feat(llm): add Doubao (Volcengine Ark) as an LLM provider

[MackDing]

What

Adds Doubao (豆包 / ByteDance) via Volcengine Ark as a first-class LLM provider in Continue, following the existing pattern used for Moonshot, Deepseek, MiniMax, zAI, and SiliconFlow.

Doubao is one of the most-used LLM families in China; Volcengine Ark (火山方舟) is its official hosted platform and exposes an OpenAI-compatible /chat/completions surface at https://ark.cn-beijing.volces.com/api/v3/. Today users have to work around the gap by using the generic OpenAI provider with apiBase and model overrides — this PR makes provider: doubao just work.

Why

• Demand: Doubao is ByteDance's flagship LLM and is widely deployed in mainland-China-facing projects.
• No new surface area: Ark is OpenAI-compatible, so the adapter is a ~30-line OpenAI subclass. Registration is the same shape used by Moonshot / Deepseek / zAI.
• Closes an implicit feature gap that existing users of China-region models have been asking about (visible in several closed Qwen/Deepseek provider-error issues where users cite Doubao as the alternative).

Changes

File	Purpose
core/llm/llms/Doubao.ts	New Doubao extends OpenAI with Ark base URL
core/llm/llms/index.ts	Register in LLMClasses
core/llm/autodetect.ts	Add to PROVIDER_HANDLES_TEMPLATING (Ark handles chat templating server-side, same as Moonshot/Deepseek)
packages/openai-adapters/src/apis/Doubao.ts	Adapter wrapping OpenAIApi with ark.cn-beijing.volces.com/api/v3/
packages/openai-adapters/src/types.ts	DoubaoConfigSchema in the discriminated union
packages/openai-adapters/src/index.ts	constructLlmApi → DoubaoApi
packages/openai-adapters/src/apis/Doubao.test.ts	3 new tests (routing / default base URL / preserved OpenAI-chat surface)
core/llm/llms/OpenAI-compatible.vitest.ts	Doubao row in the shared subclass test matrix
docs/customize/model-providers/more/doubao.mdx	User-facing config docs

Diff: 9 files, +222 / −0.

Design notes

• No default model is set. Ark addresses models by either a date-stamped model ID (e.g. doubao-seed-1-6-251015, doubao-1-5-pro-32k-250115) or a user-provisioned endpoint ID (ep-20240xxx-xxxxx). A bare alias like doubao-1-5-pro-32k would silently 404, so the adapter intentionally requires the user to pick a valid ID. The docs link to the official Ark model list so users can copy a currently valid one.
• maxStopWords = 4 matches Ark's documented limit for OpenAI-compatible chat completions.
• No FIM override. Ark does not expose a public beta/completions FIM protocol today. If that changes, we can override fimStream the way Moonshot and Deepseek do.
• Not registered in packages/llm-info — consistent with Moonshot and Deepseek, which also rely solely on the LLMClasses + openai-adapters registration.

Testing

packages/openai-adapters$ npm test
Test Files  15 passed | 2 skipped (17)
     Tests  148 passed | 5 skipped (150)

3 of those 148 are new Doubao tests covering:

1. constructLlmApi({ provider: "doubao" }) returns a DoubaoApi instance.
2. Default apiBase is the Ark cn-beijing v3 URL.
3. Inherits the OpenAI-chat surface (chatCompletionStream / chatCompletionNonStream).

TypeScript check: tsc --noEmit on packages/openai-adapters → 0 errors.

References

• Volcengine Ark docs root: https://www.volcengine.com/docs/82379
• Ark model list (date-stamped IDs): https://www.volcengine.com/docs/82379/1330310
• Ark OpenAI-SDK compatibility notes: https://www.volcengine.com/docs/82379/1330626

---

Summary by cubic

Adds Doubao (ByteDance) via Volcengine Ark as a first-class LLM provider using the OpenAI-compatible /chat/completions API. Also fixes apiBase handling so user overrides are respected over the default.

• New Features

  • New doubao provider with default apiBase: https://ark.cn-beijing.volces.com/api/v3/.
  • Registered in LLMClasses and constructLlmApi; added DoubaoConfigSchema.
  • Autodetect updated so Doubao handles chat templating server-side.
  • Docs added for configuration; tests cover routing and API surface.

• Migration

  • No default model. Set a valid Ark model ID (e.g. doubao-seed-1-6-251015) or endpoint ID (ep-...) in model.
  • If you previously used provider: openai with Ark apiBase, switch to provider: doubao and keep the same model and apiKey.

^{Written for commit b0857a5. Summary will update on new commits.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

2 issues found across 9 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="docs/customize/model-providers/more/doubao.mdx">

<violation number="1" location="docs/customize/model-providers/more/doubao.mdx:15">
P2: Configuration instructions are internally contradictory about using model aliases versus date-stamped model IDs/endpoint IDs, which can lead to invalid setup.</violation>
</file>

<file name="packages/openai-adapters/src/apis/Doubao.ts">

<violation number="1" location="packages/openai-adapters/src/apis/Doubao.ts:19">
P2: Subclass field initialization overwrites `config.apiBase` set by `OpenAIApi`, breaking custom base URL configuration.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[MackDing]

I have read the CLA document and I hereby sign the CLA

[MackDing]

Quick follow-up: all CI on my side looks green now. If the remaining blocker is the automated review feedback or pending AI checks, I’m happy to submit a follow-up patch. If there is a maintainer-preferred direction for the Doubao provider integration, I can adjust accordingly. Thanks for taking a look.

[MackDing]

Thanks for the review. Both findings addressed in b0857a5:

Doubao.ts apiBase override — Removed the class field initializer that was clobbering the parent OpenAIApi constructor's config.apiBase ?? this.apiBase resolution. The default is now passed through super() config via a DEFAULT_DOUBAO_API_BASE constant, so user-provided apiBase overrides work as documented.

doubao.mdx model ID guidance — Replaced the ambiguous "use a public model alias" wording with explicit instructions to use a date-stamped model ID or an endpoint ID, plus a warning that undated aliases are not guaranteed to resolve on Ark.

[MackDing]

Quick follow-up after rechecking the PR status: the standard GitHub checks are passing on my side, and the earlier Cubic findings were addressed in b0857a5. The remaining blockers appear to be maintainer review plus pending Continuous AI checks / two cancelled packaging jobs. Happy to make any further adjustments if there is a preferred direction for the Doubao provider integration.

[MackDing]

Additional check on the remaining VSIX blockers: both failing build-and-upload-vsix jobs appear to have built and uploaded the artifacts successfully, then failed during post-job cache cleanup with The operation was canceled. I tried rerunning the failed jobs, but GitHub requires repository admin rights for that. The remaining action needed from my side appears to be maintainer/admin rerun plus maintainer review; I do not see a code build failure in those job logs.