Sennen

	Sennen is a Claude Code / Codex plugin for ML workflows. It provides reusable skills for repo diagnosis, planning, data access, joining meetings, visualization, train/test splits, metrics, baselines, experiments, review, data defect analysis, preprocessing, and explanation. It is designed for Git, MLflow experiment tracking, DVC data versioning, and scientific data sources.

Sennen Commands

In Claude, skills are exposed as /sen-* and /db-*. In Codex, skills are available as $sen-* and $db-*.

High-Level Commands

• sen-data - connect and version scientific datasets
• sen-plan - define the prediction task, split strategy, and metric contract
• sen-experiment - run the first baseline or compare later experiments
• sen-review - review the ML approach

Specialist Commands

• sen-defects - find defects in the data itself
• sen-visualize - inspect data structure and distributions
• sen-split - create leakage-safe splits and test for leakage
• sen-metrics - define metrics, MLflow tracking, and result views
• sen-preprocess - clean data and engineer features
• sen-explain - explain model behavior and feature importance
• sen-join - join a Google Meet URL in guest mode
• sen-doctor - diagnose and repair repo readiness for the Sennen workflow

Example: Clinical Trial Enrollment Predictor

Goal: predict whether a clinical trial is still open to enrollment 12 months after launch.

Step	Command	Action
1	/sen-data	ingest all ClinicalTrials.gov interventional studies
2	/sen-data	load the data into a flat modeling table
3	/sen-plan	define the month-12 target, leakage-safe holdout, and metric contract
4	/sen-visualize	plot the enrollment survival curve
5	/sen-experiment	build a baseline predictor and corresponding survival plots
6	/sen-experiment	improve the baseline with stronger text-based models
7	/sen-experiment	try a sentence embedding model and compare it against the sparse baseline
8	/sen-explain	explain the most important variables in the best model
9	/sen-review	critique the workflow, results, and next actions

Data Versioning and Experiment Tracking

Sennen treats data versioning and experiment tracking separately:

• DVC tracks downloaded and generated datasets in data/raw/, data/processed/, and optionally data/splits/.
• MLflow is the supported experiment metrics backend.
• setup manages Python dependencies through uv and pyproject.toml only.

Install

git clone https://github.com/20minds/sennen.git
cd sennen

./setup /your/data/folder

cd /your/data/folder

claude --plugin-dir plugins/sennen

OR

codex

Repo-local setup installs:

• Claude plugin files into plugins/sennen
• Codex skills into .agents/skills

Global setup installs:

• Claude skills into ~/.claude/skills
• Codex skills into ~/.codex/skills

If you want Claude without --plugin-dir, use standalone project commands instead:

./setup --claude-standalone /your/data/folder

This installs project commands into /your/data/folder/.claude/commands/ and the matching skill bundles into /your/data/folder/.claude/skills/, so Claude can load them automatically when you start claude in that repo.

setup now also asks whether to install:

• the optional db-* scientific database skills
• the optional sen-join conference-call skill

If you choose sen-join, setup will look for TWENTYMINDS_API_KEY in the environment or repo .env, and it can prompt you to save one. The skill is still installed even if no key is available yet.

You can also run it non-interactively:

./setup . --yes
./setup /your/data/folder --yes
./setup --codex /your/data/folder --yes
./setup --codex /your/data/folder --git --uv --dvc
./setup --codex /your/data/folder --git --uv --mlflow
./setup --codex /your/data/folder --no-git --no-uv --no-dvc
./setup --claude /your/data/folder --yes
./setup --claude-standalone /your/data/folder --yes

Project Layout

Sennen uses normal repo folders for project artifacts:

.
├── config/
│   ├── exp/
│   │   ├── exp_001_baseline.yaml
│   │   ├── remix.yaml
│   │   └── tracking.yaml
│   ├── metrics/
│   │   └── metrics.yaml
│   ├── prepare/
│   │   └── pipeline.yaml
│   └── split/
│       └── split.yaml
├── data/
│   ├── raw/
│   ├── processed/
│   ├── results/
│   │   └── exp_001_baseline/
│   └── splits/
├── models/
│   ├── checkpoints/
│   └── exports/
├── results/
│   └── exp_001_baseline/
│       ├── metrics.yaml
│       └── summary.md
│   ├── data_quality/
│   ├── explanations/
│   ├── figures/
│   ├── review/
└── src/
    ├── data/
    │   └── data_001_ingest.py
    ├── exp/
    │   └── exp_001_baseline.py
    ├── lib/
    ├── prepare/
    │   └── prep_001_pipeline.py
    └── visualize/
        └── vis_001_inspect.py

Those folders are conventions used by the skills. They are created by project work as needed, not by setup.

Suggested use:

• config/exp/exp_001_baseline.yaml stores the first baseline as a normal experiment config
• later baselines can be added as config/exp/exp_00x_<label>.yaml with kind: baseline
• use reference_role: primary to mark the default baseline comparison anchor when multiple baselines exist
• config/exp/tracking.yaml stores MLflow tracking conventions such as sqlite:///mlflow.db
• results/exp_001_baseline/summary.md and later results/exp_00x_<label>/summary.md files store human-readable experiment summaries
• results/exp_001_baseline/metrics.yaml and later results/exp_00x_<label>/metrics.yaml files store Git-friendly machine-readable summaries such as metrics and runtime
• src/lib/ stores shared utilities that should not live in numbered workflow scripts
• src/exp/ stores sortable experiment entrypoints such as exp_001_baseline.py
• data/results/exp_00x_<label>/ stores large DVC-friendly prediction and evaluation artifacts
• the default Git-friendly experiment summary pair lives under results/exp_00x_<label>/

When Sennen downloads data into data/raw/ or materializes outputs into data/processed/ or data/splits/, those datasets should be tracked with DVC when DVC is available. Prefer uv run dvc add data/ only when data/ is intentionally one DVC-managed artifact boundary; otherwise track narrower directories. For experiment metrics, use MLflow to track runs in sqlite:///mlflow.db.

High-Level Workflow

These high-level commands are the recommended entry points:

Command	Purpose	Maps to lower-level commands
/sen-plan	Define the task, leakage-safe evaluation setup, and metric contract	/sen-doctor, /sen-defects, /sen-split, /sen-metrics, optionally /sen-visualize
/sen-data	Ingest, inspect, and version data	/sen-data, optionally /sen-defects, /sen-visualize, and relevant db-* skills
/sen-experiment	Run the first baseline when none exists, or run stronger models against the baseline afterward	/sen-experiment, optionally /sen-preprocess, /sen-visualize, /sen-explain
/sen-review	Critique the approach, results, and next steps	/sen-review, optionally /sen-experiment, /sen-visualize, /sen-explain

/sen-doctor is a specialist diagnosis and repair command. Use it directly when you want to prepare or validate repo structure, DVC readiness, or MLflow readiness before higher-level workflow steps.

/sen-join is a specialist utility command. It sends a request to the 20minds meeting-bot endpoint for a Google Meet URL and is intentionally separate from the main ML workflow commands.

Scientific Database Skills

Sennen also includes scientific database skills for:

Skill	Purpose
db-alphafold	Retrieve AlphaFold predicted protein structures, coordinates, and confidence metrics by UniProt ID.
db-biorxiv	Search bioRxiv preprints by topic, author, category, or date and retrieve metadata or PDFs.
db-chembl	Query ChEMBL compounds, targets, and bioactivity data for SAR and medicinal chemistry work.
db-clinicaltrials	Search ClinicalTrials.gov trials by condition, drug, status, phase, or NCT ID.
db-clinpgx	Access ClinPGx pharmacogenomics data for gene-drug interactions, allele functions, and CPIC guidance.
db-clinvar	Retrieve ClinVar variant significance and pathogenicity annotations for genomic interpretation.
db-cosmic	Query COSMIC somatic mutation, gene fusion, and cancer gene data for oncology research.
db-drugbank	Access DrugBank drug properties, targets, interactions, pathways, and pharmacology data.
db-ena	Retrieve ENA sequence records, raw reads, and assemblies by accession for genomics workflows.
db-ensembl	Query Ensembl genes, variants, orthologs, sequences, and comparative genomics data.
db-fda	Search openFDA drugs, devices, adverse events, recalls, and regulatory submission data.
db-gene	Retrieve NCBI Gene annotations, identifiers, locations, phenotypes, and related records.
db-geo	Search and download GEO expression and genomics datasets such as GSE, GSM, and GPL records.
db-gwas	Query GWAS Catalog variant-trait associations, p-values, and genetics evidence.
db-hmdb	Access HMDB metabolites, chemical properties, biomarker data, spectra, and pathways.
db-kegg	Query KEGG pathways, gene mappings, drug interactions, and identifier conversions.
db-metabolomics-workbench	Access Metabolomics Workbench studies, RefMet data, spectra, and metabolite metadata.
db-opentargets	Query Open Targets target-disease associations, tractability, safety, and drug evidence.
db-pdb	Search and download RCSB PDB macromolecular structures and metadata.
db-pubchem	Query PubChem compounds, properties, similarity, substructure, and bioactivity data.
db-pubmed	Search PubMed literature with Boolean and MeSH queries and retrieve article metadata.
db-string	Query STRING protein-protein interactions, networks, and enrichment results.
db-uniprot	Access UniProt protein records, FASTA sequences, and ID mapping workflows.
db-uspto	Search USPTO patent, trademark, assignment, and prosecution history data.
db-zinc	Access ZINC purchasable compounds and 3D-ready structures for screening and docking.