dsh-r-workbench
hougelangley/dsh-r-workbench
R data workbench with six tools: ingest CSV/TSV/text, Excel and xlsx-compatible WPS .et, ODS, JSON/NDJSON and ZIP files; clean and infer types; score whether the data can support analysis (PROCEED, CONDITIONAL or STOP); run group comparisons, correlations, association tests and calendar-month trends with effect sizes and BH correction; and write self-contained HTML, Markdown, Quarto source or DOCX reports. STOP blocks analysis and reporting.
설치
dsh plugin --profile web add github:hougelangley/dsh-r-workbenchREADME
dsh-r-workbench
A DeepSeek Harness bundle that turns the local R toolchain into six model-facing tools: acquisition, cleaning, a data-value gate, statistical analysis, reporting, and an R console.
The gate is the product: when the cleaned data cannot support a meaningful analysis, r_assess returns STOP and the analysis/report stages refuse to run. See docs/DETAIL.md for the user-facing behavior and rules.
Install
dsh plugin --profile web add github:HougeLangley/dsh-r-workbench
Then call r_env first. The plugin needs a local R installation; r_env reports missing packages and prints the exact Rscript r/install-deps.R command. RStudio and Quarto are optional and detected when present.
Layout
index.js Host plugin: six tools, run-directory lifecycle, R subprocess runner
r/common.R Shared R helpers (JSON I/O, hashing, HTML/Markdown, fonts, stage wrapper)
r/env.R Toolchain inspection
r/ingest.R Readers + cleaning
r/assess.R Six-dimension value gate
r/analyze.R Tests, effect sizes, BH correction, tables, charts
r/report.R HTML/Markdown/Quarto/DOCX composition
r/run.R Ad-hoc R console
r/install-deps.R User-library installer for required/optional packages
cordis.patch.yml Bundle patch (inserts the `r-workbench` row)
locale/{en,zh}.json Plugin Manager metadata
assets/icon-180.png Raster icon for storefronts
Stage contract
index.js invokes Rscript --vanilla r/<stage>.R <config.json> with the run directory as cwd. The config carries runDir, runId, pluginVersion, and stage-specific options. Each stage writes logs/<stage>-result.json:
{ "ok": true, "stage": "assess", "data": { "...": "..." }, "elapsedSec": 0.4 }
The Node tool reads only data; ok: false carries error and is surfaced as a tool failure. Artifacts live in the run directory, never only in the tool result, so a later turn can inspect them.
Gate enforcement
r_assesswritesassessment.jsonwith the SHA-256 ofdata/clean.csv.r_analyzeandr_reportcallrw_gate(), which rejects a missing/stale assessment, rejectsSTOP, and marksCONDITIONALruns as exploratory.r_reportadditionally requiresanalysis.jsonfor the same data hash.
Install
From a source checkout of this directory:
plugin_manager({ action: "install_bundle", target: "/absolute/path/to/r-workbench" })
The bundle has no runtime npm dependencies; it uses Node built-ins and the local Rscript/R installation. R packages are installed into the user library by Rscript r/install-deps.R.
Testing
tests/run-tests.sh exercises the R stages directly against generated fixtures (valuable CSV, GB18030 text with preamble, Excel with dates, JSON, ZIP, WPS-compatible .et, a 20-row CONDITIONAL sample, and a deliberately low-value workbook). It verifies:
- ingest/merge row counts, encoding detection, leading-zero preservation, and multi-sheet skip disclosure;
PROCEED,CONDITIONAL, andSTOPdecisions, including exploratory labeling;STOPrefusal by bothr_analyzeandr_report;- tampered
clean.rdsrejection (the gate is bound to both data artifacts); - calendar-month slope on a dataset with missing months;
- HTML report content and DOCX
word/document.xmlstructure when pandoc is available.
The real acceptance check is calling the installed tools from the Harness, not this script.