@kenkaiiii/gg-ai 5.80.1 → 5.82.0

package/dist/index.d.cts

@@ -223,6 +223,10 @@ interface StreamOptions {
     promptCacheKey?: string;
     /** OpenAI service tier for latency-sensitive requests. Only sent to first-party OpenAI API calls. */
     serviceTier?: "auto" | "default" | "flex" | "priority";
+    /** Provider speed tier. "fast" sends Anthropic `speed: "fast"` (first-party API
+     *  only) or OpenAI/Codex `service_tier: "priority"`. Set automatically by
+     *  `stream()` for fast-mode model ids (see `resolveFastModeModel`). */
+    speed?: "standard" | "fast";
     /** Codex endpoint only: send the Responses-Lite request shape (lite header,
      *  `parallel_tool_calls: false`, all-turns reasoning context). Unset follows
      *  the model family (on for gpt-5.6/gpt-6 models). Lite allows only one tool
@@ -430,7 +434,26 @@ declare function localWireModelId(id: string): string;
  * const response = await stream({ provider: "openai", model: "gpt-4.1", messages });
  * ```
  */
-declare function stream(options: StreamOptions): StreamResult;
+declare function stream(requested: StreamOptions): StreamResult;
+
+/**
+ * Fast-mode model ids: `<base model id>` + this suffix selects the same model
+ * at its provider's faster speed tier. The suffixed id is a registry/UI concept
+ * only — `stream()` rewrites it to the base id plus `speed: "fast"` before any
+ * provider sees it, so every model-id check downstream still matches.
+ */
+declare const FAST_MODE_SUFFIX = "-fast";
+interface FastModeResolution {
+    /** The real model id to send on the wire. */
+    model: string;
+    /** True when the id selected a fast speed tier. */
+    fast: boolean;
+}
+/**
+ * Split a fast-mode model id into its base model. Ids without the suffix, or
+ * whose base doesn't support fast mode for `provider`, pass through unchanged.
+ */
+declare function resolveFastModeModel(provider: Provider, model: string): FastModeResolution;
 
 /** What the embedded text will be used for. See `EmbedOptions.inputType`. */
 type EmbedInputType = "query" | "document";
@@ -1031,4 +1054,4 @@ declare function palsuEmbedding(text: string, dimensions?: number): number[];
 /** Deterministic fake relevance: the share of distinct query words found in the document (0–1). */
 declare function palsuRerankScore(query: string, document: string): number;
 
-export { type AssistantMessage, type CacheRetention, type ChatErrorReason, type ContentPart, type DoneEvent, type EmbedInputType, type EmbedOptions, type EmbedResult, type EmbedUsage, type ErrorEvent, type ErrorSource, EventStream, type FormattedChatError, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type OpenRouterProviderPreferences, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderCapability, type ProviderDiagnosticFn, type ProviderEmbedFn, type ProviderEmbedRequest, type ProviderEmbedResponse, type ProviderEntry, ProviderError, type ProviderOptions, type ProviderRerankFn, type ProviderRerankRequest, type ProviderRerankResponse, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type RerankOptions, type RerankResponse, type RerankResult, type RerankUsage, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, continuesTurn, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, rerank, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor, usesResponsesLite };
+export { type AssistantMessage, type CacheRetention, type ChatErrorReason, type ContentPart, type DoneEvent, type EmbedInputType, type EmbedOptions, type EmbedResult, type EmbedUsage, type ErrorEvent, type ErrorSource, EventStream, FAST_MODE_SUFFIX, type FastModeResolution, type FormattedChatError, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type OpenRouterProviderPreferences, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderCapability, type ProviderDiagnosticFn, type ProviderEmbedFn, type ProviderEmbedRequest, type ProviderEmbedResponse, type ProviderEntry, ProviderError, type ProviderOptions, type ProviderRerankFn, type ProviderRerankRequest, type ProviderRerankResponse, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type RerankOptions, type RerankResponse, type RerankResult, type RerankUsage, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, continuesTurn, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, rerank, resolveFastModeModel, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor, usesResponsesLite };

package/dist/index.d.ts

@@ -223,6 +223,10 @@ interface StreamOptions {
     promptCacheKey?: string;
     /** OpenAI service tier for latency-sensitive requests. Only sent to first-party OpenAI API calls. */
     serviceTier?: "auto" | "default" | "flex" | "priority";
+    /** Provider speed tier. "fast" sends Anthropic `speed: "fast"` (first-party API
+     *  only) or OpenAI/Codex `service_tier: "priority"`. Set automatically by
+     *  `stream()` for fast-mode model ids (see `resolveFastModeModel`). */
+    speed?: "standard" | "fast";
     /** Codex endpoint only: send the Responses-Lite request shape (lite header,
      *  `parallel_tool_calls: false`, all-turns reasoning context). Unset follows
      *  the model family (on for gpt-5.6/gpt-6 models). Lite allows only one tool
@@ -430,7 +434,26 @@ declare function localWireModelId(id: string): string;
  * const response = await stream({ provider: "openai", model: "gpt-4.1", messages });
  * ```
  */
-declare function stream(options: StreamOptions): StreamResult;
+declare function stream(requested: StreamOptions): StreamResult;
+
+/**
+ * Fast-mode model ids: `<base model id>` + this suffix selects the same model
+ * at its provider's faster speed tier. The suffixed id is a registry/UI concept
+ * only — `stream()` rewrites it to the base id plus `speed: "fast"` before any
+ * provider sees it, so every model-id check downstream still matches.
+ */
+declare const FAST_MODE_SUFFIX = "-fast";
+interface FastModeResolution {
+    /** The real model id to send on the wire. */
+    model: string;
+    /** True when the id selected a fast speed tier. */
+    fast: boolean;
+}
+/**
+ * Split a fast-mode model id into its base model. Ids without the suffix, or
+ * whose base doesn't support fast mode for `provider`, pass through unchanged.
+ */
+declare function resolveFastModeModel(provider: Provider, model: string): FastModeResolution;
 
 /** What the embedded text will be used for. See `EmbedOptions.inputType`. */
 type EmbedInputType = "query" | "document";
@@ -1031,4 +1054,4 @@ declare function palsuEmbedding(text: string, dimensions?: number): number[];
 /** Deterministic fake relevance: the share of distinct query words found in the document (0–1). */
 declare function palsuRerankScore(query: string, document: string): number;
 
-export { type AssistantMessage, type CacheRetention, type ChatErrorReason, type ContentPart, type DoneEvent, type EmbedInputType, type EmbedOptions, type EmbedResult, type EmbedUsage, type ErrorEvent, type ErrorSource, EventStream, type FormattedChatError, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type OpenRouterProviderPreferences, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderCapability, type ProviderDiagnosticFn, type ProviderEmbedFn, type ProviderEmbedRequest, type ProviderEmbedResponse, type ProviderEntry, ProviderError, type ProviderOptions, type ProviderRerankFn, type ProviderRerankRequest, type ProviderRerankResponse, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type RerankOptions, type RerankResponse, type RerankResult, type RerankUsage, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, continuesTurn, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, rerank, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor, usesResponsesLite };
+export { type AssistantMessage, type CacheRetention, type ChatErrorReason, type ContentPart, type DoneEvent, type EmbedInputType, type EmbedOptions, type EmbedResult, type EmbedUsage, type ErrorEvent, type ErrorSource, EventStream, FAST_MODE_SUFFIX, type FastModeResolution, type FormattedChatError, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type OpenRouterProviderPreferences, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderCapability, type ProviderDiagnosticFn, type ProviderEmbedFn, type ProviderEmbedRequest, type ProviderEmbedResponse, type ProviderEntry, ProviderError, type ProviderOptions, type ProviderRerankFn, type ProviderRerankRequest, type ProviderRerankResponse, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type RerankOptions, type RerankResponse, type RerankResult, type RerankUsage, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, continuesTurn, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, rerank, resolveFastModeModel, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor, usesResponsesLite };

package/dist/index.js

@@ -77,7 +77,8 @@ function isUsageLimitError(err) {
 }
 function isHardBillingMessage(message) {
   const lower = message.toLowerCase();
-  return lower.includes("insufficient balance") || lower.includes("insufficient credits") || lower.includes("more credits") || lower.includes("insufficient_quota") || lower.includes("exceeded your current quota") || lower.includes("quota exceeded") || lower.includes("no resource package") || lower.includes("recharge") || lower.includes("balance is too low") || lower.includes("out of credits") || lower.includes("arrears") || lower.includes("arrearage") || lower.includes("token quota") || lower.includes("exceeded_current_quota_error") || lower.includes("check your account balance") || lower.includes("does not yet include access") || lower.includes("subscription plan") || lower.includes("billing");
+  return lower.includes("insufficient balance") || lower.includes("insufficient credits") || lower.includes("more credits") || lower.includes("insufficient_quota") || lower.includes("exceeded your current quota") || lower.includes("quota exceeded") || lower.includes("no resource package") || lower.includes("recharge") || lower.includes("balance is too low") || lower.includes("out of credits") || // Anthropic fast mode on a subscription without usage credits turned on.
+  lower.includes("usage credits are required") || lower.includes("arrears") || lower.includes("arrearage") || lower.includes("token quota") || lower.includes("exceeded_current_quota_error") || lower.includes("check your account balance") || lower.includes("does not yet include access") || lower.includes("subscription plan") || lower.includes("billing");
 }
 function formatResetTime(resetsAt) {
   const when = new Date(resetsAt * 1e3);
@@ -1574,6 +1575,7 @@ function fineGrainedToolStreamingEnabled() {
   return v === "1" || v === "true" || v === "yes" || v === "on";
 }
 var SERVER_FALLBACK_BETA = "server-side-fallback-2026-07-01";
+var FAST_MODE_BETA = "fast-mode-2026-02-01";
 var anthropicServerFallback = {
   disabled: /* @__PURE__ */ new Set(),
   reset() {
@@ -1716,6 +1718,7 @@ async function* runStream(options) {
   const { system: rawSystem, messages } = toAnthropicMessages(downgradedMessages, cacheControl, {
     fallbackBlocks: useServerFallback
   });
+  const useFastMode = options.speed === "fast" && options.provider === "anthropic" && isDirectAnthropicApi(options.baseUrl);
   const system = isOAuth ? [
     {
       type: "text",
@@ -1786,6 +1789,7 @@ async function* runStream(options) {
       return contextEdits.length ? { context_management: { edits: contextEdits } } : {};
     })(),
     ...useServerFallback ? { fallbacks: "default" } : {},
+    ...useFastMode ? { speed: "fast" } : {},
     stream: useStreaming
   };
   const params = mergeProviderOptions(baseParams, options.providerOptions);
@@ -1804,7 +1808,8 @@ async function* runStream(options) {
     // so a pre-warmed cache expires before the user's first turn. cacheControl.ttl
     // is only "1h" on the first-party endpoint (see toAnthropicCacheControl).
     ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : [],
-    ...useServerFallback ? [SERVER_FALLBACK_BETA] : []
+    ...useServerFallback ? [SERVER_FALLBACK_BETA] : [],
+    ...useFastMode ? [FAST_MODE_BETA] : []
   ];
   const toRequestOptions = (betas) => ({
     signal: options.signal ?? void 0,
@@ -2189,6 +2194,9 @@ function messageToResponse(message) {
     }
   };
 }
+function isFastModeCreditsMessage(message) {
+  return /usage credits are required for fast mode/i.test(message);
+}
 function readUnifiedRateLimit(headers) {
   const status = readHeader(headers, "anthropic-ratelimit-unified-status");
   const resetRaw = readHeader(
@@ -2212,6 +2220,14 @@ function toError(err) {
     const fallbackMessage = isRawJsonErrorEcho(err.message) ? emptyProviderErrorMessage(err.status) : err.message;
     const messageCandidate = bodyMessage ?? err.message;
     const message = isRawHtmlErrorEcho(messageCandidate) ? providerHtmlErrorMessage(err.status) : bodyType && bodyMessage ? `${bodyType}: ${bodyMessage}` : bodyMessage ?? fallbackMessage;
+    if (isFastModeCreditsMessage(message)) {
+      return new ProviderError("anthropic", message, {
+        statusCode: err.status,
+        ...requestId ? { requestId } : {},
+        hint: "Fast mode bills Claude usage credits, not your plan. Turn them on at claude.ai/settings/usage, or switch to Claude Opus 5.5 (standard speed) in the model selector.",
+        cause: err
+      });
+    }
     if (err.status === 429) {
       const limit = readUnifiedRateLimit(err.headers);
       const farOff = limit.resetsAt != null && limit.resetsAt * 1e3 - Date.now() > 6e4;
@@ -2467,8 +2483,9 @@ async function* runStream2(options) {
   if (options.provider === "sakana" && options.model === "fugu-ultra" && (options.thinking === "max" || options.thinking === "ultra")) {
     params.reasoning_effort = "max";
   }
-  if (options.provider === "openai" && options.serviceTier) {
-    params.service_tier = options.serviceTier;
+  const serviceTier = options.serviceTier ?? (options.speed === "fast" ? "priority" : void 0);
+  if (options.provider === "openai" && serviceTier) {
+    params.service_tier = serviceTier;
   }
   if (isKimiK3) {
     const paramsAny = params;
@@ -2978,6 +2995,9 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
   if (options.temperature != null && !options.thinking) {
     body.temperature = options.temperature;
   }
+  if (options.speed === "fast") {
+    body.service_tier = "priority";
+  }
   body.reasoning = {
     // GPT-5.6/6 require at least low; older models still support thinking off.
     // Apply the floor here for every caller, including one-off prompt rewrites.
@@ -4713,6 +4733,18 @@ function observePreparedContext(before, after, tools) {
   };
 }
 
+// src/fast-mode.ts
+var FAST_MODE_SUFFIX = "-fast";
+var FAST_MODE_BASE_MODELS = {
+  anthropic: ["claude-opus-5-5"],
+  openai: ["gpt-6-astra", "gpt-6.1-sol", "gpt-6-luna"]
+};
+function resolveFastModeModel(provider, model) {
+  if (!model.endsWith(FAST_MODE_SUFFIX)) return { model, fast: false };
+  const base = model.slice(0, -FAST_MODE_SUFFIX.length);
+  return FAST_MODE_BASE_MODELS[provider]?.includes(base) ? { model: base, fast: true } : { model, fast: false };
+}
+
 // src/stream.ts
 var OPENAI_EMBED_BATCH_SIZE = 2048;
 var OPENROUTER_EMBED_BATCH_SIZE = 64;
@@ -4887,7 +4919,9 @@ function requireLocalBaseUrl(baseUrl) {
   }
   return baseUrl;
 }
-function stream(options) {
+function stream(requested) {
+  const fastMode = resolveFastModeModel(requested.provider, requested.model);
+  const options = fastMode.fast ? { ...requested, model: fastMode.model, speed: "fast" } : requested;
   const entry = providerRegistry.get(options.provider);
   if (!entry) {
     throw new GGAIError(
@@ -5518,6 +5552,7 @@ function palsuRerankScore(query, document) {
 }
 export {
   EventStream,
+  FAST_MODE_SUFFIX,
   GGAIError,
   ProviderError,
   REDACTED as REDACTION_MARKER,
@@ -5550,6 +5585,7 @@ export {
   redactValue,
   registerPalsuProvider,
   rerank,
+  resolveFastModeModel,
   resolveToolSchema,
   sanitizeMessagesForWire,
   setProviderDiagnostic,