Loop-Engineering v1.1.0 — Idea-to-Loop

• Added a first-class Propose mode that turns vague recurring ideas into evidence-bearing Goal Contracts and non-executable LoopProposal artifacts.
• Added a bounded intake policy: ask at most three questions that materially change routing, verification, permissions, or safety; record safe unknowns as explicit assumptions.
• Added proposal and proposal-decision schemas, templates, custom semantic/traceability validation, and CLI commands to validate, decide, and compile proposals.
• Added one domain-neutral finite Work Plan inside a Loop for composite outcomes: stable Parts, acyclic dependencies, exact Part acceptance, durable artifact bindings, and dependency-ready scheduling.
• Added optional Part-specific verification envelopes, extensible typed evidence names, and generic local/source/tool permissions so heterogeneous Parts can use domain-appropriate checks without creating domain-specific Skills.
• Added separate Part Gates and a distinct final Goal Gate; failed Parts cannot unlock dependents, final failure reopens affected Parts and descendants, and only a passing whole-Goal evaluation can complete the Loop.
• Required accepted Part artifacts to resolve to distinct regular files inside the Loop directory with recomputed SHA-256 matches; external deliverables retain a local export or receipt.
• Added immutable Part-result and Goal-run state anchors, audited part reopen / goal resolve recovery for blocked or human-required states, and exact failed-Goal-criterion tracing to affected Parts.
• Added complete finite-ledger causal replay so direct edits cannot reopen Parts, lower retry/Goal counters, resume unresolved attention, or forge completion; continuous streams retain their bounded recent ledger.
• Protected frozen state.json and loop.yaml from command-executor mutation by rejecting the draft, retaining evidence, restoring the preflight snapshot, and recording a counted failure.
• Added explicit continuous-stream versus finite-plan routing, a generic finite-project example, phase-aware adapters, and Work Plan protocol/runtime documentation without domain-specific behavior.
• Bound human proposal decisions to the exact proposal ID, revision, and canonical SHA-256; modified or unresolved proposals fail closed.
• Added digest-linked proposal revisions; revision 2 and later require the exact parent during validation, decision, and compilation.
• Made proposal compilation deterministic, persisted the complete Goal Contract, and embedded tamper-evident proposal/candidate/Goal/reconstructable-decision provenance in the resulting loop.yaml.
• Preserved approved contract identity during initialization: proposal-compiled loops require the exact metadata.name, .loop-engineering/loops root, and reviewed persistence paths.
• Preserved smaller one-shot and deterministic routes and kept proposal, evaluator, governance, runtime, and strategy-evolution trust boundaries separate.
• Added positive, negative, digest-tampering, readiness, and fresh-session behavior tests for the complete Idea-to-Loop path.

中文摘要：本版本把“模糊想法”升级为正式输入。Skill 先生成可验证 Goal 与不可执行 Loop 提案；只有人类确认假设和 prerequisites、审阅完整提案并批准精确摘要后，运行时才能确定性编译契约。有限复杂任务使用同一个 Loop 和通用 Work Plan 分 Part 推进，每个 Part 独立验收，全部通过后仍须经过最终 Goal Gate；持续任务则保留有界发现模式。编译产物保留完整 Goal Contract、可检出篡改的来源绑定和不可变持久化布局；提案设计、执行验收与权限授权互不冒充。

Install / 安装

gh skill install sheepxux/Loop-Engineering loop-engineering@v1.1.0 --agent codex
gh skill install sheepxux/Loop-Engineering loop-engineering@v1.1.0 --agent claude-code

The attached standalone Skill ZIP has SKILL.md at its archive root and can be uploaded directly to compatible Skill interfaces.

附件中的独立 Skill ZIP 在压缩包根目录直接包含 SKILL.md，可上传至兼容的 Skill 安装界面。

SHA-256 945f4e9fa4e5cdc9028bb4497a916765bd4ba7813a033e89b8a8656696883f3f

Runtime tarball / 运行时安装包

The release also includes an audited npm-compatible runtime tarball. It can be installed directly even before npm Registry publication.

本 Release 同时包含经过审计的 npm 兼容运行时安装包，即使 npm Registry 尚未发布也可以直接安装。

npm install --global https://github.com/sheepxux/Loop-Engineering/releases/download/v1.1.0/sheepxux-loop-engineering-1.1.0.tgz

Runtime tarball SHA-256 838cffb83bc9fdf5cceae7187fdc8ee5ea5fbdb39ece93f6669f96bbf26ce4a3

Loop-Engineering v1.0.2

中文

Loop-Engineering v1.0.2 是完整 Skill-first v1 的发布加固版本。它将周期性 Agent 工作变成有边界、可审计、可恢复的工程循环，并把“进化”严格限定为基准驱动的外部任务策略优化——不修改模型权重，也不能削弱安全契约。

本次重点

• 完整 canonical Agent Skill：渐进式 references、自包含 assets、黑白品牌图标、版本绑定的新会话评测证据。
• 运行时完整性：generation/契约/策略 SHA-256 绑定、可恢复多文件事务、并发安全 lease、可信 Runner 时间、不可变 evaluator 证据和失败关闭的预算/重试。
• 受控策略进化：相同 case 的 baseline/candidate 对比、逐例 arm 与策略摘要绑定、机械重算、不可变人工批准/拒绝、连续策略归档与单调回滚。
• 安全迁移：loopctl migrate 验证 v1.0.1 状态、策略父链与 lease；旧 pending 实验作废，不能复用审批。
• 发布质量：109 项测试，Node 20/22/24 CI，真实 npm tarball 空项目安装，Codex/Claude 插件安装，官方 Skill 与 GitHub 发布校验，0 个已知依赖漏洞。

安装

gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.2 --agent codex
gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.2 --agent claude-code

Codex / Claude 插件和源码安装方式见中文 README。

从 v1.0.1 升级

先停止 Runner、确认没有活跃任务并备份循环目录，然后使用 v1.0.2 运行：

npm exec --yes --package="github:sheepxux/Loop-Engineering#v1.0.2" -- loopctl migrate .loop-engineering/loops/<name>

迁移前请完整阅读 README 的 lease 与 pending experiment 说明。

English

v1.0.2 is the release-hardened complete Skill-first v1. It adds a canonical installable Agent Skill, digest-bound evidence, recoverable runtime transactions, concurrency-safe leases, trusted Runner accounting, benchmark-gated strategy evolution, immutable human approval/rejection, verified v1.0.1 migration, and clean Codex/Claude distribution.

“Evolution” means external task-strategy optimization against matched benchmarks. It does not train model weights, guarantee every candidate improves, provide an OS sandbox, or authorize high-risk external effects.

See the English README for installation, upgrade details, safety boundaries, and the complete changelog.

Loop-Engineering v1.0.1

Loop-Engineering v1.0.1

This is the release-ready patch of the complete Loop-Engineering v1 Agent Skill.

What v1 includes

• one canonical Skill for Codex and Claude Code
• progressive references, self-contained assets, scripts, and behavior evals
• Codex and Claude plugin/marketplace manifests
• portable loop.yaml protocol with fail-closed schema and safety validation
• independent evaluator artifacts bound by context ID and SHA-256
• durable state, retries, budgets, pause/resume, leases, and audit logs
• matched-benchmark task-strategy experiments with mechanical scoring
• digest-bound human approval, version archives, and monotonic rollback
• Chinese-first README with a complete English counterpart

v1.0.1 fixes

• use strict Claude marketplace manifest loading so a clean Claude Code installation reports the plugin enabled
• accept the official GitHub CLI source metadata added by both regular and --pin Skill installations
• add doctor and regression checks for both compatibility boundaries
• synchronize package, plugin, documentation, pinned runtime, and schema references at 1.0.1

Install

gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.1 --agent codex
gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.1 --agent claude-code

See the README for Codex plugin, Claude plugin, source, and optional npm installation paths.

Validation

• 43 automated tests
• CI on Node.js 20, 22, and 24 for PR, main, and tag
• Cubic AI review: no issues found on the final compatibility patch
• clean Codex home: plugin installed, enabled
• clean Claude Code home: plugin enabled
• clean GitHub Skill installs, including --pin, validate successfully
• generated npm tarball installs and completes a real dry-run from an empty directory
• official Skill, Codex plugin, Claude plugin, and GitHub Skill validation
• npm pack/publish dry-runs and production dependency audit with 0 vulnerabilities

Important boundaries

“Evolution” means benchmark-gated optimization of task instructions. It does not update model weights, weaken the immutable loop contract, or guarantee improvement. Merge, deployment, deletion, spending, external messaging, and permission changes remain human-only. The local Runner is not an operating-system sandbox, and the generated GitHub Actions workflow is a read-only preflight scaffold until an operator adds a real executor and durable state channel.

v1.0.0

Loop-Engineering v1.0.0

Superseded by v1.0.1, which fixes Claude Code marketplace loading and GitHub CLI pinned-install metadata compatibility. Use v1.0.1 for new installations.

Loop-Engineering is now a complete, installable Agent Skill with an evidence-bound local reference runtime.

Highlights

• one canonical Skill for Codex and Claude Code
• progressive references, self-contained assets, scripts, and behavior evals
• Codex and Claude plugin/marketplace manifests
• portable loop.yaml protocol with fail-closed schema and safety validation
• independent evaluator artifacts bound by context ID and SHA-256
• durable state, retries, budgets, pause/resume, leases, and audit logs
• matched-benchmark task-strategy experiments with mechanical scoring
• digest-bound human approval, version archives, and monotonic rollback
• Chinese-first README with a complete English counterpart

Install

gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.0 --agent codex
gh skill install sheepxux/Loop-Engineering loop-engineering@v1.0.0 --agent claude-code

See the README for Codex plugin, Claude plugin, source, and optional npm installation paths.

Important boundaries

“Evolution” means benchmark-gated optimization of task instructions. It does not update model weights, weaken the immutable loop contract, or guarantee improvement. Merge, deployment, deletion, spending, external messaging, and permission changes remain human-only. The local Runner is not an operating-system sandbox, and the generated GitHub Actions workflow is a read-only preflight scaffold until an operator adds a real executor and durable state channel.

Validation

• 42 automated tests across runtime, evidence, safety, evolution, installation, and rendering
• CI on Node.js 20, 22, and 24
• official Skill, Codex plugin, Claude plugin, and GitHub Skill validation
• clean tarball installation and README quick-start execution
• fresh-agent behavior checks for one-shot, recurring, and unsafe-loop requests
• npm production dependency audit: 0 vulnerabilities

Upgrade notes

The canonical Skill now lives at skills/loop-engineering/. Codex loop instances render to .agents/skills/<loop-name>/; Claude Code instances render to .claude/skills/<loop-name>/. The old duplicated static Codex and Claude adapter Skills were removed. Protocol schemas and evidence requirements are stricter than v0.1.0, so revalidate existing loop contracts before recording new runs.