agent-cleanlisted

# /agent-clean — Infrastructure-Failure Purge > Don't conflate "the agent didn't try because we ran out of credits" with "the agent tried and the code is broken." ## The signal Failed tasks in an autonomous agent fall into two completely different categories: 1. **Infra failures** — agent never got a chance: rate limit, credit exhausted, network timeout, auth expired 2. **Real failures** — agent tried, code is broken, prompt was wrong, env was missing Bulk-requeuing all failures conflates these. Bulk-dead-lettering all failures throws away real work. `/agent-clean` separates them: it purges (or requeues) only the infra failures and leaves the real failures for `/agent-triage`. ## Prerequisites | Var | Purpose | |---|---| | `AGENT_API_URL` | Base URL of the agent's HTTP API | | `AGENT_API_TOKEN` | Bearer token | ## Usage ``` /agent-clean — sweep last 24h /agent-clean 7d — sweep last 7 days /agent-clean --requeue — requeue infra failures instead of dead-lettering /agent-clean --dry-run — show what would be done, don't act ``` ## Workflow ### Phase 1: Pull failures ```bash curl -s -H "Authorization: Bearer $AGENT_API_TOKEN" \ "$AGENT_API_URL/api/tasks?status=failed&since=24h" ``` ### Phase 2: Match each failure against infra patterns For each failed task, fetch a tail of its log and match against these patterns: | Pattern | Class | |---|---| | `rate.?limit` / `429 Too Many Requests` | `rate-limit` | | `credit.