diff --git a/bug_index.md b/bug_index.md
new file mode 100644
index 0000000000..3200389d1d
--- /dev/null
+++ b/bug_index.md
@@ -0,0 +1,119 @@
+# Bug: Wire Type 4 Protobuf Error on Cursor API
+
+## Symptom
+
+- First request to Cursor (3 msgs, 34 tools) → WORKS
+- Second request (5 msgs, 34 tools, ~140KB) → FAILS with "wire type 4" protobuf parse error
+- Error comes FROM Cursor's server rejecting our encoded protobuf request
+
+## What "wire type 4" means
+
+- Valid protobuf wire types: 0=VARINT, 1=FIXED64, 2=LEN, 5=FIXED32
+- Wire type 4 = "end group" (deprecated, never valid in modern protobuf)
+- This means **corrupted protobuf data** — Cursor server reads data bytes as tag bytes
+- Root cause: a LENGTH-DELIMITED field has wrong length prefix, causing parse misalignment
+
+## Data Flow
+
+```
+Client (Claude format, 5 msgs)
+  → chatCore.translateRequest()
+    → claude-to-openai.ts (claudeToOpenAIRequest)
+    → openai-to-cursor.ts (buildCursorRequest)
+  → cursor.ts transformRequest()
+    → cursorProtobuf.ts generateCursorBody() → protobuf binary
+  → HTTP to Cursor API → "wire type 4" error
+```
+
+## Key Files
+
+- `open-sse/translator/request/claude-to-openai.ts` — Claude→OpenAI message conversion
+- `open-sse/translator/request/openai-to-cursor.ts` — OpenAI→Cursor message conversion
+- `open-sse/utils/cursorProtobuf.ts` — Protobuf binary encoder
+- `open-sse/executors/cursor.ts` — Cursor executor (HTTP + protobuf framing)
+
+## Investigation Done (Session 1)
+
+1. **Content type is NOT the issue** — `encodeField()` handles undefined/null/array gracefully (→ empty `0a 00`)
+2. **openai-to-cursor always normalizes content to string** (lines 77-87)
+3. **tool_calls on assistant msgs are silently ignored** by protobuf encoder (no encoding logic for them)
+4. **Debug logging was added** to `encodeRequest()` and `encodeMessage()` to trace message shapes
+5. **All basic varint/field encoding tests pass** — no corruption in primitives
+
+## Investigation Done (Session 2 — current)
+
+6. **Round-trip test PASSES** — a synthetic 4-msg + 34-tools + tool_results request encodes and decodes back perfectly
+7. **Large payload test PASSES** — 100KB tool result content, frame length matches perfectly (106,810 bytes)
+8. **Varint encoding tests PASS** — values up to 268M encode/decode correctly
+9. **Null bytes in content work** — no corruption
+10. **Missing content on assistant msgs works** — produces valid empty field
+
+## CRITICAL CONCLUSION
+
+The encoding logic itself is **correct for all tested patterns**. The bug is NOT in the encoder primitives.
+The issue MUST be in one of:
+
+### HIGH PRIORITY — What to investigate next
+
+#### H1: The REAL production data has something we haven't tested
+
+The user mentioned the log was already provided in a previous session. The 140,267 byte request
+with 34 tools and 5 messages has specific data that triggers the bug. We need to:
+
+- **Capture the exact translated messages** before protobuf encoding (add JSON.stringify dump)
+- **Capture the hex dump** of the protobuf output
+- **Send the hex dump to a protobuf decoder** to find where parsing breaks
+
+#### H2: assistant message with tool_calls AND content=undefined
+
+In openai-to-cursor.ts line 96-97, when assistant has tool_calls but empty content:
+
+```js
+const assistantMsg = { role: "assistant" }; // NO content property
+if (content) {
+  assistantMsg.content = content;
+} // content="" is falsy → skipped
+```
+
+Result: `msg.content` is `undefined`. In `encodeMessage`, this produces `encodeField(1, LEN, undefined)` → `0a 00`.
+This is technically valid protobuf (empty string field), BUT Cursor's server might require the content field to be present as a non-empty string for assistant messages. The FIRST request (3 msgs) might not have an assistant message at all, while the SECOND (5 msgs) does.
+
+**FIX TO TRY**: Always set `content: ""` on assistant messages in openai-to-cursor.ts.
+
+#### H3: tool_calls being passed to protobuf encoder as stray properties
+
+The Cursor message in the translated body has `{ role, content, tool_calls, tool_results }`.
+The protobuf encoder only uses `content`, `role`, and `tool_results`. The `tool_calls` array
+is silently ignored. BUT if Cursor's real protocol expects tool calls to be encoded differently
+(maybe as part of the content or as a separate protobuf field), we might be sending incomplete data.
+
+#### H4: Encoding `encodeVarint` with value from `>>>` unsigned shift on large numbers
+
+`encodeVarint` uses `value >>>= 7` which treats value as unsigned 32-bit. If a tag or length
+exceeds 2^31, the unsigned shift behavior could produce wrong bytes. However, tests show this
+works correctly up to 268M, and our payloads are ~140KB, so this is unlikely.
+
+## Recommended Fix (try first — LOW RISK)
+
+In `openai-to-cursor.ts`, always ensure assistant messages have `content: ""`:
+
+```js
+// Line 96-99, change from:
+const assistantMsg = { role: "assistant" };
+if (content) {
+  assistantMsg.content = content;
+}
+
+// To:
+const assistantMsg = { role: "assistant", content: content || "" };
+```
+
+## How to debug further
+
+Add this to `cursor.ts` `transformRequest()` BEFORE `generateCursorBody()`:
+
+```js
+console.log("[CURSOR_DEBUG] Messages to encode:", JSON.stringify(messages, null, 2).slice(0, 2000));
+```
+
+Then reproduce the error and check logs for the exact message shapes.
diff --git a/open-sse/executors/cursor.ts b/open-sse/executors/cursor.ts
index cad6ba982d..4f3e619ec7 100644
--- a/open-sse/executors/cursor.ts
+++ b/open-sse/executors/cursor.ts
@@ -230,10 +230,10 @@ export class CursorExecutor extends BaseExecutor {
   }
 
   transformRequest(model, body, stream, credentials) {
-    // Call translator to convert OpenAI format to Cursor format
-    const translatedBody = buildCursorRequest(model, body, stream, credentials);
-    const messages = translatedBody.messages || [];
-    const tools = translatedBody.tools || body.tools || [];
+    // Messages are already translated by chatCore (claude→openai→cursor)
+    // Do NOT call buildCursorRequest again — double-translation drops tool_results
+    const messages = body.messages || [];
+    const tools = body.tools || [];
     const reasoningEffort = body.reasoning_effort || null;
     return generateCursorBody(messages, model, tools, reasoningEffort);
   }
@@ -415,9 +415,18 @@ export class CursorExecutor extends BaseExecutor {
         continue;
       }
 
+      // Check for JSON error frames
       try {
         const text = payload.toString("utf-8");
         if (text.startsWith("{") && text.includes('"error"')) {
+          const hasContent = totalContent || toolCallsMap.size > 0;
+          console.log(
+            `[CURSOR BUFFER] Error frame (hasContent=${hasContent}): ${text.slice(0, 500)}`
+          );
+          // If we already have content, treat error as stream termination (not fatal)
+          if (hasContent) {
+            break;
+          }
           return createErrorResponse(JSON.parse(text));
         }
       } catch {}
@@ -426,6 +435,12 @@ export class CursorExecutor extends BaseExecutor {
       console.log(`[CURSOR DECODED] Frame ${frameCount}:`, result);
 
       if (result.error) {
+        const hasContent = totalContent || toolCallsMap.size > 0;
+        console.log(`[CURSOR BUFFER] Decoded error (hasContent=${hasContent}): ${result.error}`);
+        // If we already have content, treat error as stream termination
+        if (hasContent) {
+          break;
+        }
         return new Response(
           JSON.stringify({
             error: {
@@ -570,9 +585,19 @@ export class CursorExecutor extends BaseExecutor {
         continue;
       }
 
+      // Check for JSON error frames
       try {
         const text = payload.toString("utf-8");
         if (text.startsWith("{") && text.includes('"error"')) {
+          const hasContent = chunks.length > 0 || totalContent || toolCallsMap.size > 0;
+          // Log the full error for debugging
+          console.log(
+            `[CURSOR BUFFER SSE] Error frame (hasContent=${hasContent}): ${text.slice(0, 500)}`
+          );
+          // If we already have content, treat error as stream termination (not fatal)
+          if (hasContent) {
+            break;
+          }
           return createErrorResponse(JSON.parse(text));
         }
       } catch {}
@@ -581,6 +606,14 @@ export class CursorExecutor extends BaseExecutor {
       console.log(`[CURSOR DECODED SSE] Frame ${frameCount}:`, result);
 
       if (result.error) {
+        const hasContent = chunks.length > 0 || totalContent || toolCallsMap.size > 0;
+        console.log(
+          `[CURSOR BUFFER SSE] Decoded error (hasContent=${hasContent}): ${result.error}`
+        );
+        // If we already have content, treat error as stream termination
+        if (hasContent) {
+          break;
+        }
         return new Response(
           JSON.stringify({
             error: {
@@ -718,6 +751,56 @@ export class CursorExecutor extends BaseExecutor {
       `[CURSOR BUFFER SSE] Parsed ${frameCount} frames, toolCallsMap size: ${toolCallsMap.size}, toolCalls array: ${toolCalls.length}`
     );
 
+    // Finalize all remaining tool calls in map (stream may have ended without isLast=true)
+    for (const [id, tc] of toolCallsMap.entries()) {
+      if (!toolCalls.find((t) => t.id === id)) {
+        console.log(
+          `[CURSOR BUFFER SSE] Finalizing incomplete tool call: ${id}, isLast=${tc.isLast}`
+        );
+        const toolCallIndex = toolCalls.length;
+        toolCalls.push({
+          id: tc.id,
+          type: tc.type,
+          index: toolCallIndex,
+          function: {
+            name: tc.function.name,
+            arguments: tc.function.arguments,
+          },
+        });
+
+        // Emit SSE chunk for the finalized tool call if not already emitted
+        if (!chunks.some((c) => c.includes(tc.id))) {
+          chunks.push(
+            `data: ${JSON.stringify({
+              id: responseId,
+              object: "chat.completion.chunk",
+              created,
+              model,
+              choices: [
+                {
+                  index: 0,
+                  delta: {
+                    tool_calls: [
+                      {
+                        index: toolCallIndex,
+                        id: tc.id,
+                        type: "function",
+                        function: {
+                          name: tc.function.name,
+                          arguments: tc.function.arguments,
+                        },
+                      },
+                    ],
+                  },
+                  finish_reason: null,
+                },
+              ],
+            })}\n\n`
+          );
+        }
+      }
+    }
+
     if (chunks.length === 0 && toolCalls.length === 0) {
       chunks.push(
         `data: ${JSON.stringify({
diff --git a/open-sse/translator/request/openai-to-cursor.ts b/open-sse/translator/request/openai-to-cursor.ts
index afc0c0b100..345b427f60 100644
--- a/open-sse/translator/request/openai-to-cursor.ts
+++ b/open-sse/translator/request/openai-to-cursor.ts
@@ -1,19 +1,80 @@
 /**
  * OpenAI to Cursor Request Translator
- * Converts OpenAI messages to Cursor simple format
+ * Converts OpenAI messages to Cursor ask/agent format.
+ *
+ * Important: Cursor can loop when tool outputs are sent via protobuf tool_results
+ * with partial schema mismatches. For stability, tool outputs are represented as
+ * structured text blocks in user messages.
  */
 import { register } from "../registry.ts";
 import { FORMATS } from "../formats.ts";
 
-/**
- * Convert OpenAI messages to Cursor format with native tool_results support
- * - system → user with [System Instructions] prefix
- * - tool → accumulate into tool_results array for next user/assistant message
- * - assistant with tool_calls → keep tool_calls structure (Cursor supports it natively)
- */
+type TextPart = { type?: string; text?: string };
+type ToolUsePart = { type?: string; id?: string; name?: string; input?: unknown };
+type ToolResultPart = { type?: string; tool_use_id?: string; content?: unknown };
+
+function normalizeToolCallId(id: unknown): string {
+  return typeof id === "string" ? id.split("\n")[0] : "";
+}
+
+function extractContent(content: unknown): string {
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    return content
+      .filter((part): part is TextPart => {
+        if (!part || typeof part !== "object") return false;
+        const maybe = part as TextPart;
+        return maybe.type === "text" && typeof maybe.text === "string";
+      })
+      .map((part) => part.text as string)
+      .join("");
+  }
+  return "";
+}
+
+function sanitizeToolResultText(text: string): string {
+  // Strip non-printable control chars that can produce backend request errors.
+  return text.replace(/[\u0000-\u0008\u000B\u000C\u000E-\u001F\u007F]/g, "");
+}
+
+function buildToolResultBlock(toolName: string, toolCallId: string, resultText: string): string {
+  const cleanResult = sanitizeToolResultText(resultText || "");
+  return [
+    "<tool_result>",
+    `<tool_name>${toolName || "tool"}</tool_name>`,
+    `<tool_call_id>${toolCallId || ""}</tool_call_id>`,
+    `<result>${cleanResult}</result>`,
+    "</tool_result>",
+  ].join("\n");
+}
+
 function convertMessages(messages) {
   const result = [];
-  let pendingToolResults = [];
+  // Build a map of tool_call_id -> tool name from assistant tool calls.
+  const toolCallMetaMap = new Map();
+  const rememberToolMeta = (toolCallId: string, toolName: string) => {
+    if (!toolCallId) return;
+    const name = toolName || "tool";
+    toolCallMetaMap.set(toolCallId, { name });
+    const normalized = normalizeToolCallId(toolCallId);
+    if (normalized && normalized !== toolCallId) {
+      toolCallMetaMap.set(normalized, { name });
+    }
+  };
+
+  for (const msg of messages) {
+    if (msg.role === "assistant" && msg.tool_calls) {
+      for (const tc of msg.tool_calls) {
+        rememberToolMeta(tc.id || "", tc.function?.name || "tool");
+      }
+    }
+    if (msg.role === "assistant" && Array.isArray(msg.content)) {
+      for (const part of msg.content as ToolUsePart[]) {
+        if (part?.type !== "tool_use") continue;
+        rememberToolMeta(part.id || "", part.name || "tool");
+      }
+    }
+  }
 
   for (let i = 0; i < messages.length; i++) {
     const msg = messages[i];
@@ -21,83 +82,89 @@ function convertMessages(messages) {
     if (msg.role === "system") {
       result.push({
         role: "user",
-        content: `[System Instructions]\n${msg.content}`,
+        content: `[System Instructions]\n${extractContent(msg.content)}`,
       });
       continue;
     }
 
     if (msg.role === "tool") {
-      let toolContent = "";
-      if (typeof msg.content === "string") {
-        toolContent = msg.content;
-      } else if (Array.isArray(msg.content)) {
-        for (const part of msg.content) {
-          if (part.type === "text") {
-            toolContent += part.text;
-          }
-        }
-      }
-
-      const toolName = msg.name || "tool";
+      const toolContent = extractContent(msg.content);
       const toolCallId = msg.tool_call_id || "";
-
-      // Accumulate tool result
-      pendingToolResults.push({
-        tool_call_id: toolCallId,
-        name: toolName,
-        index: pendingToolResults.length,
-        raw_args: toolContent,
+      const toolMeta = toolCallMetaMap.get(toolCallId) || {};
+      const toolName = msg.name || toolMeta.name || "tool";
+      result.push({
+        role: "user",
+        content: buildToolResultBlock(toolName, toolCallId, toolContent),
       });
       continue;
     }
 
     if (msg.role === "user" || msg.role === "assistant") {
-      let content = "";
-
-      if (typeof msg.content === "string") {
-        content = msg.content;
-      } else if (Array.isArray(msg.content)) {
-        for (const part of msg.content) {
-          if (part.type === "text") {
-            content += part.text;
+      if (msg.role === "user" && Array.isArray(msg.content)) {
+        const parts: string[] = [];
+        for (const block of msg.content as Array<TextPart | ToolResultPart>) {
+          if (!block || typeof block !== "object") continue;
+          if (block.type === "text") {
+            if (typeof (block as TextPart).text === "string") {
+              parts.push((block as TextPart).text || "");
+            }
+            continue;
+          }
+          if (block.type === "tool_result") {
+            const tr = block as ToolResultPart;
+            const toolCallId = tr.tool_use_id || "";
+            const toolMeta =
+              toolCallMetaMap.get(toolCallId) ||
+              toolCallMetaMap.get(normalizeToolCallId(toolCallId));
+            const toolName = toolMeta?.name || "tool";
+            const toolContent = extractContent(tr.content);
+            parts.push(buildToolResultBlock(toolName, toolCallId, toolContent));
           }
         }
+        const joined = parts.filter(Boolean).join("\n");
+        if (joined) result.push({ role: "user", content: joined });
+        continue;
       }
 
-      // Keep tool_calls structure for assistant messages
+      const content = extractContent(msg.content);
+
       if (msg.role === "assistant" && msg.tool_calls && msg.tool_calls.length > 0) {
         const assistantMsg: {
           role: string;
           content?: string;
           tool_calls?: unknown;
-          tool_results?: Array<Record<string, unknown>>;
-        } = { role: "assistant" };
-        if (content) {
-          assistantMsg.content = content;
-        }
-        assistantMsg.tool_calls = msg.tool_calls;
-
-        // Attach pending tool results to assistant message with tool_calls
-        if (pendingToolResults.length > 0) {
-          assistantMsg.tool_results = pendingToolResults;
-          pendingToolResults = [];
-        }
-
+        } = { role: "assistant", content: content || "" };
+        assistantMsg.tool_calls = msg.tool_calls.map((tc) => {
+          const { index, ...rest } = tc || {};
+          return rest;
+        });
         result.push(assistantMsg);
-      } else if (content || pendingToolResults.length > 0) {
-        const msgObj: {
-          role: string;
-          content: string;
-          tool_results?: Array<Record<string, unknown>>;
-        } = { role: msg.role, content: content || "" };
+      } else if (msg.role === "assistant" && Array.isArray(msg.content)) {
+        const extractedToolCalls = (msg.content as ToolUsePart[])
+          .filter((b) => b?.type === "tool_use")
+          .map((b) => ({
+            id: b.id || "",
+            type: "function",
+            function: {
+              name: b.name || "tool",
+              arguments: JSON.stringify(b.input || {}),
+            },
+          }))
+          .filter((tc) => tc.id);
 
-        // Attach pending tool results to this message
-        if (pendingToolResults.length > 0) {
-          msgObj.tool_results = pendingToolResults;
-          pendingToolResults = [];
+        if (extractedToolCalls.length > 0) {
+          result.push({
+            role: "assistant",
+            content: content || "",
+            tool_calls: extractedToolCalls,
+          });
+        } else if (content) {
+          result.push({ role: "assistant", content });
+        }
+      } else {
+        if (content) {
+          result.push({ role: msg.role, content });
         }
-
-        result.push(msgObj);
       }
     }
   }
diff --git a/open-sse/utils/cursorProtobuf.ts b/open-sse/utils/cursorProtobuf.ts
index 39b78e2bf9..f593a6228c 100644
--- a/open-sse/utils/cursorProtobuf.ts
+++ b/open-sse/utils/cursorProtobuf.ts
@@ -28,6 +28,7 @@ const ROLE = { USER: 1, ASSISTANT: 2 };
 const UNIFIED_MODE = { CHAT: 1, AGENT: 2 };
 
 const THINKING_LEVEL = { UNSPECIFIED: 0, MEDIUM: 1, HIGH: 2 };
+const CLIENT_SIDE_TOOL_V2 = { MCP: 19 };
 
 const FIELD = {
   // StreamUnifiedChatRequestWithTools (top level)
@@ -65,6 +66,7 @@ const FIELD = {
   MSG_ID: 13,
   MSG_TOOL_RESULTS: 18,
   MSG_IS_AGENTIC: 29,
+  MSG_SERVER_BUBBLE_ID: 32,
   MSG_UNIFIED_MODE: 47,
   MSG_SUPPORTED_TOOLS: 51,
 
@@ -74,6 +76,28 @@ const FIELD = {
   TOOL_RESULT_INDEX: 3,
   TOOL_RESULT_RAW_ARGS: 5,
   TOOL_RESULT_RESULT: 8,
+  TOOL_RESULT_TOOL_CALL: 11,
+  TOOL_RESULT_MODEL_CALL_ID: 12,
+
+  // ClientSideToolV2Result (nested inside ToolResult.result)
+  CLIENT_RESULT_TOOL: 1,
+  CLIENT_RESULT_MCP_RESULT: 28,
+  CLIENT_RESULT_TOOL_CALL_ID: 35,
+  CLIENT_RESULT_MODEL_CALL_ID: 48,
+  CLIENT_RESULT_TOOL_INDEX: 49,
+
+  // MCPResult (nested inside ClientSideToolV2Result.mcp_result)
+  MCP_RESULT_SELECTED_TOOL: 1,
+  MCP_RESULT_RESULT: 2,
+
+  // ClientSideToolV2Call (nested inside ToolResult.tool_call)
+  CLIENT_CALL_TOOL: 1,
+  CLIENT_CALL_MCP_PARAMS: 27,
+  CLIENT_CALL_TOOL_CALL_ID: 3,
+  CLIENT_CALL_NAME: 9,
+  CLIENT_CALL_RAW_ARGS: 10,
+  CLIENT_CALL_TOOL_INDEX: 48,
+  CLIENT_CALL_MODEL_CALL_ID: 49,
 
   // Model
   MODEL_NAME: 1,
@@ -120,6 +144,7 @@ const FIELD = {
   TOOL_NAME: 9,
   TOOL_RAW_ARGS: 10,
   TOOL_IS_LAST: 11,
+  TOOL_IS_LAST_ALT: 15,
   TOOL_MCP_PARAMS: 27,
 
   // MCPParams
@@ -167,6 +192,13 @@ export function encodeField(fieldNum, wireType, value) {
   const tagBytes = encodeVarint(tag);
 
   if (wireType === WIRE_TYPE.VARINT) {
+    // Validate: VARINT value must be a number
+    if (typeof value !== "number" || !Number.isFinite(value) || value < 0) {
+      log(
+        "ENCODE",
+        `⚠️ VARINT field=${fieldNum} invalid value: type=${typeof value}, value=${value}`
+      );
+    }
     const valueBytes = encodeVarint(value);
     return concatArrays(tagBytes, valueBytes);
   }
@@ -202,16 +234,149 @@ function concatArrays(...arrays) {
 // ==================== MESSAGE ENCODING ====================
 
 export function encodeToolResult(toolResult) {
-  const toolCallId = toolResult.tool_call_id || "";
-  const toolName = toolResult.name || "";
-  const toolIndex = toolResult.index || 0;
+  const { toolCallId, modelCallId } = parseToolCallId(toolResult.tool_call_id || "");
+  const rawToolName = toolResult.name || "";
+  const toolName = formatCursorToolName(rawToolName);
+  const { selectedTool, serverName } = parseCursorToolName(toolName);
+  const toolIndex = toolResult.index > 0 ? toolResult.index : 1;
   const rawArgs = toolResult.raw_args || "{}";
+  const resultContent = toolResult.result || "";
+  const encodedResultMessage = encodeClientSideToolResult(
+    toolCallId,
+    modelCallId,
+    selectedTool,
+    toolIndex,
+    resultContent
+  );
+  const encodedToolCallMessage = encodeClientSideToolCall(
+    toolCallId,
+    modelCallId,
+    toolName,
+    selectedTool,
+    serverName,
+    rawArgs,
+    toolIndex
+  );
 
   return concatArrays(
     encodeField(FIELD.TOOL_RESULT_CALL_ID, WIRE_TYPE.LEN, toolCallId),
     encodeField(FIELD.TOOL_RESULT_NAME, WIRE_TYPE.LEN, toolName),
     encodeField(FIELD.TOOL_RESULT_INDEX, WIRE_TYPE.VARINT, toolIndex),
-    encodeField(FIELD.TOOL_RESULT_RAW_ARGS, WIRE_TYPE.LEN, rawArgs)
+    ...(modelCallId
+      ? [encodeField(FIELD.TOOL_RESULT_MODEL_CALL_ID, WIRE_TYPE.LEN, modelCallId)]
+      : []),
+    encodeField(FIELD.TOOL_RESULT_RAW_ARGS, WIRE_TYPE.LEN, rawArgs),
+    ...(encodedResultMessage
+      ? [encodeField(FIELD.TOOL_RESULT_RESULT, WIRE_TYPE.LEN, encodedResultMessage)]
+      : []),
+    encodeField(FIELD.TOOL_RESULT_TOOL_CALL, WIRE_TYPE.LEN, encodedToolCallMessage)
+  );
+}
+
+function parseToolCallId(toolCallIdRaw) {
+  if (typeof toolCallIdRaw !== "string" || toolCallIdRaw.length === 0) {
+    return { toolCallId: "", modelCallId: null };
+  }
+  const delimiter = "\nmc_";
+  const idx = toolCallIdRaw.indexOf(delimiter);
+  if (idx >= 0) {
+    return {
+      toolCallId: toolCallIdRaw.slice(0, idx),
+      modelCallId: toolCallIdRaw.slice(idx + delimiter.length),
+    };
+  }
+  return { toolCallId: toolCallIdRaw, modelCallId: null };
+}
+
+function formatCursorToolName(rawName) {
+  const base = typeof rawName === "string" && rawName.length > 0 ? rawName : "tool";
+
+  if (base.startsWith("mcp__")) {
+    const rest = base.slice("mcp__".length);
+    const splitIdx = rest.indexOf("__");
+    if (splitIdx >= 0) {
+      const server = rest.slice(0, splitIdx) || "custom";
+      const name = rest.slice(splitIdx + 2) || "tool";
+      return `mcp_${server}_${name}`;
+    }
+    return `mcp_custom_${rest || "tool"}`;
+  }
+
+  if (base.startsWith("mcp_")) return base;
+  return `mcp_custom_${base}`;
+}
+
+function parseCursorToolName(formattedName) {
+  if (typeof formattedName !== "string" || !formattedName.startsWith("mcp_")) {
+    return { serverName: "custom", selectedTool: formattedName || "tool" };
+  }
+
+  const tail = formattedName.slice("mcp_".length);
+  const splitIdx = tail.indexOf("_");
+  if (splitIdx < 0) {
+    return { serverName: "custom", selectedTool: tail || "tool" };
+  }
+
+  return {
+    serverName: tail.slice(0, splitIdx) || "custom",
+    selectedTool: tail.slice(splitIdx + 1) || "tool",
+  };
+}
+
+function encodeClientSideToolResult(toolCallId, modelCallId, toolName, toolIndex, resultContent) {
+  const outputText = typeof resultContent === "string" ? resultContent : "";
+  const selectedTool = typeof toolName === "string" && toolName.length > 0 ? toolName : "tool";
+
+  const mcpResult = concatArrays(
+    encodeField(FIELD.MCP_RESULT_SELECTED_TOOL, WIRE_TYPE.LEN, selectedTool),
+    encodeField(FIELD.MCP_RESULT_RESULT, WIRE_TYPE.LEN, outputText)
+  );
+
+  return concatArrays(
+    encodeField(FIELD.CLIENT_RESULT_TOOL, WIRE_TYPE.VARINT, CLIENT_SIDE_TOOL_V2.MCP),
+    encodeField(FIELD.CLIENT_RESULT_MCP_RESULT, WIRE_TYPE.LEN, mcpResult),
+    ...(toolCallId
+      ? [encodeField(FIELD.CLIENT_RESULT_TOOL_CALL_ID, WIRE_TYPE.LEN, toolCallId)]
+      : []),
+    ...(modelCallId
+      ? [encodeField(FIELD.CLIENT_RESULT_MODEL_CALL_ID, WIRE_TYPE.LEN, modelCallId)]
+      : []),
+    encodeField(FIELD.CLIENT_RESULT_TOOL_INDEX, WIRE_TYPE.VARINT, toolIndex)
+  );
+}
+
+function encodeMcpParamsForCall(toolName, rawArgs, serverName) {
+  const tool = concatArrays(
+    encodeField(FIELD.MCP_TOOL_NAME, WIRE_TYPE.LEN, toolName || "tool"),
+    encodeField(FIELD.MCP_TOOL_PARAMS, WIRE_TYPE.LEN, rawArgs || "{}"),
+    encodeField(FIELD.MCP_TOOL_SERVER, WIRE_TYPE.LEN, serverName || "custom")
+  );
+  return encodeField(FIELD.MCP_TOOLS_LIST, WIRE_TYPE.LEN, tool);
+}
+
+function encodeClientSideToolCall(
+  toolCallId,
+  modelCallId,
+  toolName,
+  selectedTool,
+  serverName,
+  rawArgs,
+  toolIndex
+) {
+  return concatArrays(
+    encodeField(FIELD.CLIENT_CALL_TOOL, WIRE_TYPE.VARINT, CLIENT_SIDE_TOOL_V2.MCP),
+    encodeField(
+      FIELD.CLIENT_CALL_MCP_PARAMS,
+      WIRE_TYPE.LEN,
+      encodeMcpParamsForCall(selectedTool, rawArgs, serverName)
+    ),
+    ...(toolCallId ? [encodeField(FIELD.CLIENT_CALL_TOOL_CALL_ID, WIRE_TYPE.LEN, toolCallId)] : []),
+    encodeField(FIELD.CLIENT_CALL_NAME, WIRE_TYPE.LEN, toolName || "tool"),
+    encodeField(FIELD.CLIENT_CALL_RAW_ARGS, WIRE_TYPE.LEN, rawArgs || "{}"),
+    encodeField(FIELD.CLIENT_CALL_TOOL_INDEX, WIRE_TYPE.VARINT, toolIndex > 0 ? toolIndex : 1),
+    ...(modelCallId
+      ? [encodeField(FIELD.CLIENT_CALL_MODEL_CALL_ID, WIRE_TYPE.LEN, modelCallId)]
+      : [])
   );
 }
 
@@ -224,6 +389,20 @@ export function encodeMessage(
   hasTools = false,
   toolResults = []
 ) {
+  // Debug: validate content type before encoding
+  if (
+    content !== null &&
+    content !== undefined &&
+    typeof content !== "string" &&
+    !(content instanceof Uint8Array) &&
+    !Buffer.isBuffer(content)
+  ) {
+    log(
+      "ENCODE",
+      `⚠️ MSG_CONTENT unexpected type: ${typeof content}, isArray=${Array.isArray(content)}, value=${JSON.stringify(content).slice(0, 200)}`
+    );
+  }
+
   return concatArrays(
     encodeField(FIELD.MSG_CONTENT, WIRE_TYPE.LEN, content),
     encodeField(FIELD.MSG_ROLE, WIRE_TYPE.VARINT, role),
@@ -311,13 +490,86 @@ export function encodeRequest(messages, modelName, tools = [], reasoningEffort =
   const isAgentic = hasTools;
   const formattedMessages = [];
   const messageIds = [];
+  const normalizedMessages = [];
 
-  // Prepare messages
+  // Guardrail: split mixed assistant payload into separate assistant messages.
   for (let i = 0; i < messages.length; i++) {
     const msg = messages[i];
+    const hasToolCalls = Array.isArray(msg?.tool_calls) && msg.tool_calls.length > 0;
+    const hasToolResults = Array.isArray(msg?.tool_results) && msg.tool_results.length > 0;
+
+    if (msg?.role === "assistant" && hasToolCalls && hasToolResults) {
+      log(
+        "ENCODE",
+        `normalizing mixed assistant tool payload at msg[${i}] (calls=${msg.tool_calls.length}, results=${msg.tool_results.length})`
+      );
+
+      // Keep assistant tool call message without embedded results
+      normalizedMessages.push({
+        ...msg,
+        tool_results: [],
+      });
+
+      // Avoid inserting duplicate assistant tool-result message if next one already matches
+      const nextMsg = messages[i + 1];
+      const nextHasToolResults =
+        nextMsg?.role === "assistant" &&
+        Array.isArray(nextMsg?.tool_results) &&
+        nextMsg.tool_results.length > 0;
+      const currentIds = new Set(
+        msg.tool_results.map((tr) => tr?.tool_call_id).filter((id) => typeof id === "string")
+      );
+      const nextIds = new Set(
+        (nextMsg?.tool_results || [])
+          .map((tr) => tr?.tool_call_id)
+          .filter((id) => typeof id === "string")
+      );
+      const sameIds =
+        currentIds.size > 0 &&
+        currentIds.size === nextIds.size &&
+        [...currentIds].every((id) => nextIds.has(id));
+
+      if (!(nextHasToolResults && sameIds)) {
+        normalizedMessages.push({
+          role: "assistant",
+          content: "",
+          tool_results: msg.tool_results,
+        });
+      }
+
+      continue;
+    }
+
+    normalizedMessages.push(msg);
+  }
+
+  // Prepare messages
+  for (let i = 0; i < normalizedMessages.length; i++) {
+    const msg = normalizedMessages[i];
     const role = msg.role === "user" ? ROLE.USER : ROLE.ASSISTANT;
     const msgId = uuidv4();
-    const isLast = i === messages.length - 1;
+    const isLast = i === normalizedMessages.length - 1;
+
+    // Debug: log message shape for diagnosis
+    const toolResultsCount = Array.isArray(msg.tool_results) ? msg.tool_results.length : 0;
+    const contentStr = typeof msg.content === "string" ? msg.content : "";
+    const contentPreview = contentStr
+      .slice(0, 120)
+      .replace(/\r/g, "\\r")
+      .replace(/\n/g, "\\n")
+      .replace(/[^\x20-\x7E]/g, "?");
+    log(
+      "ENCODE",
+      `msg[${i}] role=${msg.role} contentType=${typeof msg.content} contentLen=${contentStr.length} contentPreview="${contentPreview}" contentIsArray=${Array.isArray(msg.content)} hasToolCalls=${Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0} hasToolResults=${toolResultsCount > 0} toolResultsCount=${toolResultsCount}`
+    );
+    if (toolResultsCount > 0) {
+      for (const tr of msg.tool_results) {
+        log(
+          "ENCODE",
+          `  toolResult: callId=${tr.tool_call_id} name=${tr.name} rawArgsType=${typeof tr.raw_args} rawArgsLen=${tr.raw_args?.length || 0} resultType=${typeof tr.result} resultLen=${tr.result?.length || 0}`
+        );
+      }
+    }
 
     formattedMessages.push({
       content: msg.content,
@@ -535,8 +787,7 @@ function extractToolCall(toolCallData) {
 
   // Extract tool call ID
   if (toolCall.has(FIELD.TOOL_ID)) {
-    const fullId = new TextDecoder().decode(toolCall.get(FIELD.TOOL_ID)[0].value);
-    toolCallId = fullId.split("\n")[0]; // Cursor returns multi-line ID, take first line
+    toolCallId = new TextDecoder().decode(toolCall.get(FIELD.TOOL_ID)[0].value);
   }
 
   // Extract tool name
@@ -547,6 +798,8 @@ function extractToolCall(toolCallData) {
   // Extract is_last flag
   if (toolCall.has(FIELD.TOOL_IS_LAST)) {
     isLast = toolCall.get(FIELD.TOOL_IS_LAST)[0].value !== 0;
+  } else if (toolCall.has(FIELD.TOOL_IS_LAST_ALT)) {
+    isLast = toolCall.get(FIELD.TOOL_IS_LAST_ALT)[0].value !== 0;
   }
 
   // Extract MCP params - nested real tool info