Document Codex Responses continuity and serial tool calls

The ChatGPT Codex backend rejects store chaining, so every
turn resends full input plus encrypted reasoning.

Parallel tool calls stay off at this request layer because
backend acceptance of true is unverified, not because the
runner cannot fan out.

Author: TheGreatAxios

docs/IMPLEMENTATION.md

@@ -361,6 +361,7 @@ 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([