作答时把图内联成 base64，不再把 S3 链接交给模型

.env.local.example

@@ -17,8 +17,8 @@ KNOWHERE_BASE_URL=https://api.knowhereto.ai
 # Vercel AI Gateway key; AI SDK picks it up automatically
 AI_GATEWAY_API_KEY=vck_your_key_here
 
-# Optional override — defaults to deepseek/deepseek-v4-flash
-# CHAT_MODEL=deepseek/deepseek-v4-flash
+# Optional override — defaults to DeepSeek V4.1 Flash (native multimodal)
+# CHAT_MODEL=deepseek/deepseek-v4.1-flash
 
 # --- Auth (server-side only) ---
 #

src/agent-harness/answer-context.test.ts

@@ -1,13 +1,23 @@
 import { describe, expect, it, vi } from "vitest"
 
+import { logger } from "@/lib/logger"
+
 import { composeAnswerContext } from "./answer-context"
 import type { EvidenceLedgerSnapshot } from "./types"
 
+const revenuePng = new Uint8Array([1, 2, 3])
+const marginPng = new Uint8Array([4, 5, 6])
+
 describe("composeAnswerContext", () => {
-  it("composes retained chunks, full table HTML, images, memory, and the original question", async () => {
+  it("composes retained chunks, full table HTML, images at placeholders, memory, and the original question", async () => {
     const readTableHtml = vi.fn().mockResolvedValue(
       "<table><tr><td>Q4</td><td>24.9</td></tr></table>",
     )
+    const readImage = vi.fn(async (assetUrl: string) =>
+      assetUrl.includes("revenue.png")
+        ? { body: revenuePng, mediaType: "image/png" }
+        : { body: marginPng, mediaType: "image/png" },
+    )
     const message = await composeAnswerContext({
       ledger: makeLedger(),
       memoryItems: [
@@ -20,11 +30,18 @@ describe("composeAnswerContext", () => {
       ],
       userText: "Q4 的收入是多少？",
       readTableHtml,
+      readImage,
     })
 
     expect(readTableHtml).toHaveBeenCalledWith(
       "https://assets.example/tables/revenue.html?signature=valid",
     )
+    expect(readImage).toHaveBeenCalledWith(
+      "https://assets.example/images/revenue.png?signature=valid",
+    )
+    expect(readImage).toHaveBeenCalledWith(
+      "https://assets.example/images/margin.png?signature=valid",
+    )
     expect(message).toEqual({
       role: "user",
       content: [
@@ -38,24 +55,29 @@ describe("composeAnswerContext", () => {
             "Revenue increased in Q4.",
             "<table><tr><td>Q4</td><td>24.9</td></tr></table>",
             "",
-            '[image ref="asset:r1:result:1:image_chunk_1"]',
+            "",
           ].join("\n"),
         },
         {
           type: "image",
-          image: new URL("https://assets.example/images/revenue.png?signature=valid"),
+          image: revenuePng,
+          mediaType: "image/png",
         },
         {
           type: "text",
-          text: '[image ref="asset:r1:result:1:image_chunk_2"]',
+          text: "\n\n\n",
         },
         {
           type: "image",
-          image: new URL("https://assets.example/images/margin.png?signature=valid"),
+          image: marginPng,
+          mediaType: "image/png",
         },
         {
           type: "text",
           text: [
+            "",
+            "",
+            "",
             "## Fluid Memory",
             "",
             '[memory ref="mem:1" kind="stance"]',
@@ -69,6 +91,90 @@ describe("composeAnswerContext", () => {
     })
   })
 
+  it("drops image placeholders and warns when the image URL cannot be read", async () => {
+    const warn = vi.spyOn(logger, "warn").mockImplementation(() => {})
+    const message = await composeAnswerContext({
+      ledger: makeLedger({
+        content:
+          "Before [images/revenue.png] after [images/margin.png] end.",
+        assets: makeLedger().assets.filter((asset) => asset.type === "image"),
+      }),
+      memoryItems: [],
+      userText: "图在哪？",
+      readImage: vi.fn().mockRejectedValue(new Error("Unable to read image: 403")),
+    })
+
+    const parts = userContentParts(message)
+    const text = parts
+      .filter((part) => part.type === "text")
+      .map((part) => part.text)
+      .join("")
+    expect(text).toContain("Before  after  end.")
+    expect(text).not.toContain("[images/revenue.png]")
+    expect(text).not.toContain("[images/margin.png]")
+    expect(parts.some((part) => part.type === "image")).toBe(false)
+    expect(warn).toHaveBeenCalledWith(
+      "chat: skipped unreachable evidence asset",
+      expect.objectContaining({
+        type: "image",
+        reason: "Unable to read image: 403",
+      }),
+    )
+    warn.mockRestore()
+  })
+
+  it("drops table placeholders and warns when the table URL cannot be read", async () => {
+    const warn = vi.spyOn(logger, "warn").mockImplementation(() => {})
+    const message = await composeAnswerContext({
+      ledger: makeLedger({
+        content: "Revenue [Table: tables/revenue.html] rose.",
+        assets: makeLedger().assets.filter((asset) => asset.type === "table"),
+      }),
+      memoryItems: [],
+      userText: "收入多少？",
+      readTableHtml: vi.fn().mockRejectedValue(new Error("Unable to read table HTML: 404")),
+    })
+
+    const text = userContentParts(message)
+      .filter((part) => part.type === "text")
+      .map((part) => part.text)
+      .join("")
+    expect(text).toContain("Revenue  rose.")
+    expect(text).not.toContain("[Table: tables/revenue.html]")
+    expect(text).not.toContain("<table>")
+    expect(warn).toHaveBeenCalledWith(
+      "chat: skipped unreachable evidence asset",
+      expect.objectContaining({
+        type: "table",
+        reason: "Unable to read table HTML: 404",
+      }),
+    )
+    warn.mockRestore()
+  })
+
+  it("throws when a connected image placeholder is missing", async () => {
+    await expect(
+      composeAnswerContext({
+        ledger: {
+          ...makeLedger(),
+          chunks: [
+            {
+              ...makeLedger().chunks[0]!,
+              content: "Revenue increased in Q4.\n[Table: tables/revenue.html]",
+            },
+          ],
+        },
+        memoryItems: [],
+        userText: "Q4 的收入是多少？",
+        readTableHtml: vi.fn().mockResolvedValue("<table></table>"),
+        readImage: vi.fn().mockResolvedValue({
+          body: revenuePng,
+          mediaType: "image/png",
+        }),
+      }),
+    ).rejects.toThrow("Image placeholder")
+  })
+
   it("omits knowledge and memory sections when neither search ran", async () => {
     const message = await composeAnswerContext({
       ledger: { ...makeLedger(), retrievalCount: 0, retainedPicks: [] },
@@ -132,14 +238,31 @@ describe("composeAnswerContext", () => {
   })
 })
 
-function makeLedger(): EvidenceLedgerSnapshot {
+function userContentParts(message: { readonly content: unknown }): readonly {
+  readonly type: string
+  readonly text?: string
+}[] {
+  if (!Array.isArray(message.content)) {
+    throw new Error("expected composed message content to be an array")
+  }
+  return message.content
+}
+
+function makeLedger(
+  overrides: {
+    readonly content?: string
+    readonly assets?: EvidenceLedgerSnapshot["assets"]
+  } = {},
+): EvidenceLedgerSnapshot {
   return {
     retrievalCount: 1,
     chunks: [
       {
         ref: "r1:result:1",
         kind: "result",
-        content: "Revenue increased in Q4.\n[Table: tables/revenue.html]",
+        content:
+          overrides.content ??
+          "Revenue increased in Q4.\n[Table: tables/revenue.html]\n[images/revenue.png]\n[images/margin.png]",
         contentPreview: "Revenue increased in Q4. [Table: tables/revenue.html]",
         chunkType: "text",
         score: 0.9,
@@ -150,7 +273,7 @@ function makeLedger(): EvidenceLedgerSnapshot {
         },
       },
     ],
-    assets: [
+    assets: overrides.assets ?? [
       {
         ref: "asset:r1:result:1:table_chunk",
         chunkRef: "r1:result:1",