Testing API Contracts

Use this skill when the question is not "does the endpoint work" but "does the endpoint still promise what its consumers were built against".

A contract test checks shape and compatibility, not business behaviour. It answers: did a field disappear, did a type change, did an enum gain a value the client cannot handle, did a required field become optional. Functional API tests answer whether the value is correct. Both are needed and they fail for different reasons, which is the point of keeping them separate.

When to Use

• an API publishes an OpenAPI or JSON Schema document and nothing verifies the implementation matches it
• API tests assert status === 200 and nothing about the body
• a backend change is about to ship and no one can say which clients it breaks
• mocked fixtures need a drift guard against the real service
• two teams are arguing about whether a change is breaking
• a spec exists but has visibly diverged from what the service returns

Operating Principles

• The spec is the contract, the implementation is the claim. Test the claim against the contract, in both directions.
• Shape checks and value checks stay separate. Mixing them produces a test that fails for two unrelated reasons and reports one message.
• Additive is safe, subtractive is not. Adding an optional response field breaks nobody. Removing one, or tightening a type, breaks everybody who read it.
• Validate every response, not a sampled one. Schema validation is cheap enough to run on every API call the suite already makes.
• A breaking change is defined by consumers, not by the provider. A field no client reads can be removed freely; the argument is about evidence, not opinion.
• Undocumented behaviour is a finding. A response field absent from the spec is either a spec gap or an accidental leak. Both get reported.

Workflow

Phase 0: Locate the contract

Find the authoritative artifact and record which it is:

• an OpenAPI document in the repository, generated from code or hand-written
• an OpenAPI document served at runtime (/openapi.json, /swagger.json, /v3/api-docs)
• standalone JSON Schema files
• a GraphQL SDL
• none of the above, in which case the first deliverable is a spec derived from observed traffic, marked as derived rather than authoritative

If the spec is hand-written and the implementation is not generated from it, say so. Hand-written specs drift, and the drift is exactly what this skill exists to surface.

Phase 1: Decide which side you are testing

Side	Question	Typical artifact
Provider	Does our service return what our spec promises?	Schema validation over live responses
Consumer	Do we still work against what the provider promises?	Fixture validation plus a live sample check
Both	Are these two versions of the spec compatible?	Spec diff and breaking-change report

Choose one per run. A single output that mixes provider defects with consumer risks is unreadable to either team.

Phase 2: Wire validation into requests already being made

The cheapest contract coverage is a validator attached to the existing API test suite. Every request the suite already sends gets its response validated for free.

Recipes for ajv, zod, openapi-response-validator, and a Playwright expect extension are in ./resources/schema-validation-recipes.md.

Rules for the validator:

• Strict by default. additionalProperties: false when the spec allows it, so undocumented fields surface instead of passing silently.
• Validate the error responses too. The 4xx body shape is part of the contract and is where drift is most common.
• Report the path, not just "invalid". /items/0/price: expected number, got string is actionable; schema validation failed is not.
• Fail the test. A validator that logs a warning gets ignored within a week.

Phase 3: Cover the shape, not the sample

For each endpoint in scope, check the cases in ./resources/contract-coverage-checklist.md. The ones that catch real defects:

• empty collection - items: [] with the pagination envelope intact
• nullable fields actually null - the spec says nullable, so make one null
• optional fields absent - not present rather than present-and-null; these are different contracts
• maximum-length and boundary values - the field the spec caps at 255 characters
• every enum member - a client switch statement that has no default breaks on the value you never sent
• error responses - one per documented status code
• pagination edges - first page, last page, past the end

Phase 4: Diff the versions

When a spec has changed, produce a compatibility verdict rather than a list of edits. Classify every change using ./resources/breaking-change-checklist.md:

Class	Examples	Verdict
Safe	New optional response field, new endpoint, new optional request parameter, widened response enum a client treats as opaque	Ship
Risky	New enum value in a field clients branch on, tightened validation, changed default, changed ordering, changed pagination size	Ship with a consumer check
Breaking	Removed or renamed field, narrowed type, required field added to a request, optional response field made absent, status code change, auth requirement change	Version it or coordinate

For every Risky and Breaking item, name the affected consumers if they are knowable, and say what evidence supports the claim. "Unknown consumers" is a valid finding and raises the severity.

Phase 5: Guard the mocks

If the codebase mocks this API anywhere, the spec is also the guard for those fixtures. Validate the fixtures against the same schema:

• fixtures that fail validation are stale and are reported as defects in the test suite, not in the API
• run this check on the same cadence as the spec changes, not on every commit

This is the handoff point with mocking-network-and-time.

Phase 6: Report

Use ./resources/contract-report-template.md. A useful report contains:

• which spec, which version, which side was tested
• endpoints covered and endpoints skipped, with the reason for skipping
• validation failures grouped by endpoint, with the JSON path and both values
• the compatibility verdict per change, with severity
• undocumented fields found in responses
• what remains unverified

Common Failure Modes

• asserting only the status code and calling it an API test
• validating the happy path and never the documented error bodies
• additionalProperties left permissive, so a leaked internal field passes forever
• diffing two specs and reporting every line changed rather than classifying compatibility
• calling a change non-breaking because no test failed, when no test covered it
• treating a generated spec as proof of correctness; a spec generated from the code agrees with the code by construction
• one report that mixes provider bugs, consumer risks, and spec gaps

Resource Map

• ./resources/schema-validation-recipes.md - ajv, zod, openapi-response-validator, and a reusable Playwright matcher
• ./resources/breaking-change-checklist.md - full classification of spec changes into safe, risky, and breaking, for requests and responses
• ./resources/contract-coverage-checklist.md - the response shapes to force per endpoint
• ./resources/contract-report-template.md - report structure with a worked example

Related Skills

• api-playwright-test-developer - when the functional API tests these validators attach to are being written
• mocking-network-and-time - when the spec is being used to guard mocked fixtures against drift
• analyzing-regression-scope - when a breaking change needs its retest blast radius mapped
• verifying-acceptance-criteria - when the question is whether behaviour meets its stated contract rather than its shape
• assessing-release-readiness - when a breaking change feeds a go/no-go decision
• documenting-test-suites - when the spec gaps found here need writing up

Definition of Done

This skill is complete when:

• the authoritative spec is identified and its trustworthiness stated
• the side under test (provider, consumer, or version diff) is declared
• validation runs on every response the suite produces, including error responses, and fails the test on mismatch
• the shape cases from the coverage checklist are forced rather than hoped for
• every spec change carries a safe, risky, or breaking classification with its rationale
• undocumented response fields are reported
• mocked fixtures for this API are validated against the same schema, or their absence is noted
• the report says what was not covered