Skip to main content

​
Changelog - February 2026

Comprehensive documentation of all major changes pushed to main in February 2026.

​
🎨 Rendering & Performance

​
Instanced Rendering System (PR #946)

Impact: Massive performance improvement for resource rendering Changes:
  • Implemented GLBResourceInstancer for rocks, ores, herbs
  • Extended GLBTreeInstancer with depleted model support
  • Added InstancedModelVisualStrategy as default for GLB resources
  • Reduced draw calls from O(n) to O(1) per unique model
Performance:
  • 1000 trees: 1000 draw calls → 3 draw calls (LOD0/LOD1/LOD2)
  • Instanced depletion (stumps) for zero-cost state changes
  • Automatic LOD switching with hysteresis
API Changes:
  • ResourceVisualStrategy.onDepleted() now returns Promise<boolean>
  • Added optional getHighlightMesh() method for instanced entities
  • EntityHighlightService supports instanced hover outlines
Documentation:
  • docs/instanced-rendering.md - Complete system documentation
  • docs/api/resource-visual-strategy.md - API reference

​
🖥️ WebGPU Enforcement (Commit 47782ed)

Impact: Removed all WebGL fallback code - WebGPU is now strictly required Changes:
  • Removed isWebGLForced, isWebGLFallbackForced, isWebGLFallbackAllowed
  • Removed isWebGLAvailable, isOffscreenCanvasAvailable, canTransferCanvas
  • Changed UniversalRenderer to WebGPURenderer throughout
  • RendererBackend is now only "webgpu"
  • Removed --disable-webgpu and forceWebGL flags from streaming
Rationale:
  • All materials use TSL (Three Shading Language) which ONLY works with WebGPU
  • Post-processing effects use TSL-based node materials
  • No WebGL fallback path exists
Documentation:
  • Updated AGENTS.md with WebGPU requirements
  • Updated CLAUDE.md with browser requirements
  • Created docs/vast-ai-streaming.md for GPU streaming architecture

​
🎮 Vast.ai GPU Streaming

Impact: Production-ready GPU streaming to Twitch, Kick, X/Twitter

​
GPU Rendering Modes

Xorg with NVIDIA (preferred):
  • Requires DRI/DRM device access
  • Best performance, lowest latency
  • Falls back to Xvfb if DRI unavailable
Xvfb with NVIDIA Vulkan (fallback):
  • Virtual framebuffer + GPU rendering via ANGLE/Vulkan
  • Works in containers without DRM access
  • CDP captures from Chrome’s GPU compositor
Headless mode: NOT SUPPORTED (WebGPU requires display)

​
Audio Capture (Commits 3b6f1ee, aab66b0)

Added:
  • PulseAudio with chrome_audio virtual sink
  • FFmpeg captures from PulseAudio monitor
  • Configurable via STREAM_AUDIO_ENABLED
Configuration: 
STREAM_AUDIO_ENABLED=true
PULSE_AUDIO_DEVICE=chrome_audio.monitor
PULSE_SERVER=unix:/tmp/pulse-runtime/pulse/native

​
RTMP Multi-Streaming (Commits 5dbd239, d66d13a)

Added:
  • Simultaneous streaming to Twitch, Kick, X/Twitter
  • FFmpeg tee muxer for single-encode multi-output
  • YouTube explicitly disabled (set YOUTUBE_STREAM_KEY="")
Fixed:
  • Kick RTMP URL: rtmps://fa723fc1b171.global-contribute.live-video.net/app
  • Added KICK_STREAM_KEY and KICK_RTMP_URL to CI/CD secrets

​
Stream Quality Improvements (Commit 4c630f1)

Changed:
  • Default x264 tune: zerolatency → film (better compression)
  • Buffer multiplier: 2x → 4x bitrate (18000k for 4500k bitrate)
  • Added FLV flags for RTMP stability
  • Improved input buffering with thread_queue_size
Configuration: 
# Restore old low-latency behavior
STREAM_LOW_LATENCY=true

# Adjust buffer size
STREAM_BITRATE_BUFFER_MULTIPLIER=4

​
Deployment Validation (Commits 47782ed, 30bdaf0)

Added:
  • GPU validation: Fails if nvidia-smi fails
  • Vulkan validation: Checks ICD availability
  • Display validation: Ensures Xorg/Xvfb is accessible
  • No soft fallbacks: Deployment FAILS if WebGPU cannot initialize
Documentation:
  • docs/vast-ai-streaming.md - Complete architecture guide
  • scripts/deploy-vast.sh - Deployment script with inline docs

​
🤖 AI Agent Improvements (Commit 60a03f4)

Impact: 67% reduction in LLM API calls, 70% faster decision-making

​
Action Locks

Added:
  • Action lock prevents LLM ticks during movement
  • waitForMovementComplete() for synchronous actions
  • isMoving tracking in HyperscapeService
Benefits:
  • Reduces LLM calls by ~70%
  • Prevents decision conflicts
  • Improves responsiveness

​
Fast-Tick Mode

Added:
  • 2s interval after movement/goal changes
  • 3 fast ticks, then returns to 10s interval
Triggers:
  • Movement completed
  • Goal changed
  • Resource depleted
  • Banking completed

​
Short-Circuit LLM

Added:
  • Skip LLM for obvious decisions
  • Repeat resource gathering
  • Banking when inventory full
  • Goal execution immediately after setting
Benefits:
  • Reduces LLM calls by ~40%
  • Faster decision-making
  • More predictable behavior

​
Banking Improvements

Changed:
  • Banking actions now await movement completion
  • New banking goal type auto-restores previous goal
  • Filter depleted resources from nearby entity checks
Configuration: 
# Resource approach range (increased from 20 to 40)
RESOURCE_APPROACH_RANGE=40
Documentation:
  • docs/ai-agent-improvements.md - Complete guide

​
🔒 Security Updates

​
CSP Changes (Commits e012ed2, 1b2e230, 8626299)

Added:
  • fonts.googleapis.com to style-src (Google Fonts)
  • fonts.gstatic.com to font-src (Google Fonts)
  • static.cloudflareinsights.com to script-src (Analytics)
  • data: to script-src (WASM loading)
Removed:
  • report-uri directive (broken endpoint)
Documentation:
  • docs/security/csp-updates.md - CSP configuration guide

​
Deployment Secrets (Commits b466233, 4a6aaaf, b754d5a)

Added:
  • SOLANA_DEPLOYER_PRIVATE_KEY to secrets file
  • JWT_SECRET to secrets file
  • ARENA_EXTERNAL_BET_WRITE_KEY to secrets file
  • KICK_STREAM_KEY and KICK_RTMP_URL to CI/CD
Fixed:
  • Secrets now written to /tmp to survive git reset
  • PM2 uses pm2 kill instead of pm2 delete to pick up new env vars
  • YouTube explicitly disabled to prevent stale keys

​
🎯 Solana Markets

​
WSOL as Default (Commit 34255ee)

Changed:
  • GOLD_MINT → MARKET_MINT (defaults to WSOL)
  • Markets use native token of each chain
  • Perps oracle disabled by default (ENABLE_PERPS_ORACLE=false)
Configuration: 
# Use WSOL (default)
MARKET_MINT=So11111111111111111111111111111111111111112

# Or use GOLD
MARKET_MINT=DK9nBUMfdu4XprPRWeh8f6KnQiGWD8Z4xz3yzs9gpump
Documentation:
  • docs/solana-market-updates.md - Migration guide

​
🐛 Bug Fixes

​
Client Rendering (Commit 205f964)

Fixed:
  • Updated WebGL references to WebGPU in comments
  • SettingsPanel.tsx always shows ‘WebGPU’ instead of conditional
  • errorCodes.ts updated to mention WebGPU
  • visualTesting.ts uses 2D canvas for pixel reading (WebGPU compatible)

​
Deployment Stability (Commits 8575215, 1025071, 77403a2)

Fixed:
  • Clean up X server sockets before starting Xvfb
  • Remove /tmp/.X11-unix directory and recreate
  • Kill all X processes (Xorg, Xvfb, X) before restart
  • Added STREAM_CAPTURE_USE_EGL variable (was missing)
  • Added STREAM_CAPTURE_EXECUTABLE for custom browser path

​
PulseAudio Permissions (Commit aab66b0)

Fixed:
  • Add root user to pulse-access group
  • Create /run/pulse with proper permissions (777)
  • Export PULSE_SERVER in both deploy script and PM2 config
  • Add pactl check before using PulseAudio

​
📊 Performance Metrics

​
Instanced Rendering

  • Draw calls: 1000 trees = 3 draw calls (99.7% reduction)
  • Memory: Shared geometry/materials (1x cost vs. Nx cost)
  • LOD switching: O(n) per frame with early-exit optimization

​
AI Agents

  • LLM calls: 6/min → 2/min (67% reduction)
  • API cost: 0.02/hr→0.02/hr → 0.007/hr (65% reduction)
  • Decision latency: 10s → 3s (70% improvement)

​
Streaming

  • Encoding: Single H.264 encode for all platforms
  • Latency: Twitch ~12s, Kick ~15s, X ~18s
  • Quality: 1280x720 @ 30fps, 4500kbps bitrate

​
🔄 Breaking Changes

​
ResourceVisualStrategy API

Before: 
async onDepleted(ctx: ResourceVisualContext): Promise<void>
After: 
async onDepleted(ctx: ResourceVisualContext): Promise<boolean>
Migration: All custom visual strategies must update onDepleted() to return boolean.

​
WebGPU Requirement

Before:
  • WebGL fallback available
  • forceWebGL parameter supported
  • Headless mode worked (software rendering)
After:
  • WebGPU strictly required
  • No WebGL fallback
  • Headless mode NOT supported
  • Deployment fails if WebGPU unavailable
Migration:
  • Ensure browser supports WebGPU (Chrome 113+, Edge 113+, Safari 18+)
  • Vast.ai deployments must have NVIDIA GPU with Vulkan
  • Update CI/CD to validate WebGPU availability

​
📝 Documentation Added

​
New Files

  • docs/instanced-rendering.md - Instanced rendering system guide
  • docs/vast-ai-streaming.md - GPU streaming architecture
  • docs/ai-agent-improvements.md - AI agent optimization guide
  • docs/api/resource-visual-strategy.md - API reference
  • docs/security/csp-updates.md - CSP configuration guide
  • docs/solana-market-updates.md - Market token migration guide
  • docs/CHANGELOG-2026-02.md - This file

​
Updated Files

  • AGENTS.md - Added Vast.ai deployment architecture
  • CLAUDE.md - Updated browser requirements (Safari 18+)
  • README.md - Added recent updates section, WebGPU requirement
  • packages/server/.env.example - Added GPU streaming variables
  • packages/client/.env.example - Updated streaming canonical platform

​
🔗 Related Pull Requests

  • #946 - Instanced rendering for GLB resources
  • #948 - Instanced highlight depleted fix
  • #945 - Model agent stability audit

​
📚 Additional Resources