docs: regenerate hierarchical AGENTS.md with deep investigation

- Root: 181 lines with agent models, complexity hotspots, CI pipeline
- Hooks: 31 lifecycle hooks, execution order, patterns
- Tools: 20+ tools, LSP/AST-Grep specifics, registration
- Features: Background agents, Claude Code compat, skill MCP
- Agents: 10 agents with models, tool restrictions
- Shared: 43 utilities with usage patterns
- CLI: Commander.js entry, doctor checks, TUI framework

Generated via /init-deep with 12 parallel explore agents

src/features/AGENTS.md

@@ -1,39 +1,62 @@
 # FEATURES KNOWLEDGE BASE
 
 ## OVERVIEW
-Claude Code compatibility layer + core feature modules. Commands, skills, agents, MCPs, hooks from Claude Code work seamlessly.
+
+Core feature modules + Claude Code compatibility layer. Background agents, skill MCP, builtin skills/commands, and 5 loaders for Claude Code compat.
 
 ## STRUCTURE
+
 ```
 features/
-├── background-agent/           # Task lifecycle, notifications (1165 lines manager.ts)
-├── boulder-state/              # Boulder/Todo state persistence
-├── builtin-commands/           # Built-in slash commands (ralph-loop, refactor, init-deep)
-├── builtin-skills/             # Built-in skills (1203 lines skills.ts)
-│   ├── git-master/             # Atomic commits, history search
-│   ├── playwright/             # Browser automation
-│   └── frontend-ui-ux/         # Designer-developer skill
+├── background-agent/           # Task lifecycle (1165 lines manager.ts)
+│   ├── manager.ts              # Launch → poll → complete orchestration
+│   ├── concurrency.ts          # Per-provider/model limits
+│   └── types.ts                # BackgroundTask, LaunchInput
+├── skill-mcp-manager/          # MCP client lifecycle
+│   ├── manager.ts              # Lazy loading, idle cleanup
+│   └── types.ts                # SkillMcpConfig, transports
+├── builtin-skills/             # Playwright, git-master, frontend-ui-ux
+│   └── skills.ts               # 1203 lines of skill definitions
+├── builtin-commands/           # ralph-loop, refactor, init-deep
+│   └── templates/              # Command implementations
 ├── claude-code-agent-loader/   # ~/.claude/agents/*.md
 ├── claude-code-command-loader/ # ~/.claude/commands/*.md
-├── claude-code-mcp-loader/     # .mcp.json files with ${VAR} expansion
+├── claude-code-mcp-loader/     # .mcp.json with ${VAR} expansion
 ├── claude-code-plugin-loader/  # installed_plugins.json
 ├── claude-code-session-state/  # Session state persistence
-├── context-injector/           # AGENTS.md/README.md/Rules injection
-├── opencode-skill-loader/      # Skills from OpenCode + Claude paths
-├── skill-mcp-manager/          # MCP servers in skill YAML (stdio/http transports)
-├── task-toast-manager/         # Task status toast notifications
-└── hook-message-injector/      # Message injection into conversation streams
+├── opencode-skill-loader/      # Skills from 6 directories
+├── context-injector/           # AGENTS.md/README.md injection
+├── boulder-state/              # Todo state persistence
+├── task-toast-manager/         # Toast notifications
+└── hook-message-injector/      # Message injection
 ```
 
 ## LOADER PRIORITY
-| Loader | Priority (highest first) |
-|--------|--------------------------|
+
+| Type | Priority (highest first) |
+|------|--------------------------|
 | Commands | `.opencode/command/` > `~/.config/opencode/command/` > `.claude/commands/` > `~/.claude/commands/` |
 | Skills | `.opencode/skill/` > `~/.config/opencode/skill/` > `.claude/skills/` > `~/.claude/skills/` |
 | Agents | `.claude/agents/` > `~/.claude/agents/` |
 | MCPs | `.claude/.mcp.json` > `.mcp.json` > `~/.claude/.mcp.json` |
 
+## BACKGROUND AGENT
+
+- **Lifecycle**: `launch` → `poll` (2s interval) → `complete`
+- **Stability**: 3 consecutive polls with same message count = idle
+- **Concurrency**: Per-provider/model limits (e.g., max 3 Opus, max 10 Gemini)
+- **Notification**: Batched system reminders to parent session
+- **Cleanup**: 30m TTL, 3m stale timeout, signal handlers
+
+## SKILL MCP
+
+- **Lazy**: Clients created on first tool call
+- **Transports**: stdio (local process), http (SSE/Streamable)
+- **Environment**: `${VAR}` expansion in config
+- **Lifecycle**: 5m idle cleanup, session-scoped
+
 ## CONFIG TOGGLES
+
 ```jsonc
 {
   "claude_code": {
@@ -46,20 +69,9 @@ features/
 }
 ```
 
-## BACKGROUND AGENT
-- **Lifecycle**: `launch` → `poll` (idle/stability detection) → `complete`.
-- **Concurrency**: Per-provider/model limits in `concurrency.ts`.
-- **Notification**: Auto-injects system reminders into parent session on task completion.
-- **Cleanup**: Shutdown handler cancels pending waiters; idle tasks pruning (30m TTL).
-
-## SKILL MCP
-- **Lazy Loading**: Clients connect on first tool call via `SkillMcpManager`.
-- **Transports**: `stdio` (local process) or `http` (SSE/Streamable HTTP).
-- **Environment**: `${VAR}` expansion in config via `env-expander.ts`.
-- **Lifecycle**: Session-scoped clients; auto-cleanup after 5m idle.
-
 ## ANTI-PATTERNS
-- **Sequential Delegation**: Calling agents one-by-one; use `delegate_task` for parallel runs.
-- **Self-Report Trust**: Trusting agent's "I'm done" without verifying against session state.
-- **Main Thread Blocks**: Heavy I/O or long-running logic during loader initialization.
-- **Manual Versioning**: Updating `package.json` version field; managed exclusively by CI.
+
+- **Sequential delegation**: Use `delegate_task` for parallel
+- **Trust self-reports**: ALWAYS verify agent outputs
+- **Main thread blocks**: No heavy I/O in loader init
+- **Manual versioning**: CI manages package.json version