reverse-engineer（逆向工程师）

Reverse Engineer (Route-Aware)

Step 2 of 6 in the Reverse Engineering to Spec-Driven Development process. The 6-step process: 1. Analyze, 2. Reverse Engineer (this skill), 3. Create Specs, 4. Gap Analysis, 5. Implementation Planning, 6. Implementation.

Estimated Time: 30-45 minutes Prerequisites: Step 1 completed (analysis-report.md and route selection in .stackshift-state.json) Output: 11 documentation files in docs/reverse-engineering/

Route-Dependent Behavior:

• Greenfield: Extract business logic only (framework-agnostic)
• Brownfield: Extract business logic + technical implementation details

Output is the same regardless of implementation framework (Spec Kit, BMAD, or BMAD Auto-Pilot). The framework choice only affects what happens after Step 2.

Configuration Check

Guard: Verify state file exists before proceeding.

if [ ! -f .stackshift-state.json ]; then
  echo "ERROR: .stackshift-state.json not found."
  echo "Step 1 (Initial Analysis) must be completed first. Run /stackshift.analyze to begin."
  exit 1
fi

DETECTION_TYPE=$(cat .stackshift-state.json | jq -r '.detection_type')
ROUTE=$(cat .stackshift-state.json | jq -r '.route')

if [ "$DETECTION_TYPE" = "null" ] || [ -z "$DETECTION_TYPE" ]; then
  echo "ERROR: detection_type missing from state file. Re-run /stackshift.analyze."
  exit 1
fi
if [ "$ROUTE" = "null" ] || [ -z "$ROUTE" ]; then
  echo "ERROR: route missing from state file. Re-run /stackshift.analyze."
  exit 1
fi

echo "Detection: $DETECTION_TYPE"
echo "Route: $ROUTE"

SPEC_OUTPUT=$(cat .stackshift-state.json | jq -r '.config.spec_output_location // "."')
echo "Writing specs to: $SPEC_OUTPUT"

if [ "$SPEC_OUTPUT" != "." ]; then
  mkdir -p "$SPEC_OUTPUT/docs/reverse-engineering"
  mkdir -p "$SPEC_OUTPUT/.specify/memory/specifications"
fi

State file structure:

{
  "detection_type": "monorepo-service",
  "route": "greenfield",
  "implementation_framework": "speckit",
  "config": {
    "spec_output_location": "~/git/my-new-app",
    "build_location": "~/git/my-new-app",
    "target_stack": "Next.js 15..."
  }
}

Capture commit hash for incremental updates:

COMMIT_HASH=$(git rev-parse HEAD 2>/dev/null || echo "unknown")
COMMIT_DATE=$(git log -1 --format=%ci 2>/dev/null || date -u +"%Y-%m-%d %H:%M:%S")
echo "Pinning docs to commit: $COMMIT_HASH"

Extraction approach based on detection + route:

• detection_type determines WHAT patterns to look for (shared packages, Nx project config, monorepo structure, etc.)
• route determines HOW to document them (tech-agnostic vs tech-prescriptive)

Phase 1: Deep Codebase Analysis

Use the Task tool with subagent_type=stackshift:stackshift-code-analyzer:AGENT to perform analysis. If the agent is unavailable, fall back to the Explore agent.

Error recovery: If a subagent fails or returns empty results for a sub-phase, retry once with the Explore agent. If the retry also fails, record the gap with an [ANALYSIS INCOMPLETE] marker and continue with remaining sub-phases.

Missing components: If a sub-phase finds no relevant code (e.g., no frontend in a backend-only service), document the absence in the corresponding output file rather than skipping the sub-phase.

Launch sub-phases 1.1 through 1.6 in parallel using separate subagent invocations. Collect all results before proceeding to Phase 2.

1.1 Backend Analysis

• Find all API endpoints and record their method, route, auth requirements, parameters, and purpose.
• Catalog every data model including schemas, types, interfaces, and field definitions.
• Inventory all configuration sources: env vars, config files, and settings.
• Map every external integration: APIs, services, and databases.
• Extract business logic from services, utilities, and algorithms.

1.2 Frontend Analysis

• List all pages and routes with their purpose and auth requirements.
• Catalog all components by category: layout, form, and UI components.
• Document state management: store structure and global state patterns.
• Map the API client layer: how the frontend calls the backend.
• Extract styling patterns: design system, themes, and component styles.

1.3 Infrastructure Analysis

• Document deployment configuration: IaC tools, cloud provider, and services.
• Map CI/CD pipelines and workflows.
• Catalog database setup: type, schema, and migrations.
• Identify storage systems: object storage, file systems, and caching.

1.4 Testing Analysis

• Locate all test files and identify the testing frameworks in use.
• Classify tests by type: unit, integration, and E2E.
• Estimate coverage percentages by module.
• Catalog test data: mocks, fixtures, and seed data.

1.5 Business Context Analysis

• Read README, CONTRIBUTING, and any marketing or landing pages.
• Extract package descriptions and repository metadata.
• Identify comment patterns indicating user-facing features.
• Collect error messages and user-facing strings for persona inference.
• Analyze naming conventions to reveal domain concepts.
• Examine git history for decision archaeology.

1.6 Decision Archaeology

• Inspect dependency manifests (package.json, go.mod, requirements.txt) for technology choices.
• Analyze config files (tsconfig, eslint, prettier) for design philosophy.
• Review CI/CD configuration for deployment decisions.
• Run git blame on key architectural files to identify decision points.
• Collect comments with "why" explanations (TODO, HACK, FIXME, NOTE).
• Look for rejected alternatives visible in git history or comments.

Progress signal: After all sub-phases complete, log: "Phase 1 complete: Analysis gathered for [list which sub-phases produced results]."

Phase 2: Generate Documentation

Create docs/reverse-engineering/ directory and generate all 11 documentation files. For each file, apply the greenfield or brownfield variant as described in operations/output-file-specs.md. Read that file now for the detailed per-file specifications.

If .stackshift-docs-meta.json already exists, overwrite it completely with fresh metadata.

Step 2.1: Write metadata file FIRST

COMMIT_HASH=$(git rev-parse HEAD 2>/dev/null || echo "unknown")
COMMIT_DATE=$(git log -1 --format=%ci 2>/dev/null || date -u +"%Y-%m-%d %H:%M:%S")
GENERATED_AT=$(date -u +"%Y-%m-%dT%H:%M:%SZ")

Write docs/reverse-engineering/.stackshift-docs-meta.json:

{
  "commit_hash": "<COMMIT_HASH>",
  "commit_date": "<COMMIT_DATE>",
  "generated_at": "<GENERATED_AT>",
  "doc_count": 11,
  "route": "<greenfield|brownfield>",
  "docs": {
    "functional-specification.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "integration-points.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "configuration-reference.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "data-architecture.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "operations-guide.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "technical-debt-analysis.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "observability-requirements.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "visual-design-system.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "test-documentation.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "business-context.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" },
    "decision-rationale.md": { "generated_at": "<GENERATED_AT>", "commit_hash": "<COMMIT_HASH>" }
  }
}

Step 2.2: Add metadata header to each doc

Every generated doc starts with this header after the title:

# [Document Title]

> **Generated by StackShift** | Commit: `<short-hash>` | Date: `<GENERATED_AT>`
> Run `/stackshift.refresh-docs` to update with latest changes.

Step 2.3: Generate files with checkpoints

Generate files in this order, logging progress after each:

Batch 1 (core architecture):

1. functional-specification.md
2. data-architecture.md
3. integration-points.md
4. configuration-reference.md

After writing files 1-4, log: "Generated 4/11 files (core architecture complete)." Verify the output directory contains 4 files before continuing.

Batch 2 (operations and quality): 5. operations-guide.md 6. technical-debt-analysis.md 7. observability-requirements.md 8. visual-design-system.md 9. test-documentation.md

After writing files 5-9, log: "Generated 9/11 files (operations and quality complete)." Verify the output directory contains 9 files before continuing.

Batch 3 (context and decisions): 10. business-context.md 11. decision-rationale.md

After writing files 10-11, log: "Generated 11/11 files. Phase 2 complete."

Output structure:

docs/reverse-engineering/
├── .stackshift-docs-meta.json
├── functional-specification.md
├── integration-points.md
├── configuration-reference.md
├── data-architecture.md
├── operations-guide.md
├── technical-debt-analysis.md
├── observability-requirements.md
├── visual-design-system.md
├── test-documentation.md
├── business-context.md
└── decision-rationale.md

Success Criteria

• All 11 documentation files generated in docs/reverse-engineering/
• Comprehensive coverage of all application aspects
• Framework-agnostic functional specification (for greenfield)
• Complete data model documentation
• Business context captured with clear [INFERRED] / [NEEDS USER INPUT] markers
• Decision rationale documented with ADR format
• Integration points fully mapped with data flow diagrams
• .stackshift-docs-meta.json created with commit hash for incremental updates
• Each doc has metadata header with commit hash and generation date

Next Step

Once all documentation is generated:

For GitHub Spec Kit (implementation_framework: speckit): Proceed to Step 3 -- use /stackshift.create-specs to transform docs into .specify/ specs.

For BMAD Method (implementation_framework: bmad): Proceed to Step 6 -- hand off to BMAD's *workflow-init. BMAD's PM and Architect agents use the reverse-engineering docs as context.

For BMAD Auto-Pilot (implementation_framework: bmad-autopilot): Proceed to /stackshift.bmad-synthesize to auto-generate BMAD artifacts. The 11 reverse-engineering docs provide ~90% of what BMAD needs.

DO / DON'T

DO:

• Describe WHAT the system does, not HOW (especially for greenfield)
• Use all available signals for inference: README, comments, naming, config, git history
• Mark confidence levels: no marker = confident, [INFERRED] = reasonable inference, [NEEDS USER INPUT] = genuinely unknown
• Cross-reference between docs (e.g., tech debt informs trade-offs)
• Cite specific evidence for each inference

DON'T:

• Hard-code framework names in functional specs (greenfield)
• Mix business logic with technical implementation (greenfield)
• Fabricate business goals with no supporting evidence
• State inferences as facts without marking them
• Skip a section because it requires inference -- attempt it and mark confidence

Completeness Checklist

Verify analysis captured:

• ALL API endpoints (not just the obvious ones)
• ALL data models (including DTOs, types, interfaces)
• ALL configuration options (check multiple files)
• ALL external integrations
• ALL user-facing strings and error messages (for persona/context inference)
• ALL config files (for decision rationale inference)

Each document must be comprehensive, accurate, organized, actionable, and honest about inferred vs verified information.