refactor: optimize core engine and retrieval logic for v2.80.43

src/retrieval/lessonHelpers.ts

@@ -47,6 +47,54 @@ function parseFrontmatterType(content: string): string {
     return m ? m[1].trim().toLowerCase() : '';
 }
 
+/**
+ * Pull a specific markdown section ("## NAME ... up to the next heading") from a lesson card.
+ * Returns trimmed body text, or '' if the heading isn't found.
+ */
+function extractSection(content: string, headingRe: RegExp): string {
+    const m = content.match(headingRe);
+    if (!m || m.index === undefined) return '';
+    const after = content.slice(m.index + m[0].length);
+    const stop = after.search(/\n#{1,6}\s/);
+    const section = stop >= 0 ? after.slice(0, stop) : after;
+    return section.trim();
+}
+
+/**
+ * Slim a lesson card down to the sections that actually matter for guardrails:
+ * Mistake / Risk, Root Cause, Fix, and Prevention Checklist. Drops Situation,
+ * Applies-To, and any verbose narrative. Returned text is markdown-compatible
+ * with the original headings so the model still sees the structure.
+ *
+ * Falls back to the original content (clipped to `maxLen`) if no recognised
+ * sections are found — keeps backwards-compat for old lessons that don't
+ * follow the current template.
+ *
+ * Why: lesson cards are loaded at 2500 chars each and three cards can eat
+ * ~11K tokens. The essence sections are usually <600 chars total per card,
+ * which trims retrieval tokens by ~70% without losing the signal.
+ */
+export function extractLessonEssence(content: string, maxLen = 1200): string {
+    if (!content) return '';
+    const sections: Array<{ heading: string; body: string }> = [];
+    const want: Array<[string, RegExp]> = [
+        ['## Mistake / Risk', /^#{1,6}\s*(?:mistake\s*\/?\s*risk|mistake|risk|실수|문제)\s*$/im],
+        ['## Root Cause', /^#{1,6}\s*(?:root\s*cause|근본\s*원인|원인)\s*$/im],
+        ['## Fix', /^#{1,6}\s*(?:fix|해결|수정)\s*$/im],
+        ['## Prevention Checklist', /^#{1,6}\s*(?:prevention\s*checklist|prevention|체크리스트|예방\s*체크리스트)\s*$/im],
+    ];
+    for (const [heading, re] of want) {
+        const body = extractSection(content, re);
+        if (body && !/^<[^>]+>$/.test(body)) sections.push({ heading, body });
+    }
+    if (sections.length === 0) {
+        return content.length <= maxLen ? content.trim() : content.slice(0, maxLen).trim() + '\n…';
+    }
+    let assembled = sections.map((s) => `${s.heading}\n${s.body}`).join('\n\n');
+    if (assembled.length > maxLen) assembled = assembled.slice(0, maxLen).trim() + '\n…';
+    return assembled;
+}
+
 /** Extract the "## Prevention Checklist" bullet list from a lesson card, if present. */
 export function extractPreventionChecklist(content: string): string[] {
     if (!content) return [];