Merge pull request #784 from corbitsdev/cl-6694-track-parallel_tool_calls-and-responses-continuity-as-follow

Document Codex Responses continuity and serial tool calls

Author: TheGreatAxios

docs/IMPLEMENTATION.md

@@ -358,6 +358,7 @@ 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 (`codex-responses-adapter.ts` `buildRequest`): ChatGPT Codex is Responses-only. 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).
 
 ### State Persistence
 

src/provider/codex-responses-adapter.ts

@@ -30,8 +30,15 @@ import { PRODUCT_NAME, ENVIRONMENT_TAG_NAME } from "../branding.js";
 // Credentials and the chatgpt-account-id ride through differently: the access
 // token is injected by the harness via the bearer sentinel, while the account
 // id and session id travel in `source.defaults.providerOptions` (merged into
-// InferenceOptions.providerOptions by the harness) and are lifted into headers
-// here. Neither is placed in the request body.
+// InferenceOptions.providerOptions by the harness). Account id is headers-only
+// (`chatgpt-account-id`). Session id is the `session_id` header and
+// `prompt_cache_key` in the body — the only cache-routing signal under
+// `store: false`.
+//
+// Continuity is not Responses store chaining. The ChatGPT Codex backend
+// requires `store: false` (`store: true` → 400) and rejects
+// `previous_response_id`. Every turn resends the full `input`; encrypted
+// reasoning captured from the prior stream is resent as a `reasoning` item.
 
 export const CODEX_RESPONSES_PROVIDER = "codex-responses";
 
@@ -342,6 +349,10 @@ function buildRequest(
     store: false,
     stream: true,
     include: ["reasoning.encrypted_content"],
+    // Serial at the request layer. The reactor already executes a multi-call
+    // batch concurrently; this flag is what the ChatGPT Codex backend is sent.
+    // Do not flip without verifying the backend accepts true — unlike store /
+    // previous_response_id there is no recorded 400.
     parallel_tool_calls: false,
   };
   // The Codex backend rejects `max_output_tokens`; it is intentionally omitted.

tests/unit/codex-responses-adapter.test.ts

@@ -47,6 +47,8 @@ describe("codex-responses buildRequest", () => {
     expect(body["model"]).toBe("gpt-5-codex");
     expect(body["stream"]).toBe(true);
     expect(body["store"]).toBe(false);
+    expect(body).not.toHaveProperty("previous_response_id");
+    expect(body["parallel_tool_calls"]).toBe(false);
     expect(body["include"]).toEqual(["reasoning.encrypted_content"]);
     expect(body["prompt_cache_key"]).toBe("sess-1");
     expect(body["input"]).toEqual([