English | 한국어
Claude Code ends its turn the moment it believes it's finished, and loses the detail at the next compaction. claude-automata rebuilds it around the way human memory works:
| Plugin | Memory role |
|---|
| ploop | working memory: a single session sails autonomously for days; an independent advisor judges every completion claim |
| tx | consolidation: the screening gate into long-term memory (independent verify, CI, squash merge) |
| refine | re-grounding: large-scale workflows that eliminate technical debt |
Long-term memory is the repository's own git-tracked text, not a database. Recall is grep. Whatever never passes the gate dies with the loop, on purpose.
Getting started
Enter the following instruction in Claude Code:
Read https://raw.githubusercontent.com/clomia/claude-automata/refs/heads/main/INSTALL.md with curl -sSL and install claude-automata.
Your agent reads INSTALL.md and installs claude-automata. Needs Claude Code and uv on POSIX (macOS / Linux / WSL).
Operating the loop
/ploop:define-mission # an agent interviews you, interprets your intent, and writes the anchor
/ploop:launch [anchor text] # hand it to the loop in a fresh bypassPermissions session
Declare the mission done and a hook blocks the stop, demanding an audit by an independent advisor with access to the whole story; the loop ends only when the advisor certifies completion.
agent › Mission accomplished. Stopping.
advisor › Not yet. The mobile layout was never measured. …resuming.
Six rounds later the advisor ends the turn. Watch a full round on the site. The anchor survives every auto-compaction, and the loop fully complies with your subscription plan's terms of service.
- init sets Auto-Compact; keep it on. The loop runs unattended: open its session with
claude --permission-mode bypassPermissions (it refuses to launch otherwise), and nothing in it waits for a human — every approval it would ask for is declined and the loop moves on.
/ploop:off pauses the loop. /ploop:on resumes or restores it, even after an accidental ESC, an API error, or a session limit (interrupt with ESC first if a turn is running). Nothing else stops the loop./ploop:docent reports the loop's progress. Run it in a separate session, same directory: questions go to the docent, and the loop session stays untouched.
Change as transactions
Agents drive tx on their own. Every change lands as one verified, CI-green squash merge behind an integrity boundary, and tx blocks writes to the base branch in between. You review merged results, not work in progress.
Keep the repository lean
/refine:code [focus] · /refine:docs [focus]
Repository-wide workflows that clear technical debt: code architecture and doc-to-code alignment. They touch only the representation layer and never change behavior. A run can sweep the whole codebase and take ten hours or more; watch it with /workflows. Like the loop, it runs unattended and starts only in a bypassPermissions session.
Apache-2.0 · Recursive self-improvement: claude-automata is developed inside claude-automata. A Claude Code agent running this environment authored every contribution in this repository.
