* feat: implement desktop boot outcome architecture
- Add structured DesktopBootOutcome with target/status fields
- Implement boot outcome computation and validation
- Add desktop hosts configuration management (Tauri + TypeScript)
- Add desktop hosts probing with timeout and retry logic
- Support local/remote host classification and health checks
This provides the foundational infrastructure for desktop onboarding
flow to determine whether to show local setup, remote connection,
or recovery screens based on OpenCode availability and remote host
reachability.
* feat: add desktop onboarding UI components
Add comprehensive onboarding flow for desktop app:
- ChooserScreen: First-launch local/remote selection
- LocalSetupScreen: CLI installation guidance and manual detection
- RecoveryScreen: Recovery mode with routing to local/remote
- RemoteConnectionForm: Remote host connection with validation
- DesktopConnectionRecovery: Recovery variants and routing logic
- ConnectionSettingsPage: Manage remote connections
Components handle:
- Local vs remote choice persistence
- Recovery scenarios (unreachable, wrong-service, missing)
- Manual CLI detection (replaced auto-polling)
- Back navigation and state preservation
* feat: integrate desktop onboarding with app shell
- Update App.tsx to handle onboarding routing and recovery
- Add onboarding mode switching (first-launch/local-setup/recovery)
- Integrate desktop hosts in SettingsView
- Update DesktopHostSwitcher with recovery routing
- Add desktop shell utilities for onboarding detection
- Update web manifest for desktop app metadata
Completes the desktop onboarding feature integration,
allowing users to choose local or remote OpenCode on
first launch and recover from connection failures.
* fix: hide back button in remote connection form for first-launch chooser
In first-launch chooser mode, the back button is redundant since users
can simply click the "Local Install" tab. The back button is still shown
in recovery mode where there's no tab interface.
Changes:
- Add showBackButton prop to RemoteConnectionForm (default: true)
- Set showBackButton={false} in ChooserScreen remote tab
- Keep showBackButton={true} in RecoveryScreen for navigation
* refactor: remove Connection Settings page and simplify recovery UI
Remove the Connection Settings page as it was redundant:
- Local server is single-instance (no need to "choose")
- Remote servers are one-time setup (first-launch chooser)
- SSH Instances remain for multi-instance management
Changes:
- Remove ConnectionSettingsPage component and directory
- Remove 'connection' from Settings metadata
- Remove "Open Settings" button from recovery screens
- Remove desktopBootBypassToSettings state and logic
- Update recovery config to use 'local' icon instead of 'settings'
- Update tests to reflect removed showOpenSettings field
This simplifies the UX by focusing on:
- First-launch chooser for initial local/remote decision
- Remote Instances (SSH) for managing multiple remote machines
- No persistent "server management" needed for typical desktop usage
* fix: remove unused enableCliPolling prop and clean up TypeScript errors
Remove the obsolete enableCliPolling prop that was used for auto-
polling CLI detection. We replaced this with manual "Check and Continue"
button in a previous commit, so this prop is no longer needed.
Changes:
- Remove enableCliPolling from OnboardingScreen props and usage
- Remove enableCliPolling from App.tsx calls
- Remove unused 'connection' case from getSettingsNavIcon()
- Remove unused RiGlobalLine import
This resolves all TypeScript compilation errors reported by Copilot.
* fix: remove unused onChooseLocal prop and CLI_MISSING_ERROR_REGEX
These were left over from the refactoring:
- onChooseLocal in RecoveryScreen was defined but never used
- CLI_MISSING_ERROR_REGEX in App.tsx was leftover from removed enableCliPolling code
* fix: remove unused variables and fix React Hook dependency warnings
Remove unused memoized components and variables that were causing
lint errors in packages/ui:
- MainLayout.tsx: Remove unused MemoHeader, MemoChatView, MemoPlanView,
MemoGitView, MemoDiffView, MemoTerminalView, MemoFilesView,
MemoRightSidebarTabs, DesktopLeftSidebar, and DesktopRightPanel
- useGitHubPrStatusStore.ts: Remove unused prVisualPriority function
- useChatScrollManager.ts: Add missing markProgrammaticScroll dependency
to React.useEffect hook
These fixes resolve the CI lint failures in PR 850.
* chore: remove local claude settings from repo
* refactor(desktop): drop vibrancy code from onboarding PR
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
* fix(task): prevent subagent silent failures in session resolution and polling lifecycle
Two failure points fixed:
1. Fallback session resolution window too narrow (3s):
- resolveFallbackTaskSessionId now accepts hasRetried boolean
- First attempt uses 3s window (avoids binding wrong sessions)
- Subsequent attempts widen to 8s (handles late-appearing child sessions)
- Uses useState + useEffect instead of side effects in Zustand selector
2. Polling stops before child results are captured:
- When child session goes idle before parent sees it active, polling
would stop without fetching results
- Added final-fetch-before-stop: a one-shot delayed fetch that runs
after the settle grace period, ensuring child results are captured
- Uses taskFinalFetchDoneRef to guarantee exactly one final fetch
- Preserves existing happy path (active child → normal settle timer)
* fix(task): serialize final fetch after polling stops
Move the subagent final-fetch into a dedicated effect that runs only after
polling has stopped, avoiding races between polling writes and final-fetch
writes to the child sync store.
Also retry safely on final-fetch failure by reopening polling instead of
marking the final fetch as done before the request succeeds.
* feat(task): distinguish child session errors from normal idle
When a subagent terminates with an error, abort, timeout, or failure,
the parent session could not tell it apart from a normal completion.
Changes:
- event-reducer.ts: session.error now stores { type: 'error' } instead of
{ type: 'idle' }, so consumers can distinguish failed from completed sessions
- useSessionActivity.ts: add 'error' phase to SessionActivityPhase and
isError flag to SessionActivityResult; error phase is non-active (like idle)
but distinguishable via isError
- ToolPart.tsx: pass childSessionError to TaskToolSummary and show
'Subagent session ended with an error.' instead of the generic
'No subagent session id on task metadata.' when the child errored
Other consumers of session_status that only check 'busy'/'idle' are
unaffected — 'error' falls through to existing idle-like behavior.
* Revert "feat(task): distinguish child session errors from normal idle"
This reverts commit b3cc749bde16ed8fc55e4f304dcc455484335fba.
* fix(task): delay fallback retry window widening
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
* fix: show question content in ToolPart instead of 'Awaiting response...' after refresh
Previously, when the question tool was pending/running or completed
without parseable output, the ToolPart fell through to a generic
'Awaiting response...' message. After a page refresh or app restart,
this made questions appear empty even though the tool state still
contained the question input data.
Now the ToolPart reads question text, headers, and options from the
tool state's input field, ensuring question content persists across
refreshes regardless of QuestionCard store availability.
Fixes#879
* fix: restore QuestionCard after refresh and pause working status during active questions
Two fixes for question tool UX:
1. ChatContainer: sessionIsWorking now returns false when there are
active questions (same as it already did for permissions). This
prevents the status row from showing 'Asking question...' and
instead shows the QuestionCard.
2. sync-context: resyncDirectoryAfterReconnect now re-fetches
pending questions via listPendingQuestions(). Previously only
sessions and messages were re-fetched on SSE reconnect, so
questions asked during disconnection were lost, causing
QuestionCard to disappear after page refresh.
Refs #879
* fix: hide assistant working status while questions are pending
The assistant status hook only special-cased pending permissions, so
question tools still surfaced 'Asking question...' after refresh even
when the UI was already waiting on a QuestionCard response.
Treat pending questions like other blocking requests by clearing the
working indicator until the user answers.
Refs #879
* fix: merge question/permission stores instead of full replace on bootstrap and reconnect
The root cause of QuestionCard disappearing after refresh was a race
condition between SSE events and HTTP bootstrap. Bootstrap and reconnect
both did full replacement of state.question, wiping SSE-delivered data
that arrived between the HTTP call initiation and response arrival.
Changes:
- bootstrap.ts: question and permission stores now use merge semantics.
Only sessions present in the API response are overwritten. Sessions
absent from the response are left untouched (they may hold SSE data).
- sync-context.tsx: reconnect question resync uses the same merge pattern.
No longer clears question entries for sessions not in the API response.
- bootstrap.ts: sdk.question.list() now passes directory parameter to
scope the query correctly.
This ensures SSE-delivered question data survives the bootstrap window,
while still allowing the API response to be authoritative for sessions
it covers.
Refs #879
* fix: prune stale pending requests after reconnect
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
- Fixes header drag behavior in Desktop when connected to remote hosts
- Routes all window-drag entry points through a shared safe desktop helper
- Keeps local desktop drag behavior unchanged while avoiding remote IPC failures
Add reflect-metadata bootstrap so desktop sidecar no longer crashes on startup.
Update SDK v1.4 compatibility for model variant and diff payload handling.
Unify write/edit/apply patch expanded previews and hide write success output noise.
OpenCode 1.4.0 moved `UserMessage.variant` from the top-level to
`UserMessage.model.variant`. OpenChamber still read the legacy location,
so variant selection silently stopped being restored on sessions served
by OpenCode 1.4.0+.
Read `model.variant` first and fall back to the legacy top-level field
so both pre-1.4 and 1.4+ OpenCode servers keep working.
When a user navigates to a subagent (child) session, the parent session
node in the left sidebar was not automatically expanded, making the
child session invisible in the tree. This change adds a useEffect that
detects when the current session has a parentID and automatically adds
that parentID to expandedParents, ensuring the subagent session is
visible in the sidebar hierarchy.
* fix: restore git changes panel visibility and sidebar sync
Two independent fixes:
1. ChangesSection virtualizer returned empty rows due to useMemo caching
getVirtualItems() with a stale stable reference. First render produced
an empty array, and useMemo never recomputed because rowVirtualizer
reference never changed. Removed the useMemo to call getVirtualItems()
directly on every render. Also added a ResizeObserver to force
remeasurement on visibility transitions (defensive).
2. RightSidebarTabs now keeps git status fresh while the sidebar is open
via useRightSidebarGitSync hook (10s polling with ensureStatus).
Replaces the GitPollingProvider removed in commit d9821716.
* fix(virtualizer): prevent React error #185 from render-phase getVirtualItems
Restoring useMemo for virtualRows with totalSize as an invalidation
dependency. Calling getVirtualItems() directly during render triggers
the virtualizer's maybeNotify() → onChange() → useReducer dispatch,
causing React error #185 (https://react.dev/errors/185 — cannot
update a component while rendering a different component).
The original useMemo([rowVirtualizer, shouldVirtualize]) was removed
because rowVirtualizer is a stable useState ref that never changes,
leaving the memo permanently stale after the first empty render.
Adding totalSize (from getTotalSize()) as a dependency solves this:
it changes whenever the virtualizer recalculates after measure/scroll,
ensuring getVirtualItems() returns fresh rows while staying wrapped
in useMemo.
* fix(chat): replace hover bridge with padding to unblock desktop interactions
PR #826 removed pointer-events-none from the hover bridge div to fix
short user message revert button hover, but the transparent 44px
full-width bridge blocked all interactions (clicks, text selection)
with messages below it on Tauri desktop where the rendering difference
made the overlay consume pointer events across the entire message area.
Replacing the bridge div approach with pb-12 padding on the message
bubble container (max-w-[85%]) when sticky inline hover is active.
This extends the hover hit area downward within the bubble's own
padding, maintaining continuous hover path from bubble to action
buttons without any overlay that could block sibling elements.
The padding approach works because:
- The 48px bottom padding extends max-w-[85%]'s hover area to cover
the gap between bubble bottom and the absolute-positioned buttons
- Hovering the padding counts as hovering group/user-shell (parent)
- No absolute-positioned invisible layer means nothing blocks the
next message's interactive content
* fix(chat): reduce hover padding from pb-12 to pb-5 for tighter spacing
* fix: hide archived section and empty folders when no sessions remain
- Only push archived group in useSessionGrouping when there are archived
sessions, preventing an empty archived section from rendering
- Hide empty folders in archived bucket via shouldKeepFolder check in
SessionGroupSection (folders with no sessions and no content in
children are filtered out)
- Always filter folders through shouldKeepFolder, not just during search
* perf: memoize archived folder filtering
- Add configurable visible/instructions prompt families for commit/PR generation, PR checks/comments flows, and git conflict resolution helpers.
- Refactor prompt sending to explicit visible + synthetic parts instead of newline-based splitting, with legacy override migration for old keys.
- Polish Magic Prompts settings UX with grouped sidebar entries, tooltip-based descriptions, AI icon, and validation that visible prompts cannot be empty across web and VS Code runtimes.
Roll back aggressive chat reuse paths that froze message parts, tool state, and completed assistant bodies when sessions were inactive or backgrounded.
- Fix OpenChamber proxying to the correct OpenCode host for remote VPS setups
- Show a clear empty-state error when OpenCode is not reachable
- Harden session event routing for early or mismatched directory events
- Background polling pauses when tab is not visible
- Timer-based animations replaced with efficient frame-based updates
- Reduced CPU usage when OpenChamber is in background
- Unify send and session resumes around the latest chat tail
- Keep smooth follow active during assistant message growth
Remove staged chat rendering from the main scroll path
- Adjust package metadata and lockfiles across the workspace
- Refine chat sidebar and auto-scroll behavior in the UI
- Update PWA install handling alongside runtime package changes
- Keep older assistant messages in the active turn stable while new content streams
- Prevent existing tool rows from rerendering when new tool activity is appended
- Tighten message and tool memo comparisons to isolate render work to changed rows
- Keep assistant auto-scroll smooth during active responses without periodic snaps
- Restore live `/compact` rendering and streamed compaction summaries without refresh
- Preserve global event handling while routing session and message updates efficiently
- Show `/compact` command and streamed compaction summary live without refresh
- Handle versioned sync event types so message and part updates are processed
- Route session/message events via indexed lookup to avoid hot-path store scans
Replace live assistant-part status subscriptions with coarse session phase, retry, permission, and abort state so the composing row no longer rerenders on every text delta.
- Keep historical chat messages stable during streaming
- Isolate active stream state to the current message path
- Reduce chat container churn from sync message updates
Stop historical chat view from waking on live tail deltas by suspending detached part updates, narrowing scroll manager dependencies, and hardening sync-context HMR behavior during dev.
Split static history from live tail, stabilize chat data identities, and reduce scroll-shell churn so old viewport content stops waking on stream deltas. Also defer watcher startup and health polling so idle runtime work better matches actual usage.
- Cut broad render fanout across the app by replacing shared-store whole-object subscriptions with leaf selectors, memoizing hot chrome boundaries, and isolating disabled global providers from live session/message state. This keeps header controls, composer toolbars, side panels, and other non-hot UI surfaces from repainting on every assistant update or keystroke.
- Rework sidebar session ordering so recent, project groups, and worktree groups derive from one ordering source while avoiding streaming-time thrash. The sidebar now uses a stabilized session snapshot, preserves structural identity for unchanged rows, reads live row status/details per session, and applies a one-shot sort bump on idle->busy instead of continuously resorting during activity.
- Fix chat/input scroll instability by separating viewport-resize handling from message-growth handling, disabling conflicting native scroll anchoring, and stopping textarea autosize from collapsing on every growth keystroke. This removes the multiline typing jiggle during streaming and reduces unnecessary composer rerenders.
- Also gate voice context wiring behind voice-mode enablement and codify the learned render/scroll/order anti-patterns in AGENTS.md so future changes avoid the same classes of regressions.
- Add startup feature flag for plan mode from OPENCODE_EXPERIMENTAL* env vars.
- Stop recurring app health polling and keep a one-time startup check.
- Skip plan-mode synthetic parsing and plan tab/file work when flag is off.
- Prevent desktop chat input from inheriting mobile bottom safe-area gap
- Apply chat input safe-area classes only on mobile
- Scope standalone PWA safe-area and blur overlay CSS to mobile runtime
- Prevent header flicker to Untitled during session refresh
- Resolve header session data from global sessions first
- Pass session directory hints on more session-switch paths
* fix: resync session state after SSE reconnect to prevent stuck subagent UI
When a subagent completes while the page is in the background (common on
mobile PWA and desktop webview), the final SSE events are lost. The UI
then stays stuck on 'Waiting for subagent activity...' because:
- part.state.status never transitions to 'completed'
- session_status is never updated to 'idle'
- activeLatched remains true indefinitely
Fix:
- Add onReconnect callback to event pipeline, fired after SSE reconnect
- Add pageshow listener for bfcache restores (mobile PWA back-forward)
- On reconnect, re-fetch session list for directories with active sessions
- Pass explicit directory to useSessionActivity in ToolPart for subagents
to ensure the correct child store is queried
Closes#810
* fix(chat): resolve pending subagent task binding before metadata arrives
* fix: restore subagent activity and tool visibility after reconnect
- Resyncs session status and child session data after SSE reconnect
- Ensures child task tool messages are read from the correct directory
- Prevents stale assistant fallback from keeping sessions stuck as active
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
* feat(terminal): add resumable websocket transport
Unify terminal input and stream traffic on `/api/terminal/ws` with a v2 control-frame protocol and advertised transport capabilities.
Buffer recent PTY output on the server so rebinding clients can replay missed chunks after reconnects or startup races, while keeping SSE as a fallback stream path.
Update the web terminal client and store to negotiate the new transport, track tab lifecycle, and avoid reopening exited sessions when restoring tabs.
* fix(terminal): retry rehydrated websocket reconnects
* fix(terminal): keep reconnect retries silent
* fix(terminal): wire ws stream output and replay in runtime
* fix(web): enable ws proxy for /api in dev
---------
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
- Add error icon and colored background/border for error messages
- Add break-all and max-w-full to prevent long error text overflow
- CSS: word-break and overflow-wrap for inline code in error messages
* fix(worktree): reset IntegrateCommitsSection state when switching worktrees
Three fixes for the re-integrate commits panel getting stuck:
1. Add `key={worktreeMetadata.path}` to IntegrateCommitsSection so React
fully remounts it when switching to a different worktree, clearing any
stale `ui` state (conflict, loading, ready) from the previous session.
2. Add `cancelled` flag to the conflict-restore effect so that an async
callback started for session A cannot overwrite session B's state after
the user switches sessions. Without this guard the stale callback could
restore the old session's conflict state on top of the new session's
computed-ready state.
3. Fix off-by-one in continueIntegrate: `moved` was returning
`remaining.length` (N-1, after shifting currentCommit out) instead of
`state.remainingCommits.length` (N), undercounting the commit that was
moved by `cherry-pick --continue`.
* fix(worktree): add git-based fallback detection when store metadata is missing
Root cause: the existing worktreeMetadata resolution relies entirely on
cached store state (worktreeMap + availableWorktrees). When the store
lookup fails—due to hydrateSessionWorktreeMetadata deleting entries on
API failure, availableWorktrees being stale, or worktrees created
externally via CLI—the "Re-integrate commits" section permanently shows
"Available in worktree mode." with no way to recover.
Fix: add useDetectedWorktreeMetadata hook that performs a lightweight
git probe (`git rev-parse --absolute-git-dir --abbrev-ref HEAD`) when
the store-based lookup returns undefined. If the current directory is a
secondary git worktree, a minimal WorktreeMetadata is synthesised with
the correct projectDirectory and branch, allowing IntegrateCommitsSection
and other worktree features to function regardless of store state.
The store-based lookup remains the primary fast path; the git probe only
runs as a fallback and caches its result per directory.
* fix(worktree): fix detection command and pass current branch from git status
Two bugs in the fallback worktree detection hook:
1. `git rev-parse --absolute-git-dir --abbrev-ref HEAD` combines two
independent rev-parse options whose combined output is unreliable –
the two-line assumption (`lines.length < 2`) caused silent null
returns, meaning the fallback never actually set worktreeMetadata.
Now uses only `git rev-parse --absolute-git-dir` (single-line,
deterministic output) for worktree detection.
2. The hook was called before `useGitStatus`, so no branch was
available. Move the call to after `const status = useGitStatus(...)`
and pass `status?.current` as `currentBranch`, eliminating the need
for a second git command and keeping the branch in sync with the
already-polled git status.
Also removes `detected` from the useEffect deps array – it was an
unnecessary dep that triggered a re-run on every detected state change.
* fix(worktree): use worktree toplevel path and reset stale metadata immediately
Two bugs in useDetectedWorktreeMetadata:
1. path was set from currentDirectory (the active sub-folder) instead of the
worktree root. git rev-parse --show-toplevel now provides the actual
worktree toplevel, so operations like `git worktree remove` receive a valid
root path regardless of which sub-directory is open.
2. When currentDirectory changed with no storeMetadata, the hook kept
returning the prior detected value until the async git probe finished.
Calling setDetected(undefined) before launching the async task eliminates
the stale-metadata window.
- Add JsonTreeViewer component with collapse/expand for nested objects and arrays
- Add rainbow colors by depth level using CSS color-mix() with syntax theme tokens
- Add virtualization support (@tanstack/react-virtual) for large JSON files (>200 nodes)
- Add JsonTreeView wrapper with Expand All / Collapse All toolbar
- Integrate into FilesView: tree/text toggle button, JSON file detection
- Integrate into ToolPart: auto-detect JSON tool outputs, render as tree
- Integrate into ToolOutputDialog: JSON tree view in expanded dialog
- Add jsonTreeUtils.ts: parse, flatten, path utilities
No new dependencies - uses existing @tanstack/react-virtual.
Colors derived from existing syntax.* theme tokens - works across all themes.
Co-authored-by: Bohdan Triapitsyn <artmore@protonmail.com>
Tie macOS window transparency to the vibrancy setting so disabling vibrancy actually removes transparent backgrounds. Clarify settings copy that full transparency changes take effect for new windows or after restart.
