[BUG] Tool calling (XML) should be kept available! Proper regression testing required.

[jabbor]

Problem (one or two sentences)

Everything worked well until v. 3.36.16. For local models (which I use) tool calling needed to be set to XML, and then everthing works fine. Now: since upgrading to 3.37, I've lost hours trying all kind of things because with Ollama and LMStudio, no model works anymore. Restored v. 3.36.16, and everthing works fine, because I can use XML for tool calling.

Most probably many users would agree with me to request you to spend more time/effort on regression testing, it saves everybody a lot of frustration, and fast releases with added bugs do not help anyone in any way.

Apart from that: Merry Christmas!

Context (who is affected and when)

At least users running models locally!

Reproduction steps

Latest version RooCode 3.37 installed, Started llama.cpp, Ollama, LMStudio with devstral 2 small 2512, only API errors.

Expected result

Proper tool calling

Actual result

Error messages saying context size too small

Variations tried (optional)

Tried llama.cpp, Ollama, LMStudio.

App Version

3.37

API Provider (optional)

LM Studio

Model Used (optional)

devstral-small-2-24b-instruct-2512

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:

Root Cause Analysis

In PR #10281 (v3.37), XML tool protocol was deprecated and forced to native for all new tasks. This breaks local model users (Ollama, LMStudio, llama.cpp) whose models may not fully support native tool calling.

A fix was created (commit fda4984) to restore the Tool Protocol selector for OpenAI Compatible provider, but it hasn't been merged to main yet. The same approach needs to be extended to Ollama and LMStudio providers.

Proposed Implementation

1. Modify src/utils/resolveToolProtocol.ts - Extend the user preference check to include "ollama" and "lmstudio" providers in addition to "openai":

   // For local providers, respect user preference
   const localProviders = ["openai", "ollama", "lmstudio"]
   if (localProviders.includes(providerSettings.apiProvider) && providerSettings.toolProtocol) {
     return providerSettings.toolProtocol
   }

2. Add Tool Protocol selector to Ollama.tsx - Add a dropdown/radio selector allowing users to choose between "native" and "xml" tool protocols, similar to what was done for OpenAI Compatible

3. Add Tool Protocol selector to LMStudio.tsx - Same as above for LMStudio provider

4. Add tests - Update resolveToolProtocol.spec.ts to cover the new Ollama and LMStudio provider handling

This approach will:

• Allow local model users to select XML tool calling when their models don't support native tools
• Maintain native as the default for users who don't explicitly change the setting
• Be consistent with the pattern already established for OpenAI Compatible

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

[hannesrudolph]

@jabbor I suspect there may be some tweaks to our implementations for those providers you mentioned. Are the issues you are running into caused by the providers not talking properly to you ollama etc or are the models themselves not not compatible with NTC?

Merry Xmas to you too

[jabbor]

I'm sorry to say I can't answer your question: I'm just an end user, not a developer. As said in my bug report: RooCode gave an API error, and when I clicked on details, it said that maybe the context size was too small. The only thing I can tell you is that when switching to the previous 3.36.16 version and activating XML tool calling made everything work perfect again! Good luck with bug hunting!

…

________________________________ From: Hannes Rudolph ***@***.***> Sent: Tuesday, December 23, 2025 9:40 PM To: RooCodeInc/Roo-Code ***@***.***> Cc: jabbor ***@***.***>; Mention ***@***.***> Subject: Re: [RooCodeInc/Roo-Code] [BUG] Tool calling (XML) should be kept available! Proper regression testing required. (Issue #10312) [https://avatars.githubusercontent.com/u/49103247?s=20&v=4]hannesrudolph left a comment (RooCodeInc/Roo-Code#10312)<#10312 (comment)> @jabbor<https://github.com/jabbor> I suspect there may be some tweaks to our implementations for those providers you mentioned. Are the issues you are running into caused by the providers not talking properly to you ollama etc or are the models themselves not not compatible with NTC? Merry Xmas to you too — Reply to this email directly, view it on GitHub<#10312 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AL5TMOUXJDZKGS4ZMQ5VKWL4DGSDPAVCNFSM6AAAAACP4E46NGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMOBXHEYTEMJRGU>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[hannesrudolph]

Did this error happen in a chat that was carried on from 3.36.16 to 3.37.0? If not, can you please try that and then create a new issue with the details of that specific bug?