feat(agent): add voice input and photo attachments to assistant

Wire mic through Whisper-compatible transcriptions on LLM_BASE_URL.

Photos upload to MinIO and reach the vision model as base64 image_url parts.

tests/unit/llm-content.test.ts

@@ -0,0 +1,27 @@
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import { buildVisionContentParts, textFromMessageContent } from "../../src/lib/llm/content";
+
+describe("llm content helpers", () => {
+  it("reads plain string content", () => {
+    assert.equal(textFromMessageContent("hello"), "hello");
+  });
+
+  it("joins text parts from multimodal content", () => {
+    assert.equal(
+      textFromMessageContent([
+        { type: "text", text: "first" },
+        { type: "image_url", image_url: { url: "data:image/png;base64,abc" } },
+        { type: "text", text: "second" },
+      ]),
+      "first\nsecond",
+    );
+  });
+
+  it("builds vision parts with fallback prompt when text is empty", () => {
+    const parts = buildVisionContentParts("", ["data:image/jpeg;base64,abc"]);
+    assert.equal(parts.length, 2);
+    assert.equal(parts[0]?.type, "text");
+    assert.equal(parts[1]?.type, "image_url");
+  });
+});