All work

myFleet

One copy of my agent skills, rules, and hooks, written out to eleven coding-agent tools.

Source on GitLab Python · MPL-2.0 · v0.1.0
Isometric drawing of one source and eleven agent harnesses on the right. The source is one cyan tower of four strata, one per kind of content, each as tall as its count on one scale, from the bottom up: 46 skills, 29 policy documents, 17 hook registrations and 7 agent charters. A wire runs from the source to each of the eleven slate harness blocks, listed in the order sync-agents declares them: claude, codex, gemini, cursor, droid, opencode, kilo, kimi, antigravity, vix and zcode. Two of them, opencode and kilo, are amber because sync-agents declares no hook config file for them; the other nine take hooks. The drawing shows what the code declares, not a run of it. charters7 hooks17 policies29 skills46 claude codex gemini cursor droid opencode kilo kimi antigravity vix zcode amber: no hook config

Every coding agent keeps its instructions, skills, and settings in its own private format, so a fleet of them drifts. myFleet is the public port of the machinery that runs my own agent workspace: 46 skills, 7 agent charters, and 29 rule documents whose rules carry stable IDs, so a skill or a review can cite a rule instead of restating it. A sync engine keeps one canonical copy of each and writes it out to eleven harnesses, from Claude and Codex to Cursor, Kimi, and Kilo. It defines seventeen hooks and registers them in each harness that has a hook system.

It's a sizeable system of code and prose to keep everyone working together nicely: everything from crash recovery to worktree isolation and so much more, all motivated the hard way, of course!

Everything previews before it applies, and applying twice changes nothing. Its CI fails on anything private that slipped in. Plane, the tracker I use, shows up only as a skill and a hook that mirrors an agent's todo list; syncing a local task mirror isn't in the toolkit yet.