> For AI agents: the complete documentation index is available at /llms.txt, the full documentation bundle is available at /llms-full.txt. # Testing Choose the test layer from the contract at risk. A schema decision and a packaged executable need different evidence. ## Layers and test doubles ### Why A mocked file writer cannot prove that restoration preserves unrelated files. Disposable directories and real Git subprocesses exercise those contracts without requiring a paid service. ### Rules - MUST cover consequential pure decisions, validation failures, and failure-handling branches with unit tests. \[review] - MUST exercise filesystem, subprocess, and network-adapter contracts at an integration boundary. \[review] - MUST supply actual files in disposable directories when verifying filesystem behavior. \[review] - MUST limit test doubles to external services and inject their clients through adapter factories. \[review] - MUST exercise internal modules directly rather than replacing the asserted behavior with a fake. \[review] - MUST keep paid model calls and nondeterministic model assertions outside ordinary test runs. \[review] - MUST avoid cases that only repeat a trivial pass-through or a TypeScript guarantee. \[review] - MAY retain smoke tests against the built CLI to verify startup, packaging, and executable behavior. \[review] \[review] The configured test runner executes value and resource cases. Review checks real boundary use, meaningful failure coverage, and accurate verification reports. ## Organization and verification ### Why The failing test name should explain the broken contract. Keeping cases beside their implementation makes that contract easier to maintain. ### Rules - MUST colocate cases in `*.test.ts`; SHOULD use `*.integration.test.ts` for larger resource suites when a separate runner exists. \[review] - MUST use `describe` groups and present-tense behavior verbs in `it` names. \[review] - MUST keep each case focused on one behavior, allowing multiple assertions to establish it. \[review] - MUST invoke the consuming repository's validation commands, using workspace filters when it has workspaces. \[review] - MUST build a distributed artifact before testing that artifact's entrypoint. \[review] - MUST report a testless package as testless, rather than describing it as covered. \[review] \[review] The configured test runner executes value and resource cases. Review checks real boundary use, meaningful failure coverage, and accurate verification reports. ## Enforcement The configured test runner executes value and resource cases. Review checks real boundary use, meaningful failure coverage, and accurate verification reports.