Self-Heal

Recover autonomously from failures. Diagnose before retrying — blind retries waste tokens and amplify mistakes.

Recovery Decision Tree

Error detected
├── Transient? (network, rate limit, timeout)
│   └── → Retry with exponential backoff (max 3×)
├── Tool failure? (bad args, wrong schema)
│   └── → Fix args, retry once; if still fails → escalate
├── Logic error? (wrong output, assertion failure)
│   └── → Diagnose root cause → apply fix → verify → continue
├── Context overflow?
│   └── → Compress context → resume from last checkpoint
└── Unknown?
    └── → Log state, pause, surface to human