[BUG] azure does not support parameters: ['tool_choice']

[statichawk]

Problem (one or two sentences)

Azure hosted GPT-5 returns OpenAI completion error: 400.

Context (who is affected and when)

Roo Code user using Azure hosted GPT

Reproduction steps

1. VS Code extension with OpenAI Compatible API Provider
2. Send a request in Roo Chat

Expected result

Proper GPT response

Actual result

OpenAI completion error: 400 litellm.UnsupportedParamsError: azure does not support parameters: ['tool_choice'], for model=responses/gpt-5-chat. To drop these, set litellm.drop_params=True or for proxy: litellm_settings: drop_params: true . If you want to use these params dynamically send allowed_openai_params=['tool_choice'] in your request.. Received Model Group=gpt-5-chat Available Model Group Fallbacks=None

Variations tried (optional)

Tried our 4.1 model, works fine there.

App Version

v3.39.3

API Provider (optional)

OpenAI Compatible

Model Used (optional)

GPT-5

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

The OpenAiHandler in src/api/providers/openai.ts always includes the tool_choice parameter in requests when native tools are enabled. Azure OpenAI (especially when accessed through LiteLLM proxies) doesn't support the tool_choice parameter for certain models like GPT-5.

Proposed Solution

Add a new boolean setting openAiOmitToolChoice to the OpenAI schema that allows users to optionally omit the tool_choice parameter from API requests.

Changes Required:

1. packages/types/src/provider-settings.ts: Add openAiOmitToolChoice: z.boolean().optional() to the openAiSchema
2. src/api/providers/openai.ts: Modify the request construction to check for this.options.openAiOmitToolChoice before including tool_choice
3. webview-ui: Add a checkbox in the OpenAI provider settings UI to enable this option
4. locales: Add localization strings for the new setting

Example code change in openai.ts:

// Before:
...(metadata?.tool_choice && { tool_choice: metadata.tool_choice }),

// After:
...(metadata?.tool_choice && !this.options.openAiOmitToolChoice && { tool_choice: metadata.tool_choice }),

This approach:

• Is backward compatible (default behavior unchanged)
• Gives users control over which parameters are sent
• Works for any Azure/LiteLLM configuration that doesn't support tool_choice

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

[statichawk]

@roomote approved

[roomote]

I've seen the approval and will proceed with implementing the plan. I'll add the openAiOmitToolChoice setting to allow users to optionally omit the tool_choice parameter for Azure/LiteLLM compatibility. Will open a PR shortly to share progress.

[roomote]

Thanks for the approval! I'll proceed with implementing the openAiOmitToolChoice setting now. I'll open a PR shortly to share progress.

[roomote]

Opened a draft PR with my attempt at addressing this: #10703

The PR adds a new "Omit tool_choice parameter" checkbox to the OpenAI-compatible provider settings. When enabled, it will skip sending the tool_choice parameter in API requests, which should resolve the 400 error for Azure/LiteLLM endpoints that don't support this parameter.

Feedback welcome!

[roomote]

Opened a draft PR with my attempt at addressing this: #10704

The PR adds a new checkbox "Omit tool_choice parameter" in the OpenAI Compatible provider settings. When enabled, it will omit the tool_choice parameter from API requests, which should resolve the 400 error with Azure-hosted models.

Feedback welcome!