cascade-ai 0.53.0 → 0.54.0

package/dist/desktop-core.cjs

@@ -224247,7 +224247,7 @@ function reconcilePrice(model, live) {
 }
 
 // src/constants.ts
-var CASCADE_VERSION = "0.53.0";
+var CASCADE_VERSION = "0.54.0";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -251963,7 +251963,21 @@ var ModelSelector = class {
     this.availableProviders = availableProviders;
     this.availableModels = new Map(Object.entries(MODELS));
   }
+  /**
+   * Verdicts from earlier runs about ids that 404 on this account. Consulted
+   * BEFORE a model enters the pool, which is the only placement that helps: a
+   * T3 wave runs concurrently, so filtering after the first failure still lets
+   * every other worker in the wave call the dead id.
+   */
+  deadModels;
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    for (const [id, m3] of [...this.availableModels]) {
+      if (this.deadModels.isDead(m3.provider, m3.id)) this.availableModels.delete(id);
+    }
+  }
   addDynamicModel(model) {
+    if (this.deadModels?.isDead(model.provider, model.id)) return;
     this.availableModels.set(model.id, model);
     if (!(model.id in MODELS)) this.discovered.add(normalizeModelId(model.id));
   }
@@ -252370,6 +252384,121 @@ var LocalRequestQueue = class {
   }
 };
 
+// src/core/router/dead-models.ts
+var DEAD_MODEL_TTL_MS = 7 * 24 * 60 * 60 * 1e3;
+function deadModelKey(provider, modelId) {
+  return `${provider}:${modelId}`;
+}
+var DeadModelStore = class {
+  constructor(persistence, opts = {}) {
+    this.persistence = persistence;
+    this.ttlMs = opts.ttlMs ?? DEAD_MODEL_TTL_MS;
+    this.now = opts.now ?? (() => Date.now());
+    this.restore();
+  }
+  persistence;
+  records = /* @__PURE__ */ new Map();
+  ttlMs;
+  now;
+  restore() {
+    if (!this.persistence) return;
+    try {
+      for (const r4 of this.persistence.load()) {
+        if (!this.isExpired(r4)) this.records.set(r4.key, r4);
+      }
+    } catch {
+    }
+  }
+  isExpired(r4) {
+    return this.now() - r4.lastSeenAt > this.ttlMs;
+  }
+  flush() {
+    if (!this.persistence) return;
+    try {
+      this.persistence.save([...this.records.values()]);
+    } catch {
+    }
+  }
+  /**
+   * Record that a model was found dead. Returns true when this is the first
+   * time (so the caller can log it once rather than once per concurrent
+   * worker in the wave that discovered it).
+   */
+  record(provider, modelId, reason) {
+    const key = deadModelKey(provider, modelId);
+    const now = this.now();
+    const existing = this.records.get(key);
+    if (existing && !this.isExpired(existing)) {
+      existing.lastSeenAt = now;
+      existing.hits += 1;
+      this.flush();
+      return false;
+    }
+    this.records.set(key, {
+      key,
+      provider,
+      modelId,
+      reason,
+      firstSeenAt: now,
+      lastSeenAt: now,
+      hits: 1
+    });
+    this.flush();
+    return true;
+  }
+  /** True when this model is known dead and the verdict has not expired. */
+  isDead(provider, modelId) {
+    const r4 = this.records.get(deadModelKey(provider, modelId));
+    if (!r4) return false;
+    if (this.isExpired(r4)) {
+      this.records.delete(r4.key);
+      this.flush();
+      return false;
+    }
+    return true;
+  }
+  /** Why a model is considered dead, for /why and the models command. */
+  reasonFor(provider, modelId) {
+    const r4 = this.records.get(deadModelKey(provider, modelId));
+    return r4 && !this.isExpired(r4) ? r4.reason : null;
+  }
+  /** Live records, expired ones filtered out. */
+  list() {
+    const live = [...this.records.values()].filter((r4) => !this.isExpired(r4));
+    return live.sort((a2, b3) => b3.lastSeenAt - a2.lastSeenAt);
+  }
+  /**
+   * Forget a verdict — used when the user fixes the underlying problem (adds
+   * the model to their project, gets quota) and wants it retried now rather
+   * than after the TTL.
+   */
+  forget(provider, modelId) {
+    const removed = this.records.delete(deadModelKey(provider, modelId));
+    if (removed) this.flush();
+    return removed;
+  }
+  clear() {
+    this.records.clear();
+    this.flush();
+  }
+};
+function fileDeadModelPersistence(filePath) {
+  return {
+    load() {
+      const fs31 = __require("fs");
+      if (!fs31.existsSync(filePath)) return [];
+      const raw = JSON.parse(fs31.readFileSync(filePath, "utf-8"));
+      return Array.isArray(raw) ? raw : [];
+    },
+    save(records) {
+      const fs31 = __require("fs");
+      const path28 = __require("path");
+      fs31.mkdirSync(path28.dirname(filePath), { recursive: true });
+      fs31.writeFileSync(filePath, JSON.stringify(records, null, 2));
+    }
+  };
+}
+
 // src/utils/retry.ts
 var CascadeCancelledError = class extends Error {
   constructor(reason) {
@@ -252963,6 +253092,26 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
   localQueue;
   taskAnalyzer;
   pendingFeedbackSource;
+  /** Models known dead for this account (404 / retired / not enabled). */
+  deadModels = new DeadModelStore();
+  /**
+   * Back the dead-model memory with durable storage (a JSON file on desktop, a
+   * per-user table in the cloud). Without one the verdicts still work, but only
+   * for the life of the process — so tomorrow's run pays the same wave of 404s
+   * to learn the same fact.
+   */
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    this.selector.setDeadModelStore(store);
+  }
+  /** Models known dead for this account, for `cascade models` and /why. */
+  listDeadModels() {
+    return this.deadModels.list();
+  }
+  /** Retry a model the user has since fixed, without waiting out the TTL. */
+  forgetDeadModel(provider, modelId) {
+    return this.deadModels.forget(provider, modelId);
+  }
   worldStateDB;
   privacyPaths;
   guidanceQueue;
@@ -253327,6 +253476,8 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
       }
       if (isModelNotFoundError(errMsg)) {
         this.selector.removeModel(model.id);
+        const firstSighting = this.deadModels.record(model.provider, model.id, errMsg);
+        if (firstSighting) this.emit("model:dead", { provider: model.provider, modelId: model.id, reason: errMsg });
         const depth = (options._notFoundDepth ?? 0) + 1;
         const next = depth <= 3 ? this.selector.selectForTier(tier) : null;
         if (next && next.id !== model.id) {
@@ -257122,9 +257273,16 @@ var CAPABILITIES = [
     api: "openai-images"
   },
   {
+    // Imagen, not `gemini-3-pro-image`. That id appears on Google's PRICING
+    // page but 404s on `:predict` ("not found for API version v1beta, or is not
+    // supported for predict") — the Gemini image models are called through
+    // generateContent with an image response modality, while `:predict` is the
+    // Imagen shape this executor implements. A billing line item is not proof
+    // of a callable endpoint, which is the assumption that put the wrong id
+    // here in the first place.
     modality: "image",
     provider: "gemini",
-    modelId: "gemini-3-pro-image",
+    modelId: "imagen-4.0-generate-001",
     api: "gemini-predict"
   },
   // ── Speech (text → audio) ──
@@ -303000,6 +303158,13 @@ var Cascade = class _Cascade extends EventEmitter__default.default {
       }
     });
     this.toolRegistry = new ToolRegistry(this.config.tools, workspacePath);
+    try {
+      const nodePath = __require("path");
+      this.router.setDeadModelStore(new DeadModelStore(
+        fileDeadModelPersistence(nodePath.join(workspacePath, ".cascade", "dead-models.json"))
+      ));
+    } catch {
+    }
     this.multimodal = new MultimodalRegistry(
       (this.config.providers ?? []).map((p3) => p3.type)
     );

package/dist/index.cjs

@@ -87,6 +87,12 @@ var cron__default = /*#__PURE__*/_interopDefault(cron);
 // Cascade AI — Multi-tier AI Orchestration System
 var __defProp = Object.defineProperty;
 var __getOwnPropNames = Object.getOwnPropertyNames;
+var __require = /* @__PURE__ */ ((x) => typeof require !== "undefined" ? require : typeof Proxy !== "undefined" ? new Proxy(x, {
+  get: (a, b) => (typeof require !== "undefined" ? require : a)[b]
+}) : x)(function(x) {
+  if (typeof require !== "undefined") return require.apply(this, arguments);
+  throw Error('Dynamic require of "' + x + '" is not supported');
+});
 var __esm = (fn, res) => function __init() {
   return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
 };
@@ -1891,7 +1897,7 @@ function reconcilePrice(model, live) {
 }
 
 // src/constants.ts
-var CASCADE_VERSION = "0.53.0";
+var CASCADE_VERSION = "0.54.0";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -3704,7 +3710,21 @@ var ModelSelector = class {
     this.availableProviders = availableProviders;
     this.availableModels = new Map(Object.entries(MODELS));
   }
+  /**
+   * Verdicts from earlier runs about ids that 404 on this account. Consulted
+   * BEFORE a model enters the pool, which is the only placement that helps: a
+   * T3 wave runs concurrently, so filtering after the first failure still lets
+   * every other worker in the wave call the dead id.
+   */
+  deadModels;
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    for (const [id, m] of [...this.availableModels]) {
+      if (this.deadModels.isDead(m.provider, m.id)) this.availableModels.delete(id);
+    }
+  }
   addDynamicModel(model) {
+    if (this.deadModels?.isDead(model.provider, model.id)) return;
     this.availableModels.set(model.id, model);
     if (!(model.id in MODELS)) this.discovered.add(normalizeModelId(model.id));
   }
@@ -4111,6 +4131,121 @@ var LocalRequestQueue = class {
   }
 };
 
+// src/core/router/dead-models.ts
+var DEAD_MODEL_TTL_MS = 7 * 24 * 60 * 60 * 1e3;
+function deadModelKey(provider, modelId) {
+  return `${provider}:${modelId}`;
+}
+var DeadModelStore = class {
+  constructor(persistence, opts = {}) {
+    this.persistence = persistence;
+    this.ttlMs = opts.ttlMs ?? DEAD_MODEL_TTL_MS;
+    this.now = opts.now ?? (() => Date.now());
+    this.restore();
+  }
+  persistence;
+  records = /* @__PURE__ */ new Map();
+  ttlMs;
+  now;
+  restore() {
+    if (!this.persistence) return;
+    try {
+      for (const r of this.persistence.load()) {
+        if (!this.isExpired(r)) this.records.set(r.key, r);
+      }
+    } catch {
+    }
+  }
+  isExpired(r) {
+    return this.now() - r.lastSeenAt > this.ttlMs;
+  }
+  flush() {
+    if (!this.persistence) return;
+    try {
+      this.persistence.save([...this.records.values()]);
+    } catch {
+    }
+  }
+  /**
+   * Record that a model was found dead. Returns true when this is the first
+   * time (so the caller can log it once rather than once per concurrent
+   * worker in the wave that discovered it).
+   */
+  record(provider, modelId, reason) {
+    const key = deadModelKey(provider, modelId);
+    const now = this.now();
+    const existing = this.records.get(key);
+    if (existing && !this.isExpired(existing)) {
+      existing.lastSeenAt = now;
+      existing.hits += 1;
+      this.flush();
+      return false;
+    }
+    this.records.set(key, {
+      key,
+      provider,
+      modelId,
+      reason,
+      firstSeenAt: now,
+      lastSeenAt: now,
+      hits: 1
+    });
+    this.flush();
+    return true;
+  }
+  /** True when this model is known dead and the verdict has not expired. */
+  isDead(provider, modelId) {
+    const r = this.records.get(deadModelKey(provider, modelId));
+    if (!r) return false;
+    if (this.isExpired(r)) {
+      this.records.delete(r.key);
+      this.flush();
+      return false;
+    }
+    return true;
+  }
+  /** Why a model is considered dead, for /why and the models command. */
+  reasonFor(provider, modelId) {
+    const r = this.records.get(deadModelKey(provider, modelId));
+    return r && !this.isExpired(r) ? r.reason : null;
+  }
+  /** Live records, expired ones filtered out. */
+  list() {
+    const live = [...this.records.values()].filter((r) => !this.isExpired(r));
+    return live.sort((a, b) => b.lastSeenAt - a.lastSeenAt);
+  }
+  /**
+   * Forget a verdict — used when the user fixes the underlying problem (adds
+   * the model to their project, gets quota) and wants it retried now rather
+   * than after the TTL.
+   */
+  forget(provider, modelId) {
+    const removed = this.records.delete(deadModelKey(provider, modelId));
+    if (removed) this.flush();
+    return removed;
+  }
+  clear() {
+    this.records.clear();
+    this.flush();
+  }
+};
+function fileDeadModelPersistence(filePath) {
+  return {
+    load() {
+      const fs25 = __require("fs");
+      if (!fs25.existsSync(filePath)) return [];
+      const raw = JSON.parse(fs25.readFileSync(filePath, "utf-8"));
+      return Array.isArray(raw) ? raw : [];
+    },
+    save(records) {
+      const fs25 = __require("fs");
+      const path26 = __require("path");
+      fs25.mkdirSync(path26.dirname(filePath), { recursive: true });
+      fs25.writeFileSync(filePath, JSON.stringify(records, null, 2));
+    }
+  };
+}
+
 // src/utils/retry.ts
 var CascadeCancelledError = class extends Error {
   constructor(reason) {
@@ -4704,6 +4839,26 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
   localQueue;
   taskAnalyzer;
   pendingFeedbackSource;
+  /** Models known dead for this account (404 / retired / not enabled). */
+  deadModels = new DeadModelStore();
+  /**
+   * Back the dead-model memory with durable storage (a JSON file on desktop, a
+   * per-user table in the cloud). Without one the verdicts still work, but only
+   * for the life of the process — so tomorrow's run pays the same wave of 404s
+   * to learn the same fact.
+   */
+  setDeadModelStore(store) {
+    this.deadModels = store;
+    this.selector.setDeadModelStore(store);
+  }
+  /** Models known dead for this account, for `cascade models` and /why. */
+  listDeadModels() {
+    return this.deadModels.list();
+  }
+  /** Retry a model the user has since fixed, without waiting out the TTL. */
+  forgetDeadModel(provider, modelId) {
+    return this.deadModels.forget(provider, modelId);
+  }
   worldStateDB;
   privacyPaths;
   guidanceQueue;
@@ -5068,6 +5223,8 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
       }
       if (isModelNotFoundError(errMsg)) {
         this.selector.removeModel(model.id);
+        const firstSighting = this.deadModels.record(model.provider, model.id, errMsg);
+        if (firstSighting) this.emit("model:dead", { provider: model.provider, modelId: model.id, reason: errMsg });
         const depth = (options._notFoundDepth ?? 0) + 1;
         const next = depth <= 3 ? this.selector.selectForTier(tier) : null;
         if (next && next.id !== model.id) {
@@ -8863,9 +9020,16 @@ var CAPABILITIES = [
     api: "openai-images"
   },
   {
+    // Imagen, not `gemini-3-pro-image`. That id appears on Google's PRICING
+    // page but 404s on `:predict` ("not found for API version v1beta, or is not
+    // supported for predict") — the Gemini image models are called through
+    // generateContent with an image response modality, while `:predict` is the
+    // Imagen shape this executor implements. A billing line item is not proof
+    // of a callable endpoint, which is the assumption that put the wrong id
+    // here in the first place.
     modality: "image",
     provider: "gemini",
-    modelId: "gemini-3-pro-image",
+    modelId: "imagen-4.0-generate-001",
     api: "gemini-predict"
   },
   // ── Speech (text → audio) ──
@@ -14437,6 +14601,13 @@ var Cascade = class _Cascade extends EventEmitter__default.default {
       }
     });
     this.toolRegistry = new ToolRegistry(this.config.tools, workspacePath);
+    try {
+      const nodePath = __require("path");
+      this.router.setDeadModelStore(new DeadModelStore(
+        fileDeadModelPersistence(nodePath.join(workspacePath, ".cascade", "dead-models.json"))
+      ));
+    } catch {
+    }
     this.multimodal = new MultimodalRegistry(
       (this.config.providers ?? []).map((p) => p.type)
     );