On a held 'Claude isn't set up' pane, Cancel during the browser sign-in
released the hold with 'cancel', so the pane lost its strip and Start for
good, even after Claude was set up from Settings. Cancel during a setup now
stops only the setup; the pane is back on 'Set up once', and a setup from
anywhere flips it to 'Start Claude' (the held strip reads the live account
state). Cancel with no setup running still gives up the launch.
Bare 'npx vitest' skips run-vitest.mjs's env (e.g. ORCA_BACKGROUND_LAUNCH=1) and the workspace package
build, which made 37 tests fail spuriously in a full run on integration HEAD.
design-squad: the glossary reason ('Out of sandbox hours') is never cut. Under
Run on the line wraps (action on the same line if it fits, else below). In the
menu the 'Orca Cloud' title stays whole, the reason wraps on the second line, and
the action aligns to the top. Only in the flagged wall state. Fixtures added.
#347 (c267ae3676) broke cloud-workspace-archive-delete-race.test.ts and the premerge never ran it,
because it only mapped changed sources to sibling-named tests (foo.ts -> foo.test.ts). #435 then ran 0
targeted files for the same reason. Now each changed source also selects every test in its folder and
every src test that imports it by module name; for cloud-workspace-actions.ts that is 14 files including
the race test. PREMERGE_FULL=1 runs the whole vitest suite.
W6 saw ~750ms before a wake went live (pill still "Connecting…"): the row back to
"Asleep", the Files panel's Wake button, and the terminal's "This workspace is
asleep" note. Same race as #408 on the way in: main's ready lands before the status
re-read, and the stale "sleeping" read (or the not-yet-updated VM record) won.
#408 only fixed the row word's derivation; the terminal note and Files panel read
the status and VM record directly. One rule now, isWakeReadyAfterRead: a ready
wake newer than the last status read makes those reads stale for the dot state,
the view's VM-record override, the terminal recovery context and the Files
panel. A newer read or a failed wake still shows asleep.
A session whose first turn failed (expired sign-in, network, the harness's fake
token) has no transcript title or prompt, so the vault answers the generated
'Claude 6ebe07c5' and the tab-title sync wrote it over the tab's own title. The
sync now skips that generated name; the tab keeps 'Claude' or its launch title.
The generator moves to shared so main and the renderer agree on it.
Behind the same runOnPreflight flag: when Orca Cloud is already the selected
target (a remembered default) and a billing wall applies, the selection stays,
one line under Run on gives the reason and the one fix (members: 'Ask an owner
or admin'), and Create is disabled until it's fixed. Dev/QA only:
localStorage['orca:dev:billingFault'] = out-of-hours | payment-issue | lapsed
simulates a wall for real-app shots (import.meta.env.DEV).
On an Orca Cloud VM (orcad, ORCA_IS_CLOUD=1) the pane echoed the whole agent
launch line: claude --plugin-dir '/vercel/...' '--dangerously-skip-permissions'
'<prompt>'. Orca-built agent lines are now always staged there, so the pane shows
. /tmp/orca-launch-<hex>.sh; the script is 0600 and removes itself when the shell
runs it, before the agent starts. Local launches are unchanged (cloud-only rule).
With capability flag runOnPreflight (or localStorage orca:dev:runOnPreflight=1
for the harnesses), the composer's Orca Cloud row is disabled for an owner or
admin whose team can't start a sandbox now (out of hours, payment failed, plan
ended): it shows the reason and its one fix (Turn on overage / Update payment
method / Renew), and Enter runs that fix instead of picking a target that would
fail after Start. Members keep 'Ask an owner or admin'; a plan not started yet
(trial offer) stays pickable, it's setup. Flag off: no change.
A cloud Restart stopped the pane's host terminal, ensured a --resume in its
leaf, then reconnected. But the pane's own transport reacted to the stop
ending its stream by resubscribing (session.tabs.activate), which woke the
stopped leaf with its ORIGINAL launch before the ensure arrived; the ensure
then adopted that live terminal and the resume was lost (a new
conversation). Seen 3/3 on golden 4020131de0.
- Release the pane's transport before the stop (split releasePaneForRestart
out of handleRestartPane), so nothing reattaches on its own.
- resumeRemotePaneAgent returns the ensured terminal; on 'adopted' it stops
that terminal and ensures once more, else fails (Couldn't restart).
- Reconnect only after the host lists that exact terminal live in the
pane's tab and leaf (bounded 15 s, else Couldn't restart).
- After reconnecting, the pane must bind that handle; any other handle
shows Couldn't restart. A failed restart reattaches the pane so it is
never left empty.
A cloud create whose agent launch is planned in the pane (held, or any create the
host didn't spawn) made two host terminals in the same second (real m4air probe,
W-137): the create's startup terminal (ensureWebRuntimeWorktreeTerminalAfterWake)
and a plain one from the session-tabs initial-terminal seeder, which saw the
host's empty first answer and had its own latch. The after-wake create now also
claims the seeder's bootstrap latch, released like the seeder's own (failed: at
once; created: when the host's row is mirrored).
A send with enter on a fresh Claude/agent-teams launch waits for the launch composer (up to 8 s). When the launched agent had already exited back to its shell (a turn-less --resume printing 'No conversation found'), the composer never came, so every Enter send waited 8 s and was then refused. The launch composer watch now also watches for the agent's exit (OSC 133 launch finished + shell in front, the same check readiness waits use) and settles at once so the send goes to the shell; the send's own wait also stops when its request is cancelled.
After a --resume that finds no conversation, the fallback typed the fresh
command with terminal.send { enter: true }. On the host an Enter send waits
for the launched Claude's composer (watchLaunchComposer, 8 s default),
which never comes once Claude exited to the shell, and is then refused:
the 5 s request timed out and nothing was written, leaving a bare shell.
- Send the Enter inside the text (same bytes, no composer gate).
- Treat a refused send as a failed try; 20 s timeout, up to 3 tries, and
re-read the tail before each retry so nothing is typed twice.
- If every try fails, show Couldn't restart (hostUnresponsive) so the
person can Try again, instead of leaving the pane quiet.
In the create-with-prompt flow the agent's tab reaches the store after the
held placeholder's close resolves, so focusAgentTab found no tab and did
nothing; the placeholder's neighbour (a plain Terminal) stayed active.
Wait up to 10 s for the agent's tab, and never take focus if the person
picks another tab meanwhile.
Restart stops the pane's PTY (stopExact keepHistory) and asks the host to ensureAgentSession with the pane's provider session and tab/leaf. The stopped PTY still owned the leaf, so the host adopted it and woke it with its original argv: no --resume, a new conversation. When a launch carries a provider session and the leaf's PTY is not running, the host now names it as the stopped owner, so the leaf spawns fresh with the resume command. Live panes are still adopted; plain reconnects (no provider session) still wake the original.
Traced the Orca Cloud create path: sparse checkout IS honoured (cloud provisions in
'orca-worktree' mode; the sparse directories go to the sandbox's own git worktree
create), so Sparse/Full checkout stays. Parent worktree is NOT: every new cloud
workspace is on its own sandbox host and an explicit cross-host parent fails the
create (enforceCreateLineageHostBoundary). So with Run on = Orca Cloud the card
passes no parent handler (the shared AdvancedSection then shows no picker, no
change to it), and a parent picked before switching to Orca Cloud is not sent.
The decision lives in run-on/cloud-advanced-options.ts. Fixtures:
NewWorkspaceComposerAdvancedSection advancedLocal / advancedOrcaCloud.
DS-b: the in-app stepper now lists the web's four steps (stablyai/orca-cloud#730):
Team always follows Account (done once signed in, as the web ticks a team made for
you; a team switch keeps its own step), and the plan step is always "Plan" (a
trial is how it gets done, not a step). GitHub ticks when the repository is
readable, unchanged. NS-04 (e): the agent's glyph sits before its name in the
header context line, muted (currentColor, never the coloured mark), aria-hidden.
NS-04 (c) needs no change: the only skip line was the Agent step's italic aside,
and that step no longer renders. Fixtures: CloudSetupPage signedOut, trialPlan,
githubLeft.
design-squad: at 900 wide with both side panels open, Add project / Create
workspace stacked as two large buttons over the cloud list. The row is now
LandingActions (landing-layout.tsx); over the cloud list its column is an
@container and below 600px both buttons take small sizing on one row (wrapping
only if they still can't fit). Fixtures: landing-layout actionsOverCloudList
Wide/Narrow.
Golden source builds publish their orcad + in-VM CLI as a commit-pinned bundle (R2, uploaded by the build VM over a signed PUT; the Worker computes the stored sha256). On every boot and wake, before orcad starts and inside the existing orcad phase, a sandbox whose /opt/orca/runtime/orcad/SOURCE_COMMIT differs from its region's active golden fetches that bundle from the control plane's own host (already on the egress allowlist) with a 10-minute signed URL, verifies the sha256 the control plane passes in, checks the commit and node runtime, and swaps orcad + orca-cli atomically. Any failure keeps the old orcad and boot continues; every outcome is logged (runtime_refresh: outcome, from, to, ms). RUNTIME_REFRESH=off is the kill switch. Dev R2 bucket orca-sandbox-runtime-dev.