Campus pipeline hardening: backend parity, wall union, mullion grids, Draco tooling
Sandbox round (Linux): - detect_sheet_frames.py: multi-frame sheet splitter (column AND row layouts), grid-vote per-floor offsets, section-elevation harvesting for storey heights - dwg_dump_json.mjs + dwgjson_to_dxf.py: WASM libredwg DWG->DXF fallback (blocks/INSERT preserved, ATTRIB->TEXT so axis labels survive) - assemble_campus.py: manifest-driven campus assembly with mandatory neighbour-gap gate (overlap = exit 3, 0.1-0.3m flagged as expansion joints) - site_context_from_dump.py: site-context floors (ground/roads/parking/ fence/trees/context massings) from site-plan dumps - write_offline_viewer.py + assets/offline-viewer-libs/: fully-offline single-file viewer (vendored three r147 + pako, gzip-embedded GLB) - audit: clip-scoped coverage figures; report: multi_frame_sheet warning; audit gate floors_aligned catches the diagonal-staircase failure mode Mac round: - build_blender_showcase.py now honors slab_bounds_fallback and prefers footprint polygons for generated roofs (bounds rects made roof plates jut past the facade and buried campuses under full-extent slabs); contract now matches spec_to_glb.py exactly - merge_wall_solids (default off) in BOTH backends: shapely union in spec_to_glb.py, native delaunay_2d_cdt union in Blender; endpoint-shared wall ends pre-extended w/2 so L-corners close seamlessly (flat caps leave a notch a union alone cannot fill) while openings stay flat - glazing_mullions (default off): curtain-wall panel grid (verticals every glazing_panel_w + floor-line transoms) as pseudo wall segments in the <floor>__walls node; glass pane kept - build_showcase_variants.py keeps source_file=="site" floors intact in all variants (roads/trees live on filtered categories) - web template: instanced low-poly trees from metadata site_trees, grade-level shadow catcher (was at basement depth), data-URI favicon - tools/compress.mjs deps pinned (package.json); 15MB campus GLB -> 0.57MB - tests: 24 -> 52 checks (wall union, mullions, byte-identical defaults) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
57
SKILL.md
57
SKILL.md
@@ -48,12 +48,26 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- State whether the output is Web showcase, LOD-like massing, IFC/BIM proof, or not sufficient.
|
||||
- Include what was not modeled and what evidence supports that judgment.
|
||||
|
||||
## One-Shot Definition of Done
|
||||
|
||||
The pipeline is designed so a first pass either succeeds or STOPS at a loud gate. Do not hand over a model unless every line below has evidence (a report line, a gate status, or an image):
|
||||
|
||||
1. Read report clean: every `Warnings` entry resolved. `multi_frame_sheet` -> ran `detect_sheet_frames.py` and used its clip/offset config; `likely_xref_shell` -> user re-exported with XREFs bound.
|
||||
2. Frame alignment proven: detect report shows `grid-aligned, votes x>=3 y>=3` per floor (any `WEAK grid match` floor was verified on its overlay SVG).
|
||||
3. Audit exit code 0, and gates `floors_aligned` / `glazing_captured` / `no_annotation_layers_classified` all PASS. Never pass `--no-strict-gates` to proceed.
|
||||
4. No `WARNING topless_group` from derive (roof handled via caps or `roof_specs`), and no `WARNING mesh_from_polygons` from the GLB build (else slabs/roofs silently vanished — install mapbox_earcut).
|
||||
5. Visual check done by the Agent, not assumed: `verify_showcase.py` previews plus one rendered iso/plan image actually LOOKED AT (floors stacked, roof present, no slab fins beyond facade — default `cap_buffer` 1.3 is for colonnades, use ~0.1 otherwise).
|
||||
6. Campus placement (if 总图/定位图 exists): building envelope dims match the site rect 1:1 (e.g. 面积核算 frame), rotation taken from the rect edge angle, 180° flip disambiguated by an asymmetric feature (machine rooms / basement / entrance), and a top-view overlay image confirms the building inside its site rect. Neighbour gaps measured numerically: no overlaps, 贴建 joints 0.15-0.3 m with corroborating evidence, buildings inside 红线.
|
||||
7. `tests/test_extraction_pipeline.py` passes after ANY script change.
|
||||
|
||||
## Command Workflow
|
||||
|
||||
1. Prepare the CAD folder.
|
||||
- Work from a folder under the current workspace when possible. On macOS, avoid direct reads from `Downloads` if TCC blocks access; ask the user to move/copy the CAD folder into the workspace.
|
||||
- Preserve the original CAD folder. Put converted/intermediate files in `work/`.
|
||||
- If DWG files are present, prefer `libredwg` `dwg2dxf` for batch conversion. Use ODA/QCAD/FreeCAD only as fallback — EXCEPT for drawings containing `3DSOLID`/ACIS entities: libredwg silently drops ACIS data (R2013+ stores it in the ACDSDATA section), so 3D-solid DWGs must go through ODA File Converter (`/Applications/ODAFileConverter.app/Contents/MacOS/ODAFileConverter <abs_in_dir> <abs_out_dir> ACAD2004 DXF 0 1 "*.dwg"`). ODA reports bad arguments via GUI dialogs, not stderr: both directories must exist as absolute paths, and a silent exit-0 with no output means a dialog is blocking on screen.
|
||||
- No native converter available (Linux sandbox, no ODA/libredwg)? Use the WASM fallback: `npm install @mlightcad/libredwg-web`, then `node scripts/dwg_dump_json.mjs in.dwg dump.json --module-dir <node_modules>` and `python3 scripts/dwgjson_to_dxf.py --json dump.json --out work/dxf_raw/file.dxf`. This carries LINE/PLINE/ARC/CIRCLE/ELLIPSE/SPLINE/TEXT/INSERT + full block definitions (enough for the extraction pipeline), but drops HATCH/DIMENSION/ATTRIB — do not use it to judge hatch-only column fills or attribute-held axis numbers.
|
||||
- Multi-discipline sets (建筑/结构/给排水/电气/暖通 subfolders): model from 建筑 (architecture) only by default; pull 结构 later just for columns/beams if the showcase needs structure. Never feed all disciplines into one extraction pass.
|
||||
|
||||
3D-solid DWG branch: if the read report shows `3DSOLID`/`MESH` as the main types (typically also flagged `near_empty` for plan purposes), the floor-plan extrusion pipeline does not apply. Convert via ODA, then tessellate with `ezdxf.acis` (`load_dxf` → `mesh_from_body`, which returns a list of meshes). Only polyhedral/flat ACIS faces tessellate; count curved records (`cone-surface`, `spline-surface`, ...) in the SAT text and report them as not extractable. Export captured meshes + `MESH` entities to OBJ/GLB (OBJ is Z-up: import into Blender with `up_axis='Z'`). If curved skin dominates, tell the user to export OBJ/STL/STEP from a full CAD kernel (AutoCAD/BricsCAD) instead of promising a complete model.
|
||||
|
||||
@@ -68,7 +82,9 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- Use `references/pipeline.md` and `references/building_spec_schema.md`.
|
||||
- Map CAD layers to normalized categories: `walls`, `glazing`, `roof`, `columns`, `beams`, `railings`, `doors`, `detail_lines`, `floor_lines`, and `grid`.
|
||||
- Identify one drawing per floor or split sheets that contain multiple floor plans side by side.
|
||||
- MULTI-FRAME SHEETS (the common Chinese pattern: one DWG per building holding ALL its plans + elevations + sections stacked in modelspace): run `scripts/detect_sheet_frames.py` on the sanitized DXF. It finds `XX平面图` title texts (filtering catalog rows and title-block copies), proposes a wall-bbox-refined `clip` per floor frame, and computes per-floor `offset`s by voting-based grid-line correlation (label-free; TArch grids live on layer `DOTE`, not `AXIS`). Never model such a file as one floor (all sheets superimpose) and never stack frames without offsets (floors staircase away diagonally — the classic broken output). Read its report: `WEAK grid match` floors need overlay verification; elevation/section frames are listed for floor heights (`z`/`height` from DIM_ELEV 标高 texts like `8.100`), not as floors.
|
||||
- Align floors by structural grid or repeated control axes. Use millimeters-to-meters scale (`0.001`) unless the report proves the CAD is already in meters.
|
||||
- Floor keys: `floor_sort_key` treats a leading `B` as basement (sorted below ground). Do not invent keys like `B1F` for "Building 1 Floor" — use per-building specs instead, and set explicit `floor_order` whenever keys are nonstandard.
|
||||
- When CAD linework has enough exterior/interior boundary signal, enable generated footprints for floor slabs and roof caps. Prefer footprint polygons over rectangular floor bounds for visual/digital-twin presentation.
|
||||
- Run `scripts/derive_showcase_spec.py` to extract line/polyline geometry into `building_spec.json`.
|
||||
|
||||
@@ -77,6 +93,7 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- Run `scripts/draw_coverage_overlay.py`.
|
||||
- The audit enforces hard gates and exits with code 3 on failure (e.g. glazing evidence exists in blocks but the spec has 0 glazing segments). A failed gate means the extraction is known-incomplete: fix the config or source export and re-run; never bypass with `--no-strict-gates` just to proceed.
|
||||
- The headline coverage ratio counts modelspace linework only. Read the `Block-Locked Geometry By Layer` section too — geometry inside INSERT blocks is reported separately and a 96% coverage can coexist with an entirely missing curtain wall.
|
||||
- CLIP-SCOPE CAVEAT: audit accounting is file-global while clip-based extraction is frame-scoped. On multi-frame sheets a healthy extraction can still show ~40% coverage and a block-locked WARN, because elevations/sections/details inflate the denominator. Judge such runs by the per-floor overlays and segment counts, not the headline ratio.
|
||||
- Review the Markdown audit and SVG overlays. Iterate config if important geometry remains unclassified or hidden in high-risk entity types.
|
||||
|
||||
5. Generate Blender/GLB.
|
||||
@@ -92,6 +109,7 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- Run `scripts/build_showcase_variants.py --build` to write the filtered specs, call Blender for all three GLBs, and create `variants_manifest.json`.
|
||||
- Run `scripts/write_web_showcase.py --variants-manifest ...` to copy the template, use the standard model as `model.glb`, copy all variant GLBs, and add `model_variants` to `model_metadata.json`.
|
||||
- Use the single-GLB `write_web_showcase.py --glb ... --metadata ...` path only when the user explicitly wants a quick one-model preview.
|
||||
- Share-by-file delivery (no server, no network — mainland-China CDNs hang): `scripts/write_offline_viewer.py --spec ... --glb ... --out viewer.html` emits a fully self-contained HTML (three r147 UMD + pako inlined from `assets/offline-viewer-libs/`, GLB gzip-embedded, ACES + IBL + shadows + physical glass + instanced trees + grouped toggles + click-to-inspect, loading failures shown on screen). Smoke-test headlessly in node: decode the embedded payload and GLTFLoader().parse it.
|
||||
- Install local web deps in that output folder:
|
||||
`npm install --prefix outputs/<project>/web three@0.181.2 lucide@0.468.0`
|
||||
- Start a local server from the web output folder and give the user the URL.
|
||||
@@ -102,6 +120,33 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- Check console logs are clean, controls work, mobile panels do not overlap, and the GLB loads.
|
||||
- Report model precision honestly: showcase model, LOD-like massing/structural visualization, or IFC/BIM proof.
|
||||
|
||||
8. Raise presentation quality when the user wants "digital-twin" polish.
|
||||
- Follow `references/web_presentation_notes.md` in tier order: official template + Draco + ACES/RoomEnvironment lighting + physical glass first (cheap, big win); then geometry upgrades (merged wall solids, mullion grids, road/green surfaces, instanced trees); then interaction (click-to-inspect from GLB extras, camera tours, progressive per-building loading).
|
||||
- Polish never excuses a wrong model, and enhancements derived from plan linework must not be presented as BIM-verified elements.
|
||||
|
||||
## Multi-Building Campuses (厂区/园区 sets)
|
||||
|
||||
A folder like `建筑-1#厂房 / 2#厂房 / 3#辅房 / 总图` is a CAMPUS: several buildings plus a site plan — NOT floors of one building. Never map "one DWG = one floor" here.
|
||||
|
||||
- Model each building separately: one detect-frames pass, one extraction config, one `building_spec.json`/GLB per building DWG. Small structures (机械停车坑/水池) are usually one or two frames.
|
||||
- Assemble with `scripts/assemble_campus.py`: record each building's transform (model anchor -> surveying anchor + campus rotation) in `site_manifest.json` and let the script do the merge — it also runs the mandatory neighbour-gap check automatically.
|
||||
- Units differ by sheet type: building plans are almost always millimeters (`scale 0.001`), while 总图 site plans are often METERS at real surveying coordinates (huge x/y like 349xxx/663xxx, texts ~2-4 units tall). The extraction config has a single global `scale`, so never mix a meters site plan and mm building plans in one spec.
|
||||
- Place buildings with a site manifest: from 总图/定位图 (prefer a "带轴号(可点坐标)" version if present) measure each building's anchor point and rotation, then assemble per-building GLBs in the viewer/scene with those transforms. Record the manifest (building -> position/rotation/source) in the notes file.
|
||||
- Proven placement recipe (贝昂 case): find the building's envelope rectangle in the site drawing (layer 面积核算/建筑轮廓线) — its edge lengths must equal the model's wall-envelope bbox 1:1 (here 90.2 x 42.2 m exactly), edge angle gives rotation (7.57°), the corner matching the model's (min x, min y) gives translation. Resolve the 180° flip by densest-feature voting (e.g. machine-room walls inside a known sub-rect). Then EITHER merge at spec level (transform all segments/footprints/caps into site-local meters and append a SITE context floor) and export one GLB, OR keep per-building GLBs plus transforms in the manifest. Pick a site-local origin (用地红线 min corner), never raw surveying coordinates — float32 breaks at 349 km.
|
||||
- Anchor placement on WALL evidence, never roof/eave envelopes: 屋面 outlines include canopies and asymmetric overhangs — concentric placement against one put 2# 0.45 m off (26 cm INSIDE its neighbour). Best anchors, in order: corner coordinate callouts (X=/Y= texts), wall lines, envelope rects whose edges equal the wall bbox exactly.
|
||||
- Adjacent-building gap sanity check is mandatory: compute the wall-to-wall gap in the shared-edge frame. ~0.15-0.30 m means 贴建 with an expansion joint (变形缝) — corroborate with the neighbour's roof labeled on this building's roof plan and matching floor elevations. Several metres means fire separation. A NEGATIVE gap (overlap) is always a placement bug. Cross-check the chosen position against the boundary: if a hypothesis (e.g. a 10 m fire gap) pushes the building through 用地红线, the hypothesis is wrong.
|
||||
- Site context is a deliverable, not decoration — a campus scene with only a few boundary lines reads as broken. Run `scripts/site_context_from_dump.py` on the 总图/定位图 dumps: it emits ready-to-merge floors for ground (用地红线 footprint slab), roads + parking stalls (lines, arcs, block-expanded 车位 INSERTs), fence (围墙 as 1.8 m railings), street trees (行道树 INSERT points as green posts, custom `trees` category), and low-confidence massings for context structures drawn only in the site plan (连廊/门卫/消控室, named by nearest label). It takes one plan-copy window (site sheets usually hold several scheme copies), a per-source unit scale (总图 in meters, 定位图 copies in mm), and an exclude-quad so the real building model replaces its own site outline. Buildings that have their own drawing sets get real models placed by the manifest — never fake them as massings.
|
||||
|
||||
## TArch / 天正 t3 Export Preset (most Chinese architecture DWGs)
|
||||
|
||||
`_t3` exports use TArch layer names that the AIA-style defaults miss. Calibrated mapping:
|
||||
|
||||
- `walls: ["WALL"]`, `columns: ["COLUMN", "*COL*"]`, `glazing: ["CURTWALL", ...]`, `roof: ["屋面", "*屋顶*"]`, `detail_lines: ["STAIR", "EVTR", "平面", "开口"]`, `grid: ["DOTE", "AXIS"]` — plan grids are on `DOTE` (dash-dot); `AXIS` often holds only detail-sheet stubs.
|
||||
- Doors and windows BOTH live on layer `WINDOW` as INSERTs, and the block geometry sits on layer `0` (inherits by insert), so layer-based block probing keeps NOTHING. Use explicit expansion plus block-name classification:
|
||||
`"expand_blocks": ["$TCHSYS$*", "$DorLib2D$*"]`, `"block_name_categories": {"glazing": ["$TCHSYS$WIN2D*"], "doors": ["$DorLib2D$*"]}`.
|
||||
- Exclude annotation/frame layers inside clips: `["*-IDEN", "*图签*", "PUB_*", "DIM_*", "*TEXT*", "初审修改", "面积核算", "格子线", "辅注", "SEC-*", "立面", "剖面", "雨水管", "家具"]` (revision clouds live on `初审修改`-style layers).
|
||||
- TArch column fills may be HATCH-based; if arch plans yield suspiciously few `columns`, take columns/beams from the 结构 set instead of inventing them.
|
||||
|
||||
## Modeling Rules
|
||||
|
||||
- Prefer stable, layered massing over overfitted geometry. CAD floor plans often contain annotations, dimensions, hatches, XREF artifacts, and duplicated lines.
|
||||
@@ -125,8 +170,12 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
## Bundled Resources
|
||||
|
||||
- `scripts/cad_dxf_report.py`: batch DWG/DXF conversion, DXF sanitizing, and read report generation, including XREF-shell / near-empty sheet warnings.
|
||||
- `scripts/derive_showcase_spec.py`: configurable extraction from DXF line/polyline layers into a normalized building spec, with block-content probing for glazing/railings/doors locked inside INSERTs.
|
||||
- `scripts/audit_showcase_extraction.py`: coverage and risk audit with hard gates (exit 3 on failure) and block-locked geometry accounting, for Agent review before 3D generation.
|
||||
- `scripts/detect_sheet_frames.py`: multi-frame sheet splitter — finds plan-title frames (column AND row layouts), refines per-floor clips, grid-correlates per-floor offsets, and emits a config skeleton + report, including suggested storey elevations harvested from 剖面/立面 标高 texts.
|
||||
- `scripts/assemble_campus.py`: manifest-driven campus assembly — transforms per-building specs into site-local meters (prefixing floor keys, re-keying roof_specs), merges site-context floors (dropping massings covered by real models), and enforces the neighbour-gap check (overlap = exit 3; 0.1-0.3 m flagged as expansion-joint candidates).
|
||||
- `scripts/dwg_dump_json.mjs` + `scripts/dwgjson_to_dxf.py`: WASM libredwg fallback (npm `@mlightcad/libredwg-web`) for DWG->DXF when neither native libredwg nor ODA exists; preserves block definitions and INSERT transforms.
|
||||
- `scripts/site_context_from_dump.py`: campus site-context floors (ground/roads/parking/fence/trees/context massings) from 总图/定位图 dumps, with per-source unit scale, plan-copy windowing, and an exclude-quad for the really-modeled building.
|
||||
- `scripts/derive_showcase_spec.py`: configurable extraction from DXF line/polyline layers into a normalized building spec, with block-content probing for glazing/railings/doors locked inside INSERTs. Single-pass over expanded entities (floors share one INSERT expansion; block leaf-layer memoization skips irrelevant blocks), so multi-frame configs with many floors stay fast.
|
||||
- `scripts/audit_showcase_extraction.py`: coverage and risk audit with hard gates (exit 3 on failure) and block-locked geometry accounting (definition-walk with per-block-name caching, no per-INSERT explode), for Agent review before 3D generation.
|
||||
- `scripts/draw_coverage_overlay.py`: SVG overlays comparing source CAD vectors to extracted model geometry.
|
||||
- `scripts/verify_showcase.py`: Blender-free QA from `building_spec.json` (+ optional GLBs): per-floor segment/cap/glazing-continuity report, GLB size-budget check, and orthographic front/side/plan previews.
|
||||
- `scripts/build_blender_showcase.py`: Blender CLI renderer/exporter from `building_spec.json` to `.blend`, `.glb`, `.png`, and metadata. GLB is Draco-compressed by default.
|
||||
@@ -142,7 +191,9 @@ The Agent owns interpretation. Scripts are tools for evidence and repeatable exe
|
||||
- `references/agent_review.md`: acceptance/iteration/BIM-claim checklist for Agent judgment.
|
||||
- `references/a208_case_notes.md`: proven details from the A208 DWG/DXF/Blender/Web/IFC run.
|
||||
- `references/optimization_notes.md`: lightweighting + completeness lessons (Draco/DRACOLoader, glTF Y-up convention, block-locked glazing, glazing continuity, column-top caps, NaN sanitation, verification).
|
||||
- `assets/web-showcase-template/`: static Three.js viewer template (Y-up, registers DRACOLoader, cache-busting `?v=` build token).
|
||||
- `references/web_presentation_notes.md`: tiered visual/interaction/performance upgrade playbook for the Web showcase (template + Draco + ACES/environment lighting + physical glass; merged wall solids, mullions, site surfaces, instanced trees; click-to-inspect, tours, progressive campus loading) with honesty rules for presentation enhancements.
|
||||
- `assets/web-showcase-template/`: static Three.js viewer template (Y-up, registers DRACOLoader, cache-busting `?v=` build token; RoomEnvironment IBL, model-sized shadow frustum + ground shadow catcher, physical-glass glazing, click-to-inspect).
|
||||
- `scripts/write_offline_viewer.py` + `assets/offline-viewer-libs/`: fully-offline single-file HTML viewer generator (vendored three r147 UMD + addons + pako; gzip-embedded GLB) for share-by-file delivery where CDNs/servers are unavailable.
|
||||
- `tests/test_extraction_pipeline.py`: synthetic-DXF regression test for block-locked glazing capture, `*-IDEN` exclusion, audit gates, and XREF-shell warnings. Run it after changing extraction/audit/report scripts.
|
||||
|
||||
## BIM Branch
|
||||
|
||||
Reference in New Issue
Block a user