The filter only matched [cinematic_render] tag but Blender outputs its
own "Saved: frame_NNNN.png" lines per frame. Now also matches "Saved:"
lines, extracts the frame number, and formats as:
[cinematic_render] Frame 84/480 rendered
This shows live frame progress in the LiveRenderLog on the frontend.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaced communicate() (blocking) with selectors-based line-by-line
stdout streaming — same pattern as still render. Each frame now
streams live to the frontend:
[cinematic_render] Frame 42/480 -- 55.3s elapsed (0.76 fps)
Pipeline: Blender stdout → log_callback → emit() → Redis →
LiveRenderLog poll (2s) → frontend display
Also added log_callback parameter to cinematic render task call.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The service was sending (glb, frames_dir, width, height, engine, samples...)
but the script expected turntable order (glb, frames_dir, frame_count, degrees,
width, height, engine, samples...). Fixed by adding frame_count and degrees
placeholders to match the expected positional layout.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
System prompt rewritten with explicit RULES:
- Never ask user for info you can query yourself
- "Any product" / "beliebig" → just pick one, don't ask back
- Execute immediately, no confirmation needed
- Be concise, short answers preferred
find_product_renders tool:
- No longer requires product_name or product_id
- Call with empty params → returns any recent renders
- Enables "show me any render with transparent bg"
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Added find_product_renders tool to the AI agent:
- Searches completed renders by product name/ID
- Returns viewable image URLs (relative paths)
- Supports transparent_only filter
- AI formats results as Markdown images/links in chat
Frontend:
- ChatPanel ReactMarkdown renders <img> and <a> tags
- Images shown inline (max 200px height, rounded, bordered)
- Links open in new tab with accent color
System prompt updated to instruct AI to use Markdown image syntax
when showing renders: 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fixed: from OCP.BRepGProp import BRepGProp as brepgprop (was lowercase)
- Routed reextract_rich_metadata_task to asset_pipeline queue (render-worker
has OCC/OCP installed, worker container does not)
- Backfill verified: 45/45 products updated with volume, surface area,
part count, complexity metrics
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- All httpx tool calls use real user_id instead of uuid(int=0) for
service token — fixes 401 Unauthorized on dispatch/override
- create_order auto-submits and auto-dispatches in one step
- System prompt updated: execute immediately without asking for
confirmation, respond in user's language
- Product search returns CAD dimensions (dim_x/y/z_mm)
- query_database tool description includes cad_files schema
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- search_products now returns dim_x_mm, dim_y_mm, dim_z_mm from
cad_files.mesh_attributes->'dimensions_mm'
- query_database tool description updated with cad_files schema
including mesh_attributes->'dimensions_mm' path
- AI can now answer "what's the biggest product?" using real dimensions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Frontend: Layout extracts order/product UUID from URL path and passes
to ChatPanel as contextType/contextId. When on /orders/{uuid}, the
chat knows which order you're viewing.
Backend: System prompt now loads actual entity data for the context:
- Order: order_number, status, line counts (completed/processing/failed)
- Product: name, PIM-ID, category, STEP file status
The AI understands "this order", "current product" etc. Example:
"What's the status of this order?" → agent knows you mean SA-2026-00164
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a tool like query_database fails (e.g., bad column name), the
SQLAlchemy session enters a failed transaction state. Subsequent
operations (like saving the assistant response) then also fail with
InFailedSQLTransactionError.
Fix: rollback the session in the except block of _execute_tool().
Also improved query_database tool description with correct column
names (category_key not category) to help the AI write valid SQL.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Catch all exceptions (not just ValueError) and return meaningful
error messages from OpenAI API errors instead of generic 500.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Actionable AI assistant that uses per-tenant Azure OpenAI credentials
to execute natural language commands against the render pipeline.
Backend:
- ChatMessage model + migration (session-based conversations)
- Chat service with 10 OpenAI function-calling tools:
list_orders, search_products, create_order, dispatch_renders,
get_order_status, set_material_override, set_render_overrides,
get_render_stats, check_materials, query_database
- All tools tenant-scoped (queries filtered by tenant_id)
- Write operations use httpx to call backend API internally
- Chat API: POST /chat/messages, GET /chat/sessions, DELETE session
- Conversation history preserved in DB (last 50 messages per session)
Frontend:
- Slide-out ChatPanel (right side, w-96, animated)
- User/assistant message styling with avatars and timestamps
- Session management (new chat, session history, delete)
- Typing indicator while waiting for AI response
- Floating chat button in bottom-right corner
- Error state for unconfigured AI tenants
Example: "Render all Kugellager products as WebP at 1024x1024"
→ Agent calls search_products + create_order + dispatch_renders
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause: render-worker and render-worker-light shared the same GPU,
causing contention. Complex TRB renders went from 17s → 36s (2x slower).
Changes:
- Thumbnails back to asset_pipeline queue (not asset_pipeline_light)
- Dispatch routing always uses asset_pipeline (no queue splitting)
- render-worker-light gated behind "multi-gpu" profile — only starts with:
docker compose --profile multi-gpu up -d
- For single-GPU setups: all rendering is sequential on one worker
The dual queue approach is correct for multi-GPU machines where each
worker gets its own GPU. On single-GPU, serial execution is faster
than concurrent GPU contention.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of duplicating output types for every variation (WebP vs PNG,
different resolution), keep one canonical output type and override
specific fields per order line via render_overrides JSONB.
Backend:
- render_overrides JSONB column on OrderLine (DB migration)
- Render task merges overrides with output type settings (format, width,
height, samples, engine, denoiser, transparent_bg, cycles_device)
- POST /orders/{id}/batch-render-overrides endpoint for bulk override
- PatchLineBody accepts render_overrides for per-line patching
Frontend:
- Batch render overrides section on OrderDetail: output format dropdown
(PNG/JPG/WebP) + resolution dropdown (512-4096)
- Clear button to remove overrides
MCP:
- create_order tool: accepts product_ids, output_type, render_overrides,
material_override — enables "render all products as WebP" via Claude
- set_render_overrides tool: batch override on existing orders
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Task 4: Dual render queue
- render-worker: heavy (asset_pipeline, concurrency=1) — HQ 2048x2048, animations
- render-worker-light: light (asset_pipeline_light, concurrency=2) — thumbnails, <=1024
- Thumbnails routed to light queue automatically
- Order line renders routed by resolution at dispatch time
Task 5: GLB caching (skip re-tessellation)
- Before tessellating, check if gltf_geometry MediaAsset exists for the cad_file_id
- If found, copy to expected path — render_blender.py finds it and skips tessellation
- Saves 7-11s per re-render of the same product
Task 6: WebP output format
- New 'webp' option in output_format (OutputType admin)
- Blender renders PNG intermediate, Pillow converts to WebP (quality=90, method=4)
- 50-70% smaller files with no visible quality loss
- Correct MIME type (image/webp) in MediaAsset
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Removed from UI (saved to DB but never read by any service):
- Max Concurrent Renders, Stall Timeout, Thumbnail Format, Product Thumbnail Priority
- Render Linear/Angular Deflection (only Scene deflections are used)
- GLB Scale Factor, Smooth Normals, GLB Material Mode, PBR Roughness, PBR Metallic
Fixed Blender status check:
- Old: called is_blender_available() in backend container (Blender not installed there)
- New: dispatches Celery task on asset_pipeline queue → runs in render-worker container
- Returns: available=true, version="Blender 5.0.1", binary path
- Status card moved to System Tools tab with refresh button
Kept active: engine, device, samples, smooth angle, tessellation, scene deflections,
3D viewer zoom limits
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
create_order and add_order_line endpoints were not passing
material_override from the request body to the OrderLine constructor.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- POST /orders/{id}/batch-material-override endpoint
- Dropdown above the lines table: "Apply to all lines…"
- Options: clear all overrides, or select a library material
- Updates all order lines in one request
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add `material_override` nullable column on OrderLine (DB migration)
- Line override takes priority over OutputType override
- PATCH /orders/{id}/lines/{id} endpoint to update material_override
- Inline dropdown on each order line in the OrderDetail page
- Amber background when override is active
- Same output type, different material per line — no need to create a new output type
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The initial implementation only overrode the material_map dict in the task,
but the Blender USD primvar path bypassed it. Now:
- Added --material-override named CLI arg parsed in _blender_args.py
- Both Mode A (factory) and Mode B (template) in _blender_scene_setup.py
override usd_material_lookup and material_map when set
- Passed through full chain: task → step_processor → render_blender → CLI → Blender
- Tested: 175-part bearing rendered with single Steel-Bare material (1/1 materials)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add `material_override` nullable column on OutputType (DB migration)
- When set, ALL product parts get rendered with this single material
- Override applies after alias resolution in render_order_line task
- Admin UI: dropdown in OutputType table to select a library material
- Display: amber badge showing active override material name
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Excel preview detects when a product already has a different STEP file linked
- Excel preview detects intra-Excel conflicts (same product, different CAD model names)
- Product STEP upload warns when replacing an existing file and shows render count
- All warnings are non-blocking (amber badges, toast warnings)
- LEARNINGS.md: all open items resolved
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Per-render-position focal_length_mm/sensor_width_mm (DB → pipeline → Blender)
- FOV-based camera distance with min clamp fix for wide-angle lenses
- Unmapped materials blocking dialog on "Dispatch Renders" with batch alias creation
- Material check endpoint (GET /orders/{id}/check-materials)
- Batch alias endpoint (POST /materials/batch-aliases)
- Quick-map "No alias" badges on Materials page
- Full product hard-delete with storage cleanup (MinIO + disk files + orphaned CadFile)
- Delete button on ProductDetail page with confirmation
- Clickable product names in Media Browser (links to product page)
- Single-line render dispatch/retry (POST /orders/{id}/lines/{id}/dispatch-render)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Root cause 1: BRepBuilderAPI_Transform(shape, trsf, copy=True) destroys
all Poly_Triangulation data. The mm→m scaling was applied before export,
wiping the tessellation from BRepMesh_IncrementalMesh.
Fix: Remove BRepBuilderAPI_Transform entirely — RWGltf_CafWriter already
handles mm→m conversion and Z-up→Y-up rotation internally.
Root cause 2: RWGltf_CafWriter with MergeFaces=False (the default) fails
to find per-face tessellation from the XCAF component hierarchy, producing
degenerate meshes (~2 vertices per face instead of thousands).
Fix: SetMergeFaces(True) to compose face triangulations into proper
per-shape mesh buffers. Vertex count goes from 1,212 to 46,573.
Also bumps cache key version to v2 to invalidate broken cached GLBs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds early-exit checks in dispatch_order_line_render and
render_order_line_task to prevent rendering when order lines are
cancelled or the parent order is rejected/completed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New endpoint POST /admin/settings/regenerate-all-canonical-scenes
queues GLB + USD master export for ALL completed CAD files, replacing
existing assets. Used after pipeline changes that affect tessellation
or normals.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- @timed_step decorator with wall-clock + RSS tracking (pipeline_logger)
- Blender timing laps for sharp edges and material assignment
- MeshRegistry pattern: eliminate 13 scene.traverse() calls across viewers
- Lazy material cloning (clone-on-first-write in both viewers)
- _pipeline_session context manager: 7 create_engine() → 2 in render_thumbnail
- KD-tree spatial pre-filter for sharp edge marking (bbox-based pruning)
- Batch material library append: N bpy.ops.wm.append → single bpy.data.libraries.load
- GMSH single-session batching: compound all solids into one tessellation call
- Validate part-materials save endpoints against parsed_objects (prevents bogus keys)
- ROADMAP updated with completion status
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extract Base Color, Metallic, Roughness, Transmission, IOR from Blender
asset library materials via catalog_assets.py. Store in catalog JSON and
serve via /api/asset-libraries/pbr-map endpoint. Frontend viewers apply
PBR properties to Three.js MeshStandardMaterial using hex color strings
(avoiding Three.js ColorManagement sRGB/linear issues).
Key fixes:
- RLS bypass for material alias lookup in pbr-map endpoint
- pbrMap empty guard prevents premature grey fallback in viewers
- Cache-Control: no-cache on pbr-map requests to avoid stale data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- export_step_to_usd.py: change stage from Y-up to Z-up, keep (X,-Z,Y)
transform — matches GLB orientation exactly (verified: bounding box match)
- export_glb.py: include canonical_material in resolved_material_assignments
DB field (was being dropped during manifest parsing)
- import_usd.py: use pxr customData read (not string primvars — Blender
ignores those, confirmed by testing)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- export_step_to_usd.py: accept --material_map CLI arg, write
schaeffler:canonicalMaterialName as customData on each Mesh prim,
fix geometry transform (strip shape Location before face exploration,
apply both face_loc and shape_loc sequentially)
- import_usd.py: after Blender USD import, use pxr to read customData
directly from the USD file — builds {part_key: material_name} lookup
(Blender ignores STRING primvars and customData, but pxr reads both)
- _blender_materials.py: add apply_material_library_direct() for exact
dict-based material assignment without name-matching heuristics
- _blender_scene_setup.py: prefer direct USD lookup, fall back to
name-matching for legacy USD files without material metadata
- export_glb.py (generate_usd_master_task): resolve material_map via
material_service.resolve_material_map() and pass to subprocess;
include material hash in cache key for invalidation
- ROADMAP.md: update P5 status, add M5-M7 milestones
Tested: 3/3 parts matched (ans_lfs120), 172/175 parts matched
(F-802007.TR4-D1-H122AG). Previous: 0/25 matched.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The queue handles far more than thumbnails: OCC tessellation, USD master
generation, GLB production, order line renders, and workflow renders.
asset_pipeline better reflects its role as the render-worker's primary queue.
Updated all references in: task decorators, celery_app.py, beat_tasks.py,
docker-compose.yml worker command, worker.py MONITORED_QUEUES, admin.py,
CLAUDE.md, LEARNINGS.md, Dockerfile, helpTexts.ts, test files,
and all .claude/commands/*.md skill files.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the endpoint queued USD generation for ALL 295 completed CadFiles,
including 250 orphan CadFiles not linked to any product. Now filters to only
CadFiles referenced by at least one Product.cad_file_id, reducing the backfill
from ~285 to ~41 tasks.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- beat_tasks.py: import app.models at module level so SQLAlchemy can
resolve relationship("Template") and relationship("User") when domain
models are imported in isolation inside task functions. Fixes all
beat tasks (batch_render_notifications, recover_stuck_cad_files) that
crashed every 60s with mapper initialization error.
- _blender_materials.py: build_mat_map_lower() now adds a slug-normalized
key variant (re.sub([^a-z0-9]+, _, kl)) for each mat_map entry. OCC
part names like 'F-802007_TR4-D1-H122AG' → slug 'f_802007_tr4_d1_h122ag'
now matches USD-imported Blender objects. Existing prefix fallback
(key.startswith(part_key)) catches AF-suffix variants.
- Orders.tsx: kanban drag-to-reject implemented. submitted/processing
cards are draggable (cursor-grab). Rejected column highlights with
red ring on drag-over. Drop opens reject reason modal via createPortal.
Confirm calls rejectOrder() mutation + invalidates orders cache.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add per-file Redis SET NX EX 1800 locks to generate_gltf_geometry_task
and generate_usd_master_task — concurrent duplicates (e.g. double-click
of bulk action buttons) now log a warning and return immediately instead
of running two expensive OCC tessellation subprocesses on the same file
- Fix eng.dispose() called inside with Session() block in cache-hit path
of both tasks — moved to after the with block exits (Tasks 3+4 from plan)
- Add cad.updated_at = datetime.utcnow() in save_manual_material_overrides
(was missing vs parallel save_part_materials endpoint)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fine production settings (0.03mm/0.05rad) with GMSH → CharacteristicLengthMax
0.45mm → OOM kill on large assemblies even with per-solid iteration.
GMSH quality is algorithmic (conforming seams, no fan triangles) — a denser
mesh provides no extra UV-unwrap benefit. Cap GMSH tessellation at preview
settings (0.1mm/0.1rad) in the production fallback path.
Normal path (GMSH + valid geometry GLB) is unaffected — continues to reuse
_geometry.glb directly without re-tessellating.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Per-solid iteration prevents OOM on multi-part assemblies (25-part bearing:
2.3GB RAM when processing compound → ~100MB per solid with per-solid approach)
- Fix CharacteristicLengthMax multiplier 5× → 15× and cap MinimumCirclePoints
at 20 (prevents 63-pts/circle on angular_deflection=0.1rad → 231MB → 21MB)
- Geometry task timeout 120s → 600s for large assemblies
- Production task: reuse _geometry.glb when GMSH enabled (no re-tessellation)
and cache _production_geom.glb for OCC (mtime vs STEP check)
- Viewer now prefers production GLB when available (shows correct GMSH mesh)
- GMSH OpenMP multithreading (min(cpu_count,16)) for 4.4× speedup
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Introduces GMSH as an alternative to OCC BRepMesh for STEP→GLB tessellation.
GMSH produces conforming meshes that eliminate fan triangles at cylinder seam
edges — a structural limitation of OCC BRepMesh that cannot be fixed via
deflection parameters.
Changes:
- render-worker/Dockerfile: install gmsh>=4.15.0 + libglu1-mesa + libxft2
- export_step_to_gltf.py: --tessellation_engine occ|gmsh CLI arg +
_tessellate_with_gmsh() using BRep→GMSH→Poly_Triangulation write-back
- admin.py: tessellation_engine setting (SETTINGS_DEFAULTS, SettingsOut,
SettingsUpdate, validation)
- export_glb.py: pass tessellation_engine to export_step_to_gltf.py CLI in
both geometry and production GLB tasks
- Admin.tsx: radio button UI for OCC vs GMSH selection
Tested: 121 faces meshed, 0 BRepMesh fallback, 649K triangles on sample part.
Clean seam edges for UV unwrap — GMSH respects B-rep periodic face boundaries.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Hartmut