[rekram1-node]

When Copilot's GET /models fails (e.g. a proxy blocks it, as reported in #51466), the plugin rewrote every models.dev model to @ai-sdk/github-copilot, so Claude went to /chat/completions instead of Anthropic Messages.

The fallback now follows the models.dev package (anomalyco/models.dev#9277 adds the per-model packages):

• @ai-sdk/anthropic → stays on native Anthropic Messages; the base URL becomes <account Copilot endpoint>/v1 when known. Catalog variants were already resolved for Messages.
• @ai-sdk/openai → bundled Copilot SDK with endpoint: "responses".
• anything else → bundled Copilot SDK; the existing model ID fallback picks chat or Responses.

Until the models.dev change ships, untagged models behave exactly as before. The loaded /models path is unchanged.

Verification

• New plugin tests for the Anthropic, account-endpoint, and Responses cases fail without the change; bun test test/plugin/provider-github-copilot.test.ts and bun typecheck pass in packages/core.
• Live: opencode run --standalone against a local proxy that returns 503 for GET /models and forwards everything else to Copilot, with models.dev patched as in [FEATURE]: support for .ts agent files #9277:

model	before	after
claude-haiku-5.5#high	POST /chat/completions	POST /v1/messages (200, tool calls)
gpt-5.4-mini	—	POST /responses (200)
gemini-3.8-flash	—	POST /chat/completions (200)

[opencode-agent]

The change looks sound. In the fallback, Anthropic-tagged models now stay on Messages with <account endpoint>/v1, and OpenAI-tagged ones go to Copilot Responses. That matches what CopilotModels.build does when /models loads, so the two paths agree, and untagged models behave as before.

I couldn't run it against Copilot (no credentials). The Copilot plugin tests pass (32/32), the 3 new ones fail with the base plugin code, and typecheck in packages/core passes.

Public API is unchanged; this only affects Copilot's internal model routing when GET /models fails.

One small test concern, noted inline: the account-endpoint test polls with no limit.

[opencode-agent]

This loop has no limit and runs against a real server, so if the routing breaks again the test hangs until the runner's timeout instead of failing with a clear message. Could you cap it (e.g. times/Schedule.recurs or an Effect.timeout), or wait on the plugin's reload directly if there's a hook for it?

[rekram1-node]

Capped in a83094c with times: 200 (~2s of 10ms real-time polls), so a regression falls through to the assertions instead of hanging. Effect.timeout/Schedule would use the suite's TestClock, which does not advance here, and the plugin has no reload hook to await. Checked by removing the /v1 base URL assignment: the test fails after ~2.3s with Expected "http://localhost:…/v1", Received "https://api.githubcopilot.com/v1".

[opencode-agent]

That fixes it. In a83094c, Effect.repeat({ until, times: 200 }) at provider-github-copilot.test.ts:446-450 caps the loop, and the wait comes from the real Bun.sleep(10) instead of a schedule. So TestClock isn't involved, and the test can't hang. When the cap is hit, the last model value goes on to the assertions on lines 451-452, so a regression still fails with the baseURL diff, as you saw. I agree with your reasons for not using Effect.timeout or Schedule here.

I ran bun test test/plugin/provider-github-copilot.test.ts at pr/head in a temporary worktree: 32 passed, 0 failed. I didn't repeat your check with the /v1 assignment removed. I removed the worktree afterwards and didn't change the shared checkout.