Merge branch 'main' of https://github.com/lucide-icons/lucide into icon-usage-context

Load the Use Cases section of .github/instructions/metadata.instructions.md
into the --describe prompt so generated use-cases match the repository
style, and make useCase an array of phrases per usage group.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q2D9GYbnSV56MAVVhHPRhH

package.json

@@ -32,6 +32,7 @@
     "generate:sponsors": "node scripts/updateSponsors.mts",
     "generate:contributors": "node ./scripts/updateContributors.mts icons/*.svg",
     "generate:nextJSAliases": "node ./scripts/generateNextJSAliases.mts",
+    "icon-usage": "node ./scripts/iconUsage.mts",
     "suggest:metadata": "node ./scripts/suggestMetaData.mts",
     "suggest:metadata:debug": "node --env-file .env ./scripts/suggestMetaData.mts --debug",
     "suggest:metadata:watch": "node --env-file .env --watch ./scripts/suggestMetaData.mts",

scripts/iconUsage.md

@@ -0,0 +1,73 @@
+# Lucide Icon Usage Estimator
+
+Run with:
+
+```sh
+pnpm icon-usage trash trash-2
+```
+
+The default is a bounded first pass over `lucide-react`, with two GitHub search pages and up to 200 candidate files validated per query slice. Use `--all-packages` to include every supported package, or `--thorough` for the previous broader limits of 10 pages and 1,000 candidate files per query slice.
+
+GitHub code search requires authentication, so set `GITHUB_TOKEN` or pass `--token-env <NAME>`.
+
+Use `--verbose` or `-v` to print detailed progress logs to stderr, including queries, pages, cache hits, retries, content validation, and matches.
+
+You can pass a single icon, and icon names may be kebab-case (`trash-2`) or export names (`Trash2`, `Trash2Icon`, `LucideTrash2`). Use `--icons-file <path>` to read names from a file with one name per line, such as a list of `lucide-react` exports.
+
+## Usage In Context
+
+```sh
+pnpm icon-usage trash-2 --context
+pnpm icon-usage trash-2 --describe
+pnpm icon-usage --icons-file exports.txt --describe --max-contexts 30
+```
+
+`--context` extracts how each icon is used in every validated file: what kind of usage it is (a JSX/template element, a prop such as `icon={Trash2}`, a config object such as `{ title: 'Settings', icon: Settings }`, or a `data-lucide` attribute), the parent elements (for example `DropdownMenuItem` or `Button`), the enclosing component, label text (`aria-label`, `title`, visible text, i18n keys like `t('actions.delete')`), event handlers, links, and a short code snippet. Each usage gets a heuristic description such as `Trash2 in <DropdownMenuItem> labelled "Delete Project" within NavProjects`. Each icon gets a summary of the most common labels, parents and handlers, plus example links to the exact lines on GitHub.
+
+`--describe` does the same, then sends the collected contexts to an OpenAI model (`OPENAI_API_KEY` required; `--model` or `OPENAI_MODEL` selects the model, `gpt-5-mini` by default). The model groups them into 1 to 4 use cases per icon. Each has a `useCase` array of 1 to 3 phrases for the `use-cases` field in `icons/*.json`, written following the "Use Cases" section of `.github/instructions/metadata.instructions.md` (loaded into the prompt at runtime), a description of where the icon appears in the UI, its approximate share of the sample, and example links. The icon's existing `use-cases` are included in the prompt. Responses are cached in the cache directory.
+
+Use `--output <file>` (or `-o`) to write the results per icon to a JSON file shaped as `{ [icon]: [...usage] }`. With `--describe`, each entry is a use case (`useCase`, `description`, `sampleShare`, `examples`). Without it, each entry is an individual usage (`description`, `repository`, `url`, `kind`, `parents`, `component`, `labels`, `handlers`, `links`). `--output` implies `--context`, and the normal terminal, `--json` or `--csv` output is still printed.
+
+```sh
+pnpm icon-usage --icons-file exports.txt --describe --output usage.json
+```
+
+To keep runs bounded, searching for an icon stops once `--max-contexts` contexts (default 40) are collected, and at most two contexts are taken from each repository. When the limit is hit, the repository counts for that icon are a lower bound, and a warning says so. JSON output includes every context per match under `repositories.*.matches[].contexts`, and per-icon summaries and descriptions under `usage`.
+
+Context extraction uses heuristics, not a full parser. It works best for JSX (React, Preact, Solid), Vue, Svelte and Astro templates in the same file as the import. Angular templates in separate `.html` files are not fetched.
+
+## Method
+
+The script uses GitHub REST `GET /search/code` with package-scoped, code-extension-scoped queries, then fetches each returned file and validates common Lucide import styles before counting a repository. Counts are deduplicated by `owner/repository`.
+
+Output includes overall repository counts and per-package statistics. Per-package statistics are computed from validated matches for that package, so a repository using `trash` via `lucide-react` and `trash-2` via another package is represented correctly in both the overall totals and each package section.
+
+Supported packages are `lucide-react`, `@lucide/vue`, `@lucide/svelte`, `@lucide/angular`, `@lucide/astro`, `lucide-preact`, `lucide-solid`, `lucide`, and `@lucide/lab`. The default package is `lucide-react` because all packages plus file-scope slicing is slow for broad icons; use `--all-packages` when that broader sample is needed.
+
+Searches are split across relevant implementation file scopes using REST code search qualifiers such as `language:typescript`, `language:javascript`, `extension:vue`, and `extension:svelte`. Returned paths are still filtered by expected extensions before content is fetched, so Markdown and prompt/spec files are intentionally excluded by default because they frequently contain examples or generated implementation notes rather than live package usage.
+
+By default, queries exclude forks, archived repositories, generated files, and vendored files using GitHub code search qualifiers. Use `--include-forks`, `--include-archived`, `--include-generated`, or `--include-vendored` to include them.
+
+## GitHub API Limits
+
+GitHub REST code search is authenticated and rate-limited to 10 requests per minute. Each query can return at most 1,000 results, with up to 100 results per page. Search only covers default branches and files smaller than 384 KB.
+
+`total_count` and `incomplete_results` come from GitHub search and are recorded in JSON metadata. If a query exceeds the 1,000-result cap or returns `incomplete_results: true`, the script reports a warning and the result should be treated as incomplete rather than exhaustive.
+
+GitHub's newer code search syntax is usable for public code search queries and supports qualifiers such as `NOT is:fork`, `NOT is:archived`, `NOT is:generated`, and `NOT is:vendored`, but the documented REST code search endpoint remains capped for result retrieval.
+
+## Practical Strategy For Broad Icons
+
+For common icons that hit search caps, use the script output as lower-bound or sampled evidence. The most defensible follow-up is to narrow query slices and cite them separately, for example by package, language/path, or repository cohorts, while preserving the JSON metadata for every query performed.
+
+## Retries And Partial Results
+
+GitHub requests that time out or fail temporarily (`408`, `429`, `5xx`, rate-limit `403`, network errors) are retried up to 6 times with backoff, honouring `Retry-After` and rate-limit reset headers. If a request still fails, that search slice or file is skipped with a warning instead of stopping the run. The OpenAI client used by `--describe` retries up to 5 times.
+
+Every successful GitHub and OpenAI response is cached in the cache directory as soon as it arrives. If a run fails or is interrupted, running the same command again replays the cached responses without waiting, and only fetches what is missing.
+
+With `--output`, each icon is finished (including `--describe`) before the next one starts, and the file is rewritten after every icon. An interrupted run therefore leaves a valid JSON file with every completed icon.
+
+## Known Biases
+
+Only public GitHub repositories visible to the token are observable. Private repositories and consumers outside GitHub are absent. GitHub is not representative of every Lucide consumer. Framework/package usage can differ materially. GitHub indexing may be incomplete or delayed. Generated or vendored code can still create false positives when GitHub does not identify it. Very common icons may exceed GitHub search-result caps.

scripts/iconUsageContext.mts

@@ -0,0 +1,657 @@
+// Extracts how an icon is used inside a source file (surrounding element,
+// label text, handlers, config objects) and aggregates those usages into
+// per-icon descriptions. Used by `iconUsage.mts` with `--context`/`--describe`.
+
+export type UsageContext = {
+  line: number;
+  kind: 'element' | 'prop' | 'object' | 'reference' | 'data-attribute';
+  localName: string;
+  hostTag: string | null;
+  ancestors: string[];
+  component: string | null;
+  labels: string[];
+  i18nKeys: string[];
+  handlers: string[];
+  links: string[];
+  iconProps: string | null;
+  snippet: string;
+  description: string;
+};
+
+export type ContextMatch = {
+  repository: string;
+  path: string;
+  htmlUrl: string;
+  context: UsageContext;
+};
+
+export type IconContextSummary = {
+  contexts: number;
+  repositories: number;
+  topLabels: Array<{ value: string; count: number }>;
+  topHosts: Array<{ value: string; count: number }>;
+  topAncestors: Array<{ value: string; count: number }>;
+  topHandlers: Array<{ value: string; count: number }>;
+  kinds: Record<string, number>;
+  examples: Array<{ repository: string; htmlUrl: string; description: string }>;
+};
+
+const MAX_CONTEXTS_PER_FILE = 3;
+const SNIPPET_LINES_BEFORE = 3;
+const SNIPPET_LINES_AFTER = 3;
+const LOOKBEHIND_CHARS = 2500;
+const MAX_SNIPPET_LINE_LENGTH = 160;
+
+const escapeRegExp = (value: string) => value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+const toKebabCase = (value: string) =>
+  value
+    .replace(/([a-z0-9])([A-Z])/g, '$1-$2')
+    .replace(/([A-Z])([A-Z][a-z])/g, '$1-$2')
+    .toLowerCase();
+
+const unique = <T,>(values: T[]) => [...new Set(values)];
+
+const clean = (value: string) => value.replace(/\s+/g, ' ').trim();
+
+/**
+ * Resolves the identifiers a file uses for an icon, following `as` aliases
+ * and default imports of direct icon paths (e.g. `lucide-react/icons/trash`).
+ */
+export const localNamesFor = (
+  source: string,
+  packageName: string,
+  icon: string,
+  usageNames: string[],
+) => {
+  const names = new Set<string>();
+  const escapedPackage = escapeRegExp(packageName);
+  const namedImportRegex = new RegExp(
+    `import\\s+(?:type\\s+)?\\{([\\s\\S]*?)\\}\\s+from\\s+['"]${escapedPackage}['"]`,
+    'g',
+  );
+  for (const match of source.matchAll(namedImportRegex)) {
+    for (const specifier of (match[1] ?? '').split(',')) {
+      const parts = specifier
+        .trim()
+        .replace(/^type\s+/, '')
+        .split(/\s+as\s+/);
+      if (!usageNames.includes(parts[0]?.trim() ?? '')) continue;
+      names.add((parts[1] ?? parts[0]).trim());
+    }
+  }
+
+  const directImportRegex = new RegExp(
+    `import\\s+([A-Za-z_$][\\w$]*)\\s+from\\s+['"]${escapeRegExp(`${packageName}/icons/${icon}`)}(?:\\.js)?['"]`,
+    'g',
+  );
+  for (const match of source.matchAll(directImportRegex)) names.add(match[1]);
+
+  // Packages without imports in the template (vanilla `lucide`, Angular) are
+  // referenced by their export names directly.
+  if (names.size === 0) usageNames.forEach((name) => names.add(name));
+  return [...names];
+};
+
+/** Character ranges that are import/export statements and must not count as usage. */
+const statementRanges = (source: string) => {
+  const ranges: Array<[number, number]> = [];
+  const regex =
+    /(?:^|\n)\s*(?:import|export)\s+(?:type\s+)?(?:[\w$*\s,]*\{[\s\S]*?\}|[\w$*\s,]+?)\s*from\s*['"][^'"]+['"]/g;
+  for (const match of source.matchAll(regex)) {
+    ranges.push([match.index ?? 0, (match.index ?? 0) + match[0].length]);
+  }
+  return ranges;
+};
+
+type Tag = { name: string; start: number; end: number; attributes: string };
+
+/**
+ * Finds the end of a JSX/HTML tag that starts at `start`, skipping over `{}`
+ * expressions, strings and arrow functions so `onClick={() => a > b}` does not
+ * end the tag early.
+ */
+const findTagEnd = (source: string, start: number, limit: number) => {
+  let depth = 0;
+  let quote: string | null = null;
+  for (let index = start + 1; index < limit; index += 1) {
+    const char = source[index];
+    if (quote) {
+      if (char === quote && source[index - 1] !== '\\') quote = null;
+      continue;
+    }
+    if (char === '"' || char === "'" || char === '`') quote = char;
+    else if (char === '{') depth += 1;
+    else if (char === '}') depth = Math.max(0, depth - 1);
+    else if (char === '>' && depth === 0 && source[index - 1] !== '=') return index;
+    else if (
+      char === '<' &&
+      depth === 0 &&
+      index > start + 1 &&
+      /[A-Za-z/]/.test(source[index + 1] ?? '')
+    )
+      return -1;
+  }
+  return -1;
+};
+
+const VOID_ELEMENTS = new Set([
+  'img',
+  'input',
+  'br',
+  'hr',
+  'meta',
+  'link',
+  'source',
+  'area',
+  'col',
+  'embed',
+  'wbr',
+]);
+
+/**
+ * Scans markup before `position` and returns the still-open elements
+ * (innermost last) plus the tag `position` sits inside, if any.
+ */
+const openElementsAt = (source: string, position: number) => {
+  const windowStart = Math.max(0, position - LOOKBEHIND_CHARS);
+  const stack: Tag[] = [];
+  let host: Tag | null = null;
+  const tagRegex = /<(\/?)([A-Za-z][\w.:-]*)/g;
+  tagRegex.lastIndex = windowStart;
+  let match: RegExpExecArray | null;
+  while ((match = tagRegex.exec(source)) && match.index < position) {
+    const [, closing, name] = match;
+    // `a < b` or generics like `Array<string>` are not markup.
+    const before = source[match.index - 1] ?? '';
+    if (!closing && /[\w$)\]]/.test(before)) continue;
+    const end = findTagEnd(source, match.index, Math.min(source.length, position + 2000));
+    if (end === -1) continue;
+    const tag = { name, start: match.index, end, attributes: source.slice(match.index, end + 1) };
+    if (end >= position) {
+      if (!closing) host = tag;
+      break;
+    }
+    if (closing) {
+      const openIndex = stack.map((item) => item.name).lastIndexOf(name);
+      if (openIndex !== -1) stack.splice(openIndex);
+    } else if (source[end - 1] !== '/' && !VOID_ELEMENTS.has(name.toLowerCase())) {
+      stack.push(tag);
+    }
+    tagRegex.lastIndex = end + 1;
+  }
+  return { stack, host };
+};
+
+const LABEL_ATTRIBUTES = [
+  'aria-label',
+  'title',
+  'label',
+  'tooltip',
+  'alt',
+  'placeholder',
+  'content',
+  'text',
+  'name',
+];
+
+const attributeValues = (markup: string, names: string[]) => {
+  const values: string[] = [];
+  const pattern = names.map(escapeRegExp).join('|');
+  const regex = new RegExp(
+    `(?:^|[\\s{,])(?:${pattern})\\s*[=:]\\s*(?:\\{\\s*)?(?:"([^"]{1,80})"|'([^']{1,80})'|\`([^\`$]{1,80})\`)`,
+    'g',
+  );
+  for (const match of markup.matchAll(regex)) {
+    const value = clean(match[1] ?? match[2] ?? match[3] ?? '');
+    if (value) values.push(value);
+  }
+  return values;
+};
+
+const handlerValues = (markup: string) => {
+  const values: string[] = [];
+  const regex =
+    /(?:^|[\s(])(on[A-Z]\w*|@[\w.-]+|v-on:[\w.-]+|\(click\)|on:\w+|onclick)\s*=\s*\{?\s*["']?(?:\(\)\s*=>\s*)?([\w$.]+)/g;
+  for (const match of markup.matchAll(regex)) {
+    if (match[2] && !['function', 'async', 'e', 'event'].includes(match[2])) {
+      values.push(`${match[1]}=${match[2]}`);
+    }
+  }
+  return values;
+};
+
+const linkValues = (markup: string) =>
+  attributeValues(markup, ['href', 'to', 'url', 'path', 'route', 'routerLink']);
+
+const i18nKeys = (markup: string) => {
+  const keys: string[] = [];
+  const regex =
+    /\b(?:t|i18n\.t|\$t|translate|formatMessage|intl\.formatMessage)\(\s*(?:\{\s*id:\s*)?['"`]([\w.:-]{2,80})['"`]/g;
+  for (const match of markup.matchAll(regex)) keys.push(match[1]);
+  return keys;
+};
+
+/** Content range of an element (between its open and close tags), bounded in size. */
+const elementContent = (source: string, parent: Tag | undefined, fallbackStart: number) => {
+  const start = parent ? parent.end + 1 : fallbackStart;
+  const closing = parent ? source.indexOf(`</${parent.name}`, start) : -1;
+  const end =
+    closing === -1 ? Math.min(source.length, start + 300) : Math.min(closing, start + 600);
+  return source.slice(start, end);
+};
+
+/** Visible text inside an element, with nested tags and expressions removed. */
+const elementText = (source: string, parent: Tag | undefined, fallbackStart: number) => {
+  const content = elementContent(source, parent, fallbackStart);
+  const text = content
+    .replace(/<[^>]*>/g, ' ')
+    .replace(/\{[^{}]*\}/g, ' ')
+    .replace(/[{}]/g, ' ');
+  const value = clean(text);
+  if (!value || value.length > 80 || /[;=]|=>|\breturn\b/.test(value)) return [];
+  return [value];
+};
+
+const enclosingComponent = (source: string, position: number) => {
+  const before = source.slice(Math.max(0, position - 20000), position);
+  const regex =
+    /(?:function\s+([A-Z][\w$]*)|(?:const|let|var)\s+([A-Z][\w$]*)\s*(?::[^=]+)?=\s*(?:\([^)]*\)|[\w$]+)?\s*(?:=>|function)|class\s+([A-Z][\w$]*))/g;
+  let name: string | null = null;
+  for (const match of before.matchAll(regex)) name = match[1] ?? match[2] ?? match[3] ?? name;
+  return name;
+};
+
+const snippetAround = (lines: string[], lineIndex: number) =>
+  lines
+    .slice(Math.max(0, lineIndex - SNIPPET_LINES_BEFORE), lineIndex + SNIPPET_LINES_AFTER + 1)
+    .map((line) =>
+      line.length > MAX_SNIPPET_LINE_LENGTH ? `${line.slice(0, MAX_SNIPPET_LINE_LENGTH)}…` : line,
+    )
+    .join('\n');
+
+/** Object literal around a reference such as `{ title: 'Settings', icon: Settings }`. */
+const objectLiteralAround = (source: string, position: number) => {
+  let depth = 0;
+  let start = -1;
+  for (let index = position; index >= Math.max(0, position - 600); index -= 1) {
+    const char = source[index];
+    if (char === '}') depth += 1;
+    else if (char === '{') {
+      if (depth === 0) {
+        start = index;
+        break;
+      }
+      depth -= 1;
+    }
+  }
+  if (start === -1) return null;
+  depth = 0;
+  for (let index = start; index < Math.min(source.length, start + 1200); index += 1) {
+    const char = source[index];
+    if (char === '{') depth += 1;
+    else if (char === '}') {
+      depth -= 1;
+      if (depth === 0) return source.slice(start, index + 1);
+    }
+  }
+  return null;
+};
+
+const describe = (context: Omit<UsageContext, 'description' | 'snippet'>) => {
+  const parent = context.ancestors.at(-1);
+  const where =
+    context.kind === 'object'
+      ? 'config object'
+      : context.kind === 'prop'
+        ? `\`${context.hostTag}\` prop`
+        : parent
+          ? `<${parent}>`
+          : context.kind === 'data-attribute'
+            ? `<${context.hostTag ?? 'i'} data-lucide>`
+            : 'standalone element';
+  const parts = [`${context.localName} in ${where}`];
+  const label = context.labels[0] ?? context.i18nKeys[0];
+  if (label) parts.push(`labelled "${label}"`);
+  if (context.handlers.length > 0)
+    parts.push(`handling ${context.handlers.slice(0, 2).join(', ')}`);
+  if (context.links.length > 0) parts.push(`linking to ${context.links[0]}`);
+  if (context.iconProps && /spin|animate|loading/i.test(context.iconProps))
+    parts.push('(animated)');
+  if (context.component) parts.push(`within ${context.component}`);
+  return parts.join(' ');
+};
+
+/** Extracts up to a few usage contexts of an icon from a single source file. */
+export const extractContexts = (
+  source: string,
+  packageName: string,
+  icon: string,
+  usageNames: string[],
+): UsageContext[] => {
+  const localNames = localNamesFor(source, packageName, icon, usageNames);
+  const ranges = statementRanges(source);
+  const lines = source.split('\n');
+  const lineStarts: number[] = [];
+  lines.reduce((offset, line) => {
+    lineStarts.push(offset);
+    return offset + line.length + 1;
+  }, 0);
+  const lineAt = (position: number) => {
+    let low = 0;
+    let high = lineStarts.length - 1;
+    while (low < high) {
+      const mid = Math.ceil((low + high) / 2);
+      if (lineStarts[mid] <= position) low = mid;
+      else high = mid - 1;
+    }
+    return low;
+  };
+
+  const namePattern = unique([
+    ...localNames.map(escapeRegExp),
+    // Vue templates may use kebab-case component names.
+    ...(packageName === '@lucide/vue'
+      ? localNames.map((name) => escapeRegExp(toKebabCase(name)))
+      : []),
+  ]).join('|');
+  const occurrences: Array<{ position: number; name: string; dataAttribute: boolean }> = [];
+  for (const match of source.matchAll(
+    new RegExp(`(?<![\\w$.'"/-])(${namePattern})(?![\\w$-])`, 'g'),
+  )) {
+    occurrences.push({ position: match.index ?? 0, name: match[1], dataAttribute: false });
+  }
+  if (packageName === 'lucide') {
+    for (const match of source.matchAll(
+      new RegExp(`data-lucide\\s*=\\s*["']${escapeRegExp(icon)}["']`, 'g'),
+    )) {
+      occurrences.push({ position: match.index ?? 0, name: icon, dataAttribute: true });
+    }
+  }
+
+  const contexts: UsageContext[] = [];
+  const seenLines = new Set<number>();
+  for (const { position, name, dataAttribute } of occurrences.sort(
+    (a, b) => a.position - b.position,
+  )) {
+    if (contexts.length >= MAX_CONTEXTS_PER_FILE) break;
+    if (ranges.some(([start, end]) => position >= start && position < end)) continue;
+    const lineIndex = lineAt(position);
+    if (seenLines.has(lineIndex)) continue;
+    const lineText = lines[lineIndex];
+    if (/^\s*(\/\/|\*|\/\*)/.test(lineText)) continue;
+    // Registration calls such as `createIcons({ icons: { Trash } })` carry no context.
+    if (
+      /(?:createIcons\s*\(|LucideAngularModule\.pick\s*\(|addIcons\s*\(|icons\s*:\s*\{)[^}<]*$/.test(
+        source.slice(Math.max(0, position - 300), position),
+      )
+    )
+      continue;
+
+    const isElement = source[position - 1] === '<';
+    const { stack, host } = isElement
+      ? openElementsAt(source, position - 1)
+      : openElementsAt(source, position);
+    const ownTag = isElement
+      ? (() => {
+          const end = findTagEnd(source, position - 1, Math.min(source.length, position + 2000));
+          return end === -1 ? null : source.slice(position - 1, end + 1);
+        })()
+      : null;
+
+    let kind: UsageContext['kind'] = 'reference';
+    let hostTag: string | null = null;
+    let markupForSignals = '';
+    if (dataAttribute) {
+      kind = 'data-attribute';
+      hostTag = host?.name ?? null;
+    } else if (isElement) {
+      kind = 'element';
+    } else if (host) {
+      kind = 'prop';
+      hostTag = host.name;
+      markupForSignals = host.attributes;
+    } else {
+      const object = objectLiteralAround(source, position);
+      if (object && new RegExp(`[\\w$]+\\s*:\\s*${escapeRegExp(name)}\\b`).test(object)) {
+        kind = 'object';
+        markupForSignals = object;
+      }
+    }
+
+    const ancestors = stack.slice(-3).map((tag) => tag.name);
+    const parent = stack.at(-1);
+    const ancestorMarkup = stack
+      .slice(-2)
+      .map((tag) => tag.attributes)
+      .join(' ');
+    const signalMarkup = [markupForSignals, ancestorMarkup].join(' ');
+    const labels = unique([
+      ...(kind === 'object' ? attributeValues(markupForSignals, LABEL_ATTRIBUTES) : []),
+      ...attributeValues(
+        signalMarkup,
+        LABEL_ATTRIBUTES.filter((attr) => attr !== 'name'),
+      ),
+      ...(kind === 'element' || kind === 'data-attribute' || kind === 'prop'
+        ? elementText(source, kind === 'prop' && host ? host : parent, position)
+        : []),
+    ]).slice(0, 4);
+
+    const base = {
+      line: lineIndex + 1,
+      kind,
+      localName: name,
+      hostTag,
+      ancestors,
+      component: enclosingComponent(source, position),
+      labels,
+      i18nKeys: unique(
+        i18nKeys(
+          kind === 'object'
+            ? markupForSignals
+            : `${signalMarkup} ${elementContent(source, kind === 'prop' && host ? host : parent, position)}`,
+        ),
+      ).slice(0, 3),
+      handlers: unique(handlerValues(signalMarkup)).slice(0, 3),
+      links: unique(linkValues(signalMarkup)).slice(0, 2),
+      iconProps: ownTag
+        ? clean(ownTag.slice(name.length + 1, -1).replace(/\/$/, '')).slice(0, 120) || null
+        : null,
+    };
+    seenLines.add(lineIndex);
+    contexts.push({
+      ...base,
+      snippet: snippetAround(lines, lineIndex),
+      description: describe(base),
+    });
+  }
+  return contexts;
+};
+
+const topValues = (values: string[], limit = 8) => {
+  const counts = new Map<string, number>();
+  for (const value of values) counts.set(value, (counts.get(value) ?? 0) + 1);
+  return [...counts.entries()]
+    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))
+    .slice(0, limit)
+    .map(([value, count]) => ({ value, count }));
+};
+
+/** Aggregates the contexts collected for one icon into frequency tables. */
+export const summarizeContexts = (matches: ContextMatch[]): IconContextSummary => {
+  const contexts = matches.map((match) => match.context);
+  const kinds: Record<string, number> = {};
+  for (const context of contexts) kinds[context.kind] = (kinds[context.kind] ?? 0) + 1;
+  // Prefer one example per repository so a single project cannot dominate.
+  const examples: IconContextSummary['examples'] = [];
+  const exampleRepos = new Set<string>();
+  for (const match of matches) {
+    if (exampleRepos.has(match.repository)) continue;
+    exampleRepos.add(match.repository);
+    examples.push({
+      repository: match.repository,
+      htmlUrl: `${match.htmlUrl}#L${match.context.line}`,
+      description: match.context.description,
+    });
+    if (examples.length >= 10) break;
+  }
+  return {
+    contexts: contexts.length,
+    repositories: new Set(matches.map((match) => match.repository)).size,
+    topLabels: topValues(
+      contexts.flatMap((context) =>
+        unique([...context.labels, ...context.i18nKeys].map((value) => value.toLowerCase())),
+      ),
+    ),
+    topHosts: topValues(contexts.flatMap((context) => (context.hostTag ? [context.hostTag] : []))),
+    topAncestors: topValues(contexts.flatMap((context) => context.ancestors.slice(-1))),
+    topHandlers: topValues(
+      contexts.flatMap((context) => context.handlers.map((handler) => handler.split('=')[0])),
+    ),
+    kinds,
+    examples,
+  };
+};
+
+export type UsageDescription = {
+  useCase: string[];
+  description: string;
+  sampleShare: number;
+  examples: string[];
+};
+
+type DescribeOptions = {
+  icon: string;
+  matches: ContextMatch[];
+  summary: IconContextSummary;
+  existingUseCases: string[];
+  model: string;
+  maxPromptContexts: number;
+  readCache: (key: string) => Promise<UsageDescription[] | null>;
+  writeCache: (key: string, value: UsageDescription[]) => Promise<void>;
+};
+
+const METADATA_INSTRUCTIONS = new URL(
+  '../.github/instructions/metadata.instructions.md',
+  import.meta.url,
+);
+
+const FALLBACK_USE_CASE_GUIDELINES = `Write each use-case as a short phrase of roughly 4-12 words starting with a present participle verb (e.g. "Indicating a low battery charge level"), from the interface's perspective, naming the concrete context where useful. No trailing period, no first person, no marketing language, and never just the icon's name.`;
+
+/**
+ * Returns the "Use Cases" section of the repository's metadata instructions
+ * (including its examples), so generated use-cases follow the same style.
+ */
+const loadUseCaseGuidelines = async () => {
+  try {
+    const fs = await import('node:fs/promises');
+    const instructions = await fs.readFile(METADATA_INSTRUCTIONS, 'utf8');
+    const start = instructions.indexOf('## Use Cases');
+    if (start === -1) return FALLBACK_USE_CASE_GUIDELINES;
+    const sections = instructions.slice(start).split(/\n(?=## )/);
+    return sections
+      .filter((section, index) => index === 0 || section.startsWith('## Examples'))
+      .join('\n')
+      .trim();
+  } catch {
+    return FALLBACK_USE_CASE_GUIDELINES;
+  }
+};
+
+/**
+ * Asks an OpenAI model to cluster the collected usage contexts into distinct,
+ * real-world use cases for the icon. Responses are cached by prompt hash.
+ */
+export const describeIconUsage = async ({
+  icon,
+  matches,
+  summary,
+  existingUseCases,
+  model,
+  maxPromptContexts,
+  readCache,
+  writeCache,
+}: DescribeOptions): Promise<UsageDescription[]> => {
+  if (matches.length === 0) return [];
+  const samples = matches.slice(0, maxPromptContexts).map((match, index) => ({
+    id: index + 1,
+    repository: match.repository,
+    path: match.path,
+    component: match.context.component,
+    parents: match.context.ancestors,
+    labels: [...match.context.labels, ...match.context.i18nKeys],
+    handlers: match.context.handlers,
+    links: match.context.links,
+    snippet: match.context.snippet,
+  }));
+
+  const input = `You are maintaining the metadata for the Lucide icon library. Below are real usages of the \`${icon}\` icon found in public GitHub repositories, extracted from source code.
+
+Group these usages by what the icon means in them, and describe each group.
+
+Fields for each group:
+- useCase: an array of 1-3 use-case phrases for this group, written exactly as the use-case guidelines below describe. Add more than one phrase only when the group covers genuinely distinct uses; never rephrase the same idea. Don't repeat a phrase across groups.
+- description: one or two sentences describing how and where the icon appears in the UI (e.g. inside a ghost button in a table row, next to a "Delete" label).
+- sampleShare: the approximate fraction (0-1) of the samples that belong to this group.
+- examples: ids of up to 3 samples that best illustrate the group.
+
+Rules:
+- Return between 1 and 4 groups, most common first. Ignore samples without enough context to judge.
+- The samples are your source material: extract the function the icon serves in the interface and discard implementation details such as component names, class names or file paths.
+- Base everything on the samples; never invent use cases that are not supported by them.
+- Match the style of the existing use-cases, and avoid duplicating them.
+
+Use-case guidelines (from .github/instructions/metadata.instructions.md):
+<guidelines>
+${await loadUseCaseGuidelines()}
+</guidelines>
+
+Existing use-cases for "${icon}": ${JSON.stringify(existingUseCases)}
+
+Aggregated signals:
+${JSON.stringify({ topLabels: summary.topLabels, topParents: summary.topAncestors, topHandlers: summary.topHandlers, kinds: summary.kinds })}
+
+Samples:
+${JSON.stringify(samples, null, 1)}`;
+
+  const crypto = await import('node:crypto');
+  const cacheKey = crypto.createHash('sha256').update(`${model}\n${input}`).digest('hex');
+  const cached = await readCache(cacheKey);
+  if (cached) return cached;
+
+  const [{ default: OpenAI }, { zodTextFormat }, { default: z }] = await Promise.all([
+    import('openai'),
+    import('openai/helpers/zod'),
+    import('zod'),
+  ]);
+  const schema = z.object({
+    useCases: z.array(
+      z.object({
+        useCase: z.array(z.string()),
+        description: z.string(),
+        sampleShare: z.number(),
+        examples: z.array(z.number()),
+      }),
+    ),
+  });
+
+  // The SDK retries timeouts (408), rate limits and server errors with backoff.
+  const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY, maxRetries: 5 });
+  const response = await client.responses.create({
+    model,
+    input,
+    text: { format: zodTextFormat(schema, 'usage') },
+  });
+  const parsed = schema.parse(JSON.parse(response.output_text));
+  const descriptions = parsed.useCases.map((useCase) => ({
+    useCase: useCase.useCase,
+    description: useCase.description,
+    sampleShare: useCase.sampleShare,
+    examples: useCase.examples
+      .map((id) => matches[id - 1])
+      .filter(Boolean)
+      .map((match) => `${match.htmlUrl}#L${match.context.line}`),
+  }));
+  await writeCache(cacheKey, descriptions);
+  return descriptions;
+};