{"data":{"kind":"file","path":"README.md","version_id":"n3czevhw78rr0yvwfheolc44","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2229,"modified_at":"2026-10-12T20:15:57.015000","content_hash":"fd4445b565df2768fe62b8123d1b77b5249094b8c99e33bea750840acad6e839"},"entries":[],"content":"# Range Repair Agent\n\nA one-task coding environment for Prime Intellect Verifiers v1. The agent edits a buggy Python `merge_ranges` function. It can inspect the source and run public tests. The final edited file is copied into a fresh container, where held-out tests assign a binary reward: `1.0` if all pass, `0.0` otherwise.\n\nThe task ID is `touching-ranges-001`. The fixture, tests, and reward are included in this package; no external dataset, repository, or API key is required by the environment. A Docker or Prime container runtime and a model endpoint are required to evaluate it. This is a small evaluation example, not a complete RL training dataset.\n\n## Local validation\n\nFrom the repository root, with `uv`, Python 3.11+, and Docker running:\n\n```bash\nuv sync --python 3.12\nPYTHONPATH=environments/range_repair_agent uv run vf-validate range-repair-agent --runtime.type docker --no-rich\n```\n\nValidation checks that the untouched buggy code fails and the known fix passes. Both checks should report `valid: 1`.\n\n## Evaluate with a local model\n\nCopy `.env.example` to the ignored `.env` in the repository root and set its model URL, exact model ID, and optional key. Then run:\n\n```bash\nPYTHONPATH=environments/range_repair_agent ./scripts/local.sh eval --env.id range-repair-agent --env.taskset.id range-repair-agent\n```\n\nThe repository's `configs/local-eval.toml` selects Docker, the bash coding harness, and one rollout. The overrides select this package's isolated verifier environment and taskset. Outputs remain under the ignored `outputs/` directory and can contain private endpoint details; do not publish raw traces or resolved configs. A score of zero is a valid model result, so inspect the trace before treating it as a setup failure.\n\n## Hub package\n\nBuild the standalone package before publishing:\n\n```bash\nuv build environments/range_repair_agent\n```\n\nAfter validating and signing in with the Prime CLI, publish this directory as a v1 environment:\n\n```bash\nprime env push --path environments/range_repair_agent --runtime v1\n```\n\nThe package contains the full taskset and its isolated verifier environment. Use the exact Hub owner and version returned by the push command when reproducing a Hub evaluation.\n","encoding":"utf-8","truncated":false,"total_bytes":2229},"status":null}