fix: add timeout and retry recovery for subagent LLM streams

[timvw]

Summary

Fixes #13841 — Explore subagent hangs indefinitely with Anthropic Claude Opus 4.6 (and other providers) due to four compounding gaps in timeout/retry handling.

Changes

1. Default 300s fetch timeout (provider.ts)

The config schema documented a 300s default, but the field was .optional() with no enforced default. When unconfigured, options["timeout"] was undefined and the AbortSignal.timeout() guard was skipped entirely. Combined with Bun's socket timeout being explicitly disabled (timeout: false), there was no fallback timeout of any kind.

Now: const timeout = options["timeout"] ?? 300_000 — always applies the documented default unless explicitly set to false.

2. Stream-idle watchdog (processor.ts)

A connection that stays open but stops delivering SSE chunks was never detected. The LLM.stream() call only passed the session's AbortSignal with no idle detection.

Now: withIdleTimeout() async generator wraps stream.fullStream with a 2-minute idle timer that resets on each received chunk. On timeout, throws a StreamIdleTimeoutError (marked retryable via message-v2.ts).

3. Session-level retry cap (processor.ts)

The retry loop had no maximum attempt count — attempt++ with no ceiling. If errors stayed retryable (e.g. rate limits, overloaded), it looped forever.

Now: MAX_RETRY_ATTEMPTS = 10 — after exceeding the cap, the error is surfaced to the user instead of retrying.

4. Subagent timeout (task.ts)

The Task tool awaited SessionPrompt.prompt() with no timeout wrapper. Cancellation only happened when the parent session was manually aborted.

Now: wraps the prompt call with abortAfterAny(10 * 60 * 1000, ctx.abort) — 10-minute hard ceiling that also cascades to SessionPrompt.cancel().

Testing

• All 980 existing tests pass (0 failures)
• Changes are minimal and surgical — 59 lines added, 6 removed across 4 files

Related Issues

• Tasks/Subagents with Codex / OpenAI are frequently getting stuck with no timeout/retry, which then hangs the session forever #11865 — Same symptom with Codex/OpenAI
• Subagent AI_APICallError is silently swallowed, causing session to hang #13395 — Subagent API errors silently swallowed
• main agent hangs because of subagent (explore)  #9003 — Main agent hangs because of subagent (explore)
• Infinite retry loop when StreamIdleTimeoutError occurs during tool input generation #12234 — Infinite retry loop on StreamIdleTimeoutError

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my searches, one result stands out as potentially related:

Potential Related PR:

• fix: retry on timeout errors instead of failing #13502: fix: retry on timeout errors instead of failing (fix: retry on timeout errors instead of failing #13502)
  • This PR addresses retry logic on timeout errors, which is directly related to PR fix: add timeout and retry recovery for subagent LLM streams #13846's focus on timeout and retry recovery

However, all other searches returned only PR #13846 (the current PR), indicating there are no other open PRs with substantial overlap on subagent hangs, stream idle detection, or the specific timeout/retry recovery mechanisms being implemented in this PR.

The related PR #13502 appears to address timeout retries but may not be the exact same fix. You should verify if #13502 is already merged or if it addresses the same root causes as #13846.