fix(opencode): chatgpt subscription stream hangs due to watchdog timeout

[liuhaoyang]

Problem

ChatGPT subscription streams can hang indefinitely without timing out, leaving users with frozen connections and no feedback.

Symptoms

• LLM response streams from ChatGPT subscription endpoints stall permanently with no data
• No timeout mechanism detects or recovers from hung connections
• Users experience completely frozen sessions with no error message or recovery path

Root Cause

1. Broken provider request timeout: The existing timeout mechanism used AbortSignal.timeout() which does not work correctly in Bun's runtime. Provider HTTP requests had no effective timeout, allowing connections to hang indefinitely.

2. No stream-level watchdog: The session processor lacked any idle/first-byte timeout on LLM streams. Once a stream stopped producing chunks (or never started), there was no detection or recovery.

3. Incomplete retry classification: Transient server errors (429, 502, 503, 504) and common ChatGPT error shapes were not always classified as retryable, causing permanent failures on recoverable conditions.

4. Missing abort signal propagation: Token refresh calls in the Codex auth plugin did not propagate the caller's abort signal, so cancellation during auth refresh leaked hanging requests.

Impact

• Users on ChatGPT subscription plans experience frozen/hung sessions with no feedback
• No automatic recovery from transient server-side stalls
• Subagent tasks can run indefinitely without timeout bounds

Fix Approach

Repair the stream watchdog and timeout infrastructure across multiple layers:

• Replace broken AbortSignal.timeout() with manual setTimeout + AbortController for provider request timeouts (5 min default)
• Add a stream-level watchdog with separate first-byte (30s) and idle (120s) timeouts that wraps LLM streams
• Extend retry classification to recognize transient server errors (429/502/503/504) and deeply-nested ChatGPT error patterns
• Convert DOMException(TimeoutError) into retryable APIError for proper recovery
• Propagate abort signals through Codex OAuth token exchange and refresh paths
• Add configurable subagent timeout with input validation

[github-actions]

This issue doesn't fully meet our contributing guidelines.

What needs to be fixed:

• No issue template was used. Please use one of the three templates: Bug Report, Feature Request, or Question.
• The issue title uses a conventional commit format (fix(opencode): ...) which is meant for PRs/commits, not issue titles.
• The content appears to be AI-generated: an extensive wall of text with root cause analysis, multi-layer fix proposals, and implementation details. This level of AI-generated detail is not acceptable for issue reports. Please describe the problem you actually observed in your own words.
• If this is a bug report, please include your actual environment details, reproduction steps you tried, and what you observed — not a pre-written fix proposal.

Please edit this issue to address the above within 2 hours, or it will be automatically closed.

---

This issue might be a duplicate of existing issues. Please check:

• Explore subagent hangs indefinitely with Anthropic Claude Opus 4.6 -- no timeout or recovery #13841: Explore subagent hangs indefinitely with no timeout or recovery — covers the same AbortSignal.timeout() gap, stream watchdog absence, and subagent timeout issue
• OpenAI stream intermittently freezes in working state with high CPU and idle HTTPS socket #29129: OpenAI stream intermittently freezes in working state — same symptom (stream stalls indefinitely, no recovery, requires restart)
• [BUG] chunkTimeout does not work for AWS Bedrock and other non-SSE streaming protocols (EventStream) #26487: chunkTimeout does not work for AWS Bedrock and other non-SSE protocols — overlapping issue about stream timeout infrastructure not working correctly

If you believe this was flagged incorrectly, please let a maintainer know.

[thejasonxie]

yeah i'm experiencing lots of hangs as well right now...

[jacob-vincent-mink]

Seconded. Any logs we can provide?

[github-actions]

This issue has been automatically closed because it was not updated to meet our contributing guidelines within the 2-hour window.

Feel free to open a new issue that follows our issue templates.

[arthurfiorette]

Same happening here, anybody found a solution?

[thejasonxie]

nope besides interrupting it and ttelling it continue lol. Seems to happen with deepseek too #29416