Skip to content

Record your screen, ship a demo. Free and open-source, GPU-accelerated, no watermarks, no subscriptions. Windows, macOS, Linux. Actively maintained.

docs(testing): log the macOS and Linux GPU pass of the filled-card tilt fix

EtienneLescotd5d7523de9ab
diff --git a/technical-documentation/testing/manual-e2e-checklist.md b/technical-documentation/testing/manual-e2e-checklist.md
index 7620218b5..545325051 100644
--- a/technical-documentation/testing/manual-e2e-checklist.md
+++ b/technical-documentation/testing/manual-e2e-checklist.md
@@ -990,3 +990,5 @@ These checks are the human slice: an agent cannot click Apple's picker (see abov
| 2026-10-17 | `v2.1.0-rc.6`, `release/v2.1.0` @ `36b3814`, DMG `Openscreen-macOS-Apple-Silicon-2.1.0-rc.6.dmg` from the GitHub pre-release (build run 39010637863): 140 603 025 bytes, sha256 matches the release asset's digest; `spctl` accepted, Notarized Developer ID (M4LK7C6S84); `stapler validate` OK; `CFBundleShortVersionString` 2.1.0-rc.6. Installed over rc.5 | macOS 27.0.1 (26A434), Apple M1 (arm64), 8 GB, one 1920×1080 display at 120 Hz (1×); OBS Virtual Camera (no physical camera), BlackHole 2ch | **Partial — minimal pass toward stable: no critical bug; all 8 checked items pass** | Real OS input through computer-use, launched with `OPENSCREEN_MAC_SOURCE_PICKER=legacy`, `OPENSCREEN_DIAGNOSTIC=1` and `--remote-debugging-port` (read only: the `scripts/perf-ab/probe.js` paint probe, the playhead text and the editor's media elements). Measured with the app's bundled `ffmpeg`. Nobody else used the machine during the measures. **Smoke:** 1. The HUD appears. The tray icon (the green one in the menu bar; the aperture icon beside it is OBS's) opens Open / Check for Updates / Update Settings / Send Anonymous Usage Count / About / Save Diagnostics / Permissions… / Quit. *Open* brings back the HUD after *Hide recording bar*. 2. HUD take, 34 s, OBS webcam and microphone, six clicks on a 60 fps test pattern: the editor opens on it with the webcam PiP and two automatic zooms (1.25×, 1.80×). The take is 1920×1080 H.264 at about 56 fps, plus AAC 48 kHz stereo (mean −21.4 dB, real signal). 3. Playback: the preview moves. Probe 28.3–29.0 paints/s, p50 33.4 ms, p95 36.6. The preview's `<audio>` element plays unmuted at volume 1, in step with the playhead (4.27 → 6.34 s at 0:06.4); not judged by ear. Capture: 0 steps back, 28.8–29.5 visible changes/s. Each of two 15 s runs had one stall of about 200 ms (232 ms 11 s in, then 198 ms 2 s in; frozen on screen for 134–145 ms). rc.5's 16 s run on the same machine had none. The diff since rc.5 is export-only (#1304), so this is noted as an observation, not a regression. 4. Export MP4 1080p (1752×1080, 60 fps): 2031 frames over 33.87 s, AAC stereo at −19.2 dB mean, decodes with no error; the frame at 8 s shows the zoom and the webcam. 5. Quit, then relaunch and *Open Studio*: the project reopens intact, with its zoom, cuts, Full Camera, orange background and 15 % padding. Both project files are byte-identical (sha1) before and after the quit and the reopen. **Last fixes:** 6. Full Camera at a cut (#1294, #1304): the rc.5 cuts recreated exactly on a new held take, with the trims set to 5.043032–6.284394, 9.051596–10.292957, 13.098952–14.340314 and 17.120447–18.361809 s (app closed, values written into the project file) and Full Camera 2.961–20.5 s. Exported MP4 1868×1080, 60 fps, 948 frames. The first frame after each cut (303, 469, 637, 804) equals the next frame (mean grey diff 0.0) and differs only from the frame before the cut: no outlier, no trail. Frames 469 and 804 are the ones that ghosted on rc.5 (row above). 7. Pause on a held frame (#1296): held take (still screen, pointer moving). Paused at 0:04.2 while the pointer was moving, then padding 50 % → 15 %, then a solid orange background. The playhead still reads 0:04.2 and the pointer keeps its place over the recording. 8. Chat rewind (#1292), DeepSeek `deepseek-v4-pro`: two turns in a new conversation (a zoom at 1–3 s, then a "Rewind me" annotation at 5–7 s), then the second turn rewound straight away, in the same view, with no conversation switch or reload. Toast "Rewound to the start of that message"; the annotation leaves the timeline and the document; its text returns to the composer; `chat.json` keeps only the first exchange. No "Message not found". **Not covered:** Apple's picker; a physical camera; the Intel preview path; GIF export; the rest of the checklist (earlier rows). Takes, projects, exports and DMG moved to the Trash; HUD camera and microphone set back to Off. Side effect: choosing *Open* in the first menu-bar icon tried showed OBS's own window. |
| 2026-10-17 | `v2.1.0-rc.6`, `release/v2.1.0` @ `36b3814`, `Openscreen-Linux-2.1.0-rc.6.AppImage` from the GitHub release (build run 39010637863, all jobs green). sha512 matched `latest-linux.yml`; run directly, not installed (the `.deb` needs a sudo password). For A/B: the rc.5 AppImage from its release (sha512 matched) on the same projects | Ubuntu 24.04, GNOME Shell 46 **Wayland**, AMD Radeon 610M, 1920×1080 @ 60 Hz, USB2.0 HD UVC webcam (30 fps), built-in mic | **Partial — minimal pass toward stable: all 9 asked items pass; no critical bug found** | **Setup.** Driver: the `computer-use-linux` bridge of the rc.3 row, one RemoteDesktop grant. Clicks for auto-zoom came from Etienne pressing the touchpad: the driver's clicks do not reach `/dev/input`. **Smoke.** **1.** The AppImage launches; the HUD shows with camera, mic and system audio on; the tray menu opens (Open … Quit) and quits cleanly. **2.** A HUD take with webcam, mic and system audio. The first two attempts had no physical click (the timing of the request was unclear, not a defect). On the third, 4 touchpad presses were captured (8.15, 9.33, 10.51, 11.78 s). Stop opened the editor on the 45.1 s take with the webcam and **one auto zoom at 1.5× (7.65–13.28 s)**. **3.** Playback from 0, 12 s window, measured as in the rc.3 row (a 60 fps helper capture of the canvas): rc.6 gives 21–31 changes/s; stills of 34 ms ×185, 51 ×56, 68 ×22, 85 ×3, 102 ×1; 317 ms at start; 0 step-back. rc.5 on the same project and window gives 24–30/s with a 117 ms maximum, 300 ms at start, 0 step-back. So there is no new stutter. The webcam is AI-cut out in this project's look. Sound: the speaker sink monitor recorded during playback follows the take's audio track, with an envelope correlation of 0.99 at the expected offset. **4.** Export MP4 1080p 60 fps from the editor's dialog (save dialog → `~/Downloads`): H.264 2646×1080 (dual-frame layout), 45.07 s, AAC stereo 45.08 s (mean −25.8 dB), decodes without error, the 1.5× zoom is rendered, about 55 s to export. **5.** Tray → Quit (clean `[app] quitting`), relaunch, open the studio: the project reopens intact. The file's sha256 is unchanged, and `getCurrentDocument` is identical: one 45.12 s clip, webcam, the zoom, no trims. **Fixes.** **6. Full Camera at export (#1294, #1304): pass.** Same 45 s take, Full Camera 3.006–40 s crossed by 7 trims ending at 6.388, 10.0833, 14.7177, 19.4567, 24.2229, 29.34986635 and 34.292957. Motion blur 0.35, PiP. Exported at 1080p60 over MCP and scanned frame by frame for one-frame outliers. rc.5 gives 5 ghosted frames, exactly the first frame after cuts 3–7 (674, 889, 1108, 1359, 1590): the cut ends that round up at the millisecond, consistent with the rc.5 row's three. rc.6 on the same project gives none. **7. Pause on a held frame (#1296): pass.** VFR held-frame variant of a real take (holds of ~1 s and one of 2 s at 65.07–67.09 s; the HUD timer breaks longer ones). A paused seek to 1:06.5 inside that hold draws the cursor where the telemetry puts it at 66.5 s (≈0.44, 0.66), not at the held frame's 65.07 s (0.735, 0.739). The wallpaper and then the blur changed, and the cursor and the playhead (1:06.5) stayed put; the same held at 1:07.4. **8. Chat rewind of a message sent in the current view (#1292): pass.** Sent "Add a zoom from 20 to 22 seconds", then ↶ → Rewind: the toast reads "Rewound to the start of that message" (no "Message not found"), the text goes back to the composer, `chat.json` is rewritten without it, and the 20–22 s zoom leaves the document. **Linux.** **9. Seek into a still passage, then play (#1296): pass.** On the same held-frame take: a seek to 70 s (pointer moving) gives 28–31 changes/s with nothing above 67 ms after a 350 ms start and no step-back. A seek to 30 s (only the webcam moves) gives 23–28/s, one 167 ms still 0.5 s after the start, then ≤ 83 ms. **Also seen, not filed:** the portal no longer asks again after the first take: later HUD takes and helper captures started without a picker. The second smoke take therefore started while I was polling for a picker, so it carries driver screenshot flashes. **Not covered:** the `.deb` install path; GIF export; transcription; the rest of the checklist. Takes, projects, chat file, audio peaks, media-link entries and the exported MP4 from this pass were deleted, and `~/.config/openscreen` was restored from a backup taken before the pass. |
| 2026-10-17 | PR #1319 (`claude/fix-orbit-corner-clip-8a563f` @ `2236580`), CI artifacts of build run 39046512213 (artifacts only, no release): `Openscreen.Setup.2.1.0-x64.exe` unpacked with 7-Zip and run in place, `openscreen-linux` AppImage extracted. Reference: installed `v2.1.0` | Windows 11 Home 26200 (x64), RTX 4070 Ti; Linux: Ubuntu 24.04.2 under WSL2, llvmpipe only, so the compositor's CPU backend | **Partial — fix slice only, no defect** | MCP over stdio, isolated `--user-data-dir`, no GUI. A synthetic 10 s 1440×1080 (4:3) clip in a 16:9 format with Follow cursor, padding 50, roundness 40, a depth-4 zoom 2–8 s in `orbit` with a manual focus at (0.85, 0.85). `getFrames` at 1.0–9.5 s: on `v2.1.0` the card is cut flat by an axis-aligned edge through the whole hold, and its corners are square under `left`. On the PR build, Windows and Linux alike, the card tilts whole with four rounded corners on every frame, under `orbit` and `left`. With the zoom set to `none`, the frames are byte-identical to `v2.1.0` on Windows. **Not covered:** macOS (the SSH Mac was unreachable; the geometry test ran green on Metal in CI); the editor preview, which shares the same `plan_frame` geometry; Linux on a real GPU. |
+| 2026-10-17 | PR #1319 (`2236580`), artifact `openscreen-mac-arm64` of build run 39046512213 (artifacts only, no release), app copied out of the DMG and run in place. Reference: installed `v2.1.0-rc.6` | macOS 27.0.1 (26A434), Mac mini M1 (Macmini9,1), 8 GB | **Partial — fix slice only, no defect** | MCP over stdio, isolated `--user-data-dir`, the same project as the Windows/Linux row above (4:3 clip encoded with `h264_videotoolbox`). `orbit` (12 frames, 1.0–9.5 s) and `left` (4 s, 6 s): the whole card tilts with four rounded corners, no axis-aligned cut, inside the output; it shrinks during the hold, as designed. `none` (4 s, 6 s) and the frames outside the zoom are byte-identical to rc.6. rc.6 reproduces the bug: the top-right corner is square and the top and right edges are cut flat under `orbit`, the right corners are square under `left`. **Not covered:** the editor preview; a full MP4 export; the logs do not name the compositor backend. |
+| 2026-10-17 | PR #1319 (`2236580`), artifact `openscreen-linux` of build run 39046512213, AppImage run directly. Reference: the `v2.1.0` AppImage from the GitHub release | Ubuntu 24.04.4, GNOME 46 Wayland, AMD Radeon 610M (RADV); `[d3d] … -> backend Hardware`, no CPU fallback | **Partial — fix slice only, no defect** | Same MCP pass and project as the rows above, on the hardware Vulkan backend. `orbit` (12 frames) and `left` (4 s, 6 s): the whole card with four rounded corners, no flat cut. `v2.1.0` reproduces the bug (top and bottom edges flat, corners cut from 1.5 s). `none` and the flat frames at 1.0 and 9.5 s are byte-identical to `v2.1.0` (equal md5). **Not covered:** the editor preview; `right`; a full MP4 export. |