bugs-reproducer

name	bugs-reproducer
description	Bug reproduction for the Lexxy rich text editor. Core editing bugs use Playwright (Selenium fallback); system-level bugs use Capybara. All local.
disable-model-invocation	true

Bugs Reproducer

Reproduce bugs in Lexxy — a rich text editor built on Lexical, distributed as a Rails gem and npm package. Lexxy replaces Trix as the rich text editor for Rails Action Text.

BUG REPORT → CLASSIFY → REPRODUCE → DIAGNOSE → FIX → VERIFY
                          ^^^^^^^^^
                          core editing bug → Playwright (Selenium fallback)
                          system-level bug → Capybara

Understanding Lexxy

Lexxy wraps Lexical in a set of custom elements (<lexxy-editor>, <lexxy-toolbar>, <lexxy-table-tools>, etc.) and extends it with custom nodes, extensions, and an Action Text integration layer. Before writing reproduction steps, understand which layer the bug likely lives in:

Editor core — The <lexxy-editor> custom element (src/elements/editor.js) owns the Lexical editor instance. It is FormAssociated, manages the value (sanitized HTML) lifecycle, handles Turbo reconnection, and dispatches events (lexxy:change, lexxy:focus, lexxy:blur, lexxy:initialize).

Custom Lexical nodes (src/nodes/) — These define how content is represented and rendered:

• ActionTextAttachmentNode / ActionTextAttachmentUploadNode — file and image attachments with ActiveStorage upload lifecycle
• CustomActionTextAttachmentNode — inline custom attachments (e.g., @mentions from prompts)
• ImageGalleryNode — container that auto-collapses/splits around previewable images
• HorizontalDividerNode — <hr> rendered as a decorator node
• ProvisionalParagraphNode — invisible placeholder paragraphs around non-selectable decorator nodes so the cursor can always be placed
• WrappedTableNode — tables wrapped in a scroll container

Extensions (src/extensions/) — Plugin-like modules:

• ProvisionalParagraphExtension — inserts/removes provisional paragraphs on tree changes
• HighlightExtension — color/background-color inline styles with palette canonicalization
• TrixContentExtension — backward-compatible import rules for Trix-generated HTML (<em>, <del>, <span style>, <pre language>)
• TablesExtension — table commands + Lexical bug workarounds
• AttachmentsExtension — attachment nodes + gallery collapse logic

Clipboard & paste (src/editor/clipboard.js) — Handles paste: plain text, URLs, markdown conversion (via marked), files, and HTML with <action-text-attachment> elements.

Selection & cursor (src/editor/selection.js) — Arrow-key navigation around decorator nodes, cursor containment within the editor, node selection management.

Toolbar (src/elements/toolbar.js) — Button state synced to selection format, hotkey dispatch, overflow menu compaction via ResizeObserver.

Prompt system (src/elements/prompt.js) — Trigger-based autocomplete (@, custom triggers) with inline, deferred, and remote data sources.

Trix Compatibility

Lexxy is a drop-in replacement for Trix. Content authored in Trix must render correctly in Lexxy, and vice versa. The TrixContentExtension handles Trix HTML import rules. Bugs in this area are high priority — they affect every app migrating from Trix to Lexxy.

The test/system/trix/ directory contains Capybara tests covering both conversion directions:

• from_trix_to_lexxy_test.rb — content created in Trix, then loaded/edited in Lexxy
• from_lexxy_to_trix_test.rb — content created in Lexxy, then loaded/edited in Trix

When reproducing Trix conversion bugs, use the Capybara test suite (see "Choosing the Right Test Suite" below).

Choosing the Right Test Suite

This is the first and most important decision. Every bug falls into one of two categories, and each has its own test suite. Getting this wrong wastes all subsequent work.

1. Core editing bugs → Playwright (test/browser/)

Bugs in the editor's client-side behavior: typing, cursor movement, selection, formatting, paste handling, toolbar interactions, keyboard shortcuts, node transforms, code blocks, tables, decorator nodes (dividers, embeds), and anything that lives in the editor's JS layer.

Playwright runs against a Vite dev server serving static HTML fixtures from test/browser/fixtures/. Each fixture configures a <lexxy-editor> with different attributes (toolbar, attachments, markdown, single-line, etc.). No Rails required.

Playwright tests run across Chromium, Firefox, and WebKit, giving cross-browser coverage locally without external services. Start with one browser for fast iteration, then confirm across available browsers once the reproduction is solid.

Note: WebKit may not launch on Arch Linux due to ABI-incompatible system libraries (see home/docs/development.md). Use Chromium and Firefox locally; WebKit coverage is guaranteed by CI.

Fallback: if Playwright can't trigger the bug after 3 attempts (e.g., needs real OS-level events), fall back to Selenium WebDriver scripts.

2. System-level bugs → Capybara (test/system/)

Bugs involving anything outside the editor's JS layer: Action Text rendering and persistence, ActiveStorage uploads, Trix ↔ Lexxy conversion, form submission, SGID resolution, prompt/mention resolution with Rails-backed data, gallery display after save, page refreshes (Turbo), authenticated storage, and any scenario where data must survive a save/load roundtrip.

Capybara tests run against the dummy Rails app (test/dummy/) using selenium_chrome_headless. Stay in Capybara for all attempts — Selenium scripts add nothing over Capybara for Rails integration bugs.

Extending the Test Suite

It is encouraged to extend the test suite when reproducing bugs. Add new fixtures, helpers, or test files as needed — the test suites are designed to be extended. Keep things consistent with the existing patterns:

• Playwright fixtures: add new .html files in test/browser/fixtures/ following the existing pattern (import editor.js, configure <lexxy-editor> with the attributes needed for the scenario)
• Playwright helpers: extend EditorHandle in test/browser/helpers/editor_handle.js or add new helper modules in test/browser/helpers/
• Capybara helpers: add or extend test helpers in test/test_helpers/
• Rails fixtures: add or modify fixture data in test/fixtures/
• Dummy app routes/views: add pages to the dummy app in test/dummy/ if a specific editor configuration is needed

Reproduce Like a Human (for the Action Under Test)

The action under test — the interaction you're trying to reproduce — must go through the browser's real event pipeline. Bugs live in event side effects (mousedown → focus → selectionchange → input) — programmatic calls skip them entirely.

Setup and inspection are fine with helpers. Use editor.setValue(), editor.select(), editor.value(), and editor.flush() freely for loading initial content, positioning the cursor before the test, and reading state for assertions. These are standard test helpers, not shortcuts.

editor.paste() is a special case. It dispatches a synthetic ClipboardEvent, which is how all existing paste tests work — browsers don't allow programmatic clipboard access, so this is the correct way to reproduce paste bugs. Use editor.paste() as the action under test for paste-related bugs.

Mixed clipboard payloads need extra care. When reproducing paste bugs that include files plus HTML or plain text, Firefox may ignore synthetic file payloads on ClipboardEvent. In those cases, extend editor.paste() to attach a custom clipboardData object for file-bearing events, but keep the existing ClipboardEvent path for ordinary text/HTML paste so you do not accidentally regress the rest of the paste suite.

The reproduction step itself must be human-like:

For the action under test	Use	Don't use
Type text	editor.send("text") (PW) / element.sendKeys("text") (Selenium)	editor.setValue("text")
Press Enter	editor.send("Enter") (PW) / element.sendKeys(Key.RETURN) (Selenium)	dispatchEvent(new KeyboardEvent(...))
Focus an element	editor.click() (PW) / element.click() (Selenium)	element.focus()
Click a toolbar button	editor.clickToolbarButton("bold") or click_on "Bold"	Direct command dispatch

executeScript / evaluate is primarily for reading — inspecting DOM state, checking cursor position, reading scroll offsets. Narrow setup exceptions exist (e.g., programmatic DOM manipulation when real interaction would destroy required preconditions), but never use it to produce the action under test.

Non-Negotiable: Read Every Screenshot

After taking any screenshot, read it and describe what you actually see — not what you expect to see. The screenshot is ground truth. DOM/CSS inspection is supporting evidence; it explains what the screenshot shows, it does not override it.

This skill gets invoked repeatedly in bug-fix loops where an agent or human is iterating on a fix. Screenshot validation during those loops is just as critical as during initial reproduction — it's the quality gate that prevents wasted iterations on visually broken fixes.

After each screenshot, check:

• What's actually visible? Describe the layout, element positions, and content you see.
• Does it match expectations? Compare explicitly against the expected state.
• Any visual anomalies? Unexpected scrollbars, overflow, clipping, empty space, misalignment, elements pushed out of position.
• Something looks off? Stop and investigate before proceeding. Do not assume it's fine.

Fix validation is especially skeptical

When testing a proposed fix or change, the screenshot is the only evidence that matters:

• Don't assume the fix worked because the code looks right.
• Compare the screenshot against both the original bug state AND the expected fixed state.
• A fix that introduces new visual problems (layout shifts, unexpected scrollbars, misplaced elements, large empty areas) is not a fix.

The anti-pattern this prevents

WRONG	Take screenshot → assume it confirms expectations → move on
RIGHT	Take screenshot → read it → describe what's actually visible → compare against expected state → flag any anomaly

When screenshots matter vs. when they don't

Not every reproduction needs screenshots. The test: if a human would ask "what did it look like?", screenshots are required.

• Visual bugs (layout, rendering, toolbar state, cursor position, formatting appearance): page.screenshot() → read and describe it before proceeding. The screenshot is the evidence.
• Diagnostic bugs (DOM state, console errors, wrong HTML output, event ordering): capture the relevant diagnostic data (editor HTML, console output, Lexical state). Screenshots are optional when they wouldn't change the reader's conclusion.

Non-Negotiable: Persevere

One test is one attempt. Different attempts vary the setup: content volume, cursor positioning, page lifecycle, scroll depth, timing.

Once reproduced, stop. A single reproduction with clear evidence is the goal. The 8-attempt minimum applies only when the bug hasn't been reproduced yet.

Write at least 8 separate attempts before concluding "not reproduced." Each must target at least 2 dimensions the previous attempts didn't cover:

General dimensions

• Event race timing — vary delay between two-step interactions
• Click timing — click then immediately act vs. click, pause, then act
• Keypress speed — rapid successive keypresses vs. deliberate single presses
• Focus transitions — type or press Enter immediately after clicking, before focus/selection stabilize
• Cursor positioning — beginning, middle, end of content; inside vs. outside a node boundary
• Selection state — text selected vs. collapsed cursor vs. no selection vs. node selection
• Content volume — empty editor vs. short content vs. long content (15+ lines)
• Repetition — same action 5+ times in a row
• Page lifecycle — fresh load vs. content already present vs. navigated away and back

Lexxy-specific dimensions

• Node context — paragraph, heading, blockquote, list item, code block, table cell — bugs often manifest only in specific node types
• Decorator node boundaries — interactions at the edge of attachments, dividers, and galleries where ProvisionalParagraphNode inserts invisible placeholders and Selection intercepts arrow keys
• Gallery state — single image, multiple images in gallery, images adjacent to non-image content (triggers splitAroundInvalidChild)
• Nested structures — list inside blockquote, table inside content, code block after list
• Toolbar state transitions — format, then type, then format differently, then undo — the toolbar state sync via #updateButtonStates() can get stale
• Paste variants — plain text, URL, markdown, HTML with attachments, HTML with Trix formatting, file paste
• Highlight interactions — apply color, change color, remove color, paste colored text — the StyleCanonicalizer and highlight format sync can conflict
• Undo/redo across node changes — undo after inserting an attachment, table, or divider — history state at decorator node boundaries is fragile
• Editor reconnection — if the bug involves Turbo, test with the page refresh fixture/flow

Label each attempt (v1, v2, v3...) and log which dimensions it covers.

For core editing bugs (Playwright path): after 3 failed attempts, consider switching to Selenium if the bug might require real OS-level events that Playwright's synthetic event model doesn't trigger. Continue the 8-attempt minimum across both methods.

For system-level bugs (Capybara path): stay in Capybara for all attempts. The Rails integration is the point — Selenium scripts against the sandbox app add nothing over Capybara.

Procedure

1. Parse the Bug Report and Choose Suite

Extract from the bug report:

• Step-by-step reproduction path
• Browser/OS from report (if specified)
• Which Lexxy layer is likely involved (editor core, specific node type, extension, paste, toolbar, prompt, Action Text integration)
• Whether the bug involves persistence (save/load) or is purely client-side

Choose the test suite first. Refer to the decision guide in "Choosing the Right Test Suite" above. This determines whether you write Playwright tests, Capybara system tests, or (as a last resort) Selenium scripts. Getting this wrong wastes attempts.

2. Plan Reproduction

Problem Summary — restate the bug:

• What the user experiences vs. expected behavior
• Which Lexxy subsystem is likely involved
• Environmental conditions (browser, editor config)
• Whether this is a Trix migration issue

Reproduction Plan — design around the report specifics. Each attempt targets a specific hypothesis. List the attempts you'll write, each naming:

• The hypothesis it tests
• The dimensions it varies
• Why it differs from the previous attempt

3. Ensure Local Environment is Ready

For Playwright tests (no Rails needed):

The Playwright config starts its own Vite dev server automatically when you run the tests. No manual setup required.

For Capybara tests or Selenium scripts (needs Rails):

# Check if the dev server is running
curl -s -o /dev/null -w "%{http_code}" http://lexxy.localhost:3000/ || echo "not running"

# If not running:
cd ~/Work/basecamp/lexxy
bin/setup   # installs deps, creates DB — idempotent, safe to re-run
bin/dev &

# Wait for server (up to 30s):
for i in $(seq 1 30); do
  curl -s -o /dev/null -w "%{http_code}" http://lexxy.localhost:3000/ 2>/dev/null && break
  sleep 1
done

4a. Reproduce with Playwright (preferred)

Write test files in test/browser/tests/ following the project's existing patterns.

Test pattern:

import { test } from "../test_helper.js"
import { expect } from "@playwright/test"
import { assertEditorHtml, assertEditorContent } from "../helpers/assertions.js"

test.describe("Bug reproduction: <description>", () => {
  test.beforeEach(async ({ page }) => {
    await page.goto("/")
    await page.waitForSelector("lexxy-editor[connected]")
  })

  test("v1: <hypothesis>", async ({ page, editor }) => {
    // Setup: load initial content (programmatic is OK here)
    await editor.setValue("<p>Initial content</p>")

    // Reproduction: human-like interactions only
    await editor.send("Enter")
    await editor.send("New line text")
    await editor.flush()

    // Evidence: assert the bug manifests
    await assertEditorHtml(editor, "<p>Expected HTML</p>")
  })
})

The test_helper.js provides an editor fixture (EditorHandle instance) alongside Playwright's page. See test/browser/helpers/editor_handle.js for the full API — key methods include send(), select(), paste(), clickToolbarButton(), flush(), value(), setValue(), and uploadFile().

Assertion helpers live in test/browser/helpers/assertions.js: assertEditorHtml(), assertEditorContent(), assertEditorPlainText(). These use expect.poll with editor.flush() to account for async Lexical state updates.

HTML fixtures in test/browser/fixtures/ provide different editor configurations (toolbar enabled/disabled, attachments, markdown, single-line, plain text, etc.). Load them with page.goto("/fixture-name.html"). If a bug requires a configuration that doesn't have a fixture, create one — follow the existing fixture pattern.

Run the test:

# Single test file, single browser
yarn test:browser:chromium -- test/browser/tests/<file>.test.js

# All browsers (Chromium + Firefox + WebKit)
yarn test:browser -- test/browser/tests/<file>.test.js

# Headed (visible browser window) — useful for visual bugs
yarn test:browser:headed -- test/browser/tests/<file>.test.js

Collecting evidence:

Playwright automatically captures screenshots on failure, traces on first retry, and videos on first retry. For explicit evidence during reproduction:

// Visual bugs: screenshot at each decisive step — then READ it
await page.screenshot({ path: "/tmp/bug-step-1.png" })

// Diagnostic bugs: editor state is the evidence
const html = await editor.innerHTML()
console.log("Editor HTML:", html)

For visual bugs, take and read screenshots at each step — a screenshot without analysis is not evidence. See Non-Negotiable: Read Every Screenshot. For diagnostic bugs where HTML output or console evidence suffices, capture that instead.

4b. Reproduce with Capybara (for Rails integration bugs)

For bugs involving Action Text, uploads, Trix conversion, or persistence, write a system test in test/system/.

Follow the existing patterns: tests extend ApplicationSystemTestCase, use the EditorHandler via find_editor, and use helpers from test/test_helpers/ (editor, focus, html, toolbar, trix, console helpers).

# Run a single system test
bin/rails test test/system/<file>.rb

# Run all system tests
bin/rails test:all

Testing Action Text rendered output (non-editor view):

The dummy app (test/dummy/) has a Post model with has_rich_text :body. Use the posts CRUD to test the full save → render round-trip:

1. Visit edit_post_path(posts(:empty)) (or :hello_world for pre-existing content)
2. Compose content in the editor (use find_editor, toolbar helpers, etc.)
3. click_on "Update Post" — saves via Action Text and redirects to the show page
4. Assert against the rendered HTML on the show page (not the editor)

The show page renders @post.body inside a data-controller="code-highlight" div, which invokes highlightCode() from Lexxy via a Stimulus controller. This is the same function apps use to apply Prism syntax highlighting to rendered Action Text content.

Key distinction: bugs that only appear in the rendered view (after save, outside the editor) are different from editor bugs. The rendered view runs through Action Text sanitization (Loofah) and then client-side Prism highlighting — both can strip or transform HTML that the editor preserved. Test these by asserting on the show page DOM, not the editor.

# Example: test that highlights survive rendering
test "highlights preserved in rendered view" do
  visit edit_post_path(posts(:empty))
  find_editor.send "some text"
  find_editor.select "some"
  apply_highlight_option "background-color", 1
  click_on "Update Post"

  # Now on the show page — assert against rendered Action Text output
  assert_selector "mark[style*='background-color']", text: "some"
end

Fixtures in test/fixtures/posts.yml and test/fixtures/action_text/rich_texts.yml provide pre-seeded post data.

4c. Reproduce with Selenium (fallback)

If Playwright doesn't reproduce after 3+ attempts and the bug might require real OS-level events, fall back to Selenium WebDriver .mjs scripts.

Ensure prerequisites:

ls /tmp/node_modules/selenium-webdriver >/dev/null 2>&1 || {
  cd /tmp && npm install selenium-webdriver
}
which chromedriver >/dev/null 2>&1 || echo "Install chromedriver"

Script pattern — write .mjs scripts in /tmp/:

import webdriver from '/tmp/node_modules/selenium-webdriver/index.js';
import chrome from '/tmp/node_modules/selenium-webdriver/chrome.js';
import fs from 'fs';

async function sleep(ms) { return new Promise(r => setTimeout(r, ms)); }
async function screenshot(driver, name) {
  const data = await driver.takeScreenshot();
  fs.writeFileSync(`/tmp/${name}.png`, Buffer.from(data, 'base64'));
}

const options = new chrome.Options();
options.addArguments('--headless=new', '--no-sandbox', '--window-size=1280,900');
const driver = await new webdriver.Builder()
  .forBrowser('chrome').setChromeOptions(options).build();

try {
  await driver.get('http://lexxy.localhost:3000');
  await driver.wait(webdriver.until.elementLocated(
    webdriver.By.css('lexxy-editor[connected]')
  ), 10000);

  const content = await driver.findElement(
    webdriver.By.css('.lexxy-editor__content')
  );
  await content.click();

  // Reproduction steps here — capture evidence after each
  await content.sendKeys('Hello there');
  await screenshot(driver, 'step-1');
  // READ the screenshot: describe what's visible, compare against expected state

  // Inspect via executeScript (read only)
  const html = await driver.executeScript(
    'return document.querySelector("lexxy-editor").value'
  );
  console.log('Editor value:', html);
} finally {
  await driver.quit();
}

5. Render Verdict

Verdict	Criteria
Reproduced	Bug manifests following reported steps
Not Reproduced	Bug does not manifest after 8+ attempts varying dimensions in the chosen suite
Intermittent	Bug manifests inconsistently (note frequency, e.g., 2/5 attempts)

Verdict format:

**Reproduction Verdict: [Reproduced / Not Reproduced / Intermittent]**

**Bug:** <title>
**Confidence:** High / Medium / Low
**Method:** Playwright / Capybara / Selenium
**Browsers tested:** Chromium, Firefox, WebKit (via Playwright) / Chrome (via Selenium)
**Attempts:** <n> attempts, <n> successful reproductions

**Lexxy subsystem:** <node type / extension / clipboard / selection / toolbar / Action Text>

**Steps Executed:**
1. <step> — observed: <describe what evidence shows>
2. <step> — observed: <describe what evidence shows>
3. <step> — observed: <describe what evidence shows> ← bug manifests here

**Evidence Summary:**
- <what was actually observed — actual vs expected>
- <DOM/Lexical state if inspected>
- <console errors if present>

**Observations:**
- <any differences from reported behavior>
- <additional conditions discovered>
- <timing sensitivity if intermittent>
- <which browsers reproduce: all, or specific>

6. If Reproduced — Leave the Test

When reproduced via Playwright, keep the test file in test/browser/tests/. It serves as both evidence and a regression test for the eventual fix. Name it descriptively:

test/browser/tests/bug_<short_description>.test.js

When reproduced via Capybara, keep the test in test/system/. When reproduced via Selenium, keep the script in /tmp/ and note the path in the verdict.

Common Bug Patterns in Lexxy

These are areas where bugs tend to cluster, based on the architecture:

Decorator node navigation — Arrow keys around attachments, dividers, and galleries. The Selection class manually intercepts LEFT/RIGHT/UP/DOWN at node boundaries. Chrome has specific workarounds for fake cursor elements.

Provisional paragraph lifecycle — The invisible paragraphs inserted around decorator nodes by ProvisionalParagraphExtension. They should appear when needed, disappear when not, and convert to real paragraphs when typed into. Bugs: they don't appear, they duplicate, they don't convert, or they persist when they shouldn't.

Gallery transforms — ImageGalleryNode auto-collapses adjacent images, splits around non-image children, and unwraps when left with a single child. The transform runs per-pass, so multiple non-images embedded may need multiple passes.

Mid-update editor reads — editor.read() (and anything that calls it: el.value, #readSanitizedEditorValue, isEmpty) commits the in-flight update when called inside an editor.update() callback, which silently skips all node transforms for that update. Bugs manifest as transform-derived state missing after load or setValue — list item numbering, list merging, gallery collapse, provisional paragraphs — while the same content behaves correctly when produced by editing. Some call sites are protected only by the cachedValue being warm, so reproductions may need a cold cache (fresh connect, initial value attribute).

Sanitizer allowlist round-trip drift — Editor HTML passes through two attribute allowlists: DOMPurify when computing the form value (global ALLOWED_HTML_ATTRIBUTES plus per-tag allowedElements in src/config/dom_purify.js and extensions) and Loofah on save (lib/lexxy/engine.rb). Lexical sometimes splits state across attributes asymmetrically — e.g. ordered lists import only ol[start] but export both ol[start] and li[value] — so an allowlist that keeps the cosmetic attribute while stripping the semantic one breaks silently on the next edit, not the current one. Reproduce with a full round trip (load → read value, or edit → save → re-edit), and check both allowlists whenever a new node type or attribute is added.

Paste handling edge cases — The clipboard handler has separate code paths for: only plain text, HTML with attachments, URLs (including Safari's text/uri-list), markdown, files, and content inside code blocks (bypasses Lexxy entirely). Bugs often appear at the boundary between these paths. Cross-app paste (Trix → Lexxy) is a key variant: Trix stores raw HTML in the content attribute of <action-text-attachment>, while Lexxy stores JSON-stringified HTML. Any code that parses content must handle both formats.

Soft line break selections — Shift+Enter produces LineBreakNodes inside a single paragraph. Formatting actions that work on top-level elements can accidentally treat multiple visible lines as one block, so reproduce selection bugs with real soft line breaks (or <br> setup) whenever the report mentions composer line breaks.

Mixed clipboard payload precedence — Some apps copy the same image as both an HTML <img> snippet and a real file. Reproductions need to exercise both payloads together because browsers may expose them differently, and Lexxy should prefer the upload path when attachments are supported.

Highlight style sync — The HighlightExtension keeps Lexical's highlight format bit in sync with inline CSS styles. Two TextNode transforms run on every mutation: one for sync, one for canonical palette enforcement. Infinite loop risk if the sync logic disagrees with Lexical's internal state.

Trix HTML import — The TrixContentExtension converts Trix's HTML output (<em>, <del>, <span style>, <pre language>) to Lexxy's model. Bugs here affect every migrated document.

Turbo reconnection — The <lexxy-editor> watches its connected attribute. Rapid Turbo morphs can stack reconnections. valueBeforeDisconnect can be null if timing is wrong.

Upload lifecycle — ActionTextAttachmentUploadNode.createDOM() starts the upload as a side effect. Lexical can call createDOM() multiple times (history restore). Guard logic prevents re-upload but can falsely block.

Toolbar focus after commands — Clicking a toolbar button with the mouse moves focus to the button. If focus isn't returned to the editor, keyboard shortcuts like Ctrl+Z go to the browser instead of Lexical. The toolbar calls editor.focus() after mouse-click commands to prevent this.

External drag-and-drop — When files are dragged from the OS file manager, the browser does NOT update the DOM selection to the drop coordinates. The drag caret is purely visual. Drop handlers must use document.caretRangeFromPoint(event.clientX, event.clientY) to determine the correct insertion point. Playwright's synthetic DragEvent DOES move the DOM selection, masking this class of bugs. To reproduce, intercept the drop event, restore the stale DOM selection, and call the drop logic with the event coordinates.