cascade-ai 0.45.0 → 0.46.0

package/dist/cli.cjs

@@ -116,7 +116,7 @@ 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_CREDENTIALS_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 = "0.45.0";
+    CASCADE_VERSION = "0.46.0";
     CASCADE_CONFIG_FILE = ".cascade/config.json";
     CASCADE_DB_FILE = ".cascade/memory.db";
     CASCADE_DASHBOARD_SECRET_FILE = ".cascade/dashboard-secret";
@@ -3954,12 +3954,21 @@ var ModelSelector = class {
    * Absent for a provider ⇒ no validation ran ⇒ behave exactly as before.
    */
   validatedIds = /* @__PURE__ */ new Map();
+  /**
+   * Normalized ids that were added via discovery/registration but aren't in the
+   * bundled catalog — i.e. genuinely new models the provider reported (a Gemini
+   * flash released after the catalog was last hand-updated, an Ollama tag, etc.).
+   * getCandidatesForTier lets these compete in AUTO ranking so a newer,
+   * better-value model isn't invisible until someone edits constants.ts.
+   */
+  discovered = /* @__PURE__ */ new Set();
   constructor(availableProviders) {
     this.availableProviders = availableProviders;
     this.availableModels = new Map(Object.entries(MODELS));
   }
   addDynamicModel(model) {
     this.availableModels.set(model.id, model);
+    if (!(model.id in MODELS)) this.discovered.add(normalizeModelId(model.id));
   }
   /** Record the ids a provider's API actually serves (discovery). Empty ⇒ ignored. */
   setValidatedModels(provider, ids) {
@@ -4067,6 +4076,19 @@ var ModelSelector = class {
         candidates2.push(model);
       }
     }
+    if (this.discovered.size) {
+      const tierProviders = /* @__PURE__ */ new Set();
+      for (const key of priority) {
+        const m = MODELS[key];
+        if (m) tierProviders.add(m.provider);
+      }
+      const have = new Set(candidates2.map((c) => c.id));
+      for (const model of this.availableModels.values()) {
+        if (have.has(model.id)) continue;
+        if (!this.discovered.has(normalizeModelId(model.id))) continue;
+        if (tierProviders.has(model.provider) && this.isUsable(model)) candidates2.push(model);
+      }
+    }
     const localOnly = this.availableProviders.size > 0 && Array.from(this.availableProviders).every((p) => p === "ollama");
     if (localOnly) {
       for (const model of this.availableModels.values()) {
@@ -4778,6 +4800,14 @@ var FAMILY_MATCHERS = [
   [/gemini-?1\.5-pro/i, "gemini-1.5-pro"],
   [/gemini-?2\.0-flash-lite/i, "gemini-flash-lite"],
   [/gemini-?2\.0-flash/i, "gemini-2.0-flash"],
+  // Generic Gemini fallbacks — a model released after this table was written
+  // (e.g. a Gemini 3.x flash discovered live from the provider) still gets a
+  // sensible score by its class instead of a neutral 0.5, so it can compete in
+  // AUTO ranking. Ordered after the specific rows above; most-specific first.
+  [/gemini.*flash.?lite/i, "gemini-flash-lite"],
+  [/gemini.*pro/i, "gemini-2.5-pro"],
+  [/gemini.*flash/i, "gemini-2.5-flash"],
+  [/gemini/i, "gemini-2.5-flash"],
   [/codellama|code-llama|starcoder|stable-code/i, "codellama"],
   [/deepseek/i, "deepseek"],
   [/qwen/i, "qwen"],
@@ -4869,6 +4899,13 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
     costByFeature: {}
   };
   tierModels = /* @__PURE__ */ new Map();
+  /**
+   * Tiers the user pinned to a specific model in config (`models.tX` !== 'auto').
+   * An explicit pin always wins over Cascade Auto — selectModelForSubtask never
+   * re-selects a pinned tier per subtask. (tierModels is also filled with
+   * auto-picked defaults, so it alone can't distinguish "pinned" from "default".)
+   */
+  explicitTierModels = /* @__PURE__ */ new Set();
   config;
   sessionCostUsd = 0;
   // Per-run accounting for the hard per-task cap. Reset by beginRun() at the
@@ -4946,6 +4983,7 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
         );
       }
       this.tierModels.set(tier, model);
+      this.explicitTierModels.add(tier);
       this.ensureProvider(model, config.providers);
     }
     const azureRanked = [...azureDeployments].sort((a, b) => inferModelCapability(b.id) - inferModelCapability(a.id));
@@ -5386,6 +5424,7 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
    * Pure heuristic — no extra LLM call.
    */
   async selectModelForSubtask(tier, text, opts) {
+    if (this.explicitTierModels.has(tier)) return this.tierModels.get(tier) ?? null;
     if (!this.config?.cascadeAuto || !this.taskAnalyzer || !text.trim()) return null;
     try {
       return await this.taskAnalyzer.selectModel(text, tier, this.selector, opts);