release: v0.1.7

Since v0.1.6 — the cross-provider replay audit (Pi transform-messages
policy: every wire builder emits only items valid for its target):
- fix(responses): reasoning items without a native rs_* id dropped (the
  GPT/codex 400 Invalid 'input[N].id'); provenance gate drops foreign
  reasoning and demotes foreign backend tool calls on model switches
- fix(codex): bare rs_* reasoning references dropped (stateless backend)
- fix(wire): shared ASCII tool-call id sanitizer, symmetric call+result
  on both the Messages and Responses legs
- fix(messages): image-source whitelist (raster base64 only, http(s)
  urls only) and empty-user-content guard
- fix(conversation): char-boundary-safe code preview (CJK/emoji code
  panicked every request build)
- docs: AGENTS.md records the replay policy

AGENTS.md

@@ -220,7 +220,34 @@ edges stay deterministic Rust. The harness appends a terminal
       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.
+      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".
+- CROSS-PROVIDER REPLAY POLICY (Pi `transform-messages` pattern): the
+  conversation history is provider-agnostic and sessions switch
+  models/backends mid-history, so EACH wire builder owns emitting only
+  items valid for its target — never patch downstream except in the
+  per-backend body adapters. Concretely: the Responses input drops
+  Reasoning items without a native `rs_*` id (foreign capture is id "")
+  and provenance-gates whole turns via `transform_items_for_responses`
+  (`AssistantItem.model_id` vs the request model: foreign Reasoning
+  dropped, foreign BackendToolCall demoted to its `text_summary`);
+  the codex adapter additionally drops bare `rs_*` references (stateless
+  backend); tool-call ids pass through ONE shared ASCII
+  `sanitize_tool_call_id` symmetrically on call+result on BOTH the
+  Messages and Responses legs; Messages image sources go through
+  `parse_base64_image_data_uri` (raster whitelist, no `data:` url
+  sources) and empty user turns get a placeholder. Dangling tool calls
+  are already repaired item-level by `repair_dangling_tool_calls` on the
+  actor's build path. When a provider wire bug surfaces, fix the CLASS
+  across all three builders in the same pass — three sequential
+  single-provider fixes (thinking signature → codex system role → codex
+  reasoning id) motivated this policy.
     - `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
@@ -237,9 +264,15 @@ edges stay deterministic Rust. The harness appends a terminal
       `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. API-key `openai` Responses requests carry
+      is the shared Responses default. BODY adaptation
-      NONE of this (byte-identical). `reasoning.effort` carries the thinking
+      (`adapt_body_for_codex_backend`, same gate): the backend 400s
-      level (incl. the codex-only `ultra`). NO websocket, NO base_instructions.
+      `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` &&

Cargo.lock

@@ -5442,7 +5442,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-acp-lib"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "async-trait",
@@ -5456,7 +5456,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-agent"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -5486,7 +5486,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-agent-lifecycle"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "async-trait",
  "tokio",
@@ -5495,7 +5495,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-auth"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "async-trait",
  "http 1.4.2",
@@ -5508,7 +5508,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-bin"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "clap",
@@ -5543,7 +5543,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-chat-state"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "indexmap",
  "kigi-compaction",
@@ -5560,7 +5560,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-codebase-graph"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "ahash",
  "clap",
@@ -5596,7 +5596,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-compaction"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -5609,7 +5609,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-config"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "base64",
  "blake3",
@@ -5632,7 +5632,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-config-types"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "indexmap",
@@ -5646,7 +5646,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-crash-handler"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "backtrace",
  "libc",
@@ -5657,7 +5657,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-env"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "tracing",
  "url",
@@ -5665,7 +5665,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-fast-worktree"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "bytes",
@@ -5697,7 +5697,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-file-utils"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "aws-config",
@@ -5721,7 +5721,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-fsnotify"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "criterion",
  "dunce",
@@ -5742,7 +5742,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-gix-status"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "gix",
  "kigi-test-utils",
@@ -5752,7 +5752,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hooks"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "fastrand",
  "kigi-config",
@@ -5771,7 +5771,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hooks-plugins-types"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "serde",
  "serde_json",
@@ -5779,7 +5779,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-http"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "kigi-auth",
  "kigi-log",
@@ -5794,7 +5794,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-hunk-tracker"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "chrono",
  "dunce",
@@ -5815,14 +5815,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-interjection-core"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "serde",
 ]
 
 [[package]]
 name = "kigi-log"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "chrono",
@@ -5840,7 +5840,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-markdown"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anstyle",
  "anstyle-lossy",
@@ -5864,14 +5864,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-markdown-core"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "pulldown-cmark",
 ]
 
 [[package]]
 name = "kigi-mcp"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "async-trait",
@@ -5908,7 +5908,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-memory"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "arc-swap",
@@ -5942,7 +5942,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-mermaid"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "fontdb",
  "image",
@@ -5960,7 +5960,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-models"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "kigi-env",
  "serde",
@@ -5970,7 +5970,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-pager-minimal"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "chrono",
  "crossterm",
@@ -5987,7 +5987,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-pager-pty-harness"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "alacritty_terminal",
  "anyhow",
@@ -6012,7 +6012,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-pager-render"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "anstyle",
@@ -6064,7 +6064,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-paths"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "camino",
  "serde",
@@ -6074,7 +6074,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-prompt-queue"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "serde",
  "serde_json",
@@ -6082,7 +6082,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-proto-build"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "pbjson-build",
@@ -6093,7 +6093,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-ratatui-inline"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "ansi-width",
  "anstyle-parse 0.2.7",
@@ -6110,7 +6110,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-ratatui-textarea"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "arboard",
  "chrono",
@@ -6131,7 +6131,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sampler"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "async-openai",
  "async-stream",
@@ -6154,7 +6154,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sampling-types"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "assert_matches",
  "async-openai",
@@ -6171,7 +6171,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sandbox"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "chrono",
@@ -6192,7 +6192,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-secrets"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "regex",
  "serde_json",
@@ -6230,7 +6230,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-shell"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6367,7 +6367,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-shell-base"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "chrono",
@@ -6392,7 +6392,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-sqlite-journal"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "libc",
  "rusqlite",
@@ -6403,7 +6403,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-subagent-resolution"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "kigi-sampling-types",
  "kigi-tool-types",
@@ -6418,7 +6418,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-system-power"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "windows-sys 0.59.0",
  "zbus",
@@ -6426,7 +6426,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-test-support"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6448,7 +6448,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-test-utils"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "runfiles",
  "tracing",
@@ -6457,11 +6457,11 @@ dependencies = [
 
 [[package]]
 name = "kigi-token-estimation"
-version = "0.1.5"
+version = "0.1.7"
 
 [[package]]
 name = "kigi-tool-protocol"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "kigi-tool-types",
  "serde",
@@ -6472,7 +6472,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tool-runtime"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -6490,7 +6490,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tool-types"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "minijinja",
  "schemars 1.2.1",
@@ -6500,7 +6500,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tools"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "arc-swap",
@@ -6577,7 +6577,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tools-api"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "kigi-proto-build",
  "kigi-tool-protocol",
@@ -6590,11 +6590,11 @@ dependencies = [
 
 [[package]]
 name = "kigi-tracing-macros"
-version = "0.1.5"
+version = "0.1.7"
 
 [[package]]
 name = "kigi-tty-utils"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "libc",
  "nix 0.30.1",
@@ -6604,7 +6604,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-tui"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "ansi-to-tui",
@@ -6691,7 +6691,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-update"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "dunce",
@@ -6720,14 +6720,14 @@ dependencies = [
 
 [[package]]
 name = "kigi-version"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "semver",
 ]
 
 [[package]]
 name = "kigi-workspace"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "agent-client-protocol",
  "anyhow",
@@ -6806,7 +6806,7 @@ dependencies = [
 
 [[package]]
 name = "kigi-workspace-types"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "base64",
  "chrono",
@@ -8840,7 +8840,7 @@ dependencies = [
 
 [[package]]
 name = "ptyctl"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "alacritty_terminal",
  "anyhow",
@@ -8858,7 +8858,7 @@ dependencies = [
 
 [[package]]
 name = "ptyctl-cli"
-version = "0.1.5"
+version = "0.1.7"
 dependencies = [
  "anyhow",
  "axum",

Cargo.toml

@@ -76,7 +76,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.1.5"
+version = "0.1.7"
 edition = "2024"
 license = "Apache-2.0"
 

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

@@ -1183,6 +1183,9 @@ impl SamplingClient {
         // 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| {
@@ -1321,6 +1324,9 @@ 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

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

@@ -319,15 +319,15 @@ impl BackendToolCallItem {
                 format!("[backend x_search] {}({})", ct.name, ct.input)
             }
             BackendToolKind::CodeInterpreter(ci) => {
+                // Char-boundary-safe preview: a byte slice (`&c[..100]`)
+                // panicked on CJK/emoji code, crashing every subsequent
+                // request build on every backend.
                 let code_preview = ci
                     .code
                     .as_deref()
-                    .map(|c| {
+                    .map(|c| match c.char_indices().nth(100) {
-                        if c.len() > 100 {
+                        Some((byte_idx, _)) => format!("{}...", &c[..byte_idx]),
-                            format!("{}...", &c[..100])
+                        None => c.to_string(),
-                        } else {
-                            c.to_string()
-                        }
                     })
                     .unwrap_or_default();
                 format!("[backend code_interpreter] {code_preview}")
@@ -2188,14 +2188,86 @@ impl From<&ConversationRequest> for rs::CreateResponse {
 /// so they appear inline in the same order the model originally emitted —
 /// which is what lets the server-side prefix KV-cache hit on repeat turns.
 fn build_responses_input(req: &ConversationRequest) -> rs::InputParam {
-    let items: Vec<rs::InputItem> = req
+    let transformed = transform_items_for_responses(&req.items, req.model.as_deref());
-        .items
+    let items: Vec<rs::InputItem> = transformed
         .iter()
         .flat_map(conversation_item_to_input_items)
         .collect();
     rs::InputParam::Items(items)
 }
 
+/// Provenance gate for the Responses input (the Pi `transform-messages`
+/// pattern): opaque provider-issued items replay verbatim only when the
+/// turn that produced them ran on the SAME model this request targets.
+///
+/// Each `[Reasoning | BackendToolCall]* Assistant` run carries its
+/// provenance in `AssistantItem::model_id`. On a confirmed mismatch
+/// (both sides known, different — a mid-session `/model` switch or a
+/// cross-backend history):
+/// - `Reasoning` siblings are DROPPED — their encrypted payloads are
+///   scoped to the issuing model (OpenAI documents encrypted content as
+///   model-bound; foreign backends' blobs are undecryptable outright);
+/// - `BackendToolCall` items are DEMOTED to a synthetic assistant text
+///   summary — exactly the downgrade the Messages and ChatCompletions
+///   builders already perform — because their typed shapes carry
+///   provider-issued ids and tool names the target never declared
+///   (e.g. a grok `x_search` CustomToolCall replayed to codex).
+///
+/// Same-model runs, unknown provenance (pre-provenance histories with
+/// `model_id: None`), and trailing orphans pass through verbatim — that
+/// byte-stability is what lets the server-side prefix KV-cache hit.
+fn transform_items_for_responses(
+    items: &[ConversationItem],
+    target_model: Option<&str>,
+) -> Vec<ConversationItem> {
+    let Some(target) = target_model else {
+        return items.to_vec();
+    };
+    let mut out: Vec<ConversationItem> = Vec::with_capacity(items.len());
+    // Pending run of opaque siblings awaiting their Assistant carrier.
+    let mut run_start: usize = 0;
+    for item in items {
+        match item {
+            ConversationItem::Reasoning(_) | ConversationItem::BackendToolCall(_) => {
+                out.push(item.clone());
+            }
+            ConversationItem::Assistant(a) => {
+                let foreign = a
+                    .model_id
+                    .as_deref()
+                    .is_some_and(|producer| producer != target);
+                if foreign {
+                    // Rewrite the pending run in place.
+                    let mut rewritten: Vec<ConversationItem> = Vec::new();
+                    for pending in out.drain(run_start..) {
+                        match pending {
+                            ConversationItem::Reasoning(_) => {}
+                            ConversationItem::BackendToolCall(b) => {
+                                rewritten.push(ConversationItem::Assistant(AssistantItem {
+                                    content: b.text_summary().into(),
+                                    tool_calls: vec![],
+                                    model_id: a.model_id.clone(),
+                                    model_fingerprint: None,
+                                    reasoning_effort: None,
+                                }));
+                            }
+                            other => rewritten.push(other),
+                        }
+                    }
+                    out.extend(rewritten);
+                }
+                out.push(item.clone());
+                run_start = out.len();
+            }
+            other => {
+                out.push(other.clone());
+                run_start = out.len();
+            }
+        }
+    }
+    out
+}
+
 /// Walk a serialized Responses API request body and inject the
 /// `type: "reasoning_text"` discriminator that the API requires on
 /// `reasoning.content[*]` items.
@@ -2251,9 +2323,23 @@ fn conversation_item_to_input_items(item: &ConversationItem) -> Vec<rs::InputIte
         }
         ConversationItem::Reasoning(r) => {
             // Reasoning items round-trip back to the Responses API in their
-            // native typed form. `status` is output-only (the API rejects it
+            // native typed form — but ONLY items the Responses API itself
-            // on input), so strip it before emission; everything else
+            // produced. A native item always carries a server-issued `rs_*`
-            // (summary, content, encrypted_content, id) passes through.
+            // id; an EMPTY id marks a foreign item (Messages capture stores
+            // the Anthropic signature with id "", chat-completions and the
+            // stream-delta fallback synthesize with id "", legacy upgraders
+            // reconstruct with id ""), and the API rejects it outright:
+            // 400 "Invalid 'input[N].id': ''. Expected an ID that contains
+            // letters, numbers, underscores, or dashes". Foreign reasoning
+            // is unusable by a Responses provider anyway — drop it (the
+            // exact mirror of the Messages builder's
+            // `prune_replayed_thinking`).
+            if r.id.is_empty() {
+                return vec![];
+            }
+            // `status` is output-only (the API rejects it on input), so
+            // strip it before emission; everything else (summary, content,
+            // encrypted_content, id) passes through.
             let mut r = r.clone();
             r.status = None;
             vec![rs::InputItem::Item(rs::Item::Reasoning(r))]
@@ -2274,12 +2360,15 @@ fn conversation_item_to_input_items(item: &ConversationItem) -> Vec<rs::InputIte
                 }));
             }
 
-            // Add each tool call as a FunctionCall item
+            // Add each tool call as a FunctionCall item. The call_id is
+            // normalized to the Responses charset (foreign backends mint
+            // arbitrary ids); the ToolResult arm applies the SAME map so
+            // pairing survives.
             for tc in &a.tool_calls {
                 let arguments = sanitize_tool_arguments(&tc.id, &tc.name, tc.arguments.clone());
                 items.push(rs::InputItem::Item(rs::Item::FunctionCall(
                     rs::FunctionToolCall {
-                        call_id: tc.id.as_ref().to_owned(),
+                        call_id: sanitize_tool_call_id(&tc.id),
                         name: tc.name.clone(),
                         arguments: arguments.as_ref().to_owned(),
                         id: None,
@@ -2313,7 +2402,9 @@ fn conversation_item_to_input_items(item: &ConversationItem) -> Vec<rs::InputIte
             };
             vec![rs::InputItem::Item(rs::Item::FunctionCallOutput(
                 rs::FunctionCallOutputItemParam {
-                    call_id: t.tool_call_id.clone(),
+                    // Same normalization as the FunctionCall arm — pairing
+                    // survives because both sides map identically.
+                    call_id: sanitize_tool_call_id(&t.tool_call_id),
                     output,
                     id: None,
                     status: None,
@@ -2983,6 +3074,53 @@ pub fn dedup_duplicate_tool_results(conversation: &mut Vec<ConversationItem>) ->
 // ============================================================================
 
 /// Convert a ConversationRequest to Anthropic MessagesRequest.
+/// Normalize a tool-call id to the `[A-Za-z0-9_-]+` charset both Anthropic
+/// Messages and the OpenAI Responses API enforce ("Expected an ID that
+/// contains letters, numbers, underscores, or dashes"). Foreign backends
+/// mint arbitrary ids (chat-completions providers, UUID synthesis), so the
+/// wire builders apply this SYMMETRICALLY on the call and its result —
+/// pairing survives because both sides map through the same function.
+/// ASCII-only (the old closure used Unicode `is_alphanumeric`, letting
+/// e.g. CJK ids through to Anthropic's ASCII contract); an empty id maps
+/// to `"_"` so the mandatory field is never empty on the wire.
+fn sanitize_tool_call_id(id: &str) -> String {
+    if id.is_empty() {
+        return "_".to_string();
+    }
+    id.chars()
+        .map(|c| {
+            if c.is_ascii_alphanumeric() || c == '_' || c == '-' {
+                c
+            } else {
+                '_'
+            }
+        })
+        .collect()
+}
+
+/// The raster media types Anthropic accepts for base64 image sources.
+const ANTHROPIC_IMAGE_MEDIA_TYPES: [&str; 4] =
+    ["image/jpeg", "image/png", "image/gif", "image/webp"];
+
+/// Parse a `data:` URI into an Anthropic base64 image source
+/// `(media_type, data)`.
+///
+/// `None` for anything Anthropic would reject — non-base64 data URIs
+/// (previously leaked as `ImageSource::Url` carrying a `data:` payload:
+/// url sources must be http(s) → 400), media types outside the raster
+/// whitelist (`image/svg+xml` → 400), and param-carrying headers
+/// (`data:image/webp;name=x;base64,…` yields media type
+/// `"image/webp;name=x"` → 400). Callers degrade to a short text
+/// placeholder — never the raw URI, which for data URIs can be megabytes.
+fn parse_base64_image_data_uri(url: &str) -> Option<(String, String)> {
+    let rest = url.strip_prefix("data:")?;
+    let (media_type, data) = rest.split_once(";base64,")?;
+    let media_type = media_type.to_ascii_lowercase();
+    ANTHROPIC_IMAGE_MEDIA_TYPES
+        .contains(&media_type.as_str())
+        .then(|| (media_type, data.to_string()))
+}
+
 pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::MessagesRequest {
     use crate::messages::{
         CacheControl, ContentBlock, ImageSource, Message, MessageContent, MessageRole,
@@ -2995,19 +3133,6 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
     let mut pending_assistant: Vec<ContentBlock> = Vec::new();
     let mut pending_tool_results: Vec<ContentBlock> = Vec::new();
 
-    // Helper to sanitize tool call IDs (replace [^a-zA-Z0-9_-] with _)
-    let sanitize_tool_call_id = |id: &str| -> String {
-        id.chars()
-            .map(|c| {
-                if c.is_alphanumeric() || c == '_' || c == '-' {
-                    c
-                } else {
-                    '_'
-                }
-            })
-            .collect()
-    };
-
     // Helper to convert ContentPart to Anthropic ContentBlock
     let content_parts_to_anthropic_blocks = |parts: &[ContentPart]| -> Vec<ContentBlock> {
         parts
@@ -3018,28 +3143,9 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
                     cache_control: None,
                 },
                 ContentPart::Image { url } => {
-                    // Parse data: URI vs HTTP(S) URL
+                    if let Some((media_type, data)) = parse_base64_image_data_uri(url) {
-                    if url.starts_with("data:") {
+                        ContentBlock::Image {
-                        // data:image/png;base64,ABC123...
+                            source: ImageSource::Base64 { media_type, data },
-                        if let Some((header, data)) = url.split_once(',') {
-                            // Extract media type from header: data:image/png;base64
-                            let media_type = header
-                                .strip_prefix("data:")
-                                .and_then(|h| h.strip_suffix(";base64"))
-                                .unwrap_or("image/png")
-                                .to_string();
-                            ContentBlock::Image {
-                                source: ImageSource::Base64 {
-                                    media_type,
-                                    data: data.to_string(),
-                                },
-                            }
-                        } else {
-                            // Malformed data URI, treat as text
-                            ContentBlock::Text {
-                                text: format!("[invalid image: {}]", url),
-                                cache_control: None,
-                            }
                         }
                     } else if url.starts_with("http://") || url.starts_with("https://") {
                         ContentBlock::Image {
@@ -3047,6 +3153,13 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
                                 url: url.as_ref().to_owned(),
                             },
                         }
+                    } else if url.starts_with("data:") {
+                        // Rejected data URI (non-base64 / non-raster media
+                        // type): short placeholder, NEVER the payload.
+                        ContentBlock::Text {
+                            text: "[unsupported image]".to_string(),
+                            cache_control: None,
+                        }
                     } else {
                         // Unknown format, treat as text
                         ContentBlock::Text {
@@ -3096,7 +3209,17 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
             ConversationItem::User(u) => {
                 flush_assistant(&mut pending_assistant, &mut messages);
                 flush_tool_results(&mut pending_tool_results, &mut messages);
-                let blocks = content_parts_to_anthropic_blocks(&u.content);
+                // Anthropic rejects empty content: drop empty text parts and
+                // give an all-empty user turn a placeholder (mirrors the
+                // assistant arm's emptiness guard).
+                let mut blocks = content_parts_to_anthropic_blocks(&u.content);
+                blocks.retain(|b| !matches!(b, ContentBlock::Text { text, .. } if text.is_empty()));
+                if blocks.is_empty() {
+                    blocks.push(ContentBlock::Text {
+                        text: "[empty message]".to_string(),
+                        cache_control: None,
+                    });
+                }
                 messages.push(Message {
                     role: MessageRole::User,
                     content: MessageContent::Blocks(blocks),
@@ -3139,23 +3262,26 @@ pub fn build_messages_request(req: &ConversationRequest) -> crate::messages::Mes
                     }];
                     for img in &t.images {
                         if let ContentPart::Image { url } = img {
-                            let source = if let Some(rest) = url.strip_prefix("data:") {
+                            // Same whitelist parse as the user path; a
-                                if let Some((media_type, data)) = rest.split_once(";base64,") {
+                            // rejected data URI must never ride as
-                                    ImageSource::Base64 {
+                            // `ImageSource::Url` (url sources are http(s)
-                                        media_type: media_type.to_string(),
+                            // only — Anthropic 400s a `data:` payload).
-                                        data: data.to_string(),
+                            if let Some((media_type, data)) = parse_base64_image_data_uri(url) {
-                                    }
+                                blocks.push(ContentBlock::Image {
-                                } else {
+                                    source: ImageSource::Base64 { media_type, data },
-                                    ImageSource::Url {
+                                });
+                            } else if url.starts_with("http://") || url.starts_with("https://") {
+                                blocks.push(ContentBlock::Image {
+                                    source: ImageSource::Url {
                                         url: url.as_ref().to_owned(),
-                                    }
+                                    },
-                                }
+                                });
                             } else {
-                                ImageSource::Url {
+                                blocks.push(ContentBlock::Text {
-                                    url: url.as_ref().to_owned(),
+                                    text: "[unsupported image]".to_string(),
-                                }
+                                    cache_control: None,
-                            };
+                                });
-                            blocks.push(ContentBlock::Image { source });
+                            }
                         }
                     }
                     ToolResultContent::Blocks(blocks)
@@ -3201,6 +3327,8 @@ 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 {
@@ -3290,6 +3418,74 @@ 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`
@@ -4551,6 +4747,310 @@ mod tests {
         }
     }
 
+    /// Anthropic image sources: base64 only for whitelisted raster types;
+    /// url sources http(s) only. Previously a non-base64 `data:` URI rode
+    /// as `ImageSource::Url` (400), `image/svg+xml` passed the media type
+    /// through (400), and a param-carrying header produced
+    /// `"image/webp;name=x"` (400). Rejected images degrade to a SHORT
+    /// placeholder — never the multi-megabyte payload. Empty user turns
+    /// get a placeholder block (Anthropic rejects empty content).
+    #[test]
+    fn messages_request_guards_images_and_empty_user_content() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user_with_parts(vec![
+                ContentPart::Text { text: "".into() },
+                ContentPart::Image {
+                    url: "data:image/png;base64,AAAA".into(),
+                },
+                ContentPart::Image {
+                    url: "data:image/svg+xml;base64,PHN2Zz4=".into(),
+                },
+                ContentPart::Image {
+                    url: "data:text/plain,hello".into(),
+                },
+            ]),
+            assistant_text("a1"),
+            // Empty user turn: must not ship an empty content array.
+            ConversationItem::user(""),
+        ]);
+        let json = serde_json::to_value(build_messages_request(&req)).unwrap();
+        let messages = json["messages"].as_array().unwrap();
+
+        let first_user = &messages[0]["content"].as_array().unwrap();
+        // Valid png passes as base64.
+        assert!(
+            first_user
+                .iter()
+                .any(|b| b["type"] == "image" && b["source"]["media_type"] == "image/png"),
+            "{json:#}"
+        );
+        // svg + non-base64 rejected to short placeholders; never a
+        // data: payload in a url source, never a non-raster media type.
+        for m in messages {
+            if let Some(content) = m.get("content").and_then(|c| c.as_array()) {
+                assert!(!content.is_empty(), "empty content array: {json:#}");
+                for b in content {
+                    if b["type"] == "image" {
+                        let src = &b["source"];
+                        if src["type"] == "url" {
+                            let u = src["url"].as_str().unwrap();
+                            assert!(
+                                u.starts_with("http://") || u.starts_with("https://"),
+                                "url source must be http(s): {u}"
+                            );
+                        } else {
+                            let mt = src["media_type"].as_str().unwrap();
+                            assert!(
+                                ANTHROPIC_IMAGE_MEDIA_TYPES.contains(&mt),
+                                "media type must be whitelisted: {mt}"
+                            );
+                        }
+                    }
+                }
+            }
+        }
+        assert_eq!(
+            first_user
+                .iter()
+                .filter(|b| b["text"] == "[unsupported image]")
+                .count(),
+            2,
+            "both rejected images degrade to placeholders: {json:#}"
+        );
+        // The empty user turn carries the placeholder block.
+        let last_user = messages.last().unwrap();
+        assert_eq!(last_user["content"][0]["text"], "[empty message]");
+    }
+
+    /// Tool-result images take the same whitelist: a rejected data URI in
+    /// a tool result degrades to a text block, never an
+    /// `ImageSource::Url` carrying a `data:` payload.
+    #[test]
+    fn messages_request_guards_tool_result_images() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q"),
+            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_with_images(
+                "tc1",
+                "saw an image",
+                vec![ContentPart::Image {
+                    url: "data:text/plain,hello".into(),
+                }],
+            ),
+        ]);
+        let json = serde_json::to_value(build_messages_request(&req)).unwrap();
+        let raw = json.to_string();
+        assert!(
+            !raw.contains("data:text/plain"),
+            "rejected data URI must not reach the wire: {json:#}"
+        );
+        assert!(raw.contains("[unsupported image]"), "{json:#}");
+    }
+
+    /// Provenance gate: a turn produced by a DIFFERENT model must not
+    /// replay its opaque items to this request's target — Reasoning
+    /// (model-bound encrypted payloads) is dropped, BackendToolCall
+    /// (provider-issued ids + undeclared tool shapes, e.g. grok x_search
+    /// → codex) is demoted to the same text summary the other builders
+    /// emit. Same-model and unknown-provenance turns stay byte-verbatim
+    /// (KV-cache stability).
+    #[test]
+    fn responses_input_gates_foreign_turns_by_provenance() {
+        let custom_call: rs::CustomToolCall = serde_json::from_value(serde_json::json!({
+            "call_id": "xs_1",
+            "id": "ct_1",
+            "input": "{\"q\":\"news\"}",
+            "name": "x_search",
+        }))
+        .expect("custom tool call fixture");
+        let x_search = ConversationItem::BackendToolCall(BackendToolCallItem {
+            kind: BackendToolKind::XSearch(custom_call),
+        });
+        let foreign_assistant = ConversationItem::Assistant(AssistantItem {
+            content: "grok says hi".into(),
+            tool_calls: vec![],
+            model_id: Some("grok-4".to_string()),
+            model_fingerprint: None,
+            reasoning_effort: None,
+        });
+        let native_assistant = ConversationItem::Assistant(AssistantItem {
+            content: "codex says hi".into(),
+            tool_calls: vec![],
+            model_id: Some("gpt-5.2-codex".to_string()),
+            model_fingerprint: None,
+            reasoning_effort: None,
+        });
+        let mut req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q1"),
+            // Foreign turn: grok reasoning + x_search + assistant.
+            ConversationItem::Reasoning(rs::ReasoningItem {
+                id: "rs_grok_1".to_string(),
+                summary: vec![],
+                content: None,
+                encrypted_content: Some("grok-blob".to_string()),
+                status: None,
+            }),
+            x_search,
+            foreign_assistant,
+            ConversationItem::user("q2"),
+            // Native turn: same model as the request target.
+            ConversationItem::Reasoning(rs::ReasoningItem {
+                id: "rs_codex_1".to_string(),
+                summary: vec![],
+                content: None,
+                encrypted_content: Some("codex-blob".to_string()),
+                status: None,
+            }),
+            native_assistant,
+            ConversationItem::user("q3"),
+        ]);
+        req.model = Some("gpt-5.2-codex".to_string());
+
+        let json = serde_json::to_value(rs::CreateResponse::from(&req)).unwrap();
+        let input = json["input"].as_array().unwrap();
+
+        // Foreign reasoning + x_search gone; the summary text survives.
+        assert!(
+            !input.iter().any(|i| i["id"] == "rs_grok_1"),
+            "foreign reasoning must be dropped:\n{json:#}"
+        );
+        assert!(
+            !input.iter().any(|i| i["type"] == "custom_tool_call"),
+            "foreign backend tool call must not replay typed:\n{json:#}"
+        );
+        assert!(
+            input.iter().any(|i| i["role"] == "assistant"
+                && i["content"]
+                    .as_str()
+                    .is_some_and(|c| c.contains("[backend x_search]"))),
+            "foreign backend tool call demoted to text summary:\n{json:#}"
+        );
+        // Native reasoning verbatim.
+        assert!(
+            input
+                .iter()
+                .any(|i| i["id"] == "rs_codex_1" && i["encrypted_content"] == "codex-blob"),
+            "native reasoning must replay verbatim:\n{json:#}"
+        );
+    }
+
+    /// Both Anthropic Messages and the Responses API enforce
+    /// `[A-Za-z0-9_-]+` tool-call ids; foreign backends mint arbitrary
+    /// ones. The shared sanitizer must be ASCII-only (the old closure's
+    /// Unicode `is_alphanumeric` let CJK ids through), never emit an empty
+    /// id, and map call + result IDENTICALLY so pairing survives.
+    #[test]
+    fn tool_call_ids_sanitized_symmetrically_on_responses_leg() {
+        let weird_id = "调用#1 β";
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q"),
+            ConversationItem::Assistant(AssistantItem {
+                content: "".into(),
+                tool_calls: vec![ToolCall {
+                    id: std::sync::Arc::from(weird_id),
+                    name: "read_file".to_string(),
+                    arguments: std::sync::Arc::from("{}"),
+                }],
+                model_id: None,
+                model_fingerprint: None,
+                reasoning_effort: None,
+            }),
+            ConversationItem::tool_result(weird_id, "contents"),
+        ]);
+        let json = serde_json::to_value(rs::CreateResponse::from(&req)).unwrap();
+        let input = json["input"].as_array().unwrap();
+        let call_id = input
+            .iter()
+            .find(|i| i["type"] == "function_call")
+            .map(|i| i["call_id"].as_str().unwrap().to_string())
+            .expect("function_call present");
+        let output_id = input
+            .iter()
+            .find(|i| i["type"] == "function_call_output")
+            .map(|i| i["call_id"].as_str().unwrap().to_string())
+            .expect("function_call_output present");
+        assert_eq!(call_id, output_id, "pairing must survive sanitization");
+        assert!(
+            call_id
+                .chars()
+                .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-'),
+            "sanitized id must satisfy the wire charset: {call_id:?}"
+        );
+        assert!(!call_id.is_empty());
+        // The sanitizer itself: ASCII passthrough, unicode replaced, empty
+        // never emitted.
+        assert_eq!(sanitize_tool_call_id("toolu_01AB-cd"), "toolu_01AB-cd");
+        assert_eq!(sanitize_tool_call_id("统A1"), "_A1");
+        assert_eq!(sanitize_tool_call_id(""), "_");
+    }
+
+    /// The Responses API requires a server-issued id on every replayed
+    /// reasoning input item — an empty one 400s with "Invalid
+    /// 'input[N].id': ''" (observed on the Codex backend after a
+    /// cross-backend session switched to a GPT model). Empty-id reasoning
+    /// can only be foreign: Messages-captured (Anthropic signature,
+    /// id "") or chat-completions-synthesized (id "", no encrypted
+    /// content). Neither is usable by a Responses provider — drop them;
+    /// native `rs_*` items pass through untouched.
+    #[test]
+    fn responses_input_drops_reasoning_without_native_id() {
+        let req = ConversationRequest::from_items(vec![
+            ConversationItem::user("q1"),
+            // Messages-captured: Anthropic signature, empty id.
+            reasoning("claude thinking", Some("anthropic-sig")),
+            assistant_text("a1"),
+            ConversationItem::user("q2"),
+            // Chat-completions synthesized: empty id, nothing encrypted.
+            ConversationItem::Reasoning(synthesized_reasoning_item("kimi thinking")),
+            assistant_text("a2"),
+            ConversationItem::user("q3"),
+            // Native Responses item: server-issued id.
+            ConversationItem::Reasoning(rs::ReasoningItem {
+                id: "rs_native_1".to_string(),
+                summary: vec![],
+                content: None,
+                encrypted_content: Some("gAAAA-native".to_string()),
+                status: None,
+            }),
+            assistant_text("a3"),
+            ConversationItem::user("q4"),
+        ]);
+
+        let responses_req: rs::CreateResponse = (&req).into();
+        let json = serde_json::to_value(&responses_req).unwrap();
+        let reasoning_items: Vec<&serde_json::Value> = json["input"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .filter(|i| i.get("type").and_then(|t| t.as_str()) == Some("reasoning"))
+            .collect();
+        assert_eq!(
+            reasoning_items.len(),
+            1,
+            "only the native rs_* item may be replayed:\n{json:#}"
+        );
+        assert_eq!(reasoning_items[0]["id"], "rs_native_1");
+        assert_eq!(reasoning_items[0]["encrypted_content"], "gAAAA-native");
+        assert!(
+            !json["input"]
+                .as_array()
+                .unwrap()
+                .iter()
+                .any(|i| i.get("id").and_then(|v| v.as_str()) == Some("")),
+            "no input item may carry an empty id:\n{json:#}"
+        );
+    }
+
     #[test]
     fn test_encrypted_reasoning_included_in_responses_api_request() {
         // Test that when building a Responses API request, encrypted reasoning is included
@@ -4617,12 +5117,15 @@ mod tests {
 
     #[test]
     fn test_only_encrypted_reasoning_included_in_request() {
-        // Test that when there's only encrypted content (no visible summary),
+        // Encrypted-only reasoning replays ONLY with a native (server-issued)
-        // it's still included in the request
+        // id. An id-less encrypted blob is by construction FOREIGN (the
+        // Responses stream always captures the `rs_*` id; Messages capture
+        // stores the Anthropic signature with id "") and the API rejects
+        // empty ids — see `responses_input_drops_reasoning_without_native_id`.
         let req = ConversationRequest::from_items(vec![
             ConversationItem::user("Hello"),
             ConversationItem::Reasoning(rs::ReasoningItem {
-                id: String::new(),
+                id: "rs_hidden_1".to_string(),
                 summary: vec![],
                 content: None,
                 encrypted_content: Some("enc_hidden_thoughts".to_string()),
@@ -4655,6 +5158,7 @@ mod tests {
 
         assert_eq!(reasoning_items.len(), 1);
         let reasoning = reasoning_items[0];
+        assert_eq!(reasoning.id, "rs_hidden_1");
 
         // Encrypted content should be present
         assert_eq!(
@@ -5332,6 +5836,160 @@ 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.
@@ -9316,6 +9974,40 @@ mod tests {
         assert_prefix_stable(&req2, &req3);
     }
 
+    /// `text_summary`'s code preview truncated at BYTE 100 (`&c[..100]`) —
+    /// a panic on any non-ASCII boundary (CJK/emoji in interpreted code).
+    /// One poisoned history item then crashed every subsequent request
+    /// build on every backend. Truncation must be char-boundary safe.
+    #[test]
+    fn code_interpreter_summary_truncates_multibyte_code_safely() {
+        let item = BackendToolCallItem {
+            kind: BackendToolKind::CodeInterpreter(rs::CodeInterpreterToolCall {
+                code: Some("统计".repeat(60)),
+                container_id: "cont_1".to_string(),
+                id: "ci_1".to_string(),
+                outputs: None,
+                status: rs::CodeInterpreterToolCallStatus::Completed,
+            }),
+        };
+        let summary = item.text_summary();
+        assert!(summary.starts_with("[backend code_interpreter] 统计"));
+        assert!(
+            summary.ends_with("..."),
+            "long code must truncate: {summary}"
+        );
+        // ASCII shorter than the cap stays whole.
+        let short = BackendToolCallItem {
+            kind: BackendToolKind::CodeInterpreter(rs::CodeInterpreterToolCall {
+                code: Some("print(1)".to_string()),
+                container_id: "cont_1".to_string(),
+                id: "ci_2".to_string(),
+                outputs: None,
+                status: rs::CodeInterpreterToolCallStatus::Completed,
+            }),
+        };
+        assert_eq!(short.text_summary(), "[backend code_interpreter] print(1)");
+    }
+
     /// `BackendToolCall` items round-trip through the wire as their
     /// typed Item shape; their serialized position must be stable across
     /// turns. (This is the structural analogue of the old

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

@@ -1041,6 +1041,87 @@ pub fn patch_reasoning_effort(body: &mut Value, effort: Option<ReasoningEffort>)
     }
 }
 
+/// 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.
+/// 3. For the same reason, a replayed reasoning item WITHOUT
+///    `encrypted_content` (captured from a stateful api.openai.com session
+///    that never requested the include) references server state
+///    chatgpt.com does not have — drop it rather than 400.
+///
+/// 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);
+        }
+    }
+
+    // 3. Drop reasoning items with no encrypted payload: stateless codex
+    // cannot resolve a bare `rs_*` reference.
+    if let Some(input) = body.get_mut("input").and_then(|v| v.as_array_mut()) {
+        input.retain(|item| {
+            item.get("type").and_then(|t| t.as_str()) != Some("reasoning")
+                || item
+                    .get("encrypted_content")
+                    .and_then(|v| v.as_str())
+                    .is_some_and(|s| !s.is_empty())
+        });
+    }
+}
+
 /// 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
@@ -1551,6 +1632,98 @@ 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);
+    }
+
+    /// Stateless codex cannot resolve a bare `rs_*` reference: reasoning
+    /// input items without an encrypted payload are dropped; items WITH
+    /// one pass through untouched.
+    #[test]
+    fn codex_adapter_drops_reasoning_without_encrypted_payload() {
+        let mut body = json!({
+            "model": "gpt-5.2-codex",
+            "input": [
+                {"type": "message", "role": "user", "content": "q"},
+                {"type": "reasoning", "id": "rs_bare", "summary": []},
+                {"type": "reasoning", "id": "rs_full", "summary": [],
+                 "encrypted_content": "gAAAA-blob"},
+                {"type": "message", "role": "assistant", "content": "a"}
+            ]
+        });
+        adapt_body_for_codex_backend(&mut body);
+        let input = body["input"].as_array().unwrap();
+        assert_eq!(input.len(), 3, "bare rs_* item dropped: {body:#}");
+        assert!(
+            input
+                .iter()
+                .any(|i| i.get("id").and_then(|v| v.as_str()) == Some("rs_full")),
+            "encrypted reasoning passes through: {body:#}"
+        );
+        assert!(
+            !input
+                .iter()
+                .any(|i| i.get("id").and_then(|v| v.as_str()) == Some("rs_bare")),
+            "{body:#}"
+        );
+    }
+
+    /// 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.

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

@@ -1881,12 +1881,15 @@ impl Config {
             .default(true)
             .resolve()
     }
-    /// Graph mode (`/graph`) master switch. Default ON — the gray release
+    /// Graph mode (`/graph`) master switch. Default ON in the binary: the
-    /// (plan.md G0, `KIGI_GRAPH=1` only) is over; every install gets the
+    /// README ships graph engineering enabled for every install, but the
-    /// same commands (its absence on non-dev machines read as a platform
+    /// old `default(false)` delegated enablement to installer env plumbing
-    /// bug). `KIGI_GRAPH=0` remains the off-switch. Graph mode additionally
+    /// (`install.sh` shell-rc export vs `install.ps1` registry write) — and
-    /// requires the goal harness (nodes execute as goals), enforced at
+    /// Windows terminals don't pick up freshly-written registry env, so
-    /// availability time, not here.
+    /// `/graph` went "missing on Windows". The product default lives HERE,
+    /// not in installers. `KIGI_GRAPH=0` is the off-switch. Graph mode
+    /// additionally requires the goal harness (nodes execute as goals),
+    /// enforced at availability time, not here.
     pub(crate) fn resolve_graph(&self) -> Resolved<bool> {
         BoolFlag::env("KIGI_GRAPH").default(true).resolve()
     }

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

@@ -472,7 +472,12 @@ async fn responses_upgrade_roundtrips_reconstructed_reasoning_as_typed_input() {
             "\n",
             r#"{"type":"user","content":[{"type":"text","text":"q1"}]}"#,
             "\n",
-            r#"{"type":"assistant","content":"a1","reasoning":{"text":"legacy kigi reasoning","encrypted":"ENC_BLOB_xyz","id":"rs_kigibuild_legacy"},"model_id":"kigi"}"#,
+            // model_id matches the test client's request model: this test
+            // covers the SAME-MODEL continuation (the byte-stable
+            // SGLang-prefix path). A mismatched model_id is the provenance
+            // gate's territory (`transform_items_for_responses`) and drops
+            // the reasoning by design.
+            r#"{"type":"assistant","content":"a1","reasoning":{"text":"legacy kigi reasoning","encrypted":"ENC_BLOB_xyz","id":"rs_kigibuild_legacy"},"model_id":"test-model"}"#,
             "\n",
         ),
     )
@@ -535,13 +540,20 @@ async fn responses_upgrade_roundtrips_reconstructed_reasoning_as_typed_input() {
 /// Upgrade path, Anthropic Messages API: a legacy session whose assistant
 /// carries inline `reasoning: {text, encrypted, id}` (text = thinking,
 /// encrypted = signature) must, on load, reconstruct a sibling Reasoning
-/// item that emits a Anthropic Messages `thinking` content block (with `thinking`
+/// item — and when that turn is the ACTIVE tool-use continuation, its
-/// + `signature`) on the outgoing `/v1/messages` request.
+/// `thinking` block (text + signature) must reach the outgoing
+/// `/v1/messages` request verbatim.
+///
+/// Outside an active tool loop the block must be STRIPPED: Anthropic
+/// validates every replayed signature (model-bound), so replaying stale
+/// thinking is exactly what 400'd with "Invalid `signature` in `thinking`
+/// block" after cross-model histories (see `prune_replayed_thinking`).
 #[tokio::test]
-async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
+async fn messages_upgrade_replays_reconstructed_thinking_only_in_active_tool_loop() {
-    // 1. Seed a legacy Anthropic Messages-origin chat_history.jsonl. Anthropic Messages
+    // 1. Seed a legacy Anthropic Messages-origin chat_history.jsonl whose
-    //    thinking blocks never carried an id (stream/messages.rs sets
+    //    assistant turn issued a tool call (thinking blocks never carried an
-    //    id=""), and the signature lives in `encrypted`.
+    //    id — stream/messages.rs sets id="" — and the signature lives in
+    //    `encrypted`). The pending tool_result makes this the active loop.
     let dir = tempfile::tempdir().unwrap();
     std::fs::write(
         dir.path().join("chat_history.jsonl"),
@@ -550,7 +562,9 @@ async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
             "\n",
             r#"{"type":"user","content":[{"type":"text","text":"q1"}]}"#,
             "\n",
-            r#"{"type":"assistant","content":"a1","reasoning":{"text":"legacy anthropic thinking","encrypted":"SIGNATURE_abc","id":""},"model_id":"kigi-4.5"}"#,
+            r#"{"type":"assistant","content":"a1","reasoning":{"text":"legacy anthropic thinking","encrypted":"SIGNATURE_abc","id":""},"model_id":"kigi-4.5","tool_calls":[{"id":"tc1","name":"read_file","arguments":"{}"}]}"#,
+            "\n",
+            r#"{"type":"tool_result","tool_call_id":"tc1","content":"file contents"}"#,
             "\n",
         ),
     )
@@ -558,7 +572,7 @@ async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
 
     // 2. Load + upgrade.
     let adapter = JsonlStorageAdapter::with_root(dir.path().to_path_buf());
-    let mut items = adapter.load_chat_history_from_dir(dir.path()).unwrap();
+    let items = adapter.load_chat_history_from_dir(dir.path()).unwrap();
     assert!(
         items
             .iter()
@@ -566,20 +580,18 @@ async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
         "legacy inline reasoning must be reconstructed as a sibling on load, got {items:?}"
     );
 
-    // 3. Continue and send over the Messages API, capturing the body.
+    // 3. Send the tool-loop continuation over the Messages API.
-    items.push(ConversationItem::user("q2"));
-
     let server = MockInferenceServer::start().await.unwrap();
     server.set_response("ok");
     let client = create_test_client(&server.url(), ApiBackend::Messages);
 
     let _ = client
-        .conversation_collect(ConversationRequest::from_items(items))
+        .conversation_collect(ConversationRequest::from_items(items.clone()))
         .await
         .unwrap();
 
-    // 4. The reconstructed reasoning must emit a Anthropic Messages `thinking`
+    // 4. The active loop's reconstructed reasoning must emit an Anthropic
-    //    content block carrying the thinking text + signature.
+    //    `thinking` content block carrying the thinking text + signature.
     let body = server.request_bodies().pop().unwrap();
     let messages = body.get("messages").unwrap().as_array().unwrap();
     let thinking_block = messages
@@ -593,7 +605,7 @@ async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
         })
         .find(|b| b.get("type").and_then(Value::as_str) == Some("thinking"))
         .unwrap_or_else(|| {
-            panic!("reconstructed reasoning must emit an Anthropic thinking block; messages: {messages:#?}")
+            panic!("active-loop reasoning must emit an Anthropic thinking block; messages: {messages:#?}")
         });
     assert_eq!(
         thinking_block.get("thinking").and_then(Value::as_str),
@@ -605,6 +617,25 @@ async fn messages_upgrade_emits_reconstructed_reasoning_as_thinking_block() {
         Some("SIGNATURE_abc"),
         "signature (encrypted) preserved — required to reuse the thought server-side"
     );
+
+    // 5. A follow-up user turn CLOSES the loop: the same history plus a new
+    //    user message must replay NO thinking block at all.
+    let mut closed = items;
+    closed.push(ConversationItem::user("q2"));
+    let _ = client
+        .conversation_collect(ConversationRequest::from_items(closed))
+        .await
+        .unwrap();
+    let body = server.request_bodies().pop().unwrap();
+    let any_thinking = body["messages"].as_array().unwrap().iter().any(|m| {
+        m.get("content")
+            .and_then(Value::as_array)
+            .is_some_and(|c| c.iter().any(|b| b["type"] == "thinking"))
+    });
+    assert!(
+        !any_thinking,
+        "stale thinking must be stripped outside the active tool loop; body: {body:#?}"
+    );
 }
 
 // ============================================================================
@@ -1443,3 +1474,79 @@ async fn test_chat_completions_backend_hits_chat_endpoint_not_responses() {
         "Should NOT have called /v1/responses"
     );
 }
+
+/// ChatGPT/Codex backend body contract (`openai_codex = true`): the
+/// `/codex/responses` endpoint rejects `role: system` input outright
+/// (400 {"detail":"System messages are not allowed"}) — system content
+/// must ride the top-level `instructions` field, and stateless reasoning
+/// replay needs `include: ["reasoning.encrypted_content"]`. Ported from
+/// the official Codex CLI + Pi's api/openai-codex-responses.ts, like the
+/// identity headers.
+#[tokio::test]
+async fn codex_responses_body_hoists_system_into_instructions() {
+    let server = MockInferenceServer::start().await.unwrap();
+    server.set_response("ok");
+    let mut config = common::test_sampler_config(&server.url(), ApiBackend::Responses, &[]);
+    config.openai_codex = true;
+    let client = Client::new(config).unwrap();
+
+    let _ = client
+        .conversation_collect(ConversationRequest::from_items(vec![
+            ConversationItem::system("You are Kigi."),
+            ConversationItem::user("test"),
+        ]))
+        .await
+        .unwrap();
+
+    let body = server.request_bodies().pop().unwrap();
+    let input = body["input"].as_array().unwrap();
+    assert!(
+        input
+            .iter()
+            .all(|i| i.get("role").and_then(Value::as_str) != Some("system")),
+        "codex backend must never receive system-role input: {body:#?}"
+    );
+    assert_eq!(
+        body["instructions"].as_str(),
+        Some("You are Kigi."),
+        "system prompt must ride the instructions field: {body:#?}"
+    );
+    assert_eq!(
+        body["include"],
+        serde_json::json!(["reasoning.encrypted_content"]),
+        "stateless reasoning replay requires the include: {body:#?}"
+    );
+}
+
+/// Control: the API-key `openai` Responses path (`openai_codex = false`)
+/// stays byte-compatible — system-role input preserved, no codex fields.
+#[tokio::test]
+async fn plain_responses_body_keeps_system_role_input() {
+    let server = MockInferenceServer::start().await.unwrap();
+    server.set_response("ok");
+    let client = create_test_client(&server.url(), ApiBackend::Responses);
+
+    let _ = client
+        .conversation_collect(ConversationRequest::from_items(vec![
+            ConversationItem::system("You are Kigi."),
+            ConversationItem::user("test"),
+        ]))
+        .await
+        .unwrap();
+
+    let body = server.request_bodies().pop().unwrap();
+    assert!(
+        body["input"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .any(|i| i.get("role").and_then(Value::as_str) == Some("system")),
+        "api-key openai keeps system-role input: {body:#?}"
+    );
+    assert!(
+        body.get("instructions")
+            .map(|v| v.is_null())
+            .unwrap_or(true),
+        "no instructions hoist outside codex: {body:#?}"
+    );
+}

install.ps1

@@ -149,15 +149,13 @@ try {
         Write-Host "Run 'kigi' to get started."
     }
 
-    # Graph engineering ships enabled by default. Respect an explicit
+    # Graph engineering is enabled by default IN THE BINARY (resolve_graph
-    # user choice: only set the variable when it is not already defined
+    # defaults true) — no environment plumbing needed. The installer used
-    # (so a persisted opt-out of "0" survives reinstalls).
+    # to persist KIGI_GRAPH=1 into the User registry env, but running
-    $Graph = [Environment]::GetEnvironmentVariable("KIGI_GRAPH", "User")
+    # terminals (and new tabs of an open Windows Terminal) never pick up
-    if ($null -eq $Graph -or $Graph -eq "") {
+    # freshly-written registry variables, which made /graph "missing on
-        [Environment]::SetEnvironmentVariable("KIGI_GRAPH", "1", "User")
+    # Windows" while the shell-rc path worked on macOS/Linux. Opt out any
-        Write-Host "Enabled graph engineering (KIGI_GRAPH=1)."
+    # time with: [Environment]::SetEnvironmentVariable('KIGI_GRAPH','0','User')
-        Write-Host "Disable: [Environment]::SetEnvironmentVariable('KIGI_GRAPH','0','User')"
-    }
 } finally {
     Remove-Item -Path $TmpDir -Recurse -Force -ErrorAction SilentlyContinue
 }

install.sh

@@ -209,7 +209,6 @@ case "${SHELL:-}" in
     */zsh)
         RC_FILE="${ZDOTDIR:-$HOME}/.zshrc"
         PATH_LINE="export PATH=\"$BIN_DIR:\$PATH\""
-        GRAPH_LINE="export KIGI_GRAPH=1"
         ;;
     */bash)
         # macOS login shells read ~/.bash_profile; Linux reads ~/.bashrc.
@@ -219,7 +218,6 @@ case "${SHELL:-}" in
             RC_FILE="$HOME/.bashrc"
         fi
         PATH_LINE="export PATH=\"$BIN_DIR:\$PATH\""
-        GRAPH_LINE="export KIGI_GRAPH=1"
         ;;
     */fish)
         # fish_add_path in config.fish is fish's own idempotent way
@@ -228,12 +226,10 @@ case "${SHELL:-}" in
         mkdir -p "$FISH_CONF_DIR"
         RC_FILE="$FISH_CONF_DIR/config.fish"
         PATH_LINE="fish_add_path $BIN_DIR"
-        GRAPH_LINE="set -gx KIGI_GRAPH 1"
         ;;
     *)
         RC_FILE="$HOME/.profile"
         PATH_LINE="export PATH=\"$BIN_DIR:\$PATH\""
-        GRAPH_LINE="export KIGI_GRAPH=1"
         ;;
 esac
 
@@ -247,8 +243,8 @@ case ":$PATH:" in
         ;;
 esac
 
-# Graph engineering ships enabled by default. The KIGI_GRAPH guard makes
+# Graph engineering is enabled by default IN THE BINARY (resolve_graph
-# this idempotent AND respects an explicit user opt-out (an existing
+# defaults true) — the installer no longer writes KIGI_GRAPH=1 into shell
-# `export KIGI_GRAPH=0` line is left untouched). Disable any time with:
+# rc files (per-shell env plumbing was fragile and diverged per platform).
+# Disable any time with:
 #   echo 'export KIGI_GRAPH=0' >> <your shell rc>
-persist_line "$RC_FILE" "$GRAPH_LINE" "KIGI_GRAPH" "graph engineering (KIGI_GRAPH=1)"