Conch 🐚

Headless Terminal Driver for TUI Testing & Automation

🇯🇵 日本語ドキュメントはこちら

Conch is a robust library for programmatically controlling terminal applications. By combining node-pty for process management and @xterm/headless for accurate terminal emulation, Conch enables you to:

• Test TUI Applications: Write integration tests for interactive CLI tools (vim, k9s, inquirer, etc.) with confidence.
• Automate Terminal Tasks: Build bots that can navigate complex terminal interfaces, wait for specific states, and extract information.

Think of it as "Playwright for Terminals".

Features

• Accurate Emulation: Uses xterm.js (headless) to maintain the exact state of the terminal screen, including cursor position, colors, and alternate buffers.
• Flakiness-Free Waits: Built-in utilities like waitForText, waitForSilence, and waitForStable help you handle asynchronous terminal output reliably without random sleep().
• Human-like Input: Simulate key presses (Enter, Esc, Ctrl+C) and typing naturally.
• Snapshot Engine: Capture the "visual" state of the terminal at any moment to verify what the user actually sees.
• Pluggable Backend: Designed to support Local PTY (default), and extensible for Docker or SSH in the future.

Using Conch as an LLM/Agent Foundation (CLI/TUI that doesn’t get stuck)

LLMs are good at deciding what to do next, but they need a reliable execution substrate for terminals:

• Observation: deterministic screen state via getSnapshot() (viewport or full scrollback)
• Action: run(), pressAndSnapshot(), typeAndSnapshot()
• Wait: waitForText / waitForStable / waitForSilence instead of fragile sleeps
• Command boundaries: optional OSC 133 Shell Integration to detect prompt/command completion and exit codes

This lets you implement a robust loop: snapshot → decide → act → wait → snapshot, even for interactive TUI apps.

import { Conch } from "@ushida_yosei/conch";

const conch = await Conch.launch({
  backend: { type: "localPty", file: process.platform === "win32" ? "powershell.exe" : "bash", env: process.env },
  cols: 100,
  rows: 30,
  timeoutMs: 30_000,
  shellIntegration: { enable: true, strict: false }, // improves run() reliability
});

try {
  // (1) Bring up a TUI
  await conch.run("htop", { strict: false }); // example; pick your app

  // (2) Agent loop: observe → decide → act
  for (let step = 0; step < 20; step++) {
    const snap = conch.getSnapshot({ range: "viewport" });
    const screen = snap.text;

    // Your LLM/tooling decides the next key(s) from screen state
    const nextKey = screen.includes("Help") ? "F1" : "ArrowDown";

    await conch.pressAndSnapshot(nextKey, { wait: { kind: "change", timeoutMs: 5_000 } });
  }
} finally {
  conch.dispose();
}

Installation

Install from npm:

npm install @ushida_yosei/conch
# or
pnpm add @ushida_yosei/conch

Quick Start

Here is a simple example that spawns a shell, executes a command, and verifies the output.

import { Conch } from '@ushida_yosei/conch';

async function main() {
  // 1. Launch (backend + spawn + session)
  const conch = await Conch.launch({
    backend: { type: 'localPty', file: 'bash', args: [], env: process.env },
    cols: 80,
    rows: 24,
    timeoutMs: 30_000,
  });

  // 2. Execute a command
  conch.execute('echo "Hello Conch"');

  // 3. Wait for the output to appear on the virtual screen
  await conch.waitForText('Hello Conch');

  // 4. Inspect the screen state
  const snapshot = conch.getSnapshot();
  console.log('--- Terminal Screen ---');
  console.log(snapshot.text);

  // Cleanup
  conch.dispose();
}

main();

Documentation

• Usage Guide (USAGE.md): Detailed examples and best practices.
• API Reference (API.md): Complete API documentation for ConchSession, LocalPty, and utilities.
• Source Docs (src/README.md): Internal architecture overview.

Roadmap

• [ ] Interaction Layer: Abstract interface for connecting external agents (MCP, WebSocket servers).
• [ ] Shell Integration: Support for OSC 133 to detect command completion events.
• [ ] Telnet/SSH Server: Built-in server to allow human intervention or monitoring of automated sessions.

License

MIT