BYOK Azure: `wire_api: completions` ignored; hardcoded api-version rejected by Azure OpenAI Responses API

[davidatorres]

Summary

When the GitHub Copilot CLI is configured (via github-copilot-sdk's
ProviderConfig) to use type: azure with wire_api: completions,
the CLI ignores wire_api and sends requests to Azure OpenAI's
Responses API endpoint (/openai/responses) anyway. It also
appears to use a hardcoded api-version (per copilot/session.py:819
SDK docstring, defaults to 2024-10-21) regardless of what's passed
in azure.api_version, which Azure OpenAI rejects for the Responses
API path:

400 Azure OpenAI Responses API is enabled only for api-version
2025-03-01-preview and later

This makes BYOK against Azure OpenAI / Microsoft Foundry deployments
non-functional today.

Environment

• copilot CLI: 0.0.353 (Commit f8fd3e3), installed via Homebrew
• github-copilot-sdk: 0.3.0 (Python wrapper)
• Host: macOS 25.4.0 / arm64
• Upstream: Azure OpenAI deployment provisioned through Microsoft
  Foundry hub (endpoint https://<account>.openai.azure.com)
• Auth: Azure AD bearer token via az account get-access-token --resource https://cognitiveservices.azure.com

Reproducer

import asyncio
from copilot import CopilotClient, ProviderConfig

async def main() -> None:
    client = CopilotClient()
    provider = ProviderConfig(
        type="azure",
        wire_api="completions",       # asks for Chat Completions
        base_url="https://<your-aoai-endpoint>.openai.azure.com",
        bearer_token="<az-ad-token>",
        azure={"api_version": "2025-04-01-preview"},  # explicit modern version
    )
    session = await client.create_session(
        model="gpt-5",
        working_directory="/tmp/test",
        provider=provider,
        on_permission_request=lambda *a, **kw: {"behavior": "allow"},
    )
    response = await session.send_and_wait("Write hello.py")
    print(response)

asyncio.run(main())

Expected behavior

The Copilot CLI sends requests to
<endpoint>/openai/v1/chat/completions (Chat Completions endpoint,
no api-version constraint), OR honors the explicit
azure.api_version passed in the ProviderConfig.

Actual behavior

~/.copilot/logs/process-*.log shows:

[INFO] Using custom provider: type=azure, baseUrl=<endpoint>,
       wireApi=responses    ← wire_api was set to "completions"
[ERROR] {
  "code": "BadRequest",
  "message": "Azure OpenAI Responses API is enabled only for
              api-version 2025-03-01-preview and later"
}

Two distinct issues:

1. wire_api: completions is ignored — log line shows
   wireApi=responses even when SDK input was
   wire_api="completions".
2. azure.api_version is ignored or overridden — the request
   that reaches Azure OpenAI uses a version older than
   2025-03-01-preview despite us setting
   azure.api_version="2025-04-01-preview" in ProviderConfig.

The same error reproduces with wire_api: responses, model: gpt-5.4, and explicit azure.api_version overrides up to
2026-01-01-preview.

Suggested fixes (any one would unblock BYOK Azure paths)

1. Honor wire_api: completions for type: azure — route to
   <endpoint>/openai/v1/chat/completions instead of
   /openai/responses when set. (Cleanest; matches the SDK
   schema's wire_api: Literal['completions', 'responses'] typing.)
2. Honor azure.api_version — pass through to the request URL
   query string (?api-version=...) so operators can opt into
   newer Azure API surfaces.
3. Bump the hardcoded default to 2025-03-01-preview — at
   minimum bring the default in line with current Azure OpenAI
   Responses API requirements.

---

---

Possibly related (different concrete symptoms, same general area):

• BYOK azure provider 404 error #2576 — BYOK azure provider 404 error
• Proposal: External model provider API surface for Copilot CLI (flags/env/config + BYOK) #2710 — Proposal: External model provider API surface for Copilot CLI
• CLI headless mode: internal state accumulation causes progressive latency degradation across sessions #2755 — CLI headless mode: latency degradation under BYOK

[davidatorres]

---

Companion issue: #3209 — Hosted-mode autonomous-codegen behavior gap (the related blocker we hit on the same code-agent abstraction effort)