[BUG] XML tool calling was removed

[DaveScream]

Problem (one or two sentences)

Pelease return back XML tool calling. My local API service dont support Native tool calling so I cant work now.

Context (who is affected and when)

Native tool calling not supported for Gemini models for my local provider.

Reproduction steps

Return please

Expected result

XML tool calling

Actual result

Feature

Variations tried (optional)

No response

App Version

latest

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-3-pro-preview

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

Thanks for reporting this issue. I would like to help investigate this further.

I see that XML tool calling was intentionally removed in PR #10841 (merged earlier today) as part of an architectural decision to simplify the codebase and reduce maintenance burden. The team believed that native tool calling has been the standard for about a month and no backwards compatibility was needed.

However, I understand this has broken your workflow with your local API provider. To better understand your use case and determine the best solution, could you please clarify:

1. What specific local API provider/service are you using? (e.g., Ollama, LM Studio, llama.cpp server, a custom proxy, etc.)
2. Why does your provider not support native tool calling? Is it a limitation of the provider itself, or is there a configuration option to enable it?
3. Have you tried using the OpenAI-Compatible provider option? Some local providers support OpenAI-style native tool calling.
4. Is this affecting only Gemini models or other models as well?

Understanding your specific setup will help determine whether:

• There is an alternative provider configuration that would work
• Your local provider could be updated to support native tool calling
• A feature request to selectively restore XML support for certain providers is warranted

Please tag @roomote in your response with these details, and I will be happy to propose a solution or escalate to the maintainers.

[SongOnRoc]

@roomote can you restore XML support for OpenAI-Compatible?

[roomote]

I've seen your request to restore XML support for OpenAI-Compatible. I'll explore the codebase to understand what was removed in PR #10841 and work on a potential fix. I'll open a PR to share progress.

[roomote]

Opened PR #10887 with my attempt at addressing this. The PR adds a new openAiNativeToolsEnabled setting that defaults to true (backward compatible) but can be set to false to disable native tool calling for the OpenAI-Compatible provider. When disabled, tools are not passed to the API, allowing local providers that do not support native tool calling to work with XML-style tool invocation. Feedback welcome!

[jeepshop]

What provider/service/model are you using? You mentioned a 'local provider'?

I ask because there are work-arounds. For example I use/used ollama server with devstral-2 which RooCode rightly said doesn't support tools - and it's because the model card in ollama doesn't set the 'tool' support flag NOR does it have a supported template with tools. It depends on where you downloaded the model as well. Sometimes ollama.com models have a good tool template, sometimes not.

BUT you can generally fix it.

Run models from vllm or llama.cpp that you download from huggingface. Almost always the huggingface versions (I prefer unsloth's gguf myself) have a jinja template that supports tools.

Download a ollama.com model that has a working template, and apply it to a downloaded huggingface model that doesn't have a compatible template (ollama uses go templates, llama.cpp and vllm use jinja templates).

For example, download devstral
"ollama pull devstral-small-2"

Show the template
"ollama show --modelfile devstral-small-2:24b"

Apply the template to a bigger model, like "hf.co/unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Q6_K_XL" and save as a new model.