[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M76X0VMPTNGCX1GHT2FBCQDH":3,"skill-same:01M76X0VMPTNGCX1GHT2FBCQDH":55},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":33,"commitSha":34,"skillDirectory":35,"forksCount":36,"repositoryHtmlUrl":37,"repositoryHomepageUrl":6,"ownerHtmlUrl":38,"ownerBlogUrl":39,"ownerDescription":6,"updatedAt":40,"install":41,"resources":54},"01M76X0VMPTNGCX1GHT2FBCQDH","ray",null,"Distributed Python compute with Ray — @ray.remote tasks\u002Factors for cluster-scale parallelism, Ray Data for large-batch preprocessing, Ray Train for distributed model training (DDP\u002FFSDP\u002FDeepSpeed), Ray Tune for scalable hyperparameter search, and Ray Serve for model serving. Use when scaling a Python workload (docking screens, million-cell atlas preprocessing, hyperparameter sweeps, multi-GPU training) from a laptop to a multi-node cluster with minimal code changes. Ray Tune can use optuna as a search algorithm; Ray Train wraps pytorch-lightning-style training loops.","stanfish06\u002Fskillquarium","stanfish06","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F93099613?v=4","01M75F1CK4ZW109ZK3T26VN52F","Zhiyuan","01M75F1CM53VH6F79DZ8GXKKES",12,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fray\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: ray\ndescription: Distributed Python compute with Ray — @ray.remote tasks\u002Factors for cluster-scale parallelism, Ray Data for large-batch preprocessing, Ray Train for distributed model training (DDP\u002FFSDP\u002FDeepSpeed), Ray Tune for scalable hyperparameter search, and Ray Serve for model serving. Use when scaling a Python workload (docking screens, million-cell atlas preprocessing, hyperparameter sweeps, multi-GPU training) from a laptop to a multi-node cluster with minimal code changes. Ray Tune can use optuna as a search algorithm; Ray Train wraps pytorch-lightning-style training loops.\nlicense: Apache-2.0\nallowed-tools:\n  - Read\n  - Write\n  - Edit\n  - Bash\ncompatibility: Requires Python 3.10+ (Ray's PyPI metadata requires >=3.10; older 3.9 wheels exist only for Ray \u003C2.10) and ray 2.x (current 2.59.0). Install the extra matching your workload — ray[default] (dashboard+autoscaler), ray[data], ray[train], ray[tune], ray[serve], ray[rllib] — rather than ray[all]. Every node in a cluster must run the same Ray version and a compatible Python minor version; mismatches fail silently or crash workers.\nmetadata: {\"version\": \"1.0\", \"skill-author\": \"community\"}\n---\n\n# Ray\n\n## Overview\n\nRay is a distributed runtime for Python: `@ray.remote` turns an ordinary function or class into a task\u002Factor that Ray schedules across all CPUs\u002FGPUs on your laptop or across every node in a cluster, with the same code in both cases. On top of this \"Ray Core,\" Ray ships purpose-built libraries:\n\n- **Ray Data** — distributed, streaming data loading\u002Ftransform (`ray.data.read_parquet(...).map_batches(...)`), good for ML preprocessing pipelines that don't fit in one machine's memory.\n- **Ray Train** — distributed training wrapper (`TorchTrainer`, `ScalingConfig`) for DDP\u002FFSDP\u002FDeepSpeed, with native PyTorch Lightning and Hugging Face `Trainer` integration.\n- **Ray Tune** — scalable hyperparameter search with pluggable search algorithms (including Optuna, HyperOpt) and schedulers (ASHA, PBT, Hyperband).\n- **Ray Serve** — programmable, autoscaling model serving (`@serve.deployment`), including an LLM-serving path (`ray.serve.llm`).\n- **RLlib** — production reinforcement learning at scale.\n\nRay Core is the substrate all of the above sit on: tasks (stateless remote function calls), actors (stateful remote classes), and objects (immutable values in a shared, cluster-wide object store).\n\n## Installation\n\n```bash\n# Recommended for ML workloads\nuv pip install -U \"ray[data,train,tune,serve]\"\n\n# General-purpose distributed Python (dashboard + cluster launcher, no ML libs)\nuv pip install -U \"ray[default]\"\n\n# Minimal core only (no dashboard) — smallest footprint for worker images\nuv pip install -U \"ray\"\n\n# Reinforcement learning\nuv pip install -U \"ray[rllib]\"\n```\n\nCheck version: `python -c \"import ray; print(ray.__version__)\"` (targets the 2.5x series; APIs below are stable across recent 2.x releases). Pydantic v1 is deprecated upstream — run `pip install -U pydantic` if you see a Ray-emitted Pydantic v1 warning.\n\n## When to Use\n\n- A single-machine Python loop (embarrassingly parallel or not) needs to scale to many cores or many machines with minimal rewriting.\n- You're bottlenecked on CPU-bound preprocessing before GPU training (Ray Data) or need multi-GPU\u002Fmulti-node training (Ray Train) without hand-rolling `torch.distributed`.\n- You want to run a large hyperparameter sweep (Ray Tune) with early stopping\u002Fpruning across a cluster instead of a single machine.\n- You need to serve a model (or several) with autoscaling and custom Python routing logic (Ray Serve) rather than a static endpoint.\n\nNot a fit for simple thread\u002Fprocess-pool parallelism on one machine where `concurrent.futures` is enough — Ray's cluster machinery adds overhead that isn't worth it for that case.\n\n## Core Concepts: Ray Core\n\n```python\nimport ray\nray.init()   # starts a local cluster; omit address to run laptop-only,\n             # or ray.init(address=\"auto\") \u002F address=\"ray:\u002F\u002Fhead-ip:10001\" to join a real cluster\n\n@ray.remote\ndef process_one(sample_id):\n    ...\n    return result\n\n# Launch many tasks in parallel; futures return immediately\nfutures = [process_one.remote(sid) for sid in sample_ids]\nresults = ray.get(futures)   # blocks until all complete\n\n@ray.remote(num_gpus=1)\nclass ModelActor:\n    def __init__(self, checkpoint_path):\n        self.model = load_model(checkpoint_path)\n\n    def predict(self, batch):\n        return self.model(batch)\n\nactor = ModelActor.remote(\"model.pt\")\nprediction = ray.get(actor.predict.remote(batch))\n```\n\nKey rules:\n- `.remote()` on a function\u002Factor method returns an `ObjectRef` immediately (non-blocking); `ray.get()` resolves it.\n- Large objects passed between tasks are automatically stored in Ray's shared-memory object store and referenced, not copied — pass big arrays\u002FDataFrames as arguments rather than closing over them, for efficiency.\n- `num_cpus=`, `num_gpus=`, `resources={\"custom\": 1}` on `@ray.remote(...)` declare what each task\u002Factor needs; Ray's scheduler places work accordingly.\n\n## Ray Data: Distributed Preprocessing\n\n```python\nimport ray\n\nds = ray.data.read_parquet(\"s3:\u002F\u002Fbucket\u002Fvariants\u002F\")\nds = ds.map_batches(normalize_batch, batch_format=\"pandas\", batch_size=\"auto\")\nds = ds.filter(lambda row: row[\"qc_pass\"])\nds.write_parquet(\"s3:\u002F\u002Fbucket\u002Fvariants_clean\u002F\")\n```\n\nRay Data streams batches through the cluster rather than materializing the whole dataset in memory, and integrates directly with Ray Train (`Trainer(..., datasets={\"train\": ds})`) so preprocessing and training share the same cluster.\n\n## Ray Train: Distributed Training\n\n```python\nfrom ray.train.torch import TorchTrainer\nfrom ray.train import ScalingConfig\n\ndef train_loop_per_worker(config):\n    model = build_model()\n    model = ray.train.torch.prepare_model(model)   # wraps in DDP automatically\n    for epoch in range(config[\"epochs\"]):\n        train_one_epoch(model, ...)\n        ray.train.report({\"loss\": loss})\n\ntrainer = TorchTrainer(\n    train_loop_per_worker,\n    train_loop_config={\"epochs\": 10},\n    scaling_config=ScalingConfig(num_workers=4, use_gpu=True),\n)\nresult = trainer.fit()\n```\n\nIf you already have a PyTorch Lightning `Trainer`, use `ray.train.lightning.RayLightningEnvironment`\u002F`RayDDPStrategy` instead of rewriting the loop — Ray Train slots underneath Lightning's `Trainer(strategy=...)`.\n\n## Ray Tune: Hyperparameter Search\n\n```python\nfrom ray import tune\nfrom ray.tune.search.optuna import OptunaSearch\n\ndef trainable(config):\n    for step in range(config[\"epochs\"]):\n        acc = train_step(config[\"lr\"], config[\"batch_size\"])\n        tune.report({\"accuracy\": acc})\n\ntuner = tune.Tuner(\n    trainable,\n    tune_config=tune.TuneConfig(\n        search_alg=OptunaSearch(),      # delegates sampling to Optuna\n        metric=\"accuracy\",\n        mode=\"max\",\n        num_samples=100,\n    ),\n    param_space={\n        \"lr\": tune.loguniform(1e-4, 1e-1),\n        \"batch_size\": tune.choice([16, 32, 64]),\n        \"epochs\": 10,\n    },\n)\nresults = tuner.fit()\nbest = results.get_best_result()\n```\n\n## Ray Serve: Model Serving\n\n```python\nfrom ray import serve\n\n@serve.deployment(num_replicas=2, ray_actor_options={\"num_gpus\": 1})\nclass Predictor:\n    def __init__(self):\n        self.model = load_model()\n\n    async def __call__(self, request):\n        payload = await request.json()\n        return self.model.predict(payload[\"inputs\"])\n\nserve.run(Predictor.bind(), route_prefix=\"\u002Fpredict\")\n```\n\nRay Serve autoscales replicas based on request queue depth and supports composing multiple deployments into a pipeline (e.g., a preprocessing deployment feeding a model deployment).\n\n## Running on a Cluster\n\n```bash\nray up cluster.yaml      # provisions a cluster on AWS\u002FGCP\u002FAzure from a declarative config\nray submit cluster.yaml my_script.py\nray dashboard cluster.yaml   # tunnels the Ray dashboard to localhost\nray down cluster.yaml\n```\n\n`cluster.yaml` declares provider (aws\u002Fgcp\u002Fazure\u002Fkubernetes), head\u002Fworker node types, autoscaling min\u002Fmax workers, and a `setup_commands` list (e.g., `pip install -r requirements.txt`) run on every node — this is how you keep Python\u002Fpackage versions in sync across the cluster (see pitfalls below). On Kubernetes, use KubeRay instead of `ray up`.\n\n## Common Pitfalls\n\n- **Version skew across nodes.** Ray requires the *same* `ray` version (and generally the same Python minor version) on the driver and every worker; a mismatch causes opaque serialization or RPC errors, not a clear version-check failure. Bake versions into the cluster image or `setup_commands`.\n- **Capturing large objects in a closure.** A remote function that closes over a large local variable (instead of receiving it as an argument) gets it re-serialized on every call; pass large data as explicit `.remote()` arguments so Ray can place it in the object store once and reuse the reference.\n- **Object store spilling \u002F OOM.** Long `ray.data` pipelines or large actor state can exceed the object store's memory fraction, triggering disk spill (slow) or worker eviction. Watch `ray status` \u002F the dashboard's object store panel, and prefer streaming (`map_batches` with a small `batch_size`) over materializing everything.\n- **Forgetting `ray.get()` blocks.** Piling up thousands of unresolved `ObjectRef`s (e.g., a Python list comprehension over `.remote()` calls with no `ray.get` until the very end) can hold references and delay garbage collection; use `ray.wait()` to process results as they complete instead of `ray.get()` on everything at once for very large fan-outs.\n- **`ray[all]` bloat.** The docs explicitly discourage `pip install ray[all]`; install only the extras your workload needs to avoid dependency conflicts and slow builds.\n- **Actors are stateful — restarts lose state.** If a node dies, Ray can restart an actor (`max_restarts=`), but any in-memory state is gone unless you checkpoint it explicitly (e.g., to the object store or external storage).\n\n## Resources\n\n- Docs: https:\u002F\u002Fdocs.ray.io\u002F\n- Installation guide: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Fray-overview\u002Finstallation.html\n- Ray Train: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Ftrain\u002Ftrain.html\n- Ray Tune: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Ftune\u002Findex.html\n- Ray Serve: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Fserve\u002Findex.html\n- Source: https:\u002F\u002Fgithub.com\u002Fray-project\u002Fray\n","# Ray\n\n## Overview\n\nRay is a distributed runtime for Python: `@ray.remote` turns an ordinary function or class into a task\u002Factor that Ray schedules across all CPUs\u002FGPUs on your laptop or across every node in a cluster, with the same code in both cases. On top of this \"Ray Core,\" Ray ships purpose-built libraries:\n\n- **Ray Data** — distributed, streaming data loading\u002Ftransform (`ray.data.read_parquet(...).map_batches(...)`), good for ML preprocessing pipelines that don't fit in one machine's memory.\n- **Ray Train** — distributed training wrapper (`TorchTrainer`, `ScalingConfig`) for DDP\u002FFSDP\u002FDeepSpeed, with native PyTorch Lightning and Hugging Face `Trainer` integration.\n- **Ray Tune** — scalable hyperparameter search with pluggable search algorithms (including Optuna, HyperOpt) and schedulers (ASHA, PBT, Hyperband).\n- **Ray Serve** — programmable, autoscaling model serving (`@serve.deployment`), including an LLM-serving path (`ray.serve.llm`).\n- **RLlib** — production reinforcement learning at scale.\n\nRay Core is the substrate all of the above sit on: tasks (stateless remote function calls), actors (stateful remote classes), and objects (immutable values in a shared, cluster-wide object store).\n\n## Installation\n\n```bash\n# Recommended for ML workloads\nuv pip install -U \"ray[data,train,tune,serve]\"\n\n# General-purpose distributed Python (dashboard + cluster launcher, no ML libs)\nuv pip install -U \"ray[default]\"\n\n# Minimal core only (no dashboard) — smallest footprint for worker images\nuv pip install -U \"ray\"\n\n# Reinforcement learning\nuv pip install -U \"ray[rllib]\"\n```\n\nCheck version: `python -c \"import ray; print(ray.__version__)\"` (targets the 2.5x series; APIs below are stable across recent 2.x releases). Pydantic v1 is deprecated upstream — run `pip install -U pydantic` if you see a Ray-emitted Pydantic v1 warning.\n\n## When to Use\n\n- A single-machine Python loop (embarrassingly parallel or not) needs to scale to many cores or many machines with minimal rewriting.\n- You're bottlenecked on CPU-bound preprocessing before GPU training (Ray Data) or need multi-GPU\u002Fmulti-node training (Ray Train) without hand-rolling `torch.distributed`.\n- You want to run a large hyperparameter sweep (Ray Tune) with early stopping\u002Fpruning across a cluster instead of a single machine.\n- You need to serve a model (or several) with autoscaling and custom Python routing logic (Ray Serve) rather than a static endpoint.\n\nNot a fit for simple thread\u002Fprocess-pool parallelism on one machine where `concurrent.futures` is enough — Ray's cluster machinery adds overhead that isn't worth it for that case.\n\n## Core Concepts: Ray Core\n\n```python\nimport ray\nray.init()   # starts a local cluster; omit address to run laptop-only,\n             # or ray.init(address=\"auto\") \u002F address=\"ray:\u002F\u002Fhead-ip:10001\" to join a real cluster\n\n@ray.remote\ndef process_one(sample_id):\n    ...\n    return result\n\n# Launch many tasks in parallel; futures return immediately\nfutures = [process_one.remote(sid) for sid in sample_ids]\nresults = ray.get(futures)   # blocks until all complete\n\n@ray.remote(num_gpus=1)\nclass ModelActor:\n    def __init__(self, checkpoint_path):\n        self.model = load_model(checkpoint_path)\n\n    def predict(self, batch):\n        return self.model(batch)\n\nactor = ModelActor.remote(\"model.pt\")\nprediction = ray.get(actor.predict.remote(batch))\n```\n\nKey rules:\n- `.remote()` on a function\u002Factor method returns an `ObjectRef` immediately (non-blocking); `ray.get()` resolves it.\n- Large objects passed between tasks are automatically stored in Ray's shared-memory object store and referenced, not copied — pass big arrays\u002FDataFrames as arguments rather than closing over them, for efficiency.\n- `num_cpus=`, `num_gpus=`, `resources={\"custom\": 1}` on `@ray.remote(...)` declare what each task\u002Factor needs; Ray's scheduler places work accordingly.\n\n## Ray Data: Distributed Preprocessing\n\n```python\nimport ray\n\nds = ray.data.read_parquet(\"s3:\u002F\u002Fbucket\u002Fvariants\u002F\")\nds = ds.map_batches(normalize_batch, batch_format=\"pandas\", batch_size=\"auto\")\nds = ds.filter(lambda row: row[\"qc_pass\"])\nds.write_parquet(\"s3:\u002F\u002Fbucket\u002Fvariants_clean\u002F\")\n```\n\nRay Data streams batches through the cluster rather than materializing the whole dataset in memory, and integrates directly with Ray Train (`Trainer(..., datasets={\"train\": ds})`) so preprocessing and training share the same cluster.\n\n## Ray Train: Distributed Training\n\n```python\nfrom ray.train.torch import TorchTrainer\nfrom ray.train import ScalingConfig\n\ndef train_loop_per_worker(config):\n    model = build_model()\n    model = ray.train.torch.prepare_model(model)   # wraps in DDP automatically\n    for epoch in range(config[\"epochs\"]):\n        train_one_epoch(model, ...)\n        ray.train.report({\"loss\": loss})\n\ntrainer = TorchTrainer(\n    train_loop_per_worker,\n    train_loop_config={\"epochs\": 10},\n    scaling_config=ScalingConfig(num_workers=4, use_gpu=True),\n)\nresult = trainer.fit()\n```\n\nIf you already have a PyTorch Lightning `Trainer`, use `ray.train.lightning.RayLightningEnvironment`\u002F`RayDDPStrategy` instead of rewriting the loop — Ray Train slots underneath Lightning's `Trainer(strategy=...)`.\n\n## Ray Tune: Hyperparameter Search\n\n```python\nfrom ray import tune\nfrom ray.tune.search.optuna import OptunaSearch\n\ndef trainable(config):\n    for step in range(config[\"epochs\"]):\n        acc = train_step(config[\"lr\"], config[\"batch_size\"])\n        tune.report({\"accuracy\": acc})\n\ntuner = tune.Tuner(\n    trainable,\n    tune_config=tune.TuneConfig(\n        search_alg=OptunaSearch(),      # delegates sampling to Optuna\n        metric=\"accuracy\",\n        mode=\"max\",\n        num_samples=100,\n    ),\n    param_space={\n        \"lr\": tune.loguniform(1e-4, 1e-1),\n        \"batch_size\": tune.choice([16, 32, 64]),\n        \"epochs\": 10,\n    },\n)\nresults = tuner.fit()\nbest = results.get_best_result()\n```\n\n## Ray Serve: Model Serving\n\n```python\nfrom ray import serve\n\n@serve.deployment(num_replicas=2, ray_actor_options={\"num_gpus\": 1})\nclass Predictor:\n    def __init__(self):\n        self.model = load_model()\n\n    async def __call__(self, request):\n        payload = await request.json()\n        return self.model.predict(payload[\"inputs\"])\n\nserve.run(Predictor.bind(), route_prefix=\"\u002Fpredict\")\n```\n\nRay Serve autoscales replicas based on request queue depth and supports composing multiple deployments into a pipeline (e.g., a preprocessing deployment feeding a model deployment).\n\n## Running on a Cluster\n\n```bash\nray up cluster.yaml      # provisions a cluster on AWS\u002FGCP\u002FAzure from a declarative config\nray submit cluster.yaml my_script.py\nray dashboard cluster.yaml   # tunnels the Ray dashboard to localhost\nray down cluster.yaml\n```\n\n`cluster.yaml` declares provider (aws\u002Fgcp\u002Fazure\u002Fkubernetes), head\u002Fworker node types, autoscaling min\u002Fmax workers, and a `setup_commands` list (e.g., `pip install -r requirements.txt`) run on every node — this is how you keep Python\u002Fpackage versions in sync across the cluster (see pitfalls below). On Kubernetes, use KubeRay instead of `ray up`.\n\n## Common Pitfalls\n\n- **Version skew across nodes.** Ray requires the *same* `ray` version (and generally the same Python minor version) on the driver and every worker; a mismatch causes opaque serialization or RPC errors, not a clear version-check failure. Bake versions into the cluster image or `setup_commands`.\n- **Capturing large objects in a closure.** A remote function that closes over a large local variable (instead of receiving it as an argument) gets it re-serialized on every call; pass large data as explicit `.remote()` arguments so Ray can place it in the object store once and reuse the reference.\n- **Object store spilling \u002F OOM.** Long `ray.data` pipelines or large actor state can exceed the object store's memory fraction, triggering disk spill (slow) or worker eviction. Watch `ray status` \u002F the dashboard's object store panel, and prefer streaming (`map_batches` with a small `batch_size`) over materializing everything.\n- **Forgetting `ray.get()` blocks.** Piling up thousands of unresolved `ObjectRef`s (e.g., a Python list comprehension over `.remote()` calls with no `ray.get` until the very end) can hold references and delay garbage collection; use `ray.wait()` to process results as they complete instead of `ray.get()` on everything at once for very large fan-outs.\n- **`ray[all]` bloat.** The docs explicitly discourage `pip install ray[all]`; install only the extras your workload needs to avoid dependency conflicts and slow builds.\n- **Actors are stateful — restarts lose state.** If a node dies, Ray can restart an actor (`max_restarts=`), but any in-memory state is gone unless you checkpoint it explicitly (e.g., to the object store or external storage).\n\n## Resources\n\n- Docs: https:\u002F\u002Fdocs.ray.io\u002F\n- Installation guide: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Fray-overview\u002Finstallation.html\n- Ray Train: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Ftrain\u002Ftrain.html\n- Ray Tune: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Ftune\u002Findex.html\n- Ray Serve: https:\u002F\u002Fdocs.ray.io\u002Fen\u002Flatest\u002Fserve\u002Findex.html\n- Source: https:\u002F\u002Fgithub.com\u002Fray-project\u002Fray\n",{"name":5,"license":15,"metadata":24,"description":7,"allowed-tools":27,"compatibility":32},{"version":25,"skill-author":26},"1.0","community",[28,29,30,31],"Read","Write","Edit","Bash","Requires Python 3.10+ (Ray's PyPI metadata requires >=3.10; older 3.9 wheels exist only for Ray \u003C2.10) and ray 2.x (current 2.59.0). Install the extra matching your workload — ray[default] (dashboard+autoscaler), ray[data], ray[train], ray[tune], ray[serve], ray[rllib] — rather than ray[all]. Every node in a cluster must run the same Ray version and a compatible Python minor version; mismatches fail silently or crash workers.","master","49e49abf2f670d429e0b7e2c0349739c3fdd5cbe","skills\u002Fray",5,"https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium","https:\u002F\u002Fgithub.com\u002Fstanfish06","https:\u002F\u002Fwww.stanfish.dev\u002F","2026-11-14 21:31:37",{"source":8,"skillName":5,"commands":42,"prompt":52,"zipAvailable":53},[43,46,49],{"tool":44,"command":45},"npx","npx skills add stanfish06\u002Fskillquarium --skill ray",{"tool":47,"command":48},"bunx","bunx skills add stanfish06\u002Fskillquarium --skill ray",{"tool":50,"command":51},"pnpm","pnpm dlx skills add stanfish06\u002Fskillquarium --skill ray","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fray\u002FSKILL.md",true,[],[56,64,72,79,86,93],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":60,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M76ZJTRVA456GD7K8Y3EP82H","zarr-python","Stores and queries chunked N-D scientific arrays with Zarr-Python 3, including codecs, sharding, S3\u002FGCS storage, and NumPy\u002FDask\u002FXarray integration. Use for array layout, bounded I\u002FO, format migration, or scientific metadata preservation.","MIT","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fzarr-python\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":68,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M76ZJQZ15Q1HKC5V85NHE325","yeet","Publish local changes to GitHub by confirming scope, committing intentionally, pushing the branch, and opening a draft PR through the GitHub app from this plugin, with `gh` used only as a fallback where connector coverage is insufficient.","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fyeet\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fyeet\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M76ZJ8ASZ4QYQ70Y6SBA4Z4S","xarray","N-dimensional labeled arrays and datasets in Python. Built on top of NumPy and Dask. It introduces labels in the form of dimensions, coordinates, and attributes on top of raw NumPy-like arrays, making data analysis in physical sciences more intuitive and less error-prone. Use for working with multi-dimensional scientific data, NetCDF\u002FGRIB\u002FZarr files, climate\u002Fweather\u002Foceanographic datasets, remote sensing, geospatial imaging, large out-of-memory datasets with Dask, and labeled array operations.","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fxarray\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":60,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M76ZHFBXD6VXPQ02VCEG202B","writing-mstest-tests","Use when asked to fix, rewrite, update, improve, modernize, show corrected code for, or explain existing MSTest tests or MSTest-specific configuration. Use for \"review\" when corrected code or edits are wanted, even for one pasted assertion or passing tests with bad failure output. Covers expected\u002Factual labels; generic Boolean, collection, string, numeric, null, identity, exception, hard-cast, and object[] checks; TestContext\u002Flifecycle; timeout\u002Fcancellation; OS\u002FCI conditions, retry, cleanup, parallelization, MSTest.Sdk project setup, and MSTESTxxxx. Honor the installed MSTest version. DO NOT USE to design new test cases (code-testing), perform report-only audits, create project files rather than explain MSTest setup, run tests, migrate frameworks, or handle non-MSTest\u002Fnon-.NET code.","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fwriting-mstest-tests\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":60,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M76ZDFDF4N88RH6HSYGS7GP2","wgs-prs","End-to-end WGS to polygenic risk score pipeline. Takes paired-end FASTQ files (or a pre-existing VCF) through nf-core\u002Fsarek for variant calling, applies VCF QC (normalisation, hard filtering, Ti\u002FTv and Het\u002FHom checks), then computes polygenic risk scores via the PGS Catalog. Fills the FASTQ to VCF gap upstream of the gwas-prs skill.","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fwgs-prs\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":97,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M76ZCC1SC68J42QK4MDC8RCG","webapp-testing","Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fwebapp-testing\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fstanfish06\u002Fskillquarium\u002Fblob\u002Fmaster\u002Fskills\u002Fwebapp-testing\u002FSKILL.md",[],[]]