@ellie-ai/agent-code-exec-plugin

Code execution plugin for Ellie agents. Execute TypeScript code in a subprocess with IPC bridge to runtime tools.

Features

• Execute TypeScript code using Bun
• Scripts can call runtime tools via tool()
• Scripts can report progress via progress()
• Correlation IDs for concurrent tool calls
• Configurable timeouts

Installation

bun add @ellie-ai/agent-code-exec-plugin

Usage

import { agentPlugin } from "@ellie-ai/agent-plugin";
import { codeExecPlugin } from "@ellie-ai/agent-code-exec-plugin";

const agent = agentPlugin({
  model: anthropic({ model: "claude-sonnet-4-20250514" }),
  systemMessage: `You can write TypeScript code. Use tool() to call other tools.`,
  toolPlugins: [
    codeExecPlugin({ timeout: 60000 }),
  ],
});

SDK Functions

Inside sandbox scripts, import from ./ellie:

import { tool, progress } from "./ellie";

// Call a runtime tool
const result = await tool("google-drive-search", { query: "invoices" });

// Report progress
await progress("Processing files...");

// Concurrent calls work correctly
const [a, b] = await Promise.all([
  tool("search", { q: "a" }),
  tool("search", { q: "b" }),
]);

console.log(result);

Configuration

codeExecPlugin({
  timeout: 30000,     // Execution timeout (default: 30000ms)
  rpcTimeout: 60000,  // RPC timeout (default: 60000ms)
})

Actions

The plugin dispatches these actions for observability:

• SANDBOX_EXEC_STARTED - Execution started
• SANDBOX_EXEC_COMPLETED - Execution completed successfully
• SANDBOX_EXEC_FAILED - Execution failed
• SANDBOX_PROGRESS - Progress message from script
• SANDBOX_TOOL_REQUESTED - Tool call initiated
• SANDBOX_TOOL_COMPLETED - Tool call completed