Search

PythonReactSEOPDF
GitHub Skill

orchestrator

Multi-Agent Orchestration skill for coordinating AI agents across channels, managing sessions, distributing tasks, and leveraging distributed compute resources.

whodaniel☆ 2 Repository stars0 Repository forks

Installation

Follow the repository instructions for this skill and your agent. Review the files and dependencies before adding it to your workflow.

View source ↗Guides

Skill instructions

Orchestrator Skill

Purpose

Enable AI agents to act as Orchestrators - coordinating multiple AI agents across federated communication channels, managing sessions, distributing tasks, and aggregating results.

When to Use This Skill

Activate when you need to:

  • Coordinate multiple AI agents for collaborative work
  • Distribute tasks across parallel compute resources
  • Manage multi-agent sessions with ID assignment
  • Aggregate results from multiple sources
  • Leverage free/low-cost AI compute (Gemini tabs, Jules CLI, etc.)

Quick Start

1. Start the Relay Server
pnpm relay:start
2. Launch Persistent Orchestrator
node orchestrator-persistent.js Green  # Join Green channel
3. Session Control Commands
node session-control.js status   # Check agent status
node session-control.js pause    # Pause all agents
node session-control.js resume   # Resume operations
node session-control.js end      # End session

Protocol: DACC-v1

Distributed Agent Communication & Coordination Protocol v1

Core Principles
  1. Mandatory Signing - All agents MUST sign messages with [AGENT-XX]
  2. ID Assignment - Orchestrator assigns unique IDs to agents
  3. Structured Communication - Use formatted message templates
  4. Session Logging - All conversations logged to markdown files
Message Signing Format
[AGENT-XX] Your message content here

Acceptable alternatives:

  • ID#: AGENT-XX message
  • [GEMINI-3F-01] message (custom IDs)

Compute Resources

Available Sources (Prioritized by Cost)
Source Cost Parallelism Best For
Gemini Web Tabs FREE Multiple tabs Analysis, reasoning
Jules CLI FREE Up to 4 parallel Async coding tasks
HuggingChat FREE 1 Open source models
Perplexity FREE 1 Search + reasoning
Claude Free FREE Limited Complex reasoning
ChatGPT Free FREE Limited General tasks
Jules CLI Integration
# Check Jules status
jules version
jules remote list --session

# Submit task
jules new "Implement feature X in packages/core"

# Parallel execution
jules new --parallel 4 "Add tests to all services"

# Pull completed work
jules remote pull --session 123456 --apply
Chrome Extension (Gemini Tabs)
  1. Open multiple Gemini tabs
  2. Press Ctrl+Shift+F to open Fuse Connect
  3. Join the same channel in each tab
  4. Each tab becomes an independent agent

Session Management

Starting a Session
  1. Start Relay: pnpm relay:start
  2. Launch Orchestrator: node orchestrator-persistent.js [channel]
  3. Open Agent Tabs: Gemini tabs with Fuse extension
  4. Join Channel: All agents join same channel
Session Commands
Command Script Purpose
Pause node session-control.js pause Request all agents to pause
Resume node session-control.js resume Resume paused session
Status node session-control.js status Request status from all agents
End node session-control.js end End the current session
Session Logs

All sessions are logged to:

.agent/session-logs/YYYY-MM-DDTHH-MM-SS_Channel_persistent.md

Log contents:

  • Session metadata (ID, times, channel, protocol)
  • Participant table (assigned IDs, message counts, violations)
  • Full conversation log with timestamps
  • Signing violation tracking

Task Distribution

Task Template
═══════════════════════════════════════════════════════════════
🔧 TASK: [Task Name]
═══════════════════════════════════════════════════════════════
Task ID: [unique-id]
Priority: [high|medium|low]
Assigned To: [AGENT-XX or "all"]

📋 DESCRIPTION
[Clear description of what needs to be done]

📌 DELIVERABLES
1. [Deliverable 1]
2. [Deliverable 2]

⏱️ DEADLINE: [Time constraint if any]

🔔 Sign responses with [AGENT-XX]!
═══════════════════════════════════════════════════════════════
Parallel Task Distribution
ORCHESTRATOR assigns:
├── AGENT-01: Analyze packages/a2a-core
├── AGENT-02: Analyze packages/agent
├── JULES-01: Implement security fixes (async)
└── JULES-02: Add unit tests (async)

Scripts Reference

Script Purpose
orchestrator-persistent.js Persistent orchestrator with auto-reconnect
orchestrator-green-channel.js Single-session orchestrator
session-control.js Pause/resume/end session commands
send-improvement-task.js Send improvement tasks
scripts/list_channels.js List available channels

Self-Improvement Cycle

Daily Protocol
  1. Review - Check overnight agent conversations in session logs
  2. Identify - Find improvement candidates from agent analyses
  3. Prioritize - Rank by impact and effort
  4. Distribute - Assign tasks to available compute
  5. Implement - Agents execute tasks
  6. Verify - Check results and integrate
  7. Document - Update documentation and logs
Continuous Improvement File

Track improvement status in:

.agent/SELF_IMPROVEMENT_CYCLE.md

Integration with Other Skills

agent-discovery-protocol

Protocol specification for agent discovery and ID assignment.

jules-delegation

Delegate async coding tasks to Jules CLI.

chrome-devtools

Browser automation for testing and verification.


Troubleshooting

Agents Not Responding
  1. Check relay is running: pnpm relay:start
  2. Verify agents are on same channel
  3. Check orchestrator logs for errors
Signing Violations
  1. Orchestrator sends reminders automatically
  2. Check session log for violation count
  3. Send manual reminder: node session-control.js status
Connection Drops
  1. Persistent orchestrator auto-reconnects in 5 seconds
  2. Check relay server health: curl localhost:3001/health
  3. Agents should auto-reconnect on tab refresh

Issue Resolution Protocol

Orchestrators MUST use available tools to resolve issues autonomously.

When Issues Arise
  1. Identify the issue type
  2. Select appropriate tool(s)
  3. Execute the fix
  4. Verify the resolution
  5. Document what was done
Tool Selection Matrix
Issue Type Tools to Use Action
Build failure run_command, grep_search, view_file Find error, view code, fix
Test failure run_command, view_file, replace_file_content Run tests, view failures, fix
Agent not responding mcp_chrome-devtools_*, run_command Check browser, restart relay
Connection error run_command, read_terminal Check relay status, restart
Missing file write_to_file, run_command Create file, verify
Code error view_file, replace_file_content View, fix, verify
Documentation gap write_to_file, replace_file_content Update docs
Auto-Resolution Prompts

When encountering an error, use this mental framework:

I encountered: [ERROR DESCRIPTION]

Available tools that could help:
1. run_command - Execute shell commands to diagnose/fix
2. view_file - Inspect source code for issues
3. grep_search - Find related code or error patterns
4. replace_file_content - Fix code issues
5. write_to_file - Create missing files
6. browser_subagent - Interact with browser for testing
7. mcp_chrome-devtools_* - Direct browser control

I will:
1. Use [TOOL] to [ACTION]
2. Verify with [VERIFICATION METHOD]
3. Document the fix in session log
Common Issue Resolutions
Relay Not Running
# Check status
curl localhost:3001/health

# If down, restart
pnpm relay:start
Agent Not Signing Messages
Send reminder via session-control.js status
or broadcast direct reminder to channel
Build Errors
1. run_command: pnpm build 2>&1 | head -50
2. grep_search: Find the error file
3. view_file: Inspect the error location
4. replace_file_content: Fix the issue
5. run_command: Verify build passes
Jules Task Failed
# Check session status
jules remote list --session

# Get details
jules remote pull --session [ID]

# Retry with improved prompt
jules new "Fixed version of: [original task]"
Escalation

If tools cannot resolve the issue:

  1. Document the issue fully in session log
  2. Create a clear summary for human review
  3. Continue with other tasks that can proceed

Version History

  • v1.0.0 (2026-01-16)
    • Initial orchestrator skill
    • Persistent mode with auto-reconnect
    • Session control commands (pause/resume/end)
    • Compute resource documentation
    • DACC-v1 protocol integration

Examples

Read the usage examples in the source repository.

View source ↗

Source history

Indexed on
2026-01-20
Index updated
2026-01-26
Branch
main

Index and sync dates describe this directory. View the source repository for code history.

GitHub ↗

chat Comments (0)

chat_bubble_outline

No comments yet. Be the first to share your thoughts!