* fix(execution-host): a non-string saved host id no longer crashes the terminal workbench
Crash d1e50ca3 (1.4.259, Windows): rendering a chat name passed a non-string
executionHostId into normalizeHostPart, which threw 'e?.trim is not a function'
and took down the terminal workbench boundary. Treat it as absent, as blank is.
* fix(execution-host): take unknown at the host-id parse boundary instead of casting in the test
* fix(execution-host): a repo with a non-string host id no longer proves it is local
* fix(native-chat): guard only the chat-name host lookup, not the shared host-id parser
Reviewers found that making the shared parser read a non-string as absent turns it
into 'local' for worktree removal and trash cleanup, which could act on this
machine's disk. Restore the shared parser and check the type at the one render path
the crash report hit.
* feat(workspace-cleanup): split settings into Manual and Automatic cleanup sections
Manual cleanup gets its own section with a Review workspaces button. The rules
section is headed Automatic cleanup with an On/Off pill, a status line and Run
rules now (only while a rule is on). Turning a rule on shows its check in the
switch and under the rule, with an inline Try again when the check fails.
* fix(workspace-cleanup): move pending switch look into Switch and wrap the manual row
* fix(workspace-cleanup): narrow layouts for Activity, rule titles, Advanced and the Automatic heading
* fix(workspace-cleanup): announce a cleanup run starting and finishing to screen readers
* fix(workspace-cleanup): keep the Automatic cleanup search target and mark a pending switch unavailable
* fix(workspace-cleanup): announce every finished run, even when worded like the last one
* fix(workspace-cleanup): announce a run's result and the next start together when they arrive at once; keep main's disclosure summary
* fix(i18n): retire Simplified Chinese strings the Workspace Cleanup page no longer uses
* chore(i18n): sync protected-copy keys for the Workspace Cleanup page
* test(workspace-cleanup): check the joined result and start announcement text directly
* fix(i18n): update Simplified Chinese for the Workspace Cleanup page layout
* fix(agent-picker): wheel-scroll agent rail, warm tooltips, Blank Terminal tile
- A vertical mouse wheel scrolls the overflowing agent rail sideways.
- The rail shows a thin scrollbar on hover instead of hiding it.
- Rail tooltips share a provider with a skip delay, so moving across tabs shows labels at once.
- Blank Terminal is a tile ahead of the agents instead of a row under the card; when it is current, the dropdown opens on it with one checked row.
* fix(agent-picker): thinner rail scrollbar thumb
* fix(agent-picker): drop unused import after rebase
* test: remove low-value permission and worker projections
* test: isolate coordinator mail setup from the system heartbeat
---------
Co-authored-by: OrcaWin <293788423+OrcaWin@users.noreply.github.com>
The install status only looked for Orca's own Antigravity runtime, so a new chat that would run the user's agy showed the install screen until its session status arrived. Count a ready host program as found on this host.
* feat(cli): add orca agent installed to list installed agent CLIs
* fix(cli): agent installed omits disabled agents, marks the default, and is in the agent guide
Lead the launch picker's effort row with a Default segment where omitting
effort is valid. Picking it delegates effort, so the launch sends no effort
flag or session value and the agent runs its own level; the trigger summary
leaves the level out.
* fix(wsl): scope legacy WSL background probing to distros with legacy projects
Startup CLI registration, AI vault WSL home scans, Claude WSL profile
publish, and daily WSL telemetry no longer run wsl.exe when no local
project, folder or WSL terminal uses a distro. WSL host distros are
served by their own orcad, so Windows-side legacy probes skip them.
Refs #28000#28167#13190#13278
* fix(wsl): drop project-demand scoping; keep only harmless WSL probe skips
Review found the legacy-demand reader had false negatives (per-tab WSL
shells, worktrees outside the repo, non-persistent net use, early return)
that silently dropped running distros' sessions, and a false positive for
moved repos. wsl --list does not start a distro, and the session scan and
CLI repair already touch only confirmed-running distros, so revert that
scoping. Keep the Claude publish and telemetry skips, which only omit a
running-list read when there is nothing to do with it.
Refs #28000#28167#13190#13278
* fix(telemetry): omit distros_running when the WSL running state wasn't probed
A skipped probe was reported as a confirmed empty list, so the daily
snapshot sent distros_running=0 even with distros running. The field is
now optional and left out when no probe ran.
Refs #28167
Terminal workers without unread mail at dispatch got no check command in
the preamble, so they guessed `check --from` when Orca later reported
messages. Always include `check --terminal <handle>` for terminal
workers; chat workers keep receiving mail directly.
Fixes#43477
Co-authored-by: Neil <neil@stably.ai>
* refactor(renderer): group browser and terminal modules by owner
* refactor: consolidate floating workspace type imports
* test: mark worktree card introduction seen in completed profiles
* refactor(renderer): place upstream terminal sleep owners with tab lifecycle
Closing a host-mirrored editor tab from the tab strip retired the tab
before closeFile could find the host's tab ids, so the host never learned
of the close and restored the tab on the next workspace visit. Notify the
host per closed view from editor tab retirement, scoped to that view's
tab id, and keep the host tab for the last view of a file that was dirty
when the close began (Save with only the host's draft).
Fixes#43470
Co-authored-by: Claude <noreply@anthropic.com>
* fix(agent-picker): retry model catalog reads an older host refused before its chat host started
Hosts before 1.4.248 refuse a session-less agentSession.modelCatalog read with
structured_agent_session_unsupported (hostDisabled) until their chat host is
built. The picker treated that as permanently unsupported and fell back to
built-in models, which are empty for Pi and OMP. Treat that refusal as
transient and retry failed catalog reads a few times while the picker is open.
* fix(agent-picker): retry only the old-host refusal and dropped connections
Other catalog read errors, including timeouts, show at once as before.
The agent picker read Antigravity's model catalog under Orca's managed account,
while new chats launch on the host's agy when it is compatible. Models that agy
chats reported were saved under the host account, so the picker kept saying
"Antigravity models have not loaded". Catalog reads now resolve the account a
new chat would launch on; Orca's own sign-in still reads the managed account.
* fix(native-chat): retry a GitHub link preview when its account binding comes back
A preview lookup that started under account A is retired when the repo's
binding changes, so its late answer can't paint. If the binding returns to
A before the card re-renders, the card keeps its key and never asks again,
leaving it on "Preview unavailable". Ask again when the retired binding is
current once more, and cover both orderings with a card-level test.
* fix(native-chat): restart a preview as soon as its account binding returns
Start the fresh lookup when the retired binding is current again instead
of waiting for the retired request to settle, so a slow, hanging or
failing retired request cannot leave the card failed. Errors under an
unchanged binding still reach the card.
* ci(windows): require a trusted signature on every shipped Windows binary
The release gate checked only the binaries it had sent for signing plus a
fixed list. A Windows binary that skipped signing or arrived after staging
could ship unsigned, and Smart App Control machines refuse to load it; that
is how unsigned dev-channel builds lost their terminals on a test machine.
Release cuts now also check every .exe/.dll/.node in the shipped installer
payload (minus the non-Windows server-template binaries) for a valid
SignPath or Microsoft signature and fail with the offender list. Dev-channel
builds run the same check in advisory mode and never fail.
* ci(windows): match signers by CN and harden the payload signature gate
- Match our signer on its exact CN and Microsoft vendors on CN+O instead of
the full subject, so a certificate renewal that moves address fields
cannot block a cut; quoted, escaped and repeated attributes cannot spoof it.
- Say how to allow a vendor when the gate fails: release staging skips any
validly signed file, so a new vendor must be added to VENDOR_IDENTITIES.
- Reset LASTEXITCODE after the check so the fail-open path really warns,
and put the script's stderr into the evidence verdict when it cannot run.
- Force UTF-8 PowerShell output and time the signature read out after
10 minutes (reported as could-not-run).
- The pass message counts only the checked binaries and no longer claims
they are all SignPath-signed.
A settings row with no card around it had no edges for the tint to bleed
to, so the highlight stopped a few pixels from its text. Such rows now grow
a rounded 10px halo in the same tint, fading with it; rows inside cards and
highlighted cards keep their current look.
The desktop returns browser bridge results (such as a screenshot of a heavy
page) over the shared subscription socket. That channel only enforced the
4 MiB outbound JSON limit, while hosts cap inbound frames at 1 MiB and close
the socket with 1009 on a larger one, which ended the subscription and
detached the bridge. The channel now refuses a request whose serialized JSON
exceeds the host's inbound request limit before sending it, so the existing
desktop_browser_bridge_result_undeliverable fallback reaches the host and the
bridge stays attached.
The browser client host lease shares that channel. An unstaged result for a
host that does not stage results is now capped by the command's real request
envelope (its identity JSON plus a small allowance for id, token and method),
so a result that would overflow the request reports
browser_host_command_result_too_large instead of retiring the lease. The
allowance is larger than a real envelope, so values in a band of about 0.8 KB
just below the largest request the host accepts, which used to send, now
report too large as well. The orcad migration budget reuses the same request
envelope reserve.
This is a mitigation: a heavy-page screenshot can still be too large to send
over the bridge, and now fails with a clear undeliverable error.
Refs #43450
* Preserve path-shaped task text when generating through WSL
Keep prompt provenance through argument composition so only path operands pass through WSL translation. Carry the prompt-preservation work from #41594; tracked in STA-14217.
* Reuse the argument projection while scanning recipe options
* Scan rejected-model options once when preparing a retry
* Verify literal prompt arguments survive model-rejection retries
* Append retained retry arguments without call-argument limits
* Share recipe option compaction with rejected-model retries
* Keep Chat sends from merging with or stranding text in the agent's input line
A Chat message or model/effort change could submit together with text already typed in Claude's terminal input ("Existing draft/model opus"), and a Chat message could sit typed in Claude's prompt while the chat showed it as sent.
The clear that empties the input line went out right before the message, so Claude read both in one chunk, took the clear bytes as pasted text and dropped them ("Removed 58 invisible characters from your prompt"); model and effort commands sent no clear at all. Chat now clears only when the line holds text, waits for that clear to be acknowledged and read on its own, and sends model/effort commands through the same path as messages. Enter now waits until the input line shows the pasted message, the proof the host's prompt writer already uses, instead of a fixed 500 ms that a slow paste can outlast.
Tracked in STA-14532, STA-14534.
* Share the input-line reader between option commands and PTY sends
Keeps the composer under its line limit by passing one target object to both.
* Hold Enter for every message path until the paste stops changing
Review follow-up. Enter for plain, image and observed sends now waits until the input line, empty before the paste, shows text that stayed the same across one re-read (or becomes unreadable), instead of matching the message text, which missed sanitized escapes and accepted partial long pastes. The empty-line check that skips the clear now runs when the send starts, and cancellation clears the send's own body again.
* Press Enter only when the input line shows the whole message
Review follow-up. A stable fragment of a paste still arriving no longer counts: Enter waits until the line shows the message's text, or only the chips an agent draws for a paste or image. The 5 s cap is measured in wall time, so a hidden window cannot stretch it. Fixes the image-send write type and a test cast the gate flagged.
* Match the whole message across wrapped rows and keep captions from riding on image chips
Review follow-up. The input line is compared with all whitespace removed, so a long URL or path the agent wraps across rows still counts as shown; the expected text is prepared once per send. An image chip alone now proves only an image-only send; a caption must appear as text or as a text-paste chip. Claude's leading `!` bash-mode marker is allowed for.
* Re-project the agent's input line only after new terminal output
Council ruling on the Enter gate's polling: reads of the input line now reuse the last projection until xterm parses more output or the screen geometry changes, so a send waiting on a quiet terminal no longer rebuilds the screen projection every 100 ms.
* Drop the cached input-line read after a resize or terminal clear
Review follow-up: a resize reflows without parsed output and clear() drops lines without any, so both now invalidate the cache.
* Close gaps the bot review found in clearing and the Enter check
A cancelled send clears its whole body (up to the widest clear), not the old line's size. When text still sits on the input line after the widened clear, the send stops instead of pasting onto it. An empty line no longer proves an image-only paste landed, and agents that take images as typed @path mentions expect those mentions on the line. The cached input-line read also keys on the cursor row's text, which covers output parsed before xterm reports the write.
* Keep widened-clear sends going and read typed image mentions correctly
Review follow-up. Refusing a send when text survives the widened clear dropped non-Claude sends without a notice and would refuse clears that landed but redrew slowly, so the body is written as before. An empty line again counts for an image-only send at the old timing, for agents that draw no chip. Agents that take images as typed `@path` mentions now expect the mention text, unwrapped from its paste framing, so their Enter no longer waits the 5 s cap.
* fix(notes): keep idle terminal agents selectable in Send notes to
An idle agent sends no status events, so after 30 minutes (or right after a
restart) Send notes to greyed out live panes as 'Agent status is stale' for
every agent without an explicit idle title (Kiro, jcode, ...). Treat a pane
with an agent row and a live PTY binding as selectable regardless of row age;
the host's process/screen check still gates the send. Keep an unanswered
permission prompt disabled at any age unless a live title shows it cleared.
* fix(notes): keep permission evidence pane-scoped in Send notes to
An unanswered permission prompt now disables its pane until the agent
reports again, at any age; an idle-looking title no longer clears it,
since Claude keeps its idle glyph while a dialog is open. A tab title is
used as a pane's live title only when the tab is not split, so a split
sibling's title never speaks for another pane. Target derivation no
longer takes the current time.
* test(notes): complete the status-entry fixture
* test(notes): use current agent rows for idle send targets
* test(notes): exercise expired agent status rows
* fix(notes): never send notes into an agent's approval menu
"Send notes to" now keeps live agents selectable at any idle age, so it must never press Enter on an approval menu the host cannot see. Kiro, Amp, Command Code and Cursor keep reporting "working" from their hooks while an approval waits for a key, and the host treated that as sendable (main already did so for the first 30 minutes).
- The host recognises Kiro's approval menus on screen (rules ported from the common pattern's Kiro screen rules), so a send refuses with "needs permission" while one is up.
- A mid-turn agent whose hooks hide approvals and whose screen rules can't see them (Amp, Command Code today) is refused with "may be waiting for approval"; idle agents are unaffected. Adding a screen rule for such an agent lifts this automatically.
- Older paired hosts can't do either, so on hosts without the new terminal.send-approval-guard capability "Send notes to" disables those agents' mid-turn rows with the same reason.
Tracked in STA-15222.
* fix(notes): refuse mid-turn sends to agents whose hooks hide approvals
Both reviewers showed a screen rule can't prove an agent has no approval menu up: Cursor's file approvals and Kiro's question panel and wrapped footers slip past, and the rules made the wait-text rebuild hot. Drop the Kiro rules and the host capability. A Kiro, Amp, Command Code or Cursor pane is sendable only when that agent's own latest report ended its turn; mid-turn, or with no report at all, both the host guard and Send notes to refuse with "may be waiting for approval". The same rule on both sides works with older hosts without a capability check, and a title can no longer clear the refusal.
Tracked in STA-15222.
* fix(notes): check the running agent before trusting a pane's report
Review pass 2: the host trusted a row another agent left on the pane and missed shell-launched agents with no cached identity; a newer client on an older host could still reach Enter after a turn started. The guard now reads the pane's foreground agent (falling back to launch identity), ignores rows from another agent or terminal, and the client refuses a working report from a hook-silent agent before paste and before Enter.
Tracked in STA-15222.
* fix(notes): name a wrapped agent by its title before allowing a send
Review pass 3: an agent behind a wrapper such as node, with no hook row or launch identity, read as no agent at all, so the host guard and the client's older-host check let a send through. Both now fall back to the agent the terminal title names.
Tracked in STA-15222.
* fix(notes): don't let another agent's row vouch for a launched Kiro, Amp, Command Code or Cursor
Review pass 3: a newer client on an older host trusted a Claude row left on a pane that now runs a launch-owned hook-silent agent. The listing and the client's agent lookup now treat that pane as the launched agent.
Tracked in STA-15222.
* fix(notes): refuse hook-silent agents on hosts without the send guard
Council ruling (2 of 3, option B) after review pass 3. An older paired host has no guard between the client's last check and the Enter write, so the client can't make a send to Kiro, Cursor, Amp or Command Code safe there. New hosts advertise terminal.send-approval-guard.v1; on hosts without it "Send notes to" disables those agents with "Update Orca on this host to send to this agent" and the send refuses before any write. This replaces the client's working-status check.
Identity is per pane: the agent's own row, the tab's launch agent only on the launch PTY, or the leaf's own title; any of them naming a hook-silent agent counts, so a row another agent left behind can't vouch for it and a split sibling isn't refused for the launch agent. The renderer reads the guard as a stable set of host ids, so unrelated host-status updates don't rescan targets.
Tracked in STA-15222.
* test(notes): build send-target fixtures from accepted status rows
* fix(notes): keep the hook-silent send guard in its own runtime layer
check:changed: the guard pushed terminal-agent-status-and-submission.ts over the file-length limit, a test concatenated strings, and a fixture read the wall clock.
* fix(notes): refuse when pane identities disagree about a hook-silent agent
Review threads: a stale launch identity, a row from another hook-silent agent, or a split sibling borrowing the tab's launch agent could let the host guard or the target list trust the wrong turn-end report. One shared rule now decides for both: every identity the pane has counts, a send needs exactly one hook-silent agent whose own latest report ended its turn, and launchAgent only names the tab's original PTY.
Tracked in STA-15222.
* fix(notes): treat any paired pane as the launched agent's and cache host guard reads
Review pass 4: a paired tab's ptyId follows its active pane, so tying launchAgent to it could drop a launched Kiro's identity on an older host; on paired panes the launch agent now counts for every pane (safe direction). The guard-host set is cached by status-map identity, the host guard also matches rows keyed only by terminal handle, and a focused pane without a leaf binding is checked against the tab's PTY instead of skipped.
Tracked in STA-15222.
* wip: stored tool output format, hydration and writer
* wip: compaction page bounds, fork payload fix, type fixes
* wip: merge stored-output imports
* fix: plain client heads, writer-entry hydration, compaction passes and row isolation
* fix: forward-only compaction cursor, length guard, no casts, worker baselines
* perf: cap stored text at 2 MiB and read selected blobs in kind scans
* feat: store finished tool outputs compressed with an inline preview for older builds
* fix: ship stored-output reads with writes off; fork capture reads selected blobs
Older builds rewrite and fork rows without knowing about stored outputs, so writes
stay off until those builds leave the downgrade window.
* fix: a stored output copied without its blob keeps its preview instead of failing the write
Forks of an output whose blob could not be read no longer abort; covers the writes-off rewrite.
* chore: move the items-by-kind page's sqlite uses to its own baseline entry
* perf: compaction reads row headers lazily so a page cut by its budget never refetches the rest
Up to 511 headers per page were fetched again by the next page when large rows ended it early.
* fix(tab-bar): fit the + menu agent list to the screen and point the simulator tip at the phone
The agent list was capped at exactly 7 rows, so it cut on a row edge and
looked complete. It now grows into the room Radix gives the menu (up to 12
rows) and, only when it still overflows, cuts through a row with a fade.
Tiny windows hide the tip first, then let the whole menu scroll.
The simulator tip's arrow was fixed under the first icon; it now follows
the phone icon wherever the customizable row puts it.
* fix(tab-bar): re-measure a hidden + menu tip when the menu width changes
A tip hidden at a narrow width kept its taller wrapped height, so it could
stay hidden after the menu widened even when the shorter tip would fit.
Key the cached height to the menu's outer width (which a scrollbar the tip
causes does not change) and measure the tip again at a new width.
* refactor(tab-bar): keep a yielded + menu tip hidden until the menu reopens
Drop the width-keyed tip height cache and the restore path. Once the tip
gives its room to the agent list it stays hidden for that menu session;
the next open measures it fresh. A tip that stops being eligible simply
unmounts, so its room is freed without any extra bookkeeping.
* fix(artifacts): explain artifact API rate limits and retry once on a short Retry-After
A 429 from the artifact API now carries the server's message and its reset
time in local time to the CLI and the desktop share toast; older servers'
bare 429s say the request was rate limited and when to retry. A Retry-After
of 60s or less replays the request once with the same Idempotency-Key;
longer waits fail fast and keep the pending create intent so a later share
reuses its key.
* fix(artifacts): cancel a rate-limit replay when the account changed during the wait
A Retry-After replay waited up to 60s and then resent with the token captured
before the wait, so a sign-out, account switch or revoked session in between
could still create, update or delete an artifact under the old account. Every
artifact request now passes a check that runs immediately before the replay:
the profile, its cloud link and share lifecycle must be unchanged and, for a
stored session, the session must still be signed in. A request without the
check never replays.
The auth-context builders move unchanged into artifact-cloud-auth-context.ts to
keep the service under its line budget.
* fix(artifacts): cancel a rate-limit replay once sign-out has started
Sign-out advances the session-owner generation before it awaits remote logout,
but only marks the session signed out and unlinks the profile afterwards. A
Retry-After replay that came due during that wait still passed the account
checks and resent with the old token. The replay check now also requires the
session generation captured with the request to be unchanged; token rotation
does not advance it, so normal refreshes still replay.
* refactor(cli): keep the artifact timeout floor in the request-timeout module
Main grew the CLI client to its 300-line budget, so the artifact timeout floor
now lives beside the other per-method timeout helpers and the client reaches
them through its existing import, adding no module to the CLI's load path.
* fix(artifacts): keep one rate-limit wait per operation across a 401 session refresh
A 401 makes the session refresh re-run the whole artifact operation, and each
run got a fresh 429 wait-and-replay. So 429 -> 401 -> 429 -> 200 sent four
POSTs with the same Idempotency-Key, and a reconciled create plus its update
could wait twice and outrun the CLI's 210s request timeout.
The allowance is now created once per operation, before the refresh re-run,
and every request in the operation draws from it. 429 -> 401 -> 200 still
finishes in three calls; 429 -> 401 -> 429 surfaces the second 429 without
another wait. The CLI timeout is now the sum of the most requests one share
can send (five), the single wait and both session refreshes, still 210s.
* test(artifacts): drop timeout tests that pinned production constants and timing
Addresses FARTIFACT429-C4-B1: remove the CLI artifact timeout arithmetic test and
the slow re-run timing/work-count test, plus the imports and stub latency/update
helpers only they used. No production code changes.
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
agy names its quota buckets "Weekly Limit Remaining" / "Five Hour Limit
Remaining", so the usage roster, provider card and mobile showed those raw
names beside used percentages and truncated the reset caption.
getUsageWindowSections also appended the derived weekly rollup after
Antigravity's per-pool weekly buckets, showing the weekly value twice.
Label grouped pools from their group and window length on desktop and
mobile (as the status bar already did) through shared helpers in
grouped-usage-pool, strip the direction suffix from parsed bucket names,
and drop the weekly rollup only when it repeats a listed 10,080-minute
bucket's usage and reset.
Fixes#43453
Co-authored-by: ASH <7201776+asunghong@users.noreply.github.com>
Co-authored-by: Neil <neil@stably.ai>
* test(codex): reproduce local-date rollout publication gaps
Co-authored-by: yeonwoo <fox_93@naver.com>
* fix(codex): publish rollouts from local calendar dates and neighbors
Use local calendar keys with UTC-only calendar arithmetic, including launch, exit and abnormal-exit recovery scan targets. Probe neighboring dates for shell timezone differences while retaining bounded directory scans.
Co-authored-by: yeonwoo <fox_93@naver.com>
* test(codex): expose west-of-UTC crash recovery gap
* fix(codex): walk existing rollout folders newest first within a launch budget
Keep bounded recovery progress in the existing marker and use actual day folders as inclusive scan checkpoints. Remove predicted launch dates and calendar expansion.
Co-authored-by: yeonwoo <fox_93@naver.com>
* fix(codex): preserve fresh scan debt across bounded recovery
Resume year and month cursors at equal prefix depth, retain the original sweep checkpoint, and revisit new dates before advancing coverage. Reconcile legacy UTC markers without treating their pending date as a trusted filesystem checkpoint.
* fix(codex): persist and apply the continuation checkpoint
* test(codex): remove cancellation callback-count assertion
The filesystem walk can observe cancellation before finding any file. Callback invocation counts do not define the publication contract.
* Preserve Codex continuation scope across recovery
Co-authored-by: yeonwoo <fox_93@naver.com>
* Retain prefix replay for renewed Codex history demand
Co-authored-by: yeonwoo <fox_93@naver.com>
* Clarify backfill marker and filesystem mock types
Co-authored-by: yeonwoo <fox_93@naver.com>
* Place mock overload rationale on asserted expression
* Require unrestricted coverage before certifying Codex history
Co-authored-by: yeonwoo <fox_93@naver.com>
* Keep retained baseline certification while testing unrestricted scans
Co-authored-by: yeonwoo <fox_93@naver.com>
---------
Co-authored-by: yeonwoo <fox_93@naver.com>
When closeTerminal stops a headless single-pane tab's PTY before closing the
tab, the PTY exit can already retire the durable tab. The following
closeMobileSessionTab then throws tab_not_found, so the close failed and the
renderer kept a dead tab. Treat tab_not_found like workspace_session_unavailable
on that post-stop path and notify the renderer.
Co-authored-by: Neil <neil@stably.ai>
dev-fresh-profile spawned a bare `pnpm` with shell:false, which fails with
ENOENT on Windows where only pnpm.cmd/pnpm.exe exist. Use
resolvePnpmCliInvocation() like the other dev scripts so the child pnpm is
the one that launched the script.
Fixes#43413
Co-authored-by: Neeidaan <228572363+Neeidaan@users.noreply.github.com>
* fix(native-chat): deliver other agents' messages to chat agents directly again
#43093 sent chat agents only a pointer at `orca orchestration check` for
mail from other agents, while the chat showed the full message, so the
agent never received what the chat displayed and every message cost an
extra tool call. Chat agents again receive each orchestration message
whole, with its sender header, exactly as the chat card shows it; a
message too large for one chat message still falls back to the pointer.
The Approve once control from #43093 is unchanged.
Tracked in STA-15406.
* fix(orchestration): keep custody hold and owner wake for mail not delivered whole
Whole-delivered mail is read on settle. A pointer or in-doubt send still
holds while custody is unreadable, is handed to the mailbox's new owner
when the chat lost it, and a send made obsolete by a handoff wakes that
owner.
The intro now floats right of the sidebar as a non-modal note with Classic/New
picture tiles, a switch-back hint and Not now / Keep new (or Use classic).
It stays out of Escape and focus ownership, waits for a confirmed save before
announcing Keep new, and returns focus to where it came from.
thread/fork does not inherit the parent's model, so a side chat forked from
a Codex chat was saved on the config default and later resumed on it. Pass
the chat's selected model (falling back to the source thread's model) and
the source thread's provider to thread/fork.
Fixes#43414
* test(agent-launch): give at-rest launch waits the host rig's 10s settle window
The open-failure retry case waited for dispatch under vi.waitFor's 1s default. It took 778ms on a green
bun shard and ran out on a shard about 2x slower, though the retried child had opened.
* test(agent-launch): leave the private child-state waits untouched