Skip to main content

​
Cognitive Host Interface (CHI)

The Cognitive Host Interface (CHI) is an open standard (ADR-004) for exposing an AI agent’s internal state, intent, and memory to the host system. It transforms the agent from a “Black Box” into an “Open Box” for governance.

​
Core Pillars

​
1. Introspection

The agent must be able to answer: “What are you doing?” and “Why are you doing it?”. CHI formalizes this via the CHIAnalysis schema: 
interface CHIAnalysis {
  intent: string; // "Refactoring user login flow"
  risk_assessment: {
    level: "low" | "medium" | "high";
    factors: string[]; // ["Modifies auth code", "Deletes files"]
  };
  next_action: string; // "edit_file /src/auth.ts"
}

​
2. Output Guardrails (Vaccines)

CHI injects “vaccines” into the agent’s output stream to intercept unsafe patterns before they are executed.
  • System Prompt Leakage: Detects if the agent reveals its core instructions.
  • PII Leakage: Redacts emails, keys, and phone numbers.
  • Hallucination Check: Validates that cited files actually exist (optional).

​
3. Memory & Drift

CHI monitors the agent’s “Drift” — deviation from the original user goal.
  • Goal: “Fix the CSS bug.”
  • Action: “Deleting the database.”
  • Drift: HIGH. CHI blocks this action because it creates a semantic violation of the session goal.

​
Implementation

CHI is implemented via the ABS Kernel and enforced by the Proxy.