dev:desktop now launches its own Metro on a random port + Electron,
no dependency on a shared :8081. In worktrees, Electron gets a unique
userData path and dock badge so multiple instances run side-by-side.
Per-token trace logs from the query pump fire twice per streaming event
and contribute to memory growth under heavy use. Gate them behind an
opt-in env var so they're available for debugging but silent by default.
Trace-level logging emits two entries per streaming token per session,
causing significant memory growth and GC pressure under multi-workspace
use. Fixes#227.
On iPad, Combobox uses the desktop Modal path (no BottomSheet context)
since the breakpoint is "md"+, but ProviderSearchInput was checking
Platform.OS === "web" to choose the input component. This caused
BottomSheetTextInput to render outside a BottomSheet on iPad.
Align the check with Combobox's own isMobile breakpoint logic.
Adds logic to process file parts and convert image attachments into data URLs for the OpenCode provider, replacing the previous behavior that stripped non-text parts.
Prevents the daemon section from flashing "not running" / "PID —" while
IPC data loads. The whole card now shows a spinner until the first fetch
completes, and mutations (restart/stop/start) optimistically update the
query cache with the returned status.
Replace truncating text with a read-only TextInput + copy button pattern
so the pairing URL doesn't get cut off. Increase QR code size from
200x200 to 320x320.
- Rename findExecutable → findExecutableSync, add async findExecutable
that uses promisified execFile instead of execFileSync
- Simplify Windows PATH resolution: rely on inherited env from Explorer
instead of shelling out to PowerShell for registry PATH entries
- Add spawnProcess() helper that centralizes Windows shell/quoting concerns
- Update all agent providers and consumers to use async executable resolution
- Add windowsHide: true across all Windows child process calls
Claude Code writes local CLI command outputs (model switches, /context,
/compact, goodbye messages, etc.) as user entries in the JSONL history.
These were leaking through as user_message timeline items in Paseo.
Add a create-release job that runs first and creates the GitHub release
before any platform build jobs start. This prevents the race condition
where parallel electron-builder jobs each create their own release.
Platform jobs now depend on create-release and just upload assets to
the existing release. Single-platform retry tags (desktop-macos-v*,
desktop-linux-v*, desktop-windows-v*) skip the create-release job
since the release already exists from the original build.
Ensures renderer console.log output (including layout-debug) is
printed to stdout in packaged builds, not just the log file.
Also removes unnecessary webContents event listeners.
Pipes all renderer console.log output to the main process stdout via
electron-log, and logs did-finish-load/did-fail-load/render-process-gone
events to help diagnose layout and loading issues on Linux.
- Log fatal errors during daemon bootstrap and startup to pino (daemon.log)
instead of only writing to stderr, which is invisible on Windows desktop
- Delay process.exit to let pino async streams flush
- Make voice MCP bridge script resolution non-fatal (warn + disable instead
of crashing the daemon)
- Add PASEO_ELECTRON_FLAGS env var for passing Chromium flags on Linux
- Add layout debug logging to AppContainer for diagnosing sidebar issues
Re-add usage stamping on turn_completed events for OpenCode agents —
the extractAndResetUsage callsite was removed during context window
meter work, causing token counts and cost to be lost after each turn.
Update workspace tests to account for the new async reconciliation
that fires after workspace update fanout.
Split diff now renders two independent scrollable columns (left/right)
instead of one scroll wrapping both sides. Line number gutters are
pinned outside the scroll area in both unified and split views.
Broaden the Codex speak tool name matcher to recognize paseo_voice.speak
MCP calls alongside the existing paseo.speak convention. Wrap voice
transcription input in <spoken-input> tags so agents can distinguish
spoken from typed messages.
Stop resetting lastContextWindowUsedTokens between turns. The
result.usage fallback contains accumulated session totals which grow
incorrectly across turns. Once a task_progress arrives, its value is
the accurate context fill level and should be retained until the next
task_progress supersedes it.
Passing an arrow function to FlatList's ListHeaderComponent caused React
to treat each re-render as a new component type, unmounting and
remounting the entire header (including the bouncing dots animation).
Pass the element directly instead so React can reconcile properly.
The Silero VAD session ran for the entire voice mode lifetime without
resetting its LSTM hidden states. Over time the internal state drifted,
causing phantom speech detection on silence and getting permanently
stuck in the "speaking" phase.
Add reset() to TurnDetectionSession and call it after each completed
utterance to clear the LSTM hidden states between turns.
Settings providers section now uses useProvidersSnapshot instead of a
one-shot fetch, so it shares the same real-time data source as the
status bar and draft form. Adds a Refresh button to re-detect installed
provider CLIs without restarting the daemon.
PlanCard list_item rule wrapped children in <Text> instead of <View>,
causing layout corruption when children contained <View> elements
(nested lists, paragraphs). Also added missing paragraph rule to match
AssistantMessage renderer.
When creating a worktree, resolve the base branch by checking
origin/{branch} first, falling back to the local branch only when
the remote ref doesn't exist. This ensures worktrees start from the
latest upstream state rather than a potentially stale local branch.
Periodically fetch from remotes in the background so workspace git
watch targets pick up upstream changes without user intervention.
The manager deduplicates subscriptions per repo and triggers a
workspace refresh callback after each successful fetch.
Remember which directories and diff entries are expanded across sidebar
close/reopen cycles. State is stored per-workspace in the panel store
(zustand + AsyncStorage) with no local React state copy, following the
proven explorerTabByCheckout pattern. On desktop remount, expanded
subdirectory listings are re-fetched so the tree renders fully.
- Lower active button surface from surface3 to surface2 for dark mode
- Add gap between toolbar button groups
- Remove outer border from unified/split group, use per-button radius
emitWorkspaceUpdateForCwd and emitWorkspaceUpdatesForCwds blocked on
reconcileActiveWorkspaceRecords() (stat + reconcile on every active
workspace) before pushing the target workspace update to clients. This
caused new worktrees to appear in the sidebar with a noticeable delay
after the agent was already visible.
Emit the target workspace immediately from the registry snapshot, then
run reconciliation in the background via reconcileAndEmitWorkspaceUpdates.
Ensure the dedupeGitState early-return still triggers background
reconciliation so metadata changes (branch renames, display names)
are not silently dropped.
- Model selector is no longer disabled while providers load; opens
immediately and streams available providers as they arrive
- Selecting a non-default provider on mobile now works correctly
- Provider icon added to model trigger in mobile preferences sheet
- Thinking icon (brain) added to thinking trigger in mobile preferences
- Model descriptions for OpenCode/Pi shown inline next to model name
instead of in a tooltip
- Running agents and draft screen share a single provider/model cache,
eliminating duplicate network requests
- Provider data is prefetched when entering a workspace so the model
selector is instant on first open
The mobile preferences modal split CombinedModelSelector's onSelect into
separate onSelectProvider and onSelectModel calls, but onSelectProvider
was never passed from DraftAgentStatusBar — so selecting a non-cloud
provider was silently ignored while only the model ID updated.
Add onSelectProviderAndModel prop to ControlledStatusBar and use it on
the mobile path, matching the web path's atomic behavior.
- Add onDropdownClose callback to AgentStatusBar and wire through ControlledStatusBar
to CombinedModelSelector so running agents focus input after any dropdown closes
- Fix mobile misalignment in model selector by adding marginHorizontal: spacing[1]
to sectionHeading, drillDownRow, backButton, and providerSearchContainer to match
ComboboxItem's mobile margins
- Add CombinedModelSelector with two-level drill-down (providers → models),
favorites section, search, and single-provider auto-drill mode
- Add provider snapshot system (ProviderSnapshotManager) for cached provider/model state
- Add provider diagnostics with resolved binary paths and versions across all providers
- Add StatusBadge shared component for consistent status visual language
- Add ProviderDiagnosticSheet for detailed provider health inspection
- Add desktop auto-focus on input after any status bar interaction (model, mode,
thinking, features) using focusWithRetries
- Fix Combobox flicker on height change by switching to bottom-based CSS positioning
once initial floating-ui position resolves
- Add dynamic dropdown height: Level 1 fits content, Level 2 calculates from model count
- Remove "Other providers" section from model selector (only show available providers)
On macOS, apps launched from Finder/Dock inherit a minimal environment
(PATH is just /usr/bin:/bin:/usr/sbin:/sbin). This caused two problems:
1. Agent binaries like codex were not detected (findExecutable failed)
2. Terminals spawned by Paseo had no access to user-installed tools
(node, bun, direnv — all "command not found")
Fix: at Electron startup, spawn the user's login shell and capture its
full environment via JSON.stringify(process.env) using UUID markers.
This is the same battle-tested approach VS Code uses. The resolved
environment is merged into process.env before the daemon starts, so
all child processes — agents, terminals, git operations — inherit the
correct environment automatically.
This replaces the previous approach of invoking resolveShellEnv() (via
the shell-env npm package) at multiple scattered call sites. Now there
is a single source of truth: process.env is enriched once at startup.
Changes:
- New: packages/desktop/src/login-shell-env.ts (VS Code approach)
- Removed: resolveShellEnv(), shell-env dependency, $SHELL -lic probes
- Simplified: applyProviderEnv() and findExecutable() now trust process.env
When dictating and focused on an agent, pressing Enter now confirms
the dictation and sends the message, matching the behavior of clicking
the submit button.
Remove the "Refreshing" and "Failed to refresh agent" toasts from the
agent panel — they fire frequently but are meaningless since sync
recovers transparently. The "Reconnecting..." toast for host disconnect
is preserved.
Strip debug console.log calls across the app: render tracking in
message/stream views, dependency change tracking in workspace screen,
terminal tab slot logging, and verbose audio engine/voice runtime
bridge stats logging. Legitimate console.error/warn in catch blocks
are kept.
Same fix as the Silero VAD — sherpa-onnx returns Float32Arrays backed
by native memory which Node.js rejects as "External buffers are not
allowed". Copy into JS-managed memory before passing to audio pipeline.
Users connecting multiple daemons (e.g. local + remote via Tailscale) hit
a "belongs to X, not Y" error because the edit-host modal's Add Connection
button scoped new connections to the current host's server ID. Remove that
button and its supporting state so all connections go through the top-level
flow which has no server ID constraint.
The bundled silero_vad.onnx lives inside Electron's app.asar which
native C++ (sherpa-onnx) cannot open via OS file I/O. On first voice
activation, copy the model to ~/.paseo/models/local-speech/silero-vad/
using Node.js fs (asar-aware) and point the native code there.
buildClientConfig in test-daemon-connection.ts never set appVersion, so
probe clients (which become the live client) sent hello without it. The
daemon's version gate then hid Pi/Copilot from all these sessions.
Also update appVersion on the Session when a client reconnects with a
newer version, so stale sessions don't stay gated forever.
worktreeRoot was added as a required field in 9154f8fc, which breaks
old clients parsing new daemon responses and new clients parsing old
daemon responses. Made it .optional() with .transform() fallbacks.
Also added a critical rule to CLAUDE.md: schema changes must always
be backward-compatible in both directions.
- Add classify.ts as the single source of truth for CLI invocation
routing (discriminated union, derives known commands from Commander)
- Remove all hardcoded command lists and duplicate path detection logic
- Simplify shell wrappers to dumb pipes (zero classification)
- Fix hot-start: use `open -n -g -a` (VS Code pattern) so second-instance
event fires when app is already running
- Fix cold-start race: pull-based IPC (getPendingOpenProject) so renderer
fetches the pending path after React mounts, instead of push event that
arrived before the listener existed
- Fix asar read corruption: unpack node-entrypoint-runner.js from asar
(Node.js v24 in Electron 41 can't parse package.json inside asar)
- Strip ELECTRON_RUN_AS_NODE from env when spawning desktop app from CLI
AgentProviderSchema was z.enum() which caused old clients to reject
session messages containing unknown providers (pi, copilot), breaking
the entire session. Changed to z.string() for future clients.
For currently deployed clients (<0.1.45), the daemon now filters out
unknown providers based on the appVersion sent in the WebSocket hello
message. Clients that don't send appVersion only see claude/codex/opencode.
Mohamed Boudra