Release Loop-Engineering v1.0.0

[sheepxux]

What changed

• rebuild Loop-Engineering as a canonical, installable Agent Skill for Codex and Claude Code
• add progressive references, templates, behavior evals, plugin and marketplace manifests
• harden evidence binding, budgets, leases, state persistence, human gates, strategy promotion, and rollback
• add cross-platform Skill installation and validation commands
• rewrite Chinese-first and English documentation for the v1.0.0 release
• expand CI to Node 20, 22, and 24, including tag builds

Why

The previous release described a loop protocol, but it was not yet packaged and governed as a complete advanced Skill. This release establishes one canonical Skill source and a fail-closed reference runtime for bounded, auditable task-strategy evolution.

Validation

• 42 automated tests pass
• official Skill, Codex plugin, Claude plugin, and GitHub Skill dry-run validation pass
• npm pack and publish dry-runs pass
• generated tarball installs and executes from an empty directory
• README quick-start commands execute a real dry-run
• fresh-agent one-shot, recurring-loop, and unsafe-loop behavior checks pass
• production dependency audit reports 0 vulnerabilities

Boundaries

Strategy evolution changes task instructions only. It does not train model weights, weaken the loop contract, auto-merge, or auto-deploy. The local runner is not an operating-system sandbox, and the GitHub Actions output remains a read-only preflight scaffold.

---

Summary by cubic

Release v1.0.0. Packages Loop-Engineering as a single installable Skill with Codex and Claude Code plugins, a hardened local runtime, and matched-benchmark strategy evolution with rollback.

• New Features

  • Canonical Skill under skills/loop-engineering with shared docs, templates, and evals.
  • Platform plugins: .codex-plugin/plugin.json, .claude-plugin/plugin.json, and marketplace manifests.
  • CLI: skill install/validate flows, asset build, stronger doctor checks, and atomic writes.
  • Runtime: lease tokens and heartbeats, clearer dry-run status (does not spend budget), and fail-closed evidence binding.
  • Protocol: new approval.schema.json; evaluator identity now required; run-log includes budget; typed, unique safety gates.
  • Rendering: GitHub Actions renderer renamed to github-actions-scaffold; adapters rebuilt from the canonical source.
  • CI: modernized GitHub Actions (v6), Node 20/22/24 matrix, tag builds, explicit permissions, deterministic asset build with drift check, and production dependency audit.

• Migration

  • Update safety gate IDs to typed, kebab-case actions (e.g., merge-pull-request, deploy-production, open-pull-request).
  • Include evaluator identity in evaluator-result.json; add budget (and finishedAt) to run-log.json.
  • Pin loopctl to a release tag instead of using a floating version.
  • Replace any per-adapter SKILL copies with the canonical Skill; render loop-instance skills under project directories.
  • Use github-actions-scaffold in place of the old github-actions renderer.
  • Re-validate loops and examples with loopctl validate to target v1.0.0 schemas.

^{Written for commit 33d7306. Summary will update on new commits.}