{"data":{"kind":"file","path":"README.md","version_id":"r6chxvtr21frkyfgnujp9u3h","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1217,"modified_at":"2026-06-29T15:55:19.721000","content_hash":"679ae5022cfd7055c6330ab9a3854ef4677f30cfd29fc584d7a527c24a1f9d45"},"entries":[],"content":"# conductor-workflow\n\nConductor-RL: Multi-agent workflow orchestration environment for GRPO training.\n\n**Hub name**: `o-taisei/conductor-workflow`\n\n## What this package is\n\nPure, framework-independent core logic for the Conductor-RL verifiers environment:\n\n- **parser** -- extract and validate `\\`\\`\\`workflow` JSON blocks from model completions, with graded partial-credit `f_fmt` scoring.\n- **reward** -- tiered shaped reward: `R = w_corr*s_correct + w_fmt*f_fmt + w_exec*f_exec + w_eff*b_eff*1[correct]`.\n- **graders/** -- per-cluster verifiers:\n  - `code_exec` -- sandboxed stdin/stdout execution (subprocess + rlimit).\n  - `mcq_exact` -- letter extraction + exact match.\n  - `math_verify` -- SymPy equivalence with TinyV LLM fallback interface (stub).\n\n## Build order\n\n1. **This package** (Phase 1): pure logic, no network, no GPU, fully unit-tested.\n2. **Phase 2**: `workers.py` + `executor.py` + `load_environment()` wiring (SingleTurnEnv + Rubric + async OpenRouter calls).\n3. **Phase 3**: `prime env push` to Hub, prime-rl training config.\n\n## Development\n\n```bash\ncd environments/conductor_workflow\nuv venv && uv pip install -e \".[dev]\"\nuv run pytest -v\nuv run ruff check .\nuv run ruff format --check .\n```\n","encoding":"utf-8","truncated":false,"total_bytes":1217},"status":null}