Worker-start passed the Orca agent id straight to the shell as the worker terminal command, so `--agent cursor` ran `cursor` — which on Windows resolves to Cursor IDE's cursor.cmd and opened the desktop app, leaving a blank shell that timed out at agent_readiness. The same gap hit every agent whose CLI binary differs from its id (continue/aug/kiro/qwen-code/mistral-vibe/antigravity/trae/mimo-code/hermes/command-code/claude-agent-teams).
Adds TerminalCreateOptions.startupAgent so callers name the agent outright; createTerminal then builds the launch from the TUI agent config (command, agentCmdOverrides, default args/env, preflight trust) instead of sniffing the command string. Also covers repo-less folder workspaces, which previously skipped resolution entirely, and fails loudly instead of spawning a bare shell when an explicit agent cannot resolve.
Fixes#11926
* Honor configured shells during worktree setup
* Align setup launch paths with selected Windows shells
* Carry setup shell selection through deferred launches
* Prove Windows setup shell routing at its real adapters
* Ground remote PowerShell proof in the real writer
* Preserve Git Bash across deferred setup launches
* Harden Windows setup runner shell selection
- Resolve remote PowerShell binary without local pwsh probe: for SSH/remote
Windows worktrees, isPwshAvailable() reflects only the LOCAL host, so an
'auto' implementation could route the remote runner to a pwsh.exe the remote
lacks. Add resolveSetupRunnerShell(..., { probeLocalPwsh: false }) so remote
auto keeps the always-present powershell.exe; explicit pwsh.exe still honored.
- Preserve native exit codes in the PowerShell runner by checking
$LASTEXITCODE before $?, so a failing native command surfaces its real code
instead of a generic exit 1; $? still catches cmdlet soft-failures.
- Write the PowerShell runner with a UTF-8 BOM so Windows PowerShell 5.1 (the
new default powershell.exe) reads it as UTF-8 instead of ANSI, preventing
non-ASCII setup-script corruption.
- Add unit tests for the remote-probe behavior.
* Restore setup-shell scope narrowing over the rebase
The force-pushed rebase dropped five review-fix commits that were already
on this branch; this reapplies their combined effect on top of the new
base and the hardening commit:
- Keep SSH setup shell selection remote-owned (no local terminalWindowsShell
or pwsh routing for remote hosts; supersedes the probeLocalPwsh guard)
- Preserve cmd setup compatibility outside POSIX shells (no .ps1 runner
family, so the BOM/exit-code hardening is no longer applicable)
- Route WSL setup runners from the project runtime
- Avoid blocking PowerShell probes during setup creation
- Correct SSH and WSL background setup fixtures
* Satisfy the changed-code gates for the setup-shell runner
- createWorktreeRunnerScript took 7 positional parameters, tripping the
changed-code max-params gate; move it to a single options object.
- hooks-runner.test.ts deep-equals the createSetupRunnerScript result, so
assert the cmd shell now returned for native Windows worktrees.
* Carry the setup launch shell through observed and issue runners
- buildObservedSetupCommand takes the runner's launch shell so WSL-routed
Windows-drive setup replays use /mnt/c instead of Git Bash /c
- resolveSetupRunnerShell gates the posix runner on the same Git Bash
resolution the PTY uses, so a missing or non-MSYS bash keeps the cmd runner
- issue-command runners carry their launch shell, and the renderer passes it
when building the queued command
- treat a bare `bash` shell setting as POSIX like `bash.exe`
Co-authored-by: Orca <help@stably.ai>
* fix(worktrees): close counsel P1 gaps for Windows setup shells
Route windowless/headless creates through the shell-aware setup runner when a
PTY controller is available, existence-check explicit Git Bash paths before
committing to .sh runners, thread the resolved shell into issue-command
runners, and document the intentional Git Bash interpreter flip with a narrow
scope table.
* Convert setup env to MSYS form and harden the bare cmd runner launch
C3: a Git Bash setup runner now receives ORCA_*/CONDUCTOR_*/GHOSTX_* path
values in /c/... form, matching the runner path and the shell's own HOME/PWD.
C5: extension-less `bash` resolves to Git Bash everywhere, matching how
resolveWindowsShellStartupFamily already classifies it.
C7: runner paths carrying characters that cannot be quoted on a cmd command
line launch through a delayed-expansion PowerShell shim instead, and the batch
runner disables inherited delayed expansion so `!` in setup lines survives.
Co-authored-by: Orca <help@stably.ai>
* docs: note MSYS ORCA_* paths and bare bash Git Bash resolution
Keep the setup-shell release note aligned with C3 env conversion and C5 bare
bash resolution so the published claim matches runtime behavior.
* revert: drop windows-setup-shell doc allowlist and AGENTS link
Keep the counsel P1/P2 product fixes without expanding the docs allowlist
or AGENTS.md guidance surface.
* fix(plugins): contain Parcel unsubscribe rejections under Vitest
Dev plugin watchers fire-and-forget unsubscribe, and in-process Parcel
can reject when temp watch roots are already deleted. Catch those
rejections so they cannot fail the suite as unhandled errors.
* fix(plugins): keep in-process unsubscribe rejection surface
Swallowing Parcel unsubscribe errors broke mocked unsubscribe tests
that return non-Promises and expect rejections. Contain failures only
in PluginDevWatcher fire-and-forget paths.
---------
Co-authored-by: OrcaWin <alpha-eng@stably.ai>
Co-authored-by: Brennan Benson <79079362+brennanb2025@users.noreply.github.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: Orca <help@stably.ai>
* fix(gitlab): count diff lines whose content begins with -- or ++
countDiffLines skipped every line starting with ---/+++ as a file header,
but a removed line whose original text began with -- (SQL/Lua/Haskell
`-- comment`) becomes a diff line `---<content>`, colliding with the
`--- a/file` header — so its deletion was silently dropped from the
+N/-N shown in the GitLab MR dialog. Same collision for an added line
whose content began with ++ (+++ flag).
Track hunk state: ---/+++ are file headers only before the first @@;
inside a hunk every +/- is content, matching the unified-diff rule git
itself uses to disambiguate headers from content.
Co-authored-by: Claude <noreply@anthropic.com>
* fix(gitlab): validate countDiffLines with actual diff format
GitLab's /diffs endpoint returns json_safe_diff starting at @@ without
file headers. Add comprehensive test coverage validating the collision
fix correctly handles this format: content lines beginning with -- or ++
are counted as additions/deletions.
Tests cover binary files, empty diffs, no-newline markers, and content
beginning with @@ or C-style ++. Clarify function contract: requires
hunk headers to distinguish headers from content lines.
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(workspace-cleanup): read reflog timestamps to avoid git maintenance
Workspace activity detection now reads the reflog to find the newest HEAD movement,
avoiding false activity signals from `git gc` and `git status` restamping logs/HEAD.
Extraction of git mtime probes to exclude files that maintenance rewrites (gitdir,
index, logs/HEAD), and instead read commit markers (COMMIT_EDITMSG, ORIG_HEAD) and
reflog entry timestamps. Expands the scan with a renderer-side activity estimate to
reconcile against the Resource Manager button's fast count. Adds deletion phase
tracking (queued vs deleting) and a mismatch notice when the two counts diverge.
* fix(workspace-cleanup): parse reflog timestamps with fewer digits and im
- Regex now accepts 1-11 digit timestamps (was 9-11); trailing timezone anchor makes digit-count floor unnecessary
- Add `removalInFlight` state to prevent duplicate removal batches; UI checks this flag alongside `removalProgress`
- Filter scan errors by selected repos; only show estimate-mismatch notice when scan is complete and error-free
- Mark candidate rows as non-selectable while deleting, even if `removing` flag is omitted
The leading-icon gutter used self-start with a hand-tuned pt-0.5 nudge,
leaving status dots ~2px and lucide icons ~1px above the 20px title line
box. Give the gutter h-5 to match the line box so icons center on the
first line for single- and two-line rows alike.
* feat(voice): allow selecting a microphone for dictation
Persist a preferred audioinput device in Voice settings and pass it into
getUserMedia, falling back to the system default when the device is gone.
* fix(voice): resolve mic preference by label and detect mid-capture loss
Drop Chromium's 'default'/'communications' aliases from the picker — pinning
one behaved exactly like system default and silently defeated the setting.
Resolve a stored preference against the live device list before capturing:
a unique label match heals an id that Chromium re-salted, a known-missing
device skips the doomed getUserMedia attempt that clipped the first words,
and an unreadable list no longer reads as "unplugged".
Surface the input ending mid-dictation instead of feeding silent zeros, add
a permission affordance so the picker is not empty before mic access, and
toast the fallback once per preference rather than once per utterance.
Co-authored-by: Orca <help@stably.ai>
* add e2e tests
* add e2e tests
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
Co-authored-by: Orca <help@stably.ai>
* fix(browser): keep browser tabs rendering across worktree switches (STA-3228)
Switching away from a worktree that had a targeted background mount
unmounted BrowserPaneOverlayLayer, pulling the persistent <webview>
slots out of the DOM and killing their guests; the stale viewport cache
then kept rendering into the removed subtree, so the tab stayed blank
forever and reload threw. Keep the overlay mounted for hidden worktrees
(slots park their panes, so this stays cheap) and rebuild cached
viewports whose slot root remounted.
* fix(browser): retain live overlay slots without background churn
* fix(browser): latch overlay retention after commit
* fix(tasks): cap advertised GitHub pages at the search result window
GitHub's Search API rejects requests past its first-1000-results window
with HTTP 422, but totalPages was derived from the raw total_count, so
the pagination bar advertised pages that could never load and clicks on
them silently did nothing (#11485).
Cap per-repo advertised pages at floor(1000 / perRepoLimit), and when a
page load comes back empty, say so with a toast instead of ignoring the
click — clamping the advertised count only when no fetch threw, so
transient failures don't shrink the bar.
* fix(tasks): key pagination resets on repo selection, not array identity
The repos store installs a fresh array on every repos:changed event, so
the pagination-reset effect fired on background refreshes and bumped the
request generation, silently discarding any in-flight page navigation —
clicking an unloaded page did nothing whenever a repo refresh landed
during the fetch. Key the effect on the stable selection string instead.
* fix(tasks): distinguish end-of-data, window 422s, and failures on empty pages
Adversarial-review round 1 rework:
- fetchWorkItemsNextPage now returns issue-side envelope error types — the
channel the search-window 422 actually travels on (failedCount only
counts thrown repo calls).
- resolveEmptyPageOutcome (unit-tested) maps an empty page to
window-unreachable (clamp + toast), load-failed (toast only; may be
transient), or end-of-data (silently withdraw the speculative page the
count-fallback advertises).
- The work-items fetch effect is keyed on selectedReposKey too — its
unconditional page reset re-fired on every repos:changed array identity,
bouncing the user to page 1 mid-click. The key now includes the resolved
GitHub source context so identity changes still re-dispatch.
- Toasts carry stable ids so repeats replace instead of stack.
- Cap comment documents the conservative PR-scope tail loss; cap tests
pinned at shipped (36 → 27) and dividing (25 → 40) limits.
* fix(tasks): withdraw the speculative page when the failed count is zero
countedTotalPages of 0 comes from a swallowed count failure and routes
totalPages through the fallback, so the clamp must replace it like null.
* fix(tasks): tighten empty-page outcomes after round-2 review
- en.json's loadPageUnreachable carried the pre-reword text, and the
catalog beats the inline default — the two toasts were identical.
- end-of-data clamps only while the count is unknown/failed: the PR list
path swallows its own failures into clean-empty results, and clamping a
real count silently hid healthy pages (worse than the pre-fix no-op).
- A window 422 no longer clamps when a sibling repo's fetch threw.
- The generation effect mirrors every fetch-effect dep that resets page
state, so manual refresh/source switches invalidate in-flight clicks.
- selectedReposKey extracted as buildSelectedReposKey with stability
tests; envelope error types wire-tested through the store.
* fix(tasks): clamp against the committed count, not the click-time closure
Round-3 review: the count promise routinely resolves between click and
response, so deciding the end-of-data clamp from the closure value let a
stale null overwrite a real count. applyEmptyPageClamp now runs inside
the functional updater against the committed value, never raises an
earlier clamp, and a window 422 coinciding with a thrown sibling repo
resolves as load-failed so the toast and the clamp always agree.
* fix(tasks): only an all-window-422 empty page may clamp; harden count merges
Round-4 review: a sibling repo's envelope 403/404 arrives with
failedCount still 0, so the window branch now requires every error to be
the window 422 (non-window validation errors are demoted at the store);
the count resolution mins against an applied clamp instead of
re-advertising withdrawn pages; the generation effect mirrors
taskResumeApplied so its doc claim holds.
* fix(tasks): split the proven window limit from the count slot
Round-5 review: min-ing the count against an applied clamp pinned a
SPECULATIVE end-of-data withdrawal that raced ahead of the count,
permanently collapsing the bar for the generation. Proven window-422
limits now live in provenPageLimit (set once, only lowered, reset per
generation); the count overwrites its own slot unconditionally; and
deriveAdvertisedTotalPages (unit-tested for both arrival orders) caps
the count-or-fallback estimate with the proven limit, floored at the
loaded pages.
* fix(tasks): surface PR-side list failures so they can't read as end-of-data
Round-6 review: PartialWorkItemsResult had no PR error slot, so a
swallowed gh pr list failure reached the renderer as a clean empty page
— and with the count blocked (0) the speculative withdrawal deleted the
pagination bar with no toast and no recovery (a regression vs main's
silent no-op). PR-side errors now ride the envelope (errors.prs),
demoted so they can never join the issue-only window-422 signal;
errorTypes replaces issueErrorTypes; an empty page that a real count
said should exist now toasts instead of looking dead.
* test(tasks): cover the PR-error envelope end-to-end; neutral no-more-results toast
Round-7 review: the two literal gh-utils mocks lacked classifyListPrsError
(a PR-side rejection in those suites would TypeError instead of assert),
and the producer half of the errors.prs contract had no main-side test —
added both, plus a classifier contract test pinning the search-window
phrase the renderer keys on. The refused-clamp toast now reads the
committed count via a synchronous ref mirror instead of the click-time
closure, and says 'No more results' — nothing failed on that branch.
Both toast keys plus the new one are translated in es/ja/ko/zh.
* fix(tasks): preserve final reachable GitHub search page
* Extract GitHub search result window error pattern to shared constant
Extract the 1000-result window detection pattern to a single source of truth so
the classifier and consumer stay synchronized. The pattern is the only signal
separating a permanently unreachable page from a transient validation failure,
so drift or trimming silently demotes window 422s to generic failures and stops
capping the advertised page count (#11485).
---------
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* Allow clearing all agents from AI Vault session history filter
Add "Select all" / "Clear" buttons so users can quickly isolate one agent without unchecking each box individually. Previously, at least one agent had to remain enabled; now users can filter to zero agents and re-enable selectively.
* Address PR #12128 review feedback
- Make Select all / Clear real DropdownMenuItems so Radix roving focus reaches them by keyboard.
- Rename the zero-agent empty state to a neutral "No agents selected" now that zero agents is a valid filter.
- Use 모두 해제 for the Korean Clear label instead of 지우기 (erase).
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
Both hosts carried token-identical copies of the GitHub work-item mutation
wrappers, PR diff mapping, presentation formatters, and four components.
Move them into src/renderer/src/components/github/ so there is one source.
Behavior-neutral: getStateTone, WorkItemStateBadge, PRReviewersPanel,
PRActionsPanel, CommentReplyForm, and the per-host work-item/PR-file caches
stay in place because they genuinely differ between the two hosts.
CommentCodeContext takes loadPRFileContents as a prop so each host keeps its
own private file-contents cache.
Also folds github-issue-comment-helpers.ts into github-user-avatar.tsx and
retargets the textual boundary tests at the new modules.
* refactor(usage): share the session/daily fold between Codex and OpenCode
The Codex and OpenCode scanners each carried their own byte-identical copy of
the ~325-line aggregation pipeline (createEmptySession, the three breakdown
folds, finalizeSessions, mergeSessions, mergeDailyAggregates). Two copies means
a token-accounting fix — a bucket that double-counts, a merge that drops a
breakdown row — lands in one provider and silently not the other. The copies had
already started to drift in comments only; the next drift would have been in
arithmetic.
The providers differ in exactly one dimension: the extra metric folded alongside
the token counters (Codex `hasInferredPricing`, OpenCode `estimatedCostUsd`).
That is now injected as an empty/fromEvent/fold triple, so the shared code stays
generic without collapsing the two record schemas into a nullable union. The
clone strategy stays per-provider (`cloneSessionForMerge` vs `structuredClone`)
rather than being unified on the assumption that the difference is accidental.
`usage-provider-contract.ts` is the seam a plugin-contributed usage source will
implement. It is deliberately generic over each provider's record types: Claude
bills per turn while Codex/OpenCode bill per event, and `cachedInput` is a subset
of `input` for the latter but a peer bucket for Claude, so a single normalized
record would push nullable handling onto every consumer.
No behavior change. Emitted objects are byte-identical, including key insertion
order — verified by diffing JSON.stringify of the scan output before and after
across mixed models, mixed locations, an inferred-pricing flip, and null vs
non-null cost. Persisted field names and schemaVersion are untouched, so caches
do not invalidate.
* refactor(usage): make the provider contract load-bearing and dedupe worktree refs
Follow-up to the aggregation extraction, addressing three review points.
`UsageProvider`/`UsageScanResult` were declaration-only, which is the same
speculative-interface problem #12077 just deleted 8,900 lines of. They are now
implemented by both real providers via `satisfies`, so the seam is typechecked
against actual scan functions rather than asserted. The blocker was that codex
returns `processedFiles` and opencode returns `processedDatabases`; rather than
rename persisted-adjacent fields, the source key is a type parameter, so each
provider keeps its own on-disk name and the contract still binds. Verified the
constraint bites: swapping the key to 'processedSources' fails typecheck.
`schemaVersion` is part of provider identity in the contract, so each provider's
SCHEMA_VERSION constant (with its cache-invalidation rationale) moves into the
provider module and the store imports it. Values are unchanged (codex 5,
opencode 2) and the stores compare them exactly as before, so no cache
invalidates. This also keeps store -> provider -> scanner acyclic.
`UsageWorktreeRef` collided with the existing export in usage-worktree-metadata
(3 fields, no repoId). Two different exported types under one name in src/main
is worse than the duplication being removed, so the scan-input type is now
`UsageScanWorktreeRef`; usage-worktree-metadata is untouched.
`createWorktreeRefs` was triplicated. Codex, OpenCode, and Claude copies are
byte-identical apart from the return type name (verified by diff), and all three
ref types have the same four fields, so one shared copy replaces all three. This
is the only change to claude-usage/.
No behavior change: same functions, same arguments, same call order. The store
tests' `./scanner` mock still intercepts scanning because the provider captures
the mocked binding; their now-inert `createWorktreeRefs` mock key is dropped so
it does not read as still mocking something.
findRemoteForUrlSsh, ensureUniqueRemoteNameSsh and
configureCreatedWorktreePushTargetSsh were byte-identical to
findRemoteForUrl, ensureUniqueRemoteName and
configureCreatedWorktreePushTargetWithExec in worktree-push-target-setup.ts,
differing only in calling provider.exec instead of the injected execGit.
528a887ab5 extracted the shared module out of worktree-remote.ts and left
the SSH twin behind, so this is an unfinished extraction rather than a
deliberate split.
Feed provider.exec through the existing GitRemoteExec seam instead — the
same adapter cleanupUnusedWorktreePushTargetRemoteSsh already uses. Drops
the now-unused parseGitHubOwnerRepo import.
prepareWorktreePushTargetSsh is intentionally left alone: it uses
provider.fetchRemoteTrackingRef (which forces --no-tags and pre-validates
refs) rather than a raw fetch refspec, and carries its own check-ref-format
preamble, so it is a real behavior difference and not a clone.
No behavior change: the substituted bodies are byte-identical.
* feat(updater): add an adhoc release channel for branch builds
Hourly covers main. This covers everything that is not main yet: a
dispatchable macOS build of an unlanded branch, published to
stablyai/orca-adhoc, so the team can run an experimental feature for a
few days instead of reasoning about it from a diff.
Adhoc sits at the bottom of the version order — 'adhoc' < 'hourly' <
'rc' < stable — so no routine check can walk anyone onto somebody's
branch; only an explicit pinned jump reaches one. It gets its own repo
rather than sharing orca-hourly's, because a branch build must not
appear in the list a developer riding main is looking at.
Signed and notarized exactly like hourly, for the same reason: macOS
anchors a notarized app's TCC grants on identifier + team, so an
unnotarized build reads as a new client and silently loses file access
under Documents/Desktop/Downloads.
Tags stamp to the second rather than the minute. Hourly runs under a
concurrency group and cannot overlap itself; adhoc builds are dispatched
on demand, so two people cutting from different branches inside one
minute is ordinary — and a minute-resolution tag would collide and fail
the second build after its whole pack-and-notarize run.
Channel-specific behaviour now derives from one DEDICATED_REPO_CHANNELS
list: repo mapping, macOS-only support, and UpdateSource. The RPC schema
that validates releaseChannelOverride was a hand-copied enum missing the
new channel, which would have rejected the override on its way to the
main process; it reads the predicate now.
* fix(updater): merge the duplicated shared/types import
Co-authored-by: Orca <help@stably.ai>
* fix(ci): default the adhoc build ref to the dispatch branch
The Actions UI puts its own "Use workflow from" branch picker directly
above the ref field, and picking a branch there is what most people read
as "build this". Making the field optional means the obvious action is
also the correct one; naming a branch explicitly still wins, so main's
copy of the workflow runs rather than a stale one on an old branch.
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
mobile/src/constants/marine-creatures.ts was a hand-maintained copy of
src/shared/marine-creatures.ts, identical except for a comment header. The
copy existed because Metro only watched mobile/ and could not resolve
repo-root modules; mobile/metro.config.js:11 added src/shared to
watchFolders five weeks later, and ~195 mobile files already import from
src/shared. The renderer collapsed its copy to a re-export at the same time;
mobile was the leftover.
Point the one consumer at the shared corpus and delete the mirror, the
bespoke regex-scraping parity test that policed it, and the now-stale
max-lines baseline entry.
No behavior change: same exported symbol, byte-identical name list.
Recovery and broker open now send the optional reconnect hint so the
director admits already-assigned hosts through its bounded fast lane
(orca-cloud#212) instead of the placement queue that starved session
recovery during the 2026-08 incident. A rolled-back director that
rejects the hinted field gets one unhinted retry.
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* fix(editor): stop filing crash reports for expected lazy-chunk swaps
RichMarkdownErrorBoundary reported every caught error as a react-error-boundary
crash, including the LazyChunkLoadError sentinel that lazy-with-retry throws
after it has already exhausted its retries and its one guarded reload. That
sentinel means "the chunk hash changed under a running window" (an app update),
which is deliberate graceful degradation, not a crash.
RecoverableRenderErrorBoundary already skips reporting it (#6206); this boundary
was never updated. Crash b860def2 is exactly that path: a lazy_chunk_reload
breadcrumb ("Unexpected token ':'") fires first, then the post-reload attempt
surfaces LazyChunkLoadError and files a report.
The fallback UI is unchanged, so the pane stays usable and offers retry.
* fix(editor): prove the lazy-chunk reload landed before suppressing crash reports
- lazy-with-retry: reload guard stores the requesting document's identity, so a
vetoed reload() no longer reads as "recovery ran" (crash b860def2)
- lazy-with-retry: bound the post-reload suspension so a vetoed navigation
surfaces the real error instead of hanging the pane on a spinner
- RichMarkdownErrorBoundary: contain the LazyChunkLoadError sentinel without a
crash report, but record a lazy_chunk_boundary_degraded breadcrumb
- EditorContent: name the rich markdown chunk at the lazy call site
Co-authored-by: Orca <help@stably.ai>
* fix(editor): route lazy-chunk recovery reload through the intentional-restart path
Crash b860def2's recovery reload was requested and never landed: Terminal's
beforeunload handler preventDefault()s while any editor tab is dirty and Electron
cancels the navigation with no dialog, so chunk recovery could never run in the
common case. Take the updater's path instead — hot-exit backup, one synchronous
session checkpoint, restart latch — then reload.
- Reject on ORCA_RENDERER_UNLOAD_PREVENTED_EVENT instead of a blind, never-cleared
10s timer; keep the timer only as a backstop.
- Record a lazy_chunk_reload_vetoed breadcrumb in the same tick as the report it
now files, so the 30-entry ring cannot evict the evidence.
- Drop this document's own stale guard after a refused reload (capped in memory)
so saving the blocking tab does not forfeit recovery for the session.
- Carry reloadKey on LazyChunkLoadError and the degraded breadcrumb.
- Move renderer-restart-preparation to src/shared: it is now a renderer/preload
contract, and the composite web project cannot import preload runtime code.
Co-authored-by: Orca <help@stably.ai>
* fix(editor): clean up failed lazy chunk reload requests
* test(preload): exercise restart IPC registrations
---------
Co-authored-by: Orca <help@stably.ai>
* fix(i18n): localize status labels in settings and stats
Status pills and summaries in Settings and Stats were built from bare string
literals inside local helper functions, so they stayed English even when a
language pack was active. Neighbouring copy in the same components already went
through `translate()`, which made the panes look half-translated:
"Универсальный доступ GRANTED", "GitHub ... Connected", Russian orchestration
card titles above English summaries.
The coverage audit does not catch this: it inspects JSX attributes and object
properties, not values returned by helpers, so `verify:localization-coverage`
stays green while the strings ship untranslated.
Wrapped the remaining user-facing strings in `translate()` and let
`sync:localization-catalog` add the 31 new keys. `computerUseSummary.*` already
existed in `en.json` with identical copy but was never wired up, so those keys
are now connected instead of duplicated.
Moved the permission status helpers into `developer-permission-status.ts` to
keep `DeveloperPermissionsPane.tsx` under the 400-line gate. Agent prompts in
`orchestration-usage-examples.ts` are left in English on purpose: they are
payload sent to the agent, not UI copy.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(i18n): use plural-aware message keys for settings labels
Replace template-based pluralization (using {{value1}} for "s") with
proper i18n plural forms following _one/_other suffixes. This enables
correct pluralization across languages with distinct rules.
Also extract duplicate integration status label translation logic
into a helper function.
* CodeRabbit's nitpick: the placeholder/plural assertions in settings-status-label-localization.test.ts only read en.json, so a translated catalog could ship a stale {{value1}} or a half-translated plural family undetected.
Added a second describe block over the four shipped locale catalogs (es, ja, ko, zh), keeping the exact English-value assertions untouched and separate:
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jinjing <6427696+AmethystLiang@users.noreply.github.com>
* fix(speech): coalesce model download progress churn
A model download emits one state update per HTTP chunk (thousands for a
500MB model). Each one costs the renderer an IPC round-trip and a forced
re-render of the speech-model menu, which stays open by design while a
download runs — the Radix portal/Presence tree all four page.settings
React #185 reports crashed inside.
Emit at whole-percent granularity (what the UI renders) and keep the
modelStates array identity stable when a refresh changed nothing, so a
no-op refresh no longer forces a commit.
Adds a speech_model_state_churn breadcrumb, registered in both
coalescing sets, because no bundle in the cluster carried any speech
telemetry to confirm a download was in flight.
* fix(speech): quantise polled model state so download progress stops churning the renderer
Adversarial round 1 found the renderer-side stabilisation was inert during a real
download. The progress fan-out already coalesces to whole percent, but the renderer
discards the event payload and re-polls getModelStates, and getModelState returned
the cached downloading state verbatim - raw sub-percent progress. So every chunk
produced a fresh object, resolveModelStates never matched, and all the real benefit
came from the main-side coalescing alone.
Quantise the polled reply to the precision the UI already renders
(Math.round(progress * 100)), keeping the stored cache exact.
Also from round 1:
- the whole-status dedup swallowed downloadModel's already-downloaded branch, whose
lone 'ready' is the only notification the requesting window ever gets - a dead
click and a permanently stale pane with two windows open. Gate it on
downloading -> downloading so every one-shot transition stays unconditional.
- rename the storm test off '.react185.': it counts renders and asserts nothing
about #185, and #185 could not be reproduced at IPC-realistic pacing.
* fix(speech): stop the churn breadcrumb firing on every healthy download
Adversarial round 2. CHURN_REFRESH_THRESHOLD was 60 per 5s window, but main
clamps download progress at 0.9 and emits on whole-percent change, so one
healthy download is capped at ~91 refreshes — and a 56MB model on a fast link
lands all of them inside a single window. The breadcrumb fired on every normal
download and burned a slot in the 30-entry ring it was coalesced into to
protect. Raise it to 250 so it only fires on the per-chunk shape it was added
to detect; noOpRefreshes in the payload still separates the two causes.
Round 2 also found that every assertion round 1 added was one-sided
(toBeLessThanOrEqual), so each of the three core behaviours could be mutated
into "do nothing" with the whole suite still green:
- suppress every downloading -> downloading event: progress bar frozen at 0%
- toWholePercentState returning 0: every poll reports 0%
- resolveModelStates never adopting a same-length change: the Voice pane never
updates and a finished model never shows as ready
The suite measured that the fix reduces work, never that it still does the
work. Convert the two ceilings to exact series, assert the storm test's render
floor as well as its ceiling, and add dictation-model-state-stabilisation.test.ts
covering adoption per changed field, a full whole-percent download, and both
sides of the churn threshold.
Ruled out and deliberately not fixed: round 1's request-sequencing MEDIUM on
refreshModelStates. 50 concurrent getModelStates() settle strictly FIFO over
ipcRenderer.invoke at constant microtask depth, and migrationReady is assigned
once in the constructor, so a monotonic request id would be dead code.
* test(crash-reporting): cover speech churn breadcrumb name-coalescing
The churn breadcrumb's registration in COALESCED_RENDERER_BREADCRUMB_NAMES and
NAME_ONLY_COALESCED_BREADCRUMB_NAMES had no test: the storm test only asserted
the constant equals its own literal. Removing either registration kept the whole
suite green.
It carries no message field, so without the name-only entry
rendererBreadcrumbCoalesceKey returns undefined and every firing takes its own
ring slot — the eviction this breadcrumb exists to avoid.
* test(dictation): pin the churn threshold as a rate, not a lifetime total
Both existing churn tests stopped at exactly 250 refreshes in one window, so two
mutations survived: deleting the per-window reset (the counter degrades into a
session total, and three healthy downloads at 91 refreshes each cry wolf), and
=== to >= (fires on every refresh past the threshold, flooding the ring).
Pins Date.now rather than using fake timers so the window rule is measured, not
the machine.
* test(dictation): pin the churn clock and the window's lower bound
The two 250-refresh churn tests measured real elapsed time against a 5s
window, so a loaded runner that took longer than one window to run the
loop would roll the window and go red. Pin Date.now in both.
Pinning alone leaves CHURN_WINDOW_MS unpinned below: shrinking it 5_000 ->
50 kept all 13 tests green, yet a 50ms window can never accumulate 250
refreshes and the detector would be dead. Add a storm spread across most
of one window so the constant has to be wide enough to hold a sustained
storm, not just an instantaneous burst.
Co-authored-by: Orca <help@stably.ai>
* refactor(speech): narrow progress churn fix
---------
Co-authored-by: Orca <help@stably.ai>
* fix(dev): split the confirmation dialog so Fast Refresh can accept it
`confirmation-dialog.tsx` exported both `ConfirmationDialogProvider` and
`useConfirmationDialog`, so React Fast Refresh could never treat it as a
boundary and Vite applied every edit to it in two passes under two `?t=`
stamps. When a second file in the same subtree changed in one watcher batch,
`createContext` ran twice and the provider published one context object while
the consumer read the other — `useContext` returned null and the hook threw.
Two field crash reports hit this at `ChecksPanel`, both dev-server sessions.
The context and hook move to a new component-free `confirmation-dialog-context.ts`;
`confirmation-dialog.tsx` keeps the provider and now exports only a component, so
the refresh runtime accepts it. Not one line of the provider body changes — the 16
hook importers just point at the new module, `vi.mock` targets included, and
`App.tsx` is untouched.
* test(dev): pin the confirmation dialog Fast Refresh boundary
The split that fixed the context-identity crash had no test behind it: no
test imported ConfirmationDialogProvider, and the six vi.mock call sites
replace the hook module wholesale, so they pass just as well with the
provider and hook back in one file. Assert the module shapes the refresh
transform actually keys on -- the context module registers no component,
so it never gets an HMR footer to invalidate through.
Co-authored-by: Orca <help@stably.ai>
* test(dev): assert the refresh boundary on the module namespace
The source-regex guard did not guard. Its patterns match only declaration
forms, so `export { useConfirmationDialog } from './confirmation-dialog-context'`
in the provider module -- which restores the crash, verified in a browser --
passed it 3/3. It also failed on a comment that merely contained the word
createContext, and would fail on React 19's `<Ctx value={...}>` shorthand.
Assert on the module namespace object instead, using react-refresh's own
component criterion, so re-exports and default exports are visible. The third
test renders the provider and resolves the hook through it, which is a real
behavioural check rather than a shape one.
Co-authored-by: Orca <help@stably.ai>
* test(dev): classify boundary exports with the refresh runtime's own predicate
The hand-rolled `^[A-Z]` name check called `export class Foo {}` a component;
the runtime rejects any class whose prototype carries extra members, so that
shape restored the two-pass split undetected. Use react-refresh's exported
`isLikelyComponentType` and mirror `isCompoundComponent` instead of a third
approximation. react-refresh was already resolvable only via shamefully-hoist,
so it is now an explicit devDependency.
* test(dev): tighten confirmation dialog boundary guard
---------
Co-authored-by: Orca <help@stably.ai>
* fix(terminal): trust Pi CSI-u Shift+Enter on Windows (#9703)
Pi enables the Kitty keyboard protocol at startup and decodes CSI-u, but
TUI_AGENT_CONFIG['pi'] never set windowsShiftEnterEncoding, so on Windows
Pi could only get CSI-u via the flaky live-KKP-flag path
(isKittyKeyboardActivePane). After a tool ran a subprocess that emitted a
reset sequence, the KKP flags dropped to 0, Orca sent Esc+CR, and Pi read
it as plain Enter -> submit. It recovered on the next pane refocus.
Set windowsShiftEnterEncoding: 'csi-u' for pi, mirroring the Droid fix
(#7668), so the trusted CSI-u route covers Pi reliably independent of
KKP-flag churn from tool subprocesses.
* fix(terminal): complete Pi Windows CSI-u trust lifecycle
* test(terminal): name foreground retry timing
* test(git): accept bounded SSH remote probes
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
src/relay/relay-frame-decoder.ts and src/main/ssh/relay-frame-decoder.ts
were 264 identical lines apart from one default: the relay logs decode
faults to stderr when no handler is supplied, the SSH side stays silent.
Two copies of framing logic is exactly where a wire-format fix lands in one
and not the other.
The decoder's contract and buffer already live in src/shared, so the class
joins them there. The relay keeps a thin subclass that supplies its stderr
default, preserving behaviour for the call sites that omit onError. The
SSH copy is deleted and relay-protocol.ts points at shared directly.
Verified: pnpm typecheck, 102 tests across the 9 framing/backpressure/
handshake suites, and `pnpm build:relay` for all six platform targets plus
the WSL hook relay — the standalone bundle has no new dependencies.
#12065 started passing an AbortSignal as a third argument to provider.exec
in readRemoteUrl, and updated the Gitea, GitHub and GitLab assertions to
match. Azure DevOps and Bitbucket were missed, so main is red on three
test files.
Same expect.any(AbortSignal) shape #12065 used for the other providers.
Co-authored-by: Orca <help@stably.ai>
- RelayControlOrigin.activate rejects controls whose socket closed before
activation (hello-ack and close in one ws parser turn previously published
a dead control with no recovery path)
- RelayControlClient gains a 75s inbound-silence watchdog mirroring the
relay's ping contract, so dead or server-side-unindexed sockets terminate
and trigger origin recovery
- RelayAuthCoordinator only republishes 'registered' when the owned broker
proves a live control, and logs reconcile failures instead of swallowing
them; the origin pool logs recovery-attempt failures
- Host-proof validation reports the failing check by name (never values),
keeping main's 30s skew bounds
- isLive() plumbed client -> origin -> pool -> broker
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
* chore(dead-code): drop 2k lines of unreachable exports and orphan modules
Ran knip across every build entry (main, preload, renderer, popout, web,
cli, relay, workers, forked sidecars, config scripts) and removed what no
entry graph can reach.
- 11 orphan modules nothing imported, plus one test that only covered them
- 159 unused exports/types, with their now-dead helpers, imports and tests
Each candidate was verified against dynamic references before deletion.
42 knip hits were false positives and are kept: shared modules consumed by
the mobile/ workspace, the src/shared/plugins/** public API, vendored
shadcn primitives, and relay wire-protocol constants held for compatibility.
Adds knip.json + `pnpm audit:dead-code` so this stays measurable.
Verified: pnpm typecheck, pnpm lint, and 2081 tests across the 73 affected
test files all pass.
* chore(dead-code): move knip config under config/
Root-level additions are blocked by the root directory guard.
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
* fix(P1-D): coalesce remote-ref probes, TTL negatives, and bound unsettled keys
Keep forge resolution from stampeding git under worktree fan-out, let
remotes added mid-session be discovered without a restart, and refuse
pathological new-branch waves once the unsettled map is full.
* fix(P1-D): stop abandoned probes publishing, and split capacity refusals
A coalesced probe abandoned as stale kept running and still wrote its answer
to the cache, so a late permanent miss could land over the successor's fresher
one. Probes now publish only while they still own the in-flight key.
The hosted-review capacity refusal told brand-new branches that an earlier
attempt of their own never answered when the refusal was really the unsettled
map or the process-wide detached cap; each cap now says what it is.
Also caches stable "no such remote" SSH misses under the negative TTL instead
of re-spawning the probe on every poll.
Co-authored-by: Orca <help@stably.ai>
* Bound SSH remote URL probe with deadline to prevent hangs
The SSH branch of remote URL probes was unbounded — the relay's bounds
are per-phase and reset on every frame, so a relay dribbling output would
outlive them. Pass AbortSignal.timeout to the SSH provider's exec call to
enforce the same 30s deadline as local probes.
Treat AbortError as a transient probe error: it signals unavailable
infrastructure (deadline or cancellation), not a negative answer about
the remote.
---------
Co-authored-by: Orca <help@stably.ai>
* fix(workspaces): forget deleted remote mirrors
* fix(workspaces): tighten orphan cleanup guards
* fix(workspaces): avoid duplicate remote teardown after delete
* fix(workspaces): prevent orphaned filesystem auth on removal
When a worktree is deleted, especially from remote hosts, the filesystem
authorization cache was not being invalidated, leaving the path accessible
even though the workspace was gone. Use persisted host ownership to scope
cleanup to the correct partition and invalidate the auth cache when removing
a workspace to prevent orphaned authorization in host-partitioned scenarios.
* Fix orphaned worktree cleanup to trust persisted ownership and clean all
When a remote worktree or project is deleted, the local metadata cleanup must work even when the owning repo can no longer be resolved. The removal was incorrectly trusting a caller's potentially-stale hostId over the authoritative metadata, causing:
- SSH workspaces to be cleaned from only the local partition, stranding the remote partition with an un-bumped topology fence
- Sibling worktrees of the same repo to get rebased and lose unsaved tabs
- PTYs in orphaned workspaces to never stop when the selector can't resolve
- File watchers to keep firing events indefinitely
Now the cleanup trusts the persisted owner hostId, cleans all affected session partitions where tabs might live, intelligently gates topology fence bumps to avoid rebasing siblings, and passes the exact worktreeId to PTY sweeps that can't resolve the selector.
* Pass removal host ID to fix teardown of ownerless remote worktrees
When deleting an ownerless remote worktree, args.hostId may be absent.
Without an explicit host ID, the session teardown would incorrectly clear
the local session instead of the remote. Derive removalHostId from the
repo (the canonical owner) and pass it to every removeWorktreeMetadataAndTransientState
call to ensure the correct session is torn down.
* Scope worktree teardown to the owning host connection
- Orphaned SSH worktrees now sweep through the host's PTY provider instead of only the local one, so remote terminals die when the repo is gone
- Terminal ownership is scoped by resolved connection/runtime environment, preventing a same-id workspace on another host from being swept
- Persisted ownership beats stale live routing for in-flight keys and topology fences
- Renderer fails closed and never forgets a row whose removal route turns ambiguous mid-flight
* Fix worktree removal to scope session cleanup to the owning host
When a worktree is removed, its metadata purge must resolve the same owner
as the teardown sweep, or SSH/runtime partitions keep workspace state
forever. Additionally, materializing never-persisted host partitions
during removal can rebase sibling worktrees. Scope cleanup to owning host,
skip unwritten partitions, and detect transport-wrapped error codes that
Electron IPC re-wraps and strips causes from.
* Fix worktree removal to scope session cleanup to owning partition
- Only the owning partition may fence on emptiness; spill partitions
that never held the worktree must not claim repo authority to prevent
data loss when the renderer owns tabs elsewhere
- Tighten error code detection to require message boundaries (": " or
newline) instead of matching trailing tokens, preventing false
positives from triggering the destructive forget-local fallback
---------
Co-authored-by: gatsby74 <166927047+gatsby74@users.noreply.github.com>
* fix(P1-A): persist SSH consumer recovery without a sync store flush
rememberPtyConsumerRecovery ran on the live establish/reconnect path and
called flushOrThrow -> writeToDiskSync, parking the Electron main thread on
the profile-directory write. On a stalled or slow profile mount that freezes
the whole app during SSH recovery and reconnect.
Add Store.flushAsync(): same debounce-cancel and write serialization as
flushOrThrow, but awaits writeToDiskAsync instead of blocking. The consumer
recovery upsert/remove pair is now async and awaits it, and the SSH callers
await through to establish()/reconnect() so ownership is still durable before
relay setup continues. In-memory state still mutates synchronously (before the
first await), so no caller can observe a torn record and dispose() stays
synchronous.
* fix(P1-A): detach the SSH session when a connect attempt fails
Both failure exits in doConnect dropped the session from activeSessions
without calling detach(). claimSshPtyConsumerRecovery only reuses an existing
in-memory entry when detached === true, so the next connect attempt fell
through to minting a fresh clientInstanceId, discarding the remembered owner
lease and its resume identity.
Route both exits through abandonFailedSshSession(), which detaches (keeping
PTY ownership, unlike dispose()) before removing the session, and tolerates a
teardown throw so it can't mask the connect error being rethrown.
* fix(P1-A): await async lease persistence in SSH relay teardown
Failed connect attempts now wait for 'detached' leases to persist before
throwing, preventing reconnects from claiming them before cleanup completes.
Detach and dispose operations are now async and await store durability.
* fix(ssh): make session detach lease writes retryable on failure
Separate in-memory detach (identity recovery, provider cleanup) from lease
write persistence so rejected writes can be re-issued without re-running
provider teardown or re-minting the session identity. Introduce
flushDurableStateOrThrowAsync to flush only SSH-recovery state on the
live establish/reconnect path, avoiding snapshot writes of sidecars that
belong to quit/startup. Use Promise.allSettled in test reset to prevent
one rejected disposal from leaking state into the next test.
* fix(ssh): dispose mux on failed establish and propagate sync errors
- Dispose mux when session is disposed during establish to prevent resource leak
- Propagate synchronous errors in teardown via the completion promise instead of leaving completion undefined
- Add test coverage for terminated PTYs that exit mid-reattach and must stay dead
* fix(P1-B): recover system-SSH targets after a network drop
Two defects stopped a remote workspace auto-recovering after a blip.
runReconnectAttempt classified failures with isTransientError, which only
matches ETIMEDOUT/ECONNREFUSED/ECONNRESET by errno code or literal
substring. The system-SSH transport — the only transport FIDO2 and
ProxyUseFdpass targets can use — reports network failures as OpenSSH
prose ("System SSH connection timed out"), so the ladder published a
permanent 'error' on the first timeout and the target never came back
without a manual reconnect. isTransientReconnectError adds a
network-shaped prose table on top of isTransientError and is used only on
the reconnect path: connect() keeps the narrow classifier so an
unreachable host still fails fast instead of burning five 30s attempts
and five security-key touch prompts. Auth and passphrase failures stay
permanent on both paths.
runReconnectAttempt also had no generation fence, so a superseded attempt
published its cancellation as a permanent error over the winner's live
connection — reachable when a system-transport proc.onExit schedules a
reconnect while an attempt is still in flight. Cancellation now carries a
stable error name, and both connect() and runReconnectAttempt claim their
connectGeneration and stay silent when a newer attempt owns the state.
* fix(P1-B): retry a dropped watcher overflow marker on real capacity
emitWatcherOverflowToClient published the {kind:'overflow'} resync marker
with controlOverflow:'reject'. A full control queue rejects at admission
with no settlement callback, so the marker was silently discarded and the
remote File Explorer stayed stale until some later watcher event happened
to produce another one — for a quiet tree, possibly never.
The emitter now retains a rejected marker per (client, root) and
republishes it when the sink actually frees up. The existing
onLegacyPtyCapacity signal cannot drive that: it is gated on producer
retention, so it stays silent exactly under the dual-queue pressure that
caused the rejection. RelayDispatcher.onClientCapacity is an ungated
per-client capacity signal that fires on every writer settlement and
drain. It lives on the dispatcher rather than the writer so a retained
marker survives setWrite() replacing the primary sink, and setWrite
notifies capacity once afterwards so the marker does not wait on traffic
that may never arrive.
Retention is bounded to one marker per (client, root), released on
settlement and purged on client detach.
* fix(P1-B): address all review findings on SSH network recovery
Fix four issues from code review:
1. **Bug — admitted overflow markers lost on setWrite**: Retain markers when
settlement fails `ok: false`, not just on admission rejection. Prevents
desynced filesystem trees after SSH sink replacement.
2. **SSH error classification expanded**: Add missing OpenSSH patterns
(`ssh_exchange_identification`, `connection closed by remote`) and new
`isDefiniteSystemSshHostFailure()` classifier.
3. **ControlMaster retry optimization**: Skip second probe when first failure is
already definite host-level (network timeout, refused, unreachable). Saves
~30s per reconnect ladder step.
4. **Overflow flush under dual-queue pressure**: Gate pending marker retries on
control-lane headroom instead of re-attempting on every capacity notification.
Reduces thrash proportional to producer traffic.
Add regression tests for marker republish on sink replacement and validate auth
error detection against live OpenSSH credential rejection messages.
* rm random doc
* fix(P1-B): skip credential-failure retries and recover watcher markers o
- Auth and passphrase errors fail immediately without retry attempts
- Bare "System SSH probe failed (exit 255)" is transient only for reconnect
- Watcher markers survive client invalidation when switching SSH connections
- Add network error patterns: "lost connection", "remote end closed"
* fix(P1-D): bound hosted-review lookups with a detachable deadline
The `inflight` map in the hosted-review branch cache was only ever cleared
when the lookup settled, and nothing bounded how long that took. One wedged
provider call pinned its branch for the life of the process: every later poll
joined the same dead promise, so the card loaded forever with no in-session
recovery.
Each lookup now runs under a 120s deadline. Nothing below the funnel can be
cancelled, so the deadline detaches instead: the record is released, the
callers get the last known review (or a timeout error), and the branch enters
the existing failure backoff. The lookup keeps running and its answer is still
adopted if it lands, so a slow-but-alive host converges rather than failing
forever. A token identity keeps a detached lookup from evicting the record
that replaced it, and a wall-clock sweep expires records whose timer never
fired — main's timers are suspended across system sleep. `inflight` is capped
independently of the completed cache.
The failure backoff moves to its own module: it has a different lifetime from
the answer cache and is what a deadline records against.
* fix(P1-D): bound `git remote get-url` on the local/WSL path
`getRemoteUrlForRepo` ran the git child with no timeout, which is the one
unbounded step under the hosted-review lookup funnel: `git/runner.ts` only
arms its kill path when a timeout is passed, so a dead network mount or a
stalled WSL interop hangs the call and everything above it. The SSH branch is
already bounded by the relay mux's 30s request timeout, so it is unchanged.
* rm review doc
* rm review doc
* test(P1-D): add probe tests and transient-failure recovery verification
Add tests for coalesced-probe and remote-url-probe infrastructure. Add integration test verifying that transient Bitbucket API failures don't cache as a definitive no-review result, allowing recovery after cache TTL expiration.
* fix(P1-D): track lookups from start, prevent stale scope adoption
- Count unsettled lookups when they start, not after deadline expires: prevents multiple concurrent lookups for the same branch.
- Add evicted generation floor: prevents adopting stale results when scope is invalidated and evicted from the map.
- Consolidate duplicate repository reference cache logic into createRemoteRefProbeCache utility.
- Fix deadline wrapper in git config signature lookup: bound the caller's deadline only, not the coalesced probe itself.
* feat(P1-D): add remote-ref-probe-cache utility
Cache successful remote URL probes per repo/runtime to avoid duplicate work.
Skip caching transient errors and SSH failures so providers can retry on
reconnect, preventing stale scope adoption during the session.
* fix(agent-status): preserve Claude background work
* fix(agent-status): harden background task lifecycle
* fix(agent-status): narrow interruption retention
* fix(agent-status): scope background task authority
* fix(agent-status): isolate lifecycle inventories
* fix(agent-status): harden background evidence recovery
* fix(agent-status): reject ambiguous child authority
* perf(agent-status): skip lifecycle inventory scans
* refactor(agent-status): isolate task inventory parsing
* fix(agent-status): clear stale background evidence
* fix(agent-status): gate accepted remote evidence
* test(agent-status): pin session cron interrupts
* fix: harden Claude inventory tracking
* test: pin Claude cron drain authority
* refactor(agent-status): unify Claude turn-boundary predicate
Collapse the five inline copies of the Stop/StopFailure test into a single
isTurnBoundary constant and drop the reportedStateName/stateName alias, so a
future edit can't move one copy and leave the others behind.
Pin the two behaviors that unification now depends on: a non-interrupted
StopFailure keeps gating on live background work, and interrupted state does
not survive a mid-turn lead event that has no prompt submit.
Co-authored-by: Orca <help@stably.ai>
* fix(agent-hooks): gate local Claude background evidence
---------
Co-authored-by: Orca <help@stably.ai>
* fix(quit): stop durable state writes from parking the main thread
will-quit ran stats.flush() and store.flush() synchronously, before
preventDefault(). Both fsync and rename a multi-MB file on the profile
directory. When that directory sits on a stalled network mount the
syscall enters an uninterruptible wait: the app stops repainting and
stops responding to Force Quit, because a process blocked in the kernel
ignores SIGTERM and SIGKILL alike.
The existing 20s teardown deadline could not bound this. Its timer runs
on the very thread the syscall parked, so it never fires. The fix is to
make the quit path awaitable rather than to try to bound it — a quit
that is slow but responsive stays killable by the OS.
- preventDefault() now runs first, so every teardown step is free to await
- stats and state gain flushAsync() twins that use node:fs/promises
- both join the existing teardown barrier, which can now actually bound them
- the pass-2 will-quit re-entry returns early instead of re-running teardown
- quitFlushStarted makes the quit flush the last write, so a teardown step
touching the store cannot arm a debounce that races process exit
Making the swap async cost the atomicity of check-generation-then-rename:
a writer parked on await rename has already cleared the guard, so a later
synchronous flush could be clobbered by stale state. Both async writers now
claim their temp path, and the sync writers delete it, turning that swap
into a swallowed ENOENT.
Atomic temp+rename is unchanged, so a write cut short by the deadline
leaves the previous file whole — bounded loss, never corruption.
* fix(quit): harden async persistence finalization
* fix(persistence): bound best-effort flushes
OrcaWin
Neil
github-actions[bot]
Jinwoo HongandOrcaWin
VincentandOrcaWin
