Root Cause

Summary

Fixed a bug where retry-prompt parts were left orphaned in message history after their corresponding tool-call parts were pruned during dangling tool call cleanup. The fix generalizes the filter to remove ANY part with a tool_call_id matching a dangling ID, not just tool-call parts.

Context

When a tool call fails (e.g., web_fetch returns 403), pydantic-ai creates a retry-prompt part that references the original tool_call_id. This is different from a successful tool-return part. The sanitization logic at sanitize.py:111-142 was only filtering parts with part_kind == "tool-call", leaving retry-prompt and tool-return parts behind.

Log evidence from tunacode.log 2026-01-25:

[PRUNED] tool_call part: tool=web_fetch id=call_XoXFHWUpi1KujOYx3zHTR1KN
...
history[2].part[0] kind=retry-prompt tool=web_fetch id=call_XoXFHWUpi1KujOYx3zHTR1KN

The tool-call was pruned but the retry-prompt referencing it remained, creating an invalid message history.

Root Cause

_filter_dangling_tool_calls_from_parts() used a strict part_kind == "tool-call" check to decide what to prune. This missed:

• retry-prompt parts (pydantic-ai's error response for failed tool calls)
• tool-return parts (though these would be rare to be orphaned)

Both of these part types have a tool_call_id attribute that references the original tool call.

Changes

• sanitize.py:111-142: Changed _filter_dangling_tool_calls_from_parts() to filter ANY part that has a tool_call_id attribute matching a dangling ID, regardless of part_kind. The function now:

  1. Checks if the part has a tool_call_id attribute
  2. If so, checks if that ID is in the dangling set
  3. If so, prunes the part (logging the part_kind for debugging)

• test_tool_call_lifecycle.py: Added test_remove_dangling_removes_orphaned_retry_prompt_parts() to cover this specific scenario with a mock RetryPromptPart.

Behavioral Impact

What users notice:

• Sessions that previously failed after a tool error (like 403 from web_fetch) followed by user abort will now recover correctly on the next request
• No more "orphaned tool response" errors in the API

What didn't change:

• Normal tool call/return matching still works
• Matched tool calls are still preserved
• The detection logic (find_dangling_tool_calls) is unchanged - it correctly identifies calls without returns

Related Cards

• [[dangling-tool-calls-on-abort]] - PR #246, the original fix for dangling tool calls on UserAbortError
• [[gate-6-exception-paths]] - Design principle about exception path handling