Skip to content

[BUG] XML-style tool calls are no longer executed starting with version 3.33.0 onwards when using OpenAI Compatible provider #9385

@MasuRii

Description

@MasuRii
Image Image

Problem (one or two sentences)

Starting with version 3.33.0, perfectly valid XML-style tool calls produced by Gemini are no longer recognized or executed when using the OpenAI Compatible provider.

Context (who is affected and when)

Everyone who:

  • Uses Roo Code v3.33.0 or newer
  • Runs Gemini models (Gemini 2.5 Flasg, Gemini 2.5 Pro, Gemini 3 Pro Preview, Flash, etc.) via the OpenAI Compatible provider

Reproduction steps

  1. Update to Roo Code v3.33.0 or later.
  2. Set up OpenAI Compatible provider pointing to a Gemini model (e.g. Gemini 2.5 Pro or Gemini 3 Pro Preview) – temperature 0.7, all other settings default.
  3. Start any normal task that requires sub-task delegation (coding, research, etc.).
  4. The model correctly outputs clean XML tool calls such as:
<new_task>
  <mode>code</mode>
  <message>...</message>
</new_task>

or

<update_todo_list>
  <todos>...</todos>
</update_todo_list>
  1. Nothing happens – the tool call is completely ignored, no new sub-task is created, TODO list never updates, and the run stalls or continues incorrectly.

Expected result

XML tool calls are parsed and executed immediately (new Code/Research task appears, TODO list updates, etc.)

Actual result

Perfectly formatted XML tool calls from Gemini are ignored → agent gets stuck or produces wrong results.

Variations tried (optional)

  • Works perfectly in v3.33.0 & v3.33.1 (same Gemini model & settings)
  • Tested with Gemini 2.5 Flash, Gemini 2.5 Pro, Gemini 3 Pro Preview → all fail the same way
  • Changing temperature or other parameters has no effect

App Version

3.33.1

API Provider (optional)

OpenAI Compatible

Model Used (optional)

Google Gemini 2.5 Pro / Gemini 3 Pro Preview

Roo Code Task Links (optional)

<new_task>
<mode>code</mode>
<message>
Your task is to update the "High" thinking budget value in `browser_utils/thinking_normalizer.py`.
...
</message>
</new_task>
<update_todo_list>
<todos>
[x] Break down the main task into subtasks and create a TODO list.
[x] Delegate subtask to find the code for the 'High' thinking budget.
[-] Delegate subtask to update the token value from 24000 to 32768.
[ ] Delegate subtask to verify the change and check for related UI validation.
[ ] Synthesize results and report completion.
</todos>
</update_todo_list>

These are never processed in v3.33.0+ when running Gemini via OpenAI Compatible.

Relevant logs or errors (optional)

Thanks for looking into this!

Metadata

Metadata

Assignees

No one assigned

    Labels

    Issue - Needs InfoMissing details or unclear. Waiting on author to provide more context.bugSomething isn't working

    Type

    No type

    Projects

    Status

    Issue [Needs Info]

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions