pingchesu/hermes-curator-evolver/curator-evolution

Curator Evolution

Hermes Curator Evolver starts from evidence and keeps mutation guarded. Reports and proposals are review artifacts. auto-run can make skills actually improve, but only through low-risk bounded managed blocks, reference spillover for bulky evidence, and guarded apply.

Interpretation Checklist

1. Separate evidence from conclusions.
2. Repeated tool errors suggest a possible missing pitfall or verification step in a skill.
3. Repeated skill reads suggest the skill is active and worth keeping discoverable.
4. A single failure is not enough evidence to rewrite a skill.
5. If the skill was correct but the agent ignored it, improve triggers/descriptions only when evidence repeats.
6. Candidate search is advisory; embedding/reranker models only find candidates and do not decide edits.
7. backfill-sessions reads current Hermes state.db sessions through the read-only SessionDB API and retains explicit legacy session_*.json support; it is model-free and does not mutate skills.
8. auto-run --semantic-candidates and --rerank-candidates are explicit opt-ins that only reorder evidence-eligible candidates.
9. Guarded apply requires approval, backup, verifier/validation pass, and rollback.
10. auto-run mutates only when both --apply-low-risk and --approve-auto-apply are set.
11. Even with write flags, unattended auto-apply writes only local agent-created skills. Official/bundled skills (.bundled_manifest), hub-installed skills (.hub/lock.json), plugin-provided skills, skills.external_dirs, pinned skills, and unknown sources are skipped.
12. Core-name allowlists only operate inside that local agent-created source boundary; they do not override provenance.
13. Size guardrails target a 90k SKILL.md soft cap, spill bulky autorun evidence into references/, and skip unattended writes when the target skill is already over the 100k hard cap.
14. install-auto --enable creates a native user scheduler: systemd user timer on Linux, launchd LaunchAgent on macOS; remove it with uninstall-auto before plugin uninstall.

Safe Next Actions

• Run hermes-curator-evolver bootstrap for the default one-command setup: backfill recent sessions and enable the daily safe autorun scheduler.
• Run hermes-curator-evolver bootstrap --semantic only when the user explicitly wants model-assisted scheduler candidate ordering.
• Run hermes-curator-evolver report --days 7 for evidence.
• Run hermes-curator-evolver backfill-sessions --days 30 --format json once when existing Hermes sessions should seed evidence. Use --sessions-dir only for legacy session_*.json exports.
• Run hermes-curator-evolver propose --skill <name> --format json for a dry-run proposal.
• Run hermes-curator-evolver verify --proposal-file <proposal.json> before considering apply.
• Use hermes-curator-evolver candidates --query <text> --skills-dir <dir> for dependency-free lexical candidate search.
• Use hermes-curator-evolver auto-run --skills-dir ~/.hermes/skills --format json to preview automatic improvements.
• Use hermes-curator-evolver auto-run --skills-dir ~/.hermes/skills --semantic-candidates --rerank-candidates --format json to preview model-assisted candidate ordering.
• Use hermes-curator-evolver auto-run --skills-dir ~/.hermes/skills --apply-low-risk --approve-auto-apply for actual low-risk bounded improvement; only local agent-created skills are writable, bulky evidence can spill to references/, and already-over-hard-cap skills are skipped.
• Use hermes-curator-evolver install-auto --schedule daily --enable for plug-in daily automation without Hermes core changes; portable schedules are hourly, daily, and weekly.
• Use hermes-curator-evolver install-auto --schedule daily --enable --semantic-candidates --rerank-candidates only when the user explicitly wants model-assisted scheduler candidate ordering.
• Use hermes-curator-evolver uninstall-auto to remove the optional scheduler.
• Use guarded apply only with an exact SHA256 and a reviewed content file.