YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context
Windows native selective-redaction evidence
The final Windows backend keeps full monitor dimensions, composites isolated allowed surfaces in native z-order, and paints excluded visible regions black. It preserves allowed foreground overlap and applies native global opacity once. Unknown or changing scenes fail closed. The implementation does not add a focused-window setting or crop.
Native environment: Windows 11 Pro 24H2, build 26100.9168, Microsoft Hyper-V Video driver 10.0.26100.1150, DWM/Explorer in the interactive console session, one 1024×768 Generic PnP display. This is one Azure virtual hardware configuration, not proof across every Windows version, GPU or compositor.
Source and build identity
Final production source matches parent commit 24e1403044156e4c1d3462c773a9d8de47617818. The native worker's branch history differs because it integrated parent files before the final parent rebase; exact source hashes link the tested production code.
| File | SHA256 |
|---|---|
screenshot_privacy/windows.rs |
39e046222ddf19e354619d42adfcaae2cb7149f0b413edbc5745ceee80bd5b6d |
wgc_capture.rs |
aae27ee016b32fe65f306e5b14391aa47f32c9b25f4ae694a08f38a4dd3e69ab |
event_driven_capture.rs |
307749f49ed27be32a39fca91569ccee3d2a22b16a50b9b76943c5b33487acf9 |
| Shared probe | 07e04cbbecb0aca12aa6e8f89942304e64c695dca48b9606deb076138469d958 |
| Windows fixture | cb2e4d18383bbab3cd3101fda5512e17638453d14c4991cfe0b8dfc3444d2bca |
The final native working files use LF. Five narrow privacy unit tests passed, including the two global-opacity regressions; one live-console unit test was intentionally ignored because native fixture/probe checks exercise that behavior. Exact native unit command: cargo test -p screenpipe-screen --profile debug-dev screenshot_privacy::windows::tests -- --nocapture with CARGO_TARGET_DIR=C:\spdev, exit 0. Unit tests used debug-dev only. Final probes and the final recorder use explicit --release; build receipts record source and executable hashes and verify copied binaries against the release output path.
Final release screen checks
| Native case | Observed output and assertion | Coverage |
|---|---|---|
| Five windows, one excluded above the others | Full 1024×768, black excluded region; a separate visible ROI from each of four allowed windows retained; no excluded marker in PNG/JPEG | Final release targeted recheck |
| Strict title include list | Four allowed ROIs retained, excluded marker absent | Final release targeted recheck |
| Genuine blank title, app include policy | Resolved allowed app remains visible without a caption | Final release targeted recheck; an initial renamed-executable harness attempt was invalid and excluded |
| Configured exclude pattern with no match | Allowed content retained across the screen | Final release targeted recheck |
| No visible excluded window | Allowed content retained across the screen | Final release targeted recheck |
| Excluded window fully occluded | Allowed foreground preserved | Final release targeted recheck |
| Excluded window minimized | Allowed content retained; no excluded marker | Final release targeted recheck |
| Half-transparent allowed foreground over excluded content | All 102,000 overlap pixels are RGB(0,128,128), preserving 50% cyan over black | Final release numeric ROI |
| Excluded underlay changes magenta→red | Raw overlap changes from RGB(127,128,255) to (127,128,128); filtered ROI is byte-identical across backgrounds | Final release raw positive controls and filtered A/B |
| Allowed→excluded→allowed title transition | 192,024 allowed pixels before and after; zero while blocked | Prior guards3 native correctness run; same production logic, debug-dev profile |
The seven final stable scene rechecks use one frame per case, alongside the longer earlier native matrix. They do not pretend to be a second full matrix. The first release A/B attempt used the ordinary comparison scene, not the half-transparent fixture; only alpha-ab-release-acceptance/ is acceptance evidence.
Representative stored JPEGs
These are actual SnapshotWriter JPEGs, copied without re-encoding. Their original saved path suffixes and hashes are in the image manifest.
| Original filtered backend | Final release candidate |
|---|---|
![]() |
![]() |
Five windows, one excluded · Transparent foreground over excluded region
Clean optimized comparison
Both binaries use the same frozen probe, compiler, machine, release profile and five-window fixture scene. The baseline uses the original filtered backend at 3bb684b71567885bef406bfc3186f9da1c25a3af, with no production baseline changes. Each side runs a separate 20-frame warmup process, then a fresh 120-frame measured process. No compiler was active during measurement. This is not a same-process warmup discard.
| Metric | Original filtered baseline | Candidate |
|---|---|---|
| Saved / attempted | 120 / 120 | 119 / 120 |
| Excluded-marker leaks | 0 | 0 |
| Capture mean | 597.52 ms | 468.06 ms |
| Capture median | 594.86 ms | 463.96 ms |
| Capture p95 | 664.38 ms | 526.83 ms |
| JPEG write median / p95 | 22.04 / 24.88 ms | 22.72 / 25.18 ms |
| Saved-image read median / p95 | 10.03 / 10.53 ms | 10.08 / 11.39 ms |
| Probe CPU time | 9.09375 s | 10.234375 s |
| Peak RSS | 45,891,584 bytes | 44,404,736 bytes |
| Peak / final sampled handles | 984 / 975 | 273 / 269 |
The candidate's first attempted frame was withheld with window scene changed during privacy capture; all saved frames were 1024×768. Capture mean decreased 21.67%, while measured process CPU increased 12.54%. These results do not establish zero performance cost. The original baseline's sampled handle growth is substantial; static candidate bounds do not prove behavior under distinct-window churn.
comparison-release/recomputed-summary.json computes standard median and nearest-rank p95 from accepted frames. The native summary is preserved as a raw receipt, but used a lower percentile index. CPU/RSS/handle samples include process startup. Earlier stage2's 600-frame release run measured an older candidate against an unfiltered raw control; it is separate correctness/stability evidence, not the original-filtered comparison. Guards3's accidental debug-dev comparison is excluded from performance conclusions.
Final recorder acceptance
Watch the final release recorder session, with native desktop footage, actual stored JPEG samples, narration and captions. Video provenance.
The frozen release recorder saved 19 excluded-focused JPEGs with 19 matching canonical index rows. Every matched row omits focused identity, text and references and sets focused false. Strict include saved 17 useful JPEGs. Production db check exited 0. Native initial pixel checks sampled the excluded region. An independent host audit then scanned every pixel of all 36 stored JPEGs: zero excluded-marker pixels, more than 1,000 allowed cyan pixels per image, and every image 1024×768. The full-image receipt and audit script are included. These two runs explicitly used a 1-second idle interval and disabled visual checks, so they establish persistence and metadata behavior, not default quiet cadence.
Actual excluded-focused recorder JPEG · Actual strict-include recorder JPEG · Persistence and metadata receipt · Strict include receipt
The same frozen release binary then passed three isolated negative cases: excluded focus with URL policy, excluded focus with DRM pause policy, and a PrivacyBlocked Incognito focused title. Each ran more than 25 seconds with a 1-second idle interval and saved zero JPEGs and zero canonical frame rows. Production db check exited 0 for all three. These are synthetic policy-gate tests, not actual protected-video playback or browser URL extraction tests. Exact commands, foreground identities, native exits and empty row arrays are in the gate receipts. Info-level logs record startup configuration and capture opportunities but do not expose each rejection reason.
The valid quiet-static run omitted both visual-check and idle-interval overrides. It lasted 63.637 seconds and stored four full-monitor frames: manual, app-switch, one startup visual-change, then one idle capture 30.45 seconds later. There was no repeated visual-change loop. All four index rows contain no focused identity/text/references. An independent full-image audit passed all four stored JPEGs: full 1024×768 dimensions, useful allowed pixels and zero excluded-marker pixels. Production DB check exited 0. Six same-process resource samples span 50.065 seconds: CPU delta 0.359375 seconds, RSS 88,121,344→70,336,512 bytes, and handles 856→852 (range 852–858). This is a short observed cadence check, not a whole-process performance guarantee. Cadence receipt, samples, stored JPEG.
An earlier 44-second cadence attempt was stopped prematurely by the validator and excluded from acceptance. Raw native elapsed fields truncated some fractional start seconds; recomputed timings use the preserved ISO timestamps and final log write time. The final wrapper's export failed because the worker used the previous results folder; the completed native receipts and JPEGs were recovered by a read-only export and verified by SHA. No production source or binary changed.
The distinct-target release probe attempted 70 frames over 63.660 seconds: 61 accepted, 9 withheld, zero excluded-marker leaks. Twelve allowed targets were created over about 24 seconds, followed by 39.533 seconds of static-capture recovery sampling in the same process. First active handles 273 increased to 285 at the last recovery sample; RSS decreased from 40,755,200 to 33,439,744 bytes. Whole-run peaks were 296 handles and 41,291,776 bytes. Five withheld captures detected scene changes and four encountered native CreateFreeThreaded invalid-parameter errors during target churn. Engine compilation was active, so this lane is not a clean latency benchmark. Callback-level Closed/eviction events were not instrumented; exercising more than eight distinct targets is not direct proof of every callback. Residual handle growth remains a limitation.
Shell-preview native coverage is unavailable. UI Automation identified the visible taskbar button at (680,720) with 44×48 bounds, but two real hover attempts did not materialize a thumbnail window. Raw captures contained the excluded source window, but not an excluded thumbnail. This is not a passing preview privacy test; no further preview retries were performed.
Limits
Native multi-monitor, HiDPI, other Windows builds/GPUs, and all per-pixel layered-window variants are not established by this single-display guest. Known Explorer thumbnail/switcher surfaces are conservatively blocked when source-window ownership cannot be verified, including potentially allowed previews. That can black more than the ordinary excluded app rectangle. A live preview test needs an enumerated visible preview and a raw excluded-thumbnail positive control; an absent preview is not a pass.
Whole-scene before/after checks withhold detected geometry, identity, title, opacity and z-order changes. They cannot establish that an intermediate state which returns identically between snapshots never occurred. Final recorder negatives and target-churn resource limits must be read alongside the visual checks; neither zero marker counts nor a static screenshot alone establishes universal privacy or leak freedom.

