DevOps & Infra

golemcloud Skill

Deploying a Golem application. Use when asked to deploy agents to a Golem server, update running agents, or troubleshoot deployment issues.

golemcloud Skill

Rolling back a Golem deployment to a previous revision or version. Use when reverting a deployment, restoring a prior environment state, or recovering from a bad deploy.

bg-szy Bundle

Train or fine-tune language and vision models using TRL (Transformer Reinforcement Learning) or Unsloth with Hugging Face Jobs infrastructure. Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs package, UV scripts with PEP 723 format, dataset preparation and validation, hardware selection, cost estimation, Trackio monitoring, Hub authentication, model selection/leaderboards and model persistence. Use for tasks involving cloud GPU training, GGUF conversion, or when users mention training on Hugging Face Jobs without local GPU setup.

bg-szy Bundle

Set up an isolated Python environment for SageMaker / AWS work, with the right Python version and current boto3. Use this skill whenever Python code will be executed for a SageMaker deployment, training job, or any AWS automation — including when about to run `pip install`, when about to invoke `boto3`, when creating or activating a virtualenv, or when the user asks to "set up the environment". Never use system Python and never `pip install` into it. Always isolate. This skill prevents the most common failure modes: wrong Python version, dependency conflicts, and stale SDKs.

bg-szy Bundle

Trains and fine-tunes vision models for object detection (D-FINE, RT-DETR v2, DETR, YOLOS), image classification (timm models — MobileNetV3, MobileViT, ResNet, ViT/DINOv3 — plus any Transformers classifier), and SAM/SAM2 segmentation using Hugging Face Transformers on Hugging Face Jobs cloud GPUs. Covers COCO-format dataset preparation, Albumentations augmentation, mAP/mAR evaluation, accuracy metrics, SAM segmentation with bbox/point prompts, DiceCE loss, hardware selection, cost estimation, Trackio monitoring, and Hub persistence. Use when users mention training object detection, image classification, SAM, SAM2, segmentation, image matting, DETR, D-FINE, RT-DETR, ViT, timm, MobileNet, ResNet, bounding box models, or fine-tuning vision models on Hugging Face Jobs.

bg-szy Bundle

提供 ModelArts 全栈基础资源管理能力

bg-szy Bundle

Build and publish a Gradio demo on Hugging Face Spaces for a user-provided LoRA. Use when someone asks to create, generate, ship, or publish a Space, demo, Gradio app, or playground for a LoRA — including LoRAs for Qwen-Image, Qwen-Image-Edit, LTX-Video, Wan, FLUX, SDXL, or other diffusion base models. Also triggers when someone describes a LoRA they trained or hosts on the Hub and wants to share it. Covers picking the right base pipeline and `diffusers` inference recipe, designing a UI tailored to the LoRA's task and inputs (Union/multi-task control, edit, video, image, etc.), respecting model-card recommendations (trigger words, steps, guidance, LoRA scale, example inputs), and shipping to ZeroGPU hardware as a private Space by default.

bg-szy Bundle

Ensure a usable SageMaker execution role exists before deploying or training. Use this skill whenever about to create a SageMaker endpoint, model, training job, or any resource that requires an execution role. Use it especially when the user has not provided a role ARN explicitly, when scripts are about to call `iam:CreateRole`, or when an AccessDenied error mentions an IAM action. Never blindly call `iam:CreateRole` — always check for existing roles first. This skill prevents the most common SageMaker deployment failure: trying to create IAM resources from an SSO principal that has no IAM write permissions.

bg-szy Bundle

Pick the right serving container for a SageMaker model deployment and find its current image URI. Use this skill whenever about to deploy a model to a SageMaker endpoint and an image URI needs to be chosen — including when the user says "deploy this LLM", "host this HuggingFace model", "serve this fine-tuned model", "deploy this embedding model", "host a reranker", "serve a sentence-transformers model", or when about to hardcode any container URI in deployment code. HuggingFace-curated Deep Learning Containers are ALWAYS preferred: HuggingFace vLLM (LLMs and generative rerankers), HuggingFace vLLM-Omni (multimodal), TEI (embeddings/cross-encoder rerankers), HF Inference Toolkit (other transformers). Generic images (AWS vLLM, DJL-LMI, SGLang) are used only when no HuggingFace image is compatible — never merely because they carry a newer version. Never hardcode a container URI from memory and never default to TGI. Prevents stale-image failures and wrong-region URIs.

bg-szy Bundle

Plan and coordinate the deployment of a model to Amazon SageMaker AI. Use this skill whenever the user wants to deploy, host, serve, or expose a model on SageMaker or AWS — including phrases like "deploy a model", "host this LLM on AWS", "serve this embedding model", "deploy a reranker", "deploy a text-to-image / diffusion model", "host this for async inference", "create an endpoint", "serve my fine-tuned model", or any request that involves making a model available for inference on AWS. Use this even when the user is vague (e.g. "I just want to get this running on AWS, you figure it out"). Works for text-generation LLMs, embedding models, rerankers, classifiers, text-to-image / diffusion models — picks the right serving stack and chooses between real-time and async inference. This is the entry-point skill for SageMaker deployment work — it asks clarifying questions, picks a deployment pathway, and coordinates the other deployment skills.

bg-szy Bundle

Create a SageMaker endpoint (real-time, real-time scale-to-zero, or async) with autoscaling, CloudWatch alarms, and tagging enabled by default. Use this skill whenever about to create a SageMaker endpoint, write deployment code that calls `create_endpoint`, or finalize a deployment after the image URI and IAM role are known. Provides deploy.py for real-time endpoints, deploy_ic.py for real-time endpoints that scale to zero instances via inference components, and deploy_async.py for async endpoints (also scale-to-zero). This is the last step in the SageMaker deployment workflow. Never generate a bare `create_endpoint` call without these defaults — endpoints without autoscaling or alarms are demos, not deployments.

bg-szy Bundle

End-to-end alternative splicing analysis from FASTQ to differential splicing results for short-read bulk RNA-seq. Aligns with STAR 2-pass cohort-style, performs junction QC (RSeQC, MaxEntScan, SpliceAI), runs rMATS-turbo and leafcutter for concordant differential analysis, optionally MAJIQ V3 for complex events / heterogeneous cohorts, isoform-switching with NMD/ORF/domain consequences (IsoformSwitchAnalyzeR v2 + DRIMSeq+DEXSeq+stageR DTU), and sashimi visualizations. Use when performing comprehensive splicing analysis from raw bulk RNA-seq data; for variant-driven splice prediction see splice-variant-prediction; for rare-disease single-patient outlier detection see outlier-splicing-detection; for full-isoform PacBio/ONT analysis see long-read-splicing.

bg-szy Bundle

Cell-free DNA analysis pipeline from plasma sequencing to tumor monitoring. Preprocesses cfDNA reads, analyzes fragment patterns, estimates tumor fraction from sWGS, and optionally detects mutations from targeted panels. Use when analyzing liquid biopsy samples for cancer detection or monitoring.

robsonkades Bundle

Enums as types rather than labelled integers: instance fields instead of ordinal, constant-specific behaviour and strategy enums, extensibility through interfaces, EnumSet and EnumMap instead of bit fields and ordinal-indexed arrays, exhaustive switch and what separate compilation does to it, and what happens when an enum value crosses a database, a JSON payload or a topic. Use when int or String constants stand in for a closed set, when ordinal() appears anywhere outside a library, when @Enumerated is declared ORDINAL or left at its default, when a switch over an enum has a default branch that hides new constants, when adding a constant breaks a consumer during a rolling deploy, when values() is called in a loop, or when a set of flags is packed into an int. Does not cover annotations (java-annotations), sealed hierarchies and records as the open-data alternative (java-composition-over-inheritance), or equality and ordering contracts in general (java-object-contracts).

robsonkades Bundle

Adapter in modern Java: making an existing type usable through an interface it was not written for, and keeping a foreign model, vocabulary and failure mode from leaking inward. Covers object versus class adapters, why a lambda already adapts a single-method interface, the error-translation duty most adapters omit, when an adapter has quietly become a translator with business rules in it, and when a passthrough should be deleted. Use when integrating a vendor SDK or legacy type behind your own port, when two libraries must interoperate, when an adapter is proposed between types you own, when foreign exceptions or DTOs appear in domain code, or when reviewing a wrapper that renames methods and does nothing else. Does not cover the Kubernetes telemetry sidecar (adapter-sidecar-pattern), simplifying a subsystem you own (gof-facade), adding behaviour to the same interface (gof-decorator), controlling access to an object (gof-proxy), or layering rules in general (layering-and-boundaries).

robsonkades Bundle

Iterator in modern Java: traversing an aggregate without exposing it, and choosing between Iterator, Stream and Spliterator — external pull versus internal lazy pipeline versus the parallel decomposition primitive. Covers when a Spliterator can adapt to both, what fail-fast really promises and how weakly consistent iterators differ, streams that hold a resource and must be closed, remote pagination as iteration with page drift, and the characteristics that decide whether a stream can be sized or split. Use when exposing a collection from a type, when a custom traversal is being written, when ConcurrentModificationException appears, when a stream over a file or a result set leaks, when paging through a remote API, or when a parallel stream is not faster. Does not cover stream pipeline design and collectors in general, the tree being traversed (gof-composite), adding operations over it (gof-visitor), or database paging strategy.

robsonkades Bundle

Stream pipelines as a design decision: when a stream is clearer than a loop and when it is not, side-effect-free stages and mutable reduction with collectors, the toMap and groupingBy traps, Collection versus Stream as a return type, streams that hold an open resource, parallel streams and the shared common pool, and Gatherers for custom intermediate operations. Use when a pipeline mutates state outside itself or uses forEach to accumulate, when Collectors.toMap throws IllegalStateException or NullPointerException, when a method returns a Stream that callers iterate twice, when a stream over Files.lines or a JDBC cursor is never closed, when parallelStream() appears — especially with blocking I/O — or when a loop is being rewritten as a stream for its own sake. Does not cover lambda capture and functional interfaces (java-lambdas-and-functional-interfaces), ForkJoinPool internals (forkjoinpool-and-work-stealing), or collection choice and complexity.

robsonkades Bundle

Choosing which automated checks a change must pass, and making them cheap enough that they stay switched on: matching the gate set to the change's risk rather than running everything on everything, where each gate belongs (pre-commit, pull request, main, release), the Java toolchain that enforces each class of defect, ratcheting a gate onto a codebase that already violates it, and what to do when a gate goes red. Use when setting up or trimming a pipeline, when the build is slow enough that people push without running it, when a check is routinely bypassed or its failures ignored, when a defect class keeps reaching production, when a coverage or static-analysis threshold is being proposed, or when deciding whether a small change really needs the full pipeline. Does not cover writing the tests (java-testing-strategy), architecture rules (architecture-testing), performance thresholds (performance-regression-ci), or human review (code-review).

robsonkades Bundle

Deoptimisation and recompilation on HotSpot: uncommon traps and their reason codes, the none / maybe_recompile / reinterpret / make_not_entrant / make_not_compilable actions, jdk.Deoptimization in JFR, -XX:+TraceDeoptimization, the per-method trap limits and recompilation cutoffs, and diagnosing a method that never stabilises. Use when a method repeatedly shows "made not entrant" in the compilation log, when latency spikes correlate with class loading or a deploy, when a burst of "marked for deoptimization" follows a deploy or a plugin load, when a feature flag or APM agent is suspected of invalidating compiled code, when "made not compilable" or a flood of action "none" appears for a hot method, when someone proposes raising PerMethodRecompilationCutoff, or when -Xlog:jit+deoptimization produced an empty file. Does not cover the tiered pipeline and warm-up (jit-compilation), reading the compilation log itself (compilation-and-inlining-logs), or C2's internal representation (c2-sea-of-nodes).

robsonkades Bundle

Stating a system's fault model before designing against it: crash-stop, crash-recovery, omission, timing and Byzantine faults; partial failure and the third outcome of every remote call (unknown); gray failure and the slow node whose health check stays green; the eight fallacies as a checklist; blast radius, correlated versus independent failure, and the availability arithmetic of a dependency chain. Use when a design says "if the service is down" without defining down, when a retry is added to a call whose outcome is unknown, when a node is slow rather than dead, when replicas share a host, an AZ or a database, or when an availability target is quoted for a service built on ten others. Does not cover what the model implies about messages (delivery-semantics) or reads (consistency-models), how a failure spreads (cascading-failures), the named shapes (distributed-failure-catalogue), what an orchestrator does with a failed pod (kubernetes-service-lifecycle), or failures as types (java-exception-design).

robsonkades Bundle

How HotSpot actually executes and compiles: the runtime-generated template interpreter, C2's sea-of-nodes IR, a release-scoped diagnostic map of compilation phases, and why a given transformation fired or did not. Use when a method is believed to be "not optimised", when an allocation that looks eliminable still shows up in allocation profiling, when a hot call site reports `too large` or stays non-inlined, when `made not entrant` repeats on the same method, when someone prescribes `-XX:CompileThreshold` under tiered compilation, or when explaining why the JIT did not fix an O(n^2) loop. Does not cover the tiered pipeline, warm-up and code cache sizing (jit-compilation), reading the compiler's own decision logs end to end (compilation-and-inlining-logs), the emitted machine code (reading-jit-assembly), or the bytecode the compiler consumes (jvm-bytecode).

robsonkades Bundle

HotSpot JIT compilation and warm-up: tiered policy, C1/C2 queues and profiling, OSR, deoptimization, code-cache pressure, compiler resources in containers, and warm-up as a workload-dependent curve rather than a clock delay. Use when p99 is bad for the first minutes after a deploy, when performance degrades permanently until a restart, when "CodeCache is full" appears, when a startup probe or traffic gate needs a warm-up criterion, when -XX:-TieredCompilation or -Xcomp is proposed, when scaling out a low-traffic service makes latency worse, when a 1-2 CPU pod warms up far slower than a workstation, or when an autoscaler keeps adding cold replicas. Does not cover inlining and escape analysis (jit-inlining-and-escape-analysis), microbenchmarks (jmh-microbenchmarks), or the code cache in the memory budget (jvm-memory-regions). Reading the compiler output is compilation-and-inlining-logs, recompilation is deoptimization, and per-segment exhaustion is code-cache-segments.

robsonkades Bundle

Electing one active instance for work that must not run concurrently: the lease renewal model and the rule that failed renewal never extends the leader's conservative deadline; split-brain and resource-side fencing/idempotency; failover time as detection, election and warm-up; coordination-store leases, Kubernetes Lease objects and ShedLock rows, and what each is adequate for; and when not to elect. Use when a @Scheduled job runs once per replica after scaling out, when two instances both believe they lead, when a leader keeps working after its lease expired, when failover takes a minute nobody budgeted, or when ShedLock is described as leader election. Not lease and fencing mechanics (distributed-locks-and-leases), how the election is decided (consensus-and-quorums), why a scheduled job duplicates (stateless-service-design), splitting work by key (sharding-and-partitioning), or pod termination (kubernetes-service-lifecycle).

robsonkades Bundle

Composing a second container into the same pod to add a capability to a container you cannot or will not modify: the shared network namespace and volumes that make this different from a library, native sidecar containers (an init container with restartPolicy Always) and the startup and shutdown ordering they fix, per-container requests against pod-level QoS, and the failure matrix of a two-container pod. Use when a proxy, TLS terminator, config reloader or log shipper is added beside an application, when requests fail in the first seconds after a pod starts because the app came up before its proxy, when a Job's pod stays Running because the sidecar never exits, or when a sidecar is up but broken and the app cannot tell. Does not cover probes and graceful shutdown (kubernetes-service-lifecycle), mediating outbound traffic (ambassador-pattern), normalising what the app emits (adapter-sidecar-pattern), or JVM cgroup detection (container-awareness).

robsonkades Bundle

Deciding whether to cache, then doing it safely: saved origin work and latency, bounded size or weight, TTL and jitter, stampede and its four distinct scopes, cache-aside versus refreshAfterWrite, immutable DTOs rather than JPA entities, invalidation across instances, Redis serialisation, and why hit rate alone is a misleading metric. Use when a cache is being added or reviewed, when @Cacheable is called from within the same bean, when a cache has no size limit or no TTL, when entries are preloaded in bulk with one TTL, when hit rate is the only metric on the dashboard, when Old Gen keeps growing, when FLUSHALL appears in a deploy pipeline, or when instances disagree about a value. Does not cover the pool the cache protects (connection-pool-sizing), the queueing arithmetic (littles-law-and-queueing), or GC tuning for the resulting heap (jvm-gc-tuning).

robsonkades Bundle

The major memory-accounting domains of a JVM process — heap, Metaspace/class space, code cache, thread stacks, direct/native/JVM-internal memory and mapped/file-backed pages — and how to budget them against a container limit. Use when a pod is OOMKilled with no Java exception, when an OutOfMemoryError names something other than "Java heap space", when -Xmx is set equal to the container limit, when RSS exceeds the heap by more than expected, when a heap above 32 GB is proposed, or when sizing a JVM for Kubernetes. Does not cover collector choice and heap tuning (jvm-gc-tuning), classloader leaks (jvm-class-loading), or kernel-side memory behaviour such as page faults, swap and the OOM killer (linux-for-jvm). Metaspace internals are metaspace-internals, memory outside the heap is off-heap-memory, and heap contents are heap-dump-analysis.

robsonkades Bundle

Template Method in modern Java: fixing an algorithm's skeleton while named steps vary, and the inheritance coupling that often makes composition preferable. Covers when final protects the sequence, controlled overriding, minimal hook surfaces, the constructor-calls-an-overridable-method trap, protected hooks becoming an API you cannot change, when the pattern is genuinely right (frameworks that instantiate your subclass, contract test base classes), and how to convert one to a class taking its steps as collaborators. Use when an abstract base class with protected hooks is proposed, when a base-class change broke subclasses, when a template has grown past a handful of hooks, or when subclasses override the template method itself. Does not cover choosing a whole algorithm (gof-strategy), creating the product a template needs (gof-factory-method), the general inheritance decision (java-composition-over-inheritance), or pipeline stages contributed independently (gof-chain-of-responsibility).

robsonkades Bundle

Designing and operating always-on production profiling: question-driven signal choice, permanent overhead and coverage budgets, in-process versus host collection, context-label propagation, profile schemas, storage and cardinality, retention and incident preservation, deploy-aware comparisons, trust boundaries, and evidence-quality SLOs. Use when historical CPU/allocation/lock evidence must survive an incident, when profile cost or tenant labels can grow without bound, when a backend or agent is being selected, or when two time windows are compared as a regression claim. Does not teach one-off capture mechanics (jfr-and-async-profiler), async-profiler engines (async-profiler-advanced), JFR tuning (jfr-advanced), or graph interpretation (flame-graph-analysis).

robsonkades Bundle

Interpreting flame graphs as weighted sampled call-path aggregates: identifying the selection event and denominator, separating inclusive from leaf/self attribution, recognizing truncation, inlining, symbol and thread/task artifacts, using bottom-up and differential views, quantifying sample uncertainty, and turning a hotspot into a bounded causal experiment. Use when a graph looks CPU-heavy, idle-heavy, fragmented, changed after a deploy, or tempting to optimize by width alone. Does not collect profiles (jfr-and-async-profiler), configure engines/conversion (async-profiler-advanced), or define benchmark/latency inference (jmh-microbenchmarks, latency-statistics).

knuckles-team Skill

Installs and configures the self-hosted RustDesk client on Debian/Ubuntu hosts. Downloads the correct architecture-specific .deb package (supporting x86_64 and aarch64), resolves all required system dependencies, dynamically configures multi-profile settings pointing to the self-hosted rendezvous server, locks down configuration permissions, starts and registers the rustdesk service, and returns the machine's unique RustDesk ID. Triggers on "install rustdesk client", "setup remote desktop", "configure rustdesk", "deploy rustdesk client", "install remote access". Do NOT use for Windows or non-Debian environments.

knuckles-team Skill

Universal Docker Compose and Swarm operation and deployment atomic skill. Compiles environment templates, triggers service deploys, and monitors operational health.

knuckles-team Bundle

Deploy applications and infrastructure to Cloudflare using Workers, Pages, and related platform services. Use when the user asks to deploy, host, publish, or set up a project on Cloudflare.

knuckles-team Bundle

Master trading pipeline: Hypothesis → Debate → Backtest → Paper → Live. Orchestrates the full strategy lifecycle with human approval gates.

knuckles-team Skill

Docker Compose and Swarm configuration drift detection atomic skill. Compares running state against git configuration.

knuckles-team Bundle

Detect drift between a host's running Docker Compose/Swarm state and its git configuration, then compile and redeploy the declared configuration where the operator confirms remediation. Use when a stack's live state may have diverged from its source of truth; this workflow never redeploys without the drift report as evidence.

knuckles-team Bundle

Detect drift between the fleet's running Docker Compose/Swarm state and its git configuration across all registered stacks, then render the findings as a formatted audit report. Use when a fleet operator needs a point-in-time fleet-wide drift audit; this workflow is read-only and does not redeploy.