Summary

This PR resolves issue #5 by refactoring the failed tool call detection in the infer workflow to use proper JSON parsing instead of permissive regex patterns.

Problem

The current implementation uses regex patterns that are too permissive and can match content that isn't actually failed tool calls:

grep -i "Tool execution failed\|command not whitelisted\|error" /tmp/agent-output.txt

Solution

Replace the regex-based approach with proper JSON parsing using jq to accurately detect when "success": false occurs in tool call results.

Changes

• Extract JSON objects containing "success": false from agent output
• Use jq to parse and format failed tool calls with tool names and error messages
• Maintain backward compatibility with fallback to regex for non-JSON errors
• Properly handle edge cases and errors in JSON parsing
• Clean up temporary files properly

Implementation Details

# Extract JSON objects with "success": false from the agent output
grep -o '{"success": false[^}]*}' /tmp/agent-output.txt > /tmp/failed-tool-calls-raw.txt 2>/dev/null || true

# Format failed tool calls nicely
cat /tmp/failed-tool-calls-raw.txt | jq -r '. | "Tool: \(.tool_name // .name // "unknown"), Error: \(.error // .message // "unknown error")"' 2>/dev/null

Benefits

1. More accurate detection - Only identifies actual failed tool calls with "success": false
2. Better formatting - Provides structured output with tool names and error messages
3. Reduced false positives - Avoids matching unrelated content that contains error-like text
4. Backward compatible - Falls back to original approach if no JSON-formatted failed tool calls are found

Testing

The implementation has been tested to ensure:

• JSON-formatted failed tool calls are properly detected and formatted
• Non-JSON errors still trigger the fallback regex approach
• Edge cases (malformed JSON, missing fields) are handled gracefully
• Cleanup properly removes all temporary files

Resolves #5