[BUG] Sub-agent stream-idle timeout: next API round-trip never starts mid-conversation, no v2.1.105 watchdog retry

Preflight Checklist

• [x] I have searched existing issues — closest match is #46987 (commented). Filing a separate report because the failure mode is sub-agent specific and the v2.1.105 watchdog isn't engaging.
• [x] This is a single bug report.
• [x] I am using the latest version of Claude Code.

What's Wrong?

Background sub-agents launched via the Task tool stall mid-conversation, then ~10 minutes later die with API Error: Stream idle timeout - partial response received. The 5-minute idle-stream watchdog added in v2.1.105 doesn't engage in any of the cases observed today.

What Should Happen?

Either the watchdog fires at ~5 min and retries in non-streaming mode (as documented), or the request completes normally. Currently it does neither — the agent silently freezes for 5–10 min, then errors with no retry, losing all in-progress work.

Error Messages/Logs

API Error: Stream idle timeout - partial response received

No retry attempt visible in the sub-agent JSONL.

Steps to Reproduce

1. Spawn a long-running sub-agent (Task tool, background) on a non-trivial review or analysis prompt expected to make 30+ tool calls and run 5–15 min.
2. Observe the sub-agent JSONL grow normally for 5–10 min, then stop appending.
3. The last JSONL entry is consistently an assistant text turn ending with the model declaring its next action (e.g. "now let me write the report sections"), with no subsequent tool_use block.
4. After ~3–6 min of silence, the agent dies with the stream-idle timeout error.

Observations

• Sub-agent context (Task tool) only — main session unaffected.
• Pattern reproducible across prompts. ~10 sub-agent runs today (Sunday 19 April), 50–60% failure rate.
• Sequential vs parallel doesn't fully fix it — sequential reduces but doesn't eliminate.
• Failure mode is consistent: API round-trip after a clean assistant text turn never starts (or never returns). Not a hung tool call.
• Same prompts ran cleanly last night (18 April evening) — started failing today.
• Status page reported all components operational throughout. Real-world degraded.
• Hypothesis: weekend deploy or post-15-April-outage aftershock. Sunday low-traffic / typical maintenance window correlation.
• Cost so far: ~half a 5-hour usage window burned on retried sub-agents.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version (last night, 18 April 2026 evening, same prompts).