crewswarm 0.9.2 → 0.9.4

package/lib/crew-lead/chat-handler.mjs.bak

@@ -1,1274 +0,0 @@
-/**
- * Chat handler — extracted from crew-lead.mjs
- * Handles user messages: LLM response, dispatch, pipelines, skills, tools.
- */
-
-import fs from "node:fs";
-import path from "node:path";
-import os from "node:os";
-import crypto from "node:crypto";
-import {
-  isSharedMemoryAvailable,
-  recallMemoryContext,
-  rememberFact,
-  getMemoryStats,
-  getKeeperStats,
-  searchMemory,
-} from "../memory/shared-adapter.mjs";
-
-let _deps = {};
-
-export function initChatHandler(deps) {
-  _deps = { ...deps };
-}
-
-export async function handleChat({ message, sessionId = "default", userId = "default", firstName = "User", projectId = null }) {
-  const cfg = _deps.loadConfig();
-  let history = _deps.loadHistory(userId, sessionId);
-
-  // Fetch project context early (needed for both memory and project context injection)
-  let activeProjectOutputDir = null;
-  let projectContext = "";
-  let currentProject = null;
-  
-  if (projectId) {
-    try {
-      const projRes = await fetch(`${_deps.DASHBOARD}/api/projects`, { signal: AbortSignal.timeout(2000) });
-      const projData = await projRes.json();
-      currentProject = (projData.projects || []).find(p => p.id === projectId);
-      if (currentProject?.outputDir) {
-        activeProjectOutputDir = currentProject.outputDir;
-        const roadmapNote = currentProject.roadmapFile ? `\nROADMAP: ${currentProject.roadmapFile}` : "";
-        projectContext = `\n\n[Active project: "${currentProject.name}" at ${currentProject.outputDir}${roadmapNote}. Use this path only when dispatching (in the task or context)—do not repeat the path in every reply. When dispatching to crew-qa, specify: Write your report to ${currentProject.outputDir}/qa-report.md.]`;
-      }
-    } catch { /* project lookup failed — proceed without context */ }
-  }
-
-  // Inject shared memory once at session start — lands in history and gets
-  // prefix-cached on all subsequent calls (effectively free after first message).
-  // Now uses CLI's MemoryBroker for unified retrieval (AgentKeeper + AgentMemory + Collections).
-  if (history.length === 0) {
-    try {
-      let memoryContext = '';
-      
-      // Try CLI's MemoryBroker first (blends all memory sources)
-      if (isSharedMemoryAvailable()) {
-        const projectDir = activeProjectOutputDir || process.cwd();
-        memoryContext = await recallMemoryContext(projectDir, 'session initialization chat context', {
-          maxResults: 8,
-          includeDocs: true,
-          includeCode: false,
-          preferSuccessful: true,
-          crewId: 'crew-lead',
-          userId: userId  // NEW: user-scoped memory
-        });
-      }
-      
-      // Fallback to legacy brain.md files if shared memory not available or empty
-      if (!memoryContext) {
-        const memDir = path.join(process.cwd(), "memory");
-        const homeDir = os.homedir();
-        const readMem = (p) => { try { return fs.readFileSync(p, "utf8").trim(); } catch { return ""; } };
-
-        const brain     = readMem(_deps.BRAIN_PATH);
-        const lessons   = readMem(path.join(memDir, "lessons.md"));
-        const decisions = readMem(path.join(memDir, "decisions.md"));
-        const rules     = readMem(path.join(homeDir, ".crewswarm", "global-rules.md"));
-
-        const sections = [];
-        if (brain)     sections.push(`## Shared Brain (accumulated facts)\n${brain}`);
-        if (lessons)   sections.push(`## Lessons Learned\n${lessons}`);
-        if (decisions) sections.push(`## Key Decisions\n${decisions}`);
-        if (rules)     sections.push(`## Global Rules\n${rules}`);
-
-        if (sections.length > 0) {
-          memoryContext = `[Shared memory — injected once at session start]\n${sections.join("\n\n")}`;
-        }
-      }
-      
-      if (memoryContext) {
-        _deps.appendHistory(userId, sessionId, "system", memoryContext);
-        history = _deps.loadHistory(userId, sessionId);
-        console.log(`[crew-lead] shared memory injected into session ${userId}:${sessionId} (${memoryContext.length} chars, source: ${isSharedMemoryAvailable() ? 'MemoryBroker' : 'legacy brain.md'})`);
-      }
-    } catch (e) {
-      console.error("[crew-lead] shared memory injection failed:", e.message);
-    }
-  }
-
-  // If a project is active, inject its brain + ROADMAP + PDD once per session —
-  // all cached in the history prefix so subsequent messages get them for free.
-  // If the project changes mid-session, inject the new project context with a
-  // clear "switched to" marker so crew-lead knows which project is now active.
-  if (currentProject?.outputDir) {
-    try {
-      const projName = currentProject.name || projectId;
-      const outDir   = currentProject.outputDir;
-
-      // Check if THIS specific project's context is already in history
-      const alreadyInjected = history.some(h => h.content?.includes(`[Project memory — ${projName}`));
-
-      if (!alreadyInjected) {
-        const readSafe = (p) => { try { return fs.readFileSync(p, "utf8").trim(); } catch { return ""; } };
-
-        const projBrain = readSafe(path.join(outDir, ".crewswarm", "brain.md"));
-        const roadmap   = readSafe(path.join(outDir, "ROADMAP.md"));
-        const pddPaths  = ["PDD.md", "pdd.md", `${projName}-pdd.md`, "design.md", "DESIGN.md"];
-        const pdd       = pddPaths.map(f => readSafe(path.join(outDir, f))).find(c => c) || "";
-
-        // Detect mid-session project switch (another project was already injected)
-        const prevProject = history.find(h => h.content?.includes("[Project memory —"));
-        const isSwitch = !!prevProject;
-        const switchNote = isSwitch ? `⚠️ User switched active project to "${projName}". Previous project context above is no longer active — use this project's context from here on.\n\n` : "";
-
-        const sections = [];
-        if (projBrain) sections.push(`### Project Brain\n${projBrain}`);
-        if (roadmap)   sections.push(`### ROADMAP\n${roadmap}`);
-        if (pdd)       sections.push(`### PDD / Design\n${pdd.slice(0, 2000)}`);
-
-        const combined = `[Project memory — ${projName} at ${outDir}]\n${switchNote}${sections.join("\n\n") || "(no project files found yet)"}`;
-        _deps.appendHistory(userId, sessionId, "system", combined);
-        history = _deps.loadHistory(sessionId);
-        console.log(`[crew-lead] project context ${isSwitch ? "switched" : "injected"} for "${projName}": brain=${projBrain.length} roadmap=${roadmap.length} pdd=${pdd.length} chars`);
-      }
-    } catch (e) {
-      console.error("[crew-lead] project context injection failed:", e.message);
-    }
-  }
-
-  // ── Direct @@DISPATCH in user message — bypass LLM and execute immediately ──
-  // This prevents the LLM from intercepting user-issued dispatch commands
-  const userDispatch = _deps.parseDispatch(message, message);
-  if (userDispatch) {
-    console.log(`[crew-lead] Direct @@DISPATCH detected from user — bypassing LLM`);
-    const resolvedAgent = _deps.resolveAgentId(cfg, userDispatch.agent) || userDispatch.agent;
-    if (cfg.knownAgents.includes(resolvedAgent)) {
-      userDispatch.agent = resolvedAgent;
-      const pipelineMeta = activeProjectOutputDir ? { projectDir: activeProjectOutputDir } : null;
-      const taskId = _deps.dispatchTask(resolvedAgent, userDispatch, sessionId, pipelineMeta);
-      if (taskId) {
-        const reply = `Dispatched to ${resolvedAgent} — reply will show here when they finish.`;
-        _deps.appendHistory(userId, sessionId, "user", message);
-        _deps.appendHistory(userId, sessionId, "system", `You dispatched to ${resolvedAgent}: "${(userDispatch.task || "").slice(0, 200)}".`);
-        _deps.appendHistory(userId, sessionId, "assistant", reply);
-        _deps.broadcastSSE({ type: "chat_message", sessionId, role: "user", content: message });
-        _deps.broadcastSSE({ type: "chat_message", sessionId, role: "assistant", content: reply });
-        return { reply, dispatched: userDispatch, pendingProject: null, pipeline: null };
-      }
-    } else {
-      const reply = `Agent "${userDispatch.agent}" not found. Available: ${cfg.knownAgents.join(", ")}`;
-      _deps.appendHistory(userId, sessionId, "user", message);
-      _deps.appendHistory(userId, sessionId, "assistant", reply);
-      _deps.broadcastSSE({ type: "chat_message", sessionId, role: "user", content: message });
-      _deps.broadcastSSE({ type: "chat_message", sessionId, role: "assistant", content: reply });
-      return { reply, dispatched: null, pendingProject: null, pipeline: null };
-    }
-  }
-
-  // ── Programmatic service control — fire immediately, don't wait for LLM ──────
-  // This prevents the LLM from ever claiming it "can't" restart services.
-  const serviceIntent = _deps.parseServiceIntent(message);
-  if (serviceIntent) {
-    const { action, id } = serviceIntent;
-    const actionLabel = action === "stop" ? "stopped" : "restarted";
-    try {
-      const isSpecificAgent = id.startsWith("crew-") && id !== "crew-lead";
-      let ok = false;
-      if (isSpecificAgent) {
-        const r = await fetch(`http://127.0.0.1:${_deps.PORT}/api/agents/${id}/restart`, {
-          method: "POST",
-          headers: _deps.getRTToken() ? { authorization: `Bearer ${_deps.getRTToken()}` } : {},
-          signal: AbortSignal.timeout(8000),
-        });
-        ok = (await r.json())?.ok !== false;
-      } else {
-        const endpoint = action === "stop" ? "/api/services/stop" : "/api/services/restart";
-        const r = await fetch(`${_deps.DASHBOARD}${endpoint}`, {
-          method: "POST",
-          headers: { "content-type": "application/json" },
-          body: JSON.stringify({ id }),
-          signal: AbortSignal.timeout(8000),
-        });
-        const data = await r.json();
-        if (data?.ok === false && data?.message) {
-          const reply = `${data.message}`;
-          _deps.appendHistory(userId, sessionId, "user", message);
-          _deps.appendHistory(userId, sessionId, "assistant", reply);
-          _deps.broadcastSSE({ type: "chat_message", sessionId, role: "user", content: message });
-          _deps.broadcastSSE({ type: "chat_message", sessionId, role: "assistant", content: reply });
-          return { reply, dispatched: null, pendingProject: null, pipeline: null };
-        }
-        ok = true;
-      }
-      if (ok) {
-        const reply = `On it — **${id}** ${actionLabel}. Give it 3–5 seconds to reconnect to the RT bus, then ask me "agents online?" for a fresh count.`;
-        _deps.appendHistory(userId, sessionId, "user", message);
-        _deps.appendHistory(userId, sessionId, "assistant", reply);
-        _deps.appendHistory(userId, sessionId, "system", `Service ${id} ${actionLabel} via direct intent detection.`);
-        _deps.broadcastSSE({ type: "chat_message", sessionId, role: "user", content: message });
-        _deps.broadcastSSE({ type: "chat_message", sessionId, role: "assistant", content: reply });
-        console.log(`[crew-lead] service intent: ${action} ${id} → ok`);
-        return { reply, dispatched: null, pendingProject: null, pipeline: null };
-      }
-    } catch (e) {
-      console.error(`[crew-lead] service intent failed: ${e.message}`);
-      // fall through to normal LLM response on error
-    }
-  }
-
-  // Autonomous PM loop: user says "run until done" / "build until done" → we auto-ping PM on each handback
-  if (/run\s+until\s+done|autonomous\s+build|build\s+until\s+done/i.test(message.trim())) {
-    _deps.autonomousPmLoopSessions.add(sessionId);
-  }
-  if (/stop\s+autonomous|stop\s+(the\s+)?build/i.test(message.trim())) {
-    _deps.autonomousPmLoopSessions.delete(sessionId);
-  }
-
-  // Hard kill: "kill everything", "kill all agents", "nuke it" — SIGTERM bridges + PM loops
-  if (/\bkill\s+(everything|all|it\s+all|all\s+agents?|the\s+agents?)\b|\bnuke\s+it\b|\bnuke\s+everything\b/i.test(message.trim())) {
-    const pipelineCancelled = _deps.cancelAllPipelines(sessionId);
-    _deps.autonomousPmLoopSessions.clear();
-    let pmLoopsKilled = 0, bridgesKilled = 0;
-    try {
-      const { execSync } = await import("node:child_process");
-      const logsDir = _deps.orchestratorLogsDir;
-      if (fs.existsSync(logsDir)) {
-        for (const f of fs.readdirSync(logsDir)) {
-          if (f.startsWith("pm-loop") && f.endsWith(".pid")) {
-            fs.writeFileSync(path.join(logsDir, f.replace(".pid", ".stop")), new Date().toISOString());
-            const pid = parseInt(fs.readFileSync(path.join(logsDir, f), "utf8").trim(), 10);
-            if (pid) { try { process.kill(pid, "SIGTERM"); pmLoopsKilled++; } catch {} }
-          }
-        }
-      }
-      try { execSync(`pkill -f "gateway-bridge.mjs" 2>/dev/null`, { stdio: "ignore", shell: true }); bridgesKilled = 1; } catch {}
-    } catch {}
-    const parts = [];
-    if (pipelineCancelled > 0) parts.push(`${pipelineCancelled} pipeline(s) cancelled`);
-    if (pmLoopsKilled > 0) parts.push(`${pmLoopsKilled} PM loop(s) killed`);
-    if (bridgesKilled) parts.push("all agent bridges killed");
-    if (parts.length === 0) parts.push("nothing was running");
-    const reply = `💀 Hard kill executed: ${parts.join(", ")}. Use \`@@SERVICE restart agents\` or the Services tab to bring bridges back up.`;
-    _deps.broadcastSSE({ type: "chat", from: "crew-lead", content: reply, sessionId, ts: Date.now() });
-    _deps.broadcastSSE({ type: "kill_all", ts: Date.now(), summary: parts.join(", ") });
-    _deps.appendHistory(userId, sessionId, "assistant", reply);
-    return { reply, sessionId };
-  }
-
-  // Graceful stop: "stop everything", "stop all", "emergency stop", "halt everything"
-  if (/\b(stop|halt|abort|cancel)\s+(everything|all|it\s+all)\b|\bemergency\s+stop\b/i.test(message.trim())) {
-    const pipelineCancelled = _deps.cancelAllPipelines(sessionId);
-    const wasAutonomous = _deps.autonomousPmLoopSessions.has(sessionId);
-    _deps.autonomousPmLoopSessions.clear();
-    let pmLoopsStopped = 0;
-    try {
-      const logsDir = _deps.orchestratorLogsDir;
-      if (fs.existsSync(logsDir)) {
-        for (const f of fs.readdirSync(logsDir)) {
-          if (f.startsWith("pm-loop") && f.endsWith(".pid")) {
-            fs.writeFileSync(path.join(logsDir, f.replace(".pid", ".stop")), new Date().toISOString());
-            pmLoopsStopped++;
-          }
-        }
-      }
-    } catch {}
-    const parts = [];
-    if (pipelineCancelled > 0) parts.push(`${pipelineCancelled} pipeline(s) cancelled`);
-    if (pmLoopsStopped > 0) parts.push(`${pmLoopsStopped} PM loop(s) signalled to stop after current task`);
-    if (wasAutonomous) parts.push("autonomous mode cleared");
-    if (parts.length === 0) parts.push("nothing was running — all clear");
-    const reply = `🛑 Graceful stop: ${parts.join(", ")}. PM loops will finish their current task then halt. Say "kill everything" to hard-kill agent bridges immediately.`;
-    _deps.broadcastSSE({ type: "chat", from: "crew-lead", content: reply, sessionId, ts: Date.now() });
-    _deps.broadcastSSE({ type: "stop_all", ts: Date.now(), summary: parts.join(", ") });
-    _deps.appendHistory(userId, sessionId, "assistant", reply);
-    return { reply, sessionId };
-  }
-
-  // Pipeline cancel: "stop pipeline", "cancel pipeline", "abort", "stop everything", "kill it"
-  if (/\b(stop|cancel|abort|kill|halt)\b.*(pipeline|dispatch|task|everything|it|all|them)\b|\b(pipeline|dispatch).*(stop|cancel|abort|kill|halt)\b/i.test(message.trim())) {
-    const count = _deps.cancelAllPipelines(sessionId);
-    if (count > 0) {
-      const reply = `Cancelled ${count} running pipeline(s). All pending waves have been dropped. You can re-dispatch when ready.`;
-      _deps.broadcastSSE({ type: "chat", from: "crew-lead", content: reply, sessionId, ts: Date.now() });
-      return { reply, sessionId };
-    }
-  }
-
-  const needsSearch = _deps.messageNeedsSearch(message);
-  // Inject health snapshot broadly — lightweight pgrep + file reads, not HTTP
-  const needsHealth = message.length < 6
-    ? false
-    : /health|status|running|crashed|down\b|restart|skill|agent|workflow|pipeline|who.s.up|who.s.online|anyone.up|each.*up|up\?|is.*up|are.*up|online|services?|telegram|tg\b|opencode|project.*dir|projects?\b|registered.project|what.projects|list.project|settings|what.s.going|what.s.happening|recent.activity|rt.bus|eyes|see.what/i.test(message);
-  const needsBenchmarkCatalog = message.length > 4 && /benchmark|zeroeval|leaderboard|llm-stats|swe-bench|livecodebench|mmlu|gpqa|humaneval|gsm8k|what\.tests?|which\.tests?|available\.tests?/i.test(message);
-  let braveResults = null;
-  let codebaseResults = null;
-  let healthData = null;
-  let benchmarkCatalog = null;
-  const fetchBenchmarkCatalog = async () => {
-    const r = await fetch("https://api.zeroeval.com/leaderboard/benchmarks", { signal: AbortSignal.timeout(10000) });
-    const arr = await r.json();
-    if (!Array.isArray(arr) || !arr.length) return null;
-    const limit = 250;
-    const rows = arr.slice(0, limit).map(b => {
-      const id = b.benchmark_id || b.id || "";
-      const name = (b.name || id).slice(0, 35);
-      const desc = (b.description || "").slice(0, 60).replace(/\n/g, " ");
-      return `${id} | ${name} | ${desc}`;
-    });
-    const suffix = arr.length > limit ? `\n… and ${arr.length - limit} more (full list at api.zeroeval.com)` : "";
-    return `[Benchmark catalog from ZeroEval — use benchmark_id in @@SKILL zeroeval.benchmark {"benchmark_id":"<id>"}]\n${rows.join("\n")}${suffix}`;
-  };
-  try {
-    const [b, c, h, bc] = await Promise.all([
-    needsSearch ? _deps.searchWithBrave(message).catch(() => null) : null,
-    needsSearch ? Promise.resolve(_deps.searchCodebase(message)).catch(() => null) : null,
-    needsHealth ? (async () => {
-      try {
-        const cfgRaw    = _deps.tryRead(path.join(os.homedir(), ".crewswarm", "config.json")) || {};
-        const skillsDir = path.join(os.homedir(), ".crewswarm", "skills");
-        let skillsDetail = [];
-        try {
-          const files = fs.readdirSync(skillsDir).filter(f => f.endsWith(".json"));
-          for (const f of files) {
-            const name = f.replace(".json", "");
-            try {
-              const def = JSON.parse(fs.readFileSync(path.join(skillsDir, f), "utf8"));
-              const desc = (def.description || "").replace(/\s+/g, " ").trim().slice(0, 100);
-              const notes = def.paramNotes || "";
-              const example = `@@SKILL ${name} ${JSON.stringify(def.defaultParams || {})}`;
-              let line = `  - ${name}: ${desc}`;
-              if (notes) line += ` | Params: ${notes}`;
-              if (def.listUrl && def.listUrlIdField) {
-                try {
-                  const r = await fetch(def.listUrl, { signal: AbortSignal.timeout(5000) });
-                  const arr = await r.json();
-                  if (Array.isArray(arr) && arr.length) {
-                    const idField = def.listUrlIdField;
-                    const ids = arr.slice(0, 50).map(b => b[idField]).filter(Boolean);
-                    line += ` | IDs (live): ${ids.join(", ")}${arr.length > 50 ? ` … +${arr.length - 50} more` : ""}`;
-                  }
-                } catch {}
-              }
-              line += ` | Example: ${example}`;
-              skillsDetail.push(line);
-            } catch { skillsDetail.push(`  - ${name}: (parse failed)`); }
-          }
-        } catch {}
-        // Quick service status via pgrep (processes, not agent connections)
-        const { execSync: esc } = await import("node:child_process");
-        const isRunning = (pat) => { try { esc(`pgrep -f "${pat}"`, { stdio: "ignore" }); return true; } catch { return false; } };
-
-        // Per-agent bridge status from RT bus — who is actually connected right now
-        let rtAgentsOnline = new Set();
-        try {
-          const rtResp = await fetch("http://127.0.0.1:18889/status", { signal: AbortSignal.timeout(1500) });
-          const rtData = await rtResp.json();
-          if (Array.isArray(rtData.agents)) rtAgentsOnline = new Set(rtData.agents);
-        } catch {}
-
-        const allAgents = cfg.agentRoster.length
-          ? cfg.agentRoster
-          : cfg.knownAgents.map(id => ({ id, emoji: "🤖", model: "" }));
-
-        const agentRows = allAgents.map(a => {
-          const online = rtAgentsOnline.has(a.id);
-          const status = online ? "✅" : "❌";
-          return `  ${status} ${a.emoji||"🤖"} ${a.id}: tools=${_deps.readAgentTools(a.id).tools.join(",")||"(default)"} model=${a.model||"??"}`;
-        });
-
-        const rtBusUp = isRunning("opencrew-rt-daemon");
-        const services = [
-          `RT bus (18889): ${rtBusUp ? `✅ running — ${rtAgentsOnline.size} agents connected: ${[...rtAgentsOnline].join(", ")||"none"}` : "❌ DOWN — use @@SERVICE restart rt-bus"}`,
-          `Telegram bridge: ${isRunning("telegram-bridge") ? "✅ running" : "⚠️ stopped — use @@SERVICE restart telegram"}`,
-          `OpenCode (4096): ${isRunning("opencode serve") ? "✅ running" : "⚠️ stopped"}`,
-        ];
-
-        let projectsLine = "Registered projects (dashboard Projects tab): (none)";
-        try {
-          const projRes = await fetch(`${_deps.DASHBOARD}/api/projects`, { signal: AbortSignal.timeout(2000) });
-          if (projRes.ok) {
-            const projData = await projRes.json();
-            const projects = projData.projects || [];
-            if (projects.length) {
-              projectsLine = `Registered projects (${projects.length}): ${projects.map(p => `${p.name || p.id} → ${p.outputDir || p.roadmapFile || "?"}`).join("; ")}`;
-            }
-          }
-        } catch {}
-        const projectsSnapshot = projectsLine;
-
-        const rtActivityLog = _deps.rtActivityLog || [];
-        return [
-          `[System health snapshot — live data from your local machine, fetched right now]`,
-          `crew-lead: ${cfg.providerKey}/${cfg.modelId} | RT connected: ${!!_deps.getRtPublish()} | uptime: ${Math.floor(process.uptime()/300)*5}min`,
-          `crew-lead (this process) can create skills: use @@DEFINE_SKILL name + JSON + @@END_SKILL when the user asks. The agent list below is bridge agents only.`,
-          `OpenCode project dir: ${cfgRaw.opencodeProject || "(not set — agents write to repo root)"}`,
-          projectsSnapshot,
-          `Skills installed (${skillsDetail.length}):`,
-          ...(skillsDetail.length ? skillsDetail : ["(none)"]),
-          ``,
-          (() => {
-            let pipelineNames = [];
-            try {
-              const pipelinesDir = path.join(os.homedir(), ".crewswarm", "pipelines");
-              if (fs.existsSync(pipelinesDir)) pipelineNames = fs.readdirSync(pipelinesDir).filter(f => f.endsWith(".json")).map(f => f.replace(".json", ""));
-            } catch {}
-            return `Workflows (pipelines for cron) (${pipelineNames.length}): ${pipelineNames.length ? pipelineNames.join(", ") : "(none)"}`;
-          })(),
-          ``,
-          `Services:`,
-          ...services,
-          ``,
-          `Agents (${agentRows.length}):`,
-          ...agentRows,
-          ...(rtActivityLog.length > 0 ? [
-            ``,
-            `Recent RT activity (newest last; crew-lead sees all bus traffic):`,
-            ...rtActivityLog.slice(-25).map((e) => `  ${e.time} [${e.channel}] ${e.summary}`),
-            ``,
-          ] : []),
-          `[Use this data to answer the user's question. Do NOT say you cannot reach localhost.]`,
-        ].join("\n");
-      } catch { return null; }
-    })() : null,
-    needsBenchmarkCatalog ? fetchBenchmarkCatalog().catch(() => null) : null,
-  ]);
-    braveResults = b;
-    codebaseResults = c;
-    healthData = h;
-    benchmarkCatalog = bc;
-  } catch (e) {
-    console.error("[crew-lead] context fetch failed:", e?.message || e);
-  }
-  // Auto-search history for context when user asks about past conversations
-  const needsHistorySearch = message.length > 8 && /\b(last time|before|earlier|previously|we discussed|you said|i asked|remember when|what did|history of|past conversation|mentioned|talked about|said something|tell me again)\b/i.test(message);
-  let historyContext = null;
-  
-  if (needsHistorySearch) {
-    try {
-      // Extract key terms from the question
-      const terms = message
-        .replace(/\b(what|when|where|who|how|why|did|do|does|is|are|was|were|have|has|had|tell|show|find|search|remember|mentioned|discussed|said|talked)\b/gi, "")
-        .replace(/[^\w\s]/g, " ")
-        .split(/\s+/)
-        .filter(w => w.length > 4)
-        .slice(0, 3)
-        .join(" ");
-      
-      if (terms.trim()) {
-        const histDir = path.join(os.homedir(), ".crewswarm", "chat-history");
-        if (fs.existsSync(histDir)) {
-          const files = fs.readdirSync(histDir).filter(f => f.endsWith(".jsonl")).sort().reverse().slice(0, 10); // Last 10 sessions
-          const lq = terms.toLowerCase();
-          const hits = [];
-          
-          for (const file of files) {
-            const lines = fs.readFileSync(path.join(histDir, file), "utf8").split("\n");
-            for (const line of lines) {
-              if (!line.trim()) continue;
-              try {
-                const entry = JSON.parse(line);
-                if ((entry.content || "").toLowerCase().includes(lq)) {
-                  const snippet = (entry.content || "").slice(0, 400).replace(/\n/g, " ");
-                  hits.push(`[${entry.role}] ${snippet}`);
-                  if (hits.length >= 3) break;
-                }
-              } catch {}
-            }
-            if (hits.length >= 3) break;
-          }
-          
-          if (hits.length > 0) {
-            historyContext = `[Past conversation context - automatically retrieved]\n${hits.join("\n\n")}`;
-            console.log(`[crew-lead] Auto-retrieved ${hits.length} history matches for: ${terms}`);
-          }
-        }
-      }
-    } catch (e) {
-      console.error(`[crew-lead] Auto history search failed: ${e.message}`);
-    }
-  }
-
-  const parts = [message + projectContext];
-  if (historyContext) parts.push(historyContext);
-  if (braveResults) parts.push(`[Web context from Brave Search]\n${braveResults}`);
-  if (codebaseResults) parts.push(`[Codebase context from workspace]\n${codebaseResults}`);
-  if (healthData) parts.push(healthData);
-  if (benchmarkCatalog) parts.push(benchmarkCatalog);
-  const userContent = parts.length > 1 ? parts.join("\n\n") : (message + projectContext);
-
-  // Many chat APIs use only the first system message; agent completions (e.g. [crew-pm completed task]) are stored as "system" in history and would be dropped. Send them as "user" with a prefix so Stinki always sees them.
-  const historyAsMessages = history.map(h => {
-    if (h.role === "system") {
-      return { role: "user", content: `[Crew update — use this when answering]\n${h.content}` };
-    }
-    return { role: h.role, content: h.content };
-  });
-  const messages = [
-    { role: "system", content: _deps.buildSystemPrompt(cfg) },
-    ...historyAsMessages,
-    { role: "user", content: userContent },
-  ];
-
-  _deps.appendHistory(userId, sessionId, "user", message);
-  // Audit trail: record what Brave actually injected so later turns (and you) can verify — no "context #5" confabulation
-  if (braveResults) {
-    const count = (braveResults.match(/\n\n/g) || []).length + 1;
-    // Keep full numbered list (1. ... 2. ... 5. ...) in history so later turns can cite accurately — ~1200 chars usually includes result #5
-    const preview = braveResults.replace(/\n/g, " ").slice(0, 1200);
-    _deps.appendHistory(userId, sessionId, "system", `[Brave search] query="${message.slice(0, 60)}${message.length > 60 ? "…" : ""}" → ${count} results. Preview: ${preview}${braveResults.length > 1200 ? "…" : ""}`);
-  }
-  _deps.broadcastSSE({ type: "chat_message", sessionId, role: "user", content: message });
-
-  const llmResult = await _deps.callLLM(messages, cfg);
-  let fullReply = llmResult.reply;
-  const usedFallback = llmResult.usedFallback;
-  const activeModel = llmResult.model;
-  const fallbackReason = llmResult.reason;
-
-  // ── Direct tool execution (all crew-lead native tools) ──────────────────
-  const hasDirectTools = /@@READ_FILE[ \t]|@@WRITE_FILE[ \t]|@@WEB_SEARCH[ \t]|@@WEB_FETCH[ \t]|@@MKDIR[ \t]|@@RUN_CMD[ \t]|@@TELEGRAM[ \t]|@@WHATSAPP[ \t]|@@SEARCH_HISTORY[ \t]/.test(fullReply);
-  if (hasDirectTools) {
-    const toolResults = await _deps.execCrewLeadTools(fullReply);
-    if (toolResults.length > 0) {
-      // Follow-up LLM call: show the tool results so crew-lead can give a proper answer
-      const followUpMessages = [
-        { role: "system", content: _deps.buildSystemPrompt(cfg) },
-        ...historyAsMessages,
-        { role: "user", content: userContent },
-        { role: "assistant", content: fullReply },
-        { role: "user", content: `[Tool results]\n${toolResults.join("\n\n")}\n\nUsing only the above results, give a concise, direct answer to the user. IMPORTANT: Do NOT emit any @@ tags in your reply (no @@DISPATCH, @@PIPELINE, @@READ_FILE, @@RUN_CMD, @@WEB_SEARCH, or any other @@command). The tool phase is complete — just answer in plain text.` },
-      ];
-      try {
-        const followUp = await _deps.callLLM(followUpMessages, cfg);
-        fullReply = followUp.reply;
-      } catch (e) {
-        // fallback: append raw tool results if follow-up fails
-        fullReply = fullReply + "\n\n---\n" + toolResults.join("\n\n");
-      }
-    }
-  }
-
-  // ── @@TOOLS — permission grant/revoke command ──────────────────────────────
-  const toolsCmd = (() => {
-    const m = fullReply.match(/@@TOOLS\s+(\{[^}]+\})/);
-    if (!m) return null;
-    try { return JSON.parse(m[1]); } catch { return null; }
-  })();
-
-  // ── @@PROMPT — read/write an agent's system prompt ─────────────────────────
-  // Parsed from LLM reply OR user message so pasting @@PROMPT in chat always runs (LLM often doesn't echo it)
-  // @@PROMPT {"agent":"crew-qa","set":"You are a QA specialist..."}
-  // @@PROMPT {"agent":"crew-qa","append":"- Always use @@READ_FILE before auditing"}
-  const promptCmd = (() => {
-    const promptRe = /@@PROMPT\s+(\{[\s\S]*?\})\s*(?:\n|$)/;
-    const fromReply = fullReply.match(promptRe);
-    if (fromReply) { try { return JSON.parse(fromReply[1]); } catch {} }
-    const fromUser = (message || "").match(promptRe);
-    if (fromUser) { try { return JSON.parse(fromUser[1]); } catch {} }
-    return null;
-  })();
-
-  // ── @@CREATE_AGENT — dynamically create a new specialist agent ────────────
-  // @@CREATE_AGENT {"id":"crew-ml","role":"coder","displayName":"MLBot","description":"AI/ML and data science"}
-  const createAgentCmd = (() => {
-    const m = fullReply.match(/@@CREATE_AGENT\s+(\{[^}]+\})/);
-    if (!m) return null;
-    try { return JSON.parse(m[1]); } catch { return null; }
-  })();
-
-  // ── @@REMOVE_AGENT — remove a dynamically created agent ─────────────────
-  // @@REMOVE_AGENT crew-ml
-  const removeAgentCmd = (() => {
-    const m = fullReply.match(/@@REMOVE_AGENT\s+(crew-[a-zA-Z0-9_-]+)/);
-    return m ? m[1] : null;
-  })();
-
-  // ── @@BRAIN — append a fact to shared brain.md ─────────────────────────────
-  // @@BRAIN crew-lead: some durable fact worth remembering
-  const BRAIN_PLACEHOLDERS = /^(note text|some fact|placeholder|example|durable fact|fact here|your fact|insert fact|crew-lead: note|crew-lead: some)/i;
-  const brainCmd = (() => {
-    const m = fullReply.match(/@@BRAIN\s+([^\n]+)/);
-    if (!m) return null;
-    const entry = m[1].trim();
-    if (BRAIN_PLACEHOLDERS.test(entry) || entry.length < 10) return null; // skip template leakage
-    return entry;
-  })();
-
-  // ── @@MEMORY — search shared memory (AgentKeeper + AgentMemory + Collections)
-  // @@MEMORY search "query text"
-  // @@MEMORY stats
-  const memoryCmd = (() => {
-    const searchMatch = fullReply.match(/@@MEMORY\s+search\s+"([^"]+)"/i) || 
-                        fullReply.match(/@@MEMORY\s+search\s+([^\n]+)/i);
-    if (searchMatch) return { action: 'search', query: searchMatch[1].trim() };
-    
-    if (/@@MEMORY\s+stats\b/i.test(fullReply)) return { action: 'stats' };
-    
-    return null;
-  })();
-
-  // ── @@GLOBALRULE — append a rule to global-rules.md (injected into all agents)
-  // @@GLOBALRULE Always reply in the language the user wrote in
-  const globalRuleCmd = (() => {
-    const m = fullReply.match(/@@GLOBALRULE\s+([^\n]+)/);
-    return m ? m[1].trim() : null;
-  })();
-
-  // ── @@STOP — graceful stop: cancel pipelines + signal PM loops + clear autonomous ─
-  // PM loops finish their current task before halting. Agent bridges keep running.
-  if (/@@STOP\b/.test(fullReply) && !/@@KILL\b/.test(fullReply)) {
-    const pipelineCancelled = _deps.cancelAllPipelines(sessionId);
-    const wasAutonomous = _deps.autonomousPmLoopSessions.has(sessionId);
-    _deps.autonomousPmLoopSessions.clear();
-    let pmLoopsStopped = 0;
-    try {
-      const logsDir = _deps.orchestratorLogsDir;
-      if (fs.existsSync(logsDir)) {
-        for (const f of fs.readdirSync(logsDir)) {
-          if (f.startsWith("pm-loop") && f.endsWith(".pid")) {
-            fs.writeFileSync(path.join(logsDir, f.replace(".pid", ".stop")), new Date().toISOString());
-            pmLoopsStopped++;
-          }
-        }
-      }
-    } catch {}
-    const parts = [];
-    if (pipelineCancelled > 0) parts.push(`${pipelineCancelled} pipeline(s) cancelled`);
-    if (pmLoopsStopped > 0) parts.push(`${pmLoopsStopped} PM loop(s) signalled to stop after current task`);
-    if (wasAutonomous) parts.push("autonomous mode cleared");
-    if (parts.length === 0) parts.push("nothing was running");
-    console.log(`[crew-lead] @@STOP executed: ${parts.join(", ")}`);
-    _deps.broadcastSSE({ type: "stop_all", ts: Date.now(), summary: parts.join(", ") });
-  }
-
-  // ── @@KILL — hard kill: everything @@STOP does + SIGTERM agent bridges + PM loop procs ─
-  // Use when agents are stuck, looping, or unresponsive and you need them dead now.
-  if (/@@KILL\b/.test(fullReply)) {
-    const pipelineCancelled = _deps.cancelAllPipelines(sessionId);
-    _deps.autonomousPmLoopSessions.clear();
-    let pmLoopsKilled = 0;
-    let bridgesKilled = 0;
-    try {
-      const { execSync } = await import("node:child_process");
-      const logsDir = _deps.orchestratorLogsDir;
-      // Hard-kill each PM loop process by PID
-      if (fs.existsSync(logsDir)) {
-        for (const f of fs.readdirSync(logsDir)) {
-          if (f.startsWith("pm-loop") && f.endsWith(".pid")) {
-            fs.writeFileSync(path.join(logsDir, f.replace(".pid", ".stop")), new Date().toISOString());
-            const pid = parseInt(fs.readFileSync(path.join(logsDir, f), "utf8").trim(), 10);
-            if (pid) { try { process.kill(pid, "SIGTERM"); pmLoopsKilled++; } catch {} }
-          }
-        }
-      }
-      // Kill all agent gateway bridges (they will be respawnable via @@SERVICE restart agents)
-      try { execSync(`pkill -f "gateway-bridge.mjs" 2>/dev/null`, { stdio: "ignore", shell: true }); bridgesKilled = 1; } catch {}
-    } catch {}
-    const parts = [];
-    if (pipelineCancelled > 0) parts.push(`${pipelineCancelled} pipeline(s) cancelled`);
-    if (pmLoopsKilled > 0) parts.push(`${pmLoopsKilled} PM loop process(es) SIGTERM'd`);
-    if (bridgesKilled) parts.push("all agent bridges killed (restart with @@SERVICE restart agents)");
-    if (parts.length === 0) parts.push("nothing was running");
-    console.log(`[crew-lead] @@KILL executed: ${parts.join(", ")}`);
-    _deps.broadcastSSE({ type: "kill_all", ts: Date.now(), summary: parts.join(", ") });
-  }
-
-  // ── @@SERVICE — restart/stop a service or agent bridge ──────────────────────
-  // @@SERVICE restart telegram   @@SERVICE stop agents   @@SERVICE restart crew-coder
-  const serviceCmd = (() => {
-    const m = fullReply.match(/@@SERVICE\s+(restart|stop|start)\s+([a-zA-Z0-9_\-]+)/);
-    return m ? { action: m[1], id: m[2] } : null;
-  })();
-
-  // ── @@DEFINE_SKILL — crew-lead writes a skill JSON to ~/.crewswarm/skills/
-  // @@DEFINE_SKILL skillname\n{...json...}\n@@END_SKILL
-  const defineSkillCmds = [];
-  const defineSkillRe = /@@DEFINE_SKILL[ \t]+([a-zA-Z0-9_\-.]+)\n([\s\S]*?)@@END_SKILL/g;
-  let dsMatch;
-  while ((dsMatch = defineSkillRe.exec(fullReply)) !== null) {
-    const skillName = dsMatch[1].trim();
-    const rawJson   = dsMatch[2].trim();
-    try {
-      const def = JSON.parse(rawJson);
-      const skillsDir = path.join(os.homedir(), ".crewswarm", "skills");
-      fs.mkdirSync(skillsDir, { recursive: true });
-      fs.writeFileSync(path.join(skillsDir, skillName + ".json"), JSON.stringify(def, null, 2));
-      defineSkillCmds.push({ name: skillName, ok: true });
-      console.log(`[crew-lead] @@DEFINE_SKILL saved: ${skillName}`);
-    } catch (e) {
-      defineSkillCmds.push({ name: skillName, ok: false, error: e.message });
-    }
-  }
-
-  // ── @@DEFINE_WORKFLOW — create or replace a scheduled pipeline (stages: agent + task per stage)
-  // @@DEFINE_WORKFLOW name\n[...]\n@@END_WORKFLOW
-  const defineWorkflowCmds = [];
-  const defineWorkflowRe = /@@DEFINE_WORKFLOW[ \t]+([a-zA-Z0-9_\-]+)\n([\s\S]*?)@@END_WORKFLOW/g;
-  let dwMatch;
-  while ((dwMatch = defineWorkflowRe.exec(fullReply)) !== null) {
-    const wfName = dwMatch[1].trim();
-    const rawJson = dwMatch[2].trim();
-    try {
-      const stages = JSON.parse(rawJson);
-      if (!Array.isArray(stages)) throw new Error("stages must be a JSON array");
-      const valid = stages
-        .filter(s => s && s.agent && (s.task || s.taskText))
-        .map(s => ({ agent: s.agent, task: s.task || s.taskText || "", tool: s.tool || undefined }));
-      const pipelinesDir = path.join(os.homedir(), ".crewswarm", "pipelines");
-      fs.mkdirSync(pipelinesDir, { recursive: true });
-      fs.writeFileSync(path.join(pipelinesDir, wfName + ".json"), JSON.stringify({ stages: valid }, null, 2));
-      defineWorkflowCmds.push({ name: wfName, ok: true, stageCount: valid.length });
-      console.log(`[crew-lead] @@DEFINE_WORKFLOW saved: ${wfName} (${valid.length} stages)`);
-    } catch (e) {
-      defineWorkflowCmds.push({ name: wfName, ok: false, error: e.message });
-    }
-  }
-
-  const pipelineSteps = _deps.parsePipeline(fullReply);
-  const projectSpec = !pipelineSteps ? _deps.parseProject(fullReply) : null;
-  const dispatch = !projectSpec && !pipelineSteps ? _deps.parseDispatch(fullReply, message) : null;
-  
-  if (dispatch) {
-    console.log(`[chat-handler] parseDispatch found: agent=${dispatch.agent}, isDispatchIntended=${_deps.isDispatchIntended(message)}`);
-  }
-  // ── @@SKILL — crew-lead executes skills directly (no dispatch needed)
-  const skillCalls = [];
-  // Require skill name to be followed by JSON params or end-of-line — prevents "@@SKILL line" in prose
-  const skillRe = /@@SKILL[ \t]+([a-zA-Z0-9_\-\.]+)[ \t]*(\{[^\n]*\})?(?=[ \t]*$|[ \t]*\n|[ \t]*\{)/gm;
-  let skMatch;
-  while ((skMatch = skillRe.exec(fullReply)) !== null) {
-    const skillName = skMatch[1].trim();
-    let params = {};
-    try { if (skMatch[2]) params = JSON.parse(skMatch[2]); } catch {}
-    skillCalls.push({ skillName, params });
-  }
-
-  let cleanReply = _deps.stripThink(_deps.stripPipeline(_deps.stripProject(_deps.stripDispatch(fullReply))))
-    .replace(/@@SKILL[ \t]+[a-zA-Z0-9_\-\.]+[ \t]*(\{[^\n]*\})?(?=[ \t]*$|[ \t]*\n|[ \t]*\{)/gm, "")
-    .replace(/@@TOOLS\s+\{[^}]+\}/g, "")
-    .replace(/@@PROMPT\s+\{[\s\S]*?\}\s*(?:\n|$)/g, "")
-    .replace(/@@BRAIN\s+[^\n]+/g, "")
-    .replace(/@@MEMORY\s+(search|stats)\s*[^\n]*/gi, "")
-    .replace(/@@GLOBALRULE\s+[^\n]+/g, "")
-    .replace(/@@DEFINE_SKILL[ \t]+[a-zA-Z0-9_\-.]+\n[\s\S]*?@@END_SKILL/g, "")
-    .replace(/@@DEFINE_WORKFLOW[ \t]+[a-zA-Z0-9_\-]+\n[\s\S]*?@@END_WORKFLOW/g, "")
-    .replace(/@@SERVICE\s+(restart|stop|start)\s+[a-zA-Z0-9_\-]+/g, "")
-    .replace(/@@STOP\b/g, "")
-    .replace(/@@KILL\b/g, "")
-    .replace(/@@CREATE_AGENT\s+\{[^}]+\}/g, "")
-    .replace(/@@REMOVE_AGENT\s+crew-[a-zA-Z0-9_-]+/g, "")
-    .trim();
-
-  let dispatched = null;
-  let pendingProject = null;
-  let pipeline = null;
-
-  if (projectSpec?.name && projectSpec?.outputDir) {
-    try {
-      pendingProject = await _deps.draftProject(projectSpec, sessionId);
-      _deps.appendHistory(userId, sessionId, "system", `Roadmap drafted for "${projectSpec.name}" — awaiting user approval.`);
-      if (pendingProject) _deps.broadcastSSE({ type: "pending_project", sessionId, pendingProject });
-    } catch (e) {
-      console.error(`[crew-lead] Roadmap draft failed: ${e.message}`);
-      _deps.appendHistory(userId, sessionId, "system", `Roadmap draft failed: ${e.message}`);
-    }
-  } else if (pipelineSteps) {
-    const pipelineId = crypto.randomUUID();
-    const { steps, waves } = pipelineSteps;
-    // Use the explicitly selected project only — never infer projectDir from LLM-generated task text,
-    // which can contain stale paths from brain.md and incorrectly lock the pipeline to the wrong project.
-    const _projectDir = activeProjectOutputDir || null;
-    _deps.pendingPipelines.set(pipelineId, { steps, waves, currentWave: 0, pendingTaskIds: new Set(), waveResults: [], sessionId, projectDir: _projectDir });
-    _deps.dispatchPipelineWave(pipelineId);
-    const waveDesc = waves.length > 1 ? ` in ${waves.length} waves` : "";
-    _deps.appendHistory(userId, sessionId, "system", `Pipeline started (${steps.length} steps${waveDesc}): ${waves.map(w => w.map(s => s.agent).join("+")).join(" → ")}`);
-    const agentFlow = waves.map(w => w.length > 1 ? `[${w.map(s => s.agent).join(" ∥ ")}]` : w[0].agent).join(" → ");
-    cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Pipeline started (${steps.length} steps${waveDesc}): ${agentFlow}`;
-    pipeline = { pipelineId, steps, waves };
-  } else if (dispatch) {
-    console.log(`[chat-handler] LLM dispatch detected! Agent: ${dispatch.agent}, Task: ${(dispatch.task || "").slice(0, 60)}`);
-    const resolvedAgent = _deps.resolveAgentId(cfg, dispatch.agent) || dispatch.agent;
-    if (!cfg.knownAgents.length || cfg.knownAgents.includes(resolvedAgent)) {
-      dispatch.agent = resolvedAgent;
-    }
-    // QA always writes to projectDir/qa-report.md so crew-lead doesn't tell them a random path
-    const isQa = dispatch.agent === "crew-qa" || (dispatch.agent && dispatch.agent.includes("qa"));
-    if (isQa && activeProjectOutputDir && !/qa-report\.md|Write your report to/i.test(dispatch.task || "")) {
-      dispatch.task = (dispatch.task || "").trimEnd() + `\n\nWrite your report to ${activeProjectOutputDir}/qa-report.md (no other filename).`;
-    }
-    const pipelineMeta = activeProjectOutputDir ? { projectDir: activeProjectOutputDir } : null;
-    if (cfg.knownAgents.includes(dispatch.agent)) {
-    // Pass full dispatch spec so verify/done criteria are injected into task text
-    const ok = _deps.dispatchTask(dispatch.agent, dispatch, sessionId, pipelineMeta);
-    if (ok) {
-      dispatched = dispatch;
-      _deps.appendHistory(userId, sessionId, "system", `You dispatched to ${dispatch.agent}: "${(dispatch.task || "").slice(0, 200)}".`);
-      const dispatchLine = _deps.getRtPublish()
-        ? `\n\n↳ Dispatched to ${dispatch.agent} — reply will show here when they finish.`
-        : `\n\n↳ Dispatched to ${dispatch.agent} (via ctl — check RT Messages tab for reply).`;
-      cleanReply = (cleanReply || "").trimEnd() + dispatchLine;
-    }
-    }
-  }
-
-  // ── Detect "dispatch lie" — LLM claims it dispatched but no @@DISPATCH was parsed ──
-  if (!dispatched && !pipeline && cleanReply) {
-    // Only trigger if Stinki explicitly claims to have ALREADY dispatched (past tense with confirmation)
-    const liedPattern = /\b(?:I (?:just |already |have )?(?:dispatched|sent it|forwarded it|sicced)|consider it done|they'(?:re|ve) got it|it'?s on its way to crew-|I'?ve tasked crew-)/i;
-    if (liedPattern.test(cleanReply)) {
-      console.log(`[crew-lead] Dispatch-lie detected — auto-retrying with extraction call`);
-      let _lieRetryOk = false;
-      try {
-        const _lieRetryMsgs = [
-          ...messages,
-          { role: "assistant", content: fullReply },
-          { role: "user", content: "You described dispatching but the @@DISPATCH line was missing. Emit ONLY the @@DISPATCH JSON now. No prose, no explanation.\nFormat: @@DISPATCH {\"agent\":\"crew-X\",\"task\":\"...\"}" },
-        ];
-        const _lieResult = await _deps.callLLM(_lieRetryMsgs, cfg);
-        const _lieDispatch = !projectSpec ? _deps.parseDispatch(_lieResult.reply || "", message) : null;
-        if (_lieDispatch && _lieDispatch.agent) {
-          console.log(`[crew-lead] Dispatch-lie retry succeeded -> ${_lieDispatch.agent}`);
-          const _lieTaskId = _deps.dispatchTask(_lieDispatch.agent, _lieDispatch.task, sessionId);
-          if (_lieTaskId) {
-            dispatched = _lieDispatch;
-            cleanReply = (cleanReply || "").trimEnd() + `\n\n\u21b3 Dispatched to ${_lieDispatch.agent} \u2014 reply will show here when they finish.`;
-            _lieRetryOk = true;
-          }
-        }
-      } catch (_lieErr) {
-        console.error(`[crew-lead] Dispatch-lie retry error: ${_lieErr.message}`);
-      }
-      if (!_lieRetryOk) {
-        cleanReply = (cleanReply || "").trimEnd() + `\n\n\u26a0\ufe0f Dispatch failed even after retry \u2014 please ask me again.`;
-        _deps.appendHistory(userId, sessionId, "system", `Warning: LLM described dispatching without emitting @@DISPATCH — retry also failed.`);
-      }
-    }
-  }
-
-  // ── Execute @@TOOLS permission change ──────────────────────────────────────
-  if (toolsCmd?.agent) {
-    try {
-      const current = _deps.readAgentTools(toolsCmd.agent).tools;
-      let updated;
-      if (Array.isArray(toolsCmd.set)) {
-        updated = toolsCmd.set;
-      } else {
-        const granted = Array.isArray(toolsCmd.grant)  ? toolsCmd.grant  : [];
-        const revoked = Array.isArray(toolsCmd.revoke) ? toolsCmd.revoke : [];
-        updated = [...new Set([...current, ...granted].filter(t => !revoked.includes(t)))];
-      }
-      const saved = _deps.writeAgentTools(toolsCmd.agent, updated);
-      const note = `\n\n↳ Tool permissions updated for **${toolsCmd.agent}**: ${saved.join(", ")} — restart its bridge for changes to take effect.`;
-      cleanReply = (cleanReply || "").trimEnd() + note;
-      _deps.appendHistory(userId, sessionId, "system", `Tool permissions for ${toolsCmd.agent} updated to: ${saved.join(", ")}`);
-      console.log(`[crew-lead] @@TOOLS: ${toolsCmd.agent} → ${saved.join(", ")}`);
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to update tools for ${toolsCmd.agent}: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@PROMPT system prompt edit ────────────────────────────────────
-  if (promptCmd?.agent) {
-    try {
-      const existing = _deps.getAgentPrompts()[promptCmd.agent] || "";
-      let newPrompt;
-      if (typeof promptCmd.set === "string") {
-        newPrompt = promptCmd.set;
-      } else if (typeof promptCmd.append === "string") {
-        newPrompt = existing ? `${existing}\n${promptCmd.append}` : promptCmd.append;
-      } else {
-        newPrompt = existing;
-      }
-      _deps.writeAgentPrompt(promptCmd.agent, newPrompt);
-      const preview = newPrompt.slice(0, 120).replace(/\n/g, " ");
-      const restartNote = promptCmd.agent === "crew-lead" ? "Takes effect on your next message; no restart needed." : "Restart its bridge for changes to take effect.";
-      const note = `\n\n↳ System prompt updated for **${promptCmd.agent}**: "${preview}${newPrompt.length > 120 ? "…" : ""}" — ${restartNote}`;
-      cleanReply = (cleanReply || "").trimEnd() + note;
-      _deps.appendHistory(userId, sessionId, "system", `Prompt for ${promptCmd.agent} updated.`);
-      console.log(`[crew-lead] @@PROMPT: ${promptCmd.agent} updated (${newPrompt.length} chars)`);
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to update prompt for ${promptCmd.agent}: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@BRAIN append ──────────────────────────────────────────────────
-  // Routes to <projectDir>/.crewswarm/brain.md when a project is active, global brain.md otherwise
-  // Also stores in AgentMemory if shared memory is available
-  if (brainCmd) {
-    try {
-      const block = _deps.appendToBrain("crew-lead", brainCmd, activeProjectOutputDir || null);
-      const dest = activeProjectOutputDir ? `${path.basename(activeProjectOutputDir)}/.crewswarm/brain.md` : "memory/brain.md";
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Added to ${dest}: "${block.slice(0, 100)}"`;
-      console.log(`[crew-lead] @@BRAIN → ${dest}: ${brainCmd.slice(0, 80)}`);
-      
-      // Also store in shared AgentMemory for cross-system access
-      if (isSharedMemoryAvailable()) {
-        const factId = rememberFact('crew-lead', brainCmd, {
-          critical: brainCmd.includes('CRITICAL') || brainCmd.includes('MUST'),
-          tags: ['brain', 'crew-lead'],
-          provider: 'crew-lead-chat'
-        });
-        if (factId) {
-          console.log(`[crew-lead] Fact also stored in shared AgentMemory: ${factId}`);
-        }
-      }
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to write brain: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@MEMORY commands ───────────────────────────────────────────────
-  if (memoryCmd) {
-    try {
-      if (memoryCmd.action === 'search') {
-        if (!isSharedMemoryAvailable()) {
-          cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Shared memory not available. Run: cd crew-cli && npm run build`;
-        } else {
-          const projectDir = activeProjectOutputDir || process.cwd();
-          const hits = await searchMemory(projectDir, memoryCmd.query, {
-            maxResults: 10,
-            includeDocs: true,
-            includeCode: false,
-            preferSuccessful: true,
-            crewId: 'crew-lead'
-          });
-          
-          if (hits.length === 0) {
-            cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ **Memory search:** No results for "${memoryCmd.query}"`;
-          } else {
-            const results = hits.map(h => 
-              `- [${h.source}] **${h.title}** (score: ${h.score.toFixed(3)})\n  ${h.text.slice(0, 150)}${h.text.length > 150 ? '...' : ''}`
-            ).join('\n\n');
-            cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ **Memory search results (${hits.length}):**\n\n${results}`;
-          }
-          console.log(`[crew-lead] @@MEMORY search "${memoryCmd.query}": ${hits.length} hits`);
-        }
-      } else if (memoryCmd.action === 'stats') {
-        if (!isSharedMemoryAvailable()) {
-          cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Shared memory not available. Run: cd crew-cli && npm run build`;
-        } else {
-          const factStats = getMemoryStats('crew-lead');
-          const keeperStats = await getKeeperStats(activeProjectOutputDir || process.cwd());
-          
-          const lines = [
-            '↳ **Shared Memory Statistics:**',
-            '',
-            '**AgentMemory (cognitive facts):**',
-            `- Total facts: ${factStats?.totalFacts || 0}`,
-            `- Critical facts: ${factStats?.criticalFacts || 0}`,
-            `- Providers: ${factStats?.providers?.join(', ') || 'none'}`,
-            '',
-            '**AgentKeeper (task memory):**',
-            `- Total entries: ${keeperStats?.entries || 0}`,
-            `- Storage bytes: ${keeperStats?.bytes ? (keeperStats.bytes / 1024).toFixed(1) + 'KB' : '0KB'}`,
-            `- By tier: ${keeperStats?.byTier ? Object.entries(keeperStats.byTier).map(([k,v]) => `${k}=${v}`).join(', ') : 'none'}`,
-            `- By agent: ${keeperStats?.byAgent ? Object.entries(keeperStats.byAgent).map(([k,v]) => `${k}=${v}`).join(', ') : 'none'}`
-          ];
-          cleanReply = (cleanReply || "").trimEnd() + `\n\n${lines.join('\n')}`;
-          console.log(`[crew-lead] @@MEMORY stats: ${factStats?.totalFacts || 0} facts, ${keeperStats?.entries || 0} keeper entries`);
-        }
-      }
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Memory command failed: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@GLOBALRULE append ────────────────────────────────────────────
-  if (globalRuleCmd) {
-    try {
-      _deps.appendGlobalRule(globalRuleCmd);
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Global rule added (all agents): "${globalRuleCmd}" — restart bridges to apply.`;
-      _deps.appendHistory(userId, sessionId, "system", `Global rule added: ${globalRuleCmd}`);
-      console.log(`[crew-lead] @@GLOBALRULE: ${globalRuleCmd}`);
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to write global-rules.md: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@SERVICE control ───────────────────────────────────────────────
-  if (serviceCmd) {
-    const { action, id } = serviceCmd;
-    try {
-      // Per-agent restart: if ID looks like a specific agent, use the agents/:id/restart endpoint
-      const isSpecificAgent = id.startsWith("crew-") && id !== "crew-lead";
-      let result;
-
-      const authHeader = _deps.getRTToken() ? { authorization: `Bearer ${_deps.getRTToken()}` } : {};
-      if (action === "stop" && !isSpecificAgent) {
-        const r = await fetch(`${_deps.DASHBOARD}/api/services/stop`, {
-          method: "POST",
-          headers: { "content-type": "application/json" },
-          body: JSON.stringify({ id }),
-          signal: AbortSignal.timeout(8000),
-        });
-        result = await r.json();
-      } else if (isSpecificAgent && action !== "stop") {
-        // Single agent bridge restart via dedicated endpoint
-        const r = await fetch(`http://127.0.0.1:${_deps.PORT}/api/agents/${id}/restart`, {
-          method: "POST",
-          headers: authHeader,
-          signal: AbortSignal.timeout(8000),
-        });
-        result = await r.json();
-      } else {
-        const r = await fetch(`${_deps.DASHBOARD}/api/services/restart`, {
-          method: "POST",
-          headers: { "content-type": "application/json" },
-          body: JSON.stringify({ id }),
-          signal: AbortSignal.timeout(8000),
-        });
-        result = await r.json();
-      }
-
-      if (result?.ok === false && result?.message) {
-        cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Service **${id}** — ${result.message}`;
-      } else {
-        const actionLabel = action === "stop" ? "stopped" : "restarted";
-        cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ **${id}** ${actionLabel}. Give it 2–3 seconds to come back online.`;
-        _deps.appendHistory(userId, sessionId, "system", `Service ${id} ${actionLabel} via @@SERVICE.`);
-        console.log(`[crew-lead] @@SERVICE ${action} ${id}: ok`);
-      }
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to ${action} **${id}**: ${e.message}`;
-    }
-  }
-
-  // ── Surface @@DEFINE_SKILL results ─────────────────────────────────────────
-  for (const ds of defineSkillCmds) {
-    if (ds.ok) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Skill **${ds.name}** saved to ~/.crewswarm/skills/${ds.name}.json — agents with 'skill' permission can now call it.`;
-      _deps.appendHistory(userId, sessionId, "system", `Skill "${ds.name}" defined and saved.`);
-    } else {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to save skill **${ds.name}**: ${ds.error}`;
-    }
-  }
-
-  // ── Surface @@DEFINE_WORKFLOW results ─────────────────────────────────────
-  for (const dw of defineWorkflowCmds) {
-    if (dw.ok) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Workflow **${dw.name}** saved to ~/.crewswarm/pipelines/${dw.name}.json (${dw.stageCount} stages). Run with: \`node scripts/run-scheduled-pipeline.mjs ${dw.name}\` or add to crontab.`;
-      _deps.appendHistory(userId, sessionId, "system", `Workflow "${dw.name}" defined (${dw.stageCount} stages).`);
-    } else {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to save workflow **${dw.name}**: ${dw.error}`;
-    }
-  }
-
-  // ── Execute @@CREATE_AGENT — create a new specialist agent ────────────────
-  if (createAgentCmd?.id) {
-    try {
-      const result = _deps.createAgent(createAgentCmd);
-      const ocNote = result.useOpenCode ? `, OpenCode: ${result.useOpenCode}` : "";
-
-      // Spawn the bridge directly via start-crew.mjs --agent (no dashboard round-trip)
-      let spawnNote = "";
-      try {
-        const startScript = path.join(process.cwd(), "scripts", "start-crew.mjs");
-        const { execSync: exec2 } = await import("node:child_process");
-        exec2(`node ${startScript} --agent ${result.id}`, {
-          timeout: 10000,
-          env: { ...process.env, CREWSWARM_RT_AUTH_TOKEN: _deps.getRTToken() },
-          stdio: "pipe",
-        });
-        // Verify bridge is running
-        await new Promise(r => setTimeout(r, 1500));
-        try {
-          const psOut = exec2("ps aux", { encoding: "utf8" });
-          const running = psOut.includes(result.id);
-          spawnNote = running
-            ? " — bridge spawned and online"
-            : " — bridge spawned (verifying…)";
-        } catch {
-          spawnNote = " — bridge spawned";
-        }
-      } catch (spawnErr) {
-        console.error(`[crew-lead] Failed to spawn bridge for ${result.id}:`, spawnErr.message);
-        spawnNote = " — restart bridges to bring it online (`npm run restart-all`)";
-      }
-
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Agent **${result.id}** created (role: ${result.role}, tools: ${result.tools.join(", ")}${ocNote})${spawnNote}. You can now dispatch tasks to it.`;
-      _deps.appendHistory(userId, sessionId, "system", `Dynamic agent ${result.id} created (role: ${result.role}, openCode: ${!!result.useOpenCode}).`);
-      console.log(`[crew-lead] @@CREATE_AGENT: ${result.id} (role: ${result.role}, tools: ${result.tools.join(",")}, openCode: ${!!result.useOpenCode})`);
-      _deps.broadcastSSE({ type: "agent_created", agent: result, ts: Date.now() });
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to create agent: ${e.message}`;
-    }
-  }
-
-  // ── Execute @@REMOVE_AGENT — remove a dynamically created agent ──────────
-  if (removeAgentCmd) {
-    try {
-      _deps.removeDynamicAgent(removeAgentCmd);
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Agent **${removeAgentCmd}** removed. Restart bridges to clean up.`;
-      _deps.appendHistory(userId, sessionId, "system", `Dynamic agent ${removeAgentCmd} removed.`);
-      console.log(`[crew-lead] @@REMOVE_AGENT: ${removeAgentCmd}`);
-    } catch (e) {
-      cleanReply = (cleanReply || "").trimEnd() + `\n\n↳ Failed to remove agent: ${e.message}`;
-    }
-  }
-
-  if (usedFallback) {
-    const primaryLabel = `${cfg.providerKey}/${cfg.modelId}`;
-    const fbUrl = `${cfg.fallbackProvider.baseUrl.replace(/\/$/, "")}/chat/completions`;
-    // Strip any fallback line the model echoed (avoids duplicate banner)
-    const stripped = (cleanReply || "").replace(/^⚡\s*\*fallback:[^*]*\*[^\n]*\n?/gm, "").trimStart();
-    cleanReply = `⚡ *fallback: ${activeModel}* @ ${fbUrl} (primary ${primaryLabel} failed: ${fallbackReason})\n\n${stripped}`;
-  }
-
-  // Trim blank space left by stripped @@SKILL-only replies before appending results
-  if (skillCalls.length > 0) cleanReply = (cleanReply || "").trim();
-
-  // Execute @@SKILL calls — collect display blocks (for user) and feedback blocks (for LLM second pass)
-  const skillDisplayBlocks = [];
-  const skillFeedbackBlocks = [];
-
-  for (const { skillName, params } of skillCalls) {
-    try {
-      const result = await _deps.executeSkillFromCrewLead(skillName, params);
-      console.log(`[crew-lead] @@SKILL ${skillName} → OK`);
-      const isBenchmark = skillName === "zeroeval.benchmark" || skillName === "benchmark" || skillName === "benchmarks";
-      const _skillCount = isBenchmark && result?.models?.length ? ` · ${Math.min(Number(params?.limit ?? params?.top ?? 100), result.models.length)}/${result.models.length} models` : "";
-      const skillTag = `↳ *skill: ${skillName}${_skillCount}*`;
-
-      let displayBlock = "";
-      let feedbackBlock = "";
-
-      if (isBenchmark && Array.isArray(result) && result.length) {
-        const list = result.slice(0, 50).map(b => {
-          const id = typeof b === "object" ? b.benchmark_id : b;
-          const name = typeof b === "object" ? b.name || "" : "";
-          return `  - ${id}${name ? `: ${name}` : ""}`;
-        }).join("\n");
-        const body = `**Available benchmarks** (omit benchmark_id or use empty to list):\n${list}${result.length > 50 ? `\n  … and ${result.length - 50} more` : ""}`;
-        displayBlock = `${skillTag}\n\n${body}`;
-        feedbackBlock = `[Skill result: ${skillName} — benchmark list]\n${body}`;
-      } else if (isBenchmark && result && !result.models?.length && !Array.isArray(result)) {
-        const name = result.name || result.benchmark_id || skillName;
-        displayBlock = `${skillTag}\n\n*${name}* — no models found for this benchmark yet.`;
-        feedbackBlock = `[Skill result: ${skillName}]\nNo models found for benchmark "${name}".`;
-      } else if (isBenchmark && result?.models?.length) {
-        const limit = Number(params?.limit ?? params?.top ?? 100);
-        const top = result.models.slice(0, limit);
-        const rows = top.map(m => {
-          const pct = ((m.normalized_score ?? m.score ?? 0) * 100).toFixed(1);
-          const inC = m.input_cost_per_million ?? 0;
-          const outC = m.output_cost_per_million ?? 0;
-          const inCents = inC > 0 ? Math.round(inC * 100) + '¢' : '?';
-          const outCents = outC > 0 ? Math.round(outC * 100) + '¢' : '?';
-          const cost = (inC > 0 || outC > 0) ? ` @ ${inCents} → ${outCents}` : "";
-          const score = (m.normalized_score ?? m.score) ?? 0;
-          const centsPerPt = (inC + outC) > 0 && score > 0 ? ` → ${((inC + outC) * 100 / (score * 100)).toFixed(1)} ¢/pt` : "";
-          return `  ${m.rank}. **${m.model_name}** (${m.organization_name}) — ${pct}%${cost}${centsPerPt}`;
-        }).join("\n");
-        const showing = top.length < result.models.length ? ` (showing ${top.length} of ${result.models.length} — add "limit":N for more)` : ` (all ${result.models.length} models)`;
-        const body = `**${result.name || "Benchmark"}** — top ${top.length}${showing}:\n${rows}`;
-        displayBlock = `${skillTag}\n\n${body}`;
-        feedbackBlock = `[Skill result: ${skillName}]\n${body}`;
-      } else if (result?.output !== undefined) {
-        const out = String(result.output).trim();
-        displayBlock = `${skillTag}\n\n\`\`\`\n${out.slice(0, 3000)}${out.length > 3000 ? "\n… (truncated)" : ""}\n\`\`\``;
-        feedbackBlock = `[Skill result: ${skillName}]\n${out.slice(0, 3000)}`;
-      } else {
-        const raw = typeof result === "object" ? JSON.stringify(result) : String(result);
-        displayBlock = `${skillTag}: ${raw.slice(0, 600)}${raw.length > 600 ? "…" : ""}`;
-        feedbackBlock = `[Skill result: ${skillName}]\n${raw.slice(0, 600)}`;
-      }
-
-      // Check feedbackLoop flag from skill JSON (default: true — opt out with "feedbackLoop": false)
-      const resolvedSkillName = _deps.resolveSkillAlias(skillName);
-      const skillDefPath = path.join(os.homedir(), ".crewswarm", "skills", `${resolvedSkillName}.json`);
-      const skillDefRaw = fs.existsSync(skillDefPath) ? JSON.parse(fs.readFileSync(skillDefPath, "utf8")) : {};
-      const wantsFeedback = skillDefRaw.feedbackLoop !== false;
-
-      skillDisplayBlocks.push(displayBlock);
-      if (wantsFeedback) skillFeedbackBlocks.push(feedbackBlock);
-
-    } catch (e) {
-      console.error(`[crew-lead] @@SKILL ${skillName} failed:`, e.message);
-      const errBlock = `↳ *${skillName}* failed: ${e.message}`;
-      skillDisplayBlocks.push(errBlock);
-      skillFeedbackBlocks.push(`[Skill result: ${skillName}]\nFailed: ${e.message}`);
-    }
-  }
-
-  // Append display blocks so the user always sees the raw skill data
-  if (skillDisplayBlocks.length > 0) {
-    cleanReply = (cleanReply ? cleanReply + "\n\n" : "") + skillDisplayBlocks.join("\n\n");
-  }
-
-  // Feedback loop: second LLM call so the model actually reads the results and responds
-  if (skillFeedbackBlocks.length > 0) {
-    try {
-      const feedbackUserMsg = skillFeedbackBlocks.join("\n\n")
-        + "\n\nBased ONLY on the skill results above, respond to the user's original question. "
-        + "Be concise and specific. Do not invent numbers or models not in the results above.";
-      const feedbackMessages = [
-        ...messages,
-        { role: "assistant", content: fullReply },
-        { role: "user", content: feedbackUserMsg },
-      ];
-      console.log(`[crew-lead] Skill feedback loop — second LLM call (${skillFeedbackBlocks.length} skill(s))`);
-      const feedbackResult = await _deps.callLLM(feedbackMessages, cfg);
-      if (feedbackResult.reply?.trim()) {
-        cleanReply = cleanReply + "\n\n" + feedbackResult.reply.trim();
-      }
-    } catch (fbErr) {
-      console.error(`[crew-lead] Skill feedback loop failed:`, fbErr.message);
-      // Non-fatal — user still sees raw skill data above
-    }
-  }
-
-  // Auto-learn from conversation: extract important facts and append to brain
-  // Trigger on: user confirms something, agent discovers/fixes something, or decisions are made
-  const shouldAutoLearn = (msg) => {
-    const lower = msg.toLowerCase();
-    // Trigger phrases that indicate learning opportunities
-    return /\b(discovered|learned|figured out|found that|turns out|confirmed|fixed by|solution was|root cause|the issue is|mistake was|now i know|remember that|important:|note:|fyi:|btw:)\b/i.test(msg)
-      || /\b(always|never|every time|from now on|in future|going forward)\b/i.test(lower)
-      || (msg.includes("✅") && msg.length > 30); // Success markers with substance
-  };
-
-  // Auto-append to brain when significant facts emerge (after LLM reply)
-  if (cleanReply && shouldAutoLearn(cleanReply + " " + message)) {
-    try {
-      // Extract key fact from the conversation
-      const extractFact = (text) => {
-        // Look for sentences with learning indicators
-        const sentences = text.split(/[.!?]+/).filter(s => s.trim().length > 20);
-        for (const sent of sentences) {
-          if (shouldAutoLearn(sent)) {
-            const clean = sent.trim()
-              .replace(/^(discovered|learned|figured out|found that|turns out|confirmed|fixed by|solution was|the issue is|mistake was|now i know|remember that|important:|note:|fyi:|btw:)\s*/i, "")
-              .replace(/\s+/g, " ")
-              .slice(0, 200);
-            if (clean.length > 15) return clean;
-          }
-        }
-        return null;
-      };
-
-      const fact = extractFact(cleanReply) || extractFact(message);
-      if (fact && fact.length > 15 && fact.length < 300) {
-        const brainEntry = `crew-lead (auto): ${fact}`;
-        _deps.appendToBrain("crew-lead", brainEntry, activeProjectOutputDir || null);
-        console.log(`[crew-lead] Auto-learned to brain: ${fact.slice(0, 80)}...`);
-      }
-    } catch (e) {
-      console.error(`[crew-lead] Auto-learn failed: ${e.message}`);
-    }
-  }
-  // not conversation context. Storing it pollutes future LLM requests with noisy boilerplate.
-  const historyReply = usedFallback
-    ? (cleanReply || "").replace(/^⚡\s*\*fallback:[^*]*\*[^\n]*\n?/gm, "").trimStart()
-    : cleanReply;
-  _deps.appendHistory(userId, sessionId, "assistant", historyReply);
-  _deps.broadcastSSE({ 
-    type: "chat_message", 
-    sessionId, 
-    role: "assistant", 
-    content: cleanReply, 
-    model: activeModel,  // Always send model info
-    ...(usedFallback ? { fallbackModel: activeModel, fallbackReason } : {})
-  });
-
-  return { reply: cleanReply, dispatched, pendingProject, pipeline };
-}