Problem

The moonshotai/kimi-k2-instruct model (and similar Kimi variants) produce malformed JSON in tool call arguments, causing cascading failures:

• Unterminated string errors at char 1
• Missing closing braces/brackets
• Empty tool name fields
• Truncated key-value pairs

The current retry logic (3 retries) amplifies the problem since the same malformed JSON gets re-serialized unchanged.

Solution

This PR adds a model-specific JSON sanitizer that:

1. Detects known problematic models via is_kimi_model() check
2. Attempts repair of common malformation patterns:
   • Adds missing closing quotes/braces
   • Removes truncated key-value pairs
   • Falls back to regex extraction as last resort
3. Hooks into validation loop at run_agent.py:8820-8834
4. Fails gracefully if sanitizer unavailable (ImportError caught)

Files Changed

• tools/kimi_json_sanitizer.py - New sanitizer module
• run_agent.py - Hook into JSON validation for repair attempt before declaring invalid

Testing

Tested against the following malformed JSON patterns:

'{"tool": "", "arguments": {"command": "'  # Missing tool name
'{"tool": "terminal", "arguments": {"command": "ls -la", "timeout"'  # Unterminated
'{"tool": "read_file", "arguments": {"path": "/tmp/test.txt"'  # Missing bracket

Notes

• Only activates for models matching *kimi* pattern
• No impact on other models/providers
• Logging added for repaired vs regex-extracted cases

This enables use of capable-but-janky models without polluting the core agent loop with model-specific hacks.