fix(promptinput): keep bash-mode ! out of the local mirror (#1179)

[0xghost42]

Summary

Closes #1179. Typing ! into empty input is meant to enter bash mode and leave the prompt buffer empty (the ! only shows in the mode prefix). Before this patch, the ! stayed in the buffer and the cursor was wedged before it: e.g. !git status with the caret at position 0 instead of an empty buffer.

Root cause

useTextInput's default keystroke handler had a special case at the start of an empty buffer for the input-mode character:

if (cursor.isAtStart() && isInputModeCharacter(input)) {
  return cursor.insert(text).left()
}

That produced a cursor with text="!" and offset=0, which setValue("!", 0) then committed into the local mirror (liveValueRef="!", liveOffsetRef=0) and emitted as onChange("!").

PromptInput.detectModeEntry stripped the controlled parent value back to "" with cursor 0 — values the controlled props already held. Because the parent props ended up numerically identical to what they were before the keystroke, React did not re-render PromptInput, the useLayoutEffect in useTextInput never re-ran, and the local mirror kept ! at offset 0.

The next keystroke read liveValueRef="!" / liveOffsetRef=0, so inserts landed before the retained !, producing !git status with the cursor pinned at 0.

Fix

src/hooks/useTextInput.ts — at the same special case, emit onChange(text) as a one-shot mode-entry notification and return undefined so setValue is not called. The local mirror stays at "" / offset 0, the parent's strip remains a no-op on the controlled state, and the next character types into a clean buffer.

Test

New regression in src/components/TextInput.test.tsx:

• Mounts a controlled TextInput with a parent onChange that mirrors PromptInput's strip (setValue('') when the new value starts with !).
• Types ! then g then i then t.
• Asserts the rendered frame contains git, and does NOT contain !, !git, or git!.

Before the fix the test fails (Received: "git!"). After the fix all 4 tests in the file pass.

Validation

• bun test src/components/TextInput.test.tsx → 4/4 pass.
• bun test src/components/PromptInput/ src/hooks/ → 24/24 pass.
• bun run build clean.

Test plan

• Manual: type !git status on Ubuntu Linux 24.04 — buffer renders git status, mode prefix shows !, cursor stays at end of typed text.
• Manual: paste !ls -la into empty input — buffer renders ls -la, mode prefix shows !. (This path goes through detectModeEntry's multi-char branch and was already working.)
• CI green.

[jatmn]

Findings

• [P2] Preserve the existing buffer when ! is typed at offset 0
  src/hooks/useTextInput.ts:488
  This branch now runs whenever the cursor is at offset 0, not only when the buffer is empty. If the prompt already contains git status, the user moves the cursor to the beginning, and then types ! to switch that command into bash mode, useTextInput now calls onChange('!') instead of sending the full !git status value. PromptInput.detectModeEntry then sees prevInputLength > 0 with a one-character value, does not enter bash mode, and the existing prompt is replaced with just !. That path is already documented as supported by the detectModeEntry test for prepending ! to non-empty existing text, so please restrict the one-shot empty-mirror path to an empty buffer or otherwise preserve the full inserted value for non-empty buffers.

[Vasanthdev2004]

Code Review

PR: #1182 — fix(promptinput): keep bash-mode ! out of the local mirror
Author: @0xghost42
Files changed: 2 (+81 -1)

Blockers

src/hooks/useTextInput.ts:488 — The fix breaks ! typing at position 0 on non-empty buffers.

The condition cursor.isAtStart() && isInputModeCharacter(input) runs whenever the cursor is at offset 0, not just when the buffer is empty. If the user has git status in the buffer, moves to position 0, and types !, the fix now calls onChange('!') instead of onChange('!git status'). This replaces the entire prompt with just !.

The previous reviewer (@jatmn) already flagged this — needs a cursor.isEmpty() guard or equivalent.

Non-Blocking

• Test uses Bun.sleep(25) between keystrokes. Fine for now, but could be flaky on slow CI. Consistent with other tests in the file though.

Looks Good

• Root cause analysis in the PR description is excellent — very clear explanation of the mirror desync issue.
• The test component (BashModeStrippingTextInput) accurately mimics the real parent behavior.
• Minimal, focused change — only touches what's needed.
• Good inline comment explaining the "why" behind the fix.

---

Verdict: Changes Requested — the non-empty buffer regression is a real blocker.

[jatmn]

Please rebase before making changes, that should fix the smoke issues currently.

[gnanam1990]

Took an independent look — @jatmn's finding holds. The new branch in useTextInput.ts:488 keys off cursor.isAtStart() rather than an empty buffer, and emits onChange('!') with return undefined, so prepending ! to existing non-empty text (cursor moved to start of git status, then !) replaces the buffer with just ! instead of entering bash mode on the full value. One thing to add: the added regression test only exercises the empty-buffer path, so it wouldn't catch the non-empty-prepend regression @jatmn described — worth covering that case in the same test when you scope the fix to an empty buffer (or otherwise preserve the full inserted value). Happy to re-review once updated.