August 2026: the clip arrives by itself

In August the Windows desktop app went from v0.0.23 to v0.0.29. The month had one theme: a feature that reports success and produces nothing usable is worse than one that fails loudly, so we opened the artifacts, the MP4s on disk, instead of reading logs that said everything was fine.

The CS2 auto-clip

When you die in Counter-Strike 2, the desktop app already knows, because the game's own state integration tells it, and the clip is cut from the replay buffer without you pressing anything. On the 30 August device smoke, on the shipped v0.0.29, one CS2 sitting produced 60 detected deaths and 62 MP4s on disk, the extra two being manual captures: one to one, no silent drops. Each clip is anchored to the round start rather than a fixed last few seconds, so a 17-second clip is 17 seconds because that is how long the round ran.

Earlier in the month the clip did not contain the death. The snapshot copied back from the newest encoded frame, and the trigger fired the instant the app saw the death, so the clip ended just before the moment it was meant to capture. v0.0.25 defers the cut by two seconds: measured on a 16.26-second clip, the death lands at 13.6 to 13.8 seconds, about two and a half seconds of delivered post-roll. On the 30 August smoke we confirmed it by eye: alive at 100 HP at 13.5 seconds, the death state at 15.5.

Recording the outcome, not the setting

Two fixes in v0.0.23 and v0.0.24 are about telling the truth. Three of twelve clips in one session were digitally silent, every sample zero at -91 dB, while the record said system audio had been captured. Five of seven League of Legends clips were pure black, a mean brightness of 0.1 across sixty frames, because the Windows capture API cannot read an exclusive fullscreen window. In both cases the app had recorded the setting rather than the outcome. Now it measures the outcome: audio level and frame brightness are sampled on the encoder path, recorded truthfully, and the warning names the action to take, set the game to borderless, rather than the fault.

The capture watchdog

v0.0.26 through v0.0.28 turned those measurements into a watchdog. If the video runs dark or the audio silent for long enough, the app requests a fresh capture session and rebinds instead of asking you to restart. A minimised window reads as no opinion on both signals, because Windows sends degenerate frames and many games mute on focus loss, so alt-tabbing never trips it. On the 30 August smoke it fired live once, on a dark-video signal after 22 bad seconds, and every clip after it was bright. Nine sampled clips from the same sitting sat at a mean level of -23.0 to -34.4 dB and a mean brightness of 63.6 to 122.3; neither failure class recurred.

Valorant, in shadow

Valorant publishes no local telemetry, so nothing tells the app you died. After eight probes we found one cue that survives: the local player's health digits vanish on death and stay absent while spectating a teammate. On a pre-registered sample of 20 clips (16.9 minutes, 1,017 frames) the detector scored 4 of 4 deaths at 0.53 false positives a minute, against a control that fires at the end of every clip and scores 3 of 4 at 1.00 a minute. Four positives is a real measurement, not a precise one, so the detector ships in shadow mode: it logs and does nothing else until a few evenings of real play say whether it holds, at about 16 microseconds a frame.

Revenge Review

Revenge Review is what the death clips are for. The first idea was a Replay Coach review of every death in a session, which cannot be built as stated: a bad match has more deaths than the daily Pro allowance covers. One session report is the better product anyway: eight deaths at three frames each is 24 frames, the size of one ordinary review, and one pass over them can say you died holding the same angle four times, which no per-death review can. No death video leaves the machine; the stills go over their own authenticated route, and sending them is a separate consent, off by default.