GhostReel v0.10.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Script-chat composer shows the tool-call budget beside the brain selectors, so the research
  limit settable in Settings is adjustable directly in chat.

• Every new script-chat answer records its actual model, token count, generation speed, and
  total time. Server rates include request time when separate decode timing is unavailable.

• AgentWorld 35B-A3B GGUF model discovery/download and a RAM-expert slider for hybrid inference.

Improved

• Local script agents allow 30 research tool calls by default, count planning replies separately,
  and inspect proposed shot and quote ranges before evidence-based draft review. Incomplete
  coverage is reported explicitly, and Settings exposes the research limit.

• Hybrid MoE inference uses a resident CPU thread pool and resident host expert buffers.
  On RTX 4070, AgentWorld Q4_K_XL measured 55.3 decode tokens/s at an allocated 128k context
  with 32 of 40 expert blocks in RAM (short-prompt benchmark, not a full-context throughput claim).

• Project workspaces use the available window width; the library gives filenames room and scrolls
  horizontally in small windows instead of squeezing metadata columns.

• Local llama.cpp inference fits GPU layers and MoE experts to available memory, supports explicit
  expert-block offload, and exposes flash attention in script Settings.

• Script chat budgets against the actual loaded context, reserves output with tokenizer counts,
  and compacts old exchanges. Text inference reuses prompt prefixes across tool rounds, including
  Qwen3.5 recurrent state via bounded checkpoints.

• Local contexts respect the model's trained limit, retry smaller allocations, and fall back to
  f16 KV storage when automatic quantized attention cannot initialize.

• Opt-in NVIDIA hardware checks cover flash attention, 32k contexts, automatic and explicit MoE
  placement, real GPU memory use, and structured script generation; validated on an RTX 4070.

Fixed

• Local agent tool calls enforce the shared argument schemas, including numeric video IDs,
  instead of accepting arbitrary objects.
• Preserve GPU projector and KV-cache offload when llama.cpp auto-fit returns its full-offload
  sentinel. Explicit MoE placement also matches fused gate/up expert tensors.