Qwen models running in endless loops.

[MaliNamNam]

Branch

main

Branch (if other)

No response

Version / Tag / Commit

latest

Provider(s) Affected

Qwen

Deployment Method

From source (python)

Operating System

Ubuntu

Python Version (if running from source)

3.11.14

Bug Description

Qwen is running in an endless loop as seen in the image.

Steps to Reproduce

.

Expected Behavior

.

Actual Behavior

.

Error Logs / Messages

.

Pre-submission Checklist

• I have searched existing issues to ensure this is not a duplicate

[mirrobot-agent]

@MaliNamNam Thank you for submitting this issue. I am now beginning my analysis and will report back shortly.

[mirrobot-agent]

Initial Analysis Report

Summary: Qwen models may enter endless loops due to a combination of "dummy tool" injection for API stabilization and problematic handling of finish_reason during reasoning tag processing.

Issue Validation: Confirmed (based on codebase analysis and known behavior of agentic clients).

Reproducibility Assessment: Reproducible (can be reproduced by triggering a tool call to the injected dummy tool or ending a response with a reasoning tag).

Root Cause Analysis:

1. Dummy Tool Injection: src/rotator_library/providers/qwen_code_provider.py (line 218) injects a tool named do_not_call_me when no tools are provided to prevent Qwen API stream corruption. If the model calls this tool, the proxy correctly (but problematically) reports finish_reason: "tool_calls". Agentic clients like OpenCode will then execute this unknown tool and send the result back, which can cause the model to call it again in an infinite loop.
2. finish_reason Loss: In _convert_chunk_to_openai (line 332), chunks containing <think> or </think> tags are split, but the resulting chunks hardcode finish_reason: None. This discards the actual finish_reason (e.g., stop) if it was present in the original chunk, leaving the client waiting for a termination signal that never arrives.
3. Reasoning Context Leak: Reasoning content between tags is not tracked across chunks; if a chunk doesn't contain the tags, reasoning content is sent as standard content, which may confuse certain clients.

Suggested Labels: bug, provider:qwen, priority:high.

Proposed Next Steps:

1. Modify _convert_chunk_to_openai to preserve and pass the original finish_reason to the last split part of a chunk.
2. Investigate if the "dummy tool" can be handled internally by the proxy or if its description can be improved to further discourage model calls.
3. Implement state tracking for reasoning blocks to ensure content is correctly mapped to reasoning_content even in chunks without tags.

Missing Information (if any):

• Confirmation of which client is being used (e.g., OpenCode, standard OpenAI SDK).
• If possible, a log excerpt from logs/proxy_debug.log to confirm if a tool call to do_not_call_me is occurring.

This analysis was generated by an AI assistant.

[Mirrowel]

Fixed in latest dev