[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78Z6630D82WCF8MVX4ZJC63":3,"skill-same:01M78Z6630D82WCF8MVX4ZJC63":79},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":6,"ownerHtmlUrl":32,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":33,"install":34,"resources":46},"01M78Z6630D82WCF8MVX4ZJC63","simulation-and-rl",null,"Route LeRobot Gymnasium simulation, benchmark evaluation, and bounded RL or HIL-SERL configuration work with explicit dependency, asset, backend, and safety gates.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: simulation-and-rl\ndescription: \"Route LeRobot Gymnasium simulation, benchmark evaluation, and\n  bounded RL or HIL-SERL configuration work with explicit dependency, asset,\n  backend, and safety gates.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: Apache 2.0\n---\n\n# Simulation and RL\n\nUse this skill when a task mentions `env.type`, Gymnasium, PushT, LIBERO,\nLIBERO-plus, MetaWorld, RoboTwin\u002FRobotWin, VLABench, RoboCasa, RoboMME,\nIsaacLab Arena, RL, HIL-SERL, actor\u002Flearner, or a reward model. This skill\nroutes environment construction and evaluation; it does not silently turn a\nconfiguration check into a simulator rollout or a real-robot action.\n\n## Route first\n\n- Route generic policy training, processors, checkpoints, and inference to\n  `policy-training-inference`.\n- Route physical robot control, teleoperation, recording, and hardware safety\n  to `robot-control-data-collection`.\n- Route remote jobs, hosted services, gRPC service deployment, and extension\n  authoring to `extensions-and-services`.\n- Keep this skill responsible for simulation dispatch, benchmark evaluation,\n  RL data flow, and reward integration.\n\n## Inputs and gates\n\nCollect before changing a command or config:\n\n1. `env.type`, task\u002Fsuite, observation mode, action mode, episode horizon,\n   environment count, async preference, policy checkpoint, and evaluation\n   episode count.\n2. Whether the request is a **config\u002Fregistry check**, a CPU smoke, or an\n   actual rollout. A config check must not call `make_env()` or `reset()`.\n3. Platform, Python version, Gymnasium\u002FNumPy versions, simulator package,\n   asset-pack status, rendering mode, CUDA\u002FVulkan\u002FMuJoCo backend, and any\n   Hub credential or remote-code consent.\n4. For RL: actor\u002Flearner topology, `algorithm` (currently `sac`), policy\n   action\u002Fstate features, replay capacity, batch size, online\u002Foffline ratio,\n   warmup threshold, device, and whether human intervention is intended.\n\nUse `scripts\u002Fenvironment_probe.py` for dispatch-only inspection and\n`scripts\u002Frl_config_check.py` for bounded RL config validation. Both are\nnetwork-free, asset-free, non-actuating, and non-training.\n\n## Environment dispatch contract\n\n`make_env_config(type, **kwargs)` resolves the registered `EnvConfig` choice.\nThe config owns `create_envs(n_envs, use_async_envs)` and\n`get_env_processors()`. `make_env(config)` returns:\n\n```text\n{suite_name: {task_id: gymnasium.vector.VectorEnv}}\n```\n\nThe base config uses `gym_\u003Ctype>\u002F\u003Ctask>` and imports the package only if that\nGym ID is absent. Benchmark configs override creation for multi-task or\nthird-party simulators. `SyncVectorEnv` is the conservative default; async\nworkers are opt-in and are only selected for more than one environment. GPU,\nEGL, Vulkan, MuJoCo, SAPIEN, and asset initialization are not proved by config\nconstruction or registry inspection.\n\nRaw environment observations use `pixels`, nested camera dictionaries,\n`agent_pos`, `environment_state`, or `robot_state`. `preprocess_observation`\nnormalizes these to LeRobot observation keys; `env_to_policy_features` and the\nconfig `features_map` must agree with the policy. Evaluation depends on\n`task`, `task_description`, `_max_episode_steps`, and `info[\"is_success\"]`.\n\n## Benchmark routing\n\n| Trigger | Config route | Required gate |\n|---|---|---|\n| PushT | `pusht`, `PushT-v0` | `gym-pusht` and Pymunk; CPU smoke is reasonable |\n| Aloha | `aloha` | `gym-aloha`; verify simulator import before rollout |\n| LIBERO | `libero` | Linux, `hf-libero`, MuJoCo assets, render backend |\n| LIBERO-plus | `libero_plus` | Plus fork\u002Fassets; do not mix with vanilla LIBERO |\n| MetaWorld | `metaworld` | MetaWorld 3.0.0 and compatible Gymnasium |\n| RobotWin\u002FRoboTwin | `robotwin` | external RoboTwin tree, SAPIEN\u002FCuRobo assets, NVIDIA GPU |\n| VLABench | `vlabench` | external package, MuJoCo\u002Fdm-control, mesh assets, Linux |\n| RoboCasa | `robocasa` | editable external packages, MuJoCo, kitchen\u002Fobject assets |\n| RoboMME | `robomme` | isolated Linux ManiSkill\u002FSAPIEN\u002FVulkan environment |\n| IsaacLab Arena | `isaaclab_arena` | Hub access plus explicit trusted remote-code consent |\n| HIL simulation | `gym_manipulator` | `hilserl` extra, `gym_hil`, input device, usually NVIDIA GPU |\n\nUse the detailed matrix and known conflicts in\n[environment-overview.md](references\u002Fenvironment-overview.md) and\n[compatibility.md](references\u002Fcompatibility.md). Do not claim a runnable\nrollout from a present Python module alone.\n\n## Configuration decisions\n\nRead [configuration.md](references\u002Fconfiguration.md) before editing nested\nfields. Validate task names, observation\u002Faction dimensions, camera keys,\n`n_envs >= 1`, positive FPS\u002Fhorizon, and policy feature alignment. Keep\n`hard_reset=false` for LIBERO only when fixed initial states are enabled and\nreproducibility trade-offs are accepted. Use `use_async_envs=false` for the\nfirst smoke when a simulator has fragile rendering or fork behavior.\n\nHub environment strings or `HubEnvConfig.hub_path` are a separate path:\n`make_env` downloads\u002Fimports remote `env.py` only with\n`trust_remote_code=true`. Treat this as code execution and a credential\u002Fnetwork\nboundary, never as a local registry result.\n\n## Evaluation workflow\n\n1. Run `environment_probe.py --env-type ...` and save its JSON output.\n2. Resolve missing extras, version conflicts, backend, credentials, and assets;\n   stop on a required gap rather than substituting an unrelated simulator.\n3. For a CPU dispatch smoke, call `make_env` only for a known installed,\n   lightweight candidate; reset at most one bounded episode and close it.\n4. For a real benchmark evaluation, use `lerobot-eval` with the matching\n   policy feature names, `--eval.batch_size=1` first, explicit task and\n   episode count, and the benchmark's render\u002Fbackend settings.\n5. Verify success information and observation\u002Faction shapes before increasing\n   parallelism. Record whether the result is config-only, partial smoke, or\n   full rollout; do not label substitution as benchmark recovery.\n\nSee [workflows.md](references\u002Fworkflows.md) for safe command patterns and\nsimulation-specific evaluation notes.\n\n## RL and reward boundaries\n\nThe gRPC-free RL core exposes `ReplayBuffer`, `OnlineOfflineMixer`,\n`RLTrainer`, `RLAlgorithmConfig`, and SAC. `RLTrainer.training_step()` asks the\nalgorithm for data, and SAC consumes batches containing state, action, reward,\nnext state, done, and truncation flags. `online_ratio` must be in `[0, 1]`;\nSAC's `policy_config` must be populated before `make_algorithm`.\n\nHIL-SERL actor\u002Flearner mode is different from a local RL smoke: the actor\ncollects transitions and interventions, queues them, receives weights, and\nrequires a learner service; the learner fills replay, delays updates until\n`online_step_before_learning`, trains, checkpoints, and periodically pushes\nweights. It requires the `hilserl` extra and hardware or `gym_hil` simulation.\nDo not start actor\u002Flearner processes merely to validate JSON.\n\nReward configs route through `make_reward_model_config`,\n`make_reward_model`, and `make_reward_pre_post_processors`. Supported built-in\nnames are `reward_classifier`, `sarm`, `robometer`, and `topreward`. Reward\nmodels consume preprocessed tensor batches and emit per-sample rewards;\nclassifier threshold\u002Fsuccess-reward settings affect HIL transition processing.\nTransformer\u002FVLM-backed models may need model weights, credentials, network, and\nGPU memory. Zero-shot reward inference is not the same as trainable reward\nmodel fitting.\n\n## Recovery and handoff\n\nOn failure, classify it as missing package, version conflict, missing asset,\nbackend\u002Fdisplay failure, credential\u002Fremote-code refusal, policy feature\nmismatch, or actual simulator\u002Fruntime failure. Follow\n[troubleshooting.md](references\u002Ftroubleshooting.md); do not hide the class by\nswitching benchmarks. Report the exact command, status (`config`, `dispatch`,\n`CPU smoke`, `GPU rollout`, or `none`), unresolved gates, and whether any\nsubstitution was used.\n\n## Bundled tools\n\n- [environment_probe.py](scripts\u002Fenvironment_probe.py): registry\u002Fconfig and\n  package-presence inspection; never downloads, creates, resets, or closes an\n  external simulator.\n- [rl_config_check.py](scripts\u002Frl_config_check.py): bounded JSON schema\u002Fvalue\n  checks; never trains, contacts an actor\u002Flearner, downloads weights, or\n  initializes a simulator.\n","# Simulation and RL\n\nUse this skill when a task mentions `env.type`, Gymnasium, PushT, LIBERO,\nLIBERO-plus, MetaWorld, RoboTwin\u002FRobotWin, VLABench, RoboCasa, RoboMME,\nIsaacLab Arena, RL, HIL-SERL, actor\u002Flearner, or a reward model. This skill\nroutes environment construction and evaluation; it does not silently turn a\nconfiguration check into a simulator rollout or a real-robot action.\n\n## Route first\n\n- Route generic policy training, processors, checkpoints, and inference to\n  `policy-training-inference`.\n- Route physical robot control, teleoperation, recording, and hardware safety\n  to `robot-control-data-collection`.\n- Route remote jobs, hosted services, gRPC service deployment, and extension\n  authoring to `extensions-and-services`.\n- Keep this skill responsible for simulation dispatch, benchmark evaluation,\n  RL data flow, and reward integration.\n\n## Inputs and gates\n\nCollect before changing a command or config:\n\n1. `env.type`, task\u002Fsuite, observation mode, action mode, episode horizon,\n   environment count, async preference, policy checkpoint, and evaluation\n   episode count.\n2. Whether the request is a **config\u002Fregistry check**, a CPU smoke, or an\n   actual rollout. A config check must not call `make_env()` or `reset()`.\n3. Platform, Python version, Gymnasium\u002FNumPy versions, simulator package,\n   asset-pack status, rendering mode, CUDA\u002FVulkan\u002FMuJoCo backend, and any\n   Hub credential or remote-code consent.\n4. For RL: actor\u002Flearner topology, `algorithm` (currently `sac`), policy\n   action\u002Fstate features, replay capacity, batch size, online\u002Foffline ratio,\n   warmup threshold, device, and whether human intervention is intended.\n\nUse `scripts\u002Fenvironment_probe.py` for dispatch-only inspection and\n`scripts\u002Frl_config_check.py` for bounded RL config validation. Both are\nnetwork-free, asset-free, non-actuating, and non-training.\n\n## Environment dispatch contract\n\n`make_env_config(type, **kwargs)` resolves the registered `EnvConfig` choice.\nThe config owns `create_envs(n_envs, use_async_envs)` and\n`get_env_processors()`. `make_env(config)` returns:\n\n```text\n{suite_name: {task_id: gymnasium.vector.VectorEnv}}\n```\n\nThe base config uses `gym_\u003Ctype>\u002F\u003Ctask>` and imports the package only if that\nGym ID is absent. Benchmark configs override creation for multi-task or\nthird-party simulators. `SyncVectorEnv` is the conservative default; async\nworkers are opt-in and are only selected for more than one environment. GPU,\nEGL, Vulkan, MuJoCo, SAPIEN, and asset initialization are not proved by config\nconstruction or registry inspection.\n\nRaw environment observations use `pixels`, nested camera dictionaries,\n`agent_pos`, `environment_state`, or `robot_state`. `preprocess_observation`\nnormalizes these to LeRobot observation keys; `env_to_policy_features` and the\nconfig `features_map` must agree with the policy. Evaluation depends on\n`task`, `task_description`, `_max_episode_steps`, and `info[\"is_success\"]`.\n\n## Benchmark routing\n\n| Trigger | Config route | Required gate |\n|---|---|---|\n| PushT | `pusht`, `PushT-v0` | `gym-pusht` and Pymunk; CPU smoke is reasonable |\n| Aloha | `aloha` | `gym-aloha`; verify simulator import before rollout |\n| LIBERO | `libero` | Linux, `hf-libero`, MuJoCo assets, render backend |\n| LIBERO-plus | `libero_plus` | Plus fork\u002Fassets; do not mix with vanilla LIBERO |\n| MetaWorld | `metaworld` | MetaWorld 3.0.0 and compatible Gymnasium |\n| RobotWin\u002FRoboTwin | `robotwin` | external RoboTwin tree, SAPIEN\u002FCuRobo assets, NVIDIA GPU |\n| VLABench | `vlabench` | external package, MuJoCo\u002Fdm-control, mesh assets, Linux |\n| RoboCasa | `robocasa` | editable external packages, MuJoCo, kitchen\u002Fobject assets |\n| RoboMME | `robomme` | isolated Linux ManiSkill\u002FSAPIEN\u002FVulkan environment |\n| IsaacLab Arena | `isaaclab_arena` | Hub access plus explicit trusted remote-code consent |\n| HIL simulation | `gym_manipulator` | `hilserl` extra, `gym_hil`, input device, usually NVIDIA GPU |\n\nUse the detailed matrix and known conflicts in\n[environment-overview.md](references\u002Fenvironment-overview.md) and\n[compatibility.md](references\u002Fcompatibility.md). Do not claim a runnable\nrollout from a present Python module alone.\n\n## Configuration decisions\n\nRead [configuration.md](references\u002Fconfiguration.md) before editing nested\nfields. Validate task names, observation\u002Faction dimensions, camera keys,\n`n_envs >= 1`, positive FPS\u002Fhorizon, and policy feature alignment. Keep\n`hard_reset=false` for LIBERO only when fixed initial states are enabled and\nreproducibility trade-offs are accepted. Use `use_async_envs=false` for the\nfirst smoke when a simulator has fragile rendering or fork behavior.\n\nHub environment strings or `HubEnvConfig.hub_path` are a separate path:\n`make_env` downloads\u002Fimports remote `env.py` only with\n`trust_remote_code=true`. Treat this as code execution and a credential\u002Fnetwork\nboundary, never as a local registry result.\n\n## Evaluation workflow\n\n1. Run `environment_probe.py --env-type ...` and save its JSON output.\n2. Resolve missing extras, version conflicts, backend, credentials, and assets;\n   stop on a required gap rather than substituting an unrelated simulator.\n3. For a CPU dispatch smoke, call `make_env` only for a known installed,\n   lightweight candidate; reset at most one bounded episode and close it.\n4. For a real benchmark evaluation, use `lerobot-eval` with the matching\n   policy feature names, `--eval.batch_size=1` first, explicit task and\n   episode count, and the benchmark's render\u002Fbackend settings.\n5. Verify success information and observation\u002Faction shapes before increasing\n   parallelism. Record whether the result is config-only, partial smoke, or\n   full rollout; do not label substitution as benchmark recovery.\n\nSee [workflows.md](references\u002Fworkflows.md) for safe command patterns and\nsimulation-specific evaluation notes.\n\n## RL and reward boundaries\n\nThe gRPC-free RL core exposes `ReplayBuffer`, `OnlineOfflineMixer`,\n`RLTrainer`, `RLAlgorithmConfig`, and SAC. `RLTrainer.training_step()` asks the\nalgorithm for data, and SAC consumes batches containing state, action, reward,\nnext state, done, and truncation flags. `online_ratio` must be in `[0, 1]`;\nSAC's `policy_config` must be populated before `make_algorithm`.\n\nHIL-SERL actor\u002Flearner mode is different from a local RL smoke: the actor\ncollects transitions and interventions, queues them, receives weights, and\nrequires a learner service; the learner fills replay, delays updates until\n`online_step_before_learning`, trains, checkpoints, and periodically pushes\nweights. It requires the `hilserl` extra and hardware or `gym_hil` simulation.\nDo not start actor\u002Flearner processes merely to validate JSON.\n\nReward configs route through `make_reward_model_config`,\n`make_reward_model`, and `make_reward_pre_post_processors`. Supported built-in\nnames are `reward_classifier`, `sarm`, `robometer`, and `topreward`. Reward\nmodels consume preprocessed tensor batches and emit per-sample rewards;\nclassifier threshold\u002Fsuccess-reward settings affect HIL transition processing.\nTransformer\u002FVLM-backed models may need model weights, credentials, network, and\nGPU memory. Zero-shot reward inference is not the same as trainable reward\nmodel fitting.\n\n## Recovery and handoff\n\nOn failure, classify it as missing package, version conflict, missing asset,\nbackend\u002Fdisplay failure, credential\u002Fremote-code refusal, policy feature\nmismatch, or actual simulator\u002Fruntime failure. Follow\n[troubleshooting.md](references\u002Ftroubleshooting.md); do not hide the class by\nswitching benchmarks. Report the exact command, status (`config`, `dispatch`,\n`CPU smoke`, `GPU rollout`, or `none`), unresolved gates, and whether any\nsubstitution was used.\n\n## Bundled tools\n\n- [environment_probe.py](scripts\u002Fenvironment_probe.py): registry\u002Fconfig and\n  package-presence inspection; never downloads, creates, resets, or closes an\n  external simulator.\n- [rl_config_check.py](scripts\u002Frl_config_check.py): bounded JSON schema\u002Fvalue\n  checks; never trains, contacts an actor\u002Flearner, downloads weights, or\n  initializes a simulator.\n",{"name":5,"license":23,"metadata":24,"description":7,"disable-model-invocation":26},"Apache 2.0",{"disco-role":25},"operating",true,"main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","skills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":26},[36,39,42],{"tool":37,"command":38},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill simulation-and-rl",{"tool":40,"command":41},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill simulation-and-rl",{"tool":43,"command":44},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill simulation-and-rl","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002FSKILL.md",[47,53,57,61,65,69,75],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"references\u002Fcompatibility.md","REFERENCE","参考文档",6922,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Freferences\u002Fcompatibility.md",{"path":54,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":55,"sourceHtmlUrl":56},"references\u002Fconfiguration.md",9016,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Freferences\u002Fconfiguration.md",{"path":58,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":59,"sourceHtmlUrl":60},"references\u002Fenvironment-overview.md",9930,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Freferences\u002Fenvironment-overview.md",{"path":62,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":63,"sourceHtmlUrl":64},"references\u002Ftroubleshooting.md",7762,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Freferences\u002Ftroubleshooting.md",{"path":66,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":67,"sourceHtmlUrl":68},"references\u002Fworkflows.md",7871,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Freferences\u002Fworkflows.md",{"path":70,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":73,"sourceHtmlUrl":74},"scripts\u002Fenvironment_probe.py","SCRIPT","脚本",7624,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Fscripts\u002Fenvironment_probe.py",{"path":76,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":77,"sourceHtmlUrl":78},"scripts\u002Frl_config_check.py",9725,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Flerobot\u002Fsub-skills\u002Fsimulation-and-rl\u002Fscripts\u002Frl_config_check.py",[80,87,94,101,108,115],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":18},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":18},"01M799NS5APEC0N1FZ7XZVW6BN","passnet-solve","END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":18},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":18},"01M799NHH6NP7Z1RK77KGXAN6S","passnet-pattern-fusion","Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":114,"publishedAt":18},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":116,"name":117,"nameZh":6,"description":118,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":119,"categories":120,"keywords":121,"publishedAt":18},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]