Agent Skills: Review OpenAPI

Review OpenAPI

Review an OpenAPI spec the way its actual reader experiences it: an external API consumer — human or LLM agent — who has this one document, no access to the repo, no knowledge of the company's frameworks, tickets, or design history. The spec is a product for that reader. This skill finds and fixes the prose and examples that fail them.

It reviews the documentation quality of the contract — descriptions, error vocabularies, examples. It does not judge schema design or flag breaking changes (review-behavior-change), and it does not replace the spec linter (duh lint in DUH-RPC repos) — it runs the linter as a gate around its own edits.

The core test: the Consumer Test

For every description, example, and error vocabulary: does it serve someone integrating against this API from the outside, with only this document? If understanding a sentence requires knowing an internal framework, reading a repo doc, or having watched the design happen, it fails. If a rendered endpoint page shows information that endpoint cannot produce, it fails.

1. Resolve the spec

• Explicit argument → use it.
• Otherwise: specs touched by the current diff (git diff --name-only <base>...HEAD | grep -i 'openapi.*\.ya?ml$'), else **/api/**/openapi.yaml in the repo. Multiple hits with no argument → review each in turn.
• Locate the implementation's error emitters for Rule 4 (in this repo's Go services: the NewServiceError / NewServiceErrorWithCode call sites and the handlers that reach them).

2. The audit rules

Audit the whole spec against each rule. Fix findings in place.

Rule 1 — Self-contained: no internal context leaks

The reader has never heard of your RPC framework, your monorepo, or your tickets. Remove or rewrite, anywhere in the spec:

• Internal framework/protocol names used as explanation ("Standard DUH-RPC error body", "DUH paths are static").
• Repo document references ("see docs/duh-rpc-cheatsheet.md").
• Ticket / design-doc / roadmap ids and internal shorthand ("(D12)", "ENG-132", "slice 6", "the wire transport", invariant labels like "(I3)").
• Deferred justifications — "retryable per DUH" tells the reader nothing. State the actual, consumer-relevant reason: "Safe to retry: every operation is read-only."
• Roadmap talk — "lifecycle metadata arrives in slice 6" is internal planning. Describe what the API is today; omit what it isn't.

Rewrite test: every sentence must be justifiable to an external integrator using only concepts the spec itself defines (its schemas, its endpoints, HTTP, and the domain — e.g. git plumbing for a git API is fine; the house RPC framework is not).

Rule 2 — Don't spec-splain the API's self-correction machinery

Convenience behaviors that exist to rescue callers who have not read the spec — did-you-mean suggestions on unknown endpoints, teaching 400s that spell out the schema, endpoint listings in error bodies — do not belong in the spec. The spec reader already has the endpoint list and the schemas; documenting the recovery machinery is teaching them how the API works internally, not how to use it. Keep the behaviors implemented and surface-tested; delete their documentation from the spec, including cross-references they leave dangling (details fields, code lists, "see the top-level description" pointers).

What stays: error facts a spec-reading consumer acts on — status codes, the code discriminator values, what details carries, retry semantics.

Rule 3 — Shared schemas stay generic; specifics live where they apply

A shared error schema (Reply-style: code/message/details) is referenced by every error response on every endpoint. Any condition-specific enumeration in its property descriptions — "the codes are REPO_NOT_FOUND, REF_NOT_FOUND, … (all HTTP 404)" — renders under a 500 on an unrelated endpoint. The shared schema states only the general rule (what the field is, how to branch on it, where the specific values are listed); the specific vocabularies live on the response definitions they belong to.

Rule 4 — Per-endpoint exhaustive error vocabularies, verified against the code

Shared response components document the union of what all endpoints can return — so a repos.get renders PATH_NOT_FOUND it can never produce. Fix by splitting response components by actual vocabulary and wiring each endpoint to the variant matching exactly what its implementation emits:

1. Derive the truth from the error emitters, never from the existing spec prose. Walk each handler's reachable error paths (in Go: which err* constructors each endpoint's call graph reaches). Watch for structural facts: an endpoint without pagination cannot return a cursor-expired code; a name-anchored cursor cannot expire; a filter-not-deny endpoint has no 404.
2. Define one response component per distinct vocabulary (NotFoundRepo, NotFoundRev, NotFoundRevCursor, BadRequest, BadRequestWrongType, …), same wire schema throughout.
3. Each component's description lists its codes exhaustively and says so ("the only code here", "the code is one of: …") with a one-clause meaning per code.
4. Rewire every endpoint's responses to its exact variants.

This is documentation-only: the wire schema and generated code must not change (see Verify).

Rule 5 — Worked request/response example pairs in one consistent world

Every operation carries a request example and a 200 example, where the request example is the call that produces the response example. All examples across the spec share one example world (one fictional repo/tenant/dataset) so values cross-reference between endpoints: the id one endpoint returns is the id another endpoint's example queries.

Coherence is load-bearing — the example world's facts must actually hold:

• Graph/relational claims are real: a merge-base example yields what the example history implies; ahead/behind counts match the listed items; a "more pages exist" flag is impossible after a terminal item (e.g. a root commit).
• Weave in teaching cases: an optional field absent where the API omits it (not null-filled), a realistic opaque cursor paired with has_next_page: true, input-format variety across requests (each accepted form — symbolic name, short name, raw id — appears at least once).
• Binary/streaming endpoints get a small literal example on their media type.
• Examples are illustrations, not contract: keep them typical-shaped, not exhaustive.

3. Verify (gate every edit batch)

1. Spec lints: the repo's spec linter (duh lint <spec> here) passes.
2. Codegen-neutral: regenerate/build the generated target (bazel build …:generated or the repo's duh-gen task). Doc-only changes must produce no schema/proto/fieldmap churn.
3. Contract tests pass: any tests that read the spec (vocabulary censuses, teaching-error surface tests) still pass — Rule 2 removes documentation, never behavior.

4. Report

Summarize per rule: findings fixed (with before → after for the representative ones), findings left (with why). In a suite/pipeline context, emit findings in the host's ticket format instead.

Boundaries

• Fix prose and examples in the spec only — never change schemas, field numbers, paths, or implementation behavior to make documentation "fit". A doc/code disagreement about behavior is a finding to surface (or a review-behavior-change handoff), not a silent doc edit.
• Do not invent error codes or example values the implementation cannot produce — Rule 4's derivation is from code, and Rule 5's world must be producible by the documented API.
• Internal design docs (blueprints, ADRs, explainers) are out of scope — they are supposed to carry internal context. Only the consumer-facing spec is held to the Consumer Test.