Skip to main content

v0.1.0-beta.8

Gitmoot v0.1.0-beta.8 adds ranked SkillOpt exploration as a beta workflow for improving agent templates with broader human preference feedback before final A/B validation.

Features​

  • ranked SkillOpt review runs with explore, refine, distill, and validate modes
  • N-way review item options through repeated --option label=path artifacts
  • Markdown and GitHub feedback packets for ranked options and trait notes
  • ranked feedback import with derived pairwise preferences for optimizer-ready training packages
  • skillopt review status phase recommendations with ranking stability, pairwise preference counts, and advisory next-mode guidance
  • clean-temp-home smoke-test documentation for ranked Markdown packets, training export inspection, and optional GitHub issue or PR feedback sync
  • gitmoot skillopt train documentation for session/iteration orchestration, optimizer handoff, candidate review, explicit promote/reject decisions, and start-next gate enforcement
  • scripts/skillopt-train-smoke.sh deterministic smoke coverage with fake managed generation, fake optimizer handoff, and fake GitHub publication

Beta Limits​

  • final promotion should still use fresh A/B validation items by default
  • GitHub feedback collection still uses issues or PR comments authored by the authenticated GitHub user
  • phase recommendations are advisory and do not automatically change modes, import candidates, run optimizers, or promote templates
  • the external optimizer remains outside the main Gitmoot binary
  • train-mode smoke coverage uses fake GitHub and fake optimizer seams; real GitHub review publication still depends on authenticated gh

For current release artifacts, see GitHub releases.