[iguehring-nvidia]

Rollout spans now carry the facts needed to name a task and show its outcome, which so far lived only in rollouts.jsonl and rollout_verdicts.jsonl.

• gym.rollout.dispatch: nemo.gym.task.id (task_id, problem_id or instance_id, else task_name; omitted when the row has none) and nemo.gym.rollout.repeat.
• gym.verify: nemo.gym.rollout.reward, only for a numeric, unmasked reward.
• New gym.rollout.health span, a child of the already-ended dispatch span and in the same trace, with nemo.gym.health.verdict and the sorted distinct nemo.gym.health.checks. It is emitted for every traced dispatch, whether or not metrics export is on.

The reward is a number, not content. Per-rollout values stay on spans; the bounded gym.rollout.rewarded_total counter is unchanged.

Closes #4988.

Validation:

• Unit tests for telemetry, rollout collection, rollout health, resources server base and judge pass, except tests that also fail on main (test_supervisor_is_the_only_writer_of_the_active_servers_gauge; memory and connection-pool tests).
• Local trace check (real model server, scripted backend, pi harness, 1 rollout): dispatch has task id and repeat, health span is a child in the same trace with verdict and rollout id.
• Real run (Slurm, vLLM Nano, terminal-bench-4 mini-SWE, 10 tasks x 1 repeat, OTLP export on): 10 rollouts, 0 failures. Every dispatch span has nemo.gym.task.id and nemo.gym.rollout.repeat; every gym.verify has nemo.gym.rollout.reward and the rollout id; every rollout has one gym.rollout.health child in the same trace (3 unhealthy with their checks, 7 healthy). 10 spans of each name, so nothing is duplicated; the same attributes read back from ClickHouse.

Known limits: in the flat-row path gym.verify has no nemo.gym.rollout.id, so the reward is joined by trace id.

🤖 Generated with Claude Code

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[iguehring-nvidia]

/review

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-5052-10a22d932aa1.docs.buildwithfern.com/nemo/gym

[svcnvidia-nemo-ci]

SHIP WITH CARE — one RISK in the new regression test’s optional-dependency handling; no production-path correctness issue found.

Reviewed all 10 changed files and directly affected task-identity, verification, health-evaluation, and telemetry paths. Review was static; no tests were executed.

Review generated by codex. This was a full review. Request a full review with /review reset=true.

[svcnvidia-nemo-ci]

RISK — Preserve the SDK-only test configuration

Failure / blast radius: With OpenTelemetry SDK installed but nemo-lens absent, this test fails at health.attributes["verdict"]: the fallback setter deliberately drops attributes, so the finished span has no verdict. Span creation and ancestry still work. This introduces a false failure in the configuration this module explicitly supports; the neighboring linked-root test already distinguishes it.

Correction / preservation: Keep the production fallback unchanged and replace only this assertion with the existing conditional pattern:

if LENS_INSTALLED:
    assert health.attributes["verdict"] == "healthy"
else:
    assert not health.attributes

Retain the unconditional trace-ID, parent-span-ID, and current-context assertions. The invariant is that ended-parent linkage works with or without Lens, while attributes are required only when the real setter is available.

Acceptance (proposed, not executed): Run test_record_child_span_joins_a_parent_that_has_already_ended in both an SDK-only environment and an SDK-plus-Lens environment. Both must pass: the former exports an attribute-free child; the latter exports verdict="healthy"; both retain the parent’s trace ID and span ID without attaching a current span. Suggested code is source-traced, not execution evidence; verify before adopting.

[iguehring-nvidia]

Addressed the review finding in 10a22d9: test_record_child_span_joins_a_parent_that_has_already_ended now asserts the verdict attribute only when nemo-lens is installed, and asserts an attribute-free span otherwise, like the neighboring linked-root test. Ran it with lens installed, and with lens simulated absent (no-op attribute setter, LENS_INSTALLED forced false): both pass. The production fallback is unchanged.

[iguehring-nvidia]

/review

[svcnvidia-nemo-ci]

SHIP — LGTM — no reliability concerns in this follow-up.

The prior RISK is resolved. The ended-parent regression test now requires verdict="healthy" only with Lens installed and expects no attributes otherwise, matching the no-op fallback contract. Trace-ID, parent-span-ID, and current-context assertions remain unconditional.

Accounted for all 10 changed files and compared them with the previously reviewed head, 9404a20bddd561d16019412aa4e1bd53c1789412. The subsequent change is assertion-only; production telemetry, documentation, and integration-test wiring are unchanged. No unresolved prior findings or qualifying introduced regressions remain.

Static inspection only; no tests were executed.

Note: formal approval is not available for this request, so this review is posted as a comment only.

Review generated by codex. This was a findings-only review. Request a full review with /review reset=true.

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/40293474270