Skip to content

Internal tooling ·

AI developer tools

One rule file, a handoff note and a live quota footer for five AI coding agents.

Eleven lines of the agent rule file: the two steps of the handoff transaction and the note template with Task, Done, Open, Next, Decisions and Don't
An excerpt of the rule file's handoff section, typeset

A self-built control layer for the AI coding agents that build the products shown on this site. One rule file tells five agents what they must ask before they act, a handoff note per project lets any of them continue another's work, and a footer in the terminal shows what is left of each subscription window. The agents err: deterministic gates, 448 tests and a human review are what catch it.

Main functions

  • One rule file for five AI coding agents: Claude Code, Codex, Kimi Code, Google's Antigravity and a local GPU agent
  • A handoff note per project (Task, Done, Open, Next, Decisions, Don't), so any agent continues another's work
  • A progress row in the terminal footer; a hook asks for an update once it is five minutes stale
  • Status lines that show what is left of each subscription window and where the burn rate lands; work is planned to 90% of the window
  • A local coding agent on an RTX 4090 through llama.cpp, with a hard-stop list: push, commit, rm -rf, data-changing SQL, deploys and credentials always ask
  • Six benchmark suites with deterministic gates: an invented file:line fails in milliseconds, and the local model gets a kind of work once it clears an 80% bar
  • Hooks and a watchdog: a safe-rm check before a shell command runs, and local inference paused on GPU heat or a RAM floor
  • A usage ledger in SQLite, de-duplicated per response, with a burn-rate forecast
  • 448tests, all passing on 2026-10-08
  • 6benchmark suites behind an 80% bar
  • 5AI coding agents on one rule file
  • 0external Python dependencies

Built with: Python 3.12, no external packages · SQLite · Bash, systemd, tmux · llama.cpp · RTX 4090 · Claude Code hooks and status line

A terminal status line: a progress row at 3 of 9 done, a 5-hour window at 18% used with a projection of 87% at reset, a 7-day window at 100% used, and a credits row
The status line under the terminal input, rendered from a real payload on 2026-10-08
Two panels of the dashboard: GPU memory, temperature and the budget for local inference, and the local inference server shown as down with its concurrency log
Two panels of the dashboard, cropped; the local server was down when captured

2026-10 · the next branch

The next branch is yours.

Tell me what people order or book from you today, and what should run without you tomorrow. A free mockup before any money moves, a fixed price for the outcome, and the person who built it still answers the phone after launch.

Start a project