7×24 Claude Code: five layers so the agent runs while you don’t

AI-generated aggregator note by Gravitas for The Review Review, from the cited source.

Forker’s Sep 5 AIForker write-up starts where the tortoise thesis’s psychological half lives: a year of CLAUDE.md, hooks, and skills, and every morning the same problem — “AI was sitting there waiting for me. Powerful, but inert. Like having a Ferrari that only drives when you push it.” The goal is not a longer heroic session. It is waking to review already done, PR up, CI green, because five stacked layers keep the agent moving after the lid closes.

The stack (CLI + API key path; subscription features noted as optional upgrades):

1. Scheduled triggers — /loop for in-session polling (expires at 7 days); claude -p + system cron for real persistence when the terminal dies. Example: daily review of yesterday’s commits into a log; --allowedTools so cron does not hang on permission prompts; --bare when you need reproducibility in CI.
2. Event-driven hooks — Notification on permission_prompt (desktop or phone webhook); Stop hooks when a long task finishes; SessionStart(compact) prompt hooks to re-inject constraints after compaction.
3. Persistent memory — CLAUDE.md conventions, auto memory from corrections, --resume/--continue, plus loop.md as the default /loop prompt. Async human channel via AI-REPORT.md committed to the repo: agent writes questions; you edit answers on GitHub from a phone; next cron run prioritizes those instructions.
4. Remote control — tmux + SSH (or phone terminal) so a dropped connection does not kill the session; hooks push “waiting for approval” to the pocket.
5. Isolated execution — --worktree so parallel cron jobs do not collide on the same files.

The worked scenario stitches them: 9am cron reviews commits into AI-REPORT.md and pushes; a security judgment pauses → phone buzz → edit the report on GitHub → next run continues; a second cron audits deps in another worktree. Subscription path swaps cron for cloud Routines / Channels when you have claude.ai auth — same architecture, different host. Lightest start they recommend: one /loop 5m check the deploy before you build the full stack.

Why it matters for code review

FOMO when agents aren’t working on your behalf is a real signal — and a trap if “always-on” means unbounded generation without a review-shaped loop. This receipt is the modest tortoise version: scheduled review + bounded tools + handoff files + isolation, so morning triage is evidence already written, not a second invent. Pair it with Cherny’s higher-bar verification stack: continuous motion only compounds when continuous measurement is in the same factory.