Publish original Codex policy sandbox with pinned identity and notices

runtime/codex-policy-sandbox-4ccf7e95/Dockerfile.motion

@@ -0,0 +1,5 @@
+FROM python@sha256:528257d48c1da0dcecc2e725d1ae34498d60c965f1241e39cd6a85a8859bdf84
+RUN pip install --no-cache-dir numpy==2.2.6 mujoco==3.13.0 mink==1.3.0 quadprog==0.1.13 \
+    && pip check \
+    && pip freeze > /opt/motion-requirements.lock
+ENV OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1

runtime/codex-policy-sandbox-4ccf7e95/README.md

@@ -0,0 +1,47 @@
+# Codex policy sandbox image
+
+This optional runtime image is for the RobotAgentBench Codex example. It is
+not the Isaac Sim environment, a model checkpoint, or a simulation asset.
+
+- Platform: Linux amd64.
+- Docker image ID: `sha256:4ccf7e950d48f9b5477292bbc123b676fdbad039c6d51573d969008d8ffd5dc6`.
+- Archive: `codex-policy-sandbox-4ccf7e95.tar.gz` (144,690,151 bytes).
+- Archive SHA256: `05945dbc2e5a0e30834b75f1b7ab4cc53f3a51c682c46c9dc9087b9adb650a5d`.
+- Dependency-lock SHA256: `c74fdf44032127126f4467fad74db8e29eacb086415858243652e062fde5a822`.
+
+The archive is a tag-free `docker image save` of the original image, not a
+rebuild. Loading it preserves the image ID required by the original sandbox
+launcher. No credentials, task data, model weights or host files were added.
+Loading does not start a container; the Codex launcher applies its existing
+network, filesystem, user, CPU and memory restrictions when running policies.
+
+From a current RobotAgentBench installation, on the Linux Docker host used by
+the native worker:
+
+```sh
+python -m robotagentbench.model_examples.codex.sandbox_image --cache downloads --download
+python -m robotagentbench.model_examples.codex.sandbox_image --cache downloads \
+  --load --docker-host unix:///var/run/docker.sock
+```
+
+For offline setup, copy the archive into the cache and run only the second
+command. The installer verifies its length and SHA256 before Docker loading,
+then checks the image ID and platform. It never substitutes a rebuilt image.
+
+## Build inputs and third-party terms
+
+The original Dockerfile and exact Python dependency inventory are published
+alongside the archive. The base image is the upstream Docker Official Python
+image pinned at `python@sha256:528257d48c1da0dcecc2e725d1ae34498d60c965f1241e39cd6a85a8859bdf84`.
+It uses Python 3.11.16 and Debian bookworm. See the
+[official Python image source and license information](https://github.com/docker-library/python#license)
+and [Python source release](https://www.python.org/downloads/release/python-31116/).
+
+The image preserves Debian notices under `/usr/share/doc`, Python's license,
+and installed distributions' license/notice files under `site-packages`.
+Its dependency inventory includes NumPy, SciPy, MuJoCo, Mink, Quadprog and their
+dependencies. Their original licenses apply; **this runtime archive is not
+relicensed under the asset dataset's CC BY 4.0 license or the evaluator's
+Apache-2.0 license**. Consult the included notices and upstream package sources
+for each component. A rebuild is available for inspection but is not assumed
+to reproduce the original Docker image ID.

runtime/codex-policy-sandbox-4ccf7e95/codex-policy-sandbox-4ccf7e95.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05945dbc2e5a0e30834b75f1b7ab4cc53f3a51c682c46c9dc9087b9adb650a5d
+size 144690151

runtime/codex-policy-sandbox-4ccf7e95/requirements.lock

@@ -0,0 +1,15 @@
+absl-py==2.5.1
+daqp==0.10.3
+etils==1.14.0
+fsspec==2026.9.0
+glfw==2.11.0
+mink==1.3.0
+mujoco==3.13.0
+numpy==2.2.6
+packaging==26.3
+PyOpenGL==3.1.10
+qpsolvers==4.13.0
+quadprog==0.1.13
+scipy==1.17.1
+typing_extensions==4.16.0
+zipp==4.1.1