@kenkaiiii/gg-ai 5.88.1 → 5.90.0

package/dist/index.cjs

@@ -35,6 +35,7 @@ __export(index_exports, {
   EventStream: () => EventStream,
   FAST_MODE_SUFFIX: () => FAST_MODE_SUFFIX,
   GGAIError: () => GGAIError,
+  LOCAL_NO_API_KEY: () => LOCAL_NO_API_KEY,
   PRECONNECT_INTERVAL_MS: () => PRECONNECT_INTERVAL_MS,
   ProviderError: () => ProviderError,
   REDACTION_MARKER: () => REDACTED,
@@ -57,6 +58,7 @@ __export(index_exports, {
   isHardBillingMessage: () => isHardBillingMessage,
   isUsageLimitError: () => isUsageLimitError,
   isValidToolCallName: () => isValidToolCallName,
+  localBearerKey: () => localBearerKey,
   localWireModelId: () => localWireModelId,
   opensTurn: () => opensTurn,
   palsuAssistantMessage: () => palsuAssistantMessage,
@@ -1656,10 +1658,22 @@ function toAnthropicMessages(messages, cacheControl, options) {
           (name) => toolReferences.has(name) && !referencedTools.has(name)
         ) : [];
         if (refs.length === 0) {
+          const content = toAnthropicToolResultContent(result.content);
+          if (result.isError && typeof content !== "string" && content.some((part) => part.type !== "text")) {
+            const text = content.filter((part) => part.type === "text" && part.text !== "");
+            trailing.push({ type: "text", text: `Attachments from tool call ${toolUseId}:` });
+            for (const part of content) if (part.type !== "text") trailing.push(part);
+            return {
+              type: "tool_result",
+              tool_use_id: toolUseId,
+              content: text.length > 0 ? text : "Tool call failed; its attachments follow.",
+              is_error: true
+            };
+          }
           return {
             type: "tool_result",
             tool_use_id: toolUseId,
-            content: toAnthropicToolResultContent(result.content),
+            content,
             is_error: result.isError
           };
         }
@@ -1678,7 +1692,8 @@ ${part.text}` } : part
           type: "tool_result",
           tool_use_id: toolUseId,
           content: refs.map((name) => ({ type: "tool_reference", tool_name: name })),
-          is_error: result.isError
+          // tool_reference isn't text, so is_error would 400; the error text trails.
+          is_error: void 0
         };
       });
       out.push({
@@ -1701,8 +1716,8 @@ ${part.text}` } : part
           markAnthropicCacheBreakpoint(out, tail, cacheControl);
         } else {
           if (anchor !== void 0) markAnthropicCacheBreakpoint(out, anchor, cacheControl);
-          const longRun = isLongToolRun(messages, trajectoryStartIdx);
-          markAnthropicCacheBreakpoint(out, tail, longRun ? cacheControl : { type: "ephemeral" });
+          const blocking = calledBlockingTool(messages, trajectoryStartIdx);
+          markAnthropicCacheBreakpoint(out, tail, blocking ? cacheControl : { type: "ephemeral" });
         }
       } else {
         markAnthropicCacheBreakpoint(out, tail, cacheControl);
@@ -1738,19 +1753,13 @@ var BLOCKING_TOOL_NAMES = /* @__PURE__ */ new Set([
   "subagent",
   "ask_user"
 ]);
-var LONG_RUN_ROUNDS = 6;
-function isLongToolRun(messages, trajectoryStartIdx) {
-  let rounds = 0;
+function calledBlockingTool(messages, trajectoryStartIdx) {
   for (let i = Math.max(0, trajectoryStartIdx + 1); i < messages.length; i++) {
     const m = messages[i];
     if (m?.role !== "assistant" || typeof m.content === "string") continue;
-    let called = false;
     for (const part of m.content) {
-      if (part.type !== "tool_call") continue;
-      called = true;
-      if (BLOCKING_TOOL_NAMES.has(part.name)) return true;
+      if (part.type === "tool_call" && BLOCKING_TOOL_NAMES.has(part.name)) return true;
     }
-    if (called && ++rounds >= LONG_RUN_ROUNDS) return true;
   }
   return false;
 }
@@ -2937,6 +2946,12 @@ function getEnvironment() {
   return globalThis.process?.env;
 }
 
+// src/local-api-key.ts
+var LOCAL_NO_API_KEY = "local";
+function localBearerKey(apiKey) {
+  return apiKey && apiKey !== LOCAL_NO_API_KEY ? apiKey : void 0;
+}
+
 // src/providers/openai.ts
 function toKimiK3Effort(level) {
   switch (level) {
@@ -2981,6 +2996,8 @@ function createClient2(options) {
     const cached = openaiClientCache.get(cacheKey);
     if (cached) return cached;
   }
+  const omitAuth = options.provider === "local" && localBearerKey(options.apiKey) === void 0;
+  const defaultHeaders = omitAuth ? { ...options.defaultHeaders, Authorization: null } : options.defaultHeaders;
   const client = new import_openai.default({
     apiKey: options.apiKey,
     ...options.baseUrl ? { baseURL: options.baseUrl } : {},
@@ -2990,7 +3007,7 @@ function createClient2(options) {
     // events). The SDK default of 2 silent retries stacked under it: one 5xx
     // cost up to 3 hidden requests plus SDK sleeps before the loop even saw it.
     maxRetries: 0,
-    ...options.defaultHeaders ? { defaultHeaders: options.defaultHeaders } : {}
+    ...defaultHeaders ? { defaultHeaders } : {}
   });
   if (!options.fetch) {
     if (openaiClientCache.size >= 8) {
@@ -5747,19 +5764,25 @@ providerRegistry.register("local", {
   }),
   // Every supported local server exposes OpenAI-compatible `/v1/embeddings`
   // (the model must be an embedding model). None has an input-type parameter.
-  embed: (request) => embedOpenAICompatible(request, {
-    provider: "local",
-    baseUrl: requireLocalBaseUrl(request.baseUrl),
-    wireModel: localWireModelId(request.model)
-  }),
+  embed: (request) => embedOpenAICompatible(
+    { ...request, apiKey: localBearerKey(request.apiKey) },
+    {
+      provider: "local",
+      baseUrl: requireLocalBaseUrl(request.baseUrl),
+      wireModel: localWireModelId(request.model)
+    }
+  ),
   embedBatchSize: LOCAL_EMBED_BATCH_SIZE,
   // llama.cpp (`llama-server --rerank`) and vLLM serve `/v1/rerank`; Ollama and
   // LM Studio don't, and answer with a 404 ProviderError.
-  rerank: (request) => rerankCohereCompatible(request, {
-    provider: "local",
-    baseUrl: requireLocalBaseUrl(request.baseUrl),
-    wireModel: localWireModelId(request.model)
-  })
+  rerank: (request) => rerankCohereCompatible(
+    { ...request, apiKey: localBearerKey(request.apiKey) },
+    {
+      provider: "local",
+      baseUrl: requireLocalBaseUrl(request.baseUrl),
+      wireModel: localWireModelId(request.model)
+    }
+  )
 });
 function requireLocalBaseUrl(baseUrl) {
   if (!baseUrl) {
@@ -6133,6 +6156,13 @@ function formatChatError(error) {
       "Check your connection before trying again."
     );
   }
+  if (base.provider === "local" && (base.source === "auth" || base.statusCode === 401)) {
+    return present(
+      "auth",
+      "The local server rejected its key",
+      "Check the endpoint's key under Local models in AI Providers settings."
+    );
+  }
   if (base.source === "auth" || base.statusCode === 401) {
     return present(
       "auth",
@@ -6506,6 +6536,7 @@ function palsuRerankScore(query, document) {
   EventStream,
   FAST_MODE_SUFFIX,
   GGAIError,
+  LOCAL_NO_API_KEY,
   PRECONNECT_INTERVAL_MS,
   ProviderError,
   REDACTION_MARKER,
@@ -6528,6 +6559,7 @@ function palsuRerankScore(query, document) {
   isHardBillingMessage,
   isUsageLimitError,
   isValidToolCallName,
+  localBearerKey,
   localWireModelId,
   opensTurn,
   palsuAssistantMessage,