## Skills Library - Browse and install 36+ community skills from Anthropic & Superpowers - Local skill templates (Rails, React, testing) - Create custom project-specific skills - Beautiful UI with search and filtering - /claude-os-skills command for CLI management ## Session Insights - Parse Claude Code native .jsonl session files - Extract patterns, decisions, blockers automatically - Track tool usage and file changes - Session summary generation for LLM processing ## Documentation - Updated README with v2.3 features - Updated API_REFERENCE with new endpoints - Added SKILLS_GUIDE.md and SESSIONS_GUIDE.md - Updated GitHub Pages (docs/index.html) - Reorganized docs (internal_docs for private) ## Bug Fixes - Fixed httpx client closure in community skill install - Fixed frontend request format for skill install 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
11 KiB
11 KiB
Session Start Protocol
What Claude Should Do AUTOMATICALLY When You Start ANY Conversation
🎯 The Problem
When you start Claude Code CLI:
- You're in a specific project directory
- You may have been working on something
- There's context from previous sessions
- You have knowledge in Claude OS
But Claude starts with ZERO context.
✅ The Solution: Automatic Session Start
Every time a conversation starts, Claude should:
Phase 1: Detect Project (5 seconds)
# Check for Claude OS project
pwd
ls -la | grep -E "(CLAUDE.md|.claude-os|agent-os)"
If .claude-os/config.json exists:
- ✅ This is a Claude OS project
- → Proceed with full session start
If not:
- ℹ️ Generic mode (no project context)
- → Offer to initialize with
/claude-os-init
Phase 2: Read Project State (10 seconds)
Files to read:
CLAUDE.md- Project overview (always loaded first).claude-os/config.json- Project configurationclaude-os-state.json- Session state (if exists)agent-os/config.yml- Agent-OS config (if using specs)
Example:
Reading CLAUDE.md...
Reading .claude-os/config.json...
Reading claude-os-state.json...
✓ Project: Pistn
✓ MCPs: 4 configured
✓ Last session: Oct 29, 2025 (2 days ago)
✓ Agent-OS: Enabled (3 active specs)
Phase 3: Check Git Context (5 seconds)
git branch --show-current
git status --short
git log -3 --oneline
git remote get-url origin
Extract:
- Current branch name
- Uncommitted changes count
- Recent commits
- Whether we're ahead/behind remote
Example:
Branch: feature/appointment-redesign
Status: 3 files modified, 1 untracked
Recent commits:
- abc1234 Add sidebar navigation
- def5678 Convert panels to cards
- ghi9012 Update appointment form layout
Phase 4: Search Project Memories (15 seconds)
Search query based on:
- Last session task (from
claude-os-state.json) - Current branch name
- Recent commit messages
+recentmodifier
Example query:
"appointment dashboard redesign recent"
Search:
{project}-project_memoriesknowledge base- Return top 5 most relevant memories
- Show me titles and key insights
Example:
📚 Found 5 relevant memories:
1. Appointment Dashboard Redesign Plan (Oct 28)
→ 67-page implementation plan, 5 tabs documented
2. Current Dashboard Analysis (Oct 28)
→ All features catalogued, zero functionality loss
3. Bootstrap to Modern Cards Pattern (Oct 25)
→ Reusable conversion pattern
4. Sidebar Navigation Pattern (Oct 22)
→ localStorage persistence, responsive
5. Tekmetric Integration Notes (Oct 20)
→ Must preserve API sync on save
Phase 5: Load Coding Standards (10 seconds)
Search {project}-project_profile:
- Query: "coding standards architecture conventions"
- Load key architectural decisions
- Load tech stack preferences
Example:
📋 Coding Standards:
• Service objects for business logic
• Decorator pattern with Draper
• Fragment caching with Redis
• Bootstrap 3.x styling (no Bootstrap 4/5)
• RSpec for testing
• Concerns for shared controller logic
Phase 6: Check Kanban Board (5 seconds)
If Agent-OS is enabled:
GET /api/projects/{id}/kanban
Show:
- Total specs
- Active specs (not completed/archived)
- Current tasks in progress
- Next tasks todo
Example:
📊 Kanban Board:
• 3 specs total (52 tasks)
• 1 spec in progress: Group Account Rendering (45% complete)
• 2 tasks in progress:
- PHASE2-TASK3: Implement concern methods
- PHASE2-TASK4: Add helper methods
• 29 tasks remaining
Phase 7: Present Session Start Summary (Display to User)
Format:
═══════════════════════════════════════════════════════════════
🚀 WELCOME BACK!
═══════════════════════════════════════════════════════════════
📁 Project: Pistn (Ruby on Rails)
🌿 Branch: feature/appointment-redesign
📅 Last Session: 2 days ago (Oct 29, 2025)
⏱️ Duration: 2h 15m
═══════════════════════════════════════════════════════════════
📚 CONTEXT LOADED
═══════════════════════════════════════════════════════════════
✓ 5 relevant memories found
✓ Coding standards loaded
✓ 3 architectural patterns available
✓ Kanban board synced (45% complete)
🎯 LAST TASK: Redesign Appointment Dashboard
We completed:
✓ Sidebar navigation component
✓ Bootstrap panels → modern cards
✓ iOS-style toggle switches
Still working on:
⏳ Phase 2: Implement concern methods
⏳ Phase 2: Add helper methods
═══════════════════════════════════════════════════════════════
💡 KEY INSIGHTS
═══════════════════════════════════════════════════════════════
• 67-page implementation plan available
• Zero functionality loss requirement
• Tekmetric integration must be preserved
• Bootstrap 3.x only (no v4/v5)
═══════════════════════════════════════════════════════════════
🔄 AVAILABLE PATTERNS
═══════════════════════════════════════════════════════════════
• Sidebar navigation (from user-auth work)
• Card-based layouts (from reports redesign)
• Modern toggle switches (already in app)
• Service object pattern (business logic)
• Draper decorators (view logic)
═══════════════════════════════════════════════════════════════
⚠️ STATUS
═══════════════════════════════════════════════════════════════
Git Status:
• 3 files modified
• 1 file untracked
• ⚠️ Uncommitted changes (from last session)
Blockers:
• None! ✓
═══════════════════════════════════════════════════════════════
🎯 SUGGESTED NEXT STEPS
═══════════════════════════════════════════════════════════════
Option 1: Continue Phase 2 Implementation
→ Implement concern methods for PHASE2-TASK3
→ Should take ~30 minutes based on estimate
Option 2: Commit Previous Work
→ 3 files modified need to be committed
→ Clean git state before starting new work
Option 3: Start Something New
→ Tell me what you'd like to work on
═══════════════════════════════════════════════════════════════
What would you like to do?
🎛️ Configuration
Auto-start can be configured in .claude-os/config.json:
{
"session_management": {
"auto_start": true,
"auto_search_memories": true,
"max_memories_to_load": 5,
"search_days_back": 14,
"show_git_status": true,
"show_kanban_status": true,
"proactive_suggestions": true
}
}
🚫 When NOT to Auto-Start
Skip session start if:
- User message starts with a direct question (e.g., "What is...")
- User is clearly asking for help/docs (e.g., "How do I...")
- User explicitly says "ignore context" or "fresh start"
- No
.claude-os/directory exists
In these cases:
- Answer the question directly
- Offer to initialize Claude OS after answering
📊 Session Tracking During Conversation
Throughout the session, Claude should:
Track Key Events
- When user starts implementing something
- When they hit a blocker
- When they discover a pattern
- When they make an architectural decision
- When they solve a complex problem
Proactive Memory References
You: "I need to add caching to this endpoint"
Claude: "FYI, we have a memory about fragment caching with Redis
from the reports redesign. Want me to use that pattern?"
Auto-Save Suggestions
You: "Finally got the Tekmetric sync working with the new structure!"
Claude: "That sounds like a high-value solution. Would you like me
to save this to project memories for future reference?"
💾 Session End Protocol
When conversation ends (or user says "done"):
Option 1: User Runs /claude-os-session end
- Follow the detailed end protocol from claude-os-session.md
- Analyze work, suggest saves, update statistics
Option 2: Conversation Ends Naturally
- Claude should still provide a brief summary:
═══════════════════════════════════════════════════════════════
📊 SESSION COMPLETE
═══════════════════════════════════════════════════════════════
Duration: ~45 minutes
Work: Implemented concern methods for PHASE2-TASK3
Quick saves available:
• Concern method pattern for group accounts
• Service object integration approach
Run /claude-os-session end for full summary and saves
Or just continue working next time - I'll remember!
═══════════════════════════════════════════════════════════════
🎯 The Goal
Every session should feel like:
- Claude already knows everything
- Claude remembers what we were doing
- Claude proactively brings up relevant context
- Claude suggests next steps based on history
- Zero context loss between sessions
NOT like:
- Starting from scratch
- Repeating ourselves
- Explaining the same architecture again
- Searching for what we did before
🔄 Implementation Checklist
For Claude to do this automatically:
- Check for
.claude-os/directory at conversation start - Read
CLAUDE.mdif present - Read
.claude-os/config.jsonif present - Read
claude-os-state.jsonif present - Run git status and git log
- Search project_memories with intelligent query
- Load coding standards from project_profile
- Check Kanban board status (if Agent-OS enabled)
- Present comprehensive session start summary
- Track key events during session
- Provide session end summary
This protocol ensures you NEVER start a conversation from zero context again! 🚀