fix: preserve image tool results for Claude

src/bridge-pool.ts

@@ -3,12 +3,13 @@
  * (Cursor bridge-pool pattern).
  */
 import type { ClaudeQueryHandle } from "./query.js";
+import type { McpToolResultContent } from "./prompt.js";
 
 export type ParkedToolCall = {
   id: string;
   name: string;
   arguments: string;
-  resolve: (result: string) => void;
+  resolve: (result: McpToolResultContent[]) => void;
   reject: (error: Error) => void;
 };
 

src/prompt.ts

@@ -23,6 +23,10 @@ export type SdkUserPrompt = {
   parent_tool_use_id: null;
 };
 
+export type McpToolResultContent =
+  | { type: "text"; text: string }
+  | { type: "image"; data: string; mimeType: string };
+
 function parseDataUrl(url: string): {
   mediaType: string;
   data: string;
@@ -390,6 +394,31 @@ export function openaiContentToAnthropicBlocks(
   return blocks;
 }
 
+/**
+ * Convert OpenAI-compatible tool result content into MCP result blocks.
+ * OpenCode's read tool returns images as file attachments on the tool result;
+ * those must stay attached when the parked Claude SDK tool call resumes.
+ */
+export function openaiToolResultToMcpContent(
+  content: unknown,
+): McpToolResultContent[] {
+  const result: McpToolResultContent[] = [];
+  for (const block of openaiContentToAnthropicBlocks(content)) {
+    if (block.type === "text") {
+      result.push(block);
+      continue;
+    }
+    if (block.type === "image" && block.source.type === "base64") {
+      result.push({
+        type: "image",
+        data: block.source.data,
+        mimeType: block.source.media_type,
+      });
+    }
+  }
+  return result;
+}
+
 /**
  * Latest user turn as a Claude Agent SDK prompt (string when text-only).
  */

src/proxy.ts

@@ -56,10 +56,12 @@ import {
   buildConversationTranscript,
   extractTextContent,
   latestUserPrompt,
+  openaiToolResultToMcpContent,
   priorMessagesOf,
   promptAsStream,
   withConversationContext,
   type SdkUserPrompt,
+  type McpToolResultContent,
 } from "./prompt.js";
 import {
   detectMetaRequestKind,
@@ -325,11 +327,14 @@ async function handleRequest(req: Request): Promise<Response> {
 
 function collectToolResults(
   messages: OpenAIMessage[],
-): Map<string, string> {
-  const results = new Map<string, string>();
+): Map<string, McpToolResultContent[]> {
+  const results = new Map<string, McpToolResultContent[]>();
   for (const msg of messages) {
     if (msg.role !== "tool" || !msg.tool_call_id) continue;
-    results.set(msg.tool_call_id, extractTextContent(msg.content));
+    results.set(
+      msg.tool_call_id,
+      openaiToolResultToMcpContent(msg.content),
+    );
   }
   return results;
 }
@@ -869,16 +874,18 @@ async function buildOpenCodeMcpServer(
               resolve: () => {},
               reject: () => {},
             };
-            const resultPromise = new Promise<string>((resolve, reject) => {
-              pending.resolve = resolve;
-              pending.reject = reject;
-            });
+            const resultPromise = new Promise<McpToolResultContent[]>(
+              (resolve, reject) => {
+                pending.resolve = resolve;
+                pending.reject = reject;
+              },
+            );
             // Register before notifying so the stream consumer sees the tool.
             pendingTools.set(id, pending);
             onPark();
             const result = await resultPromise;
             return {
-              content: [{ type: "text", text: result }],
+              content: result,
             };
           },
           { alwaysLoad: true },

test/smoke.ts

@@ -354,6 +354,7 @@ async function main() {
   // Multimodal prompt conversion
   const {
     openaiContentToAnthropicBlocks,
+    openaiToolResultToMcpContent,
     latestUserPrompt: buildPrompt,
     contentHasAttachments,
   } = await import("../src/prompt.ts");
@@ -448,6 +449,24 @@ async function main() {
   ]);
   assert.equal(typeof multi === "object" && multi !== null && multi.type === "user", true);
 
+  // OpenCode read-tool attachments must survive the parked Claude SDK tool
+  // bridge. The MCP result shape uses { data, mimeType }, not Anthropic's
+  // message-level { source } shape.
+  const mcpToolResult = openaiToolResultToMcpContent([
+    { type: "text", text: "Image read successfully" },
+    {
+      type: "file",
+      file: {
+        filename: "pixel.png",
+        file_data: `data:image/png;base64,${png}`,
+      },
+    },
+  ]);
+  assert.deepEqual(mcpToolResult, [
+    { type: "text", text: "Image read successfully" },
+    { type: "image", data: png, mimeType: "image/png" },
+  ]);
+
   // Conversation key stability
   const key = conversationKeyFromMessages([
     { role: "user", content: "hello world" },