@ashlr/hub 3.10.1 → 3.11.2

package/CHANGELOG.md

@@ -339,6 +339,157 @@ superseded, not a setup procedure. Use [runtime activation authority](docs/RUNTI
 and the [current architecture boundary](docs/ARCHITECTURE.md#legacy-fleet-activation-boundary).
 Neither the historical record nor a successful test activates a resident fleet.
 
+## [3.11.2] — 2026-09-25 UTC — the Ashlr.AI mark, a lighter first paint, and a release loop measured in minutes
+
+3.11.1 was tagged but never published to npm. 3.11.2 includes everything in it.
+
+### The Ashlr.AI mark
+
+Verse now carries the Ashlr.AI keystone "A" in its rail, the macOS app icon, the menu-bar icon and on verse.ashlr.ai
+(the header and the favicon). The mark was traced from the brand's own file, `ashlar-landing`
+`public/logos/ashlar-mark.png`: navy `#1A2B3C` legs around an electric-blue `#2563EB` core. In the app, the legs take
+the ink colour so the mark reads in light and dark themes. The menu-bar icon is a dedicated template image, so it
+follows the macOS menu bar.
+
+### Faster first paint, built by Verse itself
+
+Chat's first-paint JavaScript fell from 369.9 KB to 349.7 KB, meeting the 350 KB target set in 3.10. The build now
+fails above 352 KB. The work came from Verse's own cloud self-improvement (#478).
+
+### A release loop measured in minutes
+
+- **`npm run gate`** runs the static checks in parallel with caching, then only the tests your change can reach
+  (backend and web), plus a smoke set that must always pass. Known environmental failures are listed with reasons
+  and reported, never hidden. `npm run gate:full` runs everything.
+- **`npm run ship:local`** builds, packs and installs the release as your CLI. It updates Ashlr.app with the
+  sidecar and web assets (and the native shell with `--native`), keeping the previous build aside, never deleting
+  it. It then re-signs the app, restarts the background services and waits for Verse to answer. `--dry-run` shows
+  every step.
+- See `docs/RELEASING-LOCALLY.md`.
+
+### Also from self-improvement
+
+- **#480:** the local production gate no longer fails on the `desktop/src-tauri/gen/` folder a Tauri build leaves
+  behind, while a truly unexpected file still fails it.
+- **#477:** fixes Dependabot alert #32. The vulnerable `rustls` in the desktop app is updated, and there's a
+  regression test.
+
+### Resource bar fix
+
+The rail's resource bar now includes the budget view, as the drawer does. Claude and Grok batteries show their real
+windows instead of an empty "—".
+
+### Verification
+
+Verified with the new gate itself. `npm run gate` against master took 5m24s for this 59-file change and passed: all
+static checks, first paint at 349.7 KB, 3,170 related web tests and 4,756 related backend tests (the rest were
+skipped). Separately, the full web suite (4,649 tests) and the Rust tests (173) pass.
+
+On the way, the gate caught two things, both fixed. A backend test imported a function the first-paint work had
+moved. And `compaction-point.ts`, split out of `context-math.ts`, joined the authority import closure; that
+growth was reviewed and accepted.
+
+## [3.11.1] — 2026-09-25 UTC — you can see who you're running on, all the time
+
+### Provider logos
+
+Every place Verse shows an engine now shows the provider's own mark: Anthropic's Claude, OpenAI, xAI's Grok and
+Ollama. That covers account tiles in the drawer and Apps & Accounts, the composer's seat chip, the chat header,
+⌘K, the tasks pane and chart lane labels. The account's name always sits beside the mark as text. The marks come
+from @lobehub/icons (MIT) and are inlined, so they need no network. They load with their own small chunk, never on
+chat's first paint.
+
+### Always-on resource bar
+
+The rail foot now carries a live bar with one battery per resource, showing what's left of each account's binding
+window. It covers Claude, both Codex accounts, Grok, your local models and your cloud credits. Colour tells you
+usable, low or spent. Hover or focus a row to see every window with its reset, the share kept for you, and when a
+spent account comes back. Click to open the Resources drawer. With labels shown, each row reads the full account
+name over a battery and value. With labels hidden, the rows become upright batteries under each logo. **Hide
+resource bar** is in the drawer's footer and in ⌘K, and it brings back the single capacity ring.
+
+### Verse keeps working in the background
+
+The main window no longer lets macOS suspend the page when it's hidden (macOS 14+). Until now, opening Verse
+behind another full-screen app could leave it on "Checking for an existing Verse session…" for minutes. Chats,
+Needs-you banners and resource readings also froze until you looked at it again.
+
+### Built by Verse, on its own
+
+The cloud lane's first self-improvement pull requests are in this release:
+- **#475** fixes the eight Universe campaign-recovery tests that had failed since 3.9.1. That includes a real bug:
+  a campaign whose run hit its deadline before a winner was picked was recorded "failed" instead of "completed".
+  The files that only run on macOS were verified on this Mac.
+- **#476** brings the release-policy tests (m522, m440, m468, m454, m231, npm-cli-launch) up to the current manual
+  release process. Their safety rules are kept as explicit checks, tested in both directions.
+
+The estimated cloud spend is calibrated to the real balance: $210 of $250 left.
+
+### Verification
+
+Checked locally:
+- Build, both type-checks, eslint, the real-I/O lane, docs and first paint (369.9 KB of 370) pass.
+- 4,619 web tests pass.
+- The backend suite passes 29,815 tests. The one failure is `m571`, which trips on the Tauri `gen/` folder a local
+  desktop build leaves behind; fixing that is queued in the self-improvement backlog. Every other failure that had
+  been open since 3.9.1 is fixed.
+- 173 Rust tests pass.
+
+## [3.11.0] — 2026-09-25 UTC — the cloud lane, and every resource one keystroke away
+
+Verse can now spend Claude cloud credits. It launches Claude Code cloud sessions, which keep running on your
+credits after the subscription window is spent. It tracks what each one delivers, and it uses them to improve
+itself within limits you set. Every account, local model and credit balance now sits in one drawer on the right
+of every page.
+
+### Cloud lane
+
+- **Launch from anywhere.** The chat composer's ⋯ sheet has **Run in cloud**. Command has **New cloud task** and
+  **Improve Verse**. The CLI has `ashlr cloud launch "<task>"`.
+- **How a launch works.** Verse starts each session through the Claude seat's own claude.ai login, from an isolated
+  checkout under `~/.ashlr/cloud/checkouts`, never your working copies. It marks only those checkouts as trusted.
+  The session works on `ashlr-cloud/<task>` and opens a **draft PR** with a machine-readable report. Verse follows
+  the PR with `gh`: running, then PR open, then merged or closed. Nothing in the cloud lane merges; that stays with
+  the custody gates or you.
+- **Verse improves itself.** A built-in backlog of real gaps, plus suggestions from the Leader's memos, feeds
+  self-improvement launches. The defaults are 4 a day, on `ashlrai/ashlr-hub`, pausing below $40. You can switch
+  it off in the Cloud card, in Usage, or with `ashlr cloud budget --self-improve off`. Setting
+  `ASHLR_CLOUD_AUTO=0` disables the scheduler.
+- **Spend is an estimate, and says so.** Claude doesn't expose the credit balance, so Verse counts sessions at a
+  per-session estimate you can adjust. It always links to your real balance on claude.ai.
+- **Surfaces.** The Cloud card on Command, a Cloud credits panel in Usage, a "Cloud · N running" chip on Fleet,
+  Needs-you items for ready PRs and failed launches, and `ashlr cloud list | refresh | improve | budget |
+  backlog`. See `docs/CLOUD.md`.
+- Verified end to end on this Mac: launch, session, draft PR, parsed report, closed.
+
+### Resources drawer
+
+- **On every page.** An edge tab, a rail button, **⌘.**, or "Show resources" in ⌘K opens a drawer on the right
+  with every resource you run on. That's your Claude, Codex and Grok accounts, with window meters, the reserve
+  kept for you, reset times and Reconnect / Check again. It also shows local models, with runtime state and
+  verified context windows, and your cloud credits.
+- **Overlay or pinned.** It opens as an overlay or pins as a column beside your work, and remembers which you
+  chose. A dot on the edge tab says at a glance whether everything is usable, something is tight, or an account
+  is spent or signed out.
+- It loads after first paint, so chat still opens as fast as before.
+
+### Website and docs
+
+- verse.ashlr.ai, the README and `docs/VERSE.md` now describe 3.9–3.11: verified context windows, live reasoning,
+  autonomy with custody, budget modes, the workbench, the four surfaces, burn-down history and the cloud lane.
+  `docs/CLOUD.md` is new.
+
+### Verification
+
+Built as six Claude Code cloud sessions running on Claude credits, then integrated and checked locally (GitHub
+Actions is off):
+- **Static checks:** build, both type-checks, eslint, the real-I/O lane, docs, and first paint (369.3 KB of 370).
+- **Web suite:** 4,615 tests pass.
+- **Backend suite:** 29,772 tests pass. The only failures are files that already fail on 3.9.1.
+- **Cloud lane:** 312 tests, plus a real launch on this Mac that ran from session to draft PR to parsed report.
+- **Custody guards:** they flagged the cloud modules joining the authority import closure. That growth was
+  reviewed and accepted, and those files are now owner-lane protected.
+
 ## [3.10.1] — 2026-09-24 UTC — the workbench reads cleanly, and the charts tell the truth
 
 3.10.0 landed on Mason's screen. This release fixes what showed up in the live app: labels piled on top of each