terminus-ai 2.0.0

Terminus CLI: An autonomous AI agent for terminal-based task execution

Terminus

Terminus is an agent for terminal-based task execution. It integrates with Harbor for agent evaluation and task execution.

Installation

Prerequisites

• Python >=3.12
• tmux (required for terminal session management)
• Harbor framework installed

Installing tmux

Terminus requires tmux to manage terminal sessions. Install it using your system's package manager:

macOS:

brew install tmux

Ubuntu/Debian:

sudo apt-get install tmux

Fedora:

sudo dnf install tmux

Arch Linux:

sudo pacman -S tmux

Install Terminus

For local development:

pip install -e /path/to/terminus

Usage

Command Line Interface

Terminus provides a CLI for quick testing and demonstration:

# Basic usage
terminus "Create a file hello.txt with Hello World"

# With options
terminus "Create a file hello.txt" \
  --model openai/gpt-4o \
  --logs-dir ./logs \
  --parser json \
  --temperature 0.7

# Show help
terminus --help

Note:

• The CLI runs directly on your local system using tmux (no Docker required)
• For testing and simple tasks, the CLI is quick and convenient
• For production use cases and complex evaluations, integration with Harbor is recommended (see below)

Usage with Harbor

Terminus is designed to work as an external agent with Harbor. You can use it in several ways:

Option 1: Using import path (Recommended)

When configuring your Harbor task, use the import path to load the Terminus agent:

agent:
  import_path: "terminus:Terminus"
  model_name: "anthropic/claude-sonnet-4"
  kwargs:
    parser_name: "json"  # or "xml"
    temperature: 0.7
    max_turns: 100
    enable_summarize: true

Or in Python:

from harbor import AgentFactory
from pathlib import Path

agent = AgentFactory.create_agent_from_import_path(
    import_path="terminus:Terminus",
    logs_dir=Path("./logs"),
    model_name="anthropic/claude-sonnet-4",
    parser_name="json",
    temperature=0.7,
)

Option 2: Direct instantiation

from terminus import Terminus
from pathlib import Path

agent = Terminus(
    logs_dir=Path("./logs"),
    model_name="anthropic/claude-sonnet-4",
    parser_name="json",  # or "xml"
    temperature=0.7,
    max_turns=100,
    enable_summarize=True,
)

Configuration Options

• model_name: The LLM model to use (required)
• parser_name: Response format - "json" or "xml" (default: "json")
• temperature: Sampling temperature (default: 0.7)
• max_turns: Maximum number of agent turns (default: 1000000)
• enable_summarize: Enable context summarization when limits are reached (default: True)
• api_base: Custom API base URL (optional)
• collect_rollout_details: Collect detailed token-level rollout data (default: False)