KIRA

                          ●      ●

                        ▄████◣◢████▄
                     ◥▄████▀▀  ▀▀████▄◤
                        ▀▀        ▀▀

 _____ ___ ___ __  __ ___ _  _ _   _ ___     _  _____ ___    _
|_   _| __| _ \  \/  |_ _| \| | | | / __|___| |/ /_ _| _ \  /_\
  | | | _||   / |\/| || || .` | |_| \__ \___| ' < | ||   / / _ \
  |_| |___|_|_\_|  |_|___|_|\_|\___/|___/   |_|\_\___|_|_\/_/ \_\

A smarter agent harness for Terminal-Bench, built on Terminus 2

Simple fixes, significant gains.

What is Terminus-KIRA?

Terminus-KIRA is an agent harness for Terminal-Bench, built on top of Terminus 2. It boosts frontier model performance on Terminal-Bench through a set of minimal but effective harness-level improvements — native tool calling, multimodal support, execution optimization, and smarter completion verification.

Key Features

• Native Tool Calling — Replaces ICL JSON/XML parsing with the LLM tools parameter for structured, reliable outputs
• Image Analysis (Multimodal) — image_read tool for base64-encoded image analysis directly from the terminal
• Marker-based Polling — Early command completion detection using echo markers, cutting unnecessary wait time
• Smart Completion Verification — Double-confirmation checklist covering requirements, robustness, and multi-perspective QA (test engineer, QA engineer, user)
• Prompt Caching — Anthropic ephemeral caching on recent messages to reduce latency and cost

Architecture

Terminus-KIRA extends Terminus 2 by replacing its ICL (In-Context Learning) response parsing with native LLM tool calling.

Tool definitions passed via the tools parameter:

How it works:

1. Calls litellm.acompletion directly with tools=TOOLS, bypassing the base Chat class to access native tool calling
2. The model returns structured tool calls instead of free-form text — no regex/JSON parsing needed
3. On context window overflow, automatically summarizes conversation history and retries
4. Marker-based polling appends echo '__CMDEND____' after each command; if the marker appears before the requested duration, execution moves on immediately

Evolution

Key milestones from development history:

Usage

uv run harbor run \
    --dataset terminal-bench-sample@2.0 \
    --n-tasks 1 \
    --agent-import-path "terminus_kira.terminus_kira:TerminusKira" \
    --model anthropic/claude-opus-4-6 \
    --env docker \
    -n 1

For more details, visit our blog post.

Project Structure

├── terminus_kira/
│   ├── __init__.py
│   └── terminus_kira.py        # Main agent (native tool calling)
├── prompt-templates/
│   └── terminus-kira.txt        # System prompt
├── run-scripts/
│   ├── run_docker.sh            # Local Docker execution
│   ├── run_daytona.sh           # Daytona cloud execution
│   └── run_runloop.sh           # Runloop cloud execution
├── anthropic_caching.py         # Prompt caching utility
└── pyproject.toml

Citing Us

If you found Terminus-KIRA useful, please cite us as:

@misc{terminuskira2026,
      title={Terminus-KIRA: Boosting Frontier Model Performance on Terminal-Bench with Minimal Harness },
      author={{KRAFTON AI} and {Ludo Robotics}},
      year={2026},
      url={https://github.com/krafton-ai/kira},
}

[!NOTE] TerminalBench Submission Correspondence

The agent used for the official TerminalBench submission was logged under the internal path kraftonai-terminus/terminal_bench/agents/terminus_3/terminus_3_3.py. This corresponds to terminus_kira/terminus_kira.py in this repository — the file name and path were cleaned up during open-source preparation.

Changelog

KRAFTON AI & Ludo Robotics