Design baseline: Blender_XR_Workbench_Specification.docx, version 1.0,
September 21, 2026. Implementation target: Blender 5.1.2, OpenXR, Windows,
Valve Index. This supersedes the original 4.5.11 target for the new native
track. The reference document is preserved.
XR-0001 (M0) is in progress. The native source, diagnostics, and presentation tasks are integrated separately so a working desktop diagnostic cannot be misreported as verified headset behavior.
With SteamVR closed and the Index temporarily in Windows extended-display
mode, the tester saw a stereo X/box pattern through the lenses. They then
launched the guarded 15-second Index camera preview and confirmed blue hand
outlines and the in-lens countdown. The latest hands-in-view text log records
230 stereo frames, 723 contour observations, stream_failed=0, and a normal
timer exit; no matching Windows application crash event was found. This is a
raw, uncalibrated stationary diagnostic at about 15 fps, not safe walk-around
passthrough or Blender interaction.
An isolated, source-built Windows Monado runtime now passes a bounded physical
xrCreateInstance/xrGetSystem probe on the Index with its configuration
parsed. It reports orientation support and no positional tracking, and does
not alter the registered OpenXR runtime. A subsequent camera-free physical
color test created an OpenXR session and submitted 694 stereo frames over
4.97 seconds. A hands-on test confirmed seeing the lens colors; the corner markers
were too high to inspect comfortably. The frame-test log is
native/steamvr_free/physical-probe-logs/frame-probe-20260923-160350-54240.log.
An instrumented follow-up submitted 684 frames and found orientation-valid and
position-valid flags on all 684 xrLocateViews results, while system capability
still reports no positional tracking. The Vive 3DoF path supplies a fixed
position and marks its relation valid; these flags allow Blender's current
view guard but do not prove 6DoF motion. Its log is
native/steamvr_free/physical-probe-logs/frame-probe-20260923-161105-40228.log.
The user then launched source-built Blender 5.1.2 via the isolated runtime and
confirmed seeing the scene and native task card in the Index, stable while
slowly turning their head. The run exited normally with code 0; status log:
reports/index-physical-launch/launch-20260923-161715-55fcf102.txt.
Its first attempt had failed during add-on registration before XR; the
callback-signature fix passed an isolated real-Blender smoke before the
successful retry. Camera passthrough and bare-hand Blender interaction remain
unverified. No Lighthouse base station is available for room-scale 6DoF
tracking.
The user additionally requires a SteamVR-free implementation. That is a new runtime/display-driver milestone, not a launcher flag or a completed Blender feature. The current frozen executable still uses SteamVR's OpenXR runtime. Its metadata-only headset probe reported zero connected Lighthouse references and no valid HMD pose; a hands-on test confirmed that no base station is available. Accordingly, no present software build can be marked as physically verified 6DoF Index tracking on this setup. A process-scoped Monado simulator has opened an isolated Blender XR session without SteamVR, but the Windows build has no Index display driver or independent physical tracking. It does not change the system's registered OpenXR runtime.
The repaired hands-ready Mercury camera retest processed 30 fresh stereo pairs without the former OpenCV crash, but produced zero active hands and zero valid HMD poses. This is a failed optical qualification gate, not proof that visible hands were detected or safely registered to the Blender scene.
The SteamVR-free DirectShow path read five advancing Index stereo frames, then an isolated Mercury diagnostic processed 14 non-flat, distinct pairs in a consented 15-second run. The detector produced zero scores above 0.3 and zero hand regions/keypoints. A source-level row-orientation mismatch was corrected and passed an offline mapping test; a second consented Release run still processed 14 pairs with zero scores above 0.3. A third consented Release run processed 13 distinct pairs, measured center-eye grayscale means of 61.56/67.93, and again recorded zero detector scores above 0.3. Frame acquisition averaged 0.94 seconds after the first frame, while Mercury processing averaged 2.68 ms per pair. No camera images or joint traces were saved. This is independent camera access, not a qualified optical hand provider or headset display.
A one-click stereo desktop preview now offers privacy-local, current-frame blue estimated hand contours from palm-plus-landmark agreement, with a labeled Mercury region fallback. The Release offline gate/mapping tests and a read-only smoke on the user's supplied screenshot passed with one estimated contour per camera half. The user deferred the new live preview test, so actual camera cadence, physical contour placement, and headset presentation remain unverified.
Separately, a simulator-only OpenXR client submitted three synthetic stereo projection frames through Monado's desktop compositor and verified that the system OpenXR registration did not change. A screenshot-only, in-memory bridge then submitted both annotated camera views through the same client: one blue estimated contour per eye, matching GPU texture readback, and three accepted stereo layers. These prove software texture upload and layer submission, not live passthrough, compositor output pixels, or the physical Index display.
A separate full-feature Blender 5.1.2 source build passes factory-startup CPU
Cycles rendering, USD export, release configuration and native XR API smoke.
Its SteamVR-free Monado simulator GUI test passed Select/Move/Rotate/Scale
button routing, Add Cube, Undo and Redo, Exit, and scene restoration. The fifth
primary modeling slot (More) remains disabled, and physical optical input has
not been qualified. It has not been physically presented in the Index; see
docs/full-feature-build-status.md for its exact hash and remaining features.
The separate 0.3.0 manual-install extension ZIP was rebuilt from current source, passed byte-for-byte source-parity checks, and passed Blender 5.1.2's factory-startup extension manifest validation. It contains no native runtime, camera provider, or optical model and is not a SteamVR-free headset release.
- Compiled and installed the actual Blender 5.1.2 fork; exact executable,
source and patch identity are recorded in
native/runtime-manifest.json. - Native resource-lifetime C++ regression passed against a mock OpenXR runtime. This covers creation failures and reverse-order teardown, not a real camera.
- 20 Python regression tests passed, including the original 4.5.11 tests.
- Ruff checks/format passed. The source-parity-audited 0.2.0 ZIP passed Blender's extension manifest validation and a freshly extracted install's registration, opacity, recall-matrix, geometry-cache and teardown smoke.
- The default native launcher passed its PowerShell 5.1 smoke; the actual GUI bootstrap passed with XR and camera auto-start both disabled.
- Native OpenGL desktop drawing, save/reload, no scene/image mutation, and handler cleanup passed. Stock OpenGL/Vulkan desktop drawing also passed.
- Connected Index: native OpenGL and stock Vulkan entered running sessions, invoked the XR draw callback twice, and stopped cleanly with camera off. SteamVR native states progressed 1/2/3, then 6/1/8/0 on stop; no visible or focused state was observed. Continuous frames and post-Recall visibility are explicitly unverified, not passed.
- The extended live harness first failed its continuous-draw assertion
because the connected headset remained inactive. It now records Recall
as UNOBSERVED in that case;
XRWB_REQUIRE_CONTINUOUS_DRAW=1retains the strict assertion for a worn/awake headset. Startup/cleanup gates are separate. - Following explicit user consent, Rectus v0.4.2-beta was installed. Its
current-instance loader evidence and native
fb_passthroughactive marker passed after enabling the Index cameras and restarting SteamVR. Provider initialization and shutdown passed; sustained visible frames, calibration, freshness, and timing remain unverified. The layer is globally disabled between controlled tests. No camera pixels were saved. - M2's actual source-built internal full-window editor passed a factory-scene
test: native event dispatch changed a transform, undo restored it, a search
popup composed into the editor texture, and capture/session teardown passed.
Evidence:
.tmp_xr_probe_096834ca235c4e99874eb9d5bac1bdfd. The captured editor texture was inspected visually. This is not headset visibility proof. - 68 pure Python tests currently pass. Further transaction/hand integration is in progress and is not yet part of a qualified headset release.
- The isolated stock 5.1.2 transaction GUI smoke passed object and selected-mesh exact cancel and real undo/redo, parented/unapplied transforms, snapped move/rotate/scale, stale/tracking-generation rejection, and live topology invalidation. Input was explicitly simulator provenance, not camera input.
- M2's first 100-cycle lifecycle attempt crashed in
vrclient_x64.dllafter 79 completed stop cycles. The crash alone does not identify responsibility. Source review found graphics bindings destroyed after their GL context and disabled interop unregister/close operations. The combined hand/interop candidate fixes that order and now checks failed GPU transfers. Its extracted native destructor regression and 10 guarded real camera-off restart cycles passed. Process commit was 1,415,655,424 bytes after cycle 1 and 1,392,922,624 after cycle 10; no interop cleanup errors were logged. The 100-cycle retest then completed 100 starts and 100 stops without a crash or cleanup error (.tmp_xr_probe_9b192180ad924839bc3471c60fb16ee6). Commit grew from 1,423,609,856 to 1,695,662,080 bytes, with a maximum of 1,702,625,280; cycles 20–100 remained approximately level. This is not a zero-leak proof, and the long editing soak/visible-frame gates remain unverified.
The user explicitly replaced keyboard/controller-first interaction with bare hands, no controllers, no virtual keyboard, using only the Index's built-in cameras. This overrides keyboard-specific document acceptance details for the user-facing workflow. Virtual-keyboard work was removed; controller bindings are an off-by-default developer fallback. Generic native editor key-event tests remain diagnostic tests, not a proposed virtual keyboard.
No Ultraleap service or connected tracking device was detected. Camera-only Windows hand tracking is the experimental provider track, not a completed optical capability. Passthrough imagery must never be counted as joint tracking.
The earlier delayed live test exposed invalid global Blender context access in the test harness. The live harness now uses a modal window-event context; the shipped add-on has no polling app timer. Historical failed logs remain available beside the final successful camera-off lifecycle runs.
| Milestone | Scope | Exit evidence required |
|---|---|---|
| M0 | Pinned native source, real capabilities, cube/panel, runtime recovery | Compiled fork; stereo calibration; provider coexistence; backend frame timing; camera failure |
| M1 | Independent opacity and workbench placement | AT-01–03, AT-09, AT-18 |
| M2 | Native task shell and internal editor surface | AT-05–06, AT-17, AT-21 |
| M3 | Transaction-safe object manipulation | AT-07–09, AT-13, AT-16 |
| M4 | Qualified optical hands | AT-04–06 plus measured registration and latency |
| M5 | Mesh components and topology adapters | AT-10–12, AT-14–15 |
| M6 | Retained-curve/SDF integration and heavy-scene proxies | Adapter tests and prior-gate regression |
| M7 | Packaging, lifecycle, accessibility, privacy | Supported configuration matrix; AT-19–22 |
- Blender remains the scene and undo authority. No separate imported model.
- One composition route may be active. Advertised passthrough is not proof of camera availability, calibration, or freshness.
- Optical input must carry optical provenance and live tracking evidence.
- Captured editing must preserve one owner and cancel on tracking failure.
- XR presentation helpers remain runtime data, outside saved model geometry.
- Main-thread context and native lifetimes must remain valid. The prior persistent-context-timer crash is an explicit regression constraint.
- Full UI access means an internal editor surface, including popup and text focus. The earlier SteamVR Desktop fallback does not satisfy AT-17.
No synthetic test or desktop preview qualifies stereo eye order, camera calibration, 90 Hz pacing, optical accuracy, actual tracking loss, layer coexistence, or headset comfort. These need the connected hardware and the specification's measured fixtures. The first ticket remains open until those results are recorded. The frozen launcher build and a separate full-feature source build are available; neither qualifies the physical SteamVR-free Index stack.