AgentSkillsCN

CF Plugin Cognitive Kernel

认知核插件，通过工作记忆、注意力控制、元认知，以及脚手架功能，增强 LLM 的能力。适用于管理代理工作记忆、在长跨度上下文中控制注意力、开启自我反思式推理，或构建复杂的多步骤认知任务时使用。

SKILL.md
--- frontmatter
name: "CF Plugin Cognitive Kernel"
description: "Cognitive kernel plugin for LLM augmentation with working memory, attention control, meta-cognition, and scaffolding. Use when managing agent working memory, controlling attention across long contexts, enabling self-reflective reasoning, or structuring complex multi-step cognitive tasks."

CF Plugin Cognitive Kernel

Cognitive kernel plugin for LLM augmentation providing working memory management, attention control, meta-cognition monitoring, and cognitive scaffolding for structured multi-step reasoning in agent workflows.

Quick Command Reference

TaskCommand
Enable pluginnpx @claude-flow/cli@latest plugins toggle --enable cognitive-kernel
Disable pluginnpx @claude-flow/cli@latest plugins toggle --disable cognitive-kernel
Plugin infonpx @claude-flow/cli@latest plugins info cognitive-kernel
List toolsnpx @claude-flow/cli@latest mcp tools
Check statusnpx @claude-flow/cli@latest plugins list

Installation

Via claude-flow: Already included with npx @claude-flow/cli@latest init Standalone: npx @claude-flow/plugin-cognitive-kernel@latest

Activation

bash
# Enable the plugin
npx @claude-flow/cli@latest plugins toggle --enable cognitive-kernel

# Verify activation
npx @claude-flow/cli@latest plugins info cognitive-kernel

Plugin Capabilities

Working Memory

Provides a structured working memory buffer for agents, supporting push/pop/query operations with capacity limits and priority-based eviction.

bash
npx @claude-flow/cli@latest mcp exec cognitive-kernel.working-memory \
  --action store --key "current-context" --value "refactoring auth module" \
  --ttl 300

Attention Control

Directs agent focus to relevant portions of large contexts, reducing token waste and improving response quality on long documents or codebases.

bash
npx @claude-flow/cli@latest mcp exec cognitive-kernel.attention \
  --input context.json --focus "security vulnerabilities" --max-tokens 4000

Meta-Cognition

Self-reflective monitoring layer that tracks confidence levels, detects reasoning loops, and flags when an agent should escalate or seek help.

bash
npx @claude-flow/cli@latest mcp exec cognitive-kernel.meta-cognition \
  --agent-id agent-1 --check confidence

Scaffolding

Breaks complex cognitive tasks into structured sub-steps with checkpoints, enabling systematic reasoning through difficult problems.

bash
npx @claude-flow/cli@latest mcp exec cognitive-kernel.scaffold \
  --task "Design microservice architecture" --depth 3 --checkpoints true

Common Patterns

Enhance Agent Reasoning on Complex Tasks

bash
npx @claude-flow/cli@latest plugins toggle --enable cognitive-kernel
npx @claude-flow/cli@latest mcp exec cognitive-kernel.scaffold \
  --task "Migrate monolith to microservices" --depth 4
npx @claude-flow/cli@latest mcp exec cognitive-kernel.attention \
  --input codebase-summary.json --focus "service boundaries"

Monitor Agent Confidence and Escalate

bash
npx @claude-flow/cli@latest mcp exec cognitive-kernel.meta-cognition \
  --agent-id agent-1 --check confidence --threshold 0.7 --on-low escalate

Manage Working Memory Across Sessions

bash
# Store context before session end
npx @claude-flow/cli@latest mcp exec cognitive-kernel.working-memory \
  --action store --key "progress" --value '{"step": 3, "findings": [...]}'

# Retrieve on session restore
npx @claude-flow/cli@latest mcp exec cognitive-kernel.working-memory \
  --action retrieve --key "progress"

RAN DDD Context

Bounded Context: Coherence/Interpretability

References