---

name: aura-review description: >- Checks how well an app uses Aura, Cognite's design system: what fraction of its components are Aura vs. custom-built, whether any custom components should really have been an existing Aura component, and whether Aura components got styled in ways that break the design system's own rules. Runs fully automatically by scanning the app's source and checking it against Aura's docs — no back-and-forth, so it also works unattended in CI. Writes a machine-readable aura-review/review.json (headline stats plus every finding, structured — the source of truth for any downstream consumer: a CDF upload, a metrics pipeline, a dashboard) plus a human-readable report at reviews/aura-review/feedback-round-/aura-review-report.md. Use when asked to run an Aura audit, check Aura compliance, or score how compliant an app is with the Aura design system, on an app that already exists. Do not use this while generating, scaffolding, or implementing an app — it is a downstream audit of finished work, not a build-time check. allowed-tools: Read, Glob, Grep, Bash, Write, Edit, WebFetch

Aura Review

An objective, code-level audit of Aura usage in a generated app: did the app use Aura where Aura applies, and did it use the Aura components it did reach for correctly?

This skill never asks the user anything. It is designed to run unattended — inside a CI job on a nightly schedule, or by hand — and to always finish with a machine-readable review.json plus a human-readable markdown report.

Never run this against an app that's still being built

Only invoke this once the app is functionally complete, as a separate step after building finishes — never while scaffolding or implementing it, and never in the same context as the build. This isn't a linter; it's an eval of how well a build agent reaches for Aura unprompted. A build agent that can see this skill's grading criteria will write code to satisfy those specific checks instead — a high score would then mean "read the answer key," not "uses Aura well." Keep building and reviewing as separate agents with no shared context, the way aura-eval-daily.yml runs them.

The one rule that matters more than any step below

Only flag something as wrong if it is already written down in either DESIGN.md or the public Aura docs. If you notice something that feels wrong but isn't documented anywhere, do not invent a rule to justify flagging it, that isn't a fair test of an app built against documentation that never mentioned the rule. Instead, note it as a suggestion for the docs and move on. This matters because these reports get used to tell people "Aura says you did X wrong" — that claim has to be traceable to something they could have read.

Likewise, when judging whether a custom component "could have been Aura", only say so when you can point at a specific Use when bullet for a named Aura component that matches — otherwise mark it undecided rather than guess.

Sources of truth (v1, Storybook is deliberately out of scope for now)

Source	Location	Use it for
DESIGN.md	<app-dir>/node_modules/@cognite/aura/DESIGN.md	Tokens, per-component "Use when" / "Use something else when" / "Dos and don'ts" / "Behavior", layout patterns
Aura docs site	https://docs.cognite.com/aura-design-system	Foundations (tokens, accessibility, design heuristics) and per-component Primitives pages for anything DESIGN.md doesn't cover — DESIGN.md's own "Docs reference" links here per component
Flows community components site	https://cognitedata.github.io/flows-community-components/	Aura's public shadcn-style component registry and docs site — use for components or usage guidance not shipped in DESIGN.md

Both docs URLs above are trusted, so fetch them with WebFetch directly rather than curl (curl requests to them get denied).

Do not reach for Storybook yet. It's excluded from this skill until we have a reliable index of story code (index.json / per-story JSON work).

Step 0 — locate the app and confirm prerequisites

Resolve <app-dir> from $ARGUMENTS, or default to the current directory. Print the resolved absolute <app-dir> path before doing anything else — if $ARGUMENTS was empty and this silently defaulted to the current directory, that's the only signal a user gets that they may be auditing the wrong app.

Confirm <app-dir>/node_modules/@cognite/aura/package.json and <app-dir>/node_modules/@cognite/aura/DESIGN.md both exist — if not, stop and report that npm/pnpm install needs to run first (this skill never installs dependencies itself).

This skill only ever reads the app and writes review.json plus the markdown report described in Step 6 — it never edits the app's own files or sends data anywhere.

This skill never publishes results anywhere. Getting review.json into a tracking store (CDF, a spreadsheet, a dashboard) is the caller's job, not the skill's — see cognitedata/aura-tools's aura-eval-daily.yml for how the CI eval pipeline does it.

Step 1 — run the objective scan (script, not you)

This is the part that must be deterministic, so it's a script, not something you eyeball. The script requires typescript, which lives in <app-dir>'s own node_modules (any Vite/TS scaffold has it as a devDependency) rather than next to this skill — set NODE_PATH so Node's resolver finds it there instead of copying the script into the app:

NODE_PATH="<app-dir>/node_modules" tsx <this-skill-dir>/scripts/scan-aura-usage.ts <app-dir> --out aura-review/scan.json

If tsx isn't on PATH, fall back to npx --yes tsx ... (prefer the direct binary when available — npx has been observed to resolve a stale cached path to tsx's own CLI entrypoint in some environments).

Read the resulting aura-review/scan.json. It contains:

• availableAuraComponents — Aura's full component catalog (Pascal case, e.g. Card, Accordion, ...), from the installed @cognite/aura version. Not used in this run's coverage math — kept so a future aggregate step (across many scans) can compute catalog utilization, i.e. which Aura components barely get used anywhere. See the auraCoveragePct note below for why that's a different question than this run's coverage number.
• auraUsages / nonAuraUsages — distinct component types, one entry per identifier, each with usageCount and every occurrence's locations. auraUsages only counts root-level Aura tags (e.g. Card, Button) — compound/slot sub-components (CardContent, CardHeader, AlertDescription, ...) are folded into their parent's usage, not counted as separate component types, because a Card composed of five slots is one usage of Card, not five usages of the design system. nonAuraUsages entries are tagged with importSource (relative, external, or unresolved-local for components defined in the same file) — this covers both first-party components the app built and third-party ones it chose instead (icons, react-router-dom, providers, etc.).
• excludedUsages — framework/router noise (Route, Link, Fragment, ...) and namespaced tags (React.StrictMode, property access on a namespace import, not a component the app authored). Excluded from every list and count above, kept here purely for transparency — nothing is silently dropped.
• totals.auraCoveragePct — auraUsages.length ÷ (auraUsages.length + nonAuraUsages.length) (the distinct-count lists, not the all* occurrence lists): of the different kinds of components this app reached for, what fraction were Aura. Deliberately not occurrence-based — a component reused many times would otherwise dominate the ratio and overstate coverage relative to how many different kinds of components the app actually used. Also deliberately not divided by availableAuraComponents.length — that would conflate "how many components did this app need" with "did it use Aura where it mattered" (a small app needing only 3 components would cap out far below 100% even with perfect compliance). totals.allAuraUsages / totals.allNonAuraUsages report raw occurrence counts alongside this, as context, but are not what coverage is based on.
• escapeHatches — every Aura component usage, including sub-components, whose className raw Tailwind values look like spacing, color, or arbitrary-value overrides. This is a purely mechanical regex check; it is not yet a verdict — some of these will turn out to be legitimate per Step 4.

Step 2 — report the coverage number as-is

auraCoveragePct from Step 1 goes straight into review.json's stats, with the caveat above about what it does and doesn't count. Do not adjust it — it's meant to be a stable, comparable number across runs so a future aggregate/CDF step can track it over time. Judgment happens in the two steps below, as separate numbers.

Also compute thirdPartyUsages here, mechanically, no judgment needed: it is exactly scan.json's nonAuraUsages entries whose importSource is external ({ identifier, usages: usageCount, packageName: moduleSpecifier } for each). These are skipped in Step 3 below because they're a deliberate third-party choice, not a compliance signal — but that also means there's nothing to judge, so don't re-derive this list by hand later.

(Don't confuse this with scan.json's own excludedUsages field above — that's unrelated framework/router noise. review.json's thirdPartyUsages is specifically the "Excluded from judgment" section: real components the app chose not to build in Aura, on purpose, for reasons unrelated to Aura.)

Step 3 — judge non-compliance (this is where you read DESIGN.md)

For each entry in nonAuraUsages whose importSource is relative or unresolved-local (i.e. first-party — skip external entries, they're a deliberate choice unrelated to Aura, not a compliance signal):

1. Open its first usage location and skim enough surrounding JSX to understand what the component actually renders/does.
2. Search DESIGN.md's "Primitive guidance" component entries (grep for plausible component names, or read the whole section if it's short enough) for one whose "Use when" bullets match this component's apparent purpose.
3. If you find a specific match, record a nonComplianceFindings entry: { component: identifier, usages: usageCount, couldHaveBeenAura: "<ComponentName>", evidence: "<quoted Use when bullet>" }.
4. If DESIGN.md doesn't cover it, WebFetch the matching Primitives page on https://docs.cognite.com/aura-design-system (per DESIGN.md's own "Docs reference" URL pattern) or check https://cognitedata.github.io/flows-community-components/ the same way before giving up.
5. If none of those sources support a specific match, still record the entry with couldHaveBeenAura: null and evidence explaining what the component does — do not guess, and do not drop the entry (every first-party non-Aura component gets a row, matched or not).
6. Before recording a match, confirm the named Aura component is actually importable from the installed @cognite/aura — check it appears in scan.json's availableAuraComponents, or that the package's exports map has a matching subpath (node_modules/@cognite/aura/package.json). DESIGN.md documents some components (e.g. Table) that are Storybook-only and not yet shipped as an importable component in every version — naming one of those as couldHaveBeenAura is misleading, since nobody can actually import it. If the documented match isn't shippable, record couldHaveBeenAura: null instead and say so in evidence.

couldHaveBeenAuraPct = (# entries with a non-null couldHaveBeenAura) ÷ (total entries considered, i.e. couldHaveBeenAuraConsideredCount). Report the count either side of that ratio, not just the percentage — a percentage over 2 components reads very differently than over 20.

Also report couldHaveBeenAuraOfNonAuraPct = (# entries with a non-null couldHaveBeenAura) ÷ nonAuraUsages.length (the full distinct non-Aura population from the scan, including external entries this step deliberately skipped). This is a secondary, broader-context number — it's expected to read lower than couldHaveBeenAuraPct whenever the app has legitimate third-party usage, since that usage counts in this denominator but was never a compliance candidate. Don't use it in place of couldHaveBeenAuraPct; the considered-only ratio above is still the fair compliance signal.

Step 4 — judge documented usage-quality findings

For each entry in escapeHatches, and separately for any Aura component used unusually heavily (e.g. hand-composing many raw sub-elements around it), check DESIGN.md's entry for that component's slug — and, if it doesn't cover the point, the matching page on https://docs.cognite.com/aura-design-system — for a specific "Dos and don'ts" or "Behavior" line the usage contradicts.

• If you find one, record a usageQualityFindings entry with the exact quoted rule it violates: { component: slug, location: "<file>:<line>", violates: "<short description>", quotedRule: "<exact quoted line>" }.
• If you don't, drop the escape-hatch finding from the final report — a raw Tailwind override that the docs don't actually forbid for that component isn't a usage-quality problem, it's just a mechanical hit that didn't hold up.

This step is deliberately the most expensive per-finding — it's also the one guarding against this skill's worst failure mode: grading an app against knowledge the model has but the docs don't.

Step 5 — write review.json (the one structured source of truth)

Everything computed in Steps 1–4 goes into a single aura-review/review.json — headline stats and every finding, structured. This is deliberate: this file is what any downstream consumer reads — a CDF upload, a dashboard, a spreadsheet row, anything else. None of those consumers live in this skill; they all just read this file. Do not let any of that content exist only as prose in your own output; if it's not in review.json, it isn't reusable.

{
  "appDir": "<app-dir>",
  "stats": {
    "auraCoveragePct": 0.62,
    "auraUsages": 8,
    "nonAuraUsages": 5,
    "allAuraUsages": 18,
    "allNonAuraUsages": 11,
    "couldHaveBeenAuraPct": 0.25,
    "couldHaveBeenAuraCount": 2,
    "couldHaveBeenAuraConsideredCount": 8,
    "couldHaveBeenAuraOfNonAuraPct": 0.4,
    "usageQualityFindingsCount": 3
  },
  "nonComplianceFindings": [
    { "component": "...", "usages": 1, "couldHaveBeenAura": null, "evidence": "..." }
  ],
  "usageQualityFindings": [
    { "component": "...", "location": "src/File.tsx:42", "violates": "...", "quotedRule": "..." }
  ],
  "thirdPartyUsages": [
    { "identifier": "...", "usages": 1, "packageName": "..." }
  ]
}

Keep these stats field names stable — downstream consumers index them by name, so renaming one silently breaks whatever tracks it across runs.

Step 6 — write the markdown report

Alongside review.json, always write a human-readable report, using the same method and location convention flows-design-review uses for design-review-report.md:

Choose the feedback round exactly like flows-code-review and flows-design-review do: look at <app-dir>/reviews/aura-review/. If it doesn't exist, this is round 1. Otherwise increment to the next missing feedback-round-<N>/ directory.

Write <app-dir>/reviews/aura-review/feedback-round-<N>/aura-review-report.md with this structure:

# Aura Review — <appDir> — round <N>

## Headline

| Metric | Value |
| --- | --- |
| Aura coverage | <auraCoveragePct>% (<auraUsages>/<auraUsages + nonAuraUsages> distinct component types) |
| Could have been Aura | <couldHaveBeenAuraPct>% (<couldHaveBeenAuraCount>/<couldHaveBeenAuraConsideredCount>) |
| Usage-quality findings | <usageQualityFindingsCount> |

## Non-compliance findings

| Component | Usages | Could have been | Evidence |
| --- | --- | --- | --- |
| ... | ... | ... | ... |

(If empty: "No non-compliance findings.")

## Usage-quality findings

| Component | Location | Violates | Quoted rule |
| --- | --- | --- | --- |
| ... | ... | ... | ... |

(If empty: "No usage-quality findings.")

## Excluded from judgment

Third-party components — not a design-system concern.

- <identifier> (<usages> usages, from `<packageName>`)

Populate every table straight from review.json's stats, nonComplianceFindings, usageQualityFindings, and thirdPartyUsages — don't re-derive or re-judge anything here, this is the same data restated as prose/tables for a human reader instead of a machine consumer.

Step 7 — print a one-line summary

Print to stdout, so a CI log shows the result without opening any file:

aura-review: coverage=<auraCoveragePct>% could-have-been-aura=<couldHaveBeenAuraPct>% (<n>/<total>) usage-quality-findings=<count>

Then stop. Publishing review.json or the markdown report anywhere — CDF, a spreadsheet, a dashboard — is the caller's job; this skill's output is the two files and the summary line.