VODs & Clips#

Recording System#

OpenVibe.Live automatically records your streams as VODs (Video on Demand).

Protocol-Specific Recording#

Recording Lifecycle#

1. Start: Recording begins when the stream goes live
2. Live DVR: A seekable sidecar file is generated periodically (every 60s for server-side, every 2 chunks for browser uploads) so viewers can rewind
3. Finalize: When the stream ends, the recording is remuxed for proper seeking, duration is probed, and a thumbnail is generated
4. Auto-cleanup: Recordings shorter than 10 seconds are automatically deleted (test streams, accidental go-lives)

Browser Tab Close Safety#

If you close the browser tab during a WebRTC stream, sendBeacon attempts to upload any remaining chunks. The server will auto-finalize when it detects the stream has ended.

DVR / Live Seeking#

Viewers can seek backwards in a live stream using the DVR controls:

• Click/drag the progress bar to seek
• Arrow Left — Rewind 5 seconds
• Arrow Right — Forward 5 seconds
• LIVE button — Jump back to the live edge

DVR availability appears after ~30 seconds of recording.

Clips#

Viewers can create clips from live streams:

1. Click the Clip button during a stream
2. Set the clip duration (default: 30 seconds)
3. The clip is saved from the server-side recording

Clips are unlisted by default — the stream owner can publish or delete them from the dashboard.

VOD Management#

From the dashboard:

• Toggle VODs between public and private
• Bulk delete old media by age (e.g., delete VODs older than 30 days)
• Thumbnails are auto-generated; broken thumbnails auto-regenerate on load

Chat Replay#

VODs include synchronized chat replay. Messages are stored in the database with timestamps relative to the stream start. Deleted messages are automatically excluded from replay (soft-delete with is_deleted flag).

AI moments: pastes and clips never overlap#

Two jobs turn stream moments into content: server/ai/ai-moments-job.js (image pastes for the home hero + pastes tab, every 6 h) and server/ai/auto-clip-job.js (live chat-spike clips + a VOD backfill). They used to pick the same second of the same VOD with the same title. Now:

• One shared registry — server/ai/moment-registry.js (ai_used_moments state) records every paste/clip (vod_id, stream_id, offset, scene signature, title). Both jobs ask it before creating anything: a moment is refused if it is within 2 min of a used moment on the same VOD/stream, or if its scene signature (first five words of the description) was used in the last two weeks. Legacy logs are imported once.
• Flavoured picks — findBestMoment(vod, { flavor, avoid }): paste asks for a frame that is striking on its own (a face, a gag, something odd on screen); clip asks for a beat that plays out over 25 s (a line, a reaction, a sound, chat exploding). The prompt lists the already-used timestamps and the model is told to stay away from them; if it still lands next to one, the objective signals (viewer clips, chat spikes, richest scene notes) pick a free spot.
• The paste job no longer clips the paste's own moment; it asks for a second, different beat from the same VOD for the clip. The clip backfill avoids every offset a paste used.
• Live pastes — the stream-memory vision call now also answers "is this frame screenshot-worthy?" with a caption (no extra call). When it says yes, stream-memory-job posts the frame as an image paste right away (≤ 1 per stream per 90 min, dark-frame and registry checks, setting ai_live_pastes_enabled, default on). These pastes carry metadata.live = true and no VOD link.
• Fewer tokens — VOD showcase scores are cached for 7 days (home_hero_moments.rankCache); a run only scores VODs it has never scored. Every llm.complete call keeps its kind so ai_usage shows exactly where the budget goes (moment_vod_rank, moment_pick, moment_frame, stream_memory, auto_clip_confirm).

AI-inferred category#

Go-live used to default every stream to irl, so every AI prompt (AI viewers, streamer overviews, VOD overviews, the Arena, sound detection) assumed everyone is an IRL streamer. Now the category selector defaults to Auto (stored as NULL), and the stream-memory rollup — the same summarise call that already runs — returns { overview, category, tags } with category from the fixed taxonomy (outdoors, travel, building, music, gaming, robot, desktop, irl, other). It is stored on streams.ai_category / streams.ai_tags and the channel inherits the latest read (channels.ai_category).

Everything reads the effective value: listing queries return COALESCE(ai_category, category) AS category (the self-selected value is still there as chosen_category), stream_category on VODs/clips, the AI viewers' session block ("judged from the stream itself"), streamer overviews, sound detection and the Arena all prefer ai_category. A streamer can still pick a category manually; the AI read wins once it exists.

This page is rendered from docs/vods-and-clips.md in the OpenVibe.Live repository. Found a mistake? Edit it there.