Loop Oversight

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. 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 "review → fix → re-review until clean".
5. Integration cadence. Commit per unit (atomic — keeps git log the audit trail and never trips big-change-reminder). An unattended loop opens one PR per unit and never auto-merges — the [[commit-digest]] pattern.

The per-turn contract

Paste this into the loop prompt so every iteration is self-checking:

read ledger → pick next un-done item → do it → verify (separate pass) → update the ledger row with evidence → commit.

Stop when the bound is hit, or the ledger has no un-done item two turns running.

Substrate — match it to how much you can watch

Detail and the budget/until-dry patterns: substrate-selection.md. Ready generate + verify prompts, plus a poll-until-true scaffold for "wait for a condition, then take one action" (merge once CI/main settles, confirm a deploy went live): prompt-scaffolds.md.

Completion criteria

• All five gates existed as artifacts before the loop ran (ledger file, defined unit, written bound, verify pass, commit/PR cadence).
• Every ledger row reaches a terminal state with linked evidence (commit / PR / test output), or is explicitly parked with a reason — no silent skips.
• The loop stopped at its bound, not by running out of context or being interrupted.

Red flags — stop and add a gate

• No ledger file, or the loop isn't updating it each turn → oversight is invisible.
• "I'll just let it run and check the result" with no iteration/budget ceiling → unbounded.
• The same agent writes an item and marks it done → no independent verify.
• An unattended loop merging to main → review can never happen.

Notes

• The bound lives where you write it. A Stop hook can't hard-cap a /loop (it can only nudge the next turn) and can't see token counts. The real ceiling is max iterations / until-N-empty in the prompt, or budget.remaining() in the Workflow tool.
• For a batch sweep across many targets, the dev-env-setup fleet mode (canary first → one isolated agent per target → verify each) is the supervised template — reuse that shape.
• Related: [[multi-session-plans]] owns the .claude/current_plan.md phased checkpoint this skill uses as a ledger; [[repo-review]] and ad-hoc work surface the items a loop then works.