@kenkaiiii/gg-ai 5.81.0 → 5.83.0

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 +434,43 @@ 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;
+interface FastModeRejection {
+    /** Anthropic's reason code (`unknown` when it sent none). */
+    reason: string;
+    /** Short, factual headline for the error card. */
+    headline: string;
+    /** What the user can do about it. */
+    guidance: string;
+    /** "billing": credits ran out or hit a cap. "access": credits are off or not offered. */
+    kind: "billing" | "access";
+}
+/** Describe a fast-mode rejection reason in user terms. */
+declare function describeFastModeRejection(reason: string | undefined): FastModeRejection;
+/**
+ * Recover the rejection from an error message: gg-ai's canonical message, or
+ * Anthropic's raw fast-mode wording (when it reaches a caller unnormalized).
+ */
+declare function parseFastModeRejection(message: string): FastModeRejection | undefined;
 
 /** What the embedded text will be used for. See `EmbedOptions.inputType`. */
 type EmbedInputType = "query" | "document";
@@ -1031,4 +1071,4 @@ 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 FastModeRejection, 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, describeFastModeRejection, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, parseFastModeRejection, 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 +434,43 @@ 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;
+interface FastModeRejection {
+    /** Anthropic's reason code (`unknown` when it sent none). */
+    reason: string;
+    /** Short, factual headline for the error card. */
+    headline: string;
+    /** What the user can do about it. */
+    guidance: string;
+    /** "billing": credits ran out or hit a cap. "access": credits are off or not offered. */
+    kind: "billing" | "access";
+}
+/** Describe a fast-mode rejection reason in user terms. */
+declare function describeFastModeRejection(reason: string | undefined): FastModeRejection;
+/**
+ * Recover the rejection from an error message: gg-ai's canonical message, or
+ * Anthropic's raw fast-mode wording (when it reaches a caller unnormalized).
+ */
+declare function parseFastModeRejection(message: string): FastModeRejection | undefined;
 
 /** What the embedded text will be used for. See `EmbedOptions.inputType`. */
 type EmbedInputType = "query" | "document";
@@ -1031,4 +1071,4 @@ 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 FastModeRejection, 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, describeFastModeRejection, dropInvalidToolCalls, embed, environmentSecrets, formatChatError, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, opensTurn, palsuAssistantMessage, palsuEmbedding, palsuRerankScore, palsuText, palsuThinking, palsuToolCall, parseFastModeRejection, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, rerank, resolveFastModeModel, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor, usesResponsesLite };

package/dist/index.js

@@ -1,3 +1,93 @@
+// 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 };
+}
+var FAST_MODE_REJECTION_HEADER = "anthropic-ratelimit-unified-overage-disabled-reason";
+var FAST_MODE_REJECTION_PREFIX = "Fast mode unavailable";
+var USAGE_PAGE = "claude.ai/settings/usage";
+var STANDARD_SPEED_FALLBACK = "or switch to Claude Opus 5.5 (standard speed), which uses your plan";
+function describeFastModeRejection(reason) {
+  const code = reason?.trim() || "unknown";
+  const of = (headline, guidance, kind) => ({ reason: code, headline, guidance, kind });
+  switch (code) {
+    case "overage_not_provisioned":
+    case "no_limits_configured":
+      return of(
+        "Claude usage credits are turned off",
+        `Fast mode bills your Claude usage credits, never your plan. Turn usage credits on at ${USAGE_PAGE}, ${STANDARD_SPEED_FALLBACK}.`,
+        "access"
+      );
+    case "out_of_credits":
+      return of(
+        "Your Claude usage credits have run out",
+        `Fast mode bills your Claude usage credits, never your plan, and the balance is empty. Add funds or turn on auto-reload at ${USAGE_PAGE}, ${STANDARD_SPEED_FALLBACK}.`,
+        "billing"
+      );
+    case "org_level_disabled_until":
+    case "org_spend_cap_reached":
+      return of(
+        "Claude usage credit limit reached",
+        `Fast mode bills your Claude usage credits and this period's spend limit is used up. Raise the limit at ${USAGE_PAGE}, ${STANDARD_SPEED_FALLBACK}.`,
+        "billing"
+      );
+    case "org_level_disabled":
+    case "org_service_level_disabled":
+      return of(
+        "Your organization turned Claude usage credits off",
+        `Fast mode needs usage credits. Ask an admin to turn them on, ${STANDARD_SPEED_FALLBACK}.`,
+        "access"
+      );
+    case "member_level_disabled":
+      return of(
+        "Claude usage credits are off for your account",
+        `Fast mode needs usage credits. Ask your organization's admin to turn them on for you, ${STANDARD_SPEED_FALLBACK}.`,
+        "access"
+      );
+    case "seat_tier_level_disabled":
+    case "seat_tier_zero_credit_limit":
+    case "member_zero_credit_limit":
+    case "group_zero_credit_limit":
+      return of(
+        "Claude usage credits aren't available on your plan",
+        `Fast mode needs usage credits, which your seat doesn't include. Ask your admin about access, ${STANDARD_SPEED_FALLBACK}.`,
+        "access"
+      );
+    case "fast_mode_org_disabled":
+      return of(
+        "Your organization turned fast mode off",
+        `An Owner can enable fast mode for the organization, ${STANDARD_SPEED_FALLBACK}.`,
+        "access"
+      );
+    default:
+      return of(
+        "Fast mode needs Claude usage credits",
+        `Fast mode bills your Claude usage credits, never your plan, and Anthropic couldn't charge them. Check usage credits at ${USAGE_PAGE}, ${STANDARD_SPEED_FALLBACK}.`,
+        "billing"
+      );
+  }
+}
+function fastModeRejectionMessage(rejection) {
+  return `${FAST_MODE_REJECTION_PREFIX} (${rejection.reason}): ${rejection.headline}.`;
+}
+function parseFastModeRejection(message) {
+  const canonical = /fast mode unavailable \(([a-z_]+)\)/i.exec(message);
+  if (canonical) return describeFastModeRejection(canonical[1]);
+  if (/usage credits are required for fast mode/i.test(message)) {
+    return describeFastModeRejection(void 0);
+  }
+  if (/fast mode (?:has been |is )?disabled by your organization/i.test(message)) {
+    return describeFastModeRejection("fast_mode_org_disabled");
+  }
+  return void 0;
+}
+
 // src/errors.ts
 function readHeader(headers, ...names) {
   if (!headers) return void 0;
@@ -77,7 +167,8 @@ 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 when usage credits can't pay (see fast-mode.ts).
+  lower.includes("usage credits are required") || lower.includes("fast mode unavailable") || 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);
@@ -130,6 +221,18 @@ function formatError(err) {
         guidance: "Request access via your Anthropic account team (see platform.claude.com/docs/en/about-claude/models/overview), or switch to Claude Fable 5.1 via the model selector \u2014 same underlying model, generally available."
       };
     }
+    const fastRejection = err.provider === "anthropic" ? parseFastModeRejection(err.message) : void 0;
+    if (fastRejection) {
+      return {
+        headline: `${fastRejection.headline}.`,
+        source: "provider",
+        message: cleanMessage,
+        provider: err.provider,
+        statusCode: err.statusCode,
+        ...err.requestId ? { requestId: err.requestId } : {},
+        guidance: fastRejection.guidance
+      };
+    }
     if (isUsageLimitError(err)) {
       const resetClause = err.resetsAt ? ` It resets at ${formatResetTime(err.resetsAt)}.` : "";
       return {
@@ -1306,7 +1409,7 @@ function isAdaptiveThinkingModel(model) {
 function toAnthropicThinking(level, maxTokens, model) {
   if (isAdaptiveThinkingModel(model)) {
     let effort = level;
-    if (effort === "xhigh" && !/opus-5|opus-4-8|opus-4-7|sonnet-5[-.]5|haiku-5[-.]5/.test(model)) {
+    if (effort === "xhigh" && !/fable-5|mythos-5|opus-5|opus-4-8|opus-4-7|sonnet-5[-.]5|haiku-5[-.]5/.test(model)) {
       effort = "high";
     }
     return {
@@ -1574,6 +1677,7 @@ 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 +1820,7 @@ 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 +1891,7 @@ 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 +1910,8 @@ 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,
@@ -1842,7 +1949,7 @@ async function* runStream(options) {
       yield* synthesizeEventsFromMessage(message);
       return messageToResponse(message);
     } catch (err) {
-      throw toError(err);
+      throw toError(err, { fastMode: useFastMode });
     }
   }
   const contentParts = [];
@@ -2059,7 +2166,7 @@ async function* runStream(options) {
       }
     }
   } catch (err) {
-    throw toError(err);
+    throw toError(err, { fastMode: useFastMode });
   }
   if (!receivedAnyEvent) {
     throw new ProviderError("anthropic", "Stream ended without producing any events.", {
@@ -2189,6 +2296,13 @@ function messageToResponse(message) {
     }
   };
 }
+function readFastModeRejection(err, message, fastMode) {
+  const reason = readHeader(err.headers, FAST_MODE_REJECTION_HEADER);
+  if (fastMode && err.status === 429 && reason) return describeFastModeRejection(reason);
+  const fromBody = parseFastModeRejection(message);
+  if (!fromBody) return void 0;
+  return reason ? describeFastModeRejection(reason) : fromBody;
+}
 function readUnifiedRateLimit(headers) {
   const status = readHeader(headers, "anthropic-ratelimit-unified-status");
   const resetRaw = readHeader(
@@ -2201,7 +2315,7 @@ function readUnifiedRateLimit(headers) {
   const resetsAt = Number.isFinite(resetNum) && resetNum > 0 ? resetNum : void 0;
   return { rejected: status === "rejected", ...resetsAt ? { resetsAt } : {} };
 }
-function toError(err) {
+function toError(err, context) {
   if (err instanceof ProviderError) return err;
   if (err instanceof Anthropic.APIError) {
     const errorBody = err.error;
@@ -2212,6 +2326,15 @@ 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;
+    const fastRejection = readFastModeRejection(err, message, context?.fastMode === true);
+    if (fastRejection) {
+      return new ProviderError("anthropic", fastModeRejectionMessage(fastRejection), {
+        statusCode: err.status,
+        ...requestId ? { requestId } : {},
+        hint: fastRejection.guidance,
+        cause: err
+      });
+    }
     if (err.status === 429) {
       const limit = readUnifiedRateLimit(err.headers);
       const farOff = limit.resetsAt != null && limit.resetsAt * 1e3 - Date.now() > 6e4;
@@ -2467,8 +2590,9 @@ 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;
@@ -2506,10 +2630,16 @@ async function* runStream2(options) {
 `
     );
   }
+  const requestHeaders = options.provider === "xai" ? {
+    headers: {
+      "x-grok-conv-id": normalizePromptCacheKey(options.promptCacheKey ?? "ggcoder")
+    }
+  } : {};
   if (!useStreaming) {
     try {
       const completion = await client.chat.completions.create(requestParams, {
-        signal: options.signal ?? void 0
+        signal: options.signal ?? void 0,
+        ...requestHeaders
       });
       yield* synthesizeEventsFromCompletion(completion, !!options.thinking, endpointKey);
       return completionToResponse(completion, endpointKey);
@@ -2520,7 +2650,8 @@ async function* runStream2(options) {
   let stream2;
   try {
     stream2 = await client.chat.completions.create(requestParams, {
-      signal: options.signal ?? void 0
+      signal: options.signal ?? void 0,
+      ...requestHeaders
     });
   } catch (err) {
     throw toError2(err, providerName);
@@ -2978,6 +3109,9 @@ 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.
@@ -3560,6 +3694,9 @@ var CODE_ASSIST_SUPPORTED_MODELS = /* @__PURE__ */ new Set([
   "gemma-4-31b-it",
   "gemma-4-26b-a4b-it"
 ]);
+function isGeminiApiKeyRoute(options) {
+  return options.baseUrl?.includes("generativelanguage.googleapis.com") === true;
+}
 function getGoogleProject(options) {
   const env = getEnvironment();
   return options.projectId ?? env?.GOOGLE_CLOUD_PROJECT ?? env?.GOOGLE_CLOUD_PROJECT_ID;
@@ -3592,8 +3729,28 @@ function accountGatedHint(model) {
   if (model === "gemini-3.8-flash" || model === "gemini-3.5-flash-lite") {
     return "Use /model to select Gemini 3.1 Flash Lite, or retry once Google enables this model through Code Assist for your account.";
   }
-  return `Newer Gemini models (3.7 Flash, 3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin \u2014 free/personal accounts usually can't call them. Switch to Gemini 3.1 Flash Lite (it works on this account) with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
+  return `Newer Gemini models (3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin. Try Gemini 3.1 Flash Lite with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
 }
+function isCodeAssistLicenseRequired(status, body) {
+  if (status !== 403) return false;
+  try {
+    const parsed = JSON.parse(body);
+    const details = typeof parsed === "object" && parsed !== null && "error" in parsed ? parsed.error?.details : void 0;
+    return Array.isArray(details) && details.some(
+      (detail) => typeof detail === "object" && detail !== null && detail.reason === "SUBSCRIPTION_REQUIRED"
+    );
+  } catch {
+    return false;
+  }
+}
+function isInvalidGeminiApiKey(status, body) {
+  if (status !== 400 && status !== 401 && status !== 403) return false;
+  return /"reason"\s*:\s*"API_KEY_INVALID"/.test(body) || /API key not valid/i.test(body);
+}
+var INVALID_API_KEY_MESSAGE = "Google rejected the Gemini API key: it is not valid.";
+var INVALID_API_KEY_HINT = "Create a key at aistudio.google.com/apikey and add it again under Gemini in Providers (or `ggcoder login`).";
+var LICENSE_REQUIRED_MESSAGE = "This Google account has no Gemini Code Assist licence. Google stopped serving Gemini through Code Assist to personal accounts (Google AI Pro/Ultra and the free individual tier) on June 18, 2026, so every Gemini model is refused, not just this one.";
+var LICENSE_REQUIRED_HINT = "Sign in to Gemini with a Google account that has a Gemini Code Assist Standard or Enterprise licence, or pick a model from another provider with /model.";
 function formatErrorMessage(status, body, model) {
   if (status === 404 && !CODE_ASSIST_SUPPORTED_MODELS.has(model)) {
     return `Gemini API error (404): ${body}
@@ -3817,7 +3974,29 @@ function buildCodeAssistRequest(options, request, projectId) {
     request
   };
 }
+var API_KEY_MODEL_IDS = {
+  "gemini-3-flash": "gemini-3.5-flash"
+};
+function buildApiKeyRequestPlan(options, method) {
+  const { session_id: _sessionId, ...request } = mergeProviderOptions(
+    buildGenerateRequest(options),
+    options.providerOptions
+  );
+  const base = (options.baseUrl ?? "").replace(/\/+$/, "");
+  return {
+    url: new URL(
+      `${base}/models/${encodeURIComponent(API_KEY_MODEL_IDS[options.model] ?? options.model)}:${method}`
+    ),
+    headers: {
+      "Content-Type": "application/json",
+      "x-goog-api-key": options.apiKey ?? ""
+    },
+    body: request,
+    route: "api-key"
+  };
+}
 function buildRequestPlan(options, method) {
+  if (isGeminiApiKeyRoute(options)) return buildApiKeyRequestPlan(options, method);
   if (!CODE_ASSIST_SUPPORTED_MODELS.has(options.model)) {
     throw new ProviderError("gemini", formatUnsupportedModelMessage(options.model));
   }
@@ -3831,7 +4010,8 @@ function buildRequestPlan(options, method) {
       "User-Agent": GEMINI_CLI_USER_AGENT,
       "X-Goog-Api-Client": GEMINI_CLI_API_CLIENT
     },
-    body: buildCodeAssistRequest(options, request, projectId)
+    body: buildCodeAssistRequest(options, request, projectId),
+    route: "code-assist"
   };
 }
 function normalizeGeminiStopReason(reason) {
@@ -3917,8 +4097,11 @@ async function fetchCodeAssist(plan, options) {
     if (!response.ok) {
       const text = await response.text().catch(() => "");
       const quota = parseGeminiQuota(response.status, text);
-      const accountGated = response.status === 404 && ACCOUNT_GATED_MODELS.has(options.model);
-      let message = accountGated ? accountGatedMessage(options.model) : formatErrorMessage(response.status, text, options.model);
+      const codeAssist = plan.route === "code-assist";
+      const accountGated = codeAssist && response.status === 404 && ACCOUNT_GATED_MODELS.has(options.model);
+      const licenseRequired = codeAssist && isCodeAssistLicenseRequired(response.status, text);
+      const invalidKey = !codeAssist && isInvalidGeminiApiKey(response.status, text);
+      let message = invalidKey ? INVALID_API_KEY_MESSAGE : licenseRequired ? LICENSE_REQUIRED_MESSAGE : accountGated ? accountGatedMessage(options.model) : formatErrorMessage(response.status, text, options.model);
       let resetsAt;
       if (quota?.exhausted) {
         message = `Gemini quota exhausted \u2014 usage limit reached. ${message}`;
@@ -3928,7 +4111,7 @@ async function fetchCodeAssist(plan, options) {
       throw new ProviderError("gemini", message, {
         statusCode: response.status,
         ...resetsAt !== void 0 ? { resetsAt } : {},
-        ...accountGated ? { hint: accountGatedHint(options.model) } : {}
+        ...invalidKey ? { hint: INVALID_API_KEY_HINT, source: "auth" } : licenseRequired ? { hint: LICENSE_REQUIRED_HINT, source: "auth" } : accountGated ? { hint: accountGatedHint(options.model) } : {}
       });
     }
     return response;
@@ -4815,7 +4998,7 @@ providerRegistry.register("sakana", {
 providerRegistry.register("xai", {
   // xAI's public API (console.x.ai key) is OpenAI-compatible — ride the Chat
   // Completions transport like Moonshot/DeepSeek. Grok reasoning models take
-  // top-level `reasoning_effort` (low/medium/high), which the shared thinking
+  // top-level `reasoning_effort` (low/medium/high/xhigh), which the shared thinking
   // path already sends.
   //
   // Subscription OAuth (SuperGrok / X Premium) routes to the Grok CLI chat proxy
@@ -4887,7 +5070,9 @@ 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(
@@ -5102,6 +5287,14 @@ function formatChatError(error) {
       "Update the app, or choose another model."
     );
   }
+  const fastRejection = base.provider === "anthropic" ? parseFastModeRejection(raw) : void 0;
+  if (fastRejection) {
+    return present(
+      fastRejection.kind === "billing" ? "billing" : "model_access",
+      fastRejection.headline,
+      fastRejection.guidance
+    );
+  }
   if (/insufficient.*(?:balance|credits?)|(?:balance|credits?).*(?:insufficient|exhausted|depleted)|credit balance is too low|payment_required/.test(
     text
   )) {
@@ -5518,6 +5711,7 @@ function palsuRerankScore(query, document) {
 }
 export {
   EventStream,
+  FAST_MODE_SUFFIX,
   GGAIError,
   ProviderError,
   REDACTED as REDACTION_MARKER,
@@ -5526,6 +5720,7 @@ export {
   clampProviderContextImages,
   classifyProviderError,
   continuesTurn,
+  describeFastModeRejection,
   dropInvalidToolCalls,
   embed,
   environmentSecrets,
@@ -5544,12 +5739,14 @@ export {
   palsuText,
   palsuThinking,
   palsuToolCall,
+  parseFastModeRejection,
   prewarmAnthropicCache,
   providerRegistry,
   redactText,
   redactValue,
   registerPalsuProvider,
   rerank,
+  resolveFastModeModel,
   resolveToolSchema,
   sanitizeMessagesForWire,
   setProviderDiagnostic,