cascade-ai 0.13.1 → 0.13.2

package/dist/cli.js

@@ -58,7 +58,7 @@ var __export = (target, all) => {
 var CASCADE_VERSION, CASCADE_CONFIG_FILE, CASCADE_DB_FILE, CASCADE_DASHBOARD_SECRET_FILE, GLOBAL_CONFIG_DIR, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, DEFAULT_DASHBOARD_PORT, DEFAULT_CONTEXT_LIMIT, DEFAULT_AUTO_SUMMARIZE_AT, MODELS, T1_MODEL_PRIORITY, T2_MODEL_PRIORITY, T3_MODEL_PRIORITY, VISION_MODEL_PRIORITY, COMPLEXITY_T2_COUNT, THEME_NAMES, DEFAULT_THEME, OLLAMA_BASE_URL, LM_STUDIO_BASE_URL, AZURE_BASE_URL_TEMPLATE, TOOL_NAMES, DEFAULT_APPROVAL_REQUIRED;
 var init_constants = __esm({
   "src/constants.ts"() {
-    CASCADE_VERSION = "0.13.1";
+    CASCADE_VERSION = "0.13.2";
     CASCADE_CONFIG_FILE = ".cascade/config.json";
     CASCADE_DB_FILE = ".cascade/memory.db";
     CASCADE_DASHBOARD_SECRET_FILE = ".cascade/dashboard-secret";
@@ -2989,6 +2989,8 @@ var CascadeConfigSchema = z.object({
   privacy: z.object({
     paths: z.array(z.object({ pattern: z.string().min(1), policy: z.enum(["local-only"]) })).default([])
   }).optional(),
+  /** Routing controls — forceTier pins the root tier, bypassing the classifier. */
+  routing: z.object({ forceTier: z.enum(["auto", "T1", "T2", "T3"]).default("auto") }).optional(),
   /**
    * T3→T2 reinforcement: when enabled, a worker that discovers its subtask should
    * fan out can call the `request_workers` tool to have its T2 manager spawn
@@ -4974,6 +4976,13 @@ var BaseTier = class extends EventEmitter {
   hierarchyContext = "";
   /** Propagated AbortSignal — set by the tier's `execute()` before work begins. */
   signal;
+  /**
+   * True for the run's ROOT tier (T3 for Simple, T2 for Moderate, T1 for
+   * Complex). Its own synthesis stream is the user-facing answer and is
+   * tagged `primary` so the desktop renders it live — background workers,
+   * which would interleave, are not tagged.
+   */
+  isPresenter = false;
   constructor(role, id, parentId) {
     super();
     this.role = role;
@@ -4981,6 +4990,10 @@ var BaseTier = class extends EventEmitter {
     this.parentId = parentId;
     this.label = this.id;
   }
+  /** Mark this tier as the run's presenter (root tier). */
+  setPresenter(on = true) {
+    this.isPresenter = on;
+  }
   getStatus() {
     return this.status;
   }
@@ -5716,7 +5729,7 @@ HIERARCHY CONTEXT: ${this.hierarchyContext}` : "") + textToolSuffix,
         "T3",
         options,
         (chunk) => {
-          this.emit("stream:token", { tierId: this.id, text: chunk.text });
+          this.emit("stream:token", { tierId: this.id, text: chunk.text, primary: this.isPresenter });
         }
       );
       let effectiveToolCalls = result.toolCalls ?? [];
@@ -7177,6 +7190,7 @@ ${peerOutputs}` : "";
         chunkEnd++;
       }
       i = chunkEnd;
+      const isLastChunk = chunkEnd >= completed.length;
       const prompt = `Summarize these T3 worker outputs for section "${assignment.sectionTitle}" in 2-3 sentences.
   ${currentSummary ? `
 PREVIOUS SUMMARY SO FAR:
@@ -7186,6 +7200,7 @@ NEW OUTPUTS TO INTEGRATE:
 ` : "\nOUTPUTS:\n"}${chunkText}${peerContext}`;
       const messages = [{ role: "user", content: prompt }];
       try {
+        const streamFinal = isLastChunk && this.isPresenter ? (chunk) => this.emit("stream:token", { tierId: this.id, text: chunk.text, primary: true }) : void 0;
         const result = await this.router.generate("T2", {
           messages,
           systemPrompt: this.systemPromptOverride + "You are a T2 Manager. Summarize the work of your T3 workers succinctly." + (this.hierarchyContext ? `
@@ -7193,7 +7208,7 @@ NEW OUTPUTS TO INTEGRATE:
 HIERARCHY CONTEXT: ${this.hierarchyContext}` : ""),
           maxTokens: 500,
           ...this.sectionModel ? { model: this.sectionModel } : {}
-        });
+        }, streamFinal);
         currentSummary = result.content;
       } catch (err) {
         this.log(`aggregateResults: LLM summarization failed at chunk \u2014 returning raw T3 outputs. Error: ${err instanceof Error ? err.message : String(err)}`);
@@ -7245,14 +7260,11 @@ HIERARCHY CONTEXT: ${this.hierarchyContext}` : ""),
         ...this.sectionModel ? { model: this.sectionModel } : {}
       });
       const answer = result.content.trim().toUpperCase();
-      if (answer.includes("YES")) {
-        return { requestId: req.id, approved: true, always: true, decidedBy: "T2", reasoning: "T2 LLM evaluated: consistent with section goal" };
-      }
-      if (answer.includes("NO")) {
-        return { requestId: req.id, approved: false, always: true, decidedBy: "T2", reasoning: "T2 LLM evaluated: inconsistent with section goal" };
-      }
+      const verdict = answer.includes("YES") ? "approve" : answer.includes("NO") ? "deny" : "unsure";
+      (req.trail ??= []).push({ tier: "T2", verdict, reason: `T2: ${verdict === "approve" ? "consistent with section goal" : verdict === "deny" ? "inconsistent with section goal" : "unsure"}` });
       return null;
     } catch {
+      (req.trail ??= []).push({ tier: "T2", verdict: "unsure", reason: "T2 evaluation failed" });
       return null;
     }
   }
@@ -7934,7 +7946,7 @@ Instructions:
       systemPrompt: this.systemPromptOverride + "You are a final output compiler. Summarize and format the task results clearly.",
       maxTokens: 8e3
     }, (chunk) => {
-      this.emit("stream:token", { tierId: this.id, text: chunk.text });
+      this.emit("stream:token", { tierId: this.id, text: chunk.text, primary: this.isPresenter });
     });
     return result.content;
   }
@@ -7963,14 +7975,11 @@ Reply with exactly one word: YES, NO, or UNSURE.
         temperature: 0
       });
       const answer = result.content.trim().toUpperCase();
-      if (answer.includes("YES")) {
-        return { requestId: req.id, approved: true, always: true, decidedBy: "T1", reasoning: "T1 evaluated: consistent with overall task goal" };
-      }
-      if (answer.includes("NO")) {
-        return { requestId: req.id, approved: false, always: true, decidedBy: "T1", reasoning: "T1 evaluated: not consistent with overall task goal" };
-      }
+      const verdict = answer.includes("YES") ? "approve" : answer.includes("NO") ? "deny" : "unsure";
+      (req.trail ??= []).push({ tier: "T1", verdict, reason: `T1: ${verdict === "approve" ? "consistent with overall task goal" : verdict === "deny" ? "not consistent with overall task goal" : "unsure"}` });
       return null;
     } catch {
+      (req.trail ??= []).push({ tier: "T1", verdict: "unsure", reason: "T1 evaluation failed" });
       return null;
     }
   }
@@ -11066,6 +11075,21 @@ ${last.partialOutput}` : "");
     const producesArtifact = /\b(?:create|build|implement|generate|write|refactor|rewrite|add|fix|deploy|install|migrate|scaffold|set up|save (?:a|the)|report|\.(?:pdf|md|txt|json|csv|py|js|ts|tsx|jsx|html|docx?))\b/i.test(p);
     return inquiry && !producesArtifact;
   }
+  /**
+   * Strong, explicit signals that a task needs the full hierarchy (planning +
+   * multiple sections/workers). Deliberately conservative: requires a
+   * build/implementation verb AND either an app/system-scale noun or an
+   * explicit multi-part structure, so ordinary single-file asks (handled as
+   * Simple/Moderate) don't get over-escalated.
+   */
+  looksClearlyComplex(prompt) {
+    const p = prompt.trim();
+    if (p.length < 24) return false;
+    const buildVerb = /\b(?:build|implement|create|develop|design|scaffold|refactor|migrate|architect|set up|integrate)\b/i.test(p);
+    const scaleNoun = /\b(?:app(?:lication)?|system|platform|service|api|backend|frontend|full[- ]?stack|website|dashboard|pipeline|microservices?|database schema|authentication|end[- ]to[- ]end|codebase|project|multiple files|several (?:files|modules|components)|test suite)\b/i.test(p);
+    const multiPart = /(?:\b(?:and|then|also|plus|as well as)\b.*\b(?:and|then|also)\b)|(?:^|\n)\s*(?:[-*]|\d+[.)])\s+/i.test(p);
+    return buildVerb && (scaleNoun || multiPart);
+  }
   // Cache glob scan results per workspace path to avoid repeated I/O.
   static globCache = /* @__PURE__ */ new Map();
   async countWorkspaceFiles(workspacePath) {
@@ -11149,10 +11173,16 @@ ${prompt}` : prompt;
       let verdict;
       if (match) {
         verdict = match[1] === "simple" ? "Simple" : match[1] === "moderate" ? "Moderate" : "Complex";
-        this.recordDecision("complexity", `${verdict} \u2014 classifier: ${reason || "no reason given"}`);
+        if (verdict !== "Complex" && this.looksClearlyComplex(prompt)) {
+          this.recordDecision("complexity", `Complex \u2014 heuristic floor over classifier "${verdict}": explicit multi-step build/implementation signals (T1 engaged)`);
+          verdict = "Complex";
+        } else {
+          this.recordDecision("complexity", `${verdict} \u2014 classifier: ${reason || "no reason given"}`);
+        }
       } else {
-        verdict = prompt.trim().split(/\s+/).length <= 12 ? "Simple" : "Moderate";
-        this.recordDecision("complexity", `${verdict} \u2014 classifier output unparseable; defaulted by length`);
+        const words = prompt.trim().split(/\s+/).length;
+        verdict = words <= 12 ? "Simple" : words >= 40 || this.looksClearlyComplex(prompt) ? "Complex" : "Moderate";
+        this.recordDecision("complexity", `${verdict} \u2014 classifier output unparseable; defaulted by length/signals`);
       }
       return verdict;
     } catch {
@@ -11204,7 +11234,15 @@ ${prompt}` : prompt;
       }
       escalator.resolveUserDecision(req.id, approved, always);
     });
-    const complexity = await this.determineComplexity(options.prompt, options.workspacePath || process.cwd(), options.conversationHistory);
+    const forceTier = this.config.routing?.forceTier;
+    const forced = forceTier === "T1" ? "Complex" : forceTier === "T2" ? "Moderate" : forceTier === "T3" ? "Simple" : void 0;
+    let complexity;
+    if (forced) {
+      complexity = forced;
+      this.recordDecision("complexity", `${forced} \u2014 manually forced via routing.forceTier=${forceTier}`);
+    } else {
+      complexity = await this.determineComplexity(options.prompt, options.workspacePath || process.cwd(), options.conversationHistory);
+    }
     this.telemetry.capture("cascade:session_start", {
       complexity,
       providerCount: this.config.providers.length,
@@ -11284,6 +11322,7 @@ ${prompt}` : prompt;
     try {
       if (complexity === "Simple") {
         const t3 = new T3Worker(this.router, this.toolRegistry, "root");
+        t3.setPresenter(true);
         t3.setHierarchyContext("You are the DIRECT worker for this task. There is no T1 Administrator or T2 Manager involved in this run.");
         if (identityPrompt) {
           t3.setSystemPromptOverride(identityPrompt);
@@ -11308,6 +11347,7 @@ ${prompt}` : prompt;
         this.emit("tier:status", { tierId: "t3-root", status: "COMPLETED", role: "T3" });
       } else if (complexity === "Moderate") {
         const t2 = new T2Manager(this.router, this.toolRegistry, "root");
+        t2.setPresenter(true);
         t2.setHierarchyContext("You are the ROOT Manager for this task. There is no T1 Administrator involved in this run. You are responsible for decomposing the task and managing T3 workers directly.");
         if (identityPrompt) {
           t2.setSystemPromptOverride(identityPrompt);
@@ -11357,6 +11397,7 @@ ${prompt}` : prompt;
         }
       } else {
         const t1 = new T1Administrator(this.router, this.toolRegistry, this.config);
+        t1.setPresenter(true);
         t1.setHierarchyContext("You are the top-level Administrator. You are responsible for the overall plan and supervising multiple T2 Managers.");
         if (identityPrompt) {
           t1.setSystemPromptOverride(identityPrompt);
@@ -15357,7 +15398,8 @@ var DashboardSocket = class {
       socket.on("cascade:run", (payload) => {
         if (typeof payload?.prompt === "string" && payload.prompt.trim()) {
           const sessionId = typeof payload.sessionId === "string" && payload.sessionId.trim() ? payload.sessionId.trim() : void 0;
-          callback(payload.prompt.trim(), payload.model ?? "auto", socket.id, sessionId);
+          const forceTier = ["T1", "T2", "T3"].includes(payload.forceTier) ? payload.forceTier : void 0;
+          callback(payload.prompt.trim(), payload.model ?? "auto", socket.id, sessionId, forceTier);
         }
       });
     });
@@ -15387,6 +15429,13 @@ var DashboardServer = class {
    * of runs the session performed in this server's lifetime.
    */
   sessionTaskIds = /* @__PURE__ */ new Map();
+  /**
+   * Tool-approval requests awaiting a user decision from a connected client,
+   * keyed by the request's uuid. The desktop shows a modal on
+   * `permission:user-required` and answers with `permission:decision`; this
+   * map is how that answer reaches the run that's blocked on it.
+   */
+  pendingApprovals = /* @__PURE__ */ new Map();
   port;
   host;
   workspacePath;
@@ -15445,17 +15494,18 @@ var DashboardServer = class {
       }
       this.persistConfig();
     });
-    this.socket.onCascadeRun(async (prompt, model, socketId, requestedSessionId) => {
+    this.socket.onCascadeRun(async (prompt, model, socketId, requestedSessionId, forceTier) => {
       const sessionId = requestedSessionId ?? randomUUID();
       const abortController = new AbortController();
       this.activeControllers.set(sessionId, abortController);
       const runPrompt = requestedSessionId ? this.buildContinuationPrompt(sessionId, prompt) : prompt;
       const title = this.persistRunStart(sessionId, prompt);
-      const cfg = model !== "auto" ? { ...this.config, models: { ...this.config.models, t1: model } } : this.config;
+      let cfg = model !== "auto" ? { ...this.config, models: { ...this.config.models, t1: model } } : this.config;
+      if (forceTier) cfg = { ...cfg, routing: { ...cfg.routing, forceTier } };
       const cascade = new Cascade(cfg, this.workspacePath, this.store);
       this.activeSessions.set(sessionId, cascade);
       cascade.on("stream:token", (e) => {
-        this.socket.emitToSocket(socketId, "stream:token", { sessionId, tierId: e.tierId, text: e.text });
+        this.socket.emitToSocket(socketId, "stream:token", { sessionId, tierId: e.tierId, text: e.text, primary: e.primary });
       });
       cascade.on("tier:status", (e) => {
         this.socket.emitToSocket(socketId, "tier:status", { sessionId, ...e });
@@ -15467,7 +15517,11 @@ var DashboardServer = class {
         this.socket.emitPeerMessage(e);
       });
       try {
-        const result = await cascade.run({ prompt: runPrompt, signal: abortController.signal });
+        const result = await cascade.run({
+          prompt: runPrompt,
+          signal: abortController.signal,
+          approvalCallback: this.makeApprovalCallback(sessionId)
+        });
         this.recordSessionTask(sessionId, result.taskId);
         this.persistRunEnd(sessionId, title, prompt, result.output, "COMPLETED");
         this.socket.emitToSocket(socketId, "session:complete", { sessionId, result });
@@ -15486,10 +15540,18 @@ var DashboardServer = class {
       } finally {
         this.activeSessions.delete(sessionId);
         this.activeControllers.delete(sessionId);
+        this.denyPendingApprovals(sessionId);
       }
     });
     this.socket.onSessionHalt((sessionId) => {
       this.activeControllers.get(sessionId)?.abort();
+      this.denyPendingApprovals(sessionId);
+    });
+    this.socket.onApprovalResponse(({ requestId, approved, always }) => {
+      const pending = this.pendingApprovals.get(requestId);
+      if (!pending) return;
+      this.pendingApprovals.delete(requestId);
+      pending.resolve({ approved: !!approved, always: !!always });
     });
     this.socket.onSessionSteer((message, sessionId, nodeId) => {
       const steered = this.steerSessions(message, sessionId, nodeId);
@@ -15657,6 +15719,29 @@ var DashboardServer = class {
     list.push(taskId);
     this.sessionTaskIds.set(sessionId, list);
   }
+  /**
+   * Approval bridge: cascade calls this when a dangerous tool escalates to the
+   * user. The request itself was already pushed to the client via the
+   * `permission:user-required` forward; here we just park a resolver keyed by
+   * the request id and wait for the client's `permission:decision` (handled in
+   * onApprovalResponse). Never auto-approves — an unanswered request stays
+   * pending until the client answers, the run ends, or the escalator's own
+   * timeout denies it.
+   */
+  makeApprovalCallback(sessionId) {
+    return (request) => new Promise((resolve) => {
+      this.pendingApprovals.set(request.id, { resolve, sessionId });
+    });
+  }
+  /** Deny + clear any approvals still pending for a session (run end / abort). */
+  denyPendingApprovals(sessionId) {
+    for (const [id, pending] of this.pendingApprovals) {
+      if (pending.sessionId === sessionId) {
+        this.pendingApprovals.delete(id);
+        pending.resolve({ approved: false, always: false });
+      }
+    }
+  }
   persistRuntimeRow(sessionId, title, status, latestPrompt) {
     const now = (/* @__PURE__ */ new Date()).toISOString();
     const row = { sessionId, title, workspacePath: this.workspacePath, status, startedAt: now, updatedAt: now, latestPrompt, isGlobal: false };
@@ -15845,15 +15930,6 @@ ${prompt}`;
       if (sessionId) this.socket.broadcastToRoom(`session:${sessionId}`, "session:halt", payload);
       res.json({ success: true, ...payload });
     });
-    this.app.post("/api/approve", auth, mutationLimiter, (req, res) => {
-      const body = req.body;
-      const sessionId = typeof body["sessionId"] === "string" ? body["sessionId"] : void 0;
-      const nodeId = typeof body["nodeId"] === "string" ? body["nodeId"] : void 0;
-      const payload = { sessionId, nodeId, requestedAt: (/* @__PURE__ */ new Date()).toISOString() };
-      this.socket.broadcast("session:approve", payload);
-      if (sessionId) this.socket.broadcastToRoom(`session:${sessionId}`, "session:approve", payload);
-      res.json({ success: true, ...payload });
-    });
     this.app.post("/api/inject", auth, mutationLimiter, (req, res) => {
       const body = req.body;
       const message = typeof body["message"] === "string" ? body["message"] : void 0;
@@ -16104,7 +16180,7 @@ ${prompt}`;
         const cascade = new Cascade(this.config, this.workspacePath, this.store);
         this.activeSessions.set(sessionId, cascade);
         cascade.on("stream:token", (e) => {
-          this.socket.broadcastToRoom(`session:${sessionId}`, "stream:token", { sessionId, tierId: e.tierId, text: e.text });
+          this.socket.broadcastToRoom(`session:${sessionId}`, "stream:token", { sessionId, tierId: e.tierId, text: e.text, primary: e.primary });
         });
         cascade.on("tier:status", (e) => {
           this.socket.broadcastToRoom(`session:${sessionId}`, "tier:status", { sessionId, ...e });
@@ -16116,7 +16192,11 @@ ${prompt}`;
           this.socket.emitPeerMessage(e);
         });
         try {
-          const result = await cascade.run({ prompt: runPrompt, identityId: body.identityId });
+          const result = await cascade.run({
+            prompt: runPrompt,
+            identityId: body.identityId,
+            approvalCallback: this.makeApprovalCallback(sessionId)
+          });
           this.recordSessionTask(sessionId, result.taskId);
           this.persistRunEnd(sessionId, title, prompt, result.output, "COMPLETED");
           this.socket.broadcast("cost:update", {
@@ -16134,6 +16214,7 @@ ${prompt}`;
           });
         } finally {
           this.activeSessions.delete(sessionId);
+          this.denyPendingApprovals(sessionId);
         }
       })();
     });