Codex Mode Runtime Trace — Mode Inside the Desktop App

Part of openai-research. Live capture 2026-08-26 20:57 (BRT 15:57): the user opened the Codex mode (top-level, no submodes), asked a search question ("Se a minha mãe tivesse bolas, ela seria o meu pai.\n\nPesquise online a quem se atribui essa frase.\n" — attribution found: Max Verstappen), then switched the model mid-conversation (gpt-5.6-sol → gpt-5.6-luna, effort low) and asked "que legal, e qual foi a repercussão?" (multi-turn). Evidence: app-server-stdio.jsonl, logs_2.sqlite, state_5.sqlite.

Thread metadata (state_5.sqlite — `01a03fdd-1e38-…`)

model: gpt-5.6-luna (final; Configuring session shows gpt-5.6-sol at 20:57:30 then gpt-5.6-luna at 20:57:42 — the in-conversation switch)
reasoning_effort: low (the UI "light"), sandbox_policy: {"type":"disabled"}, approval_mode: never, memory_mode: enabled
thread_source: user (NOT chatgpt_handoff — direct user-initiated thread), source: vscode
tokens_used: 626,727 — same order of magnitude as the work submode (706,805 on the Taylor thread)
Transport: same app-server; responses_websocket "connected" frames; model refresh GET /backend-api/codex/models?client_version=0.150.0 every ~3min

The tool: `webrun` → `POST /backend-api/codex/alpha/search`

handle_tool_call_with_source: dispatch_tool_call_with_code_mode_result {
  otel.name=webrun tool_name=webrun call_id="exec-a1889571-…" aborted=false }
  : endpoint_session.execute_with { http.method=POST api.path="alpha/search" }:
  Request completed POST https://chatgpt.com/backend-api/codex/alpha/search status=200 OK

The Codex mode web search is the webrun tool, executed via the code-mode host (dispatch_tool_call_with_code_mode_result), hitting alpha/search on the app-server backend — this is the /codex/alpha/search endpoint documented as "Search on demand" in codex-flow-protocol, now confirmed as the tool path of the Codex mode.
Compared to the work submode: work used web_search_call (standalone web search action.search in the Responses stream) — different execution seams.

The task-title agent

A turn-mate system prompt ran for the task created from that prompt: "You are a helpful assistant. You will be presented with a user prompt, and your job is to provide a short title for a task that will be created from that prompt.\nThe tasks typically have to do wi…" — a lightweight title generator that annotates the task/thread from the user text (the longer state_5.threads.title shows the original prompt, not the generated title in this run — the agent result feeds the creation flow).

Events per tool call

app-server event: account/rateLimits/updated and thread/tokenUsage/updated after each tool call — the UI quota/usage updates ride the same app-server event stream (also seen in the work submode).

No `x-codex-turn-state` on the desktop (correction)

x-codex-turn-state appears 0 times in the recent desktop logs (Codex mode, 20:57–21:07). The sticky turn-state routing came from the self-contained probe of 2026-08-24 (API access), not from the desktop app — the desktop uses the responses_websocket session instead. See codex-flow-protocol note.

Codex mode vs work submode (what differs)

Surface	Codex mode	Work submode
Thread source	`user`	`chatgpt_handoff` (from chat handoff) / `user` when direct
Web search tool	`webrun` → `alpha/search`	`web_search_call` (standalone) + `web__run` (CTR)
Scheduling tool	—	`automation_update` (MCP `codex_app`)
Initiation	top-level tab	chat handoff (card) or Work tab
Effort scale	same codex scale (`low` vs `high`)	same (thread `reasoning_effort`)
Quota	codex quota	codex quota (agent turns)

Attestation provider (x-oai-attestation) — same mechanism as work

The app-server attaches x-oai-attestation (header constant X_OAI_ATTESTATION_HEADER, core client generate_attestation_header_for → AttestationProvider::header_for_request) when model_provider.supports_attestation(). The desktop provider is the DeviceCheck addon: each Configuring session is preceded by AttestationGenerate { request_id: 2 } (sol) and { request_id: 3 } (luna) — the token is generated per session/thread, exactly as in the work submode. So flow A (attestation) applies to the Codex mode too (and to work); the sentinel/PoW fallback is the chat-submode/main mechanism.

Standalone web search — same service, two seams (RESOLVED)

From codex-rs tests (core/tests/suite/code_mode.rs:489, core/src/tools/spec_plan_tests.rs):

The Responses-native tool web_search (server-side, ToolSpec::WebSearch with external_web_access/indexed_web_access) is what the model sees when no web.run executor is present (work submode: web_search_call).
With Feature::StandaloneWebSearch + a web namespace tool executor (extension/MCP/CTR), the model sees web.run (not web_search) — and the tool runs as a custom_tool_call "exec" with the JS runtime:
```
const result = await tools.web__run({ search_query: [{ q: "…" }] });
text(result);
```

That runtime (code mode host) performs POST /v1/alpha/search on the app-server → desktop: POST /backend-api/codex/alpha/search, with body:

{"model": "<slug>", "commands": {"search_query": [{"q": "…"}]},
 "settings": {"allowed_callers": ["direct"], "external_web_access": true}}

So the Codex mode webrun and the work submode web_search_call are two seams of the same standalone web search service (the web search backend; the work submode runs it server-side inside the Responses stream — no client alpha/search — while the Codex mode runs it client-side via webrun → alpha/search). WebSearchMode: Live / Indexed / Cached / Disabled (per-turn config).

Task-title agent + onboarding strings

The turn-mate "You are a helpful assistant … provide a short title for a task" system prompt annotates the task being created (this ran in the handoff/creation turn). It is the "task title" service — evidence of a nested agent in the creation path, feeding the wham/tasks creation.

No "phantom tokens"

Searched the app (asar, app-server strings) and codex-rs: no phantom/token-swap endpoint exists. The only token exchange is the link-session (auth_token → __oailb 1h, see auth-and-link-session); no derived sidecar tokens beyond the OAuth family (id/access/refresh) + __oailb.

TypeScript types (code mode surface)

// ── the webrun tool call (tracing) ───────────────────────────────────────
export interface WebrunToolCallTrace {
  otel_name: 'webrun'
  tool_name: 'webrun'
  call_id: string          // 'exec-<uuid>' — the code-mode executor call id
  aborted: boolean
}
// …dispatched through: dispatch_tool_call_with_code_mode_result (code mode host) →
//   endpoint_session.execute_with { http.method: 'POST', api.path: 'alpha/search' } →
//   POST https://chatgpt.com/backend-api/codex/alpha/search  ← the ONLY client-side search seam

export interface AlphaSearchRequest {
  model: string
  commands: { search_query: Array<{ q: string }> }
  settings: { allowed_callers: ['direct']; external_web_access: boolean | 'indexed' }
}
export type WebSearchMode = 'live' | 'indexed' | 'cached' | 'disabled'
export interface WebSearchModeTarget {          // per-turn config
  web_search_mode: WebSearchMode | null         // null = server default
}

// ── the CTR (web namespace) shape ────────────────────────────────────────
export interface WebRunJs {                     // custom_tool_call 'exec' input (code mode)
  // const result = await tools.web__run({ search_query: [{ q: '<query>' }] });
  // text(result);
  search_query: Array<{ q: string }>
}

// ── the task-title agent ─────────────────────────────────────────────────
export const TASK_TITLE_PROMPT: string =
  "You are a helpful assistant. You will be presented with a user prompt, and your job is to provide a short title for a task that will be created from that prompt."
// (the run: the creation path of wham/tasks; the threads.title kept the original prompt in the observed run)

// ── code mode thread metadata (state_5) ──────────────────────────────────
export interface CodeModeThreadMeta {
  id: string
  model: 'gpt-5.6-sol' | 'gpt-5.6-luna' | string
  reasoning_effort: 'low' | 'medium' | 'high' | string
  sandbox_policy: { type: 'disabled' }
  approval_mode: 'never'
  memory_mode: 'enabled'
  thread_source: 'user' | 'chatgpt_handoff' | 'realtime_voice' | string
  source: 'vscode'
  tokens_used: number                    // 626,727 observed
}

Cross-references

codex-flow-protocol — backend endpoints, quota snapshot
work-runtime-trace — the work submode loop (the analogous trace)
chatgpt-desktop-architecture — the mode hierarchy
bezetacil — how the capture happens