Documents & drafting
Ingestion, proposals and exports exist inside product-worker. Template discovery and native tracked changes remain targets.
Document ingestion and export run in services/product-worker/internal/documentjobs, using packages/documents-go and the separate services/document-parser process. There is no independently deployed document service.
flowchart LR Upload["Authorized upload"] --> Job["Product worker document job"] Job --> Parser["Parser and OCR adapter"] Parser --> Version["Extracted version and private generation"] Version --> Context["Task context"] Context --> Proposal["Proposal for exact base version"] Proposal --> Accept["User acceptance creates version"] Accept --> Export["Export exact version"]
Implemented behavior
Section titled “Implemented behavior”The parser bounds its input and returns structured output so malformed files cannot consume an unbounded amount of worker capacity. Scanned documents use OCR; empty extracted text is not treated as a complete read. A document becomes searchable only after its embedding batch is complete.
An AI proposal is tied to the version it edited, so a late proposal cannot overwrite a newer manual change. Exports use the selected version rather than whatever version happens to be current when the download starts. Storage & indexes explains publication and cleanup; Workers & recovery explains how late jobs are rejected.
Target capabilities
Section titled “Target capabilities”Template catalog selection, section-by-section drafting with a shared outline, block-level editing, native Word tracked changes and external Word merge are not established by the current renderer. They require implementation and executable fixtures before being described as available. In particular, producing a DOCX file is not evidence that Word accept-all/reject-all semantics work.