{ "markdown": "
\n\n```\n ● ●\n\n ▄████◣◢████▄\n ◥▄████▀▀ ▀▀████▄◤\n ▀▀ ▀▀\n\n _____ ___ ___ __ __ ___ _ _ _ _ ___ _ _____ ___ _\n|_ _| __| _ \\ \\/ |_ _| \\| | | | / __|___| |/ /_ _| _ \\ /_\\\n | | | _|| / |\\/| || || .` | |_| \\__ \\___| ' < | || / / _ \\\n |_| |___|_|_\\_| |_|___|_|\\_|\\___/|___/ |_|\\_\\___|_|_\\/_/ \\_\\\n```\n\n
\n\n

\n A smarter agent harness for Terminal-Bench, built on Terminus 2\n
\n Simple fixes, significant gains.\n

\n\n

\n \"Codex\n \"Opus\n \"Gemini\n

\n\n## What is Terminus-KIRA?\n\nTerminus-KIRA is an agent harness for [Terminal-Bench](https://github.com/laude-institute/terminal-bench), built on top of [Terminus 2](https://github.com/laude-institute/terminal-bench). 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.\n\n---\n\n## Key Features\n\n- **Native Tool Calling** — Replaces ICL JSON/XML parsing with the LLM `tools` parameter for structured, reliable outputs\n- **Image Analysis (Multimodal)** — `image_read` tool for base64-encoded image analysis directly from the terminal\n- **Marker-based Polling** — Early command completion detection using echo markers, cutting unnecessary wait time\n- **Smart Completion Verification** — Double-confirmation checklist covering requirements, robustness, and multi-perspective QA (test engineer, QA engineer, user)\n- **Prompt Caching** — Anthropic ephemeral caching on recent messages to reduce latency and cost\n\n---\n\n## Architecture\n\nTerminus-KIRA extends Terminus 2 by replacing its ICL (In-Context Learning) response parsing with native LLM tool calling.\n\n**Tool definitions** passed via the `tools` parameter:\n\n| Tool | Purpose |\n|---|---|\n| `execute_commands` | Run shell commands with analysis and plan |\n| `task_complete` | Signal task completion (triggers double-confirmation) |\n| `image_read` | Analyze image files via base64 multimodal input |\n\n**How it works:**\n\n1. Calls `litellm.acompletion` directly with `tools=TOOLS`, bypassing the base `Chat` class to access native tool calling\n2. The model returns structured tool calls instead of free-form text — no regex/JSON parsing needed\n3. On context window overflow, automatically summarizes conversation history and retries\n4. Marker-based polling appends `echo '__CMDEND____'` after each command; if the marker appears before the requested duration, execution moves on immediately\n\n---\n\n## Evolution\n\nKey milestones from development history:\n\n| # | Milestone | Description |\n|---|---|---|\n| 1 | Genesis | Copy of Terminus 2 as starting point |\n| 2 | Native Tool Use | Replaced ICL JSON/XML parsing with LLM `tools` parameter |\n| 3 | Output Limiting | 30 KB cap on terminal output to prevent context bloat |\n| 4 | Autonomy & Constraints | Prompt engineering for agent autonomy and environment constraints |\n| 5 | Completion Confirmation | Include original instruction in completion check |\n| 6 | Multimodal | `image_read` tool for visual analysis of terminal screenshots |\n| 7 | Completion Checklist | Multi-perspective QA checklist (test engineer, QA, user) |\n| 8 | Execution Optimization | Marker-based polling and block timeout protection |\n| 9 | Temperature Fix | Set temperature to 1 when using reasoning effort |\n\n---\n\n## Usage\n\n```bash\nuv run harbor run \\\n --dataset terminal-bench-sample@2.0 \\\n --n-tasks 1 \\\n --agent-import-path \"terminus_kira.terminus_kira:TerminusKira\" \\\n --model anthropic/claude-opus-4-6 \\\n --env docker \\\n -n 1\n```\n\nFor more details, visit our [blog post](https://krafton-ai.github.io/blog/terminus_kira_en/).\n\n---\n\n## Project Structure\n\n```\n├── terminus_kira/\n│ ├── __init__.py\n│ └── terminus_kira.py # Main agent (native tool calling)\n├── prompt-templates/\n│ └── terminus-kira.txt # System prompt\n├── run-scripts/\n│ ├── run_docker.sh # Local Docker execution\n│ ├── run_daytona.sh # Daytona cloud execution\n│ └── run_runloop.sh # Runloop cloud execution\n├── anthropic_caching.py # Prompt caching utility\n└── pyproject.toml\n```\n\n---\n\n## Citing Us\n\nIf you found Terminus-KIRA useful, please cite us as:\n\n```bibtex\n@misc{terminuskira2026,\n title={Terminus-KIRA: Boosting Frontier Model Performance on Terminal-Bench with Minimal Harness },\n author={{KRAFTON AI} and {Ludo Robotics}},\n year={2026},\n url={https://github.com/krafton-ai/kira},\n}\n```\n\n> [!NOTE]\n> **TerminalBench Submission Correspondence**\n>\n> The agent used for the official TerminalBench submission was logged under the internal path\n> `kraftonai-terminus/terminal_bench/agents/terminus_3/terminus_3_3.py`.\n> This corresponds to [`terminus_kira/terminus_kira.py`](./terminus_kira/terminus_kira.py) in this repository —\n> the file name and path were cleaned up during open-source preparation.\n\n---\n\n## Changelog\n\n| Version | Description |\n|---|---|\n| **v1.1** | Migrated from In-Context Learning (ICL) to **native tool calling** via LLM `tools` parameter. Removed verbose JSON/XML response format instructions from system prompt — the model now receives structured tool definitions directly, resulting in a significantly shorter prompt and more reliable outputs. |\n| **v1.0** | Initial release. Fork of Terminus 2 with ICL-based JSON response parsing and full response format instructions in the system prompt. |\n\n---\nKRAFTON AI & Ludo Robotics\n", "bytes": 6079, "sha": "87c2396fb33e0eb783c807c94c028b64f53a34f44d0d5cc5a971552a1067569e", "repo_slug": "krafton-ai/kira", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/skl_krafton_ai_kira_claude_skills_scratch_pa_7f250de4/readme" }