Desktop: auto-title silently falls back to full first prompt when structured JSON title validation fails

[rawer886]

What version of the Codex App are you using (From "About Codex" dialog)?

26.903.61454

What subscription do you have?

Unknown (personal model configuration, see Additional information)

What platform is your computer?

Darwin 25.6.0 arm64 arm (macOS)

What issue are you seeing?

New Desktop threads do not get a concise auto-generated title. The sidebar keeps the user's entire first prompt as the title, e.g. a ~5000-character Chinese handoff summary appears in full in session_index.jsonl as thread_name.

Local log evidence (from logs_2.sqlite, table logs) shows the background title-generation thread does complete successfully, yet its result is never applied:

• Title task runs on model gpt-5.6-luna with reasoning_effort=low, sandbox_policy=ReadOnly { network_access: false }, turn_trigger="thread_title".
• Upstream request completes with HTTP 200.
• The task emits Output item item_type="message" with no tool calls and model_needs_follow_up=false needs_follow_up=false.
• session_index.jsonl afterwards still contains the full raw first prompt as thread_name.

The turn start options also contain a strict JSON schema:

final_output_json_schema = {
  "type": "object",
  "properties": {
    "title":       {"type": "string", "minLength": 1, "maxLength": 36},
    "description": {"type": "string", "minLength": 1}
  },
  "required": ["title", "description"],
  "additionalProperties": false
}

while the same prompt text instructs the model:

Output the title as plain text with no surrounding quotes or backticks.

These two instructions contradict each other. The observed outcome (a normal plain message is produced but no title is applied and no error is logged) is consistent with the model following the plain-text instruction, the response failing the JSON-schema check, and Codex silently falling back to the raw first prompt.

What steps can reproduce the bug?

1. Open Codex Desktop and start a new thread.
2. Paste a long, multi-section Chinese prompt (for example a handoff summary with headings, code fences and file paths; ~5000 characters).
3. Wait for background title generation to finish.
4. Check the sidebar / session_index.jsonl: the thread_name is still the full first prompt.

This reproduced 2/2 times today with the same long summary on two different threads. A related failure also occurred with a short prompt containing a commit hash: in that case the title model first ran read-only git show <hash> tools and still no title was applied.

What is the expected behavior?

Either:

• accept a plain-text title produced by the naming model, or
• if structured JSON output is strictly required, make the prompt text consistent with the JSON schema, and log a clear reason (schema validation failure / discarded title) instead of silently keeping the raw first prompt.

Additional information

• I cannot see the raw model response text: helper title threads are not persisted in thread_history_1.sqlite, and the local relay does not log response bodies. The evidence above is from Codex's own logs_2.sqlite.
• The naming model in this setup (gpt-5.6-luna) is served through a local HTTP relay (127.0.0.1:8787) that returns HTTP 200; the problem is the silent fallback, not an upstream error.
• Related but not duplicates: Codex mobile chat titles are often not summarized — the raw user prompt becomes the title #24491 (same symptom on mobile, no root cause), Thread title is auto-generated with malformed/error-like text (e.g. _invalid correction, Oops invalid JSON? I) #12163 (malformed/JSON-error-like titles are applied), Allow custom prompt or template for thread title generation #29677 (custom title template request), Thread title generation burn tokens unnecessarily by inheriting full Codex agent context and tools #40223 (title task inherits full agent context/tools).