fix(qoder): preserve user and tool-result images

CHANGELOG.md

@@ -7,6 +7,7 @@ Write each change in both `### English` and `### 中文` under `## Unreleased`.
 
 ### English
 
+- Preserve Qoder user images and image-bearing tool results, emitting tool-result images after their complete ordered tool batch.
 - Bridge Responses custom tools through function calls, restoring custom output/events and replaying tool results. Format rules are descriptive, not grammar-enforced; custom input events are emitted after argument collection.
 
 - Preserve typed upstream stream errors through the OpenAI, Anthropic, and Responses relays so invalid Devin requests do not falsely cool accounts, while transport interruptions remain retryable
@@ -21,6 +22,7 @@ Write each change in both `### English` and `### 中文` under `## Unreleased`.
 
 ### 中文
 
+- 保留 Qoder 用户消息及工具结果中的图片，并在完整、有序的工具结果批次之后发送工具图片。
 - 通过 function 调用桥接 Responses custom 工具，还原 custom 输出与事件并回放工具结果。格式规则仅作为描述传递，不强制执行语法约束；custom 输入事件在参数收集后发送。
 
 - OpenAI、Anthropic 与 Responses 流式转发会保留上游的类型化错误，避免无效的 Devin 请求被错误地冷却账号，同时传输中断仍可重试

worker/src/plaintext.mjs

@@ -235,6 +235,7 @@ function normalizeMessagesForUpstream(messages = []) {
 
   const flushToolResults = () => {
     if (!pendingToolResults.length) return;
+    const images = [];
     const order = new Map(currentBatch.map((id, index) => [id, index]));
     pendingToolResults.sort((left, right) => {
       const leftOrder = order.has(left.toolCallId) ? order.get(left.toolCallId) : Number.MAX_SAFE_INTEGER;
@@ -245,8 +246,12 @@ function normalizeMessagesForUpstream(messages = []) {
       const belongsToCurrentBatch = currentBatch.length > 0
         ? order.has(result.toolCallId)
         : callsById.has(result.toolCallId);
-      if (belongsToCurrentBatch) normalized.push(result.message);
+      if (belongsToCurrentBatch) {
+        normalized.push(result.message);
+        images.push(...result.images);
+      }
     }
+    if (images.length) normalized.push({ role: "user", content: images });
     pendingToolResults = [];
   };
 
@@ -288,14 +293,26 @@ function normalizeMessagesForUpstream(messages = []) {
         tool_call_id: toolCallId,
       };
       if (message?.name) out.name = String(message.name);
-      pendingToolResults.push({ message: out, toolCallId });
+      const images = (Array.isArray(message.content) ? message.content : []).flatMap((part) => {
+        if (part?.type === "image_url") {
+          return normalizeContentForUpstream([part]);
+        }
+        if (part?.type === "image" && typeof part.data === "string" && part.data && (part.mimeType || part.mime_type)) {
+          return [{ type: "image_url", image_url: { url: `data:${part.mimeType || part.mime_type};base64,${part.data}` } }];
+        }
+        return [];
+      });
+      pendingToolResults.push({ message: out, toolCallId, images });
       continue;
     }
 
     flushToolResults();
     currentBatch = [];
     consumedBatchIds = new Set();
-    normalized.push({ role, content: contentToString(message?.content) });
+    normalized.push({
+      role,
+      content: role === "user" ? normalizeContentForUpstream(message?.content) : contentToString(message?.content),
+    });
   }
 
   flushToolResults();

worker/test/plaintext.test.mjs

@@ -36,6 +36,52 @@ test("normalizes model ids only for stable public/settings keys", () => {
   assert.equal(canonicalModelID("GLM-5.2"), "glm-5.2");
 });
 
+test("preserves user images in the final Qoder request without changing system text", () => {
+  const images = [
+    { type: "image_url", image_url: { url: "data:image/png;base64,cHJvYmU=", detail: "high" } },
+    { type: "image_url", image_url: { url: "https://example.com/probe.png" } },
+  ];
+  const body = buildPlainChatBody({
+    model: "deepseek-flash",
+    messages: [
+      { role: "system", content: [{ type: "text", text: "Inspect images" }] },
+      { role: "user", content: [{ type: "text", text: "Compare these" }, ...images] },
+    ],
+  });
+  assert.equal(body.system, "Inspect images");
+  assert.deepEqual(body.messages[1].content, [{ type: "text", text: "Compare these" }, ...images]);
+});
+
+test("preserves tool images after the complete parallel result batch and drops orphan images", () => {
+  const image = { type: "image_url", image_url: { url: "data:image/png;base64,cHJvYmU=" } };
+  const body = buildPlainChatBody({ model: "deepseek-flash", messages: [
+    { role: "assistant", content: "", tool_calls: [
+      { id: "a", function: { name: "view", arguments: "{}" } },
+      { id: "b", function: { name: "view", arguments: "{}" } },
+    ] },
+    { role: "tool", tool_call_id: "orphan", content: [{ type: "image_url", image_url: { url: "orphan" } }] },
+    { role: "tool", tool_call_id: "b", content: [image] },
+    { role: "tool", tool_call_id: "a", content: [{ type: "image", mimeType: "image/png", data: "bWNw" }] },
+    { role: "user", content: "Describe the results" },
+  ] });
+  assert.deepEqual(body.messages.map(x => x.role), ["assistant", "tool", "tool", "user", "user"]);
+  assert.deepEqual(body.messages.slice(1, 3).map(x => x.tool_call_id), ["a", "b"]);
+  assert.deepEqual(body.messages[3].content, [
+    { type: "image_url", image_url: { url: "data:image/png;base64,bWNw" } }, image,
+  ]);
+  assert.equal(JSON.stringify(body).includes('"url":"orphan"'), false);
+});
+
+test("preserves images extracted from Responses tool output into a user message", () => {
+  const image = { type: "image_url", image_url: { url: "data:image/png;base64,cHJvYmU=" } };
+  const body = buildPlainChatBody({ model: "deepseek-flash", messages: [
+    { role: "assistant", content: "", tool_calls: [{ id: "a", function: { name: "view", arguments: "{}" } }] },
+    { role: "tool", tool_call_id: "a", content: "Image attached" },
+    { role: "user", content: [image] },
+  ] });
+  assert.deepEqual(body.messages[2], { role: "user", content: [image] });
+});
+
 test("detects reasoning flags from OpenAI-style fields", () => {
   assert.equal(wantsReasoning({ enable_thinking: true }), true);
   assert.equal(wantsReasoning({ reasoning_effort: "high" }), true);