Skip to content

v4.3.0 - Learning loop and safe upgrades

Choose a tag to compare

View all tags
@vector233 vector233 released this 09 Aug 05:16
· 17 commits to main since this release
v4.3.0
ce18b2d

Highlights

  • Adds an evidence-backed Repository Learning Loop that turns verified, reusable project experience into durable repository rules and on-demand project Skills.
  • Adds read-only upgrade analysis with --upgrade-report, safe preview with --upgrade-apply --dry-run, and transactional safe reconciliation with --upgrade-apply.
  • Safely applies upstream additions, upstream-only changes, and clean text three-way merges while preserving local customizations and reporting removals or true conflicts for manual review.
  • Keeps the installed template baseline unchanged until the upgraded project has been validated and the user explicitly accepts the new baseline.
  • Adds English and Chinese upgrade guides, deterministic learning helpers, lifecycle validation, and Fresh-Agent evaluation evidence.

Compatibility

  • Existing initialized repositories should not rerun INIT.md; install the target Companion Skill and use the upgrade workflow instead.
  • The downstream core remains usable with Markdown and Git alone. Companion Skill automation is optional.
  • Codex and Claude Code model-routing overlays remain opt-in and preserve the root-session model by default.
  • No telemetry, autonomous runtime, or automatic cross-project collection is introduced.

Upgrade outline

  1. Install the target Companion Skill from this release.
  2. Run bootstrap_knowledge.py --target <repo> --upgrade-report --ref v4.3.0.
  3. Preview with --upgrade-apply --dry-run, then apply with --upgrade-apply on a clean task branch.
  4. Validate the project and explicitly update .repo-continuity-template only after accepting the result.

Validation

  • The complete distribution lifecycle validator passes, including fresh installs, upgrade scenarios, core-only operation, opt-in adapters, and payload leakage checks.
  • GitHub Actions passed for the release commit.
  • The current Companion Skill installs successfully with skills@1.5.19.
  • A real ephemeral Fresh-Agent test using gpt-5.6-sol recovered the work route, acceptance criteria, commits, validation, risk, next action, learning candidate, and project Skill while leaving the repository unchanged.

See docs/UPGRADING.md, docs/UPGRADING.zh-CN.md, and docs/evaluations/v4.3.0-fresh-agent.md for details.

Assets 2
Loading