fix(terminal): preserve IME input, drafts and attachment readiness - #334

## Change

안녕하세요. Web UI의 라이브 터미널에서 Shift+Enter가 일반 Enter로 전달되는 문제를 확인하여 수정을
제안드립니다.

v0.3.43에서 브라우저 입력과 실제 PTY 수신 바이트를 확인해 보니, Enter와 Shift+Enter가 모두 `\r`로
전달되고 있었습니다. 줄바꿈을 기대하며 Shift+Enter를 누르면 에이전트 TUI에서는 일반 Enter와 같은 입력을 받게
됩니다. Web UI에서 에이전트가 실행 중인 pane의 터미널 보기를 열고 Shift+Enter를 누르면 재현할 수 있습니다.

이번 변경은 xterm이 IME 확정 처리를 마친 뒤 발생하는 `onKey`에서 Shift+Enter를 식별하고, 바로 이어지는
`onData`의 CR만 기존 Alt+Enter와 같은 `\x1b\r`로 변환합니다. xterm의 원래 키 처리를 끝까지
거치므로, `compositionend` 직후 전송 대기 중인 한글이 있어도 한글 → 줄바꿈 순서가 유지됩니다. 별도의 타이머나
xterm 내부 API는 사용하지 않습니다. 기존 연결 상태·입력 권한 검사와 일반 Enter·Alt+Enter, 다른 수정 키
조합은 그대로 유지합니다. 입력이 잠긴 상태에서는 매핑을 예약하지 않습니다.

기존 브라우저 회귀 테스트에 Enter·Shift+Enter·Alt+Enter의 실제 키 이벤트, 입력 프레임의 중복 전송 여부,
`isComposing` 및 `keyCode === 229` 보호 검증을 추가했습니다. 실제 xterm textarea에서
`compositionend("한")` 직후 같은 작업 안에서 Shift+Enter를 발생시켜, 전송 순서가 `["한",
"\x1b\r"]`이고 지연된 중복 전송이 없는지도 검증합니다. 변경 이력도 Unreleased 항목에 기재했습니다.

## Validation

별도로 만든 테스트 세션에서 원본과 수정본을 각각 실행했습니다. 실제 브라우저에서 세 키를 순서대로 누르고, raw-mode
PTY가 받은 바이트를 확인했습니다.

| 키 입력 | 수정 전: v0.3.43 (`1608cce`) | 수정 후 |
| --- | --- | --- |
| Enter | `0d` (CR) | `0d` (CR) |
| Shift+Enter | `0d` (CR) | `1b 0d` (ESC + CR) |
| Alt+Enter | `1b 0d` (ESC + CR) | `1b 0d` (ESC + CR) |

**수정 전 캡처**

![수정 전: Shift+Enter가 Enter와 같은 바이트를
전달](https://raw.githubusercontent.com/WOULDU-pres/herdr-web-ui/evidence/terminal-shift-enter/before.png)

**수정 후 캡처**

![수정 후: Shift+Enter가 Alt+Enter와 같은 바이트를
전달](https://raw.githubusercontent.com/WOULDU-pres/herdr-web-ui/evidence/terminal-shift-enter/after.png)

캡처는 이번 검증용으로 생성한 터미널에서 촬영했습니다. 운영 중인 터미널이나 사용자 대화는 포함하지 않았으며, 이미지와 [검증
로그](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-shift-enter/validation.txt)는
포크의 별도 자료 브랜치에 보관했습니다.

검증 환경: Linux(WSL), Herdr 0.9.3, Bun 1.4.2, lockfile에 맞는 Playwright
Chromium 153.

- `bun run generate:types --check`: 통과
- `bun run typecheck`: 통과
- `bun run build`: 통과
- `bun run test:unit`: 1,036건 통과, 3건 건너뜀, 실패 없음
- `bun scripts/ui-regression.ts`: 전체 통과(확정 직후 IME 입력 순서 회귀 테스트 포함)
- #334의 `93caf43` xterm 백포트를 적용한 별도 검증 빌드: 같은 실제 브라우저 입력 순서 테스트 통과

IME 순서 검증: [기존 구현에서 실패한
재현](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-shift-enter/ime-before.log),
[수정 후 전체 브라우저
테스트](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-shift-enter/ime-after.log),
[#334 백포트
검증](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-shift-enter/ime-backport.log).

관련 브라우저 테스트 로그([전체
원본](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-shift-enter/ui-regression.log)):

```text
PASS terminal Shift+Enter sends a newline chord once and preserves Enter, Alt+Enter and IME
PASS pending IME commit precedes Shift+Enter without duplicate text
PASS terminal clipboard paste sends text once and preserves Ctrl+C
PASS: browser frames and received PTY bytes match for all three keys
```

전체 통합 테스트도 실행했으며, 결과는 169건 통과·8건 실패였습니다. 실패는 모두 이번 PR에서 변경하지 않은
`server/codex-binding.contract.test.ts`에서 테스트용 Codex 프로세스를 찾지 못하는
경우였습니다. 수정하지 않은 원본 `1608cce`의 별도 checkout에서 대표 실패 1건을 다시 실행하여 아래와 같은 오류가
동일하게 발생하는 것을 확인했습니다. 이 환경의 통합 테스트 전체 통과 여부까지 확인한 것은 아니라는 점을 함께 남깁니다.

```text
error: no codex process in <owned-test-pane>
  at server/codex-binding.contract.test.ts:52
```

확인해 주시면 감사하겠습니다. 프로젝트에서 선호하시는 키 입력 방식이나 테스트 보완 사항이 있으면 그에 맞춰 조정하겠습니다.

---------

Co-authored-by: e2e <e2e@example.com>

Bumps package.json and herdr-plugin.toml to 0.3.44, moves the Unreleased
notes under [0.3.44] - 2026-10-03, adds the compare link, and removes a
merge marker line left in CHANGELOG.md by #351's conflict resolution.
### Added
- Native Windows x64 installation through `install.ps1`, with the same
herdr plugin startup
and updates. Windows needs Bun and Git, without Node or WSL, and keeps
using the terminal
screen mirror. ([#330](#330)
by @JJLiebig)
- Settings can choose Chat, Terminal, or Auto as the default view for
panes. Changing it resets
remembered pane views on this device. Shell panes still open in
Terminal.
([#325](#325) by
@Haeminway1)
- Choose the terminal input line or direct typing on desktop as well as
touch screens, and
customize or unbind the app's Mod+Shift shortcuts with conflict checks
and reset.
  ([#334](#334))

### Changed
- In the sidebar's By workspace view, a workspace with one pane keeps
its numbered header and
can be folded, renamed and reordered from that header, just like a
workspace with several panes.
  ([#332](#332) by @beomq)
- The top usage panel shows larger percentages, remaining-capacity
colors and compact reset
countdowns with their local time. Expired reset times stay hidden until
fresh usage arrives.
([#336](#336) by
@Haeminway1)
- Pane titles that contain a directory show its last folder, and
matching workspace and folder
names appear once in the sidebar and command palette, including Windows
paths.
([#337](#337) by
@Haeminway1)
- In the sidebar's By workspace view, a pane row no longer repeats the
workspace its header names.
Its second line shows the folder only when neither the title nor the
workspace already says it.
  ([#358](#358))

### Fixed
- GJC conversations are resolved from the foreground process directory
when it differs from
the pane directory.
([#333](#333) by @Kinetic27)
- Direct terminal input preserves rapid IME commits when punctuation
arrives before composition
timers run, with a bounded xterm 5.5 backport and Korean final-consonant
regression checks.
Native Android/Gboard checks also fixed stale editor text after
Backspace breaking the next
Hangul word. Secret entry now refuses a held or unready attachment and
failed PTY writes.
  ([#334](#334))
- New session in the browser demo opens the new session instead of
leaving a blank page.
  ([#355](#355))
- Terminal input keeps multi-character IME commits and emoji while
disconnected, waits for the
attachment before sending keys, and reports input failures. Unsent input
lines survive pane,
lens and mode changes; late acknowledgements preserve replacement edits.
Composition keeps
its Enter and Send button, and terminal key-bar taps wait until it
finishes.
  ([#334](#334))
- The background tasks list stays below the app header on a short
screen, such as a phone held
  sideways, and scrolls inside the room above the message box.
  ([#354](#354))
- The mobile message box sits closer to the home indicator while
retaining its base spacing.
([#335](#335) by
@Haeminway1)
- Adding or reconnecting a remote PC says when herdr is not running or
not answering there,
  instead of only `Bridge verification failed (500)`.
  ([#353](#353))
- Shift+Enter in the live terminal sends the same newline chord as
Alt+Enter, instead of plain
Enter that submits an agent's message. Pending IME text is sent before
the newline chord.
([#339](#339) by
@WOULDU-pres)
- On Windows, and from an install path with a space or a non-ASCII
character, the app page loads
instead of the "not built yet" notice. The Windows installer replaces a
copy from a release
  without Windows support, and says so when a release cannot run there.
  ([#356](#356))
- On an iPhone home screen app, a text field that keeps its focus with
the keyboard down no
  longer leaves a status-bar band under the message box.
([#349](#349) by
@Haeminway1)
- After `/new` in an OmO pane, the chat stops showing the conversation
before it: OmO writes the
new session's file only with its first message, and until then the chat
shows what a fresh OmO
shows. ([#351](#351) by
@WOULDU-pres)
- On a Windows PC, an omo pane shows omo's mark instead of Claude's, and
its chat finds the
conversation: process words with `bun.exe`, backslashes and a drive
letter read as omo, and the
  session folder is named as omo's engine names a Windows folder.
([#342](#342) by
@Haeminway1)
- The chat lens finds an omo pane's conversation when its process keeps
its sessions outside
`~/.omo/agent` (`OMO_CODING_AGENT_DIR`, `SENPI_CODING_AGENT_DIR` or
`PI_CODING_AGENT_DIR`).
([#350](#350) by
@Haeminway1) passes. After the merge: .

)

## Change

0.3.44 (#334) made the input line and direct typing a choice on a
desktop too. To offer it there, the key bar was shown on a fine-pointer
desktop at 769 px and wider with every key hidden except the keyboard
button that switches the two. The result is a full-width bar under the
terminal that holds one button, on a device that has a hardware keyboard
and never showed that bar before.

- **Desktop:** the bar is hidden again (`KeyBar.css`), as it was up to
0.3.43.
- **The keyboard button** is rendered only on a touch screen (`coarse`),
as before #334. A desktop window 768 px wide or narrower still shows the
key bar, as it always did, without that button.
- **Kept from #334:** Settings → Terminal input mode (Automatic, Input
line, Direct typing) works on a desktop, so the input line is still
available there. Touch screens are unchanged. The IME, draft and
readiness fixes are not touched.

`docs/terminal-input.md` and the changelog are updated.

## Validation

- `scripts/terminal-input-regression.ts` (the desktop step of
`ui-regression.ts`, 1280 wide) switched modes by clicking the removed
button. It now switches through the Settings select, and first asserts
that `.key-bar` is not visible and that no "Type straight into the
terminal" button exists.
- `bun scripts/ui-regression.ts`: 45 PASS, exit 0, including `terminal
input: desktop mode, …` and `touch terminal input line sends whole
lines, Enter alone, and yields to direct typing` (the touch button). An
earlier run of the same code stopped at `Timed out: suggestion painted`,
a step before the terminal input ones that waits on a shell `printf` in
a test pane; the rerun passed unchanged.
- Owned test session at 1280 wide: with Automatic, the terminal fills
the pane down to the bottom edge with no bar; with Input line, the input
line and its Send button show and no key bar does. No image is attached
here.
- `bun run typecheck`, `bun run build`: pass. `bun run test:unit`: 1059
pass, 3 skip.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

## Change

안녕하세요. macOS에서 Web UI의 라이브 터미널을 사용할 때 Cmd+Backspace가 한 글자만 삭제하는 문제의 수정을
제안드립니다.

현재 xterm은 Cmd+Backspace도 일반 Backspace와 같은 DEL(`0x7f`)로 전달합니다. Mac의
Command 키 의미가 반영되지 않아, 입력줄 끝에서 Cmd+Backspace를 눌러도 줄 전체가 지워지지 않습니다.

이번 변경은 macOS의 Cmd+Backspace를 터미널에서 줄 앞부분을 삭제하는 Ctrl+U(`0x15`)로 전달합니다.
커서가 줄 끝에 있으면 입력줄 전체가 삭제됩니다. xterm이 대기 중인 IME 입력을 먼저 확정한 뒤 발생하는 `onKey`에서
키를 식별하고, 바로 이어지는 `onData`만 변환하므로 한글 → 삭제 순서를 보존합니다. 입력이 잠긴 상태에서는 매핑을
예약하지 않습니다.

일반 Backspace, Ctrl/Alt/Shift가 함께 눌린 조합과 macOS 이외의 Meta+Backspace는 기존 동작을
유지합니다. Shift+Enter 수정(#339)과 분리하여 `main` 기준의 별도 브랜치에서 작성했습니다. 최신
`main`(`bb7ef50`)과의 충돌을 정리했으며, 이미 병합된 #339와 #334의 입력 준비 검사·사용자 단축키 설정을
유지했습니다. 두 키 조합을 함께 검증했습니다.

## Validation

Linux(WSL), Herdr 0.9.3, Bun 1.4.2, lockfile에 맞는 Playwright Chromium
153에서 검증했습니다. 브라우저의 별도 컨텍스트에 MacIntel 플랫폼과 Command 키 이벤트를 재현했으며, 실제
macOS/Safari 기기 테스트는 수행하지 않았습니다.

| 키 입력 | 수정 전 | 수정 후 |
| --- | --- | --- |
| Backspace | `7f` (DEL) | `7f` (DEL) |
| Cmd+Backspace | `7f` (DEL) | `15` (Ctrl+U) |
| Ctrl+U | `15` (Ctrl+U) | `15` (Ctrl+U) |

**수정 전 캡처**

![수정 전: Cmd+Backspace가 한 글자 삭제 신호를
전달](https://raw.githubusercontent.com/WOULDU-pres/herdr-web-ui/evidence/terminal-command-backspace/before.png)

**수정 후 캡처**

![수정 후: Cmd+Backspace가 Ctrl+U와 같은 줄 삭제 신호를
전달](https://raw.githubusercontent.com/WOULDU-pres/herdr-web-ui/evidence/terminal-command-backspace/after.png)

캡처는 별도로 생성한 테스트 터미널의 실제 PTY 수신 바이트입니다. WebSocket 입력 프레임과 수신 바이트가 일치하는지도
확인했습니다. [검증
자료](https://github.com/WOULDU-pres/herdr-web-ui/tree/evidence/terminal-command-backspace)는
포크의 별도 브랜치에 보관했습니다.

새 브라우저 회귀 테스트에서는 다음을 확인합니다.

- Cmd+Backspace가 Ctrl+U를 한 번만 전송하며, 다른 Backspace 조합과 비 Mac 플랫폼은 기존 시퀀스를
유지합니다.
- `compositionend("한")` 직후 Cmd+Backspace를 발생시켜 `["한", "\x15"]` 순서와 지연된
중복 전송이 없음을 확인합니다. 조합 중 플래그는 줄 삭제로 바꾸지 않습니다.
- 별도 테스트 pane의 실제 Node readline 프로세스에 `remove this entire input`을 입력한 뒤
Cmd+Backspace를 눌러, 입력 버퍼가 빈 문자열이 되는 것을 확인합니다.

검증 결과:

- `bun run generate:types --check`, `bun run typecheck`, `bun run
build`: 통과
- `bun run test:unit`: 1,059건 통과, 기존 3건 건너뜀, 실패 없음
- `bun scripts/ui-regression.ts`: 전체 통과
- 최신 `main`에 포함된 #334 xterm 백포트: 같은 Mac 키·IME 순서·실제 readline 삭제 테스트를 포함한
전체 브라우저 검사 통과

관련 [브라우저 원본
로그](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-command-backspace/ui-regression.log):

```text
PASS Cmd+Backspace clears the entire input in a real owned readline PTY
PASS macOS Cmd+Backspace sends Ctrl+U once and preserves other Backspace combinations
PASS pending IME text precedes Cmd+Backspace without duplicates; active composition stays native
PASS non-macOS Meta+Backspace keeps native DEL
```

전체 통합 테스트 재실행은 170건 통과·9건 실패였습니다. 8건은 변경하지 않은
`server/codex-binding.contract.test.ts`의 테스트용 Codex 프로세스 탐지 실패로, 수정 없는
원본에서도 대표 실패를 확인한 상태입니다. 나머지 1건은 변경하지 않은 `server/updater.test.ts`의
supervisor 종료 코드 검사가 0 대신 1을 받은 경우이며, 해당 테스트를 단독 재실행하면 통과했습니다(1건 통과, 12건
필터링, 실패 없음). 이번 PR에서 이 서버 파일들을 수정하지 않습니다. [검증
로그](https://github.com/WOULDU-pres/herdr-web-ui/blob/evidence/terminal-command-backspace/validation.txt)에
최신 통과 범위와 이 제한을 함께 기록했습니다.

확인해 주시면 감사하겠습니다. 프로젝트에서 선호하시는 키 입력 방식이나 보완할 사항이 있으면 말씀 부탁드립니다.

Co-authored-by: Hako <devswha@gmail.com>

)

## Problem

On a phone in input-line mode (the touch default), the whole key bar
goes grey and ignores taps while a word is still composing in the input
line. With Gboard that is most of the time, because every word composes
until a space; with a Korean or Japanese IME it lasts until the syllable
or phrase commits. Esc and ^C cannot stop an agent, an arrow cannot
answer a menu, Ctrl/Alt/Shift cannot be armed, and the keyboard-mode
toggle is dead.

## Cause

`PaneTerminal` passed `onComposing={setComposing}` to `TerminalInput`,
so the input line's `compositionstart`/`compositionend` wrote the same
`composingRef` and `composing` state that guard xterm's deferred IME
commit. That flag disables `<KeyBar>` and short-circuits `pressKey`,
`toggleModifier` and `toggleDirect`. The guard exists so a key-bar key
cannot overtake xterm's pending commit, but the input line's text never
goes through xterm or `onData`: it is submitted separately with
`socket.submit`. There was nothing to protect, only taps to drop.

## Fix

- `PaneTerminal.tsx`: stop wiring `TerminalInput`'s composition into
`setComposing`. The xterm host `compositionstart`/`compositionend`
listeners (and the pane reset) stay the only writers of
`composingRef`/`compositionCommitPendingRef`, so the xterm
deferred-commit gate is unchanged.
- `TerminalInput.tsx`: remove the now unused `onComposing` prop and its
unmount cleanup. The line keeps its own local `composing` ref that
guards its Send button and Enter.
- `docs/terminal-input.md`: the key bar waits only for a direct-typing
(xterm) composition; an input-line composition leaves it live.
- Review follow-up (ca26dd6): the direct-typing toggle is the one key
that still waits for an input-line composition. Switching mid-word would
unmount the line with the IME composition open, and the word could be
lost or typed into the pane. `TerminalInput` reports its composition
through `onComposing` again. `PaneTerminal` tracks it apart from the
xterm composition, `toggleDirect` returns early on it, and `KeyBar` gets
`directDisabled`, which greys out only that key.

## Tests

- New phone block in `scripts/terminal-input-regression.ts` (390x844,
touch, line mode): fill the line, dispatch `compositionstart` on
`.terminal-input-text`, assert Esc, ^C and Ctrl are enabled, tap Esc and
^C, expect exactly two `input` frames `\x1b` and `\x03`, and the line's
value unchanged.
- `bun run check run --build bun scripts/terminal-input-regression.ts`:
PASS (all three blocks). With the two source files reverted the new
block fails on "Esc stays live while the input line composes".
- `bun run typecheck`: clean. `bun run test:unit`: 3069 pass, 0 fail, 8
skip. `scripts/changelog-fold.test.ts`, `scripts/pages-gate.test.ts`,
`scripts/conventions.test.ts`: pass.
- Not run: `scripts/ui-regression.ts` as a whole (it calls
`checkTerminalInput`, so CI runs the new block there too), `check
full`/`concurrent`.
- The phone regression also checks that `[data-key="direct"]` is
disabled during the composition and that a forced tap on it leaves the
line, its text and the pane untouched. Before the follow-up fix, that
check failed with "the direct-typing toggle waits for the line's
composition".

## Notes

- No open PR changes this gate (checked the diffs of #936, #946, #943,
#822, #812 for `onComposing`/`setComposing`). #936 edits
`TerminalInput.tsx` and #943/#946/#950/#951 edit
`PaneTerminal.tsx`/`KeyBar.tsx`, so a textual conflict is possible
depending on merge order, but none duplicates this fix.
- Left out on purpose: no separate gate on the key-bar Enter while the
line composes. It sends `\r` to the pane, not the line's text, just as
it does after a commit. The direct-typing toggle is gated (see the
follow-up in Fix).
- The regression uses synthetic composition events, so it certifies the
gate, not a physical IME.
- Gate added in #334 (c50c881) for both inputs; this keeps it for xterm
only.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

## Problem

On a phone, tapping chat **Send** while the keyboard is still composing
the last word does nothing, and tapping again does nothing either.
Gboard composes every word, and Korean/Japanese/Chinese keyboards keep
the last syllable open, so this hits ordinary typing. It works only
after a space or the keyboard's own Return.

## Cause

#937 made the Send button `preventDefault` its pointerdown so focus
stays in the message box and the keyboard stays up. With focus kept,
nothing in the tap ends the composition, and `Composer.send()` returns
early while `composingRef` is set (the guard from #334). Before #937 the
tap's blur ended the composition first. #959 fixed the same thing in the
terminal input line and noted that the chat composer needed the same
treatment.

## Fix

Same approach as #959 (`TerminalInput.tsx`):
- A Send tapped during composition blurs and refocuses the message box
inside the tap. Chrome ends the composition on the focus change and
fires `compositionend` at once; refocusing in the same gesture keeps the
keyboard up.
- On that `compositionend`, the send runs a tick later (after WebKit's
committing input event).
- If the keyboard does not commit within 300 ms, the tap is withdrawn
with "Finish the word, then send." ("Finishing the word…" shows while it
waits). A later commit never sends on its own. Both strings already
exist with translations.
- Enter during composition is unchanged (it still never sends).

## Tests

- `scripts/composer-fit-demo-regression.ts`: new phone case. It types
`hello `, opens a real composition `안` via CDP
`Input.imeSetComposition`, and taps Send. It expects the box to empty,
focus to stay in the box, and the next syllable to start in an empty
box. **Fails on main** (`value "hello 안"`) and passes here.
- `bun run check fast`: 3754 unit tests, typecheck and build pass.
- Not run: the native Gboard emulator. The commit mechanism is the one
#959 verified there on Chrome 133 + Gboard. iOS keyboard flicker from
the blur/refocus is still unobserved, as in #959.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>