starred/orca

Fork 0

mirror of https://github.com/stablyai/orca.git synced 2026-09-21 16:02:20 +00:00

Files

T

NeilandOrca 17b3dff3c4 refactor(terminal): return IME composition ownership to xterm (#13128 )

* fix(terminal): return IME composition ownership to xterm

* fix(mobile): derive terminal input from native replacement ranges

* test(mobile): record iOS Japanese IME traces

* fix(mobile): preserve native IME replacement ranges

* fix(xterm): flush queued application input after IME commit

* test(terminal): pin Korean intermediate commit

* test: pin Windows IME shortcut ownership

* test: replay IBus number candidate commit

* fix: preserve native macOS input-method punctuation

* refactor(terminal): remove stale mac focus override

* fix(mobile): preserve soft keyboard deletion ranges

* fix: keep IME-owned palette chords in renderer

* fix: stop carried IME shortcuts at renderer owner

* fix: preserve carried IME shortcut dispatch

* fix: narrow main-owned shortcut actions

* test(mobile): pin Japanese IME replacement traces

* test(terminal): retain paired native IME trace

* fix(chat): preserve browser IME composition ownership

* fix(chat): retain macOS IME confirm gesture

* fix(chat): expire unmatched IME confirm carry

* fix(chat): isolate IME confirmation expiry

* fix(chat): retain active IME confirmation

* refactor(terminal): remove dead composition handler

* feat(ime): add shared Enter-ownership seams for CJK composition

The confirming Enter of a CJK composition arrives as two keydowns and the
orderings differ by platform: Windows/Linux redispatch the unmarked Enter/13
before keyup, macOS delivers keyup first. A guard reading only isComposing or
keyCode 229 misses the redispatch, so surfaces submitted on a confirm.

Adds useImeEnterGestureOwnership (carry token, next-frame expiry), a shared
ImeEnterGuardedForm for native implicit submission, and the cmdk seam covering
18 CommandInput surfaces at one site.

A chorded Enter arms the carry but is never swallowed — the reverse would eat a
user's deliberate Cmd/Ctrl+Enter. Both failure modes are pinned by
ime-enter-gesture-ownership-contract.test.ts.

Co-authored-by: Orca <help@stably.ai>

* refactor(terminal): consolidate native input listeners and parked-screen owner

Extracts the shared native-input listener installer and renames the parked-screen
detector for what it actually does, replacing per-call-site duplication. The
listener installer keeps a forgetOptionKeyLocationOnBlur flag so per-window
semantics are preserved rather than flattened.

Net deletion; no behaviour change intended.

Co-authored-by: Orca <help@stably.ai>

* test(terminal): pin recorded IME shapes as regression tests

Nine regression tests built from hashed affected-platform captures, each with a
paired ordinary negative and a discriminating mutation verified to take the file
from all-passing to exactly one failure.

Covers the Windows MS-Korean Shift family (#12179, #11878, #12151, #11946,
#12152) and the Korean TUI line-break rows (STA-3237, STA-3222, STA-3129).

STA-3237 pins the empirical 3-Shift / 2-active-composition / 2-newline ratio the
device run established — the third Shift produces nothing because Space has
already committed. That ratio is not derivable from a static capture.

Co-authored-by: Orca <help@stably.ai>

* fix(ime): guard Enter-commit surfaces against CJK confirm

Applies the Enter-ownership guards across the surfaces whose Enter commits
something: publishes, clones, pairs, installs, posts, or persists.

Tiered deliberately rather than uniformly. Irreversible and remote-effect sites
take the carry token, which also blocks the unmarked redispatch. Locally
reversible sites take the oracle check with a one-line comment naming the
residual, because a spurious commit there costs one undo.

Three numeric fields are left unguarded with the reason in-code: Chromium blanks
number inputs at compositionstart, so a confirm-Enter only ever reaches an
empty-draft reset. Measured with a CDP probe rather than assumed — a guard that
cannot fire is noise.

Co-authored-by: Orca <help@stably.ai>

* test(ime): teeth-check the Enter guards on every guarded surface

One suite per guarded surface, each verified by deleting the guard and
confirming the test fails. A green guard test without that check is unverified,
not verified.

Two shapes pass vacuously in happy-dom and are avoided here: native implicit
form submission never fires, and blur() is inert on an unfocused element. Both
made "the commit did not happen" assertions pass with the guard removed, so the
suites assert the guard's contract directly instead.

Co-authored-by: Orca <help@stably.ai>

* fix(mobile): keep iOS Korean commits whole through the live-input path

iOS Korean reports isComposing: false on every event, so it bypasses the
composition guard entirely. The strict owner rejected UIKit's transformed
post-change field and sent only the leading jamo — the reported symptom.

Prefers the authoritative same-event field text over the predicted text when the
supplied operation cannot produce it. Generic: no Korean special-case, no locale
classifier, no normalization. Adds the RN-target-keyed submit carry alongside it.

Co-authored-by: Orca <help@stably.ai>

* test(e2e): make IME capture harnesses fail loudly instead of silently

Four instruments recorded silence as success, so a void run scored as a clean
one:

- readTerminalImeBoundaryTrace returned an empty trace when the probe never
installed, making every "nothing leaked" negative pass vacuously
- summarizeLatencies([]) returned a perfect zero distribution that passed all
three latency thresholds
- the macOS Vietnamese spec pinned an input-source ID that does not exist, and
failed as though the operator had chosen the wrong source
- the expectedLineCount=1 prefix property was undocumented and one edit from
silently downgrading a PTY assertion

Input sources now resolve by enumeration and name the near-matches on failure.

Co-authored-by: Orca <help@stably.ai>

* test(terminal): cover Cangjie cancellation and fix a cross-namespace assertion

Adds #11951's recorded Cangjie cancel shape to the existing cancellation suite,
which covered Pinyin and Sogou but not Cangjie. One keystroke then Backspace
arriving as deleteContentBackward with data: null, so the stale preedit is the
only thing a fallback could replay.

Verified against the historical pre-6cd944c62b3 bundle: the positive fails with
['尸'] where [] is expected, while the ordinary negative stays green.

Also fixes the Vietnamese spec, which asserted a TIS-space input-source ID
against getKeyboardInputSourceId(). Those two Orca APIs report the same source
in different namespaces — TIS nests it under VietnameseIM, the app API does not.
The resolver stays as an installation precondition; the assertion matches the
leaf.

Co-authored-by: Orca <help@stably.ai>

* test(e2e): add a real-IME macOS arm for the Korean chord commit

The existing korean-ime-terminal-shift-enter-commit spec synthesizes composition
over CDP: Input.imeSetComposition sets the preedit directly and Input.insertText
performs the commit. Asserting the IME produced events you injected yourself is
circular, so that spec cannot certify real-IME behaviour.

This arm selects 2-Set Korean via TIS, reads it back live, and injects through
System Events key codes, so the OS owns the preedit, the commit instant, and
isComposing. PTY byte expectations are preserved verbatim.

Covers 2 of the original 4 cases by design. The other two are the Windows/Linux
redispatch-before-keyup ordering, which macOS cannot produce and which cannot be
selected -- the OS decides it. Reintroducing synthesis to "restore coverage"
would reintroduce the circularity.

Co-authored-by: Orca <help@stably.ai>

* test(e2e): assert the macOS chord arm at the PTY boundary, not the renderer

The byte expectations were transcribed from korean-ime-terminal-shift-enter-commit
:364/:383, which assert against onData -- a renderer boundary where the terminator
is CR. This spec reads the PTY child, where the tty has already converted CR to LF.

Names both forms per row rather than swapping the constant, so the conversion reads
as evidence that the capture reached past the renderer, as #11936 and #11951 record.
Ctrl+Enter's CSI-u sequence is unaffected and is identical at both boundaries.

Co-authored-by: Orca <help@stably.ai>

* test(e2e): measure composer-to-onData latency and stop dropping IME keystrokes

Two defects in the echo latency probe.

It hooked onWriteParsed and onRender but never onData, so it measured
key->parse->render echo rather than the composer-vs-onData delta the latency rows
need. Adds a third hook feeding its own sample set.

And `event.key.length !== 1` silently dropped IME keystrokes: Pinyin and Cangjie
keydowns arrive as key:'Process' (length 7). Replayed over the captured corpus,
the old filter accepted 580 of 4137 Chinese IME keydowns -- it was discarding 80%
of them. The new filter matches the shape the owner itself branches on.

Attribution charges each onData to the latest keydown rather than a FIFO head,
because composing jamo emit no onData at all and a queue would credit a whole
composition to its first keystroke. The consumer now asserts sample count before
any percentile, so a zero-sample run cannot render as a flawless distribution.

Co-authored-by: Orca <help@stably.ai>

* test(terminal): pin the WSL shifted-jamo newline shape for #11919

In Korean 2-set, Shift types ordinary letters -- the double consonants and the
compound vowels. Each such keystroke reaches Chromium as key='Process',
keyCode=229, shiftKey=true.

The v1.4.163 classifier matched exactly that pattern with no code guard, so it
called those keystrokes Enter, rewrote them to a synthetic Shift+Enter, and
injected a newline into the middle of the word -- with no Enter key pressed.
That is why the reporters said "no modifier key pressed": they had not chorded
Shift+Enter, but they had pressed Shift, to type the double consonant.

Asserts the row's own recorded capture: 40 immediate keydowns, exactly 3 of them
Shift-carrying inside a single syllable, and an onData stream with one newline
per Enter press and none mid-word. Two ordinary negatives keep it from being a
blanket mute -- the same session's non-IME keydowns still reach shortcut policy,
and an ordinary Shift+Enter still resolves through the real policy.

Co-authored-by: Orca <help@stably.ai>

* test(terminal): pin the composition commit lag that made Korean type one behind

macOS Korean 2-Set commits syllable N only when the first jamo of N+1 arrives, so
compositionend and compositionstart land in the same task. A composition-start
handler cancelled the pending finalizer that was the only path to triggerDataEvent
and ended the session without emitting bytes, so every committed syllable reached
onData exactly one syllable late and the backlog cleared only at a Space or Enter.

Types continuously with no Enter and no Space -- either would flush the backlog and
hide it -- and samples onData at every syllable boundary. Paired with a
length-matched ASCII arm that stays green throughout, so the positive is a fact
about composition rather than about timing in general.

Bisected to a single call site across five builds: pristine, 1.4.155 and 1.4.162
pass, 1.4.163 fails, removing the one call repairs it, restoring it fails
identically. That window is exactly the reporter's "started immediately after
updating".

Co-authored-by: Orca <help@stably.ai>

* test(mobile): cover the send-queue abort that silently drops queued keystrokes

One failed send in use-terminal-live-input-commit aborts every keystroke
queued behind it, with the error swallowed by .catch(() => false). The
existing test resolves(true) on every send, so the failure branch was
uncovered.

Four arms: the abort itself, an ordinary negative on the healthy path, a
throwing sender, and a liveness control proving the queue recovers once
the chain settles. Deleting the abort takes 4 passed to 3 failed, with the
ordinary negative correctly surviving.

Scope is stated in the docblock: this is a transport send-queue abort,
reachable only via a real disconnect or RPC error. REQUEST_TIMEOUT_MS is
30s, so latency alone cannot reach the branch — consistent with #7094's
symptom class, not proven to be its cause.

* test(terminal): pin that daemon snapshot/restore cannot disturb a composition

Two independent reporters attributed broken Korean composition to the
always-on PTY daemon repainting terminal state over the preedit. The
attribution is wrong on ancestry — the daemon shipped three months before
the version both call good — but the boundary was never actually tested.

Runs the real applyMainBufferSnapshot choreography against a live
composition, including the full 2J/3J/H wipe plus the resize and
alt-screen branches. textarea.value, selectionStart/End,
compositionView.textContent and .active all survive byte-identical, and
interleaving a restore between every jamo of 문제 still commits 문제 at
onData. Also pins that the uncommitted preedit is absent from the captured
snapshot: it lives in the textarea, never the buffer, so a restore has
nothing stale to echo back.

Injecting one textarea.value = '' into the restore fails exactly the three
restore-boundary tests.

* test(terminal): pin that Cmd tears down a composition where Ctrl and Shift do not

xterm's composition keydown exempts only keyCode 16/17/18 (Shift/Ctrl/Alt)
plus 20/229. macOS Meta — 91/93/224 — is absent, so a Cmd press mid-composition
takes _finalizeComposition(false): the overlay goes dark and never recovers,
because compositionstart is not re-fired. The user composes the rest of the
word blind. Linux and Windows users press Ctrl and are exempt.

xterm already has a Meta-aware modifier predicate in wasModifierKeyOnlyEvent,
so this is an internal inconsistency rather than a deliberate choice.

Owns no reported row and is version-neutral: 5/5 on both 1.4.162 and 1.4.163.
The branch is unexercised in all 328 recorded traces, so this is a hazard pin,
not a regression guard. Only the teardown is asserted; the likely duplicated
commit needs a compositionend the IME kept alive across the Cmd, which no
capture contains.

Deleting the exemption fails exactly the three paired negatives; adding Meta
to it fails exactly the two Cmd arms.

* test(native-chat): characterize preedit loss when a question card replaces the composer

An AskUserQuestion card fully replaces the composer by design, but the
in-flight composition goes with it: the composer unmounts before
compositionend reaches it, so the preedit is never committed to the draft.
The committed text survives only because the draft is cached and restored
via defaultValue. Node identity changes, value 'abc' is preserved, the 가
is gone.

Drives the real NativeChatView -> SessionGate -> InteractiveCard ->
questionActive swap -> Composer -> ComposerField, flipped by writing the
same store field an AskUserQuestion hook event writes. Flipping
questionActive to false fails exactly this test and nothing else across
639 native-chat tests, so the path was entirely unguarded.

CHARACTERIZATION TEST: it asserts the loss. Fixing the defect — committing
the preedit before the swap, or keeping the composer mounted — will make
this file fail. Update the expectations to the new contract rather than
working around them.

Owns no reported row. #12118/STA-3219 flicker is keyed to token counters,
which provably do not remount, and a question card arrives once per
question.

* test(terminal): pin the duplicated commit when Meta interrupts a composition

_finalizeComposition(false) sends textarea.value.substring(start, end) but
cannot clear the IME-owned textarea, so a later compositionend re-sends the
same range. Meta reaches that path because CompositionHelper exempts only
Shift/Ctrl/Alt; xterm's own wasModifierKeyOnlyEvent covers Meta four ways,
so the omission is an internal inconsistency rather than a choice.

Companion to the modifier-exemption guard, which deliberately pins only the
overlay teardown. This pins the data consequence.

HAZARD PIN: owns no reported row. The trigger is unverified on hardware —
no capture in the corpus contains a Meta-during-composition gesture, and
whether macOS keeps the composition alive across it is unmeasured. The
duplication follows from the code given that sequence; whether users reach
the sequence is the open half.

An earlier premise that Space (keyCode 32) reaches this path was refuted by
a corpus scan: 0 of 731 evidence files carry a keyCode-32 Space while
composing, against 171 at 229, and 229 returns early.

* test(terminal): characterize the syllable lost when the textarea blurs mid-composition

CoreBrowserTerminal._handleTextAreaBlur clears the helper textarea
unconditionally — "Text can safely be removed on blur" — while
CompositionHelper._finalizeComposition reads the committed text back out of
that same value from a deferred timeout. By the time it runs the value is
empty, the substring is '', and triggerDataEvent never sees the syllable.
xterm checks composition state in _syncTextArea and omits the same check
here.

Six cases. Blurring mid-composition loses the syllable in every ordering,
including compositionend-before-blur, which is Chromium's real order — so
it is not an ordering artifact. A bare textarea.blur() with no Orca code
loses it too, which places the owner upstream: Orca's unguarded release on
outside pointerdown is one trigger, not the cause. Committing 한 then
blurring mid-가 yields ['한'] where ['한','가'] is correct: one syllable
gone, surrounding text intact.

Teeth checked by inverting — adding an Orca-side composition guard flips
exactly the three cases that route through the release path and leaves the
bare-blur and no-blur cases green, which is the scope split: a fix in
regular-terminal-focus-ownership alone would not close this.

HAZARD PIN, but unlike the others this one has a real production injector —
clicking outside the terminal mid-composition. Owns no reported row. The
shape matches #9738's report; the injector does not, and a shape match with
a mismatched injector is not an owner.

* test(terminal): say which arm the STA-3237 fixture came from

The recorded keydowns are wave 4's A-shift-unmarked-only — the arm that
emits no PTY bytes. Nothing in the file said so, so two readers concluded
the row's events fail the owner's predicate and that STA-3237 and STA-3222
were different defects. They share an owner; the arm that fires is
Process/229+Shift, absent from this bubble-phase trace because the owner
claims it in the capture phase.

Also corrects "code-blind": the v1.4.163 policy emits \x1b\r only for a
shift-only key:'Enter', and a jamo keydown reaches that branch solely via
the isTerminalImeProcessEnter rewrite. The mock is deliberately wider so
the ownership guard stays under test if that rewrite moves.

Comments only — no assertion, fixture value, or mock behaviour changed.

* test(e2e): track the input-source selector the macOS specs shell out to

Five tracked macOS IME specs ran `swift .tmp/select-input-source.swift`, a
file that is gitignored and existed only on one machine. Anyone else
checking out the repo — or the same machine after .tmp is cleaned — could
not run them, and they are the capture drivers for the macOS rows that are
blocked waiting for exactly those runs.

Moves it to tests/e2e/ beside its callers. The chord spec now resolves it
from __dirname rather than reaching two levels up into .tmp.

* test(terminal): pin the CJK repaint decision against the reporter's own output

#12164 comment 1 and #5921 report agent output with double-width glyphs
rendering duplicated character-by-character while ASCII in the same line
stays clean. No IME, no composition, no keystroke — the user never types
the CJK.

Segmenting all three verbatim samples into maximal same-risk-class runs
gives 33 runs and zero violations of "this run is corrupted iff the
production detector flags it": 17 wide runs all corrupted, 16 narrow runs
all byte-identical. The paired negative is co-located in the same line
rather than in a separate run — the reporter supplied it without knowing.

Doubling is asserted as present, not uniform: 자바스크립트 and 시스템 each
leave a jamo undoubled, which is a repaint-region boundary artifact rather
than a per-character transform.

The discriminating arm is in the test rather than a source mutation:
be3f30e2f8 (#6890) elects a repaint for all 17 corrupted runs when the
agent types nothing, and reverting its disjunct elects none. Both
predicates agree once the user has recently typed, which is the pre-#6890
condition.

Samples inlined with per-sample SHA-256 because .tmp is gitignored and
cannot back a landed test.

* test(terminal): pin macOS period substitution landing after the composition

#11504's reporter published a DOM trace showing insertText ". " arriving
149ms after compositionend, when two spaces are typed with a CJK input
source and NSAutomaticPeriodSubstitutionEnabled is on. This replays that
trace against a real Terminal and asserts what reaches onData — bytes to
the PTY, not anything visual.

The owner is stock upstream CoreBrowserTerminal._inputEvent, not an Orca
module, confirmed at the resolved install and in the shipped bundle Vite
loads rather than in the TypeScript source.

Three mutations against that install, predictions written before the runs,
each failing exactly the arms predicted: dropping the composed/keyDownSeen
guard fails two, dropping Orca's intercept fails the one arm where the
payload arrives before the send window drains, and flipping || to && —
the candidate-fix shape — fails the arm that pins the defect itself.

CHARACTERIZATION: arm 1 asserts the broken behaviour and will fail the
moment #11504 is fixed. Update it to the new contract rather than working
around it.

composed is absent from every recorded bundle, so composed: true is the
spec-required value rather than a captured one; the test asserts it before
dispatching so a harness that dropped the field fails loudly.

* test(terminal): replay the recorded Windows Shift sessions through the IME guard

STA-3179 reports a Shift release sending Enter; #12171 reports delayed
Hangul plus doubled newlines. Both replay their own recorded Windows
MS-Korean keydowns through resolveTerminalKeyboardShortcutAction with the
shortcut policy mocked, so the assertions are about which events reach the
policy and what reaches terminal input.

STA-3179's held-Shift gesture yields exactly one newline, from the unmarked
Enter alone; its release arms nothing for the next composition, asserted
after a precondition check that the release really is keyups with shiftKey
already dropped; and an ordinary Shift press-and-release still routes every
keydown, which is the paired non-IME negative.

Teeth, verified by mutation: bypassing the isImeOwnedKeyboardEvent guard in
keyboard-handlers takes STA-3179 from 3 passed to 2 failed / 1 passed — the
survivor being the ordinary-session negative, which is correct, since a
non-IME session should not depend on that guard — and #12171 from 2 passed
to 2 failed. Source restored byte-identical.

Recorded shapes are inlined and the bundles cited in comments; nothing is
imported from .tmp, which is gitignored.

* test(native-chat): correct 61977d4517 — the preedit survives the question card

61977d4517 claimed a composed syllable vanishes silently when an
AskUserQuestion card replaces the composer, and characterized that loss.
The claim was false. Its premise was an artifact of the harness: the test
simulated a preedit with a silent textarea.value assignment and no input
event, which no IME does.

Real composition fires input with insertCompositionText on every keystroke
— the shape this repo already records in its own observed-event capture —
and React's change handler returns on input/change with no composition
gate, so onChange runs for each frame. The draft cache is written
synchronously inside the updater, so the preedit is already committed
before the card can arrive. Driven that way, it survives.

Renamed to match the contract that actually holds, and extended: Hangul
jamo-per-frame, Japanese kana accumulation followed by per-segment
conversion asserting the candidate the user was looking at survives, and a
pin on the mechanism itself — the draft cache holds the preedit while the
card is up.

Teeth: there is no fix to revert, so the mutation is the plausible wrong
one — gating onChange on isComposing(). That takes 4 passed to 3 failed,
with the English negative correctly surviving, since it has no composition
to gate.

Two consequences remain, recorded rather than fixed: the OS aborts the
composition when the field disappears, so a lone jamo returns as a
compatibility jamo the user cannot compose onto, and the remounted
composer is unfocused because the card owned focus.

* docs(native-chat): name the corrected commit and the degraded-jamo consequence

Records in the file itself that 61977d4517 is pushed and wrong, quoting
the two claims that are false, so a reader who finds it in git log reaches
the correction from the file that replaced it.

Also states the residual as a consequence rather than a curiosity: a lone
leading jamo returns as a standalone compatibility jamo (U+3131), which is
not a composable state — the user cannot resume the syllable, only delete
and retype. Preserved, but degraded into something unusable. That is the
note to find if a reporter ever describes exactly that.

The invariant these tests pin is not "the composer commits on unmount" but
"composition input events must reach React" — which is what a future IME
change would break, and is not visible from the swap site at all.

* test(terminal): replay the recorded macOS Telex commit boundaries

#6905 reports Vietnamese composed characters breaking in the terminal.
Replays the retained macOS built-in Simple Telex capture — recorded
selection and value set before each dispatch, since that is what the
commit range reads — and asserts what reaches onData: the first commit
alone, then through the real Enter, then the ASCII tail of the same run.
A code-point count would catch NFD normalisation.

ENGINE CAVEAT, stated first in the docblock: this is macOS built-in Simple
Telex, Telex only. The reporter's three named engines cannot run on the
platform they declared, and which macOS Vietnamese engine they used is
unconfirmed. This file certifies no engine, and does not imply VNI.

The owner is upstream's — CompositionHelper._finalizeComposition's
waitForPropagation branch — so the arms are copies under .tmp aliased by a
scratch config, with node_modules verified unchanged by shasum after every
run. Collapsing the range end onto its start fails all three; collapsing
the start to zero re-emits the first word into the second commit, which is
the reporter's "duplicated" direction. Different failure sets, so the
mutants are distinguishable rather than merely detectable, and the ASCII
assertion passes under both.

Falsifiability here is by mutation, not by a defective build: #6905 does
not reproduce at HEAD, so this has never been watched going red on a real
reproduction.

* docs(terminal): lead the #6905 test with its engine caveat

Comment-only. Moves the caveat above the source line so a reader meets what
the file does NOT establish before what it does — the capture is macOS
built-in Simple Telex, the reporter's named engines cannot run on the
platform they declared, and which engine they used is what gates this row.

Co-authored-by: Orca <help@stably.ai>

* docs(terminal): record that the swallow eats a keystroke after Japanese conversion

This pin framed the swallowed insertText around Cmd interrupting a
composition. A differential through Japanese multi-segment conversion shows
it is broader: type a segment, convert, then press `a`, and the `a` is
lost. No modifier, no exotic gesture. Korean surfaced it first only because
2-Set composes on nearly every keystroke.

Also records why it cannot simply be fixed. The suppression de-duplicates
IMEs that deliver their commit a task after compositionend, which a sibling
test pins; this swallow is that dedup's false positive, and the two events
differ only in payload, so no flag-timing change separates them. Both a
smaller redesign and a content-aware variant were built and measured — the
first duplicates on IBus, the second costs a reported row's test and is
blocked while the patch cannot be regenerated.

The Japanese arrays behind this are authored, not observed: no Japanese DOM
composition trace exists in the corpus.

* docs(terminal): a Japanese capture does exist — correcting dbeecb11be

That commit said no Japanese DOM composition trace exists in the corpus.
False. One does, filed under the Linux bundles rather than the bundle named
for Japanese: 30 DOM events, two にほんご->日本語 conversions, with full
selection state per event. It is retained byte-identically in three further
bundles — one capture copied four times, not four observations, checked by
hash rather than by counting files.

The claim came from checking the bundle named for Japanese, finding nothing,
and generalising to the corpus without querying the rest of it.

Replaying it emits 日本語日本語 under both sequencing extremes on all four
arms, matching its own recorded onData. So "repeated conversion is
undisturbed" is now captured rather than authored. It carries no
post-compositionend insertText, so it cannot speak to the swallow: the
a-after-conversion figure stays authored and unobserved.

Also rewords the paragraph opener. It claimed to broaden a Cmd framing, but
hazard 2 was never Cmd-framed — the lines above already say Cmd does not
reach it. The real gap was that hazard 2 named no trigger at all, which
reads as exotic when it is ordinary.

* build(xterm): land the patch regeneration harness

The five dependency patches under config/patches/ shipped with no tracked
way to regenerate any of them. The xterm one is the hard case: it is derived
from an upstream build, so no fix could be made without rebuilding, and the
tooling to rebuild lived only in one machine's scratch directory. That
blocked a measured fix for a live keystroke-loss bug, and the EditContext
reduction an OSS survey identified as the only real one available.

Adds the regenerator, the upstream pin, the hand-written source patch the
bundle hunks derive from, tests, docs, and a PR job that verifies the
shipped patches still match the pinned build. The job caches the shallow
clone keyed on the manifest, so a cold run is minutes and a warm one under
one. Round-trip verified: regenerating from a clean checkout reproduces the
shipped patch byte-for-byte.

Marks the emitted patch -diff -text. pnpm hashes it byte-for-byte, so a
CRLF checkout would break install on Windows, and its minified bundle lines
make a diff nobody can read — review the source patch instead.

Also rejects unknown flags. --check was the fallback for any unrecognised
argument, so a typo, or --help, silently triggered a full upstream build
instead of what the caller asked for.

* fix(xterm): stop swallowing a keystroke typed after an IME commit

Type a Japanese segment, convert it, then press a key one macrotask later
and that key was lost. No modifier, nothing exotic — every user who keeps
typing straight after converting. Korean surfaced it first only because
2-Set composes on nearly every keystroke.

handleCompositionInput discarded the payload unconditionally in the window
after the deferred send: _isSendingComposition stays true for one macrotask
after the timer cleared _pendingCompositionStart, and the branch substituted
'' for whatever arrived. The suppression is not itself wrong — it
de-duplicates IMEs that deliver their commit an event-loop turn late, which
terminal-stock-composition.test.ts pins. It just could not tell a duplicate
from new input, because the two events are identical apart from payload.

Now it compares against _sentComposition, the text the deferred send
actually emitted, and discards only a match. A flag-timing redesign was
measured first and rejected: it fixed this and duplicated on IBus, because
no timing change can separate events that differ only in content.

Edited in config/patches/xterm-src/ and regenerated through the harness, so
the emitted patch and the lockfile hash are derived, not hand-written.

The commit-overlap pin's swallow arm now asserts the repaired contract —
the value its own comment already named as correct and as what stock
beta.287 emits. #11504's arm at :184 flips too; it never covered that
report, as its own prior note recorded, and the reporter's +149ms arm is
untouched and still asserting the defect. Provenance hashes in three test
docblocks are updated, since regenerating changes the patch hash and with it
the resolved install directory.

* docs(terminal): re-measure the #6905 mutation citations against the new bundle

Regenerating the patch moved the resolved install, so this docblock's
patch_hash, line count, two line numbers and three mutation outcomes all
described a bundle that no longer exists. The deferred branch is one the
fix writes into, so the outcomes could not be re-pointed on reasoning.

Line numbers read off both files by diffing anchors rather than derived by
arithmetic: 201 to 205, 159 to 163. Outcomes re-run through the retained
rig, which re-resolves through the module loader and re-derives each arm
from a unique minified anchor: pristine 3 passed, m1 3 failed, m2 2 failed,
m3 3 passed — identical to the old bundle. Guard controls in both
directions exit 1, so the counts are falsifiable.

Comment-only; the assertions and expectations are unchanged.

* fix(xterm): size the preedit overlay to the cells its text will occupy

updateCompositionElements computed the overlay's left edge from the grid
but never its width, so the preedit rendered at the font's natural advance
while the committed text takes two cells per wide glyph. Measured in
Chromium 150: 가나다라 drew 48.45px as a preedit and 69.20px once committed
— the same characters, same font, 30% narrower, and drifting further with
each syllable. Every macOS mono font carrying Hangul measured 0.49–0.72 of
two cells; never 1.0.

Deriving the width from wcwidth and the cell measure moves Korean, Japanese
and Chinese to 1.000 and leaves ASCII at 1.000, which it already was:

한 12.125 -> 17.297 (17.30 expected)
가나다라 48.453 -> 69.188 (69.20)
안녕하세요 60.563 -> 86.500 (86.50)
日本語 42.000 -> 51.906 (51.90)
abcdefgh 69.234 -> 69.203 (69.20, unchanged)

Edited in config/patches/xterm-src/ and regenerated through the harness, so
the emitted patch and lockfile hash are derived rather than hand-written.

The unit test asserts the arithmetic, which is what CI can run. The pixel
consequence was measured on macOS with SF Mono in an Electron harness, not
on the Windows font stack STA-3232 reports from — so this demonstrates the
mechanism and does not stand as that row's platform evidence.

* test(e2e): pin the macOS Korean preedit as visible only while composing

#11914 reports the composing text invisible until Space. Its c3 was recorded
as unobtainable, and the reason on file was wrong: the boundary IS
assertable, but not in happy-dom, which reports display:block in BOTH the
active and inactive states and zeros for every rect. A test there passes
with the defect present.

Captured on real hardware instead: hidden and 0x0 before, .active with
display:block, a 15.84x16 rect and checkVisibility() true while composing
그, hidden again after. 39 DOM events, 2 composition starts, onData
["한","그","\r"].

Two mechanism findings are carried in the setup because both are invisible
in the result and fatal if removed. The input source must be selected AFTER
the app takes focus — focusing resets it to ABC. And the IME must be warmed
until an observed keyCode 229; typed cold it emits raw QWERTY (g k s r m)
with no composition at all, which is indistinguishable from an IME that is
not installed. Two runs were voided on exactly that signature before the
warm-up was found.

The has229 and compositionStarts assertions exist to make such a run fail
loudly rather than pass as a clean negative.

Gated on darwin plus ORCA_E2E_NATIVE_MACOS_KOREAN, like its siblings. The
final spec form has not itself been executed — the machine became
unavailable — so it carries the probe's measured values as literals rather
than a run of its own.

* docs(e2e): correct 19a8d133db — the Korean preedit spec has been executed

That commit said the landed form had never run and carried the probe's
values as literals. It has now run on real hardware: 1 passed, 9.1s, rc=0,
with the capture and log sealed under a verified hash manifest.

The teeth check was also run rather than reasoned about, and it changes
which assertion matters. Forcing the active overlay to max-width:0 with
overflow:hidden — invisible on screen — leaves the active class, the
textContent, display:block AND checkVisibility() all passing. Only
during.rect.width fails. checkVisibility() is not sufficient against this
defect; the bounding rect is the single load-bearing assertion, which the
docblock already said and this run confirms.

An earlier teeth attempt injected the CSS mid-run and tripped the
hasActiveClass poll instead, failing at the wrong assertion. It is
inconclusive and excluded from the seal rather than counted.

* test(terminal): add #12171's ordinary-English arm from a real Windows capture

c4 was recorded as unmet and the ledger sourced its control to
evidence/windows-current/, which holds 12 captures and not one English one.
The arm here comes from windows-9803-final instead — same probe, same host
geometry, same injector, en-US with no IME, replayed keydown for keydown.

Two limits are stated in the file rather than left for a reader to find. It
is a different bundle and a different run about 3.6 hours later, so it is
not a same-run arm. And it is #9803's range-active MUTANT arm: ordinary
English stays byte-exact even with that saved-range mutation live, which is
why it reads as a negative rather than as a baseline.

Bundle cited by directory with its file SHA-256; MANIFEST.sha256 verifies
21/21, rc=0. Nothing imported from .tmp.

* docs(terminal): correct #12164's grounds — the cited comments say no such thing

The rejection of #12164 from this file's family was recorded as resting on its
comment 1 (output doubling) and comment 2 (filed against 1.4.163). Checked
against the API: the issue has exactly two comments, neither of which says
that, and the string 1.4.163 appears nowhere in the thread.

The conclusion survives on better grounds. The issue BODY's repro is "Run any
CLI agent (Codex, AGY, Claude, etc.) that outputs Korean text into the Orca
terminal" — untyped output, no keystrokes, no composition — so excluding
CompositionHelper is right, and the input-path hunt was looking in the wrong
place. The body is also LLM-authored (it still contains a literal
"## 5. GitHub Submission Draft (Ready to Post)") and its Root Cause section
blames a CJK IME preedit buffer its own repro never engages, so it should not
be read as observation.

Comment-only; suite unchanged at 5/5.

* test(native-chat): make composition frames carry isComposing, not just inputType

This suite's comment claimed "Gating onChange on `isComposing` breaks here."
It did not. composeFrame() fired `input` with `inputType` but never set
`isComposing`, so a gate on `isComposing` passed all four tests untouched —
the suite asserted a discriminator it did not exercise.

Composition frames now carry both, so neither gate is exempt. Verified by
pointing the mutant at it: with an `isComposing` gate on the composer's
onChange, this suite now fails 3 of 4 (it passed 4 of 4 before), and the
ordinary-English arm correctly survives, since a composition gate should not
touch it. Production code is unchanged and stays gate-free; the mutation was
applied, measured, and reverted.

Found while excluding NativeChatView's question-card remount as the owner of
#12118 / STA-3219: the remount is real, but the preedit survives it precisely
because this write path has no composition gate.

* fix(mobile): ship the patched xterm build, matching desktop

mobile pinned @xterm/xterm 6.1.0-beta.285 while the patch is keyed to
6.1.0-beta.287, so mobile shipped stock xterm and neither IME defect fix
reached it: the swallowed keystroke after an IME commit (9506039de7) and
the preedit sized to the font rather than the grid (e04e0c88da).

Bumps the three xterm packages to the desktop versions and adds the patch
to mobile's own pnpm.patchedDependencies. No copy of the patch: pnpm
accepts the parent-relative path and records it in the lockfile against
hash 8d63166272e9040a…, byte-identical to what desktop resolves, so the
two stay in step by construction rather than by a drift check.

The workspace separation is untouched — root pnpm-workspace.yaml still
declares `packages: []` and mobile keeps its own lockfile, which is what
keeps the root's patches from failing as ERR_PNPM_UNUSED_PATCH.

Verified in the generated webview bundle rather than at the install:
alignPreeditToGrid 0->2, sentComposition 0->3, pendingInput 0->11, and the
stock-only _handleAnyTextareaChanges 2->0 and dataAlreadySent 4->0. pnpm
applies patches during linking before postinstall regenerates the bundle,
confirmed by a revert/reinstall/re-apply cycle in both directions.

Mobile suite 2971 passed, 3 skipped — identical before and after. Bundle
+1,514 B (+0.24%). Lockfile churn is xterm-only; --frozen-lockfile passes.

mobile/src/ime/ime-submit-carry.ts is NOT made redundant and is untouched:
it handles iOS firing onSubmitEditing on a React Native native TextInput
after unmarking a composition, which is outside the WebView entirely.

Known divergence left alone: desktop also patches @xterm/addon-webgl and
mobile now runs that version unpatched. That patch is glyph/texture-atlas
rendering with nothing IME-related, so it affects neither fix.

* test(terminal): pin the preedit overlay against already-committed cells

STA-3132 (arm A), STA-3170 and STA-3232 report a Korean preedit painted on
top of text already on screen. Builds v1.4.163-v1.4.166 cancel the pending
finalizer in compositionstart, so a committed syllable reaches onData one
syllable late and buffer.x is stale — the overlay lands on the cell the
flushed syllable is about to occupy.

Replays a recorded hardware trace rather than an authored one: the ordered
DOM event stream captured on Windows + MS Korean (wave5-r2 evidence, 64
events), echoing onData back as PTY output.

The load-bearing assertion is deliberately not the obvious one. Comparing
overlay style.left against cursorX is tautological — left is computed from
buffer.x. This counts committed syllables from the compositionend events
the IME fired, so the two sides are independently derived.

Discriminated by a historical re-add across seven real bundles, since the
owner is deletion-shaped: pristine beta287, v1.4.155 and v1.4.162 pass;
v1.4.163 fails; v1.4.163 with that single call removed passes; the byte
identical baseline restored fails again; head passes. Every failing arm
fails only this case — the ordinary negative stays green in all seven.

The negative asserts its own category rather than claiming it: zero
composition events, zero isComposing, zero keyCode 229, exactly 16 events,
paired against the Korean arm's 4 starts / 3 ends / 11 updates / 64 events.

Scope: cell indices, not pixels. happy-dom has no layout, so the recorded
8x16 cell metrics are supplied to the render service. This makes no claim
about pixels visually overlapping; that is affected-OS confirmation and
stays open. Covers the overlap arm only — STA-3132's auto-line-break arm
and STA-3232's half-line-capacity and a11y arms are untouched.

* test(e2e): matrix macOS period substitution against the OS preference

#11504 reports macOS inserting ". " after a Hangul Space commit. This
sweeps six arms across both states of NSAutomaticPeriodSubstitutionEnabled,
reading the preference live per run rather than asserting a literal.

Two results worth having on record.

The reporter's stated trigger did not reproduce. Their words are "There is
no second press at all. One space is enough", but korean-single-space emits
zero insertText with the preference on or off. So does word-space-word-space.

Their timing does reproduce, with different content. korean-double-space and
korean-longer-word-double-space emit a delayed insertText at +122.5-122.7ms
after compositionend — squarely the reported +149ms — but the payload is a
space, never ". ". Consistent with the double-space rule seeing two slots
under ABC and only one under Korean, where the IME commit consumes the first.

The substitution itself is real and preference-bound: latin-double-space
gives "ab . " with the preference on and "ab " with it off, on one build
with the preference as the sole variable, reproduced across two runs.

That also refutes a claim in PR #11506, which states the substitution "is
enforced outside the renderer and never reproduces in dev builds, so changes
here must be verified against a packaged app". It reproduced in the dev build
twice and did not reproduce on the signed packaged app. That claim should not
be used as a verification gate.

Gated @headful behind ORCA_E2E_NATIVE_MACOS_PERIOD, same shape as the Korean
preedit spec, so it does not run in ordinary CI. Evidence is onData and DOM
only — the PTY-child reader aborted and no packaged-app arm was stable.

* test(terminal): actually enforce the recorded jamo progression

The preedit assertion compared sample.overlayText against sample.overlayText
— the same expression on both sides. A lane proved it by mutation: corrupting
seven of the eight recorded preedit values left the suite fully green. So the
docblock's ㄱ→가→간→나→낟→다→달→라, which the matrix also cites as this row's
recorded shape, was cited and unenforced.

The first attempt at a fix was insufficient and is worth recording. Threading
stroke.preedit through to the expectation still passed on a corrupted fixture,
because that value both drives the rig and was the expectation — corrupting it
moved both sides together. Same tautology, one level down.

The expectation is now an independent literal. Verified by mutation rather
than by reading: corrupting two recorded values fails one arm; restoring them
passes 3/3.

overlayCell was never affected — it is compared against a count derived from
the compositionend events, not from the buffer, and remains the load-bearing
assertion for the overlap.

* fix(e2e): select the selectable input source, not the first match

TISCreateInputSourceList can return several entries for one input source
id. A third-party IME publishes a non-selectable parent alongside the
selectable mode, and taking sources.first can return the parent — after
which TISSelectInputSource fails with paramErr (-50) while the caller
reports success from the enable step.

Found with Qingg (com.aodaren.inputmethod.Qingg), which exposes exactly
that pair under one id. Its mode id equals the bundle id, so filtering by
name would not have helped; selectability is the discriminator.

Now filters on kTISPropertyInputSourceIsSelectCapable and falls back to
the old behaviour when nothing advertises it, so single-entry sources are
unaffected. Also enables every entry for the id rather than only the one
being selected: selecting a mode whose parent is still disabled fails the
same way.

Compile-checked, and selecting com.apple.keylayout.ABC still exits 0.

Unrelated to the enable path: on macOS 26.5.2 third-party IMEs are gated
behind a consent sheet in System Settings. TISEnableInputSource returns
noErr immediately regardless, and the enable only lands if that sheet is
answered while the requesting process is still alive.

* test(terminal): discriminate #12171 against the real shortcut policy

The prior candidate mutation for this row was correctly refused: its suite
mocked shortcut policy so Process/229 became actionable, while the real
resolveTerminalShortcutAction has no Process branch — so the kill measured
the mock. This does not mock it.

Replays a capture of this row's own gesture (d, l, Shift+T, e, k, Space,
Enter under MS Korean, committing 있다) taken on Orca 1.4.164, through the
real useTerminalKeyboardShortcuts hook, capturing bytes at terminal.input.
The earlier capture could not discriminate at all because it recorded no
shiftKey; this one records it on 10 of 10 keydowns with code populated.

One physical Shift+T produces two shifted Process/229 keydowns. Under the
pre-#12265 classifier each synthesizes {key:'Enter', shiftKey:true}, which
the real policy resolves to sendInput '\x1b\r' — twice, giving 1b0d1b0d,
the two escapes the known-bad ed96881b0d1b0d contains.

Mutation is the retained pre-12265-process-shift.patch applied to HEAD, not
an authored one: patch -p1 applies clean and diffs identical to the mutant
copy. Arms are copies; shared source hashes the same before and after.

The English arm stays clean under both modules, so the mutation
discriminates by language rather than by harness — and a real Shift+Enter
through the same rig yields exactly ['\x1b\r'] in every arm, so a silent
pristine result means the code is quiet rather than the harness dead.

Scope: 1b0d1b0d is measured at the renderer boundary. The capture recorded
no PTY bytes — window.api.pty is frozen on shipped builds and the onData
channel needs a build-time flag — so this shows the renderer producing the
two escapes that payload contains, not a re-observation of the payload.

* docs(native-chat): narrow this file's disclaimer to what is now true

It said "THIS OWNS NO REPORTED ROW". Half of that is stale: the remount site
is now the attributed owner of #12118 and STA-3219. On real Windows TSF the
questionActive swap aborts a live composition — the old node gets only a
blur and no compositionend, the text returns as committed, and the next jamo
yields 아ㄴ rather than 안.

The other half holds. This file pins the opposite property, that the text
survives, which is the half those reporters already agree with. Mutation
shows the gap rather than asserting it: deleting the unmount entirely leaves
three of four tests green, because every substantive assertion is
after.value === … and a composer that never unmounts keeps its value.

Also records why the abort cannot be asserted here. The DOM exposes no
observable separating committed text from a live preedit — value is the same
string either way, there is no EditContext, and the only composing-ness
state is a per-instance ref discarded with the node. A test pinning "no
compositionend fires" would be an anti-guard: red the day it is fixed.

The cadence objection is kept, since it is now the open question rather than
the reason for exclusion.

* refactor(terminal): drop the unread isComposing field from XtermBypassEvent

Added by #6396 for terminal IME candidate handling that this branch has since
removed. No production or test code reads it, and the policy is safe without it:
during composition `key` is 'Process', so the non-ASCII printable checks that
would care never match.

Co-authored-by: Orca <help@stably.ai>

* fix(native-chat): keep the composer mounted through an in-flight IME composition

A question card replaced the composer outright
(`{questionActive ? null : <NativeChatComposer/>}`). Unmounting the field
mid-composition aborts the composition in the OS: the node is detached before
`compositionend` can fire, the preedit returns as committed text, and a resumed
Hangul syllable degrades — 아 then ㄴ yields `아ㄴ`, never `안`.

Confirmed in rasterised pixels on Windows with a real MS Korean IME, at both
v1.4.171 and the reporter-era v1.4.164 (the swap block is byte-identical
across them): the preedit underline present before the swap, the composer
visibly absent during it, and the same glyph back afterwards WITHOUT the
underline — committed, not composing.

The swap is now deferred while a composition is in flight, which is what
editors that survive IME do: ProseMirror gates DOM work on `view.composing`,
CodeMirror protects the composing subtree from redraws. Hiding instead of
unmounting does not work — `display:none` and `visibility:hidden` both blur the
focused element and abort the composition the same way.

The hold releases on `compositionend`, which browsers also fire on blur, so
clicking into the card's own answer input yields the input region immediately;
with nothing composing the card still replaces the composer at once, so no
stray "Send a message" appears beside a question.

The existing characterization test flips to a regression guard: it pinned the
node being destroyed, which was the defect. Node identity is the load-bearing
assertion — value-only checks are trivially satisfied by a composer that never
unmounts and cannot tell a held composition from a destroyed one.

The typing-redirect handler moves to its own hook. That is not cosmetic: both
touched files sat at the 400-line cap, and `max-lines` suppressions are
forbidden, so the room had to come from a real extraction.

* fix(macos): opt Orca out of AppKit automatic period substitution

macOS "Add period with double-space" (`NSAutomaticPeriodSubstitutionEnabled`,
on by default) is applied by AppKit's text input system. Native terminals never
join that system; Chromium text fields do, so xterm's helper textarea inherits
it and a double space arrives as `". "` — a period nobody typed, handed straight
to the PTY (#11504).

Chromium answers AppKit for quote and dash substitution and defaults both off,
but declares no period accessor at all, so AppKit applies that one without
asking. This user default is the only lever: there is no per-field or
per-webContents opt-out to prefer over it. Writing the key into Orca's own
defaults domain overrides the global value for this app alone and leaves the
user's system-wide setting untouched. It necessarily covers every Orca text
field, not only terminals — AppKit offers no narrower scope, and that tradeoff
is deliberate rather than accidental.

Measured on the reporter's own build v1.4.161: with the preference ON, typing
a,b,space,space yields `onData ["a","b"," ",". "]`; with it OFF the same arm
yields two spaces.

Note the issue's causal model is wrong and this fix does not follow it. It
claims the substitution only fires with a CJK input source and never with ABC.
The measurement is the inverse — every Korean arm is clean and the ABC arm is
the one that fires — so the fix is not conditioned on input source.

NOT YET VERIFIED ON HARDWARE. The unit tests inject the writer, so they prove
the call is made on darwin and skipped elsewhere; they do not prove AppKit
honours an app-domain override for this key. That check is outstanding.

* fix(xterm): keep a live composition across a lone Cmd press on macOS

CompositionHelper.keydown exempted keyCodes 16/17/18 from tearing a composition
down, which covers Shift/Ctrl/Alt but not macOS Meta — 91/93 in Chromium, 224 in
Firefox. A lone Cmd press mid-composition therefore reached
_finalizeComposition(false), which dropped the preedit overlay's `active` class
and committed the live syllable early. macOS keeps the marked text alive across
that press, so no later compositionstart re-arms the overlay and the rest of the
word composes invisibly.

Measured on hardware (m4air, macOS 26.5.2, Apple M4, 2-Set Korean) with the Cmd
posted as a CGEventType.flagsChanged, which is what a physical modifier emits.
AppleScript `key code 55` posts nothing a browser can see — a bare `key code 56`
for Shift is equally silent — which is why no capture in the corpus ever reached
this branch. Three arms, same build otherwise: overlay live throughout with the
exemption, dark and prematurely committed without it, live again with it
restored. Evidence under
.tmp/ime-handoff/swarm-scratch/wave31-cmd-preedit/evidence/.

The fix cannot widen past a lone modifier: only a standalone press reports these
keyCodes, and a Cmd chord during composition is reported by Chromium as 229,
which was already exempt. Cmd+A still ends the composition, via the IME's own
compositionend. Ghostty draws the same line, returning early from flagsChanged
under hasMarkedText() for every modifier including Super.

Orca's terminal pane was never affected — shouldSuppressTerminalModifierKeyboardEvent
drops a standalone Meta keydown before xterm sees it, and deleting only 'Meta'
from that set is what flipped the hardware arm to broken. The popout preview
terminal and mobile's webview install no such guard and did reach the teardown.

terminal-ime-xterm-composition-commit-overlap.test.ts asked its fixer to update
the two Cmd arms to the values it named as correct; both now emit a single ['한'].

* test(native-chat): drop two byte-identical duplicate cases

`4632b86919d` copy-pasted two cases twice into the same describe block:
`retains carry across a same-frame non-Enter keyup before redispatch` and
`expires carry before a deliberate Enter after the next frame`. Each pair is
byte-identical — same title, same body — so the copies asserted nothing the
originals did not.

This is what has been failing `static analysis` on this branch since 2026-08-06:
`oxlint vitest(no-identical-title)` reports both under `--deny-warnings`, and
`verify` fails solely because it requires static analysis to pass. Every other
gate in `verify` was already green, including typecheck, xterm patch sync, the
full test shard set, and both package jobs.

12 cases still pass in the file.

* test(e2e): skip the WebGL arm when no WebGL renderer exists

The #12164 probe runs two arms, webgl and dom, and closes by asserting the
active renderer is the requested one. That assertion is right for the dom arm —
it is what proves the pane actually left WebGL, without which the arm is
meaningless — but headless CI has no GPU, xterm falls back to DOM silently, and
the webgl arm then fails.

The failure reads as a Korean rendering defect and is not one, so the webgl arm
now skips with the active renderer named. The dom arm keeps the assertion
unchanged.

This is the third of three checks that have been red on this branch since
2026-08-06. `static analysis` and `verify` were fixed in c51c6b5837e; the CI log
shows this job as 1 failed / 1 passed, the pass being the dom arm.

* chore(lint): drop five unused no-console disable directives

`check-changed-code-quality` reports unused eslint-disable directives as errors,
and these five sat above diagnostic `console.log` calls in IME test and spec
files where `no-console` is not enabled — so each suppressed nothing.

This is the second of the two static-analysis steps. `c51c6b5837e` fixed
"Enforce focused code-quality plugins" (duplicate test titles); this fixes
"Enforce changed-code quality". Both had been red on this branch since
2026-08-06, and I mistook the first for the whole job.

The diagnostic logs themselves are kept — they are what a failing IME arm prints
for a reader to inspect.

* test(e2e): cover #12164 under fractional device scale factor

Fractional display scaling was #12164's last unexplored branch, and the reason
is worth recording: earlier attempts were BLOCKED, correctly, because they
proposed mutating the Windows display scale on a remote physical machine with no
console recovery. `--force-device-scale-factor` reaches the same renderer state
per process, so nothing outside the Electron instance changes and there is
nothing to restore.

The hypothesis was specific: `프프로로젝젝트트` is what a half-pixel cell boundary
could produce on a 2-column glyph, and nothing else in the suite varies dpr.

Measured at 1.25 and 1.5, both under WebGL: ink extents 25/21/16 with identical
ink groups, matching the scale-1 run. No doubling.

The arm self-certifies before asserting — if the flag does not take, the test
fails rather than silently measuring at dpr 1. That matters here: the sibling
spec's WebGL arm went two days reporting a missing GPU as a Korean rendering
defect precisely because a silent fallback looked like a result.

* fix(terminal): match Mod+letter shortcuts by physical key, not IME-rewritten key

With a CJK input source active, macOS and Windows report the physical key through
`code` but rewrite `key` to the layout's character: Korean 2-Set turns Cmd+C into
`{ key: "ㅊ", code: "KeyC", metaKey: true }`. Every `key.toLowerCase() === 'c'`
match misses it, so the shortcut is not recognised and xterm encodes the chord as
PTY input instead — issue #13033 reports `ESC[12618;9u` and a terminal that jumps
to the bottom, because user input scrolls the viewport.

This is the same key-vs-code confusion that owned #12171, where a `Shift+T`
typing ㅆ was read as Enter for want of a `code` guard, so the fix is the same
shape: trust `code` when it is present, fall back to `key` and then the legacy
`keyCode` when it is not (Chromium omits `code` on synthetic and some keypress
events, and `keyCode` keeps its US value even when `key` is rewritten).

Applied to the four terminal-side sites, including the dashboard pop-out, which
#13033 called out specifically as having its own key handler:
pty-connection.ts Cmd/Ctrl+C copy guard
keyboard-handlers.ts Cmd+G search navigation
agent-interrupt-inference.ts interrupt inference
preview-terminal-key-handler.ts pop-out paste

Nine further `key.toLowerCase()` letter matches exist outside the terminal
(TaskPage, editor, GitHub composer, browser markup). They have the same defect
and are deliberately left for a separate change rather than widening this one.

An existing case, `matchSearchNavigate > returns null for wrong key`, overrode
only `key` and left `code: 'KeyG'`, so it began passing for the wrong reason. It
now overrides both — which is what "wrong key" means once matching is physical —
and a companion case pins the Korean-rewritten chord still matching.

#13033 was closed NOT_PLANNED; the reporter's event shapes drive the new test.

* fix(renderer): match every Mod+letter shortcut by physical key, not IME-rewritten key

Completes the previous commit. A CJK input source rewrites `event.key` while
`event.code` keeps the physical key, so `key.toLowerCase() === 'z'` and friends
silently stop matching — the shortcut is not recognised and the keystroke falls
through to whatever handles unclaimed input.

The helper moves to `@/lib/ime-latin-shortcut-key` first: it now serves the
editor, GitHub composer and browser markup, and importing terminal-pane
internals into those would be the wrong direction. `lib/` already hosts
`ime-composition-keyboard-event` for the same reason.

Nine remaining sites, all previously unreachable under Korean/Japanese/Chinese/
Vietnamese input:
TaskPage, ActivityPrototypePage, ProjectViewWrapper Cmd+F search
useMarkupKeyboardShortcuts Cmd+Z undo
GitHubMarkdownComposer, RichMarkdownLinkBubble,
rich-markdown-link-shortcut Cmd+K link
native-chat-shortcut Cmd+J
rich-markdown-key-handler Cmd+Shift+X

Six of the nine test `!== 'letter'` as early-return guards and three test
`=== 'letter'`; the negation is applied per site, since a blind substitution
would have inverted six of them.

Full suite: 4249 files pass. Three files fail locally and none is caused by this
change — the branch touches no file under `src/main/` or `src/relay/`, all four
failures reproduce on an unmodified tree or pass in isolation (the worktree
poller passes 21/21 alone, so it is full-suite parallelism), and all 16 CI test
shards are green.

* docs(ime): scope the IME composition rules to the terminal-pane directory

#11893 proposed adding these to the root `AGENTS.md`, which every agent loads on
every task regardless of what it is doing. They only bind keyboard handling, the
composer and the terminal input path, so they belong next to that code —
`tests/e2e/AGENTS.md` already establishes the nested convention here.

Kept from #11893: range-derived commits, guarding above the key dispatch,
the `attachCustomKeyEventHandler` / `CompositionHelper` interaction, no
normalization at commit, and the recorded-trace evidence bar.

Added from defects found since it was written:
- match shortcuts on `event.code`, not `event.key` (#12171, #13033)
- `keyCode === 229` means an IME owns the press
- do not unmount a field mid-composition, and hiding is not a fix because
`display:none` blurs and aborts it too (#12118, STA-3219, #11332)

The evidence bar now also names the mutation check, since a test that survives
deleting the code it guards is guarding nothing — a failure this effort hit more
than once.

* fix(terminal): gate Ctrl+Enter CSI-u on a negotiated pane, porting #12462

Found while scoping the rebase onto `main`: #12462 landed on 2026-08-06 and
fixes a real defect this branch does not carry. Ctrl+Enter emitted
`\x1b[13;5u` unconditionally, so a pane that never negotiated the kitty
keyboard protocol — local Windows ConPTY, plain shell — printed the escape
verbatim into the prompt.

This branch deletes `terminal-ime-deferred-newline.ts`, which is one of the
files #12462 touched, so a rebase resolving those conflicts by taking our side
wholesale would silently reintroduce the defect. Porting it forward now means
the fix survives the rebase however the conflicts are resolved.

Mirrors the Shift+Enter guard already here: local ConPTY falls back to the
legacy CR every emulator sends, and a negotiated pane keeps the chord, so the
fallback is scoped to panes that cannot receive CSI-u rather than to Windows.

NARROWER THAN #12462 BY ONE CONDITION, deliberately. `main` also allows CSI-u
via `hasCtrlEnterCsiUAuthority()` (trusted consumer evidence, #12329); that
helper and its plumbing do not exist on this branch. Omitting it is the
conservative direction — an authorised pane gets `\r` instead of the chord,
rather than an unnegotiated pane printing an escape — but it should be restored
when the two histories are reconciled.

Test covers both directions and is mutation-checked: forcing the gate open
fails it, so it cannot pass by construction.

* fix: reconcile two more fixtures main moved while the stack waited

Both caught by CI, not locally, and the reason the local run missed one is
worth recording:

1. `browser-toolbar-profile-dialogs.ime-enter.test.tsx` did not pass
`useNativeUserAgent` / `onUseNativeUserAgentChange`, which `main` added to
`BrowserToolbarProfileDialogsProps`.

Local `pnpm typecheck` reported 0 errors on the same commit CI failed. The
cause was a stale `config/*.tsbuildinfo` — tsc reused an incremental cache
from before the merge. Deleting it reproduced CI's error exactly. Any
"typecheck clean" during this merge should be treated as unverified unless
the cache was cleared first.

2. Localization keys for `SshDisconnectedDialog` were absent from `en.json`:
the merge took this branch's component alongside `main`'s catalog.
Regenerated with `pnpm run sync:localization-catalog` rather than hand-added.

* fix(mobile): regenerate the lockfile the merge resolved by taking one side

CI's `verify` failed with `ERR_PNPM_OUTDATED_LOCKFILE` on `mermaid (lockfile:
11.16.0, manifest: 11.16.1)`. The mismatch was in `mobile/`, not the root — the
root lockfile was consistent throughout, which is why inspecting it (and even
GitHub's merge ref) found nothing wrong.

Cause: during the merge I resolved `mobile/pnpm-lock.yaml` by taking this
branch's side wholesale rather than merging it, so it kept `mermaid 11.16.0`
while `mobile/package.json` came from `main` at `11.16.1`. Taking one side of a
lockfile is only safe when the corresponding manifest comes from the same side.

Regenerated with `pnpm install --lockfile-only`; `--frozen-lockfile` now passes
in `mobile/`. Verified the xterm patch entry survives intact — same hash
`4f1b42d268f3964d…` and the parent-relative path into `config/patches/`, which
is the desktop/mobile coupling that would silently break the mobile build.

Two earlier diagnoses of this failure were wrong and are worth recording: it was
not the root lockfile, and it was not a stale merge ref (a rerun reproduced it
exactly).

---------

Co-authored-by: Orca <help@stably.ai>

2026-08-07 22:58:40 -07:00

11 KiB

Raw Blame History

xterm Patch Regeneration

Scope

Orca ships @xterm/xterm with four source changes it needs and upstream has not taken: the IME composition hooks, the xterm-composition-* custom events they raise, the ITerminal surface those events widen, and a SortedList fix. pnpm applies them through config/patches/@xterm__xterm@<version>.patch.

That patch touches eight files. Four are hand-authored source (src/browser/CoreBrowserTerminal.ts, src/browser/Types.ts, src/browser/input/CompositionHelper.ts, src/common/SortedList.ts) and four are the build output those sources produce (lib/xterm.js, lib/xterm.mjs, and both sourcemaps). The bundle half is 7.3 MB of minified code. It is generated, and this document exists so nobody edits it by hand.

The source patch carries a fifth file, src/browser/TestUtils.test.ts, which the shipped patch does not and cannot; see The Source Patch Is a Superset.

config/patches/xterm-src/@xterm__xterm@<version>.src.patch is the source of truth. Everything else is derived from it by config/scripts/regenerate-xterm-patches.mjs, which is pinned to the exact upstream commit the published tarball was built from.

This policy covers @xterm/xterm only. The addon patches (@xterm/addon-webgl, @xterm/addon-serialize) are still hand-edited bundles and are tracked separately; see Known Gaps.

Rules

Never edit config/patches/@xterm__xterm@<version>.patch. Edit the source patch and regenerate.
Never edit lib/ inside a patched node_modules tree and re-run pnpm patch-commit. That is how bundle hunks stop matching their sources.
Every source change must land together with the regenerated bundle hunks and the pnpm-lock.yaml hash bump, in one commit.
The upstream commit lives in config/patches/xterm-upstream.json, not in a comment. A version bump that leaves it stale fails the generator, it does not silently patch the wrong tree.
Sourcemaps are deleted, not patched and not silently omitted. The patch moves the bundle, so a retained map would have to move with it; dropping only the map hunks ships offsets that point at the wrong code. Deletion is the honest form of that saving, and sourcemaps.policy in the manifest controls it.
--check is the authority on the lockfile, not pnpm install. pnpm writes the patch hash in two places — patchedDependencies and every resolution key that depends on the patched package — and on a warm store it will leave the resolution keys at their previous value while reporting success. That installs locally and drifts on CI's cold store. Always finish on step 4, and if it reports a stale hash after an install, rerun --write.

Workflow

# 1. Edit the source hunks.
$EDITOR config/patches/xterm-src/@xterm__xterm@6.1.0-beta.287.src.patch

# 2. Rebuild the bundle hunks, the full patch, and the lockfile hash.
node config/scripts/regenerate-xterm-patches.mjs --write

# 3. Reinstall so node_modules picks up the new patch hash.
pnpm install

# 4. Confirm the tree is self-consistent.
node config/scripts/regenerate-xterm-patches.mjs --check

Editing a patch file by hand is awkward for anything larger than a one-liner. For a substantial change, work in the generator's own checkout instead — after any run it is left at the pinned commit with the source patch applied:

node config/scripts/regenerate-xterm-patches.mjs --check --work-dir=/tmp/xterm
$EDITOR /tmp/xterm/upstream/src/browser/input/CompositionHelper.ts
git -C /tmp/xterm/upstream diff -- src/ > config/patches/xterm-src/@xterm__xterm@6.1.0-beta.287.src.patch
node config/scripts/regenerate-xterm-patches.mjs --write --work-dir=/tmp/xterm

--write rewrites the source patch into the canonical form it would emit on a re-diff, so a hand-produced git diff gets normalized on the first run rather than fighting --check forever.

Run the checkout outside this repository. A build tree underneath it makes tsgo walk up into Orca's own node_modules and fail with TS2300: Duplicate identifier, which is a symptom of where the tree sits and not of the patch.

How the Commit Is Known

Upstream bin/publish.js sets packageJson.commit before npm publish, so each published tarball names the commit that built it. The generator asserts that stamp against xterm-upstream.json and then compares the tarball's src/ against the checkout file by file. Only src/common/Version.ts may differ, because publish.js rewrites the version immediately before packaging; the generator applies the same stamp.

That pair of checks is what makes the rebuild trustworthy. Without them a wrong commit would still produce a plausible-looking 7 MB patch.

The Source Patch Is a Superset

Patching ICompositionHelper widens an interface, so every implementor has to follow — including MockCompositionHelper in upstream's src/browser/TestUtils.test.ts. Without that hunk the patched checkout does not type-check and npm run package never reaches webpack, so the generator cannot build the patched bundles at all.

Upstream's .npmignore strips *.test.ts, so that file is not in the published tarball. The shipped patch is a diff against the published tarball, and it therefore cannot name the file — correctly, since pnpm has nothing there to patch.

That is why the source patch is derived from the upstream checkout (git diff -- src/) and not from the emitted patch. Deriving it from the emitted patch is the trap: --write would filter the hunk out through the published file set and delete it, so the fix that makes the build work would erase itself on the first run that used it.

The two derivations are still cross-checked. assertSourceDerivationsAgree requires them to be byte-identical on every file the tarball publishes, so the carve-out stays confined to files upstream does not ship rather than becoming a place where the source patch and the shipped patch can quietly disagree. The checkout diff uses pnpm's own formatting flags minus --no-index, which is what makes that byte comparison meaningful.

Build Order

Upstream's publish path is npm ci → stamp Version.ts → npm run package. npm run package runs webpack for lib/xterm.js and then, via postpackage, bin/esbuild_all.mjs --prod for lib/xterm.mjs.

Do not run npm run setup after the packaging build. setup is the development esbuild pass with minify: false. Running it afterwards overwrites lib/xterm.mjs with an unminified bundle and a map that no longer matches, and the resulting patch is silently wrong — the failure mode is a .mjs that is 50% larger than the published one, which is easy to miss inside a 7 MB diff. forbiddenBuildScripts in the manifest encodes this and the generator refuses to run a build step that names one of those scripts.

The generator also builds the unmodified commit first and asserts that it reproduces the published lib/ byte for byte before it emits anything. A toolchain or build-order problem therefore surfaces as an explicit "did not reproduce the published bundles" error rather than as 7 MB of mystery diff.

The Lockfile Moves With the Patch

pnpm derives the patchedDependencies hash in pnpm-lock.yaml — and the .pnpm/@xterm+xterm@<version>_patch_hash=<hash>/ store directory name — from the sha256 of the patch file itself. A regenerated patch without the lockfile bump fails pnpm install --frozen-lockfile on every machine except the author's. --write makes that edit; --check fails if it is missing.

config/scripts/regenerate-xterm-patches.test.mjs asserts the same thing without a network or a build, so the ordinary test job catches lockfile drift in milliseconds even though the full rebuild runs in its own CI lane.

Toolchain Pin

toolchain in the manifest records what upstream's package-lock.json resolves at the pinned commit, and the generator fails if npm ci produces something else. The entry that matters is @typescript/native-preview (tsgo), which upstream pins to a dated development build — 7.0.0-dev.20260521.1 at the time of writing. It is a real published version and npm does not prune old releases, but it is the one dependency of this scheme that is not a stable release.

If that version ever becomes unresolvable the generator fails with a toolchain error naming it. Recovery is to move the pin to the next upstream commit whose package-lock.json resolves, re-verify that the rebuild still reproduces the published bundles, and regenerate. The committed patch keeps working the whole time — only regeneration is blocked, so this is never an outage.

Version Bumps

Bumping @xterm/xterm is:

Update the version in package.json and run pnpm install.
Rename both patch files to the new version and update patch, sourcePatch, and version in xterm-upstream.json.
Update upstream.commit to the commit field of the new tarball's package.json, and toolchain to whatever the new package-lock.json resolves.
node config/scripts/regenerate-xterm-patches.mjs --write.

Step 4 is where a real upstream conflict shows up: git apply of the source patch fails against the new tree. Resolve it in the checkout, re-diff, and rerun. The bundle hunks need no attention at any point.

Why Not Vendor a Fork

A vendored @xterm/xterm fork removes the patch entirely, but it moves Orca off the published package, so every upstream beta becomes a merge rather than a version bump, and Orca inherits responsibility for building and publishing a package it does not own. The patch is four small source hunks against a commit that reproduces byte for byte; a fork is a much larger standing cost for the same result.

Why Not Handle Composition at Runtime

CompositionHelper hooks four private call sites upstream of onData, and SortedList has no public surface at all. There is no supported extension point that reaches either, so a runtime shim would mean reaching into _core internals that upstream renames freely between betas. The patch is the smaller risk.

CI Contract

xterm_patch_sync in .github/workflows/pr.yml runs regenerate-xterm-patches.mjs --check on every PR and is part of the verify aggregate. It clones the pinned commit, installs upstream's toolchain, builds twice, and byte-compares the result against the committed patch. A warm run is about eight seconds of work around the clone and install.

config/scripts/regenerate-xterm-patches.test.mjs covers the pure pieces — pnpm's diff flags and normalization, hunk splitting, round-trip stability, the commit and build-order assertions, and lockfile coupling — with no network and no build, so they run in the ordinary test shards.

Known Gaps

@xterm/addon-webgl and @xterm/addon-serialize are still hand-edited minified bundles. Their patches carry a literal /* PATCH(orca): ... */ comment inside minified code and parser round-trip artifacts, and neither patch touches its .map file, so both addons currently ship sourcemaps whose offsets do not match the shipped bundle — the defect sourcemaps.policy now avoids for @xterm/xterm and which folding them into this manifest would also fix. Both addons build from the same pinned commit and reproduce byte for byte, so they can be folded into this manifest as additional packages entries; that change needs e2e sign-off because, unlike @xterm/xterm, it will not be a byte-for-byte no-op.

11 KiB Raw Blame History

xterm Patch Regeneration

Scope

Rules

Workflow

How the Commit Is Known

The Source Patch Is a Superset

Build Order

The Lockfile Moves With the Patch

Toolchain Pin

Version Bumps

Why Not Vendor a Fork

Why Not Handle Composition at Runtime

CI Contract

Known Gaps

11 KiB

Raw Blame History