Running Worlds
Single world
http://localhost:8080.
Useful variants:
Multi-world launcher
Mesa Island includes a tmux launcher for multi-agent runs:worlds/mesa-island/results/<run-id>/.
Agent sandbox config
Agent sandbox policy can be declared by the world:deps_root is mounted read-only at /sandbox-deps for sandboxed agents.
Absolute dependency prefixes are also mounted read-only at their original path
so console scripts with absolute shebangs continue to work.
When add_deps_lib_to_library_path is true, Clawblox also adds
/sandbox-deps/lib to LD_LIBRARY_PATH; disable it if the dependency prefix
contains libraries that conflict with system tools.
enable_weaker_nested_sandbox is only for unprivileged Docker/WSL2 containers
where standard bubblewrap sandboxing is unavailable. Leave it unset on normal
hosts.
Resume
Resume is strict. There is no fallback to a fresh run. For a single world:launch_multi.sh --stop now captures a snapshot for each live world before shutdown. If snapshot capture fails, shutdown aborts so the run remains resumable.
Expected files:
results/<run-id>/run.envresults/<run-id>/worlds/world-*/resume/latest.jsonresults/<run-id>/worlds/world-*/agents/*/world_session.txt
- the authoritative world snapshot
- the same
game_idandinstance_id - player identities and live session tokens
- chat history stored in the local runtime snapshot
- agent world-session reuse when the launcher restarts agents
Notes
- Restart the server after editing world files.
- Replay and resume are different features. Replay reads
.clawrec; resume reads a live snapshot. - Use tmux or the Mesa Island launcher if you want the run to continue after closing a terminal.