Adom Bridge

Confirmed from ab's code, and the history answers "stolen vs designed": it was DESIGNED this way, deliberately, and the design has aged badly.

What the wire does today (verified): the whole-desktop recorder - desktop_record_start/stop/status/list and desktop_recorder_open/close - is implemented in pup (HUD = a Chrome window in a recorder profile, capture = Chrome getDisplayMedia to .webm); pup's bridge.json claims the desktop_record/desktop_recorder_ prefixes and ab's CLI routes those verbs straight to the browser bridge. ab-core natively owns only desktop_record_window_* (wgc_record.rs: WGC + Media Foundation to MP4). Why: when the whole-desktop recorder was built, ab had no native monitor capture, and Chrome's screen-share gave a ready-made consent-gated capture WITH system audio for free - so ab named the verbs desktop_* (a desktop capability from the caller's view) and delegated the mechanism to pup. Native per-window capture came later and whole-monitor capture never followed it native. Your measurement on arav-rog is the exact cost: a desktop capability dying with a browser bridge.

Plan, in order: 3 (docs, DONE now, ships in the next pkg): VERBS.md and RECORDING.md say who serves desktop_record_* today (pup, via getDisplayMedia; only record_window_* is native) and point at record_window_* as the pup-down fallback. 2 (prefix ownership, shipping next release): ab refuses a bridge manifest that claims a desktop_ prefix - that namespace is ab-core's - with ab-core precedence on any conflict as the transition rule, so pup's current claim keeps working until native lands and then yields automatically. 1 (native whole-desktop recording in ab-core, the real work): WGC on a monitor (GraphicsCaptureItem from an HMONITOR - the same stack wgc_record.rs uses per HWND) + WASAPI loopback for system audio, muxed to MP4 by a two-stream Media Foundation sink writer; same verb names and args (reason, confirmDesktopNotTabRecording, monitor, fps, audio, maxDurationMs) so callers do not change; consent = ab's own always-visible overlay in the demo-panel family (no Chrome, no HUD process). Honest sizing: the video half is a small delta on existing code; the audio half (loopback capture + A/V sync in the sink writer) is the substantive piece - a solid few hours done carefully, queued as the next block. Once it lands, pup drops the prefixes (pup-bridge's half).

Update after John's direction (2026-09-05): BOTH recorders stay, split by owner, instead of "same names, pup drops them".

Why the two coexist (this is the reason the original layering happened, re-derived): Chrome's getDisplayMedia recorder offers things a native Rust recorder does not - the OS share picker as the consent surface, system audio + mic mixing for free, hardware-accelerated VP9/AV1 MediaRecorder to WebM, tab-audio capture, cursor and surface-switching options, and it runs on macOS/Linux. ab-native (WGC + Media Foundation) offers what pup cannot - no browser/Node dependency, background capture of an occluded window, no modal picker (ab's own gate), H.264 MP4, and it survives a broken pup (arav-rog). Different tools; keep both.

The naming rule (John): everything pup implements is a pup_ verb; ab-core owns desktop_*.

• pup: pup_screen_record_start/stop/status/list + pup_recorder_open/close (aliases first, old names kept until ab-native lands). Filed on pup-bridge.
• ab-core: native desktop_record_start/stop/status/list (monitor WGC + WASAPI loopback -> MP4, same args) - ask 1, next block. When it ships, pup drops the desktop_ prefix claims and the old names route to ab automatically.

Shipped in ab 2.1.31 (insiders): ask 2 - a bridge manifest claiming a verb/prefix ab-core serves natively is refused at rescan (today desktop_record_window_); any other desktop_ claim is honored with a warning until ab-core serves it; resolve_verb and the direct API's infer_app give ab-core precedence, so pup's broad desktop_record_ prefix can no longer shadow desktop_record_window_* for ah/direct-API callers (it could before). Ask 3 - VERBS.md / RECORDING.md in the 2.1.31 pkg say who serves what today plus the split-namespace direction.

pup-bridge's half is in 2.0.397 (insiders): the pup_describe catalog and every desktop_record* verb row now say "ab-core capability served by pup on its behalf until adom-bridge#143; moving out of pup", and pup's skills / README no longer present desktop recording as pup's feature (they point at adom-bridge-recording). The prefixes stay claimed until ab-core serves the verbs natively; with ab-core precedence as the transition rule, the drop is then a one-line manifest change on pup's side. Ping pup-bridge#93 when native lands and it ships the same day.

Ask 1 shipped: ab 2.1.36 (insiders) has native whole-desktop recording.

• desktop_record_start {monitor?: index|"primary"|device, fps?, codec?: h264|h265, audio?: true, maxDurationMs?: default 30 min, cap 2 h} -> WGC on the monitor (CreateForMonitor from the HMONITOR at the same index desktop_display_info reports) + WASAPI loopback of the system mix (f32/i16/i32 mix formats -> 16-bit stereo, silence padded to wall clock because loopback yields nothing while nothing plays) muxed as AAC 192 kbps into the same H.264 MP4; the encoder thread is the single writer, so audio and video share one timeline.
• Consent surface = ab's own persistent caption at the top of the screen for the whole take ("Recording your screen and system audio (Adom Bridge) ..."), hidden on stop.
• desktop_record_stop (id optional when one take runs), desktop_record_status (no id = every take; autoStopped:true when maxDurationMs hit), desktop_record_list (finished MP4s). Give hwnd/titleContains instead of monitor for single-window mode (record_window_* is the same thing by another name). desktop_recorder_open/close retired.
• GPU (John's ask): the sink writer requests hardware transforms, and the response now reports encoder + hardwareEncoder, hardware detected from the live transform's MFT_ENUM_HARDWARE_URL_Attribute. Verified on AdomLapper: "NVIDIA H.264 Encoder MFT", hardwareEncoder:true, for both monitor and window takes. The per-frame GPU->CPU staging hop is the remaining cost; the zero-copy DXGI path is #149.
• Verified live: 2560x1600@30 monitor take, 169 frames in 6 s, vide + soun tracks (avc1 + mp4a), moov finalized, 8.3 MB; caption visible in a screenshot during the take; window mode and list verified as well.

Ask 2 finished: desktop_record_ and desktop_recorder_ joined ab-core's owned prefix set, so pup's claim is refused at rescan. Per John today there are no aliases on pup's side - pup-bridge#93 now asks for the clean drop. Ask 3 (docs) updated again: RECORDING.md and VERBS.md describe the native recorder and point at pup_screen_record_* as the separate Chrome tool. Closing.