The browser now holds the window and scrolls inside itself. One bar names where the reader is, a second names the repository's sections, and a third carries whatever the open section is looking at, so no single row carries three jobs. Source and a commit each read as two panes: what to open down the left, what is open down the right. Opening a file keeps its neighbours in view. Checks group by the commit they ran against, headed by that commit's subject line rather than its sha, with a filter naming each workflow and how many runs it holds. A run has a page of its own, addressed by the record key the runner wrote it under, so one can be linked to. The badge beside a commit keeps the newest run of each workflow and folds them: a single failure decides the commit, whatever its other workflows say. Before this the newest run of any workflow spoke for all of them, so a publish that finished later hid a failed test. Sections come from the repository's config record, one cheap read, so every page carries the way back to every other. Packages, checks, files and an empty repository were each a dead end. Surfaces are panels divided by rules rather than tiles: one radius, no filled controls, identifiers in mono. Two gauges under the file pane read the repository as a whole, what it is written in and how lately it was worked on, both from walks the page already made. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Gp5VgVe393AeJL5jKARqR3