Agent workflow

For filing, extract text and call company_document_classify. One Jev request judges two groups of categories without children, each with an other option. Parent categories declare their semantic group in the template; children inherit it. One category wins directly; two category answers are compared by their selected-option probabilities. Two other answers or a probability tie become X. Register the document with the returned category and hash, then upload the original under its file name and content.txt beside it. A file belongs to a registered document, so its path names that document.

Storage is the combined result of upload and registration, not another step.

For finding, list by category or search permitted documents. Read summaries first, then text previews, then originals as needed. The requester identity passes through to row-level and file-store policies.

The agent also looks without being asked. Every recall asks company_document_recall, as the requester, for the documents a request names by title or counterpart, and for those whose meaning is near it, and reads one sentence for each beside its memory. Nearness comes from an index the host keeps in its own service state: every few minutes it lists, as each requester, the documents that changed and embeds their sentences. The index only orders candidates. What the agent reads is asked again as the requester, so the next recall sees a revised summary or a circle that no longer reads the category, and a request that only shares a document's topic stays below the bar.

A member reads through their circles, so giving a colleague access means placing them in a circle. Sharing outside the company lends what one circle reads to the recipient's email and returns their invitation link; dataroom_invitation_send emails it to them. A recipient whose own company runs internkim can accept through their agent with dataroom_invitation_accept. A circle names parent or intermediate codes. An agent edits a circle or adds a child only after checking the people whose access changes. Publication requires an explicit publishing request.

The dataroom skill ships this workflow. Its standalone helper builds a tree from the generated template and preserves originals, sidecars and catalogs. Legacy tree exports require semantic reclassification.