@yourgpt/llm-sdk 2.6.0-beta.0 → 2.6.0-beta.1

package/dist/providers/togetherai/index.js

@@ -1,5 +1,56 @@
 'use strict';
 
+// src/core/openai-parameters.ts
+function capabilities(model) {
+  const gpt = /^gpt-(\d+)(?:\.(\d+))?(?:-(.+))?$/i.exec(model);
+  const major = Number(gpt?.[1] ?? 0);
+  const minor = Number(gpt?.[2] ?? 0);
+  const chat = gpt?.[3]?.startsWith("chat") ?? false;
+  const reasoning = /^o\d+(?:-|$)/i.test(model) || major >= 5 && !chat;
+  const supportsNone = major === 5 && minor >= 1 && !chat || /^gpt-6-(?:sol|luna)(?:-\d{4}-\d{2}-\d{2})?$/i.test(model);
+  return { major, minor, reasoning, supportsNone };
+}
+function requiresOpenAIResponses(model, baseUrl, options = {}) {
+  if (options.parameterPolicy === "passthrough") return false;
+  const c = capabilities(model);
+  if (!c.reasoning || options.reasoningEffort === "none") return false;
+  if (!(c.major >= 6 || c.major === 5 && c.minor >= 4)) return false;
+  if (!baseUrl) return true;
+  try {
+    return new URL(baseUrl).hostname === "api.openai.com";
+  } catch {
+    return false;
+  }
+}
+function buildOpenAIParameters(model, maxTokens, temperature, options = {}, api = "chat") {
+  const c = capabilities(model ?? "");
+  const normalize = options.parameterPolicy !== "passthrough";
+  const effort = options.reasoningEffort;
+  if (normalize && c.reasoning && effort === "none" && !c.supportsNone) {
+    throw new Error(`${model} does not support reasoningEffort: none`);
+  }
+  if (normalize && c.major >= 6 && effort === "minimal") {
+    throw new Error(`${model} does not support reasoningEffort: minimal`);
+  }
+  const result = {
+    [api === "responses" ? "max_output_tokens" : normalize && c.reasoning ? "max_completion_tokens" : "max_tokens"]: maxTokens
+  };
+  if (effort !== void 0) {
+    if (api === "responses") result.reasoning = { effort };
+    else result.reasoning_effort = effort;
+  }
+  if (temperature !== void 0) {
+    if (!normalize || !c.reasoning || effort === "none" && c.supportsNone) {
+      result.temperature = temperature;
+    } else {
+      (options.onWarning ?? console.warn)(
+        `[llm-sdk] ${model}: omitted temperature because reasoning is enabled.`
+      );
+    }
+  }
+  return result;
+}
+
 // src/adapters/base.ts
 function stringifyForDebug(value) {
   return JSON.stringify(
@@ -106,16 +157,6 @@ function normalizeObjectJsonSchema(schema) {
   }
   return normalized;
 }
-function isOpenAIReasoningModel(modelId) {
-  if (!modelId) return false;
-  return /^(o1|o3|o4|gpt-5)/i.test(modelId);
-}
-function buildOpenAITokenParams(modelId, maxTokens, temperature) {
-  if (isOpenAIReasoningModel(modelId)) {
-    return { max_completion_tokens: maxTokens };
-  }
-  return { max_tokens: maxTokens, temperature };
-}
 function toOpenAIResponseFormat(rf) {
   if (!rf) return void 0;
   if (rf.type === "json_object") return { type: "json_object" };
@@ -447,6 +488,28 @@ function formatMessages(messages) {
   });
 }
 
+// src/core/provider-error.ts
+function providerErrorEvent(error, code) {
+  const details = error ?? {};
+  const event = {
+    type: "error",
+    message: error instanceof Error ? error.message : String(error),
+    code,
+    ...typeof details.status === "number" ? { status: details.status } : {},
+    ...typeof details.code === "string" ? { providerCode: details.code } : {},
+    ...typeof details.param === "string" ? { param: details.param } : {},
+    ...typeof details.request_id === "string" ? { requestId: details.request_id } : {}
+  };
+  Object.defineProperty(event, "cause", { value: error, enumerable: false });
+  return event;
+}
+function throwIfAborted(signal) {
+  if (!signal?.aborted) return;
+  const error = new Error("The operation was aborted");
+  error.name = "AbortError";
+  throw error;
+}
+
 // src/core/utils.ts
 function generateId(prefix = "id") {
   return `${prefix}_${Date.now()}_${Math.random().toString(36).substring(2, 9)}`;
@@ -478,39 +541,38 @@ var OpenAIAdapter = class _OpenAIAdapter {
       const { default: OpenAI } = await import('openai');
       this.client = new OpenAI({
         apiKey: this.config.apiKey,
-        baseURL: this.config.baseUrl
+        baseURL: this.config.baseUrl,
+        organization: this.config.organization,
+        defaultHeaders: this.config.headers
       });
     }
     return this.client;
   }
+  parameterOptions(request) {
+    return {
+      ...this.config,
+      reasoningEffort: request.config?.reasoningEffort ?? this.config.reasoningEffort
+    };
+  }
+  toolDefinitions(request) {
+    if (request.toolDefinitions?.length) return request.toolDefinitions;
+    return formatTools(request.actions ?? []).map((tool) => ({
+      name: tool.function.name,
+      description: tool.function.description ?? "",
+      location: "server",
+      inputSchema: tool.function.parameters
+    }));
+  }
   shouldUseResponsesApi(request) {
-    const hasTools = Array.isArray(request.toolDefinitions) && request.toolDefinitions.length > 0;
-    if (!hasTools) return false;
-    if (request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false) {
+    if (this.toolDefinitions(request).length === 0) return false;
+    if (request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false)
       return true;
-    }
-    return this.requiresResponsesApiForTools(
-      request.config?.model || this.model
+    return requiresOpenAIResponses(
+      request.config?.model || this.model,
+      this.config.baseUrl,
+      this.parameterOptions(request)
     );
   }
-  /**
-   * Whether this model rejects function tools on /v1/chat/completions.
-   *
-   * Applies to OpenAI proper only — the adapter is reused for OpenAI-compatible
-   * providers (Azure, xAI, Google, Fireworks) whose endpoints have no such
-   * restriction and may not implement /v1/responses at all.
-   */
-  requiresResponsesApiForTools(modelId) {
-    if (!modelId) return false;
-    if (!/^gpt-5\.(?:[4-9]|\d{2,})/i.test(modelId)) return false;
-    const baseUrl = this.config.baseUrl;
-    if (!baseUrl) return true;
-    try {
-      return new URL(baseUrl).hostname === "api.openai.com";
-    } catch {
-      return false;
-    }
-  }
   /**
    * Convert one message's content to Responses content parts.
    *
@@ -541,7 +603,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
     });
   }
   buildResponsesInput(request) {
-    const sourceMessages = request.rawMessages && request.rawMessages.length > 0 ? request.rawMessages : formatMessagesForOpenAI(request.messages, void 0);
+    const sourceMessages = request.rawMessages && request.rawMessages.length > 0 ? this.formatRawMessages(request.rawMessages) : formatMessagesForOpenAI(request.messages, void 0);
     const input = [];
     for (const message of sourceMessages) {
       if (message.role === "system") {
@@ -591,6 +653,26 @@ var OpenAIAdapter = class _OpenAIAdapter {
     }
     return input;
   }
+  /** Preserve Copilot screenshot attachments on every OpenAI transport. */
+  formatRawMessages(messages) {
+    return messages.map((message) => {
+      const { attachments, ...rest } = message;
+      let content = message.content;
+      if (message.role === "assistant" && Array.isArray(message.tool_calls) && message.tool_calls.length > 0 && content === "") {
+        content = null;
+      }
+      if (Array.isArray(attachments) && attachments.length > 0) {
+        const images = attachments.filter((attachment) => attachment && typeof attachment === "object").map(
+          (attachment) => attachmentToOpenAIImage(attachment)
+        ).filter((image) => image !== null);
+        if (images.length > 0) {
+          const parts = Array.isArray(content) ? content : content ? [{ type: "text", text: content }] : [];
+          content = [...parts, ...images];
+        }
+      }
+      return { ...rest, content };
+    });
+  }
   /**
    * Build the web_search tool for the Responses API, or undefined when web
    * search is off.
@@ -672,13 +754,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
     const toolCalls = Array.isArray(response?.output) ? response.output.filter((item) => item?.type === "function_call").map((item) => ({
       id: item.call_id ?? item.id ?? generateToolCallId(),
       name: item.name,
-      args: (() => {
-        try {
-          return JSON.parse(item.arguments ?? "{}");
-        } catch {
-          return {};
-        }
-      })()
+      args: JSON.parse(item.arguments ?? "{}")
     })) : [];
     return {
       content,
@@ -702,18 +778,19 @@ var OpenAIAdapter = class _OpenAIAdapter {
       instructions: request.systemPrompt,
       input: this.buildResponsesInput(request),
       tools: this.buildResponsesTools(
-        request.toolDefinitions ?? [],
+        this.toolDefinitions(request),
         request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true,
         this.buildResponsesWebSearchTool(request)
       ),
-      tool_choice: openaiToolOptions?.toolChoice === "required" ? "required" : openaiToolOptions?.toolChoice === "auto" ? "auto" : void 0,
+      tool_choice: openaiToolOptions?.toolChoice,
       parallel_tool_calls: openaiToolOptions?.parallelToolCalls,
-      // Reasoning models reject `temperature` outright — sending it would swap one
-      // 400 for another. isOpenAIReasoningModel already encodes that family test.
-      ...isOpenAIReasoningModel(modelIdForPayload) ? {} : {
-        temperature: request.config?.temperature ?? this.config.temperature
-      },
-      max_output_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+      ...buildOpenAIParameters(
+        modelIdForPayload,
+        request.config?.maxTokens ?? this.config.maxTokens,
+        request.config?.temperature ?? this.config.temperature,
+        this.parameterOptions(request),
+        "responses"
+      ),
       ...responsesTextFormat ? { text: { format: responsesTextFormat } } : {},
       stream
     };
@@ -721,10 +798,26 @@ var OpenAIAdapter = class _OpenAIAdapter {
   }
   async completeWithResponses(request) {
     const client = await this.getClient();
+    throwIfAborted(request.signal);
     const payload = this.buildResponsesPayload(request, false);
     logProviderPayload("openai", "request payload", payload, request.debug);
-    const response = await client.responses.create(payload);
+    if (!client.responses?.create)
+      throw new Error(
+        "This OpenAI client does not support Responses; install an openai version with Responses support."
+      );
+    const response = await client.responses.create(payload, {
+      signal: request.signal
+    });
+    throwIfAborted(request.signal);
     logProviderPayload("openai", "response payload", response, request.debug);
+    if (response.status && response.status !== "completed") {
+      throw Object.assign(
+        new Error(
+          response.error?.message ?? `Responses request ${response.status}`
+        ),
+        response.error
+      );
+    }
     return this.parseResponsesResult(response);
   }
   /**
@@ -742,13 +835,26 @@ var OpenAIAdapter = class _OpenAIAdapter {
    */
   async *streamWithResponses(request) {
     const client = await this.getClient();
+    throwIfAborted(request.signal);
     const payload = this.buildResponsesPayload(request, true);
     logProviderPayload("openai", "request payload", payload, request.debug);
-    const stream = await client.responses.create(payload);
+    if (!client.responses?.create)
+      throw new Error(
+        "This OpenAI client does not support Responses; install an openai version with Responses support."
+      );
+    const stream = await client.responses.create(payload, {
+      signal: request.signal
+    });
     const toolCalls = /* @__PURE__ */ new Map();
     let usage;
     const citations = [];
+    let completed = false;
     function* flushToolCalls() {
+      for (const call of toolCalls.values()) {
+        if (!call.name)
+          throw new Error("Responses tool call is missing its name");
+        JSON.parse(call.arguments || "{}");
+      }
       for (const call of toolCalls.values()) {
         yield { type: "action:start", id: call.id, name: call.name };
         yield { type: "action:args", id: call.id, args: call.arguments };
@@ -757,7 +863,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
       toolCalls.clear();
     }
     for await (const event of stream) {
-      if (request.signal?.aborted) break;
+      throwIfAborted(request.signal);
       logProviderPayload("openai", "stream chunk", event, request.debug);
       const index = event?.output_index ?? 0;
       switch (event?.type) {
@@ -787,8 +893,8 @@ var OpenAIAdapter = class _OpenAIAdapter {
           }
           break;
         }
-        case "response.completed":
-        case "response.incomplete": {
+        case "response.completed": {
+          completed = true;
           const found = this.collectResponsesCitations(
             event.response?.output,
             citations.length
@@ -804,13 +910,27 @@ var OpenAIAdapter = class _OpenAIAdapter {
           }
           break;
         }
-        case "error":
-        case "response.failed":
+        case "response.incomplete":
           throw new Error(
-            event.message ?? event.response?.error?.message ?? "Responses stream failed"
+            `Responses stream incomplete: ${event.response?.incomplete_details?.reason ?? "unknown reason"}`
           );
+        case "error":
+        case "response.failed": {
+          const details = event.response?.error ?? event;
+          throw Object.assign(
+            new Error(details.message ?? "Responses stream failed"),
+            {
+              code: details.code,
+              param: details.param,
+              status: details.status
+            }
+          );
+        }
       }
     }
+    throwIfAborted(request.signal);
+    if (!completed)
+      throw new Error("Responses stream ended without a completed response");
     yield* flushToolCalls();
     if (citations.length > 0) {
       yield { type: "citation", citations: deduplicateCitations(citations) };
@@ -819,6 +939,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
     yield { type: "done", usage };
   }
   async *stream(request) {
+    throwIfAborted(request.signal);
     if (this.shouldUseResponsesApi(request)) {
       const messageId2 = generateMessageId();
       yield { type: "message:start", id: messageId2 };
@@ -826,47 +947,14 @@ var OpenAIAdapter = class _OpenAIAdapter {
         yield* this.streamWithResponses(request);
         return;
       } catch (error) {
-        yield {
-          type: "error",
-          message: error instanceof Error ? error.message : "Unknown error",
-          code: "OPENAI_RESPONSES_ERROR"
-        };
+        yield providerErrorEvent(error, "OPENAI_RESPONSES_ERROR");
         return;
       }
     }
     const client = await this.getClient();
     let messages;
     if (request.rawMessages && request.rawMessages.length > 0) {
-      const processedMessages = request.rawMessages.map((msg) => {
-        if (msg.role === "assistant" && Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0 && msg.content === "") {
-          return { ...msg, content: null };
-        }
-        const hasAttachments = msg.attachments && Array.isArray(msg.attachments) && msg.attachments.length > 0;
-        if (hasAttachments) {
-          const content = [];
-          if (msg.content) {
-            content.push({ type: "text", text: msg.content });
-          }
-          for (const attachment of msg.attachments) {
-            if (attachment.type === "image") {
-              let imageUrl;
-              if (attachment.url) {
-                imageUrl = attachment.url;
-              } else if (attachment.data) {
-                imageUrl = attachment.data.startsWith("data:") ? attachment.data : `data:${attachment.mimeType || "image/png"};base64,${attachment.data}`;
-              } else {
-                continue;
-              }
-              content.push({
-                type: "image_url",
-                image_url: { url: imageUrl, detail: "auto" }
-              });
-            }
-          }
-          return { ...msg, content, attachments: void 0 };
-        }
-        return msg;
-      });
+      const processedMessages = this.formatRawMessages(request.rawMessages);
       if (request.systemPrompt) {
         const hasSystem = processedMessages.some((m) => m.role === "system");
         if (!hasSystem) {
@@ -886,7 +974,16 @@ var OpenAIAdapter = class _OpenAIAdapter {
         request.systemPrompt
       );
     }
-    const tools = request.actions?.length ? formatTools(request.actions) : [];
+    const tools = request.toolDefinitions?.length ? request.toolDefinitions.filter((tool) => tool.available !== false).map((tool) => ({
+      type: "function",
+      function: {
+        name: tool.name,
+        description: tool.description,
+        parameters: normalizeObjectJsonSchema(
+          tool.inputSchema
+        )
+      }
+    })) : formatTools(request.actions ?? []);
     const webSearchConfig = request.webSearch ?? this.config.webSearch;
     if (webSearchConfig) {
       const webSearchTool = {
@@ -915,26 +1012,28 @@ var OpenAIAdapter = class _OpenAIAdapter {
         tools: tools.length > 0 ? tools : void 0,
         tool_choice: tools.length > 0 ? toolChoice : void 0,
         parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
-        ...buildOpenAITokenParams(
+        ...buildOpenAIParameters(
           modelIdForPayload,
           request.config?.maxTokens ?? this.config.maxTokens,
-          request.config?.temperature ?? this.config.temperature
+          request.config?.temperature ?? this.config.temperature,
+          this.parameterOptions(request)
         ),
         response_format: toOpenAIResponseFormat(request.config?.responseFormat),
         stream: true,
         stream_options: { include_usage: true }
       };
       logProviderPayload("openai", "request payload", payload, request.debug);
-      const stream = await client.chat.completions.create(payload);
-      let currentToolCall = null;
+      const stream = await client.chat.completions.create(payload, {
+        signal: request.signal
+      });
+      const toolCalls = /* @__PURE__ */ new Map();
+      let finishReason;
       const collectedCitations = [];
       let citationIndex = 0;
       let usage;
       for await (const chunk of stream) {
         logProviderPayload("openai", "stream chunk", chunk, request.debug);
-        if (request.signal?.aborted) {
-          break;
-        }
+        throwIfAborted(request.signal);
         const delta = chunk.choices[0]?.delta;
         const choice = chunk.choices[0];
         if (delta?.content) {
@@ -957,38 +1056,19 @@ var OpenAIAdapter = class _OpenAIAdapter {
             }
           }
         }
-        if (delta?.tool_calls) {
-          for (const toolCall of delta.tool_calls) {
-            if (toolCall.id) {
-              if (currentToolCall) {
-                yield {
-                  type: "action:args",
-                  id: currentToolCall.id,
-                  args: currentToolCall.arguments
-                };
-                yield {
-                  type: "action:end",
-                  id: currentToolCall.id,
-                  name: currentToolCall.name
-                };
-              }
-              const tcExtraContent = toolCall.extra_content;
-              currentToolCall = {
-                id: toolCall.id,
-                name: toolCall.function?.name || "",
-                arguments: toolCall.function?.arguments || "",
-                ...tcExtraContent ? { extra_content: tcExtraContent } : {}
-              };
-              yield {
-                type: "action:start",
-                id: currentToolCall.id,
-                name: currentToolCall.name,
-                ...currentToolCall.extra_content ? { extra_content: currentToolCall.extra_content } : {}
-              };
-            } else if (currentToolCall && toolCall.function?.arguments) {
-              currentToolCall.arguments += toolCall.function.arguments;
-            }
-          }
+        for (const part of delta?.tool_calls ?? []) {
+          const index = part.index ?? 0;
+          const call = toolCalls.get(index) ?? {
+            id: "",
+            name: "",
+            arguments: ""
+          };
+          if (part.id) call.id = part.id;
+          if (part.function?.name) call.name += part.function.name;
+          if (part.function?.arguments)
+            call.arguments += part.function.arguments;
+          if (part.extra_content) call.extra_content = part.extra_content;
+          toolCalls.set(index, call);
         }
         if (chunk.usage) {
           usage = {
@@ -997,21 +1077,28 @@ var OpenAIAdapter = class _OpenAIAdapter {
             total_tokens: chunk.usage.total_tokens
           };
         }
-        if (choice?.finish_reason) {
-          if (currentToolCall) {
-            yield {
-              type: "action:args",
-              id: currentToolCall.id,
-              args: currentToolCall.arguments
-            };
-            yield {
-              type: "action:end",
-              id: currentToolCall.id,
-              name: currentToolCall.name
-            };
-            currentToolCall = null;
-          }
-        }
+        if (choice?.finish_reason) finishReason = choice.finish_reason;
+      }
+      throwIfAborted(request.signal);
+      if (!finishReason)
+        throw new Error("OpenAI stream ended without a finish reason");
+      if (toolCalls.size && !["tool_calls", "function_call"].includes(finishReason)) {
+        throw new Error("OpenAI stream ended with incomplete tool calls");
+      }
+      for (const call of toolCalls.values()) {
+        if (!call.id || !call.name)
+          throw new Error("OpenAI stream contains an incomplete tool call");
+        JSON.parse(call.arguments || "{}");
+      }
+      for (const call of toolCalls.values()) {
+        yield {
+          type: "action:start",
+          id: call.id,
+          name: call.name,
+          ...call.extra_content ? { extra_content: call.extra_content } : {}
+        };
+        yield { type: "action:args", id: call.id, args: call.arguments };
+        yield { type: "action:end", id: call.id, name: call.name };
       }
       if (collectedCitations.length > 0) {
         const uniqueCitations = deduplicateCitations(collectedCitations);
@@ -1020,26 +1107,18 @@ var OpenAIAdapter = class _OpenAIAdapter {
       yield { type: "message:end" };
       yield { type: "done", usage };
     } catch (error) {
-      yield {
-        type: "error",
-        message: error instanceof Error ? error.message : "Unknown error",
-        code: `${this.provider.toUpperCase()}_ERROR`
-      };
+      yield providerErrorEvent(error, `${this.provider.toUpperCase()}_ERROR`);
     }
   }
   async complete(request) {
+    throwIfAborted(request.signal);
     if (this.shouldUseResponsesApi(request)) {
       return this.completeWithResponses(request);
     }
     const client = await this.getClient();
     let messages;
     if (request.rawMessages && request.rawMessages.length > 0) {
-      const sanitized = request.rawMessages.map((msg) => {
-        if (msg.role === "assistant" && Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0 && msg.content === "") {
-          return { ...msg, content: null };
-        }
-        return msg;
-      });
+      const sanitized = this.formatRawMessages(request.rawMessages);
       if (request.systemPrompt && !sanitized.some((message2) => message2.role === "system")) {
         messages = [
           { role: "system", content: request.systemPrompt },
@@ -1054,7 +1133,16 @@ var OpenAIAdapter = class _OpenAIAdapter {
         request.systemPrompt
       );
     }
-    const tools = request.actions?.length ? formatTools(request.actions) : [];
+    const tools = request.toolDefinitions?.length ? request.toolDefinitions.filter((tool) => tool.available !== false).map((tool) => ({
+      type: "function",
+      function: {
+        name: tool.name,
+        description: tool.description,
+        parameters: normalizeObjectJsonSchema(
+          tool.inputSchema
+        )
+      }
+    })) : formatTools(request.actions ?? []);
     const openaiToolOptions = request.providerToolOptions?.openai;
     const toolChoice = openaiToolOptions?.toolChoice && typeof openaiToolOptions.toolChoice === "object" ? {
       type: "function",
@@ -1069,16 +1157,19 @@ var OpenAIAdapter = class _OpenAIAdapter {
       tools: tools.length > 0 ? tools : void 0,
       tool_choice: tools.length > 0 ? toolChoice : void 0,
       parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
-      ...buildOpenAITokenParams(
+      ...buildOpenAIParameters(
         modelIdForCompletePayload,
         request.config?.maxTokens ?? this.config.maxTokens,
-        request.config?.temperature ?? this.config.temperature
+        request.config?.temperature ?? this.config.temperature,
+        this.parameterOptions(request)
       ),
       response_format: toOpenAIResponseFormat(request.config?.responseFormat),
       stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
-    const response = await client.chat.completions.create(payload);
+    const response = await client.chat.completions.create(payload, {
+      signal: request.signal
+    });
     logProviderPayload("openai", "response payload", response, request.debug);
     const choice = response.choices?.[0];
     const message = choice?.message;