Error Classification

# Error Classification

Distinguish between retryable and permanent failures.

## Category
Tool Resilience

## Context
Error handling in any tool.

## Problem
Agents don't know whether to retry or try something different.

## Solution
Classify errors clearly:
- **Retryable**: Transient, will likely succeed later
- **Permanent**: Won't succeed without changes
- **User action needed**: Requires human input
- **Authentication**: Need to re-auth

## Example (Python)
```python
class ErrorType(Enum):
    RETRYABLE = "retryable"      # Try again later
    PERMANENT = "permanent"       # Won't work without changes
    AUTH_REQUIRED = "auth"        # Need to re-authenticate
    USER_INPUT = "user_input"     # Need human input

@tool
def api_call(endpoint: str) -> Result:
    try:
        return api.call(endpoint)
    except RateLimitError:
        raise ToolError(
            type=ErrorType.RETRYABLE,
            message="Rate limited",
            retry_after=60
        )
    except NotFoundError:
        raise ToolError(
            type=ErrorType.PERMANENT,
            message="Resource not found"
        )
    except AuthError:
        raise ToolError(
            type=ErrorType.AUTH_REQUIRED,
            message="Token expired"
        )
```

## Considerations
- Use consistent error types across tools
- Include retry-after for retryable errors
- Map external API errors to your classification

## Related Patterns
- recovery-guide
- idempotent-operation
- graceful-degradation

---
Source: https://arcade.dev/patterns/error-classification