tdd（时差）

TDD

Value: Feedback -- short cycles with verifiable evidence keep AI-generated code honest and the human in control. Tests express intent; evidence confirms progress.

Purpose

Teaches a five-step TDD cycle (RED, DOMAIN, GREEN, DOMAIN, COMMIT) that adapts to whatever harness runs it. Detects available delegation primitives and routes to guided mode (human drives each phase) or automated mode (system orchestrates phases). Prevents primitive obsession, skipped reviews, and untested complexity regardless of mode.

Practices

The Five-Step Cycle

Every feature is built by repeating: RED -> DOMAIN -> GREEN -> DOMAIN -> COMMIT.

1. RED -- Write one failing test with one assertion. Only edit test files. Write the code you wish you had -- reference types and functions that do not exist yet. Run the test. Paste the failure output. Stop. Done when: tests run and FAIL (compilation error OR assertion failure).
2. DOMAIN (after RED) -- Review the test for primitive obsession and invalid-state risks. Create type definitions with stub bodies (todo!(), raise NotImplementedError, etc.). Do not implement logic. Stop. Done when: tests COMPILE but still FAIL (assertion/panic, not compilation error).
3. GREEN -- Address the immediate error — NEVER "make the test pass" in one go. Scope check before every change: can this be fixed with ~function-scope work (~20 lines, one file)? YES → make the change, run tests, check the next error. NO → drill down by writing a failing unit test for the smallest piece needed, then route it through a standard TDD cycle with swapped roles. Only edit production files (except when drilling down). Paste output after each change. Done when: tests PASS with minimal implementation.
4. DOMAIN (after GREEN) -- Review the implementation for domain violations: anemic models, leaked validation, primitive obsession that slipped through. If violations found, raise a concern and propose a revision. Done when: types are clean and tests still pass.
5. COMMIT -- Run the full test suite. Stage all changes and create a git commit referencing the GWT scenario. Run git status after committing to verify no uncommitted files remain. This is a hard gate: no new RED phase may begin until this commit exists and the working tree is clean. Done when: git commit created, all tests passing, working tree clean.

After step 5, either start the next RED phase or tidy the code (structural changes only, separate commit).

A compilation failure IS a test failure. Do not pre-create types to avoid compilation errors. Types flow FROM tests, never precede them.

Domain review has veto power over primitive obsession and invalid-state representability. Debate continues until resolved or escalated to the human — there is no round limit.

User-Facing Modes

Guided mode (/tdd red, /tdd domain, /tdd green, /tdd commit): Each phase loads references/{phase}.md with detailed instructions for that step. For experienced engineers who want explicit phase control. Works on any harness -- no delegation primitives required. The human decides when to advance phases.

Automated mode (/tdd or /tdd auto): The system detects harness capabilities, selects an execution strategy, and orchestrates the full cycle. The user sees working code, not sausage-making. For verbose output showing phase transitions and evidence, use /tdd auto --verbose.

Capability Detection (Automated Mode)

When automated mode activates, detect available primitives in this order:

1. Subagents available? Check for Agent tool. If present, use the subagents strategy with focused per-phase agents.
2. Fallback. Use the chaining strategy -- role-switch internally between phases within a single context.

Select the most capable strategy available. Do not attempt a higher strategy when its primitives are missing.

You are the orchestrator. The agent reading this file performs capability detection and dispatches directly. Do NOT spawn a single "orchestrator" subagent to do it for you -- that hides work, bypasses strategy detection, and pre-selects the wrong strategy. Whether you were invoked by /tdd, by the pipeline, or by any other caller: you detect capabilities, you choose the strategy, you spawn the phase agents yourself.

After determining your strategy, read ONLY the entry-point file for that strategy:

Do NOT read orchestrator.md when using chaining.

orchestrator.md references references/shared-rules.md for rules that apply to all strategies (domain veto, outside-in progression, pipeline integration, pre-implementation context checklist). Read shared-rules.md when directed by your strategy's entry-point file.

Execution Strategy: Chaining (Fallback)

Used when no delegation primitives are available. The agent plays each role sequentially:

1. Load references/red.md. Execute the RED phase.
2. Load references/domain.md. Execute DOMAIN review of the test.
3. Load references/green.md. Execute the GREEN phase.
4. Load references/domain.md. Execute DOMAIN review of the implementation.
5. Load references/commit.md. Execute the COMMIT phase.
6. Repeat.

Role boundaries are advisory in this mode. The agent must self-enforce phase boundaries: only edit file types permitted by the current phase (see references/phase-boundaries.md).

Execution Strategy: Subagents

Used when the Agent tool is available for spawning focused subagents. Each phase runs in an isolated subagent with constrained scope.

• Spawn each phase agent using Agent(subagent_type="<agent-name>", prompt="...") with the prompt template in references/{phase}-prompt.md.
• The orchestrator follows references/orchestrator.md for coordination rules.
• Structural handoff schema (references/handoff-schema.md): every phase agent must return evidence fields (test output, file paths changed, domain concerns). Missing evidence fields = handoff blocked. The orchestrator does not proceed to the next phase until the schema is satisfied.
• Context isolation provides structural enforcement: each subagent receives only the files relevant to its phase.

Named Team Member Personas (Subagent Strategy)

When .claude/agents/ definitions exist (from the ensemble-team skill), the subagent strategy uses named personas for ping and pong roles. The orchestrator selects team members based on slice context, spawns them as subagents using Agent(subagent_type="<agent-name>", prompt="..."), and collects results to pass as context to the next subagent.

See references/orchestrator.md for coordination rules and references/ping-pong-pairing.md for persona selection, rotation, and pairing history.

Phase Boundary Rules

Each phase edits only its own file types. This prevents drift. See references/phase-boundaries.md for the complete file-type matrix.

If blocked by a boundary, stop and return to the orchestrator (automated) or report to the user (guided). Never circumvent boundaries.

Walking Skeleton First

The first vertical slice must be a walking skeleton: the thinnest end-to-end path proving all architectural layers connect. It may use hardcoded values or stubs. Build it before any other slice. It de-risks the architecture and gives subsequent slices a proven wiring path to extend.

Outside-In TDD

Start from an acceptance test at the application boundary -- the point where external input enters the system. Drill inward through unit tests. The outer acceptance test stays RED while inner unit tests go through their own red-green-domain-commit cycles. The slice is complete only when the outer acceptance test passes.

A test that calls internal functions directly is a unit test, not an acceptance test -- even if it asserts on user-visible behavior.

Boundary enforcement by mode:

• Pipeline mode: The CYCLE_COMPLETE evidence must include boundary_type and boundary_evidence on the acceptance test. The pipeline's TDD gate rejects evidence where the acceptance test calls internal functions directly.
• Automated mode (non-pipeline): The orchestrator checks boundary scope and re-delegates if the first test is not a boundary test. Advisory -- no gate blocks progression.
• Guided mode: The human is responsible for ensuring boundary-level tests. The skill text instructs correct behavior but cannot enforce it.

Cycle-Complete Evidence

At the end of each complete RED-DOMAIN-GREEN-DOMAIN-COMMIT cycle, produce a CYCLE_COMPLETE evidence packet containing: slice_id, acceptance_test {file, name, output, boundary_type, boundary_evidence}, unit_tests {count, all_passing, output}, domain_reviews [{phase, verdict, concerns}], commits [{hash, message}], rework_cycles, team {ping, pong, domain_reviewer}.

When pipeline-state is provided in context metadata, the TDD skill operates in pipeline mode: it receives a slice_id and stores evidence to .factory/audit-trail/slices/<slice-id>/tdd-cycles/cycle-NNN.json. When running standalone, the evidence is informational only (not stored).

See references/cycle-evidence.md for full schema.

Harness-Specific Guidance

If running on Claude Code, also read references/claude-code.md for harness-specific rules including hook-based enforcement. For maximum mechanical enforcement, ask the bootstrap skill to install optional hooks from references/hooks/claude-code-hooks.json.

Enforcement Note

• Guided mode: Advisory. The human enforces by controlling phase transitions.
• Chaining mode: Advisory. The agent self-enforces phase boundaries.
• Subagent mode: Structural. Context isolation and handoff schemas enforce phase boundaries. Missing evidence blocks handoffs.
• Pipeline mode: Gating. Evidence gates reject incomplete phase transitions.
• Optional hooks (Claude Code): Mechanical. Pre-tool-use hooks block unauthorized file edits per phase. See references/claude-code.md.

Hard constraints:

• Phase boundary violation (wrong file type in wrong phase): [H]
• Domain veto escalation (contested design decision): [RP]
• Commit gate (no new RED before prior cycle committed): [H]

See references/constraint-resolution.md in the template directory for pipeline rework budget conflicts and domain veto resolution in pipeline mode.

Constraints

• Chaining mode self-enforcement: Self-enforcement means you produce the same file-type restrictions as if separate agents were enforcing them. Writing production code during RED phase violates this constraint even though no mechanism prevents it. If you catch yourself reasoning about why a phase boundary doesn't apply in chaining mode, you are violating it.
• "~20 lines, one file" scope check: This is a judgment heuristic, not a precise threshold. The spirit is: if the change touches multiple concerns, multiple files, or requires understanding distant code, it is too large for a single cycle. Do not game this by making a 40-line change across 2 functions in one file and claiming it's "one file."

Verification

After completing a cycle, verify:

• Every failing test was written BEFORE its implementation
• Domain review occurred after EVERY RED and GREEN phase
• Phase boundary rules were respected (file-type restrictions)
• Evidence (test output) was provided at each handoff
• Commit exists for every completed RED-GREEN cycle
• GREEN phase iterated one failure at a time (not full implementation in one pass)
• Working tree clean after every COMMIT (git status verified)
• Walking skeleton completed first (first vertical slice)

HARD GATE -- COMMIT (must pass before any new RED phase):

• All tests pass
• Git commit created with message referencing the current GWT scenario
• No new RED phase started before this commit was made

Dependencies

This skill works standalone. For enhanced workflows, it integrates with:

• domain-modeling: Strengthens the domain review phases with parse-don't-validate, semantic types, and invalid-state prevention principles.
• code-review: Three-stage review (spec compliance, code quality, domain integrity) after TDD cycles complete.
• mutation-testing: Validates test quality by checking that tests detect injected mutations in production code.
• ensemble-team: Provides real-world expert personas for pair selection and mob review.

Missing a dependency? Install with:

npx skills add jwilger/agent-skills --skill domain-modeling