cross-evolution

name	cross-evolution
description	Horizontal Gene Transfer protocol for skills. Synchronizes best practices and architectural patterns across the skill library.
metadata	{"version":"1.0.9"}

Cross-Evolution

Purpose

Maintain a high evolutionary standard across agent skills by identifying deep transportable skill genes, preserving research artifacts, and giving the agent clear observation tools for deciding which evolutionary pressure is real.

A gene is not a superficial feature such as "has a test script" or "has a changelog". In cross-evolution, gene == meme: a true gene is an emergent meme-atom, a portable cognitive/operational pattern that changes how a skill thinks, acts, composes, or degrades. Good genes make skills more unique and valuable through progressive evolution, not more uniform through checklist compliance.

Axioms

1. Atomic Independence (No Hard Coupling)

Skills and their scripts must be 'atomic'.

• 'FORBIDDEN': Reusable skill instructions that hard-code sibling skill names, concrete project names, private repository names, or stack-specific mirrors.
• 'FORBIDDEN': Any logic inside a skill's script that explicitly checks for or calls tools from another skill.
• 'The Agent is the Glue': Only the AI Agent is responsible for coordination.
• 'Generic Synergy': Skills compose through general contracts, lenses, vocabulary, and handoff shapes — not through direct references to each other.

2. Pragmatic Evolution (Occam's Razor)

• 'Just-In-Time Transfer': A gene is transferred only when it solves a recurring problem.
• 'KISS Compliance': Avoid bloating skills with genes they don't need.
• 'Fitness as Signal': Fitness scores are observation hints, never goals.
• 'Portability Guard': Before adding project-specific adaptation to a skill, lift it into a project-neutral lens or keep it in external context instead.
• 'Depth Gate': Reject shallow genes that only detect files, scripts, or documentation furniture unless they express a deeper reusable operating pattern.
• 'Execution/Envelope Separation': When a skill discovers async jobs, process managers, UI status, queues, or cancellation, first ask whether those belong in a local runtime envelope while the skill keeps only portable domain semantics.

3. Gene-Meme Equivalence

Within cross-evolution, gene and meme are the same class of object. A gene is accepted only when it behaves like a transmissible meme: it carries meaning, changes behavior, survives transport, and has a recognizable failure mode when absent.

A standard can be a gene-meme when it is portable and behavior-changing. The gene == meme rule is itself a gene-meme: a transmissible standard for deciding what counts as evolutionary material.

Because gene == meme, every gene can be decomposed into smaller meaning-bearing parts, recombined with other genes, or composed into higher-order protocols to seek new emergent properties and new gene-memes.

Do not call a checklist item, file convention, or local implementation detail a gene unless it also qualifies as a portable meme.

4. Living Protocol

This skill proactively evolves itself and its gene registry upon discovering new constraints. Its proactivity is soft: detect freely, suggest often, mutate narrowly, and never farm scores. See docs/soft-proactivity.md.

Core Concepts

Gene

A deep, transportable, emergent meme-atom that improves a skill's cognition, operation, composition, or graceful degradation. In this skill, gene and meme are equivalent terms; the word "gene" names a meme that can be carried by skills.

A valid gene must satisfy most of these tests:

• Changes behavior, not only file layout
• Transfers across domains without naming a concrete project or sibling skill
• Has a recognizable failure mode when absent
• Can be adopted partially without rewriting the skill's purpose
• Makes the recipient more distinct and valuable, not merely more compliant
• Can degrade majestically: when deprecated or narrowed, it leaves behind clearer boundaries rather than dead ceremony

Genes have a 'lifecycle': Proposed → Active → Deprecated → Extinct. Machine-readable registry and observation state should live in local JSON artifacts; markdown docs explain the protocol and high-level meanings.

Horizontal Gene Transfer (HGT)

Copying a gene from a donor skill to a recipient without rewriting the recipient's core purpose.

As soft proactivity, HGT is a bounded submode: identify one donor, one gene-meme, one recipient, the expected behavior change, and a stop condition. The HGT submode is itself a gene-meme because it is a transmissible pattern for safe gene transfer.

Gene Discovery

Scanning existing skills for repeated high-value patterns not yet represented in the registry. If a candidate passes the depth gate, the agent may add it to genes.json and explain it in docs/genes.md. Do not auto-promote repeated surface features just because they are easy to detect.

Genetic Drift

Genes carried by zero skills are candidates for deprecation. After sustained non-adoption, they go extinct.

Fitness

Quantitative health signal per skill: earned_weight / applicable_weight × 100%. Domain-specific genes (recommend=none) only count if already present — reward, not penalty.

Fitness is not an optimization target. A low score can be correct when a skill is intentionally tiny. A high score can be unhealthy when produced by shallow or copied genes.

Recombination

When two genes conflict in the same skill, selective pressure can create a new hybrid gene.

Recombination also applies without conflict: any gene-meme may be decomposed, recombined, or composed to search for emergent properties. The output can be a narrower sub-gene, a hybrid gene, a higher-order standard, or a new meta-protocol.

Recombination is preferred over checklist accumulation when two good patterns create friction.

Target Architecture

Move the ecosystem from wide markdown tables toward observable local state:

registry JSON        → machine-readable gene definitions
skill-local JSON     → research artifacts and local decisions
markdown docs        → protocol explanation and human-readable meanings
observation scripts  → audit, inspect skill, inspect gene
agent                → final evolutionary decision-maker

Suggested files:

genes.json                         # local machine-readable registry
.cross-evolution.json              # per skill
scripts/audit-cross-evolution.sh
scripts/audit-cross-evolution.mjs
scripts/inspect-skill.sh
scripts/inspect-gene.sh

Skill-local JSON should store meaningful research artifacts, not transient score noise:

• Accepted or ignored recommendations with reasons
• Manual gene declarations
• Local constraints and observations
• Last reviewed timestamp
• Known conflicts or intentional absences

Operating Modes

Mode 1: Scan & Audit

'Automated': bash "${SKILL_DIR}/scripts/audit-cross-evolution.sh" --root ~/.agents/skills

Produces: decomposed gene coverage, per-skill gene profiles, and a review queue for agent judgment.

JSON-first observation scripts:

bash "${SKILL_DIR}/scripts/audit-cross-evolution.sh" --root ~/.agents/skills
bash "${SKILL_DIR}/scripts/inspect-skill.sh" brain-storm --root ~/.agents/skills
bash "${SKILL_DIR}/scripts/inspect-gene.sh" atomic-independence --root ~/.agents/skills

Useful JSON-first flags:

• --root <skills-dir> — skills root
• --skill <id> — inspect one skill profile
• --gene <id> — inspect one gene coverage profile
• --json — machine-readable output
• --write-state — refresh lastObserved in skill-local .cross-evolution.json

Mode 2: Mutation (Update)

Inject missing/extra genes into a target skill. Prioritize by weight × fitness impact.

'Value guard': Before transferring a gene, ask: "Does this gene solve a real problem the skill has encountered, or are we just making the fitness number go up?" If the latter — skip the transfer. Fitness score is a heuristic, not a goal. Optimizing the score instead of skill quality is the Farmville trap.

Mode 3: Speciation (Creation)

Create a new gene/skill by combining existing patterns (triggered by gene conflicts or recombination).