Summary

Fixes session corruption by skipping tool call extraction from assistant messages with stopReason: "aborted" or stopReason: "error".

Problem

When users interrupt requests mid-stream or errors occur during LLM responses, tool_use blocks may be incomplete or malformed. The current extractToolCallsFromAssistant() function extracts these incomplete tool calls and generates synthetic tool_result blocks for them.

This causes the Anthropic API to reject the entire session with:

HTTP 400: unexpected tool_use_id found in tool_result blocks

This triggers auth profile cooldowns and can cascade through all available profiles, making the gateway unresponsive.

Solution

Added a check at the beginning of extractToolCallsFromAssistant() to skip tool call extraction when stopReason === "aborted" or stopReason === "error":

const stopReason = (msg as { stopReason?: unknown }).stopReason;
if (stopReason === "aborted" || stopReason === "error") {
  return [];
}

Testing

All existing tests pass (10/10). The test suite already included tests for this exact scenario:

• ✅ skips tool call extraction for assistant messages with stopReason 'error' (line 117)
• ✅ skips tool call extraction for assistant messages with stopReason 'aborted' (line 140)
• ✅ drops orphan tool results that follow an aborted assistant message (line 180)

These tests were previously failing but now pass with this fix.

Impact

• High: Eliminates the most common source of session corruption
• Low Risk: Small, focused change with comprehensive test coverage
• Backwards Compatible: Only affects error handling, no API changes

Related Issues

Fixes #6788

Follow-up Work

This PR addresses the primary issue (aborted/errored messages). Future PRs could address:

• [Bug]: Tool Use/Result Pairing and Consecutive Assistant Messages After History Limiting #4650: Orphaned results after history limiting
• Claude Opus 4.6 fails with tool_use.id format validation error when switching models #10222: Tool ID format validation across models