Session replay

The SDK keeps a rolling window of what the app rendered in memory. When a report is filed that window is frozen and attached; the rest is discarded. Nothing is uploaded unless someone files a report, and there is no MediaProjection permission prompt — this is not screen recording.

Turning it on

Server-gated — enable it for the app in the dashboard, and the SDK picks it up from GET /api/config on the next start.

There is no client-side switch for replay on Android — the server gate is the only control. If you need it off for a particular build, turn it off for that app in the dashboard.

Crash-only sessions

An app can turn the on-device reporter off with EverframeConfig(..., reporterEnabled = false) to report crashes from production builds without recording every session. Each session is one of two kinds, and the Session replay card in the dashboard sets a replay coverage for each:

Each session draws once, when it starts, and keeps the result: a later rate change applies to new sessions, not to ones already running.

The on-device reporter stays off for the whole session: report.open() throws.

An error carries a replay only while its session records. At the default crash-only coverage of 0%, errors from crash-only sessions carry no replay. At 5%, errors from about 5% of crash-only sessions do. See Configuration for the switch.

Format

Android replays are everframe-vtree-v1 — a frame timeline, not video. A viewport, a list of frames, and per-frame operations on a lightweight node tree carrying role, frame, text, bg, alpha, cornerRadius and masked.

That keeps a replay in the tens of kilobytes rather than the megabytes video costs, and it means text stays text — searchable, and redactable before it leaves the device. No video encoder runs, which is the difference between a buffer you can leave on and one you cannot, especially on low-end hardware.

Both Compose and classic View hierarchies are captured; the recorder works on the rendered tree, not on your composables.

Images

Image bytes are optional and content-addressed, keyed by the first 16 hex characters of their sha256, so the same drawable across many frames costs one entry. The server’s asset mode (off, bundled, all) decides whether they travel; see Server config.

An unresolved image renders as a placeholder box rather than failing the replay.

Privacy

Sensitive views — Modifier.txSensitive(), TXSensitiveView, markSensitive() and password fields — become a single node with masked: true and no text; the subtree beneath is dropped. The masking happens in the producer, before the frame is written — masked content never exists in the buffer. See Sensitive content.

Compose text is captured: the text of a Text composable lands in the node, redacted. The contents of text fields are never read, so what a user types is not in the replay unless another composable displays it.

Cost

Frames are captured on the choreographer’s cadence and diffed, so a static screen costs almost nothing. replayDurationSec bounds the window; samplingRate decides what fraction of sessions allocate a buffer at all.