Add NeMo Gym integration

[cdreetz]

Summary

Adds Verifiers v1 support for running PyPI nemo-gym environments through NeMoGymTaskset and NeMoGymHarness.

What changed

• Adds NeMo Gym taskset/harness integration.
• Loads packaged NeMo Gym config/data by nemo_env.
• Routes NeMo Gym policy model calls through the active Verifiers rollout endpoint.
• Registers the Verifiers proxy as NeMo Gym's policy_model without spawning an extra custom model-server package/process.
• Adds environments/nemo_gym_env as a runnable example.
• Adds the optional verifiers[nemogym] dependency extra.

Validation

• uv run ruff check verifiers/v1/packages/harnesses/nemo_gym.py tests/test_v1_nemo_gym_harness.py verifiers/__init__.py verifiers/v1/packages/harnesses/__init__.py pyproject.toml
• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 uv run pytest -p pytest_asyncio.plugin tests/test_v1_nemo_gym_harness.py
• uv build --out-dir /tmp/nemo-vf-clean-pr-build
• prime eval run nemo-gym-env ... -r 100 -c 20 on example_single_tool_call: reward 1.0, 100/100 completed
• prime eval run nemo-gym-env ... -r 100 -c 20 on example_session_state_mgmt: reward 1.0, 100/100 completed

---

Note

Medium Risk
Introduces a new v1 harness/taskset plus an HTTP proxy and global-process lifecycle management for nemo-gym, which could impact rollout execution and dependency resolution (including new uv conflict rules).

Overview
Adds NeMo Gym integration for Verifiers v1 via new NeMoGymTaskset/NeMoGymHarness exports, allowing rollouts to run packaged nemo-gym JSONL tasks and configs.

The harness starts/tears down a persistent NeMo Gym server stack and injects a local OpenAI Responses-compatible proxy (NeMoGymModelProxy) that routes each rollout’s model calls back through the active Verifiers endpoint (including support for concurrent rollouts via per-rollout model routing).

Includes a runnable example environment nemo-gym-env, a new optional dependency extra verifiers[nemogym] with uv conflict constraints, expanded OpenAI Responses usage serialization (token detail fields), and a small endpoint parsing tweak to treat role="developer" as a system message; adds comprehensive proxy/runner unit tests.

^{Reviewed by Cursor Bugbot for commit 2caa939. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add NeMo Gym integration with NeMoGymHarness, NeMoGymTaskset, and a local model proxy

• Adds NeMoGymHarness and NeMoGymTaskset to run NeMo Gym tasks within the Verifiers harness lifecycle, including dataset ingestion, task normalization, and result mapping (completion, reward, metrics) back onto State.
• Introduces PersistentNeMoGymRunner that starts a long-lived NeMo Gym server stack once per harness and routes each rollout through a per-rollout proxy model, avoiding repeated process startup overhead.
• Adds NeMoGymModelProxy, an aiohttp-based local proxy exposing /v1 endpoints that dynamically routes OpenAI Responses API requests to the correct upstream server per active rollout using per-route api_key/model mappings.
• Skips the NeMo Gym policy model process via skip_nemo_gym_policy_model_process and substitutes the Verifiers proxy endpoint instead, so the external LLM provider is used directly.
• Adds a new nemo-gym-env environment package under environments/nemo_gym_env with a load_environment factory and a pyproject.toml with default eval config.
• Risk: NeMo Gym integration requires Python ≥ 3.12 and conflicts with openenv and dev extras due to dependency incompatibilities.

^{Macroscope summarized 2caa939.}

[macroscopeapp]

Approvability

Verdict: Needs human review

Unable to check for correctness in 2caa939. This PR introduces a new NeMo Gym integration with substantial new code (~1000+ lines) including async proxy infrastructure and lifecycle management. Additionally, there is an unresolved review comment identifying potential concurrency issues with shared module-level globals that could cause race conditions across event loops.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 2caa939. Configure here.}

[cursor]

Concurrent runners share mutable globals without per-loop safety

Medium Severity

_NEMO_GYM_ACTIVE_RUNNERS and _NEMO_GYM_OWNS_AIOHTTP_CLIENT are module-level mutable globals modified via global declarations in both _ensure_started and teardown. While _NEMO_GYM_GLOBALS_LOCK guards some accesses, _ensure_started reads and writes _NEMO_GYM_ACTIVE_RUNNERS and _NEMO_GYM_OWNS_AIOHTTP_CLIENT in its error handler (lines 316–318) inside the lock, but teardown acquires the same lock separately. If the module is imported in a different event loop or process (e.g. via Ray workers), the module-level asyncio.Lock won't provide cross-loop protection, letting two runners corrupt the shared counter or double-close the aiohttp client.

Additional Locations (1)

• verifiers/v1/packages/harnesses/nemo_gym.py#L245-L246

 

^{Reviewed by Cursor Bugbot for commit 2caa939. Configure here.}