* feat(cloud-backup): mirror fleet snapshots to S3-compatible storage
Add an Admiral-tier Cloud Backup feature that replicates every fleet
snapshot to off-site storage, with two provider modes that share the
same `@aws-sdk/client-s3` code path:
- Sencho Cloud Backup: zero-config, 500 MB allowance backed by
Cloudflare R2, provisioned via the sencho.io worker against the
user's Lemon Squeezy license.
- Custom S3 (BYOB): any S3-compatible bucket (AWS, MinIO, Backblaze
B2, Wasabi, R2 with own keys), with credentials encrypted via
`CryptoService` before storage.
API-triggered snapshots upload fire-and-forget so the UI returns
immediately; scheduled snapshots block on the upload so the task's
success/failure reflects cloud durability. Object keys include the
instance_id segment to prevent collisions when the same Admiral
license is activated on multiple Sencho instances.
* fix(cloud-backup): drop ES2022-only Error cause arg breaking ES2020 build
The backend tsconfig pins lib to ES2020. The two-argument
`Error(message, { cause })` form requires ES2022, so tsc rejected it
with TS2554. Revert to single-argument throw to match the
convention used elsewhere in the backend services.
* feat(files): backend foundation for stack file explorer
Install multer for multipart file upload handling. Add
isValidRelativeStackPath to validation.ts to guard client-supplied
relative paths against traversal, absolute paths, NUL bytes, backslash
injection, and double-slash segments. Add isBinaryBuffer to a new
binaryDetect.ts utility for heuristic text/binary detection via
NUL-byte fast exit and non-printable byte ratio sampling.
* fix(files): reject bare dot segments in isValidRelativeStackPath
* feat(files): add safe stack-scoped file I/O methods to FileSystemService
Adds FileEntry interface and seven new public methods to FileSystemService
for stack-scoped file operations: listStackDirectory, readStackFile,
streamStackFile, writeStackFile, deleteStackPath, mkdirStackPath, and
statStackEntry.
Each method routes through a private resolveSafeStackPath helper that
enforces two-phase path containment: a pre-realpath lexical check plus a
post-realpath symlink-escape check. ENOENT targets are handled by walking
up to the deepest existing ancestor, realpaths that ancestor, and
reattaching the remaining suffix.
Binary detection delegates to isBinaryBuffer; path safety delegates to
isPathWithinBase. Protected file names and the MIME map are module-level
constants to avoid repeated allocation.
* feat(files): frontend API wrappers and Monaco language helper
* fix(files): tighten stackFilesApi error handling and localOnly support
* fix(files): FileSystemService safety and correctness fixes
* feat(files): add file explorer API endpoints to stacks router
* feat(files): FileTree and FileTreeNode components
* fix(files): route security hardening and stream cleanup
* fix(files): FileTree accessibility, icon stroke, stale fetch guard
Add strokeWidth={1.5} to all Lucide icons in FileTreeNode to match the
design system. Add aria-expanded to directory rows for accessibility.
Guard handleDirClick .then() callbacks against stale stack name
references when the component re-renders with a new stack. Add
toast.info fallbacks when compose.yaml or .env is clicked without a
navigation callback registered.
* feat(files): FileViewer, FileUploadDropzone, NewFolderDialog, DeleteFileConfirm
* fix(files): resolve code quality findings in file explorer components
- Move editorOptions useMemo above conditional returns in FileViewer (Rules of Hooks fix)
- Fix blob download: append anchor to DOM before click, defer URL revoke 100ms
- Keep protected-file confirm input visible during NOT_EMPTY recursive retry in DeleteFileConfirm
- Remove non-functional cursor-pointer/onClick from Community upgrade pill in FileUploadDropzone
- Add success toast on folder creation in NewFolderDialog
- Switch all (e as Error).message casts to instanceof Error narrowing
* test(files): unit tests for binary detection, stack path safety, and file explorer routes
- binary-detection.test.ts: covers isBinaryBuffer edge cases (empty, NUL,
PNG header, threshold boundary, sampleBytes parameter)
- filesystem-stack-paths.test.ts: covers isValidRelativeStackPath (accepts/
rejects matrix) and FileSystemService stack methods against a real temp dir
(listStackDirectory sort and protection flags, readStackFile text/binary/
oversized paths, writeStackFile/Buffer, deleteStackPath, mkdirStackPath,
traversal guard); platform-specific empty-dir/NOT_EMPTY cases skip on Windows
- stack-files-routes.test.ts: route-level integration tests for all seven
file explorer endpoints; covers auth gating, Community-tier 403 gates,
input validation, 413 TOO_LARGE upload limit, and 204/200 happy paths
* feat(files): StackFileExplorer container with lazy tree, viewer, and action bar
* fix(files): add Download button to explorer toolbar, fix Community upgrade pill, reset state on stack change
* test(files): add missing test coverage for file explorer routes and service
* feat(files): add Files tab to EditorLayout with StackFileExplorer integration
* fix(files): add defensive activeTab guard to saveFile and discardChanges
* test(files): unit tests for FileTree expand/collapse and FileViewer render modes
Covers the three FileViewer content modes (text/Monaco, binary panel,
oversized panel) and the FileTree expand/collapse/cache cycle: first
expand fetches the subdirectory, second click collapses without a fetch,
third click re-expands from the in-memory cache without a second fetch.
* test(e2e): file explorer community and skipper+ flows
Covers the full file-explorer feature surface in two describe blocks:
Community (read-only): intercepts /api/license to simulate community
tier, confirms the upgrade pill is visible in the left pane, and
asserts that the Save button is absent after opening a text file.
Skipper+ (full CRUD): uploads a text file and confirms it appears in
the tree; edits config/app.conf and saves via Monaco; deletes an
uploaded file and asserts the tree entry is gone; issues a raw HTTP
request to the download endpoint and checks for status 200 and the
content-disposition: attachment header.
Also adds data-testid="file-action-delete" to the action bar Delete
button in StackFileExplorer for stable targeting, and exports
waitForStacksLoaded from e2e/helpers.ts to eliminate the three
identical local copies in stacks, deploy-log-panel, and stack-files
spec files.
* fix(e2e): improve test isolation and selector stability in stack-files spec
Move beforeEach seed to beforeAll/afterAll so fixtures are created once per
suite, not before every test. Extract shared seedSuite/teardownSuite helpers
to eliminate the duplicate beforeAll/afterAll blocks. Wrap teardown in
try/catch so failures log a warning rather than masking test results.
Replace waitForTimeout(500) with a deterministic expect on the file tree
sentinel. Add data-testid="anatomy-files-btn" and data-testid="delete-confirm-btn"
to replace the fragile button text/positional selectors. Assert Save button
starts disabled before editing.
* docs(files): add stack file explorer documentation
Add user-facing guide for the stack file explorer feature covering
tier access (Community read-only, Skipper+ read-write), viewing
limits, upload/download caps, protected file routing, and
troubleshooting. Update the editor page to reference the new guide
and register the page in the navigation.
* fix(docs): use canonical Skipper tier name in file explorer overview card
* fix(files): resolve lint errors blocking CI
Remove unnecessary backslash escape before double-quote in the
Content-Disposition regex (no-useless-escape). Replace five synchronous
setState resets at the top of the FileTree mount effect with a React key
prop on the FileTree element in StackFileExplorer so remounting resets
state automatically, eliminating the react-hooks/set-state-in-effect
violation.
* test(files): fix e2e seeding to work on community-tier CI
Replace the browser-side paid upload/mkdir API calls in seedTestStack with
direct Node fs writes. The upload and folder endpoints require Skipper+ so
they returned 403 on CI, which runs with no license set. Stack creation
via POST /api/stacks stays as an API call since it is community-allowed and
keeps the backend registry in sync.
Add a per-test tier check in the Skipper+ beforeEach that skips gracefully
when the instance is community, matching the pattern in auto-heal-policies.
* feat(notifications): dispatch deploy_failure alert on stack action errors
* feat(terminal): add onReady and onMessage callback props
* feat(deploy-logs): add DeployLogContext with runWithLog API
* feat(deploy-logs): add DeployLogPanel bottom drawer with resize and minimize
* feat(deploy-logs): wire DeployLogContext to App and EditorLayout action runners
* test(deploy-logs): add E2E test for deploy log panel open, failure, and minimize
* docs(deploy-logs): add user-facing and internal architecture docs
* feat(deploy-logs): redesign as opt-in modal with structured log rows
Replace the full-width bottom drawer (DeployLogPanel) with a centered
modal that streams structured log output for deploy, stop, restart,
update, install, and Git apply operations. The modal is disabled by
default; users opt in from Settings -> Appearance.
Core changes:
- New DeployFeedbackContext with runWithLog() API: if opt-in is off,
silently bypasses the UI so all call sites degrade to the existing
toast behavior without code changes.
- composeLogParser.ts: pure parser that strips ANSI escapes and
classifies compose output into stage badges (PULL, BUILD, CREATE,
START, STOP, DOWN, WARN, ERR, LOG). 15 unit tests.
- StructuredLogRow.tsx: memoized row with timestamp, stage badge, and
message. Error rows get a rose left rail; warn rows get a tinted bg.
- DeployFeedbackModal: Dialog-based, max-w-640px/max-h-70vh, elapsed
timer, auto-close 4s on success (hover cancels), persistent on
failure. Raw xterm output collapsible in footer.
- DeployFeedbackPill: minimized state anchored top-right, survives
navigation, click restores modal.
- Wires App Store install (action: install), Git apply (action: deploy),
and Git pull (action: update) in addition to the existing EditorLayout
actions.
- Fixes Terminal.tsx WS URL in generic mode (was connecting to root path
not proxied by Vite; now uses /ws).
- Settings: adds "Show deploy progress modal" checkbox to Appearance.
- Docs: renames deploy-logs.mdx to deploy-progress.mdx; updates
internal architecture doc.
* fix(deploy-logs): connect Terminal in generic mode and move pill to bottom-center
Terminal was passed stackName which routes it to the stack logs WS
(container stdout). In that mode onReady is never called, so the
deployStarted gate never resolves and the compose command never runs.
Remove stackName so Terminal uses generic WS mode, which calls onReady
on open and streams compose output.
Also reposition the minimized pill from top-right to bottom-center
(fixed bottom-6 left-1/2 -translate-x-1/2) per UX feedback.
* docs(deploy-logs): update pill position to bottom center
* test(deploy-logs): rewrite E2E spec for deploy feedback modal
The old spec targeted the removed bottom-drawer DeployLogPanel and used
the wrong field name when calling POST /api/stacks (sent 'name' but the
endpoint reads 'stackName'), causing every test to fail with a 400 before
any UI assertions ran.
Fixes:
- POST /api/stacks body now uses 'stackName' matching the API contract
- All locators updated to target the new DeployFeedbackModal and
DeployFeedbackPill components (data-testid attributes added)
- Added enableDeployFeedback helper to opt-in via localStorage before
each test that expects the modal (feature is off by default)
- Added opt-in OFF test to confirm the modal is suppressed when disabled
- Minimize/expand test now asserts the pill appears and contains the
stack name before clicking to restore the modal
* test(deploy-logs): fix compose file write endpoint in E2E helper
createStackViaApi was calling PUT /api/stacks/:name/files/docker-compose.yml
which does not exist. The correct endpoint is PUT /api/stacks/:name with
{ content } in the body.
* test(deploy-logs): use addInitScript to persist opt-in across reloads
The opt-in flag was set via page.evaluate before setupDeployStack, which
calls page.reload() and loginAs (a second navigation). Although localStorage
should persist across same-origin reloads, the React tree was reading
'false' on remount in CI. Switching to addInitScript guarantees the
localStorage value is set before any page script on every navigation, so
useDeployFeedbackEnabled's useState initializer always sees the right
value when React mounts.
* test(deploy-logs): verify localStorage and re-dispatch event before deploy
Adds syncDeployFeedbackState() called right before each deploy click in
the ON tests. It both verifies localStorage is set (failing the test
loudly with a clear message if not) and re-dispatches the
SENCHO_SETTINGS_CHANGED event to defeat any stale React state after
navigation. If the modal still does not appear with the assertion green,
the issue is downstream of localStorage and we have a clear signal.
* test(deploy-logs): wait for React re-render after dispatching opt-in event
After syncDeployFeedbackState dispatches SENCHO_SETTINGS_CHANGED, React
schedules the state update but does not flush it synchronously. The
click that follows can fire against the stale closure where isEnabled is
still false, so runWithLog takes its early-return path and the modal
never opens. A 200ms wait is enough to let React commit the new state
before the next interaction.
* test(deploy-logs): wait for stack file fetch before clicking deploy
deployStack() in EditorLayout returns early at 'if (!selectedFile)'
without calling runWithLog. selectedFile is set inside loadFile() after
GET /api/stacks/:name resolves. The previous setup clicked the stack in
the sidebar and immediately asked the test to click Deploy, racing the
fetch. CI backend logs confirmed no deploy POST ever fired for the ON
tests, while the OFF test passed only because it asserts non-existence.
Now setup awaits both the stack click and the file response together,
then verifies the action bar's deploy button is visible before returning.
* test(deploy-logs): wait for network idle and capture browser logs
Adds a networkidle wait plus a 500ms settle after the stack click so
React commits selectedFile and any follow-up env/container/backup
fetches drain before the deploy click. Also mirrors browser console
errors and pageerrors into the Playwright output so the next failure
ships with the React stack trace instead of just a 'modal not visible'
message.
* test(deploy-logs): temporary debug logging in runWithLog
Adds a console.log at the entry of runWithLog so we can see in CI logs
whether it is being called and what isEnabled value the closure has.
Also widens the test's console capture to include these debug lines.
This is diagnostic only and will be removed once the root cause of the
modal-not-opening-in-CI failure is identified.
* test(deploy-logs): debug log at deployStack entry to trace click path
Adds console.log at the first line of deployStack handler so we can
confirm in CI whether the click is reaching it at all and what
selectedFile/isStackBusy resolve to. Combined with the existing
runWithLog debug logs, this isolates whether the modal failure is in
deployStack guarding out, runWithLog early-returning, or something
else entirely.
* test(deploy-logs): drop filter, log every browser console msg
The previous filter only emitted error/warning plus the deploy-feedback
substring. The deploy-feedback debug logs never appeared, so we don't
yet know whether the log itself is firing. Remove the filter so the
full console stream shows up in CI.
* test(deploy-logs): app-level console log to verify capture pipeline
If even an unconditional log at App component render time does not
appear in CI browser logs, then the console capture listener is broken
or the dispatched logs are being filtered upstream of Playwright. This
isolates whether the issue is in the production code or the test
harness.
* test(deploy-logs): use testid locator for stack action button
Replaces the regex-based getByRole locator (/Deploy|Start/i) with
getByTestId('stack-deploy-button'). The regex matched something other
than the actual deploy button: backend logs proved no deploy POST ever
fired, and instrumentation confirmed neither deployStack nor runWithLog
ran on click despite the test claiming success.
Adds data-testid='stack-deploy-button' to both the Restart and Start
button branches in EditorLayout's action bar so the same locator works
whether the stack is running or not.
Also drops the temporary debug console.log entries in deployStack,
runWithLog, and App, and restores the test's console listener filter
to only emit error and warning messages.
* test(deploy-logs): park cursor in corner so auto-close countdown fires
After clicking the deploy button, the cursor lands inside the centered
modal. The modal pauses its 4s auto-close countdown on hover, so the
HAPPY test was waiting for a close that never happened. page.mouse.move
to (0,0) parks the cursor outside the modal before the success banner
appears, letting the countdown complete.
* test(deploy-logs): drop redundant loginAs after page.reload
page.reload preserves auth cookies, so the page lands back on the
dashboard without needing a fresh login. The loginAs call after reload
was racing on isLoginPage(): a transient login-page state during page
load made loginAs commit to filling #username, then the dashboard
committed and #username never came back. Playwright's auto-wait then
hung the fill until the test's 120s timeout, which also dragged later
stacks.spec tests down with collateral timeouts.
waitForStacksLoaded is enough to confirm we're on the dashboard with
the sidebar populated before clicking the new stack.
* test(e2e): make loginAs race-safe when login page is a false positive
isLoginPage() reports the page as a login screen if the Login button
locator reports visible at the moment of the check. Under CI load (more
real container deploys from the deploy-log-panel suite), the auth
context can render the login form for one paint, then redirect to the
dashboard. The original code committed to filling #username and hung
until the test timeout when the field was no longer there.
Now the login branch waits up to 2s for #username to actually appear
before filling. If it never appears, we fall through to the dashboard
check instead of hanging.
* feat(stacks): add per-service start/stop/restart lifecycle routes
Adds POST /:stackName/services/:serviceName/{start,stop,restart} routes
that operate on containers belonging to a single Compose service, using
the same Engine API pattern as the existing stack-level lifecycle routes.
Includes isValidServiceName validator and audit-summary entries for the
three new paths.
* test(stacks): add per-service action route tests
* test(stacks): fix test quality issues in service action tests
* feat(stacks): add per-service lifecycle menu to container cards
* fix(stacks): handle paused container state in service action menu
* docs(stacks): add per-service lifecycle actions documentation
* docs(stacks): add validation screenshots for per-service lifecycle actions
* feat(scheduler): add auto_backup, auto_stop, auto_down, auto_start actions and delete_after_run one-shot mode
Extends the scheduler with four new stack-targeted actions:
- auto_backup: backs up stack compose files and .env using the existing
FileSystemService.backupStackFiles primitive
- auto_stop: runs compose stop (containers preserved)
- auto_down: runs compose down (containers removed)
- auto_start: runs compose up -d via deployStack (universal start for both
stopped and down stacks)
Adds delete_after_run boolean column to scheduled_tasks. When enabled,
the task self-deletes after its first successful execution; failures keep
the task so the user can debug and retry.
All four new actions gate at Admiral tier, consistent with restart/snapshot/prune.
Migration is idempotent (maybeAddCol).
* docs(scheduler): update scheduled-operations doc with new lifecycle actions and delete-after-run
Adds the four new actions (Backup Stack Files, Stop Stack, Take Stack Down,
Start Stack) to the action table. Documents the delete-after-run one-shot mode
with its success-only deletion semantics. Adds the Stack Lifecycle Scheduling
section explaining stop-vs-down semantics and the local-execution boundary.
Adds three troubleshooting entries: auto-start on a missing compose folder,
auto-backup single-slot overwrite by design, and one-shot task disappearing
after successful run.
Updates the timeline description from four to five lanes. Refreshes
screenshots to show the new dialog layout with the Lifecycle lane visible.
- Add label_ids and categories columns to notification_routes via idempotent migration
- Matcher logic always evaluates routes (AND semantics across all non-empty matchers)
- getStackLabelIds skips DB call when no enabled route uses label filtering
- Extract ALL_NOTIFICATION_CATEGORIES array from NotificationService as single source of truth
- Derive VALID_CATEGORIES set from the array in the route handler
- Extract validateLabelIds and validateCategories helpers to remove POST/PUT duplication
- Extract tryAddColumn as a private DatabaseService class method (removes 5 local re-declarations)
- Extract CATEGORY_LABELS to frontend/src/lib/notificationCategories.ts (shared by NotificationPanel and NotificationRoutingSection)
- Frontend form adds label and category multiselects with AND-filter hint
- Route cards show label and category badges; empty-matcher routes show 'Matches all alerts'
- Add tests for category-only, label-only, and combined AND-semantics routing
Adds a nullable node_id column to notification_routes (null = any
node, integer = fire only when the alert originates from that specific
node). This fixes a multi-node fleet defect where a route scoped to
"my-app" would fire on every node that hosts a stack with that name.
Backend changes:
- DatabaseService: idempotent migration adds node_id INTEGER NULL and
a composite index on (node_id, enabled, priority); the two statements
are in separate try-catch blocks so the index is always created even
when the column was added in an earlier run
- NotificationService: route matcher now pre-filters by node_id before
checking stack_patterns (== null matches any node)
- notifications route: POST/PUT accept optional node_id, validated to
be null or the local node's ID; NodeRegistry guards against
cross-node misroutes
Frontend changes:
- NotificationRoutingSection: node scope Select field uses useNodes()
from NodeContext (no extra API call) to populate the local node option
- Route cards show a node badge when node_id is set
Tests: 3 new tests covering node-match, node-mismatch, and null-scope;
all 75 files (1413 tests) passing.
Introduce a NotificationCategory string-literal union (11 values) and
thread it through dispatchAlert as a required second argument. All
callers (DockerEventService, AutoHealService, ImageUpdateService,
MonitorService, PolicyEnforcement, policyGate, SchedulerService,
imageUpdates route) pass an explicit category at every call site,
giving TypeScript compile-time enforcement that no new emit site can
be added without choosing a category.
DatabaseService gains an idempotent migration that adds a nullable
category TEXT column to notification_history; existing rows keep
category=NULL (displayed as Uncategorized in the UI). The
getNotificationHistory method accepts an optional category filter
that is forwarded from the GET /api/notifications/history route via
a ?category= query param.
NotificationPanel gains a category Select dropdown so users can
filter history by category. The frontend types mirror the backend
union so API responses are type-safe end-to-end.
All 75 test files (1410 tests) updated to the new 4-arg dispatchAlert
signature and passing.
Allow a scheduled task with action='update' and target_type='fleet'
to update every eligible stack on a node in a single schedule entry.
The executor respects each stack's per-stack auto-update policy via a
single batch query, skipping stacks that have opted out. For remote
nodes the request proxies to the remote Sencho instance, which already
enforces the same policy in its /api/auto-update/execute endpoint.
Backend route validation now accepts update+fleet as a valid combo
(previously only update+stack was allowed) and requires node_id.
Frontend adds an "Auto-update All Stacks" option to the scheduled-task
creation form with a node selector and descriptive helper text.
Right-clicking a stack or opening its 3-dot menu now shows a
'Schedule task' entry in the lifecycle group (visible to paid tiers).
Clicking it navigates to Scheduled Operations and opens the New
Schedule dialog pre-filled with the stack name and active node,
removing the need to navigate there manually and re-enter the target.
- Added openScheduleTask to StackMenuCtx; wired in buildMenuCtx using
the active node from NodeContext
- Extended ScheduledOperationsView with optional prefill/onPrefillConsumed
props; a ref-guarded effect calls openCreate() with the prefill data
- openCreate refactored to accept an optional prefill arg, removing the
duplication between the effect and the existing 'New Schedule' button
* feat(auto-update): add per-stack auto-update enable/disable toggle
Paid users (Skipper and Admiral) can now opt individual stacks out of
scheduled auto-updates from the stack context menu without disabling
the global feature.
- Add stack_auto_update_settings table (node_id, stack_name) with
default enabled=true; four typed DatabaseService accessors with
parameterized queries.
- Add GET /stacks/auto-update-settings, GET /stacks/:name/auto-update,
and PUT /stacks/:name/auto-update (requirePaid + requireAdmin).
PUT broadcasts state-invalidate with action auto-update-settings-changed
so all open tabs refresh immediately.
- Stack DELETE clears the auto-update setting row alongside stack_update_status.
- autoUpdateRouter /execute skips disabled stacks before any registry
call; skip is recorded in the results array. Manual Update actions
are not affected.
- Add Auto-update: Enabled/Disabled toggle in the stack inspect group
(paid tiers only, hidden for Community, consistent with Auto-Heal).
Toggle uses optimistic update with revert-on-error toast.
- AutoUpdateReadinessView shows an Auto: Off pill and disables the
Apply now button for stacks with auto-updates off. Detection still
runs so the readiness card remains visible.
- Add 21 backend Vitest tests covering DB round-trips, endpoint auth
and tier gates, execute skip for both wildcard and named targets.
Add 3 frontend hook tests for toggle visibility and callback behavior.
* docs(auto-update): document per-stack auto-update control
Add a Per-stack control section to the auto-update readiness page
explaining how to disable and re-enable auto-updates for individual
stacks, what disabling means (scheduled apply skipped; detection still
runs; manual update unaffected), and a troubleshooting entry for
scheduled runs not applying to a specific stack.
Group readiness cards by node so updates pending on every reachable node
are visible without having to switch the active node. Apply now targets
the owning node directly, and Recheck fans out to every reachable node
in parallel; per-node cooldowns are surfaced in the toast.
Adds POST /image-updates/fleet/refresh and invalidates the fleet
aggregation cache after auto-update execute so the next read reflects
the new state immediately. A small banner appears under the hero when
some online nodes did not respond within the request timeout.
The log viewer's per-line timestamp now renders HH:mm:ss again. The
HH:mm:ss.SSS variant was busier than the actual log content needed and
made the column harder to scan at a glance. The underlying ISO
timestamp from docker logs -t is still preserved on each row, so
download / copy still carries the original precision.
The rAF-based flush from the same area stays in place; that is what
makes lines feel real-time, not the timestamp width.
Dashboard and sidebar status indicators previously only refreshed on a
5-30 second polling cadence: a container restart, a degraded -> healthy
transition, or a stack update was invisible until the next tick.
Add a lightweight, non-persisted "state-invalidate" envelope on the
existing /ws/notifications WebSocket:
Backend
- NotificationService.broadcastEvent: sibling of dispatchAlert that
pushes an arbitrary {type, ...} envelope to every subscriber WITHOUT
writing to the alerts history (these are pure ephemeral signals).
- DockerEventService.handleEvent: emit the envelope for state-changing
container actions (start/die/kill/destroy/create/restart/pause/
unpause/health_status/rename/update). Carries node id, stack name
(from the compose project label), container id, action, and
timestamp.
Frontend
- EditorLayout's two notification WebSocket handlers (local plus
per-remote-node) branch on type. On state-invalidate they re-emit a
window CustomEvent and trigger a debounced (250ms) refreshStacks so
a burst of events from compose recreating multiple services
collapses to one refetch. The refresh callback is held in a ref so
the long-lived WS effect never closes over a stale function.
- useDashboardData listens for the same window event and refetches
/stats, /system/stats, and /stacks/statuses on every signal.
Historical metrics stay on their 60s polling cadence (10-minute
trend data, not a live indicator).
Tests
- Three new docker-event-service cases assert broadcastEvent fires on
start and health_status events with the correct envelope shape, and
does not fire on non-state actions like exec_create.
- Existing 28 cases updated with the broadcastEvent mock so the
subscriber stub matches the new shape.
Polling stays as a safety net at the same intervals; the WS path is
the fast path. Multi-node fleets benefit on the local node today;
extending the remote forwarder to relay state-invalidate is a
recommended follow-up.
Previously, fetching the .env file for a stack with no env files at all
returned a 404 with a JSON error body. The frontend's secondary loader
(changeEnvFile) called res.text() without checking res.ok, which caused
the error body to be stuffed directly into the editor as if it were
file content.
Two-part fix:
Backend (routes/stacks.ts):
- For the default GET /stacks/:name/env (no ?file= query) when the
stack has no env files, respond 200 with an empty body and an
X-Env-Exists: false header instead of 404.
- For an explicit ?file= query that resolves to a missing file, keep
the 404 (the caller asked for something specific).
- Catch a TOCTOU ENOENT between access() and readFile() and return the
same friendly empty-body shape, not a generic 500.
Frontend (EditorLayout.tsx::changeEnvFile):
- Check res.ok before reading the body. On a non-OK response, clear the
editor content and surface a friendly toast instead of pasting the
server's JSON error string into the file.
When a registry pushes a new build of an image at the same tag (digest
changes, tag does not), the preview service set next_tag to the same
string as current_tag and the readiness view rendered '10.11 -> 10.11',
which reads as a UI bug.
Add an update_kind field to UpdatePreviewSummary that distinguishes:
- 'tag' - at least one image has a strictly newer tag
- 'digest' - the only updates available are same-tag rebuilds
- 'none' - nothing to apply
The frontend now branches on update_kind and renders 'Rebuild available'
next to the current tag for the digest case, leaving the version-arrow
diff for genuine tag bumps.
Three new buildSummary cases lock in the kind classification.
Two compounding issues made logs feel laggy and timestamps look duplicated:
1. State updates were batched in a 250ms setInterval, so a burst of
lines all rendered together every quarter second. Replace with a
requestAnimationFrame scheduler: lines flush on the next paint
(~16ms at 60Hz) while still collapsing a single burst into one
React commit. Cleanup uses cancelAnimationFrame.
2. The timestamp formatter rendered HH:mm:ss only, dropping the
sub-second precision that docker logs -t already emits. Two lines
logged within the same second appeared identically. Render
HH:mm:ss.SSS so successive lines remain visually distinct.
The toolbar Update button calls updateStack(), which refreshed
containers and stacks but never re-fetched the image-updates list.
The sidebar's blue "update available" dot therefore stayed visible
until the 5-minute polling interval. The right-click context menu
path (executeStackActionByFile, action='update') already calls
fetchImageUpdates() on success; mirror that call here so both paths
behave the same.
The Clipboard API requires a secure context, so navigator.clipboard is
undefined when Sencho is accessed over HTTP on a LAN IP. Most copy
buttons therefore failed silently and a few even fired success toasts
without writing anything to the clipboard.
Extract a shared copyToClipboard helper that prefers the modern API in
secure contexts and falls back to a hidden-textarea execCommand path
otherwise, then route every existing call site through it.
Updates the backend listen port, Vite dev proxy target, Docker EXPOSE,
compose port mapping, .env.example default, GitHub Actions smoke-test
default, healthcheck URLs, and every doc/example reference. Test fixtures
that include example URLs were updated for consistency, though their
assertions are port-agnostic.
The rate-limit value of 3000 in middleware/rateLimiters.ts and the
3000 entry in WEB_UI_PORTS (which detects user containers like Grafana)
are intentionally untouched.
Fresh installs land on the Community tier. The 14-day Admiral trial is now
issued by Lemon Squeezy via their hosted checkout: the user enters email +
card, receives a license key by email, and pastes it into the existing
Settings > License activation field.
Backend changes:
- LicenseService.initialize() no longer auto-creates a license_status='trial'
row on first boot. It now only ensures an instance_id exists and starts
periodic validation.
- Drop the TRIAL_DURATION_DAYS constant.
- Drop the status='trial' early-return in getVariant() so LS-issued trials
resolve through the normal variant metadata path (variant_name / product_name).
- Trial branches in getTier() and getLicenseInfo() are retained for future
work that may detect trial state from Lemon Squeezy metadata; they are
currently unreachable via the Sencho code paths.
Frontend changes:
- Settings > License surfaces a new "Try Admiral free for 14 days" CTA block
with Start monthly trial and Start annual trial buttons that open Lemon
Squeezy hosted checkout. The CTA is visible only when the user has no paid
access and is not already on a trial.
- Reserve the Admiral upgrade card for the Skipper-active upgrade path so
unlicensed users see one Admiral path (the trial CTA) instead of two.
- Pull the inline Lemon Squeezy checkout URLs into named module constants so
the Skipper, Admiral monthly, and Admiral annual endpoints are defined in
one place.
Test changes:
- license-service.test.ts covers the no-auto-trial startup path and updates
the trial-variant test to reflect the metadata-driven resolution.
- afterAll in the initialize() describe block calls destroy() so the 72-hour
validation interval does not leak into sibling test files.
Docs:
- Rewrite the Free trial section in features/licensing.mdx to document the
new LS checkout flow (email + card required, auto-converts on day 14 unless
cancelled).
- Add an operations/troubleshooting entry for cases where the trial license
key email does not arrive.
Custom OIDC stays on Community so self-hosters can wire any spec-compliant
OIDC identity provider (Authelia, Keycloak, Authentik, Zitadel, and others).
Google, GitHub, and Okta one-click presets move to Skipper. LDAP / Active
Directory and scoped RBAC are Admiral-only.
Backend enforces the split via a new requireTierForSsoProvider helper in
middleware/tierGates.ts, applied after requireAdmin in all four ssoConfig
mutation handlers. GET /sso/config (list) stays ungated so downgraded
admins can still see previously-configured providers. Invalid provider ids
now 400 before the tier check to avoid leaking tier information.
Frontend adds a compact mode to PaidGate and AdmiralGate for inline
list-item locks, and SSOSection reorders the provider cards as
Custom OIDC > Google > GitHub > Okta > LDAP to reinforce the
free-to-paid progression.
Stale 'SSO is Admiral' copy in AdmiralGate, PaidGate, and the Admiral
upgrade card on the License settings page has been replaced to reflect the
new split. User-facing licensing, SSO, overview, quickstart, and security
docs have been updated with the per-tier provider matrix.
* feat(sidebar): implement keyboard shortcuts for stack menu actions
Shortcut labels shown in the context menu and kebab menu were purely
decorative. This wires them up to the corresponding actions on the
currently selected stack.
Cmd/Ctrl shortcuts: Enter (deploy), . (stop), R (restart), Up (update),
Backspace (delete). Single-key shortcuts: a (alerts), h (auto-heal,
paid), u (check updates), p (pin/unpin).
Guards: shortcuts are blocked when an input element is focused, when the
global command palette dialog is open, when no stack is selected, or
when the stack is busy. All visibility and busy flags from the menu
context are respected.
* docs(sidebar): document keyboard shortcuts for stack actions
- Remove redundant SENCHO prefix from kicker (already shown in card header)
- Animate the left accent bar with a slow back-and-forth shimmer sweep
- Replace static status dot with animated ping pulse
* feat(security): polish scan sheets, fix CVE links, surface policy violations
Adds cveUrl helper that rewrites Trivy's 404-ing avd.aquasec.com links to
cve.org for CVE-prefixed IDs (GHSA and misconfig URLs pass through unchanged).
Redesigns both scan sheets with shadow-card-bevel chips, tracked-mono kickers,
severity row tinting with a left accent rail, and tabular-nums timestamps.
Surfaces a destructive policy-violation banner on scans whose policy_evaluation
row flags a block, and fixes the compare sheet's delta ribbon so CRITICAL
net-positive deltas render in destructive (not warning) tone. Backend parses
the JSON policy_evaluation column at the API boundary so the UI receives a
structured object.
* chore(security): suppress CVE-2026-32281 and CVE-2026-32283 in Trivy scan
Both CVEs affect Go stdlib crypto/x509 and TLS in Docker CLI 29.4.0
(Go 1.26.1) and Compose v5.1.2 (Go 1.25.8). No upstream static binary
has been released with the patched Go 1.26.2 or 1.25.9 runtimes yet.
Exposure analysis: the Docker CLI and compose plugin connect to the local
Docker socket (Unix socket, no TLS) and to public registries with well-known
CAs. Neither CVE is exploitable in this configuration. Added alongside
sibling entries already in .trivyignore for the same binary versions.
Revisit on next Docker CLI and Compose upstream release.
Scan history is now a right-side sheet that layers over the current view
(typically Resources Hub) instead of a full-page activeView branch. The
sheet opens via the existing navigation event, fetches only when open,
dismisses on Escape or overlay click, and preserves the nested scan-details
and scan-compare sheets intact via Radix portal stacking.
The fetch effect now resets selection and page state on active-node change
exactly once, avoiding a double-fetch on node switches.
Policies with block_on_deploy=1 now scan every stack image before
docker compose up runs and reject the deploy with HTTP 409 on violation.
The UI opens a dialog listing offending images; admins can override per
deploy with ?ignorePolicy=true, and every bypass is recorded in the
audit log with the originating route, actor, policy, and image list.
When Trivy is not installed on the target node the gate fails open with
a warning notification, so teams are never locked out by tooling state.
Post-deploy and scheduled scans still evaluate matching policies and
dispatch warnings on violations to surface drift on long-running stacks.
Public API additions: policy and suppression CRUD under /api/security,
plus the documented 409 block-response shape on all deploy paths.
Two polish improvements to the aggregated notifications inbox:
- Per-node filter dropdown in the bell panel (hidden on single-node
installs) so fleet operators can triage events from a specific box.
Selected node falls back to "All nodes" automatically if that node is
removed from the registry.
- 60-second safety-net poll that reconciles the list so events missed
during a WebSocket reconnect backoff appear without a manual refresh.
Uses a ref indirection to pin the interval to the latest
fetchNotifications closure.
Flip Security (Trivy), Notifications (agents + history), and App Store from
global-and-hidden-on-remote to node-scoped so operators can manage them when a
remote node is selected in the node picker. The primary instance proxies the
calls to each remote, which resolves the correct per-instance binary state,
agent config, and template registry.
Backend: key `agents` and `notification_history` by `node_id` with idempotent
column-add migrations and a `(node_id, type)` unique index on agents, matching
the Labels pattern. Thread `req.nodeId` through the /api/agents and
/api/notifications routes. Internal NotificationService and ImageUpdateService
writes resolve the middleware default via `NodeRegistry.getDefaultNodeId()` so
monitor-emitted rows share a bucket with user-facing ones (avoids split-brain
where the UI sees test notifications but not internal alerts).
Frontend: split Security on remote to render only the scanner card and hide
scan policies and CVE suppressions (those remain control-plane-only). Drop the
misleading "Always Local" badge on Developer since retention windows govern
backend jobs, not UI state. Flip the App Store registry to node-scoped.
Docs: add a "What Settings apply per node" table to multi-node, clarify
remote alert setup in alerts-notifications, and note Trivy's per-host install
in vulnerability-scanning.
Replace the static SVG fleet topology with a ReactFlow canvas laid out
via dagre. Each node renders as a rack card with status pill, type
badge, CPU/MEM/DISK bars, and stack/running counts. Pan, zoom, drag,
and minimap are enabled; user-dragged positions persist across the
30-second poll so live metric updates no longer reset layout.
* feat(fleet): reorganize overview page for clarity and density
Scope Grid/Topology to the Overview tab (moved from above the tab bar so it
no longer implies it applies to Snapshots). Move Check Updates and Refresh
onto the tab bar row, right-aligned.
Compact the overview toolbar: constrain the sort combobox to a fixed width so
it no longer stretches full-page, and collapse the Status, Type, Severity,
and Tags pill groups into a single Filters popover with an active-count
badge and an inline Clear all filters action.
Render the local node card in the same responsive grid as remote nodes
instead of a dedicated full-width row. Visual distinction is preserved
through the existing brand gradient, cyan rail, ring, and "Local" badge.
* fix(fleet): stop local card from stretching when remote expands
Merging the local and remote node cards into a single responsive grid
meant CSS grid's default align-items: stretch made every cell in a row
match the tallest one. Expanding stack details on a remote card pulled
the local card up with it.
Add items-start on the merged grid so each cell sizes to its own
content.
* feat(search): add global command palette with cross-node stacks
Press Ctrl+K from anywhere to open a search palette that jumps to
pages, switches nodes, or opens stacks on any online node in the
fleet. Trigger lives as an icon in the top bar. Replaces the former
sidebar-scoped Ctrl+K handler.
The cross-node stack search fan-out is extracted into a shared hook
so the sidebar and palette stay in sync on the same debounce +
abort shape.
* fix(search): drop render-time ref write and effect-based query reset
Replace `nodesRef.current = nodes` during render with direct use of
`nodes` in the stack select callback, and fold the query-reset logic
into a single `onOpenChange` handler so it no longer runs via an
effect. Both changes resolve react-hooks rule violations that broke
frontend lint in CI.
Labels settings are now visible on remote nodes (scope: node, no longer
hidden on remote context). The LabelsSection already routes through the
active-node proxy, so edits land on whichever node the operator is viewing.
Fleet overview's label filter previously fetched only the control-plane
node's labels and assignments, so remote stacks could never match the
filter. Rewrote aggregation to fan out /labels and /labels/assignments
to every online node via fetchForNode + Promise.allSettled with a 5s
timeout per request. The palette dedupes by (name, color) so identical
labels on multiple nodes collapse into one entry while same-name +
different-color stay distinct. The assignment map is nested by nodeId
to avoid cross-node stack-name collisions.
Keyed the label refetch effect on a stable online-node id signature
rather than the nodes array reference, so the existing 30s overview
poll (and 5s fast-poll during updates) does not cascade into repeated
fleet-wide label fetches.
Apply design audit section 13's proposed chroma/lightness values to
frontend/src/index.css. Dark-theme accents now match the audit swatches
exactly: brand 0.78 0.11 195, success 0.78 0.16 155, warning 0.82 0.14
75, destructive 0.72 0.18 20. Surface hierarchy shifts to a warm pebble
cast (hue 80, chroma 0.003 to 0.01); card anchors to the audit's
0.14 0.005 80 surface swatch; popover bumps to 0.16 to preserve the 4
percent lightness gap against card. Light theme mirrors symmetrically.
NetworkTopologyView keeps the react-flow inline-style escape hatch but
hoists the brand literal into a BRAND_COLOR constant so MiniMap and
EDGE_COLORS share one source of truth.
Also normalize em-dash comment separators to colons across index.css
and drop a duplicated dark-background literal in favor of
var(--background).
The pulsing dot signaling an available image update used --info (blue
hue 250), the same hue as --label-blue, so the two blended visually
when a stack carried a blue label. Introduce a dedicated --update
token in vibrant fuchsia (hue 320) sitting outside the label palette,
and render both the sidebar StackRow and the fleet NodeManager
indicator as a notification-dot pattern (solid stationary core with
animate-ping halo). Aligns with the design rule that static state
should not pulse.
Document the hidden constraints that PR #706 left in the codebase so they
are not accidentally removed in a future cleanup pass:
- DropdownMenuSubContent and ContextMenuSubContent must stay Portal-wrapped
so sub-menus escape ancestors with overflow-x-hidden.
- refreshLabels must stay stable via useCallback because it is captured by
buildMenuCtx's memoization and passed as a prop.
Also delete LabelAssignPopover.tsx, which had zero consumers after the
create-and-assign flow moved into the menu-layer inline form.
Portal-wrap DropdownMenuSubContent so the kebab Labels submenu renders outside
the clipped dropdown container, matching ContextMenuSubContent and fixing the
empty/broken kebab submenu.
Thread onLabelsChanged from LabelsSection through SettingsModal to
EditorLayout so label creates and deletes in Settings propagate to the
sidebar menus without a page refresh.
Add an inline "New label" form in both the kebab and context menu label
submenus that creates and assigns a label in one interaction, removing the
Settings round-trip from the label assignment flow.
The dot on the CPU sparkline added visual noise without adding
information beyond what the textual "peak X% @ HH:MM" label
already conveys. Disable showPeak on the CPU Sparkline; the peak
text label is preserved.
* chore: ignore .superpowers/ brainstorm scratch dir
* feat(sidebar): add useStackMenuItems hook with grouped menu model
Pure transform hook that converts StackMenuCtx into four ordered MenuGroup
arrays (inspect, organize, lifecycle, destructive). Shared type contract in
sidebar-types.ts gives both the ContextMenu and DropdownMenu a single source
of truth so they cannot drift. Covered by 8 unit tests.
* refactor(sidebar): stabilize useStackMenuItems memoization deps
Destructure menuVisibility flags into primitive deps so inline object
literals from callers do not defeat memoization. Add a test confirming
isBusy disables all lifecycle items.
* feat(sidebar): add usePinnedStacks hook with per-node localStorage
* refactor(sidebar): stabilize usePinnedStacks eviction signal and isPinned dep
Change evictedOldest shape to { file, seq } so consumer effects re-fire on
repeated evictions. Narrow isPinned's useCallback dep to the current node's
pinned list so it only rebinds on local changes. Add test for eviction
side-effect and a second test for the seq counter.
* feat(sidebar): add useSidebarGroupCollapse hook with per-node keys
* refactor(sidebar): tighten useSidebarGroupCollapse effect ordering
Collapse the two write/read effects into a single skip-next-write ref
pattern so switching nodes no longer writes the previous node's map under
the new key before hydration. Also skip the no-op mount write. Add test
for setCollapsed.
* feat(sidebar): add row + group-header style helpers
* feat(sidebar): add SidebarBrand with mono kicker + serif hero
* feat(sidebar): add SidebarActions wrapper for create + scan
* feat(sidebar): extract SidebarSearch with kbd pill
* feat(sidebar): add StackRow with cyan-rail active state
* refactor(sidebar): dedupe tooltip markup in StackRow, widen test coverage
Extract a local RowTooltip helper so the update and git-pending branches
share the CursorProvider scaffolding. Add four behavioral tests covering
click, keyboard activation, kebab stop-propagation, and the busy loader
branch.
* feat(sidebar): unify context + kebab menus via useStackMenuItems
* feat(sidebar): add StackGroup with collapse and pinned variant
* feat(sidebar): add StackList with pinned + label groups
* feat(sidebar): add SidebarActivityTicker with idle fallback
* feat(sidebar): add StackSidebar container composing the regions
* feat(sidebar): replace sidebar block with StackSidebar composition
* docs(sidebar): add stack sidebar feature page with screenshots
* fix(sidebar): satisfy react-hooks purity and memoization rules
* fix(sidebar): restore "Sencho Logo" alt text for E2E selector
* feat(ui): redesign host console as cockpit surface
Rework the Console view into the cockpit language introduced in #699.
The page is now a PageMasthead strip plus a terminal well plus a
floating chip strip.
The masthead shows connection state (`Connected`, `Reconnecting`,
`Disconnected`) with a cyan rail, a pulsing status dot while the
session is live, an italic state word, and a tracked-mono kicker of
`HOST CONSOLE · {node}`. The right side surfaces three metadata tiles:
shell, current viewport dimensions, and session uptime. When the
console was opened from a stack, a small back link sits beside the
state word so operators can return to the stack editor without
leaving the cockpit.
The vestigial Close Console button in the header is gone. The
floating chip strip replaces it with four controls that are more
useful in context: Copy (current selection), Clear (scrollback),
Download (full scrollback via SerializeAddon), and Reconnect (close
and reopen the WebSocket in place).
* docs(host-console): refresh screenshots for cockpit redesign
Adds masthead and chip-strip close-up screenshots to the Host Console doc
and refreshes the overview screenshot to match the new cockpit surface.
Rework the Logs view into the cockpit language: a PageMasthead strip with
cyan rail, pulsing live dot, italic state word, and tracked-mono kicker;
a SignalRail of EVENTS/MIN (with rolling 60s sparkline), ERRORS, WARNINGS,
and CONTAINERS tiles; a segmented filter strip; a day-banded feed with
severity dots and whole-row tint on ERROR/WARN; and a floating glass chip
strip for pause/resume, clear, and download.
Decouple the live log stream from the Developer Mode toggle so SSE runs by
default for everyone, with polling as an invisible fallback when
EventSource is unavailable. Drop the Standard Log Polling Rate setting
and the global_logs_refresh key it persisted; Developer Mode still gates
Real-Time Metrics and Debug Diagnostics and nothing else.
Introduces the shared PageMasthead and SignalRail primitives for reuse
across other cockpit pages.
Replace the sliding accent pill and blurred underline with material-at-
rest tabs that speak the cyan identity language used elsewhere in the
chrome. Each tab is a full-height button with a tracked-mono uppercase
label, and the active tab carries a crisp 2px cyan rail flush with the
bar's bottom edge.
Drops the <Highlight> primitive, the springs import, and the unused
navTabValue memo/prop from EditorLayout.
Extract the app header into a dedicated TopBar component and apply the
same translucent chrome language used by the sidebar: bg-sidebar +
backdrop-blur-md + inset top highlight. A new --chrome-top-highlight
token drives the highlight in both themes via the shadow-chrome-top
utility alias.
Restore the pulsing-dot unread indicator on the notification bell
(replaces the temporary count pill) and keep its aria-label dynamic so
screen readers still hear the unread count. Tighten nav a11y with
aria-current on the active item, aria-label on icon-only buttons, and a
proper <nav> landmark around the desktop Highlight group.
Swap hardcoded Tailwind padding on NodeSwitcher, UserProfileDropdown,
and NotificationPanel for var(--density-row-x) / --density-row-y /
--density-cell-y / --density-tile-y custom properties so the three
sidebar/header overlays now tighten when the Appearance setting is
switched to Compact, matching the reference pattern in ui/table.tsx.
Replace the inline Select in the sidebar with a dedicated NodeSwitcher
component that always renders as an identity card, regardless of node
count. With two or more nodes it opens a Popover listing every node
with status dot, type, version, last-seen metadata, and an active-row
accent rail, matching the design language of the user menu and
notification panel.
Extract the relative-time formatters out of NodeManager into a shared
@/lib/relativeTime module with formatTimeUntil and formatTimeAgo, and
add a 'just now' / '<1m' branch so fresh heartbeats and imminent runs
read naturally.
* feat(ui): redesign toast with accent rail and action slot
Align toast with the overlay-surface language from the user menu and
notification panel: left accent rail, tracked-mono kicker above the
message, lucide icons at 1.5px stroke, status-tinted progress bar.
Extend the toast store with an optional options object so callers can
pass an inline action button and override the default duration. Guard
removeToast against no-op dismissals so subscribers are not notified
when nothing changed, and cap the buffered queue at 20 to prevent
unbounded growth under spam.
* fix(ui): avoid impure Date.now in toast ref initializer
useRef(Date.now()) re-evaluates the impure call on every render and
trips the react-hooks/purity lint rule, even though only the first
render's value is retained. Initialize to 0 since the auto-dismiss
effect writes startRef.current = Date.now() before any read path
consumes it.
Add stack_name and container_name columns to notification_history so bell
rows can act as jump points. Producers (AutoHeal, Docker events) pass the
container context through dispatchAlert; the panel renders routable rows
as buttons that load the target stack and, when a container name is
present, open its logs modal. Non-structural notifications stay as
passive display rows.
* feat(ui): redesign user menu and notification panel per design audit
Align both floating surfaces with the stack-view design language: cyan
signal rails, identity headers, strip containers, day-banded streams,
severity rails, Instrument Serif for hero words, and tracked-mono for
metadata. Extract the notifications popover out of EditorLayout into its
own component and introduce a reusable SegmentedControl primitive for
3-way radiogroups (Auto/Light/Dark theme, All/Unread/Alerts filter).
* fix(ui): pin accessible names on avatar and bell buttons
After the redesign the avatar button's text content became the user's
initials and the bell button's text became the unread count, which
displaced the title-based accessible names and broke
getByRole('button', { name: /profile/i }). Set aria-label explicitly on
both triggers and mark the bell badge aria-hidden so screen readers and
role-based selectors see stable names regardless of dynamic content.
Anso