/redact-content

/redact-content

Apply human-guided redaction and obfuscation to a document or a batch of documents. This is not an automated PII scanner — the user specifies what to protect and which technique to use.

Arguments

$ARGUMENTS — path(s) to input file(s) plus optional flags:

• --mode=<alias|random|pattern> — default alias.
• --mapping=<path> — path to an alias mapping JSON (default: aliases/active_mapping.json).
• --scrub-metadata — also strip EXIF / document metadata from the output.
• --out=<path> — override output directory (default: output/).

Procedure

1. Resolve input files. If a directory is given, process every supported file inside (txt, md, pdf, docx, odt, images).

2. Load or create the alias mapping:

   • alias mode — maintain aliases/active_mapping.json. Reuse existing aliases so the same real name always maps to the same pseudonym across runs.
   • random mode — generate a unique per-document mapping, write it to aliases/<document-slug>.json. Do not reuse across documents.
   • pattern mode — use the pattern set in schemas/patterns.json (emails, phone numbers, IDs, addresses). Replace with labelled placeholders.

3. Ask the user for any new real-name → alias pairs not yet in the mapping. Confirm before writing.

4. Apply redaction:

   • Text/Markdown: string replacement with word-boundary regex.
   • PDFs: prefer pdftotext + re-render via Typst or qpdf for a text-layer rewrite. If true visual redaction is required, flag that to the user and recommend a purpose-built tool — this plugin does not guarantee visual-layer redaction.
   • Images: replace with metadata-scrubbed copy; do not attempt visual redaction of image content unless the user has drawn boxes in a side-car file.

5. If --scrub-metadata is set, run exiftool -all= on the output copies (never the originals).

6. Write outputs to the output/ directory (or --out). Never overwrite the input.

7. Emit a redaction log to logs/redaction-<timestamp>.jsonl with: input path, output path, mode, mapping file, count of replacements, any warnings.

Invariants

• Originals are immutable. Every redaction writes a new file.
• Alias mapping files are sensitive — remind the user not to commit them publicly if the workspace is a public repo.
• In random mode, never reuse identifiers across documents.
• Pattern mode is best-effort; always recommend human review for legally sensitive output.