fix byok-opencode: route non-OpenAI openai-protocol hosts to @ai-sdk/openai-compatible

[JackChen-me]

Fixes #5338

Why

Hit this myself using Open Design with a DeepSeek key: the built-in DeepSeek BYOK preset (Settings → Execution mode → BYOK) 404s on the first run. The daemon's OpenCode BYOK path maps protocol openai straight to @ai-sdk/openai, which calls the OpenAI Responses API ({baseURL}/responses). DeepSeek (and other OpenAI-compatible providers) only serve /chat/completions, so the request 404s. The web path already treats these as OpenAI-compatible via isOpenAICompatible(...); the OpenCode path didn't, so the two disagreed. Full trace in #5338.

What users will see

BYOK custom models on the openai protocol pointed at a non-api.openai.com endpoint (the DeepSeek preset, vLLM, and other OpenAI-compatible chat-completions hosts) now work instead of failing with Not Found. Real OpenAI (api.openai.com) is unchanged.

Surface area

• Default behavior change — for the openai BYOK protocol, non-api.openai.com hosts now use @ai-sdk/openai-compatible (/chat/completions) instead of @ai-sdk/openai (/responses). This fixes DeepSeek and also switches vLLM / local OpenAI-compatible servers to the chat-completions package (which is what they actually serve). api.openai.com keeps using @ai-sdk/openai.

Screenshots

n/a — no UI.

Bug fix verification

• Test path that reproduces the bug: apps/daemon/tests/runtimes/byok-opencode.test.ts → routes OpenAI-protocol BYOK with a non-OpenAI base URL to the OpenAI-compatible provider package
• Did the test go red on main and green on this branch? yes — on main the DeepSeek case returns @ai-sdk/openai; after the fix it returns @ai-sdk/openai-compatible. The existing vLLM test is updated to match (it is openai-compatible by nature — the test is even named that).

Validation

Ran under Node 24 / pnpm 10.33.2:

• pnpm --filter @open-design/daemon exec vitest run -c vitest.config.ts tests/runtimes/byok-opencode.test.ts → 13/13 pass (new test red before the source change, green after).
• pnpm --filter @open-design/daemon run typecheck → pass.
• pnpm guard → 78/78 pass.

Note: connection-test.test.ts / chat-route.test.ts have a few pre-existing failures on a clean main in my local sandbox — flaky withFakeAgent process-spawn flows, unrelated to this change (which only affects buildProviderEntry's provider-package selection). Left to CI.

Optional follow-ups (kept out of this PR for one-concern): (1) the proxy path normalizes DeepSeek to /v1/chat/completions; this path resolves to /chat/completions (both are served by DeepSeek) — happy to add a /v1 normalization for parity if preferred; (2) surfacing the actual failing request URL instead of a bare Not Found.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[lefarcen]

Hey @JackChen-me — thanks for the sharp write-up and the tight repro here.

Because this changes a live BYOK runtime path that users hit directly, I’m marking it for QA once the PR is merge-ready. Nothing needed from you right now; we’ll follow up here after validation.

[nettee]

@JackChen-me I reviewed the BYOK OpenCode provider routing change and the updated daemon regression coverage. The split keeps api.openai.com on @ai-sdk/openai while moving non-OpenAI openai protocol hosts, including the corrected vLLM case, onto @ai-sdk/openai-compatible, which matches the existing product behavior and directly covers the DeepSeek 404 reported in #5338. Nice focused fix.

_{🔁 Powered by Looper · runner=reviewer · agent=codex · An autonomous AI dev team for your GitHub repos.}

[AmyShang-alt]

QA validation passed for PR #5344 at head e674e44.\n\nEvidence checked:\n- Focused daemon regression test passed: tests/runtimes/byok-opencode.test.ts 13/13.\n- Daemon typecheck passed.\n- pnpm guard passed: 78/78.\n- Local PR runtime launched via tools-dev: web on 127.0.0.1:17545, daemon on 127.0.0.1:17544.\n- BYOK OpenCode with official OpenAI model gpt-5.5 succeeded, confirming the real OpenAI path is unchanged.\n- BYOK OpenCode with non-OpenAI OpenAI-compatible provider via OpenRouter + deepseek/deepseek-chat succeeded, confirming the fixed compatible-provider path works in a real run.\n- Additional local mock endpoint validation showed non-OpenAI host requests went to /chat/completions and did not hit /responses.\n\nConclusion: acceptance passed. The previous needs-validation risk is covered; this PR is ready from QA.

[AmyShang-alt]

Approved after QA validation. Automated checks passed, local PR runtime validation succeeded for official OpenAI BYOK, and non-OpenAI OpenAI-compatible BYOK via OpenRouter + deepseek/deepseek-chat succeeded.

[open-design-bot]

🎉 📡 You just leveled up to Giotto

📡 ✨ Sending steady signals.

🙌 Your contributions are sending a clear signal across the network: you care about making Open Design better. Keep transmitting.

💛 Thanks for helping Open Design move forward. Keep building in the open. 🚀

---

📊 Rank #111 among 400+ contributors

🔗 Share on X (English) · 分享到 X（中文）