Commit Graph
21999 Commits
Author SHA1 Message Date
OrcaWinandOrcaWin a6c19f592f test: remove low-value boundary and coordination tests (#43466)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 21:57:03 -08:00
Neil 65c84d4110 refactor(renderer): group app shell and activity modules by owner (#43435) 2026-11-16 21:41:41 -08:00
OrcaWinandOrcaWin 4e6b4068be test: remove low-value output, PDF and download suites (#43459)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 21:37:57 -08:00
OrcaWinandOrcaWin 354c91cd81 test: remove low-value renderer seam tests (#43452)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 21:29:34 -08:00
Neidan MartínezandNeeidaan 7a5746deb3 fix(dev): resolve pnpm via shared helper so pnpm dev:fresh starts on Windows (#43420)
dev-fresh-profile spawned a bare `pnpm` with shell:false, which fails with
ENOENT on Windows where only pnpm.cmd/pnpm.exe exist. Use
resolvePnpmCliInvocation() like the other dev scripts so the child pnpm is
the one that launched the script.

Fixes #43413

Co-authored-by: Neeidaan <228572363+Neeidaan@users.noreply.github.com>
2026-11-16 21:19:16 -08:00
Brennan Benson c2fced0c98 fix(native-chat): deliver other agents' messages to chat agents directly again (#43380)
* fix(native-chat): deliver other agents' messages to chat agents directly again

#43093 sent chat agents only a pointer at `orca orchestration check` for
mail from other agents, while the chat showed the full message, so the
agent never received what the chat displayed and every message cost an
extra tool call. Chat agents again receive each orchestration message
whole, with its sender header, exactly as the chat card shows it; a
message too large for one chat message still falls back to the pointer.
The Approve once control from #43093 is unchanged.

Tracked in STA-15406.

* fix(orchestration): keep custody hold and owner wake for mail not delivered whole

Whole-delivered mail is read on settle. A pointer or in-doubt send still
holds while custody is unreadable, is handed to the mailbox's new owner
when the chat lost it, and a send made obsolete by a handoff wakes that
owner.
2026-11-16 21:17:17 -08:00
Jinjing 07dc0e78e9 feat(sidebar): show the card style intro as a side note with picture tiles (#43440)
The intro now floats right of the sidebar as a non-modal note with Classic/New
picture tiles, a switch-back hint and Not now / Keep new (or Use classic).
It stays out of Escape and focus ownership, waits for a confirmed save before
announcing Keep new, and returns focus to where it came from.
2026-11-16 21:15:02 -08:00
Neil dca8d39701 fix(codex): pass the chat's model to thread/fork so /btw side chats keep the selected model (#43438)
thread/fork does not inherit the parent's model, so a side chat forked from
a Codex chat was saved on the config default and later resumed on it. Pass
the chat's selected model (falling back to the source thread's model) and
the source thread's provider to thread/fork.

Fixes #43414
2026-11-16 21:13:46 -08:00
OrcaWinandOrcaWin b2e62ae0f1 test: remove low-value launch history and path tests (#43439)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 21:11:29 -08:00
Brennan Benson 09eb545121 test(agent-launch): give at-rest launch waits the host rig's 10s settle window (#43433)
* test(agent-launch): give at-rest launch waits the host rig's 10s settle window

The open-failure retry case waited for dispatch under vi.waitFor's 1s default. It took 778ms on a green
bun shard and ran out on a shard about 2x slower, though the retried child had opened.

* test(agent-launch): leave the private child-state waits untouched
2026-11-16 21:02:26 -08:00
Neil 54db328114 fix(sidebar): show primary worktree delete explanation in tooltip (#43432) 2026-11-16 20:55:52 -08:00
OrcaWinandOrcaWin e39fb3af12 test: remove low-value orchestration and lifecycle test cases (#43429)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 20:43:45 -08:00
OrcaWinandOrcaWin 2f17e24a18 test: remove low-value startup and IPC test suites (#43419)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 20:26:11 -08:00
NeilandClaude bfe15312a2 fix(linear): fetch branchName for project and custom view issue lists (#43417)
Project and custom view issue lists never selected or mapped Linear's
branchName, so starting a workspace from those lists fell back to the
generated branch name instead of Linear's "Copy git branch name" value.

Fixes #43404

Co-authored-by: Claude <noreply@anthropic.com>
2026-11-16 20:17:53 -08:00
Brennan Benson 956ba85de3 Read newer saved terminal IDs without losing their bindings (#43098)
* Preserve terminal ownership when reading host-first saved IDs

Read host-qualified terminal IDs while keeping the existing writer format. Reuse workspace-scoped allocation for fallback terminals so fresh panes cannot collide after restart. Tracked in STA-14326. Rebuilt from the identity work in #40339.

* Restore qualified saved terminals through their existing physical IDs

* Read qualified catalog terminal references before live import

* Keep saved terminal regression fixtures deterministic
2026-11-16 20:11:27 -08:00
Brennan Benson 28c8b06b21 Resolve Windows drive paths inside the WSL distro for Source Control (#43052)
* Resolve Windows drive paths inside the WSL distro for Source Control

Source Control failed for a Windows-drive repository when its WSL distro mounts drives somewhere other than /mnt/ (wsl.conf [automount] root): Git ran cd /mnt/c/..., changed files never loaded and Generate stayed disabled. Worktree deletion checked the same wrong path.

Drive paths are now resolved by the distro itself at the moment of use (wslpath on the drive root), so Git, the line-ending probe, worktree removal checks, object quarantine and OpenCode file operands follow the distro's own mount root with no host-side copy to go stale. A drive with no readable mount stops the command instead of running it from the wrong directory. Git output under a custom root maps back to the drive through the existing per-distro root reader.

Tracked in STA-15054.

* Read generation recipe files from a Windows-drive workspace in WSL

Commit, PR and branch generation with a file recipe failed before launch in a WSL + Windows-drive workspace: the recipe reader passed the Windows workspace path to a runner that accepts only guest paths. The workspace and a drive recipe path now resolve under the distro's own drive mount, as Git does.

Tracked in STA-14217.

* Map Git output to drives once the distro's mount root is read

Before a drive checkout's WSL Git runs, the distro's automount root is read if it has not been since the distro started, reusing the existing root reader, so paths Git prints under a custom root map back to the drive instead of a UNC path that removal guards would not match. Mount lookups no longer leak wslpath's stderr into missing-path detection, multi-line arguments such as commit messages are never treated as drive paths, an uppercase /mnt/C directory is not a drive, and removal resolves its two paths in one guest call.

* Refuse drive Git output when the distro's mount root can't be read

A failed or timed-out root read now fails the Git call instead of reading its output against /mnt/, so a custom-root listing can never reach removal guards spelled as UNC paths; a failed read is not retried for ten seconds, so polls fail fast without another wsl.exe. Background Git is admitted before the read, so it never boots a stopped distro, and a cancelled call stops waiting for the read.

* Read the mount root once per start without adding work to warm Git calls

Per the design ruling: the root read before drive Git allows for a cold distro start (60 s) instead of failing a default-root user's first Source Control load; admission runs only when a read is needed, so warm background polls add no wsl.exe; probes and Git share the ten-second failure backoff; and a read overtaken by a distro stop reads once more, re-admitted, instead of reporting a failure.

* Recheck the mount root after admission so concurrent callers share one read

A caller that waited for admission while another caller read or failed to read the root now uses that outcome instead of starting a second read inside the failure backoff.
2026-11-16 20:10:52 -08:00
Kelvin Amoaba 993e112a56 fix(mobile): size the terminal once when it starts with a bundled font (#43410)
The document's startup wait covered only the symbol fonts, so when the
picked bundled face loaded after them the first fit measured the fallback
font, native sized the PTY from it, and the face commit resized it again.
The pending face load now lives on the scope and the startup wait also
waits for it, inside the same 2 s deadline.
2026-11-17 03:54:01 +00:00
NeilandClaude 4e7b06695f fix(browser): tolerate a missing or throwing frame list when classifying guest discard protection (#43408)
readBrowserGuestDiscardProtection read guest.mainFrame.frames.length
directly. Electron can omit the frame list or throw when the render
frame was disposed, which crashed the classifier. Treat an unreadable
frame list as possibly having child frames so the guest stays protected
(reason "frames") instead of throwing.

Fixes #43391

Co-authored-by: Claude <noreply@anthropic.com>
2026-11-16 19:52:24 -08:00
OrcaWinandOrcaWin 513e9a4bec test: remove low-value hook-root and resume test suites (#43402)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 19:52:08 -08:00
katonandNeil e9060c2039 fix(runtime): close removal watchers without reauthorizing a revoked worktree path (#43371)
Worktree removal on POSIX re-ran resolveAuthorizedPath before closing a
leftover watcher child. Once Git had dropped an external worktree's
registration, that check threw PATH_OUTSIDE_ALLOWED_DIRECTORIES and the
whole removal aborted. Fall back to the resolved spelled path for the
in-memory watcher lookup on that error; other errors still propagate.

Fixes #43368

Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
2026-11-16 19:47:06 -08:00
Neil 225baa2477 refactor(web): group API and session sync modules by feature (#43403) 2026-11-16 19:38:16 -08:00
Brennan Benson b30d12992e fix(crash-reporting): a fatal OOM crashes with a report instead of livelocking; name the worker and payload behind it (#43327)
* fix(main): crash instead of spinning on a fatal OOM, and name the worker behind it

A JS listener on SIGTRAP/SIGABRT (signal-exit, loaded by proper-lockfile) made
Chromium's fatal-OOM trap re-fire forever at full CPU instead of crashing.
Every worker thread now carries a role name, records what it posts to its parent
in a fixed shared-memory ledger, and main reports heap pressure across isolates,
so the next out-of-memory failure names the worker that caused it.

* test(worker-threads): ledger names the sender and frees its slot on exit

* fix(worker-threads): refuse fault-signal listeners synchronously, survive crashes with sender evidence

Review fixes: the guard refuses SIGTRAP/SIGABRT listeners before libuv installs a handler and also
runs in orcad and the backend process; large sends are logged to disk before posting and the last
heap sample per isolate goes into crash annotations; slot release no longer races a new claim,
a full ledger evicts its stalest slot, and the payload walk never invokes accessors.

* fix(worker-threads): cast-free guard and ledger, keyed collections count as rows, worker-only large-send log

Second review pass: refuse/forward listeners through Reflect instead of casts; drop slot eviction
(a full ledger stops recording instead of racing a release); count big keyed objects toward the
large-send threshold; scale bytes when the walk is cut short; move the large-send append into a
worker-only module with an explicit sync-fs baseline line.

* fix(worker-threads): visit each payload object once under one entry budget; no byte extrapolation

Third review pass: a wide self-referential object re-enumerated its keys per visit; the walk
now uses a seen-set and a global key/element budget. Bytes stay a lower bound, and a truncated
walk counts as large from 4 MB or 10k items instead of an extrapolated estimate.

* fix(worker-threads): split main heap into old and large-object space in crash annotations

* fix(native-chat): refuse a journal reader reply over 256 MB instead of posting it to main

A decoded whole-chat read (647 MB of JSON, ~931 MB retained) posted from the journal SQL reader
exhausted the V8 heap main shares with every worker. Replies the bounded walk finds large are
measured and refused with a failure the caller already handles; paged reads never pay for it.

* fix(native-chat): cap reader replies at 64 MB from measured decode cost; never refuse the record set

Decoding costs main 1.0-1.9x the serialized bytes, so 256 MB could still cost ~490 MB of the
shared heap. 64 MB keeps one reply under ~120 MB. The record set loads the chat store and grows
only with chat count, so it is exempt rather than refused.

* test(crash-reporting): a real process with proper-lockfile holds no SIGTRAP listener in either load order and keeps SIGTERM cleanup

* fix(native-chat): measure any reader reply the bounded walk cut short; test the reply boundary

Review: rows past the 2,048-node walk could hide a 456 MB reply. A truncated walk now always
measures the serialized size (on the reader thread, never main). The reader posts every reply
through limitJournalReaderReply, which the test exercises directly.

* chore(worker-threads): satisfy lint/format, split hang worker-send parsing, mock environment data in hang tests

* chore(worker-threads): typed calls instead of Reflect in the send ledger, signal guard and hang send parsing

* fix(native-chat): page items-by-kind visits by 8 MiB of JSON so a page of large tool outputs stays under the reader reply limit

* perf(native-chat): read items-by-kind payloads lazily so rows past the page budget are never loaded

* fix(native-chat): never refuse a reader reply that carries one stored item; no page can split it

* fix(worker-threads): list the spawn module's new imports in the web typecheck project

* fix(worker-threads): walk Map keys through built-in accessors, bound the logged kind, page items-by-kind by key lookups, record the spawned thread name, skip SIGTRAP subprocess tests on Windows

* refactor(native-chat): keep the items-by-kind payload lookup's SQLite calls in the reader's item reads
2026-11-16 19:19:04 -08:00
LITTLEandNeil 5262435ab8 fix(browser): route Paste, Copy and Select All to extension popups (#43374)
Extension action popups and chrome.windows.create popouts were not
registered as native popup windows, so the app menu sent Edit actions
to Orca's renderer over IPC, which extension pages never receive.
Mark both window kinds in the existing native popup registry so
Paste, Copy and Select All act on the focused extension field.

Fixes #43369

Co-authored-by: Neil <neil@stably.ai>
2026-11-16 19:16:26 -08:00
OrcaWinandOrcaWin 10fbb749a7 test: remove low-value HTTP clock and persistence fixtures (#43398)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 19:13:57 -08:00
Jinjing 3d3c6bc1fc fix(sidebar): line up worktree card icons beside the time (#43395)
* fix(sidebar): line up worktree card title-line icons beside the time

The new cards gave both the title-line icon group and the time group an
auto left margin, so the free space after the title was split between
them and link/automation/PR icons floated mid-row by title length. Only
the first trailing group now takes the auto margin, so icons sit in one
column next to the time.

* fix(sidebar): keep title-line icons and time in one trailing group

Wrap the icon row and the time row in a single pushed group so every
template host stays static, drop the class-asserting test, and extend
the seeded fixture with a depth-two row, a details-row icon and a
narrow sidebar.

* test(sidebar): add status-pill rows to the trailing-icons fixture
2026-11-16 19:10:07 -08:00
Brennan Benson 1bf1a3f6a9 Approve a repo's orca.yaml hooks once on mobile (#42935)
* Bring mobile setup trust to approve-a-repo-once

Mobile now counts any earlier approval as repo trust (one shared predicate
with desktop), offers Always trust / Run once / Don't run, stores nothing
for Run once, reads the setup of the ref a create checks out, and pins a
Run once create to the shown content's hash. Automations need Always trust.
Same-id repos on several hosts keep per-version approvals.

* Pin mobile Run once only when the host read the created branch

repo.hooks reports createSourceRead when it honored createSource. Without
it (older host, or an unreadable ref) the phone sends no hash, as before,
so a checkout read can't make the host skip setup silently.

* Recover skipped mobile setup and ask about branch-only setup

Run once always pins the shown setup. When the host skips setup because the
created branch differs (worktree.create or agent.launch), the phone shows the
branch's setup; Run once or Always trust runs it via worktree.runApprovedSetup.
Setup approval now applies whenever setup isn't skipped, reads the PR/MR start
point once, pins runtime-host automation setup, and carries an archive Run once
into Delete Anyway. Drops the createSourceRead reply flag.

* Pin empty mobile setup and keep skipped setup reportable

An untrusted create whose branch shows no shared setup pins empty content,
as desktop does, so setup the branch adds is re-asked. Pinned creates use
agent.launch only on hosts that report skipped setup there, else
worktree.create. Approved setup that fails to spawn reports spawn_failed.
The sheet creates from the start point it previewed, Tasks keeps Create
disabled until a review opens the workspace, and runtime automation pins
read the automation's base branch.

* Report approved setup that fails to start and finish Tasks reviews

Approved setup or default tabs that can't start now fail worktree.runApprovedSetup,
so every client's existing failure handling shows it. An empty setup pin keeps
agent.launch on older hosts. Tasks and the sheet open a reviewed workspace with a
setup-failed notice when approval can't run, and a cleared review no longer
leaves Create disabled.

* Pin empty setup for runtime-host automations

A runtime host can't trust the repo, so saving setup=run where the base branch
has no shared setup pins empty content, as desktop does. Setup the branch adds
later is skipped for review instead of running unprompted.

* Offer only per-version approval on mobile for a repo on several hosts

As desktop: repo-wide trust is ignored for a repo id on several hosts, so the
drawer hides Always trust and its scope line, and Run hooks is primary.

* Pin empty automation setup everywhere and report failed default tabs

Phone automation saves for an untrusted repo pin the shown setup, including
empty content, on paired desktops too; only repo trust leaves them unpinned.
An approved default tab that fails to start now fails the approved-setup run.
A double tap on a Tasks setup review approves once.

* Follow main's settings module move in the approved-setup test

* Keep setup approvals tied to what runs on mobile and in approved default tabs

Re-decided automation saves name the setup decision so the host replaces the saved pin. Work-item creates check setup trust for the repo they land in. Approved default-tab commands start as new tabs when the layout already holds the command-less ones. Tasks setup reviews stay busy and ignore other choices until the approval settles.

* Check a work item's own repo setup and skip repeat approved runs

A create landing in another repo inherits that repo's setup policy and is always trust-checked unless skipped. The Tasks trust drawer acts once per prompt even before its busy render. A base-only automation edit re-reviews setup. A repeated approval on the same worktree and content doesn't open its command tabs again.

* Send the user's setup choice on cross-repo creates and drop the approval memory

A work-item create in another repo now sends the form's own setup decision and only widens the trust check: it runs for that repo unless setup is explicitly skipped. The per-worktree approved-run memory is removed, and the mobile lockfile change is reverted to main's.

* Set the automation setup pin on every preparation path
2026-11-16 19:06:57 -08:00
OrcaWinandOrcaWin 51fbddeec8 test: remove low-value permission status and WSL fixtures (#43394)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 19:01:59 -08:00
Neil a130a99c3f fix(agent-status): show the approval card when Claude asks for permission in Chat UI (#43382)
Claude's PermissionRequest hook publishes a waiting row with the approval
prompt, then reports the permission block in its terminal (OSC 7501 or a
permission title). That transition invalidated the hook row, so the runtime
projection rebuilt the status from the terminal alone and dropped the
interactivePrompt, toolName and toolInput. Chat then showed "Awaiting user
input" with no Allow/Deny card.

A move into permission now records the floor of the interval it ends (at
most 5 s before the report). A waiting/blocked hook observed after that
floor owns the permission interval and keeps its prompt. Older hooks, or a
later block with no hook of its own, still fall back to the plain wait.

Fixes #39286
2026-11-16 18:55:39 -08:00
Neil f730fd05a1 fix(terminal): stop stray escape tails on Windows after returning to a busy pane (#43348)
* fix(terminal): stop stray escape tails on Windows by starting the daemon output sequence at first attach

On Windows, returning to a busy terminal could print the tail of an escape
sequence on the prompt row ("?2026h", "149C", "5;46m").

ConPTY's startup preamble is replayed into a new session before its creating
client attaches. The daemon counted those bytes in the session's output
sequence, but main counts a new session's stream from attach, so the daemon's
sequence ran 20 chars ahead of main's. A restore that used a daemon snapshot
(after a hidden-pane data gap or a renderer reload) was labelled 20 chars past
the stream position it covered; the renderer then dropped the next 20 chars of
live output, often from the middle of an escape sequence.

The output plane now counts only bytes emitted once a client has attached, so
both sides count the same stream. Protocol v57 rolls the fix into a fresh
daemon.

* test(terminal): pin the bytes the attached client receives in the pre-attach sequence test
2026-11-16 18:51:45 -08:00
OrcaWinandOrcaWin 1b10368eb2 test: remove low-value provider transcript and WSL fixtures (#43381)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 18:50:33 -08:00
OrcaWinandOrcaWin 37beab4327 test: remove low-value hook relay migration and WSL fixtures (#43378)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 18:31:12 -08:00
Brennan Benson 822a1b104a fix(usage): show the soonest reset beside the name in Detailed rows instead of a repeated top-right meter (#43359)
* fix(usage): show each window's reset in Detailed rows and drop the repeated top-right meter

* fix(usage): show the soonest reset beside the name in Detailed rows, not per window

* fix(usage): skip passed resets when picking the soonest one
2026-11-16 18:28:04 -08:00
OrcaWinandOrcaWin 7593dd1312 test: remove low-value browser catalog and shell repro fixtures (#43367)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 18:18:37 -08:00
Neil b5547ea635 fix(tab-bar): omit active strip tab from new-tab switch results (#43344)
* fix(tab-bar): omit active strip tab from new-tab switch results

* refactor(tab-bar): reuse visible tab identity for switch filtering
2026-11-16 18:09:07 -08:00
Marcelo CeolinandNeil 3f8dafa5ef fix(native-chat): attach pasted images in the web client over plain HTTP (#43356)
Insecure origins (plain HTTP) expose no async clipboard, so the web
client's clipboard image save re-read navigator.clipboard, got nothing,
and silently dropped the pasted image chip. The native chat paste path
now hands the paste event's own image to saveClipboardImageAsTempFile in
the web client; desktop keeps reading the native clipboard because a
Blob cannot cross Electron IPC.

Fixes #43355

Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
2026-11-16 18:04:48 -08:00
Jinjing 820026a532 Clarify Session History search messages (#42519)
* fix: clarify session history search copy

* fix: keep paired-client notice in session search enable prompt

Restore the notice that authenticated paired clients can search the
index, since the prompt's Enable button turns indexing on directly.
Update the panel search e2e spec for the new prompt and empty-state
copy, drop a unit test that did not fail when the copy change was
reverted, and remove the retired consent key that the rebase brought
back in the ja and zh catalogs.

* fix: localize session search prompt copy and use paired devices wording

Use the reviewed wording "Paired devices can search them too." for the
enable prompt, and add translations for the three new Session History
strings in every shipped locale so the consent stays readable outside
English.
2026-11-16 18:02:34 -08:00
Wooseong KimandNeil 18a5a94750 fix(wsl): stop launching wsl.exe when WSL is not installed (#40778)
* fix(wsl): stop launching wsl.exe when WSL is not installed

Windows 11's inbox wsl.exe opens the installer on every invocation. Remember a missing WSLService and LxssManager for this process so distro, status, and running probes do not retry.

* fix(wsl): only treat a missing service as no WSL, and refuse every wsl.exe spawn once known absent

sc.exe exits 1060 only when the service is not registered. Access denied, a
timeout or a signal kill now stay unknown instead of hiding a real install
until restart.

Once absence is confirmed, process-host refuses any wsl.exe spawn with ENOENT,
so per-distro background work (stale saved distros, hooks, git, watchers)
cannot reopen the installer either.

* fix(wsl): reuse an unknown WSL service answer for 60 s

Denied or timed-out sc.exe queries were not cached, so every running-list
read spawned another sc.exe pair even inside the wsl.exe failure backoff
(ten reads: twenty sc.exe). Unknown now holds for 60 s; confirmed absence
stays a session latch. Covers each service name alone reporting present.

* fix(wsl): never query WSL services synchronously

Sync distro and home reads now use the cached service verdict. Without one,
they return what is cached and start the async query rather than spawning
wsl.exe, so a sync read on startup can't open the installer either. The main
process starts the query at launch. List the new module in the web tsconfig.

* fix(wsl): keep sync WSL readers answering before the service verdict

Before the async sc.exe verdict, sync readers returned empty, so a default-WSL
Claude launch could pick host credentials and a daemon session could freeze
an undefined distro and skip the stopped-distro hold. They now check, once per
process and without spawning, for LxssManager.dll or wslservice.exe: present
keeps the previous sync wsl.exe path, absent returns empty. An expired unknown
verdict keeps being served while it re-checks.

* fix(wsl): read WSL service registration from the registry before sc.exe

The file-path hint missed Store MSIX and custom-INSTALLDIR layouts. Sync
readers now check HKLM\SYSTEM\CurrentControlSet\Services\{WSLService,
LxssManager} through the bundled in-process registry addon: a key proves the
service is registered (no sc.exe at all), while a missing key or unloadable
addon leaves the verdict to sc.exe's 1060, and sync readers spawn nothing
until then. Drops the sync-fs baseline entry. The stale-unknown re-check is now
an explicit peekWslInstallationAndRefreshIfDue.

* test(wsl): keep the sc.exe service-query suite independent of the host registry

On a Windows host with WSL installed, the real WSLService key answered before
the mocked sc.exe and failed 9 of the suite's tests. Mock the registry reader
as having no WSL keys. Also fix a stale function name in a comment.

---------

Co-authored-by: Neil <neil@stably.ai>
2026-11-16 17:59:37 -08:00
OrcaWinandOrcaWin 8b298a6a61 test: remove low-value routing shell and tracker fixtures (#43364)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 17:51:09 -08:00
Jinjing 5c9cae27cb fix(ui): widen split divider gaps to 16px so dividers are easy to grab (#43357)
* fix(ui): widen divider grab areas to 16px on the side away from scrollbars

* fix(ui): keep the diff tree separator on its original pixels

* fix(ui): leave sub-pixel divider crossings to the parent divider

* fix(ui): type the split handle overhang style without a cast

* fix(ui): ignore synthetic events in the divider native-drag guard

* feat(ui): widen tab and terminal split gaps to a real 16px instead of overlays

* test(ui): cover the 16px terminal divider box

* fix(ui): paint wide split gaps with the editor surface

* fix(ui): route gap file drops to the nearer pane, keep the seam under the pointer, and continue tab strips across split gaps

* fix(ui): layer the tab band border over the strip fill

* fix(ui): drop the gap drop relay; follow container moves and commit at release

* test(ui): drop the terminal divider size test that restates its formula

* ci: drop the e2e route to a spec deleted in #43014

The feature.terminal-pane-geometry route still named
issue-12656-terminal-link-tooltip, which #43014 removed, so any PR touching
terminal.css or pane-divider.ts failed 'changed e2e specs' on a missing file.
2026-11-16 17:32:15 -08:00
OrcaWinandOrcaWin 4b8a9b10c0 test: remove low-value launch and runtime inventories (#43354)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 17:22:54 -08:00
Brennan Benson be32ea71eb fix(settings): Enter in Settings search opens the first result and tighten its top gap (#43360) 2026-11-16 17:19:57 -08:00
Neil 212694c08e refactor(ipc): group handlers and PTY support by owner (#43339)
* refactor(ipc): group handlers and PTY support by owner

* chore(test): prune retired IPC timeout baseline

* fix(types): include worktree path settings in web project
2026-11-16 17:11:09 -08:00
Neil 19bb71d636 fix(markdown-preview): keep scroll position when adding a review note (#43346)
Review-note state was a dependency of the react-markdown components map, so
every note change produced new component types and React remounted the whole
preview, resetting scrollTop. Annotated block components (p, pre, blockquote,
table, li, headings) are now module-level and read annotation renderers from
a context, and links read worktree routing at click time, so the map stays
stable across note changes. Find highlights are re-applied when annotations
change because the block wrapper can detach the highlighted ranges.

Fixes #43317
2026-11-16 17:03:57 -08:00
Neil bd77b557a8 fix(terminal): show the real Windows error when conpty.node fails to load (#43337)
* fix(terminal): show the real Windows error when node-pty cannot load conpty.node

* fix(terminal): keep the first load error across the retry and cap the retry wait per module
2026-11-16 17:00:07 -08:00
Brennan Benson e0b4cc0396 fix(checks): stop a merged PR from blocking Create after new commits (#43342)
When a branch's PR is merged and new commits land on the same branch,
the Checks tab kept showing "Pull request already exists" with Open
Review pointing at the merged PR. The hosted-review cache read and the
create-eligibility lookup both ignored the worktree HEAD, so a merged
review stayed current until its cache entry expired.

Filter stale merged reviews out of the Checks tab's hosted-review cache
read, and pass the worktree HEAD into create eligibility (Checks,
Source Control, Create PR intent, and main's create preflight) so the
branch lookup re-reads merged reviews once HEAD moves.
2026-11-16 16:58:27 -08:00
OrcaWinandOrcaWin 6a830e8be3 test: remove low-value cache, lock and conversation suites (#43338)
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
2026-11-16 16:53:39 -08:00
Brennan Benson 46b91ea0e8 fix(native-chat): hide overlays of hidden chats and show the agy model (#43328)
* fix(native-chat): hide overlays of hidden chats and show the agy model

Open popovers, menus and dialogs portaled from a chat tab stayed on screen,
pinned to the top-left, after switching tabs: a hidden Activity does not hide
portals beneath an already-hidden host. Each chat pane now portals overlays into
a body-level container that hides with the pane.

Antigravity reads its model from user-level config only, so the configured
default it reports is now recorded and new chats show the model instead of
"Model".

* test: stub the act environment without a cast
2026-11-16 16:43:15 -08:00
Kelvin Amoaba fc50fb4075 feat(workspace-cleanup): delete Orca chats from deleted workspaces with their chat history (#43015)
* feat(workspace-cleanup): delete Orca's own chats of deleted workspaces with their history

* fix(workspace-cleanup): close a deleted chat's shown tab the usual way and leave its rows to the next tick

* fix(workspace-cleanup): keep a chat worker that still owes a dispatch

* chore(workspace-cleanup): baseline the forgotten-chat journal SQL, which runs in the journal workers

* fix(workspace-cleanup): follow the tab table's move after rebasing onto main

* fix(workspace-cleanup): check the gone folder after the catalog re-read, nearest the delete
2026-11-17 00:23:04 +00:00
GiorgioandClaude Opus 5.5 7ef6053588 perf(orcad): notice orcad's readiness within 50 ms of it being written (#43305)
The host-side readiness wait slept 250 ms between checks, so a launch
found orcad ready on average ~125 ms after it published readiness, and
up to 250 ms later; its fractional-sleep probe also slept 250 ms before
the first check. Every activation, recovery and rollback launch waits
in this loop.

Hosts with fractional sleep now test for the readiness file every
50 ms with the shell's builtin `[ -s ]`, which starts no process. The
checks that do (the deadline, a complete readiness line, a launch that
already exited) still run every 250 ms, and on every iteration once the
file has bytes. Hosts with integer-only sleep keep the 1 s cadence, and
Windows keeps its own wait.


Claude-Session: https://claude.ai/code/session_01GR9M8uNzGJ1Gaam7ML1Sk3

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 16:21:32 -08:00
Neil 41194fe4ae fix(pi): install Orca's managed Pi extensions on native chat launch (#43332)
Pi native chat launched Pi with PI_CODING_AGENT_DIR pointing at the chat
account dir but never wrote Orca's managed extensions there. Only the PTY
env path called installManagedExtensions, so orca-agent-status.ts was
missing (or stale after an upgrade) for chat-launched Pi and subagent rows
never appeared.

piRpcLaunchPart now refreshes the managed extensions in the chat agent dir
when agent status hooks are enabled for Pi. The gate is read from the
host's current settings via a shared structuredAgentSettingGates helper.

Fixes #43291
2026-11-16 16:18:18 -08:00