Skip to content
Back to skills

Review Gate

ASecurity

Configure Agent Looper post-verify quality review and reviewGate — Proceed / Guide / Deny / Confirm, impact-severity gating, and when not to thrash another iteration. Use when writing REVIEWS.md or tuning postQualityReview.

  • 3 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 20, 2026
ai-agentsgoawsdocumentation

Works with

  • cursor
  • cli

Security analysis

A100/100

Scanned October 20, 2026

npx -y skills add dancingteeth/agent-looper --skill review-gate --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Review Gate?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Review Gate
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/dancingteeth-review-gate/badge)](https://www.skillsdirectory.com/skills/dancingteeth-review-gate)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: review-gate
description: Configure Agent Looper post-verify quality review and reviewGate — Proceed / Guide / Deny / Confirm, impact-severity gating, and when not to thrash another iteration. Use when writing REVIEWS.md or tuning postQualityReview.
tags:
  - agentic_ai
  - agents
  - documentation
  - loops
---

# Review gate

After verify exits `0`, an optional **judge** writes `review.md`. Completion is still verify-first.

## Specs ≠ prompts

- `AGENTS.md` / skills → worker how-to
- `REVIEWS.md` → judge residual standard (not runtime worker prompt text)

Keep both sparse; delete laws models stop failing.

## Modes

| Setting | Behavior |
| --- | --- |
| `postQualityReview: false` | No judge |
| `postQualityReview: auto` | Judge when inferred risk ≠ low |
| `reviewGate: true` | Only **error + impact** findings reopen the worker (up to `maxReviewCycles`). Gating bullets win over a quoted `### Verdict — PASS` in the same `review.md`. |

On homepage / mockup / screenshot-as-hero loops, `design-loop` should set `reviewGate: true` and must not copy `false` from `*-smoke`. This skill is the `REVIEWS.md` overlay, not the only way to turn the referee on.

Cosmetic nits must not burn another iteration.

## Residual judgment vocabulary

Proceed / Guide / Deny / Confirm — Guide packets feed the next worker when the gate continues.

## Judge runtime

- Default judge: the worker's runtime (`reviewRuntime` unset → same as `runtime`, with that runtime's judge model)
- `reviewRuntime: opencode` without `reviewModel` → `opencode-go/deepseek-v4-pro` (do not copy the Flash worker slug)
- `reviewRuntime: dsh` without `reviewModel` → `deepseek-official/deepseek-v4-pro`
- `reviewRuntime: muse` without `reviewModel` → `muse-spark-1.3`
- Set `reviewRuntime` + `reviewModel` to keep the judge off Cursor quota (Pi, OpenCode BYOK, Cline, DSH, Muse, Claude, …)
- Never use Composer Fast as the judge

## Reproduce filters (opt-in)

`reviewReproduce` / `reviewReproduceAgent` drop confabulated blockers that do not cite real changed paths.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.