claude-engineering-plugin/plans/grow-your-own-garden-plugin-architecture.md

Grow Your Own Garden: Adaptive Agent Ecosystem

Issue: https://github.com/EveryInc/compound-engineering-plugin/issues/20

The Idea

Everyone grows their own garden, but we're all using the same process.

Start from a seed (minimal core: /plan, /work, /review, /compound). Each /compound loop can suggest adding agents based on what you're working on—like building up a test suite to prevent regressions, but for code review expertise.

Current Problem

• Monolithic plugin: 24 agents, users use ~30%
• No personalization (same agents for Rails dev and Python dev)
• Static collection that doesn't adapt

Proposed Solution

The Seed (Core Plugin)

4 commands + minimal agents:

Component	What's Included
Commands	/plan, /work, /review, /compound
Review Agents	security, performance, simplicity, architecture, patterns
Research Agents	best-practices, framework-docs, git-history, repo-analyst
Skills	compound-docs, file-todos, git-worktree
MCP Servers	playwright, context7

The Growth Loop

After each /compound:

✅ Learning documented

💡 It looks like you're using Rails.
   Would you like to add the "DHH Rails Reviewer"?

   [y] Yes  [n] No  [x] Never ask

Three sources of new agents:

1. Predefined - "You're using Rails, add DHH reviewer?"
2. Dynamic - "You're using actor model, create an expert?"
3. Custom - "Want to create an agent for this pattern?"

Agent Storage

.claude/agents/       → Project-specific (highest priority)
~/.claude/agents/     → User's garden
plugin/agents/        → From installed plugins

Implementation Phases

Phase 1: Split the Plugin

• Create agent-library/ with framework-specific agents (Rails, Python, TypeScript, Frontend)
• Keep compound-engineering as core with universal agents
• No breaking changes—existing users unaffected

Phase 2: Agent Discovery

• /review discovers agents from all three locations
• Project agents override user agents override plugin agents

Phase 3: Growth via /compound

• Detect tech stack (Gemfile, package.json, etc.)
• Suggest relevant agents after documenting learnings
• Install accepted agents to ~/.claude/agents/

Phase 4: Management

• /agents list - See your garden
• /agents add <name> - Add from library
• /agents disable <name> - Temporarily disable

What Goes Where

Core (seed): 11 framework-agnostic agents

• security-sentinel, performance-oracle, code-simplicity-reviewer
• architecture-strategist, pattern-recognition-specialist
• 4 research agents, 2 workflow agents

Agent Library: 10 specialized agents

• Rails: kieran-rails, dhh-rails, data-integrity (3)
• Python: kieran-python (1)
• TypeScript: kieran-typescript (1)
• Frontend: julik-races, design-iterator, design-reviewer, figma-sync (4)
• Editorial: every-style-editor (1)

Key Constraint

Claude Code doesn't support plugin dependencies. Each plugin must be independent. Users manually install what they need, or we suggest additions via /compound.

Acceptance Criteria

• Core plugin works standalone with universal agents
• /compound suggests agents based on detected tech stack
• Users can accept/decline suggestions
• /agents command for garden management
• No breaking changes for existing users