[mpatel31415]

OpenCode preserves confirmed sandbox OOM evidence and capture failures for the owning policy, rather than deciding their score.

Targets #5127 in #5127 → #5128 → #5129 → #5605. Refs #4685.

• Both OpenCode adapters report canonical stage="agent" for agent execution. Locally created setup failures use seed; existing event IDs, stages and causes are preserved.
• report_agent_sandbox_oom remains opt-in. Confirmed agent-container OOM carries its original event ID and AgentOutcome; the caller applies scoring and attribution policy.
• Missing or corrupt primary captures and uncertain command outcomes remain ungraded failures. A timeout without usable capture does not enable a zero fallback.
• Ripgrep staging bounds queue wait, upload attempts and backoff. Ambiguous uploads retry at a fresh remote path; uncertain commands are never replayed.
• Opt-in split execution gives setup and agent execution separate sandbox_timeout budgets. Plugin read-back, query/stdin/env support and converter behavior are retained.

Compatibility

Ungraded handling of missing/corrupt primary captures and timeouts without a launch marker is intentional and applies even when report_agent_sandbox_oom is disabled. Those cases do not establish a usable completed agent run; the flag controls confirmed-OOM reporting, not capture validity.

A sandbox provider_io_timeout while collecting the optional SQLite snapshot preserves a valid primary export, records the original failure, and reports observation gaps. OOM, unknown outcomes, other typed failures and primary-export failures still propagate.

SWE-bench Pro and Multilingual are not covered by the OOM-reporting flag.

Validation

Integrated selected tests, scoped pre-commit/import-linter and fresh model-backed scenarios passed as detailed in #5127, including exclusions and exact source pins. Independent Claude review and one fix pass completed. Commits are SSH-signed with DCO.

Both adapter paths and preserved origin stages have regression coverage; the live OpenCode → DeepSWE path exercises the actual canonical observation.

Full repository CI and representative real-task qualification are not claimed.

Latest review fixes: the final assembled selection at 6f247486a3ff58b72ca9d83933960f4c547e6c2d passed 949 tests; scoped pre-commit and import-linter passed. Individual fixes received independent Claude review and one fix pass. These checks do not assert full repository CI or benchmark-wide score invariance.

Final-source live snapshot check at 6f247486a3ff58b72ca9d83933960f4c547e6c2d: a real-model OpenCode → DeepSWE control retained reward 1, primary output/root usage and the original failure event after an injected optional snapshot download timeout. The preceding attempt stopped before injection with an unknown command exit; it is retained as a failed attempt, not counted as validation. Two calls on the successful retry plus one on the first attempt were captured, and all five disposable sandboxes were confirmed deleted. This is a synthetic plumbing check, not representative benchmark qualification.

[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.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-5128-12d0b6b0f783.docs.buildwithfern.com/nemo/gym

[giuliolovisotto]

Addressed the review in a7af4c7 and 1a64a1c:

• Optional snapshot I/O timeout now preserves a valid primary capture and records gaps plus the original failure; primary capture, OOM and unknown-outcome failures remain failures.
• The SWE-bench assertion compares the whole dictionary and expects stage="verification", rather than the nonexistent phase field or None. Final raw ripgrep upload timeout is typed without widening its retry budget.
• Missing/corrupt primary captures and timeouts without a launch marker are intentionally ungraded even with the OOM flag disabled: they do not establish a usable completed run. The description states this behavior and excludes SWE-bench Pro/Multilingual from OOM-flag coverage.
• error_message remains an optional shared observation field; the canonical event referenced by failure_id carries the reason, so this path does not duplicate it. verification_failure_fields() is consumed via response.model_copy(update=fields) when a report exists.

The assembled local selection passed 949 tests. Representative benchmark qualification is not claimed.

[AdamRajfer]

Looks good. Snapshot failures other than a sandbox I/O timeout propagate again, the SWE-bench assertion checks the full dict, and the ripgrep exhaustion is typed. The ungraded-without-flag behaviour is called out in the description, which is enough for me. Approving.

[AdamRajfer]

Merged main into the branch to clear the conflicts after #5127 landed. One import test was adjusted to the typed failure event the branch raises; no functional change, suites green locally.

[AdamRajfer]

/ok to test 7ecaeed

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/42099741244