dsh-vidi 0.4.2 → 0.5.0

package/lib/index.mjs

@@ -1693,6 +1693,8 @@ const AHEAD_MAX_LINES = 10;
 *
 * Events: 'state' (JobSummary), 'cue' (Cue — new or revised).
 */
+/** Drop the first `skip` seconds after decoding (an output -ss; empty when there is nothing to drop). */
+const seekAfter = (skip) => skip > .01 ? ["-ss", skip.toFixed(3)] : [];
 var LiveJob = class LiveJob extends EventEmitter {
 	opts;
 	deps;
@@ -2354,12 +2356,14 @@ var LiveJob = class LiveJob extends EventEmitter {
 		const p = this.provider;
 		this.provider = void 0;
 		await p?.stop().catch(() => {});
-		await this.translating;
-		await this.polisher.flush();
 		this.subs = {
 			state,
 			at: this.subs.at
 		};
+		this.persist();
+		this.emit("state", this.summary());
+		await this.translating;
+		await this.polisher.flush();
 		this.save();
 		this.persist();
 		this.emit("state", this.summary());
@@ -2393,7 +2397,7 @@ var LiveJob = class LiveJob extends EventEmitter {
 		this.emit("cue", cue);
 		if (cue.final) {
 			this.deps.glossary.notice(cue.src);
-			if (this.opts.mode === "realtime" && !this.provider?.translates && cue.tr === void 0 && this.targets().length) this.translate(cue);
+			if (this.opts.mode !== "polished" && !this.provider?.translates && cue.tr === void 0 && this.targets().length) this.translate(cue);
 			else {
 				this.record(cue);
 				this.polisher.add(cue);
@@ -2594,7 +2598,14 @@ var LiveJob = class LiveJob extends EventEmitter {
 	exportText(langs, format) {
 		return exportCues(this.cues, langs, format, this.title);
 	}
-	/** Best available picture source for time range [from, to]: original download, finished recording, or recorded segments. */
+	/**
+	* Best available picture source for time range [from, to]: original
+	* download, finished recording, or recorded segments. `args` are the input
+	* arguments; `skip` is how much to drop after decoding (output -ss). Files
+	* seek exactly on input; live segments do not (a seek lands between
+	* keyframes — YouTube live has one every 5 s — and decodes nothing), so they
+	* are read from the start of the first segment and trimmed after decoding.
+	*/
 	mediaFor(from, to) {
 		const original = readdirSync(this.deps.dir).find((n) => /^original\.(mp4|mkv|webm)$/.test(n));
 		if (original) return {
@@ -2604,7 +2615,7 @@ var LiveJob = class LiveJob extends EventEmitter {
 				"-i",
 				join(this.deps.dir, original)
 			],
-			offset: 0
+			skip: 0
 		};
 		if (this.captureDone && existsSync(join(this.deps.dir, "recording.mp4"))) return {
 			args: [
@@ -2613,7 +2624,7 @@ var LiveJob = class LiveJob extends EventEmitter {
 				"-i",
 				join(this.deps.dir, "recording.mp4")
 			],
-			offset: 0
+			skip: 0
 		};
 		const segs = this.segments();
 		const files = readFileSync(this.m3u8, "utf8").split("\n").filter((l) => l && !l.startsWith("#"));
@@ -2621,29 +2632,26 @@ var LiveJob = class LiveJob extends EventEmitter {
 		let last = segs.findIndex((sg) => sg.start + sg.dur >= to);
 		if (last < 0) last = segs.length - 1;
 		if (!segs.length || first > last) throw new Error("that part has not been recorded yet");
-		const list = files.slice(first, last + 1).map((f) => join(this.deps.dir, this.hlsName, f)).join("|");
 		return {
-			args: [
-				"-ss",
-				String(from - segs[first].start),
-				"-i",
-				`concat:${list}`
-			],
-			offset: 0
+			args: ["-i", `concat:${files.slice(first, last + 1).map((f) => join(this.deps.dir, this.hlsName, f)).join("|")}`],
+			skip: Math.max(0, from - segs[first].start)
 		};
 	}
 	/** One small JPEG of the picture at media time `at`, for the model to look at (undefined for audio-only or on failure). */
 	async still(at) {
 		try {
-			const { args } = this.mediaFor(at, at + 1);
+			const { args, skip } = this.mediaFor(at, at + 1);
 			const { stdout } = await run$1("ffmpeg", [
 				"-v",
 				"error",
 				...args,
+				...seekAfter(skip),
 				"-frames:v",
 				"1",
 				"-vf",
 				"scale=768:-2",
+				"-pix_fmt",
+				"yuvj420p",
 				"-q:v",
 				"5",
 				"-f",
@@ -2666,18 +2674,22 @@ var LiveJob = class LiveJob extends EventEmitter {
 		const dir = join(this.deps.dir, "frames");
 		mkdirSync(dir, { recursive: true });
 		const out = join(dir, `frame_${at.toFixed(1).replace(".", "_")}.jpg`);
-		const { args } = this.mediaFor(at, at + 1);
+		const { args, skip } = this.mediaFor(at, at + 1);
 		await run$1("ffmpeg", [
 			"-v",
 			"error",
 			"-y",
 			...args,
+			...seekAfter(skip),
 			"-frames:v",
 			"1",
+			"-pix_fmt",
+			"yuvj420p",
 			"-q:v",
 			"2",
 			out
 		], { timeout: 6e4 });
+		if (!existsSync(out)) throw new Error("no picture at that moment");
 		return out;
 	}
 	/**
@@ -2690,14 +2702,14 @@ var LiveJob = class LiveJob extends EventEmitter {
 		mkdirSync(dir, { recursive: true });
 		const name = `clip_${from.toFixed(1)}-${to.toFixed(1)}`.replace(/\./g, "_");
 		const out = join(dir, `${name}.mp4`);
-		const { args } = this.mediaFor(from, to);
+		const { args, skip } = this.mediaFor(from, to);
 		const subArgs = [];
 		const maps = [];
 		if (langs?.length) {
 			const inRange = this.cues.filter((c) => c.final && c.end > from && c.start < to).map((c) => ({
 				...c,
-				start: Math.max(0, c.start - from),
-				end: Math.min(to, c.end) - from
+				start: Math.max(0, c.start - from) + skip,
+				end: Math.min(to, c.end) - from + skip
 			}));
 			(langs.length > 1 ? [langs, ...langs.map((l) => [l])] : [langs]).forEach((ls, i) => {
 				const f = join(dir, `.${name}.${i}.srt`);
@@ -2712,6 +2724,7 @@ var LiveJob = class LiveJob extends EventEmitter {
 			"-y",
 			...args,
 			...subArgs,
+			...seekAfter(skip),
 			"-t",
 			String(to - from),
 			"-map",
@@ -2764,8 +2777,16 @@ var LiveJob = class LiveJob extends EventEmitter {
 				rows
 			});
 		}
+		const { args, skip } = this.mediaFor(from, to);
 		const pictures = /* @__PURE__ */ new Map();
 		const lines = ["ffconcat version 1.0"];
+		if (skip > 0) {
+			const blank = await drawSubtitles([], style, w, h);
+			if (!blank) throw new Error("subtitles could not be drawn on this computer");
+			const file = join(tmp, "blank.png");
+			writeFileSync(file, blank);
+			lines.push(`file '${file.replace(/'/g, "'\\''")}'`, `duration ${skip.toFixed(3)}`);
+		}
 		for (const s of spans) {
 			let file = pictures.get(s.key);
 			if (!file) {
@@ -2780,7 +2801,6 @@ var LiveJob = class LiveJob extends EventEmitter {
 		lines.push(`file '${pictures.get(spans.at(-1).key).replace(/'/g, "'\\''")}'`);
 		const list = join(tmp, "subs.ffconcat");
 		writeFileSync(list, lines.join("\n") + "\n");
-		const { args } = this.mediaFor(from, to);
 		try {
 			await run$1("ffmpeg", [
 				"-v",
@@ -2793,6 +2813,7 @@ var LiveJob = class LiveJob extends EventEmitter {
 				"0",
 				"-i",
 				list,
+				...seekAfter(skip),
 				"-t",
 				String(to - from),
 				"-filter_complex",
@@ -3400,13 +3421,21 @@ const ISO = {
 var WhisperServer = class {
 	model;
 	log;
+	audioCtx;
 	proc;
 	port;
 	starting;
 	users = 0;
-	constructor(model, log) {
+	/**
+	* @param audioCtx Whisper's encoder window in frames (1500 = 30 s). Chunks
+	* here are a few seconds long, and the encoder costs the same for 5 s as for
+	* 30 s unless the window is cut down: a window just over the longest chunk
+	* made recognition about 2.5× faster (Apple M5, large-v3-turbo).
+	*/
+	constructor(model, log, audioCtx = 0) {
 		this.model = model;
 		this.log = log;
+		this.audioCtx = audioCtx;
 	}
 	async acquire() {
 		this.users++;
@@ -3429,7 +3458,8 @@ var WhisperServer = class {
 			String(port),
 			"-t",
 			"8",
-			"-nt"
+			"-nt",
+			...this.audioCtx ? ["-ac", String(this.audioCtx)] : []
 		], { stdio: [
 			"ignore",
 			"ignore",
@@ -4596,7 +4626,7 @@ function apply(ctx, config) {
 			appendFileSync(logFile, `${(/* @__PURE__ */ new Date()).toISOString()} ${msg}\n`);
 		} catch {}
 	};
-	const whisper = new WhisperServer(config.whisperModel, log);
+	const whisper = new WhisperServer(config.whisperModel, log, Math.min(1500, Math.max(256, Math.ceil((config.maxChunkSec + 5) * 50 / 64) * 64)));
 	const jobs = /* @__PURE__ */ new Map();
 	const byConversation = /* @__PURE__ */ new Map();
 	let llm;
@@ -4753,7 +4783,7 @@ function apply(ctx, config) {
 		tgtLangs: [config.defaultTargetLang],
 		srcLang: "auto",
 		provider: defaultEngine(),
-		mode: "realtime"
+		mode: "polished"
 	};
 	/** Browser notifications: "show this video" for a session (tools, /video). */
 	const bus = new EventEmitter();
@@ -4891,13 +4921,14 @@ function apply(ctx, config) {
 	/** The conversation's video: in memory, or the latest one on disk (after a restart). */
 	function jobForConversation(conversationId) {
 		const id = byConversation.get(conversationId);
+		if (id === "") return void 0;
 		if (id && jobs.has(id)) return jobs.get(id);
 		const root = join(config.dataDir, "jobs");
 		if (!existsSync(root)) return void 0;
 		const dirs = readdirSync(root).filter((n) => /^[\w-]+$/.test(n)).sort().reverse();
 		for (const name of dirs) {
 			const file = join(root, name, "job.json");
-			if (!existsSync(file)) continue;
+			if (!existsSync(file) || existsSync(join(root, name, "closed"))) continue;
 			try {
 				if (JSON.parse(readFileSync(file, "utf8")).opts?.conversationId !== conversationId) continue;
 			} catch {
@@ -4912,6 +4943,7 @@ function apply(ctx, config) {
 			return job;
 		}
 	}
+	const modeOf = (m) => m === "polished" || m === "site" ? m : "realtime";
 	async function startJob(opts, waiting = false) {
 		const prev = waiting ? void 0 : byConversation.get(opts.conversationId);
 		if (prev) await jobs.get(prev)?.close("replaced");
@@ -4922,7 +4954,7 @@ function apply(ctx, config) {
 		job = new LiveJob(id, opts, jobDeps(dir, opts, () => job, waiting));
 		jobs.set(id, job);
 		if (!waiting) byConversation.set(opts.conversationId, id);
-		job.autoSubs = true;
+		job.autoSubs = opts.mode !== "site";
 		bus.emit("job", job);
 		job.start();
 		return job;
@@ -4986,7 +5018,7 @@ function apply(ctx, config) {
 					tgtLangs: b.tgtLangs.map(String).slice(0, 4),
 					srcLang: typeof b.srcLang === "string" ? b.srcLang : "auto",
 					provider: typeof b.provider === "string" ? b.provider : "local",
-					mode: b.mode === "polished" ? "polished" : "realtime"
+					mode: modeOf(b.mode)
 				} });
 				const source = b.source === "push" || b.source === "file" ? b.source : "url";
 				await disarm();
@@ -4994,13 +5026,25 @@ function apply(ctx, config) {
 				return send(res, 200, (await startJob({
 					source,
 					url: source === "file" && raw.startsWith("~") ? join(homedir(), raw.slice(1)) : raw,
-					mode: b.mode === "polished" ? "polished" : "realtime",
+					mode: modeOf(b.mode),
 					srcLang: typeof b.srcLang === "string" ? b.srcLang : "auto",
 					tgtLangs: Array.isArray(b.tgtLangs) ? b.tgtLangs.filter((x) => typeof x === "string").slice(0, 4) : [config.defaultTargetLang],
 					provider: typeof b.provider === "string" ? b.provider : "local",
 					conversationId: String(b.conversationId ?? "default")
 				})).summary());
 			}
+			if (route === "close" && req.method === "POST") {
+				const b = await readJson(req);
+				const job = jobs.get(String(b.jobId));
+				if (!job) return send(res, 404, { error: "no such job" });
+				await job.close("closed");
+				const conv = job.summary().conversationId;
+				if (byConversation.get(conv) === job.id) byConversation.set(conv, "");
+				try {
+					writeFileSync(join(job.summary().dir, "closed"), (/* @__PURE__ */ new Date()).toISOString());
+				} catch {}
+				return send(res, 200, { ok: true });
+			}
 			if (route === "stop" && req.method === "POST") {
 				const b = await readJson(req);
 				const job = jobs.get(String(b.jobId));
@@ -5075,7 +5119,7 @@ function apply(ctx, config) {
 					if (Array.isArray(b.tgtLangs)) opts.tgtLangs = b.tgtLangs.map(String).slice(0, 4);
 					if (typeof b.srcLang === "string") opts.srcLang = b.srcLang;
 					if (typeof b.provider === "string") opts.provider = b.provider;
-					if (b.mode === "realtime" || b.mode === "polished") opts.mode = b.mode;
+					if (b.mode === "realtime" || b.mode === "polished" || b.mode === "site") opts.mode = b.mode;
 					const svc = opts.provider ? service(opts.provider) : void 0;
 					if (svc) {
 						await keys.refresh();
@@ -5094,7 +5138,7 @@ function apply(ctx, config) {
 				return send(res, 200, await arm({
 					source: "push",
 					url: "rtmp",
-					mode: b.mode === "polished" ? "polished" : "realtime",
+					mode: modeOf(b.mode),
 					srcLang: typeof b.srcLang === "string" ? b.srcLang : "auto",
 					tgtLangs: Array.isArray(b.tgtLangs) ? b.tgtLangs.map(String).slice(0, 4) : [config.defaultTargetLang],
 					provider: typeof b.provider === "string" ? b.provider : "local",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-vidi",
-  "version": "0.4.2",
+  "version": "0.5.0",
   "description": "A video tool for DeepSeek Harness: watch live streams, videos and podcasts with real-time subtitles in any language, then clip and download",
   "type": "module",
   "license": "PolyForm-Noncommercial-1.0.0",