Claude Code 2.1.329 has been released.
130 CLI changes, 1 system prompt change
Highlights:
• Removed risky-action confirmations; gated steps no longer prompt and now run automatically
• Session /tasks shows scheduled tasks, /loop wakeups, goal+forks; 'x' prompts before stopping one task, safer
Complete details in thread ↓
@ClaudeCodeLog most tasks saved on older versions need to be rescheduled in a new signed-in session. they still run if you resume the original session.
this is another reason why we're moving to get 100% of our inference done on our own GPUs
so many of the gateway products like to talk about enterprise and security and privacy but they don't actually have control over anything
Gemini 4 Neon Fighter Jets Game
→ This is 1 shot in Three.js
→ single .html file with Zero external assets
→ Made in around 30 minutes from a basic idea
Claude is officially OUT of my workflow!
I canceled one sub a while ago, and it looks like I'm going to say goodbye to the second one too. Gemini 4 Neon is doing a really good job.
But even though it's this good, I still don't trust any single model. For now, Gemini 4 Neon is the planner, but it consults with my local models and goes back and forth until they reach an agreement.
Here's how it works:
- Gemini writes the plan.
- Then it launches 3 subagents in parallel to ask other models for reviews, fixes, and additions.
- Each subagent talks to one model:
• GLM 5.3 Flash (running locally)
• Qwen3.8-Flash-Next (running locally)
• DeepSeek V4.1 Flash (running locally)
When they're done, Neon writes a merged report combining everything. It also verifies every claim the other models make.
In most cases, the local models find issues. Each model has its own blind spots. Remember, no model is perfect, no matter which one it is. It's an absolute must to consult different models before executing a plan.
Once the final plan is set, the planner (Neon in this case) manages multiple workers to execute it. Right now I'm using GLM 5.3 Flash (locally) for all execution.
I'm also going to test Gemini 4 Hydrogen and see how it performs.
Eclipse Village
Antigravity
Gemini Neon High /Swarm
75.9 million boxes, placed and lit in one browser tab.
76 million voxel cells in the lighting grid, 8 cm each, with 30.4 million of them filled
1,064 light sources (lanterns and window lights) that switch on as the eclipse shadow arrives.
100 districts, each written by its own AI agent.
15.1 MB of code, about 332,000 lines, in a single HTML file.
2,286 model calls across 109 agent conversations.
183 million input tokens (36M fresh, 147M from cache).
6.7 million output tokens, of which 2.5 million are thinking.
Ray tracing in plain WebGL2. The village is copied into a 3D voxel texture and light is traced through it in a shader. That gives real shadows and bounce light with no WebGPU and no game engine.
The eclipse lights the scene itself. The moon's shadow is a moving ellipse that sweeps across the village. As it passes, 1,064 lights switch on house by house.
A film camera in the browser. Depth of field and motion blur are shot in the style of a miniature film: low lens, slow continuous moves.
100 agents, one village. Each agent writes one district to a shared contract, and the districts fit together with no overlaps.