Send parallel_tool_calls true on Codex Responses requests (#969)

A live probe against the ChatGPT Codex backend accepts parallel_tool_calls:true (HTTP 200 on POST /codex/responses, echoed true in response.created), so the host rewrites the field the packaged adapter still pins to false. Related to CL-7420.

Author: TheGreatAxios

docs/IMPLEMENTATION.md

@@ -405,7 +405,7 @@ Positional arguments after flags are joined into the optional initial task deliv
 
 - OpenAI-compatible chat completions, streamed via `@intx/inference`
 - JSON-schema tool definitions for director-layer tools (`ask_operator`, `present`, `submit_output`) and agent tools (`manage_tasks`, `tool_search`, `use_skill`, `search_agents`, …)
-- Codex Responses (`@corbits/codex-provider` via `codex-responses.ts`): ChatGPT Codex is Responses-only. The host system prompt is wrapped with product identity as a leading `developer` input item; conversation turns follow in order. There is no official-prompt fetch, cache, or startup refresh. It requires `store: false` (`store: true` → 400) and rejects `previous_response_id`. Multi-turn continuity is full `input` replay; encrypted reasoning captured via `include: ["reasoning.encrypted_content"]` is resent as a `reasoning` item. `prompt_cache_key` (session id) is the cache-routing signal. `parallel_tool_calls` is sent `false` (serial at this surface); the reactor already fans out a multi-call batch concurrently. `max_output_tokens` is omitted (backend rejects it).
+- Codex Responses (`@corbits/codex-provider` via `codex-responses.ts`): ChatGPT Codex is Responses-only. The host system prompt is wrapped with product identity as a leading `developer` input item; conversation turns follow in order. There is no official-prompt fetch, cache, or startup refresh. It requires `store: false` (`store: true` → 400) and rejects `previous_response_id`. Multi-turn continuity is full `input` replay; encrypted reasoning captured via `include: ["reasoning.encrypted_content"]` is resent as a `reasoning` item. `prompt_cache_key` (session id) is the cache-routing signal. `parallel_tool_calls` is sent `true` (CL-7420 live probe: backend 200s and echoes it true; the packaged adapter still pins `false`, so the host rewrites it on the wire); the reactor already fans out a multi-call batch concurrently. `max_output_tokens` is omitted (backend rejects it).
 
 ### State Persistence
 

src/provider/codex-responses.test.ts

@@ -61,6 +61,19 @@ describe("createCodexResponsesAdapter", () => {
     expect(body.prompt_cache_key).toBe("sess-1");
   });
 
+  test("sends parallel_tool_calls:true on the wire (CL-7420)", () => {
+    const adapter = createCodexResponsesAdapter(source);
+    const request = adapter.buildRequest(
+      [userTurn("hello")],
+      "gpt-5.1-codex",
+      {},
+    );
+    const body = JSON.parse(request.body) as {
+      parallel_tool_calls?: unknown;
+    };
+    expect(body.parallel_tool_calls).toBe(true);
+  });
+
   test("wraps the system prompt with host product identity as a developer item", () => {
     const adapter = createCodexResponsesAdapter(source);
     const request = adapter.buildRequest([userTurn("x")], "gpt-5.1-codex", {

src/provider/codex-responses.ts

@@ -51,9 +51,27 @@ function withHostReasoningEffort(adapter: ProviderAdapter): ProviderAdapter {
   };
 }
 
+// CL-7420 probe: the ChatGPT Codex backend accepts parallel_tool_calls:true
+// (HTTP 200 on POST /codex/responses, echoed true in response.created). The
+// packaged adapter still pins false, so the host rewrites the field on the
+// wire; the reactor already fans out a multi-call batch concurrently.
+function withParallelToolCalls(adapter: ProviderAdapter): ProviderAdapter {
+  return {
+    ...adapter,
+    buildRequest: (messages, model, options) => {
+      const request = adapter.buildRequest(messages, model, options);
+      const body = JSON.parse(request.body) as Record<string, unknown>;
+      body["parallel_tool_calls"] = true;
+      return { ...request, body: JSON.stringify(body) };
+    },
+  };
+}
+
 // CodexQuirks are required by the package factory; Responses option keys are
 // not part of that bag, so host reasoning_effort is aliased at buildRequest.
 export const createCodexResponsesAdapter: AdapterFactory = (source) =>
   withHostReasoningEffort(
-    createPackageCodexResponsesAdapter(source, CODEX_QUIRKS),
+    withParallelToolCalls(
+      createPackageCodexResponsesAdapter(source, CODEX_QUIRKS),
+    ),
   );