analyze · git:20261002.a2bb40c · 2026-10-02 · sha256 db9936be93ca2ed0
analyze git:20261002.a2bb40cA
Immutable. This exact content never changes and is served at /api/v1/blob/db9936be93ca2ed0.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_m37lx45no5k44lew/db9936be93ca2ed028f6a8a554987c92ebfdf8ede9b18d2205b0d8e84340e86f.json
Then /plugin install analyze@markdownregistry-m37lx45no5k44lew in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: analyze
description: >
Run computational analysis using the repository's registered datasets,
conventions, and validation requirements. Use for clustering, PCA/UMAP,
differential expression, survival or time-series models, statistical tests,
sensitivity analysis, continuing prior analyses, or debugging unexpected
analytical results. Do not use for merely previewing data, ingesting files,
writing reports, brainstorming, repository setup, or unrelated code fixes.
---
# Mycelium — Analyze
Resolve bundled `skills/` and `network/` paths relative to the Mycelium plugin
root—the ancestor containing `.codex-plugin/` and `.claude-plugin/`. Resolve
analysis and `.living/` paths relative to the user's repository.
Start a new analysis or continue an existing one, routing to the appropriate convention packs installed in this repository.
## Steps
1. **Read manifests** to understand the project:
- `analysis/ANALYSIS_MANIFEST.md` — existing analyses
- `data/DATA_MANIFEST.md` — available data
- `algorithms/ALGORITHM_MANIFEST.md` — available algorithms
2. **Continuing existing**: Navigate to `analysis/[name]/` and read its documentation file (UPPER_SNAKE_CASE of folder name, e.g., `SNP_ANALYSIS.md`).
3. **New analysis**: Create `analysis/[name]/` with documentation file (UPPER_SNAKE_CASE.md), `scripts/`, `outputs/`, `reports/`. If building on a parent analysis, record the lineage in the manifest entry.
4. **Route to installed conventions**:
- Read `.living/conventions/ACTIVE_CONVENTIONS.yaml` to see what's installed.
- **Always consult** the core references: `skills/core/references/analysis-conventions.md` (structure) and `skills/core/references/statistical-conventions.md` (methodology). These are the baseline.
- **If `robust-analysis` is installed** (check `.living/conventions/robust-analysis/`): Follow its conventions as the primary execution guide. Read `.living/conventions/robust-analysis/analysis-conventions.md` as the entry point — it links to detail files for strict execution, validation checks, sensitivity sweeps, null hypothesis testing, and adversarial probing. These conventions are non-negotiable when installed.
- **If a domain convention is installed** (e.g., `bioinformatics`, `image-analysis`): Consult its `analysis-conventions.md` for domain-specific methodology. Domain conventions layer on top of robust-analysis.
- **If multiple conventions apply**, follow the cascade: repo-local (`.living/conventions.md`) > domain > core.
5. **Execution standards**:
- Use marimo for exploratory work, plain Python scripts for reproducible pipelines.
- Every analysis must have a `run.sh` or `run.py` that reproduces final outputs.
- **Register reportable values with `register_value`** at the site that computes them. Any number or short phrase the analysis would expect a future report to quote (sample counts, thresholds, contrast phrases, headline metrics) goes through `skills/core/scripts/register_value.py`. The call writes `analysis/<name>/outputs/numbers.json`; the report's Phase-1 merge picks up the fragment automatically, and `scitexlintr` then catches drift between the manifest value and any quoted snapshot in the `.tex` source. Full guide: `skills/core/references/report-values-guide.md`. Skip this only for values the report will never cite.
6. **Lint with `scilintr` after writing or modifying analysis code** (non-negotiable). This is a fast static analyzer that flags silent scientific commitments — things that look like ordinary software (broad `except`, positional indexing, unchecked joins, unannotated filters, label leakage, magic thresholds, unseeded RNGs, etc.). The core principle is: *scientifically meaningful choices must not be anonymous.* Read `skills/core/references/scilintr-guide.md` before invoking it the first time in a session — it explains the philosophy, the rule catalogue, and (critically) the `ANALYSIS_OK[category]: explanation` waiver mechanism. Workflow:
- **Check the version once per session**, for the implementation the analysis uses. Python: `python3 skills/core/scripts/check_linter_versions.py scilintr` reports the installed version and the latest on PyPI; it never installs, and if it is missing, install with `pip install scilintr`. R: the Python check does not apply (the R package runs as `Rscript -e 'scilintr::main()'`); `Rscript -e 'packageVersion("scilintr")'` gives the installed version, and `Rscript -e 'op <- old.packages(repos = "https://cloud.r-project.org"); cat(if ("scilintr" %in% rownames(op)) paste("update available:", op["scilintr", "ReposVer"]) else "no newer CRAN release", "\n")'` checks CRAN without installing; if it is missing, install with `install.packages("scilintr")`. If an update is available, tell the user and offer the upgrade — run it only with their agreement, since a newer linter can add rules and change what passes mid-analysis.
- **After every code write or edit**, run `scilintr <file_or_dir>` on the touched files. The CLI is fast (sub-second on most files), so there is no reason to defer.
- **Drive findings to zero**, either by fixing the pattern (preferred — most findings have a cleaner code path) or by adding a structured waiver (`# ANALYSIS_OK[category]: explanation; where the check/assertion lives`). A waiver must answer *what / why scientifically valid / where it is recorded*. Vacuous waivers (`# ANALYSIS_OK: fine`) are not acceptable.
- **Before declaring the analysis complete**, run `scilintr` on the full analysis directory (e.g., `scilintr analysis/<name>/`) to catch cross-file rules (duplicate-parameter-source, shadow-overwrite, definition drift).
- This is complementary to `/mycelium:review`, not a substitute. Lint catches structural patterns cheaply; review catches semantic and statistical issues.
7. **Run the post-action hook protocol** after significant steps (see `/mycelium:core` for the full protocol):
- Update `analysis/ANALYSIS_MANIFEST.md`
- Update the analysis documentation file
- Log decisions to `.living/decisions.md`
- Log learnings to `.living/learnings.md`
- Run `skills/core/scripts/validate_structure.py`