Signature Project

AI Team
Orchestration

A multi-agent orchestration framework that coordinates 4 AI models (Claude, Gemini, Codex, Kimi) as a unified development team via tmux — with a formal 12-state machine, GitHub integration, and automated code review.

Python tmux Claude Gemini Codex Kimi GitHub API

Screenshots

tmux session — 4 agents working
Claude + Codex — Parallel UX review session
GitHub Project board — PRs created
GitHub Issues — Sprint tasks with auto-labels
System architecture diagram
Architecture — Orchestrator, Builder, Reviewer with 12-state machine

Impact

4

AI Models

12

PRs in 2 Hours

12

State Machine

8

Role Templates

Key Features

Multi-Model Agent Team

Coordinates Claude (Architect/Reviewer), Kimi (Deep thinking), Codex (Fast implementation), and Gemini (Research/Security) — each with specialized prompt templates.

Formal State Machine

12-state lifecycle from IDLE to DONE — INTAKE, RESEARCH, DISCUSS, USER_APPROVE, TASK_SPLIT, SPRINT, REVIEW. No skipping states, no self-approval.

tmux-based Dispatch

Sends prompts via tmux send-keys, captures output via capture-pane, detects idle/busy/crashed states via spinner and prompt indicators. Auto-restarts crashed agents.

GitHub Integration

Automated label creation, Project V2 board management, Epic/Story/Task issue hierarchy, commenting, and column management via GitHub CLI + GraphQL.

Review & QA Scoring

Automated QA scoring 0-100 with pass threshold of 85. Max 3 review cycles before escalation to human. Research debate protocol with configurable rounds.

Zero Dependencies

Pure Python 3.11+ with type hints, no external dependencies beyond stdlib. Uses tmux and GitHub CLI as system tools. Fully idempotent execution.

More Projects