@autorag/librarian 2.10.3 → 2.10.5

package/dist/index.js

@@ -2109,6 +2109,25 @@ function errorMessage(error) {
 var MANIFEST_FILENAME = "manifest.json";
 var init_bundle = () => {};
 
+// src/filesystem/path-exclusions.ts
+import { posix as posix2, win32 as win322 } from "node:path";
+function createPathExclusionMatcher(excludedPaths, platform = process.platform) {
+  const paths = platform === "win32" ? win322 : posix2;
+  const normalize = (path) => {
+    const absolute = paths.resolve(path);
+    return platform === "win32" ? absolute.toLowerCase() : absolute;
+  };
+  const excluded = excludedPaths.map((path) => {
+    const absolute = normalize(path);
+    return { absolute, prefix: absolute.endsWith(paths.sep) ? absolute : `${absolute}${paths.sep}` };
+  });
+  return (path) => {
+    const absolute = normalize(path);
+    return excluded.some((entry) => absolute === entry.absolute || absolute.startsWith(entry.prefix));
+  };
+}
+var init_path_exclusions = () => {};
+
 // src/everything/client.ts
 import { spawn as spawn2 } from "node:child_process";
 import { createHash as createHash4 } from "node:crypto";
@@ -2226,6 +2245,7 @@ class EverythingClient {
   run;
   launch;
   stateDir;
+  isExcludedPath;
   binaries;
   operation = Promise.resolve();
   constructor(options) {
@@ -2235,6 +2255,7 @@ class EverythingClient {
     this.run = options.run ?? runProcess;
     this.launch = options.launch ?? launchDetached;
     this.stateDir = join6(options.root, ".autorag", "everything");
+    this.isExcludedPath = createPathExclusionMatcher(options.excludeFolders ?? [], "win32");
   }
   isSupported() {
     return this.platform === "win32";
@@ -2254,6 +2275,9 @@ class EverythingClient {
       }
       const es = { ok: true, esPath: binaries.esPath };
       const timeoutMs = this.options.timeoutMs ?? DEFAULT_TIMEOUT_MS2;
+      if ((this.options.excludeFolders?.length ?? 0) > 0) {
+        return this.searchWithExclusions(es.esPath, request, timeoutMs);
+      }
       const args = buildEverythingSearchArgs(this.instanceName, request, timeoutMs);
       const result = await this.run(es.esPath, args, timeoutMs + 5000);
       if (result.code !== 0)
@@ -2265,6 +2289,50 @@ class EverythingClient {
       }
     });
   }
+  async searchWithExclusions(esPath, request, timeoutMs) {
+    const deadline = Date.now() + timeoutMs;
+    const offset = request.offset ?? 0;
+    const maxResults = request.maxResults ?? DEFAULT_MAX_RESULTS;
+    const results = [];
+    let skipped = 0;
+    let nativeOffset = 0;
+    const timedOut = () => ({
+      ok: false,
+      reason: "search-failed",
+      message: `Everything search exceeded ${timeoutMs}ms while filtering excluded paths; narrow the query or increase the Everything timeout.`
+    });
+    for (;; ) {
+      const remainingMs = deadline - Date.now();
+      if (remainingMs <= 0)
+        return timedOut();
+      const args = buildEverythingSearchArgs(this.instanceName, { ...request, offset: nativeOffset, maxResults: MAX_FILTER_CANDIDATES }, remainingMs);
+      const response = await this.run(esPath, args, remainingMs + 5000);
+      if (response.code !== 0)
+        return this.failure("search-failed", esPath, args, response);
+      if (Date.now() >= deadline)
+        return timedOut();
+      let entries;
+      try {
+        entries = parseEverythingJson(response.stdout);
+      } catch (error) {
+        return { ok: false, reason: "search-failed", message: error.message };
+      }
+      for (const entry of entries) {
+        if (this.isExcludedPath(entry.path))
+          continue;
+        if (skipped < offset) {
+          skipped += 1;
+          continue;
+        }
+        results.push(entry);
+        if (results.length >= maxResults)
+          return { ok: true, results };
+      }
+      if (entries.length < MAX_FILTER_CANDIDATES)
+        return { ok: true, results };
+      nativeOffset += entries.length;
+    }
+  }
   index() {
     return this.serialize(async () => {
       const es = await this.ensureRunning(true);
@@ -2338,7 +2406,8 @@ class EverythingClient {
       excludeFolders: [
         ...new Set([
           join6(this.options.root, ".autorag"),
-          ...this.options.folders.map((folder) => join6(folder, ".autorag"))
+          ...this.options.folders.map((folder) => join6(folder, ".autorag")),
+          ...this.options.excludeFolders ?? []
         ])
       ]
     }));
@@ -2432,8 +2501,9 @@ function launchDetached(command, args) {
   child.on("error", () => {});
   child.unref();
 }
-var DEFAULT_TIMEOUT_MS2 = 30000, DEFAULT_STARTUP_TIMEOUT_MS = 20000, DEFAULT_INDEX_TIMEOUT_MS, DEFAULT_POLL_INTERVAL_MS = 250, DEFAULT_MAX_RESULTS = 100, MAX_OUTPUT_BYTES, VERSION_PATTERN;
+var DEFAULT_TIMEOUT_MS2 = 30000, DEFAULT_STARTUP_TIMEOUT_MS = 20000, DEFAULT_INDEX_TIMEOUT_MS, DEFAULT_POLL_INTERVAL_MS = 250, DEFAULT_MAX_RESULTS = 100, MAX_FILTER_CANDIDATES = 1000, MAX_OUTPUT_BYTES, VERSION_PATTERN;
 var init_client = __esm(() => {
+  init_path_exclusions();
   init_bundle();
   DEFAULT_INDEX_TIMEOUT_MS = 10 * 60000;
   MAX_OUTPUT_BYTES = 8 * 1024 * 1024;
@@ -2453,6 +2523,7 @@ function compare(a, b) {
   return a < b ? -1 : a > b ? 1 : 0;
 }
 async function walkFileSearch(folders, request, options = {}) {
+  const isExcluded = createPathExclusionMatcher(options.excludeFolders ?? []);
   const maxVisited = options.maxVisited ?? DEFAULT_MAX_VISITED;
   const maxMatches = options.maxMatches ?? DEFAULT_MAX_MATCHES;
   const deadline = Date.now() + (options.deadlineMs ?? DEFAULT_DEADLINE_MS);
@@ -2475,6 +2546,8 @@ async function walkFileSearch(folders, request, options = {}) {
   outer:
     while (pending.length > 0) {
       const dir = pending.pop();
+      if (isExcluded(dir))
+        continue;
       if (Date.now() >= deadline) {
         truncated = true;
         break;
@@ -2492,6 +2565,8 @@ async function walkFileSearch(folders, request, options = {}) {
         }
         visited += 1;
         const path = join7(dir, dirent.name);
+        if (isExcluded(path))
+          continue;
         if (dirent.isDirectory()) {
           if (dirent.name === ".autorag")
             continue;
@@ -2526,6 +2601,8 @@ async function walkFileSearch(folders, request, options = {}) {
       }
     }
   for (const folder of folders) {
+    if (isExcluded(folder))
+      continue;
     if (!underPath(folder) || !matches(folder))
       continue;
     if (request.kind === "files")
@@ -2552,6 +2629,7 @@ async function walkFileSearch(folders, request, options = {}) {
 }
 var DEFAULT_MAX_VISITED = 500000, DEFAULT_DEADLINE_MS = 20000, DEFAULT_MAX_MATCHES = 1e4, SORT_ORDERS;
 var init_walk = __esm(() => {
+  init_path_exclusions();
   SORT_ORDERS = {
     "name-ascending": (a, b) => compare(a.name, b.name),
     "name-descending": (a, b) => compare(b.name, a.name),
@@ -2661,6 +2739,7 @@ class FSearchClient {
   launch;
   isAlive;
   kill;
+  isExcludedPath;
   stateDir;
   dbPath;
   pidPath;
@@ -2673,6 +2752,7 @@ class FSearchClient {
     this.launch = options.launch ?? launchDetached2;
     this.isAlive = options.isProcessAlive ?? defaultIsProcessAlive;
     this.kill = options.killProcess ?? ((pid) => process.kill(pid, "SIGTERM"));
+    this.isExcludedPath = createPathExclusionMatcher(options.excludeFolders ?? []);
     this.stateDir = join8(options.root, ".autorag", "fsearch");
     this.dbPath = join8(this.stateDir, "fsearch.db");
     this.pidPath = join8(this.stateDir, "watch.pid");
@@ -2690,6 +2770,7 @@ class FSearchClient {
       if (unavailable !== undefined || !binary.ok) {
         try {
           const walked = await walkFileSearch(this.options.folders, request, {
+            excludeFolders: this.options.excludeFolders,
             maxVisited: this.options.walkMaxVisited,
             deadlineMs: this.options.walkDeadlineMs
           });
@@ -2704,7 +2785,9 @@ class FSearchClient {
           return { ok: false, reason: "search-failed", message: error.message };
         }
       }
-      const args = buildFSearchSearchArgs(this.dbPath, this.socketPath(), request);
+      const hasExclusions = (this.options.excludeFolders?.length ?? 0) > 0;
+      const nativeRequest = hasExclusions ? { ...request, offset: 0, maxResults: MAX_RESULTS_CAP } : request;
+      const args = buildFSearchSearchArgs(this.dbPath, this.socketPath(), nativeRequest);
       const result = await this.run(binary.path, args, this.options.timeoutMs ?? DEFAULT_TIMEOUT_MS3);
       if (result.code !== 0)
         return this.failure("search-failed", binary.path, args, result);
@@ -2715,6 +2798,10 @@ class FSearchClient {
         return { ok: false, reason: "search-failed", message: error.message };
       }
       let entries = parsed.entries;
+      if (hasExclusions) {
+        const isExcluded = this.isExcludedPath;
+        entries = entries.filter((entry) => !isExcluded(entry.path));
+      }
       if (request.path !== undefined) {
         const prefix = request.path.endsWith(sep3) ? request.path : `${request.path}${sep3}`;
         entries = entries.filter((entry) => entry.path === request.path || entry.path.startsWith(prefix));
@@ -2724,7 +2811,16 @@ class FSearchClient {
       const maxResults = Math.min(request.maxResults ?? DEFAULT_MAX_RESULTS2, MAX_RESULTS_CAP);
       if (entries.length > maxResults)
         entries = entries.slice(0, maxResults);
-      return { ok: true, backend: "fsearch-cli", results: entries, total: parsed.total };
+      return {
+        ok: true,
+        backend: "fsearch-cli",
+        results: entries,
+        total: parsed.total,
+        ...hasExclusions && parsed.total !== undefined && parsed.total > parsed.entries.length && entries.length < maxResults ? {
+          truncated: true,
+          note: `FSearch checked ${parsed.entries.length} of ${parsed.total} native matches before exclusions (candidate limit ${MAX_RESULTS_CAP}); results may be incomplete. Narrow the query.`
+        } : {}
+      };
     });
   }
   index() {
@@ -2918,6 +3014,7 @@ function defaultIsProcessAlive(pid) {
 }
 var DEFAULT_TIMEOUT_MS3 = 30000, DEFAULT_STARTUP_TIMEOUT_MS2 = 1e4, STOP_TIMEOUT_MS = 2000, DEFAULT_INDEX_TIMEOUT_MS2, DEFAULT_POLL_INTERVAL_MS2 = 250, DEFAULT_MAX_RESULTS2 = 100, MAX_RESULTS_CAP = 1000, MAX_OUTPUT_BYTES2, SORT_KEYS;
 var init_client2 = __esm(() => {
+  init_path_exclusions();
   init_walk();
   DEFAULT_INDEX_TIMEOUT_MS2 = 10 * 60000;
   MAX_OUTPUT_BYTES2 = 8 * 1024 * 1024;
@@ -5045,7 +5142,20 @@ function configuredVectorDimension(workspacePath) {
   const dimension = options.dimension;
   return typeof dimension === "number" ? dimension : undefined;
 }
-function rewriteEmbedderConfig(workspacePath, embedder, chunkerConfig = {}) {
+function configuredLexicalLanguage(workspacePath) {
+  let raw;
+  try {
+    raw = readFileSync8(minSyncConfigPath(workspacePath), "utf8");
+  } catch {
+    return;
+  }
+  const lexical = parse(raw).lexical;
+  if (typeof lexical !== "object" || lexical === null || Array.isArray(lexical))
+    return "simple";
+  const language = lexical.language;
+  return typeof language === "string" ? language : "simple";
+}
+function rewriteEmbedderConfig(workspacePath, embedder, managed = {}) {
   const configPath = minSyncConfigPath(workspacePath);
   let raw;
   try {
@@ -5088,8 +5198,13 @@ function rewriteEmbedderConfig(workspacePath, embedder, chunkerConfig = {}) {
   }
   if (embedder.dimension !== undefined)
     options.dimension = embedder.dimension;
-  if (chunkerConfig.maxChunkSize !== undefined)
-    chunkerOptions.max_chunk_size = chunkerConfig.maxChunkSize;
+  if (managed.maxChunkSize !== undefined)
+    chunkerOptions.max_chunk_size = managed.maxChunkSize;
+  if (managed.lexicalLanguage !== undefined) {
+    const lexical = typeof parsed.lexical === "object" && parsed.lexical !== null && !Array.isArray(parsed.lexical) ? parsed.lexical : {};
+    parsed.lexical = lexical;
+    lexical.language = managed.lexicalLanguage;
+  }
   writeFileSync6(configPath, stringify(parsed));
   return true;
 }
@@ -5101,6 +5216,66 @@ var init_embedder_config = __esm(() => {
   };
 });
 
+// src/minsync/kiwi.ts
+import { existsSync as existsSync10 } from "node:fs";
+import { homedir as homedir2 } from "node:os";
+import { posix as posix3, win32 as win323 } from "node:path";
+function kiwiAvailable(environment = defaultEnvironment()) {
+  const { env, exists } = environment;
+  const library = env.KIWI_LIBRARY_PATH;
+  const model = env.KIWI_MODEL_PATH;
+  const libraryFound = library !== undefined && library !== "" ? exists(library) : libraryCandidates(environment).some(exists);
+  const modelFound = model !== undefined && model !== "" ? exists(model) : modelCandidates(environment).some(exists);
+  return libraryFound && modelFound;
+}
+function defaultEnvironment() {
+  let home;
+  try {
+    home = homedir2();
+  } catch {
+    home = undefined;
+  }
+  return { env: process.env, platform: process.platform, home, exists: existsSync10 };
+}
+function userRoots({ env, platform, home }) {
+  if (platform === "win32") {
+    const roots = [];
+    if (env.LOCALAPPDATA)
+      roots.push(win323.join(env.LOCALAPPDATA, "kiwi"));
+    const profile = env.USERPROFILE ?? home;
+    if (profile)
+      roots.push(win323.join(profile, "AppData", "Local", "kiwi"));
+    roots.push("C:\\kiwi", "C:\\Program Files\\Kiwi");
+    return roots;
+  }
+  return home ? [posix3.join(home, ".local", "kiwi")] : [];
+}
+function libraryCandidates(environment) {
+  const { platform } = environment;
+  if (platform === "win32")
+    return userRoots(environment).map((root) => win323.join(root, "lib", "kiwi.dll"));
+  const file = platform === "darwin" ? "libkiwi.dylib" : "libkiwi.so";
+  const system = platform === "darwin" ? ["/usr/local/lib", "/opt/homebrew/lib"] : ["/usr/local/lib", "/usr/lib"];
+  return [
+    ...userRoots(environment).map((root) => posix3.join(root, "lib", file)),
+    ...system.map((dir) => posix3.join(dir, file))
+  ];
+}
+function modelCandidates(environment) {
+  const { platform } = environment;
+  if (platform === "win32")
+    return userRoots(environment).map((root) => win323.join(root, "models", "cong", "base"));
+  const system = platform === "darwin" ? ["/usr/local/models", "/opt/homebrew/models", "/usr/local/share/kiwi/models"] : ["/usr/local/models", "/usr/local/share/kiwi/models", "/usr/share/kiwi/models"];
+  return [
+    ...userRoots(environment).map((root) => posix3.join(root, "models", "cong", "base")),
+    ...system.map((dir) => posix3.join(dir, "cong", "base"))
+  ];
+}
+var init_kiwi = () => {};
+
+// src/minsync/lexical-language.ts
+var MINSYNC_LEXICAL_LANGUAGE = "multilingual";
+
 // src/minsync/local-embedder.ts
 import { spawn as spawn7 } from "node:child_process";
 import { request } from "node:http";
@@ -5289,7 +5464,7 @@ var init_process = __esm(() => {
 });
 
 // src/minsync/client.ts
-import { existsSync as existsSync10, readFileSync as readFileSync9, writeFileSync as writeFileSync7 } from "node:fs";
+import { existsSync as existsSync11, readFileSync as readFileSync9, writeFileSync as writeFileSync7 } from "node:fs";
 import { join as join14 } from "node:path";
 
 class MinSyncClient {
@@ -5384,7 +5559,7 @@ class MinSyncClient {
     return previous === undefined || JSON.stringify(previous) !== JSON.stringify(identity);
   }
   async sync(force = false, signal) {
-    if (!existsSync10(this.binaryPath))
+    if (!existsSync11(this.binaryPath))
       throw minSyncBinaryMissingError();
     let effective = { config: {} };
     try {
@@ -5422,14 +5597,16 @@ class MinSyncClient {
       }
     }
     const spawnOptions = signal === undefined ? {} : { signal };
-    const initialized = existsSync10(minSyncConfigPath(this.workspacePath));
+    const lexical = this.resolveLexicalLanguage();
+    const initialized = existsSync11(minSyncConfigPath(this.workspacePath));
     if (!initialized) {
       const initArgs = ["init", "--format", "json"];
       if (this.embedder?.id) {
         initArgs.push("--embedder", this.embedder.id);
       }
+      initArgs.push("--language", lexical.language);
       const init = await this.spawn(initArgs, spawnOptions);
-      if (!init.ok || !existsSync10(minSyncConfigPath(this.workspacePath))) {
+      if (!init.ok || !existsSync11(minSyncConfigPath(this.workspacePath))) {
         return {
           ok: false,
           synced: 0,
@@ -5441,20 +5618,29 @@ class MinSyncClient {
     const chunkSizeChanged = this.maxChunkSize !== undefined && configuredMaxChunkSize(this.workspacePath) !== this.maxChunkSize;
     const dimensionChanged = embedder.dimension !== undefined && configuredVectorDimension(this.workspacePath) !== embedder.dimension;
     const configPath = minSyncConfigPath(this.workspacePath);
-    const shouldRewriteConfig = this.embedder !== undefined || this.maxChunkSize !== undefined || effective.identity !== undefined;
-    const originalConfig = shouldRewriteConfig ? readConfigSnapshot(configPath) : undefined;
-    const configRewritten = shouldRewriteConfig && rewriteEmbedderConfig(this.workspacePath, embedder, { maxChunkSize: this.maxChunkSize });
+    const originalConfig = readConfigSnapshot(configPath);
+    const configRewritten = rewriteEmbedderConfig(this.workspacePath, embedder, {
+      maxChunkSize: this.maxChunkSize,
+      lexicalLanguage: lexical.language
+    });
     const restoreConfig = () => {
       if (configRewritten && originalConfig !== undefined)
         writeFileSync7(configPath, originalConfig);
     };
     try {
-      return await this.checkAndSync(force || chunkSizeChanged || dimensionChanged, signal, effective, restoreConfig);
+      const result = await this.checkAndSync(force || chunkSizeChanged || dimensionChanged, signal, effective, restoreConfig);
+      return result.ok && lexical.fallback !== undefined ? { ...result, lexicalFallback: lexical.fallback } : result;
     } catch (error) {
       restoreConfig();
       throw error;
     }
   }
+  resolveLexicalLanguage() {
+    const requested = MINSYNC_LEXICAL_LANGUAGE;
+    if (kiwiAvailable() || configuredLexicalLanguage(this.workspacePath) === requested)
+      return { language: requested };
+    return { language: "simple", fallback: { requested, used: "simple" } };
+  }
   lockHeldResult(stderr) {
     return {
       ok: false,
@@ -5507,7 +5693,7 @@ class MinSyncClient {
     }
     const identityChanged = effective.identity !== undefined && this.identityMismatch(effective.identity);
     const fullReindex = forceFull || identityChanged;
-    const syncArgs = existsSync10(cursorPath) && !fullReindex ? ["sync", "--format", "json"] : ["sync", "--full", "--format", "json"];
+    const syncArgs = existsSync11(cursorPath) && !fullReindex ? ["sync", "--format", "json"] : ["sync", "--full", "--format", "json"];
     const result = await this.spawn(syncArgs, spawnOptions);
     if (!result.ok && result.code === MINSYNC_LOCK_HELD_EXIT_CODE) {
       restoreConfig();
@@ -5528,7 +5714,7 @@ class MinSyncClient {
         }
       };
     }
-    if (!existsSync10(cursorPath)) {
+    if (!existsSync11(cursorPath)) {
       restoreConfig();
       return { ok: false, synced: 0, workspacePath: this.workspacePath, reason: "not-ready: missing cursor" };
     }
@@ -5549,7 +5735,7 @@ class MinSyncClient {
     };
   }
   async query(text, topK, mode = "vector", signal) {
-    if (!existsSync10(this.binaryPath))
+    if (!existsSync11(this.binaryPath))
       throw minSyncBinaryMissingError();
     const result = await this.spawn(["query", "--format", "json", "--mode", mode, "-k", String(topK), text], {
       ...signal === undefined ? {} : { signal }
@@ -5643,6 +5829,7 @@ var MINSYNC_OLLAMA_MIGRATION_MESSAGE = "This workspace uses the legacy 768-dimen
 var init_client4 = __esm(() => {
   init_embedder_config();
   init_errors();
+  init_kiwi();
   init_local_embedder();
   init_process();
   API_KEY_ENV_PATTERN = /^[A-Za-z_][A-Za-z0-9_]*$/;
@@ -5665,14 +5852,14 @@ var init_client4 = __esm(() => {
 
 // src/minsync/installer.ts
 import { createHash as createHash9 } from "node:crypto";
-import { copyFileSync, createWriteStream as createWriteStream2, existsSync as existsSync11, mkdirSync as mkdirSync8, renameSync as renameSync4, rmSync as rmSync5 } from "node:fs";
+import { copyFileSync, createWriteStream as createWriteStream2, existsSync as existsSync12, mkdirSync as mkdirSync8, renameSync as renameSync4, rmSync as rmSync5 } from "node:fs";
 import { chmod as chmod2, mkdtemp } from "node:fs/promises";
 import { get } from "node:https";
 import { basename as basename7, delimiter as delimiter2, dirname as dirname7, join as join15 } from "node:path";
 import { pipeline } from "node:stream/promises";
 async function ensureMinSyncBinary(options) {
   const binaryPath = join15(options.root, ".autorag", "bin", executableName(options.platform ?? process.platform));
-  if (existsSync11(binaryPath))
+  if (existsSync12(binaryPath))
     return { binaryPath, version: "cached" };
   mkdirSync8(dirname7(binaryPath), { recursive: true });
   const cargoInstaller = options.cargoInstaller ?? ((destination) => installMinSyncFromCargoIfAvailable(options, destination));
@@ -5720,7 +5907,7 @@ async function installMinSyncFromCargo(options, destination, cargo) {
     throw new MinSyncReleaseError(result.stderr || `Could not install MinSync ${MINSYNC_VERSION} from crates.io`);
   }
   const installedBinary = join15(cargoRoot, "bin", executableName(options.platform ?? process.platform));
-  if (!existsSync11(installedBinary)) {
+  if (!existsSync12(installedBinary)) {
     throw new MinSyncReleaseError(`Cargo did not produce the MinSync ${MINSYNC_VERSION} binary`);
   }
   copyFileSync(installedBinary, destination);
@@ -5852,7 +6039,7 @@ function lookupCargo(env, platform) {
     if (dir.length === 0)
       continue;
     const candidate = join15(dir, execName);
-    if (existsSync11(candidate))
+    if (existsSync12(candidate))
       return candidate;
   }
   return;
@@ -5883,7 +6070,7 @@ function isGitHubAsset(value) {
 function isRecord5(value) {
   return typeof value === "object" && value !== null;
 }
-var LATEST_RELEASE_URL = "https://api.github.com/repos/NomaDamas/MinSync/releases/latest", SHA256_HEX_PATTERN, MINSYNC_VERSION = "0.4.8", MINSYNC_CARGO_INSTALL_TIMEOUT_MS, CARGO_PROBE_TIMEOUT_MS = 5000, MinSyncReleaseError;
+var LATEST_RELEASE_URL = "https://api.github.com/repos/NomaDamas/MinSync/releases/latest", SHA256_HEX_PATTERN, MINSYNC_VERSION = "0.4.10", MINSYNC_CARGO_INSTALL_TIMEOUT_MS, CARGO_PROBE_TIMEOUT_MS = 5000, MinSyncReleaseError;
 var init_installer2 = __esm(() => {
   init_process();
   SHA256_HEX_PATTERN = /^[a-f0-9]{64}$/;
@@ -5911,7 +6098,7 @@ var init_paths2 = __esm(() => {
 import { randomUUID as randomUUID4 } from "node:crypto";
 import {
   copyFileSync as copyFileSync2,
-  existsSync as existsSync12,
+  existsSync as existsSync13,
   lstatSync as lstatSync2,
   mkdirSync as mkdirSync9,
   readdirSync as readdirSync2,
@@ -5926,8 +6113,8 @@ function syncMinSyncWorkspace(root, options = {}) {
   const filesRoot = join17(workspacePath, "files");
   const index = loadMirrorIndex(root);
   const candidates = Object.values(index.entries).sort((a, b) => a.virtualPath.localeCompare(b.virtualPath));
-  const excluded = candidates.filter((entry) => normalizeVirtualPath(entry.virtualPath) !== entry.virtualPath || !existsSync12(entry.outputPath)).map((entry) => entry.virtualPath);
-  const entries = candidates.filter((entry) => normalizeVirtualPath(entry.virtualPath) === entry.virtualPath && existsSync12(entry.outputPath)).map((entry) => {
+  const excluded = candidates.filter((entry) => normalizeVirtualPath(entry.virtualPath) !== entry.virtualPath || !existsSync13(entry.outputPath)).map((entry) => entry.virtualPath);
+  const entries = candidates.filter((entry) => normalizeVirtualPath(entry.virtualPath) === entry.virtualPath && existsSync13(entry.outputPath)).map((entry) => {
     const minSyncPath = minSyncDocumentPath(workspacePath, entry.virtualPath);
     return {
       virtualPath: entry.virtualPath,
@@ -6064,7 +6251,7 @@ var init_workspace = __esm(() => {
 });
 
 // src/minsync/method.ts
-import { accessSync, constants, existsSync as existsSync13, realpathSync as realpathSync4 } from "node:fs";
+import { accessSync, constants, existsSync as existsSync14, realpathSync as realpathSync4 } from "node:fs";
 import { basename as basename9, delimiter as delimiter3, join as join18, normalize as normalize2 } from "node:path";
 function lookupInPath(env) {
   const pathEnv = env.PATH;
@@ -6160,19 +6347,19 @@ class MinSyncVectorMethod {
   }
   isBinaryMissing() {
     if (this.binaryPath !== undefined)
-      return !existsSync13(this.binaryPath);
+      return !existsSync14(this.binaryPath);
     if (lookupInPath(process.env) !== undefined)
       return false;
     const cachedBinary = join18(this.root, ".autorag", "bin", executableName(process.platform));
-    if (existsSync13(cachedBinary))
+    if (existsSync14(cachedBinary))
       return false;
     return true;
   }
   isExplicitBinaryMissing() {
-    return this.binaryPath !== undefined && !existsSync13(this.binaryPath);
+    return this.binaryPath !== undefined && !existsSync14(this.binaryPath);
   }
   isReady() {
-    return existsSync13(join18(this.workspacePath, ".minsync", "cursor.json"));
+    return existsSync14(join18(this.workspacePath, ".minsync", "cursor.json"));
   }
   async retrieve(query, options) {
     if ((workspaceSyncInFlight.get(this.workspacePath) ?? 0) > 0) {
@@ -6225,7 +6412,7 @@ class MinSyncVectorMethod {
     return results;
   }
   async resolveBinary(options = {}) {
-    if (this.binaryPath && existsSync13(this.binaryPath))
+    if (this.binaryPath && existsSync14(this.binaryPath))
       return this.binaryPath;
     if (this.binaryPath !== undefined)
       throw minSyncBinaryMissingError();
@@ -6233,7 +6420,7 @@ class MinSyncVectorMethod {
     if (pathBinary)
       return pathBinary;
     const cachedBinary = join18(this.root, ".autorag", "bin", executableName(process.platform));
-    if (existsSync13(cachedBinary))
+    if (existsSync14(cachedBinary))
       return cachedBinary;
     if (this.autoInstall && options.install !== false) {
       try {
@@ -6936,7 +7123,7 @@ var init_defaults = __esm(() => {
 });
 
 // src/parser/isolated.ts
-import { existsSync as existsSync14 } from "node:fs";
+import { existsSync as existsSync15 } from "node:fs";
 import { dirname as dirname10, join as join21 } from "node:path";
 import { fileURLToPath as fileURLToPath2, pathToFileURL } from "node:url";
 import { Worker } from "node:worker_threads";
@@ -7075,7 +7262,7 @@ function defaultWorkerUrl() {
     join21(here, "parser", "parse-worker.js"),
     join21(here, "..", "parser", "parse-worker.js")
   ]) {
-    if (existsSync14(candidate))
+    if (existsSync15(candidate))
       return pathToFileURL(candidate);
   }
   throw new Error(`The AutoRAG parse worker entry was not found near ${here}.`);
@@ -7087,7 +7274,7 @@ var init_isolated = __esm(() => {
 });
 
 // src/mirror/sync.ts
-import { existsSync as existsSync15, rmSync as rmSync7 } from "node:fs";
+import { existsSync as existsSync16, rmSync as rmSync7 } from "node:fs";
 import { opendir, readFile as readFile3, stat as stat5 } from "node:fs/promises";
 import { extname as extname3, resolve as resolve11, sep as sep6 } from "node:path";
 import { setTimeout as sleep3 } from "node:timers/promises";
@@ -7279,7 +7466,7 @@ async function syncWithRegistry(options, registry, isolated) {
     }
     const previousEntry = previous.entries[entry.virtualPath];
     const outputPath = parsedOutputPath(options.root, entry.virtualPath);
-    const unchanged = !options.force && previousEntry?.sourceMtimeNs === entry.mtimeNs && previousEntry.sourceSizeBytes === entry.sizeBytes && previousEntry.parserName === parser.name && existsSync15(outputPath);
+    const unchanged = !options.force && previousEntry?.sourceMtimeNs === entry.mtimeNs && previousEntry.sourceSizeBytes === entry.sizeBytes && previousEntry.parserName === parser.name && existsSync16(outputPath);
     if (!unchanged) {
       let bytes;
       try {
@@ -7660,7 +7847,7 @@ function removePrevious(root, index, virtualPath) {
   return 1;
 }
 function removeFile(path) {
-  if (existsSync15(path))
+  if (existsSync16(path))
     rmSync7(path, { force: true });
 }
 var DEFAULT_MAX_SOURCE_BYTES, MIRROR_CHECKPOINT_EVERY = 25, DEFAULT_MAX_DELETE_RATIO = 0.3, MASS_DELETE_FLOOR = 100, DEFAULT_MAX_SCANNED_ENTRIES = 2000000, DEFAULT_MAX_WALK_DEPTH = 64, PRODUCT_ARTIFACT_FILE_NAMES, SKIP_DIR_NAMES, PARSER_FAILURE_MESSAGE = "The registered parser failed on this file; it was skipped during indexing.", PARSER_CRASH_MESSAGE = "The registered parser crashed unexpectedly on this file; it was skipped during indexing.", DEFAULT_TRANSIENT_RETRIES = 3, DEFAULT_TRANSIENT_BACKOFF_MS = 50, TRANSIENT_ERROR_CODES, WORKSPACE_ERROR_CODES, MirrorWorkspaceError, STANDALONE_OCR_IMAGE_EXTENSIONS, RETRYABLE_SKIP_MESSAGES, SEARCH_ROOT_UNAVAILABLE_MESSAGES;
@@ -9186,12 +9373,43 @@ function digitsEndAt(text, from) {
     end += 1;
   return end;
 }
-function continuesAsProse(answer, from) {
-  if (answer[from] !== " " && answer[from] !== " ")
+function continuesAsProse(answer, openIndex, from) {
+  let cursor = from;
+  let leadingEmphasis = 0;
+  while (openIndex - 1 - leadingEmphasis >= 0 && (answer[openIndex - 1 - leadingEmphasis] === "*" || answer[openIndex - 1 - leadingEmphasis] === "_")) {
+    leadingEmphasis += 1;
+  }
+  if (leadingEmphasis > 0) {
+    let trailingEmphasis = 0;
+    while (cursor + trailingEmphasis < answer.length && trailingEmphasis < leadingEmphasis && answer[cursor + trailingEmphasis] === answer[openIndex - 1 - trailingEmphasis]) {
+      trailingEmphasis += 1;
+    }
+    if (trailingEmphasis === leadingEmphasis)
+      cursor += trailingEmphasis;
+  }
+  const tokenStart = openIndex - leadingEmphasis;
+  const linePrefix = answer.slice(answer.lastIndexOf(`
+`, tokenStart - 1) + 1, tokenStart);
+  const atHeadingStart = /^[ \t>*-]*$/u.test(linePrefix);
+  if (atHeadingStart && answer[cursor] === ":")
+    cursor += 1;
+  if (answer[cursor] !== " " && answer[cursor] !== " ")
     return false;
-  const next = answer.codePointAt(from + 1);
+  cursor += 1;
+  while (cursor < answer.length && (answer[cursor] === "*" || answer[cursor] === "_" || answer[cursor] === '"' || answer[cursor] === "'" || answer[cursor] === "“" || answer[cursor] === "‘")) {
+    cursor += 1;
+  }
+  const next = answer.codePointAt(cursor);
   return next !== undefined && /[\p{L}\p{N}]/u.test(String.fromCodePoint(next));
 }
+function isLegalPinpoint(answer, openIndex, afterCloseIndex) {
+  if (PINPOINT_RANGE_AHEAD.test(answer.slice(afterCloseIndex, afterCloseIndex + 16)))
+    return true;
+  if (PINPOINT_RANGE_BEHIND.test(answer.slice(Math.max(0, openIndex - 16), openIndex)))
+    return true;
+  const prefixWindow = answer.slice(Math.max(0, openIndex - 120), openIndex);
+  return PARA_PINPOINT_PREFIX.test(prefixWindow) || CLAUSE_START_AT_PREFIX.test(prefixWindow) || CASE_CITATION_AT_PREFIX.test(prefixWindow);
+}
 function evidenceMarkerAt(answer, index) {
   const ids = [];
   let cursor = index + 1;
@@ -9257,7 +9475,7 @@ function answerMarkers(answer, isFile = () => false) {
     }
     if (char === "[") {
       const digitsEnd = digitsEndAt(answer, index + 1);
-      if (digitsEnd > index + 1 && answer[digitsEnd] === "]" && answer[digitsEnd + 1] !== "(" && !continuesAsProse(answer, digitsEnd + 1)) {
+      if (digitsEnd > index + 1 && answer[digitsEnd] === "]" && answer[digitsEnd + 1] !== "(" && !continuesAsProse(answer, index, digitsEnd + 1) && !isLegalPinpoint(answer, index, digitsEnd + 1)) {
         markers.push({
           kind: "number",
           start: leadingStart(index),
@@ -9329,7 +9547,14 @@ function assertResultsMappingOneToOne(label, results, mapping) {
   ].filter((note) => note !== undefined);
   throw new Error(`${label}: result numbers and mapping numbers must be one-to-one, but results contain ${formatCitationList(resultNumbers)} ` + `and mapping contains ${formatCitationList(mappingNumbers)}` + (duplicateNotes.length > 0 ? ` (${duplicateNotes.join("; ")})` : "") + ". Give every result a unique number with exactly one mapping entry carrying the same number and no mapping entry without a result.");
 }
-var MAX_FILE_MARKER_CLOSERS = 8, FILE_MARKER_PREFIX = "[file:";
+var MAX_FILE_MARKER_CLOSERS = 8, FILE_MARKER_PREFIX = "[file:", PARA_PINPOINT_PREFIX, CLAUSE_START_AT_PREFIX, CASE_CITATION_AT_PREFIX, PINPOINT_RANGE_AHEAD, PINPOINT_RANGE_BEHIND;
+var init_citations = __esm(() => {
+  PARA_PINPOINT_PREFIX = /(?:^|[^\p{L}\p{N}])(?:paras?\.?|paragraphs?)(?:[ \t\u00a0]+\[\d+\][ \t\u00a0]*(?:[-–—]|,(?:[ \t\u00a0]*(?:and|or)\b)?|and\b|or\b))*[ \t\u00a0]*$/iu;
+  CLAUSE_START_AT_PREFIX = /(?:^|[\n.!?؛;:()]|[>*-])[ \t\u00a0]*At(?:[ \t\u00a0]+\[\d+\][ \t\u00a0]*(?:[-–—]|,(?:[ \t\u00a0]*(?:and|or)\b)?|and\b|or\b))*[ \t\u00a0]+$/u;
+  CASE_CITATION_AT_PREFIX = /(?:\[\d{4}\][^\n.;]*?|\b(?:v|vs\.?)\s+[^\n.;]{1,60}?|\bs(?:ection)?[ \t\u00a0]+\d+[A-Za-z]?(?:\(\d+\))?[^\n.;]{0,30}?)[ \t\u00a0]+at(?:[ \t\u00a0]+\[\d+\][ \t\u00a0]*(?:[-–—]|,(?:[ \t\u00a0]*(?:and|or)\b)?|and\b|or\b))*[ \t\u00a0]+$/iu;
+  PINPOINT_RANGE_AHEAD = /^[ \t\u00a0]*[-–—][ \t\u00a0]*\[\d+\]/u;
+  PINPOINT_RANGE_BEHIND = /\[\d+\][ \t\u00a0]*[-–—][ \t\u00a0]*$/u;
+});
 
 // src/agent/evidence-ledger.ts
 import { createHash as createHash10 } from "node:crypto";
@@ -9613,6 +9838,7 @@ function deriveResultsFromAnswer(text, ledger, options) {
 }
 var MAX_EXCERPT_CHARS = 400, MAX_CLAIM_CHARS = 280, CLAIM_LOOKBACK_CHARS = 600, SENTENCE_END, LEADING_LIST_MARK, EVIDENCE_MARKER_TEXT, LEADING_BLANK;
 var init_answer_citations = __esm(() => {
+  init_citations();
   init_evidence_ledger();
   SENTENCE_END = /[.!?。！？]/u;
   LEADING_LIST_MARK = /^[\s\-*•>]+/u;
@@ -9621,7 +9847,7 @@ var init_answer_citations = __esm(() => {
 });
 
 // src/agent/answer-guidelines.ts
-var ANSWER_IMAGE_EMBED_RULE = "**Images**: when a retrieved result is itself an image file (png, jpg, jpeg, gif, webp, svg, bmp, avif, heic, tiff) and showing it answers the question, embed it in the answer as a markdown image on its own line: `![short description](<absolute source path>)`, followed by its citation, e.g. `![Q3 revenue chart](</data/reports/q3 chart.png>) [e2]`. Use only the real source path of retrieved evidence. Never invent a path, use relative paths, or embed remote URLs. Apart from these image embeds, the no-file-paths rule stands.", ANSWER_IMAGE_DELTA_RULE = "An image already embedded in the first answer must not be embedded again.", ANSWER_CITATION_RULE = "**Citations**: cite supporting evidence inline with the evidence id shown next to each retrieved result, in square brackets (e.g. [e3], or [e3][e7] for two). Copy ids exactly; never invent one. For a local file you opened yourself with `bash`, cite `[file:<absolute path>]`. Cite the evidence that backs a sentence at the end of that sentence. Do not cite numbers, paths, or anything else.";
+var ANSWER_IMAGE_EMBED_RULE = "**Images**: when a retrieved result is itself an image file (png, jpg, jpeg, gif, webp, svg, bmp, avif, heic, tiff) and showing it answers the question, embed it in the answer as a markdown image on its own line: `![short description](<absolute source path>)`, followed by its citation, e.g. `![Q3 revenue chart](</data/reports/q3 chart.png>) [e2]`. Use only the real source path of retrieved evidence. Never invent a path, use relative paths, or embed remote URLs. Apart from these image embeds, the no-file-paths rule stands.", ANSWER_IMAGE_DELTA_RULE = "An image already embedded in the first answer must not be embedded again.", ANSWER_CITATION_RULE = "**Citations**: cite supporting evidence inline with the evidence id shown next to each retrieved result, in square brackets (e.g. [e3], or [e3][e7] for two). Copy ids exactly; never invent one. For a local file you opened yourself with `bash`, cite `[file:<absolute path>]`. Cite the evidence that backs a sentence at the end of that sentence. Do not cite numbers, paths, or anything else.", ANSWER_DELIVERABLE_RULE = '**Deliverables and artifacts**: if the user asks you to draft, write, generate, or prepare a deliverable or artifact (a reply email, letter, legal opinion, summary document, standard-format case citation block, checklist, report, or workflow document — including requests phrased as "how should I reply..." to a specific message), produce the finished, ready-to-use artifact as a file (using `write`, `edit`, or `bash` in local sessions, without overwriting existing source files) in a file format, extension, and structure aligned with the context and domain (e.g. `.eml` or `.md`/`.txt` with headers for an email reply; `.md`, `.txt`, `.docx`, `.hwp`, or `.pdf` matching workspace conventions for a legal opinion, insertable case-citation block, or summary document; `.csv`, `.xlsx`, or `.json` for structured tabular data), containing the complete unabridged artifact (including required headers, precedent passages, pinpoint paragraphs, and standard-format citations) inside the file instead of merely describing how to produce it. Do NOT dump the full artifact text into your reply: unless the user explicitly requests otherwise, keep your reply concise within the at-most-5-bullet direct-answer format, clearly stating the created file\'s exact path/location and briefly summarizing what was generated with inline citations to the underlying source evidence (`[eN]` or `[file:<source path>]`); never cite the newly created deliverable file itself as `[file:...]`.', ANSWER_DELIVERABLE_DELTA_RULE = "If a deliverable or artifact was requested and the first answer could not write the file yet (the fast phase runs without tool calls) or only gave advice about it, create the complete context-aligned deliverable file now during verification and report the created file's location and a brief summary concisely.";
 
 // src/agent/datasource-skill.ts
 import { Type as Type3 } from "typebox";
@@ -11346,6 +11572,20 @@ function interactiveKeyOf(data) {
     return "followUp";
   return;
 }
+function stripTelegramEnvelope(text) {
+  return text.replace(/^\s*\[telegram\]\s?/u, "").replace(/\n*\[time\][^\n]*/gu, "").trim();
+}
+function userInputText(event) {
+  if (event.source === "interactive")
+    return event.text.trim();
+  if (event.source === "extension" && event.text.trimStart().startsWith(TELEGRAM_INPUT_PREFIX)) {
+    return stripTelegramEnvelope(event.text);
+  }
+  if (event.source === "extension" && event.text.trimStart().startsWith(WHATSAPP_INPUT_PREFIX)) {
+    return event.text.trim();
+  }
+  return;
+}
 function createAutoRAGInteractiveExtension(getSystemPrompt, contextTransform, options, updateNotice) {
   return (pi) => {
     pi.on("before_agent_start", () => ({ systemPrompt: getSystemPrompt() }));
@@ -11376,8 +11616,8 @@ function createAutoRAGInteractiveExtension(getSystemPrompt, contextTransform, op
       });
     }
     pi.on("input", (event) => {
-      const text = event.text.trim();
-      if (event.source !== "interactive" || text.length === 0)
+      const text = userInputText(event);
+      if (text === undefined || text.length === 0)
         return;
       if (text.startsWith("/") && isSearchRunning?.() !== true)
         return;
@@ -11474,7 +11714,9 @@ async function createAutoRAGPiInteractiveRuntime(options) {
       agentDir: runtimeAgentDir,
       modelRuntime,
       settingsManager,
+      ...options.extensionFlags === undefined ? {} : { extensionFlagValues: new Map(Object.entries(options.extensionFlags)) },
       resourceLoaderOptions: {
+        ...options.extensionPaths === undefined ? {} : { additionalExtensionPaths: [...options.extensionPaths] },
         extensionFactories: [
           createAutoRAGInteractiveExtension(options.getSystemPrompt, options.contextTransform, { onInput: options.onInput, isSearchRunning: options.isSearchRunning, onKey: options.onKey }, options.updateNotice),
           ...options.extensionFactories ?? []
@@ -11508,7 +11750,7 @@ async function createAutoRAGPiInteractiveRuntime(options) {
   });
   return { runtime, dispose: () => runtime.dispose() };
 }
-var PI_BUILTIN_TOOL_NAMES, EMBEDDED_AUTH_MARKER = "<autorag-embedded-model>";
+var PI_BUILTIN_TOOL_NAMES, EMBEDDED_AUTH_MARKER = "<autorag-embedded-model>", TELEGRAM_INPUT_PREFIX = "[telegram]", WHATSAPP_INPUT_PREFIX = "Message from ";
 var init_pi_session = __esm(() => {
   init_provider();
   init_context_token_guard();
@@ -11764,11 +12006,11 @@ var init_jikji_find_tool = __esm(() => {
 
 // src/agent/local-model.ts
 import { readFileSync as readFileSync12 } from "node:fs";
-import { homedir as homedir2 } from "node:os";
+import { homedir as homedir3 } from "node:os";
 import { join as join24 } from "node:path";
 import { parse as parse2 } from "smol-toml";
 function loadLocalAutoRAGModel(options = {}) {
-  const configPath = options.configPath ?? join24(homedir2(), ".codex", "config.toml");
+  const configPath = options.configPath ?? join24(homedir3(), ".codex", "config.toml");
   const config = parse2(readFileSync12(configPath, "utf8"));
   const provider = config.model_provider;
   if (typeof provider !== "string" || provider.length === 0) {
@@ -11811,7 +12053,7 @@ var init_local_model = () => {};
 
 // src/p2p/approval-store.ts
 import {
-  existsSync as existsSync16,
+  existsSync as existsSync17,
   mkdirSync as mkdirSync11,
   readdirSync as readdirSync3,
   readFileSync as readFileSync13,
@@ -11876,7 +12118,7 @@ function savePendingPeerRequest(workspacePath, request) {
 }
 function loadPendingPeerRequest(workspacePath, id) {
   const path = requestPath(workspacePath, id);
-  if (!existsSync16(path))
+  if (!existsSync17(path))
     return;
   try {
     return parsePending(JSON.parse(readFileSync13(path, "utf8")));
@@ -11886,7 +12128,7 @@ function loadPendingPeerRequest(workspacePath, id) {
 }
 function listPendingPeerRequests(workspacePath) {
   const dir = requestsDir(workspacePath);
-  if (!existsSync16(dir))
+  if (!existsSync17(dir))
     return [];
   const requests = [];
   for (const name of readdirSync3(dir)) {
@@ -11908,7 +12150,7 @@ function writePeerRequestDecision(workspacePath, id, decision) {
 }
 function loadPeerRequestDecision(workspacePath, id) {
   const path = decisionPath(workspacePath, id);
-  if (!existsSync16(path))
+  if (!existsSync17(path))
     return;
   try {
     return parseDecision(JSON.parse(readFileSync13(path, "utf8")));
@@ -12242,7 +12484,7 @@ var init_egress_gate = __esm(() => {
 });
 
 // src/p2p/policy.ts
-import { existsSync as existsSync17, mkdirSync as mkdirSync12, readFileSync as readFileSync14, writeFileSync as writeFileSync10 } from "node:fs";
+import { existsSync as existsSync18, mkdirSync as mkdirSync12, readFileSync as readFileSync14, writeFileSync as writeFileSync10 } from "node:fs";
 import { basename as basename10, dirname as dirname13, join as join26, resolve as resolve12 } from "node:path";
 import { parse as parseToml } from "smol-toml";
 function isRecord11(value) {
@@ -12322,7 +12564,7 @@ function normalizePolicyEntries(value, workspacePath, context) {
   return entries;
 }
 function readJsonFile(path) {
-  if (!existsSync17(path))
+  if (!existsSync18(path))
     return;
   let raw;
   try {
@@ -12337,7 +12579,7 @@ function readJsonFile(path) {
   }
 }
 function readTomlFile(path) {
-  if (!existsSync17(path))
+  if (!existsSync18(path))
     return;
   let raw;
   try {
@@ -12478,7 +12720,7 @@ function globToRegExp(pattern) {
   return new RegExp(`${expression}$`, "u");
 }
 function loadSeenSources(path, workspacePath) {
-  if (!existsSync17(path))
+  if (!existsSync18(path))
     return { promoted: new Set, pending: new Set };
   let parsed;
   try {
@@ -12785,7 +13027,7 @@ var init_wire = __esm(() => {
 
 // src/p2p/simplex-query-store.ts
 import { randomUUID as randomUUID5 } from "node:crypto";
-import { existsSync as existsSync18, mkdirSync as mkdirSync13, readdirSync as readdirSync4, readFileSync as readFileSync15, renameSync as renameSync7, unlinkSync as unlinkSync4, writeFileSync as writeFileSync11 } from "node:fs";
+import { existsSync as existsSync19, mkdirSync as mkdirSync13, readdirSync as readdirSync4, readFileSync as readFileSync15, renameSync as renameSync7, unlinkSync as unlinkSync4, writeFileSync as writeFileSync11 } from "node:fs";
 import { dirname as dirname14, join as join27 } from "node:path";
 import { Value as Value2 } from "typebox/value";
 function statePath(workspacePath, directory, id) {
@@ -12821,7 +13063,7 @@ function parseState(value) {
   return value;
 }
 function readState(path) {
-  if (!existsSync18(path))
+  if (!existsSync19(path))
     return;
   try {
     return parseState(JSON.parse(readFileSync15(path, "utf8")));
@@ -12848,14 +13090,14 @@ function moveState(workspacePath, from, to, state) {
 }
 function recoverClaims(workspacePath) {
   const directory = join27(workspacePath, QUERY_DIR, PENDING_DIR);
-  if (!existsSync18(directory))
+  if (!existsSync19(directory))
     return;
   for (const name of readdirSync4(directory)) {
     if (!name.endsWith(".json.claim"))
       continue;
     const claim = join27(directory, name);
     const restored = join27(directory, name.slice(0, -".claim".length));
-    if (!existsSync18(restored))
+    if (!existsSync19(restored))
       renameSync7(claim, restored);
   }
 }
@@ -12879,7 +13121,7 @@ function loadSimplexQueryState(workspacePath, id) {
 }
 function listPendingSimplexQueryStates(workspacePath) {
   const directory = join27(workspacePath, QUERY_DIR, PENDING_DIR);
-  if (!existsSync18(directory))
+  if (!existsSync19(directory))
     return [];
   return readdirSync4(directory).filter((name) => name.endsWith(".json")).map((name) => readState(join27(directory, name))).filter((state) => state?.status === "pending").filter((state) => loadSimplexQueryState(workspacePath, state.id)?.status === "pending").sort((left, right) => left.createdAt.localeCompare(right.createdAt));
 }
@@ -13022,7 +13264,7 @@ var init_simplex_query_store = __esm(() => {
 });
 
 // src/p2p/simplex-server.ts
-import { existsSync as existsSync19, mkdirSync as mkdirSync14, readFileSync as readFileSync16, writeFileSync as writeFileSync12 } from "node:fs";
+import { existsSync as existsSync20, mkdirSync as mkdirSync14, readFileSync as readFileSync16, writeFileSync as writeFileSync12 } from "node:fs";
 import { join as join28 } from "node:path";
 import { Type as Type10 } from "typebox";
 import { Value as Value3 } from "typebox/value";
@@ -13051,7 +13293,7 @@ function rankSimplexPeerTargets(query, registry) {
 }
 function loadSimplexPeerRegistry(workspacePath) {
   const path = join28(workspacePath, PEERS_DIR, PEERS_FILENAME);
-  if (!existsSync19(path))
+  if (!existsSync20(path))
     return {};
   try {
     const parsed = JSON.parse(readFileSync16(path, "utf8"));
@@ -13957,6 +14199,7 @@ function recordStructuredResultsSession(sessionId, query, details, memory, compo
 }
 var init_search_documents = __esm(() => {
   init_memory();
+  init_citations();
 });
 
 // src/agent/search-all-tool.ts
@@ -14175,14 +14418,14 @@ var init_search_single_datasource_tool = __esm(() => {
 });
 
 // src/agent/self-config.ts
-import { existsSync as existsSync20, readFileSync as readFileSync18, rmSync as rmSync9, writeFileSync as writeFileSync14 } from "node:fs";
+import { existsSync as existsSync21, readFileSync as readFileSync18, rmSync as rmSync9, writeFileSync as writeFileSync14 } from "node:fs";
 import { dirname as dirname16, join as join29 } from "node:path";
 import { fileURLToPath as fileURLToPath3 } from "node:url";
 function findPackagedSetupSkill() {
   let dir = dirname16(fileURLToPath3(import.meta.url));
   for (let depth = 0;depth < 6; depth += 1) {
     const candidate = join29(dir, SETUP_SKILL_RELATIVE_PATH);
-    if (existsSync20(candidate))
+    if (existsSync21(candidate))
       return candidate;
     const parent = dirname16(dir);
     if (parent === dir)
@@ -14205,7 +14448,7 @@ function loadSetupSkill(skillPath) {
   return text.replace(/^---\r?\n[\s\S]*?\r?\n---(?:\r?\n)+/u, "");
 }
 function snapshotConfigFile(configPath) {
-  return existsSync20(configPath) ? readFileSync18(configPath, "utf8") : undefined;
+  return existsSync21(configPath) ? readFileSync18(configPath, "utf8") : undefined;
 }
 async function rollbackIfBroken(options, before) {
   if (options.validate === undefined)
@@ -14273,6 +14516,10 @@ function buildSystemPrompt(config) {
   const modelId = config.modelId ?? "the configured model";
   const toolLines = [
     toolLine(config, "bash", "read and inspect configured document collections with ls, find, grep, cat, and similar tools, and — in local sessions — organize local files when the user asks (rename, move, copy, create folders, move to trash)"),
+    ...config.remoteSession ? [] : [
+      toolLine(config, "write", "create a new file in local sessions — use this to produce requested deliverables and artifacts in a file format aligned with the request and context without overwriting existing source files"),
+      toolLine(config, "edit", "surgically update an existing deliverable or config file when the user asks (local sessions only)")
+    ],
     toolLine(config, "jikji_find", "local discovery through Jikji answer packs"),
     toolLine(config, "everything_search", "instant Windows file/folder name search (voidtools Everything) across the configured search folders"),
     toolLine(config, "fsearch_search", "instant macOS/Linux file/folder name search (fsearch-cli/FSearch) across the configured search folders, with a slow filesystem-walk fallback"),
@@ -14289,6 +14536,12 @@ function buildSystemPrompt(config) {
     ...config.toolNames.filter((name) => name.startsWith("search_datasource_")).map((name) => `- **${name}**: search only the ${name.slice("search_datasource_".length).replace(/_/g, "-")} datasource connection, spawning no other datasource CLIs`),
     ...config.toolNames.filter((name) => ![
       "bash",
+      "read",
+      "write",
+      "edit",
+      "grep",
+      "find",
+      "ls",
       "jikji_find",
       "everything_search",
       "fsearch_search",
@@ -14421,8 +14674,9 @@ Your answer is your final reply, written as plain text, never as a tool call. Ci
 ## Answer Guidelines
 
 - **Complete vs. delta answer**: When no first answer reached the caller, give the complete core answer. When a first answer was already delivered, return only the delta against it — corrections and newly verified findings — and never repeat its unchanged content; if nothing changed, confirm the first answer in one short line.
-- **Bullet-point core answer**: Provide the core answer to the user's question in at most 5 bullet points. If additional explanation or context is necessary, append it after the bullet points.
-- **Direct answer only**: The caller only needs the answer to their question. Never include specific file paths, datasource descriptions, or retrieval mechanics/principles in the answer (source paths are attached to each citation from the evidence ids you cite). Do not quote raw chunk text.
+- **Bullet-point core answer**: Unless the user explicitly requests a different format or length, provide the core answer to the user's question in at most 5 bullet points. If additional explanation or context is necessary, append it after the bullet points.
+- **Direct answer only**: Unless the user explicitly requests otherwise, the caller only needs the direct answer to their question. Never include specific file paths, datasource descriptions, or retrieval mechanics/principles in the answer (source paths are attached to each citation from the evidence ids you cite). Do not quote raw chunk text.
+- ${ANSWER_DELIVERABLE_RULE} ${ANSWER_DELIVERABLE_DELTA_RULE}
 - ${ANSWER_CITATION_RULE}
 - ${ANSWER_IMAGE_EMBED_RULE} ${ANSWER_IMAGE_DELTA_RULE}
 - **No per-source negative reports**: Never report individual negative findings per source (e.g. "no information found in Slack" or "checked Drive but found nothing"). Simply omit unproductive sources from the answer and focus on what was found or provide a concise overall conclusion.
@@ -16717,11 +16971,6 @@ var init_types3 = __esm(() => {
       label: "xAI",
       description: "xAI Responses web_search via the agent's xAI model credentials (XAI_API_KEY)"
     },
-    {
-      value: "perplexity",
-      label: "Perplexity",
-      description: "Anonymous Perplexity ask endpoint; no key or signup required"
-    },
     {
       value: "parallel",
       label: "Parallel",
@@ -16732,16 +16981,21 @@ var init_types3 = __esm(() => {
       label: "SearXNG",
       description: "Advanced: requires a self-hosted SEARXNG_ENDPOINT; never part of the default experience"
     },
-    {
-      value: "startpage",
-      label: "Startpage",
-      description: "Credential-free scrape of Startpage (Google-backed) results; may be bot-challenged"
-    },
     {
       value: "duckduckgo",
       label: "DuckDuckGo",
       description: "Credential-free best-effort fallback; may be bot-challenged on datacenter/shared-egress IPs"
     },
+    {
+      value: "perplexity",
+      label: "Perplexity",
+      description: "Anonymous Perplexity ask endpoint; no key or signup required"
+    },
+    {
+      value: "startpage",
+      label: "Startpage",
+      description: "Credential-free scrape of Startpage (Google-backed) results; may be bot-challenged"
+    },
     {
       value: "ecosia",
       label: "Ecosia",
@@ -18089,12 +18343,11 @@ function eventAnswer(event) {
   }
   return "";
 }
-var PERPLEXITY_ASK_URL = "https://www.perplexity.ai/rest/sse/perplexity_ask", API_VERSION2 = "2.18", DEFLECTION_PATTERN, MAX_ERROR_BYTES4, ANONYMOUS_USER_AGENT = "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36", PERPLEXITY_QUERY_SYNTAX, PerplexityProvider;
+var PERPLEXITY_ASK_URL = "https://www.perplexity.ai/rest/sse/perplexity_ask", API_VERSION2 = "2.18", MAX_ERROR_BYTES4, ANONYMOUS_USER_AGENT = "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36", PERPLEXITY_QUERY_SYNTAX, PerplexityProvider;
 var init_perplexity = __esm(() => {
   init_query();
   init_types3();
   init_utils();
-  DEFLECTION_PATTERN = /sign up|log in|create an account|verify you are human|are you a robot|가입|로그인|회원 ?가입/i;
   MAX_ERROR_BYTES4 = 8 * 1024;
   PERPLEXITY_QUERY_SYNTAX = {
     phrases: true,
@@ -18208,8 +18461,8 @@ var init_perplexity = __esm(() => {
           break;
       }
       const sources = [...sourcesByUrl.values()];
-      if (sources.length === 0 && DEFLECTION_PATTERN.test(answer)) {
-        throw new SearchProviderError("perplexity", "Perplexity anonymous ask was deflected with a sign-up wall", 403);
+      if (sources.length === 0) {
+        throw new SearchProviderError("perplexity", `Perplexity anonymous ask returned no web sources (sign-up wall or conversational deflection): ${answer.slice(0, 120)}`, 403);
       }
       const numResults = params.numSearchResults ?? params.limit;
       return {
@@ -18627,7 +18880,7 @@ var init_browser_headers = __esm(() => {
 });
 
 // src/web/search/browser-loader.ts
-import { existsSync as existsSync21 } from "node:fs";
+import { existsSync as existsSync22 } from "node:fs";
 async function closeBounded(close) {
   let timer;
   try {
@@ -18706,7 +18959,7 @@ function ensureWebSearchBrowserLoader(options = {}) {
       return true;
     const exists = options.exists ?? ((path) => {
       try {
-        return existsSync21(path);
+        return existsSync22(path);
       } catch {
         return false;
       }
@@ -18811,6 +19064,8 @@ function normalizeText(value) {
 function isChallengeResponse(page) {
   if (/\/(?:errors|captcha)\//.test(page.url) || page.url.includes("/sp/captcha"))
     return true;
+  if (ANUBIS_CHALLENGE_PATTERN.test(page.html))
+    return true;
   return page.html.includes("component---src-pages-captcha") || page.html.includes("/sp/captcha");
 }
 function parseSearchFormInputs(html) {
@@ -18915,7 +19170,7 @@ async function callStartpageHtml(params) {
     });
   }
   if (isChallengeResponse(page)) {
-    throw new SearchProviderError("startpage", "Startpage blocked the request with a CAPTCHA challenge. Startpage rate-limits automated searches from datacenter/shared-egress IPs; try another provider such as DuckDuckGo or Mojeek, or retry later.", 429);
+    throw new SearchProviderError("startpage", "Startpage blocked the request with a bot challenge (CAPTCHA or Anubis proof-of-work). Startpage rate-limits automated searches from datacenter/shared-egress IPs; try another provider such as DuckDuckGo or Mojeek, or retry later.", 429);
   }
   if (page.status < 200 || page.status >= 300) {
     const classified = classifyProviderHttpError("startpage", page.status, page.html);
@@ -18941,7 +19196,7 @@ async function searchStartpage(params) {
   }
   return { provider: "startpage", sources };
 }
-var STARTPAGE_HOME_URL = "https://www.startpage.com/", STARTPAGE_SEARCH_URL = "https://www.startpage.com/sp/search", DEFAULT_NUM_RESULTS3 = 10, MAX_NUM_RESULTS3 = 20, RECENCY_TO_STARTPAGE_WITH_DATE, StartpageProvider;
+var STARTPAGE_HOME_URL = "https://www.startpage.com/", STARTPAGE_SEARCH_URL = "https://www.startpage.com/sp/search", DEFAULT_NUM_RESULTS3 = 10, MAX_NUM_RESULTS3 = 20, RECENCY_TO_STARTPAGE_WITH_DATE, ANUBIS_CHALLENGE_PATTERN, StartpageProvider;
 var init_startpage = __esm(() => {
   init_entities();
   init_browser_page();
@@ -18954,6 +19209,7 @@ var init_startpage = __esm(() => {
     month: "m",
     year: "y"
   };
+  ANUBIS_CHALLENGE_PATTERN = /<script\b[^>]*\bid\s*=\s*["']anubis_(?:challenge|version)["']/i;
   StartpageProvider = class StartpageProvider extends SearchProvider {
     id = "startpage";
     label = "Startpage";
@@ -20171,7 +20427,7 @@ var init_web_search_tool = __esm(() => {
 
 // src/agent/agent.ts
 import { randomUUID as randomUUID7 } from "node:crypto";
-import { existsSync as existsSync22, watch as fsWatch, mkdirSync as mkdirSync16, realpathSync as realpathSync5, statSync as statSync2, writeFileSync as writeFileSync15 } from "node:fs";
+import { existsSync as existsSync23, watch as fsWatch, mkdirSync as mkdirSync16, realpathSync as realpathSync5, statSync as statSync2, writeFileSync as writeFileSync15 } from "node:fs";
 import { basename as basename11, dirname as dirname17, join as join31, resolve as resolve13 } from "node:path";
 import { clampThinkingLevel } from "@earendil-works/pi-ai/compat";
 import { getAgentDir as getAgentDir3 } from "@earendil-works/pi-coding-agent";
@@ -20302,6 +20558,9 @@ class AutoRAGAgent {
   piSessionDir;
   persistPiSessions;
   updateNotice;
+  piExtensionPaths;
+  piExtensionFlags;
+  telegram;
   boundPiRuntime;
   memoryEmbedder;
   memoryVectorPath;
@@ -20327,6 +20586,9 @@ class AutoRAGAgent {
     this.piSessionDir = options.piSessionDir;
     this.persistPiSessions = options.persistPiSessions ?? true;
     this.updateNotice = options.updateNotice;
+    this.piExtensionPaths = options.piExtensionPaths;
+    this.piExtensionFlags = options.piExtensionFlags;
+    this.telegram = options.telegram === true;
     this.datasourceSkills = options.datasourceSkills ?? [];
     this.datasourceRetry = options.datasourceRetry ?? {};
     this.datasourceVirtualScopePrefixes = this.datasourceSkills.map((skill) => normalizeVirtualPath2(`/${skill.describe().name}`));
@@ -20379,7 +20641,8 @@ class AutoRAGAgent {
       const everythingClient = new EverythingClient({
         ...options.everything ?? {},
         root: this.workspaceProjectRoot,
-        folders: this.searchPaths
+        folders: this.searchPaths,
+        excludeFolders: this.excludePaths
       });
       if (everythingClient.isSupported())
         this.everythingClient = everythingClient;
@@ -20485,8 +20748,9 @@ class AutoRAGAgent {
       return true;
     });
     this.tools = tools;
+    const builtinToolNames = this.remoteSession ? PI_BUILTIN_TOOL_NAMES.filter((name) => name !== "write" && name !== "edit") : PI_BUILTIN_TOOL_NAMES;
     const toolNames = [
-      ...PI_BUILTIN_TOOL_NAMES,
+      ...builtinToolNames,
       ...tools.map((tool) => tool.name),
       ...this.jevExtension !== undefined ? [JEV_TOOL_NAME] : []
     ];
@@ -20502,7 +20766,7 @@ class AutoRAGAgent {
     this.innerAgent = {
       state: {
         systemPrompt,
-        tools: [...PI_BUILTIN_TOOL_NAMES.map((name) => ({ name })), ...tools],
+        tools: [...builtinToolNames.map((name) => ({ name })), ...tools],
         messages: []
       },
       transformContext: (messages) => this.withMemoryContext(messages)
@@ -20796,6 +21060,7 @@ ${context.text}
   }
   async createPiInteractiveRuntime() {
     const resolved = this.configuredModel === undefined ? undefined : await this.resolveSessionModel();
+    const telegramBridge = this.telegram ? (await import("@llblab/pi-telegram")).default : undefined;
     const runtime = await createAutoRAGPiInteractiveRuntime({
       cwd: this.workspaceProjectRoot,
       agentDir: this.piAgentDir,
@@ -20816,10 +21081,14 @@ ${context.text}
       onKey: (key) => this.handleInteractiveKey(key),
       extensionFactories: [
         createEvidencePersistence(this.evidenceLedger),
-        ...this.jevExtension !== undefined ? [this.jevExtension] : []
+        ...this.jevExtension !== undefined ? [this.jevExtension] : [],
+        ...telegramBridge !== undefined ? [telegramBridge] : []
       ],
       ...this.jevExtension !== undefined ? { extensionToolNames: [JEV_TOOL_NAME] } : {},
-      ...this.updateNotice === undefined ? {} : { updateNotice: this.updateNotice }
+      remoteSession: this.remoteSession,
+      ...this.updateNotice === undefined ? {} : { updateNotice: this.updateNotice },
+      ...this.piExtensionPaths === undefined ? {} : { extensionPaths: this.piExtensionPaths },
+      ...this.piExtensionFlags === undefined ? {} : { extensionFlags: this.piExtensionFlags }
     });
     this.boundPiRuntime = runtime.runtime;
     const dispose = runtime.dispose;
@@ -21739,21 +22008,22 @@ ${result.content.replace(/\s+/gu, " ")}`).join(`
     const scope = options.scope ? ` Restrict search to virtual path scope ${options.scope}.` : "";
     return `Answer this original query immediately: ${query}${limit}${scope}
 
-` + `Baseline retrieval evidence (already gathered for you):
+Baseline retrieval evidence (already gathered for you):
 ${baseline}
 
-` + `Produce the best complete, self-contained answer you can RIGHT NOW from this evidence. Do NOT call any search, retrieval, or file-reading tools and do NOT wait for more evidence. ` + `If the query is answerable from general knowledge alone, answer directly.
+Produce the best complete, self-contained answer you can RIGHT NOW from this evidence. Do NOT call any search, retrieval, or file-reading tools and do NOT wait for more evidence. If the query is answerable from general knowledge alone, answer directly.
 
-` + `Formatting and content rules for the answer:
-` + `- Provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
-` + `- Answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
-` + `- ${ANSWER_CITATION_RULE}
-` + `- ${ANSWER_IMAGE_EMBED_RULE}
-` + `- Do not report per-source negative findings (e.g. "no information found in Slack" or "checked Drive but found nothing").
-` + `- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.
-` + `- If information is incomplete or uncertain, acknowledge it briefly without lengthy explanations, stating that it is difficult to answer fully with the given information and searching continues. If there are partial clues or leads (even if not the exact answer), mention those clues concisely.
+Formatting and content rules for the answer:
+- Unless the user explicitly requests a different format or length, provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
+- Unless the user explicitly requests otherwise, answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
+- ${ANSWER_DELIVERABLE_RULE}
+- ${ANSWER_CITATION_RULE}
+- ${ANSWER_IMAGE_EMBED_RULE}
+- Do not report per-source negative findings (e.g. "no information found in Slack" or "checked Drive but found nothing").
+- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.
+- If information is incomplete or uncertain, acknowledge it briefly without lengthy explanations, stating that it is difficult to answer fully with the given information and searching continues. If there are partial clues or leads (even if not the exact answer), mention those clues concisely.
 
-` + `Write the answer as your reply, then stop.`;
+Write the answer as your reply, then stop.`;
   }
   isConversationTurn() {
     if (this.remoteSession || this.boundPiRuntime === undefined)
@@ -21763,7 +22033,7 @@ ${baseline}
   buildDirectAnswerPrompt(query) {
     return `Answer this query directly from your own general knowledge: ${query}
 
-` + `It needs no search: it is general knowledge, simple reasoning, or conversation. Do NOT call any search, retrieval, web, or file-reading tools. ` + `Reply naturally and concisely; for small talk, just respond conversationally. Do not cite sources or mention retrieval.`;
+It needs no search: it is general knowledge, simple reasoning, or conversation. Do NOT call any search, retrieval, web, or file-reading tools. Reply naturally and concisely; for small talk, just respond conversationally. Do not cite sources or mention retrieval.`;
   }
   discoveryHint(sentence) {
     const names = this.tools.map((tool) => tool.name).filter((name) => name === JIKJI_FIND_TOOL_NAME || name === EVERYTHING_SEARCH_TOOL_NAME);
@@ -21781,52 +22051,55 @@ ${baseline}
     const scope = options.scope ? ` Restrict search to virtual path scope ${options.scope}.` : "";
     const firstAnswer = formatFirstAnswerContext(fastAnswer, fastAnswerDelivered);
     const answerRules = fastAnswerDelivered ? `Formatting and content rules for your final answer (DELTA ONLY):
-` + `- The user already has the first answer above. Your final answer MUST contain only the delta against it: (a) corrections to anything in the first answer that is wrong, unsupported, or outdated, and (b) newly verified findings that were not present in the first answer.
-` + `- NEVER restate or re-list first-answer facts that remain correct, and never repeat its bullet list.
-` + `- Mark each item clearly as a correction or as a new finding.
-` + `- If verification changed nothing and found nothing new, say so in one short line (the first answer is confirmed as-is) instead of restating it.
-` + `- ${ANSWER_CITATION_RULE} A correction or new finding that relies on evidence the first answer cited must cite that evidence's id again.
-` + `- ${ANSWER_IMAGE_EMBED_RULE} ${ANSWER_IMAGE_DELTA_RULE}
-` + `- Do not report per-source negative findings (e.g. "no information found in Slack").
-` + `- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.` : `Formatting and content rules for the final answer (COMPLETE — no first answer reached the caller):
-` + `- Provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
-` + `- Answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
-` + `- ${ANSWER_CITATION_RULE}
-` + `- ${ANSWER_IMAGE_EMBED_RULE}
-` + `- Do not report per-source negative findings (e.g. "no information found in Slack").
-` + `- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.`;
+- The user already has the first answer above. Your final answer MUST contain only the delta against it: (a) corrections to anything in the first answer that is wrong, unsupported, or outdated, and (b) newly verified findings that were not present in the first answer.
+- NEVER restate or re-list first-answer facts that remain correct, and never repeat its bullet list.
+- Mark each item clearly as a correction or as a new finding.
+- If verification changed nothing and found nothing new, say so in one short line (the first answer is confirmed as-is) instead of restating it.
+- ${ANSWER_DELIVERABLE_RULE} ${ANSWER_DELIVERABLE_DELTA_RULE}
+- ${ANSWER_CITATION_RULE} A correction or new finding that relies on evidence the first answer cited must cite that evidence's id again.
+- ${ANSWER_IMAGE_EMBED_RULE} ${ANSWER_IMAGE_DELTA_RULE}
+- Do not report per-source negative findings (e.g. "no information found in Slack").
+- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.` : `Formatting and content rules for the final answer (COMPLETE — no first answer reached the caller):
+` + `- Unless the user explicitly requests a different format or length, provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
+- Unless the user explicitly requests otherwise, answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
+- ${ANSWER_DELIVERABLE_RULE}${fastAnswer !== undefined ? ` ${ANSWER_DELIVERABLE_DELTA_RULE}` : ""}
+- ${ANSWER_CITATION_RULE}
+- ${ANSWER_IMAGE_EMBED_RULE}
+- Do not report per-source negative findings (e.g. "no information found in Slack").
+- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.`;
     return `Original query: ${query}${limit}${scope}
 
-` + `${firstAnswer}
+${firstAnswer}
 
 ` + formatSteerContext(steers) + `Now verify it rigorously. ${this.discoveryHint((tools) => `Actively use ${tools} when discovering or exploring local files and folders. `)}Check important claims against source files with bash when needed, correct anything wrong or unsupported, fill gaps with retrieval tools, and resolve conflicts and freshness. ${this.webFallbackHint()}
 
-` + `${answerRules}
+${answerRules}
 
-` + `Do not use broad grep/find or recursive filesystem scans: only inspect a path or narrow neighborhood surfaced by retrieval, and only when evidence clearly points there. ` + `Avoid spinning repeated near-identical queries against the same datasource; once additional attempts stop surfacing new evidence, conclude from the evidence available. ` + `If more search is needed, first write a brief 1–2 line progress update stating the best current hypothesis and what you are checking next, then call retrieval tools. ` + `When finished, reply with the final answer and no tool call: that last reply is what the user receives, so keep progress notes out of it.`;
+Do not use broad grep/find or recursive filesystem scans: only inspect a path or narrow neighborhood surfaced by retrieval, and only when evidence clearly points there. Avoid spinning repeated near-identical queries against the same datasource; once additional attempts stop surfacing new evidence, conclude from the evidence available. ` + `If more search is needed, first write a brief 1–2 line progress update stating the best current hypothesis and what you are checking next, then call retrieval tools. ` + `When finished, reply with the final answer and no tool call: that last reply is what the user receives, so keep progress notes out of it.`;
   }
   buildSearchPrompt(query, options, initialRetrievalContext) {
     const limit = typeof options.topK === "number" ? ` Cite at most ${options.topK} sources.` : "";
     const scope = options.scope ? ` Restrict search to virtual path scope ${options.scope}.` : "";
     return `Find and curate information for this original query: ${query}${limit}${scope}
 
-` + `Start by deciding whether this is answerable from general knowledge or memory. If it is a generic, stable question, answer it immediately without retrieval. ` + `Otherwise, baseline MinSync and Jikji retrieval is already running in parallel; do not give the final answer until its next message arrives.
+Start by deciding whether this is answerable from general knowledge or memory. If it is a generic, stable question, answer it immediately without retrieval. Otherwise, baseline MinSync and Jikji retrieval is already running in parallel; do not give the final answer until its next message arrives.
 
-` + `Baseline retrieval context:
+Baseline retrieval context:
 ${initialRetrievalContext ?? "Pending; continue only with a brief progress statement."}
 
-` + `Treat candidates as unverified evidence, verify important claims against source files when needed, and use additional tools when needed. ` + `Judge relevance, conflicts, freshness, and sufficiency in this agent loop.
+Treat candidates as unverified evidence, verify important claims against source files when needed, and use additional tools when needed. Judge relevance, conflicts, freshness, and sufficiency in this agent loop.
 
-` + `Formatting and content rules for the answer:
-` + `- Provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
-` + `- Answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
-` + `- ${ANSWER_CITATION_RULE}
-` + `- ${ANSWER_IMAGE_EMBED_RULE}
-` + `- Do not report per-source negative findings (e.g. "no information found in Slack").
-` + `- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.
-` + `- If information is incomplete or uncertain, acknowledge it briefly without lengthy explanations. If there are partial clues or leads, mention them concisely.
+Formatting and content rules for the answer:
+- Unless the user explicitly requests a different format or length, provide the core answer to the user's question in at most 5 bullet points. If additional explanation is necessary, append it after the bullet points.
+- Unless the user explicitly requests otherwise, answer the question directly. Do not include specific file paths, datasource descriptions, or retrieval mechanics in the answer text.
+- ${ANSWER_DELIVERABLE_RULE}
+- ${ANSWER_CITATION_RULE}
+- ${ANSWER_IMAGE_EMBED_RULE}
+- Do not report per-source negative findings (e.g. "no information found in Slack").
+- When evidence conflicts, treat the freshest (most recent) information as the correct source of truth.
+- If information is incomplete or uncertain, acknowledge it briefly without lengthy explanations. If there are partial clues or leads, mention them concisely.
 
-` + `${this.discoveryHint((tools) => `When exploring local files and folders, actively use ${tools} rather than exploratory bash commands. `)}If more search is needed, first write a brief 1–2 line progress update stating the best current hypothesis and what you are checking next, then call retrieval tools. ` + `Never repeat a generic status message. Do not use broad grep/find or recursive filesystem scans: only inspect a path or narrow neighborhood surfaced by retrieval, and only when evidence clearly points there. ` + `Avoid spinning repeated near-identical queries against the same datasource; once additional attempts stop surfacing new evidence, conclude from the evidence available. ` + `When finished, reply with the final answer and no tool call: that last reply is what the user receives, so keep progress notes out of it.`;
+${this.discoveryHint((tools) => `When exploring local files and folders, actively use ${tools} rather than exploratory bash commands. `)}If more search is needed, first write a brief 1–2 line progress update stating the best current hypothesis and what you are checking next, then call retrieval tools. ` + `Never repeat a generic status message. Do not use broad grep/find or recursive filesystem scans: only inspect a path or narrow neighborhood surfaced by retrieval, and only when evidence clearly points there. Avoid spinning repeated near-identical queries against the same datasource; once additional attempts stop surfacing new evidence, conclude from the evidence available. When finished, reply with the final answer and no tool call: that last reply is what the user receives, so keep progress notes out of it.`;
   }
   async refresh(force = false, opts) {
     const signal = opts?.signal;
@@ -22062,7 +22335,7 @@ ${initialRetrievalContext ?? "Pending; continue only with a brief progress state
       });
     }
     const state = this.refreshState.inFlight || persistedRunning && ownerAlive ? "indexing" : interrupted ? "failed" : persistedProgress?.state === "success" || persistedProgress?.state === "failed" ? persistedProgress.state : this.refreshState.lastOutcome === "never" ? "idle" : this.refreshState.lastOutcome;
-    const parsedMirrorReady = this.refreshState.lastOutcome === "success" || persistedProgress?.state === "success" || existsSync22(refreshReadinessPath(this.workspaceProjectRoot));
+    const parsedMirrorReady = this.refreshState.lastOutcome === "success" || persistedProgress?.state === "success" || existsSync23(refreshReadinessPath(this.workspaceProjectRoot));
     const effectiveProgress = this.refreshState.inFlight || persistedRunning ? persistedProgress : undefined;
     return {
       state,
@@ -22522,7 +22795,7 @@ function buildSteerMessage(text) {
 
 ${text}
 
-` + `Treat it as direction for the search in progress, not as a new question: adjust which sources, queries, or claims you pursue, then continue. ` + `When finished, reply with the final answer to the original question, updated for this correction, and no tool call.`;
+Treat it as direction for the search in progress, not as a new question: adjust which sources, queries, or claims you pursue, then continue. When finished, reply with the final answer to the original question, updated for this correction, and no tool call.`;
 }
 function frameSteers(messages, steerTexts) {
   if (steerTexts.size === 0)
@@ -22542,12 +22815,12 @@ function formatSteerContext(steers) {
   return `While the search ran, the user corrected it:
 ${steers.map((steer) => `- ${steer}`).join(`
 `)}
-` + `The first answer predates these corrections; follow them over it.
+The first answer predates these corrections; follow them over it.
 
 `;
 }
 function buildSteerContinuationPrompt(steers) {
-  return `${formatSteerContext(steers)}Apply the latest correction: search or verify further if it needs that, ` + `then reply with the complete final answer to the original question and no tool call.`;
+  return `${formatSteerContext(steers)}Apply the latest correction: search or verify further if it needs that, then reply with the complete final answer to the original question and no tool call.`;
 }
 function formatFirstAnswerContext(fastAnswer, delivered) {
   if (fastAnswer === undefined) {
@@ -22670,11 +22943,20 @@ function fsearchRefreshDiagnostics(fsearch) {
 function minSyncRefreshDiagnostics(minsync) {
   if (!minsync)
     return [];
+  const diagnostics = [];
   if (minsync.diagnostic)
-    return [toMinSyncDiagnostic(minsync.diagnostic, minsync.ok)];
-  if (!minsync.ok && minsync.reason)
-    return [toMinSyncReasonDiagnostic(minsync.reason)];
-  return [];
+    diagnostics.push(toMinSyncDiagnostic(minsync.diagnostic, minsync.ok));
+  else if (!minsync.ok && minsync.reason)
+    diagnostics.push(toMinSyncReasonDiagnostic(minsync.reason));
+  if (minsync.lexicalFallback) {
+    diagnostics.push({
+      code: "lexical-tokenizer-unavailable",
+      severity: "warning",
+      message: `The "${minsync.lexicalFallback.requested}" BM25 tokenizer needs the Kiwi native library and model, which were not found, so MinSync tokenizes on whitespace and a Korean word with a particle (인증서를) does not match its stem (인증서). Install Kiwi (set KIWI_LIBRARY_PATH and KIWI_MODEL_PATH, or install it under ~/.local/kiwi), then refresh again.`,
+      source: "minsync"
+    });
+  }
+  return diagnostics;
 }
 function toMinSyncDiagnostic(diag, ok) {
   const code = diag.code === "embedder-unavailable" ? "embedder-unavailable" : diag.code === "embedding-identity-mismatch" ? "embedding-identity-mismatch" : diag.code === "minsync-lock-held" ? "minsync-lock-held" : "minsync-sync-failed";
@@ -22837,7 +23119,7 @@ var init_results = __esm(() => {
     confidence: Type18.Optional(Type18.Number({ description: "Evidence confidence, 0..1", minimum: 0, maximum: 1 }))
   });
   answerSchema = Type18.String({
-    description: `Answer for the caller. At most 5 bullet points (plus optional explanation); reference results by bracketed number (e.g. [1], [2]) without file paths or raw chunk text, except the image-embed exception below. ${ANSWER_CITATION_RULE} ${ANSWER_IMAGE_EMBED_RULE}`
+    description: `Answer for the caller. Unless the user explicitly requests otherwise, provide at most 5 bullet points (plus optional explanation) and answer directly; reference results by bracketed number (e.g. [1], [2]) without file paths or raw chunk text, except the image-embed exception below. ${ANSWER_DELIVERABLE_RULE} ${ANSWER_CITATION_RULE} ${ANSWER_IMAGE_EMBED_RULE}`
   });
   warningsSchema = Type18.Optional(Type18.Array(Type18.String(), { description: "Optional warnings about this result set" }));
   curatedResultFields = {
@@ -23064,7 +23346,7 @@ var init_connector = () => {};
 
 // src/datasource/sidecar/config.ts
 import { readFileSync as readFileSync19 } from "node:fs";
-import { homedir as homedir3 } from "node:os";
+import { homedir as homedir4 } from "node:os";
 import { dirname as dirname18, isAbsolute as isAbsolute8, join as join32, resolve as resolve14, sep as sep7 } from "node:path";
 import { parse as parseToml2 } from "smol-toml";
 function resolveSidecarOptions(input) {
@@ -23104,9 +23386,9 @@ function archivePathFromCrawlerConfig(input) {
 }
 function expandHome(path) {
   if (path === "~")
-    return homedir3();
+    return homedir4();
   const prefixed = path.startsWith("~/") || sep7 === "\\" && path.startsWith("~\\");
-  return prefixed ? join32(homedir3(), path.slice(2)) : path;
+  return prefixed ? join32(homedir4(), path.slice(2)) : path;
 }
 var init_config2 = __esm(() => {
   init_connector();
@@ -23135,7 +23417,7 @@ function chatCrawler(id) {
     hierarchy: (metadata) => ["chats", str(metadata.chatName) ?? str(metadata.chatId) ?? "unknown"]
   };
 }
-var str = (value) => typeof value === "string" && value.length > 0 ? value : undefined, slack, CRAWLER_DESCRIPTORS;
+var str = (value) => typeof value === "string" && value.length > 0 ? value : undefined, slack, kakao, CRAWLER_DESCRIPTORS;
 var init_descriptors = __esm(() => {
   slack = {
     id: "slack",
@@ -23166,16 +23448,38 @@ var init_descriptors = __esm(() => {
       str(metadata.channelName) ?? str(metadata.channelId) ?? "unknown"
     ]
   };
+  kakao = {
+    id: "kakao",
+    keyExpression: "m.chunk_id",
+    chunkId: (key) => key,
+    fromClause: "chunks m",
+    liveWhere: "length(trim(coalesce(m.text, ''))) > 0 and not (json_valid(m.text) and (json_extract(m.text, '$.feedType') is not null or json_extract(m.text, '$.csIP') is not null))",
+    textExpression: "coalesce(m.text, '')",
+    titleExpression: "m.chat_name",
+    epochSecondsExpression: "unixepoch(m.started_at)",
+    fts: { table: "chunks_fts", joinOn: "chunks_fts.rowid = m.rowid" },
+    metadata: [
+      { sql: "m.chat_id", key: "chatId" },
+      { sql: "m.chat_name", key: "chatName" },
+      { sql: "m.sender_nickname", key: "senderNickname" },
+      { sql: "m.started_at", key: "startedAt" },
+      { sql: "m.ended_at", key: "endedAt" }
+    ],
+    channelIdExpression: "m.chat_id",
+    channelNameExpression: "m.chat_name",
+    hierarchy: (metadata) => ["chats", str(metadata.chatName) ?? str(metadata.chatId) ?? "unknown"]
+  };
   CRAWLER_DESCRIPTORS = {
     slack,
     telegram: chatCrawler("telegram"),
-    whatsapp: chatCrawler("whatsapp")
+    whatsapp: chatCrawler("whatsapp"),
+    kakao
   };
 });
 
 // src/datasource/skills/clawgallery/client.ts
 import { spawn as spawn10 } from "node:child_process";
-import { existsSync as existsSync23 } from "node:fs";
+import { existsSync as existsSync24 } from "node:fs";
 
 class ClawGalleryClient {
   options;
@@ -23183,7 +23487,7 @@ class ClawGalleryClient {
     this.options = options;
   }
   async bootstrap(signal) {
-    if (this.options.path !== undefined && !existsSync23(this.options.path)) {
+    if (this.options.path !== undefined && !existsSync24(this.options.path)) {
       return {
         ok: false,
         reason: "nonzero-exit",
@@ -24101,7 +24405,7 @@ var init_rclone_connector = __esm(() => {
 });
 
 // src/datasource/chunk-store.ts
-import { existsSync as existsSync24, mkdirSync as mkdirSync18, readFileSync as readFileSync21, writeFileSync as writeFileSync17 } from "node:fs";
+import { existsSync as existsSync25, mkdirSync as mkdirSync18, readFileSync as readFileSync21, writeFileSync as writeFileSync17 } from "node:fs";
 import { dirname as dirname20, join as join34 } from "node:path";
 function datasourceChunkStorePath(workspaceRoot, skillName, instanceId) {
   return join34(workspaceRoot, ".autorag", "datasources", sanitizeIdSegment(skillName), sanitizeIdSegment(instanceId), "chunks.json");
@@ -24141,7 +24445,7 @@ class DatasourceChunkStore {
     if (this.storePath === undefined)
       return false;
     try {
-      if (!existsSync24(this.storePath))
+      if (!existsSync25(this.storePath))
         return false;
       const parsed = JSON.parse(readFileSync21(this.storePath, "utf8"));
       if (parsed.version !== STORE_VERSION || !Array.isArray(parsed.chunks))
@@ -25668,7 +25972,7 @@ var init_github = __esm(() => {
 
 // src/datasource/skills/github-gist/connector.ts
 import { execFile as execFile2 } from "node:child_process";
-import { existsSync as existsSync25, mkdirSync as mkdirSync20, readFileSync as readFileSync23, writeFileSync as writeFileSync19 } from "node:fs";
+import { existsSync as existsSync26, mkdirSync as mkdirSync20, readFileSync as readFileSync23, writeFileSync as writeFileSync19 } from "node:fs";
 import { dirname as dirname22 } from "node:path";
 async function defaultGhTokenRunner() {
   return new Promise((resolve) => {
@@ -25788,7 +26092,7 @@ class GitHubGistConnector {
   }
   loadState() {
     const statePath = this.options.statePath;
-    if (statePath === undefined || !existsSync25(statePath))
+    if (statePath === undefined || !existsSync26(statePath))
       return { version: STATE_VERSION, gists: {} };
     try {
       const parsed = JSON.parse(readFileSync23(statePath, "utf8"));
@@ -25851,7 +26155,7 @@ var init_connector3 = () => {};
 
 // src/datasource/skills/github-gist/semantic.ts
 import { createHash as createHash13 } from "node:crypto";
-import { existsSync as existsSync26, mkdirSync as mkdirSync21, readFileSync as readFileSync24, writeFileSync as writeFileSync20 } from "node:fs";
+import { existsSync as existsSync27, mkdirSync as mkdirSync21, readFileSync as readFileSync24, writeFileSync as writeFileSync20 } from "node:fs";
 import { dirname as dirname23 } from "node:path";
 function contentHash(chunk) {
   return createHash13("sha256").update(`${chunk.title ?? ""}
@@ -25942,7 +26246,7 @@ ${chunk.content}`));
     if (this.loaded)
       return;
     this.loaded = true;
-    if (this.statePath === undefined || !existsSync26(this.statePath))
+    if (this.statePath === undefined || !existsSync27(this.statePath))
       return;
     try {
       const parsed = JSON.parse(readFileSync24(this.statePath, "utf8"));
@@ -26805,7 +27109,7 @@ var init_types5 = __esm(() => {
 
 // src/datasource/skills/lazykatok/client.ts
 import { spawn as spawn14 } from "node:child_process";
-import { existsSync as existsSync27 } from "node:fs";
+import { existsSync as existsSync28 } from "node:fs";
 import { delimiter as delimiter4, join as join36 } from "node:path";
 
 class LazykatokClient {
@@ -26995,7 +27299,7 @@ function lookupInPath2(executable, env) {
     if (dir.length === 0)
       continue;
     const candidate = join36(dir, executable);
-    if (existsSync27(candidate))
+    if (existsSync28(candidate))
       return candidate;
   }
   return;
@@ -27317,1436 +27621,1528 @@ var init_methods3 = __esm(() => {
   DEFAULT_KAKAO_TAGS = ["kakaotalk", "personal", "pii"];
 });
 
-// src/datasource/skills/lazykatok/skill.ts
-class LazykatokSkill {
-  client;
-  instanceId;
-  instances;
-  pollingIntervalMs;
-  tags;
-  lastIndexedAt;
-  constructor(options) {
-    this.client = options.client;
-    this.instanceId = options.instanceId ?? DEFAULT_INSTANCE_ID4;
-    this.instances = options.instances !== undefined && options.instances.length > 0 ? options.instances : [this.instanceId];
-    this.pollingIntervalMs = options.pollingIntervalMs ?? DEFAULT_POLLING_INTERVAL_MS3;
-    this.tags = options.tags ?? DEFAULT_KAKAO_TAGS2;
-    this.lastIndexedAt = options.lastIndexedAt;
-  }
-  describe() {
-    return {
-      name: KAKAO_DATASOURCE_ID2,
-      id: KAKAO_DATASOURCE_ID2,
-      type: KAKAO_SKILL_TYPE,
-      description: "KakaoTalk datasource via the external lazykatok CLI",
-      capabilities: ["chat", "external-cli", "polling", "bm25", "semantic"],
-      tags: this.tags,
-      status: "active",
-      requiresExternalCli: true,
-      datasourceId: KAKAO_DATASOURCE_ID2,
-      instanceId: this.instanceId,
-      instances: this.instances
-    };
+// src/datasource/sidecar/fusion.ts
+function encodeVector(vector) {
+  const floats = Float32Array.from(vector);
+  return new Uint8Array(floats.buffer, floats.byteOffset, floats.byteLength);
+}
+function decodeVector(blob) {
+  if (blob.byteLength % Float32Array.BYTES_PER_ELEMENT !== 0)
+    return;
+  return new Float32Array(new Uint8Array(blob).buffer);
+}
+function cosine3(a, b) {
+  if (a.length === 0 || a.length !== b.length)
+    return 0;
+  let dot = 0;
+  let normA = 0;
+  let normB = 0;
+  for (let index = 0;index < a.length; index += 1) {
+    const x = a[index] ?? 0;
+    const y = b[index] ?? 0;
+    dot += x * y;
+    normA += x * x;
+    normB += y * y;
   }
-  polling() {
-    return {
-      mode: "poll",
-      intervalMs: this.pollingIntervalMs,
-      lastIndexedAt: this.lastIndexedAt
-    };
+  if (normA === 0 || normB === 0)
+    return 0;
+  return dot / Math.sqrt(normA * normB);
+}
+function topKByCosine(query, rows, k) {
+  if (k <= 0)
+    return [];
+  const scored = [];
+  for (const row of rows) {
+    const score = cosine3(query, row.vector);
+    if (score > 0)
+      scored.push({ key: row.key, score });
   }
-  async index(options) {
-    try {
-      const doctor = await this.client.doctor(options?.signal);
-      if (!doctor.ok)
-        return this.fail("datasource-unavailable", doctor);
-      const sync = await this.client.sync(options?.signal);
-      if (!sync.ok)
-        return this.fail("datasource-index-failed", sync);
-      const indexResult = await this.client.index(options?.signal);
-      if (!indexResult.ok)
-        return this.fail("datasource-index-failed", indexResult);
-      this.lastIndexedAt = Date.now();
-      const chunkCount = "data" in indexResult && typeof indexResult.data === "object" && indexResult.data !== null && "chunkCount" in indexResult.data && typeof indexResult.data.chunkCount === "number" ? indexResult.data.chunkCount : 0;
-      return {
-        ok: true,
-        instanceId: this.instanceId,
-        skill: KAKAO_DATASOURCE_ID2,
-        chunkCount,
-        indexedAt: this.lastIndexedAt,
-        diagnostics: []
-      };
-    } catch (error) {
-      return this.fail("datasource-unavailable", {
-        ok: false,
-        reason: "spawn-error",
-        stdout: "",
-        stderr: describeRetrievalError(error),
-        code: null
-      });
+  scored.sort((a, b) => b.score - a.score || a.key.localeCompare(b.key));
+  return scored.slice(0, k);
+}
+function reciprocalRankFusion(lists) {
+  const scores = new Map;
+  for (const list of lists) {
+    const seen = new Set;
+    for (const [index, key] of list.entries()) {
+      if (seen.has(key))
+        continue;
+      seen.add(key);
+      scores.set(key, (scores.get(key) ?? 0) + 1 / (RRF_K2 + index + 1));
     }
   }
-  retrievalMethods() {
-    return [
-      new LazykatokBm25Method({ client: this.client, instanceId: this.instanceId, tags: this.tags }),
-      new LazykatokSemanticMethod({ client: this.client, instanceId: this.instanceId, tags: this.tags })
-    ];
-  }
-  describeSources() {
-    return this.instances.map((instanceId) => ({
-      source: datasourceSourcePath(KAKAO_DATASOURCE_ID2, instanceId),
-      datasourceId: KAKAO_DATASOURCE_ID2,
-      skill: KAKAO_DATASOURCE_ID2,
-      instanceId,
-      contentType: "chat",
-      metadata: {
-        datasourceId: KAKAO_DATASOURCE_ID2,
-        instanceId,
-        tags: this.tags
-      }
-    }));
-  }
-  skillManifest() {
-    const instanceSources = this.instances.map((instanceId) => `- \`${datasourceSourcePath(KAKAO_DATASOURCE_ID2, instanceId)}\``).join(`
-`);
-    const cadence = this.pollingIntervalMs > 0 ? `roughly every ${Math.round(this.pollingIntervalMs / 60000)} minute(s) when auto-refresh runs` : "on manual refresh only";
-    return {
-      name: `datasource-${KAKAO_DATASOURCE_ID2}`,
-      description: "Search indexed KakaoTalk chat history (messages, senders, room context). Use for questions about KakaoTalk conversations, decisions made in chats, or who said what.",
-      content: [
-        `# KakaoTalk datasource (${KAKAO_SKILL_TYPE})`,
-        "",
-        "This skill searches KakaoTalk chats that are indexed through the external `lazykatok` CLI. AutoRAG never reads KakaoTalk databases directly.",
-        "",
-        "## When to use",
-        "Use this skill when the question is about KakaoTalk conversations, chat participants, or content shared inside chats.",
-        "",
-        "## Indexing",
-        `Indexing is server-managed and refreshed ${cadence}. You do not trigger indexing; just search.`,
-        "",
-        "## How to search",
-        `Call the dedicated \`${datasourceSearchToolName(KAKAO_DATASOURCE_ID2)}\` tool with a natural-language \`query\` and \`topK\`. This datasource does not support per-source scope narrowing. Configured datasource instances:`,
-        instanceSources.length > 0 ? instanceSources : "- (no configured instances)",
-        "",
-        "Chat/channel filtering is owned by lazykatok.",
-        "",
-        "## Native CLI",
-        "The external `lazykatok` CLI owns the archive, index, and credentials. When the dedicated tool cannot express what you need, call lazykatok directly through `bash`:",
-        '- `lazykatok search bm25 "<query>" --json --limit 20` — lexical search (`semantic` mode also available)',
-        "- `lazykatok chunk get <chunkId> --json` — fetch one chunk by id (chunk ids appear in result metadata)",
-        "- `lazykatok chunk context <chunkId> --json` — surrounding messages of a chunk",
-        "Never pass datasource virtual paths (`/kakao/...`) to bash; they are not OS paths.",
-        "",
-        "## Output rules",
-        "Datasource source identifiers such as `/kakao/<instance>/chunks/<id>` are internal and opaque. Never put them, real file paths, account IDs, or phone numbers in the visible answer."
-      ].join(`
-`)
-    };
-  }
-  fail(code, result) {
-    const base = result.stderr.length > 0 ? `${result.reason}: ${result.stderr.trim().slice(0, 4000)}` : result.reason;
-    const message = code === "datasource-index-failed" && result.reason === "timeout" ? `${base} — ${connectorSyncTimeoutHint()}` : base;
-    const diagnostic = {
-      code,
-      severity: code === "datasource-unavailable" ? "warning" : "error",
-      message,
-      instanceId: this.instanceId,
-      source: KAKAO_DATASOURCE_ID2
-    };
-    return {
-      ok: false,
-      instanceId: this.instanceId,
-      skill: KAKAO_DATASOURCE_ID2,
-      indexedAt: Date.now(),
-      diagnostics: [diagnostic],
-      error: result.reason,
-      code,
-      message
-    };
-  }
+  return [...scores.entries()].map(([key, score]) => ({ key, score })).sort((a, b) => b.score - a.score);
 }
-var KAKAO_DATASOURCE_ID2 = "kakao", KAKAO_SKILL_TYPE = "kakaotalk", DEFAULT_INSTANCE_ID4 = "default", DEFAULT_POLLING_INTERVAL_MS3, DEFAULT_KAKAO_TAGS2;
-var init_skill7 = __esm(() => {
-  init_skip();
-  init_scope2();
-  init_methods3();
-  DEFAULT_POLLING_INTERVAL_MS3 = 15 * 60 * 1000;
-  DEFAULT_KAKAO_TAGS2 = ["kakaotalk", "personal", "pii"];
-});
-
-// src/datasource/skills/lazykatok/index.ts
-var init_lazykatok = __esm(() => {
-  init_client8();
-  init_methods3();
-  init_skill7();
-  init_types5();
-});
+var RRF_K2 = 60;
 
-// src/datasource/skills/mail-export/connector.ts
-import { readdir as readdir4, readFile as readFile5, stat as stat6 } from "node:fs/promises";
-import { basename as basename12, dirname as dirname24, extname as extname6, join as join37 } from "node:path";
-import { simpleParser as simpleParser2 } from "mailparser";
+// src/datasource/sidecar/index.ts
+import { createHash as createHash14 } from "node:crypto";
+import { mkdirSync as mkdirSync22, statSync as statSync3 } from "node:fs";
+import { dirname as dirname24 } from "node:path";
+import { DatabaseSync as DatabaseSync2 } from "node:sqlite";
+import { pathToFileURL as pathToFileURL2 } from "node:url";
 
-class MailExportConnector {
+class CrawlerSidecar {
+  descriptor;
   options;
-  constructor(options = {}) {
+  select;
+  where;
+  opened;
+  attached;
+  lastSemantic;
+  constructor(options) {
     this.options = options;
+    const { descriptor } = options;
+    this.descriptor = descriptor;
+    const metadataColumns = descriptor.metadata.map((column, index) => `${column.sql} as m${index}`);
+    this.select = [
+      `${descriptor.keyExpression} as k`,
+      `${descriptor.textExpression} as content`,
+      `${descriptor.titleExpression ?? "null"} as title`,
+      `${descriptor.epochSecondsExpression ?? "null"} as epoch`,
+      ...metadataColumns
+    ].join(", ");
+    const scope = descriptor.scopeExpression !== undefined && options.workspaceId !== undefined;
+    const allowlist = [
+      ...(options.channelIds?.length ?? 0) > 0 ? [`${descriptor.channelIdExpression} in (${options.channelIds?.map(() => "?").join(", ")})`] : [],
+      ...(options.channelNames?.length ?? 0) > 0 ? [`${descriptor.channelNameExpression} in (${options.channelNames?.map(() => "?").join(", ")})`] : []
+    ];
+    this.where = [
+      descriptor.liveWhere,
+      ...scope ? [`${descriptor.scopeExpression} = ?`] : [],
+      ...allowlist.length > 0 ? [`(${allowlist.join(" or ")})`] : []
+    ].join(" and ");
   }
-  async fetch() {
-    const paths = this.options.paths ?? [];
-    if (paths.length === 0)
-      return { ok: false, reason: "not-configured", message: "no mail export paths configured" };
-    const maxDocuments = this.options.maxDocuments ?? DEFAULT_MAX_DOCUMENTS3;
-    const maxBytes = this.options.maxBytesPerFile ?? DEFAULT_MAX_BYTES_PER_FILE2;
-    const files = [];
-    let missingPaths = 0;
-    for (const path of paths) {
-      try {
-        const info = await stat6(path);
-        if (info.isDirectory()) {
-          const entries = await readdir4(path, { recursive: true, withFileTypes: true });
-          for (const entry of entries) {
-            if (!entry.isFile())
-              continue;
-            const extension = extname6(entry.name).toLowerCase();
-            if (extension === ".eml" || extension === ".mbox") {
-              files.push(join37(entry.parentPath, entry.name));
-            }
-          }
-        } else if (info.isFile()) {
-          files.push(path);
-        }
-      } catch {
-        missingPaths += 1;
-      }
+  get db() {
+    if (this.opened !== undefined)
+      return this.opened;
+    mkdirSync22(dirname24(this.options.sidecarDbPath), { recursive: true });
+    const db = new DatabaseSync2(this.options.sidecarDbPath);
+    db.exec(`pragma busy_timeout = ${BUSY_TIMEOUT_MS}`);
+    db.exec("pragma journal_mode = wal");
+    db.exec(`
+			create table if not exists sidecar_meta (k text primary key, v text not null);
+			create table if not exists sidecar_vec (message_key text primary key, hash text not null, vec blob not null);
+		`);
+    this.opened = db;
+    return db;
+  }
+  attachCrawler() {
+    let file;
+    try {
+      const { dev, ino } = statSync3(this.options.crawlerDbPath);
+      file = { dev, ino };
+    } catch {
+      this.detachCrawler();
+      return false;
     }
-    if (missingPaths === paths.length) {
-      return { ok: false, reason: "unavailable", message: "no configured mail export paths were readable" };
+    if (this.attached !== undefined) {
+      if (this.attached.dev === file.dev && this.attached.ino === file.ino)
+        return true;
+      this.detachCrawler();
     }
-    const documents = [];
-    const warnings = [];
-    if (missingPaths > 0)
-      warnings.push(`${missingPaths} configured path(s) were unreadable`);
-    let parseFailures = 0;
-    let skippedLarge = 0;
-    for (const file of files.sort()) {
-      if (documents.length >= maxDocuments)
-        break;
-      let bytes;
-      try {
-        const info = await stat6(file);
-        if (info.size > maxBytes) {
-          skippedLarge += 1;
-          continue;
-        }
-        bytes = await readFile5(file);
-      } catch {
-        parseFailures += 1;
-        continue;
-      }
-      const stem = basename12(file, extname6(file));
-      const mailboxDir = basename12(dirname24(file));
-      const hierarchy = ["mailboxes", mailboxDir, stem];
-      const extension = extname6(file).toLowerCase();
-      const rawMessages = extension === ".mbox" ? splitMbox(bytes.toString("utf8")) : [bytes.toString("utf8")];
-      for (const [index, raw] of rawMessages.entries()) {
-        if (documents.length >= maxDocuments)
-          break;
-        try {
-          const parsed = await simpleParser2(raw);
-          const subject = parsed.subject ?? "(no subject)";
-          const from = Array.isArray(parsed.from) ? parsed.from[0]?.text : parsed.from?.text;
-          const to = Array.isArray(parsed.to) ? parsed.to[0]?.text : parsed.to?.text;
-          const body = (parsed.text ?? "").trim().slice(0, MAX_BODY_CHARS);
-          const headerBlock = [
-            `Subject: ${subject}`,
-            ...from !== undefined ? [`From: ${from}`] : [],
-            ...to !== undefined ? [`To: ${to}`] : [],
-            ...parsed.date !== undefined ? [`Date: ${parsed.date.toISOString()}`] : []
-          ].join(`
-`);
-          const content = `${headerBlock}
-
-${body}`.trim();
-          if (content.length === 0)
-            continue;
-          documents.push({
-            docId: extension === ".mbox" ? `${stem}-${index + 1}` : stem,
-            hierarchy,
-            title: subject,
-            content,
-            ...parsed.date !== undefined ? { publishedAt: parsed.date.getTime() } : {},
-            metadata: {
-              path: file,
-              subject,
-              ...from !== undefined ? { from } : {},
-              messageIndex: index
-            }
-          });
-        } catch {
-          parseFailures += 1;
-        }
-      }
+    const crawlerUri = `${pathToFileURL2(this.options.crawlerDbPath).href}?mode=ro`.replaceAll("'", "''");
+    try {
+      this.db.exec(`attach database '${crawlerUri}' as crawler`);
+    } catch {
+      return false;
     }
-    if (skippedLarge > 0)
-      warnings.push(`${skippedLarge} file(s) exceeded the size limit and were skipped`);
-    if (parseFailures > 0)
-      warnings.push(`${parseFailures} message(s) failed to parse`);
-    return { ok: true, documents, ...warnings.length > 0 ? { warnings } : {} };
+    const { descriptor } = this.options;
+    try {
+      this.db.prepare(`select ${this.select} from ${descriptor.fromClause} where ${this.where} limit 0`).all(...this.whereParams());
+      this.db.prepare(`select 1 from ${descriptor.fts.table} limit 0`).all();
+    } catch {
+      this.db.exec("detach database crawler");
+      return false;
+    }
+    this.attached = file;
+    this.lastSemantic = undefined;
+    return true;
   }
-}
-function splitMbox(text) {
-  const lines = text.split(/\r?\n/u);
-  const messages = [];
-  let current = [];
-  for (const line of lines) {
-    if (MBOX_SEPARATOR.test(line) && line.startsWith("From ")) {
-      if (current.length > 0)
-        messages.push(current.join(`
-`));
-      current = [];
-      continue;
-    }
-    current.push(line);
-  }
-  if (current.length > 0)
-    messages.push(current.join(`
-`));
-  return messages.filter((message) => message.trim().length > 0);
-}
-var DEFAULT_MAX_DOCUMENTS3 = 500, DEFAULT_MAX_BYTES_PER_FILE2, MAX_BODY_CHARS = 50000, MBOX_SEPARATOR;
-var init_connector4 = __esm(() => {
-  DEFAULT_MAX_BYTES_PER_FILE2 = 50 * 1024 * 1024;
-  MBOX_SEPARATOR = /^From .*\d{4}$/mu;
-});
-
-// src/datasource/skills/mail-export/skill.ts
-var MAIL_EXPORT_SKILL_DEFINITION, MailExportSkill;
-var init_skill8 = __esm(() => {
-  init_connector_skill();
-  init_connector4();
-  MAIL_EXPORT_SKILL_DEFINITION = {
-    skillName: "mail-export",
-    skillType: "mail-export",
-    description: "Local mail export datasource",
-    capabilities: ["email", "filesystem"],
-    defaultTags: ["mail-export", "email", "pii"],
-    contentType: "email",
-    manifestDescription: "Search indexed local mail exports (mbox/eml archives). Use for questions about archived email conversations.",
-    manifestNotes: [
-      "Only the server-configured export paths are indexed; archives are read locally and never uploaded."
-    ]
-  };
-  MailExportSkill = class MailExportSkill extends ConnectorDatasourceSkill {
-    constructor(options = {}) {
-      const { connector, connectorOptions, ...rest } = options;
-      super(MAIL_EXPORT_SKILL_DEFINITION, {
-        ...rest,
-        connector: connector ?? new MailExportConnector(connectorOptions ?? {})
-      });
-    }
-  };
-});
-
-// src/datasource/skills/mail-export/index.ts
-var init_mail_export = __esm(() => {
-  init_connector4();
-  init_skill8();
-});
-
-// src/datasource/skills/mailcrawl/config.ts
-import { chmodSync, mkdirSync as mkdirSync22 } from "node:fs";
-function validateMailcrawlInstanceId(instanceId) {
-  if (!/^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$/u.test(instanceId)) {
-    throw new Error("mailcrawl instanceId must be a safe single path segment");
+  detachCrawler() {
+    if (this.attached === undefined)
+      return;
+    this.attached = undefined;
+    this.lastSemantic = undefined;
+    try {
+      this.opened?.exec("detach database crawler");
+    } catch {}
   }
-}
-function ensurePrivateMailcrawlDataDir(path) {
-  mkdirSync22(path, { recursive: true, mode: 448 });
-  if (process.platform !== "win32")
-    chmodSync(path, 448);
-}
-var init_config3 = () => {};
-
-// src/datasource/skills/mailcrawl/client.ts
-import { spawn as spawn15 } from "node:child_process";
-
-class MailcrawlClient {
-  options;
-  constructor(options = {}) {
-    this.options = options;
-    if (options.instanceId !== undefined)
-      validateMailcrawlInstanceId(options.instanceId);
+  get semanticEnabled() {
+    return this.options.embedder !== undefined;
   }
-  async sync(signal) {
-    const args = [
-      "sync",
-      "--json",
-      ...this.options.account === undefined ? [] : ["--account", this.options.account],
-      ...this.options.mailbox === undefined ? [] : ["--mailbox", this.options.mailbox],
-      ...this.options.backend === undefined ? [] : ["--backend", this.options.backend],
-      ...this.options.source === undefined ? [] : ["--source", this.options.source],
-      ...this.options.fixture === undefined ? [] : ["--fixture", this.options.fixture],
-      ...this.options.himalayaConfig === undefined ? [] : ["--himalaya-config", this.options.himalayaConfig]
-    ];
-    const result = await this.run(args, signal, this.indexTimeoutMs());
-    if (!result.ok)
-      return failure3(result);
-    const parsed = parseObject2(result.stdout);
-    if (parsed === undefined || !hasAnyNumber(parsed, ["added", "updated", "deleted", "unchanged", "chunksAdded"])) {
-      return failure3({ ...result, ok: false, reason: "invalid-output" });
-    }
-    const messages = number(parsed, "added") ?? number(parsed, "updated") ?? number(parsed, "unchanged") ?? 0;
-    return ok4({
-      messages,
-      added: number(parsed, "added"),
-      updated: number(parsed, "updated"),
-      deleted: number(parsed, "deleted"),
-      unchanged: number(parsed, "unchanged"),
-      chunksAdded: number(parsed, "chunksAdded"),
-      archiveRevision: string(parsed, "archiveRevision")
-    }, result);
+  close() {
+    this.opened?.close();
+    this.opened = undefined;
+    this.attached = undefined;
+    this.lastSemantic = undefined;
   }
-  async search(mode, query, options) {
-    if (query.trim().length === 0)
-      return { ok: true, hits: [], stdout: "", stderr: "", code: 0 };
-    const args = [
-      "search",
-      "--mode",
-      mode === "keyword" ? "bm25" : mode,
-      ...this.options.account === undefined ? [] : ["--account", this.options.account],
-      ...this.options.mailbox === undefined ? [] : ["--mailbox", this.options.mailbox],
-      "--limit",
-      String(options?.topK ?? 20),
-      "--json",
-      query.trim()
+  whereParams() {
+    const { descriptor, workspaceId, channelIds = [], channelNames = [] } = this.options;
+    return [
+      ...descriptor.scopeExpression !== undefined && workspaceId !== undefined ? [workspaceId] : [],
+      ...channelIds,
+      ...channelNames
     ];
-    const result = await this.run(args, options?.signal, this.options.timeoutMs ?? DEFAULT_SEARCH_TIMEOUT_MS2);
-    if (!result.ok)
-      return failure3(result);
-    const parsed = parseJson3(result.stdout);
-    if (parsed === undefined)
-      return failure3({ ...result, ok: false, reason: "invalid-output" });
-    const rows = Array.isArray(parsed) ? parsed : objectArray(parsed, ["hits", "results", "items"]);
-    if (rows === undefined)
-      return failure3({ ...result, ok: false, reason: "invalid-output" });
-    const hits = rows.map((row, index) => normalizeHit2(row, mode, index));
-    if (hits.some((hit) => hit === undefined)) {
-      return failure3({ ...result, ok: false, reason: "invalid-output" });
-    }
-    return {
-      ok: true,
-      hits,
-      stdout: result.stdout,
-      stderr: result.stderr,
-      code: result.code ?? 0
-    };
   }
-  indexTimeoutMs() {
-    return this.options.indexTimeoutMs ?? this.options.timeoutMs ?? DEFAULT_INDEX_TIMEOUT_MS3;
+  searchLexical(query, topK) {
+    const match = toMatchQuery(query);
+    if (match === undefined || topK <= 0 || !this.attachCrawler())
+      return [];
+    const { descriptor } = this.options;
+    const rows = this.db.prepare(`select ${this.select}, bm25(${descriptor.fts.table}) as rank
+				from ${descriptor.fromClause} join ${descriptor.fts.table} on ${descriptor.fts.joinOn}
+				where ${descriptor.fts.table} match ? and ${this.where}
+				order by rank limit ?`).all(match, ...this.whereParams(), topK);
+    return rows.map((row) => this.toHit(this.toRow(row), -Number(row.rank)));
   }
-  async run(args, signal, timeoutMs) {
-    const violation = findRemoteEmbeddingViolation({ ...process.env, ...this.options.env ?? {} });
-    if (violation !== undefined) {
-      return { ok: false, reason: "remote-embedding-rejected", stdout: "", stderr: "", code: null };
+  async searchSemantic(query, topK) {
+    const trimmed = query.trim();
+    if (this.options.embedder === undefined || trimmed.length === 0 || topK <= 0 || !this.attachCrawler())
+      return [];
+    const ranked = await this.rankSemantic(trimmed, topK * POOL_FACTOR);
+    if (ranked.length === 0)
+      return [];
+    const live = this.hydrate(ranked.map((entry) => entry.key));
+    const hits = [];
+    for (const entry of ranked) {
+      const row = live.get(entry.key);
+      if (row !== undefined)
+        hits.push(this.toHit(row, entry.score));
+      if (hits.length >= topK)
+        break;
     }
-    const dataDir = this.options.dataDir;
-    if (dataDir !== undefined)
-      ensurePrivateMailcrawlDataDir(dataDir);
-    const env = {};
-    for (const [key, value] of Object.entries(process.env)) {
-      if (value !== undefined && (SAFE_ENV_KEYS2.has(key) || SAFE_MAILCRAWL_ENV_KEYS.has(key)))
-        env[key] = value;
+    return hits;
+  }
+  async rankSemantic(query, pool) {
+    const now = Date.now();
+    let cached = this.lastSemantic;
+    if (cached === undefined || cached.query !== query || now - cached.at > SEMANTIC_REUSE_MS) {
+      const embedder = this.options.embedder;
+      if (embedder === undefined)
+        return [];
+      cached = {
+        query,
+        at: now,
+        vector: embedder.embed([query]).then((vectors) => vectors[0], () => {
+          return;
+        }),
+        ranked: undefined
+      };
+      this.lastSemantic = cached;
     }
-    for (const [key, value] of Object.entries(this.options.env ?? {})) {
-      if (value === undefined)
-        delete env[key];
-      else if (SAFE_ENV_KEYS2.has(key) || SAFE_MAILCRAWL_ENV_KEYS.has(key))
-        env[key] = value;
+    const vector = await cached.vector;
+    if (vector === undefined) {
+      if (this.lastSemantic === cached)
+        this.lastSemantic = undefined;
+      return [];
     }
-    if (dataDir !== undefined)
-      env.MAILCRAWL_DATA_DIR = dataDir;
-    return spawnMailcrawl(this.options.binaryPath ?? DEFAULT_BINARY4, args, env, this.options, signal, timeoutMs);
-  }
-  async index(signal) {
-    const result = await this.run(["index", "--json"], signal, this.indexTimeoutMs());
-    if (!result.ok)
-      return failure3(result);
-    const parsed = parseObject2(result.stdout);
-    if (parsed === undefined || number(parsed, "embedded") === undefined && number(parsed, "reused") === undefined && string(parsed, "generation") === undefined) {
-      return failure3({ ...result, ok: false, reason: "invalid-output" });
+    if (cached.ranked === undefined || cached.ranked.pool < pool) {
+      const wide = pool * POOL_FACTOR;
+      cached.ranked = { pool: wide, keys: topKByCosine(vector, this.storedVectors(), wide) };
     }
-    return ok4({
-      embedded: number(parsed, "embedded"),
-      reused: number(parsed, "reused"),
-      generation: string(parsed, "generation"),
-      archiveRevision: string(parsed, "archiveRevision"),
-      rebuilt: typeof parsed.rebuilt === "boolean" ? parsed.rebuilt : undefined,
-      embedder: string(parsed, "embedder")
-    }, result);
+    return cached.ranked.keys.slice(0, pool);
   }
-}
-function spawnMailcrawl(binary, args, env, options, signal, timeoutMs) {
-  return new Promise((resolveResult) => {
-    const portable = portableSpawnCommand(binary, args);
-    const child = spawn15(portable.command, portable.args, {
-      env,
-      detached: process.platform !== "win32",
-      stdio: ["ignore", "pipe", "pipe"]
-    });
-    let stdout = { text: "", bytes: 0, exceeded: false }, stderr = { text: "", bytes: 0, exceeded: false }, settled = false, reason;
-    const max = options.maxBufferBytes ?? DEFAULT_MAX_BUFFER_BYTES4;
-    const timer = setTimeout(() => {
-      reason = "timeout";
-      terminate4(child);
-    }, timeoutMs);
-    const abort = () => {
-      reason = "aborted";
-      terminate4(child);
+  async searchHybrid(query, topK) {
+    if (topK <= 0)
+      return [];
+    const pool = topK * POOL_FACTOR;
+    const lexical = this.searchLexical(query, pool);
+    const semantic = await this.searchSemantic(query, pool);
+    if (semantic.length === 0)
+      return lexical.slice(0, topK);
+    const byKey = new Map;
+    for (const hit of [...semantic, ...lexical])
+      byKey.set(hit.key, hit);
+    const fused = [];
+    for (const entry of reciprocalRankFusion([lexical.map((hit) => hit.key), semantic.map((hit) => hit.key)])) {
+      const hit = byKey.get(entry.key);
+      if (hit !== undefined)
+        fused.push({ ...hit, score: entry.score });
+      if (fused.length >= topK)
+        break;
+    }
+    return fused;
+  }
+  async sync() {
+    const { embedder } = this.options;
+    if (embedder === undefined)
+      return { ok: false, message: "no embedder configured" };
+    if (!this.attachCrawler())
+      return { ok: false, message: "crawler archive not available yet" };
+    let identity;
+    try {
+      identity = await embedder.identity();
+    } catch (error) {
+      return { ok: false, message: errorMessage2(error) };
+    }
+    this.adoptIdentity(identity);
+    const pruned = this.prune();
+    const pending = this.pendingKeys();
+    const batchSize = Math.max(1, this.options.batchSize ?? DEFAULT_BATCH_SIZE);
+    const budget = Math.max(1, this.options.maxEmbedPerSync ?? DEFAULT_MAX_EMBED_PER_SYNC);
+    const target = Math.min(pending.length, budget);
+    const embed = async (rows) => {
+      const vectors = await embedder.embed(rows.map((row) => row.input));
+      if (vectors.length !== rows.length) {
+        throw new Error(`embedder returned ${vectors.length} vectors for ${rows.length} messages`);
+      }
+      const wrong = vectors.find((vector) => vector.length !== identity.dimension);
+      if (wrong !== undefined) {
+        throw new Error(`embedder returned a ${wrong.length}-dimension vector, expected ${identity.dimension}`);
+      }
+      return vectors;
     };
-    if (signal?.aborted)
-      abort();
-    signal?.addEventListener("abort", abort, { once: true });
-    child.stdout.setEncoding("utf8");
-    child.stderr.setEncoding("utf8");
-    child.stdout.on("data", (chunk) => {
-      stdout = appendOutput(stdout, chunk, max);
-      if (stdout.exceeded && reason === undefined) {
-        reason = "stdout-too-large";
-        terminate4(child);
+    let embedded = 0;
+    let skipped = 0;
+    for (let offset = 0;offset < target; offset += batchSize) {
+      const rows = [...this.hydrate(pending.slice(offset, Math.min(offset + batchSize, target))).values()];
+      if (rows.length === 0)
+        continue;
+      try {
+        this.store(rows, await embed(rows));
+        embedded += rows.length;
+      } catch (batchError) {
+        const rejected = [];
+        let lastError = batchError;
+        for (const row of rows) {
+          try {
+            this.store([row], await embed([row]));
+            embedded += 1;
+          } catch (error) {
+            rejected.push(row);
+            lastError = error;
+          }
+        }
+        if (rejected.length === 0)
+          continue;
+        if (embedded === 0)
+          return { ok: false, message: errorMessage2(lastError) };
+        this.store(rejected, undefined);
+        skipped += rejected.length;
       }
+    }
+    return { ok: true, embedded, skipped, pruned, remaining: pending.length - embedded - skipped };
+  }
+  adoptIdentity(identity) {
+    const serialized = JSON.stringify(identity);
+    const current = this.db.prepare("select v from sidecar_meta where k = ?").get(IDENTITY_KEY);
+    if (current?.v === serialized)
+      return;
+    this.transaction(() => {
+      this.db.exec("delete from sidecar_vec");
+      this.db.prepare("insert or replace into sidecar_meta (k, v) values (?, ?)").run(IDENTITY_KEY, serialized);
     });
-    child.stderr.on("data", (chunk) => {
-      stderr = appendOutput(stderr, chunk, max);
-      if (stderr.exceeded && reason === undefined) {
-        reason = "stderr-too-large";
-        terminate4(child);
+    this.lastSemantic = undefined;
+  }
+  prune() {
+    const { descriptor } = this.options;
+    const result = this.db.prepare(`delete from sidecar_vec where message_key not in (
+					select ${descriptor.keyExpression} from ${descriptor.fromClause} where ${this.where}
+				)`).run(...this.whereParams());
+    if (Number(result.changes) > 0)
+      this.lastSemantic = undefined;
+    return Number(result.changes);
+  }
+  pendingKeys() {
+    const { descriptor } = this.options;
+    const rows = this.db.prepare(`select ${this.select}, v.hash as stored_hash
+				from ${descriptor.fromClause}
+				left join sidecar_vec v on v.message_key = ${descriptor.keyExpression}
+				where ${this.where}`).iterate(...this.whereParams());
+    const pending = [];
+    for (const row of rows) {
+      if (row.stored_hash !== this.toRow(row).hash)
+        pending.push(String(row.k));
+    }
+    return pending;
+  }
+  hydrate(keys) {
+    const { descriptor } = this.options;
+    const out = new Map;
+    if (keys.length === 0)
+      return out;
+    const rows = this.db.prepare(`select ${this.select} from ${descriptor.fromClause}
+				where ${descriptor.keyExpression} in (${keys.map(() => "?").join(", ")}) and ${this.where}`).all(...keys, ...this.whereParams());
+    const byKey = new Map(rows.map((row) => [String(row.k), this.toRow(row)]));
+    for (const key of keys) {
+      const row = byKey.get(key);
+      if (row !== undefined)
+        out.set(key, row);
+    }
+    return out;
+  }
+  store(rows, vectors) {
+    const upsert = this.db.prepare("insert or replace into sidecar_vec (message_key, hash, vec) values (?, ?, ?)");
+    this.transaction(() => {
+      for (const [index, row] of rows.entries()) {
+        upsert.run(row.key, row.hash, encodeVector(vectors?.[index] ?? []));
       }
     });
-    child.on("error", (error) => {
-      if (settled)
-        return;
-      settled = true;
-      clearTimeout(timer);
-      signal?.removeEventListener("abort", abort);
-      resolveResult({
-        ok: false,
-        reason: error.code === "ENOENT" ? "binary-missing" : "spawn-error",
-        stdout: stdout.text,
-        stderr: "",
-        code: null
-      });
-    });
-    child.on("close", (code) => {
-      if (settled)
-        return;
-      settled = true;
-      clearTimeout(timer);
-      signal?.removeEventListener("abort", abort);
-      resolveResult({
-        ok: code === 0 && reason === undefined,
-        ...reason === undefined ? code === 0 ? {} : { reason: "nonzero-exit" } : { reason },
-        stdout: stdout.text,
-        stderr: stderr.text,
-        code
-      });
-    });
-  });
-}
-function terminate4(child) {
-  if (child.killed)
-    return;
-  if (process.platform !== "win32" && child.pid !== undefined) {
+    this.lastSemantic = undefined;
+  }
+  *storedVectors() {
+    const rows = this.db.prepare("select message_key, vec from sidecar_vec where length(vec) > 0").iterate();
+    for (const row of rows) {
+      const vector = row.vec instanceof Uint8Array ? decodeVector(row.vec) : undefined;
+      if (vector !== undefined)
+        yield { key: String(row.message_key), vector };
+    }
+  }
+  transaction(work) {
+    this.db.exec("begin immediate");
     try {
-      process.kill(-child.pid, "SIGTERM");
-      return;
-    } catch {}
+      work();
+      this.db.exec("commit");
+    } catch (error) {
+      this.db.exec("rollback");
+      throw error;
+    }
   }
-  child.kill("SIGTERM");
-}
-function parseJson3(text) {
-  try {
-    return JSON.parse(text.trim());
-  } catch {
-    return;
+  toRow(row) {
+    const metadata = {};
+    for (const [index, column] of this.options.descriptor.metadata.entries()) {
+      const value = row[`m${index}`];
+      if (typeof value === "string" && value.length > 0 || typeof value === "number")
+        metadata[column.key] = String(value);
+    }
+    const title = typeof row.title === "string" && row.title.length > 0 ? row.title : undefined;
+    const epoch = typeof row.epoch === "number" && Number.isFinite(row.epoch) ? row.epoch : undefined;
+    const content = String(row.content ?? "");
+    const input = `${title ?? ""}
+${content}`.slice(0, MAX_EMBED_INPUT_CHARS);
+    return {
+      key: String(row.k),
+      content,
+      title,
+      epoch,
+      metadata,
+      input,
+      hash: createHash14("sha1").update(input).digest("hex")
+    };
+  }
+  toHit(row, score) {
+    return {
+      key: row.key,
+      content: row.content,
+      score,
+      ...row.title !== undefined ? { title: row.title } : {},
+      ...row.epoch !== undefined ? { publishedAt: Math.round(row.epoch * 1000) } : {},
+      metadata: row.metadata
+    };
   }
 }
-function parseObject2(text) {
-  const value = parseJson3(text);
-  return value !== null && typeof value === "object" && !Array.isArray(value) ? value : undefined;
-}
-function objectArray(value, keys) {
-  if (value === null || typeof value !== "object" || Array.isArray(value))
-    return;
-  for (const key of keys)
-    if (Array.isArray(value[key]))
-      return value[key];
-  return;
-}
-function hasAnyNumber(value, keys) {
-  return keys.some((key) => number(value, key) !== undefined);
-}
-function normalizeHit2(value, mode, index) {
-  if (value === null || typeof value !== "object" || Array.isArray(value))
-    return;
-  const row = value;
-  const text = (key) => typeof row[key] === "string" ? row[key] : "";
-  const chunkId = text("chunkId") || text("chunk_id");
-  const messageId = text("messageId") || text("message_id");
-  const accountId = text("accountId") || text("account_id");
-  const mailbox = text("mailbox");
-  const snippet = text("snippet") || text("content") || text("text");
-  if (!isSafeOpaqueId(chunkId) || !isSafeOpaqueId(messageId) || !snippet)
-    return;
-  const to = Array.isArray(row.to) ? row.to.filter((item) => typeof item === "string") : [];
-  return {
-    chunkId,
-    messageId,
-    threadId: text("threadId") || text("thread_id"),
-    accountId,
-    mailbox,
-    subject: text("subject"),
-    from: text("from"),
-    to,
-    date: text("date"),
-    snippet,
-    score: typeof row.score === "number" ? row.score : 1 / (index + 1),
-    mode
-  };
-}
-function number(value, key) {
-  return typeof value[key] === "number" && Number.isFinite(value[key]) ? value[key] : undefined;
-}
-function string(value, key) {
-  return typeof value[key] === "string" ? value[key] : undefined;
-}
-function ok4(data, result) {
-  return { ok: true, data, stdout: result.stdout, stderr: result.stderr, code: result.code ?? 0 };
-}
-function failure3(result) {
-  return {
-    ok: false,
-    reason: result.reason ?? "nonzero-exit",
-    stdout: result.stdout,
-    stderr: boundDiagnosticText6(result.stderr),
-    code: result.code
-  };
-}
-function boundDiagnosticText6(value) {
-  if (value.length === 0)
-    return "";
-  return value.trim().slice(0, 4000);
+function toMatchQuery(text) {
+  const terms = [...new Set(text.match(/[\p{L}\p{N}_]+/gu) ?? [])];
+  return terms.length === 0 ? undefined : terms.map((term) => `"${term}"`).join(" OR ");
 }
-function appendOutput(current, chunk, maxBytes) {
-  if (current.exceeded)
-    return current;
-  const bytes = Buffer.from(chunk, "utf8");
-  const remaining = Math.max(0, maxBytes - current.bytes);
-  const prefix = truncateUtf8(chunk, remaining);
-  const total = current.bytes + bytes.byteLength;
-  return {
-    text: current.text + prefix,
-    bytes: Math.min(total, maxBytes),
-    exceeded: total > maxBytes
-  };
+function errorMessage2(error) {
+  return error instanceof Error ? error.message : String(error);
 }
-function isSafeOpaqueId(value) {
-  return value.length > 0 && value !== "." && value !== ".." && !/[\\/#\u0000-\u001f\u007f]/u.test(value);
+var DEFAULT_BATCH_SIZE = 256, DEFAULT_MAX_EMBED_PER_SYNC = 2048, MAX_EMBED_INPUT_CHARS = 8192, BUSY_TIMEOUT_MS = 5000, POOL_FACTOR = 3, SEMANTIC_REUSE_MS = 1e4, IDENTITY_KEY = "identity";
+var init_sidecar = () => {};
+
+// src/datasource/sidecar/methods.ts
+function sidecarMethodName(datasourceId, mode) {
+  return mode === "bm25" ? `${datasourceId}-sidecar-bm25` : `${datasourceId}-${mode}`;
 }
-function findRemoteEmbeddingViolation(env) {
-  for (const [key, value] of Object.entries(env)) {
-    if (value === undefined)
-      continue;
-    const lower = key.toLowerCase();
-    if (REMOTE_EMBEDDING_ENV_KEYS.has(lower))
-      return key;
-    if (lower === "mailcrawl_embedder" || lower === "mailcrawl_embed_url") {
-      if (isRemoteHttpEndpoint(value))
-        return key;
-    }
+
+class SidecarRetrievalMethod {
+  options;
+  constructor(options) {
+    this.options = options;
   }
-  return;
-}
-function isRemoteHttpEndpoint(value) {
-  const trimmed = value.trim();
-  if (!/^https?:\/\//iu.test(trimmed))
-    return false;
-  try {
-    return !LOOPBACK_HOSTS2.has(new URL(trimmed).hostname);
-  } catch {
-    return true;
+  describe() {
+    const { datasourceId, skillType, mode, tags } = this.options;
+    const { type, text, capabilities } = DESCRIPTIONS[mode];
+    return {
+      name: sidecarMethodName(datasourceId, mode),
+      type,
+      description: `${text} over ${skillType} via a read-only sidecar index`,
+      status: "active",
+      capabilities: [...capabilities, "scoped", "path-opaque-sources"],
+      datasourceId,
+      tags
+    };
   }
-}
-function truncateUtf8(value, maxBytes) {
-  if (Buffer.byteLength(value) <= maxBytes)
-    return value;
-  const decoder = new TextDecoder("utf-8", { fatal: true });
-  const bytes = Buffer.from(value, "utf8");
-  for (let end = Math.min(maxBytes, bytes.length);end >= 0; end -= 1) {
-    try {
-      return decoder.decode(bytes.subarray(0, end));
-    } catch {}
+  async retrieve(query, options) {
+    const trimmed = query.trim();
+    if (trimmed.length === 0)
+      return [];
+    const topK = options.topK ?? DEFAULT_TOP_K5;
+    const { sidecar, mode } = this.options;
+    const hits = mode === "bm25" ? sidecar.searchLexical(trimmed, topK) : mode === "semantic" ? await sidecar.searchSemantic(trimmed, topK) : await sidecar.searchHybrid(trimmed, topK);
+    const results = [];
+    for (const hit of hits) {
+      const result = this.toResult(hit, options);
+      if (result !== undefined)
+        results.push(result);
+    }
+    return results;
+  }
+  toResult(hit, options) {
+    const { descriptor, datasourceId, instanceId, mode } = this.options;
+    const chunkId = descriptor.chunkId(hit.key, hit.metadata);
+    const source = datasourceSourcePath(datasourceId, instanceId, chunkId);
+    if (!matchesDatasourceScope(source, options.scope))
+      return;
+    const method = sidecarMethodName(datasourceId, mode);
+    return {
+      id: `${datasourceId}:${instanceId}:${chunkId}`,
+      content: hit.content,
+      source,
+      score: hit.score,
+      metadata: {
+        ...hit.metadata,
+        method,
+        datasourceId,
+        instanceId,
+        backend: "sidecar",
+        ...hit.title !== undefined ? { title: hit.title } : {},
+        hierarchy: descriptor.hierarchy(hit.metadata).join("/"),
+        ...hit.publishedAt !== undefined ? { publishedAt: hit.publishedAt } : {}
+      }
+    };
   }
-  return "";
 }
-var DEFAULT_BINARY4 = "mailcrawl", DEFAULT_SEARCH_TIMEOUT_MS2 = 60000, DEFAULT_INDEX_TIMEOUT_MS3 = 1800000, DEFAULT_MAX_BUFFER_BYTES4 = 1048576, LOOPBACK_HOSTS2, SAFE_ENV_KEYS2, SAFE_MAILCRAWL_ENV_KEYS, REMOTE_EMBEDDING_ENV_KEYS;
-var init_client9 = __esm(() => {
-  init_portable_spawn();
-  init_config3();
-  LOOPBACK_HOSTS2 = new Set(["127.0.0.1", "localhost", "[::1]", "::1"]);
-  SAFE_ENV_KEYS2 = new Set(["HOME", "LANG", "LC_ALL", "PATH", "TMPDIR", "TMP", "TEMP", "NO_COLOR"]);
-  SAFE_MAILCRAWL_ENV_KEYS = new Set([
-    "MAILCRAWL_DATA_DIR",
-    "MAILCRAWL_EMBEDDER",
-    "MAILCRAWL_EMBEDDER_PROVIDER",
-    "MAILCRAWL_EMBED_URL",
-    "MAILCRAWL_EMBED_MODEL",
-    "MAILCRAWL_EMBED_DIM",
-    "MAILCRAWL_EMBED_TIMEOUT",
-    "MAILCRAWL_EMBED_BATCH_SIZE",
-    "MAILCRAWL_QUERY_PREFIX",
-    "MAILCRAWL_PASSAGE_PREFIX",
-    "MAILCRAWL_NATIVE_MODEL",
-    "MAILCRAWL_NATIVE_DEVICE",
-    "MAILCRAWL_NATIVE_DTYPE",
-    "MAILCRAWL_KIWI_MODEL",
-    "MAILCRAWL_JA_HELPER",
-    "MAILCRAWL_ZH_HELPER"
-  ]);
-  REMOTE_EMBEDDING_ENV_KEYS = new Set([
-    "mailcrawl_embedder_base_url",
-    "mailcrawl_embedding_base_url",
-    "mailcrawl_allow_remote_embeddings"
-  ]);
+var DEFAULT_TOP_K5 = 20, DESCRIPTIONS;
+var init_methods4 = __esm(() => {
+  init_scope2();
+  DESCRIPTIONS = {
+    bm25: { type: "bm25", text: "BM25-ranked FTS5 retrieval", capabilities: ["lexical", "fts5", "bm25"] },
+    semantic: { type: "vector", text: "Semantic (embedding) retrieval", capabilities: ["semantic"] },
+    hybrid: {
+      type: "hybrid",
+      text: "Hybrid BM25 + semantic retrieval (reciprocal rank fusion)",
+      capabilities: ["lexical", "semantic", "hybrid"]
+    }
+  };
 });
 
-// src/datasource/skills/mailcrawl/methods.ts
-class MailcrawlMethod {
+// src/datasource/skills/lazykatok/skill.ts
+class LazykatokSkill {
   client;
-  mode;
   instanceId;
+  instances;
+  pollingIntervalMs;
   tags;
-  account;
-  mailbox;
-  constructor(client, mode, instanceId, tags, account, mailbox) {
-    this.client = client;
-    this.mode = mode;
-    this.instanceId = instanceId;
-    this.tags = tags;
-    this.account = account;
-    this.mailbox = mailbox;
+  sidecar;
+  lastIndexedAt;
+  constructor(options) {
+    this.client = options.client;
+    this.instanceId = options.instanceId ?? DEFAULT_INSTANCE_ID4;
+    this.instances = options.instances !== undefined && options.instances.length > 0 ? options.instances : [this.instanceId];
+    this.pollingIntervalMs = options.pollingIntervalMs ?? DEFAULT_POLLING_INTERVAL_MS3;
+    this.tags = options.tags ?? DEFAULT_KAKAO_TAGS2;
+    this.lastIndexedAt = options.lastIndexedAt;
+    this.sidecar = options.sidecar === undefined ? undefined : new CrawlerSidecar({
+      ...options.sidecar,
+      channelIds: options.channelIds ?? options.sidecar.channelIds ?? [],
+      channelNames: options.channelNames ?? options.sidecar.channelNames ?? []
+    });
   }
   describe() {
     return {
-      name: `mailcrawl-${this.mode}`,
-      type: this.mode === "bm25" ? "bm25" : this.mode === "semantic" ? "vector" : "hybrid",
-      description: `${this.mode} retrieval over archived email via the external mailcrawl CLI`,
-      status: "active",
-      capabilities: ["email", this.mode, "scoped", "external-cli", "path-opaque-sources"],
-      datasourceId: "mailcrawl",
-      tags: this.tags
-    };
-  }
-  async retrieve(query, options) {
-    if (!query.trim())
-      return [];
-    const result = await this.client.search(this.mode, query, {
-      topK: options.topK,
-      signal: options.signal
-    });
-    if (!result.ok)
-      throw datasourceCliError("mailcrawl", `search --mode ${this.mode}`, result);
-    return result.hits.filter((hit) => {
-      if (this.account !== undefined && hit.accountId !== this.account)
-        return false;
-      if (this.mailbox !== undefined && hit.mailbox !== this.mailbox)
-        return false;
-      const source = datasourceSourcePath("mailcrawl", this.instanceId, hit.chunkId);
-      return matchesDatasourceScope(source, options.scope);
-    }).map((hit) => {
-      const source = datasourceSourcePath("mailcrawl", this.instanceId, hit.chunkId);
-      return {
-        id: `mailcrawl:${this.instanceId}:${hit.chunkId}`,
-        content: hit.snippet,
-        source,
-        score: hit.score,
-        metadata: {
-          method: `mailcrawl-${this.mode}`,
-          datasourceId: "mailcrawl",
-          instanceId: this.instanceId,
-          mode: this.mode,
-          chunkId: hit.chunkId,
-          messageId: hit.messageId,
-          threadId: hit.threadId,
-          accountId: hit.accountId,
-          mailbox: hit.mailbox,
-          subject: hit.subject,
-          from: hit.from,
-          to: hit.to,
-          date: hit.date
-        }
-      };
-    }).slice(0, options.topK ?? 20);
-  }
-}
-var modes, MAILCRAWL_MODES;
-var init_methods4 = __esm(() => {
-  init_errors4();
-  init_scope2();
-  modes = ["bm25", "semantic", "hybrid"];
-  MAILCRAWL_MODES = modes;
-});
-
-// src/datasource/skills/mailcrawl/skill.ts
-class MailcrawlSkill {
-  client;
-  instanceId;
-  instances;
-  interval;
-  tags;
-  account;
-  mailbox;
-  lastIndexedAt;
-  constructor(options = {}) {
-    const connectorOptions = options.connectorOptions ?? options;
-    this.instanceId = options.instanceId ?? "default";
-    this.instances = options.instances?.length ? options.instances : [this.instanceId];
-    validateMailcrawlInstanceId(this.instanceId);
-    for (const instance of this.instances)
-      validateMailcrawlInstanceId(instance);
-    this.interval = options.pollingIntervalMs ?? 15 * 60 * 1000;
-    this.tags = options.tags ?? TAGS;
-    this.account = connectorOptions.account;
-    this.mailbox = connectorOptions.mailbox;
-    this.client = options.client ?? new MailcrawlClient({ ...connectorOptions, instanceId: this.instanceId });
-  }
-  describe() {
-    return {
-      name: "mailcrawl",
-      id: "mailcrawl",
-      type: "mailcrawl-archive",
-      description: "Local email archive datasource via the external mailcrawl CLI",
-      capabilities: ["email", "external-cli", "polling", "incremental", "bm25", "semantic", "hybrid"],
-      tags: this.tags,
+      name: KAKAO_DATASOURCE_ID2,
+      id: KAKAO_DATASOURCE_ID2,
+      type: KAKAO_SKILL_TYPE,
+      description: "KakaoTalk datasource via the external lazykatok CLI",
+      capabilities: ["chat", "external-cli", "polling", "bm25", "semantic"],
+      tags: this.tags,
       status: "active",
       requiresExternalCli: true,
-      datasourceId: "mailcrawl",
+      datasourceId: KAKAO_DATASOURCE_ID2,
       instanceId: this.instanceId,
       instances: this.instances
     };
   }
   polling() {
     return {
-      mode: this.interval > 0 ? "poll" : "none",
-      ...this.interval > 0 ? { intervalMs: this.interval } : {},
+      mode: "poll",
+      intervalMs: this.pollingIntervalMs,
       lastIndexedAt: this.lastIndexedAt
     };
   }
   async index(options) {
     try {
-      const result = await this.client.sync(options?.signal);
-      if (!result.ok) {
-        return this.fail(result.reason === "binary-missing" || result.reason === "spawn-error" || result.reason === "timeout" ? "datasource-unavailable" : result.reason === "remote-embedding-rejected" ? "datasource-embedding-egress-rejected" : "datasource-index-failed", result.reason);
-      }
-      const indexed = await this.client.index(options?.signal);
-      const diagnostics = [];
-      if (!indexed.ok) {
-        diagnostics.push({
-          code: indexed.reason === "remote-embedding-rejected" ? "datasource-embedding-egress-rejected" : "datasource-index-failed",
-          severity: "warning",
-          message: `mailcrawl semantic index unavailable (${indexed.reason}); BM25 remains available`,
+      const doctor = await this.client.doctor(options?.signal);
+      if (!doctor.ok)
+        return this.fail("datasource-unavailable", doctor);
+      const sync = await this.client.sync(options?.signal);
+      if (!sync.ok)
+        return this.fail("datasource-index-failed", sync);
+      if (this.sidecar !== undefined) {
+        this.lastIndexedAt = Date.now();
+        return {
+          ok: true,
           instanceId: this.instanceId,
-          source: "mailcrawl"
-        });
+          skill: KAKAO_DATASOURCE_ID2,
+          chunkCount: syncChunkCount(sync),
+          indexedAt: this.lastIndexedAt,
+          diagnostics: await this.syncSidecar()
+        };
       }
+      const indexResult = await this.client.index(options?.signal);
+      if (!indexResult.ok)
+        return this.fail("datasource-index-failed", indexResult);
       this.lastIndexedAt = Date.now();
+      const chunkCount = "data" in indexResult && typeof indexResult.data === "object" && indexResult.data !== null && "chunkCount" in indexResult.data && typeof indexResult.data.chunkCount === "number" ? indexResult.data.chunkCount : 0;
       return {
         ok: true,
         instanceId: this.instanceId,
-        skill: "mailcrawl",
-        chunkCount: result.data.chunksAdded ?? result.data.messages,
+        skill: KAKAO_DATASOURCE_ID2,
+        chunkCount,
         indexedAt: this.lastIndexedAt,
-        diagnostics
+        diagnostics: []
       };
-    } catch {
-      return this.fail("datasource-unavailable", "mailcrawl sync failed unexpectedly");
+    } catch (error) {
+      return this.fail("datasource-unavailable", {
+        ok: false,
+        reason: "spawn-error",
+        stdout: "",
+        stderr: describeRetrievalError(error),
+        code: null
+      });
     }
   }
+  close() {
+    this.sidecar?.close();
+  }
   retrievalMethods() {
+    const cliBm25 = new LazykatokBm25Method({ client: this.client, instanceId: this.instanceId, tags: this.tags });
+    const { sidecar } = this;
+    if (sidecar === undefined) {
+      return [
+        cliBm25,
+        new LazykatokSemanticMethod({ client: this.client, instanceId: this.instanceId, tags: this.tags })
+      ];
+    }
+    const shared = {
+      sidecar,
+      descriptor: sidecar.descriptor,
+      datasourceId: KAKAO_DATASOURCE_ID2,
+      skillType: KAKAO_SKILL_TYPE,
+      instanceId: this.instanceId,
+      tags: this.tags
+    };
+    const bm25 = new SidecarRetrievalMethod({ ...shared, mode: "bm25" });
+    if (!sidecar.semanticEnabled)
+      return [cliBm25, bm25];
     return [
-      new MailcrawlMethod(this.client, "bm25", this.instanceId, this.tags, this.account, this.mailbox),
-      new MailcrawlMethod(this.client, "semantic", this.instanceId, this.tags, this.account, this.mailbox),
-      new MailcrawlMethod(this.client, "hybrid", this.instanceId, this.tags, this.account, this.mailbox)
+      cliBm25,
+      bm25,
+      new SidecarRetrievalMethod({ ...shared, mode: "semantic" }),
+      new SidecarRetrievalMethod({ ...shared, mode: "hybrid" })
     ];
   }
+  async syncSidecar() {
+    if (this.sidecar === undefined || !this.sidecar.semanticEnabled)
+      return [];
+    const base = { instanceId: this.instanceId, source: KAKAO_DATASOURCE_ID2 };
+    let result;
+    try {
+      result = await this.sidecar.sync();
+    } catch (error) {
+      result = { ok: false, message: error instanceof Error ? error.message : String(error) };
+    }
+    if (!result.ok) {
+      return [
+        {
+          code: "datasource-index-failed",
+          severity: "warning",
+          message: `semantic-unavailable: ${result.message}`.slice(0, 300),
+          ...base
+        }
+      ];
+    }
+    const diagnostics = [];
+    if (result.skipped > 0) {
+      diagnostics.push({
+        code: "datasource-index-partial",
+        severity: "warning",
+        message: `semantic index skipped ${result.skipped} message(s) the embedder rejected; they stay searchable by BM25 and are retried when edited`,
+        ...base
+      });
+    }
+    if (result.remaining > 0) {
+      diagnostics.push({
+        code: "datasource-index-partial",
+        severity: "info",
+        message: `semantic index catching up: ${result.embedded} embedded this refresh, ${result.remaining} remaining; hybrid search uses BM25 for the rest until a later refresh`,
+        ...base
+      });
+    }
+    return diagnostics;
+  }
   describeSources() {
     return this.instances.map((instanceId) => ({
-      source: datasourceSourcePath("mailcrawl", instanceId),
-      datasourceId: "mailcrawl",
-      skill: "mailcrawl",
+      source: datasourceSourcePath(KAKAO_DATASOURCE_ID2, instanceId),
+      datasourceId: KAKAO_DATASOURCE_ID2,
+      skill: KAKAO_DATASOURCE_ID2,
       instanceId,
-      contentType: "email",
-      metadata: { datasourceId: "mailcrawl", instanceId, tags: this.tags }
+      contentType: "chat",
+      metadata: {
+        datasourceId: KAKAO_DATASOURCE_ID2,
+        instanceId,
+        tags: this.tags
+      }
     }));
   }
   skillManifest() {
-    const scopes = this.instances.map((id) => `- \`${datasourceSourcePath("mailcrawl", id)}\``).join(`
+    const instanceSources = this.instances.map((instanceId) => `- \`${datasourceSourcePath(KAKAO_DATASOURCE_ID2, instanceId)}\``).join(`
 `);
+    const cadence = this.pollingIntervalMs > 0 ? `roughly every ${Math.round(this.pollingIntervalMs / 60000)} minute(s) when auto-refresh runs` : "on manual refresh only";
     return {
-      name: "datasource-mailcrawl",
-      description: "Search local email synchronized and indexed by mailcrawl.",
-      content: `# mailcrawl email datasource
-
-Use this skill for authorized local email retrieval. AutoRAG invokes the external \`mailcrawl\` CLI and never opens its SQLite archive directly.
-
-## Indexing
-Indexing is server-managed and incremental; do not trigger sync yourself.
-
-## Search
-Call the dedicated \`${datasourceSearchToolName("mailcrawl")}\` tool with a natural-language query. Available scopes:
-${scopes}
-
-The selectable retrieval modes are BM25, semantic, and hybrid. Scope arguments can only narrow trusted access. Source identifiers remain opaque and hierarchical.
-
-## Native CLI
-The external \`mailcrawl\` CLI owns the archive and index. When the dedicated tool cannot express what you need, call mailcrawl directly through \`bash\`:
-- \`mailcrawl search --mode hybrid --limit 20 --json "<query>"\` (\`bm25\` and \`semantic\` modes also available; \`--account\` / \`--mailbox\` narrow the archive)
-Never pass datasource virtual paths (\`/mailcrawl/...\`) to bash; they are not OS paths.`
+      name: `datasource-${KAKAO_DATASOURCE_ID2}`,
+      description: "Search indexed KakaoTalk chat history (messages, senders, room context). Use for questions about KakaoTalk conversations, decisions made in chats, or who said what.",
+      content: [
+        `# KakaoTalk datasource (${KAKAO_SKILL_TYPE})`,
+        "",
+        "This skill searches KakaoTalk chats that are indexed through the external `lazykatok` CLI. AutoRAG never reads KakaoTalk databases directly.",
+        "",
+        "## When to use",
+        "Use this skill when the question is about KakaoTalk conversations, chat participants, or content shared inside chats.",
+        "",
+        "## Indexing",
+        `Indexing is server-managed and refreshed ${cadence}. You do not trigger indexing; just search.`,
+        "",
+        "## How to search",
+        `Call the dedicated \`${datasourceSearchToolName(KAKAO_DATASOURCE_ID2)}\` tool with a natural-language \`query\` and \`topK\`. This datasource does not support per-source scope narrowing. Configured datasource instances:`,
+        instanceSources.length > 0 ? instanceSources : "- (no configured instances)",
+        "",
+        "Chat/channel filtering is owned by lazykatok.",
+        "",
+        "## Native CLI",
+        "The external `lazykatok` CLI owns the archive, index, and credentials. When the dedicated tool cannot express what you need, call lazykatok directly through `bash`:",
+        '- `lazykatok search bm25 "<query>" --json --limit 20` — lexical search (`semantic` mode also available)',
+        "- `lazykatok chunk get <chunkId> --json` — fetch one chunk by id (chunk ids appear in result metadata)",
+        "- `lazykatok chunk context <chunkId> --json` — surrounding messages of a chunk",
+        "Never pass datasource virtual paths (`/kakao/...`) to bash; they are not OS paths.",
+        "",
+        "## Output rules",
+        "Datasource source identifiers such as `/kakao/<instance>/chunks/<id>` are internal and opaque. Never put them, real file paths, account IDs, or phone numbers in the visible answer."
+      ].join(`
+`)
     };
   }
-  fail(code, message) {
+  fail(code, result) {
+    const base = result.stderr.length > 0 ? `${result.reason}: ${result.stderr.trim().slice(0, 4000)}` : result.reason;
+    const message = code === "datasource-index-failed" && result.reason === "timeout" ? `${base} — ${connectorSyncTimeoutHint()}` : base;
+    const diagnostic = {
+      code,
+      severity: code === "datasource-unavailable" ? "warning" : "error",
+      message,
+      instanceId: this.instanceId,
+      source: KAKAO_DATASOURCE_ID2
+    };
     return {
       ok: false,
       instanceId: this.instanceId,
-      skill: "mailcrawl",
+      skill: KAKAO_DATASOURCE_ID2,
       indexedAt: Date.now(),
-      diagnostics: [
-        {
-          code,
-          severity: code === "datasource-unavailable" ? "warning" : "error",
-          message: `mailcrawl ${message}`,
-          instanceId: this.instanceId,
-          source: "mailcrawl"
-        }
-      ],
-      error: message,
+      diagnostics: [diagnostic],
+      error: result.reason,
       code,
       message
     };
   }
 }
-var TAGS;
-var init_skill9 = __esm(() => {
+function syncChunkCount(sync) {
+  if (!("data" in sync) || typeof sync.data !== "object" || sync.data === null)
+    return 0;
+  const meta = "metadata" in sync.data && typeof sync.data.metadata === "object" && sync.data.metadata !== null ? sync.data.metadata : undefined;
+  if (typeof meta?.chunks === "number")
+    return meta.chunks;
+  if ("messageCount" in sync.data && typeof sync.data.messageCount === "number")
+    return sync.data.messageCount;
+  return 0;
+}
+var KAKAO_DATASOURCE_ID2 = "kakao", KAKAO_SKILL_TYPE = "kakaotalk", DEFAULT_INSTANCE_ID4 = "default", DEFAULT_POLLING_INTERVAL_MS3, DEFAULT_KAKAO_TAGS2;
+var init_skill7 = __esm(() => {
+  init_skip();
   init_scope2();
-  init_client9();
-  init_config3();
+  init_sidecar();
   init_methods4();
-  TAGS = ["mailcrawl", "email", "pii"];
+  init_methods3();
+  DEFAULT_POLLING_INTERVAL_MS3 = 15 * 60 * 1000;
+  DEFAULT_KAKAO_TAGS2 = ["kakaotalk", "personal", "pii"];
 });
 
-// src/datasource/skills/mailcrawl/index.ts
-var init_mailcrawl = __esm(() => {
-  init_client9();
-  init_methods4();
-  init_skill9();
+// src/datasource/skills/lazykatok/index.ts
+var init_lazykatok = __esm(() => {
+  init_client8();
+  init_methods3();
+  init_skill7();
+  init_types5();
 });
 
-// src/datasource/sidecar/fusion.ts
-function encodeVector(vector) {
-  const floats = Float32Array.from(vector);
-  return new Uint8Array(floats.buffer, floats.byteOffset, floats.byteLength);
-}
-function decodeVector(blob) {
-  if (blob.byteLength % Float32Array.BYTES_PER_ELEMENT !== 0)
-    return;
-  return new Float32Array(new Uint8Array(blob).buffer);
-}
-function cosine3(a, b) {
-  if (a.length === 0 || a.length !== b.length)
-    return 0;
-  let dot = 0;
-  let normA = 0;
-  let normB = 0;
-  for (let index = 0;index < a.length; index += 1) {
-    const x = a[index] ?? 0;
-    const y = b[index] ?? 0;
-    dot += x * y;
-    normA += x * x;
-    normB += y * y;
+// src/datasource/skills/mail-export/connector.ts
+import { readdir as readdir4, readFile as readFile5, stat as stat6 } from "node:fs/promises";
+import { basename as basename12, dirname as dirname25, extname as extname6, join as join37 } from "node:path";
+import { simpleParser as simpleParser2 } from "mailparser";
+
+class MailExportConnector {
+  options;
+  constructor(options = {}) {
+    this.options = options;
   }
-  if (normA === 0 || normB === 0)
-    return 0;
-  return dot / Math.sqrt(normA * normB);
-}
-function topKByCosine(query, rows, k) {
-  if (k <= 0)
-    return [];
-  const scored = [];
-  for (const row of rows) {
-    const score = cosine3(query, row.vector);
-    if (score > 0)
-      scored.push({ key: row.key, score });
+  async fetch() {
+    const paths = this.options.paths ?? [];
+    if (paths.length === 0)
+      return { ok: false, reason: "not-configured", message: "no mail export paths configured" };
+    const maxDocuments = this.options.maxDocuments ?? DEFAULT_MAX_DOCUMENTS3;
+    const maxBytes = this.options.maxBytesPerFile ?? DEFAULT_MAX_BYTES_PER_FILE2;
+    const files = [];
+    let missingPaths = 0;
+    for (const path of paths) {
+      try {
+        const info = await stat6(path);
+        if (info.isDirectory()) {
+          const entries = await readdir4(path, { recursive: true, withFileTypes: true });
+          for (const entry of entries) {
+            if (!entry.isFile())
+              continue;
+            const extension = extname6(entry.name).toLowerCase();
+            if (extension === ".eml" || extension === ".mbox") {
+              files.push(join37(entry.parentPath, entry.name));
+            }
+          }
+        } else if (info.isFile()) {
+          files.push(path);
+        }
+      } catch {
+        missingPaths += 1;
+      }
+    }
+    if (missingPaths === paths.length) {
+      return { ok: false, reason: "unavailable", message: "no configured mail export paths were readable" };
+    }
+    const documents = [];
+    const warnings = [];
+    if (missingPaths > 0)
+      warnings.push(`${missingPaths} configured path(s) were unreadable`);
+    let parseFailures = 0;
+    let skippedLarge = 0;
+    for (const file of files.sort()) {
+      if (documents.length >= maxDocuments)
+        break;
+      let bytes;
+      try {
+        const info = await stat6(file);
+        if (info.size > maxBytes) {
+          skippedLarge += 1;
+          continue;
+        }
+        bytes = await readFile5(file);
+      } catch {
+        parseFailures += 1;
+        continue;
+      }
+      const stem = basename12(file, extname6(file));
+      const mailboxDir = basename12(dirname25(file));
+      const hierarchy = ["mailboxes", mailboxDir, stem];
+      const extension = extname6(file).toLowerCase();
+      const rawMessages = extension === ".mbox" ? splitMbox(bytes.toString("utf8")) : [bytes.toString("utf8")];
+      for (const [index, raw] of rawMessages.entries()) {
+        if (documents.length >= maxDocuments)
+          break;
+        try {
+          const parsed = await simpleParser2(raw);
+          const subject = parsed.subject ?? "(no subject)";
+          const from = Array.isArray(parsed.from) ? parsed.from[0]?.text : parsed.from?.text;
+          const to = Array.isArray(parsed.to) ? parsed.to[0]?.text : parsed.to?.text;
+          const body = (parsed.text ?? "").trim().slice(0, MAX_BODY_CHARS);
+          const headerBlock = [
+            `Subject: ${subject}`,
+            ...from !== undefined ? [`From: ${from}`] : [],
+            ...to !== undefined ? [`To: ${to}`] : [],
+            ...parsed.date !== undefined ? [`Date: ${parsed.date.toISOString()}`] : []
+          ].join(`
+`);
+          const content = `${headerBlock}
+
+${body}`.trim();
+          if (content.length === 0)
+            continue;
+          documents.push({
+            docId: extension === ".mbox" ? `${stem}-${index + 1}` : stem,
+            hierarchy,
+            title: subject,
+            content,
+            ...parsed.date !== undefined ? { publishedAt: parsed.date.getTime() } : {},
+            metadata: {
+              path: file,
+              subject,
+              ...from !== undefined ? { from } : {},
+              messageIndex: index
+            }
+          });
+        } catch {
+          parseFailures += 1;
+        }
+      }
+    }
+    if (skippedLarge > 0)
+      warnings.push(`${skippedLarge} file(s) exceeded the size limit and were skipped`);
+    if (parseFailures > 0)
+      warnings.push(`${parseFailures} message(s) failed to parse`);
+    return { ok: true, documents, ...warnings.length > 0 ? { warnings } : {} };
   }
-  scored.sort((a, b) => b.score - a.score || a.key.localeCompare(b.key));
-  return scored.slice(0, k);
 }
-function reciprocalRankFusion(lists) {
-  const scores = new Map;
-  for (const list of lists) {
-    const seen = new Set;
-    for (const [index, key] of list.entries()) {
-      if (seen.has(key))
-        continue;
-      seen.add(key);
-      scores.set(key, (scores.get(key) ?? 0) + 1 / (RRF_K2 + index + 1));
+function splitMbox(text) {
+  const lines = text.split(/\r?\n/u);
+  const messages = [];
+  let current = [];
+  for (const line of lines) {
+    if (MBOX_SEPARATOR.test(line) && line.startsWith("From ")) {
+      if (current.length > 0)
+        messages.push(current.join(`
+`));
+      current = [];
+      continue;
     }
+    current.push(line);
   }
-  return [...scores.entries()].map(([key, score]) => ({ key, score })).sort((a, b) => b.score - a.score);
+  if (current.length > 0)
+    messages.push(current.join(`
+`));
+  return messages.filter((message) => message.trim().length > 0);
 }
-var RRF_K2 = 60;
-
-// src/datasource/sidecar/index.ts
-import { createHash as createHash14 } from "node:crypto";
-import { mkdirSync as mkdirSync23, statSync as statSync3 } from "node:fs";
-import { dirname as dirname25 } from "node:path";
-import { DatabaseSync as DatabaseSync2 } from "node:sqlite";
-import { pathToFileURL as pathToFileURL2 } from "node:url";
+var DEFAULT_MAX_DOCUMENTS3 = 500, DEFAULT_MAX_BYTES_PER_FILE2, MAX_BODY_CHARS = 50000, MBOX_SEPARATOR;
+var init_connector4 = __esm(() => {
+  DEFAULT_MAX_BYTES_PER_FILE2 = 50 * 1024 * 1024;
+  MBOX_SEPARATOR = /^From .*\d{4}$/mu;
+});
 
-class CrawlerSidecar {
-  descriptor;
-  options;
-  select;
-  where;
-  opened;
-  attached;
-  lastSemantic;
-  constructor(options) {
-    this.options = options;
-    const { descriptor } = options;
-    this.descriptor = descriptor;
-    const metadataColumns = descriptor.metadata.map((column, index) => `${column.sql} as m${index}`);
-    this.select = [
-      `${descriptor.keyExpression} as k`,
-      `${descriptor.textExpression} as content`,
-      `${descriptor.titleExpression ?? "null"} as title`,
-      `${descriptor.epochSecondsExpression ?? "null"} as epoch`,
-      ...metadataColumns
-    ].join(", ");
-    const scope = descriptor.scopeExpression !== undefined && options.workspaceId !== undefined;
-    const allowlist = [
-      ...(options.channelIds?.length ?? 0) > 0 ? [`${descriptor.channelIdExpression} in (${options.channelIds?.map(() => "?").join(", ")})`] : [],
-      ...(options.channelNames?.length ?? 0) > 0 ? [`${descriptor.channelNameExpression} in (${options.channelNames?.map(() => "?").join(", ")})`] : []
-    ];
-    this.where = [
-      descriptor.liveWhere,
-      ...scope ? [`${descriptor.scopeExpression} = ?`] : [],
-      ...allowlist.length > 0 ? [`(${allowlist.join(" or ")})`] : []
-    ].join(" and ");
-  }
-  get db() {
-    if (this.opened !== undefined)
-      return this.opened;
-    mkdirSync23(dirname25(this.options.sidecarDbPath), { recursive: true });
-    const db = new DatabaseSync2(this.options.sidecarDbPath);
-    db.exec(`pragma busy_timeout = ${BUSY_TIMEOUT_MS}`);
-    db.exec("pragma journal_mode = wal");
-    db.exec(`
-			create table if not exists sidecar_meta (k text primary key, v text not null);
-			create table if not exists sidecar_vec (message_key text primary key, hash text not null, vec blob not null);
-		`);
-    this.opened = db;
-    return db;
-  }
-  attachCrawler() {
-    let file;
-    try {
-      const { dev, ino } = statSync3(this.options.crawlerDbPath);
-      file = { dev, ino };
-    } catch {
-      this.detachCrawler();
-      return false;
-    }
-    if (this.attached !== undefined) {
-      if (this.attached.dev === file.dev && this.attached.ino === file.ino)
-        return true;
-      this.detachCrawler();
-    }
-    const crawlerUri = `${pathToFileURL2(this.options.crawlerDbPath).href}?mode=ro`.replaceAll("'", "''");
-    try {
-      this.db.exec(`attach database '${crawlerUri}' as crawler`);
-    } catch {
-      return false;
-    }
-    const { descriptor } = this.options;
-    try {
-      this.db.prepare(`select ${this.select} from ${descriptor.fromClause} where ${this.where} limit 0`).all(...this.whereParams());
-      this.db.prepare(`select 1 from ${descriptor.fts.table} limit 0`).all();
-    } catch {
-      this.db.exec("detach database crawler");
-      return false;
+// src/datasource/skills/mail-export/skill.ts
+var MAIL_EXPORT_SKILL_DEFINITION, MailExportSkill;
+var init_skill8 = __esm(() => {
+  init_connector_skill();
+  init_connector4();
+  MAIL_EXPORT_SKILL_DEFINITION = {
+    skillName: "mail-export",
+    skillType: "mail-export",
+    description: "Local mail export datasource",
+    capabilities: ["email", "filesystem"],
+    defaultTags: ["mail-export", "email", "pii"],
+    contentType: "email",
+    manifestDescription: "Search indexed local mail exports (mbox/eml archives). Use for questions about archived email conversations.",
+    manifestNotes: [
+      "Only the server-configured export paths are indexed; archives are read locally and never uploaded."
+    ]
+  };
+  MailExportSkill = class MailExportSkill extends ConnectorDatasourceSkill {
+    constructor(options = {}) {
+      const { connector, connectorOptions, ...rest } = options;
+      super(MAIL_EXPORT_SKILL_DEFINITION, {
+        ...rest,
+        connector: connector ?? new MailExportConnector(connectorOptions ?? {})
+      });
     }
-    this.attached = file;
-    this.lastSemantic = undefined;
-    return true;
-  }
-  detachCrawler() {
-    if (this.attached === undefined)
-      return;
-    this.attached = undefined;
-    this.lastSemantic = undefined;
-    try {
-      this.opened?.exec("detach database crawler");
-    } catch {}
-  }
-  get semanticEnabled() {
-    return this.options.embedder !== undefined;
+  };
+});
+
+// src/datasource/skills/mail-export/index.ts
+var init_mail_export = __esm(() => {
+  init_connector4();
+  init_skill8();
+});
+
+// src/datasource/skills/mailcrawl/config.ts
+import { chmodSync, mkdirSync as mkdirSync23 } from "node:fs";
+function validateMailcrawlInstanceId(instanceId) {
+  if (!/^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$/u.test(instanceId)) {
+    throw new Error("mailcrawl instanceId must be a safe single path segment");
   }
-  close() {
-    this.opened?.close();
-    this.opened = undefined;
-    this.attached = undefined;
-    this.lastSemantic = undefined;
+}
+function ensurePrivateMailcrawlDataDir(path) {
+  mkdirSync23(path, { recursive: true, mode: 448 });
+  if (process.platform !== "win32")
+    chmodSync(path, 448);
+}
+var init_config3 = () => {};
+
+// src/datasource/skills/mailcrawl/client.ts
+import { spawn as spawn15 } from "node:child_process";
+
+class MailcrawlClient {
+  options;
+  constructor(options = {}) {
+    this.options = options;
+    if (options.instanceId !== undefined)
+      validateMailcrawlInstanceId(options.instanceId);
   }
-  whereParams() {
-    const { descriptor, workspaceId, channelIds = [], channelNames = [] } = this.options;
-    return [
-      ...descriptor.scopeExpression !== undefined && workspaceId !== undefined ? [workspaceId] : [],
-      ...channelIds,
-      ...channelNames
+  async sync(signal) {
+    const args = [
+      "sync",
+      "--json",
+      ...this.options.account === undefined ? [] : ["--account", this.options.account],
+      ...this.options.mailbox === undefined ? [] : ["--mailbox", this.options.mailbox],
+      ...this.options.backend === undefined ? [] : ["--backend", this.options.backend],
+      ...this.options.source === undefined ? [] : ["--source", this.options.source],
+      ...this.options.fixture === undefined ? [] : ["--fixture", this.options.fixture],
+      ...this.options.himalayaConfig === undefined ? [] : ["--himalaya-config", this.options.himalayaConfig]
     ];
+    const result = await this.run(args, signal, this.indexTimeoutMs());
+    if (!result.ok)
+      return failure3(result);
+    const parsed = parseObject2(result.stdout);
+    if (parsed === undefined || !hasAnyNumber(parsed, ["added", "updated", "deleted", "unchanged", "chunksAdded"])) {
+      return failure3({ ...result, ok: false, reason: "invalid-output" });
+    }
+    const messages = number(parsed, "added") ?? number(parsed, "updated") ?? number(parsed, "unchanged") ?? 0;
+    return ok4({
+      messages,
+      added: number(parsed, "added"),
+      updated: number(parsed, "updated"),
+      deleted: number(parsed, "deleted"),
+      unchanged: number(parsed, "unchanged"),
+      chunksAdded: number(parsed, "chunksAdded"),
+      archiveRevision: string(parsed, "archiveRevision")
+    }, result);
   }
-  searchLexical(query, topK) {
-    const match = toMatchQuery(query);
-    if (match === undefined || topK <= 0 || !this.attachCrawler())
-      return [];
-    const { descriptor } = this.options;
-    const rows = this.db.prepare(`select ${this.select}, bm25(${descriptor.fts.table}) as rank
-				from ${descriptor.fromClause} join ${descriptor.fts.table} on ${descriptor.fts.joinOn}
-				where ${descriptor.fts.table} match ? and ${this.where}
-				order by rank limit ?`).all(match, ...this.whereParams(), topK);
-    return rows.map((row) => this.toHit(this.toRow(row), -Number(row.rank)));
-  }
-  async searchSemantic(query, topK) {
-    const trimmed = query.trim();
-    if (this.options.embedder === undefined || trimmed.length === 0 || topK <= 0 || !this.attachCrawler())
-      return [];
-    const ranked = await this.rankSemantic(trimmed, topK * POOL_FACTOR);
-    if (ranked.length === 0)
-      return [];
-    const live = this.hydrate(ranked.map((entry) => entry.key));
-    const hits = [];
-    for (const entry of ranked) {
-      const row = live.get(entry.key);
-      if (row !== undefined)
-        hits.push(this.toHit(row, entry.score));
-      if (hits.length >= topK)
-        break;
+  async search(mode, query, options) {
+    if (query.trim().length === 0)
+      return { ok: true, hits: [], stdout: "", stderr: "", code: 0 };
+    const args = [
+      "search",
+      "--mode",
+      mode === "keyword" ? "bm25" : mode,
+      ...this.options.account === undefined ? [] : ["--account", this.options.account],
+      ...this.options.mailbox === undefined ? [] : ["--mailbox", this.options.mailbox],
+      "--limit",
+      String(options?.topK ?? 20),
+      "--json",
+      query.trim()
+    ];
+    const result = await this.run(args, options?.signal, this.options.timeoutMs ?? DEFAULT_SEARCH_TIMEOUT_MS2);
+    if (!result.ok)
+      return failure3(result);
+    const parsed = parseJson3(result.stdout);
+    if (parsed === undefined)
+      return failure3({ ...result, ok: false, reason: "invalid-output" });
+    const rows = Array.isArray(parsed) ? parsed : objectArray(parsed, ["hits", "results", "items"]);
+    if (rows === undefined)
+      return failure3({ ...result, ok: false, reason: "invalid-output" });
+    const hits = rows.map((row, index) => normalizeHit2(row, mode, index));
+    if (hits.some((hit) => hit === undefined)) {
+      return failure3({ ...result, ok: false, reason: "invalid-output" });
     }
-    return hits;
+    return {
+      ok: true,
+      hits,
+      stdout: result.stdout,
+      stderr: result.stderr,
+      code: result.code ?? 0
+    };
   }
-  async rankSemantic(query, pool) {
-    const now = Date.now();
-    let cached = this.lastSemantic;
-    if (cached === undefined || cached.query !== query || now - cached.at > SEMANTIC_REUSE_MS) {
-      const embedder = this.options.embedder;
-      if (embedder === undefined)
-        return [];
-      cached = {
-        query,
-        at: now,
-        vector: embedder.embed([query]).then((vectors) => vectors[0], () => {
-          return;
-        }),
-        ranked: undefined
-      };
-      this.lastSemantic = cached;
+  indexTimeoutMs() {
+    return this.options.indexTimeoutMs ?? this.options.timeoutMs ?? DEFAULT_INDEX_TIMEOUT_MS3;
+  }
+  async run(args, signal, timeoutMs) {
+    const violation = findRemoteEmbeddingViolation({ ...process.env, ...this.options.env ?? {} });
+    if (violation !== undefined) {
+      return { ok: false, reason: "remote-embedding-rejected", stdout: "", stderr: "", code: null };
     }
-    const vector = await cached.vector;
-    if (vector === undefined) {
-      if (this.lastSemantic === cached)
-        this.lastSemantic = undefined;
-      return [];
+    const dataDir = this.options.dataDir;
+    if (dataDir !== undefined)
+      ensurePrivateMailcrawlDataDir(dataDir);
+    const env = {};
+    for (const [key, value] of Object.entries(process.env)) {
+      if (value !== undefined && (SAFE_ENV_KEYS2.has(key) || SAFE_MAILCRAWL_ENV_KEYS.has(key)))
+        env[key] = value;
     }
-    if (cached.ranked === undefined || cached.ranked.pool < pool) {
-      const wide = pool * POOL_FACTOR;
-      cached.ranked = { pool: wide, keys: topKByCosine(vector, this.storedVectors(), wide) };
+    for (const [key, value] of Object.entries(this.options.env ?? {})) {
+      if (value === undefined)
+        delete env[key];
+      else if (SAFE_ENV_KEYS2.has(key) || SAFE_MAILCRAWL_ENV_KEYS.has(key))
+        env[key] = value;
     }
-    return cached.ranked.keys.slice(0, pool);
+    if (dataDir !== undefined)
+      env.MAILCRAWL_DATA_DIR = dataDir;
+    return spawnMailcrawl(this.options.binaryPath ?? DEFAULT_BINARY4, args, env, this.options, signal, timeoutMs);
   }
-  async searchHybrid(query, topK) {
-    if (topK <= 0)
-      return [];
-    const pool = topK * POOL_FACTOR;
-    const lexical = this.searchLexical(query, pool);
-    const semantic = await this.searchSemantic(query, pool);
-    if (semantic.length === 0)
-      return lexical.slice(0, topK);
-    const byKey = new Map;
-    for (const hit of [...semantic, ...lexical])
-      byKey.set(hit.key, hit);
-    const fused = [];
-    for (const entry of reciprocalRankFusion([lexical.map((hit) => hit.key), semantic.map((hit) => hit.key)])) {
-      const hit = byKey.get(entry.key);
-      if (hit !== undefined)
-        fused.push({ ...hit, score: entry.score });
-      if (fused.length >= topK)
-        break;
+  async index(signal) {
+    const result = await this.run(["index", "--json"], signal, this.indexTimeoutMs());
+    if (!result.ok)
+      return failure3(result);
+    const parsed = parseObject2(result.stdout);
+    if (parsed === undefined || number(parsed, "embedded") === undefined && number(parsed, "reused") === undefined && string(parsed, "generation") === undefined) {
+      return failure3({ ...result, ok: false, reason: "invalid-output" });
     }
-    return fused;
+    return ok4({
+      embedded: number(parsed, "embedded"),
+      reused: number(parsed, "reused"),
+      generation: string(parsed, "generation"),
+      archiveRevision: string(parsed, "archiveRevision"),
+      rebuilt: typeof parsed.rebuilt === "boolean" ? parsed.rebuilt : undefined,
+      embedder: string(parsed, "embedder")
+    }, result);
   }
-  async sync() {
-    const { embedder } = this.options;
-    if (embedder === undefined)
-      return { ok: false, message: "no embedder configured" };
-    if (!this.attachCrawler())
-      return { ok: false, message: "crawler archive not available yet" };
-    let identity;
-    try {
-      identity = await embedder.identity();
-    } catch (error) {
-      return { ok: false, message: errorMessage2(error) };
-    }
-    this.adoptIdentity(identity);
-    const pruned = this.prune();
-    const pending = this.pendingKeys();
-    const batchSize = Math.max(1, this.options.batchSize ?? DEFAULT_BATCH_SIZE);
-    const budget = Math.max(1, this.options.maxEmbedPerSync ?? DEFAULT_MAX_EMBED_PER_SYNC);
-    const target = Math.min(pending.length, budget);
-    const embed = async (rows) => {
-      const vectors = await embedder.embed(rows.map((row) => row.input));
-      if (vectors.length !== rows.length) {
-        throw new Error(`embedder returned ${vectors.length} vectors for ${rows.length} messages`);
-      }
-      const wrong = vectors.find((vector) => vector.length !== identity.dimension);
-      if (wrong !== undefined) {
-        throw new Error(`embedder returned a ${wrong.length}-dimension vector, expected ${identity.dimension}`);
-      }
-      return vectors;
+}
+function spawnMailcrawl(binary, args, env, options, signal, timeoutMs) {
+  return new Promise((resolveResult) => {
+    const portable = portableSpawnCommand(binary, args);
+    const child = spawn15(portable.command, portable.args, {
+      env,
+      detached: process.platform !== "win32",
+      stdio: ["ignore", "pipe", "pipe"]
+    });
+    let stdout = { text: "", bytes: 0, exceeded: false }, stderr = { text: "", bytes: 0, exceeded: false }, settled = false, reason;
+    const max = options.maxBufferBytes ?? DEFAULT_MAX_BUFFER_BYTES4;
+    const timer = setTimeout(() => {
+      reason = "timeout";
+      terminate4(child);
+    }, timeoutMs);
+    const abort = () => {
+      reason = "aborted";
+      terminate4(child);
     };
-    let embedded = 0;
-    let skipped = 0;
-    for (let offset = 0;offset < target; offset += batchSize) {
-      const rows = [...this.hydrate(pending.slice(offset, Math.min(offset + batchSize, target))).values()];
-      if (rows.length === 0)
-        continue;
-      try {
-        this.store(rows, await embed(rows));
-        embedded += rows.length;
-      } catch (batchError) {
-        const rejected = [];
-        let lastError = batchError;
-        for (const row of rows) {
-          try {
-            this.store([row], await embed([row]));
-            embedded += 1;
-          } catch (error) {
-            rejected.push(row);
-            lastError = error;
-          }
-        }
-        if (rejected.length === 0)
-          continue;
-        if (embedded === 0)
-          return { ok: false, message: errorMessage2(lastError) };
-        this.store(rejected, undefined);
-        skipped += rejected.length;
+    if (signal?.aborted)
+      abort();
+    signal?.addEventListener("abort", abort, { once: true });
+    child.stdout.setEncoding("utf8");
+    child.stderr.setEncoding("utf8");
+    child.stdout.on("data", (chunk) => {
+      stdout = appendOutput(stdout, chunk, max);
+      if (stdout.exceeded && reason === undefined) {
+        reason = "stdout-too-large";
+        terminate4(child);
       }
-    }
-    return { ok: true, embedded, skipped, pruned, remaining: pending.length - embedded - skipped };
-  }
-  adoptIdentity(identity) {
-    const serialized = JSON.stringify(identity);
-    const current = this.db.prepare("select v from sidecar_meta where k = ?").get(IDENTITY_KEY);
-    if (current?.v === serialized)
-      return;
-    this.transaction(() => {
-      this.db.exec("delete from sidecar_vec");
-      this.db.prepare("insert or replace into sidecar_meta (k, v) values (?, ?)").run(IDENTITY_KEY, serialized);
     });
-    this.lastSemantic = undefined;
-  }
-  prune() {
-    const { descriptor } = this.options;
-    const result = this.db.prepare(`delete from sidecar_vec where message_key not in (
-					select ${descriptor.keyExpression} from ${descriptor.fromClause} where ${this.where}
-				)`).run(...this.whereParams());
-    if (Number(result.changes) > 0)
-      this.lastSemantic = undefined;
-    return Number(result.changes);
-  }
-  pendingKeys() {
-    const { descriptor } = this.options;
-    const rows = this.db.prepare(`select ${this.select}, v.hash as stored_hash
-				from ${descriptor.fromClause}
-				left join sidecar_vec v on v.message_key = ${descriptor.keyExpression}
-				where ${this.where}`).iterate(...this.whereParams());
-    const pending = [];
-    for (const row of rows) {
-      if (row.stored_hash !== this.toRow(row).hash)
-        pending.push(String(row.k));
-    }
-    return pending;
-  }
-  hydrate(keys) {
-    const { descriptor } = this.options;
-    const out = new Map;
-    if (keys.length === 0)
-      return out;
-    const rows = this.db.prepare(`select ${this.select} from ${descriptor.fromClause}
-				where ${descriptor.keyExpression} in (${keys.map(() => "?").join(", ")}) and ${this.where}`).all(...keys, ...this.whereParams());
-    const byKey = new Map(rows.map((row) => [String(row.k), this.toRow(row)]));
-    for (const key of keys) {
-      const row = byKey.get(key);
-      if (row !== undefined)
-        out.set(key, row);
-    }
-    return out;
-  }
-  store(rows, vectors) {
-    const upsert = this.db.prepare("insert or replace into sidecar_vec (message_key, hash, vec) values (?, ?, ?)");
-    this.transaction(() => {
-      for (const [index, row] of rows.entries()) {
-        upsert.run(row.key, row.hash, encodeVector(vectors?.[index] ?? []));
+    child.stderr.on("data", (chunk) => {
+      stderr = appendOutput(stderr, chunk, max);
+      if (stderr.exceeded && reason === undefined) {
+        reason = "stderr-too-large";
+        terminate4(child);
       }
     });
-    this.lastSemantic = undefined;
+    child.on("error", (error) => {
+      if (settled)
+        return;
+      settled = true;
+      clearTimeout(timer);
+      signal?.removeEventListener("abort", abort);
+      resolveResult({
+        ok: false,
+        reason: error.code === "ENOENT" ? "binary-missing" : "spawn-error",
+        stdout: stdout.text,
+        stderr: "",
+        code: null
+      });
+    });
+    child.on("close", (code) => {
+      if (settled)
+        return;
+      settled = true;
+      clearTimeout(timer);
+      signal?.removeEventListener("abort", abort);
+      resolveResult({
+        ok: code === 0 && reason === undefined,
+        ...reason === undefined ? code === 0 ? {} : { reason: "nonzero-exit" } : { reason },
+        stdout: stdout.text,
+        stderr: stderr.text,
+        code
+      });
+    });
+  });
+}
+function terminate4(child) {
+  if (child.killed)
+    return;
+  if (process.platform !== "win32" && child.pid !== undefined) {
+    try {
+      process.kill(-child.pid, "SIGTERM");
+      return;
+    } catch {}
   }
-  *storedVectors() {
-    const rows = this.db.prepare("select message_key, vec from sidecar_vec where length(vec) > 0").iterate();
-    for (const row of rows) {
-      const vector = row.vec instanceof Uint8Array ? decodeVector(row.vec) : undefined;
-      if (vector !== undefined)
-        yield { key: String(row.message_key), vector };
+  child.kill("SIGTERM");
+}
+function parseJson3(text) {
+  try {
+    return JSON.parse(text.trim());
+  } catch {
+    return;
+  }
+}
+function parseObject2(text) {
+  const value = parseJson3(text);
+  return value !== null && typeof value === "object" && !Array.isArray(value) ? value : undefined;
+}
+function objectArray(value, keys) {
+  if (value === null || typeof value !== "object" || Array.isArray(value))
+    return;
+  for (const key of keys)
+    if (Array.isArray(value[key]))
+      return value[key];
+  return;
+}
+function hasAnyNumber(value, keys) {
+  return keys.some((key) => number(value, key) !== undefined);
+}
+function normalizeHit2(value, mode, index) {
+  if (value === null || typeof value !== "object" || Array.isArray(value))
+    return;
+  const row = value;
+  const text = (key) => typeof row[key] === "string" ? row[key] : "";
+  const chunkId = text("chunkId") || text("chunk_id");
+  const messageId = text("messageId") || text("message_id");
+  const accountId = text("accountId") || text("account_id");
+  const mailbox = text("mailbox");
+  const snippet = text("snippet") || text("content") || text("text");
+  if (!isSafeOpaqueId(chunkId) || !isSafeOpaqueId(messageId) || !snippet)
+    return;
+  const to = Array.isArray(row.to) ? row.to.filter((item) => typeof item === "string") : [];
+  return {
+    chunkId,
+    messageId,
+    threadId: text("threadId") || text("thread_id"),
+    accountId,
+    mailbox,
+    subject: text("subject"),
+    from: text("from"),
+    to,
+    date: text("date"),
+    snippet,
+    score: typeof row.score === "number" ? row.score : 1 / (index + 1),
+    mode
+  };
+}
+function number(value, key) {
+  return typeof value[key] === "number" && Number.isFinite(value[key]) ? value[key] : undefined;
+}
+function string(value, key) {
+  return typeof value[key] === "string" ? value[key] : undefined;
+}
+function ok4(data, result) {
+  return { ok: true, data, stdout: result.stdout, stderr: result.stderr, code: result.code ?? 0 };
+}
+function failure3(result) {
+  return {
+    ok: false,
+    reason: result.reason ?? "nonzero-exit",
+    stdout: result.stdout,
+    stderr: boundDiagnosticText6(result.stderr),
+    code: result.code
+  };
+}
+function boundDiagnosticText6(value) {
+  if (value.length === 0)
+    return "";
+  return value.trim().slice(0, 4000);
+}
+function appendOutput(current, chunk, maxBytes) {
+  if (current.exceeded)
+    return current;
+  const bytes = Buffer.from(chunk, "utf8");
+  const remaining = Math.max(0, maxBytes - current.bytes);
+  const prefix = truncateUtf8(chunk, remaining);
+  const total = current.bytes + bytes.byteLength;
+  return {
+    text: current.text + prefix,
+    bytes: Math.min(total, maxBytes),
+    exceeded: total > maxBytes
+  };
+}
+function isSafeOpaqueId(value) {
+  return value.length > 0 && value !== "." && value !== ".." && !/[\\/#\u0000-\u001f\u007f]/u.test(value);
+}
+function findRemoteEmbeddingViolation(env) {
+  for (const [key, value] of Object.entries(env)) {
+    if (value === undefined)
+      continue;
+    const lower = key.toLowerCase();
+    if (REMOTE_EMBEDDING_ENV_KEYS.has(lower))
+      return key;
+    if (lower === "mailcrawl_embedder" || lower === "mailcrawl_embed_url") {
+      if (isRemoteHttpEndpoint(value))
+        return key;
     }
   }
-  transaction(work) {
-    this.db.exec("begin immediate");
+  return;
+}
+function isRemoteHttpEndpoint(value) {
+  const trimmed = value.trim();
+  if (!/^https?:\/\//iu.test(trimmed))
+    return false;
+  try {
+    return !LOOPBACK_HOSTS2.has(new URL(trimmed).hostname);
+  } catch {
+    return true;
+  }
+}
+function truncateUtf8(value, maxBytes) {
+  if (Buffer.byteLength(value) <= maxBytes)
+    return value;
+  const decoder = new TextDecoder("utf-8", { fatal: true });
+  const bytes = Buffer.from(value, "utf8");
+  for (let end = Math.min(maxBytes, bytes.length);end >= 0; end -= 1) {
     try {
-      work();
-      this.db.exec("commit");
-    } catch (error) {
-      this.db.exec("rollback");
-      throw error;
-    }
+      return decoder.decode(bytes.subarray(0, end));
+    } catch {}
   }
-  toRow(row) {
-    const metadata = {};
-    for (const [index, column] of this.options.descriptor.metadata.entries()) {
-      const value = row[`m${index}`];
-      if (typeof value === "string" && value.length > 0 || typeof value === "number")
-        metadata[column.key] = String(value);
-    }
-    const title = typeof row.title === "string" && row.title.length > 0 ? row.title : undefined;
-    const epoch = typeof row.epoch === "number" && Number.isFinite(row.epoch) ? row.epoch : undefined;
-    const content = String(row.content ?? "");
-    const input = `${title ?? ""}
-${content}`.slice(0, MAX_EMBED_INPUT_CHARS);
-    return {
-      key: String(row.k),
-      content,
-      title,
-      epoch,
-      metadata,
-      input,
-      hash: createHash14("sha1").update(input).digest("hex")
-    };
+  return "";
+}
+var DEFAULT_BINARY4 = "mailcrawl", DEFAULT_SEARCH_TIMEOUT_MS2 = 60000, DEFAULT_INDEX_TIMEOUT_MS3 = 1800000, DEFAULT_MAX_BUFFER_BYTES4 = 1048576, LOOPBACK_HOSTS2, SAFE_ENV_KEYS2, SAFE_MAILCRAWL_ENV_KEYS, REMOTE_EMBEDDING_ENV_KEYS;
+var init_client9 = __esm(() => {
+  init_portable_spawn();
+  init_config3();
+  LOOPBACK_HOSTS2 = new Set(["127.0.0.1", "localhost", "[::1]", "::1"]);
+  SAFE_ENV_KEYS2 = new Set(["HOME", "LANG", "LC_ALL", "PATH", "TMPDIR", "TMP", "TEMP", "NO_COLOR"]);
+  SAFE_MAILCRAWL_ENV_KEYS = new Set([
+    "MAILCRAWL_DATA_DIR",
+    "MAILCRAWL_EMBEDDER",
+    "MAILCRAWL_EMBEDDER_PROVIDER",
+    "MAILCRAWL_EMBED_URL",
+    "MAILCRAWL_EMBED_MODEL",
+    "MAILCRAWL_EMBED_DIM",
+    "MAILCRAWL_EMBED_TIMEOUT",
+    "MAILCRAWL_EMBED_BATCH_SIZE",
+    "MAILCRAWL_QUERY_PREFIX",
+    "MAILCRAWL_PASSAGE_PREFIX",
+    "MAILCRAWL_NATIVE_MODEL",
+    "MAILCRAWL_NATIVE_DEVICE",
+    "MAILCRAWL_NATIVE_DTYPE",
+    "MAILCRAWL_KIWI_MODEL",
+    "MAILCRAWL_JA_HELPER",
+    "MAILCRAWL_ZH_HELPER"
+  ]);
+  REMOTE_EMBEDDING_ENV_KEYS = new Set([
+    "mailcrawl_embedder_base_url",
+    "mailcrawl_embedding_base_url",
+    "mailcrawl_allow_remote_embeddings"
+  ]);
+});
+
+// src/datasource/skills/mailcrawl/methods.ts
+class MailcrawlMethod {
+  client;
+  mode;
+  instanceId;
+  tags;
+  account;
+  mailbox;
+  constructor(client, mode, instanceId, tags, account, mailbox) {
+    this.client = client;
+    this.mode = mode;
+    this.instanceId = instanceId;
+    this.tags = tags;
+    this.account = account;
+    this.mailbox = mailbox;
   }
-  toHit(row, score) {
+  describe() {
     return {
-      key: row.key,
-      content: row.content,
-      score,
-      ...row.title !== undefined ? { title: row.title } : {},
-      ...row.epoch !== undefined ? { publishedAt: Math.round(row.epoch * 1000) } : {},
-      metadata: row.metadata
+      name: `mailcrawl-${this.mode}`,
+      type: this.mode === "bm25" ? "bm25" : this.mode === "semantic" ? "vector" : "hybrid",
+      description: `${this.mode} retrieval over archived email via the external mailcrawl CLI`,
+      status: "active",
+      capabilities: ["email", this.mode, "scoped", "external-cli", "path-opaque-sources"],
+      datasourceId: "mailcrawl",
+      tags: this.tags
     };
   }
+  async retrieve(query, options) {
+    if (!query.trim())
+      return [];
+    const result = await this.client.search(this.mode, query, {
+      topK: options.topK,
+      signal: options.signal
+    });
+    if (!result.ok)
+      throw datasourceCliError("mailcrawl", `search --mode ${this.mode}`, result);
+    return result.hits.filter((hit) => {
+      if (this.account !== undefined && hit.accountId !== this.account)
+        return false;
+      if (this.mailbox !== undefined && hit.mailbox !== this.mailbox)
+        return false;
+      const source = datasourceSourcePath("mailcrawl", this.instanceId, hit.chunkId);
+      return matchesDatasourceScope(source, options.scope);
+    }).map((hit) => {
+      const source = datasourceSourcePath("mailcrawl", this.instanceId, hit.chunkId);
+      return {
+        id: `mailcrawl:${this.instanceId}:${hit.chunkId}`,
+        content: hit.snippet,
+        source,
+        score: hit.score,
+        metadata: {
+          method: `mailcrawl-${this.mode}`,
+          datasourceId: "mailcrawl",
+          instanceId: this.instanceId,
+          mode: this.mode,
+          chunkId: hit.chunkId,
+          messageId: hit.messageId,
+          threadId: hit.threadId,
+          accountId: hit.accountId,
+          mailbox: hit.mailbox,
+          subject: hit.subject,
+          from: hit.from,
+          to: hit.to,
+          date: hit.date
+        }
+      };
+    }).slice(0, options.topK ?? 20);
+  }
 }
-function toMatchQuery(text) {
-  const terms = [...new Set(text.match(/[\p{L}\p{N}_]+/gu) ?? [])];
-  return terms.length === 0 ? undefined : terms.map((term) => `"${term}"`).join(" OR ");
-}
-function errorMessage2(error) {
-  return error instanceof Error ? error.message : String(error);
-}
-var DEFAULT_BATCH_SIZE = 256, DEFAULT_MAX_EMBED_PER_SYNC = 2048, MAX_EMBED_INPUT_CHARS = 8192, BUSY_TIMEOUT_MS = 5000, POOL_FACTOR = 3, SEMANTIC_REUSE_MS = 1e4, IDENTITY_KEY = "identity";
-var init_sidecar = () => {};
-
-// src/datasource/sidecar/methods.ts
-function sidecarMethodName(datasourceId, mode) {
-  return mode === "bm25" ? `${datasourceId}-sidecar-bm25` : `${datasourceId}-${mode}`;
-}
+var modes, MAILCRAWL_MODES;
+var init_methods5 = __esm(() => {
+  init_errors4();
+  init_scope2();
+  modes = ["bm25", "semantic", "hybrid"];
+  MAILCRAWL_MODES = modes;
+});
 
-class SidecarRetrievalMethod {
-  options;
-  constructor(options) {
-    this.options = options;
+// src/datasource/skills/mailcrawl/skill.ts
+class MailcrawlSkill {
+  client;
+  instanceId;
+  instances;
+  interval;
+  tags;
+  account;
+  mailbox;
+  lastIndexedAt;
+  constructor(options = {}) {
+    const connectorOptions = options.connectorOptions ?? options;
+    this.instanceId = options.instanceId ?? "default";
+    this.instances = options.instances?.length ? options.instances : [this.instanceId];
+    validateMailcrawlInstanceId(this.instanceId);
+    for (const instance of this.instances)
+      validateMailcrawlInstanceId(instance);
+    this.interval = options.pollingIntervalMs ?? 15 * 60 * 1000;
+    this.tags = options.tags ?? TAGS;
+    this.account = connectorOptions.account;
+    this.mailbox = connectorOptions.mailbox;
+    this.client = options.client ?? new MailcrawlClient({ ...connectorOptions, instanceId: this.instanceId });
   }
   describe() {
-    const { datasourceId, skillType, mode, tags } = this.options;
-    const { type, text, capabilities } = DESCRIPTIONS[mode];
     return {
-      name: sidecarMethodName(datasourceId, mode),
-      type,
-      description: `${text} over ${skillType} via a read-only sidecar index`,
+      name: "mailcrawl",
+      id: "mailcrawl",
+      type: "mailcrawl-archive",
+      description: "Local email archive datasource via the external mailcrawl CLI",
+      capabilities: ["email", "external-cli", "polling", "incremental", "bm25", "semantic", "hybrid"],
+      tags: this.tags,
       status: "active",
-      capabilities: [...capabilities, "scoped", "path-opaque-sources"],
-      datasourceId,
-      tags
+      requiresExternalCli: true,
+      datasourceId: "mailcrawl",
+      instanceId: this.instanceId,
+      instances: this.instances
     };
   }
-  async retrieve(query, options) {
-    const trimmed = query.trim();
-    if (trimmed.length === 0)
-      return [];
-    const topK = options.topK ?? DEFAULT_TOP_K5;
-    const { sidecar, mode } = this.options;
-    const hits = mode === "bm25" ? sidecar.searchLexical(trimmed, topK) : mode === "semantic" ? await sidecar.searchSemantic(trimmed, topK) : await sidecar.searchHybrid(trimmed, topK);
-    const results = [];
-    for (const hit of hits) {
-      const result = this.toResult(hit, options);
-      if (result !== undefined)
-        results.push(result);
+  polling() {
+    return {
+      mode: this.interval > 0 ? "poll" : "none",
+      ...this.interval > 0 ? { intervalMs: this.interval } : {},
+      lastIndexedAt: this.lastIndexedAt
+    };
+  }
+  async index(options) {
+    try {
+      const result = await this.client.sync(options?.signal);
+      if (!result.ok) {
+        return this.fail(result.reason === "binary-missing" || result.reason === "spawn-error" || result.reason === "timeout" ? "datasource-unavailable" : result.reason === "remote-embedding-rejected" ? "datasource-embedding-egress-rejected" : "datasource-index-failed", result.reason);
+      }
+      const indexed = await this.client.index(options?.signal);
+      const diagnostics = [];
+      if (!indexed.ok) {
+        diagnostics.push({
+          code: indexed.reason === "remote-embedding-rejected" ? "datasource-embedding-egress-rejected" : "datasource-index-failed",
+          severity: "warning",
+          message: `mailcrawl semantic index unavailable (${indexed.reason}); BM25 remains available`,
+          instanceId: this.instanceId,
+          source: "mailcrawl"
+        });
+      }
+      this.lastIndexedAt = Date.now();
+      return {
+        ok: true,
+        instanceId: this.instanceId,
+        skill: "mailcrawl",
+        chunkCount: result.data.chunksAdded ?? result.data.messages,
+        indexedAt: this.lastIndexedAt,
+        diagnostics
+      };
+    } catch {
+      return this.fail("datasource-unavailable", "mailcrawl sync failed unexpectedly");
     }
-    return results;
   }
-  toResult(hit, options) {
-    const { descriptor, datasourceId, instanceId, mode } = this.options;
-    const chunkId = descriptor.chunkId(hit.key, hit.metadata);
-    const source = datasourceSourcePath(datasourceId, instanceId, chunkId);
-    if (!matchesDatasourceScope(source, options.scope))
-      return;
-    const method = sidecarMethodName(datasourceId, mode);
+  retrievalMethods() {
+    return [
+      new MailcrawlMethod(this.client, "bm25", this.instanceId, this.tags, this.account, this.mailbox),
+      new MailcrawlMethod(this.client, "semantic", this.instanceId, this.tags, this.account, this.mailbox),
+      new MailcrawlMethod(this.client, "hybrid", this.instanceId, this.tags, this.account, this.mailbox)
+    ];
+  }
+  describeSources() {
+    return this.instances.map((instanceId) => ({
+      source: datasourceSourcePath("mailcrawl", instanceId),
+      datasourceId: "mailcrawl",
+      skill: "mailcrawl",
+      instanceId,
+      contentType: "email",
+      metadata: { datasourceId: "mailcrawl", instanceId, tags: this.tags }
+    }));
+  }
+  skillManifest() {
+    const scopes = this.instances.map((id) => `- \`${datasourceSourcePath("mailcrawl", id)}\``).join(`
+`);
     return {
-      id: `${datasourceId}:${instanceId}:${chunkId}`,
-      content: hit.content,
-      source,
-      score: hit.score,
-      metadata: {
-        ...hit.metadata,
-        method,
-        datasourceId,
-        instanceId,
-        backend: "sidecar",
-        ...hit.title !== undefined ? { title: hit.title } : {},
-        hierarchy: descriptor.hierarchy(hit.metadata).join("/"),
-        ...hit.publishedAt !== undefined ? { publishedAt: hit.publishedAt } : {}
-      }
+      name: "datasource-mailcrawl",
+      description: "Search local email synchronized and indexed by mailcrawl.",
+      content: `# mailcrawl email datasource
+
+Use this skill for authorized local email retrieval. AutoRAG invokes the external \`mailcrawl\` CLI and never opens its SQLite archive directly.
+
+## Indexing
+Indexing is server-managed and incremental; do not trigger sync yourself.
+
+## Search
+Call the dedicated \`${datasourceSearchToolName("mailcrawl")}\` tool with a natural-language query. Available scopes:
+${scopes}
+
+The selectable retrieval modes are BM25, semantic, and hybrid. Scope arguments can only narrow trusted access. Source identifiers remain opaque and hierarchical.
+
+## Native CLI
+The external \`mailcrawl\` CLI owns the archive and index. When the dedicated tool cannot express what you need, call mailcrawl directly through \`bash\`:
+- \`mailcrawl search --mode hybrid --limit 20 --json "<query>"\` (\`bm25\` and \`semantic\` modes also available; \`--account\` / \`--mailbox\` narrow the archive)
+Never pass datasource virtual paths (\`/mailcrawl/...\`) to bash; they are not OS paths.`
+    };
+  }
+  fail(code, message) {
+    return {
+      ok: false,
+      instanceId: this.instanceId,
+      skill: "mailcrawl",
+      indexedAt: Date.now(),
+      diagnostics: [
+        {
+          code,
+          severity: code === "datasource-unavailable" ? "warning" : "error",
+          message: `mailcrawl ${message}`,
+          instanceId: this.instanceId,
+          source: "mailcrawl"
+        }
+      ],
+      error: message,
+      code,
+      message
     };
   }
 }
-var DEFAULT_TOP_K5 = 20, DESCRIPTIONS;
-var init_methods5 = __esm(() => {
+var TAGS;
+var init_skill9 = __esm(() => {
   init_scope2();
-  DESCRIPTIONS = {
-    bm25: { type: "bm25", text: "BM25-ranked FTS5 retrieval", capabilities: ["lexical", "fts5", "bm25"] },
-    semantic: { type: "vector", text: "Semantic (embedding) retrieval", capabilities: ["semantic"] },
-    hybrid: {
-      type: "hybrid",
-      text: "Hybrid BM25 + semantic retrieval (reciprocal rank fusion)",
-      capabilities: ["lexical", "semantic", "hybrid"]
-    }
-  };
+  init_client9();
+  init_config3();
+  init_methods5();
+  TAGS = ["mailcrawl", "email", "pii"];
+});
+
+// src/datasource/skills/mailcrawl/index.ts
+var init_mailcrawl = __esm(() => {
+  init_client9();
+  init_methods5();
+  init_skill9();
 });
 
 // src/datasource/crawler-skill.ts
@@ -29029,7 +29425,7 @@ var init_crawler_skill = __esm(() => {
   init_errors4();
   init_scope2();
   init_sidecar();
-  init_methods5();
+  init_methods4();
   DEFAULT_POLLING_INTERVAL_MS4 = 15 * 60 * 1000;
 });
 
@@ -30656,7 +31052,7 @@ var init_wacrawl = __esm(() => {
 });
 
 // src/datasource/skills/factory.ts
-import { homedir as homedir4 } from "node:os";
+import { homedir as homedir5 } from "node:os";
 import { join as join39 } from "node:path";
 function chatCrawlerArgs(descriptor, defaultCrawlerDb, config, workspaceRoot, registrationName) {
   const sidecar = resolveSidecarOptions({
@@ -30760,9 +31156,9 @@ var init_factory = __esm(() => {
   init_telecrawl();
   init_wacrawl();
   BUILDERS = {
-    telegram: (config, workspaceRoot, registrationName) => new TelecrawlSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.telegram, join39(homedir4(), ".telecrawl", "telecrawl.db"), config, workspaceRoot, registrationName)),
-    whatsapp: (config, workspaceRoot, registrationName) => new WacrawlSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.whatsapp, join39(homedir4(), ".wacrawl", "wacrawl.db"), config, workspaceRoot, registrationName)),
-    slack: (config, workspaceRoot, registrationName) => new SlackSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.slack, join39(homedir4(), ".slacrawl", "slacrawl.db"), config, workspaceRoot, registrationName)),
+    telegram: (config, workspaceRoot, registrationName) => new TelecrawlSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.telegram, join39(homedir5(), ".telecrawl", "telecrawl.db"), config, workspaceRoot, registrationName)),
+    whatsapp: (config, workspaceRoot, registrationName) => new WacrawlSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.whatsapp, join39(homedir5(), ".wacrawl", "wacrawl.db"), config, workspaceRoot, registrationName)),
+    slack: (config, workspaceRoot, registrationName) => new SlackSkill(chatCrawlerArgs(CRAWLER_DESCRIPTORS.slack, join39(homedir5(), ".slacrawl", "slacrawl.db"), config, workspaceRoot, registrationName)),
     discord: (config, workspaceRoot, registrationName, embeddingRuntime) => {
       const connector = config.connector ?? {};
       const clientOptions = {
@@ -30818,24 +31214,42 @@ var init_factory = __esm(() => {
       ...config.channels?.ids !== undefined ? { chatIds: config.channels.ids } : {},
       ...larkConnectorOptions(config.connector)
     }),
-    kakao: (config, _workspaceRoot, _registrationName) => new LazykatokSkill({
-      client: new LazykatokClient({
-        ...config.connector
-      }),
-      ...config.instanceId !== undefined ? { instanceId: config.instanceId } : {},
-      ...config.pollingIntervalMs !== undefined ? { pollingIntervalMs: config.pollingIntervalMs } : {},
-      ...config.tags !== undefined ? { tags: config.tags } : {}
-    }),
+    kakao: (config, workspaceRoot, registrationName) => {
+      const connector = config.connector ?? {};
+      const dataDir = typeof connector.workspacePath === "string" && connector.workspacePath.length > 0 ? connector.workspacePath : join39(homedir5(), "Library", "Application Support", "katok");
+      const sidecar = resolveSidecarOptions({
+        descriptor: CRAWLER_DESCRIPTORS.kakao,
+        connector: config.connector,
+        workspaceRoot,
+        registrationName,
+        instanceId: config.instanceId ?? "default",
+        defaultCrawlerDb: join39(dataDir, "archive.sqlite3")
+      });
+      return new LazykatokSkill({
+        client: new LazykatokClient(connector),
+        ...config.instanceId !== undefined ? { instanceId: config.instanceId } : {},
+        ...config.pollingIntervalMs !== undefined ? { pollingIntervalMs: config.pollingIntervalMs } : {},
+        ...config.tags !== undefined ? { tags: config.tags } : {},
+        channelIds: config.channels?.ids,
+        channelNames: config.channels?.names,
+        ...sidecar === undefined ? {} : { sidecar }
+      });
+    },
     github: (config, workspaceRoot, registrationName) => new GitHubSkill({
       ...common(config, workspaceRoot),
       skillName: registrationName,
       connectorOptions: config.connector
     }),
-    "github-gist": (config, workspaceRoot, registrationName) => new GitHubGistSkill({
-      ...common(config, workspaceRoot),
-      skillName: registrationName,
-      connectorOptions: config.connector
-    }),
+    "github-gist": (config, workspaceRoot, registrationName) => {
+      const { semantic, ...connectorOptions } = config.connector ?? {};
+      const semanticOff = semantic === false || typeof semantic === "object" && semantic !== null && semantic.enabled === false;
+      return new GitHubGistSkill({
+        ...common(config, workspaceRoot),
+        skillName: registrationName,
+        connectorOptions,
+        ...semanticOff ? { semantic: { enabled: false } } : {}
+      });
+    },
     "cloud-drive": (config, workspaceRoot, registrationName) => new CloudDriveSkill({
       ...common(config, workspaceRoot),
       skillName: registrationName,
@@ -30881,7 +31295,7 @@ var init_factory = __esm(() => {
 
 // src/cli/config.ts
 import { randomUUID as randomUUID8 } from "node:crypto";
-import { existsSync as existsSync28, mkdirSync as mkdirSync25, readFileSync as readFileSync25, renameSync as renameSync10, unlinkSync as unlinkSync5, writeFileSync as writeFileSync22 } from "node:fs";
+import { existsSync as existsSync29, mkdirSync as mkdirSync25, readFileSync as readFileSync25, renameSync as renameSync10, unlinkSync as unlinkSync5, writeFileSync as writeFileSync22 } from "node:fs";
 import { basename as basename14, dirname as dirname26, isAbsolute as isAbsolute10, join as join40, resolve as resolve16 } from "node:path";
 import { findEnvKeys, getEnvApiKey } from "@earendil-works/pi-ai/compat";
 import { getAgentDir as getAgentDir4, ModelRuntime as ModelRuntime2, SettingsManager as SettingsManager2 } from "@earendil-works/pi-coding-agent";
@@ -30906,7 +31320,7 @@ function resolveConfigPath(input) {
 function readConfigFile(configPath, explicit) {
   let exists;
   try {
-    exists = existsSync28(configPath);
+    exists = existsSync29(configPath);
   } catch {
     exists = false;
   }
@@ -30966,12 +31380,12 @@ function replaceFileAtomically(path, contents, assertCommitAllowed) {
   }
 }
 function migrateLegacyConfig(configPath, legacyPath) {
-  if (existsSync28(configPath) || !existsSync28(legacyPath))
+  if (existsSync29(configPath) || !existsSync29(legacyPath))
     return;
   const legacy = readConfigFile(legacyPath, true);
   const legacyBytes = readFileSync25(legacyPath);
   const migrated = normalizeLegacyConfigPaths(legacy ?? {}, dirname26(legacyPath));
-  const migratedPathsUnchanged = legacy?.workspacePath === migrated.workspacePath && legacy?.memoryPath === migrated.memoryPath && JSON.stringify(legacy?.searchPaths) === JSON.stringify(migrated.searchPaths);
+  const migratedPathsUnchanged = legacy?.workspacePath === migrated.workspacePath && legacy?.memoryPath === migrated.memoryPath && JSON.stringify(legacy?.searchPaths) === JSON.stringify(migrated.searchPaths) && JSON.stringify(legacy?.excludePaths) === JSON.stringify(migrated.excludePaths);
   const migratedBytes = migratedPathsUnchanged ? legacyBytes : `${JSON.stringify(migrated, null, 2)}
 `;
   mkdirSync25(dirname26(configPath), { recursive: true });
@@ -31009,15 +31423,25 @@ function resolveConfigFileReadOnly(configPath, explicit, legacyPath) {
 function resolveSearchPaths(searchPaths, origin) {
   return searchPaths.map((searchPath) => resolvePersistedPath(searchPath, origin));
 }
+function resolveExcludePaths(value, origin) {
+  if (value === undefined)
+    return;
+  if (!Array.isArray(value) || value.some((path) => typeof path !== "string" || path.length === 0)) {
+    throw new ConfigError("Config field 'excludePaths' must be a non-empty string array");
+  }
+  return resolveSearchPaths(value, origin);
+}
 function resolvePersistedPath(path, origin) {
   return isAbsolute10(path) ? path : resolve16(origin, path);
 }
 function normalizeLegacyConfigPaths(partial, origin) {
   const workspacePath = resolvePersistedPath(partial.workspacePath ?? ".", origin);
+  const excludePaths = resolveExcludePaths(partial.excludePaths, workspacePath);
   return {
     ...partial,
     workspacePath,
     ...partial.searchPaths === undefined ? {} : { searchPaths: resolveSearchPaths(partial.searchPaths, workspacePath) },
+    ...excludePaths === undefined ? {} : { excludePaths },
     ...partial.memoryPath === undefined ? {} : { memoryPath: resolvePersistedPath(partial.memoryPath, workspacePath) }
   };
 }
@@ -31472,10 +31896,7 @@ function resolveConfig(input) {
   const configOrigin = dirname26(resolve16(configPath));
   const fileWorkspacePath = typeof file.workspacePath === "string" ? resolvePersistedPath(file.workspacePath, configOrigin) : undefined;
   const fileSearchPaths = file.searchPaths ? resolveSearchPaths(file.searchPaths, fileWorkspacePath ?? configOrigin) : undefined;
-  if (file.excludePaths !== undefined && (!Array.isArray(file.excludePaths) || file.excludePaths.some((path) => typeof path !== "string" || path.length === 0))) {
-    throw new ConfigError("Config field 'excludePaths' must be a non-empty string array");
-  }
-  const fileExcludePaths = file.excludePaths ? resolveSearchPaths(file.excludePaths, fileWorkspacePath ?? configOrigin) : undefined;
+  const fileExcludePaths = resolveExcludePaths(file.excludePaths, fileWorkspacePath ?? configOrigin);
   const searchPaths = flagSearchPaths ?? envSearchPaths ?? fileSearchPaths ?? defaultSearchPaths;
   const flagWorkspacePath = flagString(flags, "workspace");
   const envWorkspacePath = envString(env, "AUTORAG_WORKSPACE");
@@ -32211,7 +32632,7 @@ function writeDefaultConfig(path, partial, opts = {}) {
   full.dupey = partial.dupey ?? { enabled: true };
   full.excludeExactDuplicates = partial.excludeExactDuplicates ?? true;
   if (partial.excludePaths !== undefined)
-    full.excludePaths = resolveSearchPaths(partial.excludePaths, workspacePath);
+    full.excludePaths = resolveExcludePaths(partial.excludePaths, workspacePath);
   if (partial.limits !== undefined)
     full.limits = normalizeLimitsConfig(partial.limits);
   if (partial.refresh !== undefined)
@@ -32230,7 +32651,7 @@ function writeDefaultConfig(path, partial, opts = {}) {
 `;
   const lock = acquireConfigWriteLock(path);
   try {
-    const exists = existsSync28(path);
+    const exists = existsSync29(path);
     if (!opts.force && exists) {
       throw new ConfigError(`Config file already exists: ${path}`);
     }