feat: migrate MCP dependencies to CLI-first + Skills architecture#51
feat: migrate MCP dependencies to CLI-first + Skills architecture#51
Conversation
Replace MCP server dependencies with CLI tools and Skills to reduce context window consumption (MCP injects 40-50% before any useful action). New files: - tools/cli-registry.json: Central registry of 24 CLI tools - tools/skills-registry.json: Registry of 12 custom + 4 community skills - tools/check-tools.sh: CLI & Skills diagnostic script - schemas/cli-registry.schema.json + skills-registry.schema.json - agents/_shared/cli-tools-protocol.md: Shared protocol (CLI > MCP) - skills/: 12 SKILL.md files (vercel, github, neon, docker, spip, asc, figma, shadcn, pgcli, pandoc, productivity, find-skill) Modified: - install.sh: Skills installation, --check-tools, --with-community-skills - CLAUDE.md: CLI Tools + Skills sections - Agents (bruce, blackemperor, robocop, brigitte, 2b3, steve, deploy:vercel, analyze:spip): Added cli-tools-protocol reference and CLI tool sections - site/data/commands.json: Added "tools" category https://claude.ai/code/session_01F1JWMm3AdYUEAqTrNWKXLu
- site/index.html: New "CLI & Skills" section with CLI vs MCP comparison, recommended tools grid, and skills install commands. Nav link added. - README.md: CLI Tools & Skills section with rules, recommended CLIs, skills list, and diagnostic commands. Version bumped to 3.5.0. - CHANGELOG.md: v3.5.0 entry documenting full CLI-first migration. - uninstall.sh: Now detects and removes ulk-* skills from ~/.claude/skills/ - install-remote.sh: Now installs skills from skills/ directory - diagnostic-skills.sh: New section 9 verifying skills installation - site/data/commands.json: Version bumped to 3.5.0, tools entries in agentFileMap https://claude.ai/code/session_01F1JWMm3AdYUEAqTrNWKXLu
… 6, App Store deploy - Model: sonnet → opus (matches peer orchestrators bruce/blackemperor) - Phase 0: auto-diagnostic, resume detection, tool verification, context-protocol - Persistent memory: .claude/agents/steve.md subagent with memory: local - Swift 6: strict concurrency (actors, Sendable, typed throws, @mainactor) - Privacy Manifest: PrivacyInfo.xcprivacy template - Modern extensions: Widgets, Live Activities, App Intents, App Clips - Phase 8: App Store deployment via asc CLI (signing, TestFlight, submission) - Roadmap: T-shirt effort estimates (XS/S/M/L/XL) on all tasks - Updated CLAUDE.md: 5→6 subagents, steve in memory table - Updated site description: reflects full Apple ecosystem scope https://claude.ai/code/session_01F1JWMm3AdYUEAqTrNWKXLu
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (36)
📝 WalkthroughWalkthroughThis PR introduces a CLI-first architecture migration from MCP-centric tooling, adds a persistent-memory "steve" agent, establishes a custom skills system (ulk-*), creates tool/skills registries with diagnostic infrastructure, and bumps the version to 3.5.0 with expanded documentation across multiple agent definitions and installation workflows. Changes
Sequence Diagram(s)sequenceDiagram
participant Agent as Steve Agent
participant FileSystem as ~/.claude/agent-memory-local/steve/
participant Context as Project Context
participant SubAgent as Sub-Agent Executor
Agent->>FileSystem: Phase 0: Read MEMORY.md (if exists)
FileSystem-->>Agent: Return steve_project_state + preferences
Agent->>Agent: Detect RESUME or scan docs/api/, docs/apple-starter-kit/
Agent->>Agent: Verify CLI tools (asc, mobicon, xcrun)
Agent->>Context: Build CONTEXTE PROJET block
Agent->>SubAgent: Pass context to 9-phase workflow
SubAgent->>SubAgent: Execute Phases 1-7 (Audit→Deploy)
SubAgent->>SubAgent: Optional Phase 8: App Store deployment via asc
SubAgent-->>Agent: Return results + state updates
Agent->>FileSystem: Update MEMORY.md with project_state + preferences
FileSystem-->>Agent: Persist memory for next session
sequenceDiagram
participant Agent as Agent
participant CheckCmd as command -v
participant CLITool as CLI Tool
participant MCPConfig as /mcp (MCP Config)
participant User as User Notification
Agent->>CheckCmd: Verify tool presence (e.g., 'gh', 'vercel')
alt Tool Available
CheckCmd-->>Agent: ✓ Found
Agent->>CLITool: Use CLI (prioritized, low context)
else Tool Not Found
CheckCmd-->>Agent: ✗ Not found
Agent->>MCPConfig: Check for configured MCP fallback
alt MCP Available
MCPConfig-->>Agent: ✓ MCP configured
Agent->>Agent: Use MCP (fallback mode)
else No MCP
MCPConfig-->>Agent: ✗ No MCP config
Agent->>User: Notify missing tool & propose installation
end
end
Estimated Code Review Effort🎯 4 (Complex) | ⏱️ ~60 minutes This PR involves heterogeneous changes across multiple domains: 12 new skill documentation files, major agent logic rewrites (steve with 9 phases + Swift 6), CLI-first protocol references in 8 agents, complex shell scripts (check-tools.sh with TUI logic and jq-based registry parsing, install.sh with skills installation and branching), JSON registry structures, installer/diagnostic enhancements, and site UI updates. While many skill files follow repetitive patterns, the logic density in diagnostic scripts, installer modifications, and steve's persistent memory control flow demand careful review across distinct codebases. Possibly Related PRs
Poem
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Replace MCP server dependencies with CLI tools and Skills to reduce
context window consumption (MCP injects 40-50% before any useful action).
New files:
figma, shadcn, pgcli, pandoc, productivity, find-skill)
Modified:
analyze:spip): Added cli-tools-protocol reference and CLI tool sections
https://claude.ai/code/session_01F1JWMm3AdYUEAqTrNWKXLu
Summary by CodeRabbit
New Features
~/.claude/skills.Documentation