Shared session-list subscriptions opened server-only event and URL-token routes against the webview origin, causing repeated 403 responses. Skip the control-stream subscription in VS Code while preserving OpenCode bridge sync and polling.
Validated with regression tests, UI type-check and lint, and the VS Code webview build. The reported gray-screen crash remains unconfirmed pending user logs.
The five sort modes sat as a chip row under the search field, which cost a
project row of list space for a setting people touch about once a month.
Put them behind an icon in the header next to reordering — the two answer
the same question about the list — and let it open the same sheet the
composer already uses to pick a model or an agent, with a check on the
active mode. The icon hides while reordering or searching, where sorting
means nothing.
The new-session button moves to the outer right edge and stays there: it
is the one action people hit without looking, so it must not slide as the
icons beside it come and go.
Testing: package type-check and lint; drove the drawer in a browser at
390px (chip row gone, sheet lists the five modes and marks the active one,
picking one reorders the list and closes the sheet, the icon disappears in
reorder mode, and the new-session button keeps the right edge with and
without a search).
Android 15+ enforces edge-to-edge, and Capacitor stops padding the WebView
once it can hand the insets to the page instead — so the app draws under
the gesture/navigation bar. Only the top inset was ever consumed: full
cover surfaces ignored the bottom one entirely, which is why a settings
list ran under the bar, and the chat shell reserved half of it, which is
enough for a home indicator but not for a bar that swallows taps.
Reserve the full inset in the shell on Android and give the surfaces that
own the bottom edge — fullscreen surfaces, the workspace drawer, bottom
sheets — an opt-in class that pads them. The sessions drawer already did
this through its footer, so it stays as it is. Capacitor reports a zero
bottom inset while the keyboard is up, so none of this stacks with the
keyboard inset. iOS deliberately keeps its softer visual inset.
Testing: package type-check and lint; drove the mobile surface in a
browser with the Android platform class and a simulated 48px inset (shell
and surfaces reserve exactly that, drawer footer unchanged, and removing
the class returns every surface to zero).
Snippets, Agents, Commands, Plugins and Skills were desktop-only, even
though the composer already reads snippets and slash commands and the chat
lets you pick an agent: you could use those features from a phone but not
manage them. They render through the same page and sidebar components, so
the mobile settings surface only had to stop filtering them out.
Making them usable needed the drill-down to be real. A split page opens
its own list on mobile, but back from an item jumped all the way out to
the settings root, and a separate header button was the only way back to
the list — except on Skills, which had a hand-rolled exception. Turn that
exception into the rule: back walks nav → page list → item in reverse on
every split page, browser and system history follow the same path, and the
extra header button goes away because back now does its job. The Android
hardware button asks Settings to step up before the shell closes it.
MCP, Providers, Usage and Magic Prompts inherit the same navigation.
Testing: package type-check, lint, settings suites under the isolated
runner; walked all six pages plus MCP in a browser at 390px (list → item →
back to list → back to root, no horizontal overflow). The hardware back
button needs a device check.
The final answer divider was too prominent and sat too far below Activity. Lower its opacity to 30% and remove the header's bottom margin in both states, matching the maintainer's visual adjustments.
Validated UI type-check, lint for both changed files, and seven existing activity and rendering tests.
The sidebar's project order (manual, A-Z, Z-A, newest, recently used) is a
shared setting, but the mobile drawer ignored it and always rendered the
manual order, so a phone could not put the project you actually work in at
the top of the list.
Read the same setting in the drawer and offer it as a chip row under the
search field, which scrolls away with it — mobile navigation has no
overflow menu to hide it in, and a control touched once a month should not
hold permanent room above the list. Picking an order persists it the way
the desktop sidebar does, so the choice follows the user across surfaces.
Drag-to-reorder is offered only under the manual order, since dragging
rewrites exactly the order the other modes ignore.
The ordering itself moves into one helper the desktop sidebar now uses
too, so both surfaces answer the setting identically.
Testing: package type-check and lint; new unit tests for the helper plus
the sidebar list suite under the isolated runner; drove the mobile surface
in a browser against an isolated server (chips fit a 390px viewport,
switching to "recent" reorders the list, the reorder button hides, and the
choice reaches the server settings).
Project rows carried an active-project dot and a session count, neither of
which answers a question the user has while browsing the drawer, and there
was no way to start a session in a project's root directory: the only
per-row action was creating a worktree.
Drop both indicators and put a "+" beside the worktree button instead. It
opens a draft already carrying the project and its directory, the same
contract as the desktop sidebar's per-project "+", so the current
directory is not switched out from under the session behind the drawer.
Project rows in search results lose the count the same way and gain the
same button; dropping it there also drops a per-project session scan that
ran on every keystroke.
Testing: package type-check and lint; drove the mobile surface in a
browser against an isolated server (rows render the button, counts gone).
Code comments required Cmd/Ctrl+Enter while quotes and browser annotations used Enter. Align desktop submission and its placeholder, preserve Shift+Enter line breaks and mobile keyboard behavior.
Validated UI type-check, UI lint, and six existing IME checks. Lint reports one existing warning in MobileChangesSurface. In-app browser validation was unavailable.
Use received frames for peer liveness and reject new requests after terminal relay failures. Keep authoritative usage snapshots separate from refresh errors, bound request lifetime, and guard concurrent refreshes and runtime switches.
Add regression coverage for relay liveness, terminal rejections, quota failures, cancellation and request deadlines. Validated with 51 focused tests, workspace type-check and lint, web build and mobile assets.
The chat opens the sessions drawer with a left-edge swipe and the
workspace drawer with a right-edge one, but neither could be closed the
same way: the only way back was the header X or the Android back button.
Reuse the edge-swipe hook on both open drawers so the mirrored gesture
closes them. To keep that free of the row swipe actions in the sessions
list, those are mirrored too: session, project and worktree rows now
reveal their actions on the LEFT with a swipe to the right, in the
mirrored button order so a partial drag still exposes delete first. When
a row has its actions out, the closing swipe puts them away first
instead of closing the drawer.
The hook grew an `enabled` flag so the listeners re-attach when a drawer
opens or a tablet rotates the workspace between panel and drawer.
Testing: package type-check and lint; gestures need a device check.
Negotiate downstream delivery credit and schedule response streams fairly before encryption. Bound queued output, preserve deltas and WebSocket close ordering, and retain legacy peer compatibility.
Validated with 81 relay tests, workspace type-check and lint, web build and mobile assets, and slow-link tests through the production relay. Confirmed on LTE by the maintainer.
* fix(i18n): align Turkish glossary for agent and prompt in telemetry strings
* fix(i18n): use literal prompt in input history scope description
---------
Co-authored-by: fitzgpt <fitzgpt@users.noreply.github.com>
Lower the compact header breakpoint from 640px to 560px while preserving file statistics on mobile. Validated with focused tests, UI type-check and lint, and the web build.
Upgrade LegendList to 3.3.10 and patch temporary padding cleanup to compare CSSOM-serialized values. Settle Activity height when React replays an interrupted layout effect.
Validated with 16 padding regressions, 3 collapse lifecycle tests, the UI suite, workspace type-check and lint, and web and VS Code builds.
The half-viewport band made return-to-end detection too broad. Restore the fixed 40px threshold without changing gesture handling or follow transitions.
Testing: 14 focused scroll tests pass; UI type-check passes; UI lint has one unrelated existing warning; dead-code report inspected. Live chat behavior not validated.
Keep live history compact without changing sorted rendering. Reuse Activity Default, preserve final answers, and summarize tool-result diffs under an animated disclosure.
Validated with the UI test suite, focused disclosure tests, UI type-check and lint, web production build, dead-code scan, and browser and animation fixtures.
Keep top padding on task tool entries and simplify the Turn stats release note to state its default without an instruction.
Testing: focused ToolPart ESLint, changelog validation, and diff checks passed. No interactive layout verification was run.
Turn stats was hidden unless users opted in. Show it by default, migrate implicit hidden lists, and preserve explicit section choices through reloads and settings sync.
Prepare App and VS Code release notes with Turn stats as the headline and BTW composer changes under improvements.
Testing: 115 focused tests passed; UI type-check passed; UI lint has one existing warning. Changelog validation and diff checks passed. Oxlint findings are limited to existing code in sections.ts; interactive app validation was not run.
Shows a separator when the last live assistant answer follows earlier visible content
Keeps render comparisons in sync with the new assistant-text context
Tweaks tool summary spacing and path text styling
Reserves scrollbar space in overlay overlays when the thumb appears
Keeps text and layout from jumping on hover
Avoids double-reserving space for nested scrollers
Updates the setting hint to explain scrollbars stay visible when the pointer leaves the scrollable area.
Refreshes the wording across all supported locales.
* fix(ui): reveal overlay scrollbar on hover or active scroll (re-port to bindScrollbar)
Re-ports #2825's container-hover reveal onto the rewritten bindScrollbar
architecture that landed on main after the original PR branch was created.
- pointerenter/pointerleave on the scroll container reveal the thumb
immediately (deliberate intent on every mouse-pointer runtime; inert
on touch via an event.pointerType === 'mouse' guard, since the Pointer
Events spec fires pointerenter on touch taps as well)
- hide-timer re-checks pointerOverContainer at fire time, so the thumb does
not vanish when the pointer moves from container onto the sibling thumb
- suppressVisibility (chat auto-follow) still suppresses hover reveal
- on hover, schedule a re-measure so the horizontal thumb reflects the
current geometry: if the container has horizontal overflow the thumb is
revealed, otherwise it stays hidden. This matches the original PR's
updateMetrics() approach and avoids the regression where onScroll (which
does not measure) would leave a legitimately overflowing horizontal
thumb hidden while hovering
- index.css: drop the Settings-specific overlay-scrollbar display:none
(thumb is never permanently visible anywhere now)
- regression tests: horizontal thumb reveals on hover when overflow exists,
stays hidden when it does not; touch pointerenter is inert; hand-off
race (container pointerleave followed by thumb pointerover) keeps the
thumb visible; pointerleave hides the thumb
* ci(ui): add overlay scrollbar interaction recording workflow
Records the hover-reveal/hide interaction of the overlay scrollbar
(PR #3219 re-port to bindScrollbar) as a webm + GIF via Playwright
recordVideo, converted with ffmpeg palettegen/paletteuse.
- scripts/record-overlay-scrollbar.mjs: drives hover in/out, forces
overflow on the first .overlay-scrollbar-target so the demo works on
a clean data dir, exports overlay-scrollbar-hover.{webm,gif}.
- .github/workflows/interaction-recording.yml: mirrors the validated
screenshots.yml pattern (auth disabled server, Playwright chromium,
15-min timeout), adds ffmpeg install step.
* docs(ui): add overlay scrollbar interaction recording
Animated GIF captured by the interaction-recording workflow (Playwright
recordVideo + ffmpeg) showing the hover-reveal/hide behavior of the
re-ported overlay scrollbar (bindScrollbar): thumb fades in on pointer
enter, hides again after the hide delay once the pointer leaves.
* feat(ui): adopt ScrollableOverlay in remaining native-scroll panels
Extends the overlay scrollbar (hover-reveal, hide-on-leave) to panels
that still used native overflow-y-auto scrolling:
- Sidebar (left nav): outer flex-1 scroll region
- ContextSidebarTab: full-height tab content
- SessionSwitcherDropdown: session list dropdown (preserves contentRef
for scrollIntoView / switcher item queries)
- HelpDialog: help content region
All four merge sizing into outerClassName (flex-1 min-h-0 / h-full /
max-h-[60vh]) and keep visual classes on className, with disableHorizontal
where the original hid horizontal overflow. Type-check passes; unit test
failures in OverlayScrollbar.test.tsx and event-pipeline.test.js are
pre-existing (reproduce on pristine HEAD).
* test(ui): flush hide timer before asserting thumb hidden
The hide path always schedules a setTimeout (hideDelayMs: 0 still
schedules a 0ms timer). happy-dom runs real timers, so the test must
let the macrotask fire before asserting dataset.visible — flushing rAF
frames alone is not enough. Fixes the one failing test in
OverlayScrollbar.test.tsx (12/13 -> 13/13).
* fix(ci): make scrollbar interaction recording hover retry across targets
The record script picked the first .overlay-scrollbar-target and moved the
pointer to its center; layout/hydration order varies between CI runs, so
the hover sometimes landed on a target whose thumb cannot reveal (empty
container), failing the run. Now it iterates targets in DOM order until
the vertical thumb actually appears (or fails after exhausting all).
* fix(ci): record workflow + i18n + overlay-chrome hide
- .github/workflows/interaction-recording.yml: drop 'ref: rework' so the
workflow checks out the PR head SHA on upstream (where 'rework' branch
does not exist). Replace with persist-credentials: true.
- packages/ui/src/lib/i18n/messages/tr.ts: add 4 missing gitView.empty
keys (parity fix, only tr.ts was behind en.ts). Translations are
approximate; the parity test only checks key existence.
- scripts/record-overlay-scrollbar.mjs: on a fresh data dir the web
build can render onboarding modals (ChooserScreen, AboutDialog,
ConfigUpdateOverlay) that float above the MainLayout with a blurred
backdrop. The thumb's isThumbVisible() returns true (DOM-mounted)
but the captured frame is dominated by the modal, so the user sees
'dialog in front, blurred background' instead of the scrollbar
reveal. hideOverlayChrome() injects CSS to hide every plausible
overlay root and best-effort closes known UI store dialogs.
* fix(ci): drop fork-specific ref in record workflow + add tr locale gitView.empty keys
- .github/workflows/interaction-recording.yml: drop 'ref: rework' so the
workflow checks out the PR head SHA on upstream (where 'rework' branch
does not exist). Replace with persist-credentials: true.
- packages/ui/src/lib/i18n/messages/tr.ts: add 4 missing gitView.empty
keys (parity fix, only tr.ts was behind en.ts). Translations are
approximate; the parity test only checks key existence.
- scripts/record-overlay-scrollbar.mjs: hide onboarding chrome (modals,
backdrops, dialogs) that float above the MainLayout when recording
against a fresh data dir, so the captured GIF shows the actual
scrollbar reveal instead of a blurred-overlay dialog screen.
- docs/interaction-recordings/overlay-scrollbar-hover.gif: regenerate
(489 KB) with overlay chrome hidden (same artifact as CI run
33507731092 which passed).
* ci: noop push to retrigger 'pr checks' on a fresh runner
The 'pr checks' check on this PR's prior head (0509212c3) failed with
'releaseJob is not a function' in packages/web/server/lib/walkthrough/
routes.test.js. This test lives in upstream main and is not touched by
this PR's diff. The same flake is currently hitting PR #3265 and
feat/scheduled-preflight-gate.
Confirmed the two latest upstream main commits (bec7a82568 sidebar
sort, 40e4b6f857 request-security) do not touch walkthrough/, so the
failure is a 20ms timing flake in the test's executor Promise, not a
code regression. This empty commit triggers a new CI run on a
different runner with a different scheduling window.
* feat(settings): adopt ScrollableOverlay in settings shell and dialogs
Several settings surfaces still rendered their scroll containers with
native browser scrollbars (overflow-y-auto / overflow-y-scroll), which
read inconsistently against the overlay scrollbar used everywhere else
in the app once content exceeded the viewport.
Wrap the relevant containers in <ScrollableOverlay>:
- SettingsView: nav sidebar (mobile), mobile fallback, mobile page
sidebar, mobile page content, and desktop split view
- DirectoryExplorerDialog: results list
- GitHubIntegrationDialog: issues / PRs list
- GitHubIssuePickerDialog, GitHubPrPickerDialog: lists
- NewWorktreeDialog: form body
The settings sub-pages (OpenChamberPage, VoiceSettings, PasskeySettings,
etc.) do not carry their own overflow — they inherit the scroll host
from SettingsView, so the shell change is sufficient for them.
type-check, lint, and ui tests (368/369, the one failure is a pre-existing
event-pipeline flake unrelated to this change) all pass.
* ci(record): re-run overlay scrollbar recording on a fresh runner window
* ci(record): give overlay thumb 1500ms to reveal on hover
The 'hover did not reveal the thumb on any target' check has been
flaky across runs since the workflow landed in this branch (about
half the runs fail with the same error). 700ms was tight on cold
GitHub-hosted runners; 1500ms absorbs the cold-start variance
without changing what the GIF captures (the thumb's hide animation
runs after pointerleave, unaffected by the longer pre-leave wait).
* ci(record): debug thumb visibility timing on hover
* ci(record): drop debug logging, keep 1500ms hover wait
Debug logging was used to identify that 200ms is enough on a
healthy runner, but 700ms was not. The flake was the OpenChamber
server being slow to initialize the OpenCode side on cold
runners — the thumb itself renders quickly once the app is up.
Keeping 1500ms absorbs that cold-start variance without slowing
successful runs by more than the GIF's existing post-hover
animation wait (1000ms hideDelayMs).
---------
Co-authored-by: sergiofspedro <sergiofspedro@users.noreply.github.com>
Co-authored-by: openchamber-ops <ops@openchamber.dev>
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
Adds a device-local setting to keep overlay scrollbars visible.
Surfaces the setting in visual settings and settings search.
Updates scrollbar behavior and tests for the new preference.
Retain the capped reasoning implementation and adopt the scroll-position release behavior from #3394. Cover pause, resume, and observer cleanup on the mounted component.
Co-authored-by: alvins82 <alvins82@gmail.com>
Complete #3227 by keeping browser completion polls on the native updater, checking the requested version, and preserving retry access after a failed restart.
simple-git 3.35/3.36 moved its unsafe-config blocklist into
@simple-git/argv-parser and expanded it to include credential.helper.
setLocalIdentity already opted in for the SSH branch (core.sshCommand)
via createGit({ allowUnsafeSshCommand: true }), but the token branch
(addConfig("credential.helper", "store")) was left without the matching
opt-in, so switching to a token-auth identity throws:
Configuring credential.helper is not permitted without enabling
allowUnsafeCredentialHelper
Forward a new allowUnsafeCredentialHelper option through createGit and
enable it in setLocalIdentity alongside the existing SSH opt-in. Cover
the token branch (and the token -> SSH cleanup) with tests mirroring
the existing SSH case.
desktopEntryMatchesApp normalized haystack values without dropping empty
ones, so a .desktop entry with no ASCII letters or digits in Name, id,
file name, or Exec (e.g. Name=抖音) normalized to "" and
needle.includes("") matched every requested app — hijacking the
installed-apps list and Open In launch specs. Empty normalized haystack
values are now filtered, matching the existing needles handling.
discovered-apps.json gains a version field (INSTALLED_APPS_CACHE_VERSION
= 2); caches written before the fix are treated as stale and refresh
through the existing TTL-expiry path instead of serving the poisoned
list for the rest of the 24h TTL.
Node's happy-eyeballs default (autoSelectFamilyAttemptTimeout=250ms)
aborts every connect attempt to provider endpoints whose TCP handshake
exceeds 250ms, so quota fetches from Node processes fail with
"fetch failed" while the Bun-based opencode CLI path succeeds (#3399).
Raise the per-attempt cap to 5s via a shared applyConnectAttemptTimeout()
helper called from the two Node entrypoints that host the quota module:
the Electron main process and the openchamber CLI. Family autoselection
stays enabled (::1 and IPv6→IPv4 fallback preserved); runtimes without
the setter are a no-op.
The OpenRouter provider called GET /api/v1/credits, which OpenRouter
documents as requiring a management key. Called with a normal inference
key it returns HTTP 200 and {"total_credits":0,"total_usage":0} instead
of an error, so the provider rendered "$0.00 left - $0.00 spent" for a
funded key and the !response.ok guard could never catch it.
Read GET /api/v1/key instead, which is documented as callable with any
valid API key. A key with a spending limit reports its own usage against
that limit in the window named by limit_reset, and a key without one
reports usage_monthly. Window usage is limit - limit_remaining rather
than usage, because usage is all-time and limit_remaining tracks the
current reset window. limit_remaining is also server-computed and
already honors include_byok_in_limit.
Bring the provider up to the deepseek.js standard while here: a 15s
timeout, 401 and 403 mapped to a session-expired message, parse failures
mapped to an invalid-response message, an explicit no-quota-data result,
and the aliases export that quota/DOCUMENTATION.md requires. Add the
missing openrouter.test.js and keep packages/vscode in sync.
Refs #3060
Share the Ollama request and parser between credential validation and quota refresh so unparseable pages cannot produce successful empty usage. Reject redirects and bound requests with a timeout while preserving both plan formats.
Tested with 94 quota tests, VS Code type-check and ESLint, and the extension build. Reviewed dead-code output; existing anti-slop findings remain outside the changed code.
Ollama Cloud's /settings page switched shapes for cost-based plans: parse
the 'Monthly usage $X of $Y used' row as a monthly window with a symmetric
'$X / $Y' money label that reads correctly in both used/remaining display
modes, and surface the Extra usage credits balance as a balance-only
credits_balance window matching the Codex/DeepSeek credits treatment
('Credits Balance' in the UI), omitted when the balance is $0. The VS Code
credential validation accepts the new page shape and the Settings cookie
placeholder shows the two-cookie format. Web and VS Code parsers kept in
sync per quota DOCUMENTATION.md.
A streaming Thinking block grew without a height cap and only moved into
the capped, scrollable box once it finished, so a long thought pushed the
whole timeline down while it streamed. The cap was dropped in June because
the box then pinned to its own end on every tick and captured the wheel,
so the chat could not be scrolled while thinking streamed.
The box is capped in every state now and follows its own end only while
streaming and only until the reader wheels or drags upward inside it;
returning to its end re-arms the follow. It is marked as a nested scroller,
so an upward wheel scrolls the box first and reaches the chat once the box
sits at its top. Growth is observed on the content box because markdown
commits asynchronously.
Testing: reasoning block tests extended for the capped, nested box while
streaming; ui type-check and lint; web build. Live reasoning stream not
exercised in a browser.
Bohdan Triapitsyn