@linxiraos/pi-catalog 1.0.9 → 1.1.0

package/dist/types/discovery/gemini-cli.d.ts

@@ -0,0 +1,33 @@
+import type { ModelSpec } from "../types.js";
+import { type VariantCollapseTable } from "../variant-collapse.js";
+/**
+ * Options for the Gemini CLI quota-based discovery fallback.
+ */
+export interface FetchGeminiCliQuotaModelsOptions {
+    /** OAuth access token sent as `Authorization: Bearer <token>`. */
+    token: string;
+    /** Cloud Code Assist endpoint. Defaults to `https://cloudcode-pa.googleapis.com`. */
+    endpoint?: string;
+    /** Pre-resolved GCP project id; otherwise discovered via `loadCodeAssist`. */
+    projectId?: string;
+    /** Optional abort signal for request cancellation. */
+    signal?: AbortSignal;
+    /** Optional fetch implementation override for tests. */
+    fetcher?: typeof fetch;
+    /** Effort-tier collapse table applied to the discovered list. */
+    collapseTable?: VariantCollapseTable;
+}
+/**
+ * Discovers the Gemini models available to a `google-gemini-cli` credential via
+ * the account's own `retrieveUserQuota` endpoint on Cloud Code Assist.
+ *
+ * This is the fallback for accounts whose credential is not authorized for the
+ * Antigravity `fetchAvailableModels` endpoint (e.g. Gemini Code Assist Standard
+ * tiers, which return HTTP 403 there). Quota buckets carry only model ids, so
+ * metadata is filled from the bundled catalog where the id is known and
+ * synthesized with Gemini CLI defaults otherwise.
+ *
+ * Returns `null` on network/payload/auth failure (the caller keeps the bundled
+ * catalog). Returns `[]` when the quota response lists no usable Gemini models.
+ */
+export declare function fetchGeminiCliQuotaModels(options: FetchGeminiCliQuotaModelsOptions): Promise<ModelSpec<"google-gemini-cli">[] | null>;

package/dist/types/discovery/index.d.ts

@@ -1,5 +1,7 @@
 export * from "./antigravity.js";
 export * from "./codex.js";
 export * from "./gemini.js";
+export * from "./gemini-cli.js";
 export * from "./gitlab-duo-workflow.js";
 export * from "./openai-compatible.js";
+export * from "./protobuf.js";

package/dist/types/discovery/protobuf.d.ts

@@ -0,0 +1,102 @@
+/**
+ * High-performance, zero-builder protobuf wire codecs for @linxiraos/pi-catalog.
+ *
+ * Schemas are declared as static IR descriptors with near-zero module load overhead
+ * and lazy compilation on first encode/decode/create invocation.
+ */
+/** JSON values carried by `google.protobuf.Value` fields. */
+export type JsonValue = null | boolean | number | string | JsonValue[] | {
+    [key: string]: JsonValue;
+};
+/** An unrecognised wire field retained for forward-compatible round-trips. */
+export interface ProtoUnknownField {
+    no: number;
+    wireType: number;
+    data: Uint8Array;
+}
+/** Shared internal metadata present on every decoded protocol message. */
+export interface ProtoMessage {
+    $typeName?: string;
+    $unknown?: ProtoUnknownField[];
+}
+/** A bidirectional codec for one protobuf message type. */
+export interface MessageCodec<T extends ProtoMessage = ProtoMessage> {
+    (value: T): Uint8Array;
+    (value: Uint8Array): T;
+    /** Creates a message with protobuf defaults for omitted fields. */
+    create(value?: Partial<T>): T;
+    /** Encodes one message into protobuf wire bytes. */
+    encode(value: T): Uint8Array;
+    /** Decodes one protobuf message from wire bytes. */
+    decode(value: Uint8Array): T;
+    /** Converts a message to its protobuf JSON representation. */
+    toJson(value: T): JsonValue;
+}
+/** Infers a message shape from a codec result. */
+export type InferMessage<TCodec> = TCodec extends MessageCodec<infer TMessage> ? TMessage : never;
+/** Erases a referenced message's concrete shape for static field descriptors. */
+export interface MessageReference {
+    encode(value: unknown): Uint8Array;
+    decode(value: Uint8Array): ProtoMessage;
+    toJson(value: unknown): JsonValue;
+}
+export type ScalarKind = "bool" | "bytes" | "double" | "enum" | "float" | "int32" | "int64" | "string" | "uint32" | "uint64";
+export type WireType = 0 | 1 | 2 | 5;
+export interface ScalarFieldDesc {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: ScalarKind;
+    readonly optional?: boolean;
+    readonly repeat?: boolean;
+}
+export interface MessageFieldDesc {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: "message";
+    readonly T: () => MessageReference;
+    readonly repeat?: boolean;
+}
+export interface EnumFieldDesc {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: "enum";
+    readonly optional?: boolean;
+    readonly repeat?: boolean;
+}
+export interface MapFieldDesc {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: "map";
+    readonly K: "string";
+    readonly V: ScalarKind | (() => MessageReference);
+}
+export type VariantDesc = {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: ScalarKind;
+} | {
+    readonly no: number;
+    readonly name: string;
+    readonly kind: "message";
+    readonly T: () => MessageReference;
+};
+export interface OneofFieldDesc {
+    readonly kind: "oneof";
+    readonly name: string;
+    readonly variants: readonly VariantDesc[];
+}
+export type FieldDesc = ScalarFieldDesc | MessageFieldDesc | EnumFieldDesc | MapFieldDesc | OneofFieldDesc;
+/** Creates a high-performance, lazy protobuf message codec from an IR field descriptor list. */
+export declare function pb<T extends ProtoMessage = ProtoMessage>(typeName: string, fields?: readonly FieldDesc[]): MessageCodec<T>;
+/** Creates a message using its codec's protobuf defaults. */
+export declare function create<TMessage extends ProtoMessage>(codec: MessageCodec<TMessage>, value?: Partial<TMessage>): TMessage;
+/** Encodes a message using its codec. */
+export declare function toBinary<TMessage extends ProtoMessage>(codec: MessageCodec<TMessage>, value: TMessage): Uint8Array;
+/** Decodes wire bytes using a message codec. */
+export declare function fromBinary<TMessage extends ProtoMessage>(codec: MessageCodec<TMessage>, value: Uint8Array): TMessage;
+/** Converts a message to protobuf JSON using its codec. */
+export declare function toJson<TMessage extends ProtoMessage>(codec: MessageCodec<TMessage>, value: TMessage): JsonValue;
+/** Encodes a JSON value as `google.protobuf.Value`. */
+export declare function encodeJsonValue(value: JsonValue): Uint8Array;
+/** Decodes `google.protobuf.Value` wire bytes into a JSON value. */
+export declare function decodeJsonValue(value: Uint8Array): JsonValue;

package/dist/types/hosts.d.ts

@@ -111,6 +111,11 @@ export declare const KNOWN_HOSTS: {
         readonly providers: readonly ["nvidia"];
         readonly urlMarkers: readonly ["integrate.api.nvidia.com"];
     };
+    /** Venice AI (`api.venice.ai`). OpenAI-compatible; drives reasoning via top-level `reasoning_effort` (and `venice_parameters.disable_thinking`), and rejects DashScope's top-level `enable_thinking` with a 400 (`additionalProperties: false` request schema). */
+    readonly venice: {
+        readonly providers: readonly ["venice"];
+        readonly urlMarkers: readonly ["api.venice.ai"];
+    };
     readonly moonshotNative: {
         readonly providers: readonly ["moonshot", "kimi-code"];
         readonly urlMarkers: readonly ["api.moonshot.ai", "api.kimi.com"];

package/dist/types/identity/family.d.ts

@@ -30,6 +30,17 @@ export declare const isClaudeModelId: (modelId: string) => boolean;
 export declare const isAnthropicNamespacedModelId: (modelId: string) => boolean;
 /** Qwen family ids (substring match — Qwen SKUs have no stable prefix shape). */
 export declare const isQwenModelId: (modelId: string) => boolean;
+/**
+ * Open-weight Qwen 3.8+ releases (`qwen3.8-27b`, `qwen3.8-2.4t-a95b`, GGUF
+ * names like `Qwen3.8-27B-UD-Q6_K_XL`) whose chat template steers thinking
+ * depth through a `reasoning_effort` template kwarg (`low`/`medium`/`xhigh`,
+ * template default `xhigh`; thinking itself cannot be disabled). Compared
+ * component-wise so `qwen3.10` sorts after `qwen3.8`. API-only `-max` SKUs are
+ * excluded — Dashscope drives them through OpenAI-style `reasoning_effort`
+ * with curated compat. The trailing guard rejects parameter-count lookalikes
+ * (`qwen-3.8b`) without breaking `qwen3.8-27b`.
+ */
+export declare const isQwen38PlusTemplateEffortModelId: (modelId: string) => boolean;
 /** Gemma open-weights family (`gemma-3-27b-it`, `google/gemma-4-E2B-it`, `gemma2-9b`). */
 export declare const isGemmaModelId: (modelId: string) => boolean;
 /** DeepSeek family by id or display name (proxies often rename the id but keep the name). */
@@ -45,6 +56,8 @@ export declare const isDeepseekModelIdOrName: (modelId: string) => boolean;
 export declare const isDeepseekV4FlashModelId: (modelId: string) => boolean;
 /** Xiaomi MiMo family by id or display name. */
 export declare const isMimoModelIdOrName: (modelId: string) => boolean;
+/** StepFun Step 3.7 Flash SKU in any namespace form (`kilo/stepfun/step-3.7-flash:free`). */
+export declare const isStep37FlashModelId: (modelId: string) => boolean;
 /** Gemini family ids in any namespace form (`gemini-*`, `google/gemini-*`, `openrouter/google/gemini-…`). */
 export declare const isGeminiModelId: (modelId: string) => boolean;
 /** Grok family ids across namespace and delimiter forms (`grok-*`, `cursor-grok-*`, `xai/grok-*`). */
@@ -52,7 +65,8 @@ export declare const isGrokModelId: (modelId: string) => boolean;
 /**
  * Grok SKUs that expose the wire `reasoning.effort` dial. Other Grok reasoners
  * (e.g. `grok-build`, `grok-4.20-0309-reasoning`) think natively but reject the
- * param, so callers must omit reasoning effort for them.
+ * param, so callers must omit reasoning effort for them. `grok-4.6` accepts
+ * `low`/`medium`/`high`/`xhigh` and 400s on `max`.
  */
 export declare const isGrokReasoningEffortCapable: (modelId: string) => boolean;
 /**
@@ -88,6 +102,13 @@ export declare const isMinimaxM3FamilyModelId: (modelId: string) => boolean;
  * and `none`.
  */
 export declare const isOpenAIGptOssModelId: (modelId: string) => boolean;
+/**
+ * Meta Muse Spark ids (`muse-spark-1.1`, `muse-spark-1.2`,
+ * `muse-spark-1.2-contributor`, `meta/muse-spark-1.2`). The Responses
+ * `reasoning.effort` wire accepts `none` (thinking-off) plus
+ * `minimal`/`low`/`medium`/`high`/`xhigh`.
+ */
+export declare const isMuseSparkModelId: (modelId: string) => boolean;
 /** OpenAI model ids (gpt-*, chatgpt-*, o1/o3/o4 SKUs, codex-*, or openai/*). */
 export declare const isOpenAIModelId: (modelId: string) => boolean;
 /**
@@ -187,6 +208,14 @@ export declare const hasOpus47ApiRestrictions: (modelId: string) => boolean;
  * @see https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages
  */
 export declare const supportsMidConversationSystemMessages: (modelId: string) => boolean;
+/**
+ * Models that reliably follow the hashline line-anchored edit dialect
+ * (`[path#TAG]` headers plus 1-indexed anchors). Kimi, MiMo, DeepSeek V4
+ * Flash, and Step 3.7 Flash miscount anchors or drop the tag header often
+ * enough that hosts fall back to the sloppy edit format for
+ * them.
+ */
+export declare const supportsHashlineEdits: (modelId: string) => boolean;
 export declare const isAnthropicFableOrMythosModel: (modelId: string) => boolean;
 /** Thinking-variant token location inside a model id. */
 export interface ThinkingVariantToken {

package/dist/types/index.d.ts

@@ -6,6 +6,7 @@ export * from "./identity/index.js";
 export * from "./model-cache.js";
 export * from "./model-manager.js";
 export * from "./model-thinking.js";
+export * from "./model-tokenizer.js";
 export * from "./models.js";
 export * from "./provider-models/index.js";
 export * from "./types.js";
@@ -15,3 +16,4 @@ export * from "./wire/codex.js";
 export * from "./wire/coreweave.js";
 export * from "./wire/gemini-headers.js";
 export * from "./wire/github-copilot.js";
+export * from "./wire/image-fetchers.js";

package/dist/types/model-manager.d.ts

@@ -31,11 +31,11 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
     /** Cached model ids whose presence forces refresh when the static or migration-policy fingerprint changes. */
     dropCachedModelIdsOnStaticMismatch?: readonly string[];
     /**
-     * Trusted, provider-wide request headers (compile-time constants, never
-     * credentials) that the cache may restore by value for any model whose live
-     * headers matched them at write time. Lets header-bearing dynamic models
-     * without a bundled static entry survive offline reads instead of being
-     * dropped as unrestorable (e.g. GitHub Copilot's User-Agent + API version).
+     * Trusted, provider-wide request headers that can be safely re-derived at
+     * cache-read time. The cache never persists their values. Models whose live
+     * headers matched this record at write time can survive offline reads without
+     * a bundled static source, including configured discovery providers whose
+     * bearer header comes from the current local config.
      */
     restorableHeaderFallback?: Record<string, string>;
     /** Optional dynamic endpoint fetcher. */

package/dist/types/model-thinking.d.ts

@@ -44,8 +44,12 @@ export declare function getSupportedEfforts<TApi extends Api>(model: ApiModel<TA
  */
 export declare function clampThinkingLevelForModel<TApi extends Api>(model: ApiModel<TApi> | undefined, requested: Effort | undefined): Effort | undefined;
 export declare function requireSupportedEffort<TApi extends Api>(model: ApiModel<TApi>, effort: Effort): Effort;
-/** Maps a normalized thinking effort to Google's `thinkingLevel` enum values. */
-export declare function mapEffortToGoogleThinkingLevel(effort: Effort): "MINIMAL" | "LOW" | "MEDIUM" | "HIGH";
+/** Maps a normalized thinking effort to Google's `thinkingLevel` enum values.
+ * When a collapsed family routes `minimal` onto the same wire id as `low`
+ * (Antigravity Gemini 3.6/3.7 Flash), emit `LOW` — Cloud Code Assist rejects
+ * `MINIMAL` on those `-low` SKUs.
+ */
+export declare function mapEffortToGoogleThinkingLevel<TApi extends Api>(effort: Effort, model?: ApiModel<TApi>): "MINIMAL" | "LOW" | "MEDIUM" | "HIGH";
 /**
  * Maps a normalized thinking effort to Anthropic adaptive effort values via
  * the model's baked `thinking.effortMap` (identity for unmapped efforts).
@@ -63,4 +67,15 @@ export declare function resolveWireModelId<TApi extends Api>(model: ApiModel<TAp
  * thinking-off requests on `thinking.requiresEffort` models.
  */
 export declare function minimumSupportedEffort<TApi extends Api>(model: ApiModel<TApi>): Effort | undefined;
+/**
+ * Clamp target for effort-less requests on `thinking.requiresEffort` models:
+ * the effort whose wire route equals the model's default wire id
+ * (`requestModelId`), so a collapsed row clamps to the tier it already
+ * advertises as its default rather than the numerically lowest supported tier
+ * (e.g. Cursor Grok 4.5/4.6 default to `medium`, the only tier the Start plan
+ * serves). Falls back to {@link minimumSupportedEffort} when no route matches
+ * the default id — families whose default already is the minimum, or that
+ * expose no routing, are unaffected.
+ */
+export declare function defaultSupportedEffort<TApi extends Api>(model: ApiModel<TApi>): Effort | undefined;
 export {};

package/dist/types/model-tokenizer.d.ts

@@ -0,0 +1,8 @@
+import type { ModelTokenizer } from "./types.js";
+/**
+ * Resolve the exact locally embedded tokenizer for a canonical model id.
+ *
+ * This is catalog policy, not a runtime caller heuristic: [`buildModel`](./build.ts)
+ * materializes the result as `Model.tokenizer`; consumers read that property.
+ */
+export declare function resolveModelTokenizer(modelId: string): ModelTokenizer | undefined;

package/dist/types/models.d.ts

@@ -1,11 +1,14 @@
 import MODELS from "./models.json";
-import type { Api, KnownProvider, Model, Usage } from "./types.js";
+import type { Api, KnownProvider, Model, ModelCost, Usage } from "./types.js";
 export type GeneratedProvider = keyof typeof MODELS;
 export declare function getBundledModel<TApi extends Api = Api>(provider: GeneratedProvider, modelId: string): Model<TApi>;
 export declare function getBundledProviders(): KnownProvider[];
 export declare function getBundledModels(provider: GeneratedProvider): Model<Api>[];
 /** Price a prompt as fully uncached input under its active context-length tier. */
-export declare function calculateUncachedInputCost(cost: Model["cost"], promptInputTokens: number): number;
+export declare function calculateUncachedInputCost(cost: ModelCost, promptInputTokens: number): number;
+/** Price one usage record from a token rate card, including active context tiers. */
+export declare function calculateUsageCost(cost: ModelCost, usage: Usage): Usage["cost"];
+/** Price one usage record from its model's token rate card. */
 export declare function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage["cost"];
 /**
  * Check if two models are equal by comparing both their id and provider.

package/dist/types/provider-models/cache-provider-id.d.ts

@@ -2,6 +2,8 @@ export interface ModelCacheProviderIdOptions {
     apiKey?: string;
     baseUrl?: string;
 }
+/** Whether a provider's model-cache namespace requires its resolved credential. */
+export declare function isCredentialScopedModelCacheProvider(providerId: string): boolean;
 export declare function getDefaultModelDiscoveryBaseUrl(providerId: string): string | undefined;
 /** Resolve an Ollama model-cache namespace scoped to the normalized discovery endpoint. */
 export declare function resolveOllamaModelCacheProviderId(providerId: string, baseUrl?: string): string;

package/dist/types/provider-models/descriptors.d.ts

@@ -205,7 +205,7 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly id: "litellm";
     readonly defaultModel: "claude-opus-4-8";
     readonly envVars: readonly ["LITELLM_API_KEY"];
-    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<import("../index.js").Api, unknown>;
     readonly catalogDiscovery: {
         readonly label: "LiteLLM";
         readonly allowUnauthenticated: true;
@@ -428,17 +428,18 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly defaultModel: "openai/gpt-oss-120b";
     readonly envVars: readonly ["COREWEAVE_API_KEY", "WANDB_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
     readonly catalogDiscovery: {
         readonly label: "CoreWeave Serverless Inference";
     };
 }, {
     readonly id: "xai";
-    readonly defaultModel: "grok-4.5";
+    readonly defaultModel: "grok-4.6";
     readonly envVars: readonly ["XAI_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-responses", unknown>;
 }, {
     readonly id: "xai-oauth";
-    readonly defaultModel: "grok-4.5";
+    readonly defaultModel: "grok-4.6";
     readonly envVars: readonly ["XAI_OAUTH_TOKEN", "XAI_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-responses", unknown>;
     readonly catalogDiscovery: {

package/dist/types/provider-models/google.d.ts

@@ -18,6 +18,8 @@ export interface GoogleAntigravityModelManagerConfig {
 }
 export interface GoogleGeminiCliModelManagerConfig {
     oauthToken?: string;
+    /** GCP project id required by Workspace/Standard credentials for quota discovery. */
+    projectId?: string;
     endpoint?: string;
     fetch?: FetchImpl;
 }

package/dist/types/provider-models/openai-compat.d.ts

@@ -166,6 +166,11 @@ export interface XaiModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * Applies xAI's long-context rate card and mirrors exact public-model prices
+ * onto matching SuperGrok catalog rows.
+ */
+export declare function applyXaiCatalogPricing(models: readonly ModelSpec[]): ModelSpec[];
 export declare function xaiModelManagerOptions(config?: XaiModelManagerConfig): ModelManagerOptions<"openai-responses">;
 export interface XaiOAuthModelManagerConfig {
     apiKey?: string;
@@ -568,12 +573,14 @@ export interface FetchLiteLLMRichModelsOptions<TApi extends Api> {
     signal?: AbortSignal;
     timeoutMs?: number;
     referenceResolver?: (modelId: string) => ModelSpec<TApi> | undefined;
+    resolveApi?: (entry: Record<string, unknown>, modelId: string) => TApi;
 }
 export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_CONTEXT_WINDOW = 128000;
 export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_MAX_TOKENS = 32768;
 export declare function normalizeLiteLLMManagementBaseUrl(baseUrl: string): string;
+export declare function resolveLiteLLMApi(entry: Record<string, unknown> | undefined, id: string, fallbackApi?: Api): Api;
 export declare function fetchLiteLLMRichModels<TApi extends Api>(options: FetchLiteLLMRichModelsOptions<TApi>): Promise<ModelSpec<TApi>[] | null>;
-export declare function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export declare function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig): ModelManagerOptions<Api>;
 export interface VllmModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -122,8 +122,8 @@ export interface Usage {
     };
 }
 export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "kimi" | "qwen" | "qwen-chat-template";
-export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false";
-export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "thinking";
+export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "venice-disable-thinking" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false";
+export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "qwen" | "thinking";
 /**
  * Compatibility settings for openai-completions API.
  * Use this to override URL-based auto-detection for custom providers.
@@ -220,6 +220,16 @@ export interface OpenAICompat {
      * Non-Qwen templates ignore the flag, so the auto-detection is safe.
      */
     qwenPreserveThinking?: boolean;
+    /**
+     * Route the requested thinking effort onto the Qwen 3.8+ chat template's
+     * `reasoning_effort` kwarg (`low`/`medium`/`xhigh`; template default
+     * `xhigh`). Emitted inside `chat_template_kwargs` for both Qwen dialects
+     * (plus the top-level field on the `qwen` dialect, which newer llama.cpp
+     * builds map natively). Without it the qwen dialects only toggle
+     * `enable_thinking` and the template always thinks at its `xhigh` default.
+     * Default: auto-detected (Qwen 3.8+ id on a local llama.cpp-style backend).
+     */
+    qwenTemplateReasoningEffort?: boolean;
     /** Whether assistant tool-call messages must include non-empty content. Default: false. */
     requiresAssistantContentForToolCalls?: boolean;
     /** Whether the provider supports the `tool_choice` parameter. Default: true. */
@@ -558,6 +568,7 @@ export interface ResolvedOpenAISharedCompat {
     allowsSyntheticReasoningContentForToolCalls: boolean;
     replayReasoningContent: boolean;
     qwenPreserveThinking: boolean;
+    qwenTemplateReasoningEffort: boolean;
     requiresThinkingAsText: boolean;
     requiresMistralToolIds: boolean;
     requiresToolResultName: boolean;
@@ -595,7 +606,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -694,19 +705,37 @@ export interface TokenCost {
     cacheWrite: number;
 }
 /**
- * Rates applied to the full request when its prompt exceeds `inputThreshold`.
- * Prompt input is the sum of uncached, cached-read, cache-write, and
- * provider-orchestration input tokens.
+ * Rates applied to the full request when its prompt exceeds `inputThreshold`,
+ * or reaches it when `inputThresholdInclusive` is true. Prompt input is the
+ * sum of uncached, cached-read, cache-write, and provider-orchestration input
+ * tokens.
  */
 export interface LongContextTokenCost extends TokenCost {
     inputThreshold: number;
+    /** Whether the long-context tier starts exactly at `inputThreshold`. */
+    inputThresholdInclusive?: boolean;
 }
 /** Base token rates plus an optional long-context tier. */
 export interface ModelCost extends TokenCost {
     longContext?: LongContextTokenCost;
 }
+/**
+ * Exact local content tokenizer family for a model.
+ *
+ * Absent means no first-party local tokenizer is known and consumers retain
+ * their estimate/default-tokenizer policy. The values name tokenizer
+ * generations rather than providers: DeepSeek V3 through V4 share
+ * `"deepseek-v3"`; Kimi K2 through K3 share `"kimi-k2"`.
+ */
+export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5";
 export interface Model<TApi extends Api = Api> {
     id: string;
+    /**
+     * Whether provider-bound private-use glyphs require reversible ASCII tokenization.
+     * Materialized by `buildModel`; request handlers read this capability instead of
+     * inferring it from the transport API.
+     */
+    requiresGlyphTokenization?: boolean;
     /**
      * Model id to send on the wire when it differs from `id`. Used by catalog
      * variants that present one upstream model under several local entries —
@@ -728,6 +757,12 @@ export interface Model<TApi extends Api = Api> {
     provider: Provider;
     baseUrl: string;
     reasoning: boolean;
+    /**
+     * Exact local tokenizer family resolved from the model identity or supplied
+     * explicitly by a catalog/discovery source. Absent leaves local counting to
+     * the consumer's fallback policy.
+     */
+    tokenizer?: ModelTokenizer;
     input: ("text" | "image")[];
     /**
      * Decoder family used for image inputs when it has narrower format support
@@ -786,6 +821,8 @@ export interface Model<TApi extends Api = Api> {
     preferWebsockets?: boolean;
     /** Codex Responses Lite transport: send the lite marker and carry instructions/tools as input items (mirrors codex-rs `use_responses_lite`). */
     useResponsesLite?: boolean;
+    /** Codex Code Mode restriction: model expects tools routed through a programmatic exec surface (mirrors codex-rs `tool_mode`). */
+    toolMode?: "code_mode_only";
     /** Preferred model to switch to when context promotion is triggered (model id or provider/id). */
     contextPromotionTarget?: string;
     /** Preferred model to use only for compaction (model id or provider/id); the active session model is unchanged. */
@@ -818,13 +855,25 @@ export interface Model<TApi extends Api = Api> {
      * `options.isOAuth = true` for the underlying provider call.
      */
     isOAuth?: boolean;
+    /**
+     * Amazon Bedrock Guardrail id or ARN attached to every Converse request for
+     * this model. Set from `providers.amazon-bedrock.guardrailIdentifier`; the
+     * streaming layer forwards it as `options.guardrailIdentifier` so accounts
+     * that gate `bedrock:InvokeModel*` on the `bedrock:GuardrailIdentifier`
+     * condition key stop returning an explicit deny.
+     */
+    guardrailIdentifier?: string;
+    /** Bedrock guardrail version. Defaults to `"DRAFT"` at request time when unset. */
+    guardrailVersion?: string;
+    /** Bedrock guardrail trace verbosity. */
+    guardrailTrace?: "enabled" | "disabled" | "enabled_full";
 }
 /**
  * A model as authored by configs, bundled catalogs, and discovery — the input
  * vocabulary of `buildModel`. Identical to `Model` except `compat` carries the
  * sparse override shape and nothing is resolved yet.
  */
-export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "compatConfig" | "supportsComputerUseConfig"> {
+export interface ModelSpec<TApi extends Api = Api> extends Omit<Model<TApi>, "compat" | "compatConfig" | "requiresGlyphTokenization" | "supportsComputerUseConfig"> {
     /** Sparse compatibility overrides; resolved into `Model.compat` by `buildModel`. */
     compat?: CompatConfigOf<TApi>;
 }

package/dist/types/variant-collapse.d.ts

@@ -21,6 +21,15 @@ export interface EffortVariantFamily {
      * when it equals the logical id).
      */
     members: readonly string[];
+    /**
+     * Preferred default wire id: overrides the member-order default for the
+     * collapsed spec's `requestModelId` when this member is live. Lets a
+     * mandatory-reasoning family advertise a canonical tier that is not its
+     * numeric floor (Cursor Grok 4.5/4.6 default to `-medium`, the only tier the
+     * Start plan serves; the `-low` floor is refused). Ignored when absent from
+     * the input or retired.
+     */
+    defaultMember?: string;
     /**
      * Wire ids upstream no longer serves (e.g. a deployment killed while
      * discovery still advertises it). Fresh collapsing never routes to them,
@@ -58,6 +67,8 @@ export declare const ANTIGRAVITY_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
 /** `google-gemini-cli` Gemini families on the official CLI's level transport. */
 export declare const GEMINI_CLI_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
 export declare const DEVIN_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
+/** `cursor` per-effort sibling families collapsed per service-tier lane: Grok 4.5/4.6 plus GPT-5.6 Luna/Sol/Terra. */
+export declare const CURSOR_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
 /** Provider id → hand collapse table. The CCA providers diverge on thinking transport. */
 export declare const VARIANT_COLLAPSE_TABLES: Readonly<Record<string, VariantCollapseTable>>;
 /**
@@ -89,10 +100,11 @@ export declare function isVariantCollapsedSpec(spec: VariantSpecLike): boolean;
  */
 export declare function collapseEffortVariants<TSpec extends VariantSpecLike>(specs: readonly TSpec[], table: VariantCollapseTable): TSpec[];
 /**
- * Collapse a full mixed-provider list: per provider, the hand table (when
- * registered) plus the automatic `X`/`X-thinking` pair rule. Used by the
- * catalog generator; the runtime equivalent lives at the model-manager merge
- * point. Output is regrouped by provider — callers re-sort.
+ * Collapse a full mixed-provider list: per provider, the hand table, Cursor's
+ * conservative live effort-sibling rule, and the automatic `X`/`X-thinking`
+ * pair rule. Used by the catalog generator; the runtime equivalent lives at
+ * the model-manager merge point. Output is regrouped by provider — callers
+ * re-sort.
  */
 export declare function collapseEffortVariantsAcrossProviders<TSpec extends VariantSpecLike>(specs: readonly TSpec[]): TSpec[];
 /**
@@ -105,29 +117,30 @@ export declare function collapseEffortVariantsAcrossProviders<TSpec extends Vari
 export declare function collapseBuiltModelVariants<TApi extends Api>(models: readonly Model<TApi>[]): Model<TApi>[];
 /**
  * Resolve a retired effort-tier variant id (collapsed member, recycled id) to
- * its replacement model id for `provider` via the hand table. Returns
- * `undefined` when the id is not a known alias; derived `X-thinking` members
- * resolve through `stripThinkingVariantToken` instead. Callers must try an
- * exact model lookup first — a live model always wins over an alias.
+ * its replacement model id for `provider` via hand-table or registered live
+ * aliases. Returns `undefined` when the id is not a known alias; derived
+ * `X-thinking` members also resolve through `stripThinkingVariantToken`.
+ * Callers must try an exact model lookup first — a live model always wins over
+ * an alias.
  */
 export declare function resolveVariantAlias(provider: Provider, modelId: string): string | undefined;
 /** Bare-id alias hit: replacement id plus the providers declaring it. */
 export interface BareVariantAliasHit {
     id: string;
-    /** Providers whose table declares the alias — candidates from these win ties. */
+    /** Providers declaring the alias — candidates from these win ties. */
     providers: readonly Provider[];
 }
 /**
- * Provider-agnostic hand-table alias lookup for bare-id selectors. Returns
- * the declaring providers so callers can prefer their models when the
- * replacement id exists on unrelated providers too (e.g. a retired Cursor
- * tier id must not resolve to `openai/gpt-5.4`).
+ * Provider-agnostic alias lookup for bare-id selectors. Returns the declaring
+ * providers so callers can prefer their models when the replacement id exists
+ * on unrelated providers too (e.g. a retired Cursor tier id must not resolve
+ * to `openai/gpt-5.4`).
  */
 export declare function resolveBareVariantAlias(modelId: string): BareVariantAliasHit | undefined;
 /**
  * Reverse alias lookup: the retired ids that resolve to `modelId` for
- * `provider` via the hand table. Used to re-key config keyed by raw member
- * ids (models.yml `modelOverrides`, suppressed selectors) onto the collapsed
- * model. Empty for providers without a table.
+ * `provider` via hand-table or registered live aliases. Used to re-key config
+ * keyed by raw member ids (models.yml `modelOverrides`, suppressed selectors)
+ * onto the collapsed model.
  */
 export declare function getVariantAliasSources(provider: Provider, modelId: string): readonly string[];