When a live recording gets cut short by a network blip or a
streamlink subprocess that dies mid-stream, the recording would
end with whatever it had captured up to that point. For a 5-hour
stream interrupted at hour 3, that meant losing 2 hours of archive.
downloadLiveStream now wraps the streamlink call in a resume loop.
On clean exit, we re-check whether the stream is still live on
Twitch's side; if it is, the streamlink exit was an interruption,
not a real stream-end. The recording continues into a new file
("..._part2.mp4", "..._part3.mp4", ...) and both parts get attached
to item.outputFiles so the user sees them as one logical recording.
Guard rails to keep the loop from misbehaving:
- Stream-still-live check before each resume. If the streamer
actually ended their broadcast, we finalize. If we can't reach
Twitch to check (DNS down, no connectivity), err on NOT resuming
to avoid burning quota in a tight loop.
- Skip resume on suspiciously short parts (<30s). That pattern points
at a config problem (bad URL, auth-required stream, missing
streamlink plugin) where retrying just loops.
- Cap at 5 resume attempts per recording. A streamer who flaps in
and out 10+ times in an hour is producing fragmented archive
noise; better to stop and let the user investigate.
- Skip resume on zero-byte parts. Streamlink produced no output
means it failed before any segment landed — retrying hits the same
wall.
- Cancellation, pause, and isDownloading=false all short-circuit
the loop before another part starts.
Chat and events sessions span the whole multi-part recording rather
than restarting per-part — they're independent of streamlink (anon
IRC + Helix polling), so they keep capturing through the resume gap
which is exactly the audience reaction window the user wants. A new
"recording_resume" event type lands in .events.jsonl so the events
viewer shows where each gap happened.
The progress meta line was rewritten to accumulate bytes across
parts. Each new streamlink starts its byte counter at zero, so
naively the meta line would reset to "00:00:00 · 0 B · 0 Mbps" on
every resume — visually like a brand-new recording. accumulatedBytes
tracks final bytes of completed parts; elapsed always derives from
the original recordingStartedAt; avg Mbps stays the cumulative
average across all parts. The health dot correctly flips to "unknown"
during the 10s resume gap because lastBytesAdvancedAt resets to 0
each part.
Settings toggle (default on). When off, behavior is identical to
4.6.12 — single part, no resume.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
7d82f70ca3