Shipkit Codebase Index

shipkit-codebase-index - Codebase Navigation Index

Purpose: Generate a lightweight index so Claude can navigate your codebase without wasteful exploration.

When to Invoke

User triggers:

• "Index the codebase"
• "Create a project map"
• "Generate codebase index"

Auto-suggested when:

• Session hook detects no index exists
• Index is older than 14 days

Process

Completion Tracking (MANDATORY)

Before starting analysis, create tasks for each index section:

1. TaskCreate: "Run generator script (base index)"
2. TaskCreate: "Detect framework"
3. TaskCreate: "Identify entry points"
4. TaskCreate: "Map concepts to files (with verification)"
5. TaskCreate: "Identify core files (with import counts)"
6. TaskCreate: "Determine skip list"
7. TaskCreate: "Write completed codebase-index.json"
8. TaskCreate: "Verify all 5 Claude-filled fields are populated"

Rules:

• TaskUpdate each task to completed only after the section has verified data (not guesses)
• The final verification task requires reading the written file and confirming: framework, entryPoints, concepts, coreFiles, and skip are all non-empty
• Do NOT declare done if any field is empty or contains placeholder values

Step 1: Run Generator Script

python ${CLAUDE_SKILL_DIR}/scripts/generate_index.py

Script provides (100% reliable data):

• scripts — from package.json
• recentlyActive — files from git history (last 14 days)
• directories — which common directories exist
• configFiles — which config files exist

Script leaves empty (Claude fills in):

• framework
• entryPoints
• concepts
• coreFiles
• skip

Step 2: Claude Analyzes and Completes Index

Read the generated index, then fill in the empty fields.

USE SUBAGENT FOR CONCEPT MAPPING - Launch Explore subagent for efficient parallel scanning:

Agent tool with subagent_type: "Explore"
Prompt: "Scan codebase to build navigation index. Find and report:

1. FRAMEWORK: Check for next.config.*, vite.config.*, nuxt.config.*, etc.
2. ENTRY POINTS: Find main app entry, layout, API routes directory, database schema
3. CONCEPTS: Map these concepts to files:
   - auth: files handling authentication, sessions, login
   - database: db connections, models, schema
   - payments: billing, subscriptions, checkout
   - api: API route handlers
   - components: reusable UI components
4. CORE FILES: Files imported by 5+ other files (high fan-in)

For each concept, list the actual file paths found.
For core files, include import count."

Why subagent: Concept mapping requires scanning multiple directories and patterns in parallel. Explore agent is optimized for this and reduces main conversation context.

Fallback (if subagent unavailable) - Manual detection:

1. Detect framework from configFiles:

   • next.config.js → Next.js
   • vite.config.ts → Vite
   • prisma/schema.prisma → uses Prisma
   • etc.

2. Identify entry points by checking which files exist:

   • src/app/page.tsx → app entry
   • src/app/layout.tsx → layout
   • src/app/api/ → API routes
   • prisma/schema.prisma → database schema

3. Map concepts to files by scanning the codebase:

   • Look for auth-related files → concepts.auth
   • Look for database files → concepts.database
   • Look for payment/billing → concepts.payments
   • etc.

4. Identify core files — files that are imported by many others

5. Skip list — derive a default skip list (no user prompt; this skill runs in fork context). Defaults: node_modules, dist, build, .next, coverage, .git, .venv, venv, __pycache__, .pytest_cache, target, vendor. If any of these directories don't exist, omit them. Add additional entries if the codebase has an obvious "legacy" or "deprecated" folder at the top level.

Step 2.5: Verification Requirements

Before claiming any index entry, verify it with tool calls:

Verification sequence for each entry type:

Config files:
  1. Glob: pattern="[config-pattern]"
  2. If empty → don't include in index
  3. If found → add to configFiles with verified path

Entry points:
  1. Glob: pattern="[entry-path]"
  2. If empty → mark as "unverified" or skip
  3. If found → Read file, confirm it exports something meaningful
  4. Add to entryPoints with status: "verified"

Concepts:
  1. Grep: pattern="[concept-keyword]" glob="**/*.{ts,tsx}"
  2. List ALL matching files
  3. If 0 matches → don't add concept
  4. If matches → add concept with verified file list

Core files:
  1. Grep: pattern="import.*from.*[filename]" glob="**/*.{ts,tsx}"
  2. Count imports per file
  3. Files with >5 imports → core files
  4. Include import count in index

Mark unverified entries: If verification cannot be completed, mark entry as status: "unverified" in index rather than guessing.

See also: shared/references/VERIFICATION-PROTOCOL.md for standard verification patterns.

Step 3: Update the Index

import json

with open('.shipkit/codebase-index.json') as f:
    index = json.load(f)

index['framework'] = 'next.js (app router)'
index['entryPoints'] = {
    'app': 'src/app/page.tsx',
    'layout': 'src/app/layout.tsx',
    'api': 'src/app/api/',
    'database': 'prisma/schema.prisma'
}
index['concepts'] = {
    'auth': ['src/lib/auth.ts', 'src/middleware.ts'],
    'database': ['src/lib/db.ts', 'prisma/schema.prisma'],
    # ... more concepts
}
index['coreFiles'] = ['src/lib/db.ts', 'src/lib/auth.ts']
index['skip'] = ['src/legacy/']

with open('.shipkit/codebase-index.json', 'w') as f:
    json.dump(index, f, indent=2)

Preserve the timestamps. The generator script sets generated, fullRefreshedAt, and mechanicalRefreshedAt. This Step loads-then-updates only the judgment fields, so those stamps carry through untouched — a full run is a fresh judgment derivation, so fullRefreshedAt correctly reflects now. Don't delete or overwrite them.

Step 4: Confirm to User

✅ Codebase index complete at .shipkit/codebase-index.json

Framework: next.js (app router)
Entry points: 4 (app, layout, api, database)
Concepts: 3 (auth, database, payments)
Recently active: 15 files
Skip: src/legacy/

I'll use this index to navigate faster.

Division of Labor

Principle: Script does 100% reliable mechanical tasks. Claude does anything requiring judgment.

Output: .shipkit/codebase-index.json

{
  "generated": "YYYY-MM-DD",
  "scripts": { "<name>": "<command>" },
  "recentlyActive": ["path/to/file.ts"],
  "directories": ["src/app", "src/components"],
  "configFiles": ["next.config.js", "tsconfig.json"],
  "framework": "next.js (app router)",
  "entryPoints": { "app": "...", "api": "...", "database": "..." },
  "concepts": { "auth": [...], "database": [...] },
  "coreFiles": ["src/lib/db.ts"],
  "skip": ["src/legacy/"]
}

Full schema reference: See references/output-schema.md

Realistic example: See references/example.json

How Claude Uses the Index

Context Files This Skill Writes

• .shipkit/codebase-index.json — Complete replacement on each run

Completion Checklist

• Script ran and created base index
• Framework detected
• Entry points identified
• Concepts mapped to files
• Core files identified
• User confirmed skip list (if any)
• Index saved

After Completion

Codebase index written to .shipkit/codebase-index.json.

Next: The index is read during execution by skills that navigate code — /shipkit-spec, /shipkit-plan, /shipkit-preflight, /shipkit-review-shipping, /shipkit-prompt-audit, /shipkit-ux-audit, and others use it for faster file/symbol lookup.

Staying fresh (you rarely need to re-run this). The index keeps itself current on two cadences:

• Mechanical layer (recentlyActive, directories, configFiles, scripts) — refreshed automatically, with no LLM, on every commit (a git commit-scoped hook) and at session start. A content-hash cache makes this near-instant and writes nothing when nothing changed.
• Judgment layer (framework, entryPoints, concepts, coreFiles, skip) — only a full /shipkit-codebase-index run re-derives these (they need Claude). The auto-refresh preserves them untouched.

So re-run /shipkit-codebase-index only when the semantic shape shifts — new modules, a new framework, a major refactor that moves where concepts live. Session start nudges you when fullRefreshedAt is older than 14 days. (Note: the commit hook fires on commits Claude makes; a commit from your own terminal is picked up at the next session start.)