cascade-ai 0.52.0 → 0.54.0

package/dist/index.d.cts

@@ -904,6 +904,16 @@ declare class ModelSelector {
      */
     private discovered;
     constructor(availableProviders: Set<ProviderType>);
+    /**
+     * 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.
+     */
+    private deadModels?;
+    setDeadModelStore(store: {
+        isDead(provider: string, modelId: string): boolean;
+    }): void;
     addDynamicModel(model: ModelInfo): void;
     /** Record the ids a provider's API actually serves (discovery). Empty ⇒ ignored. */
     setValidatedModels(provider: ProviderType, ids: string[]): void;
@@ -1077,6 +1087,60 @@ declare class TaskAnalyzer {
     static clearCache(): void;
 }
 
+interface DeadModelRecord {
+    /** `provider:modelId` — the same key shape the failover event reports. */
+    key: string;
+    provider: string;
+    modelId: string;
+    /** Provider's own words, kept verbatim so the user can act on them. */
+    reason: string;
+    firstSeenAt: number;
+    lastSeenAt: number;
+    /** How many times this model has been found dead. */
+    hits: number;
+}
+/**
+ * Where records live between runs. Desktop/CLI back this with a JSON file,
+ * the cloud server with a per-user table. Both are synchronous-optional: a
+ * store that fails to load must never stop a run.
+ */
+interface DeadModelPersistence {
+    load(): DeadModelRecord[];
+    save(records: DeadModelRecord[]): void;
+}
+declare class DeadModelStore {
+    private persistence?;
+    private records;
+    private readonly ttlMs;
+    private readonly now;
+    constructor(persistence?: DeadModelPersistence | undefined, opts?: {
+        ttlMs?: number;
+        now?: () => number;
+    });
+    private restore;
+    private isExpired;
+    private flush;
+    /**
+     * 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: string, modelId: string, reason: string): boolean;
+    /** True when this model is known dead and the verdict has not expired. */
+    isDead(provider: string, modelId: string): boolean;
+    /** Why a model is considered dead, for /why and the models command. */
+    reasonFor(provider: string, modelId: string): string | null;
+    /** Live records, expired ones filtered out. */
+    list(): DeadModelRecord[];
+    /**
+     * 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: string, modelId: string): boolean;
+    clear(): void;
+}
+
 declare class MemoryStore {
     private db;
     constructor(dbPath: string);
@@ -1509,6 +1573,19 @@ declare class CascadeRouter extends EventEmitter {
     private localQueue;
     private taskAnalyzer?;
     private pendingFeedbackSource?;
+    /** Models known dead for this account (404 / retired / not enabled). */
+    private deadModels;
+    /**
+     * 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: DeadModelStore): void;
+    /** Models known dead for this account, for `cascade models` and /why. */
+    listDeadModels(): ReturnType<DeadModelStore['list']>;
+    /** Retry a model the user has since fixed, without waiting out the TTL. */
+    forgetDeadModel(provider: string, modelId: string): boolean;
     private worldStateDB?;
     private privacyPaths?;
     private guidanceQueue?;
@@ -3797,15 +3874,8 @@ declare class CascadeToolError extends Error {
 declare function preferIpv4Host(url: string | undefined): string | undefined;
 declare function nodeHttpFetch(input: string | URL | Request, init?: RequestInit, redirectCount?: number): Promise<Response>;
 
-/**
- * Best-effort guess of the canonical base model an Azure deployment backs, from
- * its (arbitrary) deployment name. Ordered most-specific → least so "gpt-5-mini"
- * doesn't match the "gpt-5" base. Distinct point releases (gpt-5.5, gpt-5.4,
- * gpt-5.4-mini) resolve to their OWN base so their real economics + benchmark
- * scores apply — only unrecognised gpt-5.x fold into the gpt-5 base.
- * Returns null when the name gives no signal (e.g. "prod-fast") — the caller
- * then keeps neutral defaults, or the user can set an explicit base model.
- */
+/** Selectable base models, in the order a picker should show them. */
+declare const AZURE_BASE_MODELS: readonly string[];
 declare function inferAzureBaseModel(deploymentName: string): string | null;
 /**
  * The ModelInfo for one configured Azure deployment. On Azure the deployment IS
@@ -4004,4 +4074,4 @@ declare function gatherSyncBundle(config: CascadeConfig): SyncBundle;
 /** Merge a pulled bundle into a config, returning the updated config to persist. */
 declare function applySyncBundle(bundle: SyncBundle, config: CascadeConfig): CascadeConfig;
 
-export { AZURE_BASE_URL_TEMPLATE, type ApprovalRequest, type ApprovalResponse, type AuditEntry, AuditLogger, type BenchmarksConfig, type BudgetConfig, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, type CascadeConfig, type CascadeEvent, type CascadeEventType, CascadeIgnore, type CascadeMessage, CascadeRouter, type CascadeRunOptions, type CascadeRunResult, type CascadeThemeName, CascadeToolError, type Chunk, type ChunkOptions, CloudClient, type CloudConversation, type CloudMessage, type CloudSession, type CloudSessionStore, type CloudTurnInput, type CloudUser, type CodeChunkOptions, type CodeChunker, CodeSearchTool, type CompleteFn, ConfigManager, type ConversationMessage, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, type DashboardConfig, DashboardServer, type DeviceStart, type DistilledFact, type Embedder, type EncryptedBlob, type EscalationPayload, type ExtendedContextConfig, type FileManifest, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, type GenerateOptions, type GenerateResult, type GraphFactSource, GraphRetriever, type GraphSearchOptions, GraphSearchTool, type HookDefinition, type HooksConfig, HooksRunner, type Identity, type ImageAttachment, Keystore, LLMReranker, LM_STUDIO_BASE_URL, type LegacyThemeName, MODELS, type ManifestDiff, McpClient, McpOAuthProvider, type McpOAuthState, type McpOAuthStore, type McpServerConfig$1 as McpServerConfig, type MemoryConfig, MemoryStore, type Message, type MessageContent, type MessagePayload, type MessageStatus, type MessageType, type ModelInfo, type ModelOverrides, type NativeProvider, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, type OpenAIEmbedderOptions, PROVIDER_DISPLAY_NAMES, type PeerMessage, type PeerMessageEvent, type PeerSyncPayload, type PeerSyncType, type PermissionDecision, type PermissionDecisionPayload, type PermissionRequest, type PermissionTrailEntry, type PlanReviewConfig, type ProviderConfig, type ProviderType, type RankedFact, type RefreshResult, type Reranker, type RetrievalContext, type RetrievalMode, type RetrievalPlan, Retriever, type RetrieverSearchOptions, type RuntimeNode, type RuntimeNodeLog, type RuntimeRefreshPayload, type RuntimeScope, type RuntimeSession, type RuntimeSnapshotPayload, type ScheduledTask, type ScoredChunk, type SearchOptions, type Session, type SessionCheckpoint, type SessionMetadata, type SessionSubscriptionPayload, SqliteVectorStore, type StatusUpdate, type StoredMessage, type StreamChunk, type SyncBundle, T1Administrator, type T1ToT2Assignment, T1_MODEL_PRIORITY, T2Manager, type T2Result, type T2ToT3Assignment, T2_MODEL_PRIORITY, type T3Result, type T3ResultPayload, type T3SubtaskSpec, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, type TaskComplexity, TaskScheduler, Telemetry, type TelemetryConfig, type Theme, type ThemeColors, type ThemeName, type TierConfig, type TierLimits, type TierRole, type TierStatus, type TokenUsage, type ToolCall, type ToolDefinition, type ToolExecuteOptions, ToolRegistry, type ToolResult, type ToolsConfig, VISION_MODEL_PRIORITY, type VectorStore, type WebSearchConfig, type WebhookConfig, type WorkspaceConfig, WorkspaceIndex, type WorkspaceIndexOptions, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };
+export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, type ApprovalRequest, type ApprovalResponse, type AuditEntry, AuditLogger, type BenchmarksConfig, type BudgetConfig, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, type CascadeConfig, type CascadeEvent, type CascadeEventType, CascadeIgnore, type CascadeMessage, CascadeRouter, type CascadeRunOptions, type CascadeRunResult, type CascadeThemeName, CascadeToolError, type Chunk, type ChunkOptions, CloudClient, type CloudConversation, type CloudMessage, type CloudSession, type CloudSessionStore, type CloudTurnInput, type CloudUser, type CodeChunkOptions, type CodeChunker, CodeSearchTool, type CompleteFn, ConfigManager, type ConversationMessage, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, type DashboardConfig, DashboardServer, type DeviceStart, type DistilledFact, type Embedder, type EncryptedBlob, type EscalationPayload, type ExtendedContextConfig, type FileManifest, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, type GenerateOptions, type GenerateResult, type GraphFactSource, GraphRetriever, type GraphSearchOptions, GraphSearchTool, type HookDefinition, type HooksConfig, HooksRunner, type Identity, type ImageAttachment, Keystore, LLMReranker, LM_STUDIO_BASE_URL, type LegacyThemeName, MODELS, type ManifestDiff, McpClient, McpOAuthProvider, type McpOAuthState, type McpOAuthStore, type McpServerConfig$1 as McpServerConfig, type MemoryConfig, MemoryStore, type Message, type MessageContent, type MessagePayload, type MessageStatus, type MessageType, type ModelInfo, type ModelOverrides, type NativeProvider, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, type OpenAIEmbedderOptions, PROVIDER_DISPLAY_NAMES, type PeerMessage, type PeerMessageEvent, type PeerSyncPayload, type PeerSyncType, type PermissionDecision, type PermissionDecisionPayload, type PermissionRequest, type PermissionTrailEntry, type PlanReviewConfig, type ProviderConfig, type ProviderType, type RankedFact, type RefreshResult, type Reranker, type RetrievalContext, type RetrievalMode, type RetrievalPlan, Retriever, type RetrieverSearchOptions, type RuntimeNode, type RuntimeNodeLog, type RuntimeRefreshPayload, type RuntimeScope, type RuntimeSession, type RuntimeSnapshotPayload, type ScheduledTask, type ScoredChunk, type SearchOptions, type Session, type SessionCheckpoint, type SessionMetadata, type SessionSubscriptionPayload, SqliteVectorStore, type StatusUpdate, type StoredMessage, type StreamChunk, type SyncBundle, T1Administrator, type T1ToT2Assignment, T1_MODEL_PRIORITY, T2Manager, type T2Result, type T2ToT3Assignment, T2_MODEL_PRIORITY, type T3Result, type T3ResultPayload, type T3SubtaskSpec, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, type TaskComplexity, TaskScheduler, Telemetry, type TelemetryConfig, type Theme, type ThemeColors, type ThemeName, type TierConfig, type TierLimits, type TierRole, type TierStatus, type TokenUsage, type ToolCall, type ToolDefinition, type ToolExecuteOptions, ToolRegistry, type ToolResult, type ToolsConfig, VISION_MODEL_PRIORITY, type VectorStore, type WebSearchConfig, type WebhookConfig, type WorkspaceConfig, WorkspaceIndex, type WorkspaceIndexOptions, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };

package/dist/index.d.ts

@@ -904,6 +904,16 @@ declare class ModelSelector {
      */
     private discovered;
     constructor(availableProviders: Set<ProviderType>);
+    /**
+     * 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.
+     */
+    private deadModels?;
+    setDeadModelStore(store: {
+        isDead(provider: string, modelId: string): boolean;
+    }): void;
     addDynamicModel(model: ModelInfo): void;
     /** Record the ids a provider's API actually serves (discovery). Empty ⇒ ignored. */
     setValidatedModels(provider: ProviderType, ids: string[]): void;
@@ -1077,6 +1087,60 @@ declare class TaskAnalyzer {
     static clearCache(): void;
 }
 
+interface DeadModelRecord {
+    /** `provider:modelId` — the same key shape the failover event reports. */
+    key: string;
+    provider: string;
+    modelId: string;
+    /** Provider's own words, kept verbatim so the user can act on them. */
+    reason: string;
+    firstSeenAt: number;
+    lastSeenAt: number;
+    /** How many times this model has been found dead. */
+    hits: number;
+}
+/**
+ * Where records live between runs. Desktop/CLI back this with a JSON file,
+ * the cloud server with a per-user table. Both are synchronous-optional: a
+ * store that fails to load must never stop a run.
+ */
+interface DeadModelPersistence {
+    load(): DeadModelRecord[];
+    save(records: DeadModelRecord[]): void;
+}
+declare class DeadModelStore {
+    private persistence?;
+    private records;
+    private readonly ttlMs;
+    private readonly now;
+    constructor(persistence?: DeadModelPersistence | undefined, opts?: {
+        ttlMs?: number;
+        now?: () => number;
+    });
+    private restore;
+    private isExpired;
+    private flush;
+    /**
+     * 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: string, modelId: string, reason: string): boolean;
+    /** True when this model is known dead and the verdict has not expired. */
+    isDead(provider: string, modelId: string): boolean;
+    /** Why a model is considered dead, for /why and the models command. */
+    reasonFor(provider: string, modelId: string): string | null;
+    /** Live records, expired ones filtered out. */
+    list(): DeadModelRecord[];
+    /**
+     * 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: string, modelId: string): boolean;
+    clear(): void;
+}
+
 declare class MemoryStore {
     private db;
     constructor(dbPath: string);
@@ -1509,6 +1573,19 @@ declare class CascadeRouter extends EventEmitter {
     private localQueue;
     private taskAnalyzer?;
     private pendingFeedbackSource?;
+    /** Models known dead for this account (404 / retired / not enabled). */
+    private deadModels;
+    /**
+     * 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: DeadModelStore): void;
+    /** Models known dead for this account, for `cascade models` and /why. */
+    listDeadModels(): ReturnType<DeadModelStore['list']>;
+    /** Retry a model the user has since fixed, without waiting out the TTL. */
+    forgetDeadModel(provider: string, modelId: string): boolean;
     private worldStateDB?;
     private privacyPaths?;
     private guidanceQueue?;
@@ -3797,15 +3874,8 @@ declare class CascadeToolError extends Error {
 declare function preferIpv4Host(url: string | undefined): string | undefined;
 declare function nodeHttpFetch(input: string | URL | Request, init?: RequestInit, redirectCount?: number): Promise<Response>;
 
-/**
- * Best-effort guess of the canonical base model an Azure deployment backs, from
- * its (arbitrary) deployment name. Ordered most-specific → least so "gpt-5-mini"
- * doesn't match the "gpt-5" base. Distinct point releases (gpt-5.5, gpt-5.4,
- * gpt-5.4-mini) resolve to their OWN base so their real economics + benchmark
- * scores apply — only unrecognised gpt-5.x fold into the gpt-5 base.
- * Returns null when the name gives no signal (e.g. "prod-fast") — the caller
- * then keeps neutral defaults, or the user can set an explicit base model.
- */
+/** Selectable base models, in the order a picker should show them. */
+declare const AZURE_BASE_MODELS: readonly string[];
 declare function inferAzureBaseModel(deploymentName: string): string | null;
 /**
  * The ModelInfo for one configured Azure deployment. On Azure the deployment IS
@@ -4004,4 +4074,4 @@ declare function gatherSyncBundle(config: CascadeConfig): SyncBundle;
 /** Merge a pulled bundle into a config, returning the updated config to persist. */
 declare function applySyncBundle(bundle: SyncBundle, config: CascadeConfig): CascadeConfig;
 
-export { AZURE_BASE_URL_TEMPLATE, type ApprovalRequest, type ApprovalResponse, type AuditEntry, AuditLogger, type BenchmarksConfig, type BudgetConfig, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, type CascadeConfig, type CascadeEvent, type CascadeEventType, CascadeIgnore, type CascadeMessage, CascadeRouter, type CascadeRunOptions, type CascadeRunResult, type CascadeThemeName, CascadeToolError, type Chunk, type ChunkOptions, CloudClient, type CloudConversation, type CloudMessage, type CloudSession, type CloudSessionStore, type CloudTurnInput, type CloudUser, type CodeChunkOptions, type CodeChunker, CodeSearchTool, type CompleteFn, ConfigManager, type ConversationMessage, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, type DashboardConfig, DashboardServer, type DeviceStart, type DistilledFact, type Embedder, type EncryptedBlob, type EscalationPayload, type ExtendedContextConfig, type FileManifest, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, type GenerateOptions, type GenerateResult, type GraphFactSource, GraphRetriever, type GraphSearchOptions, GraphSearchTool, type HookDefinition, type HooksConfig, HooksRunner, type Identity, type ImageAttachment, Keystore, LLMReranker, LM_STUDIO_BASE_URL, type LegacyThemeName, MODELS, type ManifestDiff, McpClient, McpOAuthProvider, type McpOAuthState, type McpOAuthStore, type McpServerConfig$1 as McpServerConfig, type MemoryConfig, MemoryStore, type Message, type MessageContent, type MessagePayload, type MessageStatus, type MessageType, type ModelInfo, type ModelOverrides, type NativeProvider, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, type OpenAIEmbedderOptions, PROVIDER_DISPLAY_NAMES, type PeerMessage, type PeerMessageEvent, type PeerSyncPayload, type PeerSyncType, type PermissionDecision, type PermissionDecisionPayload, type PermissionRequest, type PermissionTrailEntry, type PlanReviewConfig, type ProviderConfig, type ProviderType, type RankedFact, type RefreshResult, type Reranker, type RetrievalContext, type RetrievalMode, type RetrievalPlan, Retriever, type RetrieverSearchOptions, type RuntimeNode, type RuntimeNodeLog, type RuntimeRefreshPayload, type RuntimeScope, type RuntimeSession, type RuntimeSnapshotPayload, type ScheduledTask, type ScoredChunk, type SearchOptions, type Session, type SessionCheckpoint, type SessionMetadata, type SessionSubscriptionPayload, SqliteVectorStore, type StatusUpdate, type StoredMessage, type StreamChunk, type SyncBundle, T1Administrator, type T1ToT2Assignment, T1_MODEL_PRIORITY, T2Manager, type T2Result, type T2ToT3Assignment, T2_MODEL_PRIORITY, type T3Result, type T3ResultPayload, type T3SubtaskSpec, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, type TaskComplexity, TaskScheduler, Telemetry, type TelemetryConfig, type Theme, type ThemeColors, type ThemeName, type TierConfig, type TierLimits, type TierRole, type TierStatus, type TokenUsage, type ToolCall, type ToolDefinition, type ToolExecuteOptions, ToolRegistry, type ToolResult, type ToolsConfig, VISION_MODEL_PRIORITY, type VectorStore, type WebSearchConfig, type WebhookConfig, type WorkspaceConfig, WorkspaceIndex, type WorkspaceIndexOptions, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };
+export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, type ApprovalRequest, type ApprovalResponse, type AuditEntry, AuditLogger, type BenchmarksConfig, type BudgetConfig, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, type CascadeConfig, type CascadeEvent, type CascadeEventType, CascadeIgnore, type CascadeMessage, CascadeRouter, type CascadeRunOptions, type CascadeRunResult, type CascadeThemeName, CascadeToolError, type Chunk, type ChunkOptions, CloudClient, type CloudConversation, type CloudMessage, type CloudSession, type CloudSessionStore, type CloudTurnInput, type CloudUser, type CodeChunkOptions, type CodeChunker, CodeSearchTool, type CompleteFn, ConfigManager, type ConversationMessage, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, type DashboardConfig, DashboardServer, type DeviceStart, type DistilledFact, type Embedder, type EncryptedBlob, type EscalationPayload, type ExtendedContextConfig, type FileManifest, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, type GenerateOptions, type GenerateResult, type GraphFactSource, GraphRetriever, type GraphSearchOptions, GraphSearchTool, type HookDefinition, type HooksConfig, HooksRunner, type Identity, type ImageAttachment, Keystore, LLMReranker, LM_STUDIO_BASE_URL, type LegacyThemeName, MODELS, type ManifestDiff, McpClient, McpOAuthProvider, type McpOAuthState, type McpOAuthStore, type McpServerConfig$1 as McpServerConfig, type MemoryConfig, MemoryStore, type Message, type MessageContent, type MessagePayload, type MessageStatus, type MessageType, type ModelInfo, type ModelOverrides, type NativeProvider, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, type OpenAIEmbedderOptions, PROVIDER_DISPLAY_NAMES, type PeerMessage, type PeerMessageEvent, type PeerSyncPayload, type PeerSyncType, type PermissionDecision, type PermissionDecisionPayload, type PermissionRequest, type PermissionTrailEntry, type PlanReviewConfig, type ProviderConfig, type ProviderType, type RankedFact, type RefreshResult, type Reranker, type RetrievalContext, type RetrievalMode, type RetrievalPlan, Retriever, type RetrieverSearchOptions, type RuntimeNode, type RuntimeNodeLog, type RuntimeRefreshPayload, type RuntimeScope, type RuntimeSession, type RuntimeSnapshotPayload, type ScheduledTask, type ScoredChunk, type SearchOptions, type Session, type SessionCheckpoint, type SessionMetadata, type SessionSubscriptionPayload, SqliteVectorStore, type StatusUpdate, type StoredMessage, type StreamChunk, type SyncBundle, T1Administrator, type T1ToT2Assignment, T1_MODEL_PRIORITY, T2Manager, type T2Result, type T2ToT3Assignment, T2_MODEL_PRIORITY, type T3Result, type T3ResultPayload, type T3SubtaskSpec, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, type TaskComplexity, TaskScheduler, Telemetry, type TelemetryConfig, type Theme, type ThemeColors, type ThemeName, type TierConfig, type TierLimits, type TierRole, type TierStatus, type TokenUsage, type ToolCall, type ToolDefinition, type ToolExecuteOptions, ToolRegistry, type ToolResult, type ToolsConfig, VISION_MODEL_PRIORITY, type VectorStore, type WebSearchConfig, type WebhookConfig, type WorkspaceConfig, WorkspaceIndex, type WorkspaceIndexOptions, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };

package/dist/index.js

@@ -41,6 +41,12 @@ import cron from 'node-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;
 };
@@ -1845,7 +1851,7 @@ function reconcilePrice(model, live) {
 }
 
 // src/constants.ts
-var CASCADE_VERSION = "0.52.0";
+var CASCADE_VERSION = "0.54.0";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -2810,22 +2816,23 @@ var OpenAIProvider = class extends BaseProvider {
 
 // src/providers/azure.ts
 var DEFAULT_AZURE_API_VERSION = "2024-12-01-preview";
+var AZURE_BASE_MODEL_RULES = [
+  [/gpt-?5\.5/, "gpt-5.5"],
+  [/gpt-?5\.4.*mini/, "gpt-5.4-mini"],
+  [/gpt-?5\.4/, "gpt-5.4"],
+  [/gpt-?5.*nano/, "gpt-5-nano"],
+  [/gpt-?5.*mini/, "gpt-5-mini"],
+  [/gpt-?5/, "gpt-5"],
+  [/gpt-?4\.1-nano/, "gpt-4.1-nano"],
+  [/gpt-?4\.1-mini/, "gpt-4.1-mini"],
+  [/gpt-?4\.1/, "gpt-4.1"],
+  [/gpt-?4o-mini/, "gpt-4o-mini"],
+  [/gpt-?4o/, "gpt-4o"]
+];
+var AZURE_BASE_MODELS = AZURE_BASE_MODEL_RULES.map(([, id]) => id);
 function inferAzureBaseModel(deploymentName) {
   const n = deploymentName.toLowerCase();
-  const rules = [
-    [/gpt-?5\.5/, "gpt-5.5"],
-    [/gpt-?5\.4.*mini/, "gpt-5.4-mini"],
-    [/gpt-?5\.4/, "gpt-5.4"],
-    [/gpt-?5.*nano/, "gpt-5-nano"],
-    [/gpt-?5.*mini/, "gpt-5-mini"],
-    [/gpt-?5/, "gpt-5"],
-    [/gpt-?4\.1-nano/, "gpt-4.1-nano"],
-    [/gpt-?4\.1-mini/, "gpt-4.1-mini"],
-    [/gpt-?4\.1/, "gpt-4.1"],
-    [/gpt-?4o-mini/, "gpt-4o-mini"],
-    [/gpt-?4o/, "gpt-4o"]
-  ];
-  for (const [re, base] of rules) if (re.test(n)) return base;
+  for (const [re, base] of AZURE_BASE_MODEL_RULES) if (re.test(n)) return base;
   return null;
 }
 function azureModelForDeployment(cfg) {
@@ -3657,7 +3664,21 @@ 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));
   }
@@ -4064,6 +4085,121 @@ 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) {
@@ -4657,6 +4793,26 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
   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;
@@ -5021,6 +5177,8 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
       }
       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) {
@@ -8816,9 +8974,16 @@ 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) ──
@@ -10947,6 +11112,20 @@ ${text}`;
   }
 };
 
+// src/tools/tool-name.ts
+var MCP_TOOL_PREFIX = "mcp__";
+function sanitizeToolNameSegment(segment) {
+  const cleaned = segment.replace(/[^a-zA-Z0-9-]+/g, "_");
+  let start = 0;
+  let end = cleaned.length;
+  while (start < end && cleaned[start] === "_") start++;
+  while (end > start && cleaned[end - 1] === "_") end--;
+  return cleaned.slice(start, end) || "_";
+}
+function mcpToolName(serverName, toolName) {
+  return `${MCP_TOOL_PREFIX}${sanitizeToolNameSegment(serverName)}__${sanitizeToolNameSegment(toolName)}`;
+}
+
 // src/tools/mcp.ts
 var McpToolWrapper = class extends BaseTool {
   name;
@@ -10960,7 +11139,7 @@ var McpToolWrapper = class extends BaseTool {
     this.mcpClient = mcpClient;
     this.serverName = serverName;
     this.toolName = toolName;
-    this.name = `mcp::${serverName}::${toolName}`;
+    this.name = mcpToolName(serverName, toolName);
     this.description = `[MCP:${serverName}] ${description}`;
     this.inputSchema = inputSchema;
   }
@@ -14376,6 +14555,13 @@ var Cascade = class _Cascade extends EventEmitter {
       }
     });
     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)
     );
@@ -19093,6 +19279,6 @@ function applySyncBundle(bundle, config) {
   return next;
 }
 
-export { AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DashboardServer, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };
+export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DashboardServer, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildManifest, buildSessionTranscript, cagCharBudget, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, discoverMcpAuthServer, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, nodeHttpFetch, parseRankOrder, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, runCascade, sessionWorthRemembering, streamCascade };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map