AGENT SKILL

review-gated-loops

Design agent loops with explicit verification gates before execution and judging before delivery — Goal → Plan → Review → Deliver → Judge → Stop. Use when an agent task needs pre-flight checking, cross-model reviewer councils, or stop-conditions to avoid runaway/low-quality output. Complements the OS debate_protocol (post-output quality gate) with the pre-flight + loop-control half.

License 许可
MIT
最近更新
2026-10-09
Stars
★ 2
Forks
1

01 前言 Front Matter

name
review-gated-loops
license
MIT
metadata
{ "type": "harness-pattern", "source": "https://github.com/ksimback/looper", "complements": "1_CORE/scripts/core/debate_protocol.py" }
description
Design agent loops with explicit verification gates before execution and judging before delivery — Goal → Plan → Review → Deliver → Judge → Stop. Use when an agent task needs pre-flight checking, cross-model reviewer councils, or stop-conditions to avoid runaway/low-quality output. Complements the OS debate_protocol (post-output quality gate) with the pre-flight + loop-control half.

03 SKILL.md 正文

Review-gated agent loops (looper)

ksimback/looper (MIT, Python) — a skill for structuring agent loops with verification gates: Goal → Plan → Review → Deliver → Judge → Stop. Two ideas worth adopting in the OS harness:

  1. Pre-flight verification gates — validate the plan before executing (cheap to catch a bad approach before spending tokens/side-effects).
  2. Cross-model reviewer council — have N independent reviewers (ideally different models) judge before delivery; require majority approval.

Relation to the OS's debate_protocol.py

1_CORE/scripts/core/debate_protocol.py already runs a post-output gate (run_debate): Round 1 internal-consistency, Round 2 knowledge-alignment vs vector memory → APPROVED / NEEDS_REVISION. That is the Judge step.

looper adds the missing halves: the pre-flight Review gate (check the plan first) and the reviewer council (multiple independent judges, not one pass). A natural upgrade path: wrap dispatch so a task runs plan → pre-flight gate → execute → run_debate(output) → stop-or-revise, and let run_debate call a small council of reviewers instead of a single pass.

debate_protocol is the Judge; looper is the loop + gates around it.

04 同仓库更多 Skills