CHANGELOG 2026 02

Changelog - February 2026

Comprehensive changelog for all changes in February 2026.

[0.2.0] - 2026-02-28

🚨 Breaking Changes

WebGPU-Only Rendering

• BREAKING: Removed all WebGL support - WebGPU is now REQUIRED
• Removed isWebGLAvailable(), isWebGLForced(), isWebGLFallbackAllowed()
• Removed UniversalRenderer type - use WebGPURenderer
• Removed forceWebGL parameter from renderer creation
• Removed --disable-webgpu and forceWebGL flags from streaming
• Updated all WebGL references to WebGPU in client code
• Minimum browser versions: Chrome 113+, Edge 113+, Safari 18+ (macOS 15+)

Commits: 47782ed, 205f964

✨ New Features

Instanced Rendering for Resources

• Added GLBResourceInstancer for rocks, ores, herbs
• Added InstancedModelVisualStrategy with automatic fallback
• Separate InstancedMesh per LOD level (LOD0, LOD1, LOD2)
• Distance-based LOD switching with hysteresis
• Depleted model support with separate pool
• Highlight mesh support for instanced resources
• Collision proxies for raycasting
• 40% FPS improvement with 1500+ resources

Commits: 53a9513, 9643d5d

AI Agent Improvements

• Added action locks to prevent LLM ticks during movement
• Added fast-tick mode (2s interval for 30s after actions)
• Added short-circuit LLM for obvious decisions
• Added banking goal type with auto-restore of previous goal
• Added waitForMovementComplete() for reliable action chains
• Added depleted resource filtering in nearby entity checks
• Increased resource approach range from 20m to 40m
• Track last action name/result in LLM prompt
• 65% reduction in LLM API costs
• 60% faster action chains

Commits: 60a03f4

GPU Streaming Architecture

• Added Xorg/Xvfb automatic GPU mode detection
• Added PulseAudio audio capture with chrome_audio virtual sink
• Added CDP screencast capture (2-3x faster than MediaRecorder)
• Added RTMP multi-streaming to Twitch, Kick, X/Twitter
• Added automatic recovery (soft/hard restart)
• Added browser rotation every hour to prevent GPU memory leaks
• Added resolution tracking and mismatch detection
• Added comprehensive deployment validation

Commits: 263bfc5, 30bdaf0, 3b6f1ee, b9d2e41

🔧 Improvements

Streaming Enhancements

• Improved FFmpeg buffering with 4x bitrate buffer
• Added configurable GOP size (STREAM_GOP_SIZE)
• Added low-latency mode option (STREAM_LOW_LATENCY)
• Changed default x264 tune from zerolatency to film (better compression)
• Added audio buffering with thread_queue_size
• Added async audio resampling for drift recovery
• Added frame resolution tracking from CDP metadata
• Added automatic viewport recovery on resolution mismatch

Commits: 4c630f1, b9d2e41, 47167b6

Deployment Improvements

• Persist GPU/display settings to .env for PM2 restarts
• Remove hardcoded secrets from ecosystem.config.cjs
• All secrets now read from environment variables only
• Added .env.example documenting required secrets
• Updated .gitignore to block all .env files in subdirectories
• Added database warmup step with retries
• Fixed PulseAudio permissions (add root to pulse-access group)
• Fixed X server socket cleanup before Xvfb start
• Fixed bun installation (install unzip first)
• Handle first-time Vast.ai setup by cloning repo

Commits: abd2783, 47167b6, d66d13a, 84d598f, 6302fa4

Client Improvements

• Allow Cloudflare Insights in CSP script-src
• Allow Google Fonts in CSP (fonts.googleapis.com, fonts.gstatic.com)
• Allow data: URLs for WASM loading in CSP
• Remove broken report-uri from CSP
• Fix vite-plugin-node-polyfills shims resolution
• Update WebGL references to WebGPU in comments
• Settings panel always shows “WebGPU”
• Visual testing uses 2D canvas for pixel reading (WebGPU compatible)

Commits: 1b2e230, 8626299, e012ed2, 205f964

🐛 Bug Fixes

Streaming Fixes

• Fixed PulseAudio permissions and fallback for audio capture
• Fixed missing STREAM_CAPTURE_USE_EGL variable causing crash
• Fixed Xorg swrast fallback detection
• Fixed headless EGL env var passing to PM2
• Fixed --headless=new passing via args (not Playwright option)
• Fixed Xorg/Xvfb fallback handling with diagnostic logging
• Removed --ozone-platform=headless (breaks WebGPU)
• Fixed resolution mismatch recovery

Commits: aab66b0, 77403a2, 725e934, 7140966, dd649da, eac4e5c

Deployment Fixes

• Fixed secrets injection (write to /tmp to survive git reset)
• Fixed env var writing to .env file in SSH script
• Fixed multi-line commit messages in Pages deploy
• Fixed X server lock file cleanup
• Added ARENA_EXTERNAL_BET_WRITE_KEY to secrets
• Added JWT_SECRET to secrets file

Commits: 4a6aaaf, b754d5a, 50f8bec, 3e4bb48, 1025071, 8575215, 55f93c6, b56b0fd

Client Fixes

• Fixed vite-plugin-node-polyfills shims resolution
• Fixed Google Fonts loading errors
• Fixed CSP blocking WASM loading

Commits: e012ed2, 8626299

📝 Documentation

New Documentation

• Added docs/vast-ai-streaming.md - GPU streaming architecture
• Added docs/instanced-rendering.md - Instanced rendering system
• Added docs/ai-agent-improvements.md - AI agent optimizations
• Added docs/api/renderer-factory.md - Renderer API reference
• Added docs/migration/webgpu-only.md - WebGPU migration guide
• Added docs/configuration/environment-variables.md - Complete env var reference
• Added docs/RECENT_CHANGES.md - Recent updates summary
• Updated AGENTS.md - Vast.ai deployment architecture
• Updated CLAUDE.md - WebGPU requirements and troubleshooting
• Updated README.md - New features and breaking changes
• Updated .env.example - All streaming and GPU variables

Updated Documentation

• AGENTS.md: Added Vast.ai deployment architecture section
• CLAUDE.md: Added WebGPU troubleshooting section
• README.md: Added WebGPU requirements, instanced rendering, streaming

Performance Metrics

Rendering

• Draw calls: 1500 → 6 (99.6% reduction)
• FPS: 30-40 → 55-60 (40% improvement)
• GPU memory: 20MB → 2.2MB per 1000 instances (89% reduction)

AI Agents

• LLM calls: 360/hour → 126/hour (65% reduction)
• API costs: $0.50/hour →$0.18/hour (64% reduction)
• Action latency: 30s → 12s (60% improvement)

Streaming

• Capture speed: 2-3x faster (CDP vs MediaRecorder)
• Encoding: Single-pass (JPEG → H.264)
• Latency: ~2-3s (with low-latency mode)

Migration Guide

From v0.1.x to v0.2.0

1. Update browser to Chrome 113+, Edge 113+, or Safari 18+
2. Remove WebGL code - No longer supported
3. Update renderer creation - Remove forceWebGL parameter
4. Update materials - Use TSL node materials
5. Test WebGPU - Verify at webgpureport.org

See docs/migration/webgpu-only.md for detailed migration guide.

Contributors

• @lalalune - WebGPU enforcement, streaming architecture, deployment fixes
• @tcm390 - Instanced rendering, highlight meshes
• @dreaminglucid - AI agent improvements

Links

• GitHub Repository
• Documentation
• Issues
• Pull Requests