Synagon — Data skill for Claude Code

Synagon

[](LICENSE) [](https://github.com/LemonAidLee/Synagon/actions/workflows/tests.yml)

Synagon is a multi-agent orchestrator built on [LangGraph](https://github.com/langchain-ai/langgraph). It drives **local command-line AI agents** — Antigravity, Claude Code, OpenCode, Codex — through a research → plan → implement → verify loop, with automatic repair when verification fails.

It uses no provider API keys. Every agent runs as a local subprocess against the CLI you have already authenticated.

**The cockpit.** `python -m orchestrator --daemon` opens a **cockpit**: one column per role, one card per agent, live as it runs. Type a goal into it, watch it flow left to right through the team, click an agent to read its terminal, approve a gate and deliver the branch without leaving the page. An **Explorer** browses the project *and every run's worktree*, so "what did the agent actually write?" is one click rather than a `cd`; `Ctrl K` opens a command palette and `Ctrl P` jumps to a file. `desktop/` wraps the same thing as a real desktop window, opened per project like an IDE. See [ARCHITECTURE.md](docs/ARCHITECTURE.md) §22–§26, which describes only what exists today.

**How far along is this?** See [Known limitations](#known-limitations) below for an honest, current account of what is not yet built — most notably that process-ownership hardening on a hard kill is proven on Windows only, and packaged installer builds are verified on Windows only.

  GOAL ──▶ decomposer ──▶ TASK ──▶ session ──▶ branch          one session per task,
                       ├▶ TASK ──▶ session ──▶ branch          in dependency order,
                       └▶ TASK ──▶ session ──▶ branch          optionally in parallel

  session:  context ▶ preflight ▶ researcher ▶ planner ▶ implementer ▶ gate ▶ verifier ▶ final