What happened
CodePawl’s GitHub release notes describe v0.20.0 as the point at which 67.js becomes Crucible. The package name is now @codepawlhq/crucible, with add-ons under @codepawlhq/crucible-*. The renamed commands are crucible-mcp and crucible-addon-check, the renderer class is CrucibleRenderer, and the skills are build-crucible-scene and render-crucible-media.
The rename includes a short migration path. The old 67js-mcp, 67js-addon-check, and Renderer67 names continue to work for one release, while printing a rename notice. Recipes, assets, and packs using the former .67js.scene, .67js.asset, .67js.pack, and 67js-pack.json names still load. That separates updating package references and command invocations from migrating existing saved scene artifacts.
The release also changes the MCP-oriented asset workflow. Its asset catalog can load a GLB model whose textures are stored next to the file and embed those textures, including a Kenney kit with Textures/colormap.png. The scene_apply operation can place the first instance of a catalog asset, allowing an agent to build a scene incrementally and add a model at any step. The release notes additionally clarify geometry guidance intended for agents: node scale is the radius for cylinders and spheres, with dimensions for a six-segment cylinder.
Native path-traced scenes using the Cycles kernel receive larger stated limits. buildTraceScene accepts maximumTriangles up to 4,000,000 unique triangles, or four times that when every copy is counted, and native render jobs on the Cycles kernel use that limit. Shared-geometry instances can reach 500,000, up from 65,536. The CPU and WebGL path tracers retain their existing lower limits. CodePawl also says scene-file rendering can handle GLB files of 200 MB and above with lower memory use during loading.
The release includes several rendering and content features. Grid volumes can take a sequence of .67vg frames with fps and loop settings, while scripts/vdb-to-grid.py --sequence converts a numbered VDB sequence in one run. Cook’s Cycles kernel gains light portals: a rectangular area light with portal: true emits no light itself and samples the sky through it, intended to help a room lit only through a window converge. A native rendering dispatch budget splits a frame into GPU dispatches sized from prior measured times; CodePawl says the image remains byte-identical to one rendered without the budget.
What the archive adds
This is a new event in a broader thread about tools making agent operations more composable. New Runtime’s earlier coverage noted that GitHub’s October 28 Copilot release made agent sessions more portable across clients and introduced an open Agent Host Protocol in its November 2 note. Crucible’s release applies the same general direction to 3D scene assembly rather than coding-agent sessions.
Trend signal
The GLB catalog import and scene_apply support agent-ready software exposes capabilities. They expose bounded actions—load an asset with its textures and place an initial instance—that another system can invoke in a scene-building sequence. This is more operational than presenting 3D generation as a single model prompt, although the release does not establish task-level reliability for agents using those actions.
Why it matters
For teams using agents to create or modify 3D scenes, asset handling is an integration problem as much as a generation problem. A model is not useful if it can select a GLB but cannot bring along its referenced textures or put the object into the active scene. Catalog support for adjacent textures and an explicit scene_apply placement operation close two concrete gaps in that workflow.
The larger native limits matter differently. They make higher-complexity Cycles-kernel scenes possible, but they are not general limits across every Crucible renderer. Teams targeting CPU or WebGL tracing should retain their existing capacity assumptions. Meanwhile, the rename is an immediate maintenance concern: update package, command, renderer-class, and skill references before the temporary compatibility path expires.
Limits and uncertainty
CodePawl’s release notes describe the implementation, but they do not establish adoption, agent completion rates, or performance on representative production scenes. The 4,000,000-triangle and 500,000-instance figures are specific to native path-traced scenes using the Cycles kernel; CPU and WebGL paths retain lower limits. The byte-identical claim applies to the dispatch-budget feature under the conditions CodePawl describes, rather than serving as a general comparison across all renderers and changes in the release. Compatibility behavior and rendering results have not been independently checked here.
Skill to build now
If a team maintains a Crucible or 67.js integration, make a small migration inventory this week. Search for the old package, command, renderer-class, and skill names; replace them with @codepawlhq/crucible, crucible-mcp, CrucibleRenderer, and the renamed skills. Then run one existing scene, asset, and pack using their old file names before changing stored artifacts. For agent workflows, test a GLB with adjacent texture files and invoke scene_apply to confirm the asset enters the scene as expected.
Reader takeaway
Crucible 0.20.0 is not only a rename. It provides a temporary compatibility bridge while adding more explicit primitives for agent-driven 3D assembly and raising limits for the native Cycles path. Update integration names now, but do not rush to rewrite existing saved scene formats.