Hollow

What's in this release

For users

• thoughts.py is now the default monitor — cleaner terminal output, no external dependencies. Just Python. Double-click install.bat or launch.bat and it opens automatically.
• stop.bat shuts everything down and clears VRAM. launch.bat brings it back. Agent memory and state survive.
• One-click Windows install handles Docker, Ollama, model downloads, and opens the monitor.

Agent improvements

• self_evaluate now works — agents can call self_evaluate() to assess their own recent work against real file evidence. Was broken in 5.4.0 due to a missing import; fixed.
• Monitor shows more — action params display up to 120 chars (was 70), results up to 160 chars (was 80). You see what agents are actually doing instead of truncated previews.
• invoke_claude — agents submit implementation requests to a human with system write access. They write the spec, you decide what to build.
• synthesize_capability dedup — if an agent deploys the same tool twice within 90 seconds, it gets "already deployed" instead of silently overwriting.
• World context grounding — agents now see factual information about their environment (Docker container, no hardware layer, actual file paths) in every existence loop. Eliminates drift into fictional frameworks.
• Futility stressor now tracks whether deployed tools actually get called in subsequent goals, not just whether goals complete.

Setup

Download the ZIP, extract anywhere, double-click install.bat. GPU optional — works on CPU at ~40s/planning call vs ~6s with NVIDIA.

Repo: https://github.com/ninjahawk/hollow-agentOS