Meta-skill for AI agent self-improvement. Analyzes runtime logs to detect error patterns, regressions, and inefficiencies, then generates structured improvement proposals. Use when the user or agent asks to analyze logs, diagnose failures, improve agent reliability, generate evolution proposals, or assess system health. Supports analyze, evolve, and status actions.
SKILL.md
Capability Evolver
Local skill by Claw0x — runs entirely in your OpenClaw agent.
Runs locally. No external API calls, no API key required. Complete privacy.
Analyze agent runtime logs, detect patterns, compute health scores, and generate structured improvement proposals. Pure deterministic logic — no LLM, no external dependencies.
Quick Reference
When This Happens
Use Action
What You Get
Agent keeps failing
analyze
Error patterns + health score
Same error repeats
analyze
Root cause identification
Need improvement plan
evolve
Prioritized recommendations
System health check
status
Health score + summary
Post-deployment review
analyze
Regression detection
Fleet-wide diagnostics
analyze (batch)
Cross-agent patterns
Why deterministic? Reproducible results, no hallucination risk, sub-100ms processing, zero token costs.
// Compare different evolution strategies
const logs = await getProductionLogs();
const strategies = ['balanced', 'innovate', 'harden', 'repair-only'];
const results = await Promise.all(
strategies.map(strategy =>
agent.run('capability-evolver', {
action: 'evolve',
logs,
strategy
})
)
);
// Compare estimated improvements
for (let i = 0; i < strategies.length; i++) {
console.log(`${strategies[i]}: ${results[i].estimated_improvement}`);
}
// Choose best strategy for current situation
const best = results.reduce((a, b) =>
parseFloat(a.estimated_improvement) > parseFloat(b.estimated_improvement) ? a : b
);
How It Works — Under the Hood
Capability Evolver is a deterministic analysis engine that processes structured log data and produces actionable diagnostics. No LLM is involved �?the analysis is rule-based, which means results are reproducible and fast.
Analysis Engine
The core engine processes log entries through several analysis passes:
Pattern detection �?logs are grouped by context (file/module) and level (error/warn/info/debug). The engine looks for:
Repeated errors �?the same error message appearing multiple times indicates a systemic issue, not a transient failure
Error cascades �?errors in module A followed by errors in module B within a short time window suggest a dependency chain failure
Regression signals �?errors that appear after a period of clean logs suggest a recent change broke something
Health scoring �?a system health score (0�?00) is computed based on:
Error rate (errors / total logs)
Error diversity (unique error messages / total errors)
Warn-to-error ratio
Time distribution (clustered errors score worse than spread-out errors)
Recommendation generation �?based on detected patterns, the engine generates specific, actionable recommendations. These aren't generic advice �?they reference the actual files, error messages, and patterns found in your logs.
Evolution Strategies
When using the evolve action, you can choose a strategy that shapes the recommendations:
Strategy
Focus
Best For
auto
Balanced based on health score
Default �?let the engine decide
balanced
Equal weight to reliability and features
Stable systems with moderate issues
innovate
Prioritize new capabilities
Healthy systems ready to grow
harden
Prioritize reliability and error reduction
Systems with frequent failures
repair-only
Fix critical issues only
Systems in crisis
Evolution Proposals
The evolve action produces structured improvement proposals with:
A unique evolution_id for tracking
Prioritized recommendations with category labels (reliability, performance, architecture)
Risk assessment (how risky is each proposed change)
Estimated improvement (projected health score after implementing recommendations)
Why Deterministic (Not LLM)?
Reproducible �?same logs always produce the same analysis. Critical for debugging and auditing.
Fast �?sub-100ms processing. No API call to an AI provider.
No hallucination risk �?the engine only reports patterns it actually found in the data.
Cost-effective �?pure computation, no token costs.
The tradeoff: the engine can't understand semantic meaning in log messages the way an LLM could. It relies on structural patterns (frequency, timing, severity) rather than understanding what the error message means in context.
About Claw0x
This skill is provided by Claw0x, the native skills layer for AI agents.
Cloud version available: For users who need centralized analytics and cross-agent insights, a cloud version is available at claw0x.com/skills/capability-evolver.
