01 / AI agent infrastructure

gongbox

Control plane for autonomous coding agents on a dedicated server

Provisioning and agent tooling for a single-user development box: a dedicated server that runs coding agents against Linear issues while the laptop is asleep. One idempotent script takes it from bare metal to a working machine. From there an issue becomes a git worktree, a tmux session and a Claude Code agent; a second agent can be opened on the resulting pull request as an independent reviewer, held to reading and commenting by allowlists rather than by instructions. Around that sit a single launcher that is the only way an agent process starts, a run recorder with structured results and retention, a one-screen health check whose JSON output is the real interface, and a deterministic Linear and GitHub reconciler that is implemented but deliberately not scheduled to run. Built for one machine and one user, not as a product.

Why it matters

Agents are easy to start and hard to keep honest. Most of this is the unglamorous half: one task per isolated worktree, one way to launch a process, a reviewer that structurally cannot push, and a record of what each run actually did.

Built with Python, Shell, systemd, tmux, git worktrees, Linear API, GitHub API.

all work