@agdp/structured-cognition

MCP server that provides deterministic reasoning middleware for AI agent systems. Six tools that catch reasoning failures — drift, noise, cascading risk, principle violations — before they compound.

Tools

Tool	What it does
signal_anchor	Classifies input noise level and flags context-dependent signals
logic_sequence	Enforces Context → Retrieval → Analysis → Action sequence
mesh_simulate	Simulates downstream impact before action execution
gate_validate	Checks actions against configurable principles
bullwhip_diagnose	Detects amplification patterns across decision history
sc_pipeline	Runs all four core tools in sequence with auto-gating

Install

npm install @agdp/structured-cognition

Or directly from GitHub:

npm install jkc3080/structured-cognition-server

Usage

As MCP Server

Add to your MCP config (Claude Desktop, OpenClaw, etc.):

{
  "mcpServers": {
    "structured-cognition": {
      "command": "npx",
      "args": ["@agdp/structured-cognition"]
    }
  }
}

Run directly

npx @agdp/structured-cognition

How it works

Every tool returns dual-block output:

• Block 1: Human-readable report with status, findings, and recommendations
• Block 2: Structured JSON for downstream agent consumption

All scoring is deterministic — same input always produces same output. No LLM calls inside the tools. The server provides the scaffolding; your agent provides the reasoning.

Tool Details

signal_anchor

Input:  raw_input (string), input_type, context_window[]
Output: noise_floor (0-1), anchor_type, context_flags[]

Classifies whether input is actionable signal or noise. Detects hedging language, vague references, and missing context.

logic_sequence

Input:  isolated_signal (string), input_type, context_window[]
Output: status (pass/flag/block), confidence, sequence trace

Forces 4-step reasoning: Context → Retrieval → Analysis → Action. Blocks if steps are skipped.

mesh_simulate

Input:  recommendation (string), action_type, risk_horizon, context_window[]
Output: risk_score (0-100), impact_map, adjusted_recommendation

Maps system nodes affected by a proposed action. Scores risk across API, database, cache, cost, and storage dimensions.

gate_validate

Input:  recommendation (string), principles[], context_window[]
Output: decision (execute/escalate/block), violations[], confidence

Checks actions against user-defined principles. Supports keyword matching (with English morphology — handles deletion, expired, overwritten, etc.) and numeric thresholds.

bullwhip_diagnose

Input:  decision_log[] (entries with input, reasoning, action, outcome)
Output: bullwhip_active (bool), severity_score, amplification layers

Detects cognitive bullwhip — when small input variations produce disproportionate action swings. Analyzes volatility, flip-flop patterns, and confidence-action gaps.

sc_pipeline

Input:  raw_input (string), input_type, principles[], context_window[]
Output: 4-stage trace with auto-gating between stages

Chains SignalAnchor → LogicStack → CausalMesh → PrincipleGate. Each stage gates the next — if one blocks, the pipeline stops.

Development

git clone https://github.com/jkc3080/structured-cognition-server.git
cd structured-cognition-server
npm install
npm run build
npm start

License

MIT