A structured governance framework for AI agent skills — observe, candidate, audit, review, promote, and observe again.
Sublation (Aufhebung / 扬弃) is a governance pipeline that turns agent execution experience into auditable skill improvements, without letting agents freely edit production skills.
Observation → Candidate → Audit → Review → Promotion → Observation Window
Core principles:
- Formal skills are read-only — agents cannot modify active skills directly
- All changes go through candidates — sandboxed copies with full audit trail
- Promotion requires review — cross-agent review + user approval before merging
- Post-promotion safety net — rollback, path verification, smoke test, fallback check
# Create an observation from a skill execution failure
python3 scripts/observe.py <skill-name> \
--skill-path ~/.hermes/skills/<category>/<skill-name> \
--session <session-id> \
--classification defect \
--reflection-type SKILL_DEFECT \
--step "script execution" \
--evidence "error message or observed behavior" \
--summary "What went wrong"
# Create a candidate from the observation
python3 scripts/candidate.py create <skill-name> \
--source-path ~/.hermes/skills/<category>/<skill-name> \
--candidate-type spec-patch \
--agent hermes
# Audit the candidate
python3 scripts/audit.py ~/.hermes/sublation/candidates/<skill>/<candidate-id>
# Check system health
python3 scripts/lifecycle.py health --warn-after-days 7- 22 audit checks (10 base + 12 strict) with
passed | conditional | failedresolution - Closed self-governance trail across runtime, lifecycle, audit, cross-skill, release, and hardening candidates
- Production sample candidates closed across NPL, Canghe, legal, GBrain, and briefing skills
- 32→11 skill consolidation via merge-driven sublation (-66%)
- 4 candidate types: spec-patch, script-enhance, infra-fix, tooling
- 3 promotion modes: human_patch, user_delegated_agent_patch, rollback
| Capability | Description |
|---|---|
| Lifecycle Management | 9-state lifecycle (active→closed) with health scanning |
| Cross-Skill Absorption | Donor→target absorption without donor modification |
| Merge-Driven Sublation | Multi-skill consolidation with review checklist |
| Darwin Evaluator Adapter | External evaluator integration (read-only, proposal-only) |
| Post-Promotion Safety Net | Rollback, path verification, smoke test, fallback check |
| Observation Window Policy | Mandatory production observation before closure |
| Legacy Migration | Plan-based migration from v2 manifests to v3 |
| Rights & Provenance | License tracking, expression copying audit |
Candidate manifests are internal runtime data and are not included in the public repo. See CHANGELOG.md for the version evolution timeline and RELEASE-v1.0.md for the v1.0 release report.
v1.0 — Maintenance Mode. The framework is complete. Future changes only from real skill practice exposing cracks — no feature development for its own sake.
MIT