You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Screenshot readback design: which routes hold a bound captureScreenshot, and the admission story for --expect-screenshot-diff #4509
Parent context: #4318 group C. The 11-08 direction asked for this design issue before the readback half of #4297 merges (the baseline half shipped in #4338).
The problem
--expect-screenshot-diff on press/click needs a bound captureScreenshot on the dispatch device. Today only the screenshot route binds one (screenshot-runtime-binding.ts); the daemon touch routes build their backend through createInteractionBackend without it, so the flag is a silent no-op on every shipped surface. The 11-08 rule: every route that admits the flag must hold a bound captureScreenshot, admitted through runtime operation admission; otherwise the request must refuse before dispatch.
Options for the decision
Admit a capture operation on touch routes. The interaction runtime facts gain a screenshot-capture operation; a request with expectScreenshotDiff admits it like any runtime operation, and the interaction backend binds the admitted capture. Cross-route operation admission may interact with fix(settle): compare post-action trees in one view across every route #4500's settle unification.
Refuse on routes without a bound capture. A typed pre-dispatch refusal (screenshot-capture-unavailable) on touch routes when the flag is set; the flag then works only where the admission can prove a capture. Smaller, but the flag stops working wherever binding does not exist yet.
Related decisions wanted
The any-one-pixel changed rule reports incidental animation as a change (review thread r4497036508): is a threshold wanted, or is any-change the honest contract for keypads where one pixel IS the reaction?
The evidence lane itself is defect-free as of 7471601's successor commits: nothing after a resolved dispatch throws out of the lane (decode/read/import/cleanup failures yield no evidence), the after-frame settle honors the request signal, and both full-resolution frames decode in the codec worker. The corruption test the review asked for pins the delivered-tap case. Live runs (iOS simulator, Android emulator) follow once the decision lands and the flag actually reports on a shipped surface.
Status update, since the band half of this issue is now decided on #4297: the excluded region is a device fact — each capture owner declares statusBarTopHeightShare on its capture facts (apple per appleOs: phones 0.06, iPad 0.02, tvOS/visionOS/macOS none; android 0.04; harmonyos 0.04; web, macOS, and TV declare nothing and compare the whole frame) — and the lane takes the band from the before-frame's own facts at the single compare call, with the ratio computed over the compared region. Live runs at that head: mid-screen dead taps on iOS and Android both read a flat zero with the warning, and an iPad header-tab reaction reads true at 0.0003.
What stays open here is the threshold question. The any-pixel rule now reports honestly (no blanket band swallowing real reactions), and the live dead taps read exactly 0 — so the concrete case that motivated a threshold (a 0.0006 clock tick reading as changed) no longer reproduces at the head. The caller-facing question left: is changedFromBefore: true at a hair-thin ratio (my iPad run: 0.00034) actionable for a PIN caller, or should the evidence carry a magnitude disclosure beside the boolean? No code ask from me on this — I would rather hear the maintainers' read first.
Activity
fix(daemon): refuse --expect-screenshot-diff on the touch routes unti…
fix(daemon): bind the admitted screenshot capture into the touch backend
fix(daemon): refuse --expect-screenshot-diff on the touch routes unti…
fix(daemon): bind the admitted screenshot capture into the touch backend
fix(interaction): keep the plan text inside the always-loaded tool bu…
fix(daemon): refuse --expect-screenshot-diff on the touch routes unti…
fix(daemon): bind the admitted screenshot capture into the touch backend
fix(interaction): keep the plan text inside the always-loaded tool bu…
strzhao commentedon Nov 16, 2026
Status update, since the band half of this issue is now decided on #4297: the excluded region is a device fact — each capture owner declares
statusBarTopHeightShareon its capture facts (apple perappleOs: phones 0.06, iPad 0.02, tvOS/visionOS/macOS none; android 0.04; harmonyos 0.04; web, macOS, and TV declare nothing and compare the whole frame) — and the lane takes the band from the before-frame's own facts at the single compare call, with the ratio computed over the compared region. Live runs at that head: mid-screen dead taps on iOS and Android both read a flat zero with the warning, and an iPad header-tab reaction reads true at 0.0003.What stays open here is the threshold question. The any-pixel rule now reports honestly (no blanket band swallowing real reactions), and the live dead taps read exactly 0 — so the concrete case that motivated a threshold (a 0.0006 clock tick reading as changed) no longer reproduces at the head. The caller-facing question left: is
changedFromBefore: trueat a hair-thin ratio (my iPad run: 0.00034) actionable for a PIN caller, or should the evidence carry a magnitude disclosure beside the boolean? No code ask from me on this — I would rather hear the maintainers' read first.fix(daemon): refuse --expect-screenshot-diff on the touch routes unti…
fix(daemon): bind the admitted screenshot capture into the touch backend
fix(interaction): keep the plan text inside the always-loaded tool bu…