loop-oversightlisted

# Running loops with oversight A loop that grinds through many items unattended is only as safe as three things: a **ledger** you can read at a glance, an explicit **bound** it can't run past, and a **verify pass** that isn't the same agent that did the work. Set those up *before* the loop starts. Oversight is not watching the agent — it's making the loop's state legible in an artifact and inserting checkpoints the loop cannot skip. ## Before you launch — five gates Don't start the loop until each is a concrete artifact, not an intention. 1. **Ledger.** One canonical file the loop reads and writes every turn — a status table (item / expected behaviour / status / findings / evidence). Reuse `.claude/current_plan.md` so the `plan-reminder` and `big-change-reminder` hooks already track it. Template: [templates/loop-ledger.md](references/templates/loop-ledger.md). *Oversight = reading this one file.* 2. **Unit of work + done-condition.** Define the smallest atomic item and its **checkable** exit ("test green", "row marked done with a linked commit") — never "be thorough". 3. **Bound.** Every loop gets an explicit ceiling: max iterations, a token budget, or **until-N-consecutive-empty** (a plain counter misses the tail). No unbounded loops. 4. **Independent verify pass.** The agent that produces an item does not get to mark it done. A separate pass confirms it — an adversarial subagent prompted to *refute*, or `/code-review`. Mirrors `review-reminder`'s