Tool Error Handling

Layered strategy:

1. Transient errors (timeout, 503): retry at code level (exponential backoff, max 3).
2. Permanent errors (404, 403): return a clear error to the model — "City 'XYZ' not found in DB. Try a valid city name or use search_city first."
3. Schema validation errors: "Parameter days must be 1-30, got -1" — model can correct its wrong parameter.
4. Fallback tools: if search_v1 fails, try search_v2.
5. Capability degradation: if a critical tool is down, send a graceful user message ("I can't search right now — could you check X manually?").

Pro tip: make error messages educational so the model learns from them. Replace "Error 500" with "Tool internal error. Retry in 30 seconds or use alternative tool X."