review-pr
This skill reviews GitHub pull requests by inspecting diffs, identifying issues, and fixing safe problems when branch permissions allow.
Identity proven by GitHub's verified organisation badge (nebius)
- Agents
- Claude Code
- License
- Apache-2.0 (permissive)
- Pin
- b1924depinned Oct 9, 2026
- Context cost
- about 2,800 tokenswhen the skill loads
- GitHub stars
- 7on nebius/nebius-ps-services
- Scripts
- None found
What it does
“Use only when explicitly asked to review a GitHub PR by number, URL, or branch: run a local exact-head review, fix safe writable findings through create-pr, revalidate, and attest readiness; never merge. In Agentic SDLC, report only.”
Generated summary from the skill's own text
This skill reviews GitHub pull requests by inspecting diffs, identifying issues, and fixing safe problems when branch permissions allow. It validates that PRs meet merge-ready criteria through local checks and CI validation. In Agentic SDLC runs, it enforces read-only review mode and routes findings back to the coordinator rather than making branch changes.
When to use it
- Review open PRs by number, URL, or branch before merging to main
- Fix safe code, test, docs, or workflow issues directly on PR branches
- Check for merge blockers, CI failures, conflicts, or unresolved review threads
- Validate PRs against requirements and evidence in Agentic SDLC runs
When not to
- When active Agentic SDLC mode requires branch changes, which must return to coordinator
- For fork or report-only reviews that cannot safely update the branch
- When conflicts require semantic judgment beyond clear merge resolution
Install
These commands install from the publisher's repository and need no account.
skills CLI
npx skills add nebius/nebius-ps-services --skill review-prInstalls from the repository's default branch, which may have changed since the commit we checked.
GitHub CLI 2.90 or later
gh skill install nebius/nebius-ps-services review-pr@b1924de70353e4e688be5017ec608ec9b89c8ff4Installs the commit we checked (b1924de).
Copy by hand
Download the folder at b1924de and put it in .claude/skills/review-pr for Claude Code and in .agents/skills/review-pr for Codex, Cursor, Gemini CLI, GitHub Copilot and OpenCode.
Copy for your agent
A prompt for any coding agent. It installs this skill from GitHub at the pinned commit, copies the files as they are, and runs nothing unless you ask.
Prompt
Install the agent skill "review-pr" by Nebius. Source, pinned to commit b1924de70353e4e688be5017ec608ec9b89c8ff4: https://github.com/nebius/nebius-ps-services/tree/b1924de70353e4e688be5017ec608ec9b89c8ff4/skills/review-pr 1. Download that folder at exactly this commit. Do not use the default branch. 2. Copy every file unchanged, keeping the folder structure, into .claude/skills/review-pr/ for Claude Code and .agents/skills/review-pr/ for Codex, Cursor, Gemini CLI, GitHub Copilot and OpenCode. Create the folders if they do not exist. 3. Do not edit the files, and do not run any script in the folder unless I ask you to. 4. When you are done, list the files you copied and where you put them. Listing and checks: https://www.agentnative.dev/skills/nebius/review-pr
SKILL.md at the pin
--- name: review-pr description: "Use only when explicitly asked to review a GitHub PR by number, URL, or branch: run a local exact-head review, fix safe writable findings through create-pr, revalidate, and attest readiness; never merge. In Agentic SDLC, report only." disable-model-invocation: true --- # Review PR ## Help For `$review-pr --help` or `$review-pr -h` (including native Claude forms), return concise help and stop before any workflow step. State the purpose and invocation policy. Show exact usage for every public action. Describe each public action, positional argument, and flag in one concise line, including `-h, --help`; say "No additional public flags" when there are no others. Use only the documented public interface. For internal or coordinator-only skills, state that boundary and that no standalone public workflow action exists. After the selected `SKILL.md` is loaded, help is report-only: do not call any additional tools, inspect project state, or modify files, private state, Git, or external systems. Never expose private helper actions or flags or treat help as workflow authorization. ## Agent Compatibility Use `$review-pr` in Codex, `/review-pr` in Claude Code, or `/skills:review-pr` in the Claude plugin. Dollar-prefixed skill examples refer to the same named skill on either host; use the native invocation syntax. Preserve the declared invocation policy, approvals and workflow ownership. Use available native tools; an unavailable required capability is a blocker, never permission to bypass a guard or claim unobserved behavior. ## Purpose Use this skill for GitHub-backed pull request review work when the goal is to inspect any provided PR, identify real issues, fix safe problems when branch permissions allow, and leave it closer to merge-ready. ## Use This Skill For - Reviewing an open PR by number, URL, or current branch against its base branch, usually `main`. - Reviewing PRs opened from the user's branch, another same-repository branch, or a fork. - Checking whether the branch is mergeable and whether CI or review blockers remain. - Fixing safe code, test, workflow, or documentation issues directly on the PR branch when the branch can be safely updated. - Resolving straightforward conflicts with `main` or the PR base branch when the correct resolution is clear and push permissions allow it. - Reporting review findings and exact blockers when the branch cannot be updated safely. - In an Agentic SDLC run, checking the PR against requirements, design, local validation, tests, evaluation, UAT, and commit evidence when that evidence is available. ## Requirements - A Git repository with the PR branch available locally, fetchable from `origin`, or available through `gh pr checkout` for same-repository or fork PRs. - GitHub CLI (`gh`) authenticated for the target repository. ## Active Agentic SDLC Review Mode When the PR maps to a matching active Agentic SDLC run, this mode overrides the generic review-and-fix behavior below. Review is findings-and-readiness-only: - Reload the active run, current checkpoint, execution coordinator, commit evidence, UAT evidence, recorded PR evidence, and current GitHub PR state. - Require the local clean `HEAD`, recorded `promoted_head`, and PR head SHA to be identical before calling the handoff reviewable. Disagreement maps to `PR_HEAD_DRIFT`. - Read the base/head diff, checks, review decision, comments, conflicts, specs, and SDLC evidence. Record a concise local readiness result when private run state is writable. - Do not check out or switch branches, edit files, run mutating formatters, commit, amend, merge, rebase, update the branch, resolve conflicts, push, or otherwise change the PR head. Any finding that requires a branch change is a coordinator input, not a repair inside `review-pr`. Classify it with `sdlc-classify-failure`, route it through `sdlc-start`, and invalidate every affected downstream gate. The workflow must validate, test, evaluate, update documentation, align, seal/promote, rerun UAT, and republish a new exact SHA before review resumes. `sdlc-merge-pr` remains the only Agentic SDLC merge path and requires a separate explicit user request. ## Sibling Skill Routing `review-pr` should stay the coordinator skill for PR review, but it should pull in the smallest relevant set of sibling skills based on the actual PR surface. Do not load unrelated skills just because they exist. Route selectively like this: - `align`: when the PR spans multiple surfaces and needs end-to-end alignment across implementation, tests, docs, help output, CI, or examples. - `github-workflows`: when the PR changes `.github/workflows/**`, reusable workflow behavior, release automation, or merge/publish gates. - `helmchart`: when the PR changes Helm charts, including `Chart.yaml`, `values.yaml`, templates, schema, or chart publication contracts. - `python-project`: when the PR centers on Python packaging, `pyproject.toml`, `src/`, CLI structure, pytest, Ruff, or general Python project hygiene. - `shell-scripting`: when the PR changes `.sh` files, shell helpers, or Bash CLI flows. - `linter`: when the PR needs shell, Markdown, or Python lint cleanup as part of making the branch merge-ready. - `nebius`: when the PR depends on live Nebius IAM, VPC, quota, MK8s, or SDK behavior. - `terraform`: when the PR is mainly about Terraform module or environment structure, interfaces, validation, or security posture. - `publish-release`, `publish-image`, or `publish-helm`: when the PR changes a release helper plus its matching publication workflow and changelog contract. If the PR is narrow, use only the matching domain skill. If the PR is broad, use `align` plus the one or two domain skills that cover the specialized surfaces. ## Workflow 1. Freeze the exact host/repository/PR, head/base branches and SHAs, author, ownership, draft state and live default. Resolve the actual PR, not an unrelated current branch. Active Agentic SDLC uses its read-only mode above. 2. For ordinary work, read the ordinary completion reference from the installed create-pr skill and the commit continuation reference from that same skill. Honor report-only restrictions and externally owned/fork branch limits. 3. Inspect the complete base-to-head diff, specs, review comments, unresolved threads, checks and conflicts. Select relevant sibling skills. Review locally; never install or run an AI agent on GitHub Actions. 4. Before editing, use the canonical owner to establish intake and select the correct PR checkout. Prove its HEAD matches the remote PR head and preserve unrelated work. Then classify findings and implement safe, supported repairs when authorized and writable. Stop for semantic ambiguity, unsafe changes, unresolved human objections or unavailable validation. Do not guess a product decision. 5. Delegate all branch switching, commits, base synchronization and pushes to create-pr's private preparation handoff, using the active grant and original root receipt. Standalone authorized review repair opens that canonical owner transaction before Git effects. It cannot call review-pr or merge-pr again. Never use raw commits, rebase, force push or GitHub update-branch here. 6. Run relevant checks, then re-review the complete final diff and fresh remote head/base after each repair. No earlier review survives a changed head or base. A newly discovered unsafe finding stops the loop and returns a blocker. 7. For an ordinary completion caller, require no unresolved findings, passing validation/CI and no outstanding human requested changes or threads. Follow the Actions protocol reference from the installed merge-pr skill to post a structured COMMENT review on the exact head. Record its review ID, head/base and validation evidence. This attestation does not approve the author's own PR. Standalone review also returns this evidence when writable and eligible; report-only/fork/non-default reviews return findings without merge admission. 8. Return findings/fixes, validation, final identity and attestation or blocker. Do not merge. The calling create-pr/publication/merge-pr owner continues only within its existing completion authorization. Standalone review closes its repair transaction after publication and verification. ## Command Reference Read `references/command-reference.md` when exact GitHub CLI or Git commands are needed for PR metadata, checkout, local base sync, conflict detection, or branch updates. ## Learning Loop When using this skill, capture durable, reusable, public-safe learnings in the narrowest appropriate surface only when the task contract allows source edits. For read-only/report-only work, or when a learning is not public-safe, evidence-backed, in scope, or free of unverified/vendor-specific claims, do not edit skill sources; report that it was skipped. Do not capture secrets, private URLs, customer data, raw logs, or one-off local state. ## Guardrails - Default behavior is review-and-fix, not report-only, unless the user asks for audit-only output or active Agentic SDLC review mode applies. - In active Agentic SDLC review mode, never mutate the exact promoted PR head; findings that require changes must return to `sdlc-classify-failure`. - A PR link or number is authoritative. Do not ignore it and review the current branch instead. - Use sibling skills selectively by changed surface; do not turn every PR review into a full-repo multi-skill pass. - Do not claim a PR is ready if checks are failing, conflicts remain, or review blockers are unresolved. - Do not claim conflicts are resolved unless the conflict-resolution commit was pushed to the PR branch or GitHub confirms the branch is mergeable after the update. - Do not clear or ignore unresolved reviewer concerns without evidence that the branch now addresses them. - Resolve conflicts automatically only when the correct merge is obvious from local context. Stop and explain when conflicts are semantic or risky. - Never rewrite the default branch. - Do not push to another contributor's branch or fork unless GitHub permits it, the user asked for branch updates, and the update is non-destructive. - Do not create a replacement PR from someone else's branch unless the user explicitly asks for that follow-up path. - Do not rebase or force-push in the ordinary review repair path. - Do not issue approval or merge effects. The Actions broker approves only after local attestation; merge-pr owns all ordinary merge effects. - Do not let a sibling skill override `review-pr`'s ownership of readiness, branch safety, or final review judgment. - Keep the PR branch aligned across code, tests, docs, and workflows before calling it ready. ## Output Contract When using this skill: 1. Review the PR against the base branch. 2. Identify whether the PR branch is same-repository, forked, writable, and safe to update. 3. Apply the smallest relevant set of sibling skills for the changed surface. 4. Fix safe issues directly on the PR branch when appropriate and writable. 5. Resolve straightforward conflicts when safe and push permissions allow it. 6. Run focused validation and report what actually ran. 7. State clearly whether the PR is ready to merge and what, if anything, still blocks it.
From nebius/nebius-ps-services at b1924de, shown as plain text under its license: Apache-2.0 (permissive). The license text is in LICENSE.
View SKILL.md on GitHub at b1924deFiles
- SKILL.md11 KB · 725f41a56522
- README.md2.5 KB · c2c0a3d15c20
- agents/
- openai.yaml887 B · 881712d5f336
- evals/
- evals.json2.3 KB · b7ce149fecc7
- trigger-prompts.csv860 B · c6ba18ebd343
- references/
- command-reference.md826 B · d9fa947fae32
6 files, 19 KB in total. Each file has a sha256 in the pin's manifest; the first 12 characters are shown.
Skill folder on GitHub at b1924deMore from Nebius
agent-development
This skill guides building the NPA chat agent backend using a grounded-first routing architecture that minimizes token spending.
Claude Code, Codex, Cursor +4 · about 5.7K tokens · pinned 6 days ago
author-npa-workflow
Helps authors write and validate NPA workflow YAML specifications, which are declarative state machines that orchestrate tools via SkyPilot.
Claude Code, Codex, Cursor +4 · about 3.3K tokens · pinned 6 days ago
cosmos
This skill enables the agent to work with Cosmos, a world model for synthetic data and video generation.
Claude Code, Codex, Cursor +4 · about 1.6K tokens · pinned 6 days ago
cosmos3-npa-workflow
This skill helps agents write, validate, and submit npa.workflow YAML specifications for Cosmos 3 image and video generation tasks.
Claude Code, Codex, Cursor +4 · about 1.7K tokens · pinned 6 days ago
cosmos3-post-training
This skill guides planning and understanding of Cosmos3 supervised fine-tuning workflows for robotic policy training on NPA infrastructure.
Claude Code, Codex, Cursor +4 · about 2.5K tokens · pinned 6 days ago
curobo
This skill enables agents to run NVIDIA cuRobo V2 motion planning for robotic arm pose planning and benchmarking on CUDA GPUs.
Claude Code, Codex, Cursor +4 · about 1.5K tokens · pinned 6 days ago
Related in Coding workflow
maintainer-review
This skill helps maintainers assess GitHub issues and PRs by evaluating whether claimed problems are real, whether existing features already solve them, and…
by OpenAIOfficial publisher
Claude Code, Codex, Cursor +4 · about 8.9K tokens · pinned 6 days ago
maintainer-review
Evaluates GitHub issues and PRs through a systematic desk review to assess demonstrated need, existing workarounds, design correctness, and maintainer impact.
by OpenAIOfficial publisher
Claude Code, Codex, Cursor +4 · about 8.3K tokens · pinned 6 days ago
dd-file-issue
This skill helps file GitHub issues to the DataDog/pup repository for bugs, feature requests, and documentation problems related to the pup CLI, agents, and…
by DatadogOfficial publisher
Claude Code, Codex, Cursor +3 · about 2.2K tokens · pinned 6 days ago
create-pr
This skill creates a GitHub pull request with a Linear-ticket-prefixed title and a decision-led narrative description for the prisma-next repository.
by PrismaOfficial publisher
Claude Code, Codex, Cursor +4 · about 4.3K tokens · pinned 6 days ago
github-review-iteration
Orchestrates an iterative GitHub PR review loop by delegating triage and implementation to sub-agents, then repeating until all actionable review items are…
by PrismaOfficial publisher
Contains scripts
Claude Code · about 1.9K tokens · pinned 6 days ago
prisma-next-feedback
Routes bug reports, feature requests, and team Q&A to the appropriate Prisma Next channel.
by PrismaOfficial publisher
Claude Code, Codex, Cursor +4 · about 4.2K tokens · pinned 6 days ago
Open roles at Nebius
5 of 83 open roles on the jobs board
- Senior Software Engineer (Capacity and Quota Management)
Amsterdam, Netherlands; Berlin, Germany; United Kingdom
- Senior Backend Software Engineer (Observability)
Amsterdam, Netherlands (Remote)
- Lead Security Automation Engineer
Israel
- Senior Software Engineer (Data Platform, C++)
London, United Kingdom
- Senior AI Infrastructure Systems Engineer (R&D / GPU / AI)
United States