Files
breeze/plans/2026-11-17-astral-projection.md
waozi ef0de86e02
Cloudflare Pages / Build and verify the site (push) Successful in 3m29s
Integrate Breeze desktop, Android, and app updates
Use the verified dependency lock across native hosts and bundled apps. Preserve saved styles, local app reactivation, and offline packages while integrating the latest practice, Habits, Lumi, and figure preview updates.
2026-11-20 02:24:34 -03:00

21 KiB
Raw Permalink Blame History

Open practice figure work — 2026-11-18

The owner wants one figure model used everywhere: Tai Chi, Sun Salutation, Workout and astral projection, native on Ziran and Kryon without raylib. It must look as good as the owner's reference sculptures (practices figure/design/reference/): smooth, correct proportions, clean colours. The default is a pretty, petite young Asian woman; a man is the second choice. Breeze colours and backdrops stay.

The MakeHuman-based figures were removed on 2026-11-18 at the owner's request; never use a MakeHuman-based body again (MakeHuman, MPFB, Anny, MB-Lab, CharMorph). The owner picked pixiv's CC0 VRoid sample models (practices shared/figure/woman.vrm and man.vrm); the preview (design/figure-preview/) plays every practice with her or him.

Codex continuation — 2026-11-18 (in progress)

Scope received from the coordinator: finish planted-foot locking, stance weight transfer and breathing, gentle head/eye follow and relaxed fingers; check every Workout and Sun Salutation step, export the studio course into canonical practices, repin the preview, then show private before/after clips. The rendering-performance lane owns Kryon's scene_gltf/scene_skin/scene_gpu and present path. Those files have not been changed here.

Implemented locally, not yet committed:

  • Grounded Tai Chi and Workout IK keeps its endpoint, blending soft reach back in only as the foot lifts. Pelvis reach limiting yields smoothly.
  • Weight transfers smoothly onto heel or ball over 14 degrees, avoiding a derivative discontinuity that jerked the whole torso in Single Whip.
  • Tai Chi has a six-second, 2 mm breath and 0.15-degree chest movement. The existing keyed load still drives weight shift and the head follows its cued gaze. VRM eyes lead that head turn by a bounded small amount.
  • Fingers have a graduated relaxed curl; a hook curls the tips separately. Both figures keep their selected models, textures and shader appearance.
  • Side-plank roll takes 2.4 seconds and computes its handover contact from the rig's actual ball distance. Approximate middle-pose dimensions were causing a 3.182 mm discontinuity when support changed feet.
  • The audit separates true floor contact (1 micrometre height tolerance) from peeling/landing; checks drift at 0.1 mm, and prints contact sample counts and maximum drift. The former 2 cm contact-height test counted low swing feet as planted. Floor, speed, acceleration and joint bounds were not relaxed.
  • The studio exporter accepts --media-dir and --course-only; canonical practices taichi/course.tch was exported without editing package caches or narration/locale text. The parity harness accepts PRACTICE_MEDIA and stages its course binary in its own test output directory.
  • design/figure-preview/review.py prepares resumable, hash-identified before/after videos and a contact sheet for both figures on private Xvfb.

Completed local verification:

  • Baseline preview audit reproduced 66 findings (34 world slides, 28 screen slides and 4 jerks).
  • Intermediate audit passed the old 5 mm limit: 255,689 contact samples, no floor/jump/knee/hand/pinned-head findings; worst drift 3.182 mm. The stricter audit isolated that residual side-plank handover and tiny IK compression during held Workout/Sun poses; these have since been edited.
  • Practices: build-queue -- ziran build --project --target=c --exe --entry figure_parts_test:main -o build/figure-parts-test tests/figure_parts_test.zi and build-queue -- ./build/figure-parts-test/figure_parts_test passed, including eye-follow and relaxed/fist/hook finger regression cases.
  • Strict audit completed successfully (desktop_run 77ca6ea365ca5e4b5ec248b7556d916251d1): every Tai Chi item, both Sun Salutation rounds, all twelve Workout exercises and all six astral techniques at 30 fps. Zero slides, floor penetrations, jerks, backwards knees, hands through the trunk, screen slides or pinned heads; 264,014 contact samples, maximum planted drift 0.000000214647 metres. These are solver-joint measurements, separate from the rendered-model review.
  • build-queue -- bash -c 'node design/tai-chi/tools/export-app.mjs --reference --media-dir ../../modules/practices && PRACTICE_MEDIA=/mnt/storage/Projects/waozixyz/modules/practices sh tests/tai_chi_zi_test.sh && make workout-test sun-salutation-pose-test' passed (desktop_run bc3ee581f32dd9ea06df89fa2dc10606ecc5). The formerly failing Tai Chi sample 11 now agrees with the exported canonical course.

Private render review completed successfully (desktop_run ccc42e1ed3ecb388b0be14c25fad0ee5e4b3):

  • build/practice-motion-review/tai-chi-before-after.mp4
  • build/practice-motion-review/sun-salutation-before-after.mp4
  • build/practice-motion-review/workout-before-after.mp4
  • build/practice-motion-review/astral-before-after.mp4
  • build/practice-motion-review/contact-sheet.jpg (woman and man).

The final clips were regenerated after fixing model retargeting (desktop_run c19b992dbb89e7a657f9fe2f925326908c2e). Both figures' wrists now use the cued palm frame, converted through the avatar's bind pose. Limbs solve with each avatar's actual segment lengths, retaining its proportions and reaching the solver's wrist/ankle targets rather than matching bone direction alone.

Selected-model regression and figure-parts tests passed (desktop_run 231362e559e293f46bf9850b64283679f87f), from canonical practices:

build-queue -- bash -c 'ziran build --project --target=c --exe --entry figure_model_pose_test:main -o build/figure-model-pose-test tests/figure_model_pose_test.zi && ./build/figure-model-pose-test/figure_model_pose_test && ziran build --project --target=c --exe --entry figure_parts_test:main -o build/figure-parts-test tests/figure_parts_test.zi && ./build/figure-parts-test/figure_parts_test'

The regression measures actual model nodes for both selected VRMs: ankle, wrist and planted drift all below 0.1 mm while the pelvis moves; palm orientation also passes. The rebuilt preview's complete solver audit still has zero findings and the same 264,014 contact samples. These numerical tests do not measure every vertex or clothing collision.

The complete rendered review covers all twelve Workout exercises at 7.5, 15 and 22.5 seconds and each step of both Sun Salutation rounds, with both figures. --review-points takes timestamps from the preview's timing rules. review.py --coverage-only resumes that part without repeating the four finished clips. Completed desktop_run 5357cb7a54ca47f56a3a671012e3a06fe79c: workout-all-poses.jpg, sun-salutation-all-poses.jpg and coverage.json beside the clips. The earlier capture failed because the script expected 15.0.png while native FormatDecimal writes 15.png; lookup now follows the native one-decimal filenames.

VLC and Gwenview opened the finished four-clip playlist and comparison sheet successfully. All 120 Workout/Sun coverage images and the four floor comparisons were regenerated after the floor/shoe follow-up, then inspected and opened. The stable native preview is running with the current checkout's figure code and media. Remaining: complete the shared integrated app validation, commit own files through private indices, push practices, update both Breeze/preview practices pins with Ziran's package commands and verify the pinned preview, then commit/push Breeze. Other tasks' styles, platform code, narration, cue JSON, Tai Chi draw and toolchain pin edits remain outside this change. No release or version bump.

Remaining outside this movement change

  • Adopt the selected VRM figures in Breeze's practice screens after the shared Kryon integration. Keep timelines, captions, narration, pause/resume and history, and expose the existing woman/man choice. This task verifies the native design preview; it does not claim the app screens use VRM yet.
  • The rendering-performance task owns Kryon scene and presentation changes. Its newer dependency pin needs its own verified handoff before adoption.
  • Complete astral narration through its owning task. The initial audit found 88 cue texts per locale but only 407/880 female clips across ten locales; that historical count has not been re-audited by this movement change.
  • Verify the owning narration task's session JSON has been pushed before claiming the pinned preview uses the final astral cue timings.
  • Load the selected VRMs in the existing three.js astral studies; they still use stand-in skeletons. No change to those studies here.
  • A BVH importer remains unstarted. These practices are hand-choreographed to narration; the diagnosed motion faults were in IK, stepping and interpolation, and were repaired there.

Owner follow-up: ground registration and plain slippers — 2026-11-18

Owner feedback relayed by the coordinator: Tai Chi looks like walking on the thin wall; Workout is too close to/off the deck; measure real soles, put feet on the floor, and replace both pairs of fashion shoes with simple flat cloth Tai Chi slippers. This extends the current unfinished change.

  • Each original painting has clear open ground. Same-name JSON files beside shared/figure/backdrops/*.webp store normalized source coordinates: Tai Chi floor (0.20,0.81)–(0.80,0.97), Sun (0.40,0.87)–(0.60,0.97), Workout (0.20,0.79)–(0.80,0.96), astral (0.18,0.78)–(0.82,0.96). Camera pitch and cover crop are stored with them. No repaint is needed.
  • preview_ground.zi transforms those areas through the same cover crop as the bitmap, fits whole-motion body/ground bounds, and anchors the shadow and feet inside the safe floor. The camera remains fixed through each item. Shadows now lie on y=0, rather than a guessed plane below the lowest current joint.
  • Complete 30 Hz audit at portrait, live-preview and landscape aspects passed: zero ground escapes over 10,589,970 projected foot/shadow samples (desktop_run 09cedc736a720fe1dd80b412373f96664ea0). The existing strict motion audit still has zero findings, 264,014 contact samples.
  • A reproducible CC0 mesh pipeline replaces only shoe primitives with matte stone/charcoal cloth slippers. Originals' shoe bounding heights were 172.56 mm (woman) and 104.62 mm (man). New upper height is 57 mm, with a 3 mm edge, one flat sole plane, no heel or platform. Original non-shoe primitives and every original bitmap/accessor byte are unchanged (tests/slippers_test.py, desktop_run f5570f1b59fb9232e661191e682db64602f5).
  • Actual underside vertices are measured at load. Foot turns use the sole plane instead of aiming the ankle-to-toe bone, and support is retargeted to the solver's heel/ball. Sole length is fitted to that span to avoid a contact jump when support changes from heel to ball. Foot-only meshes use exact affine skinning for this adjustment; other meshes retain DQ.
  • Actual skinned-vertex regression passed both figures in flat stances and heel/ball pivots (desktop_run d63a7e13dd2d907c0f85aa6d8f3426f085dd): woman floor error 0.00001659 mm, pivot error 0.00001746 mm; man floor error 0.00000222 mm, pivot error 0.00001017 mm; flatness below 0.000001 mm. Both ankle/wrist/palm regressions also pass.
  • The additional whole-timeline rendered-sole audit caught 1–9 mm errors in nearly straight Tai Chi stances (desktop_run 2254d8e3c11bf7090ae10eacd8ca4035d9fc). The retargeter now places the pelvis within both real legs' reach before solving to the measured supports. Whole-timeline verification passed as desktop_run 3294ac950ee202ea27e0df5d2b219a125d8d: 5,970 distinct poses and 324,258 planted vertices; maximum floor gap 0.00001636 mm, flatness 0.2578 mm, heel/ball error 0.00001648 mm. It samples every second and skips exactly repeated lower-body states, also checking elevated step-up support. Dense pivot/near-straight regressions and the independent 30 Hz ground/motion audits cover the other boundaries.
  • Complete Workout visual coverage revealed that this native preview omitted the app's wall and bench. It now renders those same support dimensions with shared depth, and includes their bounds in floor fitting. This fixes unsupported-looking wall sits, step-ups and triceps dips.
  • Frozen pre-floor frames are build/practice-motion-review/floor-before-*.png. review.py wrote one *-floor-before-after.jpg per practice after the final rebuild; all four were inspected and opened in Gwenview. Render/audit command e1c9c6cc16d2ebba7229b80d90200ea46c57 passed: all four comparison pictures and clips, 120 Workout/Sun poses, zero 30 Hz motion findings, zero escapes in 10,589,970 projected foot/shadow/support samples.
  • Private image inspection confirmed both figures stand on the courtyard, grassy area, deck and veranda; wall/bench supports also render correctly. It caught a clipped head in crunches: framing now follows the posed head direction and includes clothing/hair extents. Review captures additionally reject a missing silhouette or any silhouette touching an image edge.
  • Regression/native-capture/root-gate command e5e4d5ebe7f4dd4bcf143909ab8b232da742 stopped at a Ziran array-literal syntax error in the new avatar-reload reset. The reset now uses explicit element assignments. Corrected run bcf70c122718a723ed186f2fc7d73b04e00f passed all scoped regressions, renders, motion/floor audits, Tai Chi parity, Workout/Sun tests, build-laws and package-check. Its only failing pre-push gate was app bundles: Apps controller differs from the reviewed native digest at mk/apps.mk:31 (build/node-apps/library.zx). The coordinator routed this to the existing controller/artifact-review lane. No commit/push yet.
  • Reusing the model for Woman/Man now resets measured sole points before loading the next avatar; the real-mesh regression reuses one model to cover this live picker path.

Unchanged details for the owner to decide later: the woman's large white hair ribbon and flared skirt, and the man's hoodie. No other outfit changes are included.

Final scoped verdict — 2026-11-18

The movement, sole contact, floor registration, flat shoes and framing changes pass their local tests. The last full command was:

build-queue -- bash backup/figure-motion/verify.sh

The ignored script records the exact canonical-practices model tests, slipper source-preservation test, native preview rebuild, private Xvfb review/capture, Tai Chi parity and make workout-test sun-salutation-pose-test build-laws package-check, followed by the relevant pre-push plan. Overall exit was 1 solely because the shared app-bundle controller digest gate failed; every scoped motion/render test passed. No failed app validation is being bypassed and no files have been committed.

The final run checked 264,014 planted contacts at 30 Hz (maximum drift 0.000214647 mm) and 10,589,970 projected foot/shadow/support samples (zero floor-area escapes). Both avatars also passed the real-sole near-straight/pivot and live avatar-reload regressions; each measured 4,488 sole vertices. All 120 Workout/Sun images passed the silhouette edge check, including the previously clipped crunch head. Final private native capture wrote 30 frames under build/practice-motion-review/final-native/.

Final visual review inspected contact-sheet.jpg, workout-all-poses.jpg, sun-salutation-all-poses.jpg and all four *-floor-before-after.jpg. Feet and contact shadows sit below the Tai Chi wall and astral railing, inside the Workout deck and on Sun Salutation's clear ground. Workout wall/bench props are present; both figures' slippers are flat and plain. The four refreshed MP4s were opened in VLC and all four floor pictures in Gwenview; each launch reported running after two seconds.

A fresh native preview was opened from the current working tree through ignored backup/figure-motion/open-preview, which sets FIGURE_PRACTICES to canonical practices and executes the stable design/figure-preview/build/app/figure_preview. Its launch reported running (unit harmony-app-direct-open-preview-43c6e9d1, PID 2348411). This provides the locally built design preview before publication. The normal locked package still points at the earlier practices commit, so pinned-media verification and shared Breeze app integration remain pending. The app's practice screens have not been changed to claim VRM adoption.

Next integration steps, owned with codex-breeze-session-2: pass the reviewed controller digest/app-bundle gate and complete its tested root dependency graph; then publish this task's own practices files, repin root/preview practices with official package commands, verify the pinned preview and commit/push only this task's Breeze files and changelog hunks. Preserve all other lanes' current changes. No release was requested.

Owner follow-up: traditional attire and full naturalness — 2026-11-19

Authenticated coordinator handoff fd694c7ecaabc6af9a7ced6e96ffa980c7a3 relays the owner's request to implement everything, test every movement for naturalness, and use attractive traditional meditation clothing while keeping the woman's skirt. This supersedes only the earlier constraint to leave the other garments unchanged. Faces, hair, bodies, identities and flat neutral slippers remain selected.

Implementation in progress, not yet generated/built or validated:

  • Canonical figure/tools/prepare_attire.py starts from checksummed CC0 originals through the established slipper pipeline. It reuses fitted garment topology and skin weights, adds real crossed lapel/sash facings, plain matte linen colors, removes the man's hood/cord triangles and relaxes his trousers. The woman keeps the flared skirt silhouette. New tests/attire_test.py verifies geometry, identity preservation, CC0 metadata, source bytes and removed hood weights. The slipper test now delegates intentional garment changes to this explicit attire check.
  • The existing skirt visibly rotates upward in forward folds. New figure_cloth.zi and model integration constrain the hem under gravity, outside leg capsules and above the floor, retaining the skinned waist. Added fold-through-ninety-degrees/waist/capsule regression source. Actual rendered proof and clipping/stretch correction are still required.
  • The preview used to wrap directly from the last pose to time zero; it now holds the end and restarts only on an explicit Play action. Slower breakdown/follow variants are appended without changing old picker IDs. A raw-RGBA pipe mode and full_review.py retain complete videos plus native timeline-defined 30 Hz slow transition clips, with exact source/binary/media receipts and safe refusal of unresolved partials. --clothing-audit adds actual full-timeline garment surface measurements.
  • Both current pre-attire VRMs are copied as binary recovery files in ignored practices backup/figure-source/slippers/. The source originals are still in backup/figure-source/. Canonical asset bytes have not yet been regenerated by the new pipeline; the old running preview is unchanged.

New validation currently cannot start: build-queue is missing from PATH, /home/wao/.local/bin/ and the canonical Harmony scripts directory. Attempt 1ebefd984cf9daec56f7c16242072a108e39 failed with exit 127 before running anything. Report 8f50042cb7b50dfa22d20806eef899951436 was routed to Harmony and the coordinator/performance lane informed. No queue bypass, new helper, root pin edit or shared app-output regeneration is being used. The prepared next command, after the supported queue is restored, is:

build-queue -- bash backup/figure-motion/verify-attire.sh

Then inspect/correct the private clothing frames before starting the full actual-surface audit and complete/slow playback review. The former bcf70 results remain baseline evidence only; changed model/clothing/preview code needs impacted regressions. Full review must cover both avatars, every practice, transitions, repeated cadence and final holds, plus actual skirt, sleeve and body interactions. Existing native astral preview coverage also needs checking for the moving awareness double, rather than certifying only its resting physical body. No medical or biomechanical certification is implied. Root app/Moto integration and final source/asset identity stay with the existing performance/artifact lane. Everything remains uncommitted.

Further prepared source, still UNVALIDATED while the queue is unavailable:

  • Native astral playback now has an independent awareness model/skin/spring state, translucent shading and inverse-camera rigid displacement/roll. Whole-item framing includes both figures; captures reject a clipped double. This needs private rendered rope/roll/return verification and full cue-boundary continuity analysis before it can count as movement coverage.
  • Complete review capacity is 128 items and rejects overflow in both catalogue passes instead of silently dropping entries. Full videos default to 15 Hz; transition videos retain 30 Hz native samples at quarter speed without manufacturing duplicate frames. Render receipts include the Python tool hash.
  • Fresh scoped instructions now describe plain managed desktop_run commands, but the active higher-priority task instruction still requires build-queue. Requested a supported queue route/instruction correction from Harmony (1466849d7fca7f777824e01ce18f9f4319d7). No build or test was bypassed.