home / skills / robdtaylor / personal-ai-infrastructure / agentobservability
/skills/AgentObservability
This skill provides real-time multi-agent observability by visualizing swim lanes, events, and tool usage in a live dashboard.
npx playbooks add skill robdtaylor/personal-ai-infrastructure --skill agentobservabilityReview the files below or copy the command above to add this skill to your agents.
---
name: AgentObservability
description: |
Real-time observability dashboard for multi-agent Claude Code sessions.
Visualize agent interactions, tool usage, and session flows in real-time through a web dashboard. Track multiple agents running in parallel with swim lane visualization, event filtering, and live charts.
USE WHEN user says 'start observability', 'agent dashboard', 'monitor agents', 'watch agent activity', 'multi-agent monitoring', 'track subagents', or needs to debug multi-agent workflows.
**Key Features:**
- π΄ Real-time event streaming via WebSocket
- π Agent swim lanes showing parallel execution
- π Event filtering by agent, session, event type
- π Live charts for tool usage patterns
- πΎ Filesystem-based (no database required)
**Inspired by [@indydevdan](https://github.com/indydevdan)**'s work on multi-agent observability.
**Our approach:** Filesystem + in-memory streaming vs. indydevdan's SQLite database approach.
---
# Agent Observability Skill
## When to Activate This Skill
- User wants to monitor agent activity
- Need to debug multi-agent workflows
- Track parallel subagent execution
- Visualize tool usage patterns
- Analyze session flows in real-time
## Prerequisites
- Bun runtime installed
- Claude Code with hooks configured
- PAI_DIR environment variable set
## Installation
See [SETUP.md](./SETUP.md) for complete installation instructions.
**Quick Setup:**
```bash
# 1. Set environment variable
export PAI_DIR="$HOME/.claude" # Add to ~/.zshrc or ~/.bashrc
# 2. Configure hooks (merge into ~/.claude/settings.json)
cat settings.json.example
# 3. Create directory structure
mkdir -p ~/.claude/history/raw-outputs
# 4. Install dependencies
cd apps/server && bun install
cd ../client && bun install
```
## Usage
### Start the Observability Dashboard
**Terminal 1 - Server:**
```bash
cd ~/Projects/PAI/skills/agent-observability/apps/server
bun run dev
```
**Terminal 2 - Client:**
```bash
cd ~/Projects/PAI/skills/agent-observability/apps/client
bun run dev
```
**Open browser:** http://localhost:5173
### Using Claude Code
Once the dashboard is running, any Claude Code activity will appear in real-time:
1. Open Claude Code
2. Use any tool (Read, Write, Bash, etc.)
3. Launch subagents with Task tool
4. Watch events appear in the dashboard
### Event Types Captured
- **SessionStart** - New Claude Code session begins
- **UserPromptSubmit** - User sends a message
- **PreToolUse** - Before a tool is executed
- **PostToolUse** - After a tool completes
- **Stop** - Main agent task completes
- **SubagentStop** - Subagent task completes
- **SessionEnd** - Session ends
## Features
### Real-Time Visualization
- **Agent Swim Lanes:** See multiple agents (kai, designer, engineer, etc.) running in parallel
- **Event Timeline:** Chronological view of all events
- **Tool Usage Charts:** Visualize which tools are being used most
- **Session Tracking:** Track individual sessions and their lifecycles
### Filtering & Search
- Filter by agent name (kai, designer, engineer, pentester, etc.)
- Filter by event type (PreToolUse, PostToolUse, etc.)
- Filter by session ID
- Search event payloads
### Data Storage
Events are stored in JSONL (JSON Lines) format:
```
~/.claude/history/raw-outputs/YYYY-MM/YYYY-MM-DD_all-events.jsonl
```
Each line is a complete JSON object:
```jsonl
{"source_app":"kai","session_id":"abc123","hook_event_type":"PreToolUse","payload":{...},"timestamp":1234567890,"timestamp_pst":"2025-01-28 14:30:00 PST"}
```
### In-Memory Streaming
- Server keeps last 1000 events in memory
- Low memory footprint
- Fast real-time updates via WebSocket
- No database overhead
## Architecture
```
βββββββββββββββββββ
β Claude Code β Executes hooks on events
β (with hooks) β
ββββββββββ¬βββββββββ
β
βΌ
βββββββββββββββββββ
β capture-all- β Appends events to JSONL
β events.ts hook β
ββββββββββ¬βββββββββ
β
βΌ
βββββββββββββββββββββββββββββββββββββββ
β ~/.claude/history/raw-outputs/ β Daily JSONL files
β 2025-01/2025-01-28_all-events.jsonl β
ββββββββββ¬βββββββββββββββββββββββββββββ
β
βΌ
βββββββββββββββββββ
β file-ingest.ts β Watches files, streams to memory
β (Bun server) β
ββββββββββ¬βββββββββ
β
βΌ
βββββββββββββββββββ
β Vue 3 Client β Real-time dashboard visualization
β (Vite + Tail) β
βββββββββββββββββββ
```
## Configuration
### Environment Variables
**PAI_DIR:**
Path to your PAI directory (defaults to `~/.claude/`)
```bash
export PAI_DIR="/Users/yourname/.claude"
```
### Hooks Configuration
Add to `~/.claude/settings.json` (see `settings.json.example` for full template):
```json
{
"hooks": {
"PreToolUse": [{
"matcher": "*",
"hooks": [{
"type": "command",
"command": "${PAI_DIR}/skills/agent-observability/hooks/capture-all-events.ts --event-type PreToolUse"
}]
}],
// ... other hooks
}
}
```
## Troubleshooting
### No events appearing
1. Check PAI_DIR is set: `echo $PAI_DIR`
2. Verify directory exists: `ls ~/.claude/history/raw-outputs/`
3. Check hook is executable: `ls -l hooks/capture-all-events.ts`
4. Look for today's events file: `ls ~/.claude/history/raw-outputs/$(date +%Y-%m)/`
### Server won't start
1. Check Bun is installed: `bun --version`
2. Verify dependencies: `cd apps/server && bun install`
3. Check port 3001 isn't in use: `lsof -i :3001`
### Client won't connect
1. Ensure server is running first
2. Check WebSocket connection in browser console
3. Verify no firewall blocking localhost:3001
## Credits
**Inspired by [@indydevdan](https://github.com/indydevdan)**'s pioneering work on multi-agent observability for Claude Code.
**Our implementation differs** by using filesystem-based event capture and in-memory streaming instead of SQLite database persistence. Both approaches have their merits! Check out indydevdan's work for a database-backed solution with full historical persistence.
## Development
### Running in Development
```bash
# Server (hot reload)
cd apps/server
bun --watch src/index.ts
# Client (Vite dev server)
cd apps/client
bun run dev
```
### Building for Production
```bash
# Client build
cd apps/client
bun run build
bun run preview
```
### Adding New Event Types
1. Update `capture-all-events.ts` hook if needed
2. Add hook configuration to `settings.json`
3. Client will automatically display new event types
## Documentation
- [README.md](./README.md) - Complete documentation
- [SETUP.md](./SETUP.md) - Installation guide
- [history-structure/](./history-structure/) - Data storage structure
- [settings.json.example](./settings.json.example) - Hook configuration template
## License
Part of the [PAI (Personal AI Infrastructure)](https://github.com/danielmiessler/PAI) project.
## Contributing
Contributions welcome! Areas for improvement:
- Historical data persistence options
- Export functionality (CSV, JSON)
- Alert/notification system
- Advanced filtering and search
- Session replay capability
- Integration with other PAI skills
This skill provides a real-time observability dashboard for multi-agent Claude Code sessions. It visualizes agent interactions, tool usage, and session flows in a web UI so you can monitor parallel agents and debug workflows quickly. The system uses filesystem-based JSONL event capture and in-memory streaming for low overhead and fast updates.
Claude Code hooks append structured events to daily JSONL files under your PAI directory. A Bun-based server watches those files, streams the latest events over WebSocket, and retains a recent event window in memory. A Vue 3 client renders swim lanes, timelines, filters, and live charts showing tool usage and session lifecycles in real time.
Do I need a database?
No. Events are appended to JSONL files and the server keeps a recent window in memory; no database is required.
What runtime and prerequisites are required?
Bun is required for server and client scripts. Claude Code must have hooks configured and PAI_DIR must point to your ~/.claude directory.
How much history is stored?
Daily JSONL files store full history on disk; the server retains the last ~1000 events in memory for real-time streaming.