[ Changelog / 38 releases ]
Updates
Every shipped change on the road to v1.0. Latest release 1 Jul 2026.
38
Total Releases
15
Features
14
Improvements
9
Fixes
- —Unified perception, decision, and control stacks into a single runtime
- —Stabilized 40+ hour continuous play sessions with zero crashes
- —Locked the v1.0 model checkpoint (127M parameters) for public benchmarking
- —Reduced average time-to-headshot to 118ms in the Competitive profile
- —Tightened micro-correction jitter under 0.15° for held angles
- —Added per-weapon recoil compensation curves for the latest patch
- —Role-aware positioning for Duelist, Controller, Sentinel, and Initiator
- —Ability sequencing tuned per-agent from pro VOD review
- —Automatic agent-fit selection based on map and team composition
- —Correct CUDA memory release after every inference frame
- —Cut average VRAM footprint by 34%
- —Eliminated frame-time spikes during round transitions
- —Context-aware callouts derived from live game state
- —Reduced voice generation latency to 45ms
- —Added tone adaptation for clutch vs. eco rounds
- —Brightness-spike detection at 120 FPS
- —Flash-trajectory prediction for pre-emptive look-away
- —12ms look-away execution time
- —Monte Carlo economy forecasting across future rounds
- —Team-aware coordinated saves and force-buys
- —Weapon selection weighted by map and position
- —Target prioritization for consecutive kills
- —Per-weapon spray-pattern memory
- —Smoothed transfer arcs to stay in the recoil pocket
- —Fixed long-rotation timing calculations
- —Resolved pathfinding stalls near teleporters
- —Refreshed map geometry for the current patch
- —Mixed-precision (FP16) inference path
- —Dynamic batch sizing based on available memory
- —Reduced minimum VRAM from 8GB to 4GB
- —Real-time spectrum analysis of the game audio bus
- —Footstep direction detection at 87% accuracy
- —Ability-sound recognition across the full roster
- —Latency-compensated peek initiation
- —Jiggle-peek pattern generation for information gathering
- —Angle isolation to minimize simultaneous exposures
- —Recomputed smoke lineup coordinates
- —Corrected height offsets on elevated sites
- —Validated lineups across all competitive maps
- —Isolation-focused positioning in clutch scenarios
- —Spike-time-aware defuse and fake-defuse logic
- —Sound-baiting behavior to force enemy mistakes
- —Added Instant, Competitive, Human, and Fatigued profiles
- —Per-profile latency and error-injection tuning
- —Session-based fatigue curve simulation
- —OCR-free minimap icon detection
- —Teammate-position-aware rotation planning
- —Detected-enemy memory with decay over time
- —Stabilized the held-angle controller
- —Removed integrator wind-up in the aim loop
- —Re-anchored placement to common head-level pixels
- —Shared execute clock across agents
- —Flash-for-teammate timing windows
- —Post-plant utility conservation logic
- —Minimum-jerk mouse trajectories with noise injection
- —Counter-strafe timing tied to shot readiness
- —Bezier-smoothed path following for navigation
- —Constant-velocity and acceleration models blended per target
- —Occlusion-aware re-acquisition
- —Lead calculation for moving-target flicks
- —Guarded the state machine against null game-state frames
- —Added a watchdog to recover stalled inference threads
- —Improved logging around phase transitions
- —Crossfire setup generation with teammates
- —Time-to-defuse-aware repositioning
- —Off-angle selection weighted by enemy tendencies
- —TensorRT graph fusion for the policy head
- —SIMD-optimized preprocessing kernels
- —Zero-copy frame transfer from capture to GPU
- —Concise, standard, and hype callout modes
- —Adjustable callout frequency
- —Map-specific callout vocabulary
- —Corrected ground-target ray casting
- —Fixed cast cancellation on rapid state changes
- —Validated across all ground-targeted kits
- —Curriculum from bots to mirrored self-play
- —Reward shaping refined for objective play
- —Opponent pool sampling to avoid overfitting
- —Measured spray patterns per weapon and range
- —Automatic pull-down and horizontal compensation
- —First-shot-accuracy-aware tap/burst switching
- —Trade-distance maintenance in default setups
- —Stacked-site detection and adaptation
- —Lurk coordination timing
- —Double-buffered capture with fence synchronization
- —Aligned capture to the game's present call
- —Removed a race condition in the copy queue
- —Range-aware gain scheduling
- —Scoped-weapon sensitivity profiles
- —Smooth transitions to avoid overshoot
- —Per-decision confidence logging
- —Value-estimate overlays for debugging
- —Exportable decision traces for review
- —Bot-strafe tracking drills
- —Flick-to-target calibration pass
- —Cold-start latency reduction
- —Debounced file-watcher events
- —Atomic config swap between rounds
- —Validation errors now surface in the client log
- —Per-frame decision traces over the local API
- —Value-estimate and confidence streaming
- —Optional CSV export of session data
- —Learned uncertainty decay for unseen players
- —Spike-timer and utility-cooldown tracking
- —Belief-state fusion into the policy input tensor
- —Aspect-ratio-aware region mapping
- —Fixed misreads at 21:9 and 32:9
- —Fallback heuristics for HUD scaling
- —Prioritized sampling of hard opponents
- —Reduced policy collapse on narrow strategies
- —Faster convergence in stage-two RL
- —ResNet-50 backbone with feature-pyramid neck
- —YOLOv8-style enemy detection head
- —Structured game-state tensor output