fix: normalize incomplete tool requests at extraction layer

[gdeyoung]

Problem

When LLMs output valid JSON with tool_name but missing or malformed tool_args, json_parse_dirty() returned the incomplete dict as-is. This causes ValueError: Tool request must have a tool_args (type dictionary) field crashes in validate_tool_request().

This is particularly common with:

• Smaller models that omit tool_args when empty
• Models that return tool_args as a JSON string instead of a dict
• Models that use "tool" instead of "tool_name"

Root Cause

The extraction layer (json_parse_dirty()) passes through whatever DirtyJson parses without ensuring the result has the required fields for a tool request. The validation layer then crashes on incomplete input.

Fix

Normalize tool requests at the extraction layer in json_parse_dirty() before returning them:

• Missing tool_args → defaults to {}
• tool_args as string → converts to dict
• tool_args as non-dict (number, null, etc.) → defaults to {}
• "tool" key → normalized to "tool_name"

Non-tool requests (no tool_name or tool key) are left untouched.

Why This Approach

This fixes the root cause rather than band-aiding at the validation layer. Defense-in-depth (try/except at validation) is still good practice, but the extraction layer should guarantee a well-formed tool request structure.

Testing

Verified with 7 test scenarios:

• ✅ Missing tool_args → defaults to {}
• ✅ tool_args as string → converts to dict
• ✅ tool_args as number → defaults to {}
• ✅ "tool" key → normalized to "tool_name"
• ✅ Non-tool request → untouched
• ✅ Complete request → passes through unchanged
• ✅ thoughts + tool_name but no tool_args → normalized

Related: #1458, #1466