Imported from ais-one/novex-kit (.claude/skills/openapi-docs/SKILL.md). Install upstream with npx skills add ais-one/novex-kit --skill openapi-docs. Copyright stays with the author.

name: openapi-docs description: Guidance for generating per-app OpenAPI specs from this repo's existing zod DTOs (via zod-openapi's .meta({id}) + createDocument()) and publishing them behind one long-lived, multi-spec Swagger UI portal — so each REST app under apps/* is visually segregated (its own title, its own servers/base path, picked via a dropdown) while the portal itself never needs to change or redeploy when a new app is onboarded. Every operation must document every status code its controller→service→repository chain can actually produce — success plus every AppError subclass, validation failure, and the generic 500 fallback, cross-checked against common/node/errors/AppError.ts — each with a realistic example payload and a copy-pasteable curl command written directly into the operation's description. By default, every operation's description also gets a generated **Possible errors:** list of every reachable status/code, and every z.date()/z.coerce.date() field gets an explicit .meta({ override: { format: 'date-time' } }) — zod-openapi renders date schemas as a bare string with no format otherwise. Use when adding OpenAPI/Swagger documentation to a new or existing REST app (starting with vision-rest-audit), when annotating zod DTOs with .meta() for doc generation, when enumerating an endpoint's success/error responses, when extending or replacing scripts/generate-openapi.ts, or when wiring a new app into the shared docs portal/index. Not for apps with no HTTP business surface beyond the shared /health endpoint (e.g. vision-consumer-audit, a Kafka consumer — see Applying this to the apps in this repo today) and not for generating human-readable Markdown API references — see .github/skills/api-docs-generator for that unrelated Markdown-scraping skill.

OpenAPI docs: per-app specs, one shared Swagger portal

Usage

Invoke bare (/openapi-docs) for the full pattern. Invoke with an app name (/openapi-docs vision-rest-audit) to jump to that app's section under Applying this to the apps in this repo today.

This assumes the DTO vocabulary from the clean-architecture skill (controller/service/repository, request vs. response DTOs) — read that first if you haven't. It builds directly on that skill's "Response DTO (recommended)" note: define a response DTO as a zod schema with .meta({ id: '...' }) so the same schema is both the runtime shape and the documentation source.

Why one portal instead of one Swagger UI per app

Apps in this repo are deployed as separate Docker images with no shared gateway or ingress (see .github/workflows/deploy-cr.yml — one image per workflow_dispatch run, selected by app name). Mounting a live Swagger UI inside each app's own Express process would mean docs consumers need a different URL per app, and "segregated but universal" would require building a reverse proxy that doesn't exist today.

Instead: each REST app generates a static OpenAPI YAML file from its own DTOs; a single, separately-deployed docs portal (static Swagger UI, listing specs via a dropdown) reads whichever files exist. Onboarding app #2, #3, etc. means publishing a new YAML file and adding one entry to an index — the portal's own code never changes. This also sidesteps prod's network isolation: the portal serves pre-rendered static specs, so it needs no live access to an isolated-network API to render documentation — only "Try it out" (disable it in prod) would.

What already exists — build on this, don't duplicate it

• zod-openapi@^5.4.6 and @scalar/express-api-reference@^0.9.13 are already declared as dependencies of @common/node (common/compiled/node/package.json), transitively available to every app — but grep confirms zero usage anywhere in apps/* today. Nothing imports createDocument, calls .meta(, or mounts @scalar/express-api-reference. This skill is what actually wires them up.
• scripts/generate-openapi.ts is a real, working example of the intended mechanism (zod-openapi's createDocument(), no monkey-patching, no separate registry object) — but it's hardcoded against sample/template content: common/schemas/{auth,payment,notification,error}.schema.js's fictional /api/v1/auth/*, /api/v1/payments*, /api/v1/notifications* endpoints. It is not wired to any real app, and its single document.paths object assumes one merged spec for the whole repo — which is the opposite of the per-app segregation this skill asks for. Its servers block even labels http://localhost:8080 as "Local (via API Gateway)", describing a gateway that does not exist in this repo.
• npm run docs:validate (npx redocly lint docs/openapi/openapi.merged.yaml) and npm run docs:make-html (npx @redocly/cli build-docs ...) are real, reusable Redocly CLI invocations — reuse the tool, just point it at more files (see Validation below).
• The portal shell is real — docs/openapi-portal/index.html, vendored via npm run docs:portal:vendor (copies swagger-ui-dist into docs/openapi-portal/vendor/, gitignored as a build artifact) and previewable via npm run docs:portal:serve. It fetches ../openapi/index.json at load, shows an explicit "No API specs published yet" empty state when that list is empty (true today), and has an ALLOW_TRY_IT_OUT flag for the public-vs-prod deployment distinction below. See docs/openapi-portal/README.md.
• No app mounts a live Swagger UI, no per-app OpenAPI file exists yet, and scripts/generate-openapi.ts still needs rewriting to loop over onboarded apps instead of its hardcoded sample document — docs/openapi/index.json currently exists only as an empty placeholder ([]). Everything else past this section is the target shape to build, same as clean-architecture's "Applying this to the apps in this repo today" section describes target splits for apps that haven't been refactored yet.

Response completeness: every status the code can actually produce, not just the happy path

An operation's responses object must include one entry per HTTP status the request can actually receive, derived by walking the exact controller → service → repository chain that route calls — never guessed generically from what the endpoint "feels like." This repo's error surface is small and fully enumerable from common/compiled/node/errors/AppError.ts:

Plus two paths that never go through an AppError subclass at all (common/compiled/node/errors/error.middleware.ts):

• Malformed JSON body (entity.parse.failed, special-cased ahead of the AppError check) → 400 / INVALID_JSON — applies to every route accepting a JSON body (POST/PUT/PATCH).
• Any uncaught, non-AppError exception — a DB driver error, a Kafka publish failure, anything nobody wrapped — falls through normalizeError()'s final branch → 500 / INTERNAL_ERROR, message replaced with 'An unexpected error occurred' outside dev. This is always reachable on every single operation — a repository or service call can always fail unexpectedly. Document it unconditionally; an operation with only a 200 in its responses is incomplete, not simple.

Method, per operation:

1. Every route wrapped in validate('body' | 'params' | 'query', schema) (common/compiled/node/errors/validate.ts) gets a 422 VALIDATION_ERROR entry — mechanical, not app-specific.
2. Grep throw new across the exact controller/service/repository files that route's handler calls (not the whole app) and add one response entry per distinct class/status found. Don't assume a lookup failure is a 404. vision-rest-audit's publishAuditEvent service throws ValidationError('Invalid source') (services/audit-events.service.ts:43) when the request body's source key doesn't resolve to a known audit_sources row — that's 422, not 404, because an unresolved reference in the request is modeled as a bad request, not a missing resource. Read the actual throw statement; don't infer the status from what the operation conceptually sounds like.
3. If the route sits behind requireRole/requireFga/authUser (check the route file and the app's index.ts), add 401 UNAUTHORIZED. This repo has no ForbiddenError/403 subclass today — a 403 needs new AppError(msg, 403, 'FORBIDDEN') via the base class, and the response entry documents that literal code string since there's no dedicated subclass to point at.
4. Every route accepting a body adds 400 INVALID_JSON.
5. Every operation, unconditionally, adds 500 INTERNAL_ERROR.
6. A list endpoint that finds nothing returns 200 with an empty items array, never 404 — a 404 only ever comes from an explicit NotFoundError thrown by a single-resource lookup (GET /events/:event_id, never GET /events).

Every error response's content schema is the shared, already-real common/schemas/error.schema.js's ErrorResponseSchema ({ error: { code, message } }, .meta({ id: 'ErrorResponse' })-tagged) — reused via $ref across every operation and every app, never redefined per app. (scripts/generate-openapi.ts's current sample code actually imports this schema from the wrong file — auth.schema.js, which doesn't export it, a real latent bug — import it from common/schemas/error.schema.js directly, as the template below does.)

Every error code must be visible directly in the operation's description, not just buried in responses

Completing the responses object (above) is necessary but not sufficient — by default, always also surface a **Possible errors:** list in the operation's own description, one bullet per non-2xx status with its real code. Do this unless the user has explicitly asked for something else.

This isn't cosmetic. ErrorResponseSchema's message field carries its own schema-level .meta({ example: '...' }) (common/schemas/error.schema.js), and every renderer we've checked (Scalar included) shows that shared, generic example whenever a reader expands the ErrorResponse field tree — regardless of which specific response (404, 422, 500, ...) they were actually looking at. A response-level example/examples on the media type object is real and correct in the spec, but it renders in a different UI element (the response body preview), one that isn't reliably visible without extra interaction. The only place a reader is guaranteed to see the actual reachable codes for this operation, with zero clicks, is the description's own Markdown — the same reasoning the Curl examples section below already applies to request examples.

Mechanism — attach a real code/message example to every error response, then generate the description's error list from that same object so the two can never drift apart:

// apps/vision-rest-audit/src/openapi.ts
const errorExample = (code: string, message: string, details: unknown = null) => ({
  error: { code, message, ...(details !== null && { details }) },
});

const errorResponse = (description: string, example: ReturnType<typeof errorExample>) => ({
  description,
  content: { 'application/json': { schema: ErrorResponseSchema, example } },
});

// Use when one status has more than one distinct cause — OpenAPI allows only one entry per
// status code, so multiple causes become named `examples`, not multiple response entries.
const errorResponseExamples = (
  description: string,
  examples: Record<string, { summary: string; value: ReturnType<typeof errorExample> }>,
) => ({
  description,
  content: { 'application/json': { schema: ErrorResponseSchema, examples } },
});

const internalError = errorResponse('Unexpected server error', errorExample('INTERNAL_ERROR', 'An unexpected error occurred'));
const invalidJson = errorResponse('Malformed JSON in the request body', errorExample('INVALID_JSON', 'Malformed JSON in request body'));

// Loosely typed on purpose — only needs to tolerate the 200/202 success entries in the same
// responses object (which have neither `example` nor `examples`), never needs to match
// zod-openapi's own response types.
type DocumentedErrorResponse = {
  description: string;
  content?: Record<
    string,
    { schema?: unknown; example?: { error: { code: string } }; examples?: Record<string, { value: { error: { code: string } } }> }
  >;
};

// Pulls the `code` back out of whatever errorResponse()/errorResponseExamples() produced, so the
// bullet list can be generated instead of hand-written (hand-written is what drifts).
const codeOf = (response: DocumentedErrorResponse): string => {
  const media = response.content?.['application/json'];
  if (media?.example) return media.example.error.code;
  if (media?.examples) return [...new Set(Object.values(media.examples).map((e) => e.value.error.code))].join(' / ');
  return 'UNKNOWN';
};

const errorCodesSection = (responses: Record<number, DocumentedErrorResponse>): string =>
  [
    '',
    '**Possible errors:**',
    ...Object.entries(responses)
      .filter(([status]) => Number(status) >= 400)
      .map(([status, response]) => `- \`${status} ${codeOf(response)}\` — ${response.description}`),
  ].join('\n');

// Build the responses object once, as its own const, and feed it to BOTH places —
// the description's bullet list and the operation's actual `responses:` field.
const listAuditEventsResponses = {
  200: { description: 'Paginated audit events', content: { 'application/json': { schema: auditEventListResponseSchema } } },
  422: errorResponse('Request failed validation', errorExample('VALIDATION_ERROR', 'event_id: Invalid UUID')),
  500: internalError,
};

// ...inside the operation object:
description: [
  'Returns a paginated, filterable list of audit events.',
  errorCodesSection(listAuditEventsResponses),
  '',
  '**Example:**',
  '```bash',
  "curl -s 'http://localhost:3300/api/v1/audit/events?result=DENIED&page=1&limit=20'",
  '```',
].join('\n'),
responses: listAuditEventsResponses,

Renders as, e.g.:

Possible errors:

• 422 VALIDATION_ERROR — Request failed validation
• 500 INTERNAL_ERROR — Unexpected server error

See apps/vision-rest-audit/src/openapi.ts's real, complete version of errorExample/errorResponse/errorResponseExamples/codeOf/errorCodesSection and every operation's hoisted *Responses const — copy that file's pattern rather than re-deriving it.

Date/datetime fields rendered as string need an explicit format override

By default, always check: does this DTO have a z.date() or z.coerce.date() field? If so, it needs a format override — don't rely on zod-openapi to infer one.

zod-openapi@5.4.6 converts a Zod "date" schema to a bare OpenAPI { type: 'string' } with no format at all — confirmed in its own schema-override logic (case "date": ctx.jsonSchema.type = "string";, no format assignment beside it). That means a field like start_time: z.coerce.date().optional() shows up in the generated docs as an unqualified string, indistinguishable from a free-text field, even though the actual DTO coerces and validates it as a real Date at runtime. This gap is specific to z.date()/z.coerce.date() — it does not affect z.iso.datetime() (used for response fields like occurredAt), since that's a Zod string schema with its ISO format already built in, which zod-openapi already renders correctly as format: 'date-time'.

Fix with zod-openapi's override meta option — a pure documentation-layer annotation; it changes nothing about runtime parsing, coercion, or validation:

// apps/vision-rest-audit/src/dto/v1/audit-event.dto.ts
start_time: z.coerce.date().optional().meta({ override: { format: 'date-time' } }),
end_time: z.coerce.date().optional().meta({ override: { format: 'date-time' } }),

Applies the same way to audit-event-error.dto.ts's mirrored start_time/end_time query params — any future z.date()/z.coerce.date() field, request or response, gets the same override.

Curl examples: copy-pasteable, embedded in the operation itself

Don't rely on Swagger UI's "Try it out" auto-generated request-snippets panel as the only curl example — it only appears after a user expands and interacts with the operation, and it's only populated with realistic values if every field already carries a schema example. Instead, write the curl command directly into the operation's own description as a fenced code block — description renders as Markdown in both Swagger UI and Redoc, so it's visible immediately with zero interaction, in either rendering target.

Rules for the curl itself, matching .github/skills/api-docs-generator's existing "realistic fake data" rule so both doc mechanisms in this repo agree:

• Use the app's actual default port from its own src/index.ts (vision-rest-audit → 3300, matching API_PORT's fallback) as the example host — never a placeholder like <host>.
• Use realistic-looking example values everywhere a real request needs one — a real-shaped UUID for event_id, a real-looking key for source — never "string" or "<value>".
• For POST/PUT/PATCH, include the full -H 'Content-Type: application/json' -d '{...}' body inline, using the exact field names and casing the DTO expects (vision-rest-audit's request bodies are snake_case — see createAuditEventBodySchema).
• One curl per operation, demonstrating the successful call, is enough — it doesn't need to reproduce every error case; the response schemas already document what a client should branch on.

Per-app spec generation

1. Annotate DTOs with .meta({ id })

Every zod schema that should appear in the docs gets a stable id, the same pattern common/schemas/auth.schema.js already uses. In vision-rest-audit, that means src/dto/v1/audit-event.dto.ts and audit-event-error.dto.ts — e.g.:

// src/dto/v1/audit-event.dto.ts
export const auditEventSummaryResponseSchema = z
  .object({
    eventId: z.uuid(),
    sourceId: z.string(),
    // ...
    occurredAt: z.iso.datetime(),
  })
  .meta({ id: 'AuditEventSummaryResponse' });

.meta({ id }) is what lets zod-openapi emit this as a reusable components.schemas.AuditEventSummaryResponse and $ref it from every path that returns it, instead of inlining the shape at every use site.

2. One generator script per app, not a shared merged one

Add src/openapi.ts inside the app itself (colocated, not stuffed into root scripts/generate-openapi.ts alongside unrelated apps' paths):

// apps/vision-rest-audit/src/openapi.ts
import { createDocument } from 'zod-openapi';
import { ErrorResponseSchema } from '../../../common/schemas/error.schema.js';
import {
  auditEventAcceptedResponseSchema,
  auditEventDetailResponseSchema,
  auditEventSummaryResponseSchema,
  createAuditEventBodySchema,
  listAuditEventsQuerySchema,
} from './dto/v1/audit-event.dto.ts';
// ...import audit-event-error.dto.ts schemas the same way for /events/errors, /events/errors/{event_id}

const errorContent = { 'application/json': { schema: ErrorResponseSchema } };
const validationError = { description: 'Request failed validation', content: errorContent };
const internalError = { description: 'Unexpected server error', content: errorContent }; // always reachable — see Response completeness above

export const document = createDocument({
  openapi: '3.1.0',
  info: { title: 'Vision REST Audit API', version: '1.0.0' },
  servers: [{ url: 'http://localhost:3300', description: 'Local' }],
  paths: {
    '/events': {
      get: {
        tags: ['Audit Events'],
        summary: 'List audit events',
        description: [
          'Returns a paginated, filterable list of audit events. Never 404s — no matches is a 200 with an empty `items` array.',
          '',
          '**Example:**',
          '```bash',
          "curl -s 'http://localhost:3300/api/v1/audit/events?source_key=vision-mcp&result=DENIED&page=1&page_size=20'",
          '```',
        ].join('\n'),
        requestParams: { query: listAuditEventsQuerySchema },
        responses: {
          200: {
            description: 'Paginated audit events',
            content: { 'application/json': { schema: auditEventSummaryResponseSchema } },
          },
          422: validationError, // a query param fails zod — e.g. an invalid `result` enum value, a non-uuid `event_id`
          500: internalError,
        },
      },
      post: {
        tags: ['Audit Events'],
        summary: 'Record a new audit event',
        description: [
          'Accepts an event for async processing — 202 means queued, not yet persisted.',
          '',
          '**Example:**',
          '```bash',
          'curl -s -X POST http://localhost:3300/api/v1/audit/events \\',
          "  -H 'Content-Type: application/json' \\",
          '  -d \'{"event_id":"3fa85f64-5717-4562-b3fc-2c963f66afa6","feature_group":"reports","action":"report.generated","source":"vision-custom-api"}\'',
          '```',
        ].join('\n'),
        requestBody: { required: true, content: { 'application/json': { schema: createAuditEventBodySchema } } },
        responses: {
          202: {
            description: 'Accepted for async processing',
            content: { 'application/json': { schema: auditEventAcceptedResponseSchema } },
          },
          // `source` failing to resolve to a known audit_sources row is 422, NOT 404 — services/audit-events.service.ts:43
          // throws ValidationError('Invalid source'), treating an unresolved reference as a bad request, not a missing resource.
          422: {
            description: 'Request body failed validation, or `source` does not resolve to a known audit source',
            content: errorContent,
          },
          500: internalError, // e.g. the queue publish fails
        },
      },
    },
    '/events/{event_id}': {
      get: {
        tags: ['Audit Events'],
        summary: 'Get an audit event by ID',
        description: [
          'Returns full detail for one audit event, including before/after data and its source.',
          '',
          '**Example:**',
          '```bash',
          'curl -s http://localhost:3300/api/v1/audit/events/3fa85f64-5717-4562-b3fc-2c963f66afa6',
          '```',
        ].join('\n'),
        responses: {
          200: {
            description: 'Audit event detail',
            content: { 'application/json': { schema: auditEventDetailResponseSchema } },
          },
          422: validationError, // :event_id isn't a valid uuid
          404: { description: 'No audit event exists with this event_id', content: errorContent },
          500: internalError,
        },
      },
    },
  },
});

audit-event-errors.routes.ts follows the identical shape — GET /events/errors (list, same 422/500 pattern as GET /events) and GET /events/errors/{event_id} (getAuditEventErrorByEventId throws NotFoundError at services/audit-event-errors.service.ts:33, so it gets the same 200/422/404/500 set as GET /events/{event_id}). No route in vision-rest-audit currently has auth middleware attached, so none of them get a 401 yet — if requireRole/requireFga/authUser is added to a route later, add 401 UNAUTHORIZED to that operation at the same time.

A small root script writes each app's document to its own file — this replaces scripts/generate-openapi.ts's single hardcoded document with a loop over onboarded apps:

// scripts/generate-openapi.ts
import { mkdirSync, writeFileSync } from 'node:fs';
import yaml from 'js-yaml';

const APPS = ['vision-rest-audit']; // add each app here as it's onboarded — see index below

for (const app of APPS) {
  const { document } = await import(`../apps/${app}/src/openapi.ts`);
  mkdirSync('docs/openapi', { recursive: true });
  writeFileSync(`docs/openapi/${app}.yaml`, yaml.dump(document, { lineWidth: 120 }));
}

Output convention: docs/openapi/<app-name>.yaml, one file per app, filename matching the app's folder name — never one merged file across apps. A merged file is exactly what re-introduces path collisions and makes "segregated by look" impossible.

3. The index the portal reads

docs/openapi/index.json, written by the same loop, drives the portal's dropdown without the portal needing any per-app knowledge baked in:

[{ "name": "Vision REST Audit API", "file": "vision-rest-audit.yaml" }]

Regenerate it in the same script that writes the YAML files — never hand-maintain it separately, or it will drift from what's actually on disk.

Validation

Point Redocly's lint at every file, not just one:

npx redocly lint docs/openapi/*.yaml

Keep npm run docs:validate as the entry point; only its glob target changes from the single openapi.merged.yaml to docs/openapi/*.yaml.

The portal

Real, already built: docs/openapi-portal/index.html — a single static Swagger UI page that fetches ../openapi/index.json at load time and builds its urls dropdown from the result:

fetch('../openapi/index.json')
  .then((r) => r.json())
  .then((specs) => {
    window.ui = SwaggerUIBundle({
      urls: specs.map((s) => ({ url: `../openapi/${s.file}`, name: s.name })),
      'urls.primaryName': specs[0].name,
      dom_id: '#swagger-ui',
      presets: [SwaggerUIBundle.presets.apis, SwaggerUIStandalonePreset],
      layout: 'StandaloneLayout',
    });
  });

Because the dropdown list comes from index.json at runtime rather than being compiled into the page, onboarding app #2 never touches the portal's own code — only docs/openapi/*.yaml and index.json change, both produced by the generation step above. index.html shows an explicit "No API specs published yet" message rather than a blank/broken Swagger UI when index.json is [] (the state today, before any app is onboarded).

This portal is documentation tooling, not a running app — it has no build step of its own beyond npm run docs:portal:vendor (copies swagger-ui-dist's CSS/JS into docs/openapi-portal/vendor/, gitignored as a build artifact, not checked in), so it lives under docs/openapi-portal/ rather than as a new apps/* workspace. It is never mounted inside vision-rest-audit's own Express process (or any other app's) — keeping it out-of-process is what makes "universal" possible without a reverse proxy. Preview locally with npm run docs:portal:serve (serves all of docs/, not just this folder — the relative ../openapi/index.json fetch needs openapi/ and openapi-portal/ reachable as siblings, exactly as they'll be laid out once uploaded to a bucket) and open http://localhost:5050/openapi-portal/.

Deploying it fronting a prod server on the isolated network publicly? Flip ALLOW_TRY_IT_OUT to false at the top of index.html's script — it gates supportedSubmitMethods, so "Try it out" is disabled without touching anything else.

Deployment

Reuse the existing OSS/S3 upload mechanics from .github/workflows/deploy-bucket.yml (ossutil/AWS CLI steps), but as a new workflow rather than editing that one in place — deploy-bucket.yml is hard-coded to apps/${{ inputs.app }}/dist from a Vue npm run build, and this portal is neither a workspace app nor a Vite build. Per root CLAUDE.md's protected-paths rule, name it with the local- prefix (e.g. .github/workflows/local-deploy-docs.yml) so it's a downstream-only workflow, preserved across upstream template syncs, uploading docs/openapi/ (generated YAML + index.json) and docs/openapi-portal/ (the static Swagger UI shell) to the target bucket.

Two buckets, same pipeline, different reachability — not different code:

• Staging: bucket left public, matching "probably will be kept public on stag."
• Prod: same generated artifacts, published to a bucket/path reachable only inside the isolated network (VPN, internal ALB, or simply not published if prod docs aren't required yet). Never point the portal's "Try it out" at a prod server URL from a publicly-reachable deployment of the portal — disable supportedSubmitMethods: [] in the Swagger UI config for any public deployment, since static docs don't need live API access to render, but "Try it out" does.

Do / Don't

Do

• Give every response DTO meant for docs a stable .meta({ id }), matching common/schemas/*.schema.js's existing pattern.
• Generate one OpenAPI YAML per app, named after the app's own folder, under docs/openapi/.
• Regenerate docs/openapi/index.json from the same loop that writes the YAML files — never hand-edit it out of sync with what's on disk.
• Keep the portal's own code free of any per-app knowledge — it only ever reads index.json.
• Disable "Try it out" on any publicly-reachable deployment of the portal that fronts a prod server.
• Document every status code the actual controller→service→repository chain for that operation can produce, cross-checked against common/compiled/node/errors/AppError.ts's table — not what "feels" right for the endpoint.
• Always include the generic 500 INTERNAL_ERROR fallback on every operation — an uncaught exception is always possible.
• Give every operation a copy-pasteable curl command in its own description, using the app's real default port and realistic example data.
• By default, render a **Possible errors:** bullet list into every operation's description, generated from that operation's own responses object (errorCodesSection()) — a reader expanding the shared ErrorResponse schema's fields alone only ever sees that schema's generic example, never the actual per-response code.
• By default, give every z.date()/z.coerce.date() field .meta({ override: { format: 'date-time' } }) — zod-openapi renders date schemas as a bare string with no format on its own.

Don't

• Don't add a new app's paths into scripts/generate-openapi.ts's old hardcoded document.paths object — that recreates one shared, unsegregated spec.
• Don't mount @scalar/express-api-reference or swagger-ui-express live inside an app's own Express process as the primary docs surface — that defeats "one universal instance" and requires a gateway this repo doesn't have.
• Don't let the portal bake a specific app list into its HTML/JS — that means every onboarding is a portal redeploy instead of a file publish.
• Don't point a public portal's "Try it out" at a server sitting on the isolated prod network.
• Don't build an OpenAPI spec for an app with no HTTP business surface (see vision-consumer-audit below) just to have "full coverage" — there's nothing there to document.
• Don't assume a failed lookup is always a 404 — read the actual throw statement; some are deliberately ValidationError/422 (see the source-lookup example above).
• Don't skip the 500 response because it feels obvious — an operation documented with only its success status is incomplete.
• Don't rely solely on Swagger UI's auto-generated request-snippets panel as your only curl example — write one explicitly in the operation's description.
• Don't leave an operation's error codes only in the responses object — a **Possible errors:** list belongs in the description too, generated from that same responses object, not hand-typed separately.
• Don't leave a z.date()/z.coerce.date() field without .meta({ override: { format: 'date-time' } }) — the generated docs will render it as an unqualified string, indistinguishable from free text.

Applying this to the apps in this repo today

vision-rest-audit — first real app to onboard

Express 5 + zod DTOs, no OpenAPI wiring yet. src/dto/v1/audit-event.dto.ts and audit-event-error.dto.ts already have clean, well-separated request/response schemas (listAuditEventsQuerySchema, createAuditEventBodySchema, auditEventSummaryResponseSchema, auditEventDetailResponseSchema, and the mirrored *-error.dto.ts schemas for the /api/v1/audit/event-errors routes) — exactly the shapes to add .meta({ id }) to and reference from src/openapi.ts per the template above. Routes are mounted at /api/v1/audit (src/index.ts), covering audit-events.routes.ts and audit-event-errors.routes.ts. Its exact reachable status codes per route — including the source-lookup 422-not-404 case — are already enumerated in Response completeness above; use that as the literal responses object, not a re-derived guess. This is real implementation work and out of scope for setting up this skill — noted here so the target shape is unambiguous once picked up.

vision-consumer-audit — nothing to document

A Kafka consumer (src/consumers/audit-events.consumer.ts), not a REST API. Its only Express usage is a /health liveness endpoint (src/index.ts, port default 3400) — the same shared boilerplate (@common/node/health/router) every app in this repo already exposes, not a documented business API. It gets no entry in docs/openapi/ and no row in index.json. If its Kafka contract (the audit.events topic schema it consumes, which vision-rest-audit's POST /events publishes) is ever worth documenting, that's an AsyncAPI spec, not OpenAPI/Swagger — a distinct format and a distinct, not-yet-built skill.

Future REST apps

Follow vision-rest-audit's recipe: .meta({ id }) on response DTOs → src/openapi.ts → add the app's name to the APPS list in scripts/generate-openapi.ts → npm run docs:generate produces the new YAML + updates index.json → publish via local-deploy-docs.yml. No step touches the portal's own HTML/JS.

References

• Root CLAUDE.md → "Updating from upstream template" → protected paths (local- prefix convention for downstream-only workflows).
• scripts/generate-openapi.ts + docs/openapi/README.md — the real, working zod-openapi/createDocument() mechanism, currently pointed at sample content; this skill redirects it to per-app files.
• common/schemas/{auth,payment,notification,error}.schema.js — real precedent for .meta({ id }) on a zod schema, sample/template content.
• common/compiled/node/package.json — where zod-openapi and @scalar/express-api-reference are already declared but unused.
• common/compiled/node/errors/AppError.ts — the real status-code/code table this skill's completeness rule is built from (AppError/NotFoundError/ValidationError/UnauthorizedError).
• common/compiled/node/errors/error.middleware.ts — the real { error: { code, message } } envelope, the INVALID_JSON special case, and the generic-exception 500 INTERNAL_ERROR fallback every operation must document.
• common/compiled/node/errors/validate.ts — the real 422 VALIDATION_ERROR production point for every validate()-wrapped route.
• common/schemas/error.schema.js — the real, reusable ErrorResponseSchema every error response should $ref instead of redefining per app.
• apps/vision-rest-audit/src/dto/v1/{audit-event,audit-event-error}.dto.ts — the real DTOs to annotate first.
• apps/vision-rest-audit/src/index.ts — real route mount point (/api/v1/audit) the generated spec's servers/paths must match, and its API_PORT default (3300) used in the curl examples.
• apps/vision-rest-audit/src/services/{audit-events,audit-event-errors}.service.ts — real throw new NotFoundError(...)/throw new ValidationError(...) sites the completeness rule is traced from.
• apps/vision-consumer-audit/src/index.ts — real example of an app with no documentable HTTP surface.
• docs/openapi-portal/{index.html,README.md} — the real, already-built portal shell (empty-state handling, ALLOW_TRY_IT_OUT toggle, local-preview instructions).
• docs/openapi/index.json — currently a real but empty ([]) placeholder, updated by docs:generate once an app is onboarded.
• Root package.json's docs:portal:vendor / docs:portal:serve scripts — real, working vendoring and local-preview commands for the portal.
• .github/workflows/deploy-bucket.yml — real OSS/S3 upload mechanics to adapt into a new local-deploy-docs.yml, not edit in place.
• .claude/skills/clean-architecture/SKILL.md — DTO vocabulary and the original .meta({ id }) recommendation this skill builds on.
• .github/skills/api-docs-generator/SKILL.md — a separate, unrelated skill that scrapes source/OpenAPI into Markdown reference docs; not a substitute for this generation+portal mechanism.
• apps/vision-rest-audit/src/openapi.ts — the real, current errorExample/errorResponse/errorResponseExamples/codeOf/errorCodesSection implementation and every operation's hoisted *Responses const; copy this file's pattern, not the older inline-responses template further up this doc.
• apps/vision-rest-audit/src/dto/v1/{audit-event,audit-event-error}.dto.ts — real start_time/end_time fields carrying the .meta({ override: { format: 'date-time' } }) fix.
• node_modules/zod-openapi's schema-override logic (case "date": ctx.jsonSchema.type = "string") — the real source of the missing-format gap; also documented in node_modules/zod-openapi/README.md's override section.