feat: add Y-API as an OpenAI-compatible provider

[jiweiyeah]

What

Adds Y-API (y-api) as an OpenAI-compatible provider.

• Website: https://y-api.bestvirtualgoods.com
• API base: https://api.y-api.bestvirtualgoods.com/v1/
• Docs: https://y-api.bestvirtualgoods.com/docs
• Pricing: https://y-api.bestvirtualgoods.com/pricing

It is a gateway that fronts DeepSeek, GLM, Kimi, Qwen, Tencent Hunyuan, Xiaomi MiMo, and OpenAI models behind one key and one balance. Model IDs are namespaced (deepseek/deepseek-v4-pro, z-ai/glm-5.3, moonshotai/kimi-k3, …), so they pass through unchanged — no model-name translation is needed for this provider.

Changes

Follows the CONTRIBUTING checklist for adding an LLM provider:

File	Change
core/llm/llms/YApi.ts	New. Extends OpenAI; apiBase and default model set.
core/llm/llms/index.ts	Import + entry in LLMClasses.
core/llm/llms/OpenAI-compatible.vitest.ts	Coverage via the shared createOpenAISubclassTests helper.
core/llm/toolSupport.ts	PROVIDER_TOOL_SUPPORT["y-api"] — see below.
packages/openai-adapters/src/types.ts, .../index.ts	Provider literal + constructLlmApi case, so the CLI path resolves it.
docs/customize/model-providers/more/y-api.mdx	New documentation page.

The default model is deepseek/deepseek-v4-flash: it is free of charge on the service (no credit is deducted), and it is the model the service's own documentation uses in its examples.

Tool support

PROVIDER_TOOL_SUPPORT is populated deliberately: the default for an unlisted provider is false, which would silently disable agent mode for every Y-API model.

Support was measured against the live API rather than assumed — one request per model with a single function tool, checking for a tool_calls response:

• Tools work for the four DeepSeek models, the three GLM models, moonshotai/kimi-k3, tencent/hy3, and xiaomi/mimo-v2.5.
• The four openai/* models accept tools only when reasoning_effort is set explicitly. With the parameter absent the upstream returns Function tools with reasoning_effort are not supported (3/3 calls); with it present the tool call succeeds (3/3 at each of low, none, and high). The error text reads as if the two are mutually exclusive, but the behaviour is the opposite — the rejection fires on absence, not presence. Continue has no way to send reasoning_effort, so these four report no native tool support rather than advertising a capability that would 400 in agent mode.

Two related facts, for anyone touching this provider later:

• openai/gpt-6-astra requires max_completion_tokens rather than max_tokens. OpenAI.isOSeriesOrGpt5PlusModel already matches gpt-[5-9], so Continue handles this automatically — no provider-specific code was needed.
• qwen/qwen3.8-flash is advertised on the service's pricing page but is absent from GET /v1/models and returns 503 No available channel for model qwen/qwen3.8-flash under group y-api (distributor). It is not referenced anywhere in this PR.

Verification

Run locally on Node 22.22.2:

• core: tsc -p ./ --noEmit — clean.

• core: vitest run llm/llms/OpenAI-compatible.vitest.ts — 218 passed. The seven generated y-api cases all pass:

  ✓ y-api > should have correct provider name
  ✓ y-api > should have correct default API base
  ✓ y-api > streamChat should send a valid request
  ✓ y-api > chat should send a valid request
  ✓ y-api > streamComplete should send a valid request
  ✓ y-api > complete should send a valid request
  ✓ y-api > should handle embeddings
  

• packages/openai-adapters: npm run build (tsc) — clean.

• Prettier — clean on every changed file.

Not run: the jest suite. Its global setup calls fs.rmSync on core/test/.continue-test, which this sandbox blocks, and test/jest.setup-after-env.js fails to load node-fetch (SyntaxError: Cannot use import statement outside a module) under this Node build. So llm/toolSupport.test.ts is typechecked but not executed here; I'd expect CI to cover it.

CLA

I'll post the CLA signature comment on this PR.

[jiweiyeah]

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

[github-actions]

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

[jiweiyeah]

A note on the three non-green checks, since none of them come from this change.

jetbrains-tests — pre-existing and repo-wide. It fails on the current head of main-targeting branches I did not touch:

branch	run
snyk-fix-16461e5b…	35084479159
fix/issue-13223	34871704808
feat/add-api-route-provider	34850359834

The job dies in setup, before any Continue code runs — the third-party AnimMouse/setup-ffmpeg step:

tar: Error is not recoverable: exiting now
##[error]Process completed with exit code 2.
##[end-action id=__self.__AnimMouse_setup-ffmpeg.__run;outcome=failure;conclusion=failure;duration_ms=236]

require-all-checks-to-pass — a meta-check that only reflects the above.

build-and-upload-vsix (darwin, arm64, macos-latest) — reported as cancelled, not failed. The build and both artifact uploads completed first:

vsce package completed - extension created at extensions/vscode/build/continue-1.3.40-4a6f217.vsix
Artifact vscode-extension-build-darwin-arm64 has been successfully uploaded! Final size is 77444283 bytes.
Artifact download URL: …/runs/35096756936/artifacts/10446282848

##[error]The operation was canceled. lands afterwards, at 12:46:04, inside post-job cleanup while the second actions/cache save was still compressing. The runner went away mid-cache-upload; the linux, x64 and win32, x64 jobs in the same workflow finished green.

Everything this change actually touches is green: core-checks, lint, prettier-check, packages-checks (openai-adapters), all twelve test matrix jobs, all e2e, vscode-checks, binary-checks, gui-checks, plus CLA / GitGuardian / Snyk.