Fix all unresolved CodeRabbit comments on this PR:

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

src/components/ai-edition/TranscriptEditor.tsx (1)

78-81: 🚀 Performance & Scalability | 🔵 Trivial | ⚡ Quick win

isInRange makes word rendering O(n²).

isInRange calls findIndex three times per word over the full words array, so rendering the transcript is quadratic in word count. For long transcripts this re-runs on every render and will stutter. Precompute the selected id set once (you already derive selectedRange); compute the [from, to] index window in a useMemo and do an O(1) membership check per word.

🤖 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 `@src/components/ai-edition/TranscriptEditor.tsx` around lines 78 - 81, The
word selection logic in TranscriptEditor is doing an O(n²) check because
`isInRange` repeatedly scans `transcript.words` for every rendered word.
Refactor the range selection path in `TranscriptEditor` to precompute the
selected index window or selected id set once with `useMemo` (using the existing
`selectedRange`/anchor/focus state), then replace the per-word `isInRange` call
with an O(1) membership check inside the word render loop.

src/lib/ai-edition/exporter/documentExporter.ts (1)

218-218: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Move this type-only import to the top with the other imports.

import type { ExportResult } sits at the bottom of the file. It works via hoisting, but placing it with the other imports improves readability and avoids surprising future readers.

🤖 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 `@src/lib/ai-edition/exporter/documentExporter.ts` at line 218, The type-only
import for ExportResult is separated from the rest of the imports, so move it
into the main import block at the top of documentExporter.ts. Keep the existing
import grouping consistent by placing import type { ExportResult } alongside the
other import statements near the file header, without changing any runtime
behavior.

Fix all unresolved CodeRabbit comments on this PR:

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

Fix all unresolved CodeRabbit comments on this PR:

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

Fix all unresolved CodeRabbit comments on this PR:

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

electron/ai-edition/llm-call.ts-101-109 (1)

101-109: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Add a timeout to provider fetches.

Both LLM requests can remain pending indefinitely. Add an AbortController timeout so a stalled provider does not hang the chat flow.

Proposed fix

+const REQUEST_TIMEOUT_MS = 60_000;
+
+async function fetchWithTimeout(url: string, init: RequestInit): Promise<Response> {
+	const controller = new AbortController();
+	const timeout = setTimeout(() => controller.abort(), REQUEST_TIMEOUT_MS);
+	try {
+		return await fetch(url, { ...init, signal: controller.signal });
+	} finally {
+		clearTimeout(timeout);
+	}
+}

-		const res = await fetch(url, {
+		const res = await fetchWithTimeout(url, {

Apply the same replacement in callAnthropic.

Also applies to: 145-154

🤖 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 `@electron/ai-edition/llm-call.ts` around lines 101 - 109, Add an
AbortController-based timeout to the provider fetches in llm-call.ts so stalled
requests cannot hang indefinitely; update both callOpenAI and callAnthropic to
create a controller, pass its signal into fetch, and abort after the configured
timeout. Make sure the timeout is cleared once the request completes or fails,
and keep the existing request/response handling in place while wiring the signal
through the fetch call.

electron/ai-edition/chat-service.ts-47-60 (1)

47-60: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Avoid committing failed chat turns to history.

Line 55 stores the user message before the LLM call, but Line 72 returns an error without returning that message. Failed requests then silently remain in history and are included in later prompts.

Proposed fix

 	const history = messagesByProject.get(projectId) ?? [];
 	const userMessage: AiEditionChatMessage = {
 		id: uuidv4(),
 		role: "user",
 		content: message,
 		createdAt: new Date().toISOString(),
 	};
-	history.push(userMessage);
-	messagesByProject.set(projectId, history);
+	const nextHistory = [...history, userMessage];
 
 	const llmMessages: ChatMessage[] = [
 		{ role: "system", content: SYSTEM_PROMPT },
-		...history.slice(-20).map((m) => ({ role: m.role, content: m.content })),
+		...nextHistory.slice(-20).map((m) => ({ role: m.role, content: m.content })),
 	];

-	history.push(assistantMessage);
-	messagesByProject.set(projectId, history);
+	messagesByProject.set(projectId, [...nextHistory, assistantMessage]);

Also applies to: 71-82

🤖 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 `@electron/ai-edition/chat-service.ts` around lines 47 - 60, In
chat-service.ts, the send/chat flow is committing the user turn to
messagesByProject before the LLM call and leaving it there when the request
fails. Update the logic around the userMessage append and the error path that
returns early so failed turns are rolled back or only persisted after a
successful response, ensuring messagesByProject and the llmMessages history used
by the chat request do not include unsuccessful turns.

electron/ai-edition/llm-call.ts-135-143 (1)

135-143: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Keep Anthropic thinking budget below max_tokens. reasoningEffort === "high" sends thinking: { type: "enabled", budget_tokens: 4096 } with max_tokens: 1024, which Anthropic’s manual extended-thinking API rejects. Raise max_tokens for this path or lower the thinking budget.

🤖 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 `@electron/ai-edition/llm-call.ts` around lines 135 - 143, The Anthropic
request built in llm-call.ts can send an invalid extended-thinking configuration
when reasoningEffort is "high" because the thinking budget exceeds max_tokens.
Update the request construction in the body assembly so the high-thinking path
in the llm-call flow keeps budget_tokens below max_tokens, either by increasing
max_tokens for that case or reducing the thinking budget before attaching
body.thinking.

electron/ai-edition/llm-call.ts-48-49 (1)

48-49: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Fix the Gemini base URL and Anthropic thinking budget.

• Google’s OpenAI-compatible endpoint needs /v1beta/openai, otherwise requests go to .../v1beta/chat/completions and miss the compatibility path.

Proposed fix

		case "google":
-			return "https://generativelanguage.googleapis.com/v1beta";
+			return "https://generativelanguage.googleapis.com/v1beta/openai";

• callAnthropic() sets thinking: { type: "enabled", budget_tokens: 4096 } with max_tokens: 1024; keep the thinking budget below max_tokens or high-reasoning requests will fail.

🤖 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 `@electron/ai-edition/llm-call.ts` around lines 48 - 49, Update the Google
Gemini base URL in the switch that returns the provider endpoint so the
OpenAI-compatible path includes /v1beta/openai, and adjust callAnthropic() so
its thinking budget stays below max_tokens. Locate the base URL logic near the
google case and the Anthropic request payload in callAnthropic(), then change
the Anthropic thinking configuration to use a lower budget_tokens value than the
1024 max_tokens setting.

src/lib/ai-edition/store/projectStore.ts-168-181 (1)

168-181: 🚀 Performance & Scalability | 🟠 Major | ⚡ Quick win

Don't snapshot every live preview tick into undo history.

setLive() calls this path on every slider onChange, and setDocument() currently structuredClones the full document and pushes a history entry each time. That will spam undo steps and add avoidable cloning work during drags. Add a way to skip history recording for transient updates, then snapshot once on commit.

🤖 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 `@src/lib/ai-edition/store/projectStore.ts` around lines 168 - 181,
setDocument() is recording undo history for every transient live-preview update,
which causes excessive snapshots and cloning work. Update the projectStore flow
so setLive() can call setDocument() with a flag or equivalent to skip
pushHistory/structuredClone for slider onChange updates. Keep the existing
history behavior for committed document changes, and ensure the undo snapshot is
only created once when the final committed update is applied.

src/lib/ai-edition/store/useTimeline.ts-251-279 (1)

251-279: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Use source coordinates for the split point.

splitTimeSec is a timeline position, but Line 255 writes it directly into left.sourceEndSec. For any clip that starts later in the source or already sits later on the timeline, this produces the wrong source range and can even make the left half invalid. Compute the source-space split once and use it for both halves.

Proposed fix

 			const target = document.timeline.clips[targetIdx];
+			const sourceSplitSec =
+				target.sourceStartSec + (splitTimeSec - target.timelineStartSec);
 			const left = {
 				id: createId("clip"),
 				assetId: target.assetId,
 				sourceStartSec: target.sourceStartSec,
-				sourceEndSec: splitTimeSec,
+				sourceEndSec: sourceSplitSec,
 				timelineStartSec: target.timelineStartSec,
 				timelineEndSec: splitTimeSec,
 				wordRefs: [] as string[],
 				origin: "user" as const,
@@
 			const right = {
 				id: createId("clip"),
 				assetId: target.assetId,
-				sourceStartSec: target.sourceStartSec + splitTimeSec - target.timelineStartSec,
+				sourceStartSec: sourceSplitSec,
 				sourceEndSec: target.sourceEndSec,
 				timelineStartSec: splitTimeSec + duration,
 				timelineEndSec: target.timelineEndSec + duration,

🤖 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 `@src/lib/ai-edition/store/useTimeline.ts` around lines 251 - 279, The split
logic in useTimeline is mixing timeline and source coordinates, which makes the
left/right clip ranges incorrect. In the split handling block that builds left,
insert, and right, compute the source-space split point from
target.sourceStartSec plus the timeline offset before assigning clip boundaries,
then use that value for left.sourceEndSec and right.sourceStartSec so both
halves stay valid.

src/components/ai-edition/LeftPanel.tsx-256-304 (1)

256-304: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Use the selected asset’s transcript in the source transcript modal.

The readiness indicator is asset-specific via document.transcripts, but the modal always formats document.transcript, so opening a non-primary asset can show the wrong transcript.

🤖 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 `@src/components/ai-edition/LeftPanel.tsx` around lines 256 - 304, The
SourceTranscriptModal in LeftPanel is always using document.transcript, which
can show the wrong text for a non-primary asset. Update the modal data flow so
it uses the currently selected srcTranscriptAsset’s matching transcript from
document.transcripts (by assetId) instead of the global document.transcript, and
keep the existing transcript-ready indicator logic aligned with that
asset-specific lookup.

src/components/ai-edition/Bottombar.tsx-69-88 (1)

69-88: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Render the trim/skip lane that this component already receives.

skipRanges is part of BottombarProps and is passed by NewEditorShell, but it is never destructured or rendered, so trim regions cannot be selected/removed from the region lanes.

Proposed fix

 	zoomRegions,
+	skipRanges,
 	annotationRegions,
 	speedRegions,

 						{zoomRegions.length > 0 ? (
 							<LaneRow
 								label="Zoom"
...
 							/>
 						) : null}
+						{skipRanges.length > 0 ? (
+							<LaneRow
+								label="Trim"
+								kind="skip"
+								items={skipRanges.map((s) => ({
+									id: s.id,
+									startMs: Math.round(s.startSec * 1000),
+									endMs: Math.round(s.endSec * 1000),
+									label: "Trim",
+								}))}
+								selection={selection}
+								onSelect={onSelectRegion}
+								onRemove={onRemoveRegion}
+							/>
+						) : null}

Also applies to: 224-268

🤖 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 `@src/components/ai-edition/Bottombar.tsx` around lines 69 - 88, The Bottombar
component is already receiving skip ranges via BottombarProps, but they are not
being destructured or rendered, so trim/skip regions cannot appear in the lane
UI. Update Bottombar and the region-lane rendering logic to accept and display
skipRanges alongside the existing zoomRegions, annotationRegions, and
speedRegions, and wire the lane items into onSelectRegion and onRemoveRegion so
skip regions can be selected and removed.

src/components/ai-edition/Bottombar.tsx-360-410 (1)

360-410: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Avoid nesting an interactive button inside another button.

The delete control is a <button> inside the lane item <button>, which is invalid HTML and can break click/focus behavior. Make the row a non-button wrapper with separate select/delete buttons.

🤖 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 `@src/components/ai-edition/Bottombar.tsx` around lines 360 - 410, The lane
item in Bottombar currently nests the delete <button> inside the main selectable
<button>, which is invalid and can break interaction handling. Refactor the item
rendered in the lane list so the outer container is a non-interactive wrapper
(or equivalent) and move selection to a separate control while keeping the
delete action as its own button. Update the click/focus behavior around the item
rendering logic so onSelect(kind, item.id) and onRemove(kind, item.id) remain
distinct without nested interactive elements.

src/components/ai-edition/NewEditorShell.tsx-681-683 (1)

681-683: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Route titlebar New/Open through the unsaved-work prompt.

The keyboard shortcuts call promptUnsaved, but the titlebar actions open the project/new-project modals directly; selecting or creating from those modals can discard dirty work without confirmation.

Proposed direction

+const openProjectWithPrompt = useCallback(async () => {
+	const choice = await promptUnsaved("open");
+	if (choice !== "cancel") setOpenProjectOpen(true);
+}, [promptUnsaved]);
+
+const newProjectWithPrompt = useCallback(async () => {
+	const choice = await promptUnsaved("new");
+	if (choice !== "cancel") setNewProjectOpen(true);
+}, [promptUnsaved]);
+
 actions={{
-	openProject: () => setOpenProjectOpen(true),
-	newProject: () => setNewProjectOpen(true),
+	openProject: () => void openProjectWithPrompt(),
+	newProject: () => void newProjectWithPrompt(),

Also applies to: 801-812

🤖 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 `@src/components/ai-edition/NewEditorShell.tsx` around lines 681 - 683, The
titlebar actions for openProject and newProject in NewEditorShell bypass the
unsaved-work flow, so route them through the same promptUnsaved path used by the
keyboard shortcuts before opening the modals. Update the actions wiring and any
related handlers around promptUnsaved, openProject, and newProject so selecting
or creating a project always confirms/discards dirty changes consistently.

src/components/ai-edition/NewEditorShell.tsx-296-308 (1)

296-308: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Prompt before writing in Save As.

handleSaveAs saves the current document before asking for a title, so canceling the prompt still persists changes, marks the project clean, and shows “Project saved”.

Proposed fix

 const handleSaveAs = useCallback(async () => {
 	const doc = useProjectStore.getState().document;
 	if (!doc) return;
 	try {
-		const result = await nativeBridgeClient.aiEdition.save(doc);
-		if (!result.success || !result.document) {
-			throw new Error(result.error ?? "Failed to save project");
-		}
 		const title = window.prompt("Save project as", doc.project.title);
-		if (!title || title === doc.project.title) {
-			markClean();
-			toast.success("Project saved");
+		if (!title) {
 			return;
 		}
-		const renamed = { ...doc, project: { ...doc.project, title } };
-		await saveDocument(renamed);
-		toast.success(`Saved as "${title}"`);
+		const next = title === doc.project.title ? doc : { ...doc, project: { ...doc.project, title } };
+		await saveDocument(next);
+		toast.success(title === doc.project.title ? "Project saved" : `Saved as "${title}"`);
 	} catch (err) {

🤖 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 `@src/components/ai-edition/NewEditorShell.tsx` around lines 296 - 308, The
Save As flow in handleSaveAs is prompting too late, so canceling still saves and
marks the project clean. Move the title prompt before calling
nativeBridgeClient.aiEdition.save, and only proceed with the save when a valid
new title is provided; otherwise return without calling markClean or
toast.success. Keep the logic in handleSaveAs and preserve the existing success
path for when the title differs from doc.project.title.

src/components/ai-edition/Preview.tsx-124-132 (1)

124-132: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Don't layer a second transport on top of VirtualPreview.

VirtualPreview already renders play/pause, restart, a readout, and a scrubber in src/components/ai-edition/VirtualPreview.tsx:182-228. Rendering a second toolbar here creates two independent playback state machines (isPlaying there vs playing/loop here), so the UI can drift and the preview ends up with duplicate transport controls. Split VirtualPreview into a headless preview or add a showControls={false} path before wiring shell-specific controls.

Also applies to: 154-258

🤖 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 `@src/components/ai-edition/Preview.tsx` around lines 124 - 132, The Preview
component is rendering a second transport on top of VirtualPreview, which
duplicates playback controls and creates separate playback state. Update
VirtualPreview to support a headless/no-controls mode (for example via a
showControls prop) or split out the preview rendering so Preview can provide
only its shell-specific controls. Use the VirtualPreview and Preview components
to locate the transport wiring and remove the built-in play/pause, restart,
readout, and scrubber UI when embedded here.

src/components/ai-edition/Preview.tsx-62-87 (1)

62-87: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Drive seeking in virtual time, not raw video.currentTime.

These handlers compare or assign timelineStartSec/timelineEndSec and currentTimeSec directly against videoEl.currentTime, but VirtualPreview uses the underlying <video> in source-media time and translates virtual positions itself via locateVirtualPosition(...).sourceTimeSec in src/components/ai-edition/VirtualPreview.tsx:47-79. Trimmed or reordered clips will jump to the wrong frame when using previous/next, restart, or the scrubber. Route these actions through a virtual-time seek API (seekTarget or a dedicated callback) instead of mutating the DOM video element directly.

Also applies to: 97-101, 211-221

🤖 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 `@src/components/ai-edition/Preview.tsx` around lines 62 - 87, The
previous/next clip handlers in Preview are seeking against the raw video element
time instead of virtual timeline time, which can jump incorrectly for trimmed or
reordered clips. Update handlePrevClip and handleNextClip to drive seeking
through the same virtual-time translation used by VirtualPreview and
locateVirtualPosition, using seekTarget or a dedicated seek callback instead of
assigning videoEl.currentTime directly. Make sure the scrubber and restart paths
in Preview use the same virtual seek flow so timelineStartSec/timelineEndSec are
always mapped through the virtual position logic.

src/components/ai-edition/RightPanes.tsx-581-598 (1)

581-598: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Give the toggle control an accessible name.

This button only exposes pressed state. Without aria-label/aria-labelledby, screen readers have no way to tell which setting each toggle controls.

🤖 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 `@src/components/ai-edition/RightPanes.tsx` around lines 581 - 598, The Toggle
button in RightPanes.tsx only exposes pressed state and needs an accessible
name. Update the Toggle component to accept and pass through an `aria-label` or
`aria-labelledby` prop so each instance can describe the setting it controls,
while preserving the existing `checked`, `disabled`, and `onChange` behavior.

src/components/ai-edition/NewEditorShell.module.css-1197-1211 (1)

1197-1211: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

wide currently makes the modal narrower.

The base modal width is min(960px, 92vw), but .modalCard.wide overrides it to min(720px, 92vw). Every modal asking for wide therefore renders smaller than the default.

🤖 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 `@src/components/ai-edition/NewEditorShell.module.css` around lines 1197 -
1211, The .modalCard.wide rule is making the modal smaller instead of wider by
overriding the base width with a narrower value. Update the .modalCard.wide
styling in NewEditorShell.module.css so it increases the width relative to
.modalCard, and keep the change scoped to the .modalCard and .modalCard.wide
selectors.

src/components/ai-edition/ProviderSettings.tsx-91-99 (1)

91-99: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Disconnect should also clear the active provider config.

This flow removes credentials, but it never clears the persisted config, so refreshing the snapshot can leave a disconnected provider marked as the active one.

🤖 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 `@src/components/ai-edition/ProviderSettings.tsx` around lines 91 - 99, The
handleDisconnect flow removes the API key but leaves the persisted active config
behind, so the disconnected provider can still appear selected after
llmGetSnapshot(). Update the ProviderSettings handleDisconnect logic to clear
the active provider config through the aiEdition/nativeBridgeClient path before
refreshing the snapshot, then keep setSnapshot in sync with the cleared state.

src/components/ai-edition/Titlebar.tsx-499-512 (1)

499-512: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Keep LANGS aligned with the shipped locale set.

The current list exposes "de" even though no German locale is in the supplied cohort, and it omits "ar" even though Arabic locale data is present. That makes one language selectable without translations and another impossible to select at all.

🤖 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 `@src/components/ai-edition/Titlebar.tsx` around lines 499 - 512, The LANGS
list in Titlebar.tsx is out of sync with the shipped locales: it includes an
unsupported German entry and misses Arabic. Update the LANGS constant to match
the actual locale cohort by removing the extra "de" option and adding the "ar"
option, keeping the labels consistent with the other entries so the locale
picker only shows supported translations.

src/components/ai-edition/Modals.tsx-207-234 (1)

207-234: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Don’t advertise keyboard navigation you didn’t implement.

The footer says ↑↓ and Enter work here, but the list has no active index or key handlers, so keyboard users cannot do what the hint promises.

🤖 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 `@src/components/ai-edition/Modals.tsx` around lines 207 - 234, The footer in
Modals.tsx is advertising keyboard controls that the list does not actually
support. Either implement real keyboard navigation in the relevant modal/list
component by adding an active index plus key handlers for ArrowUp, ArrowDown,
and Enter, or remove the misleading hint from the footer. Use the existing
modal/footer markup in Modals.tsx to locate the text and the selection logic in
the same component tree to wire the behavior consistently.

src/components/ai-edition/Modals.tsx-483-489 (1)

483-489: 🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Clamp the crop rectangle against its far edge too.

x/y and width/height are normalized independently, so x + width or y + height can still exceed 1. That can persist out-of-bounds crop regions even though each individual field is clamped.

🤖 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 `@src/components/ai-edition/Modals.tsx` around lines 483 - 489, The crop
normalization in handleApply only clamps x, y, width, and height independently,
so the resulting CropRegion can still extend past the right or bottom edge.
Update the logic in Modals.tsx within handleApply to clamp the rectangle by its
far edge as well, ensuring x + width and y + height never exceed 1 after
normalization. Keep the fix local to the CropRegion construction so the returned
next value is always fully in bounds.

src/components/ai-edition/ExportDialog.tsx-130-137 (1)

130-137: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Use a real reveal-in-folder action here.

The toast says "Show in folder", but openExternalUrl(\file://${pickedPath}`)` opens a file URL instead of revealing the parent directory, and raw filesystem paths are not valid file URLs on every platform.

🤖 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 `@src/components/ai-edition/ExportDialog.tsx` around lines 130 - 137, The “Show
in folder” toast action in ExportDialog is using openExternalUrl with a file://
URL, which opens the file path instead of revealing the containing directory.
Update the action inside the toast.success call to use a real reveal-in-folder
capability from window.electronAPI (or add one if needed) and pass pickedPath so
the OS/file manager opens the parent folder and highlights the exported file.
Keep the change localized to the ExportDialog export success handler.

src/components/ai-edition/ProviderSettings.tsx-68-76 (1)

68-76: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Block “Save & use” until the provider can actually authenticate.

A blank API-key form still saves config, and oauth-device/pat providers can also be saved even though the form says the connect flow is not implemented. That leaves the app targeting a provider the user still cannot use, so the failure is deferred to the first chat request.

Also applies to: 327-374, 443-447

🤖 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 `@src/components/ai-edition/ProviderSettings.tsx` around lines 68 - 76, The
save flow in ProviderSettings.handleSave currently persists config even when the
provider is not actually ready to authenticate, which lets users save unusable
providers. Update the “Save & use” path so it only completes when auth is valid:
require a non-empty apiKey for active.authKind === "api-key", and block or
disable saving for oauth-device and pat providers until their connect flow is
implemented. Make the same gating consistent across the related provider
UI/actions in ProviderSettings so the app only targets providers the user can
actually use.

src/components/ai-edition/RightPanes.tsx-139-177 (1)

139-177: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Most of the background picker is non-functional right now.

The "Upload custom" CTA has no handler, and the Color tab never calls set/setLive, so users cannot choose either a custom image or a solid color from this pane.

🤖 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 `@src/components/ai-edition/RightPanes.tsx` around lines 139 - 177, The
background picker in RightPanes is mostly inert: the “Upload custom” button has
no action, and the Color tab UI only displays values without updating settings.
Wire the Upload custom CTA to a real file/image picker flow, and connect the
Color tab controls in RightPanes to call set and/or setLive so selecting a color
or gradient actually updates wallpaper. Use the existing tab handling,
subTabRow/subTabStyle, and the wallpaper-related UI in this component to locate
and patch the missing interactions.

src/components/ai-edition/RightPanes.tsx-430-447 (1)

430-447: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Make these range inputs controlled.

Both the webcam-size slider and SliderCell use defaultValue, so the thumb position only reflects the initial render. After document loads, undo/redo, or other store updates, the numeric text rerenders but the sliders stay visually stale.

Also applies to: 633-644

🤖 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 `@src/components/ai-edition/RightPanes.tsx` around lines 430 - 447, The range
inputs in RightPanes are using defaultValue, so they only initialize once and
don’t stay in sync with store updates. Update the webcam size slider and the
SliderCell range control to be controlled by binding value to the current
settings state, and keep their onChange handlers updating the store as they do
now. Make sure the relevant components in RightPanes (including the SliderCell
block referenced in the same file) re-render the thumb position when document
state, undo/redo, or other live updates change the underlying preset values.

src/components/ai-edition/Modals.tsx-251-259 (1)

251-259: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

The selected starting template is discarded.

This modal tracks template, but onCreate only receives the title and the submit path never branches on the selected option. Every card currently creates the same project.

Also applies to: 315-368

🤖 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 `@src/components/ai-edition/Modals.tsx` around lines 251 - 259, The selected
template state in NewProjectModal is never used, so every submit creates the
same project regardless of the chosen card. Update the NewProjectModal flow so
the selected Template is carried through the create action, either by changing
onCreate to accept the template or by branching in the submit handler before
calling it, and make sure the card click/update logic and submit path both use
the current template value consistently.

src/components/ai-edition/ExportDialog.tsx-93-103 (1)

93-103: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Catch picker failures before leaving the dialog in configuring.

pickExportSavePath() sits outside the try/catch. If that bridge call rejects, phase never leaves "configuring", the primary action stays disabled, and the rejection escapes uncaught.

🤖 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 `@src/components/ai-edition/ExportDialog.tsx` around lines 93 - 103, The export
picker call in ExportDialog’s handleExport flow can reject outside the existing
try/catch, leaving the dialog stuck in "configuring". Move the
window.electronAPI?.pickExportSavePath?.(suggested) call into the same guarded
path used by the export logic, and ensure any rejection sets phase back to
"idle" and records the error via setError. Use the ExportDialog component and
its handleExport/pickExportSavePath flow to locate the fix.

src/components/ai-edition/Titlebar.tsx-79-201 (1)

79-201: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Route the titlebar copy through i18n.

This component reads the current locale, but every visible label/status here is hardcoded English, so changing language only updates the badge and leaves the titlebar itself untranslated.

🤖 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 `@src/components/ai-edition/Titlebar.tsx` around lines 79 - 201, The Titlebar
component still hardcodes all visible button labels, titles, and status text in
English, so it does not respond to locale changes. Update the text in
Titlebar.tsx to use the existing i18n/locale mechanism already available in this
component, replacing strings for ProjectNameField-related status, Saved/Unsaved
copy, New recording, Return to recorder, panel toggles, Export, Theme,
Shortcuts, and Settings with translated lookups. Keep the logic in Titlebar and
its action handlers the same, but route every user-facing string through the
translation layer so the entire titlebar localizes consistently.

src/components/ai-edition/Modals.tsx-446-480 (1)

446-480: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

The crop ratio math is inverted.

detectRatio() compares height / width against presets stored as width / height, and handleRatioChange() multiplies width by that same ratio when deriving height. A 16:9 crop will be detected/applied as 9:16, and vice versa.

Suggested fix

 function detectRatio(r: CropRegion): string {
 	const candidates = CROP_RATIOS.filter((c) => c.ratio !== null);
-	const ratio = r.width === 0 ? 0 : r.height / r.width;
+	const ratio = r.height === 0 ? 0 : r.width / r.height;
 	for (const c of candidates) {
 		if (c.ratio === null) continue;
 		if (Math.abs(ratio - c.ratio) < 0.01) return c.value;
 	}
 	return "free";
 }
@@
 	const handleRatioChange = (next: string) => {
 		setRatio(next);
 		const candidate = CROP_RATIOS.find((c) => c.value === next);
 		if (candidate?.ratio && wPct > 0) {
-			const newH = Math.round(wPct * candidate.ratio);
+			const newH = Math.round(wPct / candidate.ratio);
 			setHPct(Math.min(100, Math.max(1, newH)));
 		}
 	};

🤖 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 `@src/components/ai-edition/Modals.tsx` around lines 446 - 480, The crop ratio
logic in detectRatio and handleRatioChange is using the wrong orientation,
causing width/height presets to be treated as height/width. Update detectRatio()
to compare the stored crop ratios using the same width-to-height convention as
CROP_RATIOS, and adjust handleRatioChange() so the derived dimension is
calculated from that same ratio direction. Keep the behavior in CropModal
consistent when syncing initialRegion and when applying a selected ratio.

src/lib/ai-edition/store/editorSettings.ts-138-155 (1)

138-155: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Validate enum-like legacy settings before returning a typed snapshot.

legacyEditor is an unknown envelope, but aspectRatio, cropRegion, webcamLayoutPreset, and webcamMaskShape are passed through with ?? only. A malformed project can therefore surface unsupported values into controls even though getEditorSettings() promises a typed result. Add guards for those unions/shapes instead of trusting the stored payload.

🤖 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 `@src/lib/ai-edition/store/editorSettings.ts` around lines 138 - 155, The
legacy-to-typed conversion in getEditorSettings is letting unknown enum-like
values through for aspectRatio, cropRegion, webcamLayoutPreset, and
webcamMaskShape by using direct nullish coalescing only. Update the
normalization logic in editorSettings.ts so these fields are validated against
the allowed EditorSettings unions/shapes before being returned, falling back to
DEFAULT_EDITOR_SETTINGS when legacy data is malformed, while keeping the
existing str/num/bool handling intact.

src/components/ai-edition/NewEditorShell.tsx-761-764 (1)

761-764: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Expand the right panel when compact rail tools are selected.

RightRailCompact only changes rightPane; because rightCollapsed remains true, clicking Background or Video effects does not reveal the selected pane.

Proposed fix

 <RightRailCompact
-	onChange={setRightPane}
+	onChange={(pane) => {
+		setRightPane(pane);
+		setRightCollapsed(false);
+	}}
 	onCrop={() => setCropOpen(true)}
 />

🤖 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 `@src/components/ai-edition/NewEditorShell.tsx` around lines 761 - 764, The
compact right rail updates rightPane in NewEditorShell, but the panel stays
collapsed because rightCollapsed is never cleared. Update the RightRailCompact
onChange flow so selecting a tool like Background or Video effects also expands
the right rail by setting rightCollapsed to false, alongside setRightPane, and
keep the existing onCrop behavior unchanged.

src/components/ai-edition/Bottombar.tsx-91-110 (1)

91-110: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Derive Auto Focus from persisted settings.

autoFocus is initialized once from settings.autoFocusAll; after switching projects or receiving external settings updates, the button can show/toggle stale state.

Proposed fix

-	const [autoFocus, setAutoFocus] = useState(() => settings.autoFocusAll);
...
-							on={autoFocus}
+							on={settings.autoFocusAll}
 							onClick={() => {
-								const next = !autoFocus;
-								setAutoFocus(next);
+								const next = !settings.autoFocusAll;
 								void set({ autoFocusAll: next });
 							}}

🤖 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 `@src/components/ai-edition/Bottombar.tsx` around lines 91 - 110, The Auto
Focus toggle in Bottombar is holding local state that can drift from
settings.autoFocusAll, so update it to derive from the persisted settings rather
than initialize once with useState. Use the Bottombar component’s
settings/autoFocusAll source as the single source of truth and make the button’s
on/onClick behavior reflect that value so project switches or external settings
changes stay in sync.

src/components/ai-edition/Titlebar.tsx-55-63 (1)

55-63: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

timeAgo() will go stale after the first render.

Because this text is derived from Date.now() without any ticking state, the saved indicator can stay stuck at "just now"/"5m ago" until some unrelated rerender happens.

Also applies to: 93-103

🤖 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 `@src/components/ai-edition/Titlebar.tsx` around lines 55 - 63, The saved-time
label in timeAgo() is derived from Date.now() but will not update on its own
after the initial render, so the text can become stale. Update the Titlebar
component so the displayed saved status recomputes on a timer or ticking state
(for example inside the component that uses timeAgo and the saved indicator),
and ensure any effect is cleaned up properly so the label refreshes periodically
without relying on unrelated rerenders.

src/components/ai-edition/LeftPanel.tsx (1)

390-395: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Don’t present the model label as selectable unless it changes the active LLM config.

The “Choose model” button only increments cycleIndex; send() still calls chatRun(projectId, text) with no selected model/config update, so the displayed model can diverge from the backend model. The native client contract shown for chatRun only sends { projectId, message }.

Also applies to: 568-573

🤖 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 `@src/components/ai-edition/LeftPanel.tsx` around lines 390 - 395, The model
label in LeftPanel should not imply a real selection unless it actually updates
the active LLM config used by chatRun. Update the “Choose model” flow around
modelAlternatives/modelLabel and the send() path so the displayed label always
matches the model config passed to the backend, or otherwise make it clear it is
only a preview; do the same for the related model label logic in the other
referenced block.

src/components/ai-edition/RightPanelStack.tsx (1)

11-12: 🗄️ Data Integrity & Integration | 🔵 Trivial | 🏗️ Heavy lift

Stop editing ai-edition documents through legacy casts.

This inspector imports the legacy AnnotationRegion type and mutates document.annotations / document.legacyEditor via as unknown as ... casts. That bypasses the AxcutDocument schema contract entirely, so any SSOT drift here becomes a runtime data-shape bug instead of a compile-time one.

Also applies to: 146-156, 239-347

🤖 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 `@src/components/ai-edition/RightPanelStack.tsx` around lines 11 - 12, Remove
the legacy `AnnotationRegion` dependency and stop using `as unknown as` to
mutate `document.annotations` and `document.legacyEditor` in `RightPanelStack`.
Update the inspector to work directly against the `AxcutDocument` schema by
using the proper typed fields/helpers from `AxcutDocument`, `AxcutClip`, and
`AxcutTranscript`, and adjust the affected update paths in the referenced edit
handlers so the document shape stays type-safe without casts.

src/lib/ai-edition/schema/index.test.ts (1)

170-188: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Rename this test to match what it asserts.

Line 170 says this input is rejected, but Lines 171-188 explicitly assert that documentSchema.parse(...)
does not throw. That mismatch makes the contract harder to read and regresses the same issue that was
flagged earlier.

Suggested rename

-it("documentSchema rejects v2 doc without v3 envelopes", () => {
+it("documentSchema defaults missing v3 envelopes on a v3 document", () => {

🤖 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 `@src/lib/ai-edition/schema/index.test.ts` around lines 170 - 188, The test
name in documentSchema.parse validation is misleading because it says the input
is rejected, but the assertion uses not.toThrow. Rename the test in the
documentSchema test block to match the actual behavior being asserted, and keep
the expectation aligned with the renamed description so the contract is clear
when reading the documentSchema suite.

src/components/ai-edition/VirtualPreview.tsx (1)

36-39: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

Report and clear the video element when the active source changes.

This effect only depends on onVideoElement, so after the keyed <video> is replaced for another asset, the parent can keep using the stale element for transport controls.

Suggested fix

-	// report the video element up
-	useEffect(() => {
-		onVideoElement?.(videoRef.current);
-	}, [onVideoElement]);
-
 	const isProgrammaticSeekRef = useRef(false);

 	const virtualDurationSec = useMemo(() => totalVirtualDuration(clips), [clips]);
 	const activeSource = videoSources[sourceIndex] ?? null;
+
+	useEffect(() => {
+		onVideoElement?.(videoRef.current);
+		return () => onVideoElement?.(null);
+	}, [onVideoElement, activeSource?.src]);

🤖 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 `@src/components/ai-edition/VirtualPreview.tsx` around lines 36 - 39, The video
element reporting effect in VirtualPreview only tracks onVideoElement, so the
parent can retain a stale video node after the keyed <video> is swapped for a
new active source. Update the useEffect in VirtualPreview to depend on the
active source identity used for the keyed video element, and make it notify the
parent with the current videoRef.current whenever that source changes; also
clear the previous element on cleanup or before reporting the new one so
transport controls always point at the latest video element.

Fix all unresolved CodeRabbit comments on this PR:

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