cascade-ai 0.9.7 → 0.10.3

package/dist/cli.js

@@ -9574,6 +9574,19 @@ 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);
@@ -9653,6 +9666,20 @@ 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.
@@ -10257,6 +10284,18 @@ ${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
@@ -11116,6 +11155,12 @@ var SlashCommandRegistry = class {
       description: "Explain how the last run was routed (complexity, models, failovers)",
       handler: (_args, ctx) => ({ output: ctx.onWhy(), handled: true })
     });
+    this.register({
+      command: "/rate",
+      description: "Rate the last task to improve auto-routing  /rate good | bad",
+      args: ["good", "bad"],
+      handler: (args, ctx) => ({ output: ctx.onRate(args), handled: true })
+    });
     this.register({
       command: "/auto",
       description: "Toggle autonomous (hands-off) mode  /auto [on | off | status]",
@@ -12443,6 +12488,17 @@ ${msg.content}`).join("\n\n");
         return method === "osc52" ? `\u2714 Copied ${which} (${msg.content.length} chars) via terminal escape \u2014 works over SSH if your terminal supports OSC 52.` : `\u2714 Copied ${which} (${msg.content.length} chars) to clipboard.`;
       },
       onWhy: () => formatDecisionTrail(decisionLogRef.current),
+      onRate: (args) => {
+        const cascade = cascadeRef.current;
+        if (!cascade) return "Not ready yet.";
+        const rating = (args[0] ?? "").toLowerCase();
+        if (rating !== "good" && rating !== "bad") {
+          return "Usage: /rate good | bad";
+        }
+        const recorded = cascade.rateLastRun(rating);
+        if (!recorded) return "Nothing to rate \u2014 run a task first, or auto-routing is not enabled.";
+        return rating === "good" ? "\u2714 Rated good \u2014 models used for this task type got a boost." : "\u2714 Rated bad \u2014 models used for this task type were penalised. Auto-routing will try alternatives next time.";
+      },
       onComms: () => {
         dispatch({ type: "TOGGLE_COMMS" });
         return state.showComms ? "Agent comms feed hidden. /comms to bring it back." : "Agent comms feed enabled \u2014 agent-to-agent traffic will appear during runs.";
@@ -14457,6 +14513,31 @@ 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() {
@@ -14476,6 +14557,7 @@ var DashboardServer = class {
   store;
   globalStore = null;
   broadcastTimer = null;
+  activeSessions = /* @__PURE__ */ new Map();
   port;
   host;
   workspacePath;
@@ -14495,6 +14577,30 @@ 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";
@@ -14948,6 +15054,7 @@ 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 });
@@ -14976,6 +15083,8 @@ var DashboardServer = class {
             sessionId,
             error: err instanceof Error ? err.message : String(err)
           });
+        } finally {
+          this.activeSessions.delete(sessionId);
         }
       })();
     });
@@ -15427,6 +15536,44 @@ async function telemetryCommand(action) {
   }
   console.log();
 }
+var TASK_TYPES = ["code", "analysis", "creative", "data", "mixed"];
+async function statsCommand() {
+  const tracker = new ModelPerformanceTracker();
+  await tracker.load();
+  const all = tracker.getAll();
+  if (all.size === 0) {
+    console.log(chalk11.dim("\n  No routing history yet \u2014 run some tasks first.\n"));
+    return;
+  }
+  console.log(chalk11.magenta("\n  \u25C8 Auto-Routing History\n"));
+  console.log(chalk11.dim("  Per-task-type model performance learned from past runs.\n"));
+  for (const taskType of TASK_TYPES) {
+    const entries = [];
+    for (const [key, stat] of all) {
+      if (!key.endsWith(`:${taskType}`)) continue;
+      const modelId = key.slice(0, -(taskType.length + 1));
+      const successRate = stat.sampleCount > 0 ? stat.successCount / stat.sampleCount : 0;
+      const avgCostUsd = stat.sampleCount > 0 ? stat.totalCostUsd / stat.sampleCount : 0;
+      entries.push({ modelId, successRate, samples: stat.sampleCount, avgCostUsd });
+    }
+    if (entries.length === 0) continue;
+    entries.sort((a, b) => b.successRate - a.successRate || b.samples - a.samples);
+    console.log(chalk11.bold(`  ${taskType.toUpperCase()}`));
+    const header = `  ${"Model".padEnd(36)} ${"Success".padEnd(9)} ${"Samples".padEnd(9)} Avg cost`;
+    console.log(chalk11.dim(header));
+    console.log(chalk11.dim("  " + "\u2500".repeat(62)));
+    for (const e of entries) {
+      const pct = `${Math.round(e.successRate * 100)}%`;
+      const cost = e.avgCostUsd < 1e-4 ? "<$0.0001" : `$${e.avgCostUsd.toFixed(4)}`;
+      const color = e.successRate >= 0.8 ? chalk11.green : e.successRate >= 0.5 ? chalk11.yellow : chalk11.red;
+      console.log(
+        `  ${e.modelId.padEnd(36)} ${color(pct.padEnd(9))} ${String(e.samples).padEnd(9)} ${chalk11.dim(cost)}`
+      );
+    }
+    console.log();
+  }
+  console.log(chalk11.dim("  tip: use /rate good | bad after a task to improve these scores.\n"));
+}
 
 // src/cli/index.ts
 dotenv.config();
@@ -15526,6 +15673,9 @@ program.command("export").description("Export a session conversation to Markdown
     output: opts.output
   });
 });
+program.command("stats").description("Show auto-routing history: which models work best per task type").action(async () => {
+  await statsCommand();
+});
 async function startRepl(options) {
   const workspacePath = options.workspace ?? process.cwd();
   printBanner();