v0.10.0

Added

• GitHub Actions CI (.github/workflows/tests.yml): Swift unit tests on macOS hosted runners (idb-derived FB XCFrameworks cached between runs), bridge Kotlin JVM unit tests on ubuntu, and a bridge protocol parity check — all for every push and pull request targeting main. (#26)
• make build / make test condense swift output via xcsift (TOON summary; test coverage report) when it is installed — strictly optional, plain swift output otherwise; SIM_USE_XCSIFT=0 forces plain output. (#26)
• swipe now accepts --from x,y --to x,y and positional x,y x,y coordinates on top-level, iOS, Android, and iOS batch surfaces while keeping the existing four coordinate flags. (#27, #30 — thanks @joonyeonglim!)
• tap and long-press now accept --point x,y as a coordinate-pair alternative to -x/-y on top-level, iOS, Android, and iOS batch surfaces (#25). The two forms are mutually exclusive and resolve through one shared TapCoordinateResolver, and describe-ui --point now parses through the same CoordinatePair grammar as swipe --from/--to (a malformed value fails with ArgumentParser's standard invalid-value diagnostic instead of the previous bespoke message). (#40)
• tap/long-press — and ios batch tap steps — now surface a structured advisory when a label/value selector resolves to a near-full-screen element (measured against the Application root frame), so daemon-routed calls show the warning in terminal output and --json instead of burying it in the daemon log. (#29, #36 — thanks @joonyeonglim!)

Changed

• describe-ui --point coordinates are now interpreted in the same UI space as the printed outline frames. On a rotated simulator the query is transformed onto the framebuffer before the hit-test (previously the raw point was hit-tested in framebuffer space and returned the wrong element); an upright device behaves exactly as before. (#38)
• describe-ui surfaces the calibrated interface orientation: the App: header gains a suffix tag (e.g. (landscape-right)) when the device is not upright, --json data gains an orientation field, and the alias snapshot records the orientation it was captured under. (#38)
• Swipe coordinate flags now live in a shared SwipeCoordinateOptions group, so the top-level, iOS, and Android surfaces validate identically; the swipe success line and --json data payload derive the coordinates from the execution result (data now includes a coordinates object). (#30)
• swipe --duration is capped at 10 seconds on every surface (parity with tap / multi-touch / gesture presets). The error message spells out that durations are in seconds, so a millisecond value passed by habit (0.5.x android swipe, adb shell input swipe) fails loudly instead of producing a multi-minute swipe. (#30)
• JSON output no longer emits the legacy udid key (dual-emitted since the deviceId transition); deviceId is the canonical key in devices --json, daemon stop/status --json, and Viewer API responses. Inputs (daemon wire decode, Viewer API requests) still accept udid as a deprecated alias, to be removed in a future release. (#22)
• --label/--value exact matching and --label-contains (plus the Android-only --value-contains) now fall back to whitespace-collapsed comparison when the exact pass finds nothing, on both iOS and Android, so a multi-line label (which the compact describe-ui outline renders space-joined) matches the space-joined string an agent copies back. Existing exact matches are unaffected — the fallback only runs when the exact pass matched zero elements. The Android exact pass now end-trims both query and label, matching iOS. The round-trip covers labels the outline renders untruncated (≤ 60 graphemes) and unescaped; longer labels still need the @N outline alias or describe-ui --json (raw labels). (#28, #31 — thanks @joonyeonglim!)
• Whitespace collapsing is one canonical implementation (SelectorTextMatcher in SimUseCore) shared by the iOS/Android outline renderers and both selector resolvers, so the display form and the matching form can never drift apart. The outline now also folds Unicode whitespace the old collapse missed (NBSP, U+2028/U+2029 line separators), so element lines cannot wrap on exotic line breaks. (#31)
• --wait-timeout polling now also retries while the selector matches multiple elements (previously only not-found), so a transient ambiguity during a screen transition no longer aborts the wait on the first tick; a stable ambiguity still reports multipleMatches with its disambiguation hint once the window expires. (#31)
• Daemon client now retries a command once against the same daemon when the simulator reports the post-boot transient_booting readiness gap, matching the long-documented behaviour. (#6)
• Bridge /swipe now accepts durations up to 10 s (previously silently clamped to 5 s), covering the full --duration range the CLI validates for long-press holds. Bridge versionCode bumped to 16. (#8)
• ios type builds one HID session for the whole string instead of re-initialising FBSimulatorControl per character. (#9)
• ios stream-video --format help now marks bgra as experimental (no frame count is reported for that format). (#21)

Fixed

• sim-use devices no longer hangs forever on machines with enough simulators installed to push simctl list devices -j past the ~64 KB kernel pipe buffer. SimctlDeviceLister.runSimctl now drains stdout/stderr concurrently with the child (the same readabilityHandler drain Adb.run uses on the Android path) instead of reading only after waitUntilExit(), which deadlocked once the child blocked on write(2). (#39 — thanks @JustHm!)
• The root sim-use --help abstract no longer claims the tool is iOS-Simulator-only; it now mentions Android emulators/devices as well. (#35)
• iOS taps resolved through accessibility (tap @N / #N / #<id> / the --label family, batch tap steps, paste --via-menu targets and edit-menu items) now land correctly when the simulator or the app is rotated (#34). AX frames are reported in the app's UI space while HID events are interpreted in the device-native portrait framebuffer; sim-use now self-calibrates the orientation per command with 1–3 accessibility hit-test probes and transforms AX-derived coordinates before dispatch. Explicit -x/-y coordinates keep their raw framebuffer semantics. When calibration cannot be confirmed (empty or fully symmetric screens) the command falls back to portrait and surfaces an orientation_calibration_fallback advisory. (#38)
• describe-ui quadtree recovery no longer silently drops whole regions on rotated simulators — the same #34 coordinate mismatch corrupted its coverage bookkeeping (live repro: the entire Settings sidebar vanished from the outline on an upside-down iPad). (#38)
• describe-ui --point no longer overwrites the @N alias snapshot with a single-element table, so tap @N keeps resolving against the last full outline after a point query. (#38)
• android swipe invoked directly now enforces the same coordinate rules as the other surfaces: negative coordinates and identical start/end points are rejected at validate time instead of being forwarded to the bridge. (#30)
• Swipe coordinates are validated as finite and ≤ 100000 on all surfaces, so values like inf, nan, or 1e19 fail with a clean validation error instead of trapping the daemon in the Double→Int conversion. (#30)
• Tap and long-press coordinates get the same finite / ≤ 100000 validation on every surface — tap -x inf -y 5 or -x 1e19 targeting an Android device previously trapped the Double→Int conversion instead of failing with a clean validation error. (#40)
• The top-level swipe and android swipe no longer disagree on fractional Android coordinates (truncation vs rounding); both round half away from zero via shared accessors. (#30)
• android swipe --pre-delay/--post-delay are bounded to 0–10 seconds like every other surface; the previous sign-only check let inf/nan through into the Double→UInt64 sleep conversion, which trapped the daemon. (#30)
• Android swipe rejects coordinate pairs that round to the same integer pixel (e.g. --from 10.4,10.4 --to 10.49,10.49), which previously passed the Double comparison but dispatched a degenerate same-point gesture to the bridge. (#30)
• DaemonClient.stopDaemon no longer waits on and SIGTERMs a pidfile pid that is the caller's own process (a stale pidfile can hold a recycled pid; in-process daemons in tests always do). Signalling ourselves fanned out through every live DaemonServer's SIGTERM source and tore down unrelated daemons mid-request — the main source of daemon-test flakiness under parallel load. (#26)
• Cached HID connection is now validated against the simulator's boot instance before reuse, so a simulator shut down and re-booted under the same UDID gets a fresh connection instead of hanging the daemon (or failing every command) on the dead one until restart. Additionally, any failed HID perform drops the cached connection, and failures that provably happened before delivery (dead mach port) are transparently retried once against a rebuilt session. (#23)
• ios batch --ax-cache was a complete no-op: the default perBatch never cached and every selector-based step refetched the AX tree. perBatch now resolves all steps against one snapshot, perStep refetches at each step, none never caches, and --wait-timeout poll ticks bypass the cache (updating it) so delayed elements are still found. (#20)
• Daemon client no longer tears down a healthy daemon and re-executes the command when the daemon answers with a command-level error (element not found, etc.). Failed commands now surface immediately instead of paying a full daemon respawn, and side-effecting commands are no longer executed twice. (#6)
• Daemon shutdown no longer deletes the socket/pidfile of a successor daemon that took the paths over, which previously chained invisible orphan daemons. (#7)
• Daemon base directory under /tmp is now validated on every run (symlinks, foreign owners rejected; loose permissions tightened to 0700) instead of trusting whatever was pre-created there. (#7)
• Bridge /a11y_tree_full no longer reads the active root's windowId after the node was recycled, which silently dropped popup/dialog secondary windows on Android 11–12. (#8)
• Bridge /keyboard/input no longer leaks the borrowed root AccessibilityNodeInfo on every call. (#8)
• record-video no longer hot-spins without frame pacing when screenshot frames persistently fail to decode. (#9)
• record-video no longer hangs forever when AVAssetWriter stops accepting frames; a stalled writer now fails the recording with an explicit error after 10 s. (#19)
• Daemon no longer shuts itself down in the middle of a request that runs longer than the idle timeout (e.g. tap --wait-timeout beyond the timeout, or a long batch). The idle timer now defers shutdown while a request is in flight. (#13)
• Daemon client no longer respawns and resends a command when the daemon drops the connection after receiving the request (a possible mid-execution crash). Such ambiguous outcomes now surface a dedicated error with a hint to re-observe the screen before retrying, so a side-effecting verb (tap/type/swipe) is never silently applied twice. Pre-delivery failures (connect/write) still respawn as before. (#12)
• keyboard-state now routes through the per-UDID daemon like every other verb (amortised init) and surfaces crash advisories and error Hint: lines; a vestigial run() override had silently opted it out. The soft/hidden/failure exit codes are unchanged. (#11)
• ios stream-video with the BGRA pixel format no longer exits 0 when the underlying stream fails to start or dies mid-stream; startup and mid-stream errors now terminate the streaming loop and surface as a non-zero exit instead of a stderr-only message. (#21)
• record-video's stop watchdog no longer exits 0 when video finalization overruns its grace window, which could report success for a truncated/unplayable MP4. It now warns on stderr and exits 70 (EX_SOFTWARE), and the grace window is 3 s (was 1.5 s). (#18)
• Viewer API no longer reports success when the underlying sim-use invocation exits non-zero without a parseable JSON envelope; the subprocess's stderr is now surfaced in the error response instead of a generic JSON-parse failure. (#17)
• sim-use android tap no longer crashes on every invocation with ArgumentParser's "can't read a value from a parsable argument definition" fatal. performTap's default MultiTouchOptions() was a directly-initialized ParsableArguments value, and the first .fingers read trapped before anything was dispatched; the parameter is now optional with nil meaning single-touch. The top-level sim-use tap routed to Android was unaffected (it passes its parsed options through). (#41)
• The bundled skill's preflight.py daemon-restart autofix no longer appends the session --device flag to daemon stop --all (the two are mutually exclusive), so the autofix actually stops the stale daemon instead of failing with exit 64 while reporting success. (#2 — thanks @viseator!)

Removed

• Dead .hidSwipePerformed notification posted by ios swipe — nothing ever observed it since its introduction. (#14)