tdd
Spec-Driven TDD with a strict Red-Green-Refactor cycle using context-isolated subagents. Every feature starts with a structured spec; tests are generated from numbered acceptance criteria, with post-cycle three-dimension verification (Completeness + Traceability + Coherence), severity-tiered findings (CRITICAL/WARNING/SUGGESTION), and a spec-defect escape hatch that halts when the spec is internally contradictory or unimplementable instead of silently encoding it. Use when the user says "/tdd", "test first", "use tdd", "tdd approach", "write tests first", or when implementing new features/functionality. Trigger phrases: "implement", "add feature". Do NOT use for bug fixes without new tests, documentation changes, configuration-only changes, or refactoring existing code without new behavior.
kirillgreen · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add kirillgreen/skills --skill tdd技能说明
name: tdd description: > Spec-Driven TDD with a strict Red-Green-Refactor cycle using context-isolated subagents. Every feature starts with a structured spec; tests are generated from numbered acceptance criteria, with post-cycle three-dimension verification (Completeness + Traceability + Coherence), severity-tiered findings (CRITICAL/WARNING/SUGGESTION), and a spec-defect escape hatch that halts when the spec is internally contradictory or unimplementable instead of silently encoding it. Use when the user says "/tdd", "test first", "use tdd", "tdd approach", "write tests first", or when implementing new features/functionality. Trigger phrases: "implement", "add feature". Do NOT use for bug fixes without new tests, documentation changes, configuration-only changes, or refactoring existing code without new behavior.
/tdd -- Spec-Driven Test-Driven Development
Enforce strict Red-Green-Refactor cycle with context-isolated subagents. Every feature starts with a specification — numbered acceptance criteria that drive test generation and final verification.
Flow: Spec → RED → GREEN → REFACTOR → Verify
Mode Behavior Matrix
The cycle has two modes, set during Step 1. Each step below behaves differently per mode:
| Step | locked-spec mode | prompt-only mode |
|---|---|---|
| Step 1 Spec phase | Find existing or write new spec; lock at 1c | User declines spec; pass raw feature requirement instead |
| Phase 1 RED inputs | Locked spec + AC-traceability instruction | Raw feature requirement + "name tests after behavior, no AC- prefix" |
| Phase 1 spec coverage check | Compare tests vs locked spec criteria | SKIPPED (no spec) |
| Phase 2 spec_defect signal | Implementer can raise spec_defect: true | DISABLED (no spec to be defective) |
| Step 4 Spec Verification | All three dimensions run | SKIPPED (no spec to verify against) |
| Final Report Spec line | "[N] criteria defined" | "prompt-only mode" |
| Final Report Verification block | Three-dim breakdown | "Verification: skipped (prompt-only mode)" |
Maintenance note: every mode-specific branch in this runbook is tagged <!-- mode-fork --> in the source. Grep that tag to enumerate all branches. If this matrix grows past ~8 rows or the branches become hard to keep in sync, that's the signal to split /tdd-prompt-only into its own skill.
Severity Levels
Every check in this skill emits findings at one of three tiers:
- [CRITICAL] — blocks the phase transition by default. May be acknowledged-overridden only with explicit user confirmation, in which case the finding still appears in the final report tagged
(acknowledged). In Step 4 (the last phase, no next transition to block), CRITICAL findings do not stop the cycle from being reported, but they mandate a follow-up cycle and are listed prominently in the report. - [WARNING] — should fix. Cycle continues; findings appear in the final report.
- [SUGGESTION] — informational. Surfaced once at the end, never blocks.
The final report tallies counts by tier. CRITICAL gates phase transitions (with optional acknowledged override); WARNING and SUGGESTION are advisory.
Findings Ledger
A single mutable list findings_ledger is the source of truth for all findings emitted during the cycle. The Final Report MUST be generated from this ledger, not recounted from memory.
Schema (one entry per finding):
- id: "F-001" # sequential, never reused
tier: "CRITICAL" # CRITICAL | WARNING | SUGGESTION
phase: "post-red-lint" # post-red-lint | post-green | post-refactor | step-4a | step-4b | step-4c | phase-3.5 | step-0.5 | step-1b
check: "test-isolation" # short identifier for the check that fired
message: "Test imports utils/x.ts which doesn't exist"
file: "src/utils/__tests__/x.test.ts:12" # if applicable
acknowledged: false # true once the user explicitly approves the override
acknowledged_reason: null # user's stated reason if acknowledged
acknowledged_at_phase: null # phase where the override was approved
carried_from_strike: 1 # null | 1 | 2 — for entries that survive a spec-defect cycle restart
Append rules:
- Every emitted
[CRITICAL],[WARNING],[SUGGESTION]finding gets one ledger entry. No exceptions. - IDs are sequential within a cycle, starting at F-001. Never renumber.
- Severity tier and check identifier are immutable once written.
acknowledged: trueis set ONLY when the user explicitly confirms an override (e.g., accepts an out-of-scope file). Never auto-flip.- Step 0.5 intent-check acknowledgements also go in the ledger as
tier: "SUGGESTION",check: "intent-check-bypass",phase: "step-0.5",acknowledged: true,acknowledged_reason: <user's reason or "user chose (b) proceed">. This gives bug-fix-routed-acknowledgements a record so they appear in the Final Report. - Cycle restart stale-path scrub: when a spec-defect strike discards a test file, on strike-N entry the orchestrator strips from the ledger any entry whose
file:path references the discarded test file path (and whosephaseispost-red-lint). Cross-phase findings (Post-GREEN, Step 4, Phase 3.5) are NOT stripped — they describe state that may persist. Stripped entries are logged as(stripped on strike N — file discarded)for audit but not surfaced in the Final Report.
Cycle restart rules (spec-defect strike 1 → 2 transition):
- When a
spec_defect: truediscards the test file and restarts RED, the ledger is preserved — not reset. - Existing entries gain
carried_from_strike: 1so they're distinguishable in the Final Report. - New entries on strike 2 use the next sequential ID after the last strike-1 entry.
Final Report rule:
- The Final Report's
Findings: N CRITICAL · N WARNING · N SUGGESTIONline is computed bycount by tierover the ledger.