/** * Curated skill profiles for the pi-scientific-skills package. * * Data only. The taxonomy is reproduced verbatim from the package's published * curation so that the profile ids persisted in a user's config keep meaning the * same thing across releases: renaming an id silently invalidates saved selections. * * Nothing here touches skills/ on disk — profiles are applied by filtering the * package entry in settings.json, never by editing SKILL.md frontmatter. */ /** A named bundle of skills a user can switch on or off as a unit. */ export interface SkillProfile { readonly id: string; readonly label: string; readonly description: string; readonly skills: readonly string[]; } /** A skill deliberately left out of every profile, with the reason why. */ export interface UnassignedSkill { readonly skill: string; readonly reason: string; } /** Every skill shipped by the package. Used for the "vs. all skills" delta. */ export const TOTAL_SKILL_COUNT = 176; /** * Measured cost of one skill in the system prompt, as pi renders it: an * indented, XML-escaped `` block with the name, the description, and * the absolute path of the skill's SKILL.md (`formatSkillsForPrompt` in pi's * `core/skills.js`). The tags and the path are most of the difference from a * bare description count (about 88 tokens per skill): the model pays for them * every session all the same. * Recipe (2026-09-21): tiktoken over each skill's rendered block, with the * location set to the default npm install path under a generic home * (`/home/user/.pi/agent/npm/node_modules/pi-scientific-skills/skills//SKILL.md`), * across the real 162-skill catalogue — 99,296 chars, 23,282 tokens under * cl100k_base (143.7/skill), 23,194 under o200k_base (143.2/skill). * TOKENS_PER_SKILL is the rounded average of the two. A longer home directory * or a git install adds a few tokens per skill, and a model provider's own * tokenizer may count differently; this is an estimate, not a guarantee. * Descriptions stay in context permanently, so this is a per-session floor. * `scripts/validate.mjs` re-measures the same corpus on every run. */ export const TOKENS_PER_SKILL = 143; /** Context cost of loading the package unfiltered. */ export const BASELINE_TOKEN_COST = TOTAL_SKILL_COUNT * TOKENS_PER_SKILL; /** The ten field profiles, in menu order. */ export const PROFILES: readonly SkillProfile[] = [ { id: "core", label: "Core", description: "Stats, EDA, figures, dataframes, literature and writing — the ten skills nearly every scientific user reaches for regardless of field.", skills: [ "exploratory-data-analysis", "statistical-analysis", "polars", "scientific-visualization", "matplotlib", "experimental-design", "scientific-writing", "citation-management", "paper-lookup", "scientific-critical-thinking", ], }, { id: "genomics-bioinformatics", label: "Genomics & Bioinformatics", description: "NGS day: reads, alignments, variants, genomic intervals, bulk RNA-seq, phylogenies, gene/variant databases and pipeline orchestration.", skills: [ "pysam", "biopython", "scikit-bio", "genomic-coordinates", "polars-bio", "gtars", "geniml", "tiledbvcf", "deeptools", "primer-design", "qiime2-amplicon", "mageck", "bulk-rnaseq", "pydeseq2", "pathway-enrichment", "pacsomatic", "pathogen-variant-surveillance", "genomic-intelligence", "folklore-variant-evidence", "alphagenome", "waypoint-bio", "onekgpd", "phylogenetics", "etetoolkit", "gget", "bioservices", "cobrapy", "13c-metabolic-flux", "tellurium", "depmap", "ontology-term-resolution", "database-lookup", "nextflow", "datalad", "statsmodels", ], }, { id: "ml-data-science", label: "Scientific ML & Data Science", description: "Model-building day: scikit-learn to PyTorch, transformers, GNNs, RL, forecasting and explainability, plus scaling out to big data and GPUs.", skills: [ "scikit-learn", "pytorch-lightning", "transformers", "torch-geometric", "hugging-science", "shap", "umap-learn", "aeon", "timesfm-forecasting", "stable-baselines3", "pufferlib", "pymoo", "arbor", "dask", "vaex", "zarr-python", "polars", "optimize-for-gpu", "modal", "get-available-resources", "statsmodels", "pymc", "seaborn", "networkx", ], }, { id: "scholarly-communication", label: "Writing, Literature & Presentation", description: "Manuscript-and-deadline day: systematic search, reference management, drafting, peer review, grants, slides, posters, diagrams and document conversion.", skills: [ "literature-review", "research-lookup", "exa-search", "pyzotero", "citation-management", "scientific-writing", "peer-review", "scholar-evaluation", "research-grants", "venue-templates", "scientific-slides", "pptx-posters", "latex-posters", "markitdown", "liteparse", "markdown-mermaid-writing", "scientific-schematics", "infographics", "generate-image", "open-notebook", "hypothesis-generation", "scientific-brainstorming", ], }, { id: "single-cell-omics", label: "Single-Cell Omics", description: "scanpy-centric day: AnnData objects, QC to UMAP, batch correction and integration, RNA velocity, regulatory networks, cytometry and public census data.", skills: [ "scanpy", "anndata", "scvi-tools", "scvelo", "cellxgene-census", "deepspot-m", "arboreto", "umap-learn", "pydeseq2", "pathway-enrichment", "flowio", "flowkit", "lamindb", "zarr-python", "dask", "polars", "seaborn", "scikit-learn", "ontology-term-resolution", "gget", "bioservices", ], }, { id: "drug-discovery", label: "Drug Discovery & Computational Chemistry", description: "Med-chem and structural-biology day: molecules in and out, docking, molecular dynamics, protein models, mass spec, ADMET/PK and hosted modeling platforms.", skills: [ "rdkit", "datamol", "medchem", "molfeat", "deepchem", "torchdrug", "pytdc", "diffdock", "molecular-dynamics", "relion", "rowan", "tamarind", "esm", "glycoengineering", "biopython", "pkpd-modeling", "depmap", "primekg", "ncats-arax", "matchms", "pyopenms", "nmrglue", "shap", "scikit-learn", "pymoo", ], }, { id: "clinical-translational", label: "Clinical & Translational Research", description: "Patient-data day: EHR and cohort modeling, survival analysis, PK/PD, DICOM, biomedical/regulatory evidence lookup and safety-bounded clinical documentation.", skills: [ "pyhealth", "scikit-survival", "clinical-decision-support", "clinical-reports", "treatment-plans", "pkpd-modeling", "pydicom", "imaging-data-commons", "paperclip", "iso-standards-readiness", "analytical-method-validation", "statsmodels", "statistical-power", "pymc", "scikit-learn", "shap", "primekg", "ncats-arax", "database-lookup", "pathogen-variant-surveillance", "folklore-variant-evidence", ], }, { id: "physical-sciences", label: "Physics, Astronomy, Materials & Earth Science", description: "Simulation-and-instrument day: symbolic and numerical math, quantum stacks, CFD/PIV, materials structures, astronomy, geospatial and remote sensing, with units and uncertainty tracked.", skills: [ "astropy", "qiskit", "cirq", "pennylane", "qutip", "sympy", "matlab", "uncertainty-and-units", "fluidsim", "openpiv", "simpy", "pymatgen", "pycalphad", "cantera", "pybamm", "molecular-dynamics", "geomaster", "geopandas", "marine-carbonate-chemistry", "networkx", "pymoo", "dask", "zarr-python", "optimize-for-gpu", "modal", "get-available-resources", "statsmodels", ], }, { id: "imaging-neuroscience", label: "Bioimaging & Neuroscience", description: "Pixels-and-signals day: BIDS datasets, spike sorting, physiological time series, whole-slide pathology and radiology images, and the array stack that carries them.", skills: [ "bids", "datalad", "neuropixels-analysis", "neurokit2", "nwb-conversion", "pydicom", "imaging-data-commons", "histolab", "pathml", "cellprofiler", "deepspot-m", "omero-integration", "zarr-python", "dask", "aeon", "umap-learn", "scikit-learn", "pytorch-lightning", "optimize-for-gpu", "networkx", "statsmodels", "seaborn", ], }, { id: "lab-operations", label: "Lab Operations, Automation & Platforms", description: "Bench-and-core-facility day: ELN/LIMS and protocol systems, liquid handlers and cloud labs, instrument and image data management, compute platforms and compliance evidence.", skills: [ "benchling-integration", "labarchive-integration", "protocolsio-integration", "opentrons-integration", "pylabrobot", "ginkgo-cloud-lab", "adaptyv", "primer-design", "lab-hardware-cad", "omero-integration", "lamindb", "dnanexus-integration", "latchbio-integration", "nextflow", "iso-standards-readiness", "analytical-method-validation", "relsa-severity-assessment", "flowio", "flowkit", "pyopenms", "matchms", "uncertainty-and-units", "ontology-term-resolution", "database-lookup", "get-available-resources", "statistical-power", ], }, ]; /** * Skills held out of every profile. Kept as data (not deleted) so the reasoning * is auditable and so a future release can promote one without archaeology. */ export const UNASSIGNED: readonly UnassignedSkill[] = [ { skill: "pi-agent", reason: "About the pi harness itself (installing pi, providers, authoring skills/extensions/packages), not a scientific skill. Highly relevant to this package's own audience — surface it as a standalone toggle in the picker rather than burying it in a field profile.", }, { skill: "autoskill", reason: "Requires a screenpipe daemon on localhost:3030 to observe the user's screen, with no fallback, so it fails for essentially every user. Continuous screen observation is a privacy decision a user must make explicitly; it must never arrive as a profile default.", }, { skill: "hypogenic", reason: "Thin planning/audit wrapper around one academic package (ChicagoHAI HypoGeniC/HypoRefine) that explicitly disclaims manual hypothesis work. hypothesis-generation and scientific-brainstorming in scholarly-communication cover the same moment for everyone else.", }, { skill: "consciousness-council", reason: "Generic multi-perspective 'Mind Council' deliberation on any question or creative challenge. Not scientific method; duplicates scientific-brainstorming and scientific-critical-thinking with weaker evidence bounds.", }, { skill: "what-if-oracle", reason: "Generic branch/scenario ideation (best/likely/worst/wild case) with no scientific scoping. Same ground as scientific-brainstorming, which is already in scholarly-communication.", }, { skill: "dhdna-profiler", reason: "Extracts 'cognitive patterns and thinking fingerprints' from arbitrary text. No scientific workflow fits it and its scientific standing is questionable.", }, { skill: "bgpt-paper-search", reason: "Vendor MCP service for paper search. paper-lookup (11 literature APIs, in core) plus literature-review and paperclip already cover this; a fourth search vendor is pure token cost for users without a BGPT account.", }, { skill: "paperzilla", reason: "Proprietary recommendation/chat feed requiring a Paperzilla account. Redundant with paper-lookup and research-lookup for everyone else.", }, { skill: "parallel-web", reason: "Parallel CLI web search/enrichment. research-lookup already wraps Parallel and exa-search covers general scientific web search; keeping all three loaded is duplicate capability.", }, { skill: "usfiscaldata", reason: "U.S. Treasury Fiscal Data REST API — federal financial data. Not a scientific-research skill; off-topic for every profile.", }, { skill: "market-research-reports", reason: "Business market sizing (TAM/SAM/SOM) and forecast scenario decks. Commercial analysis, not scientific research.", }, ]; /** * Held-out skills that still deserve a one-click toggle. pi-agent is about the * harness itself rather than any scientific field, so it belongs to no profile — * but it is the single most useful skill for someone configuring pi, which is * exactly what they are doing when this picker is open. */ export const STANDALONE_SKILL_IDS: readonly string[] = ["pi-agent"]; /** * Toggle ids are persisted, so standalone skills are namespaced to guarantee they * can never collide with a profile id added by a later release. */ export const STANDALONE_ID_PREFIX = "skill:"; /** Standalone skills rendered as single-skill pseudo-profiles in the picker. */ export const STANDALONE_PROFILES: readonly SkillProfile[] = STANDALONE_SKILL_IDS.map( (skill) => ({ id: `${STANDALONE_ID_PREFIX}${skill}`, label: skill, description: UNASSIGNED.find((entry) => entry.skill === skill)?.reason ?? "", skills: [skill], }), ); /** Everything the picker can toggle: field profiles first, standalone skills last. */ export const TOGGLES: readonly SkillProfile[] = [...PROFILES, ...STANDALONE_PROFILES]; /** * Inert default export. Pi discovers extensions as `extensions/*.ts` as well as * `extensions/*\/index.ts`, so this data module may be loaded as an extension in * its own right. Registering nothing keeps that harmless instead of erroring on a * missing default export. */ export default function noopExtension(): void {}