Api contract design

API Contract Design

An API is a promise. Clients — your own frontend, partners, and AI models writing integrations — build on exactly what you return, including your mistakes. Design every endpoint as if you can never change it, because for practical purposes you can't (only add).

Contract-first, always

Before writing the handler, write the contract: method, path, auth, request shape, response shape, every error case with its code. Five minutes of contract prevents the frontend and backend from building two different APIs. Keep it wherever it's enforceable — OpenAPI, typed schemas (Pydantic/zod), or at minimum a markdown table — the schema layer IS the documentation if you make it one.

Shape rules

• Resources and predictability over cleverness: plural nouns (/invoices/{id}), nesting only one level deep when ownership is intrinsic (/invoices/{id}/payments); verbs only for true actions (/invoices/{id}/cancel). A developer should guess the next endpoint correctly.
• Response envelopes are uniform across the whole API: pick one success shape and ONE error shape ({"error": {"code", "message"}}-style) and never deviate. Machine-readable code (stable, documented, for programs) + human message (for humans, changeable). Clients that parse message strings are a bug you caused.
• Every list endpoint, from day one: pagination (limit + cursor/offset), a deterministic default sort, and filters as query params. Retrofitting pagination onto a client that assumes "all rows" is a breaking change.
• Field conventions fixed project-wide: one case style, ISO-8601 timestamps in UTC, money as string-decimal or integer minor units (never JSON floats), booleans not 0/1 strings, null vs absent given meaning consistently.
• Return the resource state after mutation (the created/updated object), so clients don't need a follow-up GET.

Status codes — the short honest set

200 read/update ok · 201 created · 204 deleted/no body · 400 malformed/validation · 401 who are you · 403 you can't do that (but see threat-model-security: prefer 404 when existence itself is private) · 404 not found or hidden · 409 state conflict (duplicate, already-cancelled) · 422 semantic validation if you distinguish it · 429 rate limited · 5xx our fault, never leaks internals. Don't return 200 {"success": false} — that lies to every HTTP-aware layer between you and the client.

Idempotency — the rule everyone skips until it costs money

• GET/PUT/DELETE are naturally idempotent; keep them so.
• Any POST that creates something valuable or moves money accepts an idempotency key (client-generated, stored unique): the retry after a timeout must not create a second payment. Networks fail after the server succeeds; without this, your client's only options are "maybe duplicate" or "maybe lose it".
• Everything a queue/webhook/cron touches follows the same rule (see system-design).

Evolution without breakage

Safe: adding optional request fields, adding response fields, adding endpoints/enum-values clients were told to tolerate. Breaking: removing/renaming fields, type changes, semantics changes, tightening validation, changing defaults, changing error codes.

• Clients must ignore unknown response fields (say this in the contract); servers must reject unknown fields only where security demands it.
• When you must break: new field alongside old → dual-write/dual-read → migrate clients → deprecate with a date → remove. Version (/v2) only when the surface change is too broad for field-level evolution.
• Your own frontend counts as a client: deploys aren't atomic across app and API; the old frontend will talk to the new API for a while (and cached bundles make "a while" longer than you think).

Webhooks you emit

Sign payloads (HMAC + timestamp), retry with backoff on non-2xx, provide a redelivery/reconciliation path, and document that consumers must be idempotent. Send a thin event (type, id, minimal data) and let consumers fetch fresh state — fat payloads go stale in retry queues.

Review checklist

• Can every error the handler can produce be triggered deliberately, and does each map to the documented envelope + code?
• Does any endpoint return different shapes for the same status? (Fix it.)
• Walk a slow-network retry through every POST: what duplicates?
• Diff against the previous contract: is every change in the "safe" list? If not, where's the migration plan?
• Auth on every route confirmed server-side — including the "internal" ones that are one leaked URL away from public (threat-model-security Gate 2).