Codex reports both approved and denied automatic approval reviews on the
guardianWarning channel. Orca rendered every one as a yellow "safety review
warning", so routine approvals looked like problems. Approved reviews now
render as a neutral "approved this action" notice on desktop and mobile;
denials and timeouts keep the warning tone.
Co-authored-by: Sean876-alt <224890246+Sean876-alt@users.noreply.github.com>
Co-authored-by: Neil <neil@stably.ai>
The runtime service and failed-start cleanup now reach orca.yaml environment
recipes and plugin VM recipes through ShellRecipeDriver's recipe-shaped steps.
Same arguments, same runner, same results: records, failure copy and
diagnostics are unchanged.
* refactor: group agent launch, status, and messaging owners
* refactor: consolidate direct owner imports
* refactor: import pane ownership directly in status listeners
One lifecycle seam (create/resume/suspend/destroy/status + capability flags and a
catalogued error type) that every environment provider implements, a shared
vitest contract any driver runs from its own test, and ShellRecipeDriver, which
wraps today's orca.yaml / plugin VM recipe runner and passes the contract.
No production callers yet.
* test(e2e): give the provisioned-root VM a reachable git remote
VM creates now run the environment checks before creating the workspace
(#40152). The fixture VM's origin pointed at a GitHub URL it cannot push
to, so the git push check failed and the create paused on the build
sheet; the spec then read the VM's primary checkout before it was
adopted. Point the VM checkout at a local bare repo so the checks pass.
* test(e2e): wait for the provisioned root's adoption, not its listing
The VM's primary checkout appears under its branch name once its project
is set up, before the environment checks finish and adoption writes
ephemeralVmCheckoutMode. Poll for the adopted row instead.
Claude paints the caret over the first glyph of a dim composer suggestion
as an undimmed inverse cell. The typed-text read kept that glyph, so an
empty composer reported a one-character draft and the stock Try "..."
placeholder was never recognized. Treat a lone inverse caret glyph ahead
of a longer dim suggestion as untyped, while keeping a real typed last
glyph that sits before a dim argument hint (e.g. /commit [message]).
Fixes#40195
Co-authored-by: Neil <4138956+nwparker@users.noreply.github.com>
The browser-extension preload was a main-build Rollup input, so it could be
split into shared CJS chunks that a sandboxed preload cannot require. Build it
with the other guest preloads as a self-contained bundle, and share the
one-byte unminified output options between the preload build and sandboxed
bundles so the extension preload stays one-byte in V8.
Co-authored-by: Neil <neil@stably.ai>
Request the project on the shared issue node fields and map it in
mapRawIssueForWorkspace, so list/search rows carry the project the same
way issue detail already did. Detail now reuses the shared mapping.
Co-authored-by: Neil <neil@stably.ai>
The Codex config line scanner only treated exactly three quotes as the
multiline closer, so a value like '''printf 'ok'''' left the scanner
inside the string and later tables (e.g. [projects."..."]) were missed
when writing project trust. Match toml_parser: a run of 3-5 quotes
closes the string, with up to two quotes belonging to the content.
Co-authored-by: Neil <neil@stably.ai>
Claude reports deferred categories (MCP tools, system tools loaded on
demand) that sit outside the used total, but summarizeContextUsage
dropped the flag, so the popover listed them as window percentages next
to real usage and the rows summed to well over 100%.
Keep the deferred flag on each summary row and render deferred rows in
their own group under a divider, with token counts instead of a share of
the window and a note that they are not counted in the used total.
Fixes#39878
Co-authored-by: Neil <neil@stably.ai>
Link underlines in Native Chat, the markdown preview and the rich markdown
editor sat 2px below the baseline, so descenders (g, y, p, j) crossed them and
skip-ink broke the line into dashes. Use a 0.25em offset that scales with font
size and zoom, and override the shared renderer's underline-offset-2 in Native
Chat.
Fixes#40217
Co-authored-by: Neil <neil@stably.ai>
Claude echoes "Set model to Opus 5.5" without the "Claude" vendor prefix,
and never echoes the picker-only " alias" suffix or a raw "claude-" id
prefix. The confirmation observer required every picker-label token, so a
successful switch resolved to unknown and showed "Could not verify the
model change". Normalize the expected label before matching; context
tokens such as 1M are still required.
Fixes#40058
Co-authored-by: Neil <neil@stably.ai>
When a dashboard card has no live PTY (e.g. chat-session tabs), the agent
dialog previously showed only the "preview unavailable" notice. It now also
renders the card's stored last user/agent messages, unclamped and scrollable,
below the notice. The rendering is extracted into AgentCardLastMessages and
shared with the kanban card, which keeps its line clamps.
Fixes#40144
Co-authored-by: Neil <neil@stably.ai>
* fix(diff): persist inline / side-by-side diff view via the Default Diff View setting
The file diff and combined diff toolbars toggled inline vs side-by-side only
in local component state (plus a session-only module preference for the
combined diff), so switching worktrees or agent threads reopened the diff in
the saved inline default.
The toggle now writes the existing diffDefaultView setting, and both viewers
derive sideBySide directly from it. The per-view cache and module memory no
longer carry a separate sideBySide value.
Fixes#40241
Co-authored-by: Wooseong Kim <2222333+innocarpe@users.noreply.github.com>
* test(diff): drop the removed sideBySide field from the cached combined-diff view
---------
Co-authored-by: Neil <neil@stably.ai>
* fix(terminal): don't restart every dead saved tab when a worktree opens; log how the daemon exited
Opening a worktree mounted every saved terminal tab that had no PTY, and each
hidden mount started a fresh shell or agent. With many finished agent tabs this
started dozens of agents at once and could exhaust a host's process limit.
Saved tabs with no PTY anywhere and nothing queued to start now wait until
they are shown; the existing capped idle warm-up still admits up to four.
The main process now records the daemon's exit code or signal and names it in
the respawn log line.
Fixes#39166
* fix(daemon): forget a predecessor's exit when a replacement daemon launches
The plain-shell readiness test polled with vi.waitFor (1s default budget)
for reconcileConfig and the spawn env. Under CPU load the spawn takes over
a second and the test failed with 'expected undefined to be defined'.
Await an explicit reconcile-started signal and the spawn itself, and move
cleanup to onTestFinished so the gate is released and spies/env restored
even when a regressed spawn blocks until the test timeout.
Fixes#40078
Co-authored-by: Jerry Wu <haoyuwu@andrew.cmu.edu>
The native chat split action required the source agent-session tab to carry
a TUI agent tag, so sessions without one (e.g. untagged or non-TUI agents)
could not be split and the tour's split step stalled. Any agent-session tab
in a known group is now a valid split source.
Co-authored-by: Neil <neil@stably.ai>
The sticky External Ports header used hover:bg-accent/50, a translucent
background that let scrolled port rows show through on hover. Blend the
accent into the popover color so the hover state stays opaque.
Fixes#40056
Host session-tab snapshots rebuild native chat tabs without the
client-owned focus visit, so Cmd+J recents ranked chats last and showed
their creation time.
Fixes#39841
* test(vm): pin file saves in a recipe-created Orca server workspace (#11762)
* test(vm): wait for the save round trip, let the app reconnect itself, edit a plain-text file
* test(vm): say why the recipe save spec needs a longer timeout
* refactor: group renderer workspace ownership and navigation
* refactor: use the defining WSL rewrite store identifier
* refactor: route new upstream consumers to workspace owners
* refactor: consolidate connection owner test imports
* refactor: preserve upstream review-owner test retirement
* feat(release): publish a server-only Linux tarball for orca serve
Each release now also uploads orca-server-linux-x64.tar.gz and
orca-server-linux-arm64.tar.gz (orcad slot from the qualified template, the
pinned Node, and an orca-ide launcher) plus orca-server-sha256sums.txt, so a
Linux host can run orca serve without the desktop package.
The tarballs are packed and smoke-tested (serve --recipe-json, then pair a
second profile) on both architectures, including bare ubuntu:24.04 and
debian:12-slim containers. Upload runs only after the release is published
and nothing depends on it, so it cannot hold the desktop release.
* fix(release): keep the server tarball's launcher env out of its terminals; skip refs that predate it
- The launcher runs the CLI with ELECTRON_RUN_AS_NODE=1, as the packaged app does, and the CLI strips it
before starting orcad; ORCA_APP_EXECUTABLE and ORCA_SERVE_RUNTIME are cleared.
- Ship Node's LICENSE beside the runtime.
- Skip packing on release refs that predate the scripts, and upload only when both tarballs passed.
- Run the PR smoke when the CLI's serve-runtime selection changes; document service install.
* fix(dictation): don't throw when the paired web client has no speech catalog
* feat(i18n): translate the dictation pill, error pill and queued-send copy
* fix(i18n): say Chinese queued sends happen automatically
* fix(terminal): end a dead preserved terminal host's PTYs as host loss
After an update Orca keeps the previous version's terminal host so agents
survive. That host has no respawn, so when it later died nothing ended its
PTYs: panes stayed connected on a frozen frame and input went nowhere.
Once every reconnect attempt finds its endpoint refused or missing, each of
its PTYs now ends as host loss with history kept, and the pane's restore is
no longer refused because the dead host cannot answer.
Fixes#38432
* fix(terminal): require process-exit proof on POSIX and keep checking until proven
Review follow-ups: an unproven attempt restarts the count instead of ending
the check; on POSIX the old host's pid must also be gone; a proven-gone host
answers 'not live' for every id; one throwing exit listener cannot strand
the remaining sessions.
* fix(dictation): wait as long as the host's cloud deadline when a paired phone or browser stops
* fix(dictation): compute the host's cloud deadline from its exact sample count
* fix(agent-picker): list agents named by the search before other agents' model setups
Typing "codex" in the agent picker showed OpenCode/Cursor model setups
(e.g. "OpenCode · GPT-5 Codex") above the Codex agent itself, and Enter
picked the first setup. Direct agent matches now render first and are the
Enter target.
* test(agent-picker): assert search highlight and Enter target for named agent
* test(agent-picker): import fireEvent after rebase
* fix(dictation): store Auto as Toggle plus hold-to-record so older builds keep a working shortcut
* test(mobile): type the Auto mode settings fake instead of casting
* fix(dictation): migrate a stored Hold without hold-to-record
* fix(dictation): keep phone mode segments typed and clear hold-to-record when a phone picks Hold
* refactor(dictation): reuse the shared mode settings in the migration and phone writes
* fix(native-chat): continue every usage-limited chat at once and stop after five automatic refails
The automatic continue sent one chat per 10 s sweep, so chats stopped by the same limit resumed one by one. Its retry budget belonged to the cut, so a continuation that hit the limit again started a fresh budget and could repeat indefinitely.
Send every due chat in one pass, one at a time. Count consecutive automatic continues that ended in another usage limit from the journal: each continue's message id names the cut it answered, and the next cut's turn is joined to that submission through its key or provider item. After five, the chat shows that Orca won't continue it automatically.
* fix(native-chat): read the usage-limit continue chain from the tail, not the whole conversation
The chain walk built a map over every journal row and rescanned history for each link. Walk back from the end instead, so the work is the chain's own rows; add an N/10N contract. Replace the fan-out test's casts with typed fixtures.
* test(native-chat): count a Claude echo-opened continue in the usage-limit chain
* fix(native-chat): derive the usage-limit continue chain once per cut
A status rebuild for the same cut reuses the chain it already derived, since only history before the cut decides it; the turn's limit is read without copying its rows.
* fix(native-chat): skip slash-command turns when walking a usage-limit continue chain
A /compact run after a usage limit sits between the cut and its automatic continue; the latest-turn read skips it, so the chain walk must too. Count visited rows in the perf contracts with typed getters.
* test(native-chat): type-guard the recorded Claude capture events
* fix(native-chat): recompute the usage-limit chain when a late receipt attributes its opener
Reuse across rebuilds is keyed on the cut, its turn row's revision and the newest send, so a late attribution that completes a join is counted instead of keeping a stale zero.
* test(native-chat): guard usage-limit chain reuse at the status projection
Drives StructuredAgentSessionJournalProjections.read through real journal invalidations at 100 and 1000 turn rows: one chain derivation for an unchanged cut, again for a new cut or epoch, and a recount when a late receipt attributes the continue's opener.
* fix(native-chat): reuse a usage-limit chain only while everything it read is unchanged
The chain walk now reports the turn rows it looked up and the oldest submission it examined; the status projection keys reuse on their revisions and join evidence, so a late receipt for any link, not just the newest send, recounts the chain.
* fix(native-chat): derive the usage-limit chain fresh on each status rebuild
Reuse across rebuilds kept missing inputs the walk reads (late receipts, revised failure rows) and its key re-serialized settled history. The status projection's own cache still skips unchanged reads; each rebuild now walks only the chain's tail. Contract tests at the projection count total reads: independent of earlier history, linear in the chain's rows, one scan of the sends when no opener matches.
* fix(native-chat): find an echo-opened turn's opener without reading the turn before it
A Claude echo-opened turn names the provider's key while its opener row keeps the submission's, so the slash-command check scanned the whole previous turn and missed echo-opened commands. It now stops at the nearest root user row.
onorca.dev offers an opt-in 'mirror link' (https://downloads.onorca.dev)
next to the GitHub download for when GitHub release downloads 5xx.
Upload the onorca.dev download assets there right after each stable
desktop release and each Android release, instead of polling.
- config/scripts/release/publication/mirror-release-to-r2.sh: retrying,
resumable download from GitHub, size + sha256 check, S3 multipart
upload to R2 (no 300 MiB wrangler cap)
- release-mirror-r2.yml: reusable + manual backfill
- release-cut: run after publish-release for stable vX.Y.Z tags only
- mobile-android-release: run before bump-apk-download so the site's
mirror link never points at a missing APK (success not required)
* fix(dictation): stop a phone's cloud dictation at the 10-minute limit with a notice
* test(dictation): pin the cap flag on every phone chunk reply
* perf(dictation): skip decoding phone audio the host will drop past the cloud cap