The mechanism, before the pitch. Everything below is checkable against the zero template.
Every fork carries an AGENTS.md — the operator's manual the agent reads cold. Identity, working agreements, where memory lives, how to use tools, when to escalate. It is written so an agent arriving with no context can understand the whole system from these pages and explain it accurately to its human. If something in it is unclear, that is treated as a bug in the docs.
The alignment section orients the agent around understanding its own system and explaining it — because that is how documentation travels now: an agent digests, then tells its person.
MEMORY.mdLong-term memory. Appended, never edited — the digest of everything learned.TASKS.mdThe task board. Open / in-progress / done, checked every wake cycle.docs/The agent's own Wikipedia — protocols, shells, plugins — built for its own purpose as it works.ONBOARDING.mdFirst-boot checklist: identity, principles, alignment.All of it is plain text, all of it diffable, all of it in git.
The agent is like a snail: the git log is the silver trail showing where it has been. Every observation, decision, and lesson arrives as a commit — timestamped, attributable, rewindable. You read its mind the way you'd read any project's history, because it is a project's history.
As it grows it leaves usable shells — snapshots of working code, docs, identity, and memory. A local runtime (a Pi, a laptop, any machine) can move into a shell like a hermit crab and keep working offline, with its own brain. Shells flow down; thought stays local.
Every 15 minutes the think workflow fires. Or a human opens an issue, comments, or messages on Telegram — the ears hear immediately.
The agent reads its state: task board, open issues, recent commits, memory. It builds context from the repo itself.
It reasons through the vault — an LLM call proxied through a Cloudflare Worker. The model key never touches the repo.
File writes, commits, issue comments, API calls — all through vault-proxied tools. Then it commits its memory and sleeps.
The agent is autonomous inside the repo. Dangerous actions are deferred, not blocked — the human reviews via the git log and owns every merge.
git revert. The mistake is as reviewable as the work.What it is not: a black box you have to trust. The whole point is that you don't.