技能创建器
skill-creator
创建、改进和评估自定义技能,把专业流程转化为可重复使用的 AI 能力。
offline_recovery_harness【数据与研究】Execute a bounded D4RL-style offline recovery proxy with fixed-dataset training and normalized-score evidence.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Use this skill when full D4RL packages or datasets are unavailable but soft-mode recovery permits a declared mechanism-faithful proxy. The harness should still perform executable work: validate a fixed transition dataset, run a parameter update from that dataset, compute normalized scores, and save mechanism evidence.
module_plan.json containing the fast recovery target.environment/runtime_handoff.json describing package blockers and allowed sources.recovery_result.json with numeric metrics and target metadata.logs/generated_data_item.json with transition content and provenance.logs/training_trace.json with loss_before, loss_after, params_before, and params_after.offline_dataset_protocol.benchmark_property_taxonomy.d4rl_score_normalization.Run the harness script used by the recovery attempt, then run Distiller's validate_recovery_experiment.py on the attempt directory.
This harness is a reduced proxy, not a full D4RL benchmark run. It should be accepted only in soft mode after full recovery is blocked and the experiment gate passes.