[BUG] GLM4.5 via LMStudio as well as via an OpenAI-compatible endpoint stuck repeating file reads

[mark-ucalgary]

Problem (one or two sentences)

Using GLM4.5 to work on a query about codebase using Ask mode, at a certain point, it gets into a repeating state trying to read a file over and over again. The files are not even that long ~550 lines.

Context (who is affected and when)

Affected when using GLM4.5 model. Tried the exact same task, with MiniMax M2.1 and it worked fine.

Reproduction steps

1. Environment, latest Roo Code.
2. Backends: LM Studio and OpenAI-compatible
3. Model: GLM4.5

Expected result

Keeps working without getting into a repeating loop

Actual result

Gets into a repeating loop of reading the same file.

Variations tried (optional)

No response

App Version

3.45.0

API Provider (optional)

LM Studio

Model Used (optional)

GLM4.5

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

I've analyzed this issue and here's my proposed implementation plan:

Analysis

The ToolRepetitionDetector class in src/core/tools/ToolRepetitionDetector.ts currently detects identical consecutive tool calls by comparing the full serialized JSON (tool name + params + nativeArgs). However, if GLM4.5 is varying parameters slightly between calls (e.g., different line ranges for the same file), each call would be considered unique and bypass detection.

Looking at the screenshots, the model appears to be reading the same file (Notification.tsx) repeatedly. The current detector would only catch this if the calls are byte-for-byte identical, which they may not be if the model varies line numbers or other parameters.

Proposed Solution

Enhance the ToolRepetitionDetector to track similar read_file calls for the same file path, not just identical calls. Specifically:

1. Add path-based tracking for read_file: Track when the same file path is being read repeatedly, even with different line ranges or parameters
2. Keep existing behavior: Maintain current identical-call detection as the primary mechanism
3. Add secondary check: For read_file specifically, also track if the same file path is being read more than N times consecutively, regardless of other parameters

Implementation Details

1. Modify ToolRepetitionDetector.ts:

   • Add a previousReadFilePaths tracking mechanism
   • In the check() method, add special handling for read_file tool that extracts file paths and tracks repetitions by path
   • Reset this tracker when a different tool is used

2. Update tests in ToolRepetitionDetector.spec.ts:

   • Add tests for the new path-based detection for read_file
   • Ensure existing tests still pass

This approach will:

• Catch cases where the model reads the same file with different parameters
• Not affect other tools that legitimately need varied parameters
• Provide a clear error message guiding the model to try a different approach

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[mark-ucalgary]

@roomote The interesting thing is that this is a new problem - reading files worked perfectly with GLM4.5 and a previous version of Roo Code. Can you see if there has been any specific changes to how calls might be made to GLM models via LM Studio or OpenAI-compatible endpoints?

I've seen your comment and will investigate the recent changes. I've identified that commit ed35b09 from earlier today (Jan 29) changed the default behavior for parallel_tool_calls from false to true in the LM Studio handler. This change may be causing the issue with GLM4.5.

I'll explore a fix that makes LM Studio only include the parallel_tool_calls parameter when explicitly enabled, similar to how it was handled for LiteLLM/Bedrock compatibility in a previous fix. I'll open a PR to share progress.

I've identified the root cause: commit ed35b09 from earlier today changed the default behavior for parallel_tool_calls from false to true in the LM Studio handler. Not all models (including GLM4.5) support this parameter, and when sent to models that do not handle it properly, it can cause the model to get stuck in a loop.

Opened a PR with my attempt at addressing this: #11072

The fix only includes the parallel_tool_calls parameter when explicitly enabled, maintaining compatibility with models that do not support it. Feedback welcome!

[mark-ucalgary]

@roomote Note that I also said in the bug report that this affects both LM Studio and OpenAI-compatible endpoints - so I doubt that it is related to the commit you identified.

Thanks for the clarification! You are correct that the issue affects both LM Studio and OpenAI-compatible endpoints, so the parallel_tool_calls fix alone may not be sufficient.

I will investigate the OpenAI-compatible handler as well to check for similar recent changes or other common code paths that could be causing this behavior. I will also revisit the original repetition detection approach I proposed earlier, as it may be the more robust solution for catching this pattern regardless of which provider is being used.

I will update you once I have more findings.

[mark-ucalgary]

@roomote As I mentioned earlier, this is recent problematic behavior in the file reading with GLM4.5. I can say with certainty that this issue did not exist, and file reads worked perfect, with the version of Roo Code at the end of August 2025.

I see your comment about this being a recent regression. Since you mentioned it worked fine with the version at the end of August 2025, I will investigate changes made to the codebase since then that could affect file reading behavior for GLM4.5 via both LM Studio and OpenAI-compatible endpoints.

I will use git log to search for relevant changes in the read_file tool implementation, OpenAI handler, and any common code paths. I will report back with my findings.