Add README for vibench_agent

[deepbuilder]

vibench_agent shipped without a README.md in #2674. ng_test_all filters testable components to directories containing one, then asserts the filtered count matches the discovered count — so discovery found 174 components and tested 173, failing that assertion in all eight server-suite shards even though every selected module test passed:

AssertionError: Mismatch on the number of total modules found (174) and the number of actual modules tested (173)!
Extra candidate paths:
- responses_api_agents/vibench_agent

nemo_gym/cli/env.py:1114, on main at a2c5222.

This adds the missing README. Docs-only — no code or config changes.

Contents

• The build → harvest → verify flow, and that everything about driving OpenCode is inherited from opencode_sandboxed_agent (only sandbox acquisition and harvesting are overridden).
• Why the agent owns its sandbox instead of sharing one: reaching into the agent's box needs serialize()/connect(), which only OpenSandbox implements, so the swebench shape cannot run on Docker, Apptainer or enroot. See Sandbox server to share sandbox states needed for rollouts #2082.
• Why configs/docker.yaml must be used instead of the stock provider config — the stock 180s exec timeout kills long installs, and the harness is told the model is at http://127.0.0.1:<port>, which inside a bridged container is the container itself, so it makes zero LLM calls and exports an empty app.
• An explicit warning not to "fix" that with network: host, and the 0.0.0.0 exposure caveat for the policy-model bind.

Setup, task-row generation, reward definition and grading are not duplicated here; they stay in the resources server README.

Note on the other CI failure

The Test Wheel Use failure in that run is independent of this change: the wheel test enables prereleases/unsafe-best-match, unbounded mlflow>=3.15.1 resolved to 3.15.2, which selected sqlalchemy 2.1.0rc1, whose source metadata uv rejects (duplicate normalized extra name 'mssql-pymssql'), aborting arc_agi_resources_server startup. That needs a SQLAlchemy pin below 2.1 or prerelease containment in transitive resolution — happy to open a separate issue or PR for it if useful.

🤖 Generated with Claude Code

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[cmunley1]

/ok to test 392047f