Scientific Computing Suite¶
High-performance computing, physics/chemistry simulations, ML/DL, Julia, JAX, and data science workflows. Uses the Hub Skill architecture with 14 hubs routing to 112 sub-skills. Optimized for Claude Opus 4.7 with extended context and adaptive reasoning.
Version: 3.5.2 | 11 Agents | 2 Registered Commands | 17 Hubs → 110 Sub-skills | 5 Hook Events
Note
In v3.4.0, research-expert plus 5 methodology skills (research-methodology, research-quality-assessment, research-paper-implementation, scientific-communication, evidence-synthesis) moved to the new research-suite. This suite now focuses purely on computational work. Research-methodology delegations from science-suite agents route to the research-suite instead.
Agents¶
Agent: neural-network-master
Deep learning authority specializing in architecture design, theory, and implementation (Transformers, CNNs, diagnostics).
Model: opus
Version: 3.5.2
Agent: nonlinear-dynamics-expert
Expert in bifurcation analysis, chaos, coupled networks, pattern formation, and equation discovery (SINDy/UDE).
Model: opus
Version: 3.5.2
Agent: simulation-expert
Expert in molecular dynamics, statistical mechanics, and numerical methods (HPC/GPU).
Model: opus
Version: 3.5.2
Agent: statistical-physicist
Expert in correlation functions, non-equilibrium dynamics, and ensemble theory.
Model: opus
Version: 3.5.2
Agent: jax-pro
JAX expert — jit/vmap/pmap, sharding, VJP/JVP, XLA/HLO, Optax, Diffrax, Pallas, NumPyro. Delegates MD, bifurcation, general Bayes, and productionization to peers. Upgraded to opus in v3.5.2.
Model: opus
Version: 3.5.2
Agent: julia-pro
Julia/SciML expert — dispatch, type stability, DiffEq.jl, ModelingToolkit, SciMLSensitivity, UDE, SINDy, Turing, Optimization.jl. Delegates ML/HPC and productionization to peers. Upgraded to opus in v3.5.2.
Model: opus
Version: 3.5.2
Agent: julia-ml-hpc
Julia ML/HPC expert for Lux.jl, MLJ.jl, CUDA.jl, MPI.jl, and GNNLux. Delegates SciML/ODE to julia-pro.
Model: sonnet
Version: 3.5.2
Agent: pinn-engineer
Physics-informed AI for PINNs, NeuralPDE.jl, DeepXDE, BPINN/BNNODE, physics-constrained losses, and inverse PDEs. Renamed from ai-engineer in v3.5.2.
Model: sonnet
Version: 3.5.2
Agent: python-pro
Python systems engineer for production Python, type-driven design, PyO3/Rust extensions, async, and uv/ruff toolchain.
Model: sonnet
Version: 3.5.2
Agent: sci-workflow-engineer
Scientific LLM workflow engineer for JAX/Julia codegen prompts, experiment templates, scientific RAG, and AI-assisted pipelines. Renamed from prompt-engineer in v3.5.2.
Model: sonnet
Version: 3.5.2
Agent: ml-expert
Classical ML/MLOps with scikit-learn, XGBoost/LightGBM, Optuna, SHAP, and MLflow/W&B. Delegates DL to neural-network-master. Moved to haiku in v3.5.2.
Model: haiku
Version: 3.5.2
Registered Commands¶
Two slash commands registered in v3.5.2:
Command: md-sim
Molecular dynamics simulation setup, running, and trajectory analysis (GROMACS/OpenMM/JAX-MD).
Command: benchmark
Scientific code benchmarking across backends and hardware targets, with comparison reports.
Skill-Invoked Commands¶
These commands are triggered by skills, not directly by users:
Command: analyze-data
Analyze data files with statistical tests, visualization, and reproducible reporting.
Command: run-experiment
Design and execute computational experiments with hypothesis tracking.
The legacy paper-review command moved to research-suite and was then removed in favor of the scientific-review skill (produces a .docx deliverable with journal-specific adaptation — strictly better output).
Hub Skills¶
Skills use a hub architecture: 14 hub skills route to 117 specialized sub-skills.
Hub: nonlinear-dynamics (6 sub-skills)¶
Bifurcation analysis, chaos, coupled networks, pattern formation, and equation discovery.
bifurcation-analysis— BifurcationKit.jl: continuation, codimension-1/2 bifurcations, normal formschaos-attractors— Lyapunov exponents, attractor reconstruction, fractal dimension, recurrencenetwork-coupled-dynamics— Kuramoto synchronization, master stability, chimera states, epidemic modelspattern-formation— Turing instability, dispersion relations, spiral waves, amplitude equationsequation-discovery— SINDy, DataDrivenDiffEq.jl, PySINDy, sparse regressionjax-julia-interop— Bridge JAX and Julia SciML ecosystems via PythonCall.jl
Hub: jax-computing (6 sub-skills)¶
Core JAX, optimization, Bayesian inference, differential equations, and physics.
jax-mastery— JIT, vmap, grad, pmap, functional transformationsjax-core-programming— Pytrees, custom primitives, XLA operations, device memoryjax-optimization-pro— Optax, custom schedules, NLSQ, convergence diagnosticsjax-bayesian-pro— JAX-specific NumPyro integration and GPU-accelerated samplingjax-diffeq-pro— Diffrax solvers, neural ODEs, stiff systems, adjoint methodsjax-physics-applications— JAX-MD, JAX-CFD, PINNs, differentiable physics
Hub: julia-language (13 sub-skills)¶
Core Julia, packages, compilation, performance, testing, CI/CD, visualization, HPC, and interop.
julia-mastery— Multiple dispatch, type system, metaprogramming, performance patternscore-julia-patterns— Broadcasting, comprehensions, closures, standard librarypackage-management— Pkg.jl, Project.toml, Manifest.toml, environmentspackage-development-workflow— PkgTemplates.jl, documentation, versioning, registrationcompiler-patterns— PackageCompiler.jl, system images, standalone executablesperformance-tuning— @btime, memory allocation, SIMD, threading, type stabilityjulia-testing-patterns— Test.jl, Aqua.jl, JET.jl static analysisci-cd-patterns— GitHub Actions for Julia: test matrix, coverage, releasesvisualization-patterns— Makie.jl, Plots.jl, interactive and publication-quality figuresweb-development-julia— Genie.jl, HTTP.jl, REST APIsjulia-hpc-distributed— Distributed.jl, MPI.jl, SLURM, multi-node parallelisminterop-patterns— PythonCall.jl, RCall.jl, ccall, cross-language data exchangeecosystem-selection— Choosing optimal Julia packages for a domain
Hub: julia-ml-and-dl (9 sub-skills)¶
Julia neural networks, architectures, training, AD backends, GPU kernels, GNNs, RL, pipelines, deployment.
julia-neural-networks— Flux.jl and Lux.jl: model definition, training loopsjulia-neural-architectures— CNNs, RNNs, Transformers in Flux/Luxjulia-training-diagnostics— Loss curves, gradient norms, convergence monitoringjulia-ad-backends— Zygote.jl, Enzyme.jl, ForwardDiff.jl, DifferentiationInterface.jljulia-gpu-kernels— CUDA.jl, KernelAbstractions.jl, custom GPU kernelsjulia-graph-neural-networks— GraphNeuralNetworks.jl: GCN, GAT, message passingjulia-reinforcement-learning— ReinforcementLearning.jl: DQN, PPO, environmentsjulia-ml-pipelines— MLJ.jl: data pipelines, cross-validation, tuningjulia-model-deployment— ONNX export, HTTP.jl serving, PackageCompiler sysimages
Hub: sciml-and-diffeq (8 sub-skills)¶
SciML ecosystem, DifferentialEquations.jl, ModelingToolkit, optimization, neural PDEs.
sciml-ecosystem— Package selection guide for DifferentialEquations.jl, ModelingToolkit, Optimization.jlsciml-modern-stack— Lux.jl neural networks, SciMLSensitivity adjoint methods, UDEs, DEQdifferential-equations— ODE/SDE/PDE solvers, callbacks, ensemble simulationsmodeling-toolkit— Symbolic differential equations, automatic simplificationoptimization-patterns— Optimization.jl for parameter estimation and inverse problemsneural-pde— NeuralPDE.jl: PINNs with ModelingToolkitcatalyst-reactions— Chemical reaction networks, deterministic and stochastic simulationsjump-optimization— JuMP.jl: LP, QP, NLP, MIP with HiGHS, Ipopt
Hub: correlation-analysis (4 sub-skills)¶
Mathematical foundations, physical systems, computational methods, and experimental data.
correlation-math-foundations— Two-point functions, cumulants, Fourier/Laplace transforms, Wiener-Khinchincorrelation-physical-systems— Condensed matter, soft matter, biological, non-equilibrium correlationscorrelation-computational-methods— FFT-based autocorrelation, multi-tau correlators, JAX-accelerated GPUcorrelation-experimental-data— DLS, SAXS/SANS, XPCS, FCS, rheology data interpretation
Hub: statistical-physics-hub (8 sub-skills)¶
Equilibrium/non-equilibrium statistical mechanics, stochastic dynamics, active matter, rare events, and extremes.
statistical-physics— Ensemble theory, partition functions, phase transitions; Julia Monte Carlo idioms (v3.1.5)stochastic-dynamics— Master equations, Fokker-Planck direct PDE methods, Langevin, Green-Kubo, jump-diffusion SDEs (Fokker-Planck v3.1.5)non-equilibrium-theory— Fluctuation theorems, entropy production, linear response, BAR/Jarzynski/MBAR with pymbar worked example (BAR example v3.1.7)active-matter— Self-propelled particles, flocking, MIPS, bio-inspired materialsmultiscale-modeling— Coarse-graining, DPD, nanoscale DEMadvanced-simulations— Non-equilibrium thermodynamics, multiscale bridgingrare-events-sampling— Large-deviation theory, cloning / importance splitting, SOC / sandpile / avalanche statistics (new in v3.1.4)extreme-value-statistics— GEV/GPD/Hill/Pickands/POT, return levels, non-stationary EVT (new in v3.1.4)
Hub: deep-learning-hub (6 sub-skills)¶
Architecture design, mathematical foundations, training diagnostics, experimentation, large-scale systems.
deep-learning— Core DL: feedforward networks, backpropagation, regularizationneural-architecture-patterns— CNNs, RNNs, Transformers, diffusion models, normalizing flowsneural-network-mathematics— Universal approximation, optimization landscapes, generalization theorytraining-diagnostics— Loss curves, gradient pathologies, learning rate tuningdeep-learning-experimentation— Ablations, HPO, reproducibility, benchmarksadvanced-ml-systems— Distributed training, mixed precision, gradient checkpointing
Hub: ml-and-data-science (7 sub-skills)¶
Classical ML, data analysis, wrangling, statistics, visualization, curve fitting, experiment tracking.
machine-learning— Scikit-learn, XGBoost, LightGBM, feature engineeringdata-analysis— Pandas, descriptive statistics, correlation, hypothesis testingdata-wrangling-communication— Data cleaning, transformation, stakeholder communicationstatistical-analysis-fundamentals— Distributions, hypothesis tests, confidence intervalsscientific-visualization— Matplotlib, seaborn, plotly, domain-specific plotsnlsq-core-mastery— JAX-accelerated non-linear least squares curve fittingexperiment-tracking— MLflow, Weights & Biases, DVC
Hub: llm-and-ai (5 sub-skills)¶
LLM application development, evaluation, LangChain, RAG, and NLP.
llm-application-dev— LLM-powered apps: API integration, streaming, tool use, agentsllm-evaluation— Benchmarks, LLM-as-judge, human evaluation, quality metricslangchain-architecture— LangChain/LangGraph: chains, agents, memory, toolsrag-implementation— Vector stores, chunking, re-ranking, hybrid retrievalnlp-fundamentals— Tokenization, embeddings, NER, text classification
Hub: ml-deployment (6 sub-skills)¶
Model serving, optimization, production engineering, pipelines, infrastructure, federated learning.
model-deployment-serving— FastAPI, TorchServe, Triton, BentoML, REST/gRPCmodel-optimization-deployment— Quantization, pruning, ONNX, TensorRT, mobileml-engineering-production— Type-safe code, testing, data pipelines, monitoring, driftml-pipeline-workflow— Airflow, Prefect, Metaflow, automated retrainingdevops-ml-infrastructure— Docker, Kubernetes, GPU provisioning, cloud MLfederated-learning— Federated averaging, differential privacy, PySyft
Hub: simulation-and-hpc (10 sub-skills)¶
MD simulations, trajectory analysis, ML force fields, parallel computing, GPU, numerical methods, and applied math.
md-simulation-setup— GROMACS/LAMMPS force fields, equilibration protocolstrajectory-analysis— MDAnalysis, RMSD, RDF, free energy, clusteringml-force-fields— NequIP, MACE, DeePMD, active learning workflowsparallel-computing— MPI, OpenMP, Dask, Ray, scalability analysisgpu-acceleration— CUDA, ROCm, JAX pmap, GPU-optimized algorithmsnumerical-methods-implementation— Finite difference/element, spectral methods, iterative solverssignal-processing— FFT, filtering, spectral estimation, wavelet transformstime-series-analysis— ARIMA, state space models, changepoint detectionadvanced-optimization— Genetic algorithms, simulated annealing, basin hoppingcontrol-theory— PID, LQR, MPC, stability analysis
Hub: research-and-domains (14 sub-skills)¶
Scientific software engineering, specialized computational domains, and autonomous self-improvement research.
Note
In v3.4.0, 5 methodology sub-skills (research-methodology, research-paper-implementation, research-quality-assessment, scientific-communication, evidence-synthesis) moved to the research-suite. This hub now focuses on computational-engineering-for-science rather than methodology.
robust-testing— Property-based, metamorphic, and tolerance-aware testing for scientific codepython-development— Idiomatic Python, software engineering for sciencepython-packaging-advanced— uv workspaces, monorepos, reproducible buildsrust-extensions— PyO3/maturin high-performance Python extensionstype-driven-design— Protocols, Generics, static analysis with pyright/mypymodern-concurrency— asyncio TaskGroups, threading, multiprocessingquantum-computing— Qiskit, PennyLane, VQE/QAOAbioinformatics— Genomics, proteomics, BioPythoncomputer-vision— Image processing, detection, Vision Transformersreinforcement-learning— Gymnasium, Stable-Baselines3, RLlibsymbolic-math— SymPy, CAS, algebraic solversself-improving-ai— Research overview for autonomous self-improvement (research-framework counterpart to agent-core’s self-improving-agents)dspy-basics— Depth-skill companion for DSPy programmatic prompt optimizationrlaif-training— Depth-skill companion for Constitutional AI / RLAIF / DPO
Hub: bayesian-inference (10 sub-skills)¶
NumPyro, Turing.jl, variational inference, MCMC diagnostics, consensus / non-reversible tempering, Bayesian UDEs, Bayesian PINNs, point processes, and Bayesian SINDy equation discovery.
numpyro-core-mastery— NumPyro: NUTS/HMC, SVI, hierarchical models, GPU inferenceturing-model-design— Turing.jl: probabilistic models, Julia-native Bayesian workflowsvariational-inference-patterns— ELBO, mean-field, normalizing flows, amortized inferencemcmc-diagnostics— R-hat, ESS, BFMI, trace plots, ArviZ convergence diagnosticsconsensus-mcmc-pigeons— Scott-2016 divide-and-conquer Consensus MC and Pigeons.jl non-reversible parallel tempering (new in v3.1.4)bayesian-ude-workflow— Turing + DiffEq + Lux staged pipeline for Bayesian Universal Differential Equations (new in v3.1.4)bayesian-ude-jax— Python/JAX counterpart to Bayesian UDE via Diffrax + Equinox + NumPyro (new in v3.1.4)bayesian-pinn— BNNODE / BayesianPINN (extracted from neural-pde for budget management) (new in v3.1.4)point-processes— Hawkes processes, HSGP, Julia PointProcesses.jl, non-parametric Hawkes EM (new in v3.1.4)bayesian-sindy-workflow— Horseshoe-prior Bayesian SINDy with 5-stage Lorenz-63 worked example (NumPyro + NUTS + ArviZ PSIS-LOO), prior-sensitivity analysis, and Turing UQ-SINDy sidebar (new in v3.1.7 — extracted from equation-discovery to resolve 88% budget pressure)
Hooks¶
5 hook events with Python script implementations:
SessionStart— Detect JAX devices, GPU availability, Julia envPreToolUse— Warn before commands that could corrupt simulationsPostToolUse— NaN/Inf check on compute job output (numerical integrity)SessionEnd— Persist structured progress summary for next sessionSubagentStop— Collect results from parallel science agents
(ExecutionError was removed in v3.4.0 — not supported by the CC v2.1.113 CLI event schema.)