@jambonz/llm 0.3.4 → 0.4.1

package/dist/test-kit/index.d.cts

@@ -1,4 +1,4 @@
-import { a as AdapterFactory, d as AuthKind, A as AuthSpec, L as LlmAdapter, c as ApiKeyAuth, g as ClientOptions, P as PromptRequest, l as LlmEvent, M as Message, o as ToolCallEvent, n as ModelInfo, b as AdapterManifest } from '../types-4d2QrlAe.cjs';
+import { b as AdapterFactory, e as AuthKind, A as AuthSpec, L as LlmAdapter, d as ApiKeyAuth, h as ClientOptions, P as PromptRequest, m as LlmEvent, M as Message, p as ToolCallEvent, o as ModelInfo, c as AdapterManifest } from '../types-DxknK0Rc.cjs';
 
 /**
  * Named scenarios the harness must be able to stage. Each scenario configures

package/dist/test-kit/index.d.ts

@@ -1,4 +1,4 @@
-import { a as AdapterFactory, d as AuthKind, A as AuthSpec, L as LlmAdapter, c as ApiKeyAuth, g as ClientOptions, P as PromptRequest, l as LlmEvent, M as Message, o as ToolCallEvent, n as ModelInfo, b as AdapterManifest } from '../types-4d2QrlAe.js';
+import { b as AdapterFactory, e as AuthKind, A as AuthSpec, L as LlmAdapter, d as ApiKeyAuth, h as ClientOptions, P as PromptRequest, m as LlmEvent, M as Message, p as ToolCallEvent, o as ModelInfo, c as AdapterManifest } from '../types-DxknK0Rc.js';
 
 /**
  * Named scenarios the harness must be able to stage. Each scenario configures

package/dist/test-kit/index.js

@@ -74,7 +74,10 @@ async function* streamFromOpenAI(client, req, options) {
   }
   if (req.temperature !== void 0) body.temperature = req.temperature;
   const effectiveMaxTokens = req.maxTokens ?? options.defaultMaxTokens;
-  if (effectiveMaxTokens !== void 0) body.max_tokens = effectiveMaxTokens;
+  if (effectiveMaxTokens !== void 0) {
+    const param = options.tokensParam ?? (needsMaxCompletionTokens(req.model) ? "max_completion_tokens" : "max_tokens");
+    body[param] = effectiveMaxTokens;
+  }
   let stream;
   try {
     stream = await client.chat.completions.create(body, { signal: req.signal });
@@ -257,6 +260,9 @@ function capabilitiesFor(modelId, knownModels) {
   if (known) return known.capabilities;
   return defaultCapabilitiesForUnknown(modelId);
 }
+function needsMaxCompletionTokens(modelId) {
+  return /^(o\d|gpt-5)/i.test(modelId);
+}
 function defaultCapabilitiesForUnknown(modelId) {
   const isReasoning = /^o[13]([-/]|$)/i.test(modelId);
   if (isReasoning) {
@@ -1634,7 +1640,12 @@ var VertexOpenAIAdapter = class {
   stream(req) {
     return streamFromOpenAI(this.ensureClient(), req, {
       knownModels: vertexOpenAIManifest.knownModels,
-      defaultMaxTokens: 4096
+      defaultMaxTokens: 4096,
+      // Pin legacy `max_tokens`: Vertex's OpenAI-compat partner-model
+      // endpoint predates `max_completion_tokens` and may not recognize it.
+      // The model-id heuristic wouldn't catch a partner model named in the
+      // OpenAI style (unlikely today, but defensive).
+      tokensParam: "max_tokens"
     });
   }
   appendAssistantToolCall(history, toolCalls) {
@@ -2225,6 +2236,211 @@ var bedrockFactory = {
   create: () => new BedrockAdapter()
 };
 
+// src/adapters/azure-openai/adapter.ts
+import { AzureOpenAI } from "openai";
+
+// src/adapters/azure-openai/manifest.ts
+var AZURE_OPENAI_KNOWN_MODELS = [
+  {
+    id: "gpt-4o",
+    displayName: "GPT-4o (on Azure)",
+    capabilities: {
+      streaming: true,
+      tools: true,
+      vision: true,
+      systemPrompt: true,
+      maxContextTokens: 128e3
+    }
+  },
+  {
+    id: "gpt-4o-mini",
+    displayName: "GPT-4o mini (on Azure)",
+    capabilities: {
+      streaming: true,
+      tools: true,
+      vision: true,
+      systemPrompt: true,
+      maxContextTokens: 128e3
+    }
+  },
+  {
+    id: "gpt-4-turbo",
+    displayName: "GPT-4 Turbo (on Azure)",
+    capabilities: {
+      streaming: true,
+      tools: true,
+      vision: true,
+      systemPrompt: true,
+      maxContextTokens: 128e3
+    }
+  },
+  {
+    // Azure spells it `gpt-35-turbo` (no dot). Keep their spelling for the id
+    // so capability lookups against deployment names that echo the Azure id
+    // hit.
+    id: "gpt-35-turbo",
+    displayName: "GPT-3.5 Turbo (on Azure)",
+    capabilities: {
+      streaming: true,
+      tools: true,
+      vision: false,
+      systemPrompt: true,
+      maxContextTokens: 16385
+    }
+  },
+  {
+    id: "o1-preview",
+    displayName: "o1 preview \u2014 reasoning (on Azure)",
+    capabilities: {
+      streaming: false,
+      tools: false,
+      vision: false,
+      systemPrompt: false,
+      maxContextTokens: 128e3
+    }
+  },
+  {
+    id: "o1-mini",
+    displayName: "o1 mini \u2014 reasoning (on Azure)",
+    capabilities: {
+      streaming: false,
+      tools: false,
+      vision: false,
+      systemPrompt: false,
+      maxContextTokens: 128e3
+    }
+  }
+];
+var azureOpenAIManifest = {
+  vendor: "azure-openai",
+  displayName: "Azure OpenAI",
+  authKinds: [
+    {
+      kind: "azureOpenAIApiKey",
+      displayName: "API Key",
+      fields: [
+        {
+          name: "apiKey",
+          label: "API Key",
+          type: "password",
+          required: true,
+          help: "Azure OpenAI resource key. Find it in the Azure portal under your OpenAI resource \u2192 Keys and Endpoint."
+        },
+        {
+          name: "endpoint",
+          label: "Endpoint",
+          type: "url",
+          required: true,
+          help: "Resource endpoint URL, e.g. https://my-resource.openai.azure.com. Shown in the Azure portal alongside the keys."
+        },
+        {
+          name: "deployment",
+          label: "Deployment Name",
+          type: "text",
+          required: true,
+          help: "The name you chose for the deployment in Azure AI Studio (NOT the underlying model id \u2014 deployments can be named anything)."
+        },
+        {
+          name: "apiVersion",
+          label: "API Version",
+          type: "text",
+          required: true,
+          default: "2024-10-21",
+          help: "Data-plane api-version, e.g. 2024-10-21. Microsoft rolls these frequently; see Azure OpenAI API reference docs for the current stable GA version."
+        }
+      ]
+    }
+  ],
+  knownModels: AZURE_OPENAI_KNOWN_MODELS,
+  // Azure's list-deployments endpoint is on the control plane (ARM) and
+  // requires AAD credentials the data-plane API key doesn't carry. Return
+  // the curated set from the manifest.
+  supportsModelListing: false,
+  docsUrl: "https://learn.microsoft.com/en-us/azure/ai-services/openai/reference"
+};
+
+// src/adapters/azure-openai/adapter.ts
+var AzureOpenAIAdapter = class {
+  vendor = azureOpenAIManifest.vendor;
+  acceptedAuth = ["azureOpenAIApiKey"];
+  client;
+  init(auth, client) {
+    if (auth.kind !== "azureOpenAIApiKey") {
+      throw new Error(
+        `AzureOpenAIAdapter: unsupported auth kind '${auth.kind}'`
+      );
+    }
+    if (!auth.apiKey) {
+      throw new Error("AzureOpenAIAdapter: apiKey is required");
+    }
+    if (!auth.endpoint) {
+      throw new Error("AzureOpenAIAdapter: endpoint is required");
+    }
+    if (!auth.deployment) {
+      throw new Error("AzureOpenAIAdapter: deployment is required");
+    }
+    if (!auth.apiVersion) {
+      throw new Error("AzureOpenAIAdapter: apiVersion is required");
+    }
+    this.client = new AzureOpenAI({
+      apiKey: auth.apiKey,
+      endpoint: auth.endpoint,
+      apiVersion: auth.apiVersion,
+      deployment: auth.deployment,
+      ...client?.timeout !== void 0 ? { timeout: client.timeout } : {},
+      ...client?.maxRetries !== void 0 ? { maxRetries: client.maxRetries } : {}
+    });
+  }
+  stream(req) {
+    return streamFromOpenAI(this.ensureClient(), req, {
+      knownModels: azureOpenAIManifest.knownModels,
+      tokensParam: "max_completion_tokens"
+    });
+  }
+  appendAssistantToolCall(history, toolCalls) {
+    return appendOpenAIAssistantToolCall(history, toolCalls);
+  }
+  appendToolResult(history, toolCallId, result) {
+    return appendOpenAIToolResult(history, toolCallId, result);
+  }
+  async listAvailableModels() {
+    return [...azureOpenAIManifest.knownModels];
+  }
+  async testCredential() {
+    const client = this.ensureClient();
+    await client.chat.completions.create({
+      // Azure ignores `model` when a deployment is in the URL, but the SDK
+      // still requires it as a parameter.
+      model: "probe",
+      messages: [{ role: "user", content: "ping" }],
+      // Use `max_completion_tokens` (not legacy `max_tokens`): gpt-5 family
+      // and o-series reasoning models return 400 on the legacy parameter.
+      // The newer name is a superset — also accepted by gpt-4o, gpt-4-turbo,
+      // and gpt-3.5-turbo deployments — so it's the safe choice for a probe
+      // that has to work regardless of the underlying deployment model.
+      max_completion_tokens: 1,
+      stream: false
+    });
+  }
+  async warmup() {
+  }
+  ensureClient() {
+    if (!this.client) {
+      throw new Error(
+        "AzureOpenAIAdapter: init() must be called before stream()/listAvailableModels()"
+      );
+    }
+    return this.client;
+  }
+};
+
+// src/adapters/azure-openai/factory.ts
+var azureOpenAIFactory = {
+  vendor: azureOpenAIManifest.vendor,
+  manifest: azureOpenAIManifest,
+  create: () => new AzureOpenAIAdapter()
+};
+
 // src/adapters/index.ts
 registerAdapter(openAIFactory);
 registerAdapter(deepseekFactory);
@@ -2233,6 +2449,7 @@ registerAdapter(googleFactory);
 registerAdapter(vertexGeminiFactory);
 registerAdapter(vertexOpenAIFactory);
 registerAdapter(bedrockFactory);
+registerAdapter(azureOpenAIFactory);
 
 // src/index.ts
 async function createLlm(args) {