- Deduplicate candidates per command: 'tool' and 'tool.CMD' may resolve to
the same on-disk file and both must still be found.
- Probe known candidate names directly when readdir is denied (EACCES/EPERM):
an execute-only directory cannot be listed, but stat on known names works,
which is what the per-candidate lookup did.
- Skip the install-dir fallback's PATH re-scan: the batch already proved the
missed commands absent on the PATH it searched, and that PATH is a superset
of the fallback's (merged persisted PATH on Windows, the same filtered PATH
on a WSL host), so the fallback scans install directories only.
findCommandsOnLocalPath() replaces one isCommandOnPath() call per agent
command: it reads each absolute PATH directory once with readdir and matches
command names and PATHEXT permutations in memory, then verifies only the
hits through isExecutableFile. Symlink, directory, permission,
case-insensitivity and absolute-only PATH semantics match the per-candidate
lookup. On the machine in the issue this replaces about 24,000 stat() calls
per detection run with about 38 directory reads.
Fixes#34636
* fix(editor): restore diagram scroll after async rendering
* fix(editor): retain pending diagram scroll while hidden
* fix(editor): key Mermaid restore on rendered source and cancel it on user scroll
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(editor): preserve folds and horizontal scroll across tabs
* refactor(editor): keep reading state cache with its lifecycle
* test(editor): give stale-cache model its URI
* refactor(editor): keep editor view state cache beside diff view state
Move editorViewStateCache into scroll-cache.ts, clear it in the shared model
lifetime fixture, drop the e2e evidence spec and the duplicate LRU test.
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* Read sidebar PR and stack metadata in one GraphQL query
* Preserve direct PR error recovery without duplicate GraphQL reads
* Keep combined sidebar queries out of merge preflight
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* feat(wsl): offer Git for Windows in a Windows-drive WSL project's settings
Project settings show a Git row for a WSL-runtime project on a Windows drive (WSL
hosts on). "Use Git for Windows" checks the project, shows any blocker in plain words,
optionally sets core.autocrlf in the repo, and saves the returned runtime through the
normal project update. Once converted the row says where Git and terminals run.
Switching the distro keeps Git for Windows instead of silently dropping it.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(wsl): offer the Git choice only for a Git repo, not a folder
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Subscribe to onDidChangeModelContent directly instead of the
@monaco-editor/react onChange prop, which calls editor.getValue() before
Orca's ownership and programmatic-sync checks run. Programmatic echoes
(external reloads, shared-record commits, read-only live tail) no longer
materialize the full model text only to discard it.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Output stays buffered per task, so a long unit-test step looked hung for
30+ minutes. Print each task's completion, a 60s heartbeat naming running
tasks, and kill any step after 30 min (--timeout=<minutes> overrides).
Co-authored-by: Claude <noreply@anthropic.com>
Above the 1 MiB override cap a Markdown file can never render rich, so
skip the unsupported-syntax scan and keep those documents out of the
eligibility cache.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Conflict Next/Previous published a path-only reveal, so every editable pane
showing that file jumped, and overview mirrors never moved. Target the reveal
at the requesting file id and pane view id, and let read-only overview
mirrors accept only reveals that name them.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(editor): copy gutter URLs from the owning file
* fix(editor): scope gutter remote URL fallback to the editor owner
Look up the gutter menu's owning file by fileId, falling back to a
same-path match only within the editor's worktree so conflict-overview
editors with synthetic ids still copy a URL. Drop the e2e proof spec
and trim the unit test to the same-path/different-owner regression.
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Same-path files with different owners (runtime hosts or worktrees) shared
Monaco, rich-markdown, notebook, and PDF/media reading positions when the
tab id equals the file id. Key them by file id, matching the diff and
markdown preview keys.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(mobile): stop review notes disappearing or coming back after deletion
The phone sent its whole notes list on every add or delete, and the host
stored it as is. A note added on desktop after the phone loaded was wiped
by the phone's next edit, two quick taps could overwrite each other, and a
failed save rolled back to an old snapshot.
worktree.set now takes an optional diffCommentsBase, like linkedItemsBase:
the host applies only the writer's own adds, edits and removals to the
stored notes. The phone sends the list it started from, shows the host's
merged list from the reply, and undoes only its own change on failure.
Older hosts drop the field and replace the list as before; older phones
send no base and get the same.
Co-authored-by: Neil <neil@stably.ai>
* fix(mobile): keep loaded review notes when an early save fails
The first load was discarded once any save started, so a save that failed
before it landed left the host's notes hidden. The load now applies the
phone's changes made during the read onto the host's list instead.
* fix(mobile): send review-note requests one at a time
Two in-flight saves could reach the host in either order, so adopting the
later reply hid a note the earlier save had stored. Session notes now send
one request at a time and reconcile every reply with the phone's changes
made since it was sent, replacing the sequence counters.
* fix(mobile): drop review-note saves queued behind a failed one
A save queued behind one that failed still sent the list built on the
failed change, and an older host stores that list whole. Those saves now
undo their own change instead of sending. Also renames the shared base
type so it no longer shares a name with the renderer's mutation record.
* fix(mobile): show review notes as the host's list plus pending saves
A queued reply was reconciled against the list captured at tap time, so a
note desktop deleted meanwhile came back on the phone. The phone now keeps
the host's last list and its unanswered saves per workspace and shows the
two combined; a reply for another workspace is ignored. The review screen
also shows the host's list after a save, so an edit to a note desktop
deleted no longer looks saved.
* fix(mobile): fail queued review-note requests while disconnected
A notes request queued behind a pending load waited for a reconnect with
no timeout, holding every later save. Notes requests now fail at once when
the connection is down, like terminal input.
* fix(mobile): drop every review-note save built on a failed or dropped one
A save made while a dropped save was still pending carried its change,
so an older host could store it. Each failed or dropped save now moves
the failure count.
* fix(mobile): send each review-note save as the host's list plus its change
Each queued save now builds its payload when it is sent, from the host's
last list plus only its own change. A note added before the first load
no longer replaces the loaded notes on an older host, and a failed save's
change never rides along, so later saves no longer need dropping.
Delivered-note cleanup waits for a reconnect again so those notes are not
sent twice.
---------
Co-authored-by: Kelvin Amoaba <97001695+AmoabaKelvin@users.noreply.github.com>
The mount reconciliation already reads the full model text for its
stale-cache guard. Pass that snapshot into syncContentOnMount so a mount
reads the model once instead of twice.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
handleOpenDiffTargetFile keyed editor/markdown view-mode preferences by
the raw file path, but openFile returns owner-scoped IDs for remote
runtime files and same-path tabs in other worktrees, so Rich mode was
lost or written onto another owner's tab.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Key the notebook scroll listener and restore effects on notebook
readiness instead of raw content, so the listener attaches once an
invalid notebook becomes valid and content republishes no longer snap
the view back to a stale cached position.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Notebook cell models were anonymous and disposed whenever a cell lost
focus or the notebook view unmounted on a tab switch, so Undo did
nothing after returning to a cell.
Give cells that have a stable nbformat id a vscode-userdata model URI,
scoped to the view, so Monaco's model service keeps and restores the
disposed model's undo stack. Monaco restores only when the text is
unchanged and bounds retained history to its closed-file budget.
Id-less cells keep anonymous models because their position-based key
can move to a different cell, and a duplicate id falls back to an
anonymous model instead of colliding.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Snapshot the paste target (model, container, focus) before the context-menu
clipboard IPC read and before the yielding byte measurement of large pastes,
and re-check it afterwards so a model swapped during the await (e.g. a diff
refresh via setModel) does not receive the paste.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Monaco normalizes line endings per edit, so splitting a CR from its LF
across chunked executeEdits calls inserted an extra blank line.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(runtime): a device registry save never resurrects a device removed on disk
Every DeviceRegistry save wrote orcad's whole in-memory list. A lastSeen flush (250 ms after any
socket auth, and again at shutdown) could therefore write back a device an admin had just removed
from orca-devices.json on the host, silently undoing the revoke; it also dropped rows another
writer had added. Saves now re-read the file and write the in-memory change merged over it: rows
this process knew of that are gone from disk stay gone (in memory too), rows added on disk are
kept, and this save's own adds and removals apply. The write stays the secure-file atomic rename.
The pairing-repair e2e logs the registry's device ids after the kill and after the reconnect.
* fix(runtime): tombstone devices this process removed so a stale writer can't restore them
Review of #35483: after removeDevice(a), a stale writer (an old orcad's flush during an SSH
relaunch or upgrade overlap, an external tool) could write a back; the next save treated it as
new on disk, adopted it and let it authenticate. DeviceRegistryWriter now remembers every id
this process removed, recorded before the write so a failed load-time cleanup counts too, and
never adopts those from disk. replaceRow reports false when the merge dropped the row, and a
pending offer removed on disk is replaced instead of handed out.
* perf(orcad): spawn the terminal daemon before the runtime graph evaluates
The daemon spawn now starts right after the login env lands, before the runtime graph's imports
instead of midway through them. Bundled dynamic imports resolve as microtasks without yielding to
the event loop, so two loop turns between the first imports let a finished login env start the
daemon process, which then boots on another core while ~0.4 s of runtime modules evaluate.
* fix(orcad): register the daemon disconnect as soon as its spawn starts
An import or init step that throws between the daemon spawn and the old registration point left
no cleanup to disconnect the in-flight daemon. It is now registered right after daemonStartup,
first, so it still runs last.
* fix(clipboard): order text, image and file IPC copies
* fix(clipboard): time out stalled file path resolution in the shared copy queue
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(wsl-mover): say when a refused undo left the host's terminals asleep
* fix(wsl-mover): count only running workspaces as asleep, and flag post-sleep refusals
StopCancelled carries no backgroundTasks, so a cancelled Grok turn settled
done while the shells and subagents it left running kept going. Track
Grok's own task ids per pane and settle the cancelled row on each task's
own end.
* fix(wsl): session history reads never boot a stopped WSL distro (E9)
Transcript and session-history reads touched \\wsl.localhost shares, and ran
`wsl.exe --exec` probes, for any distro the root list named. Those touches are
the OpenCode reader, its setup, HOME and realpath probes, and the session
scanner and search index. Nothing re-checked that the distro still ran. A
`wsl --terminate` mid-scan, or a list taken while the distro was still
stopping, let background work boot it again.
Every UNC transcript touch (stat, readdir, open, access, readFile) and every
scanner `--exec` now asks a settled running read first. That read is no older
than 5 s, because the scanner process hears of no stop by itself. A user act on
the distro still passes.
* fix(wsl): opening a chat starts its stopped distro; scanner hears main's stop notes
A user-opened chat, a history page and an AI Vault resume start the
transcript's stopped WSL distro once, as a user act, so its history
shows; watch and scan reads stay background. Main forwards WSL stop and
wake notes to the session-scanner process so a terminate main saw is
honoured there at once.
* fix(wsl): type the scanner note channel structurally; test the resume wake
Keeps node:child_process out of the new scanner-notes module (the
process-host import check) and adds a test that fails if an AI Vault
resume stops starting the session's WSL distro.
L1: the orcad liveness supervisor's heartbeat/verify reads, its recovery gate, the no-boot
reconnect check and the connector's passive probe now take a settled read of their own distro, so
a list probed while the distro is still stopping cannot admit a holder spawn, identity read or
reconnect that would boot it again.
L2: a Git cancel/timeout's guest kill, the WSL move's process table and each step of the Codex
session bridge chain check the distro first. A confirmed stop counts as the group gone / an empty
table / the end of the chain; an unanswered read proves nothing and spawns nothing.
A slot whose startup qualification marker matches (same artifacts, runtime, libc, kernel and
launch env, last launch reached readiness) now also skips the synchronous node-pty load probe
child (~0.1 s on the WSL lane). The marker is now recorded only when that launch's probe
passed cleanly; schema 2 so earlier markers, which did not promise that, re-qualify once.
Key MermaidViewer's scroll cache with the pane-scoped editorViewStateKey,
matching IpynbViewer, so split panes showing the same .mmd file no longer
overwrite each other's saved diagram position.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Zoom and resize re-renders reuse the same key string instead of rebuilding
and comparing it, which matters for base64-backed images (diff, conflict
review, fallbacks) where the key embeds the whole payload.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Key the colorize effect on joined content and line count instead of the
lines array identity, so re-renders with an equal fresh array (PR comment
code context) no longer re-run monaco.editor.colorize. Check cancellation
before colorizing so excerpts replaced or unmounted while a lazy tokenizer
loads skip the work.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Monaco's CopyPasteController claims paste on the editor container in the
capture phase before Orca's listener, so Ctrl/Cmd+V bypassed the 16 MiB
cap. Listen on the editor shell instead, scoped to this editor's subtree
and text focus, and only reject over-cap pastes; smaller pastes keep
Monaco's atomic native paste. Remove the DOM chunked-paste plumbing
(save deferral, isApplyingLargePasteRef) that could never run.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Prune file/diff read generations only when a tab goes from open to closed,
replacing the append-only known-open-ID set, and fence the restored-owner
migration so a late result cannot discard a newer read's generation.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Upload up to four files concurrently within the existing 2 MiB in-flight budget, sizing each file's window from its size. A per-file failure aborts siblings and surfaces the first error after the SFTP channel closes.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
Jinjing
Chen
OrcaWinandOrcaWin
