[init-nikhil]

• send partial work to the verifier when the agent hits its call or iteration limit. motivation is we should grade these as 0 reward trajectories instead of discarding them
• remove the default model-call limit, but let users set one explicitly. This limit was too low to solve tasks on benchmarks we're interested in.
• extract result formatting and a small model-client interface for sandbox reuse, and tag model requests with the invocation they belong to. This stubs out the model client into a Python protocol, so that sandboxed implementations can pass their own implementation of it. Specifically, we want to enable constructing instances without access to Gym's configuration (since sandboxes won't have them).

Tests and validation

Current revision: df9dc50ff7c728e0042399350cbe19efd394a4f1. Rebased the existing ten-PR stack onto main 9ef5ed0f80a9f81c5d82af6588cd4cc956c4f5d1; assembled tip is 196bd5e333abd6e8ee7a93f4c8805ae33f67a66a (#4357).

Main compatibility adds only two test updates: #3846 gives the fake HTTP response its headers mapping for checkpoint-aware Environment seeding; #4088 checks serialized tool-output snapshots while retaining history, round-trip, and interpreter-object coverage. Both changes were independently reviewed. The rebase preserves upstream NOOA serialization and runtime configuration changes.

Validation on the exact rebased revisions:

• fix(nooa): grade partial output at execution limits #3843: 170 passed (one expected skip-verification aggregate-metrics warning).
• feat(nooa): use standard single-agent-turn sessions #3846: 195 passed.
• feat(nooa): stage a portable sandbox runtime #3844: 218 passed, 8 Linux-only installer skips on macOS.
• Assembled fix(nooa): preserve terminal evidence through standard cleanup #4357 tip: 373 passed, 16 platform skips (10 installer and 6 Linux supervisor tests).
• pre-commit run --all-files passed. DCO sign-offs, sequential layer ancestry, and the focused final diff were checked.

Command from each revision's responses_api_agents/nooa_agent: python -m pytest tests --import-mode=importlib -q -o addopts= -o log_cli=false --tb=short -rs. Exact early-layer tests use installed NOOA 051472343211914222e24ce36d8752f4e86bbe43; the assembled tip uses coherent core/CLI/BenchAgent 19caab169b018476ac433d040f6ae3f06aeff101. Commands and logs: /tmp/nooa-merge-main-candidate-20261016/final-tests-{3843,3846,3844,4357}/.

Prior real-model evidence for the unchanged reviewer-fix behavior remains at /tmp/nooa-review-fixes-20261016/live-final/, run on pre-rebase tip 8185eb1f405ba476aab42b00a222f92dbec4364f: real Qwen output-token and iteration exhaustion each verified once; an actual iteration limit followed by injected cleanup failure remained fatal, skipped verification, and closed both sessions. This main sync adds test compatibility rather than new agent behavior, so that smoke was not rerun and is not presented as a smoke of the rebased tree. It used the integration-test Resources fixture and was not a new SWE-bench/OpenSandbox canary.

GitHub CI is requested for the new heads. Optional cross-harness conformance failures also reproduce on main and are separate from the required Nemo_CICD_Test gate. Existing branch names, bases, draft states, native groups, and description text outside this validation section are preserved.

Required branch CI passed on this revision: 8,793 core tests, 921 sandbox tests, 95.07% coverage, and 170 NOOA tests (including Linux installer tests for #3844).

Signed-off-by: Nikhil Deshmudre nikhil.deshmudre@gmail.com

[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 81d67a4

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/38970092729

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/38970603351