Standalone Inverse Service

Highlights

docked stylus is set to moved from 135 to 180 degrees. This is the single expected visible

difference from 3.5.6 in an otherwise unchanged default session — ungated, no opt-in, every user

on handle firmware 1.3.1 or above. Detailed under Fixes.

20 Hz instead of freezing as soon as an application starts driving it (Inverse3 / Minverse

firmware 7.9 or newer, on by default).

config/torque_scaling, config/gravity_compensation and config/handedness returns the value

the device powers up with instead of a factory constant compiled into the service — so a

left-handed device saved as left-handed no longer resets to right-handed.

one shared mount or body-orientation write used to emit a session-updated per holding session,

none of them naming the device; every holding session still refreshes from its next snapshot.

Internal improvements

range 0–5000) tunes the live orientation read; the 50 ms device motion timer is an effective

floor.

devices/*inverse/query_orientation_on_idle on against old firmware tells you the knob is inert

instead of staying silent.

its last good value forever; the held value itself is unchanged, so the device stays usable.

default events channel exactly as before — only the device-scope cause was withdrawn, and

handedness, torque_scaling and gravity_compensation keep their per-session announcements.

Fixes

stylus is set to on inkwell calibration moved from 135 to 180 degrees, and the value became the

setting devices/*inverse/inkwell/dock_heading_deg. Measured 2026-09-22 on a Minverse with a

stylus at handle firmware 1.3.1: the old 135 degrees left the docked pen 45 degrees short, so the

yaw estimator had to drift it back (+42 to +50 degrees observed); the residual is now within about

5 degrees. Not stage-gated — the dock-calibration path always runs — so a docked pen's

published orientation changes by 45 degrees for every user on handle firmware 1.3.1 or above.

Below that floor the firmware ignores the value and snaps the full pose, so those users see no

change.

application started streaming forces and every consumer read a stale handshake snapshot. It is now

polled on the existing motion timer and is live in-session.

compiled-in factory constant over the device's own stored configuration, so "reset to default"

and the device's real power-on state diverged — config/handedness always snapped to right

regardless of what the device had been saved as. All three now restore the stored value, falling

back to the old constant only for a field that was never read from the device.

the payload carries only session_id, profile and scope — no device identifier — so a

subscriber could neither attribute a copy nor deduplicate the set. The device-scope announcement

is withheld pending a device-addressed replacement; the per-session refresh is untouched.

Potential breaking changes

a reset means: the reply and the written value are now the device's stored configuration rather

than a fixed service-side constant. Response shapes are unchanged and the values only differ on a

device whose stored configuration differs from the factory constant — but an integration that

asserted the literal constants (false / {enabled: true, scaling_factor: 0.0} / right) will

see different values.

was effectively static for a streaming client now moves at 20 Hz. A consumer that treated it as

constant-after-attach sees new behaviour with no opt-in.

config.body_orientation when the cause is a device-scope write, and the deprecated flat

body_orientation scope follows the same lane rather than leaking back to the default channel.

Migration: stop treating the event as the refresh trigger and read the value from the session

snapshot — it still refreshes on every holding session with no opt-in at all. The equivalent

session-scope events, and the handedness / torque_scaling / gravity_compensation

announcements, are unaffected.

bug fix and is published under Highlights and Fixes. Recorded here because the observable value

moves with no opt-in for every user on handle firmware 1.3.1 or above, so anyone who calibrated

their own application against the old 135-degree datum has real migration work. No shape changes.

Deprecations

whenever the new poll interval is non-zero, but it is deliberately not flagged deprecated in

3.5.7 — that deprecation is scheduled for 3.6.

Highlights

just the one that wrote. A second viewer refreshes on its own session-updated event instead of

polling — exactly one event per holding session, and never two for the writer.

always reporting the mount the session configured. The label that tells you which of the two you

are looking at is now truthful by construction, in every output shape.

service served a value captured when the device came up and never revisited it. It can now

refresh that reference on demand at the client's own render cadence — rate-limited service-side

so a fast client cannot flood the device link — and announce the change once the mount has

actually rotated.

Internal improvements

Wireless VerseGrip's inertial yaw drift: it fits the grip's absolute yaw against the motion of

the paired Inverse3 or Minverse it is attached to, and once a stack of quality gates all hold, it

writes a single correction back to the grip firmware. It ships disabled, is reachable only

with the experimental opt-in switch on, and carries a runtime switch that forces every live grip

back off the moment it is flipped — non-optional for something that writes to device firmware.

selection is validated — firmware capability and device readiness — before any device is changed,

so a mixed-firmware fleet can no longer end up half-switched behind an error on the second device.

two concurrent requesters could lose each other's answer, and one ambiguous value stood for both

"refused" and "nothing yet" — which could turn a refusal into a retry storm. Each request now

carries its own token, a request that never reached the wire registers nothing, and a lost

response resolves as failed instead of wedging. The sweep that does this costs nothing while no

request is outstanding.

pen performs a full pose recalibration, which is correct on every shipped pen regardless of

firmware version; that path is explicitly exempt from the capability floor introduced for

computed heading corrections, and the exemption is spelled out at the call site rather than

inferred.

Fixes

Despite its name, the scaling_factor field on config/gravity_compensation is not a

multiplier: it is the extra end-effector payload mass, in kilograms, that the device compensates

for on top of its own weight. The reset took the name at face value and restored the "neutral"

factor 1 — a phantom one-kilogram payload — so a reset left the device pushing the handle up,

which is not what a freshly connected device does. The reset now returns the real default:

compensation enabled with scaling_factor at 0, meaning the device carries its own weight and

assumes no payload. If your application had been using the reset to raise the compensation

strength, set the value you want explicitly instead.

come to disable home return while telling the caller it was enabled — a reset now genuinely

restores the factory default (on), and the value your application reads back afterwards is the

value the device is actually using.

a write that landed on the device itself — shared by every session holding that device — refreshed

only the writing client's configuration, and the device-wide routes refreshed nobody at all. Any

other client on the same device kept reporting the old value until it polled. Every holding

session is now notified.

previous dock persisted across docks, so the second dock treated its first rejection as the end

of its retry window. The retry window is now reset per dock attempt.

reassigned by the connection thread while a staging path read it under a different lock. The

identifier is now snapshotted once under the correct lock, and every path below — refusals

included — works from that copy.

message paths assembled their payload and checksum and requested the send without taking the

write lock, so two senders could interleave on the shared buffer.

Potential breaking changes

was deleted or re-pathed, the public C API has no diff, and no settings key was renamed, removed,

or given a new default. Applications that opt in to the SDK's preview or experimental surfaces

do have migration work this cycle, including a switch-tier change and a removed token set;

every step is documented in the engineering companion changelog.

route or response shape changed, but an application that relied on the reset writing a

scaling_factor of 1 now gets the factory default of 0 — compensation itself stays enabled

either way. Set the value explicitly rather than depending on the reset; see the fix above.

changed, but an application that relied on the reset to disable home return now finds it

enabled instead. Set it explicitly to off after a reset if that is what you need; see the fix

above.

session-updated event only ever arrived because of its own write will now also receive one when

another client changes a device they both hold, followed by a full configuration render. This is

the intended consequence of the fan-out above; a client that simply re-reads on the event needs no

change.

Highlights

stages before a stable release: internal alpha builds, then published beta builds that soak

before anything reaches stable. Every stable release is now preceded by a validated beta.

X.Y.Z-develop.N are now X.Y.Z-alpha.N, followed by X.Y.Z-beta.N on the beta channel.

The old naming sorted "develop" builds above "beta" builds under semantic-versioning

precedence rules; the rename fixes that ordering permanently.

existing history.

Internal improvements

deliberate, manually-triggered process gated on passing tests, and the whole new release

pipeline was rehearsed end-to-end before it shipped.

by roughly three quarters, while the full multi-platform test matrix still gates every

promotion toward a release — coverage before a release is unchanged.

1.0.228 -> 1.0.229 (lockfile-only patch bump), the release tooling's semantic-release

25.0.8 -> 25.0.9, and the Alpine container base image 3.20 -> 3.24.

Fixes

code changed. The auto-generated release notes list a handful of fixes to the new release

pipeline itself; these hardened brand-new infrastructure before it shipped and have no effect

on deployed SDKs.

3.5.4 (2026-08-26)

Internal Improvements

Bug Fixes

3.5.3 (2026-08-06)

Internal Improvements

Bug Fixes

Performance Improvements

3.5.2 (2026-06-19)

Internal Improvements

Bug Fixes

Performance Improvements

3.5.1 (2026-05-28)

Internal Improvements

Bug Fixes

Performance Improvements

Our largest release since the API rewrite a year ago. 3.5.0 introduces a unified 3D math layer, full Minverse support, a typed event stream, a complete settings API, and a cleaner HTTP surface — alongside a year of bug fixes and quality-of-life improvements.

Highlights

New features

Devices

Sessions, commands & WebSocket

HTTP

Events

Observability

Coming in 3.6

Improvements

Fixes

A year of bug fixes; the most impactful are below.

Deprecations

Still functional, but will be removed in a future major version. Responses include a deprecation_warning field where applicable.

| Deprecated | Replacement |

|---|---|

| POST /force_scale | POST /settings/devices/inverse*/filters/force_scale/gain |

| POST /gravity_compensation | POST /{device}/{id}/config/gravity_compensation |

| POST /torque_scaling | POST /{device}/{id}/config/torque_scaling |

| POST /device_handedness | POST /{device}/{id}/config/handedness |

| POST /serial_enable | POST /settings/communication/serial/enabled |

| POST /experimental/features/grip_dropped_simulation_stopper | Settings API |

| POST /experimental/features/screensaver_enable | Settings API |

| WebSocket command_data.values | position, vector, angles, torques |

| WebSocket set_angular_torque | set_angular_torques |

| WebSocket probe_cursor_position | probe_position |

| WebSocket set_coordinate_origin | configure.preset |

| WebSocket session-level set_basis | configure.basis |

Compatibility notes

No hard breaking changes against 3.4.19. Two soft notes for existing integrations:

Tutorials & documentation