API doc signature sync

The **Type:** / **类型：** blocks in website/docs/{en,zh}/api/** are hand-written, curated copies of real exported types — there is no generation and no compiler check behind them, so they can be wrong the moment they are authored, not only when the type later changes. This skill catches both: it grounds every documented signature in the actual source type and the tsc oracle instead of trusting the prose.

Core rule — verify, never recall. Read the type from packages/core/src/types/*.ts or packages/core/src/api/types.ts (and, when built, the emitted .d.ts) and let tsc decide. Never judge a signature from memory or from the doc's own prose.

When to run

• A public type in packages/core/src/types/ (e.g. api.ts, config.ts, mock.ts, runner.ts) or packages/core/src/api/types.ts changed.
• A **Type:** / **类型：** block, or the prose describing a type's fields, was edited in website/docs/**/api/**.
• Reviewing a PR that touches either side.

Scope: every page with a **Type:** block — api/runtime-api/** and api/javascript-api/**, both en and zh.

Procedure

1. en/zh parity (deterministic — run the script first)

node .agents/skills/api-doc-sync/scripts/check-type-blocks.mjs

The en and zh pages must declare structurally identical signatures (only the label and translated // comments may differ). The script enforces this and exits non-zero on any mismatch, missing counterpart, or block-count difference. Fix every reported drift before moving on. Use --json for a machine-readable inventory of all blocks.

2. Source fidelity (per changed symbol — let tsc judge)

For each documented symbol whose page changed (or whose type changed):

1. Find the source of truth. Locate the exported type in packages/core/src/types/ (start from types/index.ts re-exports) or packages/core/src/api/types.ts. Read its real shape: every overload, parameter order, optionality, generics, and union members. Prefer the built dist/**/*.d.ts when available — it is the exact surface users consume; build with pnpm --filter @rstest/core build if needed.

2. Turn the doc into compile assertions. Write a throwaway .ts file inside the repo (e.g. the repo root or packages/core/, with a temp name like __doc-probe.ts) that imports the real type and exercises every call form the docs show as a positive case, plus a // @ts-expect-error for every form the docs say is invalid (e.g. "the options object cannot be the third argument"). It must live in the repo, not the external scratchpad: tsc resolves node_modules upward from the file's own directory, so only an in-workspace file can resolve @rstest/core. Run it, then delete it:

   npx tsc --noEmit --strict --skipLibCheck <scratch>.ts

   • A positive form that fails to compile → the docs show a call that does not exist. Drift.
   • An unused @ts-expect-error → a form the docs claim is rejected is actually accepted (or vice-versa). Drift.
3. Check for under-documentation. Enumerate the overloads / fields that the source type actually has and confirm each is represented in the doc signature or prose. The bug this skill exists for was a missing overload (the (name, fn, timeout?) shorthand was dropped from the **Type:** line) — a positive-only check will not catch that, so explicitly diff the source's overload set against the documented one.
4. Check field-level claims. When the prose lists option fields (timeout, retry, …), confirm each exists on the type with the stated optionality and meaning, and that no real field is omitted.
5. Check entrypoint exports. For every named type in a **Type:** / **类型：** block, confirm it is exported from the package entrypoint that the page documents. A type that exists only in source or another entrypoint is not available to readers of that page.

6. Check named-type linkability. A signature that names another type (TestContext, TestOptions, RstestUtilities, …) can be a bare, unlinked black box. For each named type a signature references, confirm the page either links it to its canonical definition or documents it inline. Only add a link when both hold:

   • (a) the type has a canonical anchor to point at — a real heading (### TestContext → #testcontext), not a loose bullet in a list (a bullet generates no anchor); and
   • (b) the type is foreign to the page — a data structure the reader must navigate elsewhere to understand, not a fluent/chaining return type that names the very object the current page documents.

   When both hold (e.g. TestContext at /api/runtime-api/test-api/test#testcontext), add a short prose link after the signature — do not re-inline the type's members, which creates a second copy that drifts. The link goes in an adjacent sentence, since a markdown link cannot live inside the backticked **Type:** code span.

   When either test fails, treat the type as already inline-documented and skip it — no link noise. RstestUtilities is the canonical skip: no heading anchor (only a bullet gloss in types.mdx), and => RstestUtilities is a fluent self-reference to the rs/rstest object these pages already document.

3. Fix drift at the doc layer

Apply fixes to the .mdx directly. For each fix:

• Update both en and zh; keep the signature blocks structurally identical (re-run the script in step 1 to confirm).
• Preserve the curated style — friendly names like TestOptions, omitted internal generics (ExtraContext) — as long as the omission is a faithful simplification, not a missing overload or wrong arg order.
• When a brand-new API/field is documented, set <ApiMeta addedVersion="…" /> per the convention in website/AGENTS.md.

4. Re-verify

Re-run step 1 (parity) and step 2 (compile assertions) until both are clean, then pnpm exec rs fmt --check the touched .mdx files.

Notes

• The parity script is safe to wire into CI / pre-push as a hard gate — it is fully deterministic. The source-fidelity pass (step 2) is the judgement half and runs here, on demand.
• This skill does not generate signatures from source. Generation would make drift impossible but discards the curated/simplified style the docs use on purpose; that trade-off is intentionally out of scope.