[linj-glitch]

What does this PR do?

anyswe_agent runs its harness from a portable runtime bundle (CPython, NeMo Gym, Node and the CLI) inside the task image. Two leaks between that bundle and the task:

1. The agent's python, pip and node were the bundle's. agent_runner.py prepends the whole bundle bin/ to PATH, and that directory also holds python3, pip, node and npm. Every shell command the coding agent runs therefore resolves python / pip / node to the bundle instead of the task image. Agents working on numpy, pandas or Pillow tasks then import, "fix" or reinstall the bundle's copies instead of the task's. When the bundle is shared, for example a baked or bind-mounted runtime, one rollout's writes break every later sandbox.

For the CLI harnesses (claude, cline, opencode, openclaw, pi), the runner now symlinks only the harness executable into a private temp directory and puts that first on PATH. If the executable is a node script (#!/usr/bin/env node, as npm-installed CLIs are), node is linked too so the CLI still starts. Python harnesses (Hermes, NeMo Fabric) and anything unknown keep the whole bundle bin/ as before.

2. The task image's Python environment could break the runner. The runner was started with plain <runtime>/bin/python. If the task image sets PYTHONPATH (or has a user site) pointing at a checkout of numpy or pandas, the runner imports that source tree instead of the runtime's wheel and dies before its first model call ("Unable to import required dependency numpy"). It now starts with python -I. The runner already adds /nemo_gym_mount to sys.path itself, and child processes the agent spawns still get the normal environment.

The README documents both.

Validation

• python -m pytest -q responses_api_agents/anyswe_agent/tests: 24 passed.
• New TestRuntimeIsolation in tests/test_runtime_isolation.py:
  • test_unknown_harness_keeps_the_whole_bundle_bin
  • test_native_cli_hides_the_bundle_python_pip_and_node: the shim holds only claude.
  • test_node_script_cli_also_gets_node: the shim holds claude and node.
  • test_runner_python_ignores_the_task_images_python_environment: drives _run_agent_in_sandbox with a fake sandbox and checks the runner command uses python -I.
• pre-commit run --files <touched files>: all hooks pass.

Rollouts

Both problems showed up in internal GRPO training of Claude Code on R2E-Gym / SWE task images. With the shared runtime on a cluster filesystem, an agent overwrote a file inside the bundle's numpy, and every later rollout of that run failed and was masked. In numpy/pandas images the runner failed at import. With these changes the training ran several thousand rollouts without either failure. I have no public rollout numbers to share.

Compatibility and benchmark impact

• With the CLI harnesses, the agent's shell now sees the task image's python / pip / node rather than the bundle's. Tasks that need Python or Node must get them from the image, which is the intended environment. If an image has no python at all, the agent no longer silently gets CPython 3.13 from the bundle.
• The Hermes and NeMo Fabric harnesses are unchanged.
• The runner ignores PYTHON* environment variables of the image.

Checklist

• I have read the contributing guidelines.
• The change is focused; unrelated "drive-by" edits are tracked as separate issues/PRs.
• Tests added or updated and pass locally, or N/A for docs-only / non-code changes (so CI unit/server checks pass when applicable).
• Documentation added or updated, or N/A with justification.
• Pre-commit checks pass locally (pre-commit run --all-files) (so CI lint/format checks pass). (Run on the touched files.)
• New source files include the required Apache-2.0 SPDX header. (N/A: no new files.)
• All commits have DCO sign-off (git commit -s) (so the DCO check passes).

🤖 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.

[linj-glitch]

/ok to test ff5daeb

[alexandery-nvidia]

I traced the -I interplay before approving: isolated mode drops the script directory, PYTHONPATH, and the user site, which is safe here because stdlib and fastapi resolve from the bundle interpreter's own site-packages (the deps scripts install nemo_gym with dependencies into the bundle), and the deferred main() imports resolve through the explicit /nemo_gym_mount insert, which also shadows the bundle's own nemo_gym copy; nothing imports a sibling of the script. The shim fallback (unknown harness keeps the whole bin/) preserves current behavior for out-of-catalog agents, and the node-shebang detection covers the CLIs that are node scripts. Ran tests/test_runtime_isolation.py (4 passed) and tests/test_app.py (20 passed) at the PR head locally.

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/38327143475