[BUG]

[sameemqureshi]

Problem (one or two sentences)

Question was Asked to Roo Code in the Open AI Compatible Configuration profile , The Answer was given , But again attemepted t respond using the attempt_completion tool , which resulted in Model Response Incomplete
The model failed to use any tools in its response. This typically happens when the model provides only text/reasoning without calling the required tools to complete the task. error repeatedly and stuck in a loop .

Context (who is affected and when)

I experienced this on Updating to this 3.37.1 version .
I am Using the Open AI Compatible Profile with a Custom Base URL .

Reproduction steps

1. Setup the Open AI Compatible Profile with a Custom Base URL .
2. Ask Queries like "what is llm ?" from the chat .

Expected result

After the response was completed , the Loop shouldnt have happened . If there is an Issue with the Custom Base Url Underlying Service , It would be very Helpful if Roo can State the errors .

Actual result

There is a loop with not much detailed logging .

Variations tried (optional)

No response

App Version

3.37.1

API Provider (optional)

OpenAI Compatible

Model Used (optional)

gpt-4o

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

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

Analysis

The issue occurs when using OpenAI Compatible mode with a custom base URL. When the user asks a simple question like "what is llm?", the model responds with text but does not use the required attempt_completion tool. Roo Code's design requires tool usage on every response, causing a loop where:

1. Model responds with text only (no tool call)
2. Roo Code sends a "noToolsUsed" error asking the model to retry with a tool
3. Model again responds with text only
4. This continues up to the consecutiveMistakeLimit

The current "grace retry" mechanism allows 2 consecutive failures before showing an error, which means the error message appears after the second failure but the loop continues until consecutiveMistakeLimit is reached.

Proposed Solution

I propose improving the error handling to break the loop more gracefully by:

1. Better error feedback for native tool protocol: Enhance the noToolsUsed response for native protocol to provide clearer, JSON-structured guidance that the model should call attempt_completion if the task is complete.

2. Align consecutive mistake counting: Currently, consecutiveNoToolUseCount is tracked separately from consecutiveMistakeCount. After 2 consecutive no-tool failures, we should stop the loop earlier instead of continuing up to consecutiveMistakeLimit.

3. Add explicit guidance in the noToolsUsed message for native protocol: The current native protocol reminder is very generic. We should include the specific tool call syntax for attempt_completion similar to how XML protocol includes an example.

Code Changes

File: src/core/prompts/responses.ts

• Modify noToolsUsed() to provide better structured guidance for native protocol
• Add explicit attempt_completion example for native tool calling

File: src/core/task/Task.ts (around line 3288-3310)

• Consider stopping the loop after reaching consecutiveNoToolUseCount >= 2 instead of just showing an error and continuing
• Or: increase the severity of the "no tool use" handling to break out of the retry loop

This approach will:

• Help models understand they need to use attempt_completion when done
• Prevent the infinite loop by breaking earlier
• Provide clearer error messages to the user

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

[mrubens]

Any better in 3.37.1?

[sameemqureshi]

This Issue is Faced in the 3.37.1 version ,