Ports kimi-code's AgentSwarm: a `prompt_template` containing `{{item}}`
expanded over an `items` list into up to 128 subagents, run to completion
and returned as one aggregate. Kigi already exceeds upstream on planning,
verification, isolation and merge via /graph; what it lacked was cheap
immediate fan-out. Entirely client-side — no new backend surface.
The engine is three pure pieces plus a runner: `plan` validates and
expands (every fault reported before a single member starts — a
half-launched swarm is expensive to unwind), `schedule` is the launch
ramp as testable arithmetic, `run` drives it against the existing
`SubagentBackend`. Reuses the single-spawn coordinator rather than
inventing a batch API.
Load-bearing decisions, each the result of a defect found in review:
- `backgrounded` is its own outcome. A member that outlives the 600s
foreground budget is detached by the coordinator and KEEPS RUNNING;
reporting it as failed invites the model to relaunch its item, putting
a second agent on the same files. It is never offered for resume.
- `InFlightGuard` cancels live members on Drop. Send-now cancels the turn
WITHOUT cancelling subagents and aborts the task; the dropped receivers
read as "parent gone" and each child re-attaches itself. There is no
cooperative path to use instead — `Cancellation` is constructed nowhere
in the tree — so Drop is the only seam that fires.
- Retries and wall clock are both bounded. The swarm blocks the caller's
turn, so every wait needs a ceiling it cannot argue past; stragglers at
the deadline are reported as still-running, with their ids.
- `ToolKind::AgentSwarm` is its own variant: `TemplateRenderer`'s
`by_kind` map holds one tool name per kind, so sharing `Task` would
silently redirect `${{ tools.by_kind.task }}` in other tools' prompts.
- An explicitly requested model that cannot be validated is refused, as
the task tool already does — one loud error beats `items.len()` quiet
ones. Depth stays capped at 1: upstream's unlimited nesting is a
hazard, not a feature.
- `SubagentResult.rate_limited` is classified where the typed ACP error
code is still in hand; a scheduler re-deriving it from a formatted
string would stop adapting the day the wording changed.
- Aggregate output is clamped per member (head+tail, loss stated):
native tool output is truncated nowhere downstream.
42 agent_swarm tests. The fake backend awaits, so the concurrency and
ordering assertions can actually fail; the cap test also proves the
fixture can exceed the cap.
`MemoryBackendParams` carried the primary session `AuthManager` and the
session api-key provider unconditionally, while `embed_base_url` is the
CURRENT MODEL's endpoint and `AuthRetryMiddleware` stamps `Authorization`
on every request it wraps. A user who enables `[memory.embedding] model`
while running a BYOK or subscription-OAuth model therefore sent the Kimi
session bearer to that third party.
`EndpointScopedCredentials` binds the credential to the one endpoint it
may reach: `for_endpoint` drops the handle unless the caller vouches for
the URL, and `approved_for` re-checks at provider-build time in release
too, because `MemoryBackendParams` is `Clone` and callers rewrite fields
on the copy.
The shell decides through `CredentialAuthority::manager_for` rather than
a second URL predicate — it answers both whether a credential may ride
and which manager governs it, so a subscription-OAuth platform gets its
own pooled manager. The session's `SharedApiKeyProvider` is not
forwarded at all: it is hard-wired to the primary manager, so at a
pooled platform's host it would resolve the wrong bearer. A platform's
own `embed_api_key` is untouched and keeps serving its own endpoint.
The background reindex built a second provider straight from
`ApiEmbeddingProvider::from_session`, outside the chokepoint and without
401 refresh; it now embeds through the session's own params.
Test strength verified by mutation: with the guard reverted, exactly
`session_credentials_are_withheld_from_a_foreign_endpoint` and
`a_cloned_param_set_cannot_redirect_scoped_credentials` fail.
On Windows, crossterm's console backend never emits Event::Paste: a
terminal paste arrives as a per-character key burst that the event loop
reassembles with timing heuristics. Three of its bounds split real
pastes into fragments:
- the 10 ms continue window is below one Windows scheduler quantum
(~15.6 ms), so a routine mid-burst gap ended collection early;
- the 5000-event safety cap truncated any paste past ~5 KB;
- the 2 ms detection window missed bursts whose second chunk lagged,
so collection never started.
Each fragment became its own synthetic paste. The paste chip then held
only the first fragment, and pasting again could never byte-equal the
chip content, so repaste-to-expand inserted a duplicate instead of
expanding — while macOS worked, because the unix parser buffers a
bracketed paste to the end marker and always delivers one whole event.
Widen the continue window past one quantum (25 ms), raise the cap above
any real paste (200k events), and skip the detection window entirely
when the batch already holds a >=8-key burst, which typing and
auto-repeat never produce. Unix paths are untouched: batches containing
a real Event::Paste never enter this reassembly.
Covered by a paused-clock test that trickles a burst tail with
one-quantum gaps and asserts it reassembles into a single paste.
wait_and_drain treated every read() <= 0 as "queue drained", so a real
error (bad fd, EINVAL) silently ended the drain and the monitor kept
polling a broken fd. Distinguish the cases: EAGAIN ends the drain, EINTR
retries, and any other error propagates. monitor_loop logs the error
before exiting so a dead memory monitor is visible in the session log.
Sweep every first-party crate source (1956 .rs files) to the project comment
guidelines: delete redundant restatements, decorative banners, change
narration, and end-of-line comments; keep and tighten the crucial ones
(invariants, bug rationale, SAFETY blocks, ported-source attribution).
No functional code changed. Every edit is proven comment-only against the
prior tree by a comment-stripping lexer (string/char/raw-string aware) plus a
separate doctest-fence check. Where removing a comment made rustfmt or clippy
want to re-lay-out adjacent code, the minimal triggering comment is restored so
code tokens stay byte-identical.
Gates green: cargo fmt --all --check (0 diffs), cargo check and cargo clippy
--workspace --all-targets (0 warnings).
Adds scripts/check_codegen_comment_guidelines.py — the enforcement gate for
these guidelines (flags banners, end-of-line comments, change narration, and
commented-out code).
C3 (Pi openai-completions relocation): tool messages are text-only on
the OpenAI chat wire — image parts inside role:tool content 400 on
strict validators. Images from a consecutive tool-result run now batch
into ONE synthetic user message appended after the run (a user message
may not interrupt tool responses answering the same assistant's
tool_calls); an image-only result carries a '(see attached image)'
pointer placeholder. Applied at the typed conversion layer, so every
chat dialect gets the universally-valid shape.
Verified: models+sampling-types+sampler+chat-state+shell all green,
clippy clean.
C1 (decision): custom/BYOK ChatCompletions entries defaulted to the Kimi
dialect, leaking Kimi-specific body mutations (thinking:{…} control,
replayed reasoning_content, schema rewrites) to arbitrary third-party
OpenAI-compatible servers. The default is now Passthrough (vanilla
OpenAI semantics), with ONE exception mirroring Pi's base-url quirk
sniffing: entries pointed at the house/Kimi coding endpoint keep the
Kimi dialect. Registry platforms are unaffected (all declare
explicitly).
C2 (Pi mistral-conversations normalizer): Mistral requires tool-call ids
of EXACTLY nine [a-zA-Z0-9] chars; even same-session synthesized UUIDs
violate it. New ChatCompat::Mistral = StrictOpenAi behavior + the
normalizer — strip non-alphanumerics, keep exact-9 ids, else FNV-1a →
base36 (build-stable, deterministic across requests for prefix-cache
stability) with collision retry; ONE map covers tool_calls[].id and
tool_call_id so pairing survives. The mistral registry row and the
persisted 'mistral' serde value both resolve to it (pre-rename Mistral
sessions gain the contract automatically); Cerebras/NVIDIA stay on
StrictOpenAi untouched.
Verified: models+sampling-types+sampler+chat-state+shell all green,
clippy clean.
Cross-provider audit M4/M5/M6 (Anthropic Messages builder):
- Non-base64 data: URIs rode as ImageSource::Url — url sources are
http(s) only → 400; and the two image paths parsed data URIs
differently (user path split on first comma, tool-result path on
';base64,').
- No media-type whitelist: image/svg+xml and param-carrying headers
('image/webp;name=x') reached the wire → 400.
- Empty user content shipped empty arrays/text blocks → 400.
One shared parse_base64_image_data_uri (raster whitelist: jpeg/png/gif/
webp) now serves both paths; rejected images degrade to a SHORT
'[unsupported image]' placeholder (never the multi-megabyte payload);
empty user turns get '[empty message]' (mirrors the assistant guard).
Part 6 of the cross-provider replay audit.
Verified: sampling-types 292 + downstream green, clippy clean.
Cross-provider audit R5 + R2-residual: BackendToolCall items round-trip
as their typed shapes with provider-issued ids (grok x_search
CustomToolCall, web_search/code_interpreter calls) and Reasoning items
carry model-bound encrypted payloads — replaying either to a DIFFERENT
Responses target names undeclared tools / undecryptable material → 400.
New transform_items_for_responses pre-pass: each [Reasoning|
BackendToolCall]* Assistant run carries provenance in
AssistantItem.model_id; on confirmed mismatch with the request's target
model, Reasoning siblings are dropped and BackendToolCall demoted to the
same text summary the Messages/ChatCompletions builders already emit.
Same-model and unknown-provenance turns stay byte-verbatim (KV-cache
prefix stability preserved). The legacy-upgrade round-trip test now
models the same-model continuation it always described.
Part 5 of the cross-provider replay audit.
Verified: sampling-types 290 + downstream 5794 green, clippy clean.
The new truncation test's doc landed between backend_tool_call_position_
stable's doc and its #[test] attribute (clippy: duplicated attribute).
Each test owns its own doc block again.
Cross-provider audit R4+M3: the Responses leg passed tool-call ids
verbatim (call_id on both function_call and function_call_output) while
the Messages leg sanitized — and its closure used Unicode
is_alphanumeric, letting CJK ids through to Anthropic's ASCII-only
contract, with no empty-id fallback. Both providers enforce
[A-Za-z0-9_-]+ (the codex 400's own words). One module-scope
sanitize_tool_call_id now serves both builders, ASCII-only, empty → "_",
applied identically on call+result so pairing survives.
Part 4 of the cross-provider replay audit.
Verified: 6081 tests green across the four crates, clippy clean.
Cross-provider audit R3: reasoning captured on stateful api.openai.com
sessions (no include requested) carries a server-issued rs_* id but NO
encrypted_content; replaying that bare reference to the stateless
(store:false) codex backend points at server state chatgpt.com does not
have. adapt_body_for_codex_backend step 3 drops such items (encrypted
ones pass through verbatim). Capture-side include for the API-key path
is deferred: the typed CreateResponse is shared with the xai Responses
leg and changing its bytes needs separate validation.
Part 3 of the cross-provider replay audit.
Verified: sampling-types+sampler+chat-state+shell all green.
The multibyte-truncation test used 80 chars — below the 100-char cap, so
the truncation assertion failed (the previous commit's suite count was
misread; the FIX itself was correct and the panic repro held). 120 chars
now exercises both the boundary safety and the truncation.
The code-interpreter preview truncated at BYTE 100 — a guaranteed panic
on any CJK/emoji boundary in interpreted code. One poisoned history item
then crashed every subsequent request build on every backend (the
summary feeds both the Messages and ChatCompletions builders). Truncate
at 100 chars via char_indices instead. Panic pinned by test.
Part 2 of the cross-provider replay audit.
Verified: sampling-types 287 tests green.
Root cause of 400 Invalid 'input[N].id': '' on chatgpt.com/backend-api/
codex/responses: the Responses input builder replayed every stored
Reasoning item verbatim, and rs::ReasoningItem.id serializes
unconditionally — so foreign items (Messages-captured Anthropic
signatures, chat-completions-synthesized reasoning, stream-delta
fallbacks, legacy upgrades — all id '') reached the wire with an empty
id the API rejects. This also self-poisoned pure codex sessions whose
reasoning arrived only as deltas.
A native item always carries a server-issued rs_* id: empty id = foreign
= unusable by any Responses provider = dropped at the builder — the
exact mirror of the Messages builder's prune_replayed_thinking. The
encrypted-only fixture that pinned the poison shape now uses a native id
(the pass-through case it always meant to cover).
Part 1 of the cross-provider replay audit (Pi transform-messages
policy: builders emit only items valid for their target).
Verified: sampling-types 286 + sampler/chat-state/shell 5791 green.
Root cause of 400 {"detail":"System messages are not allowed"} at
chatgpt.com/backend-api/codex/responses (both platforms): the codex
adaptation covered only IDENTITY HEADERS (originator/OpenAI-Beta/UA/
chatgpt-account-id) — the BODY still carried the system prompt as
role:system input items, which the codex backend rejects outright. Its
system channel is the top-level field, and stateless
(store:false) reasoning replay requires
include:["reasoning.encrypted_content"] — both per the same reference
the headers were ported from (official Codex CLI + Pi's
api/openai-codex-responses.ts).
New adapt_body_for_codex_backend (kigi-sampling-types): hoists every
system input item into (order preserved, appended to any
existing instructions; string and parts content shapes) and requests
encrypted reasoning. Idempotent. Applied at both Responses send paths,
openai_codex-GATED — the API-key path stays byte-identical
(pinned by a control wire test).
Tests: adapter unit tests (hoist+include, no-system no-op, idempotence)
plus two mock-server wire tests (codex body has no system role,
instructions + include present; plain Responses body unchanged).
Verified: sampling-types + sampler + chat-state + shell 6076 tests
green, clippy clean.
Root cause of 'Invalid signature in thinking block' (400 at
messages.1.content.0, Claude models): build_messages_request replayed
EVERY stored Reasoning item as a thinking block with no origin check —
history synthesized by other backends (encrypted_content: None → the
mandatory signature field serialized as ""), Responses-API tco_* blobs
(signature bytes, no text), and blocks signed by a DIFFERENT model after
a mid-session /model switch. Anthropic validates every replayed
signature (model-bound), so such histories 400 deterministically. The
platform split was circumstantial: Windows sessions started on the
default model and switched to Claude; macOS sessions were Claude-native
from turn 1. Adversarially verified — no platform-divergent byte path
exists in capture, storage, or replay.
New prune_replayed_thinking pass (Pi/Claude Code replay policy): keep
exactly the final assistant message's thinking when its tool loop is
still open (request ends on the tool results — an open loop can never
span a model switch) and the block is genuinely signed; strip every
other thinking block (the API ignores valid prior-turn thinking and
rejects invalid). Assistant messages emptied by the strip (thinking-only
aborted turns) are removed — empty content arrays are rejected too.
Tests: three unit tests pin strip-outside-loop (unsigned, tco_*, stale
signed), keep-in-active-loop (verbatim text+signature at content.0), and
emptied-message removal; the legacy-upgrade integration test now proves
both wire fidelity in the active loop AND stripping once the loop
closes.
Verified: sampling-types + sampler + chat-state + shell 6072 tests
green, clippy clean.
The README always shipped graph engineering enabled; the enablement was
delegated to installer env plumbing that diverged per platform:
install.sh exported KIGI_GRAPH=1 into shell rc (worked), install.ps1
wrote the User registry variable — which running Windows terminals (and
new tabs of an open Windows Terminal) never pick up, so /graph was
'missing on Windows' despite a successful install.
With resolve_graph() defaulting true in the binary (e53a66d), the env
writes are redundant complexity: drop them from both installers, keep
KIGI_GRAPH=0 as the documented opt-out (env still beats the default),
and correct the flag comment to tell this story instead of a
'gray release' one. Both scripts syntax-checked (sh -n / pwsh parser).
Installers are served from main (raw.githubusercontent), so this takes
effect for all new installs immediately — no retag needed; the running
v0.1.5 build already carries the binary-default fix.
resolve_graph() read only the KIGI_GRAPH env var with default(false)
(plan.md G0 gate), so /graph existed solely on machines whose environment
exported the dev flag — which presented as '/graph is missing on
Windows'. There was never any platform-conditional code: the Mac worked
because the dev env var was set there.
Default is now true (matching /goal's shipped state); KIGI_GRAPH=0
remains the off-switch, and availability still requires the goal harness
(BuiltinGate::Graph). AGENTS.md updated; new test pins fresh-install-on
plus env-zero-off.
Verified: kigi-shell 5262 + kigi-tui 6874 tests green, clippy clean.
A single HTTP 502 from the GitHub release CDN killed the entire v0.1.5
tag build (one job of five, ~50 minutes wasted). The build-script
download now retries twice with backoff on 5xx/429/network errors;
genuine failures (404, offline) still fail fast with the
KIGI_SHELL_BUNDLE_RG_PATH hint, and every retry prints a cargo:warning
so flakiness stays observable.
Persist-failure ≠ switch-failure. The default_model rollback arm re-ran
set_default_model_inner(prev) AND issued a reverse SwitchModel whenever
the config.toml write failed — the only mechanism in the codebase that
deliberately re-shows the ORIGINAL model after a successful pick. On
Windows, where AV/indexer file locks routinely failed that write (until
27d009c), every /model selection appeared to not take.
The session switch succeeds independently and reports its own failures
via handle_switch_model_complete; a disk-persist failure now keeps the
live model, logs, and surfaces only the 'Could not save' toast — the
same policy PersistPreferredModel already ships ('still active for this
session').
Verified: kigi-tui 6874 tests green, clippy clean.
Root cause of 'model+effort switch works on Mac, not on Windows': the
switch APPLIES in-session (the dispatch/apply chain is platform-identical,
verified adversarially) but its persistence never sticks on Windows.
Every tmp+rename atomic write except auth/storage.rs committed with a
bare fs::rename, and Windows MoveFileExW(REPLACE_EXISTING) fails with a
sharing violation whenever AV/search-indexer/cloud-sync transiently holds
the destination open. Consequences: [models].default never persisted
(next launch = original model), the session summary's current model never
persisted (resume = original model), and the models cache went silently
stale (all its write errors were swallowed).
- New kigi_shell_base::util::fs::replace_file — THE commit step for
tmp+rename: plain rename on Unix; on Windows delete-first + two short
backoffs (the pattern auth/storage.rs shipped first), tmp cleaned on
failure, error always returned. Windows branch type-checked against
x86_64-pc-windows-msvc.
- Adopted at every replace site: config.toml (save_config /
atomic_write_string / mcp saves), models cache (plus unique tmp
suffixes and tracing::warn on failure — writes were fully silent),
session storage (summary/current-model, jsonl, plan/signals/
announcement/goal/graph state), auth.json, active-sessions registry,
prompt history, claude/kimi import, campaigns state, goal artifacts.
Directory-move renames (worktree pool, corrupt-file backups) keep
plain rename — their destinations don't pre-exist.
Verified: kigi-shell + kigi-shell-base 5318 tests green, clippy clean,
msvc-target check of the new cfg(windows) code clean.
The shell stamps meta.provider (platform display name via
parse_managed_model_key) on every managed {platform}/{model} catalog
entry in to_acp_model_info; user-defined [model.*] entries stay
provider-less. The /model dropdown surfaces it in the description column
(a model's own description still wins), so the picker reads as 'choose a
model from your connected providers'.
Verified: kigi-shell 5261 + kigi-tui 6873 tests green, clippy clean.
A required-args slash command submitted bare (/model, or /m) ran straight
into its usage error. The completeness contract already existed —
is_command_complete's documented 'Blocks' row — but had zero non-test
callers; both dispatch paths ran the command unconditionally.
Both the session prompt path and the dashboard dispatch box now consult
is_command_complete before running: an incomplete required-args command
re-opens the input as '/{command} ' with the cursor in the args phase, so
the existing suggestion dropdown lists the choices (for /model: the
catalog — by construction the connected providers' models, after
1bb10ef). Generalizes to every future required-args command; no new UI.
Verified: kigi-tui 6872 tests green, clippy clean.
Root cause of 'connected Claude, still shows Kimi / unknown model': every
fetch-plan decision consulted only the primary (Kimi) session and API keys
— never the stored subscription-OAuth sessions:
- on_auth_changed's wipe guard: a claude-pro-max-only login satisfied
'no session, no keys' → catalog wiped, early return BEFORE the fetch
and BEFORE notify_models_updated. The TUI kept an empty picker and the
prompt bar rendered 'unknown'. Guard decision extracted into the pure
should_wipe_catalog_on_auth_change (matrix-tested); a stored OAuth
session now vetoes the wipe, so the fetch runs, the first real catalog
reselects the default model (first entry = the connected provider's),
and kigi/models/update reaches the client.
- Startup prefetch: the arming gate ignored stored OAuth sessions and the
prefetch thread passed an empty token map — a claude-only user booted
onto the bundled Kimi table until a later refresh. The gate now takes
has_stored_oauth and the thread resolves each stored session's bearer
(refresh-on-expiry) via a current-thread runtime.
- Cache origins: from_config's startup cache load and cache_origin()
computed the fetch-plan origin with an empty token map, so a
claude-inclusive cached catalog never matched at startup. Both now use
presence-only stubs (stored_oauth_token_stubs — names only, no bearers)
proven equal to the real-token origin by test.
New probes in models_fetch: stored_oauth_platforms / stored_oauth_token_
stubs (sync auth.json scope scan; no AuthManager, no secrets).
Verified: kigi-shell 5260 tests green, clippy clean.
/login previously fired the resolved method's flow immediately — there was
no way to see providers or their status. It now lands on the provider
picker; already-connected providers show a green 'connected' badge in the
key column.
Shell: initialize() probes stored credentials once (primary session flag,
auth.json oauth/<provider> scopes, resolved platform keys env>auth.json>
config) and stamps _meta.connected on each advertised method
(connected_method_ids + stamp_connected_meta, pure and unit-tested).
Display state only — never an authorization input.
TUI:
- PendingMenuItem::connected() reads the badge from method meta; the
picker renders it green (accent_success), replacing the shortcut hint.
- New Action::OpenLoginPicker: /login shows the picker; mid-session it
stashes the view like dispatch_login, and starts no flow by itself.
- Mid-session picker: last row reads 'Cancel' and dispatches CancelLogin
(clicking it must not exit the app); Esc also returns to the session.
- After a successful login, the just-authenticated method is stamped
connected in the TUI's advertised-methods copy (auth_in_flight_method →
AuthComplete), so a later /login shows the badge without re-initialize.
Verified: kigi-shell 5256 + kigi-tui 6870 tests green, clippy clean.
Root cause: PendingMenuItem::Login carried only a label, so every provider
row collapsed to the id-less Action::Login, and dispatch_login resolved
the FIRST advertised interactive method — the Kimi device flow. Selecting
Grok/Claude/Copilot/Codex all opened kimi.com. (The shell side was already
correct: authenticate() dispatches each method id to its own OAuth flow.)
- PendingMenuItem::Login now carries the advertised method id (None only
on the no-interactive-method fallback row).
- New Action::LoginWith(AuthMethodId); the picker dispatches it with the
row's own id. Action::Login keeps its meaning (resolved/default method)
for /login, auto-login, and re-auth.
- dispatch_login_with resolves the id against the advertised methods and
FAILS CLOSED on an unknown id — no silent first-method fallback — then
adopts the method's label and start mode.
Tests: picker rows pinned to their method ids; LoginWith(claude-pro-max)
must authenticate with claude-pro-max even when kimi-code was previously
resolved; unknown ids surface an error and start nothing.
Verified: kigi-tui 6867 tests green, clippy clean.
Root cause (issues: unscrollable list + half-blocked moon): the stacked
welcome layout gave the menu one Length row per item. With ~30 advertised
providers the vertical layout overflowed and the constraint solver
squeezed every Length — clipping the logo — while rows past the fold were
silently unreachable.
- WelcomeLayout::compute_inner caps the menu to the rows genuinely left
over after logo/prompt/version, so the chrome never gets squeezed.
- render_menu is now a minimal-scroll viewport: it scrolls only when the
selection exits it (stable under hover — no re-centering feedback
loop), draws the picker-style scrollbar, and returns index-aligned hit
rects (zero Rect for off-screen rows) plus the offset, which AppView
feeds back next frame.
- Mouse wheel moves the login-picker selection (clamped, not wrapping).
Test: pending_menu_scrolls_and_never_clips_the_moon renders the full
method list at 40 rows and asserts the whole moon (10 braille rows), the
version badge, and that selecting Quit scrolls it into view.
Verified: kigi-tui 6864 tests green, clippy clean.
One authority answers 'which credential may ride this request':
credential_class / manager_for / credential_for / bearer_resolver_for,
keyed by (platform, base_url). SessionCredential is an opaque type with
no production constructor, so a new call site cannot re-introduce the
session-bearer leak. Platform-scoped tests extended across all bearer
channels (session, aux, summary, subagent override).
Verified: cargo check --workspace --all-targets clean; kigi-shell and
kigi-tui suites green (6611+ tests).
29th platform `openai-codex` (uses_oauth, Responses wire). PKCE-localhost login at
auth.openai.com (client app_EMoamEEZ73f0CkXaXp7hrann, redirect localhost:1455/auth/
callback, form token exchange, fresh-random state) reusing the claude-pro-max flow;
OAuthFlow::PkceLocalhost gained a redirect_path and OAuthConfig an authorize_extra
(empty elsewhere, so claude/xai/copilot authorize URLs stay byte-identical).
Codex-specific: the access token is a JWT carrying chatgpt_account_id, which becomes
the `chatgpt-account-id` inference header. It is derived STATELESSLY from whichever
bearer rides each request (so a rotated token needs no persisted field), and BOTH
login and refresh fail fast when the claim is absent — gated on the explicit
OAuthConfig.requires_chatgpt_account_id fact, never inferred from the token-body
encoding (a plain form endpoint is the OAuth norm and must not inherit this).
Inference rides the existing Responses wire at chatgpt.com/backend-api/codex →
/responses, with codex headers (chatgpt-account-id, originator, OpenAI-Beta
responses=experimental, codex UA) gated on SamplerConfig.openai_codex so API-key
`openai` stays byte-identical; store:false was already the global Responses default.
Catalog is HARDCODED (no live endpoint exists for this backend; read from the
official Codex CLI's model cache): gpt-5.6-sol/terra/luna + gpt-5.5, ctx 272000,
each with its real reasoning levels (low..ultra — ReasoningEffort gained Ultra).
Excluded: gpt-5.3-codex-spark (supported_in_api=false), gpt-5.4/-mini and
codex-auto-review (hidden) — they would list but fail at inference. The fetch
short-circuits before any HTTP; Kigi never shells out to the codex CLI or reads
~/.codex.
Security review fixes: redact any `account-id` header from request logs (it was
reaching debug logs), strict 3-segment JWT check (fail closed), refresh no longer
fails open on a missing claim. Inherits leak-safe pooled routing (scope
oauth/openai-codex) — never the Kimi token. Full gate green (234 suites, 0 warnings).
28th platform `github-copilot` (uses_oauth, ChatCompletions wire). Two-stage auth:
RFC-8628 GitHub device flow (client Iv1.b507a08c87ecfe98, scope read:user, errors
in a 200 body) mints the DURABLE github token; a GET api.github.com/copilot_internal/
v2/token exchange re-mints the SHORT-LIVED copilot token. Persisted as key=copilot
token, refresh_token=github token, expires_at=copilot expiry; the "refresh" is a
copilot-token re-mint (not a refresh_token grant), dispatched via
OAuthTokenBody::GithubCopilotExchange in the generic refresher.
VS Code editor-identity headers on /models + /chat/completions, gated on
SamplerConfig.github_copilot / PlatformId::sends_copilot_editor_headers() so every
other ChatCompletions provider stays byte-identical. Live /models filtered
(parse_github_copilot_listing) to the openai-completions-served models: keep iff
model_picker_enabled && policy.state!="disabled" && tool_calls!=false AND not a
claude-4.x/5.x (messages) or gpt-5/oswe/mai- (responses-only) id — those need
per-model wire routing (documented debt), excluded rather than mis-routed.
Inherits the leak-safe pooled routing (scope oauth/github-copilot); its bearer/
refresh/api_key never touch the Kimi token (regression test added). Fail-fast on
an out-of-range copilot expires_at (would otherwise silently 401 mid-session).
Adversarial security review: GO, no CRITICAL/HIGH. Known limitation: Pi's
per-model policy-enablement POST is not ported (documented in AGENTS.md).
27th registry variant, 2nd subscription-OAuth provider. Log in with a Claude
Pro/Max subscription via PKCE authorization-code + S256 (loopback callback on
127.0.0.1:53692, with a manual code-paste fallback), then use it against
api.anthropic.com — reusing the existing Anthropic Messages wire + Anthropic
listing + the multi-provider OAuth foundation (dbce6bf). Sourced from Pi
(earendil-works/pi auth/oauth/anthropic.ts): client 9d1c250a..., authorize
claude.ai/oauth/authorize, token platform.claude.com/v1/oauth/token, scope
'…user:inference user:sessions:claude_code…'.
New machinery (foundation handles token routing — claude-pro-max is a
uses_oauth platform so its bearer/refresh/api_key already route to its own
pooled manager, never Kimi):
- OAuthConfig gains flow{DeviceCode|PkceLocalhost} + token_host + token_body
{Form|JSON}; xai/kimi rows unchanged (DeviceCode/Form).
- auth/oauth_pkce.rs: PKCE S256 wire — loopback listener with STRICT state
validation (CSRF, fail-closed), manual-paste fallback, JSON code→token
exchange + rotating-refresh. Never logs code/verifier/tokens.
- Messages OAuth adaptation gated on SamplerConfig.anthropic_oauth (true only
for a claude-pro-max managed key): Authorization: Bearer + anthropic-beta
oauth + user-agent claude-cli + x-app cli, and the required 'You are Claude
Code' system prefix. API-key anthropic/minimax Messages requests are
BYTE-IDENTICAL (regression-guarded).
- Live /models under the OAuth Bearer + oauth-beta headers (Anthropic listing,
enriched from models.dev anthropic); persistent 401 → 0 models + WARN, NO
hardcoded fallback list (honest failure).
Adversarial review: no blocking findings (secret handling, CSRF/state, the
anthropic_oauth gate, token routing, non-regression all CONFIRMED). Full gate
green. Registry at 27; picker updated. Residual (unverifiable without a real
Claude Pro/Max account): whether GET /v1/models accepts the OAuth bearer, and
the real endpoint's acceptance of the OAuth Messages request.
First subscription-OAuth provider beyond Kimi Code (26th registry variant).
Log in with a Grok/SuperGrok/X subscription via RFC-8628 device-code OAuth
(auth.x.ai), then use it against api.x.ai/v1 — reusing the existing xai wire
(ChatCompletions + OpenAI listing + Passthrough + restrict + models_dev_id
xai). Sourced from Pi (earendil-works/pi auth/oauth/xai.ts): client
b1a00492..., scope 'openid profile email offline_access grok-cli:access
api:access', standard Bearer (no x-xai-token-auth).
Foundation (generalizes Kigi's Kimi-singleton OAuth to per-provider, root
cause, not a patch):
- Registry: OAuthConfig on PlatformSpec (client_id/host/device+token
paths/scope/scope_key); XAI_OAUTH_CONFIG + XAI_GROK_SPEC (uses_oauth, method
id 'xai-grok', an interactive login after kimi-code).
- Generic device-code wire (auth/oauth_device.rs) + GenericDeviceRefresher,
sharing the RFC-8628 core with Kimi; Kimi's bespoke flow is byte-identical
(X-Msh headers, KIMI_CODE_OAUTH_SCOPE, keyring gating unchanged).
- Per-provider AuthManager via a process-global pool (auth/oauth_registry.rs):
build-on-demand with start_proactive_refresh, keyed by scope. The session
resolves the AuthManager for the ACTIVE model's platform for bearer/refresh/
401-recovery/api_key — an oauth-platform model always uses its OWN token,
never the primary.
- Live /models under OAuth; base routes oauth().is_some() -> platform.base_url()
(kimi-code stays on proxy_url).
Security: adversarial review + a systematic token-leak audit found and closed
FIVE channels where the primary Kimi token could reach api.x.ai (bearer
resolver, api_key stamping, aux summary/classifier/image-describe models, and
subagent model-override). Each fix routes through the platform-aware resolver
(the oauth model's pooled token or None, NEVER the primary) and is revert-to-red
verified. No access/refresh token is ever logged.
Registry at 26; picker updated (xai-grok interactive login row); TUI
context-window already auto-updates per model. Full gate green (234 suites,
fmt, clippy -D warnings, deny). GPT/Claude/Grok officially permit third-party
subscription use.
Providers 21-22 (24th & 25th registry variants), sourced from Pi
(earendil-works/pi). Pi drives MiniMax through its Anthropic-COMPATIBLE
surface (baseUrl .../anthropic), so Kigi reuses the existing Anthropic
Messages machinery (wire_api=Messages, listing=Anthropic, key_header=XApiKey
x-api-key+anthropic-version) rather than the OpenAI path. Global:
api.minimax.io/anthropic, MINIMAX_API_KEY, models.dev minimax. China:
api.minimaxi.com/anthropic, MINIMAX_CN_API_KEY, models.dev minimax-cn.
The base carries the /v1 suffix (.../anthropic/v1) since Kigi appends bare
paths → listing .../anthropic/v1/models?limit=1000, inference
.../anthropic/v1/messages (matches the live-probed x-api-key-gated endpoint).
restrict_to_enriched=FALSE: the 7 MiniMax-M* models are clean (no pollution)
and restrict would drop launch-day models not yet in models.dev.
Also HARDENS parse_anthropic_listing to tolerate a bare array in addition to
the {data:[...]} envelope (mirrors parse_openai_listing's sniff that Together
taught us) — so MiniMax's Anthropic-compatible /models can't silently empty
the catalog if it serves a bare array. A bare object without data still errors.
Review found no defects (5 areas CONFIRMED incl. the /v1 non-doubling, the
additive parser change, restrict=false rationale). Residual (logged): the live
200 body / anthropic-version acceptance is unverifiable without a key; the
bare-array tolerance + restrict=false hedge most shapes.
Tests: e2e mocks the x-api-key-gated Anthropic listing (proves the auth header
+ enrichment + keying under minimax/); both validation tests reject 401 with
the per-variant console host; new parser test covers envelope + bare array +
bare-object-errors. Registry at 25; picker 26 rows.
Providers 19-20 (22nd & 23rd registry variants), sourced from Pi
(earendil-works/pi). Pi's xiaomi.ts / xiaomi-token-plan-cn.ts use plain
openAICompletionsApi (no thinking dialect) → Kigi Passthrough. Global: MiMo
at api.xiaomimimo.com/v1, XIAOMI_API_KEY, models.dev xiaomi. China token plan:
token-plan-cn.xiaomimimo.com/v1, XIAOMI_TOKEN_PLAN_CN_API_KEY, models.dev
xiaomi-token-plan-cn. Both Bearer + OpenAI listing + ChatCompletions +
restrict_to_enriched; /models auth-gated → validator.
id-match proven vs Pi's static ids (mimo-v2.5-pro, mimo-v2-flash, ...). The
CN token plan lists 4 mimo TTS models (tool_call=false); restrict's tool_call
cut drops them, keeping only the 3 chat models — the e2e proves this
non-vacuously with the real mimo-v2-tts. Review found no defects.
Registry at 23; picker 24 rows (welcome picker verified rendering all rows at
a taller viewport).
Providers 17-18 (20th & 21st registry variants), sourced authoritatively
from Pi (earendil-works/pi), the open-source agent whose provider list is
being mirrored. Pi's zai.ts / zai-coding-cn.ts use plain openAICompletionsApi
(NO special thinking dialect — overturns the matrix's 'thinking:{type} → new
dialect' concern), so Kigi maps them to Passthrough. Global: api.z.ai/api/
coding/paas/v4, ZAI_API_KEY, models.dev zai-coding-plan. China (Zhipu
BigModel): open.bigmodel.cn/api/coding/paas/v4, ZAI_CODING_CN_API_KEY,
models.dev zhipuai-coding-plan. Both Bearer + OpenAI listing + ChatCompletions
+ restrict_to_enriched; /models is auth-gated → validator.
id-match is PROVEN (not just assumed like Qwen): Pi's static model ids
[glm-4.5-air, glm-4.7, glm-5-turbo, glm-5.1, glm-5.2, glm-5v-turbo] are
byte-identical to the models.dev zai-coding-plan keys, all tool_call=true, so
restrict keeps every model with no silent-empty risk. Review found no defects.
GLM thinking is not lost (reasoning_content is parsed regardless of dialect).
Tests: e2e proves enrichment-supplied context + non-vacuous restrict (a
non-enriched wire model is dropped) + Passthrough; both variants' validation
tests hit /models and assert the per-variant console host (z.ai vs
open.bigmodel.cn). Registry at 21; picker 22 rows.
Also fixes the welcome login-picker test for the now-taller menu (renders at
a taller viewport to verify content coverage) and logs the real menu-overflow
UX debt: the picker clips rows past the fold with no scroll (q/l shortcuts
still work; only shown when unauthenticated) — deferred to its own cycle.
Provider 16 (19th registry variant). Same endpoint + models + Kimi dialect
as the existing OAuth kimi-code platform (api.kimi.com/coding/v1 via the
KIGI_CODE_BASE_URL override), but authenticated with a static KIMI_API_KEY
instead of the device flow — for users who have a Kimi For Coding key rather
than an OAuth subscription. Bearer, OpenAI listing, ChatCompletions,
ChatCompat::Kimi, wire_serves_metadata=true (Kimi /models self-serves
context/thinking), restrict_to_enriched=false (clean 3-model catalog).
/coding/v1/models is auth-gated (401) so it doubles as the validator.
No collision: KIMI_API_KEY was previously unused (grep-verified), and the
house BYOK reads only KIGI_API_KEY/XAI_API_KEY/legacy. kimi-code (OAuth) and
kimi-coding (static key) are independently gated (OAuth-token vs key) and
their models get distinct managed keys (kimi-code/k3 vs kimi-coding/k3) — a
user with both simply sees each Kimi model twice; no dedup collision, no crash.
Review (6 areas): no blocking defects; confirmed the spec correctly mirrors
KIMI_CODE_SPEC (differing only in uses_oauth/api_key_envs/console_host/labels)
and the coexistence is benign. Strengthened the e2e's dialect assertion (Kimi
is the default ChatCompat, so it did not discriminate a parse failure) by
also asserting parse_managed_model_key attributes the key to KimiCoding.
Tests: e2e proves wire-served context (1_048_576 from the wire) with the
models.dev fetch SKIPPED (all-wire-metadata provider, .expect(0)), bare-id
round-trip under kimi-coding/, Kimi dialect; validation test rejects a 401
from /models. Registry at 19; picker 20 rows; snapshot already bundles
kimi-for-coding.
Providers 14-15 (17th & 18th registry variants), Alibaba DashScope
compatible-mode. Global: token-plan.ap-southeast-1.maas.aliyuncs.com,
QWEN_TOKEN_PLAN_API_KEY, models.dev alibaba-token-plan. China:
token-plan.cn-beijing.maas.aliyuncs.com, QWEN_TOKEN_PLAN_CN_API_KEY,
alibaba-token-plan-cn. Both Bearer + OpenAI listing + ChatCompletions +
Passthrough (stream_options.include_usage is documented-supported).
/models is auth-gated (401 without a key) so it doubles as the validator;
metadata from models.dev enrichment. The token plan is a multi-vendor
catalog (deepseek/kimi/minimax/glm/qwen); restrict_to_enriched keeps the 15
tool-calling chat models and drops the 4 qwen-image/wan image generators.
Review downgraded two flagged concerns: the enable_thinking non-streaming
400 cannot occur (Kigi never issues non-streaming ChatCompletions in
production — all inference streams), and Qwen thinking is NOT invisible
(reasoning_content is parsed regardless of the Passthrough dialect). No
defects. Residual (logged): restrict does an exact id-match of live /models
ids vs the models.dev keys; a mismatch fails safe (0 models) — verify with a
real key. Snapshot already bundles both providers (regenerated in ebf1105).
Tests: global e2e proves enrichment-supplied context (wire carries none),
non-vacuous tool_call restriction (qwen-image dropped), bare-id round-trip,
Passthrough; both variants' validation tests hit /models (401 reject) and
assert the correct per-variant console host. Registry at 18; picker 19 rows.
13th provider (16th registry variant). Also reconciles a naming collision
the matrix flagged: this fork is house-branded "xai" (cf. xai.dev metadata,
KIGI_CODE_XAI_API_KEY legacy env), so XAI_API_KEY + method xai.api_key were
the GENERIC house BYOK, not x.ai/Grok. The provider table wants xai/XAI_API_KEY
for Grok.
Resolution (user-approved): XAI_API_KEY now keys the x.ai/Grok provider; the
house BYOK primary env moves to KIGI_API_KEY, keeping XAI_API_KEY and
KIGI_CODE_XAI_API_KEY as back-compat fallbacks (read_xai_api_key_env checks
KIGI_API_KEY first). The xai.api_key method id is unchanged (persisted-session
compat); the platform method id is the bare "xai", distinct from it.
xAI spec: api.x.ai/v1, Bearer, OpenAI listing + ChatCompletions, Passthrough
(docs confirm stream_options.include_usage accepted). /v1/models is minimal
(ids only) and requires auth, so it doubles as the key validator (401 on bad
key, no override) and metadata comes from models.dev enrichment. Live ids
match the models.dev "xai" keys byte-for-byte, so restrict_to_enriched keeps
the 5 tool-calling chat models (grok-4.5/4.3/4.20-0309-*/build-0.1) and drops
the grok-imagine-* generators + the non-tool multi-agent model. Snapshot
regenerated to include the xai provider (was stale; gen script already listed
it in TARGETS).
Env migration is comprehensive to avoid keying the xai platform (which would
trigger a live api.x.ai fetch) or leaving house-key reads stranded: routed
the trace CLI resolver + acp_agent/auth.json bridge + paste-key ext handler
through the new primary; moved all leader/pager/e2e harness setters to
KIGI_API_KEY; made every house-key isolation test unset KIGI_API_KEY too;
updated user-facing hints to name KIGI_API_KEY.
Tests: e2e proves enrichment-supplied context (wire carries none), non-vacuous
tool_call restriction, bare-id round-trip under xai/, Passthrough; validation
tests hit /models (401 reject, 200 accept); house_env_var_takes_precedence_over_xai
pins the new precedence. Registry at 16; picker 17 rows; 4 auth arrays + xai.
Review (16 findings, all fixed): caught a missed else-branch env clear in the
paste-key handler (would leak the house key past a clear) and a non-hermetic
credential-priority test; both fixed.
12th provider. API-key via AI_GATEWAY_API_KEY, Bearer, OpenAI listing +
ChatCompletions, Passthrough dialect. Second wire-metadata provider but
takes the enrichment path instead: Vercel serves context under
context_window, which WireModel ignores (reads context_length), so
wire_serves_metadata=false + restrict_to_enriched pulls context/limits
from the models.dev "vercel" snapshot (302/306 live ids match snapshot
keys byte-for-byte, so restrict keeps essentially the whole catalog).
/models is public (200 for any key), so login validation targets
/credits (key_validation_path) which 401s on a bad bearer — avoids
false-accepting invalid keys against the public listing.
Tests: e2e proves enrichment-wins (wire context_window=999 distinct from
enrichment context=400000, asserts 400000) and non-vacuous tool_call
restriction; validation test proves /credits (not /models) is hit.
Registry at 15 (ordinal/VARIANT_COUNT/ALL), 4 auth arrays + 16-row picker.
The 14th registry row: id "nvidia", NVIDIA_API_KEY > auth.json "nvidia"
scope, https://integrate.api.nvidia.com/v1 with KIGI_NVIDIA_BASE_URL
override, Bearer, ChatCompletions, enrichment-backed metadata
(models_dev_id nvidia — the quirk matrix's earlier 'absent' claim was
wrong; models.dev has 84 nvidia models), restrict_to_enriched=true (the
NIM listing mixes chat/embedding/rerank/vision/image; keep the 45
tool-calling chat models). Slashed org/model ids
(nvidia/meta/llama-3.3-70b-instruct) round-trip via the first-slash split;
the native id rides the wire.
NIM exposes raw vLLM behavior and stream_options support varies per model
(some strict vLLM backends 4xx on it), so chat_compat=StrictOpenAi strips
stream_options — streaming works across the whole fleet, usage falls back
to estimation. Snapshot reasoning models carry no effort menus, so kigi
sends no reasoning_effort (which an unsupported strict validator would
400 on). Review: no defects. Logged note: a key lacking the org 'Public
API Endpoints' permission passes /models validation but 403s on chat
(user-fixable edge case).
The 13th registry row: id "cerebras", CEREBRAS_API_KEY > auth.json
"cerebras" scope, https://api.cerebras.ai/v1 with KIGI_CEREBRAS_BASE_URL
override, Bearer, ChatCompletions, enrichment-backed metadata
(models_dev_id cerebras).
Cerebras' catalog is all chat LLMs (no embedding/tts pollution) and its
/models is minimal (ids only), so restrict_to_enriched=FALSE: keep every
live model, enrich the known ones (context + effort menus low/medium/high),
unknown ones keep the default context. The e2e pins this enrich-without-
restrict path (new — prior enrichment providers all used restrict=true).
Review caught a likely-DOA defect: Cerebras uses strict
additionalProperties:false validation (confirmed 400-rejecting store,
maxTokens, thinking, nested reasoning_content), and stream_options is not
in its schema — so Passthrough (which keeps the stream_options.include_usage
kigi injects on every streaming request) would very likely 400 all
streaming. Generalized ChatCompat::Mistral -> ChatCompat::StrictOpenAi
(serde alias "mistral" keeps pre-rename persisted sessions loading), which
strips stream_options + private fields for any strict OpenAI-compat
validator; both Mistral and Cerebras now map to it. Future strict-validator
candidates (NVIDIA/Azure/Xiaomi/OpenCode) noted for the same check.
reasoning_effort (incl. "none") passes through; /v1/models requires auth
so key validation works; console cloud.cerebras.ai.
The 12th registry row: id "together", TOGETHER_API_KEY > auth.json
"together" scope, https://api.together.xyz/v1 with KIGI_TOGETHER_BASE_URL
override, Bearer, ChatCompletions + Passthrough, enrichment-backed
metadata (models_dev_id togetherai) with the tool-calling listing
restriction (Together's listing mixes chat/embedding/rerank/image types).
Together's GET /v1/models returns a BARE JSON ARRAY, not the OpenAI
{object:list,data:[]} envelope. New shared parser parse_openai_listing
tolerates both shapes (sniffs the top-level [ vs { for accurate
diagnostics); every OpenAI-listing provider now routes through it, with
byte-equivalent envelope behavior (verified: Groq/Google/OpenRouter
unchanged) and non-silent errors. Together's org/Model ids match the
togetherai snapshot keys exactly (verified), so restrict_to_enriched keeps
the ~27 tool-calling models without the id-shape trap.
Review: ship-ready, shared-parser change proven strictly-additive and safe,
registry/e2e/dialect all correct. Backlog logged: kigi ignores the wire
per-model field, so a live Together chat model absent from
models.dev is dropped until indexed — a future wire- filter would
unlock the fresh full catalog.
The 11th registry row and the first THIRD-PARTY wire_serves_metadata=true
provider: id "openrouter", OPENROUTER_API_KEY > auth.json "openrouter"
scope, https://openrouter.ai/api/v1 (note /api/v1) with
KIGI_OPENROUTER_BASE_URL override, Bearer, OpenAI listing +
ChatCompletions + Passthrough.
OpenRouter's public /models serves context_length for every model
(verified live: 340/340), so it needs NO enrichment: models_dev_id=None,
wire_serves_metadata=true, restrict_to_enriched=false. An OpenRouter-only
user makes zero models.dev calls; context comes straight from the listing.
Slashed ids (anthropic/claude-opus-4.8) round-trip through the managed key
via the first-slash split; the native id rides the wire. The e2e pins all
of this with the models.dev refresh disabled.
Review-confirmed defect fixed (and independently re-verified with live
curls): OpenRouter's /models is PUBLIC — GET /models returns 200 for ANY
key — so login key-validation would false-accept a bad key, deferring the
failure to the first chat 401. New spec field key_validation_path lets a
public-listing platform validate against an auth-requiring endpoint;
OpenRouter uses /key (401s for bad keys). Reusable for Vercel (also
public). Tests pin the /key validation and no regression to the default
/models path.
Gate caught a fixture regression: the kimi_import test used openrouter.ai
to represent a CUSTOM provider, which now correctly dedupes to the builtin
OpenRouter — moved the fixture to a reserved llm.example.test host that no
future platform can shadow.
The 10th registry row: id "google", GEMINI_API_KEY > auth.json "google"
scope, https://generativelanguage.googleapis.com/v1beta/openai (Gemini's
OpenAI-compatibility shim) with KIGI_GOOGLE_BASE_URL override, Bearer,
OpenAI listing + ChatCompletions + Passthrough, enrichment-backed
metadata (models_dev_id google), tool-calling listing restriction.
Review caught a ship-blocking defect: Gemini's compat /models returns
-PREFIXED ids (confirmed via Google's own cookbook), but the
models.dev snapshot keys and the chat endpoint use the BARE id. Without
normalization the enrichment lookup misses and restrict_to_enriched
silently empties the Gemini catalog (login works, zero models
selectable). A doc WebFetch had hidden this — the compat docs show bare
INPUT ids to retrieve/chat but never print the list OUTPUT.
Fix: new spec field strip_listing_id_prefix, applied in the fetch before
filter/enrich/keying. Google sets Some("models/") (defensive: a no-op if
an id is already bare, so correct regardless of the live shape); all other
rows None. The e2e now feeds the REAL prefixed listing ids and asserts
they survive as the bare managed key google/gemini-2.5-pro with the bare
id on the wire (chat rejects the prefix); a registry test pins the config.
Note: models.dev models Gemini reasoning as budget_tokens (not
effort-type), so no auto effort-menu — models still fully work; reasoning
is dynamic. Gemini compat applies default safety filters (no BLOCK_NONE).
The 9th registry row, pure Groq pattern: id "fireworks",
FIREWORKS_API_KEY > auth.json "fireworks" scope,
https://api.fireworks.ai/inference/v1 (note /inference/v1) with
KIGI_FIREWORKS_BASE_URL override, OpenAI listing + ChatCompletions +
Passthrough dialect, enrichment-backed metadata (models_dev_id
fireworks-ai) with the tool-calling listing restriction.
Fireworks native ids are deeply slashed (accounts/fireworks/models/glm-5p2);
the e2e pins the full round-trip: the managed key
(fireworks/accounts/fireworks/models/glm-5p2) parses back on the first
slash, and — the 404-risk property — the NATIVE id rides the inference
wire (entry.model and the resolved SamplerConfig.model) while the
fireworks/ prefix stays internal routing only.
Review (models.dev provider.toml + Fireworks docs): row facts confirmed,
registry integrity at 9, counts complete, e2e strong on all axes, zero
defects. One tradeoff logged as debt: restrict_to_enriched drops
fine-tuned/account-scoped deployed models (a headline Fireworks feature)
that can never be in models.dev.
The 8th registry row: id "mistral", MISTRAL_API_KEY > auth.json
"mistral" scope, https://api.mistral.ai/v1 with KIGI_MISTRAL_BASE_URL
override, enrichment-backed metadata with the tool-calling listing
restriction (embed/moderation/OCR noise).
Mistral is NOT a pure-pattern provider — an adversarial review found two
doc-confirmed blockers that no test exercises (no e2e covers a chat POST),
so the gate-green registry row alone would have shipped it DOA. A research
workflow pinned the exact wire shapes against the mistralai/client-python
SDK source (adversarially verified), then both were fixed:
1. stream_options 422: Mistral's strict Pydantic validator rejects the
stream_options.include_usage field kigi injects on every streaming
request (the SDK's request model has no such field). New
ChatCompat::Mistral dialect strips it (plus the kigi-private message
fields, like Passthrough). Streaming usage falls back to token
estimation.
2. Reasoning content arrays: Mistral reasoning models return content as
Union[str, List[ContentChunk]] on both streaming and non-streaming,
which the flat Option<String> path could not decode -> aborted turn.
A UNIVERSAL lenient deserializer (#[serde(from = "Raw..")] on
ChatResponseMessage + ChatChunkDelta) accepts string-or-array, routing
{type:text} chunks to the answer and the nested text of {type:thinking}
chunks to reasoning_content, tolerant of the OPEN chunk union (unknown
types ignored, never fatal). String content stays byte-identical for
every other provider (kimi/deepseek/groq/BYOK).
Review refuted all seven attack lines (no regression, no crash, exhaustive)
and flagged one coverage gap, now closed: a stream-consumer integration
test drives a full thinking -> transition -> answer chunk sequence and
proves it yields the same reasoning-sibling + assistant-answer result as
the reasoning_content string path.
Also folds a verified quirk matrix for all 23 remaining API providers into
providers-plan.md, tiered by real difficulty (self-enriching OpenRouter/
Vercel; bare-array Together listing; Messages-dialect MiniMax reusing the
Anthropic machinery; non-Bearer Azure/Bedrock; router wildcards; the OAuth
block).
The 7th registry row and the first pure-pattern cycle: id "groq",
GROQ_API_KEY > auth.json "groq" scope, https://api.groq.com/openai/v1
with KIGI_GROQ_BASE_URL override, OpenAI listing + ChatCompletions +
Passthrough dialect (Groq accepts the OpenAI-style reasoning_effort
scalar verbatim), enrichment-backed metadata with the tool-calling
restriction (the listing carries whisper/tts/guard noise — 8 of 15
enrichment entries are non-chat).
Review verdict: faithful pattern repeat, zero blocking findings — row
facts verified against live Groq docs, registry integrity at 7, e2e
proven strong on both axes (restriction + dialect mapping). Added the
recommended pin: managed keys split on the FIRST slash, so Groq's
provider-native slashed ids (openai/gpt-oss-120b and 10 more) round-trip
as groq/openai/gpt-oss-120b → (Groq, openai/gpt-oss-120b).
Also fixes a pre-existing test race this cycle surfaced: the enterprise-
endpoints test asserted moonshot's fixed base while non-serial, racing
serial tests that legitimately point KIGI_MOONSHOT_CN_BASE_URL at
wiremock; now serial + env-unset like its documented siblings. Two
display-only advisories logged as tracked debt (Groq's delta.reasoning
field invisible in the TUI; platform-generic rate-limit copy).
ZacharyZhang-NY