Skip to content

Latest commit

 

History

10 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Skill Surgery · 元技能封装器

License: MIT Hermes Agent agentskills.io GitHub last commit

All skills will be encapsulated. This skill does the encapsulating. And it too, is being encapsulated.

Inject self-evolution lifecycles (on_load / on_use / on_error / on_unload) into any Hermes Agent SKILL.md — turning static instruction sets into self-improving components.

skill-surgery is not an alternative to hermes-agent-self-evolution, hermes-dojo, or curator-evolver.
Self-evolution = GEPA mutates skill parameters. Dojo = finds which skills are weak. Curator = decides whether to approve changes.
skill-surgery = opens the SKILL.md itself and implants a nervous system. It changes how a skill runs, not what it says.

What it does

Not a skill that writes articles or searches data. Its target is other SKILLs:

Input:  a plain SKILL.md (no self-evolution capability)
Output: the same SKILL.md + injected "self-evolution encapsulation" section
        + AGENTS.md dispatch discipline
        + skill_logger.db call baseline

An encapsulated skill can sense its own errors, log invocations, trigger auto-correction, and report back validation results.

How it's different

Project What it does Relation to skill-surgery
EvoSkill Generates new skills from failure trajectories It's the "midwife" — births new skills. skill-surgery is the "surgeon" — implants evolution nerves into existing ones
upskill Adds learnings.md staging area to a SKILL It mounts an external module. skill-surgery does injection — opens SKILL.md and writes the evolution chapter inside
SkillForge Q-value ranking + effectiveness tracking It's the "inspector" — tells you which skill works. skill-surgery is the "modifier" — makes bad skills improve themselves

skill-surgery uniqueness:

  1. Meta-skill — operates not on user tasks, but on other skills
  2. Injection-based encapsulation — writes back into target SKILL.md, not external decoration
  3. Recursive bootstrapping — the encapsulator must pass a "triple gate" before it can encapsulate others

Three encapsulation modes

Mode When Depth
light Simple skills (search/fetch/query) SKILL.md injection + call baseline
standard Skills with dispatch sub-agents + on_error signal recognition
full Multi-step pipelines + 30-day behavioral validation + error branch wiring + auto-rollback

Bootstrap status (Triple Gate)

Gate 1: Encapsulated at least 2 skills ✅ (6 skills done)
Gate 2: Error frequency dropped in both within 30 days ⏳ (pending)
Gate 3: Root cause analyzed → template improved ⏳ (pending)

Current: Gate 1 passed, Gates 2/3 pending.

When to use

  • ✅ 10+ skills in production with recurring error patterns
  • ✅ Need for "every SKILL has its own evolution loop"
  • ❌ Beginners with few skills (manual fixes are enough)
  • ❌ Light users with 3-5 skills

Install

# Clone to Hermes skills directory
cp -r skill-surgery/ ~/.hermes/skills/skill-surgery/

# First use — encapsulate itself
# See SKILL.md bootstrap guide for details

Quick demo

# Encapsulate a skill with light mode
hermes skills dispatch skill-surgery "encapsulate hengzong-analysis light"

# Check encapsulation status
bash lifecycle-gate.sh status hengzong-analysis

Links

License

MIT

About

Meta-skill that injects self-evolution lifecycles (on_load/on_use/on_error/on_unload) into any Hermes Agent SKILL.md — turning static instruction sets into self-improving components.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors