@kenkaiiii/gg-ai 5.23.3 → 5.25.0

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
@@ -507,6 +517,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     provider?: string;
     thinking?: boolean;
     supportsImages?: boolean;
+    /** Wire name for reasoning on assistant messages. Defaults to `reasoning_content`. */
+    reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
 /**
@@ -600,4 +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
@@ -507,6 +517,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     provider?: string;
     thinking?: boolean;
     supportsImages?: boolean;
+    /** Wire name for reasoning on assistant messages. Defaults to `reasoning_content`. */
+    reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
 /**
@@ -600,4 +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

@@ -501,6 +501,39 @@ function normalizeRootForAnthropic(schema) {
   return out;
 }
 
+// src/providers/reasoning-field.ts
+var REASONING_FIELD_ALIASES = [
+  "reasoning_content",
+  "reasoning",
+  "reasoning_text"
+];
+var DEFAULT_REASONING_FIELD = REASONING_FIELD_ALIASES[0];
+function readReasoning(obj) {
+  if (!obj) return void 0;
+  for (const field of REASONING_FIELD_ALIASES) {
+    const value = obj[field];
+    if (typeof value === "string" && value) return { field, text: value };
+  }
+  return void 0;
+}
+function reasoningFieldKey(provider, baseUrl, model) {
+  return `${provider}|${baseUrl ?? ""}|${model}`;
+}
+var MAX_REMEMBERED_ENDPOINTS = 64;
+var detectedFields = /* @__PURE__ */ new Map();
+function rememberReasoningField(key, field) {
+  if (detectedFields.get(key) === field) return;
+  detectedFields.set(key, field);
+  while (detectedFields.size > MAX_REMEMBERED_ENDPOINTS) {
+    const oldest = detectedFields.keys().next();
+    if (oldest.done) break;
+    detectedFields.delete(oldest.value);
+  }
+}
+function getReasoningField(key) {
+  return detectedFields.get(key) ?? DEFAULT_REASONING_FIELD;
+}
+
 // src/providers/transform.ts
 function hasValidThinkingSignature(part) {
   return typeof part.signature === "string" && part.signature.trim().length > 0;
@@ -921,6 +954,7 @@ function remapToolCallId(id, idMap) {
   return mapped;
 }
 function toOpenAIMessages(messages, options) {
+  const reasoningField = options?.reasoningField || DEFAULT_REASONING_FIELD;
   const out = [];
   const idMap = /* @__PURE__ */ new Map();
   const mergeToolResultText = options?.provider === "glm";
@@ -987,9 +1021,9 @@ function toOpenAIMessages(messages, options) {
         ...hasToolCalls ? { tool_calls: toolCalls } : {}
       };
       if (thinkingParts) {
-        assistantMsg.reasoning_content = thinkingParts;
+        assistantMsg[reasoningField] = thinkingParts;
       } else if (options?.thinking && hasToolCalls && options.provider !== "glm") {
-        assistantMsg.reasoning_content = " ";
+        assistantMsg[reasoningField] = " ";
       }
       out.push(assistantMsg);
       continue;
@@ -1067,6 +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")) {
@@ -1862,7 +1900,9 @@ function streamOpenAI(options) {
 async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
   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;
@@ -1885,7 +1925,8 @@ async function* runStream2(options) {
     // disabled K3 must NOT carry placeholder reasoning_content (mirrors the
     // official CLI: reasoning is preserved only while thinking is enabled).
     thinking: isKimiK27 || !!options.thinking,
-    supportsImages: options.supportsImages
+    supportsImages: options.supportsImages,
+    reasoningField: getReasoningField(endpointKey)
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
   const effectiveTemp = options.temperature ?? defaultTemp;
@@ -1897,7 +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 } } : {}
@@ -1911,6 +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;
   }
@@ -1947,8 +1993,8 @@ async function* runStream2(options) {
       const completion = await client.chat.completions.create(params, {
         signal: options.signal ?? void 0
       });
-      yield* synthesizeEventsFromCompletion(completion, !!options.thinking);
-      return completionToResponse(completion);
+      yield* synthesizeEventsFromCompletion(completion, !!options.thinking, endpointKey);
+      return completionToResponse(completion, endpointKey);
     } catch (err) {
       throw toError2(err, providerName);
     }
@@ -1983,11 +2029,12 @@ async function* runStream2(options) {
         finishReason = choice.finish_reason;
       }
       const delta = choice.delta;
-      const reasoningContent = delta.reasoning_content;
-      if (typeof reasoningContent === "string" && reasoningContent) {
-        thinkingAccum += reasoningContent;
+      const reasoning = readReasoning(delta);
+      if (reasoning) {
+        rememberReasoningField(endpointKey, reasoning.field);
+        thinkingAccum += reasoning.text;
         if (options.thinking) {
-          yield { type: "thinking_delta", text: reasoningContent };
+          yield { type: "thinking_delta", text: reasoning.text };
         }
       }
       if (delta.content) {
@@ -2072,16 +2119,17 @@ async function* runStream2(options) {
   yield { type: "done", stopReason };
   return response;
 }
-function* synthesizeEventsFromCompletion(completion, thinkingEnabled) {
+function* synthesizeEventsFromCompletion(completion, thinkingEnabled, endpointKey) {
   const choice = completion.choices?.[0];
   if (!choice) {
     yield { type: "done", stopReason: normalizeOpenAIStopReason(null) };
     return;
   }
   const msg = choice.message;
-  const reasoning = msg.reasoning_content;
-  if (typeof reasoning === "string" && reasoning && thinkingEnabled) {
-    yield { type: "thinking_delta", text: reasoning };
+  const reasoning = readReasoning(msg);
+  if (reasoning) {
+    rememberReasoningField(endpointKey, reasoning.field);
+    if (thinkingEnabled) yield { type: "thinking_delta", text: reasoning.text };
   }
   if (typeof msg.content === "string" && msg.content) {
     yield { type: "text_delta", text: msg.content };
@@ -2109,15 +2157,16 @@ function* synthesizeEventsFromCompletion(completion, thinkingEnabled) {
   }
   yield { type: "done", stopReason: normalizeOpenAIStopReason(choice.finish_reason ?? null) };
 }
-function completionToResponse(completion) {
+function completionToResponse(completion, endpointKey) {
   const choice = completion.choices?.[0];
   const contentParts = [];
   let textAccum = "";
   if (choice) {
     const msg = choice.message;
-    const reasoning = msg.reasoning_content;
-    if (typeof reasoning === "string" && reasoning) {
-      contentParts.push({ type: "thinking", text: reasoning });
+    const reasoning = readReasoning(msg);
+    if (reasoning) {
+      rememberReasoningField(endpointKey, reasoning.field);
+      contentParts.push({ type: "thinking", text: reasoning.text });
     }
     if (typeof msg.content === "string" && msg.content) {
       textAccum = msg.content;
@@ -3469,6 +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) {
@@ -3872,6 +3943,7 @@ export {
   formatErrorForDisplay,
   isHardBillingMessage,
   isUsageLimitError,
+  localWireModelId,
   palsuAssistantMessage,
   palsuText,
   palsuThinking,