[anwithk]

Summary

Encodes the NeMo Gym Code Component Ownership map in .github/CODEOWNERS so GitHub auto-requests the owning team on every PR, instead of leaving review assignment to whoever happens to notice. Today the file only covers CI paths, so everything else lands with no owner.

The split follows the doc's principle: Core owns the contracts and the substrate, Frontier Eval owns the implementations and the evaluation outputs built on top of them. base_responses_api_agent.py (the interface) is Core while every harness under responses_api_agents/ is Frontier Eval; rollout_collection.py (the engine) is Core while reward_profile.py (what consumes rollouts to make decisions) is Frontier Eval.

Teams used, all confirmed to exist with write access to this repo (a team without write access is silently ignored by CODEOWNERS):

Team	Lane
@NVIDIA-NeMo/gym_env_benchmarks	Frontier Eval
@NVIDIA-NeMo/gym_core	Core-Arch + RL
@NVIDIA-NeMo/gym_architects	Advisors, backstop and escalation
@NVIDIA-NeMo/automation	CI, build, release (unchanged)

Member lists are deliberately not duplicated into the file — GitHub team membership is the source of truth, maintained by each team's maintainers.

Open questions for reviewers

• docs/ and fern/ are intentionally left unassigned in this PR, so they fall to the gym_architects backstop. The ownership table assigns docs to TME, but the callouts say both lanes should review. There is a docs_team in the org with write access if we want to route there. This is the main thing to settle before this leaves draft.
• Files not in the ownership doc are mapped by analogy and grouped under a clearly marked section. Roughly 14 nemo_gym modules postdate the doc: streaming dialects, the discovery/registry modules, orchestration/, and the rollout observability helpers went to Core as substrate; judge.py, skills.py, and rollout_reverification.py went to Frontier Eval as verification and evaluation-output surfaces. rollout_reverification.py is the least obvious call — it is a rollout_* sibling of the Core engine but it consumes rollouts to recompute rewards, which is the Frontier Eval side of the line.
• tests/ goes to Core, since the core library tests encode integration expectations for the Core contract. Per-server tests live under each server directory and already inherit that server's owner.
• Two policies in the doc cannot be expressed in CODEOWNERS and are recorded as comments instead: breaking changes to shared contracts and user-facing CLI need sign-off from both lanes, and well-isolated changes may be reviewed by frequent contributors without the code owner.

example_environments/ from the doc is omitted because that directory does not exist in the repo.

Test plan

• All 58 path patterns resolve to files or directories that exist in the repo, and every rule is syntactically valid.
• Every team handle verified via the GitHub API to exist and hold write access on this repo.
• Simulated GitHub's last-match-wins resolution against a spread of real paths. Spot checks: responses_api_agents/simple_agent/app.py to Frontier Eval, nemo_gym/sandbox/api.py to Core, nemo_gym/config_types.py to both lanes, responses_api_agents/verifiers_agent/app.py to both lanes, .github/workflows/unit-tests.yml to automation, README.md to the architects backstop.
• GitHub's own CODEOWNERS validation API reports zero errors on this branch, confirming no rule or team handle is being silently dropped.

Note that reviewer routing on this PR itself is governed by the CODEOWNERS on main, not the version in this branch, so the new rules only take effect after merge.

[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.

[marta-sd]

This is in the co-owned section, but points to the core team. What was the intention?