Loading...

The Service Runtime and Runtime SDK for Haply devices. Bundled in a convenient installer. Also known as the "external" Service.
3.5.8 is a patch release focused on hot-plug and reconnect and on mount presets. The public
C API is unchanged from 3.5.7 and no route was removed. Every change that reaches a client
running on defaults is a fix for a bug that shipped in 3.5.7, plus clearer diagnostics. The new work
of this cycle sits behind the SDK's opt-in stage switches, which still ship off, and is invisible to a
default client.
Unlike 3.5.7, this release does change what a default client observes, in eight places. Five are
broad: a device is omitted from a session until that session has registered it,
state.current_cursor_position moves to application space, the force gate becomes per session, mount
writes require a unit rotation, and the centered presets land on the device's current rotation and can
report custom or defaults. Three are narrower: on a Minverse, arm_front and defaults reset the
session mount and state.body_orientation values change, and device-connection-loop can fire for a
receiver that names no device. Each is listed under Potential breaking changes with who notices and
how to migrate.
An engineering companion changelog (__CODE_code-1__)
carries the stage-gated API detail, the internal mechanisms behind these fixes, design-record pointers
and issue references.
anything. It joins open sessions the moment it is ready, keeps the session's configuration across
an unplug, and arrives in a full snapshot. The same reconnect story, per device:
calibration, the dock heading sent to the paired stylus and the calibration events resume, and the
device-safety-triggered event is delivered again. In 3.5.7 they were lost until the service
restarted.
1.5 s forever: retries back off up to 30 s apart, clients get one device-connection-loop event
per episode, and a replug brings the grip back at once.
configures no longer changes the driver's value, and the default returns when the driver lets go.
boot. The preset is resolved for the device as it sits when you apply it; rotating it afterwards
turns the workspace with it, and re-applying re-centers.
over HTTP and WebSocket alike.
state.body_orientation and the device mount reference are no longer rotated aquarter-turn about Y: the device reads identity on its calibration pose again.
state.current_cursor_position is now reported in the session's application space, likecursor_position beside it, so a commanded position echoes back in the coordinates it was sent in.
detection pass (up to about 0.5 s).
device not found warning, a reminder every10 s with the dropped-command count, and one device found again line — instead of one line per
command (16 939 lines in 2.5 s on the bench).
the device rotation, so a centered preset's state stays accurate.
POST /ports/{port}/reset also cancels a silent receiver's backoff, so the next reopen attempt isimmediate.
its code, a lost response, or a request replaced by a newer one), with the stylus motion against its
gate; giving up is logged once. Log only: retries, gates, timings, events and wire fields are
unchanged.
device, and warned about only if it is still missing after 10 s (rate-limited, with a count).
Hot-plug and reconnect
with raw hardware values and no session configuration, for up to about half a second after a
hot-plug or wake. It is now omitted until registered, then arrives in a full snapshot (v3.0 and
v3.1, WebSocket and GET /devices?session=).
mid-update. The list a reply reads is now a consistent snapshot.
session_id targeting) only received that session'sconfiguration after a manual full-state request, both after the target changed and after
switching target. The next reply is now a full one in both cases.
old controller and no full snapshot announced the return. This was a documented known limitation in
3.5.7.
start overwrote it, and a device registered after the session started kept the compile-time floor
instead of the default.
calibration results, the dock heading to the paired stylus and the calibration events stopped after
the first drop.
every 1.5 s with no end (1,562 times in 53 minutes in one field log), flooding the log while the
grip simply vanished for clients. Reopens now back off to 30 s and the episode is announced once.
although no reset exists there; the log now says what the reset really did (only Linux resets the
port, and a failed reset is reported as such). Log only.
unregistered device (2,178 Unable to serialize lines in about one second on the bench, 11 in 3 ms
on one receiver replug).
Mount and presets
target into inf/NaN, while HTTP refused valid scales outside [1, 5.5]. Both transports now apply
the same rule: any finite scale that is not near zero on any axis.
center; a session DELETE config/mount left config.preset naming the old preset, so re-applying
it brought the deleted mount back; and on a Minverse, arm_front / defaults kept the current
mount instead of resetting it.
already re-zeroes the orientation on its calibration pose, so its state.body_orientation and its
stored mount reference read a quarter-turn about Y off (a left-handed unit reported its identity
reference as a +90 degree rotation about Y).
Device configuration
reflected in the reported config. Toggling gravity compensation or moving the tool-weight slider
on the Inverse3x screen now updates config.gravity_compensation (v3.0 and v3.1) and
GET /{device_type}/{id}/config/gravity_compensation within one serial round trip, and each session
holding the device gets one full reply on its next message. In 3.5.7 the device's change notice
was ignored and the old value stayed until the device reconnected. Earlier firmware does not send
the notice, so nothing changes there.
Wire values
cursor_position,cursor_velocity and current_cursor_force in the same object were in application space. With a
non-identity basis the echo came back permuted (XZ-Y: sent (x, 0.090, 0.197), echoed
(x, -0.197, 0.090)). It is now converted with the session basis and mount in v3.1 full and
streaming frames. Present since 3.5.0.
34) whilesession.session_id in the same frame carried the real id. It is now the full 32-bit id, matching
the documented uint32 schema. Present since 3.5.0.
and v3.1 arrays, full and streaming, WebSocket and GET /devices?session=). In 3.5.7 it appeared
bare during the hot-plug window. The JSON shape is unchanged, and the default output keeps
custom_verse_grip: [] when every custom grip is omitted. Who notices: clients that index devices
by array position across frames, treat a device missing from one frame as unplugged, or read the
bare values of that first frame. Migration: key devices by device_id; treat the first frame that
contains a device as its arrival; detect unplug from the events channel, not array membership.
state.current_cursor_position changes value for every v3.1 session with a non-identity basis,mount or preset. Field and type are unchanged; sessions with an identity basis and mount see the
same numbers. Who notices: clients that compare or plot the field, and any client that converted
it from device to application space itself. Migration: drop that client-side conversion; the field
now matches cursor_position and the commanded target.
device, and it returns to the default on release, close or lock expiry.
GET/POST/DELETE /{device_type}/{id}/config/filters/force_gate and
config.filters.force_gate.gain report the session's own value (default 0.3), not the device's
live gain. Who notices: a client that set the gate from a watching or configuring session, or read
it back expecting the device's live value. Migration: set the force gate from the session that
drives; read the driving session to see the effective value.
POST|PATCH config/mount answers400 invalid mount transform (non-finite, non-unit rotation or ~zero scale axis) for a quaternion
whose squared norm is off by more than 1e-3, and WebSocket configure.mount logs a warning and keeps
the previous mount; a scale outside [1, 5.5] that 3.5.7 refused is now accepted. Who notices: a
client that sent unnormalised quaternions, or relied on the service to cap the scale. Migration:
normalise quaternions before sending; enforce any preferred scale range in the client UI.
arm_front_centered / led_front_centered resolvetheir position on the current device rotation when applied; config.preset reads custom (with a
session-updated for config.preset) once the device has turned more than 1 degree since the apply
— including after a replug or an inkwell re-calibration, with no client write; and it reads
defaults after a session-scope DELETE config/mount. Who notices: a client that keys behaviour
on the config.preset name. Migration: re-apply the preset to re-center after moving the device;
treat custom as "the preset no longer describes the pose".
arm_front_centered position moves from x = -0.13 to x = -0.12 (hand-tuned on hardware). *Who
notices:* Minverse users who relied on those presets keeping a custom mount, or calibrated against
the old center. Migration: re-check Minverse calibrations made against the old center.
case device_id is the last id seen ready on the port and is empty if there never was one,
attempts counts failed attempts in the episode and window_ms is the time since its first
failure. The payload shape is unchanged. Who notices: a client that assumed device_id is never
empty. Migration: handle an empty device_id.
notices: a client that compensated the quarter-turn itself. Migration:* remove that compensation,
or it will double-correct.
All available downloads for this release: