feat: azure adapter and OpenAI Files API uploads

README.md

@@ -18,6 +18,7 @@ Models and Adapters for these providers are shipped out of the box:
 
 - Anthropic
 - OpenAI
+- Azure OpenAI
 - Gemini
 - Vertex AI
 - Tributary
@@ -51,7 +52,7 @@ const calculator = new Tool({
 });
 
 const agent = new Agent({
-  model: openAIModel({ model: "gpt-4.1-mini" }), // Or via shorthand "openai:gpt-4.1-mini"
+  model: openAIModel({ model: "gpt-4.1-mini" }), // Or via shorthand "openai:gpt-4.1-mini", "azure:gpt-4.1-mini"
   instructions: "You are a helpful assistant. Use the calculator when math is needed.",
   tools: [calculator],
 });
@@ -361,6 +362,40 @@ console.log(result.outputText);
 
 This works for file inputs like PDFs, images, and CSVs, depending on the model provider.
 
+### Files API
+
+`openAIModel` and `azureOpenAIModel` accept a `fileStore`. With one set, images and PDFs are uploaded to the provider's
+Files API once and referenced by `file_id` on later calls instead of being sent as base64 every time.
+
+```ts
+import { Agent } from "jsr:@alphaxiv/agents";
+import { openAIModel } from "jsr:@alphaxiv/agents/openai";
+
+const agent = new Agent({
+  model: openAIModel({
+    model: "gpt-5.4-mini",
+    fileStore: {
+      get: (url) => db.getProviderFile(url),
+      set: (url, fileId, expiresAt) => db.upsertProviderFile(url, fileId, expiresAt),
+      delete: (url, fileId) => db.deleteProviderFile(url, fileId),
+    },
+  }),
+  instructions: "You are a helpful assistant.",
+});
+```
+
+The store maps a source URL to a provider file id and lives wherever you keep it:
+
+- `get` returns the row for a URL, or nothing
+- `set` stores a new id and returns the one that won, keeping a live row's id and replacing an expired one
+- `delete` removes a row only while it still holds the id you pass
+
+Rows expire 7 days after upload and are never extended. The adapter only replaces an expired row when that URL comes up
+again, so sweep expired rows on your own schedule and delete their files on the provider with your own client.
+
+If the provider rejects a file id, the adapter uploads again and resends the request inside the same call, so a run sees
+one uninterrupted stream.
+
 ## Tracing
 
 Tracing is optional, but built in. You can:

deno.json

@@ -1,6 +1,6 @@
 {
   "name": "@alphaxiv/agents",
-  "version": "0.6.16",
+  "version": "0.7.0",
   "license": "MIT",
   "fmt": {
     "lineWidth": 120
@@ -20,6 +20,7 @@
     "./openrouter": "./src/adapters/openrouter/adapter.ts",
     "./tributary": "./src/adapters/tributary/adapter.ts",
     "./openai": "./src/adapters/openai/adapter.ts",
+    "./azure-openai": "./src/adapters/azure_openai/adapter.ts",
     "./sid": "./src/adapters/sid/adapter.ts",
 
     "./sid/tools": "./src/adapters/sid/tools.ts"

mod.ts

@@ -82,6 +82,7 @@ export type {
   ChatItemToolUse,
   ChatLike,
   ContextSummaryStartEvent,
+  ProviderFileStore,
   ProviderStreamMetadata,
   StreamItem,
   TokenUsage,

src/adapters/anthropic/adapter.ts

@@ -167,6 +167,7 @@ ${JSON.stringify(structuredOutput.originalJsonSchema, null, 2)}
       // caches the tools with it, and one on the conversation tail caches the turn so far.
       if (cacheControl) applyAnthropicCacheBreakpoint(anthropicHistory, cacheControl);
 
+      yield { type: "request_start", index: 0 };
       const response = client.beta.messages.stream({
         model: options.model,
         // An empty text block is rejected, so a blank prompt stays a bare string (and has nothing to cache).

src/adapters/azure_openai/adapter.ts

@@ -0,0 +1,50 @@
+import { requireEnv } from "../../util.ts";
+import {
+  type OpenResponsesClient,
+  openResponsesModel,
+  type OpenResponsesServiceTier,
+} from "../open_responses/adapter.ts";
+import type { Adapter } from "../adapter.ts";
+import { getOpenAISupportedMimeTypes } from "../openai/mimes.ts";
+import {
+  getModelModalities,
+  type OpenAIModels,
+  resolveOpenAIReasoning,
+  type SupportedReasoningEffort,
+} from "../openai/models.ts";
+import type { ProviderFileStore } from "../../types.ts";
+
+/**
+ * OpenAI models on Azure Foundry, addressed by model name rather than by deployment name.
+ * A deployment named after a model id takes that name over, so never create one.
+ *
+ * With a `fileStore`, files are capped at 50 MB each, and Azure never expires an upload on its own.
+ * The store's sweep is the only thing that deletes them.
+ */
+export function azureOpenAIModel<zO, zI, TModel extends OpenAIModels>(options: {
+  model: TModel;
+  apiKey?: string;
+  endpoint?: string;
+  serviceTier?: OpenResponsesServiceTier;
+  effort?: SupportedReasoningEffort<TModel>;
+  parallelToolCalls?: boolean;
+  client?: OpenResponsesClient;
+  /** When set, images and PDFs are uploaded to the Files API and sent as `file_id`. */
+  fileStore?: ProviderFileStore;
+}): Adapter<zO, zI> {
+  return openResponsesModel({
+    provider: "Azure",
+    model: options.model,
+    supportedMimeTypes: getOpenAISupportedMimeTypes(getModelModalities(options.model)),
+    client: options.client,
+    openAIOptions: options.client ? undefined : {
+      apiKey: options.apiKey ?? requireEnv("AZURE_OPENAI_API_KEY"),
+      baseURL: `${(options.endpoint ?? requireEnv("AZURE_OPENAI_ENDPOINT")).replace(/\/$/, "")}/openai/v1`,
+    },
+    reasoning: resolveOpenAIReasoning(options.model, options.effort),
+    parallelToolCalls: options.parallelToolCalls,
+    serviceTier: options.serviceTier,
+    // Responses on Azure only reads ids uploaded with this purpose, and Azure rejects `expires_after`.
+    files: options.fileStore ? { purpose: "assistants", store: options.fileStore } : undefined,
+  });
+}

src/adapters/google_genai/adapter.ts

@@ -122,6 +122,7 @@ export function googleGenerateContentAPIModel<zO, zI>(
         inlineFiles: options.googleGenAIOptions?.vertexai === true,
       });
 
+      yield { type: "request_start", index: 0 };
       const stream = await client.models.generateContentStream({
         model: options.model,
         contents: googleHistory,

src/adapters/model_resolver.ts

@@ -1,5 +1,6 @@
 import type { Adapter } from "./adapter.ts";
 import { anthropicModel } from "./anthropic/adapter.ts";
+import { azureOpenAIModel } from "./azure_openai/adapter.ts";
 import type { AnthropicModels } from "./anthropic/models.ts";
 import { geminiModel } from "./gemini/adapter.ts";
 import type { GoogleModels } from "./google_genai/models.ts";
@@ -16,7 +17,8 @@ import { openAIModel } from "./openai/adapter.ts";
  *
  * Type-safe prefixes ensure autocomplete works for known model names:
  * - `"anthropic:claude-sonnet-4-5"`
- * - `"openai:gpt-4o"`
+ * - `"openai:gpt-5.6-luna"`
+ * - `"azure:gpt-5.6-luna"`
  * - `"gemini:gemini-2.5-pro"`
  * - `"vertex:gemini-2.5-pro"`
  * - `"openrouter:<any-model-path>"`
@@ -26,6 +28,7 @@ import { openAIModel } from "./openai/adapter.ts";
 export type ModelString =
   | `anthropic:${AnthropicModels}`
   | `openai:${OpenAIModels}`
+  | `azure:${OpenAIModels}`
   | `gemini:${GoogleModels}`
   | `vertex:${GoogleModels}`
   | `openrouter:${OpenRouterModels}`
@@ -60,6 +63,8 @@ export function resolveModel(model: AdapterLike): Adapter<unknown, unknown> {
       return anthropicModel({ model: modelName as AnthropicModels });
     case "openai":
       return openAIModel({ model: modelName as OpenAIModels });
+    case "azure":
+      return azureOpenAIModel({ model: modelName as OpenAIModels });
     case "gemini":
       return geminiModel({ model: modelName as GoogleModels });
     case "vertex":