The shared Mac hit load 60 from parallel premerges and vitest runs, starving the e2e harness. In the merged copy
every branch re-runs (#447), after the merge and before install/typecheck/lint/tests, the script now takes one of
PREMERGE_MAX_CONCURRENT (default 2) flock'd slot files under PREMERGE_LOCK_DIR (default /tmp/vdrv), held by a python
parent for the run (a killed run frees its slot). Queued runs print 'premerge: queued for one of 2 slots …, waited Ns'
every 30 s and 'got slot i/2 after Ns'. Tested: 3 concurrent runs with a limit of 2, the third queued 10 s.
Two defects the real-app proof found:
- ~400 ms blank pane on attach: the woken terminal's fresh-spawn viewport
blanking ran before the snapshot drain. A pane with pre-sleep output to put
back now holds rendering (DEC 2026) there instead of blanking; it keeps its
greyed history until the restore rewrites it.
- History cut to ~880 rows: by "asleep" the transport had resubscribed and
replaced the buffer with the host's 1000-row snapshot, and the capture could
re-run on it. It now captures when sleep is asked for (going to sleep), once
per sleep (a later look only refines the start time), and arms only after
the workspace was actually asleep.
Sleep reboots the VM; the woken pane attaches to a new terminal whose snapshot
replay erases scrollback (`\x1b[3J`), so everything the pane showed was gone.
- Capture (asleep/waking, before the new terminal attaches): the pane's normal
buffer rendered from its cells as text + SGR only (never its escape stream),
capped at terminalScrollbackRows. Shell panes keep their visible rows too;
agent / full-screen panes (the #451 rule) keep only the scrollback above, so
no frozen TUI frame comes back. No live buffer (never mounted, restarted
while asleep): the plain-text read-only copy saved at sleep, uncoloured; none
at all: nothing restored, nothing said.
- Restore (once the workspace left "asleep"; the woken terminal's snapshot
drain): after the snapshot paints, a client-side rewrite of restored output,
the dim divider, then the woken screen, cursor put back with relative moves.
Nothing reaches the PTY (replay writer, replay guard); a final sanitizer keeps
only text and CSI…m (OSC 52/0, DECSET, bracketed paste, cursor moves dropped).
Pinned to the bottom unless the person had scrolled up. One divider per sleep.
- The kept output is dropped once restored, or when its pane / tab / workspace
goes away.
On cut 68321452c9 a shell pane running `sleep 900` woke with no "Woke up"
bar. The command printed nothing, so its prompt line was the last text and the
detector read it as idle / half-typed (a rule from the first JA-15 detector).
The capture now notes whether the cursor sat below the last text: Enter was
pressed on that prompt line, so the silent command is running and gets Rerun. A
half-typed line (cursor still on it) and a bare prompt still offer nothing; agent
launches and the orca-launch script are still never offered.
#402's fallback sent the raw settings.agentDefaultArgs string as agentArgs for a
pane with nothing recorded. The host takes explicit agentArgs verbatim and only
projects its launch profile (configured Arguments + permission choice) when none
are sent, so a pane launched as 'claude --dangerously-skip-permissions --model
opus --effort high' restarted as 'claude --model opus --effort high' (real-app,
pane's own process by marker, before/after). Project the defaults through
resolveTuiAgentLaunchRawArgs (the same projection a fresh launch uses), so the
permission choice is kept; with no default args nothing is sent, as before.
filter-base-list-tests.sh is the one place premerge drops base-list tests (logged to stderr). The test covers:
a base-list test is dropped and logged while an off-list test is still selected; shared prefix/suffix/folder
never skips; entries are literal (no regex) and comments are ignored; a missing list keeps everything.
After a wake the host relaunches the pane's agent itself (WAKE-1, #450, ~7 s),
so #449's "Restart Claude" would sit over a live, relaunched Claude (K3).
On an agent pane the woke-up strip now:
- disappears for good once the agent reports in after the sleep (hook
evidenceObservedAt newer than the asleep capture; never updatedAt, which a
reconnect replay restamps);
- shows nothing for a 15 s grace;
- then offers "Restart <agent>" only over a bare shell (e.g. "No conversation
found"); a screen that may be the agent shows nothing and is re-checked on
each terminal draw (no polling) for up to 60 s after the wake, then the hint
lets go.
Codex is relaunched by the same host path, so the same rule applies. Shell panes
are unchanged.
w6's #450 premerge pulled in 4 agent-hooks tests through the importer selection; all 4 fail identically on
integration without the change (and on upstream at the merge base) under run-vitest.mjs, so premerge failed
on environment-only reds. The base list (41 files, derived on upstream e295b0d37c) now lives in the repo
and is subtracted from the targeted run.
After a wake, a Claude pane's "Woke up · processes stopped while asleep" strip
offered `Rerun Try "create a util logging.py that..."`: Claude's input
placeholder, scraped as the last prompt line.
- Shell-command detection rejects TUI chrome (box-drawing rows), bare ">"/"›"
input rows, agent launch lines and Orca's self-deleting
`. /tmp/orca-launch-*.sh` script. Alternate-screen panes get no hint.
- A Claude/Codex pane (agent status row, not back at a shell prompt) offers
"Restart Claude" through the existing RA-11 restart path (#429/#425: resume,
fresh fallback on "No conversation found"), factored into one shared helper
used by the sign-in strip too. Interim until WAKE-1 relaunches agents on wake.
- Shell panes keep Rerun with their real command.
design-squad: the real shot offered 'Add a GitHub repository' while Hello-World (a GitHub project) was in the picker. hasRepos counted only GitHub repositories with no clone here; it now also counts added git projects with a GitHub remote (or one not read yet).
A Claude pane that never took a turn has no transcript, so the host's wake relaunch (and an explicit ensureAgentSession resume) ran claude --resume <id>, which printed 'No conversation found' and left a bare shell. When the hook-reported transcript path is local to the host and not on disk, the host now launches claude --session-id <id> with the pane's recorded args (model, effort) instead. Same id, so resume claims, holds and readers still match. A session with a transcript, a non-Claude agent, an SSH host, or no recorded path keep the plain resume. #425 stays the desktop backstop.
A branch cut before #437 ran its own old copy of premerge-check.sh (the integration merge only happened in
the temp worktree), so #440's first gate selected 0 tests although the merged gate selects the parity test.
The script now merges, then execs the merged copy; it also prints every targeted test file.
design-squad (c): nothing switches on its own. One line under Project ('Orca Cloud needs a GitHub repository.') with 'Choose repository' (opens the Project picker at its GitHub repositories) or 'Add a GitHub repository' when there are none. Create is disabled with the same reason (title + aria-describedby), keyboard submit too, so the create can't fail with 'needs a GitHub repository'. Picking a GitHub repository clears it.
structured-worker-cli 'keeps a refused first task inspectable' armed a global
'reject the next turn/start' right after 'agent launch' returned. launch returns
at input_accepted, before the coordinator's own first turn reaches the agent,
and #191's pre-turn snapshot widened that gap, so the coordinator's turn
sometimes consumed the rejection and the worker's task was accepted. Product
refusal was never lost: the logged turn/start order shows the coordinator
refused instead. The fixture now refuses only the turn carrying the task text.
Checkpoints also discover the repo once per folder (not-a-repo rechecked after
60 s, cache dropped on any capture failure), so later sends add no git spawn
before the turn starts. 20/20 runs green (8/8 tests each).