# API Contract Compatibility > Design, change, or review APIs and serialized interfaces while making compatibility, versioning, errors, and migration behavior explicit. Use for HTTP, RPC, CLI, event, library, configuration, and file-format interfaces used by other code or people. - Skill: `chloevpin/api-contract-compatibility` (Agent Skill) - Install (CLI): `npx skillmds@latest add chloevpin/api-contract-compatibility` - Raw SKILL.md: https://api.skillmd.com/api/skills/chloevpin/api-contract-compatibility/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Integrations & APIs - Author: ChloeVPin (https://skillmd.com/u/chloevpin) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/chloevpin/api-contract-compatibility --- Lifecycle: `draft` # API Contract and Compatibility ## Purpose and scope Make an interface’s observable contract explicit and prevent accidental breakage for callers, consumers, operators, and stored data. This skill covers interface design and evolution; it does not replace domain requirements, security review, performance analysis, or a service owner’s release policy. ## Triggers and prerequisites Trigger when adding or changing endpoints, function signatures, commands, schemas, events, configuration, serialized data, error responses, or compatibility behavior. Prerequisites: consumers and ownership, current contract or examples, versioning policy, supported environments, data lifecycle, and the intended change. ## Decision criteria - An interface includes successful results, errors, validation, defaults, ordering, side effects, permissions, limits, timing expectations, and serialization, not only a method name or URL. - Compatibility is relative to identified consumers and versions; “backward compatible” without a consumer set is an unsupported claim. - Prefer additive, tolerant, and reversible evolution when consumers cannot migrate atomically; reject ambiguity rather than silently reinterpret data. - A deprecation or version signal is actionable only with an effective boundary, migration path, and owner. - Preserve distinctions between omitted, explicitly null, defaulted, and disabled values, as well as exception/error behavior; these are observable contract elements for client libraries and serialized interfaces. - For CLI interfaces, include exit status, stdout/stderr routing, encoding, help/error text, interactive behavior, and terminal capability assumptions in the contract. ## Procedure 1. Inventory consumers, producers, stored representations, generated clients, documentation, monitoring, and dynamic or out-of-repository callers. Identify who owns each boundary. 2. Write the current and proposed contract: inputs, outputs, required/optional fields, defaults, validation, errors, ordering, idempotency, side effects, limits, permissions, and version semantics. 3. Classify the change for each consumer: additive, behavior change, source incompatible, binary/serialization incompatible, operationally incompatible, or ambiguous. Do not assume an apparently optional field is safe for strict parsers. 4. Define decision boundaries for unknown, missing, null, extra, malformed, duplicate, stale, and future-version values. Specify whether clients reject, ignore, preserve, or migrate them. 5. Choose an evolution path: compatible addition, adapter, explicit version, staged rollout, dual read/write, migration, deprecation, or intentional breaking release. Record why alternatives do not fit. 6. Check security and resource implications at the boundary: authorization scope, data exposure, validation, amplification, rate/size limits, and error detail. Route deeper concerns to the relevant security skill. 7. Verify with representative existing consumers, contract examples, malformed and boundary inputs, old/new version combinations, retries, partial rollout, and rollback or migration recovery where applicable. 8. Compare the changed interface against the baseline and inspect generated artifacts, schemas, clients, documentation, observability, and operational runbooks for drift. 9. Publish the contract and migration/deprecation instructions with owner, effective dates or version boundary, compatibility limits, and review trigger. Preserve old behavior until the stated migration condition is met. ## Examples and counterexamples Good: Add a response field only after checking strict consumers, documenting its optionality, testing old clients against new responses, and defining when it becomes required. Bad: Add a field and call it backward compatible because the server still compiles. Good: Change an event schema through versioned messages or an additive transition with dual consumers and a rollback plan. Bad: Rename a serialized field and rely on every consumer upgrading at the same instant. ## Failure modes and recovery If consumers are unknown, treat compatibility as uncertain and use telemetry, repository search, or an explicit version boundary before breaking the interface. If old and new behavior cannot coexist, stage the migration or require an intentional breaking release. If documentation and runtime behavior disagree, report the contract as unresolved and fix the boundary before claiming compatibility. If rollback cannot restore the old contract, document the irreversible migration and require stronger review. ## Validation evidence and provenance Claims in this skill map to graded findings in [`docs/research.md`](../../docs/research.md): - HTTP semantics (C5, Strong fact): RFC 9110 defines the standardized behavior contracts that changes must respect. - Version-change semantics (C5, Moderate): semver communicates incompatibility classes; ecosystem effectiveness evidence is mixed, so consumer verification is required rather than version-number trust. - API evolution practice (C5, Moderate): Google's API design guide reflects operating internet-scale APIs. Source boundary: compatibility classification for a specific interface is always verified against actual consumers, not inferred from sources. Confidence: medium-high for explicit-contract and staged-evolution principles; medium for any specific classification. Freshness: review when consumers, version policy, serialization standards, or migration tooling change. Disconfirmation: widespread evidence that version-number discipline alone suffices without consumer verification would simplify the procedure. ## Related skills and conflicts Related: `requirements-to-acceptance`, `behavior-preserving-refactoring`, `secure-coding-review`, `performance-regression-analysis`, `regression-test-design`, `knowledge-maintenance`, and `repository-change-verification`. This skill does not authorize breaking unknown consumers, exposing data, or treating a version number as a substitute for migration evidence.