DEP + Evo-Cubed: Skill Entity Architecture (SEA)

DEP + Evo-Cubed: Skill Entity Architecture (SEA)

> Deprecation note (2026-05-13): Mac3 was the Tier 2 worker host at the time this phase shipped. Mac3 has since been retired. References in this document reflect the Feb-2026 architecture and are kept for historical accuracy. The current home for Tier 2 scoring is Mac4:8100 (cognitive twin). See SOOP-2 launch memory for migration plan.

Concept

An inverted skill system where creative/philosophical skills become autonomous daemon entities.

### Current Problem
- 105+ skills in [home-path] — most are static files loaded on-demand
- Creative/philosophical skills (phi:, art:, nav:*) are powerful but interfere with normal operations
- Skills have no memory of when they've been useful before

### Proposed Architecture
1. Discord Category: Skill Entities — each skill gets its own channel
2. Each channel has the skill SKILL.md as system prompt + rolling conversation history
3. Router middleware intercepts every prompt/response in main channels
4. Relevance scoring — broadcasts to each skill-entity: Score 0-1 how relevant are you
5. Threshold activation — any skill above threshold (0.7) gets full context and generates injection
6. Injection gets appended as skill perspective before final response

### Model Strategy
- Scorer (hot path): MiniMax-3B-v0.1 local model at localhost:18080 (free, fast) for relevance classification
- Activator (cold path): Heavier model only when threshold crossed
- Each skill evolves its own activation patterns based on accumulated history

### Key Questions
- How does the router avoid latency bottlenecks?
- How do skills accumulate context without unbounded growth?
- Injection format? (prepend, sidebar, footnote?)
- How do skills learn when NOT to activate?
- What happens when multiple skills activate simultaneously?
- How does this integrate with existing Clawdbot gateway architecture?
- MiniMax local vs embedding-based scoring vs regex triggers?
- Discord channel per skill vs internal routing vs webhook-based?

### Implementation Constraints
- Must work within Clawdbot existing gateway/session architecture
- Should use MiniMax for scoring layer
- Mac1 (M4 Air), Mac3 (M1 8GB), Mac4 (M4 Mini 16GB) available
- Must not increase main session latency significantly

Task

1. Run DEP: Audit concept, score categories 0-10, gap list with severity, TIE techniques, commitment/uncertainty analysis
2. Run Evo-Cubed Stage 1 EXPLORE: 4-6 different architecture paths
3. Evo-Cubed Stage 2 COMPOUND: Build best elements into unified system
4. Evo-Cubed Stage 3 EXPAND + MASTER PLAN: Stress-test, produce phased execution checklist
5. Save to: Desktop/skill-entity-architecture/CREATIVE_EVOLUTION_SEA_v1.md

Reference skill docs:
- DEP: [home-path]
- Evo-Cubed: [home-path]