huskies: done 496_bug_hard_rate_limit_without_reset_at_never_auto_schedules_retry

2026-04-08 00:04:28 +00:00
parent dea410149a
commit ea8e12190b
2 changed files with 34 additions and 25 deletions
@@ -0,0 +1,34 @@
+---
+name: "Hard rate limit without reset_at never auto-schedules retry"
+---
+
+# Bug 496: Hard rate limit without reset_at never auto-schedules retry
+
+## Description
+
+When the API returns a hard rate limit block (`status=rejected`) without a `reset_at` timestamp, `pty.rs` downgrades it to a `RateLimitWarning` instead of a `RateLimitHardBlock`. The auto-scheduler only listens for `RateLimitHardBlock` events, so no timer is set and the agent is never restarted. The agent sits idle until the 300s inactivity timeout kills it, and the story is stuck.
+
+In practice, most hard blocks come without `reset_at` (as seen in the logs: "no reset_at in rate_limit_info"). This means the auto-resume feature from story 423 almost never fires.
+
+Fix: when there's a hard block without `reset_at`, either:
+1. Send `RateLimitHardBlock` with a default backoff time (e.g. `Utc::now() + 5 minutes`)
+2. Or add a separate retry mechanism that doesn't depend on knowing the exact reset time
+
+## How to Reproduce
+
+1. Run an agent that hits the API rate limit
+2. Observe logs show "no reset_at in rate_limit_info"
+3. Agent gets killed by inactivity timeout
+4. Story sits in current with no agent, never restarted
+
+## Actual Result
+
+Hard block without reset_at is downgraded to RateLimitWarning. No timer set. Agent dies and story is stuck.
+
+## Expected Result
+
+Hard block without reset_at triggers a retry with a default backoff (e.g. 5 minutes). Agent is automatically restarted when the backoff expires.
+
+## Acceptance Criteria
+
+- [ ] Bug is fixed and verified