Skip to content

chief-of-staff.md

Latest commit

 

History

History
121 lines (96 loc) · 4.5 KB

chief-of-staff.md

File metadata and controls

121 lines (96 loc) · 4.5 KB

Chief of Staff

Autonomous agent manager that watches task files, spawns sub-agents, and maintains system health.

Architecture

  • Task Parser (server/lib/taskParser.js): Parses TASKS.md and COS-TASKS.md formats
  • CoS Service (server/services/cos.js): State management, health monitoring, task evaluation
  • Task Watcher (server/services/taskWatcher.js): File watching with chokidar
  • Sub-Agent Spawner (server/services/subAgentSpawner.js): Claude CLI execution with MCP
  • CoS Routes (server/routes/cos.js): REST API endpoints
  • CoS UI (client/src/pages/ChiefOfStaff.jsx): Tasks, Agents, Health, Config tabs

Features

  1. Dual Task Lists: User tasks (TASKS.md) and system tasks (COS-TASKS.md)
  2. Autonomous Execution: Auto-approved tasks run without user intervention
  3. Approval Workflow: Tasks marked APPROVAL require user confirmation
  4. System Health Monitoring: PM2 process checks, memory usage, error detection
  5. Sub-Agent Spawning: Claude CLI with --dangerously-skip-permissions and MCP servers
  6. Self-Improvement: Can analyze performance and suggest prompt/config improvements
  7. Script Generation: Creates automation scripts for repetitive tasks
  8. Report Generation: Daily summaries of completed work

Task File Format

# Tasks
## Pending
- [ ] #task-001 | HIGH | Task description
  - Context: Additional context
  - App: app-name

## In Progress
- [~] #task-002 | MEDIUM | Another task
  - Agent: agent-id
  - Started: 2024-01-15T10:30:00Z

## Completed
- [x] #task-003 | LOW | Done task
  - Completed: 2024-01-14T15:45:00Z

System Task Format

- [ ] #sys-001 | HIGH | AUTO | Auto-approved task
- [ ] #sys-002 | MEDIUM | APPROVAL | Needs user approval

Data Storage

./data/cos/
├── state.json           # Daemon state and config
├── agents/{agentId}/    # Agent prompts and outputs
├── reports/{date}.json  # Daily reports
└── scripts/             # Generated automation scripts

Model Selection Rules

The selectModelForTask function routes tasks to appropriate model tiers:

Tier Trigger Example Tasks
heavy Critical priority, visual analysis, complex reasoning Architect, refactor, security audit, long context
medium Standard development tasks, default Most coding tasks, bug fixes, feature implementation
light Documentation-only tasks Update README, write docs, format text

Important: Light model (haiku) is NEVER used for coding tasks. Tasks containing keywords like fix, bug, implement, test, feature, api, component, etc. are automatically routed to medium tier or higher.

Configuration

Setting Default Description
evaluationIntervalMs 60000 Task evaluation interval (1 minute)
healthCheckIntervalMs 900000 Health check interval (15 minutes)
maxConcurrentAgents 3 Max parallel agents (global)
maxConcurrentAgentsPerProject 2 Max parallel agents per project
maxProcessMemoryMb 2048 Memory alert threshold
autoStart false Start on server boot
selfImprovementEnabled true Allow self-analysis
proactiveMode true Always find work when idle
comprehensiveAppImprovement true Apply full analysis to managed apps
avatarStyle svg CoS UI avatar style: svg, cyber, sigil, or ascii

API Endpoints

Route Description
GET /api/cos Get CoS status
POST /api/cos/start Start daemon
POST /api/cos/stop Stop daemon
GET/PUT /api/cos/config Configuration
GET /api/cos/tasks Get all tasks
POST /api/cos/evaluate Force evaluation
GET /api/cos/health Health status
POST /api/cos/health/check Run health check
GET /api/cos/agents List agents
POST /api/cos/agents/:id/terminate Terminate agent
GET /api/cos/reports List reports
GET /api/cos/learning Get learning insights
GET /api/cos/digest Get weekly digest

Prompt Templates

Template Purpose
cos-agent-briefing Brief sub-agent on task
cos-evaluate Evaluate tasks and decide actions
cos-report-summary Generate daily summary
cos-self-improvement Analyze and suggest improvements

Related Features