Skill Evaluation Engine

§1 Prompt Analysis

Entrada: user_prompt, active_project, available_skills[]

Algoritmo:

1. Tokenizar el prompt en keywords
2. Para cada skill disponible: a. Calcular keyword_score = matched_keywords / total_keywords * 100 b. Calcular context_score = project_type_match * 100 c. Calcular history_score = previous_activations_success_rate * 100 d. final_score = keyword_score * 0.4 + context_score * 0.3 + history_score * 0.3
3. Filtrar skills con final_score > threshold (default 30)
4. Ordenar por final_score descendente
5. Retornar top-5

Salida: Lista de skills recomendados con scores y razones

§2 Context Detection

Tipos de proyecto detectables:

• software: presencia de package.json, .sln, Cargo.toml, pom.xml
• research: presencia de experiments/, bibliography/, datasets/
• hardware: presencia de hardware/, bom.json, revisions/
• legal: presencia de legal/, deadlines.json, court-calendar.json
• healthcare: presencia de quality/, pdca/, incidents/
• nonprofit: presencia de impact/, volunteers/
• education: presencia de curricula/, classroom/

Mapping proyecto→skills:

• software → architecture-intelligence, developer-experience
• research → diagram-generation, knowledge-graph
• hardware → regulatory-compliance, cost-management
• legal → cost-management, regulatory-compliance
• healthcare → regulatory-compliance, enterprise-analytics
• nonprofit → executive-reporting, cost-management

§3 Instinct Integration

Cuando un instinto de categoría "context" tiene confianza >70%, boost el score de los skills asociados en +20 puntos.

§4 Feedback Loop

Cada activación registra:

• skill_name, timestamp, prompt_summary, user_accepted (bool)
• Si accepted → +2 al history_score futuro
• Si rejected → -3 al history_score futuro
• Registry: .opencode/skills/eval-registry.json