The fork button on a stopped assistant message was inert. Frozen
interrupted nodes carry a flow-ordering seq of turnEnd.seq - 0.9, and
session.fork takes a non-negative integer on the wire, so every such
request was rejected as invalid-params before reaching the host — where
an aborted turn's logged turn/end has always made it forkable.
SessionsService.fork floors atSeq at the wire boundary. Flooring stays
inside the anchor's own turn (every turn opens with turn/start), so the
host's first-turn/end-at-or-after cut still closes on that turn.
The PTY capture does continue past the terminal-takeover bytes with the
fatal diagnostic; only the reset never follows. State that precisely in
both notes.
Document on FailLoudProcess.exit that callers treat it as the end of the
run, matching how the release path already relies on it.
Review of the previous commit found two defects in the release path, both
reproduced against the implementation:
- The timeout guarding a never-settling release was unref'ed. An
unhandledRejection listener suppresses Node's default fatal exit, so with
nothing else referenced the process reached an empty event loop and exited
0 on the very failure it was reporting. Keep the timer referenced and clear
it once the race settles.
- The handler uninstalled itself before awaiting the release. A second
concurrent rejection then became uncaught and killed the process
mid-teardown, stranding exactly the terminal state this restores. Replace
the uninstall with a latch: the first rejection is the reported one, and
later rejections (teardown's own included) fall through to the pending exit.
Add the PTY regression the fake-process tests cannot express: boot the shipped
tree over a fixture whose llm-pi-ai providers value is list-shaped, expect exit
1, and assert the captured bytes carry both the diagnostic and ESC[?2004l.
Against the pre-fix source the stream ends at ESC[?2004h ESC[>7u ESC[?u ESC[c
with no reset and the case fails, so it pins the actual bug.
Split the two-shape formatting test into one install per case; a latched
handler reports once by design.
A dsh launch whose config failed validation returned the user to a broken
shell: typing was invisible and the next command was mangled by a stray
Device Attributes reply (1;2;4cecho ...).
The Loader mounts entries concurrently, so ui-tui can already hold the
terminal (raw mode, bracketed paste, keyboard protocol, plus an in-flight
DA query) when a sibling entry rejects on its own config. installFailLoud
wrote its diagnostic and exited immediately, so nothing disposed the tree
and ProcessTerminal.stop() never ran.
Give installFailLoud an optional release teardown, awaited between the
diagnostic and the exit and bounded by FAIL_LOUD_RELEASE_TIMEOUT_MS. The
TUI launcher passes one that disposes the root context, reaching the same
shutdown() the /exit path already uses (drainInput() + ui.stop()). The
context is captured in boot()'s prepare hook because the rejection arrives
while boot() is still in flight.
Bins that pass no release keep the previous behavior exactly.
Review findings from ds-review-bot: closing the loading picker now
aborts the scan through the AbortSignal both query methods accept, a
signal-ignoring backend's late settlement is dropped by a staleness
check, one catch spans listing and projection so a projection failure
closes the overlay instead of stranding the loading placeholder,
setCandidates clears a stale still-loading error, and the batch
comment no longer overstates the win as scaling with session count.
ConversationSession seeds a stored draft in its own mount effect, and a parent's mount effect runs after its children's — so the unlock effect measured an empty mirror and never ran again for the draft that then appeared, leaving a restored long draft showing its head with the caret at its end. The effect now depends on the draft being non-empty; clearing on send and typing the first character flip it too, where both the focus and the reveal are no-ops. Pre-existing (the old geometry did not scroll for a programmatic value change either), fixed here because the reveal now exists.
Also from review: the golden's paste goes back to a block NOT ending in a newline, so the collapsed branch keeps a real engine under it while the standalone case owns the after-newline branch; the shared line-height rule names the reveal as its third consumer.
This PR moved the 14-line cap declaration off the auto-grow mirror and onto the box that scrolls the draft, and retargeted the approval scenario's cap probe from the textarea to that box. The sibling note stated both as they were; implemented notes have to state what shipped, in the change that ships it.
A caret straight after a newline is on a line with nothing to measure — the shape a trailing-newline draft ends in. chromium returns no client rects at all for the collapsed position (an all-zero box, which sent the reveal upward instead), firefox reports the line above, WebKit the right one. Measure the newline the caret just left and step one line down: all three then land on 649 of 652 with the caret's line at 315 inside the 336px box. The browser case now pastes a newline-terminated block, and fails 'expected 0 to be greater than 0' without the rule.
The selector overlay opens as soon as the command dispatches: the
picker renders a loading placeholder over an undefined candidate set,
owns terminal input from its first frame, answers Enter with a
still-loading error, and cancels on Escape exactly like the loaded
list. The finished scan swaps rows in through setCandidates without
replacing the overlay; a scan failure closes it and keeps the existing
notice.
preventScroll stopped the browser from revealing a focus nobody gestured for, and took the caret with it. The composer DOM is reused across sessions, so switching to a longer draft keeps the previous offset while the value swap puts the caret at the new draft's end: measured on all three engines, 940px below a box sitting at 0, where the old geometry reached the caret at 628. The unlock effect now suppresses the browser's walk and reveals the caret in its own scrollport, landing at 625 of 628. The focus-keeping mousedown gets preventScroll too and needs no reveal. revealCaret returns early when the box cannot scroll: the draft fits, so every caret is already in view.
The selector called readSession per listed session under an unbounded
Promise.all: each call re-listed the whole persistence store (O(N^2)
listings), decompressed and parsed the complete log, replay-validated
every event, and deep-cloned it up to three times, only to derive one
row's title, activity time, turn label, route, and goal phase. On a
real 185-session / 87 MB store the selector took tens of seconds.
Candidate rows now come from one projectSessions batch over borrowed
logs; a rejected projection degrades to the same disabled unreadable
row. Preflight still replay-validates the single chosen session through
readSession, which is already live-preferred, so its redundant live
shortcut is gone.
Pasting a long block left the view where it was while the caret sat at the
end of what was pasted. Paste, ctrl/meta-Enter newline and cut all suppress
the native edit — the machine owns the draft and the undo log — and restore
the caret with `setSelectionRange`, which reveals nothing: measured in
chromium and WebKit, before this branch as well as on it. Firefox happened to
reveal it, in the old geometry only.
The three restores now share one helper that measures the caret against the
hidden mirror — same draft, same metrics, same wrap width, so a Range
collapsed at the caret's index reports where the caret is without a caret API
— and scrolls the scrollport the minimum that brings the line inside, which
is what the browser does for typing. One scrollport is what makes this
possible at all: the reveal is finally a single offset to move.
Also from review: the composer's own focus() on unlock and session switch
passes preventScroll, so a session switch cannot move the transcript through
the taller textarea's reveal chain.
The composer paints its draft in two layers — the textarea owns the value,
the selection and the caret, the backdrop paints every visible glyph — and
they had one scroll offset each, kept equal by a `scroll` listener. That
holds at rest and not in motion: a wheel gesture scrolls the textarea on the
compositor, the listener runs afterwards, and for those frames the caret sits
at the new offset with the words at the old one. Measured on a harness of the
same geometry, a 200px offset change separates caret from glyphs by ~200px
(chromium 203, firefox 202, WebKit 203) until a later frame — the caret
flying out of its own text when a user swipes a long draft quickly.
Both layers now ride one scrollport: `[data-input-scroll]` carries the
14-line cap, the auto-grow stack inside it is as tall as the whole draft, and
the textarea holds no scrollable overflow of its own. The browser applies one
offset to both layers in the same frame, so the coupling is structural rather
than maintained. The backdrop's trailing-line sentinel and the cross-engine
wrap-width premise go with the mirror: the layers now share a containing
block, which closes the WebKit 768-against-776 gap by construction.
Review follow-up: the llm/adapters-updated listener's disposer was
discarded, leaving it firing (harmlessly, behind isDisposed()) between
TUI shutdown and fiber disposal, asymmetric with the sibling channel
listeners. The controller now exposes detach(), and the channel's
detachListeners() calls it on both the dispose() and startup-failure
paths.
