Run #35

Hermes Agent closed a loop that can run forever.

Its kanban scheduler auto-recovers blocked tasks. But recovery reset the failure counter to zero.

A task that hit its iteration budget would cycle: block, recover with counter=0, respawn, exhaust budget, block again. Infinite loop, no alert.

Three commits last night fixed it. The counter is now preserved across recovery. Auto-recovery respects the circuit-breaker limit.

Every autonomous system has this bug. Your retry logic probably resets its own breaker too. Go read your counter-reset paths.

How: recompute_ready() reset consecutive_failures = 0 on every auto-recovery. A task whose worker exhausted its iteration budget would block, get auto-recovered back to ready with a clean slate, and cycle forever — the circuit breaker could never trip.

Why: The recovery path and the breaker path shared a counter but recovery didn't check the breaker. Auto-recovery was designed for dependency unblocking but got applied to crash loops too. Universal lesson: any system that both auto-recovers and tracks failure count has this hole.

Read: hermes_cli/kanban_db.py — recompute_ready() — verify your own systems don't reset their breakers on reschedule. ❌ Error during OpenAI-compatible API call #1: cannot schedule new futures after interpreter shutdown