@kenkaiiii/gg-ai 5.23.3 → 5.25.0

package/dist/index.cjs

@@ -42,6 +42,7 @@ __export(index_exports, {
   formatErrorForDisplay: () => formatErrorForDisplay,
   isHardBillingMessage: () => isHardBillingMessage,
   isUsageLimitError: () => isUsageLimitError,
+  localWireModelId: () => localWireModelId,
   palsuAssistantMessage: () => palsuAssistantMessage,
   palsuText: () => palsuText,
   palsuThinking: () => palsuThinking,
@@ -561,6 +562,39 @@ 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;
@@ -981,6 +1015,7 @@ 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";
@@ -1047,9 +1082,9 @@ 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;
@@ -1127,6 +1162,10 @@ 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")) {
@@ -1922,7 +1961,9 @@ 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;
@@ -1945,7 +1986,8 @@ 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;
@@ -1957,7 +1999,7 @@ 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 } } : {}
@@ -1971,6 +2013,11 @@ 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;
   }
@@ -2007,8 +2054,8 @@ 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);
     }
@@ -2043,11 +2090,12 @@ 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) {
@@ -2132,16 +2180,17 @@ 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 };
@@ -2169,15 +2218,16 @@ 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;
@@ -3529,6 +3579,28 @@ 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) {
@@ -3933,6 +4005,7 @@ function registerPalsuProvider(config) {
   formatErrorForDisplay,
   isHardBillingMessage,
   isUsageLimitError,
+  localWireModelId,
   palsuAssistantMessage,
   palsuText,
   palsuThinking,