infinite looping via iflow

[redzrush101]

Branch

main

Branch (if other)

No response

Version / Tag / Commit

20260123-1-bf7ab7e

Provider(s) Affected

Iflow

Deployment Method

Binary (exe/executable)

Operating System

Fedora Linux 43 Kinoite same issue when launching the binary on a ubuntu podman

Python Version (if running from source)

No response

Bug Description

Glm via iflow is endlessly looping when using opencode and writing a simple message

Steps to Reproduce

1. Launch the proxy binary.
2. Log in to Iflow via OAuth.
3. Add the following snippet to opencode.jsonc so the model appears:

"iflow-glm-4-7": {
  "id": "iflow/glm-4.7",
  "name": "Iflow GLM 4.7",
  "limit": {
    "context": 200000,
    "output": 128000
  },
  "reasoning": true,
  "modalities": {
    "input": ["text"],
    "output": ["text"]
  }
}

I got the information from the curl dump

4. Open Opencode, select the model, and type “hey”.

Attached files

• [response.json](https://github.com/user-attachments/files/24827677/response.json)
• [metadata.json](https://github.com/user-attachments/files/24827678/metadata.json)
• [request.json](https://github.com/user-attachments/files/24827681/request.json)
• [final_response.json](https://github.com/user-attachments/files/24827679/final_response.json)
• [response_stream.log](https://github.com/user-attachments/files/24827675/response_stream.log)
• [request_payload.json](https://github.com/user-attachments/files/24827676/request_payload.json)

If you want it more formal (e.g., for docs or a GitHub issue), I can tune the tone too.

Expected Behavior

It shouldnt loop, claude code works fine

Actual Behavior

The agent is looping

Error Logs / Messages

Attached below

Pre-submission Checklist

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

[mirrobot-agent]

@redzrush101 Thank you for submitting this issue. I am now beginning my analysis and will report back shortly.

[mirrobot-agent]

Initial Analysis Report

Summary: GLM-4 via Iflow enters an infinite loop when used with Opencode (Claude Code) because the terminal finish_reason: "stop" is lost during chunk processing, causing the agent to assume the response is incomplete and triggering retries.

Issue Validation: Confirmed
Reproducibility Assessment: Reproducible

Root Cause Analysis:
The investigation identified a conflict between the iflow provider's chunking logic and the proxy's core streaming wrapper:

1. Split Chunk Logic: src/rotator_library/providers/iflow_provider.py (lines 276-298) splits the final chunk into two:
   • Chunk 1: Has choices (with finish_reason) but no usage data.
   • Chunk 2: Has usage data but no choices.
2. Core Wrapper behavior: The _safe_streaming_wrapper in src/rotator_library/client.py (lines 966-987) only allows a finish_reason to pass through if the chunk ALSO contains usage data (specifically completion_tokens > 0).
3. The Result:
   • Chunk 1 has its finish_reason stripped because it lacks usage data.
   • Chunk 2 is ignored by the finish reason logic because it lacks choices.
   • The client receives [DONE] but never sees finish_reason: "stop".

This behavior is identical in qwen_code_provider.py and likely affects both.

Suggested Labels: bug, provider:iflow, provider:qwen_code

Proposed Next Steps:

1. Combine Final Chunks: Refactor iflow_provider.py and qwen_code_provider.py to yield a single chunk containing both choices and usage when both are present in the raw provider response.
2. Strengthen Core Wrapper: Update the core streaming wrapper to be more resilient to providers that send usage data in separate chunks.
3. Handle Reasoning Tags: (Bonus) Add handling for <thought> or <think> tags in the Iflow provider similar to the Qwen provider to ensure reasoning content is correctly mapped to the thinking block in Anthropic-compatible responses.

This analysis was generated by an AI assistant.

[Mirrowel]

Fixed in latest dev

[redzrush101]

Fixed in the latest dev build.

It now depends on the model glm-4.7 works ig. I used the latest commit, but it seems to break OAuth login, so I logged into iFlow using the stable release first, then switched to the latest dev build after getting the credentials.

[Mirrowel]

So some work and some not? This is.. great. This means format changes based on the model. I'll look