[BUG] Azure Foundry with OpenAI compatability

[NWessel]

Problem (one or two sentences)

Seems like they are migrating to the new openai endpoint, which instead of "completions" is now "responses"

https://{our instance url}.openai.azure.com/openai/responses?api-version=2025-04-01-preview

In the body, they changed from "messages" to "input" and
from "max_completion_tokens" to "max_output_tokens"

I think roo needs to be updated to support this?
See link: https://platform.openai.com/docs/api-reference/responses/create

Context (who is affected and when)

People using azure foundry with newer models i think?

Reproduction steps

Setup azure foundry with a deployment modal: gpt-5.2-codex

Fill data in as seen in image before

Expected result

Model works and can be prompted

Actual result

API call to model fails because body is wrong

Variations tried (optional)

No response

App Version

Version: 3.41.1 (d49bd9a)

API Provider (optional)

None

Model Used (optional)

gpt-5.2-codex

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

/time: 2026-01-16T08:38:06.712Z
Extension version: 3.41.1
Provider: openai (proxy)
Model: gpt-5.2-codex

400
OpenAI completion error: 400 Unsupported parameter: 'messages'. In the Responses API, this parameter has moved to 'input'. Try again with the new parameter. See the API documentation for more information: https://platform.openai.com/docs/api-reference/responses/create.

[roomote]

I have analyzed this issue and here is my proposed implementation plan:

Problem Analysis

The error occurs because Azure Foundry is serving newer models (like gpt-5.2-codex) through the OpenAI Responses API instead of the traditional Chat Completions API. The key differences are:

Chat Completions API	Responses API
Endpoint: /chat/completions	Endpoint: /responses
Parameter: messages	Parameter: input
Parameter: max_completion_tokens	Parameter: max_output_tokens

Currently, the openai provider (used for OpenAI-compatible endpoints including Azure) only supports the Chat Completions API format, which causes the 400 error when Azure Foundry expects the Responses API format.

Proposed Solution

Add a new configuration option openAiUseResponsesApi to the openai provider that allows users to opt-in to the Responses API format. When enabled:

1. Modify src/api/providers/openai.ts:

   • Add detection for Responses API mode via new setting openAiUseResponsesApi
   • When enabled, convert the request to use input instead of messages
   • Use max_output_tokens instead of max_completion_tokens
   • Use the SDK's responses.create() method or fallback to manual fetch to /responses endpoint

2. Update src/shared/api.ts:

   • Add openAiUseResponsesApi?: boolean to ProviderSettings

3. Update webview UI (settings panel):

   • Add a toggle for "Use Responses API" in the OpenAI provider settings
   • Show this option when OpenAI-compatible provider is selected

4. Leverage existing code from openai-native.ts:

   • Reuse the formatFullConversation() logic for converting messages to Responses API input format
   • Reuse the stream response handling for Responses API events

Alternative Considered

I also considered auto-detecting the Responses API based on the error response, but this would add latency (failed request + retry) and could be confusing. An explicit user toggle is cleaner and gives users control.

Impact

• No breaking changes to existing users (opt-in feature)
• Azure Foundry users can enable this to use newer models
• Minimal code changes by reusing existing Responses API implementation from openai-native.ts

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

[NWessel]

I tried the "normal", "openai" provider as well beside the "openai compatible".
It seems just being able to disable the "prompt_cache_retention" would do the work @roomote

Date/time: 2026-01-16T13:37:39.048Z
Extension version: 3.41.1
Provider: openai-native
Model: gpt-5.2-codex

Error processing response stream: Responses API error: prompt_cache_retention is not supported on this model