tdd-migration-pipeline（TDD 迁移管道）

TDD Migration Pipeline

Orchestrator-only workflow for migrating or rewriting codebases. You do NOT read files, write code, or validate anything yourself. You only instruct agents and pipe context (paths, not contents).

When to Use

• Migrating codebase from one language/framework to another
• Rewriting a system with TDD guarantees
• Large-scale refactoring with behavioral contracts
• When you want zero context growth in the orchestrator

Core Principles

1. ZERO orchestrator execution - only instruct and pipe context
2. All work done by agents - you never read/write/validate
3. Context window stays flat - pass paths, not contents
4. New code in separate directory - never modify source

Your Constraints

• Never say "let me read..." or "looking at..."
• Only say "Agent X: do Y with Z"
• Your context should NOT grow during execution
• All agents must use qlty and tldr skills

Pipeline Phases

Phase 1: SPEC

Instruct spec-agent (use scout or architect):
- Analyze {source_path} using tldr-skill
- Output: spec.md with behavioral contracts, types, edge cases

Agent prompt template:

Analyze the codebase at {source_path} using tldr commands (tldr structure, tldr extract, tldr calls).
Create spec.md with:
- All behavioral contracts (what each function/class promises)
- Input/output types
- Edge cases and invariants
- Dependencies between components
Write to: {target_dir}/spec.md

Phase 2: FAILING TESTS

Instruct test-agent (use arbiter):
- Read spec.md
- Write failing tests in {target_dir}/tests/
- Tests should define expected behavior before implementation

Instruct review-agent (use critic):
- Validate tests cover spec completely
- No gaps in behavioral coverage

Phase 3: ADVERSARIAL (x3 iterations)

Instruct premortem-agent (use premortem skill):
- Review spec + tests
- Identify failure modes, race conditions, edge cases
- DO NOT ASK - just add mitigations directly to spec
- Run 3 passes with fresh perspective each time

Key: Each pass should find NEW issues, not repeat previous ones.

Phase 4: PHASED PLAN

Instruct planner-agent (use architect or plan-agent):
- Input: spec.md + tests + mitigations
- Output: phased-plan.yaml
- Requirements:
  - Dependency-ordered phases
  - Each phase = one testable unit
  - Clear inputs/outputs per phase

Phase 5: BUILD LOOP (per phase)

For each phase in phased-plan.yaml:

  Instruct builder-agent (use kraken or spark):
  - Write code to pass tests for this phase
  - Use qlty for quality checks
  - Run tests after each change

  Instruct review-agent (use critic or judge):
  - Validate implementation matches spec
  - Check for regressions in previous phases
  - Verify no breaking changes

Phase 6: INTEGRATION VALIDATION

Instruct integration-agent (use atlas or validator):
- Use tldr to diff against {reference_repo}
- Check for:
  - No race conditions
  - No hangs or deadlocks
  - No breaking changes vs original
  - All behavioral contracts preserved
- Output: validation-report.md

Invocation

When invoking this workflow, specify:

SOURCE: {path to source code}
TARGET_DIR: {new folder for migrated code}
TARGET_LANG: {typescript|python|go|rust|etc}
REFERENCE_REPO: {url or path for final diff comparison}
SKILLS: [tldr-code, qlty-check, {domain-specific}]

Agent Mapping

Example Orchestration

# Phase 1
Task(scout): "Analyze /src/old-system using tldr structure and tldr extract.
              Create spec.md at /migration/spec.md with all behavioral contracts."

# Phase 2
Task(arbiter): "Read /migration/spec.md. Write failing tests to /migration/tests/
               that define expected behavior."

Task(critic): "Review /migration/spec.md vs /migration/tests/.
              Report any behavioral gaps."

# Phase 3 (x3)
Task(premortem): "Review /migration/spec.md and /migration/tests/.
                 Identify failure modes. Add mitigations directly to spec. Pass 1/3."
[repeat with "Pass 2/3", "Pass 3/3"]

# Phase 4
Task(architect): "From /migration/spec.md and /migration/tests/,
                 create /migration/phased-plan.yaml with dependency-ordered phases."

# Phase 5 (loop)
Task(kraken): "Implement phase 1 from /migration/phased-plan.yaml.
              Code goes in /migration/src/. Run tests after."

Task(critic): "Review /migration/src/ against /migration/spec.md.
              Check for spec compliance and regressions."
[repeat for each phase]

# Phase 6
Task(atlas): "Run full integration tests on /migration/src/.
             Use tldr to diff against /src/old-system.
             Output /migration/validation-report.md."

Anti-Patterns (DO NOT)

• Reading files into your context ("let me check the code...")
• Writing code directly ("I'll implement this function...")
• Validating anything yourself ("looking at the tests, I see...")
• Modifying the source directory
• Skipping the adversarial phase
• Running build without tests first

Success Criteria

• All tests pass
• qlty reports clean
• tldr diff shows no breaking changes
• No race conditions or hangs
• validation-report.md confirms behavioral equivalence