v4.0.1

libaoahid 4.0.1

No library code changed. Hardware-verified on a Samsung Galaxy Tab S11 and a
POCO F6 Pro (Windows 10 x64 and Arch Linux): the profiles and paths marked
Verified in docs/TARGET_MATRIX.md; every other row is still unverified.

Documentation

• README: a Demo section with a video of aoahid_player, an app built on this
  library.
• docs/TARGET_MATRIX.md: rows 1, 2, 3 and 6 of the accessory mode and Bulk
  Channel table are verified on hardware (Samsung Galaxy Tab S11 and POCO F6
  Pro, on Windows 10 x64 and Arch Linux).
• docs/TARGET_MATRIX.md: Keyboard, Mouse, Consumer Toggle, all three
  Gamepad D-pad forms, and fixed-MT Touchscreen took effect on the same
  hardware (Application API column).
• docs/PORTING.md: the Samsung-on-Windows driver note is a hardware
  observation, not a user report.

v4.0.0

libaoahid 4.0.0

Changed (breaking)

• aoahid_device_open sends no AOA request. It no longer sends request 51 to
  re-check the version, so it no longer returns AOAHID_ERR_NOT_AOA or a
  version error. A device without AOA 2.0 HID opens and then fails at
  aoahid_node_open with AOAHID_ERR_STALL.
• The library applies no AOA version policy. aoahid_discover still sends
  request 51 to each device and lists those with a nonzero version; the new
  aoahid_device_info.protocol_version reports that version, and the caller
  decides what to open. A caller that fills aoahid_device_info itself sends
  no request 51 at all, which also avoids Android's accessory-handshake
  broadcast (SOURCE_CONFLICTS.md T-12).
• aoahid_device_info gains protocol_version. It fits in existing padding,
  so the size and the offsets of serial and product are unchanged.
• aoahid_device_options drops every member that did nothing:
  startup_mode, accept_future_protocol_versions,
  reenumeration_timeout_ms, first_report_attempts,
  first_report_backoff_us, accessory_strings, and
  enable_deprecated_audio_mode. validate_reports moves to the end so the
  structure has no interior padding. Rebuild every caller against this header.
• Removed: aoahid_device_protocol_version(), aoahid_startup_mode,
  AOAHID_START_CURRENT_USB_MODE, AOAHID_START_ACCESSORY_MODE, and
  AOAHID_ERR_VERSION (code 5 is now unassigned; no other code changes). Use
  aoahid_accessory_start to switch a device to accessory mode.
• The shared library is libaoahid.so.4 with the AOAHID_4.0 symbol node.
• The Python, C#, and Rust bindings follow the new layouts and drop the
  removed names. The C#, Python, Rust, and C examples print each device's AOA
  version and refuse a version below 2 themselves.

Documentation

• The guide, API, design, and example documents describe probe-free open and
  caller-side version checks, and no longer describe ABI placeholders. The
  guide's first-report policy now matches the library, which has never retried
  since 2.0.0.

Verification status

Fake-backend tests only (the dev, TSan, and ASan/UBSan configurations).
Nothing is hardware-verified.

v3.0.5

libaoahid 3.0.5

Changed

• Reports are serialized a byte at a time instead of a bit at a time, about
  2.9 times faster for a 16-contact touchscreen report (766 ns to 267 ns on
  one x86-64 machine). The bytes on the wire are unchanged; the new encoder
  was checked against the old one over 2,000,000 random fields.
• Internal cleanups with no behavior change: unused parameters, a duplicated
  protocol-version check, and setup-packet fields rewritten on every submit
  although acquiring the transfer slot had already written them.

Fixed

• examples/c/verify: a failure could print "no field" and "no reason"
  instead of the failing call's diagnostic, because aoahid_result_name()
  clears the last-error record and C leaves argument evaluation order
  unspecified. The record is now read first.
• On Linux, the shared library now relinks when cmake/aoahid.map changes,
  as the Apple build already did for its export list.
• The integration tests compile with Clang 22: the replacement
  operator delete overloads in tests/integration/test_transport.cpp no
  longer take top-level const parameters.

Documentation

• README: the ADB section no longer suggests adb connect localhost:5555, a
  port the adb server scans for emulators, and describes how
  aoahid_adb_proxy works (a Channel on the application's own USB handle).
• QUICKSTART: AOAHID_USE_FAKE_LIBUSB is OFF unless a test configuration
  turns it on, and aoahid_verify_all is listed with the other verify
  programs. examples/README.md lists touchpad.c.

Verification status

Fake-backend tests only (the dev, TSan, and ASan/UBSan configurations, and a
Clang build). Nothing is hardware-verified.

v3.0.4

libaoahid 3.0.4

Changed

• The repository is renamed from nemarpuc/Libaoa_hid to
  nemarpuc/libaoahid, matching the library (libaoahid), header
  (aoahid.h), and CMake package (aoahid) names. Package metadata, release
  validation (AOAHID_EXPECTED_REPOSITORY), and documentation links use the
  new name. GitHub redirects the old URLs; the Pages URL is now
  https://nemarpuc.github.io/libaoahid/. The library is unchanged.
• 3.0.3 was tagged, but its release did not publish because the repository
  was renamed while the release job ran. 3.0.4 contains all of 3.0.3.

Verification status

No library code changed. Nothing is hardware-verified.

v3.0.2

libaoahid 3.0.2

Fixed

• On Windows, a Channel on the interface that libusb's WinUSB backend
  auto-claims for control transfers (typically ADB) could lose its claim when
  a HID report transfer in flight during aoahid_channel_open completed.
  aoahid_channel_open now waits, within the Device's close_drain_timeout_ms,
  until no report transfer is in flight before it claims the interface, and
  returns AOAHID_ERR_TIMEOUT if one still is. aoahid_channel_close waits
  the same way (best effort) before releasing it. See FACT_AUDIT.md A-25.

Verification status

Fake-backend tests only. Nothing is hardware-verified.

v3.0.1

libaoahid 3.0.1

Packaging

• Linux archives record how the bundled libusb was built from the bundled
  source (libusb_build in share/doc/libaoahid/build-metadata.json), and
  THIRD_PARTY_NOTICES.md points to it. The library itself is unchanged.

Verification status

No library code changed. Nothing is hardware-verified.

v3.0.0

libaoahid 3.0.0

Added

• aoahid_channel_options.read_mode and aoahid_channel_read_mode.
  AOAHID_CHANNEL_READ_STREAM (zero) keeps the read-ahead byte stream.
  AOAHID_CHANNEL_READ_REQUEST submits no read-ahead. Instead, a read with
  nothing buffered submits one IN transfer of its capacity, rounded up to
  wMaxPacketSize and capped at transfer_bytes, the way host adb reads a
  length-prefixed payload. A request completes as soon as it is full, so a
  packet-aligned message with no zero-length packet no longer waits for more
  data. A timeout leaves the request pending for the next read, so no byte is
  lost.
• The Python, C#, and Rust bindings expose the new field and constants.

Changed (breaking)

• aoahid_channel_options grows by one 32-bit field, so its struct_size
  changes. Rebuild every caller against this header. Zero keeps the 2.x
  behavior.

Documentation

• docs/API.md explains when a stream-mode read waits (a Bulk IN transfer
  completes only when full or on a short packet), with AOSP adb references.

Verification status

Verified with fake-libusb tests (request sizing, pending timeout, buffered
remainder, zero-length packet, unplug), plus ASan/UBSan and ThreadSanitizer.
Nothing is hardware-verified.

v2.0.1

libaoahid 2.0.1

Removed

• docs/ARCHITECTURE_USB_HUB.md, which still described a rejected automatic
  design and an unimplemented ADB-server plan. docs/API.md now cites the
  Android 10-second accessory-request timeout directly from AOSP
  UsbDeviceManager.java.
• An unused internal Channel query and two fake-libusb test helpers left over
  from that rejected design. No public function, structure, value, or behavior
  changes; the ABI is identical to 2.0.0.

Verification status

Documentation and dead-code removal only. Nothing is hardware-verified; the
accessory-mode and Channel hypotheses in docs/TARGET_MATRIX.md remain
[unverified on hardware], and every profile remains not hardware-verified.

v2.0.0

libaoahid 2.0.0

Changed (breaking)

• The library never retries a transfer. The first report after
  aoahid_node_open used to be resent automatically after a STALL (default 20
  attempts, 1 ms apart); it is now sent once and AOAHID_ERR_STALL is returned.
  A STALL means the target refused request 57 (f_accessory stalls while the
  HID ID is not yet registered), so the refused state now stays pending and the
  caller's next submit resends it; timeouts, cancellations, and I/O errors still
  consume the submitted state because delivery is unknown.
  first_report_attempts and first_report_backoff_us stay in
  aoahid_device_options for layout compatibility and are ignored.
  examples/c/verify/verify_common.h shows a bounded caller-side resend.
• SOVERSION 2 (libaoahid.so.2) with the ELF symbol-version node
  AOAHID_2.0, because the change above alters documented behavior.

Added

• aoahid_accessory_start with aoahid_accessory_options: sends AOA requests
  51, 52 (the caller's aoahid_aoa_strings; a nonempty manufacturer and model
  are required, each at most 256 bytes including its NUL as AOA 1.0 specifies,
  checked before any request), and 53 to a device in its current USB mode, then
  closes it and returns. It never waits for re-enumeration and never retries. The
  application rediscovers the accessory-mode device (VID 0x18D1, PID
  0x2D00-0x2D05) and opens it with the unchanged aoahid_device_open. A
  device already in accessory mode receives no request. There is no call that
  leaves accessory mode.
• Bulk Channels on a Device: aoahid_channel_open, aoahid_channel_read,
  aoahid_channel_write, aoahid_channel_close, and aoahid_channel_options.
  A Channel selects configuration 1 only on an unconfigured device (AOA 1.0),
  claims the first interface with the requested class triple (for example ADB
  0xFF/0x42/0x01) on the Device's own USB handle, keeps IN
  transfers submitted ahead of the reader, uses a transfer pool separate from
  HID, reads as a byte stream, back-pressures writes, and can end a write with
  an explicit zero-length packet that behaves the same on every backend. Read
  and write take no lock and allocate nothing. Channels follow the Node close
  rules, and aoahid_device_close also closes them.
• examples/c/verify/verify_accessory.c: switches the first phone to
  accessory mode, types one letter, and opens an ADB Channel when USB debugging
  is on.
• C, Python, C#, and Rust declarations for every new function and structure.

Changed

• Every USB handle is owned by one internal Port shared by the Device and its
  Channels. aoahid_discover probes a device that an open Device already holds
  through that handle instead of opening it a second time (WinUSB refuses a
  second open). aoahid_accessory_start likewise borrows an open Device's
  handle.
• docs/API.md now describes the aoahid_toggle release rule introduced in
  1.0.0 (usage must be 0 or the pressed Usage).
• udev/51-aoahid.rules comments describe the accessory-mode VID/PID range
  that aoahid_accessory_start produces; the rules themselves are unchanged.

Verification status

The new accessory start and Bulk Channels are tested only against the
deterministic fake libusb backend (including ThreadSanitizer runs and an
allocation probe of Channel read/write). None of it is hardware-verified: the
ten accessory-mode and Channel hypotheses in docs/TARGET_MATRIX.md,
including re-enumeration, HID in accessory mode, ADB next to HID on one handle
on Linux and Windows, and the first-report STALL behavior without library
retry, are recorded as [unverified on hardware]. No profile's status in
TARGET_MATRIX.md changes: every profile remains not hardware-verified.

v1.0.0

libaoahid 1.0.0

Added

• examples/c/verify/verify_toggle.c and verify_battery.c, joining
  verify_keyboard/verify_touch/verify_mouse/verify_all as
  human-observable real-device verification programs (see
  docs/QUICKSTART.md). verify_toggle fires four Consumer Control Usages
  in turn (Play/Pause, Volume Increment, Mute, "AC New") for a media app to
  react to, deliberately excluding System Control Usages that could suspend
  or power off the phone mid-test. verify_battery sends a low/mid/high
  Battery Strength sequence and the explicit unknown (Null) state; unlike
  every other program here it produces no getevent output by design, so
  confirming it means checking adb shell dumpsys battery and
  /sys/class/power_supply instead.

Fixed

• aoahid_toggle(node, usage, 0) (release) silently ignored usage
  entirely, always releasing whichever Usage was currently pressed
  regardless of what was passed. A caller that lost track of which Usage it
  had pressed could release the wrong one -- or release nothing, if nothing
  was pressed -- with no indication anything was wrong. usage on release
  must now be either 0 (release whichever Usage is pressed, for a caller
  that never tracked it -- the prior behavior, kept as an explicit opt-in)
  or the Usage actually pressed; a different nonzero Usage is rejected with
  AOAHID_ERR_PARAM instead of silently doing the wrong thing.
• examples/c/verify/verify_toggle.c did not set
  aoahid_toggle_options.expected_linux_event_types/expected_linux_codes,
  two of the four required parallel arrays (see examples/c/profiles/toggle.c
  for the reference shape), so aoahid_spec_create_toggle always failed with
  AOAHID_ERR_UNSET_FIELD and the program could never run.

Verification status

verify_toggle and verify_battery were each run once against one physical
Samsung Android target after the fixes above, confirming both now run to
completion without error; verify_toggle's Play/Pause, Volume Increment, and
Mute were also each observed changing that target's foreground media app.
This falls well short of TARGET_MATRIX.md's "Minimum per-profile cases" for
either profile (every allowed Usage under foreground/background/screen-off,
zero re-arm, and -- for Battery -- the declared minimum/midpoint/maximum plus
power_supply/application-API association), so this release changes no
profile's status in TARGET_MATRIX.md: Toggle and Battery Strength remain
not hardware-verified, same as every other profile. The usage-on-release
fix and both new example programs touch neither the wire format nor the
generated descriptor of any profile.