feat(skill): BM25-based smart skill retrieval for large catalogues

.changeset/skill-search-bm25.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": minor
+---
+
+Add BM25-based skill search for large catalogues. When >80 skills are installed, the system prompt switches from a full listing to a compact name-only format and the model discovers skills via the Skill tool's new `action: "search"` endpoint. Startup memory reduced ~95% via lazy content loading.

packages/agent-core/src/profile/default/system.md

@@ -148,9 +148,12 @@ Skills are grouped by scope (`Project`, `User`, `Extra`, `Built-in`) so you can
 
 ## How to use skills
 
-Identify the skills that are likely to be useful for the tasks you are currently working on, read the skill file for detailed instructions, guidelines, scripts and more.
+When you need a skill, follow this two-step process:
 
-Only read skill details when needed to conserve the context window.
+1. **Search**: Call the `Skill` tool with `action: "search"` and relevant keywords to find matching skills. The search returns ranked results instantly.
+2. **Load**: Once you identify the right skill from search results, call the `Skill` tool with `action: "load"` and the skill name to load its full instructions into context.
+
+Only read skill details when needed to conserve the context window. Do NOT guess skill names — always search first when the skill listing above does not contain enough detail.
 
 # Ultimate Reminders
 

packages/agent-core/src/skill/index.ts

@@ -2,4 +2,5 @@ export * from './builtin';
 export * from './parser';
 export * from './registry';
 export * from './scanner';
+export * from './search';
 export * from './types';

packages/agent-core/src/skill/parser.ts

@@ -1,3 +1,4 @@
+import { createReadStream } from 'node:fs';
 import { readFile } from 'node:fs/promises';
 import path from 'pathe';
 
@@ -8,6 +9,13 @@ import type { SkillDefinition, SkillMetadata, SkillSource } from './types';
 import { isSupportedSkillType } from './types';
 import { escapeXmlTags } from '../utils/xml-escape';
 
+/**
+ * Sentinel stored in SkillDefinition.content when only frontmatter was
+ * parsed at startup.  renderSkillPrompt() checks for this to decide
+ * whether to lazy-load the full body from disk.
+ */
+export const LAZY_CONTENT_SENTINEL = '\u0000LAZY';
+
 export class FrontmatterError extends Error {
   constructor(message: string, cause?: unknown) {
     super(message);
@@ -79,6 +87,56 @@ export async function parseSkillFromFile(options: ParseSkillOptions): Promise<Sk
   return parseSkillText({ ...options, text });
 }
 
+/**
+ * Read only the frontmatter from a SKILL.md file, leaving `content` empty.
+ * The body is not read from disk — callers can load it later via
+ * `readFile` + `parseSkillText` when the full content is actually needed.
+ *
+ * This avoids loading the full body of thousands of SKILL files into memory
+ * at startup when only the index (name, description) is needed.
+ */
+export async function parseSkillMetaFromFile(options: ParseSkillOptions): Promise<SkillDefinition> {
+  const stream = createReadStream(options.skillMdPath, { encoding: 'utf8', highWaterMark: 4096 });
+  let buffer = '';
+  let fenceCount = 0;
+
+  try {
+    for await (const chunk of stream) {
+      buffer += chunk;
+      const fences = buffer.match(/^---\s*$/gm);
+      if (fences !== null && fences.length >= 2) {
+        fenceCount = 2;
+        break;
+      }
+    }
+  } finally {
+    stream.close();
+  }
+
+  if (fenceCount < 2) {
+    return parseSkillFromFile(options);
+  }
+
+  // M1 fix: find second fence in the original buffer to handle CRLF correctly.
+  // split(/\r?\n/) strips \r\n as one separator but offset counting must
+  // account for the original byte positions.
+  let fencesFound = 0;
+  let offset = 0;
+  const lines = buffer.split('\n');
+  for (const line of lines) {
+    const trimmed = line.endsWith('\r') ? line.slice(0, -1) : line;
+    if (/^---\s*$/.test(trimmed)) {
+      fencesFound++;
+      if (fencesFound === 2) break;
+    }
+    offset += line.length + 1; // +1 for the \n that split removed
+  }
+
+  const frontmatterOnly = buffer.slice(0, offset + 3);
+  const result = parseSkillText({ ...options, text: frontmatterOnly });
+  return { ...result, content: LAZY_CONTENT_SENTINEL };
+}
+
 export function parseFrontmatter(text: string): ParsedFrontmatter {
   const lines = text.split(/\r?\n/);
   if (lines[0]?.trim() !== FENCE) {

packages/agent-core/src/skill/registry.ts

@@ -1,11 +1,28 @@
-import { expandSkillParameters, skillArgumentNames } from './parser';
+import { readFileSync } from 'node:fs';
+
+import { LAZY_CONTENT_SENTINEL, expandSkillParameters, skillArgumentNames, parseSkillMetaFromFile, parseSkillText } from './parser';
 import { discoverSkills, type DiscoverSkillsOptions } from './scanner';
+import { SkillSearchIndex, type SkillSearchResult } from './search';
 import type { SkillDefinition, SkillRoot, SkillSource, SkippedSkill } from './types';
 import { isInlineSkillType, normalizeSkillName } from './types';
 import { escapeXmlAttr } from '../utils/xml-escape';
 
 const LISTING_DESC_MAX = 250;
 
+/**
+ * Above this threshold, getModelSkillListing() switches to a compact
+ * name-only listing and tells the model to use the `skill_search` tool.
+ * Below it, the legacy full listing is injected into the system prompt
+ * (cheaper for prompt caching with small catalogues).
+ */
+const COMPACT_LISTING_THRESHOLD = 80;
+
+/**
+ * Above this threshold, the compact listing drops descriptions entirely
+ * and lists only skill names.
+ */
+const NAMES_ONLY_LISTING_THRESHOLD = 300;
+
 export class SkillNotFoundError extends Error {
   readonly skillName: string;
 
@@ -30,6 +47,9 @@ export class SkillRegistry {
   private readonly discoverImpl: typeof discoverSkills;
   private readonly onWarning: (message: string, cause?: unknown) => void;
   readonly sessionId?: string;
+  private readonly searchIndex = new SkillSearchIndex();
+
+  private indexDirty = false;
 
   constructor(options: SkillRegistryOptions = {}) {
     this.discoverImpl = options.discover ?? discoverSkills;
@@ -42,8 +62,13 @@ export class SkillRegistry {
       if (!this.roots.includes(root.path)) this.roots.push(root.path);
     }
 
+    // Only parse frontmatter at startup (name, description, whenToUse).
+    // The full body is loaded on demand when renderSkillPrompt() is called.
+    // This saves ~95% memory for large skill catalogues.
+
     const skills = await this.discoverImpl({
       roots,
+      parse: parseSkillMetaFromFile,
       onWarning: this.onWarning,
       onSkippedByPolicy: (skill) => this.skipped.push(skill),
       onDiscoveredSkill: (skill) => {
@@ -54,6 +79,10 @@ export class SkillRegistry {
     for (const skill of skills) {
       this.byName.set(normalizeSkillName(skill.name), skill);
     }
+
+    // Build the BM25 search index so the model can discover skills
+    // via the `skill_search` tool instead of scanning a full listing.
+    this.searchIndex.build(this.listInvocableSkills());
   }
 
   registerBuiltinSkill(skill: SkillDefinition): void {
@@ -64,6 +93,7 @@ export class SkillRegistry {
     const key = normalizeSkillName(skill.name);
     if (options.replace === true || !this.byName.has(key)) {
       this.byName.set(key, skill);
+      this.indexDirty = true;
     }
     this.indexPluginSkill(skill, options);
   }
@@ -88,8 +118,22 @@ export class SkillRegistry {
   }
 
   renderSkillPrompt(skill: SkillDefinition, rawArgs: string): string {
+    // Lazy content loading: when compact mode parsed only frontmatter,
+    // the body is empty. Read the full file now (sync, only for activated skills).
+    let content = skill.content;
+    if (content === LAZY_CONTENT_SENTINEL && skill.path.length > 0) {
+      const text = readFileSync(skill.path, 'utf8');
+      const full = parseSkillText({
+        skillMdPath: skill.path,
+        skillDirName: skill.dir.split('/').pop() ?? skill.dir,
+        source: skill.source,
+        text,
+      });
+      content = full.content;
+    }
+
     const argumentNames = skillArgumentNames(skill.metadata);
-    const content = expandSkillParameters(skill.content, rawArgs, {
+    content = expandSkillParameters(content, rawArgs, {
       skillDir: skill.dir,
       sessionId: this.sessionId,
       argumentNames,
@@ -129,16 +173,47 @@ export class SkillRegistry {
     return rendered.length === 0 ? 'No skills' : rendered;
   }
 
+  /**
+   * Search skills by free-text query. Delegates to the BM25 index.
+   * Lazily rebuilds the index if skills were registered since the last build.
+   */
+  searchSkills(query: string, limit?: number): readonly SkillSearchResult[] {
+    if (this.indexDirty) {
+      this.searchIndex.build(this.listInvocableSkills());
+      this.indexDirty = false;
+    }
+    return this.searchIndex.search(query, limit);
+  }
+
   getModelSkillListing(): string {
-    const lines = ['DISREGARD any earlier skill listings. Current available skills:'];
-    const listing = renderGroupedSkills(
-      this.listInvocableSkills().filter((skill) => skill.metadata.isSubSkill !== true),
-      formatModelSkill,
+    const invocable = this.listInvocableSkills().filter(
+      (skill) => skill.metadata.isSubSkill !== true,
     );
-    if (listing.length > 0) {
-      lines.push(listing);
+
+    // Auto-detect: small catalogue → legacy full listing.
+    // Large catalogue → compact/names-only + search-first.
+    if (invocable.length <= COMPACT_LISTING_THRESHOLD) {
+      const lines = ['DISREGARD any earlier skill listings. Current available skills:'];
+      const listing = renderGroupedSkills(invocable, formatModelSkill);
+      if (listing.length > 0) lines.push(listing);
+      return lines.length === 1 ? '' : lines.join('\n');
     }
-    return lines.length === 1 ? '' : lines.join('\n');
+
+    // Tier 2+3: Large catalogue — search-first.
+    const count = invocable.length;
+    const format = count > NAMES_ONLY_LISTING_THRESHOLD
+      ? formatNameOnlySkill
+      : formatCompactSkill;
+    const lines = [
+      `You have access to ${String(count)} registered skills.`,
+      'To find relevant skills, call the `Skill` tool with `action: "search"` and keywords from the user\'s request.',
+      'Do NOT guess skill names — always search first, then load with `action: "load"`.',
+      '',
+      'Skill names by scope:',
+    ];
+    const listing = renderGroupedSkills(invocable, format);
+    if (listing.length > 0) lines.push(listing);
+    return lines.join('\n');
   }
 }
 
@@ -182,6 +257,16 @@ function formatModelSkill(skill: SkillDefinition): readonly string[] {
   return lines;
 }
 
+/** Compact format: name + 80-char description, no path. */
+function formatCompactSkill(skill: SkillDefinition): readonly string[] {
+  return [`- ${skill.name}: ${truncate(skill.description, 80)}`];
+}
+
+/** Minimal format: name only. Used for catalogues > 200 skills. */
+function formatNameOnlySkill(skill: SkillDefinition): readonly string[] {
+  return [`- ${skill.name}`];
+}
+
 function truncate(value: string, max: number): string {
   return value.length > max ? value.slice(0, max) : value;
 }