[init-nikhil]

Current stack revision — November 3, 2026

Head: af4a32ac815b71eee5007e43d84baf28a699a7f4. Stack refreshed onto main 3880fa5024bf97199cba96ade2e0fa6ec9eb6fbf with current import locations and merge conflicts resolved. Each PR remains a separate layer; landing uses native stack merge.

#4088 now packages the launching Gym checkout using its unchanged root pyproject.toml and full core dependencies. The separate responses_api_agents/nooa_agent/runtime/pyproject.toml was removed. Tracked source and package assets are copied with their current contents; untracked files, Git metadata, and caches are excluded. This requires Git on the controller. Sandbox installation stays in its private Python, and source changes invalidate prepared offline wheelhouses. #4813 retains its workflow change and code-owner review requirement.

Validation: 87 focused runtime, offline-bundle, and task-network-policy tests passed after the refresh; all-files pre-commit passed (private guardword catalog unavailable). A real Gym wheel built successfully from the new archive format, and its NOOA entrypoint, resource dispatcher, and core package asset were verified. No model rollout or Linux installation was run for this packaging change. Earlier evidence below is historical and does not establish fresh results for this head.

GitHub dismissed earlier reviews after this update; current-head reviews and required CI are still needed.

Stack refresh — 2026-11-02

Current head: bb9b21707ea8d54f3406e2151cf5ac79fee729a3. Stack refreshed onto main 2b8efc987af03951a80f941f2a278317badfe085. This PR remains a separate layer in the intended stack merge: #4088 → #4820 → #4821 → #4351 → #4357. #4813 is excluded.

Reconstructed each layer from its final reviewed patch instead of replaying foundation commits already merged into main. Resolved current package-import and legacy-environment renames while retaining upstream checkpoint handling and newer harness support. #4088 also exports its networking helpers through the current public API and mechanically updates existing callers; #4351 exports its newly used response types.

Fresh validation on the combined refreshed stack:

• Full NOOA suite plus conformance runner/probe and harness-behavior tests: 913 passed, 10 Linux-only skips with pinned native NOOA 19caab169b018476ac433d040f6ae3f06aeff101.
• Public-API facade suite and corrected GDPval recipe fixture: 372 passed.
• Approved-prefix NOOA/network/recipe/metrics run: 671 passed, 10 skipped, and one GDPval fixture failure caused by the renamed legacy environment config; corrected and passed in the separate recipe run above.
• pre-commit run --all-files: passed, including package layering, guardwords, lint and format.
• Main advanced during validation; its intervening paths were disjoint from this stack. The final rebase was conflict-free, with the validated stack files unchanged.

No fresh real-model benchmark was run for this conflict/import refresh. Prior live-rollout evidence below is historical. Fresh reviews and required GitHub CI remain outstanding; this is not a merge-readiness claim.

---

Earlier implementation and validation (before this refresh)

• run NOOA through the existing GDPval resources server and tools
• use the canonical harness-neutral prompt while retaining file export, finish and judging
• provide a small recipe with explicit benchmark split and private override order
• correctly count prepared gdpval tasks so source changes still trigger metrics conflicts

Tests and validation

Head bc6a7e7976b29e08af1d9986489c2aabcdeea82c; base 9b9c186321d0defb19f1e10edf16150bfb1a9ee4. The existing stack tracks this work; no separate issue is needed.

40 metrics/core checks and the standalone canonical recipe fixture passed. Three focused metrics/recipe tests and independent CLI/comparator probes also passed: Stirrup-to-NOOA and reverse source reuse work, source bytes remain unchanged, and source-statistics/rendered-template conflicts are still detected.

python -m pytest -q -p no:cacheprovider --import-mode=importlib \
  tests/unit_tests/test_gdpval_prompt_metrics.py \
  benchmarks/gdpval/tests/test_nooa_gdpval.py

One fresh canonical Property Manager task 1e5a1d7f-12c1-48c6-afd9-82257b3f2409 using policy model openai/openai/gpt-5.5 completed through the real CLI, GDPval resources server, Docker sandbox, NOOA policy, finish(reason, paths) and native TaskResult. Eight distinct policy response IDs were recorded. The canonical rubric workflow sampled GPT-5.5 and returned 0.96, with a gradable row and no infrastructure or cleanup failure. The submitted PM_Weekly_Task_Schedule.docx passed structural readback and a visual check of its two-page rendering; the task container, services and temporary controller were removed.

The sanitized evidence bundle contains the report, recorded commands, resolved nonsecret settings, deliverable and previews. Upload and metadata readback in the shared NeMo Gym Drive were verified; individual recipient access was not tested.

Reproduce on this revision with the task extracted from the canonical dataset and a private overlay containing reachable model/judge endpoints, credentials and judge_media_mode=images_and_text. The evidence bundle records the original command arguments and resolved nonsecret values; the command below substitutes portable filenames for those run-local paths:

gym env start --config benchmarks/gdpval/nooa.yaml \
  --config responses_api_models/vllm_model/configs/vllm_model.yaml \
  --config private-overlay.yaml
gym eval run --no-serve --config benchmarks/gdpval/nooa.yaml \
  --config responses_api_models/vllm_model/configs/vllm_model.yaml \
  --config private-overlay.yaml --input gdp-pm-schedule-original.jsonl \
  --output rollouts.jsonl --limit 1 --concurrency 1 --num-repeats 1 \
  --agent gdpval_nooa_agent \
  --prompt-config benchmarks/gdpval/prompts/harness-neutral.yaml

The run used an explicit one-task input and judge_media_mode=images_and_text on arm64; the shipped --split benchmark recipe was checked separately. Only GPT-5.5 was sampled from the canonical judge panel. Web-search authentication, x86_64 parity and the full suite were not validated. Persisted per-call/turn objects were unavailable, and fallback terminal output is explicitly non-trainable. This fresh 0.96 result is separate from the historical 0.95 run.

pre-commit run --all-files passed. Fresh official Codex review returned no actionable findings; the independent bug hunt passed for this head and base.

CI snapshot (2026-10-24 05:44 UTC) on this head: 9 checks remain pending. Current checks. The all-green readiness gate remains unmet; no final verification-success post or new review request is implied.

[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.

[init-nikhil]

/ok to test f4f26df

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-4821-af4a32ac815b.docs.buildwithfern.com/nemo/gym

[init-nikhil]

CODEX SELF REVIEW

Reviewed head f4f26df2c19a433fb7de44d3f97aa225bb6a9924 against actual stacked base 10b5752d81c6c5098f1dc48cf35c21b4a81b571e with the official Codex CLI 0.154.0 (gpt-6-astra, fresh read-only review process). No actionable findings were returned.

The engine performed static source inspection; it did not execute tests. Local, live-model, and CI validation are tracked separately. This is not a readiness or correctness guarantee.

[init-nikhil]

/ok to test 7f2a164

[init-nikhil]

CODEX SELF REVIEW

Reviewed head 7f2a164947f2acca32f5a8734b07c1080e390d47 against actual stacked base 15e0fb21f83dfc8879ed1ef74e595f91044d94e4 with official Codex CLI 0.154.0 (gpt-6-astra, fresh read-only review process).

No actionable findings were returned.

The engine performed static source inspection; it did not execute tests. Model, benchmark and CI validation are tracked separately.

[init-nikhil]

/ok to test f5ae09c

[init-nikhil]

CODEX SELF REVIEW

Reviewed head f5ae09ce24e13d77cff23c62b71e7ba0fff0c5f2 against actual stacked base 5f686c347465e2041335ea46153eba463b60bb4f with official Codex CLI 0.154.0 (gpt-6-astra, fresh read-only review process).

No actionable findings were returned.

The engine performed static source inspection; it did not execute tests. Model, benchmark and CI validation are tracked separately.

[furgalep]

LGTM

[furgalep]

LGTM

[init-nikhil]

/ok to test af4a32a

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/41766410424