kiku-agent 1.3.1 → 1.3.3

package/server/agent/eval/modules/utils/smokes.js

@@ -0,0 +1,119 @@
+import fs from 'node:fs/promises';
+import path from 'node:path';
+import os from 'node:os';
+import { REPO_ROOT, DEFAULTS } from '../core/config.js';
+import { estimateTextTokens } from '../../../contextCompactor.js';
+
+export async function runWebSmoke({ out = DEFAULTS.webSmokeOut, timeoutMs = 60000 } = {}) {
+  const outPath = path.isAbsolute(out) ? out : path.resolve(REPO_ROOT, out);
+  await fs.mkdir(path.dirname(outPath), { recursive: true });
+  process.env.KIKU_DATA_DIR = await fs.mkdtemp(path.join(os.tmpdir(), 'kiku-web-eval-'));
+  const { executeTool } = await import('../../../tools.js');
+  const checks = [
+    ['web_search', { query: 'Ju Jingyi Legend of Yunxi' }],
+    ['search_chinese_web', { query: '鞠婧祎 芸汐传' }],
+    ['fetch_webpage', { url: 'https://example.com' }],
+    ['web_extract', { query: 'Ju Jingyi Legend of Yunxi', maxResults: 2 }],
+  ];
+  const rows = [];
+  try {
+    for (const [tool, parameters] of checks) {
+      const start = performance.now();
+      let row;
+      try {
+        const result = await executeTool(tool, parameters, { signal: AbortSignal.timeout(timeoutMs) });
+        row = { tool, parameters, result, durationMs: Math.round(performance.now() - start) };
+      } catch (error) {
+        row = { tool, parameters, error: error.message, durationMs: Math.round(performance.now() - start) };
+      }
+      rows.push(row);
+      await fs.writeFile(outPath, JSON.stringify({ checkedAt: new Date().toISOString(), scope: 'Real external tool availability, not answer-quality scoring. Success status does not establish result relevance.', rows }, null, 2));
+      console.log(`${tool}: ${row.result?.status || row.error}; ${row.durationMs}ms`);
+    }
+  } finally {
+    await fs.rm(process.env.KIKU_DATA_DIR, { recursive: true, force: true });
+  }
+  return { out: outPath, rows };
+}
+
+export async function runResearchSmoke({ out = DEFAULTS.researchSmokeOut, model = DEFAULTS.model, baseUrl = DEFAULTS.baseUrl, timeoutMs = 180000 } = {}) {
+  const outPath = path.isAbsolute(out) ? out : path.resolve(REPO_ROOT, out);
+  const fsSync = await import('node:fs');
+  const dir = fsSync.mkdtempSync(path.join(os.tmpdir(), 'kiku-research-eval-'));
+  process.env.KIKU_DATA_DIR = dir;
+  const [{ LLMClient }, { ResearchOrchestrator }, { scoreResearchComplexity, shouldAutoResearch }] = await Promise.all([
+    import('../../../llmClient.js'), import('../../../research.js'), import('../../../agentCore.js'),
+  ]);
+  const question = 'Compare In Blossom and Sword and Fairy 4 using the Kiku catalog and public sources. Verify the release years, the characters played by Ju Jingyi, and episode counts. Explain any source disagreement, and clearly distinguish unverified information. Provide a concise evidence report with sources.';
+  const client = new LLMClient({ provider: 'lmstudio', model, apiKey: 'lm-studio', baseUrl });
+  const started = Date.now();
+  const report = {
+    startedAt: new Date().toISOString(), model: client.model, question, timeoutMs,
+    complexity: scoreResearchComplexity(question), autoEligible: shouldAutoResearch(question, { mode: 'auto' }),
+    scope: 'Direct real ResearchOrchestrator run, default 3 workers and 3 iterations. Auto eligibility is checked separately; this does not test the main agent decision to invoke research.',
+    state: 'running', maxConcurrentRequests: 0, requests: [], progress: [],
+  };
+  fsSync.mkdirSync(path.dirname(outPath), { recursive: true });
+  const save = () => { report.elapsedMs = Date.now() - started; fsSync.writeFileSync(outPath, JSON.stringify(report, null, 2)); };
+  let active = 0;
+  const chat = client.chat.bind(client);
+  client.chat = async options => {
+    const row = { startedMs: Date.now() - started };
+    report.requests.push(row);
+    report.maxConcurrentRequests = Math.max(report.maxConcurrentRequests, ++active);
+    save();
+    try { const result = await chat(options); row.usage = result?.usage; return result; }
+    catch (error) { row.error = error.message; throw error; }
+    finally { active--; row.endedMs = Date.now() - started; save(); }
+  };
+  const research = new ResearchOrchestrator({ client });
+  const hardStop = setTimeout(() => {
+    report.state = 'timeout'; report.error = 'Hard timeout after cancellation grace period'; save(); process.exit(1);
+  }, timeoutMs + 15000);
+  try {
+    report.result = await research.run(question, {
+      signal: AbortSignal.timeout(timeoutMs),
+      onProgress: event => { report.progress.push({ elapsedMs: Date.now() - started, ...event }); save(); console.log(JSON.stringify(event)); },
+    });
+    report.state = 'completed';
+  } catch (error) {
+    report.state = 'error'; report.error = error.message; process.exitCode = 1;
+  } finally {
+    clearTimeout(hardStop); research.cancel(); save(); fsSync.rmSync(dir, { recursive: true, force: true });
+  }
+  console.log(JSON.stringify({ state: report.state, elapsedMs: report.elapsedMs, requests: report.requests.length, maxConcurrentRequests: report.maxConcurrentRequests }));
+  return report;
+}
+
+export async function runTokenCalibration({ out = DEFAULTS.tokenCalibrationOut, model = DEFAULTS.model, baseUrl = DEFAULTS.baseUrl } = {}) {
+  const outPath = path.isAbsolute(out) ? out : path.resolve(REPO_ROOT, out);
+  const samples = {
+    th: 'จวีจิ้งอีเป็นนักร้องและนักแสดงชาวจีน แฟนคลับต้องการตรวจสอบข้อมูลผลงานจากแหล่งอ้างอิง หากไม่มีข้อมูลเพียงพอควรบอกตรงไปตรงมาและไม่แต่งรายละเอียดเพิ่มเติม ',
+    en: 'The assistant should verify the title, character and release year against the catalog. If evidence is missing, explain the uncertainty instead of inventing details. ',
+    zh: '请根据资料库核实电视剧名称、角色和发行年份。如果没有足够证据，请明确说明不确定性，不要编造细节。',
+    mixed: 'ข้อมูล作品 In Blossom (花间令) ปี 2024: {"role":"Yang Caiwei","episodes":32,"verified":true}. ',
+  };
+  const rows = [];
+  for (const [language, sample] of Object.entries(samples)) {
+    const measurements = [];
+    for (const repeats of [3, 6]) {
+      const content = sample.repeat(repeats).trim();
+      const started = Date.now();
+      const response = await fetch(`${baseUrl.replace(/\/$/, '')}/chat/completions`, {
+        method: 'POST', headers: { 'Content-Type': 'application/json' }, signal: AbortSignal.timeout(90000),
+        body: JSON.stringify({ model, messages: [{ role: 'user', content }], max_tokens: 1, temperature: 0, stream: false }),
+      });
+      if (!response.ok) throw new Error(`Calibration HTTP ${response.status}`);
+      const result = await response.json();
+      if (!Number.isFinite(result.usage?.prompt_tokens)) throw new Error('Provider did not report prompt token usage');
+      measurements.push({ repeats, characters: content.length, estimate: estimateTextTokens(content), promptTokens: result.usage.prompt_tokens, durationMs: Date.now() - started });
+    }
+    const estimatedDelta = measurements[1].estimate - measurements[0].estimate;
+    const measuredDelta = measurements[1].promptTokens - measurements[0].promptTokens;
+    rows.push({ language, sample, measurements, estimatedDelta, measuredDelta, ratio: measuredDelta > 0 ? estimatedDelta / measuredDelta : null });
+    await fs.mkdir(path.dirname(outPath), { recursive: true });
+    await fs.writeFile(outPath, JSON.stringify({ model, checkedAt: new Date().toISOString(), method: 'Real prompt usage delta between 3 and 6 repeats; chat-template overhead cancels. Four samples only, no universal tokenizer bound or o200k claim.', rows }, null, 2));
+    console.log(`${language}: estimated delta ${estimatedDelta}, measured ${measuredDelta}, ratio ${rows.at(-1).ratio}`);
+  }
+  return { out: outPath, rows };
+}

package/server/agent/eventBus.js

@@ -83,6 +83,7 @@ export const Events = Object.freeze({
   AGENT_TOOL_START:    'agent:tool:start',
   AGENT_TOOL_RESULT:   'agent:tool:result',
   AGENT_TOOL_ERROR:    'agent:tool:error',
+  AGENT_TOOL_ESCALATED:'agent:tool:escalated',
 
   AGENT_PLAN_START:       'agent:plan:start',
   AGENT_PLAN_STEP_UPDATE: 'agent:plan:step:update',

package/server/agent/ink/main.jsx

@@ -2,194 +2,199 @@ import React from 'react';
 import { render } from 'ink';
 import '../youtubeAuth.js';
 import { App } from './App.jsx';
-import { KikuAgent } from '../agentCore.js';
-import { LMStudioClient } from '../llmClient.js';
-import { toolsSchema } from '../tools.js';
-import { config } from '../config.js';
-import { COMMAND_DEFINITIONS } from '../commandRegistry.js';
-import { C, UI_ICONS } from '../ui.js';
-import { existsSync } from 'node:fs';
-import { spawn } from 'node:child_process';
-import { PACKAGE_VERSION } from '../version.js';
-
-function parseCliArgs() {
-  const args = process.argv.slice(2);
-  const options = {
-    query: null,
-    model: 'google/gemma-4-e4b',
-    url: null,
-    showTools: false,
-    showStats: false,
-    thinking: 'show',
-    verbose: false,
-    help: false,
-    version: false
-  };
-
-  for (let i = 0; i < args.length; i++) {
-    const arg = args[i];
-    if (arg === '--help' || arg === '-h') options.help = true;
-    else if (arg === '--version' || arg === '-V') options.version = true;
-    else if (arg === '--tools' || arg === '-t') options.showTools = true;
-    else if (arg === '--stats' || arg === '-s') options.showStats = true;
-    else if (arg === '--verbose' || arg === '-v') { options.verbose = true; options.thinking = 'show'; }
-    else if (arg === '--thinking') options.thinking = args[++i] || 'show';
-    else if (arg === '--model' || arg === '-m') options.model = args[++i] || options.model;
-    else if (arg === '--url' || arg === '-u') options.url = args[++i];
-    else if (arg === '--query' || arg === '-q') options.query = args[++i];
-    else if (!arg.startsWith('-') && !options.query) { options.query = args.slice(i).join(' '); break; }
-  }
-  return options;
-}
-
-async function waitForLMStudio(baseUrl, timeoutMs = 30000) {
-  const client = new LMStudioClient({ baseUrl });
-  const start = Date.now();
-  while (Date.now() - start < timeoutMs) {
-    const health = await client.checkConnection();
-    if (health.connected) return health;
-    await new Promise(r => setTimeout(r, 1000));
-  }
-  return null;
-}
-
-async function attemptAutoStartLMStudio(baseUrl, model) {
-  const isWin = process.platform === 'win32';
-  const isMac = process.platform === 'darwin';
-
-  const candidates = isWin
-    ? [
-        `${process.env.LOCALAPPDATA}\\LM Studio\\LM Studio.exe`,
-        `${process.env.LOCALAPPDATA}\\Programs\\LM Studio\\LM Studio.exe`,
-        'C:\\Program Files\\LM Studio\\LM Studio.exe'
-      ]
-    : isMac
-      ? ['/Applications/LM Studio.app/Contents/MacOS/LM Studio']
-      : ['lmstudio', '/usr/local/bin/lmstudio', '/opt/lmstudio/bin/lmstudio'];
-
-  const exe = candidates.find(p => {
-    if (p.includes('/') || p.includes('\\')) {
-      try { return existsSync(p); } catch { return false; }
-    }
-    return true;
-  });
-
-  if (!exe) return null;
-
-  try {
-    let port = '1234';
-    try { const parsedUrl = new URL(baseUrl); if (parsedUrl.port) port = parsedUrl.port; } catch {}
-
-    const child = spawn(exe, isWin
-      ? ['--model', model, '--serve', '--port', port]
-      : ['serve', '--model', model, '--port', port],
-      { detached: true, stdio: 'ignore', windowsHide: true }
-    );
-    child.unref();
-
-    return await waitForLMStudio(baseUrl, 45000);
-  } catch { return null; }
-}
-
-function printHelpText() {
-  const commands = COMMAND_DEFINITIONS
-    .map(command => `  ${C.emerald}${command.usage.padEnd(30)}${C.reset} ${command.desc}`)
-    .join('\n');
-  console.log(`\n  ${C.bold}${C.gold}${UI_ICONS.help} Kiku AI Agent — ${COMMAND_DEFINITIONS.length} Commands${C.reset}\n${commands}\n  ${C.dim}Tip: Press [Tab] for auto-completion.${C.reset}\n`);
-}
-
-async function printStatsText() {
-  const { executeTool } = await import('../tools.js');
-  const result = await executeTool('get_system_stats', {});
-  const s = result.stats;
-  console.log(JSON.stringify(s, null, 2));
-}
-
-async function main() {
-  const opts = parseCliArgs();
-
-  if (opts.version) { console.log(PACKAGE_VERSION); process.exit(0); }
-  if (opts.help) { printHelpText(); process.exit(0); }
-  if (opts.showStats) { await printStatsText(); process.exit(0); }
-
-  if (opts.showTools) {
-    console.log(`\n  ${C.bold}${C.sakura}Tools (${toolsSchema.length})${C.reset}\n`);
-    for (const [i, t] of toolsSchema.entries()) {
-      console.log(`  ${C.gold}${String(i + 1).padStart(2)}${C.reset} ${C.cyberCyan}${t.function.name}${C.reset} — ${C.dim}${t.function.description.slice(0, 60)}${C.reset}`);
-    }
-    process.exit(0);
-  }
-
-  const baseUrl = opts.url || config.lmStudio.baseUrl;
-  const client = new LMStudioClient({ baseUrl, model: opts.model });
-  const health = await client.checkConnection();
-
-  if (!health.connected) {
-    console.log(`  ${C.amber}!! LM Studio offline${C.reset}`);
-    console.log(`  ${C.dim}• ${health.error}${C.reset}\n`);
-
-    const autoHealth = await attemptAutoStartLMStudio(baseUrl, opts.model || config.lmStudio.model);
-    if (autoHealth) {
-      console.log(`  ${C.emerald}[OK] LM Studio started.${C.reset}\n`);
-    } else {
-      console.log(`  ${C.sakura}Setup: Open LM Studio → load google/gemma-4-e4b → Start Server :1234${C.reset}\n`);
-      process.exit(1);
-    }
-  }
-
-  const agent = new KikuAgent({
-    baseUrl,
-    model: health.activeModel || opts.model,
-    thinkingMode: opts.thinking,
-  });
-
-  await agent._ready;
-
-  if (opts.query) {
-    const { executeTool: execTool } = await import('../tools.js');
-    const { sanitizeQuery } = await import('../sanitizer.js');
-    const { sanitized } = sanitizeQuery(opts.query);
-    const prompt = sanitized || opts.query;
-
-    console.log(`  ${C.bold}${C.gold}QUERY${C.reset} ${prompt}\n`);
-
-    await agent.ask(prompt, {
-      onReasoningStart: () => process.stdout.write('  THINKING '),
-      onReasoningDelta: (chunk) => process.stdout.write(chunk),
-      onReasoningEnd: () => process.stdout.write('\n'),
-      onToolStart: (name) => process.stdout.write(`  [RUNNING] ${name}... `),
-      onToolResult: () => process.stdout.write(`[OK]\n`),
-      onTokenDelta: (chunk) => process.stdout.write(chunk),
-      onFinalAnswer: () => {},
-      onError: (err) => console.error(`\n  [ERROR] ${err.message}`),
-    });
-    process.exit(0);
-  }
-
-  const helpFn = (sub) => { printHelpText(); };
-  const adFn = () => { console.log(`  ${UI_ICONS.sponsor} SPONSOR  DashSecurity — https://dashsecurity.kpltgroup.com/`); };
-
-  const { waitUntilExit } = render(
-    <App
-      agent={agent}
-      client={client}
-      baseUrl={baseUrl}
-      printHelp={helpFn}
-      printAd={adFn}
-    />,
-    {
-      exitOnCtrlC: true,
-      patchConsole: false,
-    }
-  );
-
-  await waitUntilExit;
-
-  try { await agent.endSession(); } catch {}
-  console.log('\n  Session archived. Goodbye. (鞠躬 ~ 谢谢!)\n');
-}
-
-main().catch(err => {
-  console.error(`\n  ${C.red}Fatal: ${err.message}${C.reset}\n`);
-  process.exit(1);
-});
+import { KikuAgent } from '../agentCore.js';
+import { LMStudioClient } from '../llmClient.js';
+import { toolsSchema } from '../tools.js';
+import { config } from '../config.js';
+import { COMMAND_DEFINITIONS } from '../commandRegistry.js';
+import { C, UI_ICONS } from '../ui.js';
+import { existsSync } from 'node:fs';
+import { spawn } from 'node:child_process';
+import { PACKAGE_VERSION } from '../version.js';
+import { resolveDynamicWindowsAppPaths } from '../appPaths.js';
+import { registerBackgroundProcess } from '../processManager.js';
+
+function parseCliArgs() {
+  const args = process.argv.slice(2);
+  const options = {
+    query: null,
+    model: 'google/gemma-4-e4b',
+    url: null,
+    showTools: false,
+    showStats: false,
+    thinking: 'show',
+    verbose: false,
+    help: false,
+    version: false
+  };
+
+  for (let i = 0; i < args.length; i++) {
+    const arg = args[i];
+    if (arg === '--help' || arg === '-h') options.help = true;
+    else if (arg === '--version' || arg === '-V') options.version = true;
+    else if (arg === '--tools' || arg === '-t') options.showTools = true;
+    else if (arg === '--stats' || arg === '-s') options.showStats = true;
+    else if (arg === '--verbose' || arg === '-v') { options.verbose = true; options.thinking = 'show'; }
+    else if (arg === '--thinking') options.thinking = args[++i] || 'show';
+    else if (arg === '--model' || arg === '-m') options.model = args[++i] || options.model;
+    else if (arg === '--url' || arg === '-u') options.url = args[++i];
+    else if (arg === '--query' || arg === '-q') options.query = args[++i];
+    else if (!arg.startsWith('-') && !options.query) { options.query = args.slice(i).join(' '); break; }
+  }
+  return options;
+}
+
+async function waitForLMStudio(baseUrl, timeoutMs = 30000) {
+  const client = new LMStudioClient({ baseUrl });
+  const start = Date.now();
+  while (Date.now() - start < timeoutMs) {
+    const health = await client.checkConnection();
+    if (health.connected) return health;
+    await new Promise(r => setTimeout(r, 1000));
+  }
+  return null;
+}
+
+async function attemptAutoStartLMStudio(baseUrl, model) {
+  const isWin = process.platform === 'win32';
+  const isMac = process.platform === 'darwin';
+
+  const candidates = isWin
+    ? resolveDynamicWindowsAppPaths('LM Studio', 'LM Studio.exe')
+    : isMac
+      ? ['/Applications/LM Studio.app/Contents/MacOS/LM Studio']
+      : ['lmstudio', '/usr/local/bin/lmstudio', '/opt/lmstudio/bin/lmstudio'];
+
+  const exe = candidates.find(p => {
+    if (p.includes('/') || p.includes('\\')) {
+      try { return existsSync(p); } catch { return false; }
+    }
+    return true;
+  });
+
+  if (!exe) return null;
+
+  try {
+    let port = '1234';
+    try { const parsedUrl = new URL(baseUrl); if (parsedUrl.port) port = parsedUrl.port; } catch {}
+
+    const child = spawn(exe, isWin
+      ? ['--model', model, '--serve', '--port', port]
+      : ['serve', '--model', model, '--port', port],
+      { detached: true, stdio: 'ignore', windowsHide: true }
+    );
+    if (child.pid) {
+      registerBackgroundProcess('lmstudio', child, {
+        command: exe,
+        args: isWin ? ['--model', model, '--serve', '--port', port] : ['serve', '--model', model, '--port', port],
+        metadata: { model, port },
+      });
+    }
+    child.unref();
+
+    return await waitForLMStudio(baseUrl, 45000);
+  } catch { return null; }
+}
+
+function printHelpText() {
+  const commands = COMMAND_DEFINITIONS
+    .map(command => `  ${C.emerald}${command.usage.padEnd(30)}${C.reset} ${command.desc}`)
+    .join('\n');
+  console.log(`\n  ${C.bold}${C.gold}${UI_ICONS.help} Kiku AI Agent — ${COMMAND_DEFINITIONS.length} Commands${C.reset}\n${commands}\n  ${C.dim}Tip: Press [Tab] for auto-completion.${C.reset}\n`);
+}
+
+async function printStatsText() {
+  const { executeTool } = await import('../tools.js');
+  const result = await executeTool('get_system_stats', {});
+  const s = result.stats;
+  console.log(JSON.stringify(s, null, 2));
+}
+
+async function main() {
+  const opts = parseCliArgs();
+
+  if (opts.version) { console.log(PACKAGE_VERSION); process.exit(0); }
+  if (opts.help) { printHelpText(); process.exit(0); }
+  if (opts.showStats) { await printStatsText(); process.exit(0); }
+
+  if (opts.showTools) {
+    console.log(`\n  ${C.bold}${C.sakura}Tools (${toolsSchema.length})${C.reset}\n`);
+    for (const [i, t] of toolsSchema.entries()) {
+      console.log(`  ${C.gold}${String(i + 1).padStart(2)}${C.reset} ${C.cyberCyan}${t.function.name}${C.reset} — ${C.dim}${t.function.description.slice(0, 60)}${C.reset}`);
+    }
+    process.exit(0);
+  }
+
+  const baseUrl = opts.url || config.lmStudio.baseUrl;
+  const client = new LMStudioClient({ baseUrl, model: opts.model });
+  const health = await client.checkConnection();
+
+  if (!health.connected) {
+    console.log(`  ${C.amber}!! LM Studio offline${C.reset}`);
+    console.log(`  ${C.dim}• ${health.error}${C.reset}\n`);
+
+    const autoHealth = await attemptAutoStartLMStudio(baseUrl, opts.model || config.lmStudio.model);
+    if (autoHealth) {
+      console.log(`  ${C.emerald}[OK] LM Studio started.${C.reset}\n`);
+    } else {
+      console.log(`  ${C.sakura}Setup: Open LM Studio → load google/gemma-4-e4b → Start Server :1234${C.reset}\n`);
+      process.exit(1);
+    }
+  }
+
+  const agent = new KikuAgent({
+    baseUrl,
+    model: health.activeModel || opts.model,
+    thinkingMode: opts.thinking,
+  });
+
+  await agent._ready;
+
+  if (opts.query) {
+    const { executeTool: execTool } = await import('../tools.js');
+    const { sanitizeQuery } = await import('../sanitizer.js');
+    const { sanitized } = sanitizeQuery(opts.query);
+    const prompt = sanitized || opts.query;
+
+    console.log(`  ${C.bold}${C.gold}QUERY${C.reset} ${prompt}\n`);
+
+    await agent.ask(prompt, {
+      onReasoningStart: () => process.stdout.write('  THINKING '),
+      onReasoningDelta: (chunk) => process.stdout.write(chunk),
+      onReasoningEnd: () => process.stdout.write('\n'),
+      onToolStart: (name) => process.stdout.write(`  [RUNNING] ${name}... `),
+      onToolResult: () => process.stdout.write(`[OK]\n`),
+      onTokenDelta: (chunk) => process.stdout.write(chunk),
+      onFinalAnswer: () => {},
+      onError: (err) => console.error(`\n  [ERROR] ${err.message}`),
+    });
+    process.exit(0);
+  }
+
+  const helpFn = (sub) => { printHelpText(); };
+  const adFn = () => { console.log(`  ${UI_ICONS.sponsor} SPONSOR  DashSecurity — https://dashsecurity.kpltgroup.com/`); };
+
+  const { waitUntilExit } = render(
+    <App
+      agent={agent}
+      client={client}
+      baseUrl={baseUrl}
+      printHelp={helpFn}
+      printAd={adFn}
+    />,
+    {
+      exitOnCtrlC: true,
+      patchConsole: false,
+    }
+  );
+
+  await waitUntilExit;
+
+  try { await agent.endSession(); } catch {}
+  console.log('\n  Session archived. Goodbye. (鞠躬 ~ 谢谢!)\n');
+}
+
+main().catch(err => {
+  console.error(`\n  ${C.red}Fatal: ${err.message}${C.reset}\n`);
+  process.exit(1);
+});

package/server/agent/jobManager.js

@@ -3,8 +3,9 @@ function errorMessage(error) {
 }
 
 export class AgentJobManager {
-  constructor(limit = 50) {
+  constructor(limit = 50, maxConcurrentHeavy = 2) {
     this.limit = limit;
+    this.maxConcurrentHeavy = maxConcurrentHeavy;
     this.jobs = [];
   }
 
@@ -28,6 +29,15 @@ export class AgentJobManager {
   }
 
   async run(type, label, runner, options = {}) {
+    const isHeavy = type === 'research' || type === 'verify';
+    const limit = options.maxConcurrent ?? (isHeavy ? this.maxConcurrentHeavy : Infinity);
+    if (Number.isFinite(limit)) {
+      const activeCount = this.jobs.filter(j => (j.type === type || (isHeavy && (j.type === 'research' || j.type === 'verify'))) && j.state === 'running').length;
+      if (activeCount >= limit) {
+        throw new Error(`Too many concurrent ${type} tasks running (${activeCount}/${limit}). Please wait for the active task to complete.`);
+      }
+    }
+
     const job = this.create(type, label);
     const abortParent = () => job.controller.abort(options.signal?.reason);
     options.signal?.addEventListener('abort', abortParent, { once: true });