Self-review

Catch issues before a maintainer does, over your whole PR diff. You are already on the branch with the project conventions available, so: get the diff → review it against the rubric and the guides for the areas you touched → report → iterate with the contributor until it is ready, then remind them to put the notes on the PR.

Report-only — do not edit, commit, or push as part of reviewing. Once the contributor makes fixes, load commit-and-pr for the commit/PR conventions and run-cpu-tests for the tests you cite.

1. Get the diff

Identify the remote that points to verl-project/verl-omni (do not assume it is named origin), then diff the whole branch against its base:

git remote -v
git fetch <upstream-remote> main
git diff <upstream-remote>/main...HEAD        # whole branch, not just the last commit

If the branch trails main and the diff looks polluted with unrelated merged files, scope to your own commits instead: git log <upstream-remote>/main..HEAD --oneline, then git show <commit>.

2. Read the rules for what you touched

Review against the project rules, not a remembered copy. AGENTS.md is the top-level contract; then read the guide for each area you changed:

3. verl-omni rubric

Beyond generic correctness, these project-specific traps are the ones a generic review misses — each has bitten a real PR and none is obvious from the diff alone:

• Silent field loss / wire compatibility — the rollout request/output path threads private engine keys (prompt_token_ids, the dual-written multi_modal_data, extra_fields). Keep valid requests byte-identical on the wire and fail closed on conflicting/unsupported fields rather than dropping or overwriting them.
• Shape guessing — flag new ndim==5 / shape[1]==3-style modality or layout inference; the declared media contract (DiffusionIOSpec, media_kind) should drive it instead.
• Sanity gates the CPU job skips — changed config dataclasses need the regenerated _generated_*.yaml (scripts/generate_trainer_config.sh); check_dataproto_usage.py, check_device_api_usage.py (no literal cuda / nccl / .cuda), check_docstrings.py, check_license.py, and validate_structure.py each run as their own job.
• Pin-induced false failures — a red test may be a local verl / vllm-omni pin mismatch, not the diff. Confirm the env matches .github/*_pin.txt before treating it as a finding.
• Don't overstate GPU evidence — "reached engine init" is not "end-to-end passed"; cite only what a run actually completed.
• Surgical diff — every changed line should trace to the stated goal; flag unrelated refactors, reformatting, and orphaned dead code (unused imports, variables, functions left by your own change).

4. Report

• Blocking — numbered. Each: title → explanation → file.py:line → impact. Cite the rule, e.g. Per .agents/rules/config.md: regenerate the yaml.
• Non-blocking — same format, lower severity: raise with the reviewer rather than guess at now.
• Dead code (advisory) — a short table: path:line · Likely-dead / Used · reason.
• Summary — a short synthesis and a verdict (READY / NEEDS CHANGES), spelling out what to fix before submitting vs what to leave for the actual review.

Be concrete, cite the rule, review the whole diff, and don't invent issues or flag pure style the formatter already enforces.

Evidence rule (mandatory)

Self-review is prone to confident guessing. Every finding and every claim must be grounded, or explicitly marked as ungrounded:

• Tag each finding with its evidence basis: [verified] (you read the exact file.py:line and it shows the problem), [likely] (inferred from the diff but not confirmed at the source), or [unchecked] (rubric item you did not actually inspect). A file.py:line with no matching read is not [verified].
• Never claim a check passed that you did not run. State the exact command and its result, or write "not run". Do not report "CPU tests pass", "ruff clean", or "config regenerated" from assumption.
• Separate verified vs unverified in the summary. List what actually ran (command + outcome) apart from what remains unverified (e.g. GPU end-to-end, untested pipelines). A READY verdict must name what it does not cover.
• If you could not open a file or run a check, say so plainly rather than producing a plausible-sounding finding. An honest gap beats a fabricated one.

5. Iterate, then share

Expect several rounds: the contributor fixes findings, you review again. Keep going until the verdict is READY — only the leave for the actual review items should reach the reviewer unresolved. End by reminding the contributor to put the final summary in the PR description or a comment; it saves the reviewer a round-trip. Never commit the review notes as part of the diff.