@linxiraos/pi-catalog 1.0.0

package/dist/types/discovery/openai-compatible.d.ts

@@ -0,0 +1,74 @@
+import type { Api, FetchImpl, ModelSpec, Provider } from "../types.js";
+/**
+ * Minimal OpenAI-style model entry shape consumed by discovery.
+ *
+ * Providers may return additional fields; this type only captures
+ * fields that are useful for generic normalization.
+ */
+export interface OpenAICompatibleModelRecord {
+    id?: unknown;
+    name?: unknown;
+    object?: unknown;
+    owned_by?: unknown;
+    [key: string]: unknown;
+}
+/**
+ * Tolerant envelope for OpenAI-compatible `/models` responses.
+ *
+ * Common providers return `{ data: [...] }`, but variants such as
+ * `{ models: [...] }`, `{ result: [...] }`, or direct arrays are also
+ * accepted during extraction.
+ */
+export interface OpenAICompatibleModelsEnvelope {
+    data?: unknown;
+    models?: unknown;
+    result?: unknown;
+    items?: unknown;
+    [key: string]: unknown;
+}
+/**
+ * Context passed to custom OpenAI-compatible model mappers.
+ */
+export interface OpenAICompatibleModelMapperContext<TApi extends Api> {
+    api: TApi;
+    provider: Provider;
+    baseUrl: string;
+}
+/**
+ * Options for fetching and normalizing OpenAI-compatible `/models` catalogs.
+ */
+export interface FetchOpenAICompatibleModelsOptions<TApi extends Api> {
+    /** API type assigned to normalized models. */
+    api: TApi;
+    /** Provider id assigned to normalized models. */
+    provider: Provider;
+    /** Provider base URL used for both fetch and normalized model records. */
+    baseUrl: string;
+    /** Optional bearer token for Authorization header. */
+    apiKey?: string;
+    /** Additional request headers. */
+    headers?: Record<string, string>;
+    /** Optional AbortSignal for request cancellation; caller owns its lifecycle. */
+    signal?: AbortSignal;
+    /** Optional cancellable request timeout used when `signal` is omitted. */
+    timeoutMs?: number;
+    /** Optional fetch implementation override for testing/custom runtimes. */
+    fetch?: FetchImpl;
+    /**
+     * Optional post-normalization filter.
+     * Return false to skip a model.
+     */
+    filterModel?: (entry: OpenAICompatibleModelRecord, model: ModelSpec<TApi>) => boolean;
+    /**
+     * Optional mapper override for provider-specific quirks.
+     * Return null to skip a model.
+     */
+    mapModel?: (entry: OpenAICompatibleModelRecord, defaults: ModelSpec<TApi>, context: OpenAICompatibleModelMapperContext<TApi>) => ModelSpec<TApi> | null;
+}
+/**
+ * Fetches and normalizes an OpenAI-compatible `/models` catalog.
+ *
+ * Returns `null` on transport/protocol failures.
+ * Returns `[]` only when the endpoint responds successfully with no usable models.
+ */
+export declare function fetchOpenAICompatibleModels<TApi extends Api>(options: FetchOpenAICompatibleModelsOptions<TApi>): Promise<ModelSpec<TApi>[] | null>;

package/dist/types/effort.d.ts

@@ -0,0 +1,10 @@
+/** User-facing thinking levels, ordered least to most intensive. */
+export declare const enum Effort {
+    Minimal = "minimal",
+    Low = "low",
+    Medium = "medium",
+    High = "high",
+    XHigh = "xhigh",
+    Max = "max"
+}
+export declare const THINKING_EFFORTS: readonly Effort[];

package/dist/types/fireworks-model-id.d.ts

@@ -0,0 +1,23 @@
+export declare function toFireworksPublicModelId(modelId: string): string;
+export declare function toFireworksWireModelId(modelId: string): string;
+/**
+ * Fire Pass exposes its Kimi K2.6 Turbo subscription through a dedicated router
+ * endpoint at `accounts/fireworks/routers/<id>` rather than the `models/` namespace.
+ * We keep a friendly public id (e.g. `kimi-k2.6-turbo`) in the catalog and translate
+ * to the wire form (`accounts/fireworks/routers/kimi-k2p6-turbo`) at request time.
+ */
+export declare function toFirepassPublicModelId(modelId: string): string;
+export declare function toFirepassWireModelId(modelId: string): string;
+/**
+ * Public-id suffix marking a Fireworks "Fast" serving-path variant. Fast is a
+ * higher-throughput route (100+ tok/s) exposed under a dedicated router id
+ * (`accounts/fireworks/routers/<id>-fast`), not a separate model — same weights,
+ * higher price, no Priority tier. We keep a friendly `<id>-fast` public id and
+ * translate it to the router wire form at request time (compat
+ * `wireModelIdMode: "firepass"`). See https://docs.fireworks.ai/serverless/serving-paths.
+ */
+export declare const FIREWORKS_FAST_SUFFIX = "-fast";
+/** True for a Fireworks public model id that selects the Fast serving path. */
+export declare function isFireworksFastModelId(modelId: string): boolean;
+/** Strip the Fast suffix to recover the base (Standard-tier) model id. */
+export declare function toFireworksBaseModelId(modelId: string): string;

package/dist/types/hosts.d.ts

@@ -0,0 +1,164 @@
+/**
+ * Known model-endpoint host classification — the single vocabulary for the
+ * `provider === id || baseUrl.includes(marker)` idiom that gates wire-level
+ * behavior (compat detection, routing, header shaping, watchdog floors).
+ *
+ * Markers are case-insensitive substrings matched against the base URL, NOT
+ * parsed hostnames: proxies regularly embed the upstream host in a path
+ * segment, and the historical call sites all used substring semantics.
+ * Callers that need strict hostname matching — where a substring false
+ * positive is dangerous, e.g. the Anthropic official-endpoint OAuth gate —
+ * parse the URL and compare the hostname themselves.
+ */
+export declare const KNOWN_HOSTS: {
+    readonly openai: {
+        readonly providers: readonly ["openai"];
+        readonly urlMarkers: readonly ["api.openai.com"];
+    };
+    readonly azureOpenAI: {
+        readonly providers: readonly ["azure"];
+        readonly urlMarkers: readonly [".openai.azure.com", "azure.com/openai", "models.inference.ai.azure.com"];
+    };
+    readonly openrouter: {
+        readonly providers: readonly ["openrouter"];
+        readonly urlMarkers: readonly ["openrouter.ai"];
+    };
+    readonly vercelAIGateway: {
+        readonly providers: readonly ["vercel-ai-gateway"];
+        readonly urlMarkers: readonly ["ai-gateway.vercel.sh"];
+    };
+    readonly githubCopilot: {
+        readonly providers: readonly ["github-copilot"];
+        readonly urlMarkers: readonly ["githubcopilot.com", "copilot-api."];
+    };
+    readonly anthropic: {
+        readonly providers: readonly ["anthropic"];
+        readonly urlMarkers: readonly ["api.anthropic.com"];
+    };
+    /** DeepSeek's first-party API only — gates direct-API quirks (max_tokens field, thinking extraBody). */
+    readonly deepseekDirect: {
+        readonly providers: readonly ["deepseek"];
+        readonly urlMarkers: readonly ["api.deepseek.com"];
+    };
+    /** Any DeepSeek-operated host (first-party API, web-chat fronts). Wider than `deepseekDirect` on purpose. */
+    readonly deepseekFamily: {
+        readonly providers: readonly ["deepseek"];
+        readonly urlMarkers: readonly ["deepseek.com"];
+    };
+    readonly cerebras: {
+        readonly providers: readonly ["cerebras"];
+        readonly urlMarkers: readonly ["cerebras.ai"];
+    };
+    readonly zai: {
+        readonly providers: readonly ["zai"];
+        readonly urlMarkers: readonly ["api.z.ai"];
+    };
+    readonly zhipu: {
+        readonly providers: readonly ["zhipu-coding-plan"];
+        readonly urlMarkers: readonly ["open.bigmodel.cn"];
+    };
+    readonly kilo: {
+        readonly providers: readonly ["kilo"];
+        readonly urlMarkers: readonly ["api.kilo.ai"];
+    };
+    readonly alibabaDashscope: {
+        readonly providers: readonly ["alibaba-coding-plan", "alibaba-token-plan"];
+        readonly urlMarkers: readonly ["dashscope", "token-plan."];
+    };
+    readonly umans: {
+        readonly providers: readonly ["umans"];
+        readonly urlMarkers: readonly ["api.code.umans.ai"];
+    };
+    readonly xiaomi: {
+        readonly providers: readonly ["xiaomi"];
+        readonly providerPrefixes: readonly ["xiaomi-token-plan-"];
+        readonly urlMarkers: readonly ["xiaomimimo.com"];
+    };
+    readonly xai: {
+        readonly providers: readonly ["xai"];
+        readonly urlMarkers: readonly ["api.x.ai"];
+    };
+    readonly mistral: {
+        readonly providers: readonly ["mistral"];
+        readonly urlMarkers: readonly ["mistral.ai"];
+    };
+    readonly together: {
+        readonly providers: readonly ["together"];
+        readonly urlMarkers: readonly ["api.together.xyz"];
+    };
+    readonly baseten: {
+        readonly providers: readonly ["baseten"];
+        readonly urlMarkers: readonly ["baseten.co"];
+    };
+    /** URL-only on purpose: the `fireworks`/`firepass` providers route per-model and not every model is Fireworks-shaped. */
+    readonly fireworks: {
+        readonly urlMarkers: readonly ["fireworks.ai"];
+    };
+    readonly groq: {
+        readonly providers: readonly ["groq"];
+        readonly urlMarkers: readonly ["api.groq.com"];
+    };
+    readonly minimax: {
+        readonly providers: readonly ["minimax", "minimax-code", "minimax-code-cn"];
+        readonly urlMarkers: readonly ["api.minimax.io", "api.minimaxi.com"];
+    };
+    readonly qwenPortal: {
+        readonly providers: readonly ["qwen-portal"];
+        readonly urlMarkers: readonly ["portal.qwen.ai"];
+    };
+    /** NVIDIA NIM (`integrate.api.nvidia.com`). Qwen NIM endpoints take `chat_template_kwargs.enable_thinking`, never top-level `enable_thinking`. */
+    readonly nvidia: {
+        readonly providers: readonly ["nvidia"];
+        readonly urlMarkers: readonly ["integrate.api.nvidia.com"];
+    };
+    readonly moonshotNative: {
+        readonly providers: readonly ["moonshot", "kimi-code"];
+        readonly urlMarkers: readonly ["api.moonshot.ai", "api.kimi.com"];
+    };
+    /** Google AI Studio's OpenAI-compatible shim (`/v1beta/openai`) — a subset of chat-completions; rejects `store` with a 400. Native Gemini uses `google-generative-ai` api instead. */
+    readonly googleAistudio: {
+        readonly providers: readonly [];
+        readonly urlMarkers: readonly ["generativelanguage.googleapis.com"];
+    };
+    readonly opencode: {
+        readonly providers: readonly ["opencode-go", "opencode-zen"];
+        readonly urlMarkers: readonly ["opencode.ai"];
+    };
+    /** ZenMux's Anthropic-compatible proxy (`zenmux.ai/api/anthropic`) forwards to signature-enforcing Anthropic. */
+    readonly zenmux: {
+        readonly providers: readonly ["zenmux"];
+        readonly urlMarkers: readonly ["zenmux.ai"];
+    };
+    readonly chutes: {
+        readonly urlMarkers: readonly ["chutes.ai"];
+    };
+};
+export type KnownHost = keyof typeof KNOWN_HOSTS;
+/** URL-only host check (for call sites that have no provider id, e.g. raw env config). */
+export declare function hostMatchesUrl(baseUrl: string | undefined, host: KnownHost): boolean;
+/** Provider-or-URL host check — the canonical `provider === id || baseUrl.includes(marker)` idiom. */
+export declare function modelMatchesHost(model: {
+    provider: string;
+    baseUrl: string;
+}, host: KnownHost): boolean;
+/**
+ * Hostname for a Vertex AI GenerateContent / rawPredict / OpenAI-compat
+ * request for the given location.
+ *
+ * - `global` → global endpoint (`aiplatform.googleapis.com`)
+ * - `eu` / `us` multi-regions → REP endpoints (`aiplatform.{eu|us}.rep.googleapis.com`)
+ * - every other location → regional (`{location}-aiplatform.googleapis.com`)
+ *
+ * Multi-region codes do NOT follow the regional `{location}-aiplatform` pattern;
+ * interpolating them that way yields hosts like `eu-aiplatform.googleapis.com`
+ * that 404.
+ */
+export declare function resolveVertexEndpointHost(location: string): string;
+/** Vertex AI express-mode OpenAI-compatible endpoint (`…/endpoints/openapi`). */
+export declare function isVertexExpressOpenAIUrl(baseUrl: string): boolean;
+/** Vertex AI Anthropic raw-predict endpoints (`:streamRawPredict` / `:rawPredict`). */
+export declare function isVertexRawPredictUrl(baseUrl: string): boolean;
+/** Azure OpenAI deployment-scoped path (`…/deployments/<name>/…`). */
+export declare function isAzureDeploymentsUrl(baseUrl: string): boolean;
+/** Alibaba DashScope consumer `compatible-mode` endpoint (rejects multimodal arrays for some text-only SKUs). */
+export declare function isDashscopeCompatibleModeUrl(baseUrl: string): boolean;

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

@@ -0,0 +1,3 @@
+import { type ModelReferenceIndex } from "./reference.js";
+/** Proxy-reference index over the bundled catalog. */
+export declare function getBundledModelReferenceIndex(): ModelReferenceIndex;

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

@@ -0,0 +1,69 @@
+/**
+ * Model-id classification: parse a model id into its family (gemini / anthropic /
+ * openai), kind/variant, and version. This is the shared layer both catalog
+ * policy rules (`model-thinking.ts`) and downstream consumers build on —
+ * classification lives here, the rules that consume it stay with their domain.
+ */
+export type SemVer = {
+    major: number;
+    minor: number;
+    patch: number;
+};
+export type GeminiKind = "pro" | "flash";
+export type AnthropicKind = "opus" | "sonnet" | "fable" | "mythos";
+export type OpenAIVariant = "base" | "codex" | "codex-max" | "codex-mini" | "codex-spark" | "mini" | "max" | "nano";
+export type GlmVariant = "base" | "air" | "turbo" | "flash" | "flashx" | "preview";
+export interface GeminiModel {
+    family: "gemini";
+    kind: GeminiKind;
+    version: SemVer;
+}
+export interface AnthropicModel {
+    family: "anthropic";
+    kind: AnthropicKind;
+    version: SemVer;
+}
+export interface OpenAIModel {
+    family: "openai";
+    variant: OpenAIVariant;
+    version: SemVer;
+}
+export interface GlmModel {
+    family: "glm";
+    /** Suffix variant (`-air`, `-turbo`, `-flash`, `-flashx`, `-preview`); `base` when none. */
+    variant: GlmVariant;
+    /** Vision SKU — the `v` that attaches directly to the version (`glm-4v`, `glm-4.5v`). */
+    vision: boolean;
+    version: SemVer;
+}
+export interface UnknownModel {
+    family: "unknown";
+    id: string;
+}
+export type ParsedModel = GeminiModel | AnthropicModel | OpenAIModel | UnknownModel;
+export declare function bareModelId(modelId: string): string;
+export declare function parseKnownModel(modelId: string): ParsedModel;
+export declare const parseGeminiModel: (modelId: string) => GeminiModel | null;
+export declare const parseAnthropicModel: (modelId: string) => AnthropicModel | null;
+export declare const parseOpenAIModel: (modelId: string) => OpenAIModel | null;
+/**
+ * Parse a GLM (Zhipu / Z.AI) model id into family + variant + vision + version.
+ * Shape: `glm-<version>[v][-<variant>]` — e.g. `glm-4.5`, `glm-4.5-air`,
+ * `glm-5-turbo`, `glm-4.5v`, `glm-5-preview`. The `v` (vision) attaches to the
+ * version; other variants are `-` suffixes. Standalone like `parseAnthropicModel`
+ * is used in family.ts — GLM needs no global thinking policy, so it stays out of
+ * `parseKnownModel`.
+ */
+export declare const parseGlmModel: (modelId: string) => GlmModel | null;
+export declare function isFableOrMythos(kind: AnthropicKind): boolean;
+/**
+ * Returns true if the parsed Anthropic model is part of the adaptive-thinking
+ * Claude generation at or above a specific capability threshold.
+ * - Opus has a configurable minimum version floor (e.g. "4.6", "4.7", "4.8").
+ * - Sonnet, Fable, and Mythos all require version 5 or higher.
+ */
+export declare function isAnthropicAdaptiveGenAtLeast(parsed: AnthropicModel, opusMin: "4.6" | "4.7" | "4.8"): boolean;
+export declare function parseSemVer(version: string): SemVer | null;
+export declare function semverGte(left: SemVer | string, right: SemVer | string): boolean;
+export declare function semverEqual(left: SemVer | string, right: SemVer | string): boolean;
+export declare function compareSemVer(left: SemVer | string | null, right: SemVer | string | null): number;

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

@@ -0,0 +1,3 @@
+export type Dialect = "glm" | "hermes" | "kimi" | "xml" | "anthropic" | "deepseek" | "harmony" | "qwen3" | "gemini" | "gemma" | "minimax";
+export declare const FALLBACK_DIALECT: Dialect;
+export declare function preferredDialect(modelId: string): Dialect;

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

@@ -0,0 +1,181 @@
+/**
+ * Model-family id predicates: the shared vocabulary for "is this id a member
+ * of family X" checks that gate wire-level behavior across hosts (a Kimi or
+ * DeepSeek model keeps its quirks no matter which OpenAI-compatible proxy
+ * serves it). Looser per-feature heuristics (e.g. stream-markup healing)
+ * deliberately keep their own patterns — only provably-shared matchers live
+ * here.
+ */
+/** Kimi family ids in any namespace form (`moonshotai/kimi-*`, `kimi-k2.6`, `vendor/kimi.x`). */
+export declare const isKimiModelId: (modelId: string) => boolean;
+/** Kimi K2.6 specifically, including router ids that spell the version `k2p6`. */
+export declare const isKimiK26ModelId: (modelId: string) => boolean;
+/**
+ * Kimi K3 in any namespace form (`kimi-k3`, `kimi-k3.1`, `kimi-k3-turbo`,
+ * `moonshotai/kimi-k3`). K3 always reasons and drives thinking via OpenAI-style
+ * `reasoning_effort: "max"`, not the K2.x binary `thinking: { type }` block —
+ * see the moonshot discovery mapper and `buildOpenAICompat`.
+ */
+export declare const isKimiK3ModelId: (modelId: string) => boolean;
+/**
+ * Claude ids in any namespace form: bare (`claude-*`), path-namespaced
+ * (`anthropic/claude.x`), or dot-prefixed (`us.anthropic.claude-…`,
+ * `global.anthropic.claude-…`, `au.anthropic.claude-…` — Bedrock cross-region
+ * inference profiles). Necessary because {@link parseAnthropicModel} only
+ * classifies kinds enumerated in its regex, so any dotted profile whose kind
+ * (e.g. `haiku`) is not enumerated would otherwise slip past this fallback.
+ */
+export declare const isClaudeModelId: (modelId: string) => boolean;
+/** `anthropic/`-namespaced ids (aggregator catalogs like OpenRouter). */
+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;
+/** 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). */
+export declare const isDeepseekModelIdOrName: (modelId: string) => boolean;
+/**
+ * DeepSeek V4 Flash SKU in any host/namespace form (`deepseek-v4-flash`, dated
+ * `deepseek-v4-flash-0731`, `deepseek-ai/DeepSeek-V4-Flash`). Flash is the only
+ * V4 model whose `reasoning_effort` accepts the `low` tier; V4 Pro tops out at
+ * `high`/`max`. See https://api-docs.deepseek.com/api/create-chat-completion.
+ */
+export declare const isDeepseekV4FlashModelId: (modelId: string) => boolean;
+/** Xiaomi MiMo family by id or display name. */
+export declare const isMimoModelIdOrName: (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.
+ */
+export declare const isGrokReasoningEffortCapable: (modelId: string) => boolean;
+/**
+ * MiniMax M2-generation family (M2, M2.1, M2.5, M2.7, including `-highspeed`/
+ * `-lightning`/`-her`/`-turbo` variants, dotless aliases like `minimax-m21`,
+ * and short `minimax/m2-…` ids on aggregator hosts). Underlying model accepts
+ * only `low|medium|high` for `reasoning_effort` and 400s on `minimal`,
+ * `xhigh`, or `none` — so hosts whose default effort map otherwise lowers
+ * `minimal` to `none` (Fireworks) or expects the full 5-tier scale must
+ * clamp instead. Excludes M1, M3, MiniMax-Text-01, music, hailuo, voice ids.
+ */
+export declare const isMinimaxM2FamilyModelId: (modelId: string) => boolean;
+/** MiniMax M3 family ids in bundled/default and aggregator namespace forms. */
+export declare const isMinimaxM3FamilyModelId: (modelId: string) => boolean;
+/**
+ * OpenAI gpt-oss family (`gpt-oss-20b`, `gpt-oss-120b`, `gpt-oss:120b`,
+ * `vendor/gpt-oss-…`). The Harmony reasoning format only accepts
+ * `low|medium|high` for `reasoning_effort` and rejects `minimal`, `xhigh`,
+ * and `none`.
+ */
+export declare const isOpenAIGptOssModelId: (modelId: string) => boolean;
+/** OpenAI model ids (gpt-*, chatgpt-*, o1/o3/o4 SKUs, codex-*, or openai/*). */
+export declare const isOpenAIModelId: (modelId: string) => boolean;
+/**
+ * OpenAI Codex models that honor `reasoning.context: "all_turns"` (full
+ * cross-turn reasoning replay). The `reasoning.context` field itself exists for
+ * the whole gpt-5/o-series family, but the `all_turns` value is only accepted
+ * from gpt-5.4 onward; earlier ids (`gpt-5.1-codex`, `gpt-5.3-codex`, and
+ * `gpt-5.3-codex-spark`) reject it with
+ * `Unsupported value: 'all_turns' is not supported with this model`. Version
+ * floor (not an allowlist) so 5.6/6.x inherit support automatically. Callers
+ * fall back to omitting `context`, letting the server default to `current_turn`.
+ */
+export declare const supportsAllTurnsReasoningContext: (modelId: string) => boolean;
+/**
+ * OpenAI Codex models that accept `reasoning.summary`. Shares the gpt-5.4 wire
+ * floor with {@link supportsAllTurnsReasoningContext}: earlier Codex ids
+ * (`gpt-5.1-codex`, `gpt-5.3-codex`, `gpt-5.3-codex-spark`) reject the field
+ * with `Unsupported parameter: 'reasoning.summary' is not supported with this
+ * model`. Callers omit `summary` for unsupported ids, letting the server skip
+ * the human-readable summary stream.
+ */
+export declare const supportsCodexReasoningSummary: (modelId: string) => boolean;
+/**
+ * OpenAI proprietary models whose serving path rejects explicit sampling
+ * parameters (`temperature`, `top_p`, `top_k`, …) with
+ * `400 Unsupported parameter: 'temperature' is not supported with this model`.
+ * Covers the o-series and the entire gpt-5+ generation — base, `mini`, `nano`,
+ * `codex*`, the `luna`/`sol`/`terra` SKUs, and the `-chat-latest` variants,
+ * since even the non-reasoning gpt-5 chat models reject sampling params (see
+ * litellm#13781). Holds regardless of which OpenAI-serving host proxies the
+ * model (official, Azure, GitHub Copilot). Version floor (not an allowlist) so
+ * 6.x inherits automatically. Issue #5606.
+ */
+export declare const isOpenAISamplingRestrictedModelId: (modelId: string) => boolean;
+/**
+ * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
+ * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning
+ * `-flash`/`-flashx`/`-preview` variants, and pre-4.5 ids. Matching the family
+ * keeps newly-bumped integers (`glm-5.3`, `glm-6`, …) covered without a per-id
+ * allowlist.
+ */
+export declare const isReasoningGlmModelId: (modelId: string) => boolean;
+/** GLM-5.2+ coding SKUs accept `reasoning_effort` in addition to binary thinking. */
+export declare const isGlm52ReasoningEffortModelId: (modelId: string) => boolean;
+/** GLM vision SKUs — the `v` that attaches to the version (`glm-4v`, `glm-4.5v`). */
+export declare const isGlmVisionModelId: (modelId: string) => boolean;
+/**
+ * Coarse vendor-lineage token for "are two models the same family?" checks
+ * (e.g. picking a cross-family reviewer). All Claude point releases share a token,
+ * Claude and GPT differ; namespace prefixes and aggregator mirrors fold onto the
+ * lineage via {@link parseKnownModel}'s `bareModelId` normalization. Opaque and
+ * comparison-only — not a stable key to persist, since the vocabulary tracks new
+ * releases. Returns `""` for ids it cannot classify; callers fall back to the provider.
+ *
+ * Vendor-only by design: a model's kind/variant (opus vs sonnet, codex vs base) is
+ * collapsed onto the single vendor token; use {@link parseKnownModel} for finer breakdowns.
+ */
+export declare const modelFamilyToken: (modelId: string) => string;
+/**
+ * True for Claude generations that support extended thinking: Sonnet/Opus 3.7+,
+ * every 4.x/5+ Opus/Sonnet, and the Fable/Mythos generation. Pre-thinking
+ * models (Claude 3.5 and older) are excluded so no thinking effort dial is
+ * fabricated for a model that rejects thinking parameters. Classifier-based, so
+ * dotted and dashed version forms both match; ids the classifier does not parse
+ * (e.g. Haiku, bare dated ids) return false.
+ */
+export declare const anthropicModelSupportsThinking: (modelId: string) => boolean;
+/**
+ * Adaptive thinking `display` is supported starting with Claude Opus 4.7+,
+ * Sonnet 5+, and the Claude Fable/Mythos 5 generation. Older adaptive-thinking
+ * models (Opus 4.6, Sonnet 4.6) reject the field. Classifier-based, so dotted
+ * and dashed version forms both match while bare dated ids
+ * (`claude-opus-4-20250514` = Opus 4.0) stay excluded.
+ */
+export declare const supportsAdaptiveThinkingDisplay: (modelId: string) => boolean;
+/**
+ * Returns true for Anthropic models with Opus 4.7+, Sonnet 5+, and Fable/Mythos 5+
+ * API restrictions:
+ * - Sampling parameters (temperature/top_p/top_k) return 400 error
+ * - Thinking content is omitted by default (needs display: "summarized")
+ */
+export declare const hasOpus47ApiRestrictions: (modelId: string) => boolean;
+/**
+ * Mid-conversation `role: "system"` messages (system instructions appended at
+ * non-first positions in the `messages` array) are supported starting with
+ * Claude Opus 4.8+, Sonnet 5+, and the Claude Fable/Mythos 5 generation.
+ * Earlier Claude models reject the role.
+ * @see https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages
+ */
+export declare const supportsMidConversationSystemMessages: (modelId: string) => boolean;
+export declare const isAnthropicFableOrMythosModel: (modelId: string) => boolean;
+/** Thinking-variant token location inside a model id. */
+export interface ThinkingVariantToken {
+    index: number;
+    length: number;
+}
+/**
+ * Locates the first thinking-variant token (`-thinking`, `-reasoner`,
+ * `-reasoning`; trailing or infix) in a model id. The token ends at the id
+ * end or any non-alphanumeric boundary, and negated forms (`non-thinking`,
+ * `no-thinking`) never match — those name the NON-thinking SKU.
+ */
+export declare function findThinkingVariantToken(modelId: string): ThinkingVariantToken | undefined;
+/**
+ * Removes the located thinking-variant token: `kimi-k2-thinking` → `kimi-k2`,
+ * `mimo-v2-flash-thinking-original` → `mimo-v2-flash-original`,
+ * `grok-4.1-fast-reasoning` → `grok-4.1-fast`. Returns `undefined` when no
+ * token exists or nothing would remain. Callers MUST verify the result names
+ * a live model.
+ */
+export declare const stripThinkingVariantToken: (modelId: string) => string | undefined;

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

@@ -0,0 +1,12 @@
+export declare function getModelLikeIdSegments(modelId: string): string[];
+export declare function getLongestModelLikeIdSegment(modelId: string): string | undefined;
+/**
+ * Strip reseller / wrapper tags that are injected as bracketed affixes around an
+ * upstream model id, e.g.
+ *   "[Kiro] claude-opus-4-8"                -> "claude-opus-4-8"
+ *   "[gcli转] gemini-3.1-pro-preview [假流]" -> "gemini-3.1-pro-preview"
+ *
+ * Candidates are returned most-stripped first: both ends, then leading-only, then trailing-only.
+ */
+export declare function getBracketStrippedModelIdCandidates(modelId: string): string[];
+export declare function stripBracketedModelIdAffixes(modelId: string): string | undefined;

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

@@ -0,0 +1,8 @@
+export * from "./bundled.js";
+export * from "./classify.js";
+export * from "./dialect.js";
+export * from "./family.js";
+export * from "./id.js";
+export * from "./markers.js";
+export * from "./priority.js";
+export * from "./reference.js";

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

@@ -0,0 +1,4 @@
+/** Marker pattern used by canonical-id resolution (`search` excluded). */
+export declare const CANONICAL_TRAILING_MARKER_PATTERN: RegExp;
+/** Marker pattern used by proxy-reference lookup (`search` included). */
+export declare const REFERENCE_TRAILING_MARKER_PATTERN: RegExp;

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

@@ -0,0 +1 @@
+export declare function buildModelProviderPriorityRank(configuredProviderOrder?: readonly string[]): Map<string, number>;

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

@@ -0,0 +1,29 @@
+/**
+ * Proxy/reseller reference lookup: given a custom model id served through a
+ * proxy (`[Kiro] claude-opus-4-8`, `gpt-5.4:cloud`, `vendor/claude-sonnet-4-6-thinking`),
+ * find the bundled upstream model so missing pricing/capability metadata can be
+ * inherited while keeping the custom transport.
+ *
+ * Kept separate from canonical-id resolution (`./equivalence`): this lookup
+ * may strip `search`-style markers and prefers cache-pricing-complete
+ * references, both of which would be wrong for canonical coalescing.
+ */
+import type { Api, Model, ThinkingConfig } from "../types.js";
+export interface ModelReferenceIndex {
+    exact: Map<string, Model<Api>>;
+    suffixAlias: Map<string, Model<Api>>;
+}
+export declare function isZeroCostXaiOAuthReference(candidate: Model<Api>): boolean;
+/**
+ * Build a reference index from a model catalog (typically the bundled models).
+ * Pure: callers are responsible for memoizing the result.
+ */
+export declare function buildModelReferenceIndex(models: Iterable<Model<Api>>): ModelReferenceIndex;
+/**
+ * Inherit bundled reference thinking only for same-provider matches. Wire routing
+ * (`effortRouting`) is provider-specific; cross-provider inheritance can rewrite
+ * gateway ids (e.g. Portkey `@modal/GLM-5-2-FP8` → devin `glm-5-2`).
+ */
+export declare function inheritReferenceThinking(modelThinking: ThinkingConfig | undefined, reference: Pick<Model<Api>, "provider" | "thinking"> | undefined, provider: string): ThinkingConfig | undefined;
+/** Resolve a (possibly proxied/affixed) model id to its bundled upstream reference. */
+export declare function resolveModelReference(modelId: string, index: ModelReferenceIndex): Model<Api> | undefined;

package/dist/types/index.d.ts

@@ -0,0 +1,17 @@
+export * from "./compat/openai.js";
+export * from "./discovery/index.js";
+export * from "./effort.js";
+export * from "./fireworks-model-id.js";
+export * from "./identity/index.js";
+export * from "./model-cache.js";
+export * from "./model-manager.js";
+export * from "./model-thinking.js";
+export * from "./models.js";
+export * from "./provider-models/index.js";
+export * from "./types.js";
+export * from "./utils.js";
+export * from "./variant-collapse.js";
+export * from "./wire/codex.js";
+export * from "./wire/coreweave.js";
+export * from "./wire/gemini-headers.js";
+export * from "./wire/github-copilot.js";

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

@@ -0,0 +1,23 @@
+import type { Api, Model, ModelSpec } from "./types.js";
+interface CacheEntry<TApi extends Api = Api> {
+    models: ModelSpec<TApi>[];
+    fresh: boolean;
+    authoritative: boolean;
+    updatedAt: number;
+    /** Model ids whose live headers were intentionally omitted from disk. */
+    headerOmittedModelIds: readonly string[];
+    /** Header-bearing model ids that cannot be rebuilt from the static source. */
+    unrestorableHeaderModelIds: readonly string[];
+    /** Whether unrestorable markers predate request-model header matching. */
+    legacyHeaderRestoreMarkers: boolean;
+    /**
+     * Hash of the static catalog slice that was merged into `models` when this
+     * row was written. `resolveProviderModels` compares against the current
+     * static fingerprint and bypasses the static+cache re-merge when they
+     * match — the cache already incorporates the same static state.
+     */
+    staticFingerprint: string;
+}
+export declare function readModelCache<TApi extends Api>(providerId: string, ttlMs: number, now: () => number, dbPath?: string): CacheEntry<TApi> | null;
+export declare function writeModelCache<TApi extends Api>(providerId: string, updatedAt: number, models: Model<TApi>[], authoritative: boolean, staticFingerprint: string, dbPath?: string, staticHeaderSources?: readonly Model<TApi>[], restorableHeaderFallback?: Record<string, string>): void;
+export {};