@mariozechner/pi-ai 0.71.0 → 0.72.0

package/dist/types.d.ts

@@ -2,9 +2,11 @@ import type { AssistantMessageEventStream } from "./utils/event-stream.js";
 export type { AssistantMessageEventStream } from "./utils/event-stream.js";
 export type KnownApi = "openai-completions" | "mistral-conversations" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" | "anthropic-messages" | "bedrock-converse-stream" | "google-generative-ai" | "google-vertex";
 export type Api = KnownApi | (string & {});
-export type KnownProvider = "amazon-bedrock" | "anthropic" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway";
+export type KnownProvider = "amazon-bedrock" | "anthropic" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi";
 export type Provider = KnownProvider | string;
 export type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh";
+export type ModelThinkingLevel = "off" | ThinkingLevel;
+export type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;
 /** Token budgets for each thinking level (token-based providers only) */
 export interface ThinkingBudgets {
     minimal?: number;
@@ -13,7 +15,7 @@ export interface ThinkingBudgets {
     high?: number;
 }
 export type CacheRetention = "none" | "short" | "long";
-export type Transport = "sse" | "websocket" | "auto";
+export type Transport = "sse" | "websocket" | "websocket-cached" | "auto";
 export interface ProviderResponse {
     status: number;
     headers: Record<string, string>;
@@ -245,8 +247,6 @@ export interface OpenAICompletionsCompat {
     supportsDeveloperRole?: boolean;
     /** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */
     supportsReasoningEffort?: boolean;
-    /** Optional mapping from pi-ai reasoning levels to provider/model-specific `reasoning_effort` values. */
-    reasoningEffortMap?: Partial<Record<ThinkingLevel, string>>;
     /** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */
     supportsUsageInStreaming?: boolean;
     /** Which field to use for max tokens. Default: auto-detected from URL. */
@@ -382,6 +382,11 @@ export interface Model<TApi extends Api> {
     provider: Provider;
     baseUrl: string;
     reasoning: boolean;
+    /**
+     * Maps pi thinking levels to provider/model-specific values.
+     * Missing keys use provider defaults. null marks a level as unsupported.
+     */
+    thinkingLevelMap?: ThinkingLevelMap;
     input: ("text" | "image")[];
     cost: {
         input: number;

package/dist/types.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAE3E,YAAY,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAE3E,MAAM,MAAM,QAAQ,GACjB,oBAAoB,GACpB,uBAAuB,GACvB,kBAAkB,GAClB,wBAAwB,GACxB,wBAAwB,GACxB,oBAAoB,GACpB,yBAAyB,GACzB,sBAAsB,GACtB,eAAe,CAAC;AAEnB,MAAM,MAAM,GAAG,GAAG,QAAQ,GAAG,CAAC,MAAM,GAAG,EAAE,CAAC,CAAC;AAE3C,MAAM,MAAM,aAAa,GACtB,gBAAgB,GAChB,WAAW,GACX,QAAQ,GACR,eAAe,GACf,QAAQ,GACR,wBAAwB,GACxB,cAAc,GACd,UAAU,GACV,gBAAgB,GAChB,KAAK,GACL,MAAM,GACN,UAAU,GACV,YAAY,GACZ,mBAAmB,GACnB,KAAK,GACL,SAAS,GACT,SAAS,GACT,YAAY,GACZ,YAAY,GACZ,eAAe,GACf,aAAa,GACb,WAAW,GACX,UAAU,GACV,aAAa,GACb,aAAa,GACb,uBAAuB,GACvB,uBAAuB,CAAC;AAC3B,MAAM,MAAM,QAAQ,GAAG,aAAa,GAAG,MAAM,CAAC;AAE9C,MAAM,MAAM,aAAa,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,CAAC;AAE5E,yEAAyE;AACzE,MAAM,WAAW,eAAe;IAC/B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,IAAI,CAAC,EAAE,MAAM,CAAC;CACd;AAGD,MAAM,MAAM,cAAc,GAAG,MAAM,GAAG,OAAO,GAAG,MAAM,CAAC;AAEvD,MAAM,MAAM,SAAS,GAAG,KAAK,GAAG,WAAW,GAAG,MAAM,CAAC;AAErD,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAChC;AAED,MAAM,WAAW,aAAa;IAC7B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;OAGG;IACH,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB;;;OAGG;IACH,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC;;;;OAIG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,SAAS,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,KAAK,OAAO,GAAG,SAAS,GAAG,OAAO,CAAC,OAAO,GAAG,SAAS,CAAC,CAAC;IACxG;;;OAGG;IACH,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,gBAAgB,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACrF;;;;OAIG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACnC;AAED,MAAM,MAAM,qBAAqB,GAAG,aAAa,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AAG5E,MAAM,WAAW,mBAAoB,SAAQ,aAAa;IACzD,SAAS,CAAC,EAAE,aAAa,CAAC;IAC1B,4EAA4E;IAC5E,eAAe,CAAC,EAAE,eAAe,CAAC;CAClC;AAUD,MAAM,MAAM,cAAc,CAAC,IAAI,SAAS,GAAG,GAAG,GAAG,EAAE,QAAQ,SAAS,aAAa,GAAG,aAAa,IAAI,CACpG,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,QAAQ,KACd,2BAA2B,CAAC;AAEjC,MAAM,WAAW,eAAe;IAC/B,CAAC,EAAE,CAAC,CAAC;IACL,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,CAAC,EAAE,YAAY,GAAG,cAAc,CAAC;CACtC;AAED,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,aAAa,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC/B,IAAI,EAAE,UAAU,CAAC;IACjB,QAAQ,EAAE,MAAM,CAAC;IACjB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B;;gDAE4C;IAC5C,QAAQ,CAAC,EAAE,OAAO,CAAC;CACnB;AAED,MAAM,WAAW,YAAY;IAC5B,IAAI,EAAE,OAAO,CAAC;IACd,IAAI,EAAE,MAAM,CAAC;IACb,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,QAAQ;IACxB,IAAI,EAAE,UAAU,CAAC;IACjB,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;IAC/B,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,KAAK;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE;QACL,KAAK,EAAE,MAAM,CAAC;QACd,MAAM,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;QACnB,KAAK,EAAE,MAAM,CAAC;KACd,CAAC;CACF;AAED,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,QAAQ,GAAG,SAAS,GAAG,OAAO,GAAG,SAAS,CAAC;AAE7E,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,GAAG,CAAC,WAAW,GAAG,YAAY,CAAC,EAAE,CAAC;IACjD,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,gBAAgB;IAChC,IAAI,EAAE,WAAW,CAAC;IAClB,OAAO,EAAE,CAAC,WAAW,GAAG,eAAe,GAAG,QAAQ,CAAC,EAAE,CAAC;IACtD,GAAG,EAAE,GAAG,CAAC;IACT,QAAQ,EAAE,QAAQ,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,KAAK,EAAE,KAAK,CAAC;IACb,UAAU,EAAE,UAAU,CAAC;IACvB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,iBAAiB,CAAC,QAAQ,GAAG,GAAG;IAChD,IAAI,EAAE,YAAY,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,OAAO,EAAE,CAAC,WAAW,GAAG,YAAY,CAAC,EAAE,CAAC;IACxC,OAAO,CAAC,EAAE,QAAQ,CAAC;IACnB,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,MAAM,OAAO,GAAG,WAAW,GAAG,gBAAgB,GAAG,iBAAiB,CAAC;AAEzE,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,SAAS,CAAC;AAEvC,MAAM,WAAW,IAAI,CAAC,WAAW,SAAS,OAAO,GAAG,OAAO;IAC1D,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,WAAW,CAAC;CACxB;AAED,MAAM,WAAW,OAAO;IACvB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,QAAQ,EAAE,OAAO,EAAE,CAAC;IACpB,KAAK,CAAC,EAAE,IAAI,EAAE,CAAC;CACf;AAED;;;;;;;GAOG;AACH,MAAM,MAAM,qBAAqB,GAC9B;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC5C;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACvE;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACtF;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACtF;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC3E;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,cAAc,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC3E;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,cAAc,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,QAAQ,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC7F;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,OAAO,CAAC,UAAU,EAAE,MAAM,GAAG,QAAQ,GAAG,SAAS,CAAC,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACvG;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,OAAO,CAAC,UAAU,EAAE,SAAS,GAAG,OAAO,CAAC,CAAC;IAAC,KAAK,EAAE,gBAAgB,CAAA;CAAE,CAAC;AAEhG;;;GAGG;AACH,MAAM,WAAW,uBAAuB;IACvC,wFAAwF;IACxF,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,yGAAyG;IACzG,qBAAqB,CAAC,EAAE,OAAO,CAAC;IAChC,yFAAyF;IACzF,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,yGAAyG;IACzG,kBAAkB,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,aAAa,EAAE,MAAM,CAAC,CAAC,CAAC;IAC5D,qIAAqI;IACrI,wBAAwB,CAAC,EAAE,OAAO,CAAC;IACnC,0EAA0E;IAC1E,cAAc,CAAC,EAAE,uBAAuB,GAAG,YAAY,CAAC;IACxD,sFAAsF;IACtF,sBAAsB,CAAC,EAAE,OAAO,CAAC;IACjC,2HAA2H;IAC3H,gCAAgC,CAAC,EAAE,OAAO,CAAC;IAC3C,4HAA4H;IAC5H,sBAAsB,CAAC,EAAE,OAAO,CAAC;IACjC,wJAAwJ;IACxJ,2CAA2C,CAAC,EAAE,OAAO,CAAC;IACtD,wWAAwW;IACxW,cAAc,CAAC,EAAE,QAAQ,GAAG,YAAY,GAAG,UAAU,GAAG,KAAK,GAAG,MAAM,GAAG,oBAAoB,CAAC;IAC9F,4FAA4F;IAC5F,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;IACtC,iGAAiG;IACjG,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,0GAA0G;IAC1G,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,2FAA2F;IAC3F,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,iMAAiM;IACjM,kBAAkB,CAAC,EAAE,WAAW,CAAC;IACjC,mLAAmL;IACnL,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,oLAAoL;IACpL,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,wDAAwD;AACxD,MAAM,WAAW,qBAAqB;IACrC,qIAAqI;IACrI,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,oFAAoF;IACpF,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,qEAAqE;AACrE,MAAM,WAAW,uBAAuB;IACvC;;;;;;OAMG;IACH,+BAA+B,CAAC,EAAE,OAAO,CAAC;IAC1C,+GAA+G;IAC/G,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED;;;;;GAKG;AACH,MAAM,WAAW,iBAAiB;IACjC,0EAA0E;IAC1E,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,4GAA4G;IAC5G,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,yJAAyJ;IACzJ,eAAe,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACnC,+EAA+E;IAC/E,GAAG,CAAC,EAAE,OAAO,CAAC;IACd,+EAA+E;IAC/E,wBAAwB,CAAC,EAAE,OAAO,CAAC;IACnC,2GAA2G;IAC3G,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;IACjB,0EAA0E;IAC1E,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,kIAAkI;IAClI,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,yHAAyH;IACzH,IAAI,CAAC,EACF,MAAM,GACN;QACA,4DAA4D;QAC5D,EAAE,CAAC,EAAE,MAAM,CAAC;QACZ,0DAA0D;QAC1D,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KACzB,CAAC;IACL,8CAA8C;IAC9C,SAAS,CAAC,EAAE;QACX,uCAAuC;QACvC,MAAM,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACzB,2CAA2C;QAC3C,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QAC7B,uBAAuB;QACvB,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACxB,4BAA4B;QAC5B,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACxB,yBAAyB;QACzB,OAAO,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;KAC1B,CAAC;IACF,oIAAoI;IACpI,wBAAwB,CAAC,EACtB,MAAM,GACN;QACA,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;KACZ,CAAC;IACL,2HAA2H;IAC3H,qBAAqB,CAAC,EACnB,MAAM,GACN;QACA,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;KACZ,CAAC;CACL;AAED;;;;GAIG;AACH,MAAM,WAAW,oBAAoB;IACpC,mGAAmG;IACnG,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,8EAA8E;IAC9E,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;CACjB;AAGD,MAAM,WAAW,KAAK,CAAC,IAAI,SAAS,GAAG;IACtC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,GAAG,EAAE,IAAI,CAAC;IACV,QAAQ,EAAE,QAAQ,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,OAAO,CAAC;IACnB,KAAK,EAAE,CAAC,MAAM,GAAG,OAAO,CAAC,EAAE,CAAC;IAC5B,IAAI,EAAE;QACL,KAAK,EAAE,MAAM,CAAC;QACd,MAAM,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;KACnB,CAAC;IACF,aAAa,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,kGAAkG;IAClG,MAAM,CAAC,EAAE,IAAI,SAAS,oBAAoB,GACvC,uBAAuB,GACvB,IAAI,SAAS,kBAAkB,GAC9B,qBAAqB,GACrB,IAAI,SAAS,oBAAoB,GAChC,uBAAuB,GACvB,KAAK,CAAC;CACX","sourcesContent":["import type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\";\nexport type Provider = KnownProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Optional mapping from pi-ai reasoning levels to provider/model-specific `reasoning_effort` values. */\n\treasoningEffortMap?: Partial<Record<ThinkingLevel, string>>;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n}\n"]}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAE3E,YAAY,EAAE,2BAA2B,EAAE,MAAM,yBAAyB,CAAC;AAE3E,MAAM,MAAM,QAAQ,GACjB,oBAAoB,GACpB,uBAAuB,GACvB,kBAAkB,GAClB,wBAAwB,GACxB,wBAAwB,GACxB,oBAAoB,GACpB,yBAAyB,GACzB,sBAAsB,GACtB,eAAe,CAAC;AAEnB,MAAM,MAAM,GAAG,GAAG,QAAQ,GAAG,CAAC,MAAM,GAAG,EAAE,CAAC,CAAC;AAE3C,MAAM,MAAM,aAAa,GACtB,gBAAgB,GAChB,WAAW,GACX,QAAQ,GACR,eAAe,GACf,QAAQ,GACR,wBAAwB,GACxB,cAAc,GACd,UAAU,GACV,gBAAgB,GAChB,KAAK,GACL,MAAM,GACN,UAAU,GACV,YAAY,GACZ,mBAAmB,GACnB,KAAK,GACL,SAAS,GACT,SAAS,GACT,YAAY,GACZ,YAAY,GACZ,eAAe,GACf,aAAa,GACb,WAAW,GACX,UAAU,GACV,aAAa,GACb,aAAa,GACb,uBAAuB,GACvB,uBAAuB,GACvB,QAAQ,CAAC;AACZ,MAAM,MAAM,QAAQ,GAAG,aAAa,GAAG,MAAM,CAAC;AAE9C,MAAM,MAAM,aAAa,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,CAAC;AAC5E,MAAM,MAAM,kBAAkB,GAAG,KAAK,GAAG,aAAa,CAAC;AACvD,MAAM,MAAM,gBAAgB,GAAG,OAAO,CAAC,MAAM,CAAC,kBAAkB,EAAE,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC;AAElF,yEAAyE;AACzE,MAAM,WAAW,eAAe;IAC/B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,GAAG,CAAC,EAAE,MAAM,CAAC;IACb,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,IAAI,CAAC,EAAE,MAAM,CAAC;CACd;AAGD,MAAM,MAAM,cAAc,GAAG,MAAM,GAAG,OAAO,GAAG,MAAM,CAAC;AAEvD,MAAM,MAAM,SAAS,GAAG,KAAK,GAAG,WAAW,GAAG,kBAAkB,GAAG,MAAM,CAAC;AAE1E,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAChC;AAED,MAAM,WAAW,aAAa;IAC7B,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;OAGG;IACH,SAAS,CAAC,EAAE,SAAS,CAAC;IACtB;;;OAGG;IACH,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC;;;;OAIG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,SAAS,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,KAAK,OAAO,GAAG,SAAS,GAAG,OAAO,CAAC,OAAO,GAAG,SAAS,CAAC,CAAC;IACxG;;;OAGG;IACH,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,gBAAgB,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,KAAK,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACrF;;;;OAIG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC;;;OAGG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACnC;AAED,MAAM,MAAM,qBAAqB,GAAG,aAAa,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AAG5E,MAAM,WAAW,mBAAoB,SAAQ,aAAa;IACzD,SAAS,CAAC,EAAE,aAAa,CAAC;IAC1B,4EAA4E;IAC5E,eAAe,CAAC,EAAE,eAAe,CAAC;CAClC;AAUD,MAAM,MAAM,cAAc,CAAC,IAAI,SAAS,GAAG,GAAG,GAAG,EAAE,QAAQ,SAAS,aAAa,GAAG,aAAa,IAAI,CACpG,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAClB,OAAO,EAAE,OAAO,EAChB,OAAO,CAAC,EAAE,QAAQ,KACd,2BAA2B,CAAC;AAEjC,MAAM,WAAW,eAAe;IAC/B,CAAC,EAAE,CAAC,CAAC;IACL,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,CAAC,EAAE,YAAY,GAAG,cAAc,CAAC;CACtC;AAED,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,aAAa,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC/B,IAAI,EAAE,UAAU,CAAC;IACjB,QAAQ,EAAE,MAAM,CAAC;IACjB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B;;gDAE4C;IAC5C,QAAQ,CAAC,EAAE,OAAO,CAAC;CACnB;AAED,MAAM,WAAW,YAAY;IAC5B,IAAI,EAAE,OAAO,CAAC;IACd,IAAI,EAAE,MAAM,CAAC;IACb,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,QAAQ;IACxB,IAAI,EAAE,UAAU,CAAC;IACjB,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;IAC/B,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,KAAK;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE;QACL,KAAK,EAAE,MAAM,CAAC;QACd,MAAM,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;QACnB,KAAK,EAAE,MAAM,CAAC;KACd,CAAC;CACF;AAED,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,QAAQ,GAAG,SAAS,GAAG,OAAO,GAAG,SAAS,CAAC;AAE7E,MAAM,WAAW,WAAW;IAC3B,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,GAAG,CAAC,WAAW,GAAG,YAAY,CAAC,EAAE,CAAC;IACjD,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,gBAAgB;IAChC,IAAI,EAAE,WAAW,CAAC;IAClB,OAAO,EAAE,CAAC,WAAW,GAAG,eAAe,GAAG,QAAQ,CAAC,EAAE,CAAC;IACtD,GAAG,EAAE,GAAG,CAAC;IACT,QAAQ,EAAE,QAAQ,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,KAAK,EAAE,KAAK,CAAC;IACb,UAAU,EAAE,UAAU,CAAC;IACvB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,iBAAiB,CAAC,QAAQ,GAAG,GAAG;IAChD,IAAI,EAAE,YAAY,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,QAAQ,EAAE,MAAM,CAAC;IACjB,OAAO,EAAE,CAAC,WAAW,GAAG,YAAY,CAAC,EAAE,CAAC;IACxC,OAAO,CAAC,EAAE,QAAQ,CAAC;IACnB,OAAO,EAAE,OAAO,CAAC;IACjB,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,MAAM,OAAO,GAAG,WAAW,GAAG,gBAAgB,GAAG,iBAAiB,CAAC;AAEzE,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,SAAS,CAAC;AAEvC,MAAM,WAAW,IAAI,CAAC,WAAW,SAAS,OAAO,GAAG,OAAO;IAC1D,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,WAAW,CAAC;CACxB;AAED,MAAM,WAAW,OAAO;IACvB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,QAAQ,EAAE,OAAO,EAAE,CAAC;IACpB,KAAK,CAAC,EAAE,IAAI,EAAE,CAAC;CACf;AAED;;;;;;;GAOG;AACH,MAAM,MAAM,qBAAqB,GAC9B;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC5C;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACvE;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACtF;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACtF;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC3E;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,cAAc,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC3E;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,cAAc,CAAC;IAAC,YAAY,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,QAAQ,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GAC7F;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,OAAO,CAAC,UAAU,EAAE,MAAM,GAAG,QAAQ,GAAG,SAAS,CAAC,CAAC;IAAC,OAAO,EAAE,gBAAgB,CAAA;CAAE,GACvG;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,OAAO,CAAC,UAAU,EAAE,SAAS,GAAG,OAAO,CAAC,CAAC;IAAC,KAAK,EAAE,gBAAgB,CAAA;CAAE,CAAC;AAEhG;;;GAGG;AACH,MAAM,WAAW,uBAAuB;IACvC,wFAAwF;IACxF,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,yGAAyG;IACzG,qBAAqB,CAAC,EAAE,OAAO,CAAC;IAChC,yFAAyF;IACzF,uBAAuB,CAAC,EAAE,OAAO,CAAC;IAClC,qIAAqI;IACrI,wBAAwB,CAAC,EAAE,OAAO,CAAC;IACnC,0EAA0E;IAC1E,cAAc,CAAC,EAAE,uBAAuB,GAAG,YAAY,CAAC;IACxD,sFAAsF;IACtF,sBAAsB,CAAC,EAAE,OAAO,CAAC;IACjC,2HAA2H;IAC3H,gCAAgC,CAAC,EAAE,OAAO,CAAC;IAC3C,4HAA4H;IAC5H,sBAAsB,CAAC,EAAE,OAAO,CAAC;IACjC,wJAAwJ;IACxJ,2CAA2C,CAAC,EAAE,OAAO,CAAC;IACtD,wWAAwW;IACxW,cAAc,CAAC,EAAE,QAAQ,GAAG,YAAY,GAAG,UAAU,GAAG,KAAK,GAAG,MAAM,GAAG,oBAAoB,CAAC;IAC9F,4FAA4F;IAC5F,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;IACtC,iGAAiG;IACjG,oBAAoB,CAAC,EAAE,oBAAoB,CAAC;IAC5C,0GAA0G;IAC1G,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,2FAA2F;IAC3F,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,iMAAiM;IACjM,kBAAkB,CAAC,EAAE,WAAW,CAAC;IACjC,mLAAmL;IACnL,0BAA0B,CAAC,EAAE,OAAO,CAAC;IACrC,oLAAoL;IACpL,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,wDAAwD;AACxD,MAAM,WAAW,qBAAqB;IACrC,qIAAqI;IACrI,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,oFAAoF;IACpF,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED,qEAAqE;AACrE,MAAM,WAAW,uBAAuB;IACvC;;;;;;OAMG;IACH,+BAA+B,CAAC,EAAE,OAAO,CAAC;IAC1C,+GAA+G;IAC/G,0BAA0B,CAAC,EAAE,OAAO,CAAC;CACrC;AAED;;;;;GAKG;AACH,MAAM,WAAW,iBAAiB;IACjC,0EAA0E;IAC1E,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,4GAA4G;IAC5G,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B,yJAAyJ;IACzJ,eAAe,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;IACnC,+EAA+E;IAC/E,GAAG,CAAC,EAAE,OAAO,CAAC;IACd,+EAA+E;IAC/E,wBAAwB,CAAC,EAAE,OAAO,CAAC;IACnC,2GAA2G;IAC3G,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;IACjB,0EAA0E;IAC1E,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,6DAA6D;IAC7D,MAAM,CAAC,EAAE,MAAM,EAAE,CAAC;IAClB,kIAAkI;IAClI,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,yHAAyH;IACzH,IAAI,CAAC,EACF,MAAM,GACN;QACA,4DAA4D;QAC5D,EAAE,CAAC,EAAE,MAAM,CAAC;QACZ,0DAA0D;QAC1D,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KACzB,CAAC;IACL,8CAA8C;IAC9C,SAAS,CAAC,EAAE;QACX,uCAAuC;QACvC,MAAM,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACzB,2CAA2C;QAC3C,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QAC7B,uBAAuB;QACvB,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACxB,4BAA4B;QAC5B,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;QACxB,yBAAyB;QACzB,OAAO,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;KAC1B,CAAC;IACF,oIAAoI;IACpI,wBAAwB,CAAC,EACtB,MAAM,GACN;QACA,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,oDAAoD;QACpD,GAAG,CAAC,EAAE,MAAM,CAAC;KACZ,CAAC;IACL,2HAA2H;IAC3H,qBAAqB,CAAC,EACnB,MAAM,GACN;QACA,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;QACb,yDAAyD;QACzD,GAAG,CAAC,EAAE,MAAM,CAAC;KACZ,CAAC;CACL;AAED;;;;GAIG;AACH,MAAM,WAAW,oBAAoB;IACpC,mGAAmG;IACnG,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;IAChB,8EAA8E;IAC9E,KAAK,CAAC,EAAE,MAAM,EAAE,CAAC;CACjB;AAGD,MAAM,WAAW,KAAK,CAAC,IAAI,SAAS,GAAG;IACtC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,GAAG,EAAE,IAAI,CAAC;IACV,QAAQ,EAAE,QAAQ,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,OAAO,CAAC;IACnB;;;OAGG;IACH,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,KAAK,EAAE,CAAC,MAAM,GAAG,OAAO,CAAC,EAAE,CAAC;IAC5B,IAAI,EAAE;QACL,KAAK,EAAE,MAAM,CAAC;QACd,MAAM,EAAE,MAAM,CAAC;QACf,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;KACnB,CAAC;IACF,aAAa,EAAE,MAAM,CAAC;IACtB,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,kGAAkG;IAClG,MAAM,CAAC,EAAE,IAAI,SAAS,oBAAoB,GACvC,uBAAuB,GACvB,IAAI,SAAS,kBAAkB,GAC9B,qBAAqB,GACrB,IAAI,SAAS,oBAAoB,GAChC,uBAAuB,GACvB,KAAK,CAAC;CACX","sourcesContent":["import type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"xiaomi\";\nexport type Provider = KnownProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n}\n"]}

package/dist/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\";\nexport type Provider = KnownProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Optional mapping from pi-ai reasoning levels to provider/model-specific `reasoning_effort` values. */\n\treasoningEffortMap?: Partial<Record<ThinkingLevel, string>>;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n}\n"]}
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"xiaomi\";\nexport type Provider = KnownProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n}\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"toolcall_end\"; contentIndex: number; toolCall: ToolCall; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n}\n"]}

package/dist/utils/overflow.d.ts

@@ -26,6 +26,8 @@ import type { AssistantMessage } from "../types.js";
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
  *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
  * - Ollama: May truncate input silently for some setups, but may also return explicit
  *   overflow errors that match the patterns above. Silent truncation still cannot be
  *   detected here because we do not know the expected token count.

package/dist/utils/overflow.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAiEpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6CG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CAmB5F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAoEpD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,gBAAgB,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,OAAO,CA6B5F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,IAAI,MAAM,EAAE,CAE9C","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/dist/utils/overflow.js

@@ -21,6 +21,9 @@
  * - Cerebras: "400/413 status code (no body)"
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
  * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
+ *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
+ *   input filling the context window.
  * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
  */
 const OVERFLOW_PATTERNS = [
@@ -86,6 +89,8 @@ const NON_OVERFLOW_PATTERNS = [
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
  *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
  * - Ollama: May truncate input silently for some setups, but may also return explicit
  *   overflow errors that match the patterns above. Silent truncation still cannot be
  *   detected here because we do not know the expected token count.
@@ -121,6 +126,15 @@ export function isContextOverflow(message, contextWindow) {
             return true;
         }
     }
+    // Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input
+    // to fit the context window, leaving no room for output. Returns stopReason "length"
+    // with output=0 and input+cacheRead filling the context window.
+    if (contextWindow && message.stopReason === "length" && message.usage.output === 0) {
+        const inputTokens = message.usage.input + message.usage.cacheRead;
+        if (inputTokens >= contextWindow * 0.99) {
+            return true;
+        }
+    }
     return false;
 }
 /**

package/dist/utils/overflow.js.map

@@ -1 +1 @@
-{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6CG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AAAA,CACb;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,GAAa;IAC/C,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAAA,CAC9B","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../src/utils/overflow.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,MAAM,iBAAiB,GAAG;IACzB,qBAAqB,EAAE,2BAA2B;IAClD,oBAAoB,EAAE,kDAAkD;IACxE,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,uCAAuC;IACtE,yCAAyC,EAAE,kBAAkB;IAC7D,+BAA+B,EAAE,aAAa;IAC9C,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,4BAA4B;IACrE,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,mBAAmB;IAC1D,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,kBAAkB;IACjD,sDAAsD,EAAE,UAAU;IAClE,gCAAgC,EAAE,yDAAyD;IAC3F,oDAAoD,EAAE,iCAAiC;IACvF,gCAAgC,EAAE,mBAAmB;IACrD,kBAAkB,EAAE,mBAAmB;IACvC,uBAAuB,EAAE,mBAAmB;IAC5C,+CAA+C,EAAE,iCAAiC;CAClF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,qBAAqB,GAAG;IAC7B,2CAA2C,EAAE,oFAAoF;IACjI,aAAa,EAAE,wBAAwB;IACvC,oBAAoB,EAAE,yBAAyB;CAC/C,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAAW;IAC7F,uCAAuC;IACvC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;QAC5D,oFAAoF;QACpF,MAAM,aAAa,GAAG,qBAAqB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,aAAa,IAAI,iBAAiB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,YAAa,CAAC,CAAC,EAAE,CAAC;YACpF,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,8EAA8E;IAC9E,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,MAAM,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,GAAG,aAAa,EAAE,CAAC;YACjC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,sFAAsF;IACtF,qFAAqF;IACrF,gEAAgE;IAChE,IAAI,aAAa,IAAI,OAAO,CAAC,UAAU,KAAK,QAAQ,IAAI,OAAO,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACpF,MAAM,WAAW,GAAG,OAAO,CAAC,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,SAAS,CAAC;QAClE,IAAI,WAAW,IAAI,aAAa,GAAG,IAAI,EAAE,CAAC;YACzC,OAAO,IAAI,CAAC;QACb,CAAC;IACF,CAAC;IAED,OAAO,KAAK,CAAC;AAAA,CACb;AAED;;GAEG;AACH,MAAM,UAAU,mBAAmB,GAAa;IAC/C,OAAO,CAAC,GAAG,iBAAiB,CAAC,CAAC;AAAA,CAC9B","sourcesContent":["import type { AssistantMessage } from \"../types.js\";\n\n/**\n * Regex patterns to detect context overflow errors from different providers.\n *\n * These patterns match error messages returned when the input exceeds\n * the model's context window.\n *\n * Provider-specific patterns (with example error messages):\n *\n * - Anthropic: \"prompt is too long: 213462 tokens > 200000 maximum\"\n * - Anthropic: \"413 {\\\"error\\\":{\\\"type\\\":\\\"request_too_large\\\",\\\"message\\\":\\\"Request exceeds the maximum size\\\"}}\"\n * - OpenAI: \"Your input exceeds the context window of this model\"\n * - Google: \"The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)\"\n * - xAI: \"This model's maximum prompt length is 131072 but the request contains 537812 tokens\"\n * - Groq: \"Please reduce the length of the messages or completion\"\n * - OpenRouter: \"This endpoint's maximum context length is X tokens. However, you requested about Y tokens\"\n * - llama.cpp: \"the request exceeds the available context size, try increasing it\"\n * - LM Studio: \"tokens to keep from the initial prompt is greater than the context length\"\n * - GitHub Copilot: \"prompt token count of X exceeds the limit of Y\"\n * - MiniMax: \"invalid params, context window exceeds limit\"\n * - Kimi For Coding: \"Your request exceeded model token limit: X (requested: Y)\"\n * - Cerebras: \"400/413 status code (no body)\"\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow\n * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason \"length\"\n *   with output=0 (no room left to generate). Detected via stopReason \"length\" + zero output +\n *   input filling the context window.\n * - Ollama: Some deployments truncate silently, others return errors like \"prompt too long; exceeded max context length by X tokens\"\n */\nconst OVERFLOW_PATTERNS = [\n\t/prompt is too long/i, // Anthropic token overflow\n\t/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)\n\t/input is too long for requested model/i, // Amazon Bedrock\n\t/exceeds the context window/i, // OpenAI (Completions & Responses API)\n\t/input token count.*exceeds the maximum/i, // Google (Gemini)\n\t/maximum prompt length is \\d+/i, // xAI (Grok)\n\t/reduce the length of the messages/i, // Groq\n\t/maximum context length is \\d+ tokens/i, // OpenRouter (all backends)\n\t/exceeds the limit of \\d+/i, // GitHub Copilot\n\t/exceeds the available context size/i, // llama.cpp server\n\t/greater than the context length/i, // LM Studio\n\t/context window exceeds limit/i, // MiniMax\n\t/exceeded model token limit/i, // Kimi For Coding\n\t/too large for model with \\d+ maximum context length/i, // Mistral\n\t/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text\n\t/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error\n\t/context[_ ]length[_ ]exceeded/i, // Generic fallback\n\t/too many tokens/i, // Generic fallback\n\t/token limit exceeded/i, // Generic fallback\n\t/^4(?:00|13)\\s*(?:status code)?\\s*\\(no body\\)/i, // Cerebras: 400/413 with no body\n];\n\n/**\n * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).\n * Error messages matching any of these are excluded from overflow detection\n * even if they also match an OVERFLOW_PATTERN.\n *\n * Example: Bedrock formats throttling errors as \"ThrottlingException: Too many tokens,\n * please wait before trying again.\" which would match the /too many tokens/i overflow\n * pattern without this exclusion.\n */\nconst NON_OVERFLOW_PATTERNS = [\n\t/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)\n\t/rate limit/i, // Generic rate limiting\n\t/too many requests/i, // Generic HTTP 429 style\n];\n\n/**\n * Check if an assistant message represents a context overflow error.\n *\n * This handles two cases:\n * 1. Error-based overflow: Most providers return stopReason \"error\" with a\n *    specific error message pattern.\n * 2. Silent overflow: Some providers accept overflow requests and return\n *    successfully. For these, we check if usage.input exceeds the context window.\n *\n * ## Reliability by Provider\n *\n * **Reliable detection (returns error with detectable message):**\n * - Anthropic: \"prompt is too long: X tokens > Y maximum\" or \"request_too_large\"\n * - OpenAI (Completions & Responses): \"exceeds the context window\"\n * - Google Gemini: \"input token count exceeds the maximum\"\n * - xAI (Grok): \"maximum prompt length is X but request contains Y\"\n * - Groq: \"reduce the length of the messages\"\n * - Cerebras: 400/413 status code (no body)\n * - Mistral: \"Prompt contains X tokens ... too large for model with Y maximum context length\"\n * - OpenRouter (all backends): \"maximum context length is X tokens\"\n * - llama.cpp: \"exceeds the available context size\"\n * - LM Studio: \"greater than the context length\"\n * - Kimi For Coding: \"exceeded model token limit: X (requested: Y)\"\n *\n * **Unreliable detection:**\n * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),\n *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.\n * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason \"length\" with\n *   output=0. Pass contextWindow param to detect via the \"filled context + zero output\" signal.\n * - Ollama: May truncate input silently for some setups, but may also return explicit\n *   overflow errors that match the patterns above. Silent truncation still cannot be\n *   detected here because we do not know the expected token count.\n *\n * ## Custom Providers\n *\n * If you've added custom models via settings.json, this function may not detect\n * overflow errors from those providers. To add support:\n *\n * 1. Send a request that exceeds the model's context window\n * 2. Check the errorMessage in the response\n * 3. Create a regex pattern that matches the error\n * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or\n *    check the errorMessage yourself before calling this function\n *\n * @param message - The assistant message to check\n * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)\n * @returns true if the message indicates a context overflow\n */\nexport function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {\n\t// Case 1: Check error message patterns\n\tif (message.stopReason === \"error\" && message.errorMessage) {\n\t\t// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)\n\t\tconst isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));\n\t\tif (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context\n\tif (contextWindow && message.stopReason === \"stop\") {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens > contextWindow) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\t// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input\n\t// to fit the context window, leaving no room for output. Returns stopReason \"length\"\n\t// with output=0 and input+cacheRead filling the context window.\n\tif (contextWindow && message.stopReason === \"length\" && message.usage.output === 0) {\n\t\tconst inputTokens = message.usage.input + message.usage.cacheRead;\n\t\tif (inputTokens >= contextWindow * 0.99) {\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\n/**\n * Get the overflow patterns for testing purposes.\n */\nexport function getOverflowPatterns(): RegExp[] {\n\treturn [...OVERFLOW_PATTERNS];\n}\n"]}

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@mariozechner/pi-ai",
-	"version": "0.71.0",
+	"version": "0.72.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",