Scientific Computing Suite ========================== High-performance computing, physics/chemistry simulations, ML/DL, Julia, JAX, and data science workflows. Uses the :term:`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 :doc:`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 :description: Deep learning authority specializing in architecture design, theory, and implementation (Transformers, CNNs, diagnostics). :model: opus :version: 3.5.2 .. agent:: nonlinear-dynamics-expert :description: Expert in bifurcation analysis, chaos, coupled networks, pattern formation, and equation discovery (SINDy/UDE). :model: opus :version: 3.5.2 .. agent:: simulation-expert :description: Expert in molecular dynamics, statistical mechanics, and numerical methods (HPC/GPU). :model: opus :version: 3.5.2 .. agent:: statistical-physicist :description: Expert in correlation functions, non-equilibrium dynamics, and ensemble theory. :model: opus :version: 3.5.2 .. agent:: jax-pro :description: 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 :description: 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 :description: 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 :description: 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 :description: 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 :description: 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 :description: 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 :description: Molecular dynamics simulation setup, running, and trajectory analysis (GROMACS/OpenMM/JAX-MD). .. command:: benchmark :description: 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 :description: Analyze data files with statistical tests, visualization, and reproducible reporting. .. command:: run-experiment :description: 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 forms - ``chaos-attractors`` — Lyapunov exponents, attractor reconstruction, fractal dimension, recurrence - ``network-coupled-dynamics`` — Kuramoto synchronization, master stability, chimera states, epidemic models - ``pattern-formation`` — Turing instability, dispersion relations, spiral waves, amplitude equations - ``equation-discovery`` — SINDy, DataDrivenDiffEq.jl, PySINDy, sparse regression - ``jax-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 transformations - ``jax-core-programming`` — Pytrees, custom primitives, XLA operations, device memory - ``jax-optimization-pro`` — Optax, custom schedules, NLSQ, convergence diagnostics - ``jax-bayesian-pro`` — JAX-specific NumPyro integration and GPU-accelerated sampling - ``jax-diffeq-pro`` — Diffrax solvers, neural ODEs, stiff systems, adjoint methods - ``jax-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 patterns - ``core-julia-patterns`` — Broadcasting, comprehensions, closures, standard library - ``package-management`` — Pkg.jl, Project.toml, Manifest.toml, environments - ``package-development-workflow`` — PkgTemplates.jl, documentation, versioning, registration - ``compiler-patterns`` — PackageCompiler.jl, system images, standalone executables - ``performance-tuning`` — @btime, memory allocation, SIMD, threading, type stability - ``julia-testing-patterns`` — Test.jl, Aqua.jl, JET.jl static analysis - ``ci-cd-patterns`` — GitHub Actions for Julia: test matrix, coverage, releases - ``visualization-patterns`` — Makie.jl, Plots.jl, interactive and publication-quality figures - ``web-development-julia`` — Genie.jl, HTTP.jl, REST APIs - ``julia-hpc-distributed`` — Distributed.jl, MPI.jl, SLURM, multi-node parallelism - ``interop-patterns`` — PythonCall.jl, RCall.jl, ccall, cross-language data exchange - ``ecosystem-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 loops - ``julia-neural-architectures`` — CNNs, RNNs, Transformers in Flux/Lux - ``julia-training-diagnostics`` — Loss curves, gradient norms, convergence monitoring - ``julia-ad-backends`` — Zygote.jl, Enzyme.jl, ForwardDiff.jl, DifferentiationInterface.jl - ``julia-gpu-kernels`` — CUDA.jl, KernelAbstractions.jl, custom GPU kernels - ``julia-graph-neural-networks`` — GraphNeuralNetworks.jl: GCN, GAT, message passing - ``julia-reinforcement-learning`` — ReinforcementLearning.jl: DQN, PPO, environments - ``julia-ml-pipelines`` — MLJ.jl: data pipelines, cross-validation, tuning - ``julia-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.jl - ``sciml-modern-stack`` — Lux.jl neural networks, SciMLSensitivity adjoint methods, UDEs, DEQ - ``differential-equations`` — ODE/SDE/PDE solvers, callbacks, ensemble simulations - ``modeling-toolkit`` — Symbolic differential equations, automatic simplification - ``optimization-patterns`` — Optimization.jl for parameter estimation and inverse problems - ``neural-pde`` — NeuralPDE.jl: PINNs with ModelingToolkit - ``catalyst-reactions`` — Chemical reaction networks, deterministic and stochastic simulations - ``jump-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-Khinchin - ``correlation-physical-systems`` — Condensed matter, soft matter, biological, non-equilibrium correlations - ``correlation-computational-methods`` — FFT-based autocorrelation, multi-tau correlators, JAX-accelerated GPU - ``correlation-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 materials - ``multiscale-modeling`` — Coarse-graining, DPD, nanoscale DEM - ``advanced-simulations`` — Non-equilibrium thermodynamics, multiscale bridging - ``rare-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, regularization - ``neural-architecture-patterns`` — CNNs, RNNs, Transformers, diffusion models, normalizing flows - ``neural-network-mathematics`` — Universal approximation, optimization landscapes, generalization theory - ``training-diagnostics`` — Loss curves, gradient pathologies, learning rate tuning - ``deep-learning-experimentation`` — Ablations, HPO, reproducibility, benchmarks - ``advanced-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 engineering - ``data-analysis`` — Pandas, descriptive statistics, correlation, hypothesis testing - ``data-wrangling-communication`` — Data cleaning, transformation, stakeholder communication - ``statistical-analysis-fundamentals`` — Distributions, hypothesis tests, confidence intervals - ``scientific-visualization`` — Matplotlib, seaborn, plotly, domain-specific plots - ``nlsq-core-mastery`` — JAX-accelerated non-linear least squares curve fitting - ``experiment-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, agents - ``llm-evaluation`` — Benchmarks, LLM-as-judge, human evaluation, quality metrics - ``langchain-architecture`` — LangChain/LangGraph: chains, agents, memory, tools - ``rag-implementation`` — Vector stores, chunking, re-ranking, hybrid retrieval - ``nlp-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/gRPC - ``model-optimization-deployment`` — Quantization, pruning, ONNX, TensorRT, mobile - ``ml-engineering-production`` — Type-safe code, testing, data pipelines, monitoring, drift - ``ml-pipeline-workflow`` — Airflow, Prefect, Metaflow, automated retraining - ``devops-ml-infrastructure`` — Docker, Kubernetes, GPU provisioning, cloud ML - ``federated-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 protocols - ``trajectory-analysis`` — MDAnalysis, RMSD, RDF, free energy, clustering - ``ml-force-fields`` — NequIP, MACE, DeePMD, active learning workflows - ``parallel-computing`` — MPI, OpenMP, Dask, Ray, scalability analysis - ``gpu-acceleration`` — CUDA, ROCm, JAX pmap, GPU-optimized algorithms - ``numerical-methods-implementation`` — Finite difference/element, spectral methods, iterative solvers - ``signal-processing`` — FFT, filtering, spectral estimation, wavelet transforms - ``time-series-analysis`` — ARIMA, state space models, changepoint detection - ``advanced-optimization`` — Genetic algorithms, simulated annealing, basin hopping - ``control-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 :doc:`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 code - ``python-development`` — Idiomatic Python, software engineering for science - ``python-packaging-advanced`` — uv workspaces, monorepos, reproducible builds - ``rust-extensions`` — PyO3/maturin high-performance Python extensions - ``type-driven-design`` — Protocols, Generics, static analysis with pyright/mypy - ``modern-concurrency`` — asyncio TaskGroups, threading, multiprocessing - ``quantum-computing`` — Qiskit, PennyLane, VQE/QAOA - ``bioinformatics`` — Genomics, proteomics, BioPython - ``computer-vision`` — Image processing, detection, Vision Transformers - ``reinforcement-learning`` — Gymnasium, Stable-Baselines3, RLlib - ``symbolic-math`` — SymPy, CAS, algebraic solvers - ``self-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 optimization - ``rlaif-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 inference - ``turing-model-design`` — Turing.jl: probabilistic models, Julia-native Bayesian workflows - ``variational-inference-patterns`` — ELBO, mean-field, normalizing flows, amortized inference - ``mcmc-diagnostics`` — R-hat, ESS, BFMI, trace plots, ArviZ convergence diagnostics - ``consensus-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 env - ``PreToolUse`` — Warn before commands that could corrupt simulations - ``PostToolUse`` — NaN/Inf check on compute job output (numerical integrity) - ``SessionEnd`` — Persist structured progress summary for next session - ``SubagentStop`` — Collect results from parallel science agents (``ExecutionError`` was removed in v3.4.0 — not supported by the CC v2.1.113 CLI event schema.)