Recording

Recording

When to use

Reach for this skill any time you touch the recorder: the record button, the in-progress toolbar, permission prompts, chunked upload flow, or the camera bubble. If you're adding support for a new source (e.g. tab capture, iPhone continuity camera) or changing how chunks are finalized server-side, this is your map.

Data model touched

• recordings — the row gets created as soon as the user presses Record or imports a source. Native/file recordings transition uploading → processing → ready (or failed). videoUrl, durationMs, videoSizeBytes, width, height, hasAudio, hasCamera are populated as the upload streams in. Loom imports use import-loom-recording and become ready with either a Clips-hosted MP4 or a Loom embed when MP4 export is unavailable.
• application_state.record-intent — the agent writes this when it wants to start a recording. The UI reads and clears it, then prompts for permission.
• application_state.navigation — set to { view: "record" } while the recorder is active.

Binary uploads hit the custom API routes (/api/uploads/:id/chunk and /api/uploads/:id/abort) rather than actions, because actions aren't the right tool for binary streaming bodies. The final chunk calls finalize-recording. Loom URL imports are metadata-only and should go through the import-loom-recording action.

Some recordings are linked to a meeting — when meeting_id is non-null on the recording row, it was created via start-meeting-recording and both the recording and meetings skills apply. See the meetings skill for the bidirectional link.

Lifecycle

1. Intent. Either the user clicks Record (global Cmd+Shift+L) or the agent calls pnpm action start-recording --mode=screen. The agent version writes record-intent to application state; the UI picks it up and initiates the same flow as a user click.
2. Permission. Call navigator.mediaDevices.getDisplayMedia({ video, audio }) for screen, getUserMedia({ video, audio }) for camera. Do not prompt without a user gesture. The agent path relies on the UI's button — we never bypass the browser's permission model.
3. Create row. As soon as the stream is granted, call create-recording to insert the row with status: "uploading" and a pre-generated id. That id is used for every subsequent chunk upload.
4. Record. Start a MediaRecorder with mimeType: "video/webm;codecs=vp9,opus" (fallback to vp8, then browser default). Use timeslice: 2000 so chunks arrive every 2s.
5. Upload each chunk. ondataavailable POSTs the chunk bytes to /api/uploads/chunk with headers X-Recording-Id and X-Chunk-Index. Don't retry inline — buffer failed chunks in IndexedDB and let a background worker re-send.
6. Live transcription. Alongside the MediaRecorder, useLiveTranscription runs the Web Speech API to accumulate transcript text in real time. On stop, the client calls save-browser-transcript to persist the result immediately — no API key needed. Desktop recordings use local Whisper/macOS speech first when available, and fall back to Web Speech in the webview on non-mac before relying on upload transcription.
7. Finalize. On stop, send the final chunk to /api/uploads/:id/chunk?isFinal=1. The route calls finalize-recording, which stitches chunks, makes the media seekable (see below), uploads the finished media when storage is configured, transitions status to ready, then kicks off request-transcript for higher-quality output (see ai-video-tools).
8. Navigate immediately. Desktop recorders open /r/:id as soon as Stop starts finalization. The recording row already exists, so the page can show the title, share link, and upload progress while it polls from uploading or processing to ready. Do not wait for the upload/finalize response before opening the page.

Mobile companion lifecycle

The Agent-Native mobile app uses the same recording rows and binary upload routes with native capture primitives:

1. Camera video/import uses expo-camera / the system photo picker; meeting audio uses expo-audio with background recording enabled.
2. The native file is copied into the documents directory and a typed AsyncStorage capture job is written before upload starts.
3. create-recording receives a stable client-generated id, sourceAppName: "Agent-Native Mobile", and the container MIME type.
4. Upload reads at most 3 MiB through an Expo FileHandle and persists the next chunk index after every acknowledged POST. The 4 MiB server cap still applies.
5. Foreground/resume reconciliation polls /api/uploads/:id/status; retryable failures back off and never discard the local file. Completion can post a local notification and the recording appears in Clips everywhere.

Mobile audio M4A is an MP4 container and uses the recording route's MP4 MIME alias. The bytes are not converted or loaded whole into JavaScript memory.

Because the phone persists each audio/video file before any network work and resumes bounded chunk uploads from its durable queue, a transient upload failure never loses the capture. Do not ask users to keep a capture screen open or to re-record after a failed upload; tell them to reconnect Clips and retry the saved job from mobile Home.

Mobile meeting capture is microphone-only; never claim it has the desktop mic-plus-system-audio fidelity. See the meetings skill for what that means for attendee attribution.

Seekable playback (don't ship raw MediaRecorder output)

Raw MediaRecorder files are not friendly to progressive HTTP playback, which shows up as "clip takes minutes to load" and "re-buffers every time I seek" even though the file downloads fine:

• MP4 is written with the moov metadata atom after mdat, so a player must fetch the whole file before it can start or seek.
• WebM is a live stream with no Cues (seek index) and an unknown Segment duration, so Chrome won't honor currentTime = X and has to scan/download.

finalize-recording fixes this before upload: MP4 gets pure-TS faststart (server/lib/faststart.ts), WebM gets a lossless ffmpeg -c copy remux that writes a SeekHead + Cues + real duration (server/lib/video-remux.ts). Both are best-effort — on failure we upload the original, never block finalize. Recordings above CLIPS_INLINE_REMUX_MAX_BYTES (default 200 MB) skip the inline pass and are repaired in the background.

The streaming/resumable upload path forwards raw bytes straight to the provider and cannot rewrite them inline, so finalize-recording schedules a background ensureRecordingSeekable pass for those.

To repair clips uploaded before this existed (or via streaming), call the reprocess-recording action: --id, --ids='[...]', or --all --limit=N. It re-fetches provider media, rewrites it, re-uploads, and repoints the row. It's idempotent (already-seekable clips are skipped unless --force) and only touches provider-hosted clips owned by the caller. This is the right tool when a user reports a specific slow/buffering clip.

Seekability remuxing cannot repair a recording whose audio continues while the video track has a large timestamp gap (common when a mobile browser suspends the camera after the user switches apps). For a clip that freezes or appears to stop before its declared duration, call reprocess-recording with --normalizeTimeline=true. That explicit mode uses the same owner-scoped fetch and upload flow but fully transcodes to a constant-30-fps faststart MP4 (H.264 + AAC). It preserves audio and duplicates the last decoded video frame through missing-frame gaps. The action uploads to a new media object and atomically repoints the row only after verified output is stored; any transcode, audio verification, upload, or concurrent-update failure leaves the original URL and format untouched.

Loom import

Use import-loom-recording for Loom share or embed URLs. The action validates the Loom URL, reads Loom oEmbed metadata from Loom's public endpoint, and creates a ready recording with Loom's embed URL, thumbnail, title, duration, and dimensions. When Loom exposes a signed public transcript JSON URL on the share page, the action imports that transcript into Clips and stores normalized segments; never store Loom's signed CDN URLs.

When Loom exposes a downloadable public MP4, import-loom-recording downloads it, reuploads the bytes to Clips storage, and creates a ready, playable Clips-hosted recording, importing Loom's public transcript when the share page exposes one. If Loom allows public playback but does not expose an MP4, the action keeps a ready, embed-backed recording and still imports the transcript when available.

Embed-backed Loom imports render a Loom iframe and hide the native Clips editor. Reuploaded Loom imports support Clips-native trimming, exports, frame extraction, and upload-based transcription; ask the user to upload the original video file when those capabilities are needed for an embed-backed import.

Browser diagnostics and recorder install options

Browser recordings can save bounded, redacted diagnostics through createBrowserDiagnosticsCapture: console messages plus fetch/XHR method, URL, status, duration, and errors. The browser recorder only captures activity from the recorder page itself. The Clips Chrome extension is the active-tab path for browser logs: it launches /record with an extension capture session and passes developerLogs=1/0, then saves diagnostics with source extension.

The Web Store listing is live, so the public Chrome extension option shows by default. UI prompts that otherwise say "Download desktop app" use the shared install-choice component (CaptureInstallButton / CaptureInstallInlineLink) to offer two options: Chrome extension for browser logs, and desktop app for the most seamless native capture. Set VITE_CLIPS_CHROME_EXTENSION_ENABLED=0 to hide the Chrome option again, or VITE_CLIPS_CHROME_EXTENSION_URL to point at a different listing.

save-browser-diagnostics is UI/internal. It stores bounded console logs plus fetch/XHR method, URL path/query keys, status, and duration. It never captures headers, bodies, cookies, or network URL query values. Console text keeps useful non-secret values while redacting credential-looking keys and headers. Use get-recording-player-data for the full diagnostics payload when you have editor access; see the video-sharing skill for the narrower redacted stream that public agent context exposes.

Chrome extension

The extension lives in chrome-extension/. It launches /record with clipsExtensionId and clipsCaptureSessionId, and the recorder sends CLIPS_CAPTURE_START / CLIPS_CAPTURE_STOP / CLIPS_CAPTURE_CANCEL back to the extension. It uses the Chrome debugger API only on the tab the user launched from, only while a recording is active, and returns the same redacted diagnostics shape saved by save-browser-diagnostics.

The extension also enhances GitHub issue and PR markdown: a narrow github.com content script detects Clips /r/, /share/, and /embed/ links, then renders the existing /embed/:id player in an extension-owned preview iframe so the video is playable without leaving GitHub. Keep this scoped to GitHub unless there is a deliberate permission review.

Folders, spaces, and bulk moves

Use move-recording for both single and bulk folder moves. Pass id for one clip or ids for the selected clips, and folderId: null to move them to the library or space root.

Pause / resume

MediaRecorder.pause() / .resume() are supported in all evergreen browsers. Keep a single MediaRecorder instance across pauses — don't tear down the stream, or the permission prompt will fire again. While paused, the upload worker keeps draining its buffer so we catch up before the user stops.

Camera bubble

When mode is screen+camera, "the bubble" is two different things:

• On-screen self-view. CameraBubble renders a plain, unrecorded <video> element the user drags around to frame themselves during countdown/recording. It is never captured — it's just local framing UI.
• Recorded composite. Display capture can't include that DOM element once the user records another window/app, so the saved video has to bake the camera circle in before MediaRecorder ever sees a frame. createCameraCompositeStream (app/lib/camera-composite.ts) draws the display video plus a clipped, mirrored camera circle onto an offscreen <canvas> and feeds canvas.captureStream() into a single MediaRecorder. There is no second MediaRecorder and no server-side ffmpeg stitching — the composite happens client-side, live, before upload.

Because that draw loop runs continuously for the whole recording, keep its CPU/GPU cost bounded: a Worker-based timer drives the draw loop at the capture frame rate (SCREEN_CAPTURE_FRAME_RATE, 24fps — falling back to requestAnimationFrame only if Worker creation fails, e.g. under a strict CSP), the canvas is hard-capped to 1080p-class dimensions (1920px on its longest edge) even if the source display track is Retina/4K, and the bubble's drop shadow is pre-rendered into a small cached sprite (keyed by bubble size) instead of re-blurring with shadowBlur every frame.

Restart (desktop)

Restart throws the current take away and immediately starts another one. It must never re-acquire the screen: the toolbar click reaches the popover through async Tauri IPC, which carries no user activation, so a second getDisplayMedia would throw. Instead the dying session hands its live display and mic streams to the replacement — RecorderHandle.discardForRestart() returns a RestartHandoff that startRecording accepts as preAcquiredDisplayStream / preAcquiredAudioStream.

Ownership is the opposite of the camera's. The popover owns the camera stream for the whole session and re-hands it unchanged, so restart must not bump bubbleSessionEpoch or clear bubbleStreamTransferredToRecorder. The handed-off display and mic streams belong to the recorder, so the new session stops them on its own stop/cancel, and whoever asked for the restart must stop them if the new session never comes up.

cancel() and discardForRestart() share one discardTake(forRestart) per backend, and the difference is which teardown they run. Cancel ends the session: hide_overlays (which destroys the camera bubble window too) and clearRecordingState(). A restart ends only the take, so it uses hide_recording_chrome, which spares the bubble, and leaves the recording state active.

Stop, cancel and restart are mutually exclusive terminal transitions, so each gets its own promise slot — never reuse cancelPromise for a discard, or a cancel arriving mid-restart returns the take-level teardown and the session never ends. A cancel that lands after a discard still owes the session half, plus stopping the capture the retake never took ownership of. On the app side restartInFlightRef latches synchronously so stop and cancel events cannot act on the recorder a restart is already tearing down.

Native full-screen backends re-acquire capture in Rust and hand off nothing. resolveRestartHandoff vets the inherited streams before anything is acquired, and it treats the two kinds of capture differently: an ended display share is fatal and fails with RESTART_CAPTURE_ENDED_MESSAGE, because re-acquiring it would surface an activation error that names the wrong cause, while an ended microphone is just dropped and re-acquired normally. A restart also has to wait for the old take's transcriptionCapture.cancel() — the engine is process-global (audio_transcription_stop / native_speech_stop), so a late cancel would stop the replacement's engine. And stop() after a discard must throw rather than answer with the discarded take's id, or an aborted recording gets published as a finished one. The recording-flow latches (recordingFlowGateRef, recordingFlowActive, clipsForceAlive, set_recording_state) stay held across the restart; releasing them the way cancel does lets the popover blur auto-hide fire mid-restart.

Holding those latches has a catch. The show_toolbar effect is keyed on isRecording || recordingFlowActive, so it does not re-run when the flow never leaves — but the discard already closed the toolbar window. Restart therefore bumps recordingChromeEpoch to rebuild it. The countdown needs no such nudge; the recorder recreates it on every start.

Error recovery

Code sketch

// app/hooks/use-recorder.ts
export function useRecorder() {
  const start = async (mode: "screen" | "camera" | "screen+camera") => {
    const stream =
      mode === "camera"
        ? await navigator.mediaDevices.getUserMedia({
            video: true,
            audio: true,
          })
        : await navigator.mediaDevices.getDisplayMedia({
            video: true,
            audio: true,
          });

    const { id } = await callAction("create-recording", {
      title: "Untitled recording",
    });

    const rec = new MediaRecorder(stream, {
      mimeType: "video/webm;codecs=vp9,opus",
    });
    let chunkIndex = 0;
    rec.ondataavailable = async (e) => {
      if (!e.data.size) return;
      const params = new URLSearchParams({
        index: String(chunkIndex++),
        total: "unknown-until-stop",
        isFinal: "0",
      });
      await fetch(`/api/uploads/${id}/chunk?${params.toString()}`, {
        method: "POST",
        headers: { "Content-Type": "application/octet-stream" },
        body: e.data,
      });
    };
    rec.onstop = async () => {
      // Send the final chunk with isFinal=1; the route calls finalize-recording.
    };
    rec.start(2000);
    return {
      id,
      stop: () => rec.stop(),
      pause: () => rec.pause(),
      resume: () => rec.resume(),
    };
  };

  return { start };
}

Rules

• Never start a MediaRecorder without a user gesture (or a user-initiated record-intent).
• Never re-prompt for permissions on pause/resume — reuse the stream.
• Never fire the upload from the main thread if the chunks are large — prefer a web worker for anything longer than ~60s.
• Never read a complete mobile recording into JavaScript memory. Use bounded FileHandle reads and checkpoint every acknowledged chunk.
• The recordings row must exist before the first chunk is sent.
• On every lifecycle change, write navigation → { view: "record" } → { view: "recording", recordingId } so the agent can see what's happening.
• All AI generated during/after recording goes through the agent chat — see ai-video-tools.

Related skills

• ai-video-tools — transcription kicks off when upload completes.
• video-editing — after recording, users edit via non-destructive editsJson.
• server-plugins — why the upload is an /api/ route, not an action.
• real-time-sync — how the UI learns about status transitions from uploading → ready.