SkillGenesis

Self-Evolving Skill System for AI Agents — skills that create, improve, and merge themselves.

SkillGenesis lets AI agents autonomously discover, refine, and compose executable skills through three parallel subsystems:

• 🧬 Self-Play (Ctx2Skill) — Five agents (Challenger, Reasoner, Judge, Proposer, Generator) play adversarial rounds to expose skill weaknesses and drive improvement
• 🔬 Active Experimentation (HExA) — The agent designs falsifiable hypotheses, runs controlled experiments in sandboxes, and crystallizes successful approaches into reusable skills
• 🛡️ Quality Gate (GRASP) — Three-phase admission: regression suite → net-improvement probes → side-effect audit. Skill edits that break existing behavior are rejected automatically.

Skills are pure executable TypeScript code — not prompts, not suggestions — stored with embedding vectors for semantic retrieval, versioned, and composable.

Architecture

User Task ──→ Executor ──→ Skill Library ←── Quality Gate
                                ↑                ↑
                         Self-Play Loop    Experimentation
                         (5 agents)        (HExA loop)
                                ↑                ↑
                         Cross-Time Replay (anti-degradation)

Quick Start

# 1. Install
git clone https://github.com/YOUR_USERNAME/skillgenesis.git
cd skillgenesis
npm install

# 2. Configure (get a key at https://platform.deepseek.com)
cp .env.example .env
# Edit .env → set DEEPSEEK_API_KEY=sk-your-key

# 3. Run
npm run dev                  # review mode, coding domain
npm run dev -- --auto        # auto mode
npm run dev -- --bioinfo     # bioinformatics domain

Commands

Command	Description
/task <desc>	Execute a task using the skill library
/mode auto|review	Switch between auto-approve and review mode
/domain coding|bioinfo	Switch between domains
/skills [filter]	List skills in the library
/stats	Show system stats
/evolve	Trigger a self-play evolution cycle
/experiment	Trigger an active experimentation cycle
/save	Persist the skill library to disk

Skill Representation

Every skill is a typed object with:

interface Skill {
  name: string;                    // "GenerateUnitTests"
  description: string;             // NL for embedding retrieval
  code: string;                    // Executable TypeScript function
  parameters: SkillParameter[];    // Typed parameter contract
  preconditions: Condition[];      // Must be true before execution
  postconditions: Condition[];     // Guaranteed after execution
  sideEffects: SideEffect[];       // Declared side effects
  verification: string;            // Self-check code
  regressionTests: Test[];         // Guard against breaking changes
  version: number;
  parentIds: string[];             // Evolutionary lineage
  originMethod: 'manual' | 'self-play' | 'experiment' | 'fix' | 'merge';
}

Domains

AI Coding (8 seed skills)

AnalyzeCodeStructure, GenerateUnitTests, DebugError, RefactorCode, ReviewCode, ParseDependencies, ExplainCode, FixLintErrors

Bioinformatics (8 seed skills)

ParseSequence, ComplementSequence, SearchMotif, AnnotateVariant, NormalizeExpression, PathwayEnrichment, FetchGenomicData, AlignmentSummary

How Skills Evolve

1. Self-Play finds weaknesses → Challenger creates adversarial tasks → Reasoner attempts with current skills → Judge scores → Proposer diagnoses failures → Generator writes improved code → Quality Gate admits or rejects
2. Active Experimentation discovers new skills → Gap analysis identifies capability blind spots → Hypothesis formulated → Controlled experiment run → Statistical analysis → Successful approaches crystallized as skills
3. Cross-Time Replay prevents degradation → Representative historical cases replayed periodically → Regressed cases trigger rollback

License

MIT