iliyarka87

Aion Kit — Development skill for Claude Code

Development community

Mechanical guardrails for unattended agent work: checklist gate, director/executor/judge loop, AF_UNIX door into Claude Code, failure registries.

How to install Aion Kit

This entry records only its repository, not the path inside it, so there is no exact command to give. Open iliyarka87/aion-kit and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Aion Kit does

Mechanical guardrails for unattended agent work: checklist gate, director/executor/judge loop, AF_UNIX door into Claude Code, failure registries. MIT.

Alternatives in Development

Executor 24.1k ★
Bolt Executor — Use the @bolt-executor agent to implement the following: $ARGUMENTS 523 ★
Evolving Agents — A director agent that learns how your team works: it plans by walking your CLAUDE.md and memory as linked stat 453 ★

README

AION kit

Mechanical guardrails that let a coding agent (Claude Code, Codex, …) work through a long checklist **unattended** — overnight, without a human clicking "allow", without skipped items and without "done" that nobody verified.

Five parts, each usable on its own:

part	what it is	start here
`aion-checklist/`	the gate: a mechanical lock on a checklist — one item in work, dependencies enforced, `DONE` only by the controller and only with evidence, every refusal journaled	`python3 sobrat.py` → `python3 proby.py`
`aion-loop/`	the loop: director → executor → judge → gate, driven by a small "hand" process; registries of failures, decisions, invariants; a change gate with eight refusal conditions; a PreToolUse hook for Claude Code; stable/candidate/rollback	`aion-loop/README.md`, `python3 proverit-komplekt.py`
`aion-socket/`	the door: a direct AF_UNIX link into a live Claude Code window on the same machine — the hand drops a task into the window as an ordinary message and reads the reply from an exact offset	`aion-socket/README.md`, `python3 -m unittest discover -s tests`
`svyaz/`	the link: director → executor contract — task packet with mandatory fields, a shared vocabulary of terminal markers, idempotent delivery keyed by `item:attempt`, recovery after restart	`svyaz/README.md`
`aion-router/`	the router (optional): one stable plan per task, model/effort chosen by task family with spend protection, separate lanes per chat	`aion-router/README.md`

What problem this solves

An agent left alone with a 69-item plan will (we measured this on ourselves): forget where it stopped after a context reset, report "done" without a check, jump ahead of unmet dependencies, quietly widen the task, and stall on a permission popup u