cave-harness 0.1.0

Code Agent Virtualization Environment - Virtualize any terminal-based code agent

CAVE — Code Agent Virtualization Environment

Package: cave-harness

CAVE virtualizes any terminal-based code agent (e.g. Claude Code) behind an HTTP server, so it can be driven, observed, and orchestrated as a service instead of an interactive terminal. It is the agent runtime that PromptWorld and the other *Worlds are built on.

What it gives you

• An HTTP server around a code agent — start a long-lived agent process and talk to it over HTTP instead of a TTY: inject prompts, stream output, and run it inside a loop.
• A hook harness — a registry + dispatch layer for Claude Code hooks (session-start / per-turn / stop), so behavior can be layered onto the agent from outside.
• Loop infrastructure — the autopoiesis / guru loops that keep an agent working across turns.

Install

pip install cave-harness

Optional extras:

pip install "cave-harness[sdna]"   # + sanctuary-dna integration
pip install "cave-harness[dev]"    # + pytest / pytest-asyncio

Run the server

cave-server

This launches the HTTP server (cave.server.http_server:run_server, FastAPI + uvicorn). Point your client at it to register hooks, inject prompts, and drive the virtualized agent.

Where it sits

CAVE is the runtime layer of the SANCREV / GNOSYS stack: the *Worlds (PromptWorld, …) compile their agents and run them on top of the CAVE harness. See the ecosystem docs for the full picture.

Docs

Full documentation and the ecosystem map: https://sancovp.github.io/aisaac/

License

GPBL-1.0 — see LICENSE.