Test Plan

/skill:test-plan — TDD-First Test Planning

Purpose: Generate test specifications from task requirements BEFORE implementation. Tests serve as executable specifications that define what the code must do. Designed for iterative use by team agents in a TDD workflow. Project-agnostic — adapts to any TypeScript monorepo.

When to Use

• Starting TDD on a new task — generate it.todo() stubs before implementation
• Validating whether an existing test suite covers the task requirements
• Refreshing a plan after requirements change

When NOT to Use

• Implementation already exists and passes — use /review-tests to audit coverage
• Non-TypeScript repo without an equivalent test runner convention
• Exploratory spike code that will be discarded

Test Philosophy

Tests should verify behavior through public interfaces, not implementation details. Code can change entirely; tests shouldn't break.

Good tests exercise real code paths through public APIs and describe what the system does. A good test reads like a specification — "user can checkout with valid cart." These survive refactors because they don't care about internal structure.

Bad tests are coupled to implementation: they mock internal collaborators, test private methods, or assert on call counts. The warning sign: your test breaks when you refactor but behavior hasn't changed.

// Good — tests observable behavior through the interface
test("createUser makes user retrievable", async () => {
  const user = await createUser({ name: "Alice" });
  const retrieved = await getUser(user.id);
  expect(retrieved.name).toBe("Alice");
});

// Bad — bypasses the interface to verify internals
test("createUser saves to database", async () => {
  await createUser({ name: "Alice" });
  const row = await db.query("SELECT * FROM users WHERE name = ?", ["Alice"]);
  expect(row).toBeDefined();
});

Anti-pattern: horizontal slices. Do NOT implement all it.todo() stubs in bulk, then fill them in. That's "RED all, GREEN all" — it produces tests that verify imagined behavior, test data shapes instead of capabilities, and go stale quickly.

WRONG (horizontal):  todo → todo → todo → todo → impl → impl → impl → impl
RIGHT (vertical):    todo → impl → todo → impl → todo → impl

Work through stubs one at a time: write the minimal implementation to pass one test, then move to the next.

Mocking: Mock at system boundaries only — external APIs, third-party services, time, randomness. Never mock your own modules or internal collaborators. Prefer real substitutes (PGLite for Postgres, in-memory filesystem) over mocks where they exist.

Implementation Cycle Checklist

Before moving from one stub to the next, verify:

[ ] Test describes behavior, not implementation
[ ] Test uses public interface only
[ ] Test would survive an internal refactor
[ ] Implementation is minimal for this test
[ ] No speculative features added

After all stubs pass, check for refactor candidates:

• Extract duplication
• Deepen modules (move complexity behind simpler interfaces)
• Apply SOLID where it emerges naturally — never anticipate it

Never refactor while RED. Get to GREEN first, always.

Usage

/skill:test-plan --task="Implement vendor CRUD service"   # Plan + generate test files
/skill:test-plan --task="..." --ontology docs/prds/vendors/ONTOLOGY.md   # Use an explicit ontology
/skill:test-plan --validate                                # Run tests, report pass/fail
/skill:test-plan --refresh                                 # Update plan after implementation changes

Modes

Full Plan (default — --task="...")

Runs the complete orchestrator FSM. Produces test files with it.todo() stubs, a plan JSON file, and a markdown report. Use when starting a new feature.

Validate (--validate)

Runs generated test files and reports results:

• Reads .test-plan/plan-latest.json to find generated test files
• Executes tests via Vitest: pnpm vitest run --reporter=json {files}
• Updates plan with pass/fail/todo counts
• Does NOT regenerate test files or write a new markdown report

Refresh (--refresh)

Re-analyzes the task after implementation changes:

• Reads current implementation to discover new patterns and changed signatures
• Updates specs for evolved code
• Generates updated test files
• Archives old plan, writes new plan-latest.json and markdown report

Ontology input (--ontology <path>)

Optional in every mode. Points at an ONTOLOGY.md whose format is defined in _internal/ontology-readiness/SKILL.md. When one resolves, its settled rows become the primary source of test specs and prose re-derivation becomes the fallback.

Default resolution when the flag is omitted:

1. The sibling ONTOLOGY.md of the requirements artifact, when the task names one or one is discoverable (docs/prds/<slug>/PRD.md → docs/prds/<slug>/ONTOLOGY.md).
2. Otherwise the most recently modified docs/prds/*/ONTOLOGY.md.
3. Otherwise none — analysts re-derive constraints from prose exactly as they do today. Absence of an ontology is not an error and never BLOCKED.

Only rows whose Status column reads settled are used. deferred and unknown rows are ignored with a logged note in the report naming the row and its state — see _internal/ontology-readiness/SKILL.md ### Item states. An unsettled row is an open decision, not a test target.

Project Discovery (Run Once at Start)

Before entering the FSM, load project context:

1. Read CLAUDE.md from repo root — learn project conventions, error handling, shared package, test framework
2. Read .test-plan/config.json if it exists — get team mappings, analyst overrides (see foundations/project-config.md)
3. Read workspace config — pnpm-workspace.yaml or root package.json workspaces field to discover package boundaries
4. Identify shared package — from config sharedPackage field, or auto-detect

This context is passed to every analyst.

Orchestrator FSM (7 States)

State 1: DISCOVER_ASK

Parse the task description to understand what's being built.

Inputs: --task argument, current branch name, agent prompt (if team agent)

Extract:

• Entity: The primary domain entity (vendor, rfi, task, change-order)
• Package: Target package (packages/api, packages/db, apps/web)
• Layer: Target layer (service, router, schema, component)
• Operations: What operations are being built (CRUD, transitions, calculations)

Layer detection: Use the task description first. If ambiguous, use file path patterns from .test-plan/config.json layerDetection field, or defaults:

• **/services/** → service
• **/routers/** → router
• **/schema/** → schema
• **/*.tsx → component

For --validate mode: Skip. Read existing plan from .test-plan/plan-latest.json.

For --refresh mode: Read existing plan, then re-parse to find changes.

State 2: GATHER_CONTEXT

Read all relevant source files to build the specification context.

Read in order: 0. ONTOLOGY.md — the path from --ontology, or the default resolution in ## Modes above. Read it first when one resolves: its settled rows constrain how every schema below is interpreted. Parse ## Entity Types (reference schemes), ## Fact Types (Constraints and Modality), ## Lifecycles (Total flag and transition rows), and ## Temporality. Keep each row's Status; drop anything that is not settled and log it.

1. Validation schemas from the project's shared/types package for the target entity — input schemas, output types, enums

2. DB schema from the database package for the target entity — columns, types, defaults, relations, status enums

3. Enum definitions from the shared package — status values, categories

4. Reference implementation — find an existing service/router for a similar entity to learn the established patterns

5. CLAUDE.md conventions — error handling patterns, logging, boundary rules, testing expectations

6. Classify dependencies using _internal/dependency-classification/SKILL.md.

   For each dependency in scope of the target, record <name> — <category> — <chosen test approach> in the plan output. Flag miscategorizations (e.g., Postgres mocked rather than substituted with PGLite) as findings. Never mock in-process or local-substitutable dependencies — the rubric explains why.

Ontology slicing. When step 0 produced an ontology, split its settled rows into one slice per analyst. A row may land in more than one slice; a row that lands in none is carried into the report as unsliced context.

Each slice is a verbatim copy of the matching rows with their # or heading, so the analyst can cite the row a spec came from. The integration-surface slice is the only one that carries rows from two sections: its edge-crossing fact-type rows travel with the ## Entity Types row of every entity they name, so the analyst has each entity's reference scheme alongside the fact.

For --validate mode: Skip. Context already in plan file.

State 3: SELECT_ANALYSTS

Choose which analysts to run based on the target layer:

Check .test-plan/config.json analysts field for overrides (enabled/disabled).

State-lifecycle exception: Include state-lifecycle if the entity has a status enum, or if the ontology declares a ## Lifecycles section for the target entity — an ontology-declared lifecycle selects the analyst even when no status enum exists in code yet, and its specs then define the enum the implementation must add. If neither holds, skip even for service layer.

For --validate mode: Skip. Analysts determined by existing plan.

State 4: DISPATCH_ANALYSTS

For each selected analyst:

1. Read test-plan/analysts/{name}/SKILL.md
2. Pass the task context, Zod schemas, DB schema, reference implementation, and CLAUDE.md conventions
3. Pass that analyst's ontology slice from State 2, when an ontology resolved — for integration-surface that is the edge-crossing ## Fact Types rows plus the ## Entity Types rows (reference schemes) of every entity those fact types name. Omit the item when no ontology resolved or the analyst's slice is empty — the analyst then re-derives from prose as before.
4. Analyst produces TestSpecification[] conforming to the canonical schema in foundations/test-case-schema.md. Specs that violate the schema are dropped at State 5 with a warning — the analyst is responsible for valid output.

Analysts run conceptually in parallel — each operates on the same context independently.

Analyst unavailable / disabled: If .test-plan/config.json disables a selected analyst, skip it with a logged note in the report. The plan is still produced from the remaining analysts; the verdict downgrades to PARTIAL if the missing analyst would have contributed P1 specs.

For --validate mode: Skip. Specifications already in plan file.

State 5: CONSOLIDATE_PLAN

Merge analyst outputs into a unified plan:

0. Validate against schema: drop any spec that violates foundations/test-case-schema.md. Record the count and reason in the report; do not silently keep malformed specs.

1. Deduplicate: If two analysts produced specs for the same behavior (same testFile + describeBlock + similar testName), keep the higher-priority one. If same priority, keep the one with more assertions. Conflict resolution is fully defined here — analysts do not coordinate.

2. Assign to test files: Group specs by testFile path. Each unique path becomes one generated file.

3. Calculate summary: Count totals by priority, analyst, and category.

4. Determine verdict:

   • READY: All P1 specs generated, context was complete
   • PARTIAL: P1 specs generated but some context was missing (e.g., no Zod schema found)
   • BLOCKED: Critical context missing (no entity found, no schemas, can't determine layer)

For --validate mode: Skip. Proceed to running tests.

State 6: GENERATE_OUTPUT

Full/Refresh Mode

1. Generate test files: Pass grouped specs to the test writer (see test-writer/SKILL.md). The test writer creates .test.ts files with it.todo() stubs.

2. Write plan JSON: Write .test-plan/plan-latest.json conforming to foundations/plan-schema.md.

   • Archive existing plan-latest.json to plan-{planId}.json first (if exists)

3. Write markdown report: Use template from foundations/report-template.md. Save to docs/test-plans/.

4. Update report index: Prepend to docs/test-plans/index.md. Update docs/test-plans/latest.md.

Validate Mode

1. Run tests: Execute generated test files:

   pnpm vitest run --reporter=json {test-file-paths}
   

2. Parse results: Count passed, failed, todo tests.

3. Calculate validation verdict:

   • ALL_PASS: All non-todo tests pass
   • PARTIAL_PASS: Some pass, some fail
   • FAILING: >50% of non-todo tests fail

4. Update plan: Add validation field to .test-plan/plan-latest.json. Bump timestamp.

State 7: RESPOND

Present summary to the caller.

Full/Refresh Mode Response

Test Plan: {entity}.{layer} — {verdict}

Task: "{task description}"
Analysts: {list}

Specs: {total} ({p1} P1, {p2} P2, {p3} P3)

Generated files:
  {path} — {testCount} tests ({todoCount} todos)

Plan: .test-plan/plan-latest.json
Report: {reportPath}

Next steps:
  1. Review generated test files
  2. Implement the feature, making tests pass in P1 → P2 → P3 order
  3. Run /skill:test-plan --validate when ready

Validate Mode Response

Test Plan Validation: {entity}.{layer} — {validation verdict}

Results: {passed} passed, {failed} failed, {todo} todo

{If FAILING: list failed tests with error messages}

Plan updated: .test-plan/plan-latest.json

Next steps:
  {ALL_PASS: "Proceed to /skill:domain-review"}
  {PARTIAL_PASS: "Fix failing tests, then re-validate"}
  {FAILING: "Major implementation issues — review failing tests"}

Critical: Always include the plan file path in your response. The calling team agent needs this to track progress.

Team Agent Integration

This skill is designed for TDD use by team agents. The typical lifecycle:

task received → /skill:test-plan --task="..."    (generate specs)
implementing  → fill in it.todo() stubs   (TDD)
"done"        → /skill:test-plan --validate      (verify tests pass)
refactored    → /skill:test-plan --refresh       (update specs)
passing       → /skill:domain-review               (quality check)
clean         → create PR

See foundations/team-agent-protocol.md for the complete workflow.

Integration with Code Review

The test-plan and domain-review skills are complementary:

Both must pass before creating a PR. See foundations/team-agent-protocol.md for the combined workflow.

Relationship to /skill:execute-plan and /skill:execute-prd

/skill:execute-plan verifies each task's acceptance criteria and reports coverage holes as unproved rows in its final report. It does NOT call /skill:test-plan.

Run /skill:test-plan manually before invoking /skill:execute-plan when you want stronger TDD discipline: analyst-generated specs, classified dependency handling, and it.todo() stubs that pre-declare the behavior coverage. The stubs then feed directly into execute-plan's coverage tracking — a test that was generated by /skill:test-plan and passes is unambiguously proved.

If you skip /skill:test-plan, execute-plan still verifies acceptance criteria per task but without the upfront spec generation. This is fine for straightforward features; use /skill:test-plan first when the behavior surface is complex or when unproved rows on the final report are unacceptable.

Relationship to native skills

This skill generates the stubs that superpowers:test-driven-development's red phase consumes; invoke it from within that process, not instead of it.

Contract

• Inputs: task description and target entity name; optional .test-plan/config.json (analyst overrides, team mappings, paths); optional --ontology <path> (an ONTOLOGY.md in the format defined by _internal/ontology-readiness/SKILL.md, defaulting to the requirements artifact's sibling then the most recently modified docs/prds/*/ONTOLOGY.md); Zod schemas, DB schema, reference implementation, CLAUDE.md conventions. Calls each test-plan/analysts/<name>/SKILL.md selected by State 3 and test-plan/test-writer/SKILL.md in State 6. Spec schema canonical at test-plan/foundations/test-case-schema.md.
• Preconditions: TypeScript monorepo with a Vitest-shaped test convention; foundations/* files readable; .test-plan/ directory writable; entity exists in repo (DB and Zod schemas locatable).
• Outputs: .test-plan/plan-latest.json (conforming to foundations/plan-schema.md); markdown report under docs/test-plans/ (per foundations/report-template.md); .test.ts files with it.todo() stubs from test-writer (does not overwrite existing files).
• Postconditions: plan verdict is READY, PARTIAL, or BLOCKED; existing tests are preserved; index docs/test-plans/index.md and latest.md updated.
• Failure modes: no ontology resolves → analysts re-derive from prose, plan proceeds (not an error); deferred / unknown ontology rows → ignored with a logged note naming the row and its state, per _internal/ontology-readiness/SKILL.md ### Item states; critical context missing (no entity / schemas / layer) → BLOCKED; analyst disabled in .test-plan/config.json → continue and downgrade to PARTIAL if its P1s are missed; spec violating canonical schema → drop at State 5 with a logged note (do not silently keep).