Skip to content

API Contracts

Principle · api-contracts

Loads automatically whenever the agent touches code this principle governs. Install the skills · View source

A contract is the only part of a service that is genuinely hard to change, because you do not control who depends on it or when they will redeploy. Everything behind it is refactorable; everything in it is a promise. Design it deliberately, and evolve it additively.

The contract is written before the implementation, and the specification is the source of truth — not a document generated afterwards from whatever the code happened to return.

Config Resolution

  1. Read .msskills/config.yaml in the repo root; look up paths.api.
  2. A custom document exists at that path → read its frontmatter mode:
  3. override → use it alone as the API standard; ignore the defaults below.
  4. overlay (default, or no mode key) → read defaults first, then apply the custom document's sections on top, matched by exact heading; new sections append.
  5. A path is configured but no file exists there → say which path is missing, then use the defaults.
  6. No config file or no paths.api key → use defaults.
  7. .msskills/stack.md exists → use its protocol and tooling: REST, gRPC, or GraphQL; its serialisation format; its schema registry and compatibility mode; its spec linter.
  8. An existing specification is in the repository → its conventions outrank these defaults. Match the API that exists rather than introducing a second style beside it.

Self-Validation Checklist

STOP before presenting any new or changed contract. Verify every check. Fix failures first.

  1. SPEC FIRST: Does a specification fragment exist — OpenAPI, AsyncAPI, protobuf, JSON Schema — before the implementation? If the code came first → write the spec now and reconcile.
  2. CONSUMER-SHAPED: Does each operation serve a real consumer use case, or does it expose the internal model? An endpoint that mirrors a table row by row → reshape it around what the caller is trying to do.
  3. BACKWARD COMPATIBLE: For a change to an existing contract, is it purely additive — new optional field, new endpoint, new enum value the consumer may ignore? If it removes, renames, narrows a type, tightens validation, or changes a default → it is breaking. STOP and route through collaborative-judgment.
  4. ERRORS SPECIFIED: Is every failure enumerated with a status code, a stable machine-readable code, and guidance on whether retrying helps? A contract with only the success case is half written.
  5. IDEMPOTENCY: Can every non-GET operation be safely retried? If not → define the idempotency key, its scope, and how long it is honoured (see resilience-patterns).
  6. NO UNBOUNDED COLLECTION: Does any response return a list with no limit? Every collection is paginated with an explicit default and maximum page size.
  7. TYPES EXPLICIT: Are money, time, and identifiers unambiguous — currency alongside amount, timestamps in UTC with offset, IDs typed rather than bare strings? Ambiguity here becomes a production incident.
  8. NO INTERNAL LEAKAGE: Does the payload expose database identifiers, internal enum names, stack traces, or fields no consumer needs? Strip them (see secure-service).
  9. EVENTS ARE FACTS: Is each published event named in the past tense, describing something that happened, and does it carry what consumers need without requiring a call back? Events named as commands (SendEmail) belong on a queue, not an event stream.
  10. COMPATIBILITY MODE: For an event schema, is the registry's compatibility mode stated, and does the change satisfy it?
  11. VERIFIABLE: Is it clear which contract test will hold this promise (see contract-testing)?

All checks pass → state "Contract holds: , , errors specified."

Active Anti-Pattern Scan

Any box you can check is a defect. Fix it before presenting.

  • [ ] Code-First Spec: the specification is generated from handlers after the fact, so the contract is whatever the code does → author the spec deliberately.
  • [ ] CRUD Leakage: endpoints that mirror tables, forcing consumers to orchestrate business operations → model the operation the consumer actually performs.
  • [ ] Chatty Contract: the consumer must call three endpoints to render one screen → give it one purpose-built representation.
  • [ ] Silent Breaking Change: a field removed, renamed, or narrowed without a version or a migration window → additive change, or an explicit deprecation cycle.
  • [ ] Envelope Errors: HTTP 200 with {"success": false} → use the status code; it is the part every proxy, client, and dashboard already understands.
  • [ ] Stringly-Typed Payload: dates, money, enums, and booleans passed as free text → use typed fields with stated formats.
  • [ ] Unbounded List: a collection endpoint with no pagination → it will time out the day the data grows.
  • [ ] Chatty Event: an event carrying only an ID, forcing every consumer to call back → include the facts consumers need, and say why anything is deliberately omitted.
  • [ ] Command as Event: SendConfirmationEmail published to a topic → that is a command with one owner, not a fact.
  • [ ] Version in Every Path: /v2/ minted for an additive change → version only on a genuine break.
  • [ ] Undocumented Nullability: optional and required unstated, so consumers guess → mark every field.
  • [ ] Leaked Internals: database IDs, internal status names, or exception text in responses → map to the public vocabulary.

Ambiguity Signals

Route these through collaborative-judgment. Each has two defensible answers.

  • Break now or carry the field forever. A clean break costs a coordinated migration; carrying a deprecated field costs permanent complexity. Which is cheaper depends on how many consumers exist and whether you can see their traffic.
  • Versioning strategy. URL path, header, or media type — each trades discoverability against cache and routing behaviour, and the right answer is usually "whatever the rest of the estate already does".
  • Event granularity. Fine-grained facts are flexible and chatty; coarse aggregate events are efficient and couple consumers to a bigger shape.
  • Thin versus fat events. Carrying state removes callbacks and duplicates data that can go stale; carrying only an ID keeps one source of truth and adds a synchronous dependency.
  • Whether a change is really breaking. Tightening validation, changing a default, or adding a required field to a request breaks some consumers. Deciding whether those consumers exist is a judgment call, not a lookup.