@ai-sdk/azure provider does not work with /chat/completions since v1.3.4

[chunglam2525]

Description

@ai-sdk/azure provider does not work since v1.3.4

here is my opencode.json config

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "foundary": {
      "npm": "@ai-sdk/azure",
      "options": {
        "baseURL": "https://custom-domain/openai",
        "apiKey": "xxxxxxxx",
        "apiVersion": "2025-04-01-preview",
        "useDeploymentBasedUrls": true,
        "useCompletionUrls": true
      },
      "models": {
        "gpt-5.4": {
          "name": "GPT-5.4"
        }
      }
    }
  },
  "autoupdate": false
}

I have looked up a lot of similar issue regarding azure openai api, seems like everyone is confuse about azure multiple supporting API format.

Expect:
POST: https://custom-domain/openai/deployments/{deployment-id}/chat/completions?api-version={api-version}

{
  "model": "gpt-5.4",
  "messages": [
    ...
  ],
  ...
}

Actual:
POST: https://custom-domain/openai/deployments/{deployment-id}/responses?api-version={api-version}

{
  "model": "gpt-5.4",
  "input": [
    ...
  ],
  ...
}

This line of code sdk.chat(modelID) should direct me to the chat completion method with useCompletionUrls options based on the logic

changing @ai-sdk/azure version to 3.0.49 in opencode v1.3.4 might cause this issue to appear

Plugins

No response

OpenCode version

v1.3.4

Steps to reproduce

1. prepare a azure api that have (only) this API format available - https://custom-domain/openai/deployments/{deployment-id}/chat/completions?api-version={api-version}
2. use opencode version v1.3.3 with this config - it works
3. use opencode version v1.3.4 with same config - it does not works

Screenshot and/or share link

No response

Operating System

macOS

Terminal

WezTerm

[rekram1-node]

are u getting a 404 not found? I managed to hit azure completions just fine

[chunglam2525]

are u getting a 404 not found? I managed to hit azure completions just fine

I did get Resource Not Found error in response.

I try running debug log and saw it was hitting /response endpoint, which indicates it was using response method from @ai-sdk/azure provider.

I also try forcing the api path with similar to this method, but later found out that the request body json is in different formats with the different error message I received.

One thing that I am uncertain is that in previous opencode versions, I did not even input useCompletionUrls option at all and the api still works fine, hitting the completion endpoint correctly. But based on the code logic, it should default to using response instead.

[rekram1-node]

Oh I think I found it

[rekram1-node]

Wait why is ur provider called "foundary"

If it was called "azure" I think it'd work but it's not?

[rekram1-node]

"useCompletionUrls" is something we have it's only tied to that oen provider name

[chunglam2525]

Wait why is ur provider called "foundary"

If it was called "azure" I think it'd work but it's not?

It does work after I change the namespace from foundary to azure (tested on opencode v1.3.13)

  "provider": {
    "azure": {
      "npm": "@ai-sdk/azure",
      "options": {
        "baseURL": "https://custom-domain/openai",
        "apiKey": "xxxxxxxx",
        "apiVersion": "2025-04-01-preview",
        "useDeploymentBasedUrls": true,
        "useCompletionUrls": true
      },
      "models": {
        "gpt-5.4": {
          "name": "GPT-5.4"
        }
      }
    }
  }

I thought this is just a display name tag in TUI model select, separating custom provider with default provider, didn't know it have actual mapping functionality

Does this mean provider[key] and provider[key].npm always need to pair up correctly?

[rekram1-node]

No, it's just that deployment / completion urls thing is kinda an old artifact, im not even sure if it is documented there has got to be a cleaner solution.

But it has been hard coded specifically for that provider this whole time