@kenkaiiii/gg-ai 5.24.0 → 5.26.0

package/dist/index.d.cts

@@ -2,7 +2,9 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
+ | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -293,6 +295,14 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
 
+/**
+ * Local model ids are namespaced by endpoint (`local/<endpointId>/<rawId>`) so
+ * the same model name served by two machines stays distinct in the registry.
+ * The server only knows the raw id, so strip the routing prefix here — at the
+ * one place that talks to the wire. Counterpart to gg-core's
+ * `formatLocalModelId`/`parseLocalModelId`.
+ */
+declare function localWireModelId(id: string): string;
 /**
  * Unified streaming entry point. Returns a StreamResult that is both
  * an async iterable (for streaming events) and thenable (await for
@@ -602,4 +612,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

package/dist/index.d.ts

@@ -2,7 +2,9 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
+ | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -293,6 +295,14 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
 
+/**
+ * Local model ids are namespaced by endpoint (`local/<endpointId>/<rawId>`) so
+ * the same model name served by two machines stays distinct in the registry.
+ * The server only knows the raw id, so strip the routing prefix here — at the
+ * one place that talks to the wire. Counterpart to gg-core's
+ * `formatLocalModelId`/`parseLocalModelId`.
+ */
+declare function localWireModelId(id: string): string;
 /**
  * Unified streaming entry point. Returns a StreamResult that is both
  * an async iterable (for streaming events) and thenable (await for
@@ -602,4 +612,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

package/dist/index.js

@@ -1101,6 +1101,10 @@ function toOpenAIToolChoice(choice) {
   if (choice === "required") return "required";
   return { type: "function", function: { name: choice.name } };
 }
+function toLocalReasoningEffort(level) {
+  if (level === "max" || level === "ultra" || level === "xhigh") return "max";
+  return level;
+}
 function toOpenAIReasoningEffort(level, model) {
   const effort = level === "max" || level === "ultra" ? "xhigh" : level;
   if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
@@ -1898,6 +1902,7 @@ async function* runStream2(options) {
   const useStreaming = options.streaming !== false;
   const endpointKey = reasoningFieldKey(providerName, options.baseUrl, options.model);
   const client = createClient2(options);
+  const isLocal = options.provider === "local";
   const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
   const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").endsWith("/coding/v1") === true;
   const k3Effort = options.thinking ? toKimiK3Effort(options.thinking) : void 0;
@@ -1933,7 +1938,7 @@ async function* runStream2(options) {
     ...effectiveTemp != null && !options.thinking && !hasFixedKimiSampling ? { temperature: effectiveTemp } : {},
     ...options.topP != null && !hasFixedKimiSampling ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
-    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
+    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 && !isLocal ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
     ...options.tools?.length ? { tools: toOpenAITools(options.tools) } : {},
     ...options.toolChoice && options.tools?.length ? { tool_choice: toOpenAIToolChoice(options.toolChoice) } : {},
     ...useStreaming ? { stream_options: { include_usage: true } } : {}
@@ -1947,6 +1952,11 @@ async function* runStream2(options) {
       paramsAny.prompt_cache_retention = "24h";
     }
   }
+  if (isLocal && options.thinking) {
+    params.reasoning_effort = toLocalReasoningEffort(
+      options.thinking
+    );
+  }
   if (options.provider === "openai" && options.serviceTier) {
     params.service_tier = options.serviceTier;
   }
@@ -3508,6 +3518,28 @@ providerRegistry.register("minimax", {
     serverTools: void 0
   })
 });
+function localWireModelId(id) {
+  const match = /^local\/[^/]+\/(.+)$/.exec(id);
+  return match?.[1] ?? id;
+}
+providerRegistry.register("local", {
+  // Locally hosted OpenAI-compatible servers (Ollama, LM Studio, llama.cpp,
+  // vLLM). There is no default endpoint: the baseUrl comes from the endpoint
+  // credential the discovery layer wrote, so a missing one is a wiring bug, not
+  // something to paper over with a guess at someone else's port.
+  stream: (options) => {
+    if (!options.baseUrl) {
+      throw new GGAIError(
+        "Local provider requires a baseUrl (e.g. http://127.0.0.1:11434/v1). No local endpoint was resolved for this model \u2014 re-scan for local models."
+      );
+    }
+    return streamOpenAI({
+      ...options,
+      model: localWireModelId(options.model),
+      webSearch: false
+    });
+  }
+});
 function stream(options) {
   const entry = providerRegistry.get(options.provider);
   if (!entry) {
@@ -3911,6 +3943,7 @@ export {
   formatErrorForDisplay,
   isHardBillingMessage,
   isUsageLimitError,
+  localWireModelId,
   palsuAssistantMessage,
   palsuText,
   palsuThinking,