Knowledge base

Review, edit, and promote the durable pages a project keeps

Guide#user #knowledge-base Download Markdown

Every project has one knowledge base: a tree of markdown pages under /fs/ that outlives every session, shared by every agent with write access. You read and review it as a wiki under the project's Wiki tab. This page is the human's half; the page format, paths, limits, and the routes are in Knowledge base pages, and the agent's tools are in MCP tools.

The wiki

The tree is the whole knowledge base in one listing, drawn as a full tree on a desktop and one drill-in level on a phone. A page row carries its title and a mono line of its type, status, and derived trust; a stale page adds the word "stale", and a directory row carries its child count. Selecting a page renders it: the frontmatter block is metadata, shown as the page's type, status, and tags rather than as body text, and the backlinks are listed under it.

A page

A page is markdown with a YAML frontmatter block. The hub reads it, and the wiki shows type, status, and tags as metadata; title labels a listing, and stale_after is the date after which the page reads as stale. type is the one field a page needs. Any other top-level key is kept and ignored, so a page can carry fields of its own. A page with no block is stored as sent and flagged as needing one, so nothing is lost to a missing field. The full key list is in Knowledge base pages.

Review and trust

A page's trust is derived, never stored: unverified when nobody has reviewed it, human_reviewed when the newest review was yours, machine_confirmed when it was another agent's, and edited_since_review when the page no longer holds the bytes that were verified.

Review stamps the page you have read, carrying the version you read. If the page changed under you, the review is refused and nothing is stamped. Needs review lists the pages the hub judges not human-reviewed, and a row opens the page.

Editing a page

Edit opens the editor, whose Save carries the version token the read carried. A page that changed under the editor is refused rather than overwritten, and the refusal is shown in place with Reload theirs and Keep mine. New page writes a page that is not there yet.

History, revert, and restore

A page's History, opened from the reader's control row, lists its versions newest first with who, when, and what each did, and marks the version the page holds now. A version opens as a line diff against the page as it is now. Revert to this writes that version back after a confirmation; a revert deletes nothing, so it is not a danger action and can itself be reverted. A diff too long to work out line by line says so and shows the version itself.

A deleted page's address says it was deleted and links to its history, where Restore this writes it back the same way. Forget history, in the history's header, drops the kept bytes after a dialog that names the page; the rows stay, so the history still says who wrote what and when, but the versions can no longer be read or restored. Forgetting a history is an admin action, like prune.

Promoting session knowledge

A session's working state is pruned with the session, so durable knowledge is promoted out. A session brain entry in the session detail offers Save to wiki, which copies it into the knowledge base as a page that cites the session it came from and leaves the source where it was. An agent does the same with the brain_promote tool.

Housekeeping

The Wiki home carries the page, needs-review, and stale counts, and holds:

Recent changes, which reads who did what to which page and when, grouped by day, with rows that open that page's history;
Lint, the whole tree's findings, with a Re-check that walks it again.

What the wiki does not do

There is no rename or move: a move is a read, a write at the new path, and a delete, and it starts a new history because history is keyed by path. There is no rendered-compare merge; a conflict offers the hub's copy or yours. Pages carry no comment threads. Wiki links ([[page]]) are not links to the hub; only markdown links are followed. Offline reading is not shipped: what the service worker caches is the app shell and its assets, not page content, so reading pages needs a connection.

You should now see

The project's wiki, with a page tree, a reader that shows metadata and backlinks, and the housekeeping screens for review, history, and lint.