[WIP] after human cancel on claude code cli, the copilot-api keep return 400

[Copilot]

• Analyzed the issue: "after human cancel on claude code cli, the copilot-api keep return 400"
• Explored existing codebase and found tool conversation cleanup logic already exists
• Built and tested current implementation - all existing tests pass
• Identified edge case: current logic only removes tool_use blocks when followed by user messages
• Created test case that reproduces the issue: tool_use followed by assistant message without tool_result
• Update cleanup logic to handle assistant messages after abandoned tool_use
• Ensure all tests pass with the updated logic
• Verify the fix resolves the 400 error scenario

The Fix:
Modified validateAndCleanupToolConversations function to detect abandoned tool_use blocks more comprehensively:

• Before: Only removed tool_use blocks when followed by subsequent user messages
• After: Removes tool_use blocks when followed by any subsequent messages (user or assistant)

This fixes the issue where Claude Code CLI cancellations could leave tool_use blocks in an invalid state, causing subsequent 400 errors from the Anthropic API.

Changes Made:

1. Updated logic in tool-conversation-cleanup.ts to check for any subsequent messages, not just user messages
2. Added comprehensive test cases for edge scenarios including assistant messages after tool_use
3. Added specific test with exact tool ID from error message (toolu_vrtx_01GG3T6npgARqCjxkn9Fm9KW)
4. Added test for long conversations (216+ messages) to verify the fix works in complex scenarios
5. Updated documentation to reflect the broader scope of abandonment detection
6. All existing tests continue to pass, ensuring backward compatibility

Root Cause:
The issue occurred when a tool_use block was followed by an assistant message instead of a user message. The previous logic only considered user messages as indicators of tool abandonment, but assistant messages can also indicate that a tool call was cancelled or bypassed.

Test Coverage:

• ✅ Original Claude Code CLI cancellation scenario (existing test)
• ✅ Tool_use followed by assistant message without tool_result (new test)
• ✅ Long conversations with abandoned tools (new test)
• ✅ Exact tool ID from error message (new test)
• ✅ Mixed scenarios with partial completions (existing test)
• ✅ Valid pending tool calls at end of conversation (existing test)

Fixes #1.

---

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.

[hao1939]

please take the fix: https://github.com/qike-ms/copilot-api/pull/1/files

[hao1939]

try again

[Copilot]

please take the fix: https://github.com/qike-ms/copilot-api/pull/1/files

I can't access external URLs directly. Could you please share the specific changes from that PR that you'd like me to implement? You can either:

1. Copy/paste the relevant code changes here
2. Describe what specific fix you want me to apply

The current implementation handles abandoned tool_use blocks by detecting when they're followed by any subsequent messages without tool_result blocks. If there's a different approach or additional handling needed, I'm happy to implement it once I understand what changes you'd like me to make.

[Copilot]

try again

I can't access external URLs directly. Could you please share the specific changes from that PR that you'd like me to implement? You can either:

1. Copy/paste the relevant code changes here
2. Describe what specific fix you want me to apply

The current implementation handles abandoned tool_use blocks by detecting when they're followed by any subsequent messages without tool_result blocks. If there's a different approach or additional handling needed, I'm happy to implement it once I understand what changes you'd like me to make.