Pr Review

Review a pull request against my review standards using specialized sub-agents that run in parallel, then merge and deduplicate findings.

Input: The argument after /pr-review is the GitHub PR URL. Optionally, a second sentence can provide focus context (e.g., "focus on the new dialog component").

Voice — include in every sub-agent prompt:

Write like a friendly senior engineer reviewing a teammate's PR. Be direct, concise, and conversational. Say "this" not "this code block." Vary your phrasing — no two comments should sound templated. When you're unsure, express it naturally ("I think," "looks like," "not sure if this was intentional") — never use structured confidence markers. Explain "why" only when the reason isn't obvious or common knowledge — don't bloat comments with explanations everyone already knows. No praise, no emojis. One or two sentences per comment, max.

Steps

1. Parse the PR URL to extract owner, repo, and pr_number.

2. Fetch PR context — run these in parallel:

   # PR metadata (title, description, labels, branch, HEAD SHA)
   gh pr view {pr_number} --repo {owner}/{repo} --json title,body,labels,headRefName,headRefOid
   
   # Changed files with diffs
   gh api repos/{owner}/{repo}/pulls/{pr_number}/files --paginate --jq '.[] | {filename, status, patch}' > /tmp/pr-review-diff.json
   

   Then annotate every diff line with its real new-file line number, so sub-agents copy line numbers instead of counting from @@ headers:

   python3 - <<'PY'
   import json, re
   files = [json.loads(l) for l in open('/tmp/pr-review-diff.json') if l.strip()]
   out = []
   for f in files:
       out.append(f"### FILE: {f['filename']} ({f['status']})")
       newln = None
       for l in (f.get('patch') or '').split('\n'):
           m = re.match(r'^@@ -\d+(?:,\d+)? \+(\d+)', l)
           if m:
               newln = int(m.group(1)); out.append(l); continue
           if newln is None:
               out.append(l); continue
           if l.startswith('-'):              # removed line — no new-file number
               out.append(f"      {l}")
           else:                              # added/context line — prefix real new-file line number
               out.append(f"{newln:>6} {l}"); newln += 1
       out.append('')
   open('/tmp/pr-review-annotated.txt', 'w').write('\n'.join(out))
   PY
   

   Pass the annotated diffs (from /tmp/pr-review-annotated.txt) to sub-agents.

3. Filter to web-relevant files only — skip files outside web/ unless the PR is explicitly about root config. Focus on .ts, .tsx, .css, .mjs, .json changes. Skip generated files (*.gen.*).

4. Build an intent brief — a short summary (3-5 sentences) of:

   • What this PR is doing (from title, body, labels)
   • Which areas of the codebase it touches
   • Any focus context the user provided

   This brief is passed to every sub-agent so they understand the PR's purpose.

5. Classify changed files into two groups:

   • Test files: *.test.*, *.spec.*
   • Source files: everything else

6. Spawn review sub-agents in parallel. Each sub-agent receives the Voice instructions, the intent brief, its assigned annotated file diffs, the shared severity/format definitions below, and only the review rules for its domain.

   • Code Quality Agent — gets source file diffs + Code Quality Rules
   • Intent & UX Agent — gets source file diffs + Intent & UX Rules
   • Testing Agent — gets test file diffs + Testing Rules. Skip entirely if no test files in the PR.
   • Architecture Agent (readonly: true) — gets all file diffs + codebase access. No rules from this file — its job is to search the codebase for existing solutions, patterns, or utilities that overlap with or duplicate what the PR introduces. Flag only when the approach itself is wrong or when existing codebase utilities/patterns should be used instead.

   Tool restrictions for sub-agents:

   • Code Quality, Intent & UX, and Testing agents must include this instruction: "All diffs you need are provided below. Do NOT use Read, Grep, or Glob tools — work exclusively from the diffs in this prompt."
   • Architecture Agent must include this instruction: "The PR diffs are provided below for reference. You MUST search the broader codebase to find existing patterns, utilities, or solutions that overlap with what the PR introduces. Use Read, Grep, and Glob freely."

   Each sub-agent prompt must include:

   Severity tags (internal — used by sub-agents only, translated at merge time):

   • [blocking] — Must fix. Bugs, incorrect patterns, violations of core conventions.
   • [suggestion] — Should consider. Better abstractions, cleaner patterns, architectural improvements.
   • [nit] — Minor. Naming, formatting, small simplifications.
   • [question] — Needs clarification. Intent unclear, seems unrelated, or potentially unintentional.

   Output format for each sub-agent — return findings grouped by file path, each as: - [severity] Line N: description N is the number printed to the left of the line you're flagging in the annotated diff — copy it verbatim, do not count or compute it. If the code isn't in any diff hunk (no number on the left), use [not-in-diff] instead of a line number.

7. Merge findings from all sub-agents:

   • If two agents flagged the same code for the same underlying issue, merge into one finding with the higher severity and note both perspectives.
   • Sort by severity: blocking first, then suggestions, nits, questions.
   • Translate severity tags to final format: strip the [blocking] and [suggestion] prefixes entirely (no prefix in output). Convert [nit] to nit: and [question] to Question:.

8. Output the review in the format specified at the bottom. Prepend a metadata header as the very first line of the review output (before any findings):

   <!-- pr:{owner}/{repo}/{pr_number} sha:{headSha} -->
   

   This is invisible in chat and used by /post-pr-review to extract PR metadata.

Code Quality Rules

Assigned to the Code Quality Agent. Covers TypeScript patterns, React Query, component design, and codebase conventions.

TypeScript Strictness

• No as assertions. Flag every use of as. Suggest Zod parsing, type guards, or fixing the upstream type instead.
• No any. Flag every any. Suggest proper types, generics, or unknown with narrowing.
• No unnecessary unknown. If the shape is known, type it properly.
• Type constants and config objects so typos are caught at compile time, not runtime.
• Prefer exhaustive switch over if/else chains so TypeScript catches unmatched cases when types expand.
• No fallback on exhaustively-typed lookups. If a const map covers all possible keys, don't add ?? fallback — it's dead code that hides type gaps.

React Query Is the Only Way to Call APIs

• Never call SDK/API directly in components or event handlers. Always use useQuery / useMutation from React Query.
• Use select to transform query data instead of useMemo on the raw response.
• Use variables from useMutation to track submitted values instead of a separate useState.
• Use combine with useQueries when merging multiple queries.
• Prefer query invalidation over optimistic updates unless there's a clear UX justification.
• Don't create custom query keys or hand-write a queryFn when generated queryOptions already provide them.

Component Architecture

• Break large conditional blocks into sub-components for readability.
• Reduce state count. Flag components with many useState calls. Ask:
  • Can related states be combined into one object?
  • Can React Query handle loading/error/data states instead?
  • Can state be derived from existing data instead of synced?
• No boolean props to switch behavior. Prefer enums or mode: 'view' | 'edit' so the pattern scales when a third option appears.
• Keep domain logic out of shared/common components. Flag domain-specific branches in a component meant to be generic — push them to the caller.
• Make required props required. Don't make a prop optional and pass a no-op — make it optional or split the component.
• Don't pass hooks as props. Flag and question this pattern.
• Hooks should never be called conditionally. This violates the Rules of Hooks.

Avoid useEffect When React Patterns Suffice

• Adjust state during render instead of syncing with useEffect — see React docs: "You Might Not Need an Effect".
• Use key to reset component state instead of useEffect watching a prop.
• Derive values during render instead of storing them in state and syncing with effects.

Use Existing Abstractions

• Before writing any utility function, check the project's existing dependencies. These libraries cover most common needs: es-toolkit, react-use, use-debounce, date-fns / date-fns-tz, fast-equals.
• Prefer a schema-driven form abstraction (react-hook-form + Zod style) over manual form state built from useState/useRef + hand-rolled validation.
• Single source of truth for derived values. If the same non-trivial derivation or fallback is repeated across call sites, extract it into one helper/selector — or compute it once at the parent/data boundary and pass it down — so a site can't be missed.

Key internal abstractions:

• Domain form fields (InputField, RichTextEditorField, CheckboxField, etc. from @/components/form-field/) — not raw Radix inputs inside FormField.
• Flag any case where existing codebase helpers or dependencies are being reinvented.

Naming Conventions

• use prefix is reserved for hooks. Never name a regular function useSomething.
• Schemas start with capital letters (e.g., const UserSchema = z.object({...})).
• Constants use SCREAMING_SNAKE_CASE.
• UI-only keys use camelCase, even if BE sends snake_case.
• Destructure and rename loading states from different queries so it's clear which query they belong to (e.g., isPending: isAccountLoading and isPending: isContactLoading).
• Name functions/types for their domain, not their implementation. Avoid generic names like isIAMPending when isAdminPermissionsLoading is clearer.
• Flag confusingly similar names in the same scope — sibling functions or variables should be immediately distinguishable without reading their implementations.

Static Values Outside Render

• Move constant arrays, objects, regex, and templates outside components/functions if they don't depend on props/state. They get recreated on every render otherwise.
• Flag useMemo wrapping static data that should just be a module-level constant.

React Query State Semantics

• isLoading vs isPending vs isFetching — flag redundant checks. isFetching is true during all fetches; isPending || isFetching is redundant unless enabled: false is used.
• Use isLoading (which is isPending && isFetching) for initial load states instead of combining flags manually.

Don't Mutate Props

• If a function receives an array or object parameter, copy it before modifying. Never mutate arguments directly.

Prefer Modern APIs

• toSorted() over sort() (avoids mutation).
• replaceAll() over chained replace().
• Array.from() over push loops when building arrays.
• array.at(-1) over array[array.length - 1].
• Destructuring over repeated property access.
• Default parameters over ?? chains in function bodies — push defaults to the signature or type definition.

Intent & UX Rules

Assigned to the Intent & UX Agent. Covers questioning intent, code hygiene, UX patterns, and URL state.

Questioning Intent

• Unrelated changes — flag files/hunks that seem unrelated to the PR's purpose as [question].
• Removed code — if code, comments, or components are deleted, ask if it was intentional.
• Changed logic flow — if a component's rendering conditions or data flow changed and it seems unrelated to the PR's purpose, ask if it was intentional. Do NOT question visual/styling changes.
• Templated content — when a file is clearly derived from another adapter/domain, flag user-facing strings, variable names, or domain references that weren't updated to match the new context.

Code Hygiene

• Remove console.log and debug statements.
• Remove commented-out code. Use version control, not comments, to preserve old code.
• Flag unnecessary re-exports. If a module's only purpose is export { X } from './other' and callers could import directly, question whether the indirection layer is justified.
• Don't keep "will use later" code. Dead code should be deleted.
• Fix typos in user-facing strings and variable names.
• Add comments for non-obvious code — regex patterns, workarounds, and intentional refetchOnWindowFocus: false overrides should have a comment explaining why.

UX Awareness

• Inline errors for forms, not toasts. Toast is for async operations, not form validation.
• !important in Tailwind — always flag and ask why it's needed. It almost never is.
• Arbitrary Tailwind values like [13px] — prefer extending the Tailwind config with design tokens.

NUQS (URL State)

• Use nuqs for all URL state. Don't use useSearchParams or manual query string parsing.
• Use the correct parser — e.g., parseAsBoolean for booleans instead of comparing strings like edit === "true".
• Define query state schemas adjacent to the route file.
• setQueryStates supports partial updates — no need to spread ...prev.

Testing Rules

Assigned to the Testing Agent. Only used when the PR contains test files (*.test.*, *.spec.*).

• Never mock components. Tests should render real components.
• Never mock helper functions. Let the real implementation run.
• Avoid getByTestId, querySelector, and data-testid. Prefer accessible queries: getByRole, getByText, getByLabelText, getByPlaceholderText.
• No manual timeouts (setTimeout, hardcoded delays) in tests. Use expect.poll, waitFor, or Vitest's async utilities.
• Prefer userEvent.type over fireEvent.change for typing interactions.
• No mocking useRouter, useParams unless verified as the recommended approach.

Output Format

Group findings by file path. Within each file, list findings in order of severity (blocking first, then suggestions, nits, questions). Use this format:

## `web/components/example/example-component.tsx`

- Line 42: Using `as Status` here. A type guard would be safer.

- Line 18: This would read cleaner with `select` in useQuery instead of the extra `useMemo`.

- nit: Line 7: Schema name should be `FormSchema`.

- Question: Line 55: The progress bar got removed. Was this intentional?

If a file has no findings, skip it entirely.