@ashlr/hub 3.10.0 → 3.11.0

package/CHANGELOG.md

@@ -339,6 +339,162 @@ superseded, not a setup procedure.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.
 
+## [3.11.0] — 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
+other, raw timestamps and file paths where people expect words, the model named twice in the composer, and a Claude
+burn-down chart that forgot everything on every reload.
+
+### Charts
+
+- Axis labels no longer collide. When the start, "Resets …" and end labels don't fit, they shorten together
+  ("Fri 11:46 PM", then "Sep 18") before the lowest-priority label is dropped. Within one day, the axis shows clock
+  time only.
+- Ticks come from the step: a 0–1 axis reads 0, 0.2 … 1.0 rather than rounding 0.25 to "0.3", and count data gets
+  whole-number ticks.
+- Placeholder timestamps (epoch 0, anything before 2000) never set a chart's range, so the Growth harness chart no
+  longer starts on "Dec 31". A burst of readings shows as a cluster, not three stacked labels.
+- End-of-line labels move apart, or fall back to the legend. Heatmap digits are drawn without the halo that garbled
+  them. An all-zero funnel shows empty tracks.
+
+### Command
+
+- **Burn-down history survives a reload.** Seat window readings are recorded to
+  `~/.ashlr/routing/capacity-history.jsonl` (0600, at least 8 days kept, capped at 2 MiB, flat runs compressed) by
+  the daemon and by the Verse server every minute, and served at `GET /api/verse/budget/history`. The weekly
+  cards draw the whole window from it.
+- The Claude card shows its full weekly window on a 0–100% axis, with the even-pace line, projection and the 40%
+  reserve. The chart places Claude's reset, which Claude gives only in words ("Sep 25 at 6:59pm
+  (America/New_York)"), on the axis. It does this only for that exact form and only when the time falls within
+  one window of now; otherwise the words stay words.
+- Needs-you rows read like sentences: "Patch · Claude run", a title that ends on a whole word, "2 files · +384 −0",
+  "Test-and-repair loop" rather than "TITRR", and "38 days ago".
+
+### Fleet
+
+- "Why this seat" is one short sentence plus a held-back list. Reasons are structured on the server, so no more
+  ".;" joins and no raw ISO times. "Eligible again" shows the actual reopening time: a reset, "when you switch
+  autonomy on", or Claude's own words.
+- Lane chips read "Local · off", with the shared reason shown once.
+- "Fleet dark since" has one server definition: the last sign of fleet life, set only while the fleet is dark.
+  Every surface uses it. Growth says "No fleet runs or proposals since …" when it means that.
+
+### Chat
+
+- **Composer:** a seat chip names the account and one picker names the model. Nothing truncates; at narrow widths
+  the footer folds into icons and then into the ⋯ sheet. Effort shows only when the model supports it ("Effort:
+  High"). The context ring matches the header's and its tooltip gives exact tokens and the compaction point.
+- **Transcript:** a centred reading column. User turns are a quiet block on the right, and the assistant writes
+  plain prose. Duration and failure links sit in a muted turn footer. File and tool rows show project-relative
+  paths, with the full path on hover.
+- **Chat list:** folder names as they are on disk, with an explained "!" badge and ↑/↓/Enter navigation. "Disconnect
+  from hub" asks first, because it clears unsent drafts.
+
+### Accounts, approvals and panels
+
+- Every account row leads with its state ("Connected · usable now", "Spent · resets Sat 11:46 PM · usable again
+  in 1d 7h", "Signed out · reconnect to use it"), orders usable accounts first, and has a zero-cost **Check again**.
+- One rule for percentages and one prose tidier across the Usage, Budget, Health, Autonomy, Context, Dock and Git
+  panels. Raw ISO times, absolute paths, ".;" joins, "1 repos", untruncatable pills and icon buttons with no
+  tooltip are fixed throughout.
+
+### Shell
+
+- Only the current item in the left rail looks active. Each surface's data and code are warmed at idle after first
+  paint, so first visits skip the skeleton.
+- Keys pressed within a frame of opening the Settings menu no longer land on a different item from the highlighted
+  one, and Escape returns focus to the gear.
+
+### Hardening from review
+
+Before release, an eight-lens review ran over the whole change, with skeptics verifying each finding. All confirmed
+findings are fixed:
+
+- **Capacity history:**
+  - When it grows too large, compaction thins older readings per seat window instead of cutting every seat's first
+    days, so each series keeps its full 8 days.
+  - A Codex weekly-only window stays on its weekly line near its reset.
+  - `ASHLR_CAPACITY_HISTORY=0` now stops every recorder.
+  - The file is opened non-blocking, so a pipe at that path can't hang the server.
+- **Composer:** focus stays put when you change a setting in the ⋯ sheet or when the footer folds, Esc still stops a
+  turn, resizing never re-renders per pixel, and Send shows a sending state.
+- **Warm-up:** it waits for a visible, idle window, steps aside while chat reads are in flight, and never refetches
+  surfaces you haven't opened. Chat first paint stays at 367.1 KB, inside its 370 KB budget.
+- **Charts:** a placeholder reset is shown as unknown, not as "Resets now" with a green verdict. Label widths follow
+  the Display size setting. Heatmap digits and the funnel track meet contrast targets in both themes.
+- **Dates:** day buckets on the Mind, Growth, Usage and Spend charts use your local calendar day. Time-dependent
+  tests pass in every timezone and locale checked, and a test that would have failed on Sep 26 is fixed at the
+  source.
+- **Text:** the prose tidier no longer rewrites `..` inside paths or git ranges. Seat and model ids are shown by
+  name. One percent rule ("<1%", never a rounded "100%") applies everywhere a used share is printed.
+- **Accessibility:** the Disconnect warning, spoken diff stats, the header status chip and the icon-only mode picker
+  are all announced or reachable by keyboard.
+
+### Verification
+
+Checked locally, since GitHub Actions is off:
+- Build, both type-checks, eslint, the real-I/O lane, docs and the first-paint budget pass.
+- 4,476 web tests pass. The Verse and chart suites also pass under America/Los_Angeles and Asia/Tokyo
+  (2,382 tests each).
+- The backend suite passes 29,454 tests. Its 22 failures are the files that already fail on 3.9.1, plus four
+  load-timing flakes that pass when run on their own.
+- Rust and Swift code are unchanged since 3.10.0.
+
 ## [3.10.0] — 2026-09-24 UTC — autonomy with custody, and a workbench you can live in
 
 The fleet could propose but never finish: every change still waited on Mason, and the only way to let it merge