[BUG] XML tool calls are no longer supported. Remove any XML tool markup

[multivac2x]

Problem (one or two sentences)

I am getting several times this error. I am normally using GROK FAST 1 with VS CODE LM API but it's happening also with other models, I did not tried with other providers.

The context length is quite small (near 20K) and I was not able to find a specific pattern to reproduce the issue.

The solution that I found is to start a new task.

The error that I get is the following:
Roo said
<

Error
XML tool calls are no longer supported. Remove any XML tool markup (e.g. <read_file>...</read_file>) and use native tool calling instead.Details

API Request
$0.0000

Roo said
<

Error
XML tool calls are no longer supported. Remove any XML tool markup (e.g. <read_file>...</read_file>) and use native tool calling instead.Details

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.Details

Roo is having trouble...
This may indicate a failure in the model's thought process or inability to use a tool properly, which can be mitigated with some user guidance (e.g. "Try breaking down the task into smaller steps").Details

Context (who is affected and when)

I did not find a reproducible situation.

Reproduction steps

I did not find it yet.

Expected result

Normal Roo Code behaviours to parse and complete the different tasks

Actual result

Blocking error that force to start a new task

Variations tried (optional)

I tried to compress the context and continue, but most of the time that I recall was unsuccessful.

Sometimes I send a new command and I am able to continue the chat.

App Version

3.45.0

API Provider (optional)

None

Model Used (optional)

GROK FAST 1, but also with other models

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I have analyzed this issue and here is my proposed implementation plan:

Analysis

The error "XML tool calls are no longer supported" is triggered in src/core/assistant-message/presentAssistantMessage.ts when the containsXmlToolMarkup() function detects XML-style tool tags (like <read_file>...</read_file>) in the model's text output.

Current behavior:

1. When XML tool markup is detected in a text response, the code increments consecutiveMistakeCount
2. Sets didAlreadyUseTool = true, which stops further processing for that turn
3. Shows the error message to the user
4. The conversation effectively stalls, requiring a new task

Root cause: Some models (especially via VS Code LM API with Grok or other providers) occasionally output XML-style tool markup in their text responses instead of using native tool calling, even when configured for native tools. This can happen sporadically when the model "forgets" the native tool format mid-conversation.

Proposed Fix

I propose improving the error recovery to provide a clearer instruction that guides the model back to using native tools, rather than just stopping. The change would:

1. Modify the error handling in presentAssistantMessage.ts to provide a more instructive error message that explicitly tells the model HOW to call tools properly
2. Include specific guidance about using the native tool calling mechanism instead of XML markup

Specific change in src/core/assistant-message/presentAssistantMessage.ts (around line 296-303):

// Current error message
const errorMessage = "XML tool calls are no longer supported. Remove any XML tool markup (e.g. <read_file>...</read_file>) and use native tool calling instead."

// Proposed change - more instructive message
const errorMessage = "XML tool calls are no longer supported. The model emitted XML-style tool markup in its text response. To call tools, use the native tool calling mechanism provided by the API (function calling), not XML tags in text. Please retry the request using proper tool calls."

This would provide a clearer explanation of what went wrong and how to fix it, which can help the model self-correct on the next turn.

Questions

Before proceeding, I have a few clarifying questions:

1. Would you prefer more aggressive error recovery? For example, we could automatically retry the request when XML markup is detected, rather than just showing an error. However, this could create loops if the model keeps making the same mistake.

2. Should we add model-specific handling? Some providers/models may need different treatment. Would it be valuable to track which models frequently trigger this error to inform future improvements?

3. Is the current detection logic appropriate? The containsXmlToolMarkup() function already excludes content inside code blocks (both fenced and inline). Should we make it more lenient (e.g., only trigger if the XML markup looks like it's actually trying to call a tool with parameters)?

Please mention @roomote in a comment saying that you approve this plan (or with any modifications), and I will proceed with the implementation.

[sameemqureshi]

Any Update on this Issue ?
This issue is also faced in the Latest Version 3.47.3 when Using the Gemini Models

  "error": {
    "timestamp": "2026-02-13T09:02:56.653Z",
    "version": "3.47.3",
    "provider": "gemini",
    "model": "gemini-2.5-pro",
    "details": "Gemini generate context stream error: No output generated. Check the stream for errors."
  },

[sameemqureshi]

@roomote

[roomote]

I've seen your comment and will explore a potential fix for the XML tool call error. I'll look into improving the error recovery so the model can self-correct, and open a PR to share progress. Thanks for reporting this is still occurring with Gemini models on 3.47.3.

[roomote]

Opened PR #11480 with my attempt at addressing this. The fix improves the error recovery when models produce XML-style tool markup by providing comprehensive guidance (tool instructions + next steps) in the error message sent back to the model, making it much more likely to self-correct on the next turn. Feedback welcome!