ai-conductor: Skill for Claude Code

Write Tests

Design tests for this repository. Use the provider-neutral tdd skill separately when implementing behavior test-first.

1. Classify the test

Choose the narrowest level that proves the behavior.

• Unit: Exercise one function, class, transition, or adapter contract. Inject dependencies and mock every process, network, LLM, GitHub, filesystem boundary not under test.
• Integration: Exercise real collaboration between selected internal components. Use real temporary files or local Git only when they are the boundary under test. Fake all third-party services.
• Acceptance: Exercise an observable story or gate across the minimum real internal path. Fake third-party boundaries; create only the authoritative artifacts the scenario requires.
• Smoke: Exercise a real executable or third-party service. Put it under src/conductor/test/smoke/ or name it *.smoke.test.ts. Smoke tests are excluded from the ordinary suite.

Only explicit smoke tests may call real Claude, Codex, GitHub, package registries, HTTP services, or other third parties.

2. Select the seam before writing the fixture

Name the exact behavior and terminal observation first. Test that seam directly.

• Prefer exported pure helpers for selection, classification, ordering, and state calculations.
• For an engine-native boundary, invoke the smallest method that owns the behavior.
• Use Conductor.run() only when the test must prove orchestration across multiple steps.
• Never use a full conductor run merely to prove plugin registration, provider invocation, marker contents, configuration resolution, event emission, or one state transition.

If the assertion can be made without completing the full workflow, the test MUST stop at that boundary.

3. Bound every Conductor fixture

Before calling Conductor.run(), write down:

1. The first step that may run.
2. The exact steps expected to dispatch.
3. The condition that ends the run.
4. Every completion artifact or gate verdict required before that endpoint.