Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,12 @@
- **The view stays where you were reading.** Pages land beneath the divider when you read from the top, and above your viewport when you read the tail. The divider survives the Tools / Thinking / AI / User filters, because it is how the rest of the session is reached. Snapshots from older daemons still render their prose marker the same way.
- **Verified:** headless browser run on a real 1144-message Claude Code session: 1210 not loaded on open, one auto page on scroll (1015 left, no cascade), one click (813 left), Load all to the session start with the divider removed; 17 transcript tests pass.

### Fix: the hosted Quality tab shows your real grade instead of "Nothing to grade yet" (carries #5121) (2026-09-02)
- **Who this reaches:** every hosted user who opened Quality on app.clawmetry.com and saw an empty week under a machine that had graded runs all week. Same node, same runtime: the local tab said 117 tasks with 4 rough runs; the hosted tab said there was nothing to grade.
- **Two faults, both fixed.** The grade never rode the encrypted snapshot, so the cloud had nothing to serve. And the hosted container answered anyway from a DuckDB of its own, empty but perfectly able to reply, so a working machine was reported as idle. The sync daemon now bakes the report card into the snapshot, node-wide and one card per runtime (a quiet runtime still gets a card saying so in its own name, never another runtime's grade under its filter), and the hosted process no longer reads its own store for this tab: with no slice it says the grade is computed on your machine, which is true.
- **Cheap on your node.** Read once, compose many: three store reads per snapshot cycle whatever the runtime count, calibration carried once per slice instead of inside fourteen cards (85 kB to 61 kB), and one report-card composer shared by the daemon and the request path so local and hosted never disagree.
- **Verified:** 10 new tests pin the slice shape, the read count, the quiet-runtime rule, and the unreachable-vs-empty distinction; 67 existing quality tests pass; live snapshot for the founder node decrypted to confirm the slice arrives. The hosted interceptor ships in the matching cloud release (clawmetry-cloud #2245).

### Feature: the dashboard grows up visually, and light mode is back (carries #5420) (2026-09-01)
- **Who this reaches:** everyone who opens the dashboard, and especially the person seeing it for the first time while deciding whether to trust it. The shell now reads like the enterprise consoles it sits next to: a flat near-black dark theme with hairline borders instead of blue-tinted panels and a glow, and a sidebar whose 10+ entries finally group under plain-word section labels — Observe (Agents, Activity, Sessions), Analyze (Cost, Quality, Harness Engineering), Govern (Approvals, Alerts, Notifications).
- **Consistent icons, quieter selection.** Navigation icons are now one matched set of line icons instead of mixed text glyphs, and the selected entry is marked with a calm raised state rather than a loud blue pill. Nothing moved behind your back: every tab keeps its name, its tooltip, and its deep link.
Expand Down
Loading