Repository navigation
GPT-5.5 subagent calls fail with ms=0 / resp_bytes=0 — model crashes immediately #131
Description
Bug Summary
GPT-5.5 subagent API calls fail immediately with zero response time and zero response bytes. The model "crashes out and quits" without generating any output.
Evidence from harness.trace.jsonl
4 API calls failed with is_error: true, ms: 0, resp_bytes: 0, context_tokens: 0:
| t | agent | model | req_bytes | resp_bytes | ms |
|---|---|---|---|---|---|
| 893922 | Frontend signup UX | gpt-5.5 | 4636 | 0 | 0 |
| 893946 | Backend signup behavior | gpt-5.5 | 4747 | 0 | 0 |
| 960805 | Backend signup behavior | gpt-5.5 | 4747 | 0 | 0 |
| 1125313 | judge | gpt-5.5 | 7115 | 0 | 0 |
All failures are subagent calls (workflow/judge), not the main agent. The ms: 0 indicates the request was rejected/errored at the transport layer before any tokens were generated.
Impact
- Subagent workflows silently lose entire branches of analysis
- The main agent gets no results from the failed subagents and may produce incomplete work
- User has to retry or switch models
Suggested root causes to investigate
- Rate limiting on OpenAI GPT-5.5 for concurrent subagent calls
- API key/auth issues specific to the gpt-5.5 model routing
- Request payload validation failure (e.g. content policy, context length)
- Subagent spawning race condition causing duplicate/aborted requests
Environment
- Models: gpt-5.5 (subagents), kimi-k2.6 (main)
- Session: lawbook/lawplain repo, ultracode workflow mode
- File: harness.trace.jsonl
Activity
Merge pull request #150 from justrach/fix/tls-ca-bundle-race-131
fix(behavior): unblock the local stream and make both sinks account f…