Based on the approved spec, write complete test cases. Tests are the ONLY executable contract between human and agent. Once approved, the agent's sole objective is to make these tests pass.

Step 1: Read Spec

Read .harness/specs/$ARGUMENTS*.md. Extract all AC (Acceptance Criteria) and INV (Invariants).

If spec status is not approved, STOP and tell user to run /proposal first.

Step 2: Detect Test Framework

Auto-detect the project's test framework:

Check package.json for jest/vitest/mocha
Check for pytest.ini, setup.cfg, pyproject.toml for pytest
Check for *_test.go files for Go
Check for Cargo.toml for Rust
Fall back to asking the user

Step 3: Generate Three-Layer Tests

For each AC-N, generate:

V1 Unit Test: Pure logic validation (data transforms, validation rules, computations)
V2 Integration Test: Module interaction (state management, API calls, event flow)
V3 E2E Test: User perspective (operation flow, UI state, screenshots)

For each INV-N, generate V1 invariant tests.

Naming Convention

[F001] Feature Name
  [AC-1] Acceptance Criteria Description
    [T001-V1] should... (unit)
    [T002-V2] should... (integration)
    [T003-V3] should... (e2e)
  [AC-2] ...
    [T004-V1] should...
  [INV-1] Invariant Description
    [T010-V1] must always...

Step 4: Test Quality Requirements

Each test has exactly one assertion focus
Test descriptions express expected behavior in plain language
Boundary cases covered: null, empty, extreme values, error inputs
Test IDs are globally unique — they will directly map to atomic task IDs
Every test must be independently runnable

Step 5: Confirm All RED

Run all tests. Confirm ALL FAIL. This proves tests are valid — they don't accidentally pass without implementation.

Step 6: Output

Test files written to project's test directory
Output test case list: each test's ID + description + layer
Update .harness/progress.md with test alignment results
Wait for human approval

Key Principle: Once human approves these tests, they become the SOLE standard for implementation. The agent's goal is to make them GREEN.

Next Step — Auto-Navigate

Read the flow field from .harness/specs/{FXXX}-*.md
Display progress visualization:

📋 SDD 进度 — {FXXX}: {feature name}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✅ proposal  →  ✅ tdd-align  →  ⬜ {next_phase}  →  ...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
当前阶段: tdd-align (测试对齐) ✅ 已完成
下一阶段: {next_phase} ({description})

Ask naturally: "测试对齐完成，共 {N} 个测试（V1: {x}, V2: {y}, V3: {z}），全部 RED。下一步是 {next phase}（{description}）。是否继续？"
If user confirms → execute next phase's skill logic directly
If flow says next is implement (SMALL complexity, no decompose) → execute implementation directly without task DAG