Ian Rodgers
All projects

everythingdev

A self-hosted development environment where AI agents build, review and ship code, and I make the decisions.

What it does

One repository provisions my Linux machine and installs a layer of hooks, skills and reviewer agents into Claude Code. An orchestrator dispatches AI worker sessions into a pool of git worktrees, gates every change through an independent adversarial review, merges only when that review and CI pass, and reaches me through a single Telegram message when a decision is mine.

It manages nine projects, Gleam Eats and Truth Seeker among them. The code is written by AI agents working under my direction; I set the direction, pick between designs and answer the decisions they raise.

Architecture

  1. Intent
  2. Worker in a leased worktree
  3. Adversarial review
  4. Merge check + CI
  5. Deploy
  • Hooks. 7 Claude Code hooks: a guard before every tool call, a sync of per-project state at session end, and a watcher that wakes the orchestrator.
  • Orchestration. 37 Python modules: a file-based job queue for escalations, decisions and review receipts; a worker dispatcher; a scheduler; a systemd watchdog; a lease pool of git worktrees.
  • Review. 4 reviewer lenses and a fixer agent. A change lands only when every lens routed to it has passed at that exact commit, a merge check exits 0, and CI is green.
  • State. Per-project manifests and backlogs live in a separate fleet repository, one entry per project.

Hardest problems

  1. CI with no free minutes left

    When GitHub's free CI minutes ran out in August 2026, all CI moved to self-hosted runners running as user services. The host has no sudo and no Docker, so LXD supplies throwaway containers, with podman inside one for the image build.

  2. A scheduler that died unnoticed

    The scheduler stopped for 83 minutes without anyone knowing, because its monitor lived inside a terminal session. It now runs under a systemd watchdog with bounded backoff, a standard-library phone alarm and passes that never block.

  3. Code comments nobody could check

    1.9 MB of comments made claims about the code that nothing verified. A precision-first checker now pins every referenced path, qualified name and annotated count to the code; three broader rules were measured and removed.

Screenshots

A rendering of a real decision message on my phone, asking where the Gleam Eats review form's bar should sit on wide screens, with lettered buttons and a recommended answer quoting part of the decision's recorded reasoning.A rendering of the live terminal tab list, with eight AI worker tabs open, seven of them working.