Owner-only. Reviews honojs/hono PRs or local branches with two checks — tests (do they pass, is the approach right) and breaking change (public API, types, hc).
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Owner-only. Reviews honojs/hono PRs or local branches with two checks — tests (do they pass, is the approach right) and breaking change (public API, types, hc).
Use this skill to help Yusuke review a PR or local branch on honojs/hono (core repo). The framing is collaborative — the output is material a maintainer skims and then forms their own opinion from, not a verdict handed down. Owner-only: it reflects how Yusuke reviews the core, not a generic PR reviewer.
Scope.honojs/hono only. Sibling repos (middleware, honox, vite-plugins, , , ...) are not covered — they each have different concerns and would warrant their own skill.
create-hono
node-server
Confirm the target first
Do not start reviewing automatically. Always begin by asking the user what to review. Do not auto-pick the current branch, guess from context, or jump straight into gh pr diff.
Ask a single short question, e.g.:
What should I review? Reply with a PR number, "current branch", or a specific path.
Phrase the question in whichever language the user is using in the session. Wait for the user's reply before fetching anything. If the user already supplies a PR number or branch in the request — whether as a slash-command argument (/hono-review 3845) or in natural language (a "what do you think of PR #3845?" / "look at this branch" style message) — treat that as the answer and proceed without asking again. Only ask when the target is genuinely ambiguous.
When to invoke
Activate not just on the slash command but whenever Yusuke makes a review-flavoured request about a PR or branch on honojs/hono. Typical shapes (in any language he is using):
/hono-review <PR#> or /hono-review
"what do you think of PR #XXXX?"
"let's review PR XXXX together"
"give PR XXXX a quick review"
"look at this branch" / "review this diff"
Common situations:
A community PR is open against honojs/hono and needs a maintainer-flavoured triage pass.
A local branch in the honojs/hono checkout is about to be pushed or merged.
A backport or security-fix branch needs a sanity check before landing.
Inputs
Once the user has confirmed the target, pick whichever fits:
PR number: gh pr diff <N> --repo honojs/hono for the patch, gh pr view <N> --repo honojs/hono for context. If anything beyond a static review is needed (running tests, type-perf check), gh pr checkout <N> locally first.
Local branch: git diff main...HEAD for the patch, git log main..HEAD --oneline for the commit shape. Confirm the working tree is the honojs/hono repo.
Narrow review: when the diff is wide, prioritise src/**, then runtime-tests/**, then everything else.
Static review by default
The default mode is static review. Do not check the PR out locally just to start the review — gh pr diff <N> and gh pr view <N> (or git diff main...HEAD for the current branch) give enough context for the two checks below. CI already runs the test matrix and format/lint, so re-running them locally on every review is wasteful.
Escalate to a local checkout (gh pr checkout <N>) only when one of these applies:
Yusuke explicitly asks to run things locally (e.g. "run it locally", "actually try it", or the equivalent in whichever language he is using).
The static review surfaced something that needs running code to confirm (a suspected runtime edge case, a type-inference question that needs tsserver).
CI is red and the failure isn't self-evident from the logs.
If a local checkout exists, the mechanical commands are:
bun run test
bun run format:fix && bun run lint:fix
Surface the result inside Issues: test failures block merge; format:fix / lint:fix diffs are listed as files the contributor needs to commit.
Review checklist
Two checks. Everything else is "mention it only if you happen to notice it while doing these two" — never a starting point.
1. Tests
This is about the test approach, not just whether they pass. Even when CI is green, the wrong tests can pass.
Edge cases covered, not just the happy path (empty body, wrong Content-Type, invalid Accept-Encoding, malformed headers, zero-length input, ...). Match the edge cases nearby tests already cover.
Runtime-sensitive code has a runtime-tests/<runtime>/ case. compress, body parsers, streaming, and cookie handling are the usual suspects.
Public-type changes have a *.test-d.ts reflecting the new inference. If missing, ask for one.
Test style matches the surrounding files (no random new helper / framework / abstraction introduced just for the diff).
No .skip / .only / commented-out assertions left behind.
If the diff adds behaviour without a test, that itself is an issue.
2. Breaking change
Anything exported from src/index.ts or a preset (src/preset/*) — values and types — is the public contract.
Renames, removals, signature changes, or changes to the hc-inferred shape are breaking even when runtime behaviour is identical. TS users hit it on the next compile.
If a break is intentional, the deprecation flow should be visible in the diff: keep the old name as a re-export, mark it @deprecated, schedule removal for the next major.
hc (RPC client) is especially load-bearing — its inferred shape is the public API.
Things worth mentioning if you notice them
These are not checklist items — do not go looking for them. If, while doing the two checks above, one of these jumps out, fold it into Issues:
Module-level mutable state (Map, WeakMap, lazy singletons, memoization layers) in src/. Core must stay isolate-safe; in-process caches can leak between requests/tenants on Workers-style runtimes.
New import that drags in unrelated code or top-level side effects (breaks tree-shaking; bundle size is a Hono headline).
Generic-chain additions on Hono / Context / Handler / Env / Schema that look like they would slow tsserver visibly across user route chains.
Anything that smells like a real security issue (parsing, header smuggling, cookie scope, CORS reflection, body re-read attacks, prototype pollution). Stop and hand it to Yusuke — do not write security findings into a public PR comment. Follow the GHSA runbook.
How to present findings
Keep the output short. Four sections, in this order:
Summary — one or two short sentences, always present. Lead with the user-facing impact (what was broken, what wasn't working, what the PR enables) in plain language a non-expert could follow. Then add a brief clue about how the fix works — file path is enough; do not stack three protocol/API names in a row to explain the mechanism. The maintainer hasn't read the diff yet; you are orienting them, not summarising the code.
Good — 1–3 bullets on what the PR does right. Skip the section if there is nothing notable.
Issues — what needs to change before merge. Most important first. Skip the section if none. For each issue:
Lead with a one-line headline in bold that names the problem at a glance — a noun phrase (e.g. "missing precedence test case", "node:fs import in core path"), not a full sentence. The reader should be able to scan headlines and immediately know what is wrong without reading the body. Match the language of the rest of the review.
Below the headline, two short lines:
What: the specifics in one sentence, citing file:line.
Fix: what to change (one sentence or a short code sketch).
Do not add a "Why" line. The headline + What + Fix should make the problem self-evident. If a reader can read all three and still ask "why is this a problem?", the headline or What is doing its job poorly — rewrite those rather than adding an explanation.
Verdict — exactly one line: merge OK, merge OK after issues fixed, or do not merge.
The two-check structure is for your internal analysis, not for the output shape. Do not group findings by check, and do not use "blockers / suggestions / notes" tiers. Stop after the verdict line.
Tone is collaborative — you are handing material to a maintainer who will form their own judgment, not pronouncing a verdict for them. "merge OK" is a suggestion, not a decision.
Voice
Write the review like you are explaining the PR to Yusuke over coffee, not delivering a technical report. He has the deep context; what he needs from the skill is something easy to skim.
Short sentences. Cut anything that does not add information. If a sentence runs three lines on screen, split it.
Lead with the human-readable consequence, not the mechanism. "Fixes a bug where Chrome refuses the WebSocket connection" lands instantly; "Implements echo-back of Sec-WebSocket-Protocol on the upgrade response" only lands after the reader has parsed three protocol names.
Jargon only when it actually shortens the sentence. Terms like subprotocol, "echo back", "spread order" are fine when they are the most economical option — but pair them with a plain-language clue rather than chaining them. "Makes sure the user-supplied options.protocol wins" beats "places ...options after the header-derived field in the spread so it wins".
Don't stack three technical nouns in a row. It forces the reader to context-switch on every word.
Friendly, not formal. Reviewer notes, not a CHANGELOG entry. Light chat-tone wording is welcome — match whichever language the user is using in the session.
The line between AI and Yusuke is at remote-mutating actions:
AI does (locally, when Yusuke asks): gh pr checkout, edit files, run bun run test / format:fix / lint:fix, show git status -sb / git diff.
Yusuke does (remote / commit-side): git add, git commit, git push (incl. force-push), gh pr review, gh pr comment, anything that updates the PR on GitHub.
When Yusuke patches the contributor's branch
For minor fixes — typo, missing test case, format/lint diff the contributor forgot — Yusuke sometimes patches the contributor's branch directly. The skill handles the local prep so the path from "I want to fix this" to "ready to commit" is one step; Yusuke takes it from there.
Trigger: only act after Yusuke explicitly says he wants to patch the branch — phrasings along the lines of "I'll just fix it", "let me edit the branch", "add the test for me", "apply the format fix", or equivalents in whichever language he is using. Do not propose this unsolicited.
Steps the skill does:
gh pr checkout <N> — switch to the contributor's branch with the correct upstream. Confirm with git status -sb and verify the branch matches the PR. If the PR is from a fork without maintainer-edit permission, surface that immediately — there is no clean recovery.
Apply the edit Yusuke described. Stay within what he asked for; this is a light touch, not a rewrite.
Run the mechanical checks locally:
bun run test
bun run format:fix && bun run lint:fix
Show git status -sb and the relevant git diff so Yusuke can see exactly what will be committed.
Then stop.git add / git commit / git push are Yusuke's. Do not stage, do not draft a commit message and run it, do not push. If he asks for a commit-message suggestion, give it as text — let him run the command.
The Fix lines in the review output should already be concrete enough that this whole flow is a small step ("apply Fix #1, commit, push") rather than a discussion. Vague Fix lines like "add a test case" defeat the point — rewrite them to something paste-ready.
What this skill is not
Not a publish helper — np handles that locally.
Not a CI replacement. CI runs the test matrix; this skill catches what CI cannot judge (test approach, breaking-change manners, things "worth mentioning if you notice them").
