Token导航 LogoToken导航TokenDH.com
待分类权限需确认github未标认证来源可访问许可证需确认审计通过

feedback-loop反馈回路

Agent Skill

feedback-loop 用于处理 GitHub 仓库、Issue、Pull Request 和代码协作信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要围绕仓库状态、代码变更或协作事项进行整理时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

总安装

220

周安装

9

GitHub Stars

511

下载量

71
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:feedback-loop(反馈回路)
来源仓库:https://github.com/dmmulroy/.dotfiles
仓库路径:skills/feedback-loop
安装命令:
npx skills add https://github.com/dmmulroy/.dotfiles --skill feedback-loop
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/dmmulroy/.dotfiles --skill feedback-loop

简介

feedback-loop 暂无可用说明文档,仅可通过 GitHub 查看原始代码结构。

  • 适用于探索性学习或参考他人技能组织方式,支持 Codex、Claude、Cursor、Gemini CLI。
  • 通过 npx skills add 命令安装,但无 SKILL.md 指导,需自行分析实现逻辑。
  • 使用前应谨慎评估安全性,避免执行未知脚本或修改系统配置。
  • 适用宿主包括 Codex、Claude、Cursor、Gemini CLI,接入前应确认版本、权限和运行环境要求。

SKILL.md

name
feedback-loop
description
Self-validate work through deterministic feedback loops with repro, measurement, and exit criteria. Use when stuck, outcomes unclear, validation needs automation across code, UI/visual, or data pipeline tasks.

Feedback Loop

Create fast, deterministic feedback loops so the agent can self-validate and converge.

When to Load

  • Task stalled or unclear success criteria
  • Need reproducible failures or measurable outcomes
  • UI/visual or data outputs hard to validate

Step 0: Discover Repro

Before any loop, find or create the repro command. In an unfamiliar codebase:

  1. Check AGENTS.md, README.md, CONTRIBUTING.md for test/dev/build commands
  2. Inspect runner configs: package.json scripts, Makefile, justfile, Taskfile, Cargo.toml, pyproject.toml
  3. Check CI: .github/workflows/, .gitlab-ci.yml, Jenkinsfile for test/build steps
  4. Search for test files near affected code: glob **/*test*.*, **/*spec*.*
  5. If no test exists, write a minimal reproduction script or test first — this is your first loop iteration

Loop Selector

SignalLoop typeRecipe
Failing tests, errors, type errors, regressionsDebuggingloop-recipes.md#debugging
Visual output, animation, layout, renderingUI/visualloop-recipes.md#uivisual
Metrics drift, ETL, data qualityData pipelineloop-recipes.md#data-pipeline
Build/compile errors, dependency issuesDebuggingUse build command as repro

If multiple match, prefer the loop with the most deterministic measurement.

Core Loop (every iteration)

  1. Sense: capture current output + signal (run repro, read watcher output)
  2. Hypothesize: state suspected cause and what you expect to change
  3. Change: smallest fix to test hypothesis (one thing at a time)
  4. Measure: run deterministic check (test, snapshot, metrics query)
  5. Decide: keep if green, revert if no signal change, iterate with new hypothesis

Principles

  • Convert human-centric signals into text or structured artifacts
  • Encode state in parameters so results are deterministic
  • Prefer fast, headless runs over slow UI paths
  • Allow the agent to add logs/metrics to expose signal
  • Separate exploration (agent) from acceptance (human)

Required Artifacts (non-negotiable)

ArtifactWhatExample
ReproDeterministic command or URLnpm test -- auth, agent-browser open localhost:3000
Expected vs observedExplicit comparison"expected 200, got 401"
Measurement methodRe-runnable checktest runner, agent-browser snapshot -i, metrics query

Full artifact matrix per loop type: loop-recipes.md

Persistent Observability (tmux)

If tmux is available, set up persistent feedback sources in adjacent panes:

SourceStart commandRead signal
Test watchernpm run test:watch / pytest-watchtmux capture-pane -p -J -t {pane} -S -50 then grep PASS/FAIL
Dev servernpm run dev / cargo watch -x runtmux capture-pane -p -J -t {pane} -S -50 then grep ERROR
Build watchertsc --watch / cargo watch -x checktmux capture-pane -p -J -t {pane} -S -50 then grep error
Browseragent-browser open localhost:3000agent-browser snapshot -i

Read from adjacent pane: tmux capture-pane -p -J -t {pane} -S -50 Wait for pattern: while ! tmux capture-pane -p -t {pane} -S -20 | grep -q "pattern"; do sleep 1; done

This gives continuous feedback without re-running full commands each iteration.

Artifact Persistence

  • During loop: keep artifacts in conversation context (fast, ephemeral)
  • On rescope or escalation: persist to a markdown file for tracking
  • On completion: include final repro + resolution in commit message or PR description

Exit Criteria

  • Expected output matches observed and measurement is green
  • Repro command passes deterministically

Rescoping (when loop stalls)

If no new signal after 2 iterations:

  1. State what you tried and why it didn't produce signal
  2. Question your loop type — wrong category? Re-run selector
  3. Question your repro — is it actually exercising the bug? Widen scope
  4. Improve observability — add more logging, check adjacent systems, use tmux watchers
  5. Reduce scope — find a smaller, more isolated failing case
  6. If still stuck: escalate to user (see below)

Loop Switching

If a bug spans domains (e.g., visual symptom but root cause in data/logic):

  • Start with the most deterministic measurement
  • If 2 iterations produce no signal, switch loop type
  • Derive a text/assertion proxy from visual symptoms where possible (e.g., check computed styles, DOM structure, API response instead of screenshot)
  • Carry artifacts forward — repro command and observations transfer between loops

Escalation Triggers

Stop and ask the user when:

  • Agent cannot execute measurement (no screenshot tool, no DB access, no metrics)
  • Fix requires changes outside agent scope (infra, permissions, external service)
  • 3 rescope attempts with no convergence
  • Measurement is subjective (visual "looks right", UX feel)

When escalating, provide: repro command, last hypothesis, all artifacts collected.

Handling Flaky Repros

If the repro is non-deterministic:

  • Pin randomness (seed values, --seed flags)
  • Freeze time (mocks, faketime, test fixtures with fixed dates)
  • Mock external dependencies (network, APIs, filesystems)
  • Run N times to distinguish signal from noise
  • Reduce concurrency / isolate the test

In This Skill

FilePurpose
loop-recipes.mdSteps, tooling, artifacts, checklists per loop type
examples.mdFull worked loop iterations with hypotheses + decisions

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

34.99%
按下载量换算25

Claude

28.09%
按下载量换算20

Cursor

19.46%
按下载量换算14

Gemini CLI

9.14%
按下载量换算6

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

权限需确认

当前来源未能明确判断权限范围,默认进入异常复核队列。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills