tdd（时差）

Test-Driven Development

Treat executable evidence as the source of truth. Use this workflow to prevent five common agent failures:

• guessing instead of proving
• writing many tests before learning anything
• testing shapes or internals instead of behavior
• over-mocking code under your control
• letting the feedback loop get so slow that TDD stops working

Start the Session

Before editing code:

1. Read repo-local instructions such as AGENTS.md, CLAUDE.md, and package or test scripts.
2. Inspect nearby production code and nearby tests to infer naming, seams, fixtures, and the fastest targeted command.
3. Identify two commands up front:

- the fastest command that runs the single target test - the broader command that validates the changed boundary before finishing

1. Choose the smallest public or near-public seam that can prove the behavior. See seams.md.
2. Ask the user only when the public contract, intended behavior, or required coverage scope is materially ambiguous.

Follow repo-local instructions if they are stricter than this skill.

Work Vertically

Do not batch all tests first and all implementation later.

Wrong:
  RED:   test1, test2, test3
  GREEN: impl1, impl2, impl3

Right:
  RED -> GREEN -> REFACTOR: test1 -> impl1
  RED -> GREEN -> REFACTOR: test2 -> impl2
  RED -> GREEN -> REFACTOR: test3 -> impl3

Write one failing test. Make it pass with the smallest sensible change. Refactor only on green. Repeat.

Choose the Entry Path

New feature

1. Start with the smallest user-visible or caller-visible behavior worth shipping.
2. Write one tracer-bullet test at the chosen seam.
3. Make it fail for the right reason.
4. Implement the thinnest slice that turns the test green.
5. Add the next behavior only after the current one is proven.

Bug fix or regression

1. List 2-3 plausible hypotheses before changing code.
2. Reproduce the bug with the smallest failing regression test that would have caught it.
3. Narrow the reproduction with logs, assertions, or a lower seam if the first repro is noisy.
4. Fix the code under that failing test.
5. Add one neighboring test only if it proves the fix is specific rather than accidental.

See bugfixes.md for the detailed mini-loop.

Legacy code or refactor

1. Freeze current behavior with a characterization test before reshaping internals.
2. Refactor behind that safety rail until the design exposes a better seam.
3. Replace overly broad characterization coverage with tighter behavioral tests when the seam improves.
4. Use the green state to deepen modules and simplify interfaces.

See seams.md, deep-modules.md, and refactoring.md.

Run the Core Loop

For each cycle:

1. RED: Write or tighten one test that proves one behavior. Confirm it fails.
2. GREEN: Write the minimum production change that makes only that behavior pass.
3. REFACTOR: Clean up duplication, naming, and structure while staying green.
4. Re-run the smallest relevant command first, then widen verification as confidence grows.

If the test cannot fail, the loop is invalid. Break it on purpose, lower the seam, or add the missing observability before trusting it.

Keep Feedback Fast

Use this ladder:

1. Run the single target test while iterating.
2. Run the surrounding file, package, or focused suite after a small cluster of green cycles.
3. Run broader verification before finishing: the relevant integration suite, typecheck, lint, or full tests for the touched boundary.
4. Call out any verification gap explicitly if time, tooling, or environment prevents broader checks.

If the loop feels slow, the seam is probably too high. Move down a level unless the behavior truly lives in the browser or across system boundaries.

Choose Assertions That Survive Refactors

• Assert observable outcomes, not helper calls.
• Prefer public interfaces over internal collaborators.
• Mock only system boundaries you do not control. See mocking.md.
• Treat tests as specifications for behavior, not snapshots of implementation shape.
• Keep each test about one behavior, even if that behavior needs more than one assertion.

See tests.md for examples and rewrites.

Use Subagents Carefully

When other agents help:

• Keep ownership of the failing test, the red/green loop, and final verification in the main thread.
• Let workers explore implementation ideas or refactors under the existing failing test.
• Reject fixes that only pass by weakening the test unless the test was proving the wrong behavior.

Avoid These Anti-Patterns

• Editing production code for the target behavior before seeing a meaningful failing test
• Writing the whole test plan up front
• Solving multiple behaviors in one red/green cycle
• Using browser or end-to-end tests for logic that could be proven faster elsewhere
• Mocking modules you own just to make the test convenient
• Leaving debug scaffolding, speculative branches, or unverified refactors behind

Finish with Proof

Consider the task done only when:

• the changed behavior is covered by a test that failed before the change
• the targeted tests pass
• the relevant broader verification command has run, or the gap is documented clearly
• refactors happened only while green
• the final summary states what behavior is now proven and what still relies on manual verification

Read More Only As Needed

• tests.md for durable behavior-level assertions and bad-to-better rewrites
• mocking.md for boundary-only mocking rules
• seams.md for seam selection, test levels, and fast feedback
• bugfixes.md for regression-test-first debugging
• interface-design.md for testable interfaces
• deep-modules.md for hiding complexity behind small APIs
• refactoring.md for green-only cleanup targets