* Enable accessibility tree (`ax`) command on iOS emulator sessions
Fetch the accessibility tree from serve-sim's /ax endpoint, which requires an
active session but provides the same UI snapshot capability as Android's
uiautomator output. Derive the endpoint from the stream URL when not explicitly
provided by the helper, and route through the bridge to pass session context to
the backend.
* Add ax command routing and backend integration tests
Tests verify accessibility tree routes through EmulatorBridge,
Android backend ignores iOS-specific ax URLs, and ax endpoints
are derived from serve-sim stream URLs.
* perf(runtime): idle websocket heartbeat without clients
Co-authored-by: Orca <help@stably.ai>
* fix(runtime): probe immediately when the WS heartbeat arms
Arming the heartbeat on the first accepted connection started a fresh interval,
so the first liveness ping was a full interval (~15s) out — a socket that died
right after connecting went unprobed for that window. Run one sweep synchronously
in start() so the first ping goes out at arm time; the seeded socket is pinged
(never reaped on the arm sweep) and reaped on the next tick only if it never pongs.
Tests updated for the earlier first probe.
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
* fix(daemon): retire macOS daemons whose login session died (#7936)
A daemon that survives a full macOS logout is unsalvageable: its PAM
context can no longer host login(1) spawns (every new PTY becomes a
'Login incorrect' prompt zombie) and its Mach bootstrap namespace has
lost the system DNS resolver, so terminals it hosts have no egress.
Today it also keeps the #9301 preflight's cached 'accepted' verdict, so
it keeps wrapping spawns in login(1) forever; only a manual daemon
restart recovers.
GUI-spawned daemons now watch for login-session death from the inside:
a fresh cache-bypassing PAM probe (triggered by PTY-exit bursts, fresh
client hellos, and a slow periodic timer) must conclusively reject
three consecutive times AND the in-process system resolver must be
degraded; then the daemon exits crash-style so session meta stays
unclean and the replacement daemon cold-restores scrollback. A
conclusive rejection also flips the spawn-wrapper cache off
immediately.
Headless serve/SSH daemons never get the watch (they must survive their
spawning session ending), and a session that never conclusively
accepted login(1) never arms it — a PAM anomaly alone can't kill a
healthy daemon (fast user switching keeps accepting, so switched-away
sessions are preserved).
* test(daemon): e2e seam to drive login-session death oracles from a verdict file
A dead macOS login session cannot be fabricated without root (PAM owns
audit-session teardown), so live lifecycle QA drives the death watch's
probe and resolver oracles from ORCA_E2E_LOGIN_SESSION_PROBE_FILE:
'alive' → accepted/healthy, 'dead' → rejected/unhealthy, anything else
inconclusive — with compressed watch timing. Mirrors the existing
ORCA_E2E_DAEMON_INIT_DELAY_MS seam; inert unless the env var is set.
* fix(daemon): close the hang-shaped gap in login-session death detection
The conclusive-PAM-verdict trigger had one blind failure shape: login(1)
hanging at the prompt past the probe bound (killed → inconclusive
forever → the watch never fires). Three changes close it:
- The death-watch probe gets its own 4s bound (the 500ms preflight bound
exists for spawn-path latency, which doesn't apply off-path), so a
slow-but-answering PAM stack isn't misread as a hang.
- An inconclusive pipe probe escalates to a PTY-hosted probe via
script(1) — a dead session's PAM stack may only misbehave under a real
tty (the pipe-vs-PTY fidelity limit the preflight documents).
- A streak of timeout-killed probes (which a live session never
produces) is a second retirement trigger, at a higher threshold (5)
and still gated on the degraded resolver, logged with a distinct
cause so field logs discriminate the two paths.
Every dead-session behavior — fast reject, prompt-then-EOF, or hang —
now fires retirement; all inconclusive states still fail toward
preserving the daemon.
* fix(daemon): keep login-session retirement conclusive
* fix(daemon): stop login watch before clean shutdown
* fix(daemon): make login-session PTY probe reliable
* fix(daemon): close login-session watch races
* fix(daemon): ignore health probes for login watch activity
Production crash diagnostics measured ~128 `git worktree list` execs/min
(9,400 in one 80-minute session, ~16% of wall-clock in git subprocesses):
the resolved-worktree scan fans out over every registered repo on a 30s
cache TTL, and most registered repos on the affected installs were
agent-CLI scratch repos (~/.codex-tmp capsules, vendor imports, skill
checkouts) that need no freshness.
Classify agent-scratch repo roots with a curated shared matcher and stamp
their scan-cache entries with a 5-minute TTL instead of 30s. Orca-driven
mutations still bypass the TTL via the per-repo generation bump, so only
passive pickup of external changes slows for scratch repos. Expected
steady-state reduction on the measured install: ~82% fewer git spawns.
Zone.js patches the global Promise with a non-native thenable. When a bare
`new Promise(...)` crosses the Electron executeJavaScript boundary, it's
serialized as-is, losing { page, target } and exposing __zone_symbol__*
fields instead. Wrap in an async IIFE to return a native promise that
Electron always unwraps correctly.
* feat(mobile-pairing): surface unpaired-device auth failures instead of silent 4001 loops
Desktop: when a phone repeatedly fails direct-transport E2EE auth with a
token missing from the device registry (pre-v1.4.106 pairing-path bug left
desktops that regenerated their registry rejecting paired phones forever),
throttle to one notification per session and show an actionable toast
pointing at Settings -> Mobile to re-pair.
Mobile: map a bare 4001 close onto the existing auth retry budget (the
encrypted e2ee_error is undecryptable when the desktop keypair changed, so
the close code is the only surviving signal) instead of looping the generic
reconnect forever, and make the auth-failed verdict say 'Pairing invalid -
re-pair with your desktop' instead of a bare 'Auth failed'.
* fix(mobile-pairing): handle stale keys and startup notification races
* fix(mobile-pairing): isolate auth notification failures
* fix(mobile-pairing): keep recovery alert actionable
* fix(mobile): persist per-device tab selection so worktree return restores the last open tab
A phone's tab selection lived only in the host's in-memory
ClientSessionTabSelectionStore. Any host restart wiped it, and the
per-device projection then fell back to deterministic topology, so
returning to a worktree on mobile always landed on the first tab
instead of the tab last opened on the phone.
Persist the per-device selections in the Store (keyed deviceId ->
worktreeId), hydrate them when the runtime constructs, and guard
projection so an early empty snapshot after restart cannot wipe a
hydrated selection before tabs arrive. Selections are pruned with the
worktree/repo and on device revoke, and malformed persisted payloads
degrade to empty instead of throwing.
* fix(mobile): harden persisted tab selection cleanup
* fix(mobile): preserve tab selection across worktree rename
* fix(agent-status): keep Claude in-process teammates visible as idle sidebar rows
Claude Code 2.1.21x runs named Agent-tool agents as turn-based in-process
teammates: SubagentStop and TeammateIdle fire at every TURN end while the
teammate stays alive awaiting mail (verified live on 2.1.217). Treating
those events as finish signals deleted the child row seconds after each
burst, so the sidebar showed no subagents for most of a teammate's life.
Root-cause fix: the roster now tracks a working/idle state per child.
- One-shot children (hyphen-free ids) keep remove-on-stop: their
SubagentStop is a true finish.
- Teammate-shaped rows park as idle on SubagentStop/TeammateIdle and
revive to working via the next SubagentStart (same lifecycle id,
first-observed startedAt preserved).
- Idle rows never gate the pane 'working' (#8825's done-gate rule).
- Only TeammateIdle-confirmed idle rows survive a complete lead-Stop
fold; a stopped workflow lane wearing a teammate-shaped id is reaped
there (or immediately, once a fold tagged it listedAsSubagentTask), so
the pre-#8825 idle pile cannot rebuild.
- At the wire cap, the oldest idle row is evicted to admit a working
spawn; working children are never displaced.
- Hydrate keeps pruning idle snapshots: idle-teammate liveness cannot be
proven across a restart, and a live teammate re-earns its row.
* fix(agent-status): restore inventory-confirmed workflow lanes
* perf(rate-limits): throttle statusline usage posts to one curl per pane per 15s
The managed Claude statusline script posted on every statusLine tick
carrying rate_limits — ~3 curl spawns/sec per streaming pane, multiplied
across concurrent panes. The service drops same-value posts inside its
30s dedupe window anyway, so most spawns bought nothing.
Gate the post on a per-pane stamp file: POSIX compares date +%s against
the stamp; Windows uses an all-builtin seconds-of-day parse of %TIME%
(octal-safe, no extra process). Both fail open — unparseable time,
garbage stamp, or midnight wrap posts rather than darkening the live
usage feed — and the stamp only advances when a post actually fires, so
skipped ticks never defer the next one.
Measured: 30 rapid rate-limit ticks spawn 1 curl (was 30).
* perf(rate-limits): keep throttled statusline ticks process-free
Use Claude's monotonic session duration for the POSIX throttle so skipped ticks do not replace curl churn with date churn, while retaining a fail-open date fallback. Key temp files by the stable leaf UUID so path-like or long host tab IDs cannot disable the throttle.
* fix(rate-limits): preserve throttle isolation across upgrades
* fix(rate-limits): reject leading-zero stamp values before arithmetic
All-digits validation still admitted values like 008, which are invalid
octal inside $(( )) — and dash treats that expansion error as fatal, so
the script died before rewriting the stamp and the pane's live usage
feed stayed dark until the file was deleted. Allow-list canonical
decimals (same pattern as the duration parse) on both the stamp and the
computed clock so malformed values fail open to posting.
Verified under dash: the old digits-only check aborts at the arithmetic
(Illegal number: 008); the allow-list survives and posts.
* perf(rate-limits): eliminate POSIX statusline cat churn
* test(rate-limits): exercise overlapping statusline ticks
The stamp check/write is deliberately lock-free (a lock could wedge the
feed closed; fail-open is the contract), so a truly concurrent burst may
post more than once, bounded by overlap width — an exact at-most-one
assertion would be flaky by design. Assert the invariants that do hold:
every overlapping run exits 0, the raced stamp lands valid, and it
throttles the following ticks.
* fix(naming): remove identifier-first name post-processing
Workspace display names and tab titles were being rewritten after
generation by prompt-scanning heuristics from #8238 — a stray "#1" in
prose became a workspace named "#1 - Fix", and the rewrite ran
downstream of generation so user naming instructions couldn't override
it. Per the same principle as #9088, naming defaults stay minimal and
user overrides own the style.
- Delete work-item-reference.ts and display-name-from-work.ts (+tests).
- Auto-rename display names return to the humanized branch slug; tab
titles return to the cleaned first prompt clause.
- Explicit create-from-work-item naming returns to action-first
("Review PR 1234").
- Keep #8238's URL-before-markdown strip-order bugfix in tab titles,
with regression tests adjusted to the natural expectations.
* test(naming): pin incidental marker regression
* fix(browser): stop failing goto when a redirect or download aborts the load
Electron's loadURL rejects with ERR_ABORTED (-3) when the initial
navigation is superseded — a client-side/meta redirect (common in SSO
flows) or a download-triggered load. Since #9633 drives goto through
wc.loadURL directly, that rejection surfaced as a spurious
'Failed to navigate' browser_error even though the page landed fine.
Treat ERR_ABORTED like offscreen-browser-backend already does: resolve
with the page's actual URL/title. Every other loadURL failure still
fails closed.
* fix(browser): settle replacement navigation after abort
* fix(browser): clean up aborted navigation destruction race
* feat: continue agent work in a new session
* fix: harden new-session continuation
* fix: source last prompt from provider-authenticated transcript records
Preview user entries can be tool results or harness-injected skill text,
so the continuation prompt's last-prompt hint now comes from the vault
scanner's provider-authenticated lastUserPrompt. Also softens the
continuation instructions for already-complete tasks and adds a cost
warning to the full-transcript option.
* fix: move Continue in New Session row action into the hover group
Edge-usage action; keep the resting row at three icons and reveal it with
the other session actions on hover, matching Resume's gating.
---------
Co-authored-by: jz.feng <jz.feng@aftership.com>
Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>
* fix(dashboard-popout): zoom the pop-out window, not the main window behind it
The View menu's zoom handlers always sent terminal:zoom to the main
window, so zooming while the dashboard pop-out was focused zoomed the
window behind it. The pop-out also never applied the persisted
uiZoomLevel, so it always rendered at 100% in a zoomed app.
The pop-out now applies uiZoomLevel on dom-ready and follows app-zoom
changes while open (main-window zoom, settings control, mobile ui.set).
Menu zoom routes to the pop-out when it is the focused window, stepping
its own webContents zoom, and a narrow before-input-event handler
resolves the zoom.in/out/reset chords (honoring keybinding overrides)
since the pop-out has no renderer-side shortcut plumbing. The step/clamp
constants move to shared/ui-zoom-level.ts so main and renderer share one
definition.
* fix(dashboard-popout): isolate window zoom
* fix(dashboard-popout): deny unused permissions
* docs(dashboard-popout): clarify wheel zoom path
* fix(dashboard-popout): wire terminal copy/paste in the popped-out window
The Edit menu's Paste is a custom item that routes Cmd/Ctrl+V to the
focused window as ui:appMenuPaste, and only the main window's React root
listens for it — the popout dropped it, so paste silently did nothing.
The copy chord similarly resolves in the main window's before-input-event
handler, which the popout window never registers; the menu's role:'copy'
no-ops on xterm's empty hidden textarea.
AgentTerminalPreview now subscribes to onAppMenuPaste (guarded on focus
inside the preview) and pastes via terminal.paste(), which xterm flags as
user input so the existing preview->PTY routing and main-process input
limits apply. A custom key handler honors the terminal.copySelection and
terminal.paste keybindings, skipping plain Mod+V since the menu
accelerator owns that chord (handling it twice would paste double). The
popout bootstrap fetches keybinding overrides so custom bindings apply.
* fix(dashboard-popout): harden terminal clipboard routing
* fix(dashboard-popout): authorize terminal clipboard text
* test(dashboard-popout): harden terminal paste coverage
* fix(terminal): normalize streamed paste newlines
* fix(dashboard-popout): forward macOS IME native-text commits in the preview terminal (#9771)
* fix(dashboard-popout): forward macOS IME native-text commits in the preview terminal
The preview terminal enables xterm's kitty keyboard protocol (via
buildDefaultTerminalOptions), whose encoder can encode and cancel a
printable keydown before Chromium commits the real IME/native text —
silently dropping macOS input-source commits and synthetic Unicode
injection. Main-window panes guard this with the IME native-text
forwarder; the pop-out preview never installed it.
Install the composition tracker + forwarder (macOS-only, mirroring
TerminalPane) and claim native-text key events at the top of the
preview's custom key handler so the committed glyph reaches the PTY via
terminal.input() and the existing user-input routing.
* fix(dashboard-popout): prewarm IME input source
* fix(skills): preserve released history across new tags (#9778)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(skills): record latest Linear release history (#9777)
---------
Co-authored-by: OrcaWin <alpha-eng@stably.ai>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
---------
Co-authored-by: OrcaWin <alpha-eng@stably.ai>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(codex): count per-account homes in usage and state removal blast radius
- usage scanner now includes codex-accounts/*/home/sessions so multi-account
usage is no longer silently undercounted (audit F2)
- account-removal dialog copy now states that session history and MCP logins
are permanently deleted with the managed home (audit F1 mitigation)
* fix(codex): harden account usage discovery
* feat(linear): add MCP-style save issue
* fix(linear): harden save issue parity
* fix(linear): close save issue contract gaps
* docs(linear): bundle project discovery with save issue
* fix(rate-limits): unstick Claude "Limited" usage and feed live usage from session statuslines
The OAuth usage endpoint's 429 Retry-After (~50 min) was ignored, so the
30s-15min automated retry lanes kept landing inside the throttle window and
the status bar stayed on a bare "Limited" indefinitely while Claude itself
worked fine.
- Respect Retry-After on 429: carry it through usageMetadata.retryAtMs and
gate automated refetches (activation lane, poll cycles) until it expires;
user-directed refreshes still bypass.
- Keep the last-known usage snapshot visible through rate-limited windows
(24h) instead of dropping it after the generic 30-minute stale threshold.
- Add a managed Claude statusLine command that forwards each session's
rate_limits (Claude Code >=2.1.80) to a new /statusline/claude loopback
route, feeding live usage windows with zero usage-endpoint calls; OAuth
polling pauses while the live feed is fresh. User-owned statusLine
settings are never overwritten.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(rate-limits): keep last-known window when a statusline post carries only one
Statusline payloads may report five_hour and seven_day independently; a
partial post must not wipe the other bar to null. Also document the
seconds-vs-ms epoch heuristic.
Addresses CodeRabbit review on #9617.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(rate-limits): unstick Claude usage with live statusline feed
The OAuth polling endpoint is rate-limited; Claude's status often shows
"Limited" until the next poll cycle, even when quota remains. Live posts
from the statusline command update usage within 100ms, eliminating false
"Limited" displays during active sessions.
Manages install lifecycle via marker to respect user deletions. Handles
Windows payload buffering and guards before curl spawn. Protects against
live-post/OAuth-fetch races and cross-attribution during account switches.
Gracefully tolerates schema drift in statusline parsing.
* test(rate-limits): assert stale outgoing post doesn't affect incoming
Capture usedPercent before ingesting and assert it remains unchanged,
rather than checking for a specific value. This is more precise and less
brittle when testing session switch isolation.
---------
Co-authored-by: Dzmitry Bachko <dbachko@users.noreply.github.com>
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* perf(runtime): gate PTY path-candidate extraction on mobile-connected
The 3-regex path extractor ran on every PTY chunk of every session for a mobile-only consumer. Now gated behind a sticky mobile-connected flag; first onReady synchronously backfills candidates from the retained 64KB window per original chunk boundary (lazy safety-net fallback). ~28.5us/chunk (~99.8%) saved on the desktop-only hot path; mobile parity preserved (Tier-1 raw-window read unchanged + backfill before any file RPC). 5-round review converged 2-consecutive-clean; independently verified tsc+902 tests.
Co-authored-by: Orca <help@stably.ai>
* fix(runtime): guard path-candidate activation on the remote runtime proxy
MobileSocketWiring.onReady called this.runtime.activateRecentPtyPathCandidateTracking() unconditionally, throwing on the E2EE/remote proxy (which lacks the method) and breaking the remote runtime connection (CI verify failure). Guarded with optional chaining (runtime-rpc.ts:900) — activation is a local-host concern. Mutation-verified E2EE-proxy regression test. Integration test + 3 typechecks green.
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
* perf(mobile): route graph-sync snapshot fanout through the notify coalescer
syncWindowGraph no longer unconditionally re-serializes+fans every worktree's mobile snapshot to every client. Serve-hydrate fast path skips the build-then-drop loop when no serve-owned pty exists (desktop case); changed-worktree gating (strict superset of snapshot inputs) routes only real changes through the existing coalescer. No-op syncs 720->0 emits; single-worktree churn 720->120 (~83% fewer bytes to phone). Real changes still propagate. 11-round GPT-5.6-Sol xHIGH review converged 2-consecutive-clean; independently verified tsc+830 tests.
Co-authored-by: Orca <help@stably.ai>
* fix(mobile): bump snapshotVersion on preserved prune frames + carry split-group layout
Pre-merge adversarial review found two stale-phone regressions: (1) preserved-headless snapshots did not bump snapshotVersion, so a pruned-but-serve-preserved worktree change was dropped by clients' same-epoch freshness gate (orca-runtime.ts:22316); (2) serve-only no-op hydrate dropped the phone's split-group layout (orca-runtime.ts:3731). Both fixed + mutation-verified regression tests. 3 typechecks + suites green.
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
* fix(native-chat): add reconciliation to catch missed transcript updates
Periodic reconciliation catches file changes that fs.watch may miss on
remote filesystems. Event-driven drains use a bounded max-wait timer to
ensure responsiveness. Extracts timer logic into a scheduler module for
testability and maintainability.
* fix(native-chat): add reconciliation to catch missed transcript updates
Detects same-size file rewrites and native watcher failures by reconciling
transcript state even when fs.watch events don't fire. Extracts watcher
binding into a dedicated module that gracefully degrades to
reconciliation-only mode when native watch is unavailable. Adds test
coverage for watcher rebind failures, parent-directory replacements, and
unsubscribe races.
* Document transcript watcher reconciliation contract
Clarify that bind() is best-effort with reconciliation fallback,
invalidate() detaches from stale identities, and the caller's polling
loop owns transcript liveness. Supports resilience on remote filesystems.
* fix(github): prevent unscoped work item queries
* fix(review): translate non-English rationale comment to English (client.ts:1151)
Restores the English '// Why:' convention and re-states the allSettled partial-results rationale that was dropped when the comment was replaced. Flagged by Ce-code-review (maintainability/previous-comments) and a prior PR reviewer. No behavior change.
Co-authored-by: Orca <help@stably.ai>
* fix(tasks): surface repos with no resolvable GitHub source
Since #9660 an unresolvable GitHub source returns an empty null-source
envelope instead of an unscoped search. That empty was indistinguishable
from a genuine zero-result query — no log, no telemetry, no UI signal.
Surface it in the Tasks list: a per-repo info row with Retry for each
fetched repo that resolved neither an issue nor a PR source (told apart
from genuine-zero and not-yet-fetched purely via the cached `sources`).
Renderer-only — the signal was already threaded through the envelope.
- selectTaskPageUnresolvedSourceRepos selector (+ unit tests)
- per-repo row reusing the existing banner region + handleRetryIssuesFetch
- suppress the generic empty state while those rows show
- assert the null-source envelope contract in client-work-items test
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: Orca <help@stably.ai>
* feat(dashboard): open-worktree dialog control + finished-time parity
Pop-out agent dashboard polish:
- Terminal dialog: replace the footer Close button with the standard top-right
X, rename the primary action from Focus worktree to Open worktree, and prefix
it with an open-in-window icon.
- Card time column: finished agents now read time-since-finish (matching the
left worktree sidebar) instead of time-since-start. Extracts the sidebar's
finished-timestamp logic into a shared helper so the two surfaces cannot
drift, and threads finishedAt through the dashboard snapshot contract.
* fix(dashboard): thread finishedAt through payload validator + test fixture
CI typecheck caught a DashboardCard fixture in dashboard-payload-validation.test.ts
missing the new finishedAt field; also validate finishedAt in the relay guard.
* feat(dashboard): swap pending-question emoji for MessageCircleQuestion icon
The amber pending-question chip used a raw hand emoji (renders differently
per-OS and off the design system); use the lucide chat-with-question icon.
xterm.js has no bold-color theme slot (upstream xtermjs/xterm.js#6032), so the Bold Text override was silently dropped — a false affordance. Makes the Settings copy and the Ghostty importer honest that it isn't applied yet, while preserving any persisted value for a future real implementation. No terminal-rendering change.
Closes#8595
Orca spread its whole process.env into every spawned terminal, so a dev build leaked NODE_ENV=development into integrated terminals (breaking next build, Vitest, etc.). Strips the inherited value at all three PTY-spawn sites (daemon, local provider, relay); a value the user sets themselves still wins.
Supersedes #9058, which patched only the relay path and so missed the reported (local daemon) repro.
Closes#9057
Co-authored-by: John Debay <john@debay.net>
* Access the Floating Workspace from mobile (#8405)
Surface the desktop Floating Workspace (the global, repo-less scratchpad
of terminal tabs under the synthetic `global-floating-terminal` id) on the
mobile app so a Claude session left running there is reachable from a phone.
Adds a terminal-icon button to the mobile host header (phone + tablet
sidebar) that opens the existing Session screen for the floating id. The
sentinel already had host-side RPC support (#5946: local runtime, homedir
cwd, explicit-id fast paths in session.tabs.*); this wires up the mobile
surface and gates it on a new `floatingWorkspaceEnabled` status flag so the
entry hides on hosts that predate it or where the feature is disabled.
The Session screen learns an `isFloatingWorkspaceRoute` flag (mirroring the
existing `folder:` route pattern) that hides repo-backed surfaces — Files,
Source Control, PR/checks, agent history — skips the diff-comment and GitHub
probes, routes terminal URL taps to the phone browser, and limits the New
Tab drawer to terminals + agents (browser/markdown creation resolves a real
worktree host-side and stays desktop-only). useLiveWorktreeName
short-circuits for the sentinel so it no longer polls worktree.show forever.
Extracted the host status.get gating into a useHostStatusGates hook to keep
the host screen under the max-lines ratchet.
* Harden mobile Floating Workspace routing
* Fix mobile host gate reuse race
* Harden floating mobile session polling
* fix(mobile): harden floating workspace route reuse
* fix(mobile): skip floating workspace repo lookup
* fix(mobile): clarify floating workspace header action
## Description
Every durable save serialized the ~1.5 MB durable state **twice**, synchronously on the main Electron thread: once in `computeStateHash()` (the plaintext hash for the no-op-write guard) and again in `buildStateToSave()` (the encrypted on-disk payload). The guard couldn't hash the payload directly because `encrypt()` on random-IV platforms produces different bytes each save for identical state.
This PR collapses the two into **one** `JSON.stringify`. `buildStateToSave()` now returns `{ payload, stateHash }`: each encrypted secret slot is serialized as a **fresh per-slot random sentinel** (`orca-secret-slot-<uuid>`), and after the single stringify each sentinel is substituted exactly once — to its ciphertext for the on-disk `payload`, and to its plaintext for the guard `hashInput`. The guard hash is therefore a pure function of the plaintext state (`hashInput === JSON.stringify(plaintext durable state)`), and the on-disk `payload` is byte-identical to the previous two-stringify output.
No change to the save API, debounce timings, guard semantics, or on-disk format (compact JSON, same 3 encrypted fields).
### Why sentinels (not a string replace of the ciphertext)
The substitution must be **position-exact**. A naive `payload.replace(ciphertext, plaintext)` — or even a `"key":"ciphertext"`-anchored replace — can be **mimicked by user-controlled state**, which would substitute the wrong site and let two *distinct* states normalize to the same guard hash → a real change is silently **not written** (data loss). This is reachable on deterministic-IV platforms (macOS / legacy-Linux OSCrypt use a constant IV, so a user can read their own ciphertext out of `orca-data.json`):
- **Value vector**: a plaintext free-text field (e.g. `httpProxyBypassRules`) whose value equals a secret's ciphertext.
- **Key vector**: `agentDefaultEnv` lets the user name an env var exactly after a secret field (e.g. `browserKagiSessionLink`) with a ciphertext value, producing a `"browserKagiSessionLink":"<ciphertext>"` token inside `settings` (before `ui`).
A per-slot random sentinel is minted **after** all user data is captured, so it cannot occur anywhere else in the serialized state and matches only the intended secret slot — closing the entire class. (Both vectors were found by an adversarial fable→gpt-5.6-sol review loop; each has a store-level regression test that fails on the earlier approaches and passes now.)
The empty-secret / `safeStorage`-unavailable / encrypt-failure cases (`blob === plaintext`) get no sentinel and reproduce `main`'s behavior byte-for-byte.
## Evidence
- Full-state `JSON.stringify` calls per changed save: **2 → 1** (call-counting test).
- Frontier-review benchmark on an 8 MiB real-method state: **no perf regression** (median 12.73 ms sentinel vs 13.77 ms two-stringify baseline).
- Independent review modeled 625 adversarial durable states: every payload byte-identical to the pre-PR implementation, and every distinct state produced a distinct guard hash (no dropped-write collision).
## No-Regression Proof
`src/main/persistence-single-serialize.test.ts` (nondeterministic **and** deterministic cipher mocks):
- identical state with secrets set → no-op (inode unchanged); real change (incl. secret rotation) → writes
- on-disk payload is ciphertext (plaintext appears nowhere), round-trips through reload/decrypt
- empty secret + `isEncryptionAvailable() === false` → plaintext payload, guard still skips
- sync `flushOrThrow()` also skips on identical state
- exactly one full-state serialization per save
- **regression (value vector)**: a plaintext field equal to a secret's ciphertext, then swapped → write is not skipped, round-trips
- **regression (key vector)**: an `agentDefaultEnv` var named after a secret field holding its ciphertext, then swapped → write is not skipped, round-trips
Suites: `persistence-single-serialize.test.ts` + `persistence.test.ts` all green; `pnpm run typecheck:node` clean; oxlint clean.
## ELI5
Orca saves your app state whenever something changes, and it used to build the whole (large) state twice each time — once to check "did anything actually change?" and once to write the file. Now it builds it once and reuses that copy for both jobs. To compare states safely it swaps each encrypted secret for a random one-time marker while checking, so no value you type can ever be mistaken for a secret and trick it into skipping a real save. What lands on disk is exactly the same, and your secrets stay encrypted.
Made with [Orca](https://github.com/stablyai/orca) 🐋
Jinjing
NeilandOrca
OrcaWin
Brennan Benson
Cal
