[Bug]: Codex V2 run stays "running" after the native turn fails immediately (modelNotFound) #28032

Version: T3 Code Nightly 0.1.6-nightly.20261031.3956, headless t3 serve on Linux amd64. Codex CLI app-server, ChatGPT account.

What happened

A thread was created over MCP (t3_thread_launch) with an explicit Codex model the account cannot use (gpt-5-codex). Codex rejected the turn about 2 s after it started and reported the turn as failed. T3 never finalized the run:

• The thread status stayed running.
• The run's status stayed running, with completedAt: null.
• The thread got no error item. Its only items were the user message and "Workspace ready".
• So nothing watching the run, such as MCP wait/notification clients, saw a terminal state.

The run stayed like this until it was interrupted by hand about 2 minutes later.

Expected

The run ends as failed, with Codex's failure message shown on the thread ("The 'gpt-5-codex' model is not supported when using Codex with a ChatGPT account.").

Native event sequence (provider event log, abridged)

07:16:12.577  → turn/start
07:16:12.586  ← warning  "Model metadata for `gpt-5-codex` not found. Defaulting to fallback metadata…"
07:16:12.616  ← thread/status/changed
07:16:12.616  ← turn/started        turn 01a190d0-e8a6-7d81-98fa-60dfe00d8656
07:16:14.682  ← thread/status/changed
07:16:14.683  ← error              {"error":{"message":"The 'gpt-5-codex' model is not supported when using Codex with a ChatGPT account.","codexErrorInfo":"modelNotFound"},"willRetry":false,"turnId":"01a190d0-e8a6-…"}
07:16:14.690  ← thread/readState/changed
07:16:14.690  ← turn/completed      {"turn":{"id":"01a190d0-e8a6-…","status":"failed","error":{"message":"…not supported…","codexErrorInfo":"modelNotFound"},"durationMs":2098}}
… nothing for this run until …
07:18:05.981  → turn/interrupt   (manual stop)

T3 thread state read about 70 s after turn/completed:

{"status":"running","activeRunId":"run:…:ordinal:1",
 "recentRuns":[{"status":"running","startedAt":"2026-10-31T07:16:12.475Z","completedAt":null}],
 "itemCount":2}

Where it may be

apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts at the tag:

• The turn/completed handler returns early when activeTurns has no context for the turn id. It doesn't log anything, and it never finalizes.
• The error handler (willRetry: false) only records latestProviderFailure after awaitActiveTurn(turnId).

If the turn's context isn't in activeTurns when turn/completed arrives, the terminal event is dropped silently. That can happen if registration races the fast failure, or if the error handler is still waiting in awaitActiveTurn. Not verified; the event timing above is the evidence.

Related but different: #15197 (Claude run stays running after result/success), #27348 (OpenCode turn stuck).

Repro

1. On a Codex instance signed in with a ChatGPT account, launch a thread with modelSelection: {provider: "codex", model: "gpt-5-codex"} (any model the account rejects).
2. Observe the native turn/completed with status: "failed".
3. Observe that the T3 run remains running indefinitely.