{root}/sessions/{urlencoded(cwd)}/{session_id}/`
@@ -289,7 +299,7 @@ impl JsonlStorageAdapter {
         }
         let tmp = path.with_extension("jsonl.tmp");
         tokio::fs::write(&tmp, &content).await?;
-        tokio::fs::rename(&tmp, &path).await
+        replace_file_async(tmp, path).await
     }
     fn read_jsonl<T: serde::de::DeserializeOwned>(&self, path: PathBuf) -> io::Result<Vec<T>> {
         if !path.exists() {
@@ -378,7 +388,7 @@ impl JsonlStorageAdapter {
             .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
         let tmp = summary_path.with_extension("json.tmp");
         std::fs::write(&tmp, &bytes)?;
-        std::fs::rename(&tmp, &summary_path)
+        crate::util::fs::replace_file(&tmp, &summary_path)
     }
     fn read_summary_sync(&self, info: &Info) -> io::Result<Summary> {
         let path = self.summary_file(info);
@@ -1057,7 +1067,7 @@ impl StorageAdapter for JsonlStorageAdapter {
         let target = self.plan_mode_state_file(info);
         let tmp = target.with_extension("json.tmp");
         tokio::fs::write(&tmp, json).await?;
-        tokio::fs::rename(&tmp, &target).await
+        replace_file_async(tmp, target).await
     }
     async fn write_signals(
         &self,
@@ -1069,7 +1079,7 @@ impl StorageAdapter for JsonlStorageAdapter {
         let target = self.signals_file(info);
         let tmp = target.with_extension("json.tmp");
         tokio::fs::write(&tmp, signals_json).await?;
-        tokio::fs::rename(&tmp, &target).await
+        replace_file_async(tmp, target).await
     }
     async fn write_announcement_state(
         &self,
@@ -1081,7 +1091,7 @@ impl StorageAdapter for JsonlStorageAdapter {
         let target = self.announcement_state_file(info);
         let tmp = target.with_extension("json.tmp");
         tokio::fs::write(&tmp, json).await?;
-        tokio::fs::rename(&tmp, &target).await
+        replace_file_async(tmp, target).await
     }
     async fn write_goal_mode_state(
         &self,
@@ -1096,7 +1106,7 @@ impl StorageAdapter for JsonlStorageAdapter {
         }
         let tmp = target.with_extension("json.tmp");
         tokio::fs::write(&tmp, json).await?;
-        tokio::fs::rename(&tmp, &target).await
+        replace_file_async(tmp, target).await
     }
     async fn write_graph_mode_state(
         &self,
@@ -1119,7 +1129,7 @@ impl StorageAdapter for JsonlStorageAdapter {
         }
         let tmp = target.with_extension("json.tmp");
         tokio::fs::write(&tmp, json).await?;
-        tokio::fs::rename(&tmp, &target).await
+        replace_file_async(tmp, target).await
     }
     async fn load_session(&self, info: &Info) -> io::Result<PersistedData> {
         let summary = self.read_summary_sync(info)?;

crates/codegen/kigi-shell/src/session/storage/summary_write.rs

@@ -230,7 +230,10 @@ fn write_summary_atomic(summary_path: &Path, summary: &Summary) -> io::Result<()
         .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
     let tmp = summary_path.with_extension("json.tmp");
     std::fs::write(&tmp, &bytes)?;
-    std::fs::rename(&tmp, summary_path)
+    // Windows-safe replace: this is the write that persists a session's
+    // CURRENT MODEL — a bare rename made a switched model silently revert
+    // on resume whenever AV/indexer held summary.json open on Windows.
+    crate::util::fs::replace_file(&tmp, summary_path)
 }
 
 #[cfg(test)]

crates/codegen/kigi-shell/src/test_support/lsp_runtime.rs

@@ -44,7 +44,11 @@ pub(crate) fn ctx_with_toggle(toggle: HashMap<String, bool>) -> SubagentSpawnCon
             temperature: None,
             top_p: None,
             api_backend: Default::default(),
+            chat_compat: Default::default(),
             auth_scheme: Default::default(),
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
             extra_headers: Default::default(),
             context_window: 256_000,
             force_http1: false,

crates/codegen/kigi-shell/src/trace_classifier/mod.rs

@@ -1044,7 +1044,9 @@ pub async fn resolve_api_key(explicit: Option<&str>, kigi_home: &Path) -> Result
     if let Some(k) = explicit.map(str::trim).filter(|s| !s.is_empty()) {
         return Ok(k.to_owned());
     }
-    let env_key = std::env::var("XAI_API_KEY").ok();
+    // Canonical house-key resolution: KIGI_API_KEY, then the back-compat
+    // XAI_API_KEY / KIGI_CODE_XAI_API_KEY (see `read_xai_api_key_env`).
+    let env_key = crate::agent::auth_method::read_xai_api_key_env().ok();
     if let Some(k) = env_key.as_deref().map(str::trim).filter(|s| !s.is_empty()) {
         return Ok(k.to_owned());
     }
@@ -1052,7 +1054,7 @@ pub async fn resolve_api_key(explicit: Option<&str>, kigi_home: &Path) -> Result
         return Ok(key);
     }
     Err(anyhow!(
-        "no API key: pass --api-key, set XAI_API_KEY, or run `kigi login` to populate \
+        "no API key: pass --api-key, set KIGI_API_KEY, or run `kigi login` to populate \
          <kigi-home>/auth.json. An expired OIDC token is auto-refreshed when a refresh_token \
          is present; if not, re-login is required."
     ))
@@ -2344,7 +2346,7 @@ mod tests {
         .await;
         assert!(
             msg.contains("--api-key")
-                && msg.contains("XAI_API_KEY")
+                && msg.contains("KIGI_API_KEY")
                 && msg.contains("kigi login")
                 && msg.contains("auth.json"),
             "error names all three sources: {msg}",
@@ -2445,7 +2447,9 @@ mod tests {
     ///   refresher that could mint credentials independent of the
     ///   fixture.
     const ISOLATED_ENV_KEYS: &[&str] = &[
+        "KIGI_API_KEY",
         "XAI_API_KEY",
+        "KIGI_CODE_XAI_API_KEY",
         "KIGI_AUTH",
         "KIGI_AUTH_PATH",
         "KIGI_AUTH_PROVIDER_COMMAND",

crates/codegen/kigi-shell/src/util/config/campaigns.rs

@@ -114,9 +114,7 @@ fn dismiss_campaign_ids_at(
     let nonce = DISMISS_TMP_NONCE.fetch_add(1, Ordering::Relaxed);
     let tmp = path.with_extension(format!("json.{}.{}.tmp", std::process::id(), nonce));
     std::fs::write(&tmp, &json)?;
-    std::fs::rename(&tmp, &path).inspect_err(|_| {
-        let _ = std::fs::remove_file(&tmp);
-    })
+    crate::util::fs::replace_file(&tmp, &path)
 }
 
 /// `KIGI_CAMPAIGNS_OVERRIDE` JSON array replaces all sources (`[]` = none; beats

crates/codegen/kigi-shell/src/util/config/mcp.rs

@@ -367,7 +367,9 @@ pub async fn save_mcp_disabled_tools(server_name: &str, disabled_tools: &[String
         let _ = tokio::fs::create_dir_all(parent).await;
     }
     tokio::fs::write(&tmp, &toml_str).await?;
-    tokio::fs::rename(&tmp, &path).await?;
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &path))
+        .await
+        .map_err(std::io::Error::other)??;
     Ok(())
 }
 
@@ -419,7 +421,9 @@ pub async fn save_mcp_server_enabled(server_name: &str, enabled: bool) -> Result
         let _ = tokio::fs::create_dir_all(parent).await;
     }
     tokio::fs::write(&tmp, &toml_str).await?;
-    tokio::fs::rename(&tmp, &path).await?;
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &path))
+        .await
+        .map_err(std::io::Error::other)??;
     Ok(())
 }
 
@@ -476,7 +480,10 @@ pub async fn save_mcp_server_config_at(
         let _ = tokio::fs::create_dir_all(parent).await;
     }
     tokio::fs::write(&tmp, &toml_str).await?;
-    tokio::fs::rename(&tmp, &path).await?;
+    let dest = path.to_path_buf();
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &dest))
+        .await
+        .map_err(std::io::Error::other)??;
     Ok(())
 }
 
@@ -553,7 +560,12 @@ pub async fn delete_mcp_server_config_at(
         let _ = tokio::fs::create_dir_all(parent).await;
     }
     tokio::fs::write(&tmp, &toml_str).await?;
-    tokio::fs::rename(&tmp, &path).await?;
+    {
+        let dest = path.to_path_buf();
+        tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &dest))
+            .await
+            .map_err(std::io::Error::other)??;
+    }
 
     // Clean up OAuth credentials for the deleted server.
     if let Ok(mut cred_store) = kigi_mcp::credentials::McpCredentialStore::load_default() {

crates/codegen/kigi-shell/src/util/config/persist.rs

@@ -88,17 +88,15 @@ pub async fn save_config(config: &Config) -> Result<()> {
     }
     let _ = prior_mode;
 
-    tokio::fs::rename(&tmp, &path).await?;
+    // Windows-safe replace (delete-first + retry on sharing violations) —
+    // a bare rename made `/model` persistence silently fail on Windows
+    // whenever AV/indexer/cloud-sync held config.toml open.
+    tokio::task::spawn_blocking(move || crate::util::fs::replace_file(&tmp, &path))
+        .await
+        .map_err(|e| anyhow::anyhow!("config replace task: {e}"))??;
     Ok(())
 }
 
-/// Acquire the `config.toml` write lock used by [`save_config`], so callers that
-/// mutate the file directly (marketplace add/remove) can't interleave with a
-/// settings save and clobber it.
-pub(crate) async fn lock_config_writes() -> tokio::sync::MutexGuard<'static, ()> {
-    SAVE_LOCK.lock().await
-}
-
 /// Read a file, treating only `NotFound` as empty. Hard read errors (EACCES,
 /// EIO) propagate so callers don't clobber an unreadable file on the next write.
 pub(crate) fn read_to_string_or_empty(path: &std::path::Path) -> std::io::Result<String> {
@@ -145,11 +143,8 @@ pub(crate) fn atomic_write_string(path: &std::path::Path, content: &str) -> std:
     }
     let _ = prior_mode;
 
-    if let Err(e) = std::fs::rename(&tmp, path) {
-        let _ = std::fs::remove_file(&tmp);
-        return Err(e);
-    }
-    Ok(())
+    // Windows-safe replace; cleans up the tmp file on failure itself.
+    crate::util::fs::replace_file(&tmp, path)
 }
 
 /// Merge `[toolset.ask_user_question]` into the root table. `[toolset]` is

crates/codegen/kigi-shell/tests/common/mod.rs

@@ -38,6 +38,10 @@ pub fn test_sampler_config(
         top_p: None,
         api_backend,
         auth_scheme: Default::default(),
+        anthropic_oauth: false,
+        github_copilot: false,
+        openai_codex: false,
+        chat_compat: Default::default(),
         extra_headers: extra_headers
             .iter()
             .map(|(k, v)| (k.to_string(), v.to_string()))

crates/codegen/kigi-shell/tests/git_contention_e2e.rs

@@ -433,7 +433,7 @@ fn git_rebase_refresh_storm_e2e() {
         std::env::set_var("KIGI_SHARE_DIR", kigi_home.path());
         std::env::set_var("KIGI_CODE_BASE_URL", server.url());
         std::env::set_var("KIGI_API_BASE_URL", server.url());
-        std::env::set_var("XAI_API_KEY", "test-key-for-ci");
+        std::env::set_var("KIGI_API_KEY", "test-key-for-ci");
         std::env::set_var("KIGI_TELEMETRY_ENABLED", "false");
         std::env::set_var("KIGI_FEEDBACK_ENABLED", "false");
         std::env::set_var("KIGI_TRACE_UPLOAD", "false");