[Yushu-Li]

MultiAgentPolicy.forward accepts a current-observation batch, but its unmasked-observation path unconditionally accesses tmp_batch.obs_next. Ordinary inference batches have no next observation yet, so dispatch fails with AttributeError: 'dict' object has no attribute 'obs_next' before the agent policy runs.

I guarded the optional next-observation access. Existing next observations are still unwrapped, and the masked-observation path is unchanged.

Refs #1244: this covers the missing-obs_next inference failure described in that discussion. It does not claim to resolve every multi-agent/on-policy preprocessing problem in that issue.

Validation on CPU:

• Four regression cases use the existing real continuous/discrete PPO fixture, with and without next observations. Outputs match direct deterministic policy inference, and the input batch is preserved.
• On the original source, both missing-next-observation cases fail; both supplied-next-observation controls pass.
• Policy, Batch, Collector and return tests: 90 passed, 1 skipped.
• Full Ruff lint and formatting checks passed.
• Changed-file Mypy reports 8 existing diagnostics, also present on unchanged f240205 with the same uncached --follow-imports=silent command; there are no added diagnostic headers. This is not a clean type-check pass.
• No full training, learning-curve benchmark or full project test run was performed.

I used Codex assistance for implementation and local regression checks.

• Change described and relevant issue linked
• Changelog updated
• Regression tests added
• Ruff formatting and lint checked
• Related local tests run
• Clean type check (baseline diagnostics documented above)
• Full project tests and documentation build

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T07:19:29.905370Z	142fb66	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.