The plan for Claudeboard was written on 13 October and the owner’s instruction for the first slice was two words: go, parallelize. So the slice was cut in half and given to two builders at once.
The collector
The first half reads what is already on the machine: the desktop app’s session store, one record per session, and the command-line tool’s process files, which say which sessions are busy and which are idle. It folds them into a local index and serves the resulting graph, nodes and edges, over a live connection to whatever is drawing it. Run against the real machine, it resolves well over a hundred nodes and every live session the app’s own sidebar lists.
The board
The second half is the page: a canvas that draws the graph as a force layout, with sessions as nodes and the flows between them as edges. It was built against the collector’s contract rather than the collector, so the two could be reviewed and tested apart.
The rule that got written down
Version one serves the page locally and the owner wants to glance at it from the phone over a private link. The reviewer’s decision, recorded here because it will govern everything after it: content never leaves the machine. A client that is not on the machine itself gets metadata, names, states, counts, and nothing else. Status text is stripped before it goes out.
Both halves are open as pull requests, waiting on the owner. A new manager session took over the project today; the first picture of the board against the real sidebar follows the merge.