🤖 Sanitize malformed tool inputs in chat history

[ammar-agent]

Fixes workspace errors when chat history contains corrupted tool inputs (e.g., string instead of object).

Problem

The httpjail-coder workspace produced this error:

messages.1.content.3.tool_use.input: Input should be a valid dictionary

Root cause: A message in history had a malformed tool input where the AI generated invalid JSON:

{
  "type": "dynamic-tool",
  "input": "{\"script\" timeout_secs=\"10\": \"ls\"}"  // String instead of object
}

When loaded from history and sent to the API, this was rejected.

Solution

Created sanitizeToolInputs() that runs before sending messages to the API:

• Scans all assistant messages for tool parts
• Replaces invalid inputs (string/null/array) with empty objects {}
• Preserves valid object inputs unchanged
• Original history remains untouched

Testing

• 10 tests including reproduction with the actual problematic message
• All edge cases covered (string, null, array, valid objects)
• No existing tests broken

Generated with cmux