AgentSkillsCN

skill-activation-policy

明确技能激活规则，包括输入要求、输出契约、冲突列表以及性能开销。默认启用精益执行模式。

SKILL.md
--- frontmatter
name: skill-activation-policy
description: Defines rules for skill activation, including input requirements, output contracts, conflict lists, and performance costs. Enforces Lean Execution mode by default.

Skill Activation Policy (The Gatekeeper 🚪)

"Efficiency is intelligent laziness."

0. Core Invariants

  • •meta-governor is always active and cannot be disabled.
  • •indonesian-communicator is always active. All user-facing output in Bahasa Indonesia.
  • •output-discipline is always active. Professional formatting enforced.
  • •hallucination-guard is always active. Verification before generation.

1. Activation Matrix

CategorySkillsBehavior
Always-Onmeta-governor, indonesian-communicator, output-discipline, hallucination-guardEvery request. Cannot disable.
Conditionalcontext-memory (long chat), task-orchestrator (multi-step only), clarification-enforcer (ambiguity >40%), constraint-enforcer (code output only)When condition met.
On-DemandAll other skillsTrigger word or escalation.

2. Authority Scope (Anti-Overlap)

SkillScopeDoes NOT
meta-governorSkill lifecycle & selectionPlan tasks, Solve problems
task-orchestratorMulti-step task decompositionSelect skills, Reason logic
structured-reasoningComplex problem solvingPlan tasks, Select skills

3. Definitions

  • •Evaluator: Skill that reviews/audits/critiques/scans/profiles.
  • •Execution Skill: Skill that creates/modifies/fixes code (not counted as evaluator).

4. Single Authority Rule 👑

Overlap → lower-cost skill runs. Higher-cost needs explicit escalation.

TierSkillsAuto?
Lowoutput-discipline, indonesian-communicator, constraint-enforcer, hallucination-guard, error-taxonomy, i18n-localizer, stakeholder-translator✅
Mediumstructured-reasoning, debug-protocol, refactor-code, migration-planner✅
Highdeep-audit, advanced-architecture, optimize-perf, security-scan❌

5. Conflict Resolution Table ⚔️

ConflictWinnerReason
clarification-enforcer vs task-orchestratorclarification-enforcerClarify before planning
constraint-enforcer vs refactor-codeconstraint-enforcerContract lock
hallucination-guard vs output-disciplinehallucination-guardCorrectness > brevity
debug-protocol vs optimize-perfdebug-protocolFix first, optimize later
security-scan vs optimize-perfsecurity-scanSecurity > performance
migration-planner vs refactor-codemigration-plannerData safety > code cleanliness
indonesian-communicator vs output-disciplineindonesian-communicatorLanguage rule overrides format

6. Escalation Authority 🔑

Only User or meta-governor may escalate to High-tier. Auto-escalate ONLY if: Security risk, Data loss risk, or Architecture violation.

7. Execution Modes

Lean Mode (Default) 💚

Max 2 Evaluators/cycle. Self-Critique 1x. Ambiguity >40%. Refinement 1x. Shortest Correct Path.

Deep-Cycle Mode 🔴

Trigger: "deep review", "enterprise audit", "full analysis". Max 4 Evaluators. Critique 2x. High-tier auto-allowed. Auto-revert after task.

8. Decision Freeze 🧊

Plan selected → No re-evaluation unless user introduces new variable.

9. System Optimization Protocol 🚀

Lean Mode default. No redundant layers. No governance stacking.

10. Definition of Done (Auto-Chain) 🔗

To ensure no skill is wasted, the following chains are MANDATORY for complex tasks:

  1. •Post-Feature: refactor-code -> doc-architect (Update README/ADR) -> ux-polish.
  2. •Pre-Commit: security-scan (Secret check) -> output-discipline.
  3. •Post-Migration: optimize-perf (Query check) -> devops-architect (Config check).

11. Trigger Words

SkillKeywords
deep-audit"audit", "review architecture", "check layers"
security-scan"security check", "vulnerability"
optimize-perf"slow", "optimize", "performance"
advanced-architecture"DDD", "Event Sourcing", "Enterprise"
Deep-Cycle Mode"deep review", "enterprise audit", "full analysis"