Fix Korean 2-Set terminal arrows

[austinywang]

Summary

• add a Korean 2-Set post-composition marked-selection regression for Left/Right arrows
• pass the physical key and input source into the IME suppression decision
• let plain Korean 2-Set Left/Right arrows reach Ghostty while preserving suppression for IME-owned marked-text changes
• route Korean marked-selection arrow key equivalents into the terminal keyDown path so AppKit window handling cannot swallow them first

Fixes #4093

Verification

• Reproduced locally in cmux NIGHTLY 0.64.4-nightly.2579225123401 with macOS input source com.apple.inputmethod.Korean.2SetKorean: after typing Korean text, Left/Right did not move the terminal cursor.
• Added regression test: CJKIMEMarkedSelectionTests.testKeyDownForKoreanPostCompositionHorizontalArrowsForwardsToTerminal.
• Did not run local tests per repository policy; CI is the test runner.
• Ran git diff --check.
• Built and launched tagged dev app with CMUX_SKIP_ZIG_BUILD=1 ./scripts/reload.sh --tag issue-4093-korean-2set-arrow-keys --launch.
• Manual dogfood in tagged dev app with Korean 2-Set selected: physical 2-set input stayed Korean, Left moved the insertion point before the final character, and Right moved it back after the final character.

Build

App path:
/Users/austinwang/Library/Developer/Xcode/DerivedData/cmux-issue-4093-korean-2set-arrow-keys/Build/Products/Debug/cmux DEV issue-4093-korean-2set-arrow-keys.app

Notes

• Cloud Mac video artifact was not captured because the debugging flow was redirected to local cmux NIGHTLY repro and tagged dev-app verification before the fix.
• The working tree had a pre-existing dirty ghostty submodule (aef980e27..fe972c095) before this change; it was not staged or committed.

---

Note

Medium Risk
Modifies macOS IME key routing and marked-text suppression logic, which can subtly affect keyboard shortcuts and text input behavior across input methods. Adds input-source–specific branching and new preedit buffering paths, increasing the chance of regressions in edge-case key handling.

Overview
Fixes a regression where Korean 2-Set post-composition Left/Right arrows were being swallowed by AppKit by ensuring they flow through the terminal keyDown path and are explicitly allowed to forward even when the IME updates marked-selection.

Makes IME suppression decisions event + input-source aware (passing NSEvent and keyboard source ID), adds allow-list logic for Korean 2-Set arrow forwarding, and keeps IME-owned composition command keys from also moving the terminal cursor. Additionally buffers Zhuyin/Bopomofo component insertText updates as editable preedit (instead of committing to the terminal) and expands CJKIMEMarkedSelectionTests to cover the new Korean/Zhuyin behaviors.

^{Reviewed by Cursor Bugbot for commit 50b0c0e. Bugbot is set up for automated code reviews on this repo. Configure here.}

---

Summary by cubic

Fixes a regression where Left/Right arrows didn’t move the cursor after Korean 2‑Set input and restores correct Zhuyin (Bopomofo) IME routing. During composition, non‑Cmd key equivalents bypass the window so keyDown handles IME input; idle Zhuyin navigation keys reach the terminal; and Zhuyin preedit is buffered until commit (Fixes #4093).

• Bug Fixes
  • Korean 2‑Set: forward only plain Left/Right after post‑composition selection changes; don’t reroute their key equivalents; clear stale key‑up suppression when forwarding.
  • Composition routing: while marked text is visible, do not handle non‑Cmd key equivalents at the window; keep IME composition commands (arrows, paging, space, return, tab, delete, home/end, escape) inside text input for IME sources; when no marked text, do not suppress Zhuyin Down/PageUp/PageDown/Space so they reach the terminal.
  • Zhuyin preedit: buffer raw Bopomofo components and tone marks as marked text and commit only the chosen candidate; refined replacement/selection handling and input‑source/event‑aware suppression.
  • Added tests for Korean arrow forwarding, idle Zhuyin navigation pass‑through, Zhuyin command routing with and without marked text, and Zhuyin preedit buffering/commit.

^{Written for commit 50b0c0e. Summary will update on new commits.}

Summary by CodeRabbit

• Bug Fixes

  • Prevent non-Command keys from being claimed during active IME composition and fix stale key-up handling.

• New Features

  • Input-source–aware IME gating with special handling for Korean 2-Set arrow navigation.
  • Zhuyin (Bopomofo) preedit buffering and improved insertion/replacement behavior for composed text.

• Tests

  • Expanded CJK IME tests covering Korean arrow forwarding, Zhuyin buffering, and post-composition behaviors.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cmux	Ready	Preview, Comment	May 13, 2026 1:09pm
cmux-staging	Building	Preview, Comment	May 13, 2026 1:09pm

[coderabbitai]

📝 Walkthrough

Walkthrough

Makes IME suppression event- and input-source-aware, adds Korean “2Set” arrow detection and routing, captures keyboard layout in keyDown, alters AppDelegate key-equivalent handling for IME composition, and extends CJK IME tests.

Changes

Korean 2-Set IME arrow key routing

Layer / File(s)	Summary
IME suppression logic and Korean/Bopomofo helpers Sources/GhosttyNSView+IMEComposition.swift	shouldSuppressGhosttyKeyForwardingAfterIMEHandling now accepts event and inputSourceId; adds Korean “2Set” input-source allow-list, input-source predicates, modifier/key allow-lists, arrow/keyCode recognition, Bopomofo scalar validation, and updates DEBUG test helper signature.
keyDown integration and keyboard ID capture Sources/GhosttyTerminalView.swift	keyDown captures KeyboardLayout.id before interpretation, passes inputSourceId into suppression checks, records imeConsumedKeyUps on suppression and clears stale entries on forwarding, and adds Bopomofo preedit buffering helpers used by insertText(_:replacementRange:).
AppDelegate window-level behavior Sources/AppDelegate.swift	Swizzled NSWindow.cmux_performKeyEquivalent(with:) early branch now returns false for Ghostty first responder with marked text and non-Command events instead of delegating through.
CJK IME marked-selection tests cmuxTests/CJKIMEMarkedSelectionTests.swift	Adds Carbon.HIToolbox import, KoreanArrowProbe, and tests validating Korean arrow forwarding, Zhuyin/Bopomofo buffering and suppression, and commit behavior.

Sequence Diagram

sequenceDiagram
  participant User
  participant Window as AppDelegate.cmux_performKeyEquivalent
  participant View as GhosttyNSView.keyDown
  participant Supp as GhosttyNSView.shouldSuppressGhosttyKeyForwardingAfterIMEHandling
  participant Terminal as Ghostty terminal
  User->>Window: press arrow/key
  Window->>View: window-level check (marked text, modifiers)
  alt window-level return false
    Window-->>User: return false (block performKeyEquivalent)
  else
    Window->>View: forward to keyDown
  end
  View->>View: capture KeyboardLayout.id (keyboardIdBefore)
  View->>View: interpretKeyEvents / handleTextInputKeyEvent
  View->>Supp: call(before, after, accumulatedText, event, inputSourceId)
  alt suppression true
    Supp-->>View: return true (suppress)
    View->>View: record imeConsumedKeyUps and return
  else
    Supp-->>View: return false (allow)
    View->>View: clear stale imeConsumedKeyUps
    View->>Terminal: forward key to terminal
  end

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related issues

• Korean 2-set IME: arrow keys swallowed in terminal (regression from #3867) #4093: Addresses the same IME suppression/routing behavior by adding event/input-source-aware suppression and window-level forwarding.

Possibly related PRs

• manaflow-ai/cmux#2529: Also adjusts GhosttyTerminalView IME/composition key handling to prevent composing-time events from being forwarded or committed.
• Narrow IME candidate key suppression #3867: Overhauls IME key-equivalent routing and suppression logic and adds related tests.
• Restore Zhuyin IME candidate marked-text handling #3574: Modifies IME composition/key-forwarding behavior in GhosttyNSView/GhosttyTerminalView with marked-text selection and suppression tweaks.

"🐰 I hopped through IME fields so wide,
arrows find the terminal, nowhere to hide,
source and event I sniff with care,
preedit stitched and commits laid bare,
keys and caret now dance side by side."

🚥 Pre-merge checks | ✅ 14 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 10.71% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (14 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main fix: enabling Korean 2-Set terminal arrow keys to work correctly after IME composition. It is specific and directly related to the core issue being resolved.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Cmux Swift Actor Isolation	✅ Passed	No Swift 6 actor isolation violations. Changes are isolated to UI/MainActor methods. No problematic value models, service protocols, shared mutable Sendable types, or cross-isolation access.
Cmux Swift Blocking Runtime	✅ Passed	No blocking/timing synchronization primitives introduced in IME fix. Changes use decision logic, state tracking, and helper methods only.
Cmux No Hacky Sleeps	✅ Passed	Check inapplicable. Rule excludes Swift code and applies to TypeScript/JavaScript/shell/build scripts only. PR contains only Swift and test infrastructure changes.
Cmux Swift Concurrency	✅ Passed	No legacy async patterns introduced. Changes add synchronous IME helper methods with no new DispatchQueue, Combine, completion-handlers, or fire-and-forget Tasks.
Cmux Swift @Concurrent	✅ Passed	All changes are synchronous pure helpers for IME composition logic. No async functions, invalid @concurrent, or heavy operations violate the rules.
Cmux Swift File And Package Boundaries	✅ Passed	Repository initialization with 1,742 new files. Check allows repo initializations. GhosttyNSView+IMEComposition.swift under 400 lines with focused IME responsibility (AppKit glue).
Cmux Swift Logging	✅ Passed	No logging violations found in production Swift code changes. IME files contain no print/NSLog. Test and CLI output are allowed per rules.
Cmux Swiftui State Layout	✅ Passed	No new SwiftUI state introduced. Changes are AppKit IME key handling. No @Published/@observable added. Pre-existing state only touched incidentally per allowed exceptions.
Cmux Architecture Rethink	✅ Passed	Adds pure functional IME suppression logic without timing delays, mutable state, or lifecycle splits. Clear ownership and single source of truth for decisions. No architectural violations.
Cmux Swift Auxiliary Window Close Shortcuts	✅ Passed	PR modifies IME key handling logic only. No user-visible NSWindow or NSPanel creation or changes. Test-only window fixture is allowed per review rules.
Description check	✅ Passed	The PR description is comprehensive and addresses all required template sections with substantive content.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch issue-4093-korean-2set-arrow-keys

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[greptile-apps]

Greptile Summary

Fixes a regression where Korean 2-Set Left/Right arrows were swallowed after Hangul composition, and improves Zhuyin (Bopomofo) preedit buffering so raw phonetic components are held in marked-text until a candidate is committed.

• AppDelegate: performKeyEquivalent now returns false (instead of recursing into cmux_performKeyEquivalent) for non-Command keys during IME composition, letting AppKit dispatch naturally through keyDown.
• IME suppression logic: shouldSuppressGhosttyKeyForwardingAfterIMEHandling gains event and inputSourceId parameters; Korean 2-Set plain Left/Right are allow-listed to forward to the terminal even when the marked-text selection changes; a new shouldKeepIMECompositionCommandInsideTextInput guard suppresses navigation keys that belong to IME candidate menus (e.g., Zhuyin Down/Space).
• Bopomofo preedit buffering: insertText now intercepts Bopomofo-scalar strings during key accumulation and routes them into insertBopomofoPreeditText, keeping them as editable preedit until the IME commits a chosen character; a imeConsumedKeyUps.remove call clears stale key-up suppression when a key is forwarded.

Confidence Score: 5/5

Safe to merge. The IME key-routing changes are localized to the Korean 2-Set and Zhuyin paths, existing suppression logic for all other input sources is preserved, and new regression tests cover the core scenarios.

The Korean 2-Set forwarding allow-list is a narrow, well-guarded change: only plain Left/Right on exactly com.apple.inputmethod.Korean.2SetKorean bypass suppression. The Bopomofo preedit buffering is similarly constrained to Bopomofo-range Unicode scalars. The AppDelegate change (returning false instead of recursing into cmux_performKeyEquivalent) cleanly removes an unnecessary detour. The new imeConsumedKeyUps.remove correctly cleans stale suppression on forwarded keys. No existing suppression paths are widened beyond their intended scope.

No files require special attention. The most complex logic lives in GhosttyNSView+IMEComposition.swift and is directly exercised by the new test suite.

Important Files Changed

Filename	Overview
Sources/GhosttyNSView+IMEComposition.swift	New IME suppression helpers: Korean 2-Set allow-list, Bopomofo source detection, and shouldKeepIMECompositionCommandInsideTextInput. Logic is correct; isInputMethodSource and isBopomofoInputSource correctly use en_US_POSIX locale with .caseInsensitive for ASCII bundle IDs.
Sources/GhosttyTerminalView.swift	Adds Bopomofo preedit buffering in insertText, insertBopomofoPreeditText, and effectiveBopomofoPreeditReplacementRange; changes keyboardIdBefore to always capture (not just when no marked text); adds imeConsumedKeyUps.remove on forwarded keys. All logic correct; insertionLocation uses NSString.length safely since all Bopomofo/modifier-letter scalars are BMP.
Sources/AppDelegate.swift	Single-line change: returns false instead of cmux_performKeyEquivalent(with:) during IME composition for non-Command keys, letting AppKit dispatch through keyDown rather than re-entering key-equivalent machinery. Intent and effect are correct.
cmuxTests/CJKIMEMarkedSelectionTests.swift	New test cases cover Korean 2-Set Left/Right forwarding, idle Zhuyin key pass-through, Zhuyin Down suppression during composition, Bopomofo component buffering, mid-preedit insertion, and committed-candidate accumulator flow. Coverage is well-targeted to the changed paths.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[NSWindow.performKeyEquivalent] -->|hasMarkedText AND no Command| B[return false → AppKit dispatches via keyDown]
    A -->|no marked text OR has Command| C[normal key equivalent routing]
    B --> D[GhosttyNSView.keyDown]
    D --> E[capture keyboardIdBefore]
    E --> F[interpretKeyEvents]
    F -->|Bopomofo scalar + keyTextAccumulator active| G[insertBopomofoPreeditText: buffer in preedit]
    F -->|other text| H[insertText → unmarkText + accumulate]
    F --> I[shouldSuppressGhosttyKeyForwardingAfterIMEHandling]
    I -->|accumulated text present| J[return false — don't suppress]
    I -->|no marked text before or after| J
    I -->|marked text changed| K[return true — suppress]
    I -->|selection changed: Korean 2-Set Left/Right plain| J
    I -->|selection changed: other| K
    I -->|no change + inputmethod source + IME nav key| K
    I -->|no change + Korean 2-Set Left/Right plain| J
    I -->|no change + other| J
    J --> L[imeConsumedKeyUps.remove keyCode]
    L --> M[forward key event to Ghostty terminal]
    K --> N[imeConsumedKeyUps.insert keyCode: return early]

Loading

_{Reviews (6): Last reviewed commit: "Address IME suppression review feedback" | Re-trigger Greptile}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cmuxTests/CJKIMEMarkedSelectionTests.swift`:
- Around line 218-298: The test currently calls surfaceView.keyDown(with:)
directly which bypasses AppKit's performKeyEquivalent/window/AppDelegate
routing; change the test to dispatch the generated NSEvent through the normal
AppKit event path (e.g. call window.sendEvent(event) or
NSApplication.shared.sendEvent(event)) after setting
window.makeFirstResponder(surfaceView) so performKeyEquivalent and the
window/AppDelegate reroute are exercised; keep the cjkIMEInterpretKeyEventsHook,
GhosttyNSView.debugGhosttySurfaceKeyEventObserver and the keyEvent(...) helper,
but replace the surfaceView.keyDown(with: event) call with sending the event via
window.sendEvent(event) (or NSApp.sendEvent) so the KoreanArrowProbe cases go
through performKeyEquivalent.

In `@Sources/GhosttyNSView`+IMEComposition.swift:
- Around line 119-122: The current isInputMethodSource(_:) uses
localizedCaseInsensitiveContains("inputmethod") which permissively matches any
bundle id containing that substring; to make it stricter, update
isInputMethodSource(_ inputSourceId: String?) to check for the dot-anchored
token (e.g., use a case-insensitive contains(".inputmethod.") or a
regex/hasSuffix/contains pattern that matches ".inputmethod."), so it only
returns true for standard Apple-style IM bundle IDs; keep the nil guard and
behavior otherwise the same.

In `@Sources/GhosttyTerminalView.swift`:
- Around line 13284-13293: In insertBopomofoPreeditText, compute the caret
offset using the original NSRange UTF-16 coordinates before mutating markedText:
after confirming replacementRange.location != NSNotFound and that
Range(replacementRange, in: markedText.string) exists, derive the new caret
location as replacementRange.location + chars.count (or compute distance from
start using the old range.lowerBound) first, then assign markedText =
NSMutableAttributedString(string: next) and finally call markedSelectedRange =
normalizedMarkedSelectionRange(NSRange(location: location, length: 0),
markedLength: markedText.length); this avoids using an index from the old string
against the new string and prevents the trap/crash.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: acd45826-e52a-44b5-9a82-4a7b8e879cae

📥 Commits

Reviewing files that changed from the base of the PR and between f206799 and 9795cb4.

📒 Files selected for processing (4)

• Sources/AppDelegate.swift
• Sources/GhosttyNSView+IMEComposition.swift
• Sources/GhosttyTerminalView.swift
• cmuxTests/CJKIMEMarkedSelectionTests.swift

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cmuxTests/CJKIMEMarkedSelectionTests.swift`:
- Around line 218-299: The test currently only asserts that arrow key events are
forwarded to Ghostty but doesn't verify the IME-updated marked selection; after
each window.sendEvent(event) in
testKeyDownForKoreanPostCompositionHorizontalArrowsForwardsToTerminal (inside
the probes loop within withExtendedLifetime) add an assertion that
surfaceView.selectedRange() (or equivalent accessor) equals probe.selectionAfter
so the test validates both the IME selection update and the forwarding behavior
handled by cjkIMEInterpretKeyEventsHook.

In `@Sources/GhosttyNSView`+IMEComposition.swift:
- Around line 5-7: The korean2SetInputSourceIDs Set currently only contains
"com.apple.inputmethod.Korean.2SetKorean" which is too narrow; update the Set
(korean2SetInputSourceIDs) to either include other Apple Korean input-source IDs
known to use the same IME flow (e.g., 3SetKorean, 390Sebeolsik,
GongjinCheongRomaja, HNCRomaja) or, if you intend to scope only to 2-Set for
now, add a concise comment next to private static let korean2SetInputSourceIDs
documenting that other Korean layouts are intentionally deferred and add a
TODO/issue reference for a follow-up to cover the family; ensure the symbol name
korean2SetInputSourceIDs is updated accordingly if you broaden semantics (e.g.,
rename to koreanInputSourceIDs) to reflect the wider coverage.

In `@Sources/GhosttyTerminalView.swift`:
- Around line 13227-13230: The buffered Bopomofo preedit path ignores NSNotFound
semantics and always appends instead of inserting at the current marked
insertion point; update the logic in insertText(_:replacementRange:) /
insertBopomofoPreeditText(chars, replacementRange:) so that when
replacementRange == NSNotFound and a marked range exists you use
markedSelectedRange as the actual replacement/insertion range (convert
markedSelectedRange to an NSRange as needed), otherwise preserve
replacementRange; ensure the code performs a replace-in-range operation rather
than appending to the end so in-place edits within keyTextAccumulator behave
correctly (refer to keyTextAccumulator, shouldBufferBopomofoInsertedPreedit(_:),
insertBopomofoPreeditText(_:replacementRange:), insertText(_:replacementRange:),
markedSelectedRange, and NSNotFound).

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: a45bad3a-f9de-4da6-84d2-dd166e45d9eb

📥 Commits

Reviewing files that changed from the base of the PR and between 9795cb4 and 2090f4d.

📒 Files selected for processing (3)

• Sources/GhosttyNSView+IMEComposition.swift
• Sources/GhosttyTerminalView.swift
• cmuxTests/CJKIMEMarkedSelectionTests.swift

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 81a237a. Configure here.}

[austinywang]

@coderabbitai review

[coderabbitai]

✅ Actions performed

Review triggered.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[austinywang]

@coderabbitai resume

Stale CodeRabbit review: all inline threads from this review are resolved and superseded by later commits/checks.

Stale CodeRabbit review: all inline threads from this review are resolved and superseded by later commits/checks.

[coderabbitai]

Actionable comments posted: 1

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (2)

cmuxTests/CJKIMEMarkedSelectionTests.swift (2)

19-19: 🧹 Nitpick | 🔵 Trivial | ⚡ Quick win

Remove the empty deinit.

Per repo convention, deinit {} should only be included when there is a specific lifecycle reason (e.g., releasing resources, removing observers). This test class has no teardown logic. Based on learnings, the empty deinit should be removed unless it serves a purpose.

♻️ Proposed fix

-    deinit {}
-

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cmuxTests/CJKIMEMarkedSelectionTests.swift` at line 19, Remove the empty
deinitializer from the CJKIMEMarkedSelectionTests class: delete the `deinit {}`
declaration (no other changes required) so the class follows repo convention of
only defining deinit when it performs teardown work like removing observers or
releasing resources.

---

1-474: 🧹 Nitpick | 🔵 Trivial | 💤 Low value

File length approaching the coherence threshold.

This test file is 474 lines, which is acceptable given its clear single responsibility (CJK IME marked-selection testing). However, it's approaching the 800-line threshold for coherent files. Future additions may warrant splitting into separate test files (e.g., Korean-specific and Zhuyin-specific test classes).

As per coding guidelines, "A new production Swift file must not exceed 400 lines without a clear single responsibility, or 800 lines even when the responsibility is mostly coherent."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@cmuxTests/CJKIMEMarkedSelectionTests.swift` around lines 1 - 474, The file
CJKIMEMarkedSelectionTests.swift is nearing the 800-line guideline threshold;
split it into smaller, focused test files to keep each file under the policy
limits: extract Korean-specific tests (e.g.,
testKeyDownForKoreanPostCompositionHorizontalArrowsForwardsToTerminal and
KoreanArrowProbe) into a new CJKIMEKoreanTests.swift and Zhuyin-specific tests
(e.g., testSuppressesTerminalForwardingWhenZhuyinStartsMarkedText,
testBuffersZhuyinComponentInsertTextAsPreedit, etc.) into
CJKIMEZhuyinTests.swift, relocate shared helpers and types
(makeHostedTerminalWindow, keyEvent(_:keyCode:windowNumber:),
HostedTerminalWindow, and any debug setup/teardown logic) into a common test
helper file or base test class (e.g., GhosttyTestHelpers or CJKIMETestBase) that
both new test files import, and ensure each new file keeps its
`@MainActor/XCTestCase` annotations and restores any debug overrides in defer
blocks as in the original tests.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@Sources/GhosttyTerminalView.swift`:
- Around line 7755-7760: The call to
shouldSuppressGhosttyKeyForwardingAfterIMEHandling is passing translationEvent
(which has Ghostty-modified modifiers) but the suppression predicate needs the
raw physical AppKit key event; replace the translationEvent argument with the
original/unmodified physical NSEvent (the raw AppKit event variable used earlier
in this scope) so shouldSuppressGhosttyKeyForwardingAfterIMEHandling(before:
markedStateBefore, after: (markedText.string, markedSelectedRange),
accumulatedText: accumulatedText, event: <physical NSEvent>, inputSourceId:
keyboardIdBefore) receives the unmodified event for correct IME suppression
decisions.

---

Outside diff comments:
In `@cmuxTests/CJKIMEMarkedSelectionTests.swift`:
- Line 19: Remove the empty deinitializer from the CJKIMEMarkedSelectionTests
class: delete the `deinit {}` declaration (no other changes required) so the
class follows repo convention of only defining deinit when it performs teardown
work like removing observers or releasing resources.
- Around line 1-474: The file CJKIMEMarkedSelectionTests.swift is nearing the
800-line guideline threshold; split it into smaller, focused test files to keep
each file under the policy limits: extract Korean-specific tests (e.g.,
testKeyDownForKoreanPostCompositionHorizontalArrowsForwardsToTerminal and
KoreanArrowProbe) into a new CJKIMEKoreanTests.swift and Zhuyin-specific tests
(e.g., testSuppressesTerminalForwardingWhenZhuyinStartsMarkedText,
testBuffersZhuyinComponentInsertTextAsPreedit, etc.) into
CJKIMEZhuyinTests.swift, relocate shared helpers and types
(makeHostedTerminalWindow, keyEvent(_:keyCode:windowNumber:),
HostedTerminalWindow, and any debug setup/teardown logic) into a common test
helper file or base test class (e.g., GhosttyTestHelpers or CJKIMETestBase) that
both new test files import, and ensure each new file keeps its
`@MainActor/XCTestCase` annotations and restores any debug overrides in defer
blocks as in the original tests.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 14f0e44b-e84b-429a-a143-ba000a548ecb

📥 Commits

Reviewing files that changed from the base of the PR and between f2a1955 and afbfbe0.

📒 Files selected for processing (4)

• Sources/AppDelegate.swift
• Sources/GhosttyNSView+IMEComposition.swift
• Sources/GhosttyTerminalView.swift
• cmuxTests/CJKIMEMarkedSelectionTests.swift

[coderabbitai]

✅ Actions performed

Reviews resumed.

Stale CodeRabbit review: all inline threads from this review are resolved and superseded by later CodeRabbit approval on the current head commit.