brownfield（brownfield 搜索）

Brownfield Setup

Configures Claude Code for an existing codebase by discovering what's already there. Conventions are detected from the actual project — not assumed from templates.

Run the workflow skill after this to install commands and agents:

npx skills add squirrelsoft-dev/agent-skills@workflow

Step 1 — Preflight

Check for existing setup and available execution mode:

ls .claude/ 2>/dev/null && echo "EXISTS" || echo "FRESH"
echo "${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-disabled}"
claude --version 2>/dev/null || echo "unknown"
command -v jq &>/dev/null && echo "jq: installed" || echo "jq: MISSING — required for brownfield analysis"

If .claude/ already exists, ask the developer:

"A .claude/ directory already exists. Running this will overwrite it. Continue, or run the greenfield skill instead for a fresh setup?"

Wait for confirmation before proceeding.

If jq is not installed, stop and instruct:

"The brownfield skill requires jq for JSON processing. Install it first: sudo apt-get install jq (Debian/Ubuntu) or brew install jq (macOS)."

Step 2 — Parallel Analysis

Run all four analysis scripts. Use subagents in parallel if available, otherwise run sequentially. All scripts are read-only — no files are written in this phase.

bash "${CLAUDE_SKILL_DIR}/scripts/analyze-repo.sh"        > /tmp/bf-repo.json
bash "${CLAUDE_SKILL_DIR}/scripts/analyze-conventions.sh" > /tmp/bf-conventions.json
bash "${CLAUDE_SKILL_DIR}/scripts/analyze-git.sh"         > /tmp/bf-git.json
bash "${CLAUDE_SKILL_DIR}/scripts/analyze-tests.sh"       > /tmp/bf-tests.json

Verify all four JSON files are valid:

jq . /tmp/bf-repo.json /tmp/bf-conventions.json /tmp/bf-git.json /tmp/bf-tests.json >/dev/null

Step 3 — Human Review Checkpoint

Synthesize findings from the four JSON outputs and present this summary. Do NOT write any files yet.

Load reference files for context:

• Rules content guidelines: ${CLAUDE_SKILL_DIR}/references/rules-templates.md
• Hook recommendations: ${CLAUDE_SKILL_DIR}/references/hook-detection.md

Present findings:

## Brownfield Analysis — [Project Name from basename $PWD]

### Stack & Runtime
[language, framework, versions, package manager from bf-repo.json]

### Conventions Detected
[naming, imports, formatting from bf-conventions.json — with real file examples from the codebase]

### Testing
[framework, file locations, exact run command from bf-tests.json]

### Git & Collaboration
[commit style with 3 real examples from bf-git.json, branch pattern, CI platform]

### ⚠️ Conflicts & Ambiguities
[list any competing conventions that need a human decision]
[e.g. "Files in /legacy use snake_case, files in /src use camelCase — which is current?"]

### Tech Debt Signals
[TODO/FIXME density, files untouched 12+ months, deprecated patterns]

Detect and report quality tools:

command -v gitleaks &>/dev/null && echo "gitleaks: installed" || echo "gitleaks: missing"
command -v semgrep  &>/dev/null && echo "semgrep: installed"  || echo "semgrep: missing"
command -v gh       &>/dev/null && echo "gh CLI: installed"   || echo "gh CLI: missing"

For any missing tools, list what they do and how to install them:

Recommended tools (install before running /workflow):

  gitleaks — Secret/credential detection
    https://github.com/gitleaks/gitleaks#installing

  semgrep — SAST security scanner
    https://semgrep.dev/docs/getting-started/

  gh CLI — Required for /triage, /pr, /squash-pr, /address-pr-comments
    https://cli.github.com/

Do NOT offer to install these tools automatically.

Check for Previous Configuration

Before asking interview questions, check if this skill has been run before:

SKILL_NAME="brownfield" \
bash "${CLAUDE_SKILL_DIR}/scripts/load-config.sh"

If the output is a non-empty JSON object (not {}), previous answers exist. Present the analysis summary above and the saved answers:

Previous brownfield configuration found:

  Agent teams:   [AGENT_TEAMS]
  Commit style:  [COMMIT_STYLE]
  CI/CD:         [CI_CD]
  Corrections:   [CORRECTIONS or "none"]

Would you like to:
  a) Update — use these settings with the fresh analysis above and regenerate artifacts
  b) Reconfigure — answer all questions from scratch

Use AskUserQuestion and wait for the response.

• If Update: load saved values into environment variables. If the fresh analysis reveals new conflicts or ambiguities not covered by saved corrections, ask only about those. Then skip to Step 3b.
• If Reconfigure: proceed to the full interview below.

If no previous config exists, ask all questions:

1. Does this match your understanding? Correct anything that's wrong.
2. For any conflicts listed above — which convention is the current standard?
3. Enable agent teams? (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS) — yes / no
4. Commit style going forward? — detected style / conventional commits / free-form
5. Scaffold GitHub Actions CI if none detected? — yes / no

Wait for developer response before proceeding.

Save Configuration

After the interview (or after loading previous config for an update), save the current answers:

SKILL_NAME="brownfield" \
CONFIG_JSON='{"AGENT_TEAMS":"[val]","COMMIT_STYLE":"[val]","CI_CD":"[val]","CORRECTIONS":"[val]"}' \
bash "${CLAUDE_SKILL_DIR}/scripts/save-config.sh"

Replace [val] with actual values. CORRECTIONS should be a JSON-escaped string of any corrections the developer made to the analysis. This writes to .claude/skill-config.json so future runs can reuse these answers.

Proceed to Step 3b.

Step 3b — Architectural Analysis

After the developer confirms script findings, run a deeper analysis of the actual codebase to identify patterns that scripts can't detect.

Load the investigation guide: ${CLAUDE_SKILL_DIR}/references/architecture-analysis.md

Use an Explore subagent (context: fork) to read the codebase. The subagent should use the confirmed stack, org pattern, and file locations from the analysis JSON as a starting point to know where to look.

What to investigate

1. Dependency and call patterns — centralized clients, import graph, service layers
2. Architectural boundaries — top-level directory responsibilities, frozen/legacy areas
3. Error handling patterns — centralized handlers, custom error classes, consistent shapes
4. Auth and permissions patterns — middleware, guards, decorators, role checks
5. State management patterns — local vs global state, server state (React Query/SWR)
6. Testing patterns — shared utilities, custom matchers, mock strategies
7. Git history patterns — co-change coupling, high-churn areas (if gh CLI available)

Output

Generate additional .claude/rules/ files only when genuine patterns are found. Do not pad with generic advice. Example files:

• architecture.md — if clear architectural layers or centralized patterns found
• legacy.md — if frozen/deprecated directories found
• gotchas.md — if project-specific traps discovered

Developer review

Present generated rules for review before writing:

## Additional Rules Generated from Codebase Analysis

I identified these architectural patterns and generated corresponding rules:

**architecture.md** — centralized API client pattern, feature boundary enforcement
**gotchas.md** — 3 project-specific traps found

Preview:
[show first few lines of each file]

Are these accurate? I can:
- Accept all and proceed
- Edit any file before committing
- Skip specific files

Wait for developer confirmation. Then proceed to Step 4.

Step 4 — Generate All Artifacts

After the developer confirms findings and answers questions, generate everything. Use subagents in parallel if available, otherwise sequential.

Pass confirmed answers as environment variables to each script.

4a — Rules

STACK="[detected]" \
CONVENTIONS_JSON="/tmp/bf-conventions.json" \
TESTS_JSON="/tmp/bf-tests.json" \
REPO_JSON="/tmp/bf-repo.json" \
bash "${CLAUDE_SKILL_DIR}/scripts/generate-rules.sh"

4b — Hooks

STACK="[from bf-repo.json]" \
PKG_MANAGER="[from bf-repo.json]" \
FORMATTER="[from bf-repo.json]" \
TEST_RUNNER="[from bf-repo.json, e.g. vitest/jest/pytest/go-test]" \
TYPECHECKER="[python only — ty/pyright/mypy, ask at interview if unset]" \
AGENT_TEAMS="[from developer answer]" \
bash "${CLAUDE_SKILL_DIR}/scripts/generate-hooks.sh"

4c — Settings

PROJECT_DIR="$PWD" \
PKG_MANAGER="[from bf-repo.json]" \
AGENT_TEAMS="[from developer answer]" \
LINT_CMD="[from bf-repo.json]" \
TEST_CMD="[from bf-repo.json]" \
bash "${CLAUDE_SKILL_DIR}/scripts/generate-settings.sh"

4d — CLAUDE.md

REPO_JSON="/tmp/bf-repo.json" \
GIT_JSON="/tmp/bf-git.json" \
PROJECT_NAME="[detected or asked]" \
COMMIT_STYLE="[from developer answer]" \
bash "${CLAUDE_SKILL_DIR}/scripts/generate-claude-md.sh"

4e —.gitignore

PROJECT_DIR="$PWD" \
bash "${CLAUDE_SKILL_DIR}/scripts/generate-gitignore.sh"

Content guidelines for rules: ${CLAUDE_SKILL_DIR}/references/rules-templates.md Hook recommendations: ${CLAUDE_SKILL_DIR}/references/hook-detection.md

Step 5 — Commit

git add .claude/ CLAUDE.md
git commit -m "chore: add Claude Code configuration for brownfield onboarding"

If conventional commits were detected or chosen:

git commit -m "chore(claude): add brownfield Claude Code configuration"

Step 6 — Completion

✅ Brownfield Setup Complete — [Project Name]

Generated:
- CLAUDE.md
- .claude/settings.json        (agent teams: yes/no)
- .claude/settings.local.json
- .claude/rules/               [list each file]
- .claude/hooks/               [list each script]

Key conventions captured:
- [most important naming rule with example]
- [commit format with example]
- [test location and run command]

Next steps:
1. npx skills add squirrelsoft-dev/agent-skills@workflow
2. Launch Claude Code
3. Run /workflow to install commands and agents
4. Restart Claude Code

Reference Files

• Rules content guidelines: ${CLAUDE_SKILL_DIR}/references/rules-templates.md
• Hook recommendation logic: ${CLAUDE_SKILL_DIR}/references/hook-detection.md