--- name: model-inspection description: "Use when someone wants their own Blender model checked for problems: any .blend (a path, or already open) or a character or prop they've made. Run it before they bake, export, rig or send to Unity, and again after a fix. The request may be a plain \"is this mesh clean?\" or \"audit/sanity-check my model\", or a symptom: dark or odd shading patches, flipped or inside-out faces, holes, non-manifold or doubled geometry, a surface that looks off from some angles, a dent the viewport shows but renders don't, parts clipping through each other, or which part sits in front. It inspects the scene's real geometry and depth, so it can name which faces are bad and where, instead of guessing from screenshots. No tool name needed. Don't use for general Blender how-to (retopology technique, shortcuts, bevels, render passes, install or launch trouble), pretty renders, or fitting a model to reference images." --- # Blender model inspection Flat, orthographic renders hide flipped faces, dents, holes and parts that pass through each other. The `.blend` holds the exact topology and depth, so read it with these tools instead of judging from a screenshot. Every tool runs Blender headless with `--factory-startup --background`, never overwrites the `.blend`, and needs Blender; if it is missing or unverified, use the sibling [`bootstrap`](../bootstrap/SKILL.md) skill first. ## Delegate the analysis Hand any inspection, audit or review request to the `blender-inspect:blender-model-inspector` agent through the Agent tool (`subagent_type: "blender-inspect:blender-model-inspector"`), not the main thread and not a general-purpose agent. Give it the `.blend` path, the parts or suspicion the user named, and any pairs that may touch. It runs the loop below and returns the report. It is read-only, so the main thread keeps the fix step (4); after a fix, delegate the re-audit to the agent again. A single quick count (`blender_mesh.py`) may run inline. If you are that agent, skip this section and run the loop yourself. Resolve `` two levels above this file. Tools live at `/scripts/blender_*.py` and run under any Python 3.11+. ## The loop 1. **Audit** with `blender_mesh_audit.py`. It gives exact counts and world-space locations for ten defect classes. Start here: it is the only tool that counts defects. 2. **Look** with `blender_inspect_render.py` at the locations the audit named, or at a suspicion the audit cannot test (a dent, a soft surface problem). 3. **Order** with `blender_depth_order.py` when the question is which part is in front, or whether layers sit on each other or pierce each other. 4. **Fix** in the scene. That is the user's modeling step; inspecting never edits. 5. **Re-audit**, and re-render the same views. A fix is shown only when the counts or images change. Never report a fix from the edit alone. Stop when every requested check has a payload or a named refusal. Do not run all three tools on every scene; choose by the question. ## Which tool for which question | Question | Tool | |---|---| | Object, polygon, vertex counts, materials, bounds | `blender_mesh.py` | | Holes, loose or doubled geometry, flipped or crossing faces | `blender_mesh_audit.py` | | Do objects A and B pass through each other | `blender_mesh_audit.py --pairs A:B` | | Can I see the dent, seam, flipped face or crossing | `blender_inspect_render.py` | | Which part is in front, by how much, how much do they overlap | `blender_depth_order.py` | | Garment clearance against a body | `blender_fit.py --mode probe` | | A plain named-view render, or a render compared with a reference | `blender_render.py`, `blender_multiview_render.py` | `blender_fit.py` and the two plain renderers are the earlier tools. Its `apply-copy` mode writes a new `.blend`, so use `probe` unless the user authorised a bounded edit. ## Read each payload Every payload carries `interpretation_limits`; read them before a claim. Refusals exit 2 with empty stdout and one stderr line: report the line, never infer a result. **`blender_mesh_audit`** (`mesh-audit-v1`; `--max-locations` defaults to 50): per object `defects..count` and `locations`, plus `clean`, and a `summary`. The ten classes are `non_manifold_edges` (3+ faces), `boundary_edges`, `wire_edges`, `loose_verts`, `duplicate_verts`, `degenerate_faces`, `ngons`, `poles`, `inconsistent_normals` and `self_intersections`. - `count` is exact; `locations` is cut at the limit, lowest index first. - Read the counts, not only `clean`: a cylinder cap (n-gon) or a sphere pole makes `clean` false without a bug. Report deliberate topology as such. - Locations are world-space scene units. An edge's is its midpoint; a crossing's is near, not on, the crossing. Indices refer to the **evaluated** mesh (render modifiers applied), so they match the base mesh only without modifiers. - It does not see: a uniformly inverted island or negative-scaled object (count 0), a shell wholly inside another, long sliver faces above the area threshold, or whether a defect matters. **`blender_inspect_render`** (`inspect-render-v1`; needs `--output-dir` and `--views`: presets `front-left-high`, `front-right-high`, `back-left-high`, `back-right-high`, or a `render-views-v1` manifest; `--modes` default to all five; perspective, 50 mm by default): per view, `files.` paths, `camera`, `coverage`, `depth_range`, and `RENDER_BLOCKED` for a view with no geometry. Framing is locked across views and modes. Open the PNGs and describe what you see. | Mode | Use it for | Blind spot | |---|---|---| | `matcap` | dents, ridges, seams, creases (cavity shading on); a flipped face renders as a **hole** (back-face culling on) | culling hides a single-sided quad seen from behind | | `normal` | crossings, faceting, sudden direction changes | **does not show a flipped face**: back faces are turned toward the camera | | `depth` | relief and distance: near is warm, far is cool; the PNG's bottom strip carries the near/far legend in scene units; the `.npy` is float32 planar depth, NaN for background | depth is along the camera axis, not the distance from the camera centre | | `ao` | contact, crevices and gaps (convex shapes show almost none) | not a defect detector on its own | | `object-id` | which object owns each pixel; the JSON maps colour to object | it is coverage, not geometry | The matcap and normal renders can disagree about a flipped face; that disagreement is itself the evidence. Use several views: a defect facing away from one camera shows in another. **`blender_depth_order`** (`depth-order-v1`; same `--views`, no output directory): per view, `objects[]` with `depth_range`, `screen_bbox` and `visible_fraction`, and `pairs[]` sorted by overlap. Read each pair's `summary` line first, then check before trusting it: - `front_fraction` near 1.0 means a clean order. A low value means the two interleave or interpenetrate; follow up with `blender_mesh_audit --pairs`. - `method: passes` is per-pixel; `method: vertices` is a bounding-box estimate. Say which. - `depth_gap` is the surface separation (median over shared pixels); `range_gap` is the empty space between the extents. A part resting on another has a negative `range_gap` and a small positive `depth_gap`: that is normal. - Overlap compares silhouettes, including pixels another object hides; `visible_fraction` is what is seen. A preset's direction runs from subject to camera with front = -Y and up = +Z, and a thin part can be edge-on to one. **`blender_mesh`**: counts come from the base mesh, so unapplied modifiers change the render-visible density; bounds are in scene units. ## What each layer does not establish - Counts and locations do not say a defect is visible or matters. Pair them with a render. - Renders and depth do not prove topology: a sealed-looking surface can hide doubled vertices. They cover only the views requested. - Depth order is from the camera's view. It does not establish physical clearance or contact; use `--pairs` or `blender_fit --mode probe`. - Nothing here establishes that the model matches a concept image or is stylistically right. For the concept side use the `pil-agent-plugin` tools (relative concept depth is its `pil_depth.py`). ## Report For each finding give the object, the class or pair, the count, the world-space location, the view and image path that shows it, and the limit that applies. State which checks were not run. After a fix, cite the before and after counts.