- name
SKILL.md
Файл навыка остаётся в репозитории издателя.
Вставьте скопированную команду в терминал. Она установит этот SKILL.md в выбранного агента.
npx skills add mariadb-operator/mariadb-operator --skill mariadb-operator-pr-review -a cursor -yВставьте скопированную команду в терминал. Она установит этот SKILL.md в выбранного агента.
npx skills add mariadb-operator/mariadb-operator --skill mariadb-operator-pr-review -a claude-code -yВставьте скопированную команду в терминал. Она установит этот SKILL.md в выбранного агента.
npx skills add mariadb-operator/mariadb-operator --skill mariadb-operator-pr-review -a codex -yВставьте скопированную команду в терминал. Она установит этот SKILL.md в выбранного агента.
npx skills add mariadb-operator/mariadb-operator --skill mariadb-operator-pr-review -a grok -yОтправьте скопированное сообщение в чат Grok Bot. Бот загрузит этот SKILL.md с GitHub и добавит его к навыкам, общим для всех ваших ботов.
Install this skill: https://github.com/mariadb-operator/mariadb-operator/tree/HEAD/.agents/skills/mariadb-operator-pr-reviewОт издателя
You are a maintainer of mariadb-operator performing an initial review of a pull request. Evaluate the PR for correctness, safety, pitfalls, backwards compatibility and code quality, then give an overall assessment.
Read AGENTS.md at the repository root before reviewing. It is the authoritative description of this
project's architecture, patterns, gotchas and guardrails. This skill deliberately does not restate those
rules — it tells you how to run a review that checks a diff against them. Throughout, "§ " refers to a
section of AGENTS.md. When a check below says "verify against § X", open that section and compare the diff to it.
Whenever this skill calls GitHub — fetching PR context or delivering a review — pick the access method in this order, falling through only when the previous one is unavailable:
mcp__github-mariadb-operator__*). These may show up as
deferred tools — if so, load their schema with ToolSearch (e.g.
ToolSearch({query: "select:mcp__github-mariadb-operator__get_file_contents", max_results: 1})) before
calling them.gh CLI with the project-specific token, if the mariadb-operator MCP server isn't connected. Use
GITHUB_MARIADB_OPERATOR_TOKEN explicitly (GH_TOKEN="$GITHUB_MARIADB_OPERATOR_TOKEN" gh ...) rather than
the ambient gh auth session.mcp__github__*), if neither of the above is available.gh CLI with default credentials (plain gh auth, no explicit token) as the last resort, if none of the above work.Given a PR URL or number, collect it via the access method chosen in the GitHub credentials
section (the gh invocations below assume that method — add the GH_TOKEN override or swap in the equivalent
MCP tool as appropriate):
gh pr view <n> --json title,body,author,state,additions,deletions,files,labels,baseRefName
gh pr diff <n> # full unified diff
gh pr view <n> --json reviews,comments # existing review discussion
Then, from the local checkout, pull the surrounding context the diff alone can't show:
api/v1alpha1/, internal/controller/, internal/webhook/,
pkg/controller/ — you need the before to judge behavioral change:
git fetch origin <baseRefName>
git show origin/<baseRefName>:<path/to/file.go>
git grep for callers of any new or changed function signature, and for users of any
renamed/removed field or type.docs/<feature>.md (§ Feature Map) — cheaper than reading code to learn intended semantics.For large PRs, spend attention proportional to risk: read HA/backup/CRD/webhook/builder changes line by line first,
then controllers, then everything else. Diff stats (gh pr diff <n> --name-only) tell you where to start.
Don't read generated files to review their contents (§ Token Savers) — for review, only their presence
or absence in the diff matters: a change to api/v1alpha1/ types with no regenerated artifacts is a quick reject.
pkg/replication, pkg/galera),
backup/restore/PITR, CRD schemas, webhooks, or pkg/builder.AGENTS.md, skills, READMEs) the content is claims
about the repo: sample only the few a reader would act on and that would cause real harm if wrong; do not
exhaustively re-verify every statement.Quick rejects — if any fire, surface it prominently up front:
api/v1alpha1/ changed, but the regenerated artifacts are missing from the diff
(§ Codegen and generated files).For each dimension give a verdict of PASS / CONCERN / FAIL. On PASS, state the verdict and stop — a
one-line reason at most, and even that names the area at a high level, never the individual checks, file:lines,
or AGENTS.md rules you verified. Do not enumerate what you checked or restate the diff. Spend words only where
there is something to fix: every CONCERN/FAIL gets specific file:line references, the failure scenario, and a
brief restatement of the AGENTS.md rule that was violated (name the § section) so the author sees which convention
the finding is grounded in. A wall of green justifications is noise the author has to read past — the checks below
are what you run, not what you report back.
On PASS, do not explicitly enumerate what you checked. This holds even when the checks were interesting or non-trivial to run: listing the facts you verified, the files or lines you spot-checked, the claims that held, or the rules that were satisfied is a PASS violation regardless of how it is phrased. Write a single one-line summary naming the area at a high level — the ✅ in the section heading already carries the verdict, so do not repeat ✅ PASS in the body. The enumeration belongs in your working notes, never in the output.
Quality bar for findings. A review's value comes from a few findings the author will act on, not from volume.
Before raising anything, ask: would a maintainer block or comment on this? Every CONCERN/FAIL needs (a) a
file:line reference, (b) the concrete failure scenario — what input or cluster state makes it go wrong — and
(c) severity honestly stated. If you cannot articulate the failure scenario, it's an observation, not a finding —
either verify it in the code (read callers, check the pre-change version) or drop it. Style opinions that
golangci-lint doesn't enforce are not findings.
Each dimension below names what to assess and which AGENTS.md section(s) hold the rules. AGENTS.md is the source of truth — read those sections and check the diff against them. Restate a rule in the output only when a finding relies on it, and then only briefly (see below); never recap the rules you checked on a PASS.
Does the code do what the PR claims, with sound control flow, valid API usage, and correct handling of edge cases (nil/empty/zero, boundaries, races, pointer/generic/interface use)? Where the diff touches a pattern documented in § Architecture and Code Patterns, verify it follows that pattern rather than re-implementing or bypassing it.
Check the diff against the applicable subsections of § Safety Guardrails, with line-by-line scrutiny of the areas it flags as dangerous.
Confirm the PR steps on none of § Gotchas and Non-obvious Rules, and re-check it against § Kubernetes best practices. Look for load-bearing assumptions that hold today but may not, and test scenarios the diff omits (§ Testing).
Check the diff against § Safety Guardrails → Backward compatibility. Classify each finding as additive (safe),
behavioral (risky), or breaking — v1alpha1 permits change, but breaking changes still need communication
and migration guidance.
Is the code well-written, maintainable and consistent with the surrounding codebase? Don't hand-verify what golangci-lint/CI already checks (§ CI — what a PR must pass); flag only what lint cannot see. Confirm tests are present and tiered per § Testing, and that the change is appropriately scoped (no dead code, leftover debug, or over-engineering).
Pick one:
Assign a risk level: LOW / MEDIUM / HIGH. Anything touching HA sequencing, backup/restore/PITR, or CRD compatibility starts at MEDIUM and rises with blast radius.
docs/ or examples/ updates without matching code.By default, reply the review back to the user in your response — do not post it to GitHub. Post it (e.g.
gh pr comment, using the access method from the GitHub credentials section) only when
the user explicitly asks you to.
Render the review as GitHub-flavored Markdown — clean enough to drop straight into a PR comment. Use the
emoji legend below so the verdict is scannable at a glance. For any dimension that is PASS, write a single
one-line summary naming the area at a high level — the ✅ in the section heading already carries the verdict, so
do not repeat ✅ PASS in the body. No file:line lists, no recap of what passed. Reserve justification,
references and failure scenarios for CONCERN / FAIL.
Emoji legend:
Use this exact structure:
## 🔍 PR Review
<1-2 sentences: what the PR does and why>
### Verdict at a glance
| Dimension | Result |
|---|---|
| Correctness | ✅ PASS |
| Safety | ✅ PASS |
| Pitfall Detection | ⚠️ CONCERN |
| Backwards Compatibility | ✅ PASS |
| Code Quality | ✅ PASS |
**Overall:** 📝 Approve with notes — **Risk:** 🟡 MEDIUM
---
### ✅ Correctness
<verdict; on CONCERN/FAIL add justification with `file:line` references and failure scenario>
### ✅ Safety
<verdict; on CONCERN/FAIL add justification with `file:line` references and failure scenario>
### ⚠️ Pitfall Detection
<verdict; on CONCERN/FAIL add justification with `file:line` references and failure scenario>
### ✅ Backwards Compatibility
<verdict; on CONCERN/FAIL bullets classifying each finding as additive / behavioral / breaking>
### ✅ Code Quality
<verdict; on CONCERN/FAIL add justification with `file:line` references>
---
### 📋 Suggested before merge
<actionable items as a checklist, or "None">
Match the emoji in each dimension heading and in the summary table to that dimension's actual verdict, and set the Overall and Risk emoji accordingly — the table and the section headings must agree.
>
---name: mariadb-operator-pr-reviewdescription: >---# .agents/skills/mariadb-operator-pr-review/SKILL.mdFAQ
Скачайте .agents/skills/mariadb-operator-pr-review/SKILL.md у издателя или скопируйте путь и откройте репозиторий. PlugMCP только показывает навык в каталоге и не хранит файл.
Издатель хранит его в https://github.com/mariadb-operator/mariadb-operator.
Claude, Cursor, Codex и Grok Build могут читать SKILL.md, который вы добавляете из репозитория издателя. Grok Bot устанавливает его, когда вы отправляете в чат ссылку на папку навыка на GitHub.