// Reviews a pull request from a PR URL by directly fetching the URL content (no `gh` dependency) and verifies compliance with CONTRIBUTING.md. Use when the user asks for a PR review, to review changes before merge, or to check contribution guidelines.
| name | review-pr |
| description | Reviews a pull request from a PR URL by directly fetching the URL content (no `gh` dependency) and verifies compliance with CONTRIBUTING.md. Use when the user asks for a PR review, to review changes before merge, or to check contribution guidelines. |
Reviews the changes in a specific GitHub pull request. The primary focus is code correctness and design-pattern consistency with the existing codebase. PR formatting, commit conventions, and user-facing documentation are checked but should not dominate the review. See CONTRIBUTING.md for the contribution rules referenced below.
Require a PR URL (for example: https://github.com/RLinf/RLinf/pull/123).
Fetch PR details directly from the URL:
<PR_URL>.diff (preferred)<PR_URL>.patch (fallback)All cross-references must be against origin/main, not the local working tree. The current checkout may be on an unrelated branch or contain WIP changes; Glob/Grep over the working tree does not show the upstream state. Before any cross-reference lookup:
git fetch origin main to refresh.git show origin/main:<path> (or browse https://github.com/RLinf/RLinf/blob/main/<path>).git log origin/main..<pr-head> to scope what the PR actually adds.If the PR page is private and URL fetch is blocked, report that access is unavailable and ask the user to provide exported diff/details.
Findings are ordered by severity. Categories are listed in priority order — most of the review should be on (a) and (b); (c)–(e) are checked but should not pad the output.
For every changed function, branch, and config path, look for:
enable_offload), file handles, Ray actor lifetime, missed cleanup on exceptions, double-init.except, exceptions on hot paths, missing input validation at boundaries; conversely, over-validation of internally-trusted state.origin/main version and the new version side-by-side; partial refactors often change semantics by accident.Cite file:line in the diff and the matching origin/main:file:line when relevant.
origin/main — primaryThe PR must match how RLinf already does things. Mismatches are usually defects from writing in isolation, not style nits.
origin/main — comparable model under rlinf/models/embodiment/, env under rlinf/envs/, runner under rlinf/runners/, worker under rlinf/workers/, advantage/loss/reward under rlinf/algorithms/. Compare structure, naming, registration, base-class usage, and config wiring.register_advantage / register_policy_loss / register_reward; new model/env must extend SupportedModel / SupportedEnvType and update get_env_cls() and validate_cfg. Flag ad-hoc bypasses.Worker, implement initialize, use self.log_info / log_warning / log_error (not print or stdlib logging), launch via create_group(...).launch(...).BasePolicy and implement the documented forwards (default_forward, predict_action_batch, plus algorithm-specific). Flag re-implementations of base behavior.examples/ and follow the same key hierarchy; no calculations or dynamic values in YAML; fields read-only in code.rlinf/utils/ (placement, checkpoint, distributed, data-iter, logging), point to the existing helper with file:line.Both directions matter, and EN/ZH parity must be checked explicitly:
origin/main + this PR. Verify with git show origin/main:rlinf/.... Stale references = finding.docs/source-en/ and docs/source-zh/ must agree on setup commands, paths, env vars, config keys, supported models/envs/algorithms, capability claims, reported numbers (metrics, table values, dataset sizes, trial counts), and section structure/order. If only one side is updated, name the matching file that also needs the change.opensora.rst) for section naming/order, code-block conventions, table/link style.Mention only if there are real issues; do not pad the review.
pre-commit run --all-files.xxx != yyy restatements).logging / self.log_* not print.Signed-off-by; messages follow Conventional Commits <type>(<scope>): <description>.Before listing issues, start with a brief explanation of the PR:
Keep this concise (3-6 bullets), then move to findings.
Severity + Area/File: issue summarySuggested fix: concrete actionReference: file:line in the diff, plus origin/main:file:line when the finding came from a main-branch cross-checkFor a concise checklist, see reference.md.
Adds install command in install script, Docker build stage in Dockerfile, and CI jobs for docker build and embodied e2e test when introducing a new model or environment in RLinf. Use when adding a new embodied model (e.g. dexbotic), new env (e.g. maniskill_libero), or new model+env combination that should be installable, dockerized, and tested in CI.
Cross-checks RLinf documentation against code and other docs, including English-Chinese parity checks. Use when adding or editing docs, reviewing doc PRs, validating commands/config keys/model-env names, or ensuring EN and ZH docs stay consistent.
Adds a new publication page to the RLinf Sphinx docs (EN + ZH) and wires it into the Publications index/toctree and navbar. Use when the user asks to add a publication/paper page, convert a paper README/markdown into docs, or update the publications listing.
Adds example documentation for a new model or environment in RLinf (RST pages in the docs gallery for both English and Chinese). Use when adding a new embodied or reasoning example, or new benchmark (e.g. LIBERO, ManiSkill).