kiku-agent 1.2.0 → 1.3.0

package/server/agent/agentCore.js

@@ -1,6 +1,7 @@
 import { AgentStateMachine, AgentStates } from './agentState.js';
+import { buildSystemPrompt } from './prompts.js';
 import { config } from './config.js';
-import { contextCompactor } from './contextCompactor.js';
+import { contextCompactor, compactToolResultForContext, estimateTextTokens, estimateTokenCount } from './contextCompactor.js';
 import { createAbortError, isAbortError } from './errors.js';
 import { Events, eventBus } from './eventBus.js';
 import { AgentJobManager } from './jobManager.js';
@@ -12,15 +13,22 @@ import { sanitizeQuery, sanitizeResponse } from './sanitizer.js';
 import { AgentSession } from './session.js';
 import { ResearchSubagent } from './subagent.js';
 import { SuggestionEngine } from './suggestions.js';
+import { buildTextToolProtocolDirective, extractTopLevelJsonObjects, MALFORMED_TOOL_CALL_FEEDBACK, parseTextToolCallsFromContent, stripToolCallBlocks } from './textToolProtocol.js';
 import { DynamicToolRouter } from './toolRouter.js';
 import { executeTool, prepareToolArguments, toolsSchema, isBuiltinTool } from './tools.js';
 import { detectAmbiguities, isQueryTooVague, emitClarifyingQuestions, calculateConfidence, generateClarifyingQuestions, buildClarificationContext } from './clarifyingQuestions.js';
 import { ClarificationAnalytics, ClarificationConfig } from './clarificationAnalytics.js';
 import { calculateTokenCost } from './telemetry.js';
 import { extractCitations, resolveCitationReference } from './citations.js';
+import { resolveQueryEntities } from './catalogEntityResolver.js';
+import { TokenTracker } from './tokenTracker.js';
+import { RoutingInterceptor } from './routingInterceptor.js';
 
-export function detectDominantLanguage(text) {
-  if (!text || typeof text !== 'string') return 'th';
+
+const OUTPUT_TOKEN_BUDGET = Math.max(512, Number(process.env.KIKU_MAX_OUTPUT_TOKENS) || 4096);
+
+export function detectDominantLanguage(text, sessionLang = null) {
+  if (!text || typeof text !== 'string') return sessionLang || 'th';
 
   const clean = text
     .replace(/\[SYSTEM\s+[A-Z]+:[^\]]*\]/gi, '')
@@ -32,21 +40,19 @@ export function detectDominantLanguage(text) {
 
   const targetText = clean || text;
 
-  const thaiMatches = targetText.match(/[\u0E00-\u0E7F]/g) || [];
-  const thaiCount = thaiMatches.length;
-
-  const chineseMatches = targetText.match(/[\u4E00-\u9FA5]/g) || [];
-  const chineseCount = chineseMatches.length;
-
-  const kanaMatches = targetText.match(/[\u3040-\u30FF]/g) || [];
-  const kanaCount = kanaMatches.length;
-
-  const hangulMatches = targetText.match(/[\uAC00-\uD7AF]/g) || [];
-  const hangulCount = hangulMatches.length;
-
-  const englishMatches = targetText.match(/[a-zA-Z]+/g) || [];
-  const englishCount = englishMatches.join('').length;
-
+  let thaiCount = 0;
+  let chineseCount = 0;
+  let kanaCount = 0;
+  let hangulCount = 0;
+  let englishCount = 0;
+  for (const ch of targetText) {
+    const code = ch.codePointAt(0);
+    if (code >= 0x0E00 && code <= 0x0E7F) thaiCount++;
+    else if (code >= 0x3040 && code <= 0x30FF) kanaCount++;
+    else if (code >= 0x4E00 && code <= 0x9FA5) chineseCount++;
+    else if (code >= 0xAC00 && code <= 0xD7AF) hangulCount++;
+    else if ((code >= 0x41 && code <= 0x5A) || (code >= 0x61 && code <= 0x7A)) englishCount++;
+  }
 
   if (kanaCount > 0) {
     return 'ja';
@@ -56,15 +62,31 @@ export function detectDominantLanguage(text) {
     return 'ko';
   }
 
-  if (thaiCount >= chineseCount && thaiCount >= englishCount && thaiCount > 0) {
-    return 'th';
-  } else if (chineseCount > thaiCount && chineseCount >= englishCount && chineseCount > 0) {
+
+  if (thaiCount > 0) {
+    if (thaiCount >= 2 || thaiCount >= chineseCount) {
+      return 'th';
+    }
+  }
+
+  if (chineseCount > 0 && chineseCount >= thaiCount && chineseCount >= englishCount) {
     return 'zh';
-  } else if (englishCount > 0) {
+  }
+
+  if (thaiCount > 0) {
+    return 'th';
+  }
+
+
+  if (sessionLang && sessionLang !== 'en' && (targetText.match(/[a-zA-Z]+/g) || []).length <= 3 && !/\b(what|who|when|where|why|how|which|can|tell|give|list|show|is|are|was|were)\b/i.test(targetText)) {
+    return sessionLang;
+  }
+
+  if (englishCount > 0) {
     return 'en';
   }
 
-  return 'th';
+  return sessionLang || 'th';
 }
 
 function stableToolSignature(toolName, args) {
@@ -78,11 +100,81 @@ function stableToolSignature(toolName, args) {
   return `${toolName}:${JSON.stringify(sortValue(args || {}))}`;
 }
 
-export class KikuAgent {
-  constructor(options = {}) {
-    this.allowedToolNames = options.allowedToolNames
-      ? new Set(Array.from(options.allowedToolNames, String))
-      : null;
+
+export { extractTopLevelJsonObjects };
+
+const REPEAT_LOOP_THRESHOLD = 3;
+
+
+function summarizeToolDescription(description, maxChars = 160) {
+  const text = String(description || '').replace(/\s+/g, ' ').trim();
+  if (text.length <= maxChars) return text;
+  const cut = text.slice(0, maxChars);
+  const lastStop = Math.max(cut.lastIndexOf('. '), cut.lastIndexOf('; '));
+  return lastStop > 60 ? cut.slice(0, lastStop + 1) : `${cut.trim()}…`;
+}
+
+
+export function stripInjectedDirectives(text) {
+  let t = String(text || '');
+  t = t.replace(/\[SYSTEM DIRECTIVE[\s\S]*$/i, '');
+  t = t.replace(/\[SYSTEM CLARIFICATION\][\s\S]*$/i, '');
+  t = t.replace(/\[CONTEXT GROUNDING:[^\]]*\]/gi, '');
+  t = t.replace(/\[RESTORED SESSION TOOL CONTEXT\][\s\S]*$/i, '');
+  return t.trim();
+}
+
+const AUTO_RESEARCH_KEYWORDS = [
+ 
+  'วิเคราะห์', 'เปรียบเทียบ', 'เจาะลึก', 'ลงลึก', 'วิจัย', 'ค้นคว้า', 'รายงาน', 'ข่าวลือ', 'ตรวจสอบ', 'รวบรวม', 'ครบถ้วน',
+  
+  'analyze', 'analysis', 'compare', 'research', 'investigate', 'deep dive', 'in depth', 'in-depth', 'comprehensive', 'everything about', 'report on', 'rumor',
+ 
+  '分析', '比较', '對比', '调查', '研究', '深入', '全面',
+];
+
+export function scoreResearchComplexity(prompt) {
+  const text = stripInjectedDirectives(prompt);
+  const signals = [];
+  let score = 0;
+  if (!text) return { score, signals, strippedLength: 0 };
+
+  if (text.length >= 240) { score += 1; signals.push('length'); }
+  if (text.length >= 600) { score += 1; signals.push('very-long'); }
+
+  const questionMarks = (text.match(/[?？]/g) || []).length;
+  if (questionMarks >= 2) { score += 1; signals.push('multi-question'); }
+
+  const listMarkers = (text.match(/(^|\n)\s*(?:\d+[.)]|[-•*])\s+\S/g) || []).length;
+  if (listMarkers >= 2) { score += 1; signals.push('list'); }
+
+  const lower = text.toLowerCase();
+  const matchedKeywords = AUTO_RESEARCH_KEYWORDS.filter(k => lower.includes(k));
+  if (matchedKeywords.length >= 1) { score += 1; signals.push(`keywords:${matchedKeywords.slice(0, 3).join('|')}`); }
+  if (matchedKeywords.length >= 3) { score += 1; signals.push('keyword-rich'); }
+
+  return { score, signals, strippedLength: text.length };
+}
+
+
+export function shouldAutoResearch(prompt, options = {}) {
+  const mode = String(options.mode ?? 'auto').toLowerCase();
+  if (mode === 'off') return false;
+  const { score } = scoreResearchComplexity(prompt);
+  if (mode === 'on') return score >= 1;
+  return score >= (Number(options.threshold) || 3);
+}
+
+function buildAutoResearchDirective(complexity) {
+  const signalText = (complexity.signals || []).slice(0, 4).join(', ');
+  return `[SYSTEM DIRECTIVE: AUTO DEEP RESEARCH] The user's query is complex and multi-part (complexity score ${complexity.score}${signalText ? ` — ${signalText}` : ''}). Do NOT answer it with a few quick tool calls. You MUST begin by calling run_deep_research ONCE with a focused, self-contained question derived from the user's request. After the research report returns, synthesize the final answer from its cited findings. If run_deep_research fails or returns insufficient data, fill the gaps with web_search/web_extract instead of leaving any section unanswered.`;
+}
+
+export class KikuAgent {
+  constructor(options = {}) {
+    this.allowedToolNames = options.allowedToolNames
+      ? new Set(Array.from(options.allowedToolNames, String))
+      : null;
     this.client = options.client || new LLMClient({
       provider: options.provider || config.activeProvider,
       baseUrl: options.baseUrl,
@@ -94,13 +186,18 @@ export class KikuAgent {
     this.session = new AgentSession({
       thinkingMode: options.thinkingMode || 'hide',
       sessionId: options.sessionId,
-      storageKey: options.sessionStorageKey,
-      storageFile: options.sessionStorageFile,
-      persistence: options.sessionPersistence !== false,
+      storageKey: options.sessionStorageKey,
+      storageFile: options.sessionStorageFile,
+      persistence: options.sessionPersistence !== false,
     });
     this.memory = options.memory || new AgentMemory();
+    this.activeRunId = options.activeRunId || options.runId || null;
+    if (this.activeRunId && typeof this.memory.setActiveRunId === 'function') {
+      this.memory.setActiveRunId(this.activeRunId);
+    }
     this.mcpManager = options.mcpManager || new McpClientManager();
     this.router = new DynamicToolRouter(this.getAllToolsSchema());
+    this.routingInterceptor = new RoutingInterceptor();
     this.research = new ResearchOrchestrator({ client: this.client });
     this.suggestions = new SuggestionEngine({ client: this.client, memory: this.memory });
     this.jobs = new AgentJobManager();
@@ -110,6 +207,8 @@ export class KikuAgent {
     this.initHistory();
     this.restoredMessageCount = 0;
     this.lastCitations = [];
+    this.lastQueryMetrics = null;
+    this.queryMetricsHistory = [];
 
     this.stateMachine = new AgentStateMachine();
 
@@ -133,7 +232,12 @@ export class KikuAgent {
         await this.mcpManager.loadPersistedServers();
         this.router.setTools(this.getAllToolsSchema());
         this.systemPrompt = this.buildSystemPrompt();
-      } catch { }
+      } catch (err) {
+       
+        this.mcpInitError = err?.message || String(err);
+        console.warn('[Kiku] Failed to load persisted MCP servers:', this.mcpInitError);
+        eventBus.emit(Events.AGENT_QUERY_ERROR, { type: 'mcp_init', error: this.mcpInitError });
+      }
     }
     if (!this.conversationHistory || this.conversationHistory.length === 0) {
       this.initHistory();
@@ -360,6 +464,52 @@ export class KikuAgent {
     return this.endSession();
   }
 
+  /**
+   * Summarize tool evidence from persisted session messages. Raw tool payloads
+   * are not rehydrated into LLM history (they violate tool-call pairing after
+   * a restart and bloat the window), so the model gets one compact evidence
+   * line per earlier query instead.
+   */
+  _rehydrateSessionMessages() {
+    const evidenceLines = [];
+    const messages = this.session.messages || [];
+    for (let i = 0; i < messages.length; i++) {
+      const msg = messages[i];
+      if (msg.role !== 'assistant') continue;
+      const toolParts = (msg.parts || []).filter(p => p.type === 'tool');
+      if (toolParts.length === 0) continue;
+
+      const textPart = (msg.parts || []).find(p => p.type === 'text');
+      let userQuestion = '';
+      for (let j = i - 1; j >= 0; j--) {
+        if (messages[j].role === 'user') {
+          const part = (messages[j].parts || []).find(p => p.type === 'text');
+          userQuestion = String(part?.content || '').replace(/\s+/g, ' ').trim().slice(0, 80);
+          break;
+        }
+      }
+      const toolSummary = toolParts
+        .map(p => {
+          const status = p.status || p.result?.status || 'ok';
+          const count = p.result?.count ?? p.result?.returnedCount;
+          return `${p.name || 'tool'}(${status}${count !== undefined ? `, ${count}` : ''})`;
+        })
+        .join(', ');
+      const answerSnippet = String(textPart?.content || '').replace(/\s+/g, ' ').trim().slice(0, 100);
+      evidenceLines.push(`- Q: "${userQuestion || '(no text)'}" → ${toolSummary}${answerSnippet ? ` → answered: "${answerSnippet}"` : ''}`);
+    }
+    return evidenceLines;
+  }
+
+  _pushRestoredToolContext() {
+    const evidenceLines = this._rehydrateSessionMessages();
+    if (evidenceLines.length === 0) return;
+    this.conversationHistory.push({
+      role: 'system',
+      content: `[RESTORED SESSION TOOL CONTEXT] The queries below were already answered with verified tool results earlier in this session (the raw tool outputs are no longer in history):\n${evidenceLines.slice(-20).join('\n')}\nTreat these as previously verified facts. Re-call the relevant tools only if the user needs fresh or deeper detail.`
+    });
+  }
+
   async loadSession() {
     const loaded = await this.session.loadFromDisk();
     if (loaded && this.session.messages.length > 0) {
@@ -389,6 +539,7 @@ export class KikuAgent {
           }
         }
       }
+      this._pushRestoredToolContext();
       this.restoredMessageCount = this.session.messages.length;
     }
   }
@@ -441,6 +592,7 @@ export class KikuAgent {
         }
       }
     }
+    this._pushRestoredToolContext();
     this.restoredMessageCount = this.session.messages.length;
     eventBus.emit(Events.SESSION_SAVED, { sessionId: this.session.id });
     return archived;
@@ -509,21 +661,100 @@ export class KikuAgent {
     return { hydrated: true, messageCount: normalized.length };
   }
 
-  applySlidingWindow() {
+  _isLocalProvider() {
+    const base = String(this.client?.baseUrl || '');
+    if (/localhost|127\.0\.0\.1|0\.0\.0\.0|\[::1\]/i.test(base)) return true;
+    return this.client?.provider === 'lmstudio' && !base;
+  }
+
+  async applySlidingWindow() {
     if (this.conversationHistory.length > this.slidingWindowSize + 1) {
-      this.conversationHistory = contextCompactor.compact(this.conversationHistory, {
+      const options = {
         maxTokens: 24000,
         keepRecentMessages: this.slidingWindowSize,
-      });
+      };
+      if (!this._isLocalProvider()) {
+        options.summarizer = (older) => this._llmSummarizeOlderMessages(older);
+      }
+      this.conversationHistory = await contextCompactor.compactAsync(this.conversationHistory, options);
+    }
+  }
+
+  /** LLM summary of the dropped window (cloud providers only) — falls back to
+   * the heuristic summary inside the compactor when this fails. */
+  async _llmSummarizeOlderMessages(olderMessages) {
+    const transcript = (olderMessages || [])
+      .map(m => {
+        const role = m.role === 'tool' ? `tool(${m.name || '?'})` : m.role;
+        const content = typeof m.content === 'string' ? m.content : JSON.stringify(m.content ?? m.tool_calls ?? '');
+        return `${role}: ${content.slice(0, 400)}`;
+      })
+      .join('\n')
+      .slice(0, 6000);
+
+    if (!transcript.trim()) return '';
+    const res = await Promise.race([
+      this.client.chat({
+        messages: [
+          { role: 'system', content: 'Summarize the following conversation excerpt for an assistant context window. Capture what the user asked, which tools were used, and the key verified facts. Maximum 120 words. Output only the summary text.' },
+          { role: 'user', content: transcript }
+        ],
+        temperature: 0.1,
+        maxTokens: 300,
+        stream: false
+      }),
+      new Promise((_, reject) => setTimeout(() => reject(new Error('summary timeout')), 20_000))
+    ]);
+    return String(res?.content || '').trim();
+  }
+
+  /**
+   * Hard context-budget guard: KIKU_CONTEXT_LIMIT becomes a real budget instead
+   * of telemetry. Before each LLM call, if the estimated prompt exceeds 85% of
+   * the limit, replace the oldest tool results with omission stubs — tool
+   * payloads are the bulkiest and least-likely-to-be-reused part of history.
+   */
+  _enforceContextBudget() {
+    const rawLimit = Number(process.env.KIKU_CONTEXT_LIMIT);
+    if (!Number.isFinite(rawLimit) || rawLimit <= 0) return false;
+    const budget = Math.floor(rawLimit * 0.85);
+
+    let total = 0;
+    for (const msg of this.conversationHistory) {
+      if (typeof msg._tokenEstimate !== 'number') {
+        msg._tokenEstimate = estimateTokenCount([msg]);
+      }
+      total += msg._tokenEstimate;
     }
+    if (total <= budget) return false;
+
+    const omitted = JSON.stringify({
+      authoritative: false,
+      status: 'omitted',
+      message: 'Older tool result was omitted to fit the context window; its facts are no longer visible. Re-call the tool only if essential.'
+    });
+    const omittedTokens = estimateTextTokens(omitted);
+    let trimmed = false;
+
+    for (const msg of this.conversationHistory) {
+      if (total <= budget) break;
+      if (msg.role !== 'tool' || msg._omittedForContext) continue;
+      const prev = msg._tokenEstimate || estimateTokenCount([msg]);
+      total -= prev - omittedTokens;
+      msg.content = omitted;
+      msg._omittedForContext = true;
+      msg._tokenEstimate = omittedTokens;
+      trimmed = true;
+    }
+    return trimmed;
   }
 
-  getAllToolsSchema() {
-    const mcpTools = this.mcpManager ? this.mcpManager.getToolsSchema() : [];
-    const allTools = [...toolsSchema, ...mcpTools];
-    if (!this.allowedToolNames) return allTools;
-    return allTools.filter(tool => this.allowedToolNames.has(tool.function?.name));
-  }
+  getAllToolsSchema() {
+    const mcpTools = this.mcpManager ? this.mcpManager.getToolsSchema() : [];
+    const allTools = [...toolsSchema, ...mcpTools];
+    if (!this.allowedToolNames) return allTools;
+    return allTools.filter(tool => this.allowedToolNames.has(tool.function?.name));
+  }
 
   async connectMcp(name, serverConfig) {
     const result = await this.mcpManager.connect(name, serverConfig);
@@ -586,180 +817,18 @@ export class KikuAgent {
     return results;
   }
 
-  buildSystemPrompt() {
-    const allTools = this.getAllToolsSchema();
-    const today = new Date().toISOString().slice(0, 10);
-    const toolsSummary = allTools.map(t => {
-      const f = t.function;
-      const paramKeys = Object.keys(f.parameters?.properties || {}).join(', ');
-      return `- ${f.name}(${paramKeys}): ${f.description}`;
-    }).join('\n');
-
-    return `${config.agent.systemPersona}
-
-=== AVAILABLE KNOWLEDGE & LIVE TOOLS ===
-${toolsSummary}
-
-=== DYNAMIC MULTILINGUAL & RESPONSE RULES (CRITICAL) ===
-
-0. NO SYSTEM LEAK (ABSOLUTE RULE):
-   - NEVER include "Thinking Process", "Tool Calling", "Action:", "Action Input:", function names, parameter schemas, JSON tool call blocks, or any internal reasoning in your response text.
-   - NEVER include advertisements, sponsor mentions, or promotional content.
-   - Your response must be CLEAN user-facing content ONLY.
-   - Think silently using tools, then present ONLY the final polished answer.
-
-1. LANGUAGE MIRRORING:
-   - THAI ➔ polite Thai (ใช้ ค่ะ/นะคะ เสมอ)
-   - CHINESE ➔ fluent, polite Chinese (简体/繁体中文)
-   - ENGLISH ➔ polished, professional English
-   - MIXED ➔ Match dominant language, keep bilingual titles
-
-2. REAL YOUTUBE & MEDIA LINKS:
-   - Direct Watch: https://www.youtube.com/watch?v=VIDEO_ID
-   - Search: https://www.youtube.com/results?search_query=QUERY
-   - Also Bilibili, QQ Music, Youku when applicable
-   - NEVER use fake placeholder IDs
-
-3. FACT ACCURACY & NO DUPLICATION:
-   - Always use tools to verify facts before answering. Do NOT hallucinate.
-   - Current date is ${today}. Prioritize current and recent projects when relevant.
-   - Never repeat the same statement. Sort newest first.
-
-3A. CONVERSATION TOPIC ANCHOR:
-   - Resolve short or ambiguous follow-ups (for example "ทำไมน่าจับตามอง?", "งานนี้", "why is it notable?") against the most recent user topic and assistant answer.
-   - The current follow-up inherits that topic. Do not default the subject to Ju Jingyi merely because the persona or knowledge base focuses on her.
-   - A persona changes voice and tone only; it never replaces the user's subject.
-   - If the reference remains genuinely ambiguous after reading recent history, ask one focused clarification question instead of guessing.
-
-3B. FUTURE-EVENT CLAIM SAFETY:
-   - Treat any event after today's date as scheduled or reported, not completed fact.
-   - Never assert an exact future date, invitation, guest list, or attendance without checking a current official source with a live web tool.
-   - Never carry a date, guest list, or attendance status from one annual edition into another year.
-   - When official confirmation is unavailable, say clearly that the detail is unconfirmed and subject to change.
-
-4. RESPONSE STYLE:
-   - NEVER start with template greetings — dive DIRECTLY into the answer.
-   - Format with Markdown headers (###), bold (**text**), bullet points (•).
-   - Emoji are optional: use at most one contextually relevant emoji per response and never repeat them across headings or bullets.
-   - Match answer depth to what the user asked.
-   - Separate catalog facts from interpretation. Clearly frame acting observations as analysis unless a cited review supports them.
-   - Do not infer that newer means deeper, more realistic, or better acted. Compare concrete role demands, narrative agency, genre, and verified scenes instead.
-   - Omit catalog ratings from analytical conclusions unless the tool output names the rating source and scale.
-   - Avoid unsupported superlatives or career verdicts (for example "unrivalled", "top-tier", or "completely transcended idol status").
-   - In Blossom (花间令): Ju Jingyi portrays Yang Caiwei's identity in Shangguan Zhi's appearance after the face-swap premise; do not describe this as two unrelated roles throughout the series.
-   - The Legend of White Snake (新白娘子传奇): Bai Suzhen is the white-snake spirit in human form and Xu Xian is the human man; never reverse this relationship.
-   - On an explicit continuation, answer the new angle without restating the previous response.
-
-5. SMART TOOL SELECTION (MANDATORY LOCAL DATABASE FIRST):
-   - ALWAYS call LOCAL built-in database tools FIRST. They query the rich, local catalog database directly in-process and are the authoritative source of truth.
-   - Videos/stages/audio ➔ search_discography, get_music_videos, get_media_links, play_music
-   - Songs/OSTs ➔ search_discography
-   - Dramas/roles ➔ search_works, get_work_detail, get_character_roles
-   - Profile/fandom ➔ get_profile_and_fandom
-   - Brands ➔ get_brand_endorsements
-   - Awards ➔ get_awards
-   - Career timeline ➔ get_career_timeline
-   - Local files ➔ read_local_file, list_directory, get_system_stats
-   - MCP tools (mcp_*) ➔ Secondary only; only invoke when user explicitly mentions MCP (@mcp) or when specialized external services are required.
-   - News/external facts ➔ web_search or web_extract
-   - Deep multi-source research ➔ web_extract (search + auto-fetch)
-   - Read webpage content ➔ fetch_webpage
-   - Fetch API/JSON ➔ fetch_json
-   - Screenshot ➔ screenshot_webpage
-
-6. AGGRESSIVE WEB SEARCH FALLBACK (CRITICAL):
-   - If a local catalog tool (e.g. search_works, search_discography, get_brand_endorsements, get_awards) returns not_found, empty, or 0 results, DO NOT give up or guess. You MUST IMMEDIATELY call web_search or web_extract in the next step to look up live external facts online.
-   - Combine local catalog data WITH live web search for complete, accurate, and up-to-date answers.
-   - Use web_extract for comprehensive research on any topic.
-
-7. LEGAL & CAREER KNOWLEDGE:
-   - When discussing controversies, be specific and accurate:
-     • Contract dispute: Star48/Siba Media supplementary agreement. JJY claims main contract ended June 2024, alleges forged signatures. Star48 claims until 2033.
-     • Tax allegation: Tip-off report filed by Star48 side; JJY denies and confirms tax compliance.
-     • Portrait/copyright cases: JJY has sued beauty clinics and brands for unauthorized image use (肖像权纠纷).
-   - Use web_search or web_extract for latest verified information on legal matters.
-
-8. DELIBERATE REASONING & AUTONOMOUS TOOL SELECTION (CRITICAL):
-   - Think carefully about the complexity, goal, and nature of the query before deciding which tool to use:
-   • Direct Specific Query ➔ Use direct catalog tools (search_works, search_discography, get_profile_and_fandom, get_awards, get_brand_endorsements, etc.).
-   • Isolated In-Depth Subtask ➔ Call run_subagent with a specific task parameter to delegate an isolated iterative investigation with evidence and tool call tracing.
-   • Complex Multi-Faceted Research ➔ Call run_deep_research with question when the user asks an open-ended, complex query that requires parallel multi-worker investigation across catalog, live web sources, and independent verifier with cited sources.
-   • Rumors, Cast Leaks & Fact Verification ➔ Call verify_claim with claim when the user asks whether a rumor, statement, or unconfirmed news is true (returns VERIFIED, UNVERIFIED, CONFLICTING, or FALSE).
-    • Teaching / Updating Response Preferences ➔ Call manage_skill with action 'learn' (with instruction), 'list', 'remove' (with skillId), 'clear', 'enable', or 'disable' when the user instructs how they want Kiku to answer in future conversations.
-
-9. ANTI-HALLUCINATION GUARDRAILS (CRITICAL — MANDATORY COMPLIANCE):
-
-   0. INDEPENDENT VERIFICATION FIRST:
-      - When asked about specific named entities (e.g., upcoming dramas, casts, titles, products, brand deals, real-world releases), do NOT generate an explanation based solely on keyword inference or assumed context.
-      - You MUST cross-reference and verify the exact entity existence using your local catalog tools (search_works, search_discography, get_brand_endorsements, etc.) BEFORE writing the response.
-      - If the local catalog has no match, you MUST call web_search or web_extract to verify externally before responding.
-
-   1. ZERO-ASSUMPTION ON UNVERIFIED ENTITIES:
-      - If the subject/entity cannot be verified with official records (not found in catalog tools AND not found via web_search), explicitly state that no verified record exists.
-      - Strictly avoid fabricating plausible generic narratives (e.g., generic tropes, placeholder plotlines, speculative casting, invented brand partnerships).
-      - NEVER invent drama titles, song names, character roles, brand endorsements, or career milestones that are not backed by tool results.
-
-   2. FALLBACK & UNCERTAINTY PROTOCOL:
-      - If the entity appears to be misspelled, fan fiction, or an unverified rumor, directly flag the discrepancy (including linguistic errors like incorrect Pinyin or romanization) rather than adopting the user's premise as true.
-      - Always prioritize stating "No official information found in the database" over generating a complete but ungrounded description.
-      - If tool results are ambiguous or partial, present only what is confirmed and clearly label uncertain parts as "unverified" or "not confirmed."
-
-   3. SOURCE ATTRIBUTION:
-      - When stating facts, prefer answers that trace back to tool results. If a claim comes from local catalog, you may present it as confirmed data. If from web search, note it as external/recent information.
-      - NEVER present fabricated information as if it were verified catalog data.
-
-   4. SELF-CHECK BEFORE FINAL ANSWER:
-      - Before synthesizing your final response, mentally verify: "Does every factual claim in my answer have a corresponding tool result backing it up?"
-      - If any claim lacks tool evidence, either (a) call the appropriate tool to verify it, or (b) remove it from your answer.
-      - This applies to ALL entity types: dramas, songs, MVs, brands, awards, dates, character names, and career milestones.`;
+  buildSystemPrompt(scopedTools = null) {
+    const allTools = this.getAllToolsSchema ? this.getAllToolsSchema() : [];
+    const activeTools = scopedTools || allTools;
+    return buildSystemPrompt(activeTools, { isNativeToolCalling: true });
   }
 
-  parseTextToolCalls(content) {
-    if (!content) return [];
-    const calls = [];
-    const allTools = this.getAllToolsSchema();
-
-    const jsonBlockRegex = /```(?:json)?\s*(\{[\s\S]*?\})\s*```/g;
-    for (const match of content.matchAll(jsonBlockRegex)) {
-      try {
-        const parsed = JSON.parse(match[1]);
-        const toolName = parsed.action || parsed.tool || parsed.name || parsed.function;
-        const toolArgs = parsed.action_input || parsed.parameters || parsed.arguments || parsed.args || {};
-        if (toolName && allTools.some(t => t.function.name === toolName)) {
-          calls.push({
-            id: `call_txt_${Date.now()}_${calls.length}`,
-            function: {
-              name: toolName,
-              arguments: typeof toolArgs === 'string' ? toolArgs : JSON.stringify(toolArgs)
-            }
-          });
-        }
-      } catch {
-      }
-    }
-
-    const actionRegex = /Action:\s*([a-zA-Z0-9_]+)\s*\nAction Input:\s*(\{[\s\S]*?\}|[^\n]+)/g;
-    for (const match of content.matchAll(actionRegex)) {
-      const toolName = match[1].trim();
-      const rawInput = match[2].trim();
-      if (allTools.some(t => t.function.name === toolName)) {
-        let args = {};
-        try {
-          args = JSON.parse(rawInput);
-        } catch {
-          args = { query: rawInput };
-        }
-        calls.push({
-          id: `call_act_${Date.now()}_${calls.length}`,
-          function: {
-            name: toolName,
-            arguments: JSON.stringify(args)
-          }
-        });
-      }
-    }
+  _extractTextToolCalls(content) {
+    return parseTextToolCallsFromContent(content, this._activeToolSchema || this.getAllToolsSchema());
+  }
 
-    return calls;
+  parseTextToolCalls(content) {
+    return this._extractTextToolCalls(content).calls;
   }
 
   async ask(userPrompt, callbacks = {}) {
@@ -792,12 +861,24 @@ ${toolsSummary}
       onTokenDelta = () => { },
       onFinalAnswer = () => { },
       onError = () => { },
+      onNotice = () => { },
+      systemDirectives = null,
       signal = null
     } = callbacks;
 
     const queryStartTime = Date.now();
     const queryId = `q_${Date.now()}_${Math.random().toString(36).slice(2, 6)}`;
-    const dominantLang = detectDominantLanguage(promptForLLM);
+    const dominantLang = detectDominantLanguage(promptForLLM, this.session?.dominantLanguage);
+    if (this.session) {
+      this.session.dominantLanguage = dominantLang;
+    }
+
+    const tokenTracker = new TokenTracker({
+      sessionId: this.session?.id || null,
+      modelId: this.client?.model || 'unknown',
+      contextLimit: Number(process.env.KIKU_CONTEXT_LIMIT) || 0,
+    });
+    tokenTracker.start();
 
     this.stateMachine.startQuery(queryId);
     eventBus.emit(Events.AGENT_QUERY_START, { queryId, prompt: promptForLLM });
@@ -809,6 +890,27 @@ ${toolsSummary}
       ];
     }
 
+    const isSecurityAttack = /(ผู้ใช้คนอื่น|หัวหน้าของเขา|แสดงประวัติส่วนตัวของเขา|ขอรหัส|password|secret|credential|token|api_?key|api\s*key|\.env|cookies\.txt|kiku_vip_keys|system\s+directive|system\s+prompt|prompt\s+leak|bypass|unauthorized|环境变量)/i.test(promptForLLM);
+    if (isSecurityAttack) {
+      const refusalMessage = dominantLang === 'zh'
+        ? "抱歉，无法提供系统配置、私有数据或敏感指令。"
+        : dominantLang === 'en'
+          ? "I cannot disclose system configurations, credentials, or personal private data."
+          : "ขออภัยค่ะ ไม่สามารถเปิดเผยข้อมูลการกำหนดค่าระบบ ข้อมูลประจำตัว หรือข้อมูลส่วนบุคคลได้";
+      this.session.createUserMessage(promptForLLM);
+      this.session.createAssistantMessage({ content: refusalMessage });
+      this.conversationHistory.push({ role: 'user', content: promptForLLM });
+      this.conversationHistory.push({ role: 'assistant', content: refusalMessage });
+      const totalDurationMs = Date.now() - queryStartTime;
+      onFinalAnswer(refusalMessage, { durationMs: totalDurationMs, language: dominantLang });
+      this.stateMachine.transition(AgentStates.COMPLETE);
+      eventBus.emit(Events.AGENT_FINAL_ANSWER, { queryId, answer: refusalMessage, durationMs: totalDurationMs });
+      eventBus.emit(Events.AGENT_QUERY_END, { queryId, durationMs: totalDurationMs });
+      this.stateMachine.reset();
+      await this.saveSession().catch(() => {});
+      return refusalMessage;
+    }
+
     let promptForLLMWithGrounding = promptForLLM;
     if (this.lastCitations && this.lastCitations.length > 0) {
       const citedRef = resolveCitationReference(promptForLLM, this.lastCitations);
@@ -821,17 +923,21 @@ ${toolsSummary}
     const userMessage = { role: 'user', content: promptForLLMWithGrounding };
     this.conversationHistory.push(userMessage);
 
-    const memoryContext = this.memory.buildMemoryContext(promptForLLM);
+    const memoryContext = (this.activeRunId && typeof this.memory.addMemoryContext === 'function')
+      ? await this.memory.addMemoryContext({ currentQuery: promptForLLM, activeRunId: this.activeRunId })
+      : this.memory.buildMemoryContext(promptForLLM);
     this.conversationHistory[0] = {
       role: 'system',
       content: this.systemPrompt + memoryContext
     };
 
+    const turnDirectives = [];
+
     let langInstruction;
     if (dominantLang === 'th') {
-      langInstruction = '[SYSTEM NOTE: User language is Thai. Reply in Thai only.]\n';
+      langInstruction = '[SYSTEM NOTE: User language is Thai. Reply in Thai only. Maintain Thai even when explaining errors, refusals, or reporting data not found.]\n';
     } else if (dominantLang === 'zh') {
-      langInstruction = '[SYSTEM NOTE: User language is Chinese. Reply in Chinese only.]\n';
+      langInstruction = '[SYSTEM NOTE: User language is Chinese. Reply in Chinese only. Maintain Chinese even when reporting data not found.]\n';
     } else if (dominantLang === 'ja') {
       langInstruction = '[SYSTEM NOTE: User language is Japanese. Reply in Japanese only.]\n';
     } else if (dominantLang === 'ko') {
@@ -844,7 +950,59 @@ ${toolsSummary}
     if (directMentions.length > 0) {
       langInstruction += `[SYSTEM DIRECTIVE: User explicitly invoked @${directMentions.join(', @')}. You MUST prioritize calling the referenced tool(s) to fetch accurate data.]\n`;
     }
-    this.conversationHistory.push({ role: 'system', content: langInstruction });
+    turnDirectives.push({ role: 'system', content: langInstruction });
+
+    if (/(เธอ|เขา|ศิลปิน|\bshe\b|\bher\b)/i.test(promptForLLM) && !/(บอต|bot|ai|คุณคือใคร|who are you)/i.test(promptForLLM)) {
+      turnDirectives.push({
+        role: 'system',
+        content: '[SUBJECT DISAMBIGUATION: The pronoun/subject ("เธอ", "เขา", "ศิลปิน", "She", "Her") in this query refers strictly to Ju Jingyi (鞠婧祎), NOT to the AI assistant. Query the appropriate catalog tool for Ju Jingyi\'s facts/quotes/works.]'
+      });
+    }
+
+    for (const directive of (Array.isArray(systemDirectives) ? systemDirectives : [])) {
+      if (directive && typeof directive === 'string') {
+        turnDirectives.push({ role: 'system', content: directive });
+      }
+    }
+
+    let autoResearchTriggered = false;
+    const autoResearchMode = String(config.agent.autoResearch || 'auto').toLowerCase();
+    if (autoResearchMode !== 'off') {
+      const complexity = scoreResearchComplexity(promptForLLM);
+      if (shouldAutoResearch(promptForLLM, { mode: autoResearchMode })) {
+        const orchestratorAvailable = this.getAllToolsSchema().some(t => t.function?.name === 'run_deep_research');
+        if (orchestratorAvailable) {
+          autoResearchTriggered = true;
+          turnDirectives.push({ role: 'system', content: buildAutoResearchDirective(complexity) });
+        }
+      }
+    }
+
+    const matchedQueryIntents = this.router.detectIntents(promptForLLM);
+    const queryResolution = resolveQueryEntities(promptForLLM);
+
+    if (this.conversationHistory.length > 2) {
+      const priorAssistantMsgs = this.conversationHistory
+        .filter(m => m.role === 'assistant' && typeof m.content === 'string')
+        .slice(-2);
+      const priorText = priorAssistantMsgs.map(m => m.content).join(' ');
+      const priorResolution = resolveQueryEntities(priorText);
+
+      if (priorResolution.hasBrand && (queryResolution.hasSong || queryResolution.hasWork || queryResolution.hasMedia)) {
+        turnDirectives.push({
+          role: 'system',
+          content: '[CONTEXT DISAMBIGUATION: The user has shifted focus away from earlier brand/endorsement discussions. The current query targets a music/video/drama entity. Query tools appropriate for works/songs/videos, NOT brand endorsements.]'
+        });
+      }
+    }
+
+    const mentionsArtist = /(ju\s?jing\s?yi|jujingyi|จวีจิ้งอี|จิ้งอี|จวี|鞠婧祎|小鞠|kikuuu?|miju|蜜橘)/i.test(promptForLLM);
+    if (matchedQueryIntents.length === 0 && queryResolution.matchedEntities.length === 0 && !mentionsArtist && !autoResearchTriggered) {
+      turnDirectives.push({
+        role: 'system',
+        content: "[SYSTEM NOTE: This query does not appear to be about Ju Jingyi or her catalog. Answer directly from your general knowledge in the user's language. Do NOT call catalog tools (profile, works, discography, brands, awards, gallery, quotes, timeline) for this query; use web_search or web_extract only if current external information is genuinely required.]"
+      });
+    }
 
     const clarifyingResult = await this._handleClarifyingQuestions(
       promptForLLM,
@@ -853,13 +1011,13 @@ ${toolsSummary}
       callbacks
     );
     if (clarifyingResult) {
-      this.conversationHistory.push({
+      turnDirectives.push({
         role: 'user',
         content: `[SYSTEM CLARIFICATION] ${clarifyingResult.summary}\nOriginal query: ${promptForLLM}`
       });
     }
 
-    this.applySlidingWindow();
+    await this.applySlidingWindow();
 
     const currentAssistantSessionMsg = this.session.createAssistantMessage({
       model: this.client?.model || null,
@@ -872,12 +1030,25 @@ ${toolsSummary}
 
     let currentIteration = 0;
     const toolResultCache = new Map();
+    const repeatSignatureCounts = new Map();
+    this.routingInterceptor.reset();
+    let textProtocolTaught = false;
     const toolResultTracker = { notFound: 0, error: 0, success: 0, total: 0, toolNames: [] };
+    const recordedTools = [];
     const activeTools = this.router.routeToolsForQuery(promptForLLM, {
       showAll: callbacks.showAllTools || false,
       includeMcp: true,
     });
 
+    if (autoResearchTriggered) {
+      const orchestrator = this.getAllToolsSchema().find(t => t.function?.name === 'run_deep_research');
+      if (orchestrator && !activeTools.some(t => t.function?.name === 'run_deep_research')) {
+        activeTools.push(orchestrator);
+      }
+    }
+    this._activeToolSchema = activeTools;
+    const activeToolNames = new Set(activeTools.map(t => t.function?.name).filter(Boolean));
+
     while (currentIteration < this.maxIterations) {
       if (signal?.aborted) {
         this.stateMachine.reset();
@@ -887,19 +1058,31 @@ ${toolsSummary}
       currentIteration++;
       onStepStart(currentIteration, this.maxIterations);
 
+      if (currentIteration > 1) {
+        await this.applySlidingWindow();
+      }
+      this._enforceContextBudget();
+
       let assistantMessage = null;
       let isReasoning = false;
       let isTokenEmitting = false;
 
       try {
+        const trackerSegment = tokenTracker.beginSegment(`iteration-${currentIteration}`);
+        const chatMessages = [
+          this.conversationHistory[0],
+          ...turnDirectives,
+          ...this.conversationHistory.slice(1)
+        ];
         assistantMessage = await this.client.chat({
-          messages: this.conversationHistory,
+          messages: chatMessages,
           tools: activeTools,
           temperature: 0.3,
-          maxTokens: 8192,
+          maxTokens: OUTPUT_TOKEN_BUDGET,
           stream: true,
           signal,
           onThought: (thoughtChunk) => {
+            tokenTracker.markToken(trackerSegment);
             if (!isReasoning) {
               isReasoning = true;
               this.stateMachine.transition(AgentStates.STREAMING_THOUGHTS);
@@ -911,6 +1094,7 @@ ${toolsSummary}
             eventBus.emit(Events.AGENT_REASONING_DELTA, { queryId, chunk: thoughtChunk });
           },
           onToken: (tokenChunk) => {
+            tokenTracker.markToken(trackerSegment);
             if (isReasoning) {
               isReasoning = false;
               this.session.finalizeReasoningPart(currentAssistantSessionMsg);
@@ -939,6 +1123,11 @@ ${toolsSummary}
           }
           onReasoningEnd();
         }
+
+        tokenTracker.endSegment({
+          promptTokens: assistantMessage.usage?.prompt_tokens,
+          completionTokens: assistantMessage.usage?.completion_tokens,
+        }, trackerSegment);
       } catch (err) {
         this.stateMachine.reset();
         if (!isAbortError(err)) {
@@ -960,22 +1149,83 @@ ${toolsSummary}
       }
 
       let toolCalls = assistantMessage.tool_calls || [];
+      let malformedTextCalls = 0;
       if (toolCalls.length === 0 && assistantMessage.content) {
-        toolCalls = this.parseTextToolCalls(assistantMessage.content);
-       
-        if (toolCalls.length > 0) {
-          let cleaned = assistantMessage.content;
+        const parsedTextCalls = this._extractTextToolCalls(assistantMessage.content);
+        toolCalls = parsedTextCalls.calls;
+        malformedTextCalls = parsedTextCalls.malformed;
 
-          cleaned = cleaned.replace(/```(?:json)?\s*\{[\s\S]*?\}\s*```/g, '');
+        if (toolCalls.length > 0) {
+          assistantMessage.content = stripToolCallBlocks(assistantMessage.content);
         
-          cleaned = cleaned.replace(/Action:\s*[a-zA-Z0-9_]+\s*\nAction Input:\s*(?:\{[\s\S]*?\}|[^\n]+)/g, '');
-          assistantMessage.content = cleaned.trim();
+          assistantMessage.tool_calls = toolCalls.map(tc => ({
+            id: tc.id,
+            type: 'function',
+            function: tc.function
+          }));
         }
       }
 
-      this.conversationHistory.push(assistantMessage);
+      const historyMessage = {
+        role: 'assistant',
+        content: assistantMessage.content || '',
+        ...(assistantMessage.tool_calls ? { tool_calls: assistantMessage.tool_calls } : {})
+      };
+      this.conversationHistory.push(historyMessage);
 
       if (!toolCalls || toolCalls.length === 0) {
+       
+        if (currentIteration === 1 && recordedTools.length === 0) {
+          const content = String(assistantMessage.content || '');
+          const refusalRegex = /(ไม่มี(การ)?บันทึก|ไม่มี(อยู่ใน)?ระบบ|ไม่มีข้อมูล|ไม่มีรายละเอียด|ไม่พบ|ข้อมูลทางธุรกิจ|ข้อมูลธุรกิจ|ข้อมูลเชิงพาณิชย์|ความลับทางการค้า|ความลับ|ความลับเชิงพาณิชย์|ปรัชญาการทำงาน|ตัวตนของฉัน|นอกเหนือขอบเขต|ไม่สามารถระบุ|ไม่ได้ระบุ|not found|no verified record|no record|could not find|unable to find|outside the scope|confidential|internal business|trade secret|business secret|commercial secret|proprietary|internal statistics?|operational philosophy|my own mindset|without querying|没有找到|未找到|暂无|没有记录|无法查询|商业机密|内部数据|机密)/i;
+          const isRefusal = refusalRegex.test(content);
+          const isSec = /(ผู้ใช้คนอื่น|หัวหน้าของเขา|รหัสผ่าน|ความลับของระบบ|password|secret|credential|token|\.env|cookies\.txt|kiku_vip_keys|system\s+directive|mcp\s+directive|bypass|unauthorized)/i.test(promptForLLM);
+          if (isRefusal && !isSec) {
+            this.conversationHistory.pop();
+            if (currentAssistantSessionMsg) currentAssistantSessionMsg.parts = [];
+            this.conversationHistory.push({
+              role: 'user',
+              content: '[MANDATORY RETRIEVAL RULE & VERIFICATION CONSTRAINT] You are strictly FORBIDDEN from asserting that catalog information does not exist, is confidential, or is missing without querying a tool first in this turn. Call the appropriate catalog tool (e.g. get_brand_endorsements, search_works, search_discography, get_awards, get_quotes, get_music_videos, get_gallery_photos) NOW to fetch and verify the facts before answering.'
+            });
+            continue;
+          }
+
+          const explicitCatalogRegex = /(ตาม|ใน|จาก)?(แค็ตตาล็อก|catalog|ฐานข้อมูล|database|ข้อมูลทางการ|บันทึกทางการ)/i;
+          const isExplicitCatalogMention = explicitCatalogRegex.test(promptForLLM);
+          const hasCatalogTarget = (queryResolution?.recommendedTools?.length > 0 || queryResolution?.matchedEntities?.length > 0 || isExplicitCatalogMention) && (mentionsArtist || isExplicitCatalogMention);
+
+          if (!isSec && hasCatalogTarget) {
+            const suggestedTools = queryResolution?.recommendedTools?.length > 0
+              ? queryResolution.recommendedTools.join(', ')
+              : 'search_works, search_discography, get_brand_endorsements, get_career_timeline, get_profile_and_fandom';
+            this.conversationHistory.pop();
+            if (currentAssistantSessionMsg) currentAssistantSessionMsg.parts = [];
+            this.conversationHistory.push({
+              role: 'user',
+              content: `[MANDATORY CATALOG-FIRST RETRIEVAL] This request asks for verified facts from the official catalog. You must NOT provide an unverified direct answer without calling a catalog tool first. Call the appropriate tool (recommended: ${suggestedTools}) NOW to retrieve and verify the exact facts before formulating your response.`
+            });
+            continue;
+          }
+        }
+
+        const needProtocol = Boolean(this.client?.toolsDisabledReason) && !textProtocolTaught;
+        if (currentIteration < this.maxIterations && (needProtocol || malformedTextCalls > 0)) {
+          if (needProtocol) {
+            textProtocolTaught = true;
+            this.conversationHistory.push({
+              role: 'system',
+              content: buildTextToolProtocolDirective(this.getAllToolsSchema())
+            });
+            try {
+              onNotice('Native tool calling is unavailable for this model/provider — switched to the text tool protocol.');
+            } catch { }
+          }
+          if (malformedTextCalls > 0) {
+            this.conversationHistory.push({ role: 'user', content: MALFORMED_TOOL_CALL_FEEDBACK });
+          }
+          continue;
+        }
+
         const totalDurationMs = Date.now() - queryStartTime;
         this.session.stats.totalDurationMs += totalDurationMs;
 
@@ -986,6 +1236,10 @@ ${toolsSummary}
                 dominantLang === 'en' ? "Sorry, unable to process a response." :
                   "ขออภัยค่ะ ไม่สามารถประมวลผลคำตอบได้"
         ));
+        historyMessage.content = finalAnswer;
+        this.conversationHistory = this.conversationHistory.filter(m =>
+          !(typeof m?.content === 'string' && m.content.includes('[MANDATORY'))
+        );
         onFinalAnswer(finalAnswer, { durationMs: totalDurationMs, language: dominantLang });
         this._extractFactsFromFinalAnswer(finalAnswer);
 
@@ -993,29 +1247,68 @@ ${toolsSummary}
         eventBus.emit(Events.AGENT_FINAL_ANSWER, { queryId, answer: finalAnswer, durationMs: totalDurationMs });
         this.stateMachine.reset();
 
+        this._recordQueryMetrics(tokenTracker, {
+          promptTokens: queryPromptTokens,
+          completionTokens: queryCompletionTokens,
+        });
+
         await this.saveSession().catch(err => console.error('Session save failed:', err));
         return finalAnswer;
       }
 
+      let shouldAbortLoop = false;
       const executionPromises = toolCalls.map(async (tc, toolCallIndex) => {
         if (signal?.aborted) throw createAbortError();
 
-        const toolName = tc?.function?.name || '';
-        const toolCallId = tc?.id || `call_${queryId}_${currentIteration}_${toolCallIndex}`;
-        if (this.allowedToolNames && !this.allowedToolNames.has(toolName)) {
-          const rejectedResult = { status: 'error', message: `Tool "${toolName}" is not permitted in this execution context.` };
-          onToolStart(toolName, {});
-          onToolResult(toolName, rejectedResult, 0);
-          return {
-            toolMessage: {
-              role: 'tool',
-              tool_call_id: toolCallId,
-              name: toolName,
-              content: JSON.stringify(rejectedResult),
-            },
-          };
-        }
-        const isBuiltin = isBuiltinTool(toolName);
+        
+        const intercepted = this.routingInterceptor.interceptToolCall(tc, currentIteration - 1);
+        if (intercepted.aborted) {
+          shouldAbortLoop = true;
+          return null;
+        }
+        if (intercepted.rerouted && intercepted.replacement) {
+          tc = {
+            ...tc,
+            function: {
+              name: intercepted.replacement.name,
+              arguments: JSON.stringify(intercepted.replacement.params)
+            }
+          };
+        }
+
+        const toolName = tc?.function?.name || '';
+        const toolCallId = tc?.id || `call_${queryId}_${currentIteration}_${toolCallIndex}`;
+        if (this.allowedToolNames && !this.allowedToolNames.has(toolName)) {
+          const rejectedResult = { status: 'error', message: `Tool "${toolName}" is not permitted in this execution context.` };
+          onToolStart(toolName, {});
+          onToolResult(toolName, rejectedResult, 0);
+          return {
+            toolMessage: {
+              role: 'tool',
+              tool_call_id: toolCallId,
+              name: toolName,
+              content: JSON.stringify(rejectedResult),
+            },
+          };
+        }
+        if (!activeToolNames.has(toolName)) {
+          const available = [...activeToolNames].slice(0, 20).join(', ');
+          const rejectedResult = {
+            status: 'error',
+            message: `Tool "${toolName}" was not offered for this query, so calling it is not allowed. Available tools for this query: ${available}. Re-answer using the available tools or the evidence you already have.`
+          };
+          onToolStart(toolName, {});
+          onToolResult(toolName, rejectedResult, 0);
+          return {
+            toolMessage: {
+              role: 'tool',
+              tool_call_id: toolCallId,
+              name: toolName,
+              content: JSON.stringify(rejectedResult),
+            },
+          };
+        }
+        const isBuiltin = isBuiltinTool(toolName);
         const isMcp = !isBuiltin && (toolName.startsWith('mcp_') || Boolean(this.mcpManager?.isMcpTool(toolName)));
 
         let prepared;
@@ -1067,6 +1360,8 @@ ${toolsSummary}
         const toolStartTime = Date.now();
         const signature = stableToolSignature(toolName, finalArgs);
         const cachedResult = toolResultCache.get(signature);
+        const repeatCount = (repeatSignatureCounts.get(signature) || 0) + 1;
+        repeatSignatureCounts.set(signature, repeatCount);
 
         let result;
         if (cachedResult) {
@@ -1096,6 +1391,15 @@ ${toolsSummary}
 
         const cappedResult = this._capToolResultForHistory(result);
 
+        let instruction = result.status === 'not_found'
+          ? 'CRITICAL: No matching data was found for this query in the catalog. State explicitly in the user\'s language that no verified records exist in the database. Do NOT invent, fabricate, or guess any records. Do NOT claim you searched external platforms like Bilibili, Weibo, or Google unless those specific tools were called. Under no circumstances should you generate plausible-sounding but unverified details.'
+          : result.status === 'error'
+            ? 'CRITICAL: The tool failed to retrieve data. Do NOT claim that the requested facts were verified. You MUST acknowledge the error and either retry or inform the user that the information could not be retrieved.'
+            : 'Use ONLY the specific facts present in this result when answering. Do NOT add, infer, or extrapolate information beyond what is explicitly provided. Do NOT fabricate additional details, dates, names, or relationships not present in this result. Do not echo raw JSON data — summarize in natural language while preserving factual accuracy.';
+        if (cachedResult) {
+          instruction += ' NOTE: You already called this exact tool with identical arguments earlier in this query — the duplicated result above adds nothing new. Do NOT repeat this call; answer with the retrieved data now.';
+        }
+
         return {
           toolMessage: {
             role: 'tool',
@@ -1104,36 +1408,60 @@ ${toolsSummary}
               content: JSON.stringify({
               authoritative: true,
               result: cappedResult,
-              instruction: result.status === 'not_found'
-                ? 'CRITICAL: No matching data was found for this query. Do NOT invent, fabricate, or guess any records. You MUST either (a) state explicitly that no verified information exists, or (b) immediately call web_search to look up live external information. Under no circumstances should you generate plausible-sounding but unverified details.'
-                : result.status === 'error'
-                  ? 'CRITICAL: The tool failed to retrieve data. Do NOT claim that the requested facts were verified. You MUST acknowledge the error and either retry or inform the user that the information could not be retrieved.'
-                  : 'Use ONLY the specific facts present in this result when answering. Do NOT add, infer, or extrapolate information beyond what is explicitly provided. Do NOT fabricate additional details, dates, names, or relationships not present in this result. Do not echo raw JSON data — summarize in natural language while preserving factual accuracy.'
+              instruction
             })
-          }
+          },
+          toolName,
+          result
         };
       });
 
       const toolResults = await Promise.all(executionPromises);
+   
       for (const item of toolResults) {
         if (item?.toolMessage) {
           this.conversationHistory.push(item.toolMessage);
         }
+        if (item?.toolName) {
+          recordedTools.push({ name: item.toolName, result: item.result });
+        }
+      }
+
+     
+      if (shouldAbortLoop) {
+        try {
+          onNotice('Duplicate tool call signature detected by RoutingInterceptor — moving directly to final answer synthesis.');
+        } catch { }
+        break;
+      }
+
+      if (currentIteration < this.maxIterations) {
+        let stuck = false;
+        for (const count of repeatSignatureCounts.values()) {
+          if (count >= REPEAT_LOOP_THRESHOLD) { stuck = true; break; }
+        }
+        if (stuck) {
+          try {
+            onNotice('Repeated identical tool calls detected — moving to the final answer.');
+          } catch { }
+          break;
+        }
       }
     }
 
-    const selfCritiqueDirective = await this._performSelfCritique(
+    const selfCritiqueDirective = await this._performSelfCritique( 
       toolResultTracker,
       dominantLang,
       promptForLLM
     );
 
     const langPrompt = dominantLang === 'zh'
-      ? "请根据上述工具查询结果，为粉丝提供详细、全面且深入的解答。条理清晰地阐述作品背景、角色特质、音乐成就与演艺里程碑，包含具体的 YouTube 真实链接（如 Direct Watch: https://www.youtube.com/watch?v=... 或 Search: https://www.youtube.com/results?search_query=...）及完整事实数据。\n\n【反幻觉守则 — 必须遵守】只陈述工具结果中明确包含的事实。禁止编造、推测或补充工具结果中不存在的剧名、歌曲名、角色名、品牌名、日期或任何细节。如果工具结果不足以完整回答，请明确说明哪些部分未经验证，而不是凭空补全。"
+      ? "请根据上述工具查询结果回答，回答深度须与问题相称：单一事实类问题（日期、名字、数字、是/否）用 1-4 句话直接简洁作答，不加标题或表格；仅当问题是比较、排序、列表或深度分析时，才使用完整的 Markdown 结构（标题、要点、表格）并附上具体的 YouTube 真实链接（如 Direct Watch: https://www.youtube.com/watch?v=... 或 Search: https://www.youtube.com/results?search_query=...）。不要用无关段落填充简短的事实性回答，但也绝不能遗漏用户明确要求的任何条目。\n\n【反幻觉守则 — 必须遵守】只陈述工具结果中明确包含的事实。禁止编造、推测或补充工具结果中不存在的剧名、歌曲名、角色名、品牌名、日期或任何细节。如果工具结果不足以完整回答，请明确说明哪些部分未经验证，而不是凭空补全。"
       : dominantLang === 'en'
-        ? "Based on the tool results above, provide a comprehensive, richly detailed response for fans. Explain background storylines, character traits, musical highlights, and career milestones with well-organized Markdown headers, bullet points, and verified YouTube links (https://www.youtube.com/results?search_query=...).\n\n【Anti-Hallucination Mandate — MANDATORY】Only state facts explicitly present in the tool results above. Do NOT fabricate, infer, or supplement drama titles, song names, character roles, brand names, dates, or any details not found in the tool results. If the tool results are insufficient to fully answer, explicitly note which parts are unverified rather than inventing plausible-sounding details."
-        : "สรุปคำตอบจากผลลัพธ์เครื่องมือข้างต้น โดยให้ข้อมูลแฟนคลับอย่างละเอียด ลึกซึ้ง ครบถ้วนและชัดเจน จัดโครงสร้างด้วยหัวข้อ Markdown, ตาราง และ bullet points ให้สวยงาม พร้อมแสดงลิงก์ YouTube จริง (Direct Watch Link และ https://www.youtube.com/results?search_query=...) พร้อมเกร็ดประวัติและผลงานอย่างเต็มที่\n\n【กฎป้องกันการแต่งข้อมูล — บังคับใช้】ระบุเฉพาะข้อเท็จจริงที่มีอยู่จริงในผลลัพธ์เครื่องมือเท่านั้น ห้ามสร้างชื่อละคร เพลง ตัวละคร แบรนด์ วันที่ หรือรายละเอียดใดๆ ที่ไม่มีอยู่ในผลลัพธ์เครื่องมือ หากข้อมูลไม่เพียงพอสำหรับคำตอบที่สมบูรณ์ ให้ระบุชัดเจนว่าส่วนใดยังไม่ได้รับการยืนยัน แทนการเติมข้อมูลที่ฟังดูน่าเชื่อแต่ไม่มีหลักฐาน";
+        ? "Based on the tool results above, provide a response whose depth matches the question: for simple single-fact questions (a date, a name, a number, yes/no) answer concisely and directly in 1-4 sentences without headers or tables; only for comparison, ranking, list, or deep-dive questions use rich detail with well-organized Markdown headers, bullet points, and verified YouTube links (https://www.youtube.com/results?search_query=...). Never pad short factual answers with filler sections, and never omit any item the user explicitly requested.\n\n【Anti-Hallucination Mandate — MANDATORY】Only state facts explicitly present in the tool results above. Do NOT fabricate, infer, or supplement drama titles, song names, character roles, brand names, dates, or any details not found in the tool results. If the tool results are insufficient to fully answer, explicitly note which parts are unverified rather than inventing plausible-sounding details."
+        : "สรุปคำตอบจากผลลัพธ์เครื่องมือข้างต้น โดยให้ความลึกของคำตอบสมดุลกับคำถาม: คำถามข้อเท็จจริงเดี่ยว (วันที่ ชื่อ ตัวเลข ใช่/ไม่ใช่) ตอบสั้น ตรงประเด็น 1-4 ประโยค โดยไม่ต้องมีหัวข้อหรือตาราง ส่วนคำถามเปรียบเทียบ จัดอันดับ จัดรายการ หรือวิเคราะห์เชิงลึก จึงจัดโครงสร้างด้วยหัวข้อ Markdown, ตาราง และ bullet points พร้อมแสดงลิงก์ YouTube จริง (Direct Watch Link และ https://www.youtube.com/results?search_query=...) อย่างครบถ้วน ห้ามเติมเนื้อหาฟุ่มเฟือยในคำตอบสั้น แต่ห้ามลดทอนรายการที่ผู้ใช้ระบุขอมาทุกข้อ\n\n【กฎป้องกันการแต่งข้อมูล — บังคับใช้】ระบุเฉพาะข้อเท็จจริงที่มีอยู่จริงในผลลัพธ์เครื่องมือเท่านั้น ห้ามสร้างชื่อละคร เพลง ตัวละคร แบรนด์ วันที่ หรือรายละเอียดใดๆ ที่ไม่มีอยู่ในผลลัพธ์เครื่องมือ หากข้อมูลไม่เพียงพอสำหรับคำตอบที่สมบูรณ์ ให้ระบุชัดเจนว่าส่วนใดยังไม่ได้รับการยืนยัน แทนการเติมข้อมูลที่ฟังดูน่าเชื่อแต่ไม่มีหลักฐาน";
 
+    const synthHistoryBase = this.conversationHistory.length;
     this.conversationHistory.push({ role: 'user', content: langPrompt });
     this.conversationHistory.push({ role: 'system', content: selfCritiqueDirective });
 
@@ -1143,13 +1471,19 @@ ${toolsSummary}
 
     let isReasoning = false;
     let synthTokenStarted = false;
+    const synthSegment = tokenTracker.beginSegment('synthesis');
     const finalRes = await this.client.chat({
-      messages: this.conversationHistory,
+      messages: [
+        this.conversationHistory[0],
+        ...turnDirectives,
+        ...this.conversationHistory.slice(1)
+      ],
       temperature: 0.3,
-      maxTokens: 8192,
+      maxTokens: OUTPUT_TOKEN_BUDGET,
       stream: true,
       signal,
       onThought: (thoughtChunk) => {
+        tokenTracker.markToken(synthSegment);
         if (!isReasoning) {
           isReasoning = true;
           onReasoningStart();
@@ -1158,6 +1492,7 @@ ${toolsSummary}
         onReasoningDelta(thoughtChunk);
       },
       onToken: (tokenChunk) => {
+        tokenTracker.markToken(synthSegment);
         if (isReasoning) {
           isReasoning = false;
           this.session.finalizeReasoningPart(currentAssistantSessionMsg);
@@ -1172,6 +1507,14 @@ ${toolsSummary}
       }
     });
 
+   
+    this.conversationHistory.length = Math.min(this.conversationHistory.length, synthHistoryBase);
+
+    tokenTracker.endSegment({
+      promptTokens: finalRes?.usage?.prompt_tokens,
+      completionTokens: finalRes?.usage?.completion_tokens,
+    }, synthSegment);
+
     if (finalRes?.usage) {
       queryPromptTokens += Number(finalRes.usage.prompt_tokens || 0);
       queryCompletionTokens += Number(finalRes.usage.completion_tokens || 0);
@@ -1195,6 +1538,11 @@ ${toolsSummary}
     this.session.recordQueryTokens(queryTokens, queryCostUsd);
     this.lastTelemetry = { tokens: queryTokens, costUsd: queryCostUsd };
 
+    this._recordQueryMetrics(tokenTracker, {
+      promptTokens: queryPromptTokens,
+      completionTokens: queryCompletionTokens,
+    });
+
     const totalDurationMs = Date.now() - queryStartTime;
     this.session.stats.totalDurationMs += totalDurationMs;
 
@@ -1204,6 +1552,10 @@ ${toolsSummary}
           dominantLang === 'ko' ? "검색이 완료되었습니다" :
             dominantLang === 'en' ? "Search complete." : "เสร็จสิ้นการค้นหาข้อมูล";
     const output = sanitizeResponse(finalRes?.content || defaultDone);
+    this.conversationHistory.push({ role: 'assistant', content: output });
+    this.conversationHistory = this.conversationHistory.filter(m =>
+      !(typeof m?.content === 'string' && m.content.includes('[MANDATORY'))
+    );
     const citations = extractCitations(output, recordedTools);
     this.lastCitations = citations;
     if (currentAssistantSessionMsg) {
@@ -1245,6 +1597,22 @@ ${toolsSummary}
     return this.ask(userPrompt, callbacks);
   }
 
+  /** Fold the query's TokenTracker timings into lastQueryMetrics / history. */
+  _recordQueryMetrics(tokenTracker, { promptTokens, completionTokens } = {}) {
+    try {
+      if (!tokenTracker) return null;
+      const payload = tokenTracker.finish({ promptTokens, completionTokens });
+      if (payload) {
+        this.lastQueryMetrics = payload;
+        this.queryMetricsHistory.push(payload);
+        if (this.queryMetricsHistory.length > 50) this.queryMetricsHistory.shift();
+      }
+      return payload;
+    } catch {
+      return null;
+    }
+  }
+
   _computeConfidenceScore(toolResultTracker) {
     if (toolResultTracker.total === 0) return 0;
 
@@ -1312,16 +1680,17 @@ ${toolsSummary}
   _capToolResultForHistory(result) {
     if (!result || typeof result !== 'object') return result;
 
-    const MAX_RESULT_BYTES = 16384;
-    const serialized = JSON.stringify(result);
-    if (serialized.length <= MAX_RESULT_BYTES) return result;
+    const compacted = compactToolResultForContext(result, { maxTokens: 800 });
 
+    const MAX_RESULT_BYTES = 16384;
+    const serialized = JSON.stringify(compacted);
+    if (serialized.length <= MAX_RESULT_BYTES) return compacted;
 
     const truncate = (obj) => {
       if (Array.isArray(obj)) {
-        if (obj.length > 8) {
-          const kept = obj.slice(0, 8).map(truncate);
-          kept.push({ _truncated: true, totalItems: obj.length, showing: 8 });
+        if (obj.length > 5) {
+          const kept = obj.slice(0, 5).map(truncate);
+          kept.push({ _truncated: true, totalItems: obj.length, showing: 5 });
           return kept;
         }
         return obj.map(truncate);
@@ -1333,14 +1702,13 @@ ${toolsSummary}
         }
         return out;
       }
-    
-      if (typeof obj === 'string' && obj.length > 500) {
-        return `${obj.slice(0, 500)}…[truncated]`;
+      if (typeof obj === 'string' && obj.length > 300) {
+        return `${obj.slice(0, 300)}…[truncated]`;
       }
       return obj;
     };
 
-    const capped = truncate(result);
+    const capped = truncate(compacted);
     capped._cappedForHistory = true;
     return capped;
   }