Hollow

v5.4.0

Three local LLMs with persistent psychological states running on your hardware. See the README for what actually happened when we let them run overnight.

What's new

For agents

• invoke_claude — agents submit implementation requests to a human with system write access. They design, you decide whether to build.
• check_claude_status — agents check whether their requests were fulfilled.
• self_evaluate — agents assess their own work against actual file evidence using their own model.
• /agentOS/design/ — writable space for agent specs and architectural proposals.
• Agents now receive accurate world context in every existence prompt (what environment they're actually in, what capabilities actually do). Previously they invented hardware that doesn't exist.
• Futility stressor now checks whether deployed tools actually get called in subsequent goals, not just whether goals complete.
• synthesize_capability: 90s dedup guard stops spin loops. Auto-test on deployment surfaces broken code immediately.

Bug fixes

• autonomy_loop: shell commands that exit 0 with no stdout now count as successful artifacts.
• fs_write: non-string content auto-coerced to JSON instead of 422.

For users

• stop.bat — one-click shutdown. Stops containers, clears VRAM. Run launch.bat to resume. State preserved.
• GPU block in docker-compose.yml documented as optional.
• config.example.json updated to qwen3.5:9b.

Setup

Download the ZIP, extract anywhere, double-click install.bat. Everything else is automatic.