JoseAntonioNuevo/test-suite-doctor

test-suite-doctor 是做什么的？

Shrink an oversized test suite (e.g. 2,000+ tests, much of it AI-generated slop) to a small, high-quality suite while retaining a configurable share of real coverage — measured, not guessed.

Hard rules

1. Metrics before opinions. Never propose deleting a test before collect and minimize have produced a plan. Reading test code and judging it "useless" is not evidence — a trivial-looking test can be the only cover for a branch.
2. The plan is a proposal, not a verdict. Coverage cannot see API contracts, documented regression cases, or property invariants. Every drop candidate gets reviewed (step 3) before deletion.
3. Work on a branch. Before deleting anything, verify the working tree is clean and create a branch (e.g. test-doctor/minimize). Never delete tests on a dirty tree or the default branch.
4. Verify or revert. The job is not done until verify exits 0. If it fails, regenerate tests for the reported gaps and re-verify — or restore the deleted tests.

Requirements check (do this first)

• Node.js ≥ 22 and a package.json using a locally installed Vitest or Jest (node <skill-root>/dist/cli.mjs collect --help).
• A coverage provider must be installed (Vitest: @vitest/coverage-v8; Jest has one built in). If missing, offer to install it as a dev dependency.
• Mutation verification additionally needs @stryker-mutator/core configured in the target repo — it is opt-in, do not install it unprompted.
• The suite should be green before starting. If the baseline run reports failures, tell the user and get agreement on how to handle them before minimizing (failing tests contribute unreliable coverage).

All commands below run from the target repo root; $SKILL is this skill's directory. Artifacts land in .test-doctor/ (suggest gitignoring it).

Workflow

1. MEASURE (script — always first)

node "$SKILL/dist/cli.mjs" collect --cwd .

Detects the runner, runs the full suite once as the coverage baseline, then re-runs each test file in isolation recording covered lines/branches and runtime. Writes .test-doctor/report.json.

• Default granularity is per test file (per-test isolation is exact but runs every test separately — offer --granularity test only for small or already-reduced suites, or scoped with --filter).
• On a 2,000-test suite this takes a while; run it in the background and report progress. Take collectionErrors in the report seriously — those units block minimization by default. If the user explicitly chooses --keep-unmeasured, every affected unit is a mandatory keep and the plan is marked untrusted.
• --filter creates a deliberately scoped baseline: collection, minimization, and verification cover only matching test files. State that scope in every recommendation; it is not a whole-repository guarantee.

2. MINIMIZE (script)

node "$SKILL/dist/cli.mjs" minimize --coverage-floor 0.97 --target-count 200

Greedy weighted-sum selection: repeatedly keeps the unit with the most newly covered lines/branches per millisecond of runtime until the coverage floor is met. Writes .test-doctor/plan.json and a human-readable .test-doctor/plan.md with a justification per kept and dropped unit.

• Ask the user for their thresholds if not stated; defaults are a 0.97 line floor, no branch floor, no target count. --help lists all knobs.
• Force-keep known contract/regression tests up front: --keep 'contract' --keep 'regression' (regex on unit id).
• Heed the plan's warnings array — e.g. the floor being unreachable, or requiring more units than the target count. Relay them to the user verbatim.

3. REVIEW (judgment — yours)

Read references/slop-patterns.md, then review the drop list in plan.md:

• Rescue any test that verifies behavior invisible to coverage: public API contracts, regressions referencing an issue/bug ID, error-message wording relied on by consumers, property/invariant checks. Re-run minimize with additional --keep patterns rather than editing the plan by hand.
• Confirm the rest are safe deletions; classify them by slop pattern so the user sees why each dies (the patterns file has the catalog).

Then review the keep list against references/quality-rules.md: kept tests that smell (mock-the-mock, tautologies, hollow snapshots) get rewritten in place — they earn their coverage, but they should earn it cleanly.

Only after this review, delete the dropped test files/cases on the branch. Delete whole files when all their units are dropped; edit files when only some tests within them die (test granularity).

4. REGENERATE (judgment — yours)

Run verify once (expect it may fail) to get the coverage-gap report, then write new tests for the worst-hit source files listed by verify, following references/quality-rules.md: AAA structure, behavior-driven names, one behavior per test, mocks only at external boundaries. Prefer one strong test covering a real user-visible behavior over three shallow ones.

5. VERIFY (script — gates completion)

node "$SKILL/dist/cli.mjs" verify --coverage-floor 0.97
# optionally, for critical modules only (slow):
node "$SKILL/dist/cli.mjs" verify --mutation --mutate "src/billing/**/*.ts" --mutation-floor 80

Re-runs the suite, requires it green, and requires line retention vs the baseline to meet the floor. Exit 0 = done; exit 1 = fix and re-run (regenerate more tests, or restore deletions); exit 2 = environment problem, stop and report. Mutation testing is opt-in per module because it is slow — offer it for the modules the user calls critical.

Version 2 baselines fingerprint covered source, runner configuration, lockfile, runner, and coverage provider. If verify reports provenance drift, collect a new baseline. Use --allow-provenance-drift only with explicit user agreement and report the resulting verdict as untrusted. Legacy v1 baselines similarly require the explicit --allow-legacy-baseline escape hatch.

Reporting back

Summarize for the user: units before → after, line/branch retention, runtime before → after, slop patterns found (with counts), tests rescued from the drop list and why, tests regenerated, and the final verify verdict. Include the path to plan.md for the full audit trail.

Failure modes

• Runner not detected / ambiguous → the CLI exits 2 with guidance; pass --runner or --cwd (monorepos: run per package).
• Incomplete collection → fix the isolated failures. Do not filter them away merely to make a plan pass; a filter changes the declared baseline scope. --keep-unmeasured is the conservative escape hatch.
• Verify keeps failing after regeneration → stop looping after ~3 attempts; present the residual gap file-by-file and let the user decide between lowering the floor and keeping more tests.
• Suite was red before starting → do not "fix" unrelated failures silently; report them and agree on scope first.