mlmm_toolkit

Install recipes for mlmm-toolkit core + MLIP (UMA / MACE / Orb / AIMNet2) / DFT / xTB / AmberTools backends, with CUDA / PyTorch / e3nn / aarch64 quirks. TRIGGER on install / setup / `pip install` / `conda env` / `ImportError` / CUDA-version mismatch / "GPU not detected" / `huggingface` auth / e3nn version conflict / AmberTools / `tleap` not-found questions. SKIP when mlmm imports cleanly and the user is invoking subcommands — the CLI skill covers usage.

Where the source code lives in `mlmm-toolkit`. 6 physical layer directories (`cli` / `workflows` / `domain` / `backends` / `io` / `core`) + 3 repo-internal forks (`pysisyphus` / `thermoanalysis` / `hessian_ff`). Tells an agent which directory to grep for a given concern (Click option, ONIOM stage runner, MLIP backend, hessian-ff analytical MM Hessian, output writer, chemistry default, link-atom math) before touching code. TRIGGER on questions like "where is X implemented", "which file defines flag Y", "how is the repo organised", "what's safe to refactor". SKIP for usage questions — those belong to `mlmm-cli` / `-overview`.

Per-subcommand reference for mlmm-toolkit's 22 CLI subcommands (extract / mm-parm / define-layer / opt / tsopt / freq / irc / dft / scan / path-search / all / …). SKILL.md is the orientation + cross-cutting flag conventions + canonical recipes; each subcommand has its own md (extract.md / mm-parm.md / tsopt.md / …) for flags, validation, and caveats. TRIGGER on questions about a specific subcommand, flag, or shell invocation. SKIP for install / HPC / output-parsing / structure-format-editing / overview questions.

Fallback skill for detecting the current compute environment (local / PBS / SLURM, CPU architecture, GPU, CUDA, conda env). Use only when the env is unknown — other skills assume placeholders are filled by the user or context.

PBS (Torque / PBSPro) and SLURM submission for mlmm-toolkit — generic preamble templates with placeholders, walltime budgeting, CPU vs GPU choice, job monitoring, and the dynamic-dispatch (`flock` + `pbsdsh`) recipe in `dynamic-dispatch.md`. TRIGGER on cluster submission / `qsub` / `sbatch` / walltime / preamble / multi-job dispatch / `pbsdsh` / `flock` / many-system batch questions. SKIP for local single-machine runs, install setup, or output parsing. Note: mlmm-toolkit runs as a single-GPU job per invocation.

How to drive `mlmm-toolkit` from an MCP-speaking agent (Claude Desktop / Claude Code / Cursor / Codeium / custom Python or TypeScript MCP SDK clients) via the bundled `mlmm-mcp` server. Lists the 22 MCP tools (one per CLI subcommand, including mlmm-specific topology / ONIOM-layer / ONIOM-input tools) and the shared `SubcmdResult` schema. TRIGGER on questions about MCP setup, agent integration, tool names, return-value shape, or "how does an agent invoke mlmm". SKIP for direct CLI usage — that's `mlmm-cli`.

PDB / XYZ / GJF / Amber parm7+rst7 input-file reference for mlmm-toolkit, plus the charge / multiplicity decision workflow and the B-factor layer encoding (ML=0 / movable-MM=10 / frozen-MM=20) that defines a three-layer ONIOM system from a single prepared PDB. TRIGGER on editing or inspecting a structure file, deciding `-q` / `-l` / `-m`, interpreting residue / charge / spin in an input, or assigning B-factor layers. SKIP for subcommand syntax, output parsing, install, or HPC questions.

Decision know-how for ML/MM enzyme reaction-barrier campaigns — precision (fp32 vs fp64 by GPU class), TS-candidate routes (path-search MEP vs distance-restrained scan), fixing wrong imaginary-frequency counts at TS-opt (fp64 / --coord-type dlc), reading a barrier when the scan started from Product (reverse direction), staged vs concerted scans, and the same-atom-set rule for controlled mutant-vs-WT comparisons (B-factor layer transplant + --detect-layer). TRIGGER on "barrier", "imaginary frequency", "wrong saddle", "fp64 / precision", "MEP vs restraint", "scan from product", "staged vs concerted scan", "mutant comparison", or "controlled experiment". SKIP for install / pure structure-format editing / MCP-transport questions.

Output parsing and multi-step workflow selection for mlmm-toolkit — `summary.json` schema, R/TS/P/IM canonical paths, `bond_changes` interpretation, energy-diagram conventions, and the cluster + 1-step / multistep / scan-list / endpoint-MEP / TS-only / DFT//ML/MM / stage-by-stage (subcommand-only, gate each stage) recipes used to extract numerical results. TRIGGER on output parsing (`summary.json`, `result.json`, `seg_NN/`), extracting barriers / ΔE / Gibbs for a paper, choosing between multi-input / scan-list / endpoint-MEP / TS-only modes, or running the pipeline subcommand-by-subcommand with a success check at each stage (instead of one `all` run). SKIP for single-subcommand syntax (CLI skill) or install / HPC questions.

Orientation for mlmm-toolkit — what it is, when to use it, and how it differs from generic ML/MM MD packages (three-layer ML/movable-MM/frozen ONIOM via PDB B-factor encoding, analytical hessian_ff full-system Hessian, microiteration with link-atom Jacobian coupling, AmberTools-driven MM parameterization). TRIGGER on first-touch / "what is mlmm-toolkit" / "should I use it" / "how does it compare to OpenMM / GROMACS / Sire" questions. SKIP when the user has already named a subcommand, an install issue, an output file, a structure format, or a cluster — sibling skills cover those.