prepare repos for looping.

npx @acoyfellow/loop init

AGENTS.md              # how to build/test (you edit this)
tasks.md               # what to do (you edit this)
loop.json              # config
.loop/
  progress.md          # what's done (append-only)
  errors.md            # what went wrong (append-only)
  failures             # consecutive failure count
  PAUSED               # kill switch (exists = stopped)
.github/workflows/
  loop.yml             # CI trigger

npx @acoyfellow/loop init     # scaffold files
npx @acoyfellow/loop enter    # one iteration
npx @acoyfellow/loop watch    # loop until paused
npx @acoyfellow/loop status   # show state
npx @acoyfellow/loop pause    # stop
npx @acoyfellow/loop resume   # continue

{
  "agent": "claude",
  "maxFailures": 5,
  "context": ["AGENTS.md", "tasks.md"]
}

agents: claude, opencode, aider, custom

for custom:

{
  "agent": "custom",
  "customCommand": "my-agent --prompt"
}

1. something must change - if agent produces no diff, it's a failure
2. failure limit - after N consecutive failures, loop pauses
3. kill switch - .loop/PAUSED file stops everything
4. error log - errors persist so next iteration can avoid them
5. progress log - progress persists so loop knows where it left off

you don't run loops. you prepare vessels to enter loops.

the repo IS the rig. git gives you:

• persistence (survives context resets)
• versioning (every change tracked)
• triggers (actions, webhooks)
• portability (clone anywhere)

you're not controlling the work. you're orchestrating the conditions under which loops can safely run.