cascade-ai 0.9.7 → 0.10.3

package/dist/index.d.cts

@@ -691,6 +691,13 @@ declare class ModelSelector {
     private resolveDynamicModel;
 }
 
+interface ModelStat {
+    successCount: number;
+    failureCount: number;
+    totalRetries: number;
+    totalCostUsd: number;
+    sampleCount: number;
+}
 declare class ModelPerformanceTracker {
     private stats;
     private readonly statsFile;
@@ -699,6 +706,13 @@ declare class ModelPerformanceTracker {
     load(): Promise<void>;
     save(): Promise<void>;
     record(modelId: string, taskType: TaskType, outcome: 'success' | 'failure', retries?: number, costUsd?: number): void;
+    /**
+     * Record an explicit user rating (good/bad). Counts as 3 automatic samples
+     * so user feedback carries significantly more weight than auto-detected outcomes.
+     */
+    recordExplicit(modelId: string, taskType: TaskType, rating: 'good' | 'bad', costUsd?: number): void;
+    /** Returns all stats keyed by "modelId:taskType" — used by `cascade stats`. */
+    getAll(): Map<string, ModelStat>;
     /**
      * Returns 0.05–1.0; defaults to 0.5 (neutral prior) when no history exists.
      * High retry counts penalise the score.
@@ -754,6 +768,12 @@ declare class TaskAnalyzer {
      * during this session and persist stats to disk.
      */
     recordRunOutcome(outcome: 'success' | 'failure', costByTier: Record<string, number>): void;
+    /**
+     * Record an explicit user rating (good/bad) for the last run's selected models.
+     * Explicit ratings carry 3× the weight of auto-detected outcomes.
+     * Does NOT clear lastSelectedModels — the auto record already did that.
+     */
+    recordExplicitRating(rating: 'good' | 'bad'): boolean;
     private scoreModel;
     private costEfficiency;
     private taskMatchScore;
@@ -1521,6 +1541,13 @@ declare class Cascade extends EventEmitter {
     resumeRun(opts?: {
         maxTokens?: number;
     }): Promise<CascadeRunResult | null>;
+    /**
+     * Record an explicit user rating for the last completed run.
+     * Explicit ratings carry 3× the weight of auto-detected outcomes so user
+     * feedback meaningfully shifts future routing decisions.
+     * Returns false when called before any task has run in this session.
+     */
+    rateLastRun(rating: 'good' | 'bad'): boolean;
     /**
      * Rough pre-execution cost estimate for a plan: ~3 T2 calls per section
      * plus ~4 T3 calls per subtask at typical token volumes. A ballpark for
@@ -1988,6 +2015,15 @@ declare class DashboardSocket {
     emitApprovalRequest(request: PermissionRequest): void;
     onApprovalResponse(callback: (data: PermissionDecisionPayload) => void): void;
     private setupHandlers;
+    onSessionRate(callback: (sessionId: string, rating: 'good' | 'bad') => void): void;
+    onConfigUpdate(callback: (data: {
+        keys?: Record<string, string>;
+        models?: Record<string, string>;
+        budget?: {
+            maxCostPerRun?: number;
+            autoBias?: string;
+        };
+    }) => void): void;
     close(): void;
 }
 
@@ -2000,6 +2036,7 @@ declare class DashboardServer {
     private store;
     private globalStore;
     private broadcastTimer;
+    private activeSessions;
     private port;
     private host;
     private workspacePath;

package/dist/index.d.ts

@@ -691,6 +691,13 @@ declare class ModelSelector {
     private resolveDynamicModel;
 }
 
+interface ModelStat {
+    successCount: number;
+    failureCount: number;
+    totalRetries: number;
+    totalCostUsd: number;
+    sampleCount: number;
+}
 declare class ModelPerformanceTracker {
     private stats;
     private readonly statsFile;
@@ -699,6 +706,13 @@ declare class ModelPerformanceTracker {
     load(): Promise<void>;
     save(): Promise<void>;
     record(modelId: string, taskType: TaskType, outcome: 'success' | 'failure', retries?: number, costUsd?: number): void;
+    /**
+     * Record an explicit user rating (good/bad). Counts as 3 automatic samples
+     * so user feedback carries significantly more weight than auto-detected outcomes.
+     */
+    recordExplicit(modelId: string, taskType: TaskType, rating: 'good' | 'bad', costUsd?: number): void;
+    /** Returns all stats keyed by "modelId:taskType" — used by `cascade stats`. */
+    getAll(): Map<string, ModelStat>;
     /**
      * Returns 0.05–1.0; defaults to 0.5 (neutral prior) when no history exists.
      * High retry counts penalise the score.
@@ -754,6 +768,12 @@ declare class TaskAnalyzer {
      * during this session and persist stats to disk.
      */
     recordRunOutcome(outcome: 'success' | 'failure', costByTier: Record<string, number>): void;
+    /**
+     * Record an explicit user rating (good/bad) for the last run's selected models.
+     * Explicit ratings carry 3× the weight of auto-detected outcomes.
+     * Does NOT clear lastSelectedModels — the auto record already did that.
+     */
+    recordExplicitRating(rating: 'good' | 'bad'): boolean;
     private scoreModel;
     private costEfficiency;
     private taskMatchScore;
@@ -1521,6 +1541,13 @@ declare class Cascade extends EventEmitter {
     resumeRun(opts?: {
         maxTokens?: number;
     }): Promise<CascadeRunResult | null>;
+    /**
+     * Record an explicit user rating for the last completed run.
+     * Explicit ratings carry 3× the weight of auto-detected outcomes so user
+     * feedback meaningfully shifts future routing decisions.
+     * Returns false when called before any task has run in this session.
+     */
+    rateLastRun(rating: 'good' | 'bad'): boolean;
     /**
      * Rough pre-execution cost estimate for a plan: ~3 T2 calls per section
      * plus ~4 T3 calls per subtask at typical token volumes. A ballpark for
@@ -1988,6 +2015,15 @@ declare class DashboardSocket {
     emitApprovalRequest(request: PermissionRequest): void;
     onApprovalResponse(callback: (data: PermissionDecisionPayload) => void): void;
     private setupHandlers;
+    onSessionRate(callback: (sessionId: string, rating: 'good' | 'bad') => void): void;
+    onConfigUpdate(callback: (data: {
+        keys?: Record<string, string>;
+        models?: Record<string, string>;
+        budget?: {
+            maxCostPerRun?: number;
+            autoBias?: string;
+        };
+    }) => void): void;
     close(): void;
 }
 
@@ -2000,6 +2036,7 @@ declare class DashboardServer {
     private store;
     private globalStore;
     private broadcastTimer;
+    private activeSessions;
     private port;
     private host;
     private workspacePath;

package/dist/index.js

@@ -8018,6 +8018,19 @@ var TaskAnalyzer = class {
     this.lastSelectedModels.clear();
     void this.tracker.save();
   }
+  /**
+   * Record an explicit user rating (good/bad) for the last run's selected models.
+   * Explicit ratings carry 3× the weight of auto-detected outcomes.
+   * Does NOT clear lastSelectedModels — the auto record already did that.
+   */
+  recordExplicitRating(rating) {
+    if (!this.tracker || !this.lastProfile) return false;
+    const taskType = this.lastProfile.type;
+    for (const [, model] of this.lastSelectedModels) {
+      this.tracker.recordExplicit(model.id, taskType, rating, 0);
+    }
+    return this.lastSelectedModels.size > 0;
+  }
   scoreModel(model, profile) {
     const perf = this.tracker?.performanceScore(model.id, profile.type) ?? 0.5;
     const costEff = this.costEfficiency(model, profile.complexity);
@@ -8097,6 +8110,20 @@ var ModelPerformanceTracker = class {
       sampleCount: s.sampleCount + 1
     });
   }
+  /**
+   * Record an explicit user rating (good/bad). Counts as 3 automatic samples
+   * so user feedback carries significantly more weight than auto-detected outcomes.
+   */
+  recordExplicit(modelId, taskType, rating, costUsd = 0) {
+    const outcome = rating === "good" ? "success" : "failure";
+    this.record(modelId, taskType, outcome, 0, costUsd);
+    this.record(modelId, taskType, outcome, 0, 0);
+    this.record(modelId, taskType, outcome, 0, 0);
+  }
+  /** Returns all stats keyed by "modelId:taskType" — used by `cascade stats`. */
+  getAll() {
+    return new Map(this.stats);
+  }
   /**
    * Returns 0.05–1.0; defaults to 0.5 (neutral prior) when no history exists.
    * High retry counts penalise the score.
@@ -8701,6 +8728,18 @@ ${last.partialOutput}` : "");
     if (!prompt) return null;
     return this.run({ prompt });
   }
+  /**
+   * Record an explicit user rating for the last completed run.
+   * Explicit ratings carry 3× the weight of auto-detected outcomes so user
+   * feedback meaningfully shifts future routing decisions.
+   * Returns false when called before any task has run in this session.
+   */
+  rateLastRun(rating) {
+    if (!this.taskAnalyzer) return false;
+    const recorded = this.taskAnalyzer.recordExplicitRating(rating);
+    if (recorded) void this.perfTracker?.save();
+    return recorded;
+  }
   /**
    * Rough pre-execution cost estimate for a plan: ~3 T2 calls per section
    * plus ~4 T3 calls per subtask at typical token volumes. A ballpark for
@@ -10497,6 +10536,31 @@ var DashboardSocket = class {
         const { sessionId } = normalizeSessionSubscriptionPayload(payload);
         socket.leave(`session:${sessionId}`);
       });
+      socket.on("session:rate", (payload) => {
+        const sessionId = typeof payload?.sessionId === "string" ? payload.sessionId : "";
+        const rating = payload?.rating === "good" || payload?.rating === "bad" ? payload.rating : null;
+        if (sessionId && rating) {
+          this.io.emit("session:rate", { sessionId, rating });
+        }
+      });
+    });
+  }
+  onSessionRate(callback) {
+    this.io.on("connection", (socket) => {
+      socket.on("session:rate", (payload) => {
+        const sessionId = typeof payload?.sessionId === "string" ? payload.sessionId : "";
+        const rating = payload?.rating === "good" || payload?.rating === "bad" ? payload.rating : null;
+        if (sessionId && rating) callback(sessionId, rating);
+      });
+    });
+  }
+  onConfigUpdate(callback) {
+    this.io.on("connection", (socket) => {
+      socket.on("config:update", (payload) => {
+        if (typeof payload === "object" && payload !== null) {
+          callback(payload);
+        }
+      });
     });
   }
   close() {
@@ -10513,6 +10577,7 @@ var DashboardServer = class {
   store;
   globalStore = null;
   broadcastTimer = null;
+  activeSessions = /* @__PURE__ */ new Map();
   port;
   host;
   workspacePath;
@@ -10532,6 +10597,30 @@ var DashboardServer = class {
     });
     this.setupMiddleware();
     this.setupRoutes();
+    this.socket.onSessionRate((sessionId, rating) => {
+      this.activeSessions.get(sessionId)?.rateLastRun(rating);
+    });
+    this.socket.onConfigUpdate((data) => {
+      if (data.keys) {
+        for (const [type, apiKey] of Object.entries(data.keys)) {
+          if (!apiKey) continue;
+          const provider = this.config.providers.find((p) => p.type === type);
+          if (provider) provider.apiKey = apiKey;
+          else this.config.providers.push({ type, apiKey });
+        }
+      }
+      if (data.models) {
+        this.config.models = { ...this.config.models, ...data.models };
+      }
+      if (data.budget) {
+        if (typeof data.budget.maxCostPerRun === "number") {
+          this.config.budget.maxCostPerRunUsd = data.budget.maxCostPerRun;
+        }
+        if (data.budget.autoBias === "balanced" || data.budget.autoBias === "quality" || data.budget.autoBias === "cost") {
+          this.config.autoBias = data.budget.autoBias;
+        }
+      }
+    });
   }
   async start() {
     const isLoopback = this.host === "127.0.0.1" || this.host === "::1" || this.host === "localhost";
@@ -10985,6 +11074,7 @@ var DashboardServer = class {
       res.json({ sessionId, status: "ACTIVE" });
       void (async () => {
         const cascade = new Cascade(this.config, this.workspacePath, this.store);
+        this.activeSessions.set(sessionId, cascade);
         cascade.on("stream:token", (e) => {
           this.socket.broadcast("stream:token", { sessionId, tierId: e.tierId, text: e.text });
           this.socket.broadcastToRoom(`session:${sessionId}`, "stream:token", { sessionId, tierId: e.tierId, text: e.text });
@@ -11013,6 +11103,8 @@ var DashboardServer = class {
             sessionId,
             error: err instanceof Error ? err.message : String(err)
           });
+        } finally {
+          this.activeSessions.delete(sessionId);
         }
       })();
     });