mirror of
https://github.com/meshy-dev/meshy-3d-agent.git
synced 2026-09-19 08:09:05 +08:00
boliu
b9db44b566
Align the outward-facing skills with docs.meshy.ai. Every change is sourced from the public docs or the public, unauthenticated Animation Library endpoint; nothing internal is used. - Animation was a dead end: `action_id: 1 # from Animation Library` with no way to find a real ID. Now points at GET https://api.meshy.ai/web/public/animations/resources (no key, 680 actions, ?category= to narrow) with a working lookup snippet. IDs are not 1..N — the catalog includes -2, -1, 0 — so the old hardcoded 1 was never "the first animation". - model_type "lowpoly" is deprecated and "smart-topology" was missing: documented on image-to-3d with meshy-t2/meshy-t1 and both mutual-exclusion rules. Scoped correctly — smart-topology is image-to-3d only; text-to-3d keeps standard/lowpoly and multi-image-to-3d has no model_type at all. - hd_texture is deprecated: replaced with texture_resolution (2k/4k/8k) across refine, image-to-3d, multi-image-to-3d, retexture. - Rigging examples fed it untextured meshes: preconditions block (textured input, bipedal humanoid, 300k face cap via input_task_id, +Z for model_url) plus a source -> "rig this task ID" table routing text-to-3d through refine. - multi_view_thumbnails promoted to SKILL.md as the default inspection move, so agents stop pulling a 50-200 MB GLB just to look at a result. - FAILED tasks report consumed_credits: 0 (refunded) — a transient failure is a lost wait, not lost credits. scripts/build.py --check and scripts/validate_skills.py both pass. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>