crewswarm 0.9.2 → 0.9.4

package/lib/contacts/identity-linker.mjs

@@ -185,6 +185,7 @@ export function unlinkIdentity(contactId) {
     SET platform_links = NULL
     WHERE contact_id = ?
   `).run(contactId);
+  return true;
 }
 
 /**

package/lib/crew-judge/judge.mjs

@@ -150,25 +150,26 @@ function getJudgeModel() {
 
 async function callJudgeModel(model, prompt) {
   // Import LLM caller from crew-lead
-  const { callLLMForText } = await import('../crew-lead/llm-caller.mjs');
-  
+  const { callLLM } = await import('../crew-lead/llm-caller.mjs');
+  const { loadProviderMap } = await import('../runtime/config.mjs');
+
   try {
-    const response = await callLLMForText({
-      provider: model.split('/')[0],
-      model: model.split('/').slice(1).join('/'),
-      messages: [
-        {
-          role: 'system',
-          content: 'You are crew-judge. Return ONLY valid JSON: {"decision": "CONTINUE"|"SHIP"|"RESET", "reasoning": "...", "confidence": 0.0-1.0}'
-        },
-        {
-          role: 'user',
-          content: prompt
-        }
-      ],
-      temperature: 0.3, // Lower temperature for more consistent decisions
-      maxTokens: 500
-    });
+    const providerKey = model.split('/')[0];
+    const modelId = model.split('/').slice(1).join('/');
+    const providerMap = loadProviderMap();
+    const provider = providerMap[providerKey] || null;
+
+    const messages = [
+      {
+        role: 'system',
+        content: 'You are crew-judge. Return ONLY valid JSON: {"decision": "CONTINUE"|"SHIP"|"RESET", "reasoning": "...", "confidence": 0.0-1.0}'
+      },
+      {
+        role: 'user',
+        content: prompt
+      }
+    ];
+    const { reply: response } = await callLLM(messages, { provider, modelId, providerKey }, { temperature: 0.3, maxTokens: 500 });
 
     // Parse JSON response
     const jsonMatch = response.match(/\{[\s\S]*\}/);

package/lib/crew-lead/agent-manager.mjs

@@ -79,7 +79,7 @@ export function createAgent({ id, role, displayName, prompt, description, model
   const defaultOcModel = (() => {
     const existingCoder = swarm.agents.find(a => a.opencodeModel && a.useOpenCode);
     if (existingCoder) return existingCoder.opencodeModel;
-    return process.env.CREWSWARM_OPENCODE_MODEL || "openai/gpt-5.3-codex";
+    return process.env.CREWSWARM_OPENCODE_MODEL || "openai/gpt-5.4";
   })();
 
   const agentEntry = {

package/lib/crew-lead/background.mjs

@@ -28,6 +28,19 @@ const _timeoutLog = [];
 let _lastBgConsciousnessAt = 0;
 let _bgLoopInterval = null;
 
+/** Reset internal state for test isolation. */
+export function resetForTesting() {
+  _agentTimeoutCounts.clear();
+  _timeoutLog.length = 0;
+  _lastBgConsciousnessAt = 0;
+  if (_bgLoopInterval) { clearInterval(_bgLoopInterval); _bgLoopInterval = null; }
+}
+
+/** Stop the background loop interval (prevents process from hanging in tests). */
+export function stopBackgroundLoop() {
+  if (_bgLoopInterval) { clearInterval(_bgLoopInterval); _bgLoopInterval = null; }
+}
+
 export function initBackground({
   broadcastSSE,
   appendHistory,
@@ -267,4 +280,4 @@ export function getRateLimitFallback(agentId) {
   return "crew-main";
 }
 
-export const RATE_LIMIT_PATTERN = /429|rate\s*limit|throttl|quota\s*exceeded|too\s*many\s*requests|resource_exhausted|overloaded/i;
+export const RATE_LIMIT_PATTERN = /429|rate[\s_]*limit|throttl|quota[\s_]*exceeded|too[\s_]*many[\s_]*requests|resource_exhausted|overloaded/i;

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

@@ -750,7 +750,10 @@ Reply with your answers and I'll turn this into a concrete build plan with file
   // IMPORTANT: Do not gate on keywords only. Casual replies ("yeah no shit", "ok cool")
   // used to skip the snapshot while long contradictory assistant history stayed in context,
   // so the LLM invented swarm-wide offline/online flips. Fetch for substantive messages.
-  const needsHealth = message.trim().length >= 6;
+  // Custom personas (displayName set to something other than default "crew-lead") are
+  // personality/chat personas — they don't need operational system state, skip snapshot.
+  const isCustomPersona = cfg.displayName && cfg.displayName.toLowerCase() !== "crew-lead";
+  const needsHealth = !isCustomPersona && message.trim().length >= 6;
   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(
@@ -2577,6 +2580,40 @@ Reply with your answers and I'll turn this into a concrete build plan with file
           ...(usedFallback ? { fallbackReason } : {}),
         },
       });
+
+      // Route @mentions in assistant replies — crew-lead can autonomously dispatch
+      if (replyMentions.length && channelMode) {
+        try {
+          const {
+            handleAutonomousMentions: routeMentions,
+            detectMentionTargets,
+          } = await import("../chat/autonomous-mentions.mjs");
+          const mentionTargets = detectMentionTargets(historyReply);
+          const mentionRoute = classifySharedChatMention(historyReply);
+          if (mentionTargets.length && mentionRoute.mode === "dispatch") {
+            void routeMentions({
+              message: { content: historyReply },
+              sender: "crew-lead",
+              channel: sharedChannel,
+              projectId: sharedChannel,
+              sessionId,
+              projectDir: explicitProjectDir || activeProjectOutputDir || null,
+              originMessageId: assistantMessageId,
+              originThreadId: sharedThreadId,
+              chatHistory: loadProjectMessages(sharedChannel, {
+                limit: 10,
+                threadId: sharedThreadId,
+                excludeDirect: true,
+              }),
+              broadcastSSE: _deps.broadcastSSE,
+            }).catch((err) => {
+              console.warn(`[chat-handler] Assistant mention routing failed: ${err.message}`);
+            });
+          }
+        } catch (mentionErr) {
+          console.warn(`[chat-handler] Assistant mention dispatch error: ${mentionErr.message}`);
+        }
+      }
     } catch (e) {
       console.warn(
         `[chat-handler] Failed to save assistant message to project store: ${e.message}`,

package/lib/crew-lead/http-server.mjs

@@ -8,7 +8,7 @@ import fs from "node:fs";
 import crypto from "node:crypto";
 import path from "node:path";
 import os from "node:os";
-import { spawn } from "node:child_process";
+import { spawn, execFileSync } from "node:child_process";
 import { executeCLI } from "../bridges/cli-executor.mjs";
 import { applySharedChatPromptOverlay } from "../chat/shared-chat-prompt-overlay.mjs";
 import { classifySharedChatMention } from "../chat/mention-routing-intent.mjs";
@@ -20,7 +20,7 @@ import {
 import { shouldSkipGeminiPassthroughLine } from "../gemini-cli-passthrough-noise.mjs";
 import { applyProjectDirToPipelineSteps } from "../dispatch/parsers.mjs";
 import { normalizeProjectDir } from "../runtime/project-dir.mjs";
-import { CREWSWARM_REPO_ROOT } from "../runtime/config.mjs";
+import { CREWSWARM_REPO_ROOT, loadSwarmConfig } from "../runtime/config.mjs";
 import { resolveCursorLaunchSpec } from "../engines/cursor-launcher.mjs";
 
 let _deps = {};
@@ -60,6 +60,25 @@ function resolveCliBinary(configured, candidates = []) {
   return configured;
 }
 
+function isClaudeOauthAuthenticated() {
+  try {
+    const bin = resolveCliBinary("claude", [
+      path.join(os.homedir(), ".local", "bin", "claude"),
+      "/usr/local/bin/claude",
+      "/opt/homebrew/bin/claude",
+    ]);
+    const output = execFileSync(bin, ["auth", "status"], {
+      encoding: "utf8",
+      stdio: ["ignore", "pipe", "ignore"],
+    }).trim().toLowerCase();
+    if (!output) return false;
+    if (/"loggedin"\s*:\s*true/.test(output)) return true;
+    return output.includes("logged in");
+  } catch {
+    return false;
+  }
+}
+
 /** Strip SGR / OSC ANSI sequences (OpenCode and CLIs often emit colored TTY junk on stdout). */
 function stripAnsiPassthrough(text) {
   return String(text || "")
@@ -94,6 +113,7 @@ function shouldDropPassthroughStderrLine(engine, line) {
   if (/rmcp::/i.test(l)) return true;
   if (/error decoding response body.*initialized notification/i.test(l))
     return true;
+  if (/\[Executor\]\s+\w+\s+\((OAuth|API)\)/i.test(l)) return true;
   if (
     engine === "codex" &&
     /worker quit with fatal/i.test(l) &&
@@ -3532,6 +3552,9 @@ export function createAndStartServer(PORT) {
           sessionId: reqSessionId,
           injectHistory,
           model: reqModel,
+          permissionMode: reqPermissionMode,
+          sandbox: reqSandbox,
+          forceL2: reqForceL2,
         } = JSON.parse(body || "{}");
         if (!message) {
           json(res, 400, { ok: false, error: "message required" });
@@ -3730,7 +3753,7 @@ export function createAndStartServer(PORT) {
           const ocModel =
             reqModel ||
             process.env.CREWSWARM_OPENCODE_MODEL ||
-            "groq/moonshotai/kimi-k2-instruct-0905";
+            "openai/gpt-5.4";
           args = ["run", finalMessage, "--model", ocModel];
           // Add workspace directory context
           if (projectDir) args.push("--dir", projectDir);
@@ -3753,7 +3776,7 @@ export function createAndStartServer(PORT) {
 
           if (priorCodexThread) {
             // `codex resume <thread_id> "prompt"` continues the prior session
-            args = ["resume", priorCodexThread, finalMessage, "--json"];
+            args = ["resume", priorCodexThread, finalMessage];
             if (codexModel) args.push("--model", codexModel);
             if (projectDir) args.push("-C", projectDir);
           } else {
@@ -3762,7 +3785,7 @@ export function createAndStartServer(PORT) {
               "never",
               "exec",
               "--sandbox",
-              "danger-full-access",
+              reqSandbox || "danger-full-access",
               "--skip-git-repo-check",
               "--color",
               "never",
@@ -3803,10 +3826,10 @@ export function createAndStartServer(PORT) {
           const crewCliModel =
             reqModel || process.env.CREWSWARM_CREW_CLI_MODEL || null;
           if (bin.endsWith(".js")) {
-            args = [bin, "chat", finalMessage, "--direct", "--json"];
+            args = [bin, "chat", finalMessage, "--direct", "--json", "--apply"];
             bin = process.execPath;
           } else {
-            args = ["chat", finalMessage, "--direct", "--json"];
+            args = ["chat", finalMessage, "--direct", "--json", "--apply"];
           }
           if (crewCliModel) args.push("--model", crewCliModel);
           if (projectDir && projectDir !== process.cwd())
@@ -3859,12 +3882,16 @@ export function createAndStartServer(PORT) {
             "-p",
             "--setting-sources",
             "user",
-            "--dangerously-skip-permissions",
             "--output-format",
             "stream-json",
             "--verbose",
             "--include-partial-messages",
           ];
+          if (reqPermissionMode) {
+            args.push("--permission-mode", reqPermissionMode);
+          } else {
+            args.push("--dangerously-skip-permissions");
+          }
           if (projectDir) args.push("--add-dir", projectDir);
           if (reqModel) args.push("--model", reqModel);
           // Prefer --resume <session-id> for per-project isolation; fall back to --continue (most recent global)
@@ -3877,7 +3904,10 @@ export function createAndStartServer(PORT) {
           } else if (continueSession) {
             args.push("--continue");
           }
-          args.push(finalMessage);
+          // Skip user MCP servers to avoid 30s+ init hangs
+          args.push("--strict-mcp-config", "--mcp-config", path.join(os.homedir(), ".crewswarm", "config", "empty-mcp.json"));
+          // -- separates flags from prompt (--mcp-config is variadic and eats positional args)
+          args.push("--", finalMessage);
         }
 
         send({ type: "start", engine, message: message.slice(0, 80) });
@@ -3901,9 +3931,41 @@ export function createAndStartServer(PORT) {
         const spawnCwd =
           engine === "claude" && projectDir ? "/tmp" : projectDir;
 
+        // Merge crewswarm.json env vars into spawn env (crew-cli L2, model overrides, etc.)
+        const spawnEnv = { ...process.env };
+        try {
+          const _sysCfg = loadSwarmConfig();
+          if (_sysCfg?.env && typeof _sysCfg.env === "object") {
+            for (const [k, v] of Object.entries(_sysCfg.env)) {
+              if (!spawnEnv[k] && v != null) spawnEnv[k] = String(v);
+            }
+          }
+        } catch {}
+        // Strip API keys AFTER config merge so CLI tools use OAuth (free) not API credits
+        if (engine === "claude") {
+          delete spawnEnv.ANTHROPIC_API_KEY;
+          delete spawnEnv.CLAUDECODE;
+          delete spawnEnv.CLAUDE_CODE;
+          delete spawnEnv.CREWSWARM_CLAUDE_CODE;
+        }
+        if (engine === "codex") {
+          delete spawnEnv.OPENAI_API_KEY;
+        }
+        if (engine === "gemini" || engine === "gemini-cli") {
+          delete spawnEnv.GEMINI_API_KEY;
+          delete spawnEnv.GOOGLE_API_KEY;
+        }
+        if (engine === "opencode" || engine === "antigravity") {
+          delete spawnEnv.OPENAI_API_KEY;
+          delete spawnEnv.ANTHROPIC_API_KEY;
+          delete spawnEnv.GEMINI_API_KEY;
+        }
+        // Force L2 planner for crew-cli when requested (enhance-prompt planner path)
+        if (reqForceL2) spawnEnv.CREW_FORCE_L2 = "true";
+
         const proc = _spawn(bin, args, {
           cwd: spawnCwd,
-          env: process.env,
+          env: spawnEnv,
           stdio: ["ignore", "pipe", "pipe"],
         });
         let lineBuffer = "";
@@ -3916,7 +3978,7 @@ export function createAndStartServer(PORT) {
         let passthroughGracefulEnd = false;
 
         const passthroughTimeoutMs = Number(
-          process.env.CREWSWARM_PASSTHROUGH_TIMEOUT_MS || "240000",
+          process.env.CREWSWARM_PASSTHROUGH_TIMEOUT_MS || "300000",
         );
         const passthroughWatchdog = setTimeout(() => {
           send({
@@ -4206,38 +4268,39 @@ export function createAndStartServer(PORT) {
 
           // crew-cli special handling: extract JSON from output (skipping logs) and send only response field
           if (engine === "crew-cli" && fullOutput.trim()) {
+            let extracted = false;
             try {
-              // crew-cli may emit logs before JSON - use regex to find the JSON object
-              // Pattern matches: crew chat --json outputs "chat.result"
-              const jsonMatch = fullOutput.match(
-                /\{[\s\S]*"kind":\s*"chat\.result"[\s\S]*\}/,
-              );
-              if (!jsonMatch) {
-                throw new Error("No chat.result JSON found in output");
-              }
-              const parsed = JSON.parse(jsonMatch[0]);
-              if (parsed.response) {
-                const responseText = typeof parsed.response === 'string'
-                  ? parsed.response
-                  : JSON.stringify(parsed.response, null, 2);
-                console.error(
-                  `[crew-cli] ✅ Extracted response: ${responseText.slice(0, 80)}...`,
-                );
-                // NOW send the actual response as a chunk
-                send({ type: "chunk", text: responseText });
-                fullOutput = responseText; // Replace fullOutput for notifications
-              } else {
-                console.error(`[crew-cli] ⚠️ No response field in JSON`);
-                send({ type: "chunk", text: fullOutput }); // Send raw output as fallback
+              // Strategy 1: cross-line regex to find JSON envelope
+              const jsonMatch = fullOutput.match(/\{[\s\S]*"kind":\s*"[^"]+\.result"[\s\S]*\}/);
+              if (jsonMatch) {
+                const parsed = JSON.parse(jsonMatch[0]);
+                if (parsed.response) {
+                  const responseText = typeof parsed.response === "string"
+                    ? parsed.response
+                    : JSON.stringify(parsed.response, null, 2);
+                  console.log(`[crew-cli] ✅ Extracted response: ${responseText.slice(0, 80)}...`);
+                  send({ type: "chunk", text: responseText });
+                  fullOutput = responseText;
+                  extracted = true;
+                }
               }
-            } catch (err) {
-              console.error(
-                `[crew-cli] ❌ Failed to extract JSON: ${err.message}`,
-              );
-              console.error(
-                `[crew-cli] First 200 chars: ${fullOutput.slice(0, 200)}`,
-              );
-              // Send raw output as fallback
+            } catch { /* fall through */ }
+            if (!extracted) {
+              // Strategy 2: regex extract just the response field
+              try {
+                const respMatch = fullOutput.match(/"response":\s*"((?:[^"\\]|\\.)*)"/);
+                if (respMatch) {
+                  const responseText = respMatch[1].replace(/\\n/g, "\n").replace(/\\"/g, '"').replace(/\\t/g, "\t");
+                  console.log(`[crew-cli] ✅ Extracted response (regex): ${responseText.slice(0, 80)}...`);
+                  send({ type: "chunk", text: responseText });
+                  fullOutput = responseText;
+                  extracted = true;
+                }
+              } catch { /* fall through */ }
+            }
+            if (!extracted) {
+              console.error(`[crew-cli] ❌ Failed to extract response from crew-cli output`);
+              console.error(`[crew-cli] First 200 chars: ${fullOutput.slice(0, 200)}`);
               send({ type: "chunk", text: fullOutput });
             }
           }
@@ -4495,22 +4558,8 @@ export function createAndStartServer(PORT) {
             );
             enabled = cfg.claudeCode === true;
           } catch { }
-          const hasKey = !!(
-            process.env.ANTHROPIC_API_KEY ||
-            (() => {
-              try {
-                return JSON.parse(
-                  fs.readFileSync(
-                    path.join(os.homedir(), ".crewswarm", "crewswarm.json"),
-                    "utf8",
-                  ),
-                )?.providers?.anthropic?.apiKey;
-              } catch {
-                return null;
-              }
-            })()
-          );
-          json(res, 200, { ok: true, enabled, hasKey });
+          const hasAuth = isClaudeOauthAuthenticated();
+          json(res, 200, { ok: true, enabled, hasKey: hasAuth, hasAuth });
           return;
         }
         if (req.method === "POST") {

package/lib/crew-lead/llm-caller.mjs

@@ -128,14 +128,11 @@ async function callGeminiCliForChat(messages) {
 }
 
 export async function _callLLMOnce(baseUrl, apiKey, modelId, providerKey, messages, options = {}) {
-  // OpenRouter requires full model ID (e.g. openrouter/hunter-alpha), not bare "hunter-alpha"
-  const isOpenRouter = providerKey === "openrouter" || baseUrl.includes("openrouter.ai");
-  if (isOpenRouter && modelId && !modelId.startsWith("openrouter/")) {
-    modelId = "openrouter/" + modelId;
-  }
+  modelId = normalizeExternalModelId(modelId, providerKey, baseUrl);
   const isAnthropic = providerKey === "anthropic" || baseUrl.includes("anthropic.com");
   const enableStreaming = options.stream || false;
   const onStreamToken = options.onStreamToken || null;
+  const requestTimeoutMs = Number(options.timeoutMs || _LLM_TIMEOUT) || _LLM_TIMEOUT;
   
   const headers = { "content-type": "application/json" };
   if (isAnthropic) {
@@ -150,9 +147,10 @@ export async function _callLLMOnce(baseUrl, apiKey, modelId, providerKey, messag
   const isOpenAI = providerKey === "openai" || baseUrl.includes("openai.com");
   
   // Gemini 2.5 Flash: input 1,048,576 tokens / output 65,536 tokens
-  // Anthropic Claude: output typically capped at 8,192 tokens
+  // Anthropic Claude varies by model; Haiku is stricter at 4,096 output tokens
   // All others: 4,096 safe default
-  const maxOutputTokens = isGemini ? 16384 : isAnthropic ? 8192 : 4096;
+  const anthropicMaxOutputTokens = /haiku/i.test(modelId) ? 4096 : 8192;
+  const maxOutputTokens = isGemini ? 16384 : isAnthropic ? anthropicMaxOutputTokens : 4096;
   
   let body, endpoint;
   
@@ -230,7 +228,7 @@ export async function _callLLMOnce(baseUrl, apiKey, modelId, providerKey, messag
     method: "POST",
     headers,
     body: JSON.stringify(body),
-    signal: AbortSignal.timeout(_LLM_TIMEOUT),
+    signal: AbortSignal.timeout(requestTimeoutMs),
   });
 
   if (!res.ok) {
@@ -321,6 +319,24 @@ export async function _callLLMOnce(baseUrl, apiKey, modelId, providerKey, messag
   return reply;
 }
 
+export function normalizeExternalModelId(modelId, providerKey, baseUrl = "") {
+  let normalized = String(modelId || "").trim();
+  if (!normalized) return normalized;
+
+  const isOpenRouter = providerKey === "openrouter" || baseUrl.includes("openrouter.ai");
+  const isPerplexity = providerKey === "perplexity" || baseUrl.includes("api.perplexity.ai");
+
+  if (isOpenRouter) {
+    normalized = normalized.replace(/^openrouter\//i, "");
+  }
+
+  if (isPerplexity) {
+    normalized = normalized.replace(/^perplexity\//i, "");
+  }
+
+  return normalized;
+}
+
 function _recordCrewLeadTokens(modelId, providerKey, usage) {
   if (!usage) return;
   const p = Number(usage.prompt_tokens || usage.input_tokens || 0);

package/lib/crew-lead/prompts.mjs

@@ -41,6 +41,12 @@ export function initPrompts({
 let _sysPromptCache = null;
 let _sysPromptKey = "";
 
+/** Clear the system prompt memoization cache (for test isolation). */
+export function resetPromptCache() {
+  _sysPromptCache = null;
+  _sysPromptKey = "";
+}
+
 export function buildSystemPrompt(cfg) {
   // Memoize — only rebuild when config files or agent prompts change
   const keyParts = [cfg.providerKey, cfg.modelId, cfg.displayName];
@@ -137,7 +143,7 @@ export function buildSystemPrompt(cfg) {
     "DEFAULT: CHAT. You are a conversational assistant first.",
     "- Questions, explanations, status, clarifications, follow-ups, 'how does X work', 'what is X', 'can you', 'show me' → ANSWER. Never dispatch.",
     "- Short messages under ~8 words ('i mean X', 'no, X', 'about X') → corrections, not directives. ANSWER.",
-    "- Greetings ('hi', 'yo', 'hey') → reply briefly. No health dump unless asked.",
+    "- Greetings ('hi', 'yo', 'hey', 'hello', 'sup', 'what's up') → reply with 1-2 SHORT sentences. Do NOT report system status, agent counts, RT bus, skills, or health data. Just be conversational. Only report health/status when the user EXPLICITLY asks ('status', 'health', 'how is the system', 'what's running').",
     "- Self-audit requests (find bugs, review, inspect) → DO IT YOURSELF with @@READ_FILE / @@RUN_CMD. No asking permission. Only dispatch to crew-qa/crew-coder if user explicitly says 'have X do it'.",
     "",
     "DISPATCH: when user gives explicit action language.",
@@ -150,6 +156,13 @@ export function buildSystemPrompt(cfg) {
     `- With acceptance criteria: @@DISPATCH {"agent":"crew-coder","task":"Write JWT auth middleware","verify":"@@READ_FILE src/auth.ts","done":"exports verifyToken, returns 401 on invalid"}`,
     "- You MUST emit the @@DISPATCH line. Describing a dispatch in prose without the line = NOTHING happens.",
     "",
+    "ENGINE ROUTING: tasks are routed to CLI engines or direct-llm based on keyword matching.",
+    "- Tasks with coding keywords (create, build, implement, fix, refactor, add, update, modify, code, function, class, component, api, endpoint, route, test, bug, error, issue, file, script, module, package) → CLI engine (native file I/O, shell, full repo context).",
+    "- Tasks WITHOUT coding keywords → direct-llm (LLM API call; can still write files via @@WRITE_FILE markers, but no shell or repo context).",
+    "- CLI engines are better for complex multi-file coding. direct-llm is fine for research reports, single-file writes, and simple tasks.",
+    "- When dispatching build tasks, use verbs like 'Create', 'Build', 'Implement', 'Fix' to ensure CLI routing.",
+    "- This is automatic keyword matching — no LLM call. See docs/ORCHESTRATION-PROTOCOL.md for the full keyword list.",
+    "",
     "PIPELINE: when user wants multi-agent coordinated work.",
     "- Triggers: 'build me X', 'create X', 'kick off', 'rally the crew', 'dispatch the crew'.",
     "- Before firing for complex tasks (3+ agents / 2+ waves): show plan in 1-3 lines, ask 'Fire it?'. Skip confirmation for single-agent or explicit 'go build'.",

package/lib/crew-lead/tools.mjs

@@ -154,10 +154,11 @@ export async function execCrewLeadTools(reply) {
         } catch { return null; }
       })();
       
-      // Fallback: check search-tools.json
+      // Fallback: check search-tools.json (same dir as config file)
       if (!braveKey) {
         try {
-          const searchTools = JSON.parse(fs.readFileSync(path.join(os.homedir(), ".crewswarm", "search-tools.json"), "utf8"));
+          const cfgDir = _crewswarmCfgFile ? path.dirname(_crewswarmCfgFile) : path.join(os.homedir(), ".crewswarm");
+          const searchTools = JSON.parse(fs.readFileSync(path.join(cfgDir, "search-tools.json"), "utf8"));
           braveKey = searchTools.brave?.apiKey;
         } catch {}
       }

package/lib/crew-lead/wave-dispatcher.mjs

@@ -377,7 +377,17 @@ export function dispatchPipelineWave(pipelineId) {
   // this path: when Cursor CLI fails, orchestrator falls back to plain LLM text
   // with @mentions — subagents never run and the pipeline advances on garbage.
   // Direct per-agent dispatch is reliable for project pipelines.
-  if (_deps._cursorWavesEnabled && waveSteps.length > 1 && !pipeline.projectDir) {
+  // Only route through orchestrator if it has a CLI engine that can actually dispatch.
+  // Without a CLI engine, the orchestrator just describes dispatching in text — useless.
+  const orchestratorHasCliEngine = (() => {
+    try {
+      const agents = (_deps.loadAgentList || (() => []))();
+      const orch = agents.find(a => a.id === "crew-orchestrator");
+      if (!orch) return false;
+      return !!(orch.engine || orch.useClaudeCode || orch.useCursorCli || orch.useCursor || orch.useCodex || orch.useCrewCLI || orch.useGeminiCli || orch.useOpenCode);
+    } catch { return false; }
+  })();
+  if (_deps._cursorWavesEnabled && orchestratorHasCliEngine && waveSteps.length > 1 && !pipeline.projectDir) {
     const waveManifest = {
       wave: currentWave + 1,
       projectDir: pipeline.projectDir || "",
@@ -525,10 +535,12 @@ export function checkWaveQualityGate(pipeline, pipelineId) {
       const hasWriteFile = /@@WRITE_FILE/.test(result);
       const wroteFile = /(?:wrote|created|saved|written|updated|enhanced|implemented|added)\s+(?:to\s+)?(?:\/\S+|[a-zA-Z][\w/.-]+\.(?:html|css|js|ts|py|json|md|txt|sh|yaml|yml))/i.test(result);
       const opencodeWrote = /(?:←\s*Write|Wrote file|Write\s+\.\.[\w/.-]+\.(?:html|css|js|ts|py|json|md)|Created\s+`\/)/i.test(result);
+      // Cursor outputs "`filename` is in place at:" or "Created the file at:\n`/path`"
+      const cursorInPlace = /`[\w/.-]+`\s+is\s+in\s+place\s+at|Created\s+the\s+file\s+at:/i.test(result);
       const explicitDone = /Done\.\s+(?:Created|Updated|Enhanced|Implemented|The\s+(?:file|prototype|component))/i.test(result);
       // OpenCode agents write silently — check if any src files changed in last 20 min as fallback
       let opencodeFilesChanged = false;
-      if (!hasWriteFile && !wroteFile && !opencodeWrote && !explicitDone) {
+      if (!hasWriteFile && !wroteFile && !opencodeWrote && !cursorInPlace && !explicitDone) {
         try {
           const cutoff = Date.now() - 20 * 60 * 1000;
           // Use dynamic project directory from pipeline metadata or environment
@@ -555,7 +567,7 @@ export function checkWaveQualityGate(pipeline, pipelineId) {
           opencodeFilesChanged = dirs.some(checkDir);
         } catch {}
       }
-      if (!hasWriteFile && !wroteFile && !opencodeWrote && !explicitDone && !opencodeFilesChanged) {
+      if (!hasWriteFile && !wroteFile && !opencodeWrote && !cursorInPlace && !explicitDone && !opencodeFilesChanged) {
         issues.push(`${agent} (builder) did not write any files`);
       }
     }
@@ -702,8 +714,10 @@ export function dispatchTask(agent, task, sessionId = "owner", pipelineMeta = nu
     task = _deps.writeTaskBrief?.(agent, task, projectDir) ?? task;
   }
 
-  const rp = typeof _deps.getRtPublish === "function" ? _deps.getRtPublish() : null;
-  console.log(`[wave-dispatcher] getRtPublish exists: ${typeof _deps.getRtPublish === "function"}, rp is: ${typeof rp}`);
+  let rp = typeof _deps.getRtPublish === "function" ? _deps.getRtPublish() : null;
+  if (!rp && typeof _deps.getRtPublish === "function") {
+    console.warn(`[wave-dispatcher] RT publish not available for ${agent} — will use fallback`);
+  }
   if (rp) {
     try {
       // Build extraFlags: start with global settings, override with pipeline-specific settings