sixb-ai/sixb/documents

The Sixb agent image has ReportLab, pypdf, pdfplumber, python-pptx, python-docx, openpyxl, pandas, PyArrow, NumPy, matplotlib, Pillow, PptxGenJS, LibreOffice, Poppler, Pandoc, qpdf, and DejaVu/Liberation/Noto fonts.

Use python3 for Python scripts. For JavaScript slide authoring, use a .cjs script with const pptxgen = require("/opt/sixb/js/node_modules/pptxgenjs"); this works from any workspace. Use the user's source facts, formatting, and template. Keep requested charts and tables editable. Preserve distinctions such as proposed versus approved costs and response versus completion times.

For digital PDF tables, use pdfplumber and verify extracted rows and totals against the source. For scans, use pdf2image or pdftoppm followed by pytesseract/Tesseract; English OCR data is installed. OCR output can misread numbers and punctuation. Check source page images before using uncertain values. pypdf and pdfplumber alone do not OCR image-only pages.

Pandoc can read DOCX as Markdown or create DOCX from Markdown. For PDF output, render the generated DOCX with sixb-render; this image does not include a TeX engine. qpdf --check checks PDF structure, not visual layout. Use pandas/PyArrow for Parquet data; keep source units and null values intact.

After authoring, run:

sixb-render draft.pptx --output-dir previews-1

The helper accepts PDF, PPTX, DOCX, and XLSX. It creates a PDF, page PNGs, and a JSON manifest. Always use a new output directory after revisions. Rendering has a 30-page default limit; use --first-page and --last-page for larger documents and inspect every required page across batches. The manifest's complete field says whether previews cover the entire document. Pass --timeout (seconds per conversion step, default 60) when a large document times out.

Inspect the PNG paths with view_file when image input is supported. Check clipping, overlap, contrast, fonts, chart labels, and units. Repair and render again when needed. Text extraction alone does not verify layout. A model without image input can check text/structure but must not claim visual inspection. LibreOffice is a compatibility check; it does not prove exact PowerPoint layout.

For spreadsheets, inspect formulas and values separately from printed-page previews. openpyxl does not calculate formulas. Use LibreOffice on a copy when recalculation is needed and verify the result.

Create previews inside the workspace so view_file can read them, and delete a preview directory once you have inspected it: the workspace can outlive the turn. Keep source scripts when another turn may need to edit the document. When you have publish_artifact, deliver each requested file with it as a file artifact; otherwise return the file as your instructions say. Preview PNGs and validation logs are working files unless the user requests them.