diff --git a/src/adapters/cursor/protobuf-request.ts b/src/adapters/cursor/protobuf-request.ts
index 39a18bd6eb..18d5957eab 100644
--- a/src/adapters/cursor/protobuf-request.ts
+++ b/src/adapters/cursor/protobuf-request.ts
@@ -188,10 +188,12 @@ function assistantRootText(
 }
 
 // Cursor builds the actual model prompt from rootPromptMessagesJson (turns[] is UI/display metadata),
-// so prior history — including assistant tool calls and tool results — must be replayed here or a
-// ResumeAction has nothing model-visible to continue from. The active user message is excluded
-// because it travels in the action. Tool results are assistant-role text with a [Tool Result]
-// or [Tool Error] marker so Cursor does not wrap them as `<user_query>` (#1992). Each entry is a SHA-256 blob ID.
+// so prior history must be replayed here or a ResumeAction has nothing model-visible to continue from.
+// The active user message is excluded because it travels in the action. When the continuation cannot
+// rely on native MCP turn state, tool results stay assistant-role text with a [Tool Result] /
+// [Tool Error] marker so Cursor does not wrap them as `<user_query>` (#1992). Native resume models
+// already carry the paired MCP result on turns[], so that marker is omitted from root replay — Auto
+// few-shot-mimics it as chat text otherwise. Each entry is a SHA-256 blob ID.
 function rootPromptMessages(request: CursorRunRequest, requestScope: CursorBlobRequestScopeToken): {
   ids: Uint8Array[];
   byteLength: number;
@@ -212,6 +214,7 @@ function rootPromptMessages(request: CursorRunRequest, requestScope: CursorBlobR
   }
 
   const externalModel = isCursorExternalWireModel(request.modelId);
+  const echoToolResultInRoot = cursorNeedsExternalToolContinuation(request.modelId);
   const lastRawIsToolResult = messages.at(-1)?.role === "toolResult";
   const activeUserIndex = lastRawIsToolResult ? -1 : lastActionIndex(messages);
 
@@ -243,6 +246,10 @@ function rootPromptMessages(request: CursorRunRequest, requestScope: CursorBlobR
       }
       // Assistant tool CALLS are intentionally NOT replayed as visible "[Tool Call]" text here.
     } else if (message.role === "toolResult") {
+      // Native resume models already receive the paired MCP result through turns[]. Replaying
+      // the same payload as assistant-role "[Tool Result]" / "[tool_result]" text teaches Auto
+      // to echo that envelope as chat instead of continuing from the structured result.
+      if (!echoToolResultInRoot) continue;
       // #1920: the prefix must reflect the NORMALIZED error state (an empty
       // node_repl result is an error even when the runtime said isError=false).
       const prefix = normalizedToolResult(message, contentToText(message.content)).isError ? "[Tool Error]" : "[Tool Result]";
diff --git a/tests/cursor-blob.test.ts b/tests/cursor-blob.test.ts
index 1139ea79fb..cf37282c80 100644
--- a/tests/cursor-blob.test.ts
+++ b/tests/cursor-blob.test.ts
@@ -913,6 +913,47 @@ describe("Cursor blob handshake", () => {
     const run = msg.message.case === "runRequest" ? msg.message.value : undefined;
 
     expect(run?.action?.action.case).toBe("resumeAction");
+    const roots = decodeRootMessages(bytes) as Array<{ role?: string; content?: unknown }>;
+    const serialized = JSON.stringify(roots);
+    expect(serialized).toContain("read a file");
+    expect(serialized).not.toContain("[Tool Result]");
+    expect(serialized).not.toContain("[tool_result]");
+  });
+
+  test("native Auto Intelligence omits assistant-role [Tool Result] root replay", () => {
+    const bytes = encodeCursorRunRequest({
+      modelId: "auto-intelligence",
+      conversationId: "c-auto-intel",
+      system: ["You are helpful."],
+      messages: [{ role: "tool", content: "[tool_result]\ncall_id: call_1\nname: read_file\nis_error: false\noutput:\ncontents" }],
+      rawMessages: [
+        { role: "user", content: "read a file", timestamp: 1 },
+        {
+          role: "assistant",
+          model: "cursor/auto-intelligence",
+          timestamp: 2,
+          content: [{ type: "toolCall", id: "call_1", name: "read_file", arguments: { path: "a.txt" } }],
+        },
+        { role: "toolResult", toolCallId: "call_1", toolName: "read_file", content: "contents", isError: false, timestamp: 3 },
+      ],
+    });
+    const msg = fromBinary(AgentClientMessageSchema, bytes);
+    const run = msg.message.case === "runRequest" ? msg.message.value : undefined;
+    expect(run?.action?.action.case).toBe("resumeAction");
+    const roots = decodeRootMessages(bytes) as Array<{ role?: string; content?: unknown }>;
+    const serialized = JSON.stringify(roots);
+    expect(roots.some(root => root.role === "assistant")).toBe(false);
+    expect(serialized).not.toContain("[Tool Result]");
+    expect(serialized).not.toContain("[tool_result]");
+    expect(serialized).toContain("read a file");
+    const turnIds = run?.conversationState?.turns ?? [];
+    expect(turnIds).toHaveLength(1);
+    const turn = fromBinary(ConversationTurnStructureSchema, blobData(turnIds[0]!));
+    expect(turn.turn.case).toBe("agentConversationTurn");
+    const steps = turn.turn.case === "agentConversationTurn" ? turn.turn.value.steps : [];
+    expect(steps).toHaveLength(1);
+    const step = fromBinary(ConversationStepSchema, blobData(steps[0]!));
+    expect(step.message.case).toBe("toolCall");
   });
 
   test("drives composer-2.5 tool-result continuations as userMessageAction", () => {
diff --git a/tests/cursor-tool-continuation.test.ts b/tests/cursor-tool-continuation.test.ts
index 6880d741cb..2772245035 100644
--- a/tests/cursor-tool-continuation.test.ts
+++ b/tests/cursor-tool-continuation.test.ts
@@ -39,7 +39,7 @@ describe("363-B: tool result reaches the model via rootPromptMessagesJson", () =
     { role: "toolResult", toolCallId: "call_1", toolName: "read_file", toolNamespace: "mcp__fs", content: "FILE CONTENTS HERE", isError: false, timestamp: 3 },
   ];
 
-  test("tool result text is present in rootPromptMessagesJson, not only in turns[]", () => {
+  test("external-continuation tool result text is present in rootPromptMessagesJson, not only in turns[]", () => {
     const bytes = encodeCursorRunRequest({
       modelId: "composer-2.5",
       conversationId: "c1",
@@ -49,14 +49,29 @@ describe("363-B: tool result reaches the model via rootPromptMessagesJson", () =
     });
     const roots = decodeRoots(bytes);
     const serialized = JSON.stringify(roots);
-    // The model prompt (rootPromptMessagesJson) MUST carry the tool result, or ResumeAction has
-    // nothing model-visible to resume from. Reference: danger-pi buildRootPromptMessagesJson.
+    // composer-2.5 still continues as userMessageAction, so the model prompt must carry the
+    // tool result. Reference: danger-pi buildRootPromptMessagesJson.
     expect(serialized).toContain("FILE CONTENTS HERE");
     expect(serialized).toContain("call_1");
     // The prior user turn must also be replayed (not system-only).
     expect(serialized).toContain("read a file");
   });
 
+  test("native resume models keep tool results on turns[], not as assistant-role root text", () => {
+    const bytes = encodeCursorRunRequest({
+      modelId: "auto-intelligence",
+      conversationId: "c-auto",
+      system: ["You are helpful."],
+      messages: [{ role: "tool", content: "[tool_result]\ncall_id: call_1\nname: mcp__fs__read_file\nis_error: false\noutput:\nFILE CONTENTS HERE" }],
+      rawMessages,
+    });
+    const serialized = JSON.stringify(decodeRoots(bytes));
+    expect(serialized).toContain("read a file");
+    expect(serialized).not.toContain("[Tool Result]");
+    expect(serialized).not.toContain("[tool_result]");
+    expect(serialized).not.toContain("FILE CONTENTS HERE");
+  });
+
   test("rootPromptMessagesJson still leads with the system prompt blob", () => {
     const bytes = encodeCursorRunRequest({
       modelId: "composer-2.5",
@@ -82,11 +97,25 @@ describe("363-B: tool result reaches the model via rootPromptMessagesJson", () =
     // "[Tool Call]" text. The model few-shot-mimics that marker and emits later parallel/mixed tool
     // calls as inert text instead of real tool frames (halting multi-tool continuations).
     expect(serialized).not.toContain("[Tool Call]");
-    // ...but the tool's model-visible continuation context (call id + output) must still survive via
-    // the paired tool RESULT echo, so the model can continue from it.
+    // composer-2.5 still needs the paired tool RESULT echo in the model-visible prompt.
     expect(serialized).toContain("FILE CONTENTS HERE");
     expect(serialized).toContain("call_1");
   });
+
+  test("native resume models do not few-shot [Tool Result] as assistant chat", () => {
+    const bytes = encodeCursorRunRequest({
+      modelId: "composer-2.5-fast",
+      conversationId: "c1",
+      system: ["You are helpful."],
+      messages: [{ role: "tool", content: "[tool_result]\ncall_id: call_1\nname: mcp__fs__read_file\nis_error: false\noutput:\nFILE CONTENTS HERE" }],
+      rawMessages,
+    });
+    const serialized = JSON.stringify(decodeRoots(bytes));
+    expect(serialized).not.toContain("[Tool Call]");
+    expect(serialized).not.toContain("[Tool Result]");
+    expect(serialized).not.toContain("[tool_result]");
+    expect(serialized).toContain("read a file");
+  });
 });
 
 import { create as createPb } from "@bufbuild/protobuf";