#5 Release Loop-Engineering v1.0.2

What changed

• Completes the canonical Agent Skill with black-and-white brand assets, progressive references, version-bound eval evidence, and clean Codex/Claude installation.
• Hardens the local runtime with generation/contract/strategy digests, crash-recoverable transactions, concurrency-safe leases, trusted Runner time, immutable evaluator evidence, and fail-closed budget/retry handling.
• Adds benchmark-gated strategy evolution with exact arm/case attribution, digest-bound approval and rejection, safe migration, continuous strategy archives, and monotonic rollback.
• Updates the Chinese-default and English documentation, v1.0.2 upgrade path, changelog, package smoke test, and SHA-pinned Node 20/22/24 CI.

Why

The prior release established the Skill-first shape, but did not yet provide the integrity, migration, packaging, evaluation, and release-governance guarantees needed for a public v1-quality distribution. This release closes those gaps while keeping the claim precise: evolution improves external task strategy through matched benchmarks; it does not train model weights or weaken the safety contract.

Validation

• npm run build
• npm run smoke — 109 tests passing
• npm run package:smoke — real tarball and dual-Skill install
• npm pack --dry-run
• npm publish --dry-run --access public
• npm audit --omit=dev — 0 vulnerabilities
• official Skill quick validator
• gh skill publish --dry-run
• claude plugin validate . --strict
• clean Codex and Claude plugin installs report v1.0.2 enabled
• 20/20 repeated stale-lock concurrency stress passes

---

Summary by cubic

Release v1.0.2 hardens the Skill and local runtime with recoverable state, fail‑closed leases, and digest‑bound records. It also upgrades strategy evolution with benchmark‑gated, per‑case evidence and adds new assets, schemas, and CI/package checks.

• New Features

  • Recoverable multi-file transactions for run/state and experiment approve/reject/promote/rollback.
  • Fail‑closed, concurrency‑safe Runner leases with renew/recover; trusted dry‑run/command boundaries and overrun accounting.
  • Digest bindings for loop generation, contract, and active/archived strategies (SHA‑256); immutable run records with a 1,000‑entry ledger.
  • Evolution: trigger‑gated experiments, exact arm/case attribution, evaluator identity on every case, digest‑bound approvals, archives, and monotonic rollback.
  • Schemas hardened (new decision.schema.json, stricter IDs/unique arrays, run/budget fields); validator and examples updated.
  • Skill/UI: black‑and‑white brand assets, safer routing prompt, fresh‑session eval evidence, and updated CN/EN docs.
  • Tooling/CI: npm run package:smoke, CI runs it on Node 22, and SHA‑pinned actions/checkout and actions/setup-node.

• Migration

  • Version bump to 1.0.2 across manifests and docs; workflows pin GitHub tag v1.0.2.
  • Run loopctl migrate <loop-dir> to add generation/contract/strategy digests, ledgers, and integrity locks; do not hand‑edit state.
  • Rebuild any v1.0.1 pending experiments under a new ID; approvals must match v1.0.2 digests.