fix(agents): parse prose and rendered HTML anchors

tools/skills/agent-guidance-markdown.mjs

@@ -1,4 +1,5 @@
 import { Marked, Tokenizer } from 'marked';
+import { parseFragment } from 'parse5';
 
 // Lex only: no Markdown is rendered and no HTML is sanitized or re-emitted.
 const markdownLexer = new Marked({
@@ -34,14 +35,50 @@ function inlineText(tokens) {
         .map((token) => {
             if (token.type === 'html') return '';
             if (token.tokens) return inlineText(token.tokens);
-            return token.text ?? '';
+            return token.type === 'text'
+                ? decodeEntities(token.text ?? '')
+                : (token.text ?? '');
         })
         .join('');
 }
 
+function decodeEntities(text) {
+    return text.replace(
+        /&(?:#(?:x[0-9a-f]+|[0-9]+)|[a-z][a-z0-9]+);/giu,
+        (entity) => parseFragment(entity).childNodes[0]?.value ?? entity
+    );
+}
+
+function htmlAnchors(html) {
+    const found = [];
+    function visit(node) {
+        if (['script', 'style', 'template'].includes(node.tagName)) return;
+        for (const attribute of node.attrs ?? []) {
+            if (
+                attribute.name === 'id' ||
+                (node.tagName === 'a' && attribute.name === 'name')
+            )
+                found.push(attribute.value);
+        }
+        for (const child of node.childNodes ?? []) visit(child);
+    }
+    visit(parseFragment(html));
+    return found;
+}
+
+export function guidanceProse(markdown) {
+    function prose(token) {
+        if (['code', 'codespan', 'html'].includes(token.type)) return '';
+        if (token.items) return token.items.map(prose).join('\n');
+        if (token.tokens) return token.tokens.map(prose).join('');
+        return token.text ?? '';
+    }
+    return markdownLexer.lexer(markdown).map(prose).join('\n');
+}
+
 export function guidanceAnchors(markdown) {
     const found = new Set();
-    const explicit = new Set();
+    const html = [];
     markdownLexer.walkTokens(markdownLexer.lexer(markdown), (token) => {
         if (token.type === 'heading') {
             const slug = inlineText(token.tokens)
@@ -53,14 +90,9 @@ export function guidanceAnchors(markdown) {
             while (found.has(unique)) unique = `${slug}-${++suffix}`;
             found.add(unique);
         }
-        if (token.type === 'html') {
-            for (const match of token.raw.matchAll(
-                /<(?:a|[a-z][\w-]*)\b[^>]*\b(?:id|name)=["']([^"']+)["']/giu
-            ))
-                explicit.add(match[1]);
-        }
+        if (token.type === 'html') html.push(token.raw);
     });
-    return new Set([...found, ...explicit]);
+    return new Set([...found, ...htmlAnchors(html.join('\n'))]);
 }
 
 function isLiteralRepositoryPath(token) {

tools/skills/project.json

@@ -12,7 +12,7 @@
                 "{projectRoot}/*.mjs",
                 "{projectRoot}/project.json",
                 {
-                    "externalDependencies": ["marked"]
+                    "externalDependencies": ["marked", "parse5"]
                 }
             ],
             "options": {

tools/skills/validate-agent-guidance.mjs

@@ -3,6 +3,7 @@ import { dirname, isAbsolute, relative, resolve, sep } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import {
     guidanceAnchors as anchors,
+    guidanceProse,
     guidanceReferences as references,
 } from './agent-guidance-markdown.mjs';
 
@@ -21,27 +22,6 @@ function within(root, path) {
     );
 }
 
-// Deliberately scoped to authored guidance, not a general Markdown crawler.
-function withoutFences(markdown) {
-    let fence;
-    return markdown
-        .split(/\r?\n/u)
-        .map((line) => {
-            const match = /^\s{0,3}(`{3,}|~{3,})/u.exec(line);
-            if (match) {
-                if (!fence) fence = match[1];
-                else if (
-                    match[1][0] === fence[0] &&
-                    match[1].length >= fence.length
-                )
-                    fence = undefined;
-                return '';
-            }
-            return fence ? '' : line;
-        })
-        .join('\n');
-}
-
 async function validateReference(
     rootDir,
     source,
@@ -109,7 +89,7 @@ export async function validateAgentGuidance({ rootDir }) {
                 diagnostics.push(
                     `${source}: at most ${maxBytes} UTF-8 bytes allowed (received ${bytes})`
                 );
-            const unfenced = withoutFences(markdown);
+            const unfenced = guidanceProse(markdown);
             const imports = [...unfenced.matchAll(/^\s*@([^\s]+)\s*$/gmu)].map(
                 (match) => match[1]
             );

tools/skills/validate-agent-guidance.test.mjs

@@ -454,3 +454,47 @@ for (const [encoded, filename] of [
         );
     });
 }
+
+for (const source of ['AGENTS.md', 'CLAUDE.md']) {
+    test(`${source}: indented decorators are not imports`, async (t) => {
+        assert.deepEqual(
+            await diagnostics(t, {
+                [source]:
+                    (source === 'CLAUDE.md' ? '@AGENTS.md\n\n' : '') +
+                    '    @Injectable()\n    @docs/missing.md\n',
+            }),
+            []
+        );
+    });
+}
+for (const body of [
+    '<!-- <a id="old"></a> -->',
+    `<script>const s = '<a id="old"></a>';</script>`,
+    '<template><a id="old"></a></template>',
+]) {
+    test(`non-rendered HTML does not define anchors: ${body}`, async (t) => {
+        assert.match(
+            (
+                await diagnostics(t, {
+                    'AGENTS.md': '[Old](docs/example.md#old)',
+                    'docs/example.md': body,
+                })
+            ).join('\n'),
+            /missing anchor/
+        );
+    });
+}
+for (const [heading, anchor] of [
+    ['A &amp; B', 'a--b'],
+    ['Caf&eacute; &#x41; &#66;', 'café-a-b'],
+]) {
+    test(`heading entities produce rendered anchors: ${heading}`, async (t) => {
+        assert.deepEqual(
+            await diagnostics(t, {
+                'AGENTS.md': `[Heading](docs/example.md#${anchor})`,
+                'docs/example.md': '# ' + heading,
+            }),
+            []
+        );
+    });
+}