Ship it everywhere, and know that it landed.

Drawing a connection

Distribution is a canvas. On one side are the things you can ship: documents you wrote or uploaded, and skills. On the other are the destinations: your Spaces, the repositories bound to them, and the people in them. You drag from one to the other.

There is no publish button, and there is no confirm step. The arrow you draw is the record, so a connection is live the moment you make it and gone the moment you delete it. That is worth knowing before your first drag: this is a canvas that saves as you move, not a diagram you commit at the end.

Seeing the page at all needs knowledge.distribute, which owners, admins and curators hold at the workspace level. A Space admin does not get it, because every control here writes across the whole workspace. See who sees what, and who decides.

What lands, and where

Harbor writes whole files at its own paths. It never writes a region inside a file you authored, and it never edits your CLAUDE.md or AGENTS.md. It used to, through a channel that wrote a managed block inside a customer’s file, and that channel was deleted; the cleanup that strips those old markers is still running.

Beside them sits a .milestone-managed manifest listing the files Harbor put there, with a checksum for each. Only files on that list are ever removed. A markdown file you drop into .claude/docs/ yourself is left alone forever, and the directory survives as long as anything of yours is in it.

There is no pull request, and that is deliberate

A file does not arrive as a pull request for you to approve. The daemon on each machine writes it into the working tree, which means it shows up in git status alongside whatever else you are doing, and it becomes part of the repository when somebody commits it. Nothing Harbor runs commits or pushes on your behalf.

Harbor did open pull requests, and that design was dropped once the arithmetic was clear: one document into twenty repositories is twenty review requests, and every edit to that document is another twenty. Nobody approves that, so the file would simply stop being updated, which is the exact failure the feature exists to prevent.

Whether it actually landed

Drawing an arrow is a statement of intent. Whether the file is on somebody’s machine is a separate question, and the honest answer to it is often we cannot tell yet. Harbor answers with one of seven states rather than a checkmark.

Only delivered is green, and it cannot be earned by a row existing. It requires a machine to have reported holding those exact bytes, recently enough that the report still counts. A connection you drew thirty seconds ago is pending, not delivered, and a machine that stopped reporting takes its files to stale rather than leaving them green. Four of the seven states are grey on purpose: they are not findings about your files, they are statements about the limits of what Harbor watched.

This is the same discipline as served versus cited. A tool that renders every unproven thing as a tick is a tool you cannot use to answer a question that matters.

What this does not do

Two things it would be reasonable to expect, which are not what ships:

• It does not curate the document’s contents for you. Nothing in the distribution path reads cite rates, adds rules that proved useful, or removes rules agents ignored. What ships is the document as you last left it. The evidence-driven part is real, but it is on Knowledge, as two lists of candidates for removal that a person acts on.
• It does not reach a machine nobody has set up. Delivery rides on the CLI. A teammate who has never run harbor init has no route, and the canvas says so once for the workspace rather than greying out every card.