cc-claw 0.3.11 → 0.4.1

package/dist/agents/mcp-server.js

@@ -72,7 +72,7 @@ Typical workflow:
       description: z.string().optional().describe("Short description of what this agent does"),
       model: z.string().optional().describe("Model override for the sub-agent"),
       skills: z.array(z.string()).optional().describe("Skills to inject into the sub-agent"),
-      permMode: z.string().optional().describe("Permission mode: yolo, safe, readonly, plan (defaults to chat's mode)"),
+      permMode: z.string().optional().describe("Permission mode: yolo, safe, plan (defaults to chat's mode)"),
       role: z.string().optional().describe("Agent role. Presets: worker, reviewer, planner, researcher, writer, analyst, debugger, critic, synthesizer. Custom roles accepted as labels."),
       persona: z.string().optional().describe("Custom system prompt for the agent. Replaces the role's default prompt while preserving orchestrator tools and task."),
       allowedTools: z.array(z.string()).optional().describe("Restrict agent to specific tools (Claude-only). E.g., ['Read', 'Grep', 'Glob']"),

package/dist/cli.js

@@ -48,7 +48,7 @@ var VERSION;
 var init_version = __esm({
   "src/version.ts"() {
     "use strict";
-    VERSION = true ? "0.3.11" : (() => {
+    VERSION = true ? "0.4.1" : (() => {
       try {
         return JSON.parse(readFileSync(join2(process.cwd(), "package.json"), "utf-8")).version ?? "unknown";
       } catch {
@@ -922,12 +922,14 @@ __export(store_exports3, {
   ALL_TOOLS: () => ALL_TOOLS,
   addHeartbeatWatch: () => addHeartbeatWatch,
   addUsage: () => addUsage,
+  appendMessageLog: () => appendMessageLog,
   cancelJobById: () => cancelJobById,
   checkBackendLimits: () => checkBackendLimits,
   cleanExpiredWatches: () => cleanExpiredWatches,
   clearAllSessions: () => clearAllSessions,
   clearBackendLimit: () => clearBackendLimit,
   clearCwd: () => clearCwd,
+  clearMessageLog: () => clearMessageLog,
   clearModel: () => clearModel,
   clearSession: () => clearSession,
   clearSummarizer: () => clearSummarizer,
@@ -935,6 +937,7 @@ __export(store_exports3, {
   clearUsage: () => clearUsage,
   completeJobRun: () => completeJobRun,
   deleteBookmark: () => deleteBookmark,
+  determineEscalationTarget: () => determineEscalationTarget,
   findBookmarksByPrefix: () => findBookmarksByPrefix,
   forgetMemory: () => forgetMemory,
   getActiveJobs: () => getActiveJobs,
@@ -961,14 +964,18 @@ __export(store_exports3, {
   getMemoriesWithoutEmbeddings: () => getMemoriesWithoutEmbeddings,
   getMode: () => getMode,
   getModel: () => getModel,
+  getPendingEscalation: () => getPendingEscalation,
   getRecentBookmarks: () => getRecentBookmarks,
   getRecentMemories: () => getRecentMemories,
+  getRecentMessageLog: () => getRecentMessageLog,
   getSessionId: () => getSessionId,
   getSessionStartedAt: () => getSessionStartedAt,
   getSessionSummariesWithoutEmbeddings: () => getSessionSummariesWithoutEmbeddings,
   getSummarizer: () => getSummarizer,
   getThinkingLevel: () => getThinkingLevel,
   getToolsMap: () => getToolsMap,
+  getUnsummarizedChatIds: () => getUnsummarizedChatIds,
+  getUnsummarizedLog: () => getUnsummarizedLog,
   getUsage: () => getUsage,
   getVerboseLevel: () => getVerboseLevel,
   incrementJobFailures: () => incrementJobFailures,
@@ -977,10 +984,13 @@ __export(store_exports3, {
   insertJobRun: () => insertJobRun,
   listMemories: () => listMemories,
   listSessionSummaries: () => listSessionSummaries,
+  markMessageLogSummarized: () => markMessageLogSummarized,
   openDatabaseReadOnly: () => openDatabaseReadOnly,
   pruneJobRuns: () => pruneJobRuns,
+  pruneMessageLog: () => pruneMessageLog,
   removeChatAlias: () => removeChatAlias,
   removeHeartbeatWatch: () => removeHeartbeatWatch,
+  removePendingEscalation: () => removePendingEscalation,
   resetJobFailures: () => resetJobFailures,
   resetTools: () => resetTools,
   saveMemory: () => saveMemory,
@@ -989,6 +999,7 @@ __export(store_exports3, {
   saveSessionSummaryWithEmbedding: () => saveSessionSummaryWithEmbedding,
   searchMemories: () => searchMemories,
   searchMemoriesReadOnly: () => searchMemoriesReadOnly,
+  searchMessageLog: () => searchMessageLog,
   searchSessionSummaries: () => searchSessionSummaries,
   setBackend: () => setBackend,
   setBackendLimit: () => setBackendLimit,
@@ -1003,6 +1014,8 @@ __export(store_exports3, {
   setSummarizer: () => setSummarizer,
   setThinkingLevel: () => setThinkingLevel,
   setVerboseLevel: () => setVerboseLevel,
+  storePendingEscalation: () => storePendingEscalation,
+  toFts5Query: () => toFts5Query,
   toggleTool: () => toggleTool,
   touchBookmark: () => touchBookmark,
   updateHeartbeatTimestamps: () => updateHeartbeatTimestamps,
@@ -1084,6 +1097,7 @@ function initDatabase() {
       mode TEXT NOT NULL DEFAULT 'yolo'
     );
   `);
+  db.prepare("UPDATE chat_mode SET mode = 'plan' WHERE mode = 'readonly'").run();
   db.exec(`
     CREATE TABLE IF NOT EXISTS chat_verbose (
       chat_id TEXT PRIMARY KEY,
@@ -1395,6 +1409,52 @@ function initDatabase() {
       PRIMARY KEY (chatId, alias)
     )
   `);
+  db.exec(`
+    CREATE TABLE IF NOT EXISTS message_log (
+      id         INTEGER PRIMARY KEY AUTOINCREMENT,
+      chat_id    TEXT    NOT NULL,
+      role       TEXT    NOT NULL,
+      content    TEXT    NOT NULL,
+      backend    TEXT,
+      model      TEXT,
+      session_id TEXT,
+      summarized INTEGER NOT NULL DEFAULT 0,
+      created_at TEXT    NOT NULL DEFAULT (datetime('now'))
+    )
+  `);
+  db.exec(`
+    CREATE INDEX IF NOT EXISTS idx_message_log_chat
+      ON message_log (chat_id, created_at DESC)
+  `);
+  db.exec(`
+    CREATE INDEX IF NOT EXISTS idx_message_log_unsummarized
+      ON message_log (chat_id, summarized, created_at ASC)
+  `);
+  db.exec(`
+    CREATE VIRTUAL TABLE IF NOT EXISTS message_log_fts
+      USING fts5(content, content='message_log', content_rowid='id')
+  `);
+  db.exec(`
+    CREATE TRIGGER IF NOT EXISTS message_log_ai
+      AFTER INSERT ON message_log BEGIN
+        INSERT INTO message_log_fts(rowid, content) VALUES (new.id, new.content);
+      END
+  `);
+  db.exec(`
+    CREATE TRIGGER IF NOT EXISTS message_log_ad
+      AFTER DELETE ON message_log BEGIN
+        INSERT INTO message_log_fts(message_log_fts, rowid, content)
+          VALUES ('delete', old.id, old.content);
+      END
+  `);
+  db.exec(`
+    CREATE TRIGGER IF NOT EXISTS message_log_au
+      AFTER UPDATE ON message_log BEGIN
+        INSERT INTO message_log_fts(message_log_fts, rowid, content)
+          VALUES ('delete', old.id, old.content);
+        INSERT INTO message_log_fts(rowid, content) VALUES (new.id, new.content);
+      END
+  `);
   try {
     db.exec("ALTER TABLE memories ADD COLUMN embedding TEXT");
   } catch {
@@ -1674,7 +1734,8 @@ function getMode(chatId) {
   const row = db.prepare(
     "SELECT mode FROM chat_mode WHERE chat_id = ?"
   ).get(chatId);
-  return row?.mode ?? "yolo";
+  const raw = row?.mode ?? "yolo";
+  return raw === "readonly" ? "plan" : raw;
 }
 function setMode(chatId, mode) {
   db.prepare(`
@@ -1683,6 +1744,30 @@ function setMode(chatId, mode) {
     ON CONFLICT(chat_id) DO UPDATE SET mode = ?
   `).run(chatId, mode, mode);
 }
+function storePendingEscalation(chatId, message) {
+  pendingEscalations.set(chatId, { message, createdAt: Date.now() });
+  setTimeout(() => pendingEscalations.delete(chatId), ESCALATION_TTL_MS);
+}
+function getPendingEscalation(chatId) {
+  const entry = pendingEscalations.get(chatId);
+  if (!entry) return void 0;
+  if (Date.now() - entry.createdAt > ESCALATION_TTL_MS) {
+    pendingEscalations.delete(chatId);
+    return void 0;
+  }
+  return entry.message;
+}
+function removePendingEscalation(chatId) {
+  pendingEscalations.delete(chatId);
+}
+function determineEscalationTarget(chatId, currentMode) {
+  if (currentMode === "plan") {
+    const tools2 = getEnabledTools(chatId);
+    if (tools2.length > 0 && tools2.length < ALL_TOOLS.length) return "safe";
+    return "yolo";
+  }
+  return "yolo";
+}
 function getUsage(chatId) {
   const row = db.prepare(
     "SELECT input_tokens, output_tokens, cache_read_tokens, request_count, last_input_tokens, last_cache_read_tokens FROM chat_usage WHERE chat_id = ?"
@@ -2166,7 +2251,77 @@ function getAllChatAliases() {
 function removeChatAlias(alias) {
   return db.prepare("DELETE FROM chat_aliases WHERE alias = ?").run(alias.toLowerCase()).changes > 0;
 }
-var db, STOP_WORDS, ALL_TOOLS, JOB_SELECT;
+function appendMessageLog(chatId, role, content, backend2, model2, sessionId) {
+  getDb().prepare(`
+    INSERT INTO message_log (chat_id, role, content, backend, model, session_id)
+    VALUES (?, ?, ?, ?, ?, ?)
+  `).run(chatId, role, content, backend2, model2, sessionId);
+}
+function getRecentMessageLog(chatId, limit) {
+  return getDb().prepare(`
+    SELECT * FROM message_log
+    WHERE chat_id = ?
+    ORDER BY created_at DESC, id DESC
+    LIMIT ?
+  `).all(chatId, limit);
+}
+function getUnsummarizedLog(chatId) {
+  return getDb().prepare(`
+    SELECT * FROM message_log
+    WHERE chat_id = ? AND summarized = 0
+    ORDER BY created_at ASC, id ASC
+  `).all(chatId);
+}
+function searchMessageLog(chatId, query, limit = 20) {
+  const sanitized = toFts5Query(query);
+  if (!sanitized) return [];
+  try {
+    return getDb().prepare(`
+      SELECT ml.* FROM message_log ml
+      JOIN message_log_fts fts ON fts.rowid = ml.id
+      WHERE ml.chat_id = ? AND message_log_fts MATCH ?
+      ORDER BY fts.rank
+      LIMIT ?
+    `).all(chatId, sanitized, limit);
+  } catch {
+    return [];
+  }
+}
+function markMessageLogSummarized(chatId) {
+  getDb().prepare(`
+    UPDATE message_log SET summarized = 1
+    WHERE chat_id = ? AND summarized = 0
+  `).run(chatId);
+}
+function clearMessageLog(chatId) {
+  getDb().prepare(`DELETE FROM message_log WHERE chat_id = ?`).run(chatId);
+}
+function getUnsummarizedChatIds() {
+  const rows = getDb().prepare(
+    `SELECT DISTINCT chat_id FROM message_log WHERE summarized = 0`
+  ).all();
+  return rows.map((r) => r.chat_id);
+}
+function pruneMessageLog(retentionDays = 30, rowCapPerChat = 2e3) {
+  const db3 = getDb();
+  db3.prepare(`
+    DELETE FROM message_log
+    WHERE created_at < datetime('now', '-' || ? || ' days')
+  `).run(retentionDays);
+  const chats2 = db3.prepare(
+    `SELECT DISTINCT chat_id FROM message_log`
+  ).all();
+  for (const { chat_id } of chats2) {
+    db3.prepare(`
+      DELETE FROM message_log
+      WHERE chat_id = ? AND id NOT IN (
+        SELECT id FROM message_log WHERE chat_id = ?
+        ORDER BY created_at DESC, id DESC LIMIT ?
+      )
+    `).run(chat_id, chat_id, rowCapPerChat);
+  }
+}
+var db, STOP_WORDS, pendingEscalations, ESCALATION_TTL_MS, ALL_TOOLS, JOB_SELECT;
 var init_store4 = __esm({
   "src/memory/store.ts"() {
     "use strict";
@@ -2291,6 +2446,8 @@ var init_store4 = __esm({
       "near",
       "or"
     ]);
+    pendingEscalations = /* @__PURE__ */ new Map();
+    ESCALATION_TTL_MS = 5 * 60 * 1e3;
     ALL_TOOLS = ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "WebFetch", "WebSearch", "Agent", "AskUserQuestion"];
     JOB_SELECT = `
   SELECT id, schedule_type as scheduleType, cron, at_time as atTime, every_ms as everyMs,
@@ -2444,15 +2601,14 @@ var init_claude = __esm({
         switch (opts.permMode) {
           case "plan":
             args.push("--permission-mode", "plan");
-            break;
-          case "readonly":
-            args.push("--permission-mode", "bypassPermissions");
-            args.push("--allowedTools", "Read,Glob,Grep");
+            args.push("--strict-mcp-config");
             break;
           case "safe":
             args.push("--permission-mode", "bypassPermissions");
             if (opts.allowedTools.length > 0) {
-              args.push("--allowedTools", opts.allowedTools.join(","));
+              args.push("--tools", opts.allowedTools.join(","));
+            } else {
+              args.push("--tools", "Read,Glob,Grep");
             }
             break;
           case "yolo":
@@ -2611,11 +2767,10 @@ var init_gemini = __esm({
         if (opts.model) args.push("-m", opts.model);
         switch (opts.permMode) {
           case "plan":
-          case "readonly":
             args.push("--approval-mode", "plan");
             break;
           case "safe":
-            args.push("--approval-mode", "auto_edit");
+            args.push("--approval-mode", "plan");
             break;
           case "yolo":
           default:
@@ -2697,21 +2852,34 @@ var init_codex = __esm({
           thinking: "adjustable",
           thinkingLevels: ["low", "medium", "high", "extra_high"],
           defaultThinkingLevel: "medium"
+        },
+        "gpt-5.1-codex-max": {
+          label: "GPT-5.1 Codex Max \u2014 deep reasoning",
+          thinking: "adjustable",
+          thinkingLevels: ["low", "medium", "high", "extra_high"],
+          defaultThinkingLevel: "medium"
+        },
+        "gpt-5.1-codex-mini": {
+          label: "GPT-5.1 Codex Mini \u2014 cheaper, faster",
+          thinking: "adjustable",
+          thinkingLevels: ["low", "medium", "high", "extra_high"],
+          defaultThinkingLevel: "low"
         }
       };
       defaultModel = "gpt-5.4";
-      // Internal-only model for summarization — not shown in /model picker
       summarizerModel = "gpt-5.1-codex-mini";
       pricing = {
         "gpt-5.4": { in: 2.5, out: 10, cache: 0.63 },
         "gpt-5.3-codex": { in: 2, out: 8, cache: 0.5 },
         "gpt-5.2-codex": { in: 2, out: 8, cache: 0.5 },
+        "gpt-5.1-codex-max": { in: 1.25, out: 10, cache: 0.125 },
         "gpt-5.1-codex-mini": { in: 0.3, out: 1.2, cache: 0.08 }
       };
       contextWindow = {
         "gpt-5.4": 2e5,
         "gpt-5.3-codex": 2e5,
         "gpt-5.2-codex": 2e5,
+        "gpt-5.1-codex-max": 4e5,
         "gpt-5.1-codex-mini": 2e5
       };
       // Codex has no single result event with final text -- the caller (spawnQuery)
@@ -2753,11 +2921,10 @@ var init_codex = __esm({
         if (opts.model) args.push("-m", opts.model);
         switch (opts.permMode) {
           case "plan":
-          case "readonly":
             args.push("--sandbox", "read-only");
             break;
           case "safe":
-            args.push("--full-auto");
+            args.push("--sandbox", "read-only");
             break;
           case "yolo":
           default:
@@ -2915,6 +3082,31 @@ function mergeAndScore(allItems, vectorScores, ftsScores, getDays, decayRate, to
   results.sort((a, b) => b.score - a.score);
   return results.slice(0, topK);
 }
+function setPendingContextBridge(chatId, bridge) {
+  pendingContextBridges.set(chatId, bridge);
+}
+function consumeContextBridge(chatId) {
+  const bridge = pendingContextBridges.get(chatId) ?? null;
+  pendingContextBridges.delete(chatId);
+  return bridge;
+}
+function buildContextBridge(chatId, pairs = 15) {
+  const rows = getRecentMessageLog(chatId, pairs * 2).reverse();
+  if (rows.length === 0) return null;
+  const lines = [`[Conversation history \u2014 continued from previous session]`];
+  let totalChars = lines[0].length;
+  for (const row of rows) {
+    const backendLabel = row.backend ?? "unknown";
+    const date = row.created_at.slice(0, 10);
+    const roleLabel = row.role === "user" ? "You" : "Assistant";
+    const line = `[${backendLabel} \xB7 ${date}] ${roleLabel}: ${row.content}`;
+    if (totalChars + line.length > MAX_BRIDGE_CHARS) break;
+    lines.push(line);
+    totalChars += line.length;
+  }
+  lines.push(`[End of recent history \u2014 you are continuing this conversation]`);
+  return lines.join("\n");
+}
 async function injectMemoryContext(userMessage) {
   const provider = resolveProvider();
   const useVectors = !!provider;
@@ -3007,7 +3199,7 @@ async function injectMemoryContext(userMessage) {
 ${lines.join("\n")}
 [End memory context]`;
 }
-var MEMORY_DECAY_RATE, SESSION_DECAY_RATE, VECTOR_TOP_K, FTS_TOP_K, FINAL_TOP_K_MEMORIES, FINAL_TOP_K_SESSIONS, MAX_MEMORY_CHARS, MAX_SESSION_CHARS;
+var MEMORY_DECAY_RATE, SESSION_DECAY_RATE, VECTOR_TOP_K, FTS_TOP_K, FINAL_TOP_K_MEMORIES, FINAL_TOP_K_SESSIONS, MAX_MEMORY_CHARS, MAX_SESSION_CHARS, MAX_BRIDGE_CHARS, pendingContextBridges;
 var init_inject = __esm({
   "src/memory/inject.ts"() {
     "use strict";
@@ -3021,6 +3213,8 @@ var init_inject = __esm({
     FINAL_TOP_K_SESSIONS = 3;
     MAX_MEMORY_CHARS = 500;
     MAX_SESSION_CHARS = 800;
+    MAX_BRIDGE_CHARS = 48e3;
+    pendingContextBridges = /* @__PURE__ */ new Map();
   }
 });
 
@@ -3205,9 +3399,12 @@ function searchContext(userMessage) {
   }
   return null;
 }
-async function assembleBootstrapPrompt(userMessage, tier = "full", chatId) {
+async function assembleBootstrapPrompt(userMessage, tier = "full", chatId, permMode) {
   const sections = [];
   syncNativeCliFiles();
+  if (permMode && permMode !== "yolo") {
+    sections.push(buildPermissionNotice(permMode));
+  }
   if (tier === "full") {
     const ctx = searchContext(userMessage);
     if (ctx) {
@@ -3215,6 +3412,12 @@ async function assembleBootstrapPrompt(userMessage, tier = "full", chatId) {
 ${ctx}`);
     }
   }
+  if (chatId && tier !== "slim") {
+    const bridge = consumeContextBridge(chatId);
+    if (bridge) {
+      sections.push(bridge);
+    }
+  }
   if (tier !== "slim") {
     const memory2 = await injectMemoryContext(userMessage);
     if (memory2) {
@@ -3232,6 +3435,12 @@ ${ctx}`);
   log(`[bootstrap] Assembled prompt: tier=${tier}, sections=${sections.length}, totalChars=${result.length}`);
   return result;
 }
+function buildPermissionNotice(mode) {
+  if (mode === "plan") {
+    return "[Permission notice]\nYou are in restricted mode (plan). You can read and analyze files but CANNOT edit, write, or execute commands.\nIf the user's request requires modifications or execution, include the marker [NEED_PERMISSION] at the end of your response. Briefly explain what you would need to do and why elevated access is required. Do NOT attempt workarounds or refuse the task \u2014 signal the need and the system will ask the user to approve.\n[End permission notice]";
+  }
+  return "[Permission notice]\nYou are in restricted mode (safe). Only a subset of tools are available to you.\nIf the user's request requires tools you don't have access to, include the marker [NEED_PERMISSION] at the end of your response. Briefly explain what you would need to do and why elevated access is required. Do NOT attempt workarounds or refuse the task \u2014 signal the need and the system will ask the user to approve.\n[End permission notice]";
+}
 function truncateToTokenBudget(text, budget) {
   const charBudget = budget * 4;
   if (text.length <= charBudget) return text;
@@ -3296,66 +3505,62 @@ var init_loader = __esm({
 });
 
 // src/memory/session-log.ts
-function appendToLog(chatId, userMessage, assistantResponse) {
-  if (!logs.has(chatId)) logs.set(chatId, []);
-  const entries = logs.get(chatId);
+function appendToLog(chatId, userMessage, assistantResponse, backend2, model2, sessionId) {
   const now = Date.now();
-  entries.push({ role: "user", text: userMessage, timestamp: now });
-  entries.push({ role: "assistant", text: assistantResponse, timestamp: now });
-  const maxEntries = MAX_PAIRS * 2;
-  if (entries.length > maxEntries) {
-    entries.splice(0, entries.length - maxEntries);
+  appendMessageLog(chatId, "user", userMessage, backend2 ?? null, model2 ?? null, sessionId ?? null);
+  appendMessageLog(chatId, "assistant", assistantResponse, backend2 ?? null, model2 ?? null, sessionId ?? null);
+  const existing = cache.get(chatId) ?? [];
+  existing.push(
+    { role: "user", text: userMessage, timestamp: now },
+    { role: "assistant", text: assistantResponse, timestamp: now }
+  );
+  if (existing.length > CACHE_SIZE * 2) {
+    existing.splice(0, existing.length - CACHE_SIZE * 2);
   }
+  cache.set(chatId, existing);
 }
 function getLog(chatId) {
-  return logs.get(chatId) ?? [];
+  const rows = getUnsummarizedLog(chatId);
+  return rows.map((r) => ({
+    role: r.role,
+    text: r.content,
+    timestamp: (/* @__PURE__ */ new Date(r.created_at + (r.created_at.includes("Z") ? "" : "Z"))).getTime()
+  }));
+}
+function getMessagePairCount(chatId) {
+  const cached = cache.get(chatId);
+  if (cached && cached.length > 0) {
+    return Math.floor(cached.length / 2);
+  }
+  return Math.floor(getUnsummarizedLog(chatId).length / 2);
 }
 function clearLog(chatId) {
-  logs.delete(chatId);
+  markMessageLogSummarized(chatId);
+  cache.delete(chatId);
 }
 function getLoggedChatIds() {
-  return [...logs.keys()].filter((id) => (logs.get(id)?.length ?? 0) > 0);
+  return getUnsummarizedChatIds();
 }
-function getMessagePairCount(chatId) {
-  return Math.floor((logs.get(chatId)?.length ?? 0) / 2);
-}
-var logs, MAX_PAIRS;
+var CACHE_SIZE, cache;
 var init_session_log = __esm({
   "src/memory/session-log.ts"() {
     "use strict";
-    logs = /* @__PURE__ */ new Map();
-    MAX_PAIRS = 100;
+    init_store4();
+    CACHE_SIZE = 20;
+    cache = /* @__PURE__ */ new Map();
   }
 });
 
 // src/memory/summarize.ts
 var summarize_exports = {};
 __export(summarize_exports, {
+  CONTEXT_BRIDGE_PROMPT: () => CONTEXT_BRIDGE_PROMPT,
   summarizeAllPending: () => summarizeAllPending,
-  summarizeSession: () => summarizeSession
+  summarizeSession: () => summarizeSession,
+  summarizeWithFallbackChain: () => summarizeWithFallbackChain
 });
 import { spawn } from "child_process";
 import { createInterface } from "readline";
-function resolveSummarizerAdapter(chatId) {
-  const config2 = getSummarizer(chatId);
-  if (config2.backend === "off") return null;
-  let adapter;
-  if (config2.backend) {
-    adapter = getAdapter(config2.backend);
-  } else {
-    try {
-      adapter = getAdapterForChat(chatId);
-    } catch {
-      try {
-        adapter = getAdapter("claude");
-      } catch {
-        return null;
-      }
-    }
-  }
-  const model2 = config2.model ?? adapter.summarizerModel;
-  return { adapter, model: model2 };
-}
 function buildTranscript(entries) {
   const lines = [];
   let totalLen = 0;
@@ -3373,19 +3578,7 @@ function buildTranscript(entries) {
   }
   return lines.join("\n\n");
 }
-async function summarizeSession(chatId) {
-  const pairCount = getMessagePairCount(chatId);
-  if (pairCount < MIN_PAIRS) {
-    clearLog(chatId);
-    return false;
-  }
-  const resolved = resolveSummarizerAdapter(chatId);
-  if (!resolved) {
-    clearLog(chatId);
-    return false;
-  }
-  const { adapter, model: model2 } = resolved;
-  const entries = getLog(chatId);
+async function attemptSummarize(chatId, adapter, model2, entries) {
   const transcript = buildTranscript(entries);
   const prompt = `${SUMMARIZE_PROMPT}
 
@@ -3404,7 +3597,6 @@ ${transcript}`;
     let inputTokens = 0;
     let outputTokens = 0;
     let cacheReadTokens = 0;
-    const SUMMARIZE_TIMEOUT_MS = 6e4;
     await new Promise((resolve) => {
       const proc = spawn(config2.executable, config2.args, {
         env,
@@ -3414,7 +3606,7 @@ ${transcript}`;
       proc.stderr?.resume();
       const rl2 = createInterface({ input: proc.stdout });
       const timeout = setTimeout(() => {
-        warn(`[summarize] Timeout after ${SUMMARIZE_TIMEOUT_MS / 1e3}s for chat ${chatId} \u2014 killing process`);
+        warn(`[summarize] Timeout (${adapter.id}:${model2}) for chat ${chatId}`);
         rl2.close();
         proc.kill("SIGTERM");
         setTimeout(() => proc.kill("SIGKILL"), 2e3);
@@ -3429,9 +3621,7 @@ ${transcript}`;
         }
         const events = adapter.parseLine(msg);
         for (const ev of events) {
-          if (ev.type === "text" && ev.text) {
-            accumulatedText += ev.text;
-          }
+          if (ev.type === "text" && ev.text) accumulatedText += ev.text;
           if (ev.type === "usage" && ev.usage) {
             inputTokens += ev.usage.input;
             outputTokens += ev.usage.output;
@@ -3463,11 +3653,9 @@ ${transcript}`;
     if (inputTokens + outputTokens > 0) {
       addUsage(chatId, inputTokens, outputTokens, cacheReadTokens, model2);
     }
+    if (!resultText) resultText = accumulatedText;
     if (!resultText) {
-      resultText = accumulatedText;
-    }
-    if (!resultText) {
-      warn(`[summarize] ${adapter.displayName} returned empty result for chat ${chatId} \u2014 keeping log for retry`);
+      warn(`[summarize] ${adapter.id}:${model2} returned empty result for chat ${chatId}`);
       return false;
     }
     const summaryMatch = resultText.match(/SUMMARY:\s*(.+?)(?=\nKEY_DETAILS:|\nTOPICS:|$)/s);
@@ -3475,20 +3663,72 @@ ${transcript}`;
     const topicsMatch = resultText.match(/TOPICS:\s*(.+)/);
     let summary = summaryMatch?.[1]?.trim() ?? resultText.trim();
     const keyDetails = keyDetailsMatch?.[1]?.trim();
-    if (keyDetails) {
-      summary += `
+    if (keyDetails) summary += `
 Key details: ${keyDetails}`;
-    }
     const topics = topicsMatch?.[1]?.trim() ?? "";
-    saveSessionSummaryWithEmbedding(chatId, summary, topics, pairCount);
-    log(`[summarize] Saved session summary for chat ${chatId} via ${adapter.id}:${model2} (${pairCount} pairs, topics: ${topics})`);
-    clearLog(chatId);
+    saveSessionSummaryWithEmbedding(chatId, summary, topics, Math.floor(entries.length / 2));
+    log(`[summarize] Saved summary via ${adapter.id}:${model2} for chat ${chatId}`);
     return true;
   } catch (err) {
-    warn(`[summarize] Failed for chat ${chatId}: ${errorMessage(err)} \u2014 keeping log for retry`);
+    warn(`[summarize] ${adapter.id}:${model2} failed for ${chatId}: ${errorMessage(err)}`);
     return false;
   }
 }
+async function summarizeWithFallbackChain(chatId, targetBackendId) {
+  const pairCount = getMessagePairCount(chatId);
+  if (pairCount < MIN_PAIRS) {
+    clearLog(chatId);
+    return false;
+  }
+  const entries = getLog(chatId);
+  if (entries.length === 0) return false;
+  const tried = /* @__PURE__ */ new Set();
+  try {
+    const config2 = getSummarizer(chatId);
+    if (config2.backend !== "off") {
+      const adapter = config2.backend ? getAdapter(config2.backend) : getAdapterForChat(chatId);
+      const model2 = config2.model ?? adapter.summarizerModel;
+      const key = `${adapter.id}:${model2}`;
+      tried.add(key);
+      if (await attemptSummarize(chatId, adapter, model2, entries)) {
+        clearLog(chatId);
+        return true;
+      }
+    }
+  } catch {
+  }
+  if (targetBackendId) {
+    try {
+      const targetAdapter = getAdapter(targetBackendId);
+      const model2 = targetAdapter.summarizerModel;
+      const key = `${targetAdapter.id}:${model2}`;
+      if (!tried.has(key)) {
+        tried.add(key);
+        if (await attemptSummarize(chatId, targetAdapter, model2, entries)) {
+          clearLog(chatId);
+          return true;
+        }
+      }
+    } catch {
+    }
+  }
+  for (const adapter of getAllAdapters()) {
+    const model2 = adapter.summarizerModel;
+    const key = `${adapter.id}:${model2}`;
+    if (!tried.has(key)) {
+      tried.add(key);
+      if (await attemptSummarize(chatId, adapter, model2, entries)) {
+        clearLog(chatId);
+        return true;
+      }
+    }
+  }
+  warn(`[summarize] All fallback attempts failed for chat ${chatId} \u2014 raw log preserved`);
+  return false;
+}
+async function summarizeSession(chatId) {
+  return summarizeWithFallbackChain(chatId);
+}
 async function summarizeAllPending() {
   const chatIds = getLoggedChatIds();
   if (chatIds.length === 0) return;
@@ -3497,7 +3737,7 @@ async function summarizeAllPending() {
     await summarizeSession(chatId);
   }
 }
-var MIN_PAIRS, USER_MSG_LIMIT, AGENT_MSG_LIMIT, TRANSCRIPT_CAP, SUMMARIZE_PROMPT;
+var MIN_PAIRS, USER_MSG_LIMIT, AGENT_MSG_LIMIT, TRANSCRIPT_CAP, SUMMARIZE_TIMEOUT_MS, SUMMARIZE_PROMPT, CONTEXT_BRIDGE_PROMPT;
 var init_summarize = __esm({
   "src/memory/summarize.ts"() {
     "use strict";
@@ -3509,18 +3749,20 @@ var init_summarize = __esm({
     USER_MSG_LIMIT = 4e3;
     AGENT_MSG_LIMIT = 6e3;
     TRANSCRIPT_CAP = 1e5;
+    SUMMARIZE_TIMEOUT_MS = 6e4;
     SUMMARIZE_PROMPT = `You are summarizing a conversation session for long-term episodic memory. This summary will be injected into future conversations to provide context, so it must contain enough specific detail to be useful weeks or months later.
 
 Instructions:
 1. SUMMARY: Write a detailed summary (5-15 sentences depending on session complexity). Capture:
    - What the user asked for and why
    - What was done, decided, or accomplished
-   - Specific details that would be hard to rediscover: names, dates, values, locations, references, sources, outcomes
-   - Anything left unfinished or planned for later
+   - Specific details that would be hard to rediscover: names, dates, values, locations, references, sources, decisions, outcomes \u2014 across any domain
+   - The state at session end: what's done, what's in progress, what's planned next
    - Any preferences, corrections, or instructions the user expressed
 
 2. KEY_DETAILS: List the 3-8 most important specific facts as bullet points. These should be concrete and retrievable \u2014 not vague.
    Good: "User's quarterly report deadline is March 15"
+   Good: "Flight booked to NYC on April 3, confirmation #AB1234"
    Bad: "Discussed deadlines"
 
 3. TOPICS: Extract 5-10 specific topic keywords that would help retrieve this session in a future search. Use proper nouns, specific terms, and distinguishing phrases \u2014 NOT generic words like "discussion" or "help".
@@ -3531,6 +3773,14 @@ KEY_DETAILS:
 - <detail 1>
 - <detail 2>
 TOPICS: <specific-keyword1, specific-keyword2, ...>`;
+    CONTEXT_BRIDGE_PROMPT = `You are handing off an active work session to a new AI assistant. Write a concise handoff brief (3-8 sentences) covering:
+- What the user is actively working on RIGHT NOW
+- Exact current state: what's done, what's in progress, what's blocked
+- Specific details needed to continue: names, dates, decisions, references, links, preferences, or any domain-specific context relevant to the work
+- The immediate next step the assistant should be ready to take
+- Any critical constraints or preferences expressed by the user
+
+Be specific. The new assistant has no prior context. Cover any domain \u2014 personal, research, scheduling, content, technical \u2014 whatever applies to this session.`;
   }
 });
 
@@ -4448,7 +4698,7 @@ function startNextQueued(chatId) {
     description: agent.description ?? void 0,
     model: agent.model ?? void 0,
     skills: JSON.parse(agent.skills),
-    permMode: agent.permMode === "inherit" ? void 0 : agent.permMode,
+    permMode: agent.permMode === "inherit" ? getMode(chatId) : agent.permMode,
     role: agent.role ?? "worker",
     persona: agent.persona ?? void 0,
     maxRuntimeMs: agent.maxRuntimeMs,
@@ -4954,6 +5204,10 @@ function startDashboard() {
         if (body.cwd && !existsSync8(body.cwd)) {
           return jsonResponse(res, { error: `Directory not found: ${body.cwd}` }, 400);
         }
+        if (!body.permMode || body.permMode === "inherit") {
+          const { getMode: getMode2 } = await Promise.resolve().then(() => (init_store4(), store_exports3));
+          body.permMode = getMode2(body.chatId);
+        }
         const result = await spawnSubAgent(body.chatId, body);
         return jsonResponse(res, result);
       } catch (err) {
@@ -5206,7 +5460,10 @@ function startDashboard() {
         const { getMode: getMode2, getCwd: getCwd2, getModel: getModel2, addUsage: addUsage2, getBackend: getBackend2 } = await Promise.resolve().then(() => (init_store4(), store_exports3));
         const { getAdapterForChat: getAdapterForChat2 } = await Promise.resolve().then(() => (init_backends(), backends_exports));
         const chatId = body.chatId;
-        const mode = body.mode ?? getMode2(chatId);
+        const PERM_LEVEL = { plan: 0, safe: 1, yolo: 2 };
+        const storedMode = getMode2(chatId);
+        const requestedMode = body.mode ?? storedMode;
+        const mode = (PERM_LEVEL[requestedMode] ?? 2) <= (PERM_LEVEL[storedMode] ?? 2) ? requestedMode : storedMode;
         const cwd = body.cwd ?? getCwd2(chatId);
         const model2 = body.model ?? getModel2(chatId) ?? (() => {
           try {
@@ -5858,11 +6115,11 @@ function askAgent(chatId, userMessage, opts) {
 async function askAgentImpl(chatId, userMessage, opts) {
   const { cwd, onStream, model: model2, backend: backend2, permMode, onToolAction, bootstrapTier, timeoutMs } = opts ?? {};
   const adapter = backend2 ? getAdapter(backend2) : getAdapterForChat(chatId);
-  const mode = permMode ?? "yolo";
+  const mode = permMode ?? getMode(chatId);
   const thinkingLevel = getThinkingLevel(chatId);
   const resolvedCwd = cwd ?? WORKSPACE_PATH;
   const tier = bootstrapTier ?? "full";
-  const fullPrompt = await assembleBootstrapPrompt(userMessage, tier, chatId);
+  const fullPrompt = await assembleBootstrapPrompt(userMessage, tier, chatId, mode);
   const existingSessionId = getSessionId(chatId);
   const allowedTools = getEnabledTools(chatId);
   const mcpConfigPath = getMcpConfigPath(chatId);
@@ -5923,12 +6180,14 @@ async function askAgentImpl(chatId, userMessage, opts) {
     clearSession(chatId);
   }
   if (result.resultText) {
-    appendToLog(chatId, userMessage, result.resultText);
+    appendToLog(chatId, userMessage, result.resultText, adapter.id, model2 ?? null, result.sessionId ?? null);
     const AUTO_SUMMARIZE_THRESHOLD = 30;
     const pairCount = getMessagePairCount(chatId);
     if (pairCount >= AUTO_SUMMARIZE_THRESHOLD) {
       log(`[agent] Auto-summarizing chat ${chatId} after ${pairCount} turns`);
-      summarizeSession(chatId).catch((err) => {
+      summarizeWithFallbackChain(chatId).then((saved) => {
+        if (saved) opts?.onCompaction?.(chatId);
+      }).catch((err) => {
         warn(`[agent] Auto-summarize failed for chat ${chatId}: ${err}`);
       });
     }
@@ -6488,7 +6747,8 @@ async function runWithRetry(job, model2, runId, t0) {
           model: currentModel,
           backend: currentBackend,
           bootstrapTier: "slim",
-          timeoutMs
+          timeoutMs,
+          permMode: job.sessionType === "main" ? getMode(job.chatId) : "yolo"
         });
         if (isFallback) {
           response.text = `[Fallback: ran on ${currentBackend}:${currentModel}]
@@ -6616,7 +6876,7 @@ function wrapBackendAdapter(adapter) {
       const config2 = adapter.buildSpawnConfig({
         prompt: opts.prompt,
         sessionId: opts.sessionId,
-        permMode: opts.permMode ?? "yolo",
+        permMode: opts.permMode ?? "plan",
         allowedTools: opts.allowedTools ?? [],
         cwd: opts.cwd
       });
@@ -8063,14 +8323,6 @@ var init_heartbeat = __esm({
 });
 
 // src/format-time.ts
-function formatLocalDate(utcDatetime) {
-  const d = parseUtcDatetime(utcDatetime);
-  if (!d) return utcDatetime.split("T")[0] ?? utcDatetime.split(" ")[0];
-  const year = d.getFullYear();
-  const month = String(d.getMonth() + 1).padStart(2, "0");
-  const day = String(d.getDate()).padStart(2, "0");
-  return `${year}-${month}-${day}`;
-}
 function formatLocalDateTime(utcDatetime) {
   const d = parseUtcDatetime(utcDatetime);
   if (!d) return utcDatetime;
@@ -9253,7 +9505,7 @@ async function handleCommand(msg, channel) {
     case "help":
       await channel.sendText(
         chatId,
-        "Hey! I'm CC-Claw \u2014 your personal AI assistant on Telegram.\n\nI use AI coding CLIs (Claude, Gemini, Codex) as my brain. Just send me a message to get started.\n\nCommands:\n/backend [name] - Switch AI backend (or /claude /gemini /codex)\n/model - Switch model for active backend\n/summarizer - Configure session summarization model\n/status - Show session, model, backend, and usage\n/cost - Show estimated API cost (use /cost all for all-time)\n/usage - Show usage per backend with limits\n/limits - Configure usage limits per backend\n/newchat - Start a fresh conversation\n/summarize - Save session to memory (without resetting)\n/summarize all - Summarize all pending sessions (pre-restart)\n/cwd <path> - Set working directory\n/cwd - Show current working directory\n/memory - List stored memories\n/forget <keyword> - Remove a memory\n/voice - Toggle voice responses\n/imagine <prompt> - Generate an image (or /image)\n/cron <description> - Schedule a task (or /schedule)\n/cron - List scheduled jobs (or /jobs)\n/cron cancel <id> - Cancel a job\n/cron pause <id> - Pause a job\n/cron resume <id> - Resume a job\n/cron run <id> - Trigger a job now\n/cron runs [id] - View run history\n/cron edit <id> - Edit a job\n/cron health - Scheduler health\n/skills - List skills from all backends\n/skill-install <url> - Install a skill from GitHub\n/setup-profile - Set up your user profile\n/chats - List authorized chats and aliases\n/heartbeat - Proactive awareness (on/off/interval/hours)\n/history - List recent session summaries\n/stop - Cancel the current running task\n/tools - Configure which tools the agent can use\n/permissions - Switch permission mode (yolo/safe/readonly/plan)\n/verbose - Tool visibility (off/normal/verbose)\n/agents - List active sub-agents\n/tasks - Show task board for current orchestration\n/stopagent <id> - Cancel a specific sub-agent\n/stopall - Cancel all sub-agents in this chat\n/runners - List registered CLI runners\n/mcps - List registered MCP servers\n/help - Show this message",
+        "Hey! I'm CC-Claw \u2014 your personal AI assistant on Telegram.\n\nI use AI coding CLIs (Claude, Gemini, Codex) as my brain. Just send me a message to get started.\n\nCommands:\n/backend [name] - Switch AI backend (or /claude /gemini /codex)\n/model - Switch model for active backend\n/summarizer - Configure session summarization model\n/status - Show session, model, backend, and usage\n/cost - Show estimated API cost (use /cost all for all-time)\n/usage - Show usage per backend with limits\n/limits - Configure usage limits per backend\n/newchat - Start a fresh conversation\n/summarize - Save session to memory (without resetting)\n/summarize all - Summarize all pending sessions (pre-restart)\n/cwd <path> - Set working directory\n/cwd - Show current working directory\n/memory - List stored memories\n/forget <keyword> - Remove a memory\n/voice - Toggle voice responses\n/imagine <prompt> - Generate an image (or /image)\n/cron <description> - Schedule a task (or /schedule)\n/cron - List scheduled jobs (or /jobs)\n/cron cancel <id> - Cancel a job\n/cron pause <id> - Pause a job\n/cron resume <id> - Resume a job\n/cron run <id> - Trigger a job now\n/cron runs [id] - View run history\n/cron edit <id> - Edit a job\n/cron health - Scheduler health\n/skills - List skills from all backends\n/skill-install <url> - Install a skill from GitHub\n/setup-profile - Set up your user profile\n/chats - List authorized chats and aliases\n/heartbeat - Proactive awareness (on/off/interval/hours)\n/history - List recent session summaries\n/stop - Cancel the current running task\n/tools - Configure which tools the agent can use\n/permissions - Switch permission mode (yolo/safe/plan)\n/verbose - Tool visibility (off/normal/verbose)\n/agents - List active sub-agents\n/tasks - Show task board for current orchestration\n/stopagent <id> - Cancel a specific sub-agent\n/stopall - Cancel all sub-agents in this chat\n/runners - List registered CLI runners\n/mcps - List registered MCP servers\n/help - Show this message",
         "plain"
       );
       break;
@@ -9313,7 +9565,7 @@ async function handleCommand(msg, channel) {
           toolButtons.push(row);
         }
         toolButtons.push([{ label: "Reset to defaults (all on)", data: "tool:reset" }]);
-        const modeNote = currentMode === "readonly" ? "\n\nNote: In read-only mode, tool list is ignored (Read/Glob/Grep only)." : currentMode === "plan" ? "\n\nNote: In plan mode, no tools execute." : currentMode === "yolo" ? "\n\nNote: In YOLO mode, tool list is ignored (all tools allowed)." : "";
+        const modeNote = currentMode === "plan" ? "\n\nNote: In plan mode, tool list is ignored (read-only)." : currentMode === "yolo" ? "\n\nNote: In YOLO mode, tool list is ignored (all tools allowed)." : "";
         await channel.sendKeyboard(
           chatId,
           `Configure allowed tools (mode: ${currentMode})${modeNote}
@@ -9349,7 +9601,15 @@ Tap to toggle:`,
       } else {
         const pairs = getMessagePairCount(chatId);
         if (pairs < 2) {
-          await channel.sendText(chatId, "Not enough conversation to summarize (need at least 2 exchanges). Session log is preserved.", "plain");
+          if (getSessionId(chatId)) {
+            await channel.sendText(
+              chatId,
+              "Session log was cleared (auto-compact or service restart). Your conversation history is preserved in episodic memory. Continue chatting normally, or use /newchat to start fresh.",
+              "plain"
+            );
+          } else {
+            await channel.sendText(chatId, "Not enough conversation to summarize (need at least 2 exchanges).", "plain");
+          }
           break;
         }
         await channel.sendText(chatId, `Summarizing current session (${pairs} exchanges)...`, "plain");
@@ -9882,20 +10142,38 @@ ${lines.join("\n")}`, "plain");
       break;
     }
     case "history": {
-      const historyLimit = commandArgs ? parseInt(commandArgs, 10) || 5 : 5;
-      const summaries = listSessionSummaries(historyLimit, chatId);
-      if (summaries.length === 0) {
-        await channel.sendText(chatId, "No session history yet. History is saved automatically when you use /newchat.", "plain");
-        return;
+      const query = commandArgs?.trim() ?? "";
+      if (query) {
+        const results = searchMessageLog(chatId, query, 20);
+        if (results.length === 0) {
+          await channel.sendText(chatId, `No matching history found. Try /memories ${query} for older sessions.`, "plain");
+          break;
+        }
+        const lines = [`\u{1F4CB} History: "${query}"`, "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501"];
+        for (const row of results.slice(0, 10)) {
+          const roleLabel = row.role === "user" ? "You" : "Assistant";
+          const date = row.created_at.slice(0, 10);
+          const preview = row.content.slice(0, 300) + (row.content.length > 300 ? "\u2026" : "");
+          lines.push(`[${row.backend ?? "?"} \xB7 ${date}] ${roleLabel}: ${preview}`);
+        }
+        lines.push("\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501");
+        await channel.sendText(chatId, lines.join("\n"), "plain");
+      } else {
+        const rows = getRecentMessageLog(chatId, 20).reverse();
+        if (rows.length === 0) {
+          await channel.sendText(chatId, "No conversation history found.", "plain");
+          break;
+        }
+        const lines = ["\u{1F4CB} Recent conversation (last 10 exchanges):", "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501"];
+        for (const row of rows) {
+          const roleLabel = row.role === "user" ? "You" : "Assistant";
+          const date = row.created_at.slice(0, 10);
+          const preview = row.content.slice(0, 300) + (row.content.length > 300 ? "\u2026" : "");
+          lines.push(`[${row.backend ?? "?"} \xB7 ${date}] ${roleLabel}: ${preview}`);
+        }
+        lines.push("\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501");
+        await channel.sendText(chatId, lines.join("\n"), "plain");
       }
-      const lines = summaries.map((s) => {
-        const date = formatLocalDate(s.created_at);
-        const shortSummary = s.summary.length > 200 ? s.summary.slice(0, 200) + "\u2026" : s.summary;
-        return `${date} (${s.message_count} msgs)
-${shortSummary}
-Topics: ${s.topics}`;
-      });
-      await channel.sendText(chatId, lines.join("\n\n"), "plain");
       break;
     }
     case "skills": {
@@ -10516,7 +10794,7 @@ async function handleText(msg, channel) {
     await channel.sendText(chatId, limitMsg, "plain");
     return;
   }
-  if (isChatBusy(chatId)) {
+  if (isChatBusy(chatId) && !bypassBusyCheck.delete(chatId)) {
     if (typeof channel.sendKeyboard === "function") {
       pendingInterrupts.set(chatId, { msg, channel });
       await channel.sendKeyboard(chatId, "\u23F3 Agent is working on a request\u2026", [
@@ -10544,8 +10822,29 @@ async function handleText(msg, channel) {
     const tMode = getMode(chatId);
     const tVerbose = getVerboseLevel(chatId);
     const tToolCb = tVerbose !== "off" ? makeToolActionCallback(chatId, channel, tVerbose) : void 0;
-    const response = await askAgent(chatId, text, { cwd: getCwd(chatId), model: model2, permMode: tMode, onToolAction: tToolCb });
+    const response = await askAgent(chatId, text, {
+      cwd: getCwd(chatId),
+      model: model2,
+      permMode: tMode,
+      onToolAction: tToolCb,
+      onCompaction: (cid) => {
+        channel.sendText(cid, "\u{1F4BE} Context saved to memory.").catch(() => {
+        });
+      }
+    });
     if (response.usage) addUsage(chatId, response.usage.input, response.usage.output, response.usage.cacheRead, model2);
+    if (response.text.includes("[NEED_PERMISSION]") && tMode !== "yolo") {
+      const cleanText = response.text.replace(/\[NEED_PERMISSION\]/g, "").trim();
+      await sendResponse(chatId, channel, cleanText, msg.messageId);
+      const targetMode = determineEscalationTarget(chatId, tMode);
+      storePendingEscalation(chatId, text);
+      if (typeof channel.sendKeyboard === "function") {
+        await channel.sendKeyboard(chatId, `Switch to ${targetMode} mode?`, [
+          [{ label: "Allow", data: `perm:escalate:${targetMode}` }, { label: "Deny", data: "perm:deny" }]
+        ]);
+      }
+      return;
+    }
     await sendResponse(chatId, channel, response.text, msg.messageId);
   } catch (err) {
     error("[router] Error:", err);
@@ -10731,7 +11030,28 @@ async function sendResponse(chatId, channel, text, messageId) {
       ]);
     }
   }
-  const afterFiles = await processFileSends2(chatId, channel, afterUpdates);
+  let afterHistory = afterUpdates;
+  const historySearchMatch = afterHistory.match(/\[HISTORY_SEARCH:([^\]]+)\]/);
+  if (historySearchMatch) {
+    afterHistory = afterHistory.replace(/\[HISTORY_SEARCH:[^\]]+\]/g, "").trim();
+    const hsQuery = historySearchMatch[1].trim();
+    const hsResults = searchMessageLog(chatId, hsQuery, 10);
+    if (hsResults.length > 0) {
+      const hsLines = [`\u{1F4CB} History: "${hsQuery}"`, "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501"];
+      for (const row of hsResults.slice(0, 8)) {
+        const label2 = row.role === "user" ? "You" : "Assistant";
+        const date = row.created_at.slice(0, 10);
+        hsLines.push(`[${row.backend ?? "?"} \xB7 ${date}] ${label2}: ${row.content.slice(0, 250)}${row.content.length > 250 ? "\u2026" : ""}`);
+      }
+      hsLines.push("\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501");
+      channel.sendText(chatId, hsLines.join("\n")).catch(() => {
+      });
+    } else {
+      channel.sendText(chatId, `No history found for "${hsQuery}". Try /memories ${hsQuery} for older sessions.`).catch(() => {
+      });
+    }
+  }
+  const afterFiles = await processFileSends2(chatId, channel, afterHistory);
   const cleanText = await processImageGenerations(chatId, channel, afterFiles);
   if (!cleanText) return;
   if (isVoiceEnabled(chatId)) {
@@ -10757,16 +11077,17 @@ async function sendBackendSwitchConfirmation(chatId, target, channel) {
     await channel.sendText(chatId, `Already using ${targetAdapter.displayName}.`, "plain");
     return;
   }
-  const currentLabel = current ? getAdapter(current).displayName : "current backend";
   if (typeof channel.sendKeyboard === "function") {
     await channel.sendKeyboard(
       chatId,
-      `\u26A0\uFE0F Switching to ${targetAdapter.displayName} will summarize and reset your current session.
-
-What would you like to do?`,
+      `\u{1F504} Switch to ${targetAdapter.displayName}?
+Your conversation history is preserved. ${targetAdapter.displayName} will receive a summary of recent context and can access your full history on request.`,
       [
-        [{ label: `Stay on ${currentLabel}`, data: "backend_cancel" }],
-        [{ label: `Switch to ${targetAdapter.displayName} + summarize`, data: `backend_confirm:${target}` }]
+        [
+          { label: `Switch to ${targetAdapter.displayName}`, data: `backend_confirm:${target}` },
+          { label: `Switch + Clear History`, data: `backend_confirm_clear:${target}` }
+        ],
+        [{ label: "Cancel", data: `backend_cancel:${target}` }]
       ]
     );
   } else {
@@ -10774,21 +11095,21 @@ What would you like to do?`,
   }
 }
 async function doBackendSwitch(chatId, backendId, channel) {
-  summarizeSession(chatId).catch(() => {
-  });
+  const targetAdapter = getAdapter(backendId);
+  const summarized = await summarizeWithFallbackChain(chatId, backendId);
+  if (summarized) {
+    await channel.sendText(chatId, `\u{1F4BE} Context saved \u2014 ${targetAdapter.displayName} summarized your session.`, "plain");
+  }
+  const bridge = buildContextBridge(chatId);
+  if (bridge) {
+    setPendingContextBridge(chatId, bridge);
+  }
   clearSession(chatId);
   clearModel(chatId);
   clearThinkingLevel(chatId);
   setBackend(chatId, backendId);
-  const adapter = getAdapter(backendId);
-  logActivity(getDb(), { chatId, source: "telegram", eventType: "config_changed", summary: `Backend switched to ${adapter.displayName}`, detail: { field: "backend", value: backendId } });
-  await channel.sendText(
-    chatId,
-    `Backend switched to ${adapter.displayName}.
-Default model: ${adapter.defaultModel}
-Session reset. Ready!`,
-    "plain"
-  );
+  logActivity(getDb(), { chatId, source: "telegram", eventType: "config_changed", summary: `Backend switched to ${targetAdapter.displayName}`, detail: { field: "backend", value: backendId } });
+  await channel.sendText(chatId, `\u2705 Switched to ${targetAdapter.displayName}. Ready!`, "plain");
 }
 async function handleCallback(chatId, data, channel) {
   if (data.startsWith("backend:")) {
@@ -10801,14 +11122,17 @@ async function handleCallback(chatId, data, channel) {
       return;
     }
     await sendBackendSwitchConfirmation(chatId, chosen, channel);
+  } else if (data.startsWith("backend_confirm_clear:")) {
+    const target = data.slice(21);
+    if (!getAllBackendIds().includes(target)) return;
+    clearMessageLog(chatId);
+    await doBackendSwitch(chatId, target, channel);
   } else if (data.startsWith("backend_confirm:")) {
     const chosen = data.slice(16);
     if (!getAllBackendIds().includes(chosen)) return;
     await doBackendSwitch(chatId, chosen, channel);
-  } else if (data === "backend_cancel") {
-    const current = getBackend(chatId);
-    const label2 = current ? getAdapter(current).displayName : "current backend";
-    await channel.sendText(chatId, `No change. Staying on ${label2}.`, "plain");
+  } else if (data.startsWith("backend_cancel:") || data === "backend_cancel") {
+    await channel.sendText(chatId, "Switch cancelled.", "plain");
   } else if (data.startsWith("model:")) {
     const chosen = data.slice(6);
     let adapter;
@@ -10863,7 +11187,8 @@ Select thinking/effort level:`,
       await channel.sendText(chatId, `Summarizer pinned to ${bk}:${mdl ?? "default"}.`, "plain");
     }
   } else if (data.startsWith("perms:")) {
-    const chosen = data.slice(6);
+    let chosen = data.slice(6);
+    if (chosen === "readonly") chosen = "plan";
     if (!PERM_MODES[chosen]) return;
     const previous = getMode(chatId);
     if (chosen === previous) {
@@ -10878,6 +11203,20 @@ Select thinking/effort level:`,
 ${PERM_MODES[chosen]}`,
       "plain"
     );
+  } else if (data.startsWith("perm:escalate:")) {
+    const targetMode = data.slice(14);
+    if (!PERM_MODES[targetMode]) return;
+    setMode(chatId, targetMode);
+    logActivity(getDb(), { chatId, source: "telegram", eventType: "config_changed", summary: `Permission escalated to ${targetMode}`, detail: { field: "permissions", value: targetMode } });
+    await channel.sendText(chatId, `Switched to ${targetMode} mode. Re-processing your request...`, "plain");
+    const pending = getPendingEscalation(chatId);
+    if (pending) {
+      removePendingEscalation(chatId);
+      await handleMessage({ text: pending, chatId, source: "telegram" }, channel);
+    }
+  } else if (data === "perm:deny") {
+    removePendingEscalation(chatId);
+    await channel.sendText(chatId, "Keeping current mode.", "plain");
   } else if (data.startsWith("verbose:")) {
     const chosen = data.slice(8);
     if (!VERBOSE_LEVELS[chosen]) return;
@@ -10965,6 +11304,7 @@ ${PERM_MODES[chosen]}`,
       await handleMessage(pending.msg, pending.channel);
     } else if (action === "queue" && pending) {
       pendingInterrupts.delete(targetChatId);
+      bypassBusyCheck.add(targetChatId);
       await channel.sendText(chatId, "\u{1F4E5} Message queued \u2014 will process after current task.", "plain");
       handleMessage(pending.msg, pending.channel).catch(() => {
       });
@@ -10981,8 +11321,9 @@ ${PERM_MODES[chosen]}`,
       await channel.sendText(chatId, "OK \u2014 you can switch manually with /backend when ready.", "plain");
     } else if (pendingMsg) {
       pendingFallbackMessages.delete(targetChatId);
-      const { summarizeSession: summarizeSession2 } = await Promise.resolve().then(() => (init_summarize(), summarize_exports));
-      await summarizeSession2(targetChatId);
+      await summarizeWithFallbackChain(targetChatId, targetBackend);
+      const bridge = buildContextBridge(targetChatId);
+      if (bridge) setPendingContextBridge(targetChatId, bridge);
       clearSession(targetChatId);
       setBackend(targetChatId, targetBackend);
       const adapter = getAdapter(targetBackend);
@@ -11202,7 +11543,7 @@ Use /skills <page> to navigate (e.g. /skills 2)` : "";
   const header2 = totalPages > 1 ? `${skills2.length} skills (page ${safePage}/${totalPages}). Select one to invoke:` : `${skills2.length} skills available. Select one to invoke:`;
   await channel.sendKeyboard(chatId, header2, buttons);
 }
-var PERM_MODES, VERBOSE_LEVELS, pendingInterrupts, pendingFallbackMessages, CLI_INSTALL_HINTS, BLOCKED_PATH_PATTERNS2, ALLOWED_REACTION_EMOJIS, SKILLS_PER_PAGE;
+var PERM_MODES, VERBOSE_LEVELS, pendingInterrupts, bypassBusyCheck, pendingFallbackMessages, CLI_INSTALL_HINTS, BLOCKED_PATH_PATTERNS2, ALLOWED_REACTION_EMOJIS, SKILLS_PER_PAGE;
 var init_router = __esm({
   "src/router.ts"() {
     "use strict";
@@ -11218,6 +11559,8 @@ var init_router = __esm({
     init_stt();
     init_store4();
     init_summarize();
+    init_inject();
+    init_store4();
     init_session_log();
     init_backends();
     init_cron();
@@ -11235,8 +11578,7 @@ var init_router = __esm({
     PERM_MODES = {
       yolo: "YOLO \u2014 all tools, full autopilot",
       safe: "Safe \u2014 only my allowed tools",
-      readonly: "Read-only \u2014 Read, Glob, Grep only",
-      plan: "Plan \u2014 plan only, no execution"
+      plan: "Plan \u2014 read and analyze only"
     };
     VERBOSE_LEVELS = {
       off: "Off \u2014 no tool visibility",
@@ -11244,6 +11586,7 @@ var init_router = __esm({
       verbose: "Verbose \u2014 full details"
     };
     pendingInterrupts = /* @__PURE__ */ new Map();
+    bypassBusyCheck = /* @__PURE__ */ new Set();
     pendingFallbackMessages = /* @__PURE__ */ new Map();
     CLI_INSTALL_HINTS = {
       claude: "Install: npm install -g @anthropic-ai/claude-code",
@@ -11980,7 +12323,7 @@ var index_exports = {};
 __export(index_exports, {
   main: () => main
 });
-import { mkdirSync as mkdirSync7, existsSync as existsSync16, renameSync, statSync as statSync2 } from "fs";
+import { mkdirSync as mkdirSync7, existsSync as existsSync16, renameSync, statSync as statSync2, readFileSync as readFileSync10 } from "fs";
 import { join as join17 } from "path";
 import dotenv from "dotenv";
 function migrateLayout() {
@@ -12020,14 +12363,21 @@ function rotateLogs() {
 }
 async function main() {
   rotateLogs();
-  log("[cc-claw] Starting up...");
+  let version = "unknown";
+  try {
+    const pkgPath = new URL("../package.json", import.meta.url);
+    version = JSON.parse(readFileSync10(pkgPath, "utf-8")).version;
+  } catch {
+  }
+  log(`[cc-claw] Starting v${version}`);
   initDatabase();
+  pruneMessageLog(30, 2e3);
   bootstrapBuiltinMcps(getDb());
-  const SUMMARIZE_TIMEOUT_MS = 3e4;
+  const SUMMARIZE_TIMEOUT_MS2 = 3e4;
   try {
     let timer;
     const timeoutPromise = new Promise((_, reject) => {
-      timer = setTimeout(() => reject(new Error("timeout")), SUMMARIZE_TIMEOUT_MS);
+      timer = setTimeout(() => reject(new Error("timeout")), SUMMARIZE_TIMEOUT_MS2);
     });
     try {
       await Promise.race([
@@ -12553,13 +12903,13 @@ var init_daemon = __esm({
 });
 
 // src/cli/resolve-chat.ts
-import { readFileSync as readFileSync11 } from "fs";
+import { readFileSync as readFileSync12 } from "fs";
 function resolveChatId(globalOpts) {
   const explicit = globalOpts.chat;
   if (explicit) return explicit;
   if (_cachedDefault) return _cachedDefault;
   try {
-    const content = readFileSync11(ENV_PATH, "utf-8");
+    const content = readFileSync12(ENV_PATH, "utf-8");
     const match = content.match(/^ALLOWED_CHAT_ID=(.+)$/m);
     if (match) {
       _cachedDefault = match[1].split(",")[0].trim();
@@ -12585,12 +12935,12 @@ __export(api_client_exports, {
   apiPost: () => apiPost,
   isDaemonRunning: () => isDaemonRunning
 });
-import { readFileSync as readFileSync12, existsSync as existsSync18 } from "fs";
+import { readFileSync as readFileSync13, existsSync as existsSync18 } from "fs";
 import { request as httpRequest } from "http";
 function getToken() {
   if (process.env.CC_CLAW_API_TOKEN) return process.env.CC_CLAW_API_TOKEN;
   try {
-    if (existsSync18(TOKEN_PATH)) return readFileSync12(TOKEN_PATH, "utf-8").trim();
+    if (existsSync18(TOKEN_PATH)) return readFileSync13(TOKEN_PATH, "utf-8").trim();
   } catch {
   }
   return null;
@@ -12915,8 +13265,8 @@ async function doctorCommand(globalOpts, localOpts) {
   }
   if (existsSync20(ERROR_LOG_PATH)) {
     try {
-      const { readFileSync: readFileSync17 } = await import("fs");
-      const logContent = readFileSync17(ERROR_LOG_PATH, "utf-8");
+      const { readFileSync: readFileSync18 } = await import("fs");
+      const logContent = readFileSync18(ERROR_LOG_PATH, "utf-8");
       const recentLines = logContent.split("\n").filter(Boolean).slice(-100);
       const last24h = Date.now() - 864e5;
       const recentErrors = recentLines.filter((line) => {
@@ -13032,7 +13382,7 @@ var logs_exports = {};
 __export(logs_exports, {
   logsCommand: () => logsCommand
 });
-import { existsSync as existsSync21, readFileSync as readFileSync13, watchFile as watchFile2, unwatchFile as unwatchFile2 } from "fs";
+import { existsSync as existsSync21, readFileSync as readFileSync14, watchFile as watchFile2, unwatchFile as unwatchFile2 } from "fs";
 async function logsCommand(opts) {
   const logFile = opts.error ? ERROR_LOG_PATH : LOG_PATH;
   if (!existsSync21(logFile)) {
@@ -13040,7 +13390,7 @@ async function logsCommand(opts) {
     process.exit(1);
   }
   const maxLines = parseInt(opts.lines ?? "100", 10);
-  const content = readFileSync13(logFile, "utf-8");
+  const content = readFileSync14(logFile, "utf-8");
   const allLines = content.split("\n");
   const tailLines = allLines.slice(-maxLines);
   console.log(muted(`  \u2500\u2500 ${logFile} (last ${tailLines.length} lines) \u2500\u2500`));
@@ -13050,7 +13400,7 @@ async function logsCommand(opts) {
     let lastSize = Buffer.byteLength(content, "utf-8");
     watchFile2(logFile, { interval: 500 }, () => {
       try {
-        const newContent = readFileSync13(logFile, "utf-8");
+        const newContent = readFileSync14(logFile, "utf-8");
         const newSize = Buffer.byteLength(newContent, "utf-8");
         if (newSize > lastSize) {
           const newPart = newContent.slice(content.length);
@@ -14055,7 +14405,7 @@ __export(config_exports, {
   configList: () => configList,
   configSet: () => configSet
 });
-import { existsSync as existsSync29, readFileSync as readFileSync14 } from "fs";
+import { existsSync as existsSync29, readFileSync as readFileSync15 } from "fs";
 async function configList(globalOpts) {
   if (!existsSync29(DB_PATH)) {
     outputError("DB_NOT_FOUND", "Database not found.");
@@ -14141,7 +14491,7 @@ async function configEnv(_globalOpts) {
     outputError("ENV_NOT_FOUND", `No .env file at ${ENV_PATH}. Run cc-claw setup.`);
     process.exit(1);
   }
-  const content = readFileSync14(ENV_PATH, "utf-8");
+  const content = readFileSync15(ENV_PATH, "utf-8");
   const entries = {};
   const secretPatterns = /TOKEN|KEY|SECRET|PASSWORD|CREDENTIALS/i;
   for (const line of content.split("\n")) {
@@ -14271,7 +14621,8 @@ async function permissionsGet(globalOpts) {
   output({ mode }, (d) => d.mode);
 }
 async function permissionsSet(globalOpts, mode) {
-  const valid = ["yolo", "safe", "readonly", "plan"];
+  if (mode === "readonly") mode = "plan";
+  const valid = ["yolo", "safe", "plan"];
   if (!valid.includes(mode)) {
     outputError("INVALID_MODE", `Invalid mode "${mode}". Valid: ${valid.join(", ")}`);
     process.exit(1);
@@ -14797,11 +15148,11 @@ __export(chat_exports, {
   chatSend: () => chatSend
 });
 import { request as httpRequest2 } from "http";
-import { readFileSync as readFileSync15, existsSync as existsSync37 } from "fs";
+import { readFileSync as readFileSync16, existsSync as existsSync37 } from "fs";
 function getToken2() {
   if (process.env.CC_CLAW_API_TOKEN) return process.env.CC_CLAW_API_TOKEN;
   try {
-    if (existsSync37(TOKEN_PATH2)) return readFileSync15(TOKEN_PATH2, "utf-8").trim();
+    if (existsSync37(TOKEN_PATH2)) return readFileSync16(TOKEN_PATH2, "utf-8").trim();
   } catch {
   }
   return null;
@@ -15227,7 +15578,7 @@ var init_completion = __esm({
 
 // src/setup.ts
 var setup_exports = {};
-import { existsSync as existsSync38, writeFileSync as writeFileSync8, readFileSync as readFileSync16, copyFileSync as copyFileSync2, mkdirSync as mkdirSync10, statSync as statSync6 } from "fs";
+import { existsSync as existsSync38, writeFileSync as writeFileSync8, readFileSync as readFileSync17, copyFileSync as copyFileSync2, mkdirSync as mkdirSync10, statSync as statSync6 } from "fs";
 import { execFileSync as execFileSync4 } from "child_process";
 import { createInterface as createInterface5 } from "readline";
 import { join as join19 } from "path";
@@ -15311,7 +15662,7 @@ async function setup() {
   if (envSource) {
     console.log(yellow(`  Found existing config at ${envSource} \u2014 your values will be preserved`));
     console.log(yellow("  unless you enter new ones. Just press Enter to keep existing values.\n"));
-    const existing = readFileSync16(envSource, "utf-8");
+    const existing = readFileSync17(envSource, "utf-8");
     for (const line of existing.split("\n")) {
       const match = line.match(/^([^#=]+)=(.*)$/);
       if (match) env[match[1].trim()] = match[2].trim();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cc-claw",
-  "version": "0.3.11",
+  "version": "0.4.1",
   "description": "CC-Claw: Personal AI assistant on Telegram — multi-backend (Claude, Gemini, Codex), sub-agent orchestration, MCP management",
   "type": "module",
   "main": "dist/cli.js",
@@ -39,7 +39,7 @@
   "license": "MIT",
   "dependencies": {
     "@modelcontextprotocol/sdk": "^1.27.1",
-    "better-sqlite3": "^12.6.2",
+    "better-sqlite3": "^12.8.0",
     "commander": "^14.0.3",
     "croner": "^9.0.0",
     "dotenv": "^16.4.7",
@@ -47,9 +47,6 @@
     "picocolors": "^1.1.1",
     "zod": "^4.3.6"
   },
-  "optionalDependencies": {
-    "groq-sdk": "^0.15.0"
-  },
   "devDependencies": {
     "@types/better-sqlite3": "^7.6.13",
     "@types/bun": "^1.3.10",