/agent-triage — Smart Failure Triage

Failed tasks fall into a small number of buckets. Classify, then act.

Prerequisites

Endpoints used:

• GET /api/tasks?status=failed&since=<window> — list failed tasks
• GET /api/tasks/:id/log — full log for one task
• POST /api/tasks/:id/requeue — requeue
• POST /api/tasks/:id/dead-letter — dead-letter

Workflow

Phase 1: Pull failures

curl -s -H "Authorization: Bearer $AGENT_API_TOKEN" \
  "$AGENT_API_URL/api/tasks?status=failed&since=24h"

Phase 2: Classify each failure

For each failed task, fetch its log/last error and assign one classification:

Phase 3: Present grouped summary

TRIAGE — <N failures> in last 24h
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

INFRA           <count>  → bulk requeue
  • <task-id> <slug>  — <one-line cause>
  • <task-id> <slug>  — <one-line cause>

AUTH            <count>  → fix credentials, then requeue
  • <task-id> <slug>  — <missing/expired creds>

CODE            <count>  → requeue with amendment OR dead-letter
  • <task-id> <slug>  — <error type>  — <attempt N/3>

PROMPT          <count>  → operator amends prompt
  • <task-id> <slug>  — <pattern: looped / hallucinated / max-iter>

ENV             <count>  → fix worker env, then requeue
  • <task-id> <slug>  — <missing dep / wrong version>

WONTFIX         <count>  → dead-letter
  • <task-id> <slug>  — <reason>

UNKNOWN         <count>  → /agent-inspect each
  • <task-id> <slug>

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PROPOSED ACTIONS
  Bulk requeue:    <N> tasks (infra)
  Bulk dead-letter: <N> tasks (wontfix)
  Operator action:  <N> tasks (auth/prompt/env)
  Inspect deeper:   <N> tasks (unknown)

Phase 4: Bulk action

After presenting, offer bulk operations. Do NOT execute without explicit user approval.

# Bulk requeue infra failures (after user confirms)
for tid in $INFRA_TASK_IDS; do
  curl -s -X POST -H "Authorization: Bearer $AGENT_API_TOKEN" \
    "$AGENT_API_URL/api/tasks/$tid/requeue"
done

Heuristics for log → classification

Common patterns to grep in logs:

When in doubt, classify as unknown and recommend /agent-inspect.

Notes

• Default window is 24h. For longer windows (/agent-triage 7d), sweep is heavier — paginate accordingly.
• Don't auto-classify based on task slug or project name — read the actual error.
• A single task may match multiple patterns. Pick the most specific (auth beats code, env beats prompt).
• Never bulk-act without showing the user the count + class breakdown first. A misclassification that bulk-requeues 50 prompt failures will create 50 more failures.