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
Commit 3a64ee7
Browse filesBrowse the repository at this point in the historyBrowse files
fix(daemon): surface proven no-effect gestures to the agent (#1601)
* fix(daemon): surface proven no-effect gestures to the agent (#1600)
The #1542 post-gesture stabilization loop already PROVES when a scroll,
swipe, or pan moved nothing: its accept-stale verdict fires only when the
quiet post-gesture capture still equals the pre-gesture baseline after the
distrust cap. But the verdict went to the diagnostics stream only — the
agent-facing response reported plain success, and the benchmark showed the
cost: element-18 burned ~40 tool calls re-issuing scrolls the daemon knew
did nothing ("Scrolled up by 1200px" fifteen times over a byte-identical
tree) before stumbling onto raw swipe.
The stabilization loop now returns the verdict alongside the capture, and
the snapshot handler appends a warning through the existing annotations
channel: it names the exact gesture, admits the at-edge ambiguity the
platform cannot resolve, and hands over the raw-drag escape hatch that
moved the stuck list in the field.
Live-verified on the seeded Bluesky feed: scroll up at top warns, scroll
down with real room moves silently, a fling at the bottom edge warns —
all three truthful.
* fix: require full-surface evidence before the no-effect claim (#1601 review P1)
accept-stale alone is subset-tolerant by design: a successful scroll that
replaced every list cell under fixed chrome still classifies 'unchanged' on
the shared chrome alone, and #1573 has live evidence of that shape. The
agent-facing claim now additionally requires every discriminating entry of
the quiet capture to match the baseline exactly, in both directions
(haveIdenticalDiscriminatingSurfaces): any appeared or vanished real
element — including scope drift — vetoes the warning. Silence is the safe
failure mode for a message that steers the agent's next move.
Red evidence: the new fixed-chrome + replaced-list regression test fails on
the previous PR commit (gestureNoEffect wrongly present), passes with the
gate.
* refactor: extract the corroborated-accept result builder
capturePostGestureStabilizedResult crossed the complexity threshold (14
cyclomatic) after the #1601 review gate; the accept/veto construction now
lives in buildAcceptedStabilizedResult, which also keeps the corroboration
rule stated in one place.
0 commit comments