Shared document-to-text extraction. One script, no state: reads an input file, prints JSON to stdout, writes nothing to disk (PHI-safe — no caches, no temp files; callers own any caching).

cd <this skill dir> && bun install

This pulls liteparse (the bin, used for PDF/DOCX/XLSX/PPTX, OCR included) and rtf-to-text (RTF). Without it, PDFs still work via a fallback if poppler is installed; other binary formats require liteparse.

bun <this skill dir>/scripts/extract.ts <input-file> [--content-type <mime>]

Output on stdout:

{ "text": "...", "method": "liteparse | pdftotext | rtf-to-text | passthrough", "pages": 12 }

Tables with multiple value columns (option A vs option B, in-tier vs out-of-tier) can interleave columns line-by-line in the extracted text: fragments of adjacent cells alternate, and a cell's text can even land mid-sentence inside a neighboring column. Values usually survive, but which column a value belongs to can become ambiguous. When an answer comes from one column of a multi-column table and the document has no redundant restatement of the value elsewhere, verify it by reading the original page directly before treating it as ground truth. The extracted text's anchor tells you which page: pass it to the Read tool's parameter (e.g. pages: "37") to render just that page to vision instead of the whole document.

Import the functions instead of shelling out when you're already in bun TS:

import { extract, resolveLit } from "../doc-extract/scripts/extract";
const lit = resolveLit([myRoot]); // also checks myRoot/node_modules/.bin/lit
const text = extract(lit, "/path/to/file.pdf"); // string | null

The contracts skill consumes it this way (its ingest caching stays on the contracts side).