@agent-monitor/agent

Action interception, security guardrails, safe tool definitions, approval management, and reference DeepSeek autonomous agent runtime.

Overview

@agent-monitor/agent sits between the AI model and the operating system. It provides the runtime interception harness, containment guardrails, safe tools, and the DeepSeek client implementation.

Key Modules

• ActionInterceptor:
  • Intercepts tool calls before execution.
  • Enforces authoritative pre- and post-approval Kill Switch circuit breaker.
  • Enforces workspace containment and symlink checks.
  • Evaluates deterministic policies and coordinates human approval.
  • Emits strictly sequenced audit events.
• ApprovalManager:
  • Manages pending approval requests.
  • Handles asynchronous resolution and expiration timeouts.
• tools/:
  • Safe tools: readFileTool, writeFileTool, listFilesTool, runCommandTool.
  • Guardrails: resolveSafeWorkspacePath (iterative URL decoding, POSIX and Windows backslash normalization, canonical symlink validation).
  • Child Process Isolation: runCommandTool sanitizes process.env to prevent accidental credential leakage to child processes.
• deepseek/:
  • DeepSeekClient: Lightweight HTTP client for DeepSeek Chat API.
  • DeepSeekCodingAgent: Reference autonomous coding agent loop.

Installation

npm install @agent-monitor/agent @agent-monitor/core

Example Usage

import {
  ActionInterceptor,
  readFileTool,
  writeFileTool,
  runCommandTool,
} from "@agent-monitor/agent";
import { PolicyEngine } from "@agent-monitor/core";

const interceptor = new ActionInterceptor({
  sink: {
    emit: async (event) => console.log("Event:", event.type),
  },
  policyEngine: new PolicyEngine(),
});

interceptor.registerTool(readFileTool);
interceptor.registerTool(writeFileTool);
interceptor.registerTool(runCommandTool);

const output = await interceptor.invoke(
  "read_file",
  { path: "package.json" },
  {
    sessionId: "ses_123",
    agentId: "my-agent",
    workspaceRoot: process.cwd(),
  },
);