fix: harden AI subtitle fix response parsing

2026-03-20 03:16:46 -07:00 · 2026-03-08 16:01:40 -07:00
parent 8e319a417d
commit 93cd688625
22 changed files with 641 additions and 55 deletions
--- a/Harden-AI-subtitle-fix-against-non-SRT-model-responses.md
+++ b/Harden-AI-subtitle-fix-against-non-SRT-model-responses.md
@@ -0,0 +1,56 @@
+---
+id: TASK-117.1
+title: Harden AI subtitle fix against non-SRT model responses
+status: Done
+assignee:
+  - '@codex'
+created_date: '2026-03-08 08:22'
+updated_date: '2026-03-08 08:25'
+labels: []
+dependencies: []
+references:
+  - >-
+    /Users/sudacode/projects/japanese/SubMiner/launcher/youtube/subtitle-fix-ai.ts
+  - /Users/sudacode/projects/japanese/SubMiner/launcher/youtube/srt.ts
+  - >-
+    /Users/sudacode/projects/japanese/SubMiner/launcher/youtube/subtitle-fix-ai.test.ts
+parent_task_id: TASK-117
+---
+
+## Description
+
+<!-- SECTION:DESCRIPTION:BEGIN -->
+Prevent optional YouTube AI subtitle post-processing from bailing out whenever the model returns usable cue text in a non-SRT wrapper or text-only format. The launcher should recover safe cases, preserve original timing, and fall back cleanly when the response cannot be mapped back to the source cues.
+<!-- SECTION:DESCRIPTION:END -->
+
+## Acceptance Criteria
+<!-- AC:BEGIN -->
+- [x] #1 AI subtitle fixing accepts safe AI responses that omit SRT framing but still provide one corrected text payload per original cue while preserving original cue timing.
+- [x] #2 AI subtitle fixing still rejects responses that cannot be mapped back to the original cue batch without guessing and falls back to the raw subtitle file with a warning.
+- [x] #3 Automated tests cover wrapped-SRT and text-only AI responses plus an unrecoverable invalid response case.
+<!-- AC:END -->
+
+## Implementation Plan
+
+<!-- SECTION:PLAN:BEGIN -->
+1. Add failing tests in launcher/youtube/subtitle-fix-ai.test.ts for three cases: wrapped valid SRT, text-only one-block-per-cue output, and unrecoverable invalid output.
+2. Extend launcher/youtube/subtitle-fix-ai.ts with a small response-normalization path that first strips markdown/code-fence wrappers, then accepts deterministic text-only cue batches only when they map 1:1 to the original cues without changing timestamps.
+3. Keep existing safety rules: preserve cue count and timing, log a warning, and fall back to the raw subtitle file when normalization cannot recover a trustworthy batch.
+4. Run focused launcher unit tests for subtitle-fix-ai and SRT parsing; expand only if the change affects adjacent behavior.
+<!-- SECTION:PLAN:END -->
+
+## Implementation Notes
+
+<!-- SECTION:NOTES:BEGIN -->
+Implemented deterministic AI subtitle-response recovery for fenced SRT, embedded SRT payloads, and text-only 1:1 cue batches while preserving original timing and existing fallback behavior.
+
+Verification: bun test launcher/youtube/*.test.ts passed; bun run typecheck passed; repo-wide format check still reports unrelated pre-existing warnings in launcher/youtube/orchestrator.ts and scripts/build-changelog*.
+<!-- SECTION:NOTES:END -->
+
+## Final Summary
+
+<!-- SECTION:FINAL_SUMMARY:BEGIN -->
+Hardened the launcher AI subtitle-fix path so it can recover deterministic non-SRT model responses instead of immediately falling back. Added `parseAiSubtitleFixResponse` in `launcher/youtube/subtitle-fix-ai.ts` to normalize markdown-fenced or embedded SRT payloads first, then accept text-only responses only when they map 1:1 onto the original cue batch and preserve source timings. Added regression coverage in `launcher/youtube/subtitle-fix-ai.test.ts` for fenced SRT, text-only cue batches, and unrecoverable invalid output, plus a changelog fragment in `changes/task-117.1.md`.
+
+Verification: `bun test launcher/youtube/*.test.ts`, `bun run typecheck`, `bunx prettier --check launcher/youtube/subtitle-fix-ai.ts launcher/youtube/subtitle-fix-ai.test.ts`, and `bun run changelog:lint` passed. Repo-wide `bun run format:check:src` still reports unrelated pre-existing warnings in `launcher/youtube/orchestrator.ts` and `scripts/build-changelog*`.
+<!-- SECTION:FINAL_SUMMARY:END -->
--- a/changes/task-117.1.md
+++ b/changes/task-117.1.md
@@ -0,0 +1,4 @@
+type: fixed
+area: launcher
+
+- Hardened YouTube AI subtitle fixing so fenced SRT output and text-only one-cue-per-block responses can still be applied without losing original cue timing.
--- a/changes/task-130.md
+++ b/changes/task-130.md
@@ -0,0 +1,5 @@
+type: fixed
+area: launcher
+
+- Keep the background SubMiner process running after a launcher-managed mpv session exits so the next mpv instance can reconnect without restarting the app.
+- Reuse prior tokenization readiness after the background app is already warm so reopening a video does not pause again waiting for duplicate warmup completion.
--- a/launcher/smoke.e2e.test.ts
+++ b/launcher/smoke.e2e.test.ts
@@ -295,7 +295,7 @@ test('launcher mpv status returns ready when socket is connectable', async () =>
 });

 test(
-  'launcher start-overlay run forwards socket/backend and stops overlay after mpv exits',
+  'launcher start-overlay run forwards socket/backend and keeps background app alive after mpv exits',
  { timeout: LONG_SMOKE_TEST_TIMEOUT_MS },
  async () => {
    await withSmokeCase('overlay-start-stop', async (smokeCase) => {
@@ -310,7 +310,6 @@ test(
      const appStartPath = path.join(smokeCase.artifactsDir, 'fake-app-start.log');
      const appStopPath = path.join(smokeCase.artifactsDir, 'fake-app-stop.log');
      await waitForJsonLines(appStartPath, 1);
-      await waitForJsonLines(appStopPath, 1);

      const appStartEntries = readJsonLines(appStartPath);
      const appStopEntries = readJsonLines(appStopPath);
@@ -325,7 +324,7 @@ test(
      assert.match(result.stdout, /Starting SubMiner overlay/i);

      assert.equal(appStartEntries.length, 1);
-      assert.equal(appStopEntries.length, 1);
+      assert.equal(appStopEntries.length, 0);
      assert.equal(mpvEntries.length >= 1, true);

      const appStartArgs = appStartEntries[0]?.argv;
@@ -337,9 +336,6 @@ test(
      assert.equal((appStartArgs as string[]).includes(smokeCase.socketPath), true);
      assert.equal(appStartEntries[0]?.subminerMpvLog, smokeCase.mpvOverlayLogPath);

-      const appStopArgs = appStopEntries[0]?.argv;
-      assert.deepEqual(appStopArgs, ['--stop']);
-
      const mpvFirstArgs = mpvEntries[0]?.argv;
      assert.equal(Array.isArray(mpvFirstArgs), true);
      assert.equal(
--- a/launcher/youtube/subtitle-fix-ai.test.ts
+++ b/launcher/youtube/subtitle-fix-ai.test.ts
@@ -0,0 +1,126 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { applyFixedCueBatch, parseAiSubtitleFixResponse } from './subtitle-fix-ai';
+import { parseSrt } from './srt';
+
+test('applyFixedCueBatch accepts content-only fixes with identical timing', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんいちは
+
+2
+00:00:03,000 --> 00:00:04,000
+世界
+`);
+  const fixed = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんにちは
+
+2
+00:00:03,000 --> 00:00:04,000
+世界
+`);
+
+  const merged = applyFixedCueBatch(original, fixed);
+  assert.equal(merged[0]?.text, 'こんにちは');
+});
+
+test('applyFixedCueBatch rejects changed timestamps', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんいちは
+`);
+  const fixed = parseSrt(`1
+00:00:01,100 --> 00:00:02,000
+こんにちは
+`);
+
+  assert.throws(() => applyFixedCueBatch(original, fixed), /timestamps/i);
+});
+
+test('parseAiSubtitleFixResponse accepts valid SRT wrapped in markdown fences', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんいちは
+
+2
+00:00:03,000 --> 00:00:04,000
+世界
+`);
+
+  const parsed = parseAiSubtitleFixResponse(
+    original,
+    '```srt\n1\n00:00:01,000 --> 00:00:02,000\nこんにちは\n\n2\n00:00:03,000 --> 00:00:04,000\n世界\n```',
+  );
+
+  assert.equal(parsed[0]?.text, 'こんにちは');
+  assert.equal(parsed[1]?.text, '世界');
+});
+
+test('parseAiSubtitleFixResponse accepts text-only one-block-per-cue output', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんいちは
+
+2
+00:00:03,000 --> 00:00:04,000
+世界
+`);
+
+  const parsed = parseAiSubtitleFixResponse(
+    original,
+    `こんにちは
+
+世界`,
+  );
+
+  assert.equal(parsed[0]?.start, '00:00:01,000');
+  assert.equal(parsed[0]?.text, 'こんにちは');
+  assert.equal(parsed[1]?.end, '00:00:04,000');
+  assert.equal(parsed[1]?.text, '世界');
+});
+
+test('parseAiSubtitleFixResponse rejects unrecoverable text-only output', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんいちは
+
+2
+00:00:03,000 --> 00:00:04,000
+世界
+`);
+
+  assert.throws(
+    () => parseAiSubtitleFixResponse(original, 'こんにちは\n世界\n余分です'),
+    /cue block|cue count/i,
+  );
+});
+
+test('parseAiSubtitleFixResponse rejects language drift for primary Japanese subtitles', () => {
+  const original = parseSrt(`1
+00:00:01,000 --> 00:00:02,000
+こんにちは
+
+2
+00:00:03,000 --> 00:00:04,000
+今日はいい天気ですね
+`);
+
+  assert.throws(
+    () =>
+      parseAiSubtitleFixResponse(
+        original,
+        `1
+00:00:01,000 --> 00:00:02,000
+Hello
+
+2
+00:00:03,000 --> 00:00:04,000
+The weather is nice today
+`,
+        'ja',
+      ),
+    /language/i,
+  );
+});
--- a/launcher/youtube/subtitle-fix-ai.ts
+++ b/launcher/youtube/subtitle-fix-ai.ts
@@ -0,0 +1,213 @@
+import type { LauncherAiConfig } from '../types.js';
+import { requestAiChatCompletion, resolveAiApiKey } from '../../src/ai/client.js';
+import { parseSrt, stringifySrt, type SrtCue } from './srt.js';
+
+const DEFAULT_SUBTITLE_FIX_PROMPT =
+  'Fix transcription mistakes only. Preserve cue numbering, timestamps, and valid SRT formatting exactly. Return only corrected SRT.';
+
+const SRT_BLOCK_PATTERN =
+  /(?:^|\n)(\d+\n\d{2}:\d{2}:\d{2},\d{3} --> \d{2}:\d{2}:\d{2},\d{3}[\s\S]*)$/;
+const CODE_FENCE_PATTERN = /^```(?:\w+)?\s*\n([\s\S]*?)\n```$/;
+const JAPANESE_CHAR_PATTERN = /[\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Han}]/gu;
+const LATIN_LETTER_PATTERN = /\p{Script=Latin}/gu;
+
+export function applyFixedCueBatch(original: SrtCue[], fixed: SrtCue[]): SrtCue[] {
+  if (original.length !== fixed.length) {
+    throw new Error('Fixed subtitle batch must preserve cue count.');
+  }
+
+  return original.map((cue, index) => {
+    const nextCue = fixed[index];
+    if (!nextCue) {
+      throw new Error('Missing fixed subtitle cue.');
+    }
+    if (cue.start !== nextCue.start || cue.end !== nextCue.end) {
+      throw new Error('Fixed subtitle batch must preserve cue timestamps.');
+    }
+    return {
+      ...cue,
+      text: nextCue.text,
+    };
+  });
+}
+
+function chunkCues(cues: SrtCue[], size: number): SrtCue[][] {
+  const chunks: SrtCue[][] = [];
+  for (let index = 0; index < cues.length; index += size) {
+    chunks.push(cues.slice(index, index + size));
+  }
+  return chunks;
+}
+
+function normalizeAiSubtitleFixCandidates(content: string): string[] {
+  const trimmed = content.replace(/\r\n/g, '\n').trim();
+  if (!trimmed) {
+    return [];
+  }
+
+  const candidates = new Set<string>([trimmed]);
+  const fenced = CODE_FENCE_PATTERN.exec(trimmed)?.[1]?.trim();
+  if (fenced) {
+    candidates.add(fenced);
+  }
+
+  const srtBlock = SRT_BLOCK_PATTERN.exec(trimmed)?.[1]?.trim();
+  if (srtBlock) {
+    candidates.add(srtBlock);
+  }
+
+  return [...candidates];
+}
+
+function parseTextOnlyCueBatch(original: SrtCue[], content: string): SrtCue[] {
+  const paragraphBlocks = content
+    .split(/\n{2,}/)
+    .map((block) => block.trim())
+    .filter((block) => block.length > 0);
+  if (paragraphBlocks.length === original.length) {
+    return original.map((cue, index) => ({
+      ...cue,
+      text: paragraphBlocks[index]!,
+    }));
+  }
+
+  const lineBlocks = content
+    .split('\n')
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0);
+  if (lineBlocks.length === original.length) {
+    return original.map((cue, index) => ({
+      ...cue,
+      text: lineBlocks[index]!,
+    }));
+  }
+
+  throw new Error('Fixed subtitle batch must preserve cue count.');
+}
+
+function countPatternMatches(content: string, pattern: RegExp): number {
+  pattern.lastIndex = 0;
+  return [...content.matchAll(pattern)].length;
+}
+
+function isJapaneseLanguageCode(language: string | undefined): boolean {
+  if (!language) return false;
+  const normalized = language.trim().toLowerCase();
+  return normalized === 'ja' || normalized === 'jp' || normalized === 'jpn';
+}
+
+function validateExpectedLanguage(
+  original: SrtCue[],
+  fixed: SrtCue[],
+  expectedLanguage: string | undefined,
+): void {
+  if (!isJapaneseLanguageCode(expectedLanguage)) return;
+
+  const originalText = original.map((cue) => cue.text).join('\n');
+  const fixedText = fixed.map((cue) => cue.text).join('\n');
+  const originalJapaneseChars = countPatternMatches(originalText, JAPANESE_CHAR_PATTERN);
+  if (originalJapaneseChars < 4) return;
+
+  const fixedJapaneseChars = countPatternMatches(fixedText, JAPANESE_CHAR_PATTERN);
+  const fixedLatinLetters = countPatternMatches(fixedText, LATIN_LETTER_PATTERN);
+  if (fixedJapaneseChars === 0 && fixedLatinLetters >= 4) {
+    throw new Error('Fixed subtitle batch changed language away from expected Japanese.');
+  }
+}
+
+export function parseAiSubtitleFixResponse(
+  original: SrtCue[],
+  content: string,
+  expectedLanguage?: string,
+): SrtCue[] {
+  const candidates = normalizeAiSubtitleFixCandidates(content);
+  let lastError: Error | null = null;
+
+  for (const candidate of candidates) {
+    try {
+      const parsed = parseSrt(candidate);
+      validateExpectedLanguage(original, parsed, expectedLanguage);
+      return parsed;
+    } catch (error) {
+      lastError = error as Error;
+    }
+  }
+
+  for (const candidate of candidates) {
+    try {
+      const parsed = parseTextOnlyCueBatch(original, candidate);
+      validateExpectedLanguage(original, parsed, expectedLanguage);
+      return parsed;
+    } catch (error) {
+      lastError = error as Error;
+    }
+  }
+
+  throw lastError ?? new Error('AI subtitle fix returned empty content.');
+}
+
+export async function fixSubtitleWithAi(
+  subtitleContent: string,
+  aiConfig: LauncherAiConfig,
+  logWarning: (message: string) => void,
+  expectedLanguage?: string,
+): Promise<string | null> {
+  if (aiConfig.enabled !== true) {
+    return null;
+  }
+
+  const apiKey = await resolveAiApiKey(aiConfig);
+  if (!apiKey) {
+    return null;
+  }
+
+  const cues = parseSrt(subtitleContent);
+  if (cues.length === 0) {
+    return null;
+  }
+
+  const fixedChunks: SrtCue[] = [];
+  for (const chunk of chunkCues(cues, 25)) {
+    const fixedContent = await requestAiChatCompletion(
+      {
+        apiKey,
+        baseUrl: aiConfig.baseUrl,
+        model: aiConfig.model,
+        timeoutMs: aiConfig.requestTimeoutMs,
+        messages: [
+          {
+            role: 'system',
+            content: aiConfig.systemPrompt?.trim() || DEFAULT_SUBTITLE_FIX_PROMPT,
+          },
+          {
+            role: 'user',
+            content: stringifySrt(chunk),
+          },
+        ],
+      },
+      {
+        logWarning,
+      },
+    );
+    if (!fixedContent) {
+      return null;
+    }
+
+    let parsedFixed: SrtCue[];
+    try {
+      parsedFixed = parseAiSubtitleFixResponse(chunk, fixedContent, expectedLanguage);
+    } catch (error) {
+      logWarning(`AI subtitle fix returned invalid SRT: ${(error as Error).message}`);
+      return null;
+    }
+
+    try {
+      fixedChunks.push(...applyFixedCueBatch(chunk, parsedFixed));
+    } catch (error) {
+      logWarning(`AI subtitle fix validation failed: ${(error as Error).message}`);
+      return null;
+    }
+  }
+
+  return stringifySrt(fixedChunks);
+}
--- a/plugin/subminer/lifecycle.lua
+++ b/plugin/subminer/lifecycle.lua
@@ -62,9 +62,7 @@ function M.create(ctx)
 		hover.clear_hover_overlay()
 		process.disarm_auto_play_ready_gate()
 		if state.overlay_running or state.texthooker_running then
-			subminer_log("info", "lifecycle", "mpv shutting down, stopping SubMiner process")
-			show_osd("Shutting down...")
-			process.stop_overlay()
+			subminer_log("info", "lifecycle", "mpv shutting down, preserving SubMiner background process")
 		end
 	end

--- a/scripts/get-mpv-window-macos.swift
+++ b/scripts/get-mpv-window-macos.swift
@@ -20,6 +20,11 @@ private struct WindowGeometry {
    let height: Int
 }

+private struct WindowState {
+    let geometry: WindowGeometry
+    let focused: Bool
+}
+
 private let targetMpvSocketPath: String? = {
    guard CommandLine.arguments.count > 1 else {
        return nil
@@ -136,7 +141,11 @@ private func geometryFromAXWindow(_ axWindow: AXUIElement) -> WindowGeometry? {
    return geometry
 }

-private func geometryFromAccessibilityAPI() -> WindowGeometry? {
+private func frontmostApplicationPid() -> pid_t? {
+    NSWorkspace.shared.frontmostApplication?.processIdentifier
+}
+
+private func windowStateFromAccessibilityAPI() -> WindowState? {
    let runningApps = NSWorkspace.shared.runningApplications.filter { app in
        guard let name = app.localizedName else {
            return false
@@ -144,6 +153,8 @@ private func geometryFromAccessibilityAPI() -> WindowGeometry? {
        return normalizedMpvName(name)
    }

+    let frontmostPid = frontmostApplicationPid()
+
    for app in runningApps {
        let appElement = AXUIElementCreateApplication(app.processIdentifier)
        if !windowHasTargetSocket(app.processIdentifier) {
@@ -173,7 +184,10 @@ private func geometryFromAccessibilityAPI() -> WindowGeometry? {
            }

            if let geometry = geometryFromAXWindow(window) {
-                return geometry
+                return WindowState(
+                    geometry: geometry,
+                    focused: frontmostPid == windowPid
+                )
            }
        }
    }
@@ -181,11 +195,12 @@ private func geometryFromAccessibilityAPI() -> WindowGeometry? {
    return nil
 }

-private func geometryFromCoreGraphics() -> WindowGeometry? {
+private func windowStateFromCoreGraphics() -> WindowState? {
    // Keep the CG fallback for environments without Accessibility permissions.
    // Use on-screen layer-0 windows to avoid off-screen helpers/shadows.
    let options: CGWindowListOption = [.optionOnScreenOnly, .excludeDesktopElements]
    let windowList = CGWindowListCopyWindowInfo(options, kCGNullWindowID) as? [[String: Any]] ?? []
+    let frontmostPid = frontmostApplicationPid()

    for window in windowList {
        guard let ownerName = window[kCGWindowOwnerName as String] as? String,
@@ -226,14 +241,19 @@ private func geometryFromCoreGraphics() -> WindowGeometry? {
            continue
        }

-        return geometry
+        return WindowState(
+            geometry: geometry,
+            focused: frontmostPid == ownerPid
+        )
    }

    return nil
 }

-if let window = geometryFromAccessibilityAPI() ?? geometryFromCoreGraphics() {
-    print("\(window.x),\(window.y),\(window.width),\(window.height)")
+if let window = windowStateFromAccessibilityAPI() ?? windowStateFromCoreGraphics() {
+    print(
+        "\(window.geometry.x),\(window.geometry.y),\(window.geometry.width),\(window.geometry.height),\(window.focused ? 1 : 0)"
+    )
 } else {
    print("not-found")
 }
--- a/src/core/services/cli-command.test.ts
+++ b/src/core/services/cli-command.test.ts
@@ -201,7 +201,7 @@ function createDeps(overrides: Partial<CliCommandServiceDeps> = {}) {
  return { deps, calls, osd };
 }

-test('handleCliCommand ignores --start for second-instance when overlay runtime is already initialized', () => {
+test('handleCliCommand reconnects MPV for second-instance --start when overlay runtime is already initialized', () => {
  const { deps, calls } = createDeps({
    isOverlayRuntimeInitialized: () => true,
  });
@@ -209,11 +209,9 @@ test('handleCliCommand ignores --start for second-instance when overlay runtime

  handleCliCommand(args, 'second-instance', deps);

-  assert.ok(calls.includes('log:Ignoring --start because SubMiner is already running.'));
-  assert.equal(
-    calls.some((value) => value.includes('connectMpvClient')),
-    false,
-  );
+  assert.ok(calls.includes('setMpvClientSocketPath:/tmp/subminer.sock'));
+  assert.equal(calls.some((value) => value.includes('connectMpvClient')), true);
+  assert.equal(calls.some((value) => value.includes('initializeOverlayRuntime')), false);
 });

 test('handleCliCommand processes --start for second-instance when overlay runtime is not initialized', () => {
--- a/src/core/services/cli-command.ts
+++ b/src/core/services/cli-command.ts
@@ -259,10 +259,9 @@ export function handleCliCommand(
    deps.setLogLevel?.(args.logLevel);
  }

-  const ignoreSecondInstanceStart =
+  const reuseSecondInstanceStart =
    source === 'second-instance' && args.start && deps.isOverlayRuntimeInitialized();
-  const shouldStart =
-    (!ignoreSecondInstanceStart && args.start) || args.toggle || args.toggleVisibleOverlay;
+  const shouldStart = args.start || args.toggle || args.toggleVisibleOverlay;
  const needsOverlayRuntime = commandNeedsOverlayRuntime(args);
  const shouldInitializeOverlayRuntime = needsOverlayRuntime || args.start;

@@ -285,8 +284,8 @@ export function handleCliCommand(
    return;
  }

-  if (ignoreSecondInstanceStart) {
-    deps.log('Ignoring --start because SubMiner is already running.');
+  if (reuseSecondInstanceStart) {
+    deps.log('Reusing running SubMiner instance for --start.');
  }

  if (shouldInitializeOverlayRuntime && !deps.isOverlayRuntimeInitialized()) {
--- a/src/core/services/overlay-shortcut-handler.test.ts
+++ b/src/core/services/overlay-shortcut-handler.test.ts
@@ -6,6 +6,7 @@ import {
  OverlayShortcutRuntimeDeps,
  runOverlayShortcutLocalFallback,
 } from './overlay-shortcut-handler';
+import { shouldActivateOverlayShortcuts } from './overlay-shortcut';

 function makeShortcuts(overrides: Partial<ConfiguredShortcuts> = {}): ConfiguredShortcuts {
  return {
@@ -279,3 +280,36 @@ test('runOverlayShortcutLocalFallback returns false when no action matches', ()
  assert.equal(result, false);
  assert.equal(called, false);
 });
+
+test('shouldActivateOverlayShortcuts disables macOS overlay shortcuts when tracked mpv is unfocused', () => {
+  assert.equal(
+    shouldActivateOverlayShortcuts({
+      overlayRuntimeInitialized: true,
+      isMacOSPlatform: true,
+      trackedMpvWindowFocused: false,
+    }),
+    false,
+  );
+});
+
+test('shouldActivateOverlayShortcuts keeps macOS overlay shortcuts active when tracked mpv is focused', () => {
+  assert.equal(
+    shouldActivateOverlayShortcuts({
+      overlayRuntimeInitialized: true,
+      isMacOSPlatform: true,
+      trackedMpvWindowFocused: true,
+    }),
+    true,
+  );
+});
+
+test('shouldActivateOverlayShortcuts preserves non-macOS behavior', () => {
+  assert.equal(
+    shouldActivateOverlayShortcuts({
+      overlayRuntimeInitialized: true,
+      isMacOSPlatform: false,
+      trackedMpvWindowFocused: false,
+    }),
+    true,
+  );
+});
--- a/src/core/services/overlay-shortcut.ts
+++ b/src/core/services/overlay-shortcut.ts
@@ -27,6 +27,20 @@ export interface OverlayShortcutLifecycleDeps {
  cancelPendingMineSentenceMultiple: () => void;
 }

+export function shouldActivateOverlayShortcuts(args: {
+  overlayRuntimeInitialized: boolean;
+  isMacOSPlatform: boolean;
+  trackedMpvWindowFocused: boolean;
+}): boolean {
+  if (!args.overlayRuntimeInitialized) {
+    return false;
+  }
+  if (!args.isMacOSPlatform) {
+    return true;
+  }
+  return args.trackedMpvWindowFocused;
+}
+
 export function registerOverlayShortcuts(
  shortcuts: ConfiguredShortcuts,
  handlers: OverlayShortcutHandlers,
--- a/src/main/overlay-shortcuts-runtime.ts
+++ b/src/main/overlay-shortcuts-runtime.ts
@@ -6,9 +6,10 @@ import {
 import {
  refreshOverlayShortcutsRuntime,
  registerOverlayShortcuts,
+  shouldActivateOverlayShortcuts,
  syncOverlayShortcutsRuntime,
  unregisterOverlayShortcutsRuntime,
-} from '../core/services';
+} from '../core/services/overlay-shortcut';
 import { runOverlayShortcutLocalFallback } from '../core/services/overlay-shortcut-handler';

 export interface OverlayShortcutRuntimeServiceInput {
@@ -16,6 +17,8 @@ export interface OverlayShortcutRuntimeServiceInput {
  getShortcutsRegistered: () => boolean;
  setShortcutsRegistered: (registered: boolean) => void;
  isOverlayRuntimeInitialized: () => boolean;
+  isMacOSPlatform: () => boolean;
+  isTrackedMpvWindowFocused: () => boolean;
  showMpvOsd: (text: string) => void;
  openRuntimeOptionsPalette: () => void;
  openJimaku: () => void;
@@ -89,7 +92,12 @@ export function createOverlayShortcutsRuntimeService(
    };
  };

-  const shouldOverlayShortcutsBeActive = () => input.isOverlayRuntimeInitialized();
+  const shouldOverlayShortcutsBeActive = () =>
+    shouldActivateOverlayShortcuts({
+      overlayRuntimeInitialized: input.isOverlayRuntimeInitialized(),
+      isMacOSPlatform: input.isMacOSPlatform(),
+      trackedMpvWindowFocused: input.isTrackedMpvWindowFocused(),
+    });

  return {
    tryHandleOverlayShortcutLocalFallback: (inputEvent) =>
--- a/src/main/runtime/mpv-main-event-actions.test.ts
+++ b/src/main/runtime/mpv-main-event-actions.test.ts
@@ -58,6 +58,7 @@ test('media path change handler reports stop for empty path and probes media key
    ensureAnilistMediaGuess: (mediaKey) => calls.push(`guess:${mediaKey}`),
    syncImmersionMediaState: () => calls.push('sync'),
    scheduleCharacterDictionarySync: () => calls.push('dict-sync'),
+    signalAutoplayReadyIfWarm: (path) => calls.push(`autoplay:${path}`),
    refreshDiscordPresence: () => calls.push('presence'),
  });

@@ -74,6 +75,34 @@ test('media path change handler reports stop for empty path and probes media key
  ]);
 });

+test('media path change handler signals autoplay-ready fast path for warm non-empty media', () => {
+  const calls: string[] = [];
+  const handler = createHandleMpvMediaPathChangeHandler({
+    updateCurrentMediaPath: (path) => calls.push(`path:${path}`),
+    reportJellyfinRemoteStopped: () => calls.push('stopped'),
+    restoreMpvSubVisibility: () => calls.push('restore-mpv-sub'),
+    getCurrentAnilistMediaKey: () => null,
+    resetAnilistMediaTracking: (mediaKey) => calls.push(`reset:${String(mediaKey)}`),
+    maybeProbeAnilistDuration: (mediaKey) => calls.push(`probe:${mediaKey}`),
+    ensureAnilistMediaGuess: (mediaKey) => calls.push(`guess:${mediaKey}`),
+    syncImmersionMediaState: () => calls.push('sync'),
+    scheduleCharacterDictionarySync: () => calls.push('dict-sync'),
+    signalAutoplayReadyIfWarm: (path) => calls.push(`autoplay:${path}`),
+    refreshDiscordPresence: () => calls.push('presence'),
+  });
+
+  handler({ path: '/tmp/video.mkv' });
+
+  assert.deepEqual(calls, [
+    'path:/tmp/video.mkv',
+    'reset:null',
+    'sync',
+    'dict-sync',
+    'autoplay:/tmp/video.mkv',
+    'presence',
+  ]);
+});
+
 test('media title change handler clears guess state and syncs immersion', () => {
  const calls: string[] = [];
  const handler = createHandleMpvMediaTitleChangeHandler({
--- a/src/main/runtime/mpv-main-event-actions.ts
+++ b/src/main/runtime/mpv-main-event-actions.ts
@@ -40,6 +40,7 @@ export function createHandleMpvMediaPathChangeHandler(deps: {
  ensureAnilistMediaGuess: (mediaKey: string) => void;
  syncImmersionMediaState: () => void;
  scheduleCharacterDictionarySync?: () => void;
+  signalAutoplayReadyIfWarm?: (path: string) => void;
  refreshDiscordPresence: () => void;
 }) {
  return ({ path }: { path: string | null }): void => {
@@ -58,6 +59,7 @@ export function createHandleMpvMediaPathChangeHandler(deps: {
    deps.syncImmersionMediaState();
    if (normalizedPath.trim().length > 0) {
      deps.scheduleCharacterDictionarySync?.();
+      deps.signalAutoplayReadyIfWarm?.(normalizedPath);
    }
    deps.refreshDiscordPresence();
  };
--- a/src/main/runtime/mpv-main-event-bindings.ts
+++ b/src/main/runtime/mpv-main-event-bindings.ts
@@ -50,6 +50,7 @@ export function createBindMpvMainEventHandlersHandler(deps: {
  maybeProbeAnilistDuration: (mediaKey: string) => void;
  ensureAnilistMediaGuess: (mediaKey: string) => void;
  syncImmersionMediaState: () => void;
+  signalAutoplayReadyIfWarm?: (path: string) => void;

  updateCurrentMediaTitle: (title: string) => void;
  resetAnilistMediaGuessState: () => void;
@@ -105,6 +106,7 @@ export function createBindMpvMainEventHandlersHandler(deps: {
      maybeProbeAnilistDuration: (mediaKey) => deps.maybeProbeAnilistDuration(mediaKey),
      ensureAnilistMediaGuess: (mediaKey) => deps.ensureAnilistMediaGuess(mediaKey),
      syncImmersionMediaState: () => deps.syncImmersionMediaState(),
+      signalAutoplayReadyIfWarm: (path) => deps.signalAutoplayReadyIfWarm?.(path),
      scheduleCharacterDictionarySync: () => deps.scheduleCharacterDictionarySync?.(),
      refreshDiscordPresence: () => deps.refreshDiscordPresence(),
    });
--- a/src/main/runtime/mpv-main-event-main-deps.ts
+++ b/src/main/runtime/mpv-main-event-main-deps.ts
@@ -33,6 +33,7 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
  maybeProbeAnilistDuration: (mediaKey: string) => void;
  ensureAnilistMediaGuess: (mediaKey: string) => void;
  syncImmersionMediaState: () => void;
+  signalAutoplayReadyIfWarm?: (path: string) => void;
  scheduleCharacterDictionarySync?: () => void;
  updateCurrentMediaTitle: (title: string) => void;
  resetAnilistMediaGuessState: () => void;
@@ -82,6 +83,7 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
    maybeProbeAnilistDuration: (mediaKey: string) => deps.maybeProbeAnilistDuration(mediaKey),
    ensureAnilistMediaGuess: (mediaKey: string) => deps.ensureAnilistMediaGuess(mediaKey),
    syncImmersionMediaState: () => deps.syncImmersionMediaState(),
+    signalAutoplayReadyIfWarm: (path: string) => deps.signalAutoplayReadyIfWarm?.(path),
    scheduleCharacterDictionarySync: () => deps.scheduleCharacterDictionarySync?.(),
    updateCurrentMediaTitle: (title: string) => deps.updateCurrentMediaTitle(title),
    resetAnilistMediaGuessState: () => deps.resetAnilistMediaGuessState(),
--- a/src/main/runtime/overlay-shortcuts-runtime-main-deps.test.ts
+++ b/src/main/runtime/overlay-shortcuts-runtime-main-deps.test.ts
@@ -13,6 +13,8 @@ test('overlay shortcuts runtime main deps builder maps lifecycle and action call
      calls.push(`registered:${registered}`);
    },
    isOverlayRuntimeInitialized: () => true,
+    isMacOSPlatform: () => true,
+    isTrackedMpvWindowFocused: () => false,
    showMpvOsd: (text) => calls.push(`osd:${text}`),
    openRuntimeOptionsPalette: () => calls.push('runtime-options'),
    openJimaku: () => calls.push('jimaku'),
@@ -40,6 +42,8 @@ test('overlay shortcuts runtime main deps builder maps lifecycle and action call
  })();

  assert.equal(deps.isOverlayRuntimeInitialized(), true);
+  assert.equal(deps.isMacOSPlatform(), true);
+  assert.equal(deps.isTrackedMpvWindowFocused(), false);
  assert.equal(deps.getShortcutsRegistered(), false);
  deps.setShortcutsRegistered(true);
  assert.equal(shortcutsRegistered, true);
--- a/src/main/runtime/overlay-shortcuts-runtime-main-deps.ts
+++ b/src/main/runtime/overlay-shortcuts-runtime-main-deps.ts
@@ -8,6 +8,8 @@ export function createBuildOverlayShortcutsRuntimeMainDepsHandler(
    getShortcutsRegistered: () => deps.getShortcutsRegistered(),
    setShortcutsRegistered: (registered: boolean) => deps.setShortcutsRegistered(registered),
    isOverlayRuntimeInitialized: () => deps.isOverlayRuntimeInitialized(),
+    isMacOSPlatform: () => deps.isMacOSPlatform(),
+    isTrackedMpvWindowFocused: () => deps.isTrackedMpvWindowFocused(),
    showMpvOsd: (text: string) => deps.showMpvOsd(text),
    openRuntimeOptionsPalette: () => deps.openRuntimeOptionsPalette(),
    openJimaku: () => deps.openJimaku(),
--- a/src/window-trackers/base-tracker.ts
+++ b/src/window-trackers/base-tracker.ts
@@ -21,13 +21,17 @@ import { WindowGeometry } from '../types';
 export type GeometryChangeCallback = (geometry: WindowGeometry) => void;
 export type WindowFoundCallback = (geometry: WindowGeometry) => void;
 export type WindowLostCallback = () => void;
+export type WindowFocusChangeCallback = (focused: boolean) => void;

 export abstract class BaseWindowTracker {
  protected currentGeometry: WindowGeometry | null = null;
  protected windowFound: boolean = false;
+  protected focusKnown: boolean = false;
+  protected windowFocused: boolean = false;
  public onGeometryChange: GeometryChangeCallback | null = null;
  public onWindowFound: WindowFoundCallback | null = null;
  public onWindowLost: WindowLostCallback | null = null;
+  public onWindowFocusChange: WindowFocusChangeCallback | null = null;

  abstract start(): void;
  abstract stop(): void;
@@ -40,6 +44,19 @@ export abstract class BaseWindowTracker {
    return this.windowFound;
  }

+  isFocused(): boolean {
+    return this.focusKnown ? this.windowFocused : this.windowFound;
+  }
+
+  protected updateFocus(focused: boolean): void {
+    const changed = !this.focusKnown || this.windowFocused !== focused;
+    this.focusKnown = true;
+    this.windowFocused = focused;
+    if (changed) {
+      this.onWindowFocusChange?.(focused);
+    }
+  }
+
  protected updateGeometry(newGeometry: WindowGeometry | null): void {
    if (newGeometry) {
      if (!this.windowFound) {
@@ -58,6 +75,12 @@ export abstract class BaseWindowTracker {
        if (this.onGeometryChange) this.onGeometryChange(newGeometry);
      }
    } else {
+      const focusChanged = this.focusKnown && this.windowFocused;
+      this.focusKnown = false;
+      this.windowFocused = false;
+      if (focusChanged) {
+        this.onWindowFocusChange?.(false);
+      }
      if (this.windowFound) {
        this.windowFound = false;
        this.currentGeometry = null;
--- a/src/window-trackers/macos-tracker.ts
+++ b/src/window-trackers/macos-tracker.ts
@@ -22,9 +22,56 @@ import * as fs from 'fs';
 import * as os from 'os';
 import { BaseWindowTracker } from './base-tracker';
 import { createLogger } from '../logger';
+import type { WindowGeometry } from '../types';

 const log = createLogger('tracker').child('macos');

+export interface MacOSHelperWindowState {
+  geometry: WindowGeometry;
+  focused: boolean;
+}
+
+export function parseMacOSHelperOutput(result: string): MacOSHelperWindowState | null {
+  const trimmed = result.trim();
+  if (!trimmed || trimmed === 'not-found') {
+    return null;
+  }
+
+  const parts = trimmed.split(',');
+  if (parts.length !== 4 && parts.length !== 5) {
+    return null;
+  }
+
+  const x = parseInt(parts[0]!, 10);
+  const y = parseInt(parts[1]!, 10);
+  const width = parseInt(parts[2]!, 10);
+  const height = parseInt(parts[3]!, 10);
+  if (
+    !Number.isFinite(x) ||
+    !Number.isFinite(y) ||
+    !Number.isFinite(width) ||
+    !Number.isFinite(height) ||
+    width <= 0 ||
+    height <= 0
+  ) {
+    return null;
+  }
+
+  const focusedRaw = parts[4]?.trim().toLowerCase();
+  const focused =
+    focusedRaw === undefined ? true : focusedRaw === '1' || focusedRaw === 'true';
+
+  return {
+    geometry: {
+      x,
+      y,
+      width,
+      height,
+    },
+    focused,
+  };
+}
+
 export class MacOSWindowTracker extends BaseWindowTracker {
  private pollInterval: ReturnType<typeof setInterval> | null = null;
  private pollInFlight = false;
@@ -173,33 +220,12 @@ export class MacOSWindowTracker extends BaseWindowTracker {
          return;
        }

-        const result = (stdout || '').trim();
-        if (result && result !== 'not-found') {
-          const parts = result.split(',');
-          if (parts.length === 4) {
-            const x = parseInt(parts[0]!, 10);
-            const y = parseInt(parts[1]!, 10);
-            const width = parseInt(parts[2]!, 10);
-            const height = parseInt(parts[3]!, 10);
-
-            if (
-              Number.isFinite(x) &&
-              Number.isFinite(y) &&
-              Number.isFinite(width) &&
-              Number.isFinite(height) &&
-              width > 0 &&
-              height > 0
-            ) {
-              this.updateGeometry({
-                x,
-                y,
-                width,
-                height,
-              });
-              this.pollInFlight = false;
-              return;
-            }
-          }
+        const parsed = parseMacOSHelperOutput(stdout || '');
+        if (parsed) {
+          this.updateFocus(parsed.focused);
+          this.updateGeometry(parsed.geometry);
+          this.pollInFlight = false;
+          return;
        }

        this.updateGeometry(null);
--- a/src/window-trackers/x11-tracker.test.ts
+++ b/src/window-trackers/x11-tracker.test.ts
@@ -1,6 +1,7 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import { parseX11WindowGeometry, parseX11WindowPid, X11WindowTracker } from './x11-tracker';
+import { parseMacOSHelperOutput } from './macos-tracker';

 test('parseX11WindowGeometry parses xwininfo output', () => {
  const geometry = parseX11WindowGeometry(`
@@ -52,3 +53,27 @@ Height: 360`;
  release();
  await new Promise((resolve) => setTimeout(resolve, 0));
 });
+
+test('parseMacOSHelperOutput parses geometry and focused state', () => {
+  assert.deepEqual(parseMacOSHelperOutput('120,240,1280,720,1'), {
+    geometry: {
+      x: 120,
+      y: 240,
+      width: 1280,
+      height: 720,
+    },
+    focused: true,
+  });
+});
+
+test('parseMacOSHelperOutput tolerates unfocused helper output', () => {
+  assert.deepEqual(parseMacOSHelperOutput('120,240,1280,720,0'), {
+    geometry: {
+      x: 120,
+      y: 240,
+      width: 1280,
+      height: 720,
+    },
+    focused: false,
+  });
+});