Answer from your Yandex Wiki knowledge base

The agent works with the knowledge base through the API at api.wiki.yandex.net: it walks the tree of sections and pages and reads texts, attachments and dynamic grids. On that basis it answers from your internal policies and always names the source page, and in a section audit it shows what has not been updated in a long time and collects a list of duplicates — a question the Wiki interface only answers by paging through the tree by hand.

A Wiki page is addressed by its path: the agent requests it by slug and lists the fields it needs explicitly, because the API does not hand over the whole content by default. The practical consequence: if a page has been moved to another section its address changes, and the agent searches the tree and the title again instead of concluding that the document is gone. The link it puts in the answer points to the page the fact came from.

On the write side the agent creates a page in the right section, lays out outcomes and decisions point by point, attaches the files from the discussion, refreshes the rows of a dynamic grid on a schedule and adds a summary at the end. Editing a page replaces its content, so before saving the agent re-reads the current version and lays the new text over that rather than over a stale copy. Deleting subpages is confirmed with you separately.

The agent sees exactly what the connected account sees, and it edits only the pages that account may write to. It walks the section tree in portions instead of pulling the whole knowledge base down: the table of contents first, then the content of the pages and attachments it actually needs, so a large section does not turn into hundreds of requests in a row. It holds the request rate itself, and it calls a restricted page restricted instead of answering as though it did not exist.