[BUG] Code mode does not work with devstral2 123b 2512 (NVIDIA NIM provider)

[Hullabalo]

Problem (one or two sentences)

Expected roo to use tools without error (without blocking error).

Context (who is affected and when)

happens when giving a task to devstral2 123b 2512 model, using Nvidia NIM provider

Reproduction steps

Using devstral2 123b 2512 with Nvidia NIM provider

• ask for a basic coding task in Code mode (same problem occurs when starting in Ask mode)
  -> model returns a first assistant message (e.g. "To create a basic Python code that writes "hi, buddy" 10 times in the console, I can write a simple script for you. Let me proceed with writing the code.")
  -> tool is used correctly (file written or edited)
  -> then, get an error message :
  OpenAI completion error: 400 Unexpected role 'user' after role 'tool' Unexpected role 'user' after role 'tool'

Expected result

model must returns a message or use another tool

Actual result

returns the same error message repeatedly

Variations tried (optional)

starting the same task in Ask mode, there is this added error message after roo switch to Code mode (and then this is the same behavior, like when starting in Code mode)

"I will switch to "💻 Code" mode to write the Python code for you."

then :

Date/time: 2026-01-08T08:58:24.686Z
Extension version: 3.38.3
Provider: openai
Model: mistralai/devstral-2-123b-instruct-2512

The model provided text/reasoning but did not call any of the required tools. This usually indicates the model misunderstood the task or is having difficulty determining which tool to use. The model has been automatically prompted to retry with proper tool usage.

App Version

3.38.3

API Provider (optional)

OpenAI Compatible

Model Used (optional)

mistralai/devstral-2-123b-instruct-2512

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

{
  "error": {
    "timestamp": "2026-01-08T08:49:47.510Z",
    "version": "3.38.3",
    "provider": "openai",
    "model": "mistralai/devstral-2-123b-instruct-2512",
    "details": "400\nOpenAI completion error: 400 Unexpected role 'user' after role 'tool' Unexpected role 'user' after role 'tool'"
  },
  "history": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "<task>\ncreate a basic python code that write 10 times 'hi, buddy' in console\n</task>"
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\nhello_world.py\n\n# VSCode Open Tabs\nhello_world.py\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2026-01-08T08:45:27.836Z\nUser time zone: Europe/Paris, UTC+1:00\n\n# Current Cost\n$0.00\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>mistralai/devstral-2-123b-instruct-2512</model>\n<tool_format>native</tool_format>\n\n\n# Current Workspace Directory (/home/seb/Desktop/TEST) Files\nhello_world.py\nYou have not created a todo list yet. Create one with `update_todo_list` if your task is complicated or involves multiple steps.\n</environment_details>"
        }
      ],
      "ts": 1767861927851
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "text",
          "text": "I'll create a basic Python script that writes \"hi, buddy\" 10 times to the console. Let me write this code to the `hello_world.py` file."
        }
      ],
      "ts": 1767861930979
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "[ERROR] You did not use a tool in your previous response! Please retry with a tool use.\n\n# Reminder: Instructions for Tool Use\n\nTools are invoked using the platform's native tool calling mechanism. Each tool requires specific parameters as defined in the tool descriptions. Refer to the tool definitions provided in your system instructions for the correct parameter structure and usage examples.\n\nAlways ensure you provide all required parameters for the tool you wish to use.\n\n# Next Steps\n\nIf you have completed the user's task, use the attempt_completion tool.\nIf you require additional information from the user, use the ask_followup_question tool.\nOtherwise, if you have not completed the task and do not need additional information, then proceed with the next step of the task.\n(This is an automated message, so do not respond to it conversationally.)"
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\nhello_world.py\n\n# VSCode Open Tabs\nhello_world.py\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2026-01-08T08:45:31.207Z\nUser time zone: Europe/Paris, UTC+1:00\n\n# Current Cost\n$0.00\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>mistralai/devstral-2-123b-instruct-2512</model>\n<tool_format>native</tool_format>\nYou have not created a todo list yet. Create one with `update_todo_list` if your task is complicated or involves multiple steps.\n</environment_details>"
        }
      ],
      "ts": 1767861931210
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "tool_use",
          "id": "Sm5LHOyfs",
          "name": "write_to_file",
          "input": {
            "path": "hello_world.py",
            "content": "for _ in range(10):\n    print('hi, buddy')"
          }
        }
      ],
      "ts": 1767861933933
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "Sm5LHOyfs",
          "content": "{\"path\":\"hello_world.py\",\"operation\":\"modified\",\"notice\":\"You do not need to re-read the file, as you have seen all changes Proceed with the task using these changes as the new baseline.\"}"
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\nhello_world.py\n\n# VSCode Open Tabs\nhello_world.py\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2026-01-08T08:45:35.509Z\nUser time zone: Europe/Paris, UTC+1:00\n\n# Current Cost\n$0.00\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>mistralai/devstral-2-123b-instruct-2512</model>\n<tool_format>native</tool_format>\nYou have not created a todo list yet. Create one with `update_todo_list` if your task is complicated or involves multiple steps.\n</environment_details>"
        }
      ],
      "ts": 1767861935513
    }
  ]
}

[roomote]

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

Root Cause Analysis

The error "Unexpected role 'user' after role 'tool'" occurs because NVIDIA NIM (and some other strict OpenAI-compatible APIs) do not allow a "user" role message directly after a "tool" role message. Looking at the conversation history in the logs:

1. After a tool_result is sent (which becomes a role: "tool" message)
2. The environment_details is sent as a separate role: "user" message
3. NVIDIA NIM rejects this sequence

Solution

The codebase already has a solution for this exact problem: the mergeToolResultText option in convertToOpenAiMessages(). This option merges text content (like environment_details) into the last tool message instead of creating a separate user message. It's already used by:

• DeepSeek (for thinking models)
• Z.ai provider
• MiniMax (via a similar approach)

Proposed Implementation

Option 1 (Recommended): Add a new setting for OpenAI Compatible providers

Add a new setting openAiStrictToolMessageOrdering (or similar name) that, when enabled, passes mergeToolResultText: true to convertToOpenAiMessages(). This approach:

• Doesn't break existing users
• Allows users with strict providers like NVIDIA NIM to opt-in
• Is explicit about its purpose

Files to modify:

1. packages/types/src/provider-settings.ts - Add the new setting type
2. src/api/providers/openai.ts - Use the setting when calling convertToOpenAiMessages()
3. Webview settings UI - Add toggle for the new setting

Option 2: Always enable mergeToolResultText for OpenAI-compatible providers

This would be simpler but more invasive - always merge text content into tool messages for all OpenAI-compatible calls. This is technically more compatible but changes behavior for all users.

---

I believe Option 1 is the safer approach. Before implementing, I'd like to confirm:

1. Is adding a new user-facing setting acceptable, or would you prefer Option 2 (always enable)?
2. Any preference on the setting name? Suggestions: openAiStrictToolMessageOrdering, openAiMergeToolResultText, or openAiCompatibilityMode

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

[jeepshop]

Can confirm that Devstral-Small-2507 running in Ollama is also now broken - when previously it ran with tools just fine.

I really miss having Devstral because it's one of the better local coding models. Qwen3 might score better on coding tasks, but I found that it's tool calling support in previous Roo builds would fail ~15% of the time whereas Devstral would only fail ~5% of the time.

Other local models - that previously ALL worked with native tool calling in previous RooCode builds.

Works: Mistral-Small-3.1-24B-Instruct-2503
Works: Qwen3-Coder-30B-A3B-Instruct
Works: Nemotron-Orchestrator-8B
Fails: Nemotron-3-Nano-30B-A3B
Fails: Devstral-Small-2507
Fails: Devstral-Small-2-24B-Instruct

// Please share this file with Roo Code Support (support@roocode.com) to diagnose the issue faster
// Just make sure you're OK sharing the contents of the conversation below.

{
"error": {
"timestamp": "2026-01-09T15:57:08.328Z",
"version": "3.39.1",
"provider": "openai",
"model": "Devstral-2507-128k:Q6",
"details": "400\nOpenAI completion error: 400 registry.ollama.ai/library/Devstral-2507-128k:Q6 does not support tools"
},
"history": [
{
"role": "user",
"content": [
{
"type": "text",
"text": "\nWhat kind of model are you?\n"
},
{
"type": "text",
"text": "<environment_details>\n# VSCode Visible Files\n..\..\..\Program Files\Microsoft VS Code\tasks,..\..\..\Users\xxx\AppData\Local\Temp\roo-diagnostics-f80a6349-1767974200046.json\n\n# VSCode Open Tabs\nAI Instruct/AspDocumentationPlan.md,AI Knowledge/Notes.md,../../../Users/xxx/AppData/Local/Temp/roo-diagnostics-f80a6349-1767974200046.json\n\n# Current Mode\nask\n❓ Ask\nFTI-Devstral-2507-128k:Q6\n<tool_format>native</tool_format>\n\n\n# Current Workspace Directory (c:/Development/Lockheed/FTI.Afirm.Legacy.Doc) Files\n.roo/\n.vscode/\nAI Instruct/\nAI Knowledge/\nDatabases/\nDatabases - Unreferenced/\nDocumentation/\nImport Batch Files/\nLM Releases/\nMCP Server/\nOld Notes/\n\n(File list truncated. Use list_files on specific subdirectories if you need to explore further.)\nYou have not created a todo list yet. Create one with update_todo_list if your task is complicated or involves multiple steps.\n</environment_details>"
}
],
"ts": 1767974225045
}
]
}

[jeepshop]

Is it possible that nether Devstral 1 or 2 ever worked with Native tool calling, even though I had set to Native. Did RooCode silently switch to XML mode when Native didn't work previously?

If I try a simple direct tool call sample against Ollama directly I get the exact same error that RooCode is reporting;

curl http://localhost:11434/api/chat -d '{ "model": "FTI-Devstral-2507-128k:Q6",  "messages": [    { "role": "user", "content": "What's the weather in Paris?" }  ],  "tools": [    {      "type": "function",      "function": {        "name": "get_weather",        "description": "Get weather for a city",        "parameters": {          "type": "object",          "properties": { "city": { "type": "string" } },          "required": ["city"]        }      }    }  ]}'

Additionally if I view the extract the model card, indeed there is NO tool support in the template (at least the unsloth GUFF models from hugging face)

{{ if .System }}<s>[SYSTEM_PROMPT]{{ .System }}[/SYSTEM_PROMPT]{{ end }}{{ if .Prompt }}[INST]{{ .Prompt }}[/INST]{{ end }}{{ .Response }}</s>

So I did some MORE investigation and discovered that the Devstral2 that is hosted on Ollama.com DOES have a complete template that includes tool calling. And while it DOES correctly answer the curl query above, it still fails in RooCode.

{{- $lastUserIndex := -1 }}
{{- $hasSystemPrompt := false }}
{{- range $index, $_ := .Messages }}
{{- if eq .Role "user" }}{{ $lastUserIndex = $index }}{{ end }}
{{- if eq .Role "system" }}{{ $hasSystemPrompt = true }}{{ end }}
{{- end }}

//System prompt removed for brevity

{{- range $index, $_ := .Messages }}
{{- if eq .Role "system" }}[SYSTEM_PROMPT]{{ .Content }}[/SYSTEM_PROMPT]
{{- else if eq .Role "user" }}
{{- if and (eq $lastUserIndex $index) $.Tools }}[AVAILABLE_TOOLS]{{ $.Tools }}[/AVAILABLE_TOOLS]
{{- end }}[INST]{{ .Content }}[/INST]
{{- else if eq .Role "assistant" }}
{{- if .Content }}{{ .Content }}
{{- if and (not .ToolCalls) (not (eq (len (slice $.Messages $index)) 1)) }}</s>
{{- end }}
{{- end }}
{{- if .ToolCalls }}[TOOL_CALLS]
{{- range .ToolCalls }}{{ .Function.Name }}[ARGS]{{ .Function.Arguments }}
{{- end }}</s>
{{- end }}
{{- else if eq .Role "tool" }}[TOOL_RESULTS]{{ .Content }}[/TOOL_RESULTS]
{{- end }}
{{- end }}

UPDATE:
I also attempted to use Devstral 2 via open router and get the following errors "Unexpected role 'user' after role 'tool' Unexpected role 'user' after role 'tool'".

[bngfavored]

@roomote Option 1 looks best to allow for turning it on. I approve.

[roomote]

Thanks for the approval! I'll proceed with implementing Option 1 - adding a new setting openAiStrictToolMessageOrdering for OpenAI Compatible providers. I'll open a PR shortly to share progress.