bug: tool exceptions are caught and replaced with generic error messages — lose root cause

Problem

When a tool raises an exception during execution, CrewAI catches it and replaces the full traceback with a generic error string like "Error: tool execution failed". The original exception type, message, and traceback are discarded.

This makes debugging impossible — users can't tell whether the failure was:

• A transient network error (should retry)
• A parameter validation error (should fix input)
• A tool bug (should report upstream)
• A rate limit / auth error (should adjust config)

Steps to Reproduce

1. Create a custom tool that raises ValueError("invalid parameter: count must be positive")
2. Configure a crew to use this tool
3. The agent receives only "Error: tool execution failed" with no indication of what went wrong

Expected Behavior

The full exception should be available to:

1. The agent (so it can adapt its strategy — retry, fix input, skip tool)
2. The user/logs (for debugging)
3. At minimum, a structured error object with type, message, and traceback fields

Proposed Fix

# Instead of:
return f"Error: tool execution failed"

# Return:
return json.dumps({
    "error": True,
    "type": type(e).__name__,
    "message": str(e),
    "retryable": isinstance(e, (TimeoutError, ConnectionError)),
})

Related

• #5930 (PDF tool errors are swallowed)
• #5990 (Deepseek format errors are opaque)

Start by tracing the tool execution path where exceptions are caught and generic errors are returned; the issue does not name a file or test. Review related issues #5930 and #5990 for existing behavior, then verify that exception type, message, and traceback information reach the agent and user or logs without losing useful error context.