Responses adapter hygiene: reasoning drop, dedupe, indexer leak, tool-name codec, double SSE parse (CL-6912)

src/provider/codex-responses-adapter.test.ts

@@ -4,6 +4,8 @@ import { PRODUCT_NAME } from "../branding.js";
 import {
   createCodexResponsesAdapter,
   isResponsesStreamTerminal,
+  signatureForModel,
+  tagSignature,
 } from "./codex-responses-adapter.js";
 
 const source: LastCycleSource = {
@@ -134,6 +136,158 @@ describe("createCodexResponsesAdapter usage parsing", () => {
   });
 });
 
+describe("signatureForModel", () => {
+  const turnWithModel = (model: string | undefined): ConversationTurn =>
+    ({
+      role: "assistant",
+      model,
+      content: [],
+      timestamp: 0,
+    }) as unknown as ConversationTurn;
+
+  test("replays a signature on a turn with no persisted model", () => {
+    const signature = tagSignature("codex-responses", "cipher");
+    const result = signatureForModel(
+      turnWithModel(undefined),
+      "gpt-5.1-codex",
+      "codex-responses",
+      signature,
+    );
+    expect(result).toBe("cipher");
+  });
+
+  test("drops a signature when the turn's model genuinely differs", () => {
+    const signature = tagSignature("codex-responses", "cipher");
+    const result = signatureForModel(
+      turnWithModel("gpt-5.0-codex"),
+      "gpt-5.1-codex",
+      "codex-responses",
+      signature,
+    );
+    expect(result).toBeUndefined();
+  });
+
+  test("replays a signature when the model matches", () => {
+    const signature = tagSignature("codex-responses", "cipher");
+    const result = signatureForModel(
+      turnWithModel("gpt-5.1-codex"),
+      "gpt-5.1-codex",
+      "codex-responses",
+      signature,
+    );
+    expect(result).toBe("cipher");
+  });
+});
+
+describe("createCodexResponsesAdapter orphaned function_call suppression", () => {
+  test("drops a function_call whose reasoning signature could not be replayed", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      { role: "user", timestamp: 0, content: [{ type: "text", text: "hi" }] },
+      {
+        role: "assistant",
+        model: "gpt-5.0-codex",
+        timestamp: 0,
+        content: [
+          { type: "thinking", thinking: "ponder", signature: tagSignature("codex-responses", "c") },
+          { type: "tool_call", id: "call_1", name: "shell", arguments: {} },
+        ],
+      },
+    ] as unknown as ConversationTurn[];
+
+    const request = adapter.buildRequest(turns, "gpt-5.1-codex", {});
+    const body = JSON.parse(request.body) as { input: { type: string }[] };
+
+    expect(body.input.some((item) => item.type === "reasoning")).toBe(false);
+    expect(body.input.some((item) => item.type === "function_call")).toBe(false);
+  });
+
+  test("keeps the function_call when its reasoning signature replays cleanly", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      { role: "user", timestamp: 0, content: [{ type: "text", text: "hi" }] },
+      {
+        role: "assistant",
+        model: "gpt-5.1-codex",
+        timestamp: 0,
+        content: [
+          { type: "thinking", thinking: "ponder", signature: tagSignature("codex-responses", "c") },
+          { type: "tool_call", id: "call_1", name: "shell", arguments: {} },
+        ],
+      },
+    ] as unknown as ConversationTurn[];
+
+    const request = adapter.buildRequest(turns, "gpt-5.1-codex", {});
+    const body = JSON.parse(request.body) as { input: { type: string }[] };
+
+    expect(body.input.some((item) => item.type === "reasoning")).toBe(true);
+    expect(body.input.some((item) => item.type === "function_call")).toBe(true);
+  });
+});
+
+describe("createCodexResponsesAdapter tool-name codec", () => {
+  test("encodes a non-wire-safe tool name on the outgoing function tool definition", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      { role: "user", timestamp: 0, content: [{ type: "text", text: "hi" }] },
+    ];
+
+    const request = adapter.buildRequest(turns, "gpt-5.1-codex", {
+      tools: [
+        {
+          name: "@intx/tools-posix/sidecar-bundle:run_shell",
+          description: "run a shell command",
+          inputSchema: {},
+        },
+      ],
+    } as never);
+    const body = JSON.parse(request.body) as { tools: { name: string }[] };
+
+    expect(body.tools[0]?.name).toMatch(/^[A-Za-z_][A-Za-z0-9_-]*$/);
+    expect(body.tools[0]?.name).not.toBe("@intx/tools-posix/sidecar-bundle:run_shell");
+  });
+
+  test("decodes an encoded tool_call.start name back to the internal id", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const encoded = "IX_-40intx-2Ftools-2Dposix-2Fsidecar-2Dbundle-3Arun_shell";
+    const sseData = JSON.stringify({
+      type: "response.output_item.added",
+      item: { type: "function_call", id: "item_1", call_id: "call_1", name: encoded },
+    });
+
+    const events = adapter.parseResponse(sseData);
+    const start = events.find((e) => e.type === "inference.tool_call.start");
+
+    expect((start?.data as { name?: string })?.name).not.toBe(encoded);
+  });
+});
+
+describe("createCodexResponsesAdapter block indexer reset", () => {
+  test("resets block indices on a new buildRequest instead of accumulating across requests", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      { role: "user", timestamp: 0, content: [{ type: "text", text: "hi" }] },
+    ];
+
+    adapter.buildRequest(turns, "gpt-5.1-codex", {});
+    adapter.parseResponse(
+      JSON.stringify({ type: "response.output_text.delta", item_id: "item_1", delta: "a" }),
+    );
+    adapter.parseResponse(
+      JSON.stringify({ type: "response.output_text.delta", item_id: "item_2", delta: "b" }),
+    );
+
+    // A new request (a fresh HTTP round trip) with a brand-new item id should
+    // start indexing from 0 again, not continue accumulating from the prior
+    // request's indexer state.
+    adapter.buildRequest(turns, "gpt-5.1-codex", {});
+    const secondRequestDelta = adapter.parseResponse(
+      JSON.stringify({ type: "response.output_text.delta", item_id: "item_3", delta: "c" }),
+    );
+    expect((secondRequestDelta[0]?.data as { index?: number })?.index).toBe(0);
+  });
+});
+
 describe("isResponsesStreamTerminal", () => {
   test("is true for the Responses end-of-turn events", () => {
     for (const type of ["response.completed", "response.incomplete", "response.done"]) {

src/provider/codex-responses-adapter.ts

@@ -1,8 +1,11 @@
 import {
   BEARER_CREDENTIAL_SENTINEL,
   ProtocolMismatchError,
+  decodeToolName,
+  encodeToolName,
   type BuiltRequest,
   type ProviderAdapter,
+  type ToolNameLimit,
 } from "@intx/inference";
 import type {
   ContentBlock,
@@ -157,7 +160,11 @@ export function signatureForModel(
   requestProvider: string,
   signature: string,
 ): string | undefined {
-  if (turn.model !== requestModel) return undefined;
+  // `model` is optional on the persisted turn schema; a turn saved before that
+  // field existed (or otherwise missing it) is not evidence of a model
+  // switch — treat the absence as benign and fall through to the provider
+  // check, rather than dropping reasoning that never actually crossed models.
+  if (turn.model !== undefined && turn.model !== requestModel) return undefined;
   const tagged = untagSignature(signature);
   if (tagged === undefined) return undefined;
   return tagged.provider === requestProvider ? tagged.encryptedContent : undefined;
@@ -171,6 +178,14 @@ export function signatureForModel(
 // (held in a thinking block's signature) AND that backend is the one this
 // request is going to — replaying it to a different provider gets a 400 it
 // cannot recover from.
+// Wire-charset limit for function names on the Responses surface (Codex,
+// Grok, and the generic OpenAI Responses adapter all share OpenAI's
+// `^[a-zA-Z0-9_-]{1,64}$` function-name charset).
+export const RESPONSES_TOOL_NAME_LIMIT: ToolNameLimit = {
+  provider: "responses",
+  maxLength: 64,
+};
+
 function toResponsesItems(
   turn: ConversationTurn,
   requestModel: string,
@@ -180,11 +195,20 @@ function toResponsesItems(
   const textKind: "input_text" | "output_text" =
     turn.role === "assistant" ? "output_text" : "input_text";
   const textParts: ResponsesContentPart[] = [];
+  // A reasoning block whose signature we could not replay (foreign provider,
+  // model switch, or a missing/untagged signature) leaves any function_call
+  // it produced without the reasoning item the Responses API expects to
+  // precede it — the exact orphaned shape that degenerates reasoning models.
+  // Suppress function_call items until the next text or successfully-replayed
+  // reasoning item re-establishes a clean turn shape; tool results are
+  // unaffected since they never need a preceding reasoning item.
+  let suppressOrphanedCalls = false;
 
   const flushText = (): void => {
     if (textParts.length > 0) {
       items.push({ type: "message", role: turn.role, content: [...textParts] });
       textParts.length = 0;
+      suppressOrphanedCalls = false;
     }
   };
 
@@ -206,15 +230,17 @@ function toResponsesItems(
         } as ResponsesContentPart);
       }
     } else if (block.type === "tool_call") {
+      if (suppressOrphanedCalls) continue;
       flushText();
       items.push({
         type: "function_call",
-        name: block.name,
+        name: encodeToolName(block.name, RESPONSES_TOOL_NAME_LIMIT),
         arguments: JSON.stringify(block.arguments ?? {}),
         call_id: block.id,
       });
     } else if (block.type === "tool_result") {
       flushText();
+      suppressOrphanedCalls = false;
       items.push({
         type: "function_call_output",
         call_id: block.callId,
@@ -234,6 +260,9 @@ function toResponsesItems(
       );
       if (encryptedContent !== undefined) {
         items.push({ type: "reasoning", summary: [], encrypted_content: encryptedContent });
+        suppressOrphanedCalls = false;
+      } else {
+        suppressOrphanedCalls = true;
       }
     }
   }
@@ -260,7 +289,7 @@ function toResponsesTools(options: InferenceOptions): unknown[] | undefined {
   // `type`, not nested under a `function` key (unlike Chat Completions).
   return options.tools.map((t) => ({
     type: "function",
-    name: t.name,
+    name: encodeToolName(t.name, RESPONSES_TOOL_NAME_LIMIT),
     description: t.description,
     parameters: t.inputSchema,
   }));
@@ -466,7 +495,7 @@ export function parseResponse(
               seq,
               data: {
                 callId,
-                name,
+                name: decodeToolName(name),
                 partial: EMPTY_PARTIAL,
                 index: blockIndexFor(indexer, itemId, "tool_call"),
               },
@@ -608,13 +637,15 @@ export function isResponsesStreamTerminal(sseData: string): boolean {
 }
 
 export function createCodexResponsesAdapter(source: LastCycleSource): ProviderAdapter {
-  const indexer: CodexBlockIndexer = {
-    nextIndex: 0,
-    items: new Map<string, { index: number; kind: CodexBlockKind }>(),
-  };
+  // Re-created per request in buildRequest, not just once here — otherwise
+  // block indices accumulate across every request the adapter instance ever
+  // serves, growing the map for the life of the conversation.
+  let indexer: CodexBlockIndexer = createResponsesBlockIndexer();
   return {
-    buildRequest: (messages, model, options) =>
-      buildRequest(messages, model, options, source.provider),
+    buildRequest: (messages, model, options) => {
+      indexer = createResponsesBlockIndexer();
+      return buildRequest(messages, model, options, source.provider);
+    },
     parseResponse: (sseData) => parseResponse(sseData, indexer, source),
     parseJSONResponse,
     isStreamTerminal: isResponsesStreamTerminal,

src/provider/grok-responses-adapter.test.ts

@@ -98,6 +98,52 @@ describe("createGrokResponsesAdapter", () => {
     expect(body.reasoning).toEqual({ effort: "low", summary: "detailed" });
   });
 
+  test("keeps the latest function_call_output on a duplicate call_id", () => {
+    const adapter = createGrokResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      {
+        role: "user",
+        timestamp: 0,
+        content: [
+          { type: "tool_result", callId: "call_1", content: [{ type: "text", text: "stale" }] },
+          { type: "tool_result", callId: "call_1", content: [{ type: "text", text: "fresh" }] },
+        ],
+      },
+    ] as unknown as ConversationTurn[];
+
+    const request = adapter.buildRequest(turns, "grok-4.5", {});
+    const body = JSON.parse(request.body) as {
+      input: { type: string; call_id?: string; output?: string }[];
+    };
+    const outputs = body.input.filter((item) => item.type === "function_call_output");
+
+    expect(outputs).toHaveLength(1);
+    expect(outputs[0]?.output).toBe("fresh");
+  });
+
+  test("dedupes a duplicate function_call on the same call_id", () => {
+    const adapter = createGrokResponsesAdapter(source);
+    const turns: ConversationTurn[] = [
+      {
+        role: "assistant",
+        timestamp: 0,
+        content: [
+          { type: "tool_call", id: "call_1", name: "shell", arguments: { a: 1 } },
+          { type: "tool_call", id: "call_1", name: "shell", arguments: { a: 2 } },
+        ],
+      },
+    ] as unknown as ConversationTurn[];
+
+    const request = adapter.buildRequest(turns, "grok-4.5", {});
+    const body = JSON.parse(request.body) as {
+      input: { type: string; call_id?: string; arguments?: string }[];
+    };
+    const calls = body.input.filter((item) => item.type === "function_call");
+
+    expect(calls).toHaveLength(1);
+    expect(calls[0]?.arguments).toBe(JSON.stringify({ a: 2 }));
+  });
+
   test("does not invent high when no reasoning_effort is set", () => {
     const adapter = createGrokResponsesAdapter(source);
     const turns: ConversationTurn[] = [