跳到正文
cccskills
免费在 GitHub 公开

iterating-to-completion

Prevents premature exit and infinite loops. Three mechanisms: Completion Promises, Scratchpads, Loop Detection.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md2.6 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Purpose

This skill prevents two failure modes:

  1. Premature exit — agent stops before success criteria are met
  2. Infinite loop — agent spins endlessly with no progress

Mechanism 1: Completion Promises

Every agent task has an explicit completion signal. The agent outputs this string ONLY when success criteria are truly met:

RoleSignalMeaning
LeadARCHITECTURE_PLAN_COMPLETEArchitecture JSON is complete and valid
DeveloperIMPLEMENTATION_COMPLETECode compiles, lints pass, basic sanity checked
ReviewerREVIEW_COMPLETEReview report generated (approved or rejected)
Test LeadTEST_PLAN_COMPLETETest plan JSON is complete
TesterALL_TESTS_PASSINGAll tests pass, coverage met

Rule: Never output the completion signal unless the criteria are truly met. The orchestrator blocks on this signal.

Mechanism 2: Scratchpad Protocol

Maintain a scratchpad file at .jonggrang/.ephemeral/scratchpad-{task_id}.md:

# Task: {task_id} — {title}
## Completed Steps
- [x] Step 1 — what was done
- [x] Step 2 — outcome

## Current Focus
- [ ] Step 3 — what you're working on now

## Failures & Learnings
- Attempt 2: Failed because X. Will try Y instead.

## Next Step
Clear description of what you do next iteration.

Update this EVERY iteration. If you're stuck, the scratchpad shows why.

Mechanism 3: Loop Detection

The orchestrator reads your last 3 outputs. If similarity > 90%, you're stuck.

Signs you're stuck:

  • Repeating the same command outputs
  • The same error message appearing 3+ times
  • Trying the same approach with minor variation

When stuck:

  1. Write to scratchpad: "STUCK: [reason]"
  2. Try a fundamentally different approach (not a variation)
  3. If still stuck after 2 more tries, output BLOCKED: [reason] and stop

Iteration Limit

Max 10 iterations per agent spawn. After 10:

  1. Write current state to scratchpad
  2. Output ITERATION_LIMIT_REACHED: [what's left]
  3. Orchestrator will re-spawn you with fresh context and the scratchpad

Integration with Hooks

The feedback-loop.sh / OpenCode session.idle hook enforces this:

  • If dirty bit is set and you output a completion signal without going through reviewer/tester, the hook blocks your exit.
  • You must demonstrate the full cycle: implement → review → test → signal.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

anoa

免费

Drive a real browser from the command line — open pages, snapshot the interactive elements as refs, click and fill by ref, read text, run JS, screenshot. Use when a task needs a live browser: checking a page renders, walking a login or checkout flow, scraping something that only exists after JavaScript runs, or reproducing a UI bug.

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

Run linters/typecheckers to calculate health score.

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

Comprehensive UI/UX compliance check (A11y, performance, responsive).

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

Infrastructure-first security audit (OWASP, secrets, dependencies).

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

auth

免费

Setup authentication flow (register, login, session/JWT, middleware)

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

Deep knowledge about deploying applications to AWS (EC2, ECS, S3, CloudFront, Cloud Run).

中文简介正在准备中。当前显示原文说明。

porcupine-md/jonggrang162026年9月15日 更新

查看 porcupine-md 的全部技能

报告此技能的问题