{"data":{"kind":"file","path":"README.md","version_id":"ar6rw61o8d54pnyy1huz8rty","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2648,"modified_at":"2026-10-23T13:42:39.666000","content_hash":"d45839809469e5e241e7f3d81fee4ebf55d556e4c14967dbf2c994d5cbed40ee"},"entries":[],"content":"# ComputerRL-AgentBench 0.1.0\n\nA small, deterministic, in-memory desktop prototype inspired by\n[THUDM/ComputerRL](https://github.com/thudm/ComputerRL). Not affiliated with\nTHUDM or Prime Intellect. Includes a Verifiers v0.1.8 multi-turn adapter.\n\n## Scope and limitations\n\nFour fixed tasks: move a file, change theme, enter a spreadsheet formula,\nand append text to a document. Both `gui` and `api` are labels for the same\nmock state mutations. No screenshots, real GUI, VM, network, or host filesystem\naccess. This is not a ComputerRL reproduction or a general computer-use benchmark.\nTasks are public, simple, and can be solved in one action. There is no hidden\nholdout, meaningful task diversity, or real-model benchmark result.\n\n## Install and run\n\nRequires Python 3.11 or 3.12.\n\n```bash\npython -m pip install -e .\npython examples/run_demo.py\npython -m pip install pytest\npython -m pytest tests/ -q\n```\n\n```python\nfrom computerrl_agentbench import ComputerAgentEnv, load_environment\n\nenv = ComputerAgentEnv(seed=7, max_steps=12)\nobservation, info = env.reset(task_id=\"settings_theme\")\nobservation, reward, terminated, truncated, info = env.step({\n    \"type\": \"gui\", \"operation\": \"set_setting\", \"key\": \"theme\", \"value\": \"dark\"\n})\n\nhub_env = load_environment(seed=7, max_steps=12)\n```\n\n## Prime / Verifiers interface\n\n`load_environment(seed=0, max_steps=12, **kwargs)` returns a `vf.MultiTurnEnv`\nwith all four tasks. The assistant emits one JSON action per turn. Observations\nand execution feedback are returned as user messages. State is independent per\nrollout. Successful completion or exhausted action budget ends the episode;\nthe final permitted action is executed before termination checks.\n\nActions use `type` (`gui` or `api`) and `operation`:\n\n- `move_file`: `source`, `destination`\n- `set_setting`: `key`, `value`\n- `set_cell`: `sheet`, `cell`, `value`\n- `append_document`: `document`, `text`\n\nThe Gym-like interface returns +1 on success, -0.01 on a valid unfinished step,\nand -0.05 on invalid input. The Hub rubric returns binary terminal success\n(1 or 0); intermediate step penalties appear in feedback but are not summed\ninto the Hub reward. Spreadsheet grading checks the literal formula string,\nnot a calculation engine. Document grading checks a suffix.\n\nTests include deterministic scripted-client rollouts through Verifiers,\nnot paid inference or real-model evaluation. Remote CI status is separate\nfrom local checks.\n\n## Licensing\n\nThe supplied archive did not declare a license. No open-source license is\ngranted by this packaging adaptation; metadata uses LicenseRef-Proprietary.\nAsk the rights holder for reuse permission.\n","encoding":"utf-8","truncated":false,"total_bytes":2648},"status":null}