From a scanned page and a recording to a searchable document

Loomcycle versions v1.122 through v1.124 add pure-Go extraction of DOCX, XLSX, PPTX, PDF, scanned pages and audio recordings into Markdown without the model ever holding the source bytes, the project announced. The release builds a Document tagged untrusted with pictures stored as image chunks by page and slide, routes text-layer-free pages through vision-model OCR that both operator and agent must opt into with page caps, a ledger purpose and a dry-run cost estimate, and sends recordings through any OpenAI-audio-shaped whisper server cut into overlapping pieces under minute budgets. Search now filters by path, document, type, status, tag and as_of act inside every retrieval leg rather than after it, with chunks carrying validity periods and documents_scope keeping notes and references apart.

What lands on a document-based reference system is DOCX, XLSX, PPTX, PDF, scans and recordings. loomcycle v1.122 to v1.124 reads them without the model ever holding the bytes: pure-Go extraction to Markdown with bounded zip and XML parsers, PDF text layers with CMaps and Cyrillic, and an import that builds a Document tagged untrusted with pictures as image chunks by page and slide. Pages with no text layer go through OCR by a vision model, opted into by both operator and agent, with page caps, a ledger purpose and a dry-run cost estimate. Recordings go through any OpenAI-audio-shaped whisper server, cut into overlapping pieces under minute budgets. Search filters by path, document, type, status, tag and as of act inside every retrieval leg rather than after it, chunks carry validity periods, and documents scope keeps notes and references apart.