Handle untagged continuation lines in rolling auto-captions

Short continuations (e.g. a single trailing word) carry no inline timing
tags, so filtering by tags missed them. Instead, in files with inline
timing tags, drop leading cue lines that repeat the previous cue's last
line — covers both tagged and untagged new text.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Handoff.md

@@ -0,0 +1,41 @@
+# yt-dlf — Handoff
+
+> Übergabedokument für die Fortsetzung in einer Claude-Code-Session.
+> Stand: 2026-08-29 — Web-Frontend für yt-dlp/ffmpeg mit Login, Ad-Spalten
+> und User-Verwaltung; implementiert, Deploy-Status unbekannt.
+> Dauerhafte Fakten stehen in `README.md`/`INSTALL.md`; hier steht nur, was
+> **nicht** aus Code und Git-Historie ersichtlich ist.
+
+**🔜 NÄCHSTE SESSION:** kein vorgeplanter Auftrag
+
+## Offene Punkte
+
+- [ ] Deploy-Status klären — `build/` und `users.txt` liegen lokal vor
+  (beide gitignored), aber ob und wo die App produktiv läuft, ist nirgends
+  festgehalten.
+
+## Stolperfallen
+
+- **Git-Konten-Mix** — Bestandsprojekt: Remote `origin` zeigt direkt auf
+  `git@gitea.101010.cloud:stwaidele/yt-dlf.git` und die Repo-Config steht
+  auf Stefans Konto. Claude-Commits daher per Author-Override
+  (`git -c user.name="Claude <Modell>" -c user.email="<slug>@waidele.info" commit …`),
+  Repo-Config nicht umbiegen. Push läuft über Stefans Default-Key.
+
+## Session 2026-08-29 — Handoff angelegt (Claude Fable 5)
+
+Projekt hatte noch keine `Handoff.md`; per `/handoff init` aus README,
+INSTALL.md und Git-Historie erstellt. Kein Projekt-`CLAUDE.md` vorhanden.
+Beide bisherigen Commits (16.05.) stammen von Stefan; seither keine
+Änderungen im Working Tree.
+
+## Typische Handgriffe
+
+Dev-Server, Deployment und User-Verwaltung sind vollständig in `README.md`
+und `INSTALL.md` dokumentiert — dort nachschlagen, hier nicht duplizieren.
+
+## Verwandte Handoffs
+
+_keine_
+
+<!-- Ältere Sessions: git log -p Handoff.md -->

scripts/subtitle_to_markdown.py

@@ -9,9 +9,18 @@ def strip_html(text):
     return re.sub(r'<[^>]+>', '', text)
 
 
+# YouTube auto-captions use rolling cues: each cue repeats the previous
+# cue's last line as plain text before the new text (which usually, but not
+# always, carries inline word timings like <00:00:00.200> or <c>…</c>).
+def has_inline_timing_tags(text):
+    return re.search(r'<\d{2}:\d{2}:|<c[.\w]*>', text) is not None
+
+
 def parse_vtt(content):
+    rolling = has_inline_timing_tags(content)
     blocks = re.split(r'\n{2,}', content.strip())
     cues = []
+    prev_last = None
     for block in blocks:
         lines = block.strip().splitlines()
         if not lines:
@@ -23,11 +32,14 @@ def parse_vtt(content):
         ts_idx = next((i for i, l in enumerate(lines) if '-->' in l), None)
         if ts_idx is None:
             continue
-        text = ' '.join(
+        text_lines = [strip_html(l).strip() for l in lines[ts_idx + 1:]]
-            strip_html(l).strip()
+        if rolling:
-            for l in lines[ts_idx + 1:]
+            while text_lines and prev_last and text_lines[0] == prev_last:
-            if strip_html(l).strip()
+                text_lines.pop(0)
-        )
+        non_empty = [l for l in text_lines if l]
+        if rolling and non_empty:
+            prev_last = non_empty[-1]
+        text = ' '.join(non_empty)
         if text:
             cues.append(text)
     return cues

src/lib/subtitle.js

@@ -5,9 +5,18 @@ function stripHtml(text) {
 	return text.replace(/<[^>]+>/g, '');
 }
 
+// YouTube auto-captions use rolling cues: each cue repeats the previous
+// cue's last line as plain text before the new text (which usually, but not
+// always, carries inline word timings like <00:00:00.200> or <c>…</c>).
+function hasInlineTimingTags(text) {
+	return /<\d{2}:\d{2}:|<c[.\w]*>/.test(text);
+}
+
 function parseVtt(content) {
+	const rolling = hasInlineTimingTags(content);
 	const blocks = content.split(/\n{2,}/);
 	const cues = [];
+	let prevLast = null;
 	for (const block of blocks) {
 		const lines = block.trim().split('\n');
 		if (!lines.length) continue;
@@ -15,11 +24,15 @@ function parseVtt(content) {
 		if (lines[0].startsWith('NOTE') || lines[0].startsWith('STYLE')) continue;
 		const tsIdx = lines.findIndex(l => l.includes('-->'));
 		if (tsIdx === -1) continue;
-		const text = lines
+		const textLines = lines.slice(tsIdx + 1).map(l => stripHtml(l).trim());
-			.slice(tsIdx + 1)
+		if (rolling) {
-			.map(l => stripHtml(l).trim())
+			while (textLines.length && prevLast && textLines[0] === prevLast) {
-			.filter(Boolean)
+				textLines.shift();
-			.join(' ');
+			}
+		}
+		const nonEmpty = textLines.filter(Boolean);
+		if (rolling && nonEmpty.length) prevLast = nonEmpty[nonEmpty.length - 1];
+		const text = nonEmpty.join(' ');
 		if (text) cues.push(text);
 	}
 	return cues;

src/routes/api/download/+server.js

@@ -105,12 +105,18 @@ async function runDownload(videoUrl, audioOnly, send, signal) {
 		}
 	}
 
+	const writtenMd = new Set();
 	for (const subPath of subtitlePaths) {
 		if (!existsSync(subPath)) continue;
 		send('info', `Converting: ${basename(subPath)}`);
 		try {
 			const md = convertSubtitleToMarkdown(subPath);
 			const mdPath = subPath.replace(/\.(vtt|srt)$/, '.md');
+			if (writtenMd.has(md)) {
+				send('info', `Skipped (identical): ${basename(mdPath)}`);
+				continue;
+			}
+			writtenMd.add(md);
 			writeFileSync(mdPath, md, 'utf-8');
 			send('info', `Saved: ${basename(mdPath)}`);
 		} catch (err) {