eschatch 0.2.0

ESChatch is a true pty wrapper that transparently logs every byte of input/output from any terminal application (zsh, vim, python, SSH sessions, etc.), then lets you hit Ctrl+X, type a natural-language task, and have an LLM instantly generate and inject the exact keystrokes back into the running program.

The screen-scrape context makes it magically context-aware across all applications.

Features

• Universal terminal copilot - Works in shells, editors, REPLs, SSH sessions, anywhere
• Transparent logging - Logs all I/O to configurable session directories
• Context-aware - LLM sees your terminal screen and recent input history
• Configurable - Escape keys, LLM providers, models, prompts via TOML config
• Safety rails - Preview mode and destructive command detection
• Multi-provider - Supports OpenAI, Ollama, vLLM, litellm proxy, and more
• Chat mode - Multi-turn conversations with context retention
• Special commands - /explain, /debug, /chat, /clear, /help

Quick Start

Installation

# Install from source
pip install -e .

# Or with pipx (recommended)
pipx install git+https://github.com/day50/ESChatch.git

Usage

# Start with default shell (bash)
eschatch

# Use zsh
eschatch -e zsh

# Start Python REPL
eschatch -e python

# Open vim with a file
eschatch -e "vim myfile.py"

# SSH session
eschatch -e "ssh user@host"

Basic Workflow

1. Start ESChatch with your desired command
2. Work normally in the terminal
3. Press Ctrl+X to enter escape mode
4. Type your task (e.g., "find all .py files modified today")
5. Press Enter
6. The LLM generates and injects the appropriate command

Special Commands

ESChatch supports special slash commands:

• /chat - Enable multi-turn conversation mode (press Enter twice to exit)
• /explain - Explain what's happening in the current terminal state
• /debug - Analyze errors and suggest fixes
• /clear - Clear conversation history
• /help - Show available commands

Example:

Ctrl+X → /explain
→ "You're in a Python REPL with pandas loaded. The last operation filtered a DataFrame..."

Ctrl+X → /debug
→ "The error shows a FileNotFoundError. Check that 'data.csv' exists in your current directory..."

Configuration

Config File

Create ~/.config/eschatch/config.toml:

[general]
escape_key = "ctrl+x"  # Options: ctrl+x, ctrl+space, ctrl+c, ctrl+d, escape
log_dir = "~/.eschatch/logs"
session_dir = "~/.eschatch/sessions"

[llm]
provider = "litellm"  # Options: litellm, ollama, openai, vllm, simonw
model = "gpt-4o-mini"
base_url = "http://localhost:4000"
api_key = "sk-..."  # Optional for local providers
temperature = 0.0
max_tokens = 500

[context]
max_bytes = 2000  # Amount of I/O history to include
sliding_window = true

[prompt]
system = """You are an expert Linux/terminal operator..."""

[safety]
preview_mode = false  # Show commands before injecting
confirm_destructive = true  # Warn about rm -rf, dd, etc.

Environment Variables

Override config with environment variables:

export ESCHATCH_MODEL="gpt-4o"
export ESCHATCH_BASE_URL="http://localhost:4000"
export ESCHATCH_API_KEY="sk-..."

Generate Default Config

eschatch --install-config

LLM Providers

OpenAI / litellm Proxy

[llm]
provider = "litellm"
model = "gpt-4o-mini"
base_url = "http://localhost:4000"
api_key = "sk-..."

Ollama (Local)

[llm]
provider = "ollama"
model = "llama3.1"
base_url = "http://localhost:11434"

vLLM

[llm]
provider = "vllm"
model = "meta-llama/Llama-3.1-8B"
base_url = "http://localhost:8000/v1"

simonw/llm CLI

[llm]
provider = "simonw"

Examples

Shell Commands

Task: "find all Python files modified in the last 2 days"
→ find . -name "*.py" -mtime -2 -type f

Task: "show disk usage sorted by size"
→ du -sh * | sort -rh

Inside Vim

Task: "format this Python function"
→ :%!black -

Task: "search for all occurrences of 'TODO'"
→ /TODO

Python REPL

Task: "load the json file and parse it"
→ import json; data = json.load(open("file.json"))

SSH Session

Works transparently over SSH - the LLM sees the remote terminal state.

Safety Features

Preview Mode

Enable to review generated commands before injection:

eschatch --preview

Or in config:

[safety]
preview_mode = true

Destructive Command Detection

Automatically warns about dangerous patterns like:

• rm -rf / or rm -rf ~
• dd if=...
• mkfs
• Fork bombs
• Direct disk writes

Architecture

┌─────────────────────────────────────────────────────────┐
│                     User Terminal                        │
└─────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────┐
│                    ESChatch (pty)                        │
│  ┌─────────────┐  ┌──────────────┐  ┌────────────────┐  │
│  │ Input Log   │  │ Output Log   │  │ Escape Handler │  │
│  └─────────────┘  └──────────────┘  └────────────────┘  │
└─────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────┐
│                      LLM Client                          │
│  (OpenAI / Ollama / vLLM / litellm / simonw)            │
└─────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────┐
│              Injected Command → Application              │
└─────────────────────────────────────────────────────────┘

Command Line Options

usage: eschatch [-h] [--exec EXEC_COMMAND] [--config CONFIG]
                [--model MODEL] [--base-url BASE_URL] [--preview]
                [--install-config] [--verbose]

ESChatch - LLM-powered terminal command injection

options:
  -h, --help            show this help message and exit
  --exec, -e EXEC_COMMAND
                        Command to execute (default: bash)
  --config, -c CONFIG   Path to config file
  --model, -m MODEL     LLM model to use
  --base-url BASE_URL   LLM base URL
  --preview             Preview mode - show generated commands
  --install-config      Create default config file
  --verbose, -v         Enable verbose logging

Limitations

• Unix-only - Uses pty, so Linux/macOS only
• Single-threaded - One session at a time
• Raw terminal - Requires tty (won't work in some IDE terminals)
• Early alpha - Actively developed, may have bugs

Troubleshooting

LLM connection fails

• Check your base_url and api_key in config
• Ensure your LLM provider is running (Ollama, litellm proxy, etc.)
• Run with --verbose for detailed logs

Escape key not working

• Try different escape keys in config: ctrl+space, escape
• Some terminals may intercept certain key combinations

Commands not injecting correctly

• Check that the LLM understands the current context
• Review the screen scrape in verbose logs
• Try a more specific task description

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Format code
black .

# Lint
ruff check .

License

MIT License - see LICENSE

Contributing

Contributions welcome! Please read CONTRIBUTING.md first.

Related Projects

• sidechat - LLM side panel for terminals
• llm - CLI tool for LLM access
• litellm - Unified LLM API

Acknowledgments

ESChatch was inspired by the idea of making LLMs a seamless part of the terminal workflow, working at the pty level to be truly application-agnostic.