[IDE Extension] Runaway task remains active for 12+ hours and consumes 50% of Pro usage

What version of the IDE extension are you using?

openai.chatgpt 26.901.22334 (bundled codex-cli 0.153.0)

What subscription do you have?

ChatGPT Pro

Which IDE are you using?

Visual Studio Code 1.136.1 (x64)

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What issue are you seeing?

A relatively small Unity UI/prefab task remained actively running in the Codex VS Code sidebar for more than 12 hours. At capture time the UI showed Working for 12h 2m 47s, a Stop button, and 105 files changed (+5,812 / -9,349).

The actual affected session was still being written at approximately 2026-09-09 23:55 KST. Its local JSONL transcript had grown to 5,805,728,952 bytes (about 5.8 GB). The latest event was response_item.reasoning; there was no completion event for the active turn.

The latest token_count event recorded 365,386,654 total tokens (364,550,617 input; 358,629,120 cached input; 836,037 output; 276,965 reasoning output). Most input tokens were cached, so this counter may not map 1:1 to billed usage, but the scale strongly indicates runaway execution. The account usage meter showed 50% of the Pro weekly allowance consumed.

The live status message said that compilation passed, a validation check failed because a nested prefab was treated as a normal asset path, and Codex would immediately modify the check and run again. This shows the turn was still performing automatic fix-and-retry work rather than merely displaying stale UI.

Model shown in the extension: GPT-5.6 Sol Ultra.

What steps can reproduce the bug?

1. On Windows, open the Codex IDE extension in VS Code and work locally in an existing Unity project.
2. Select GPT-5.6 Sol Ultra.
3. In an existing conversation, request a targeted UI/prefab consistency change across a few related prefabs, including checking for duplicate elements.
4. Let the agent run and perform its own compile/validation cycles.
5. Observe the task continuing for 12+ hours, repeatedly reporting another fix and rerun, while producing a very large diff and transcript.
6. Observe hundreds of millions of session tokens and substantial Pro usage consumption without a terminal condition.

I cannot yet provide a minimal deterministic reproduction. The affected private session ID and redacted logs can be shared with OpenAI maintainers through a private channel.

What is the expected behavior?

A targeted task should reach a terminal state or stop with a clear error after bounded retries. Codex should not continue automatic fix-and-rerun cycles for 12+ hours, generate hundreds of millions of tokens, or expand to 105 changed files without asking the user. Stop/cancellation should work reliably, and usage caused by a runaway loop should be reviewable.

Additional information

Correction: the initial issue body accidentally referenced a different completed session (18 million tokens). After matching the exact live status text, I identified the actual affected session and replaced those incorrect details with the measurements above.

Please investigate the runaway retry behavior, provide a safe termination/recovery path, and review/restore the Pro usage consumed by this apparent bug.

For privacy, this public report omits the repository name, file names, account identifier, full prompt, screenshot, and session ID. I can provide the affected session ID, screenshot, and a redacted transcript privately to OpenAI maintainers.

Start with the VS Code sidebar task lifecycle, especially the Stop/cancellation path and the automatic compile/validation retry cycle described here. Reproduce with the supplied Windows, VS Code, extension, and Unity setup if possible; done means bounded retries, reliable cancellation, and a terminal state without runaway transcript or usage growth.