Frame Loop

Turn a green candidate into a loop an agent can run unattended without cheating and without you babysitting it. Four branches: run them in order for a full handoff, or one alone when only that piece is missing.

The governing seam is the Hart split, made mechanical. The agent edits the attempt; a separate controller owns the evaluator. The change-gate enforces it — before any score counts, the candidate's diff must be a subset of the declared editable paths. [references/boundary.md](references/boundary.md) is the single source for protected paths and the change-gate; every branch defers to it and none restates it.

Branches

1. Frame the contract. Re-confirm the candidate is green per the gates skill, with the causal gate and the Goodhart pre-mortem explicitly satisfied — a contract for a causally-invalid target is a clean way to optimize the wrong thing. (If the gates skill is not installed, apply those two checks from first principles; they are the keystone.) Then fill the six primitives:

• TARGET — metric, better-direction, bound.
• BATTERY — the fixed task and data: a frozen snapshot, never live state.
• JUDGE — the command that scores one attempt; if scoring needs judgment, the JUDGE is a calibrated judge-panel, not a bare prompt.
• LEDGER — git, one commit per experiment, keep or revert.
• BUDGET — cost per trial and a total-run cap; the cap is what bounds gaming.
• UPDATE-RULE — keep iff the held-out number improves and no guard regresses.

Done when: all six are executable as written and the causal gate is re-confirmed.

1. Build the harness. Draft the frozen evaluator for the BATTERY using [references/harness.md](references/harness.md): fixed cases, a machine-readable result schema, baseline capture, a retest band for noise, and the three states — baseline, incumbent, candidate.

Done when: the harness has fixed cases, a result schema, a baseline run, a keep/revert rule, a retest band, and at least one regression case.

1. Select loop mode. Classify with [references/loop-mode.md](references/loop-mode.md) as overnight, short ratchet, needs harness, manual only, or reject, and write the loop brief with its stop conditions. Return a brief only for overnight or short ratchet; otherwise return the no-go reason and the smallest missing prerequisite.

Done when: a mode is assigned and either a loop brief or a named blocker is returned.

1. Set the goal contract. If the loop runs under a persistent goal (Codex Goal Mode, or a Claude Code background run), bind it with [references/goal-control.md](references/goal-control.md). The goal is task state, not the evaluator; budget exhaustion is not completion.

Done when: the goal names the artifact, metric, gates, and stop conditions, and completion is defined as a requirement-by-requirement audit.

Output

Full handoff: the contract (six primitives), the harness spec, the loop-mode decision with its brief, the goal policy if relevant, and the open risks with the smallest next action. Single branch: that branch's artifact plus any prerequisite it needs from the others.

Hard rules

• Reject any loop where the agent can reach the evaluator, fixtures, expected outputs, scoring code, baseline records, prior logs, secrets, production data, deploy config, or billing config.
• The change-gate runs before scoring, every trial. A protected-path diff reverts the run.
• Crashes, missing metrics, suspicious jumps, and noise-band wins are retest or revert, never keep.
• Keep the evaluator boring: deterministic first; a judge-panel only when it carries a current agreement receipt and is frozen.
• Money and compliance paths stay human-ratified. The loop never auto-merges them.
• Delete a bad candidate rather than carry it with caveats.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TylerJNewman
• Source: TylerJNewman/skills
• License: MIT

Install and usage instructions live in the source repository linked above.