Codebase Adapter

Pre-injected project snapshot (verify in phase 2 — do not skip discovery):

Working directory: !pwd

Lockfile detected: !ls bun.lock pnpm-lock.yaml yarn.lock package-lock.json 2>/dev/null | head -1 || echo "none"

Agents installed: !ls .claude/agents/ 2>/dev/null | tr '\n' ' ' || ls agents/ 2>/dev/null | tr '\n' ' ' || echo "(not found)"

MCP servers configured: !python3 -c "import json; s=json.load(open('.claude/settings.json')); print(', '.join(s.get('mcpServers',{}).keys()) or '(none)')" 2>/dev/null || echo "(not found)"

package.json: !cat package.json 2>/dev/null || echo "(not found)"

You MUST NOT change:

• Any <phase_*> logic or step sequences
• Dispatch protocols, decomposition rules, execution mode decisions
• Core agent principles, role descriptions, or behavior narratives
• Output template formats or section headers
• Skill descriptions or their trigger conditions
• XML tag structures (<workflow>, <domain_areas>, <skill_map>, etc.)

Store the resolved plugin root. All edits in phases 3–5 use paths relative to it.

Discover with Read, Glob, Grep — verify everything, assume nothing.

Package manager — detect from lockfile: bun.lock → bun, pnpm-lock.yaml → pnpm, yarn.lock → yarn, package-lock.json → npm.

Monorepo — check for workspaces in package.json, pnpm-workspace.yaml, or turbo.json. Record workspace names and their paths.

Tech stack — read package.json dependencies in root and each workspace:

• HTTP framework: Fastify, Express, Hono, Elysia, or none
• ORM/DB: Drizzle, Prisma, Kysely + database type (PostgreSQL, MySQL, SQLite)
• Frontend: Next.js (detect App Router via app/ directory), Vite + React, or none
• Validation: Zod (check version — v3 imports from "zod", v4 imports from "zod/v4"), Valibot
• Queue: BullMQ, or none
• Cache: ioredis, @upstash/redis, or none
• Testing: Vitest or Jest; Playwright or Cypress
• Docker: check for Dockerfile or docker-compose.yml

Project paths (verify with Glob — record exact paths):

• Backend source root (e.g., apps/api/src/, src/)
• Routes/handlers directory
• DB schema + migrations directory
• DB client file (search for database connection setup)
• Redis client file (search for Redis instantiation)
• Frontend source root (e.g., apps/web/, src/)
• Shared package: directory path and package name (e.g., packages/shared/, @acme/shared)

Quality gates — read scripts from each workspace's package.json:

• typecheck, lint, test, build command names
• Compose the full run command per workspace (e.g., pnpm --filter @acme/api typecheck)

Available skills — search across all scopes in order:

1. [plugin-root]/skills/ (plugin or project-local)
2. .claude/skills/ in the project root (project scope)
3. ~/.claude/skills/ (personal scope) Merge and deduplicate — plugin-root takes priority.

Available agents — search across all scopes in order:

1. [plugin-root]/agents/ and [plugin-root]/.claude/agents/ (plugin or project-local)
2. .claude/agents/ in the project root (project scope)
3. ~/.claude/agents/ (personal scope) Merge and deduplicate — plugin-root takes priority.

Available MCPs — read .claude/settings.json in project root (and ~/.claude/settings.json as fallback). Extract mcpServers keys.

<domain_areas> block (lines between the opening and closing tags)

Rebuild the Backend and Frontend tables:

• "Agent type" column: only include agents whose .md file exists in [plugin-root]/.claude/agents/
• "Key skills" column: only include skills whose directory exists in [plugin-root]/skills/ AND are relevant to the discovered tech stack
• Remove rows for sub-areas with no matching tech (e.g., remove "Queues" row if BullMQ is absent; remove "Pages/routing" row if no frontend framework is detected)
• Keep the "When" column text unchanged
• Keep the surrounding narrative text and the Cross-cutting specialists table unchanged

<skill_map> block

Update each sub-area's skill list:

• Remove skills not present in [plugin-root]/skills/
• Remove skills not relevant to discovered tech (e.g., remove /bullmq if BullMQ not installed)
• Keep the exact slash-command format (e.g., /drizzle-pg, /fastify-best-practices)

<available_mcps> block

Replace the bullet list with only MCPs found in mcpServers. Keep the same format: - [name] — [one-line purpose]. If context7 is configured, always keep its usage instructions below the list intact.

<required_skills> example block

Update the example Skill() calls to reflect the project's actual primary backend skills (e.g., replace fastify-best-practices if Fastify is not the framework).

skills: frontmatter list Remove skills not installed in [plugin-root]/skills/ or not relevant to the discovered tech stack. Do not add skills not already listed. If all skills remain valid, skip this file.

Quality gate command examples Find patterns like yarn workspace @myapp/api test or yarn workspace @myapp/api tsc. Replace with the actual workspace run command composed in phase 2. Use real workspace package names where known.

Project path references Find and replace generic placeholder paths with actual discovered paths:

Package name references Replace placeholder names with actual package names:

Tech-specific notes

• Zod import: update to match discovered version (from "zod" for v3, from "zod/v4" for v4)
• ESM flag: match "type": "module" presence in the workspace's package.json

Skip any agent where none of the above patterns are found — do not force edits.

Find the workspace command pattern example (text like yarn workspace <name> <script>, pnpm --filter <name> <script>). Replace all occurrences with the actual package manager syntax discovered in phase 2. Use <name> and <script> as placeholders where the real values would vary.

Use the data already gathered in phases 1–2 to fill in the values:

Agent count — count .md files in the project's .claude/agents/ directory (this includes both plugin-installed agents and any locally defined agents).

Skill count — count directories in the project's .claude/skills/ directory (this includes both plugin-installed skills and any locally defined skills).

agent-memory — check whether agent-memory/ exists at the project root. If it does, count its subdirectories and note which agents are excluded (agents without a matching subdirectory).

Paths — use the actual plugin-root-relative paths discovered in phase 2 (e.g., real skills/, agents/, .claude/ locations).

Generate the section using this template, substituting bracketed placeholders with real values:

## Claude Code Workflow Architecture

### Content Layout

\`\`\`
.claude/
  agents/          ← [AGENT_COUNT] agent definitions (markdown, YAML frontmatter)
  skills/          ← [SKILL_COUNT] skill directories (symlinked from skills/)
  settings.local.json  ← Permission allowlist for Claude Code

.claude-plugin/
  plugin.json      ← Claude Code plugin manifest
  marketplace.json ← Marketplace listing metadata

agents/            ← Published copy of .claude/agents (included in npm package)
[AGENT_MEMORY_LINE]
skills/            ← [SKILL_COUNT] skill directories (each has SKILL.md + optional references/ and scripts/)
\`\`\`

### Agent Definitions (\`.claude/agents/*.md\`)

Each agent is a single markdown file with YAML frontmatter:

\`\`\`yaml
---
name: agent-name
color: named-color       # Visual identification in Claude Code UI
description: "..."
skills:                  # Optional — skills loaded when agent is invoked
  - skill-slug
---
\`\`\`

Body contains: role description, core principles, workflow steps, quality gates, output format, and a reference to \`agent-memory/<name>/MEMORY.md\`.

### Skill Definitions (\`skills/<name>/\`)

Each skill directory contains:
- \`SKILL.md\` — main content with YAML frontmatter (\`name\`, \`description\`, optional \`argument-hint\`, optional \`allowed-tools\`)
- \`references/\` — deep-dive sub-files linked from SKILL.md
- \`scripts/\` — optional executables Claude can run (Python, bash) for visual output or automation

Where [AGENT_MEMORY_LINE] is:

• If agent-memory/ exists: agent-memory/ ← [SUBDIR_COUNT] persistent memory directories (one per agent[EXCLUSION_NOTE])
  • [EXCLUSION_NOTE] = , excluded: [list agents without a subdirectory] if any are missing, otherwise omit
• If agent-memory/ does not exist: omit that line entirely