BMASS v0.7.0 — Native Runtime
This is a major experimental architecture update.
BMASS now uses a native C++ runtime built with the llama.cpp libraries, removing much of the earlier Python orchestration and moving the model closer to the operating-system layer.
Major changes
- Added native
bmassC++ runtime - Added controlled shell executor
- Added model → shell → Linux → observation → model loop
- Shell output is returned to the model as new operational information
- Added external system-prompt loading
- Reduced dependence on Python orchestration
- Added updated installation instructions
- Added compiled experimental BMASS executable
- Added demonstration screenshot
Architecture
User request
↓
Local model
↓
<shell>command</shell>
↓
BMASS shell executor
↓
Linux
↓
Real command output
↓
Local model
↓
Corrected action or final response
The model can now observe real system state, respond to command failures and issue a revised command within the same operation.
Reference environment
Development and testing used:
Alpine Linux
Qwen3 0.6B GGUF
Low-end Intel Celeron laptop
4 GB RAM
Local llama.cpp build
Offline-first operation
Installation warning
Paths in the supplied source and scripts reflect the reference BMASS installation.
Please inspect and adjust:
model path
llama.cpp source and build paths
BMASS executable path
shared-library path
system-prompt path
shell-executor path
startup and login configuration
Please Do not copy the executable blindly between incompatible systems. Recompiling from source is recommended.
Experimental status
BMASS remains an experimental research project.
The shell interface should be tested in a controlled environment.
Do not provide unrestricted privileges, sensitive data or network access without understanding the consequences.
Current limitations include:
small-model command errors
verbose reasoning output
limited context length
incomplete context management
evolving shell restrictions
architecture-specific compiled binaries