One-shot installer that takes a FRESH clone of SceneConductor to a fully runnable pipeline — initializes submodules, downloads Blender 4.2.1, builds all five conda envs (./setup.sh, including the GroundingDINO torch-compat patch + _C verification), downloads every model checkpoint (GroundedSAM, GALP, SAM3D, Qwen), wires the SAM3D symlink, and verifies everything. Use this skill WHENEVER the user wants to install / set up / provision / bootstrap SceneConductor, is starting from a fresh git clone, hits setup errors such as a missing `groundingdino._C` CUDA extension, a conda env that cannot be found, or absent model checkpoints, or asks "how do I get this running". Trigger on "/sceneconductor-setup", "set up SceneConductor", "install the pipeline", "provision the environment", "bootstrap a fresh clone". Prefer this over hand-running setup.sh — it also fetches checkpoints and validates the install, which setup.sh alone does not.
Initialize a 3D scene from a single indoor `image.png` — produces segmentation masks, deduped textured GLBs, and GALP outputs (layout_prediction.json + .glb) ready for Blender import. Accepts scene-dir (required), --gpu N (default 0), and --force (delete prior outputs + re-run everything). Trigger on "/stage1-initialize-scene".
Interior look-dev pass that enhances ONLY the lights and stage wall/floor/ceiling colors of an existing Blender `.blend` to match a reference image, leaving object materials and geometry untouched. Trigger on "/stage2-sub-env-enhance".
Build a separable stage (Floor, Wall_NN, Ceiling) inside a `.blend` by committing a rectilinear floor polygon in 2D from a pointmap + image, then extruding in 3D. Also vendors the layout_prediction.json → blender_scene.json → .blend converters (`src/convert.py`, `src/build.py`). Trigger on "/stage2-sub-pointmap-to-separable-stage".
Stage 3 orchestrator: scene-analyze-prepare (auto pre-step if needed) → auto-pass (no LLM) → planner-review (1 LLM call, revises operation plan + adds deletes) → apply revised plan → render → validation (1 LLM call) → island-refiner (needed groups only, N-iter Opus per group (default 20, configurable via num_max_iter) via stage3-sub-island-refiner) → merge-back → final render. Self-contained — auto-dispatches scene-analyze-prepare if its outputs are missing. Trigger on "/stage3-scene-refinement" or "/scene-refinement".
Per-group island refiner. Dispatched once per problem group (identified by validation) via the Task tool. Each dispatch drives the FULL iteration refinement loop (default 20 iters) inside a single `stage3-island-refiner` Opus subagent for ONE relation-group island (anchor + members). The agent refines the WHOLE GROUP PATTERN per iter — multi-member translation moves for spacing, and per-member yaw targets (NEVER uniform yaw deltas) for facing direction. Triggered indirectly via stage3-scene-refinement; do not invoke directly.
Given a scene_dir (image + .blend), extract three planning JSONs — `json/object_state.json`, `json/blend_info.json`, and `inputs/relation_graph.json` — without producing any operation plan or Blender edits. Trigger on "/scene-analyze-prepare".
Single-shot Stage 3 auto-pass — heuristic + graph_tool → operation_plan.json, apply, render. No LLM. Trigger ONLY on direct invocation ("/scene-refiner"), never auto-triggered from "fix the scene".