Add changelog entry for PR #9253

changelog.txt

@@ -0,0 +1 @@
+Add message normalization for Ollama model compatibility with MCP tool calling (fixes Mistral Large 3, Ministral 3, Gemma3 27B)

extensions/cli/src/stream/streamChatResponse.ts

@@ -19,6 +19,7 @@ import {
   withExponentialBackoff,
 } from "../util/exponentialBackoff.js";
 import { logger } from "../util/logger.js";
+import { normalizeMessagesForModel } from "../util/messageNormalizer.js";
 import { validateContextLength } from "../util/tokenizer.js";
 
 import { getRequestTools, handleToolCalls } from "./handleToolCalls.js";
@@ -257,6 +258,15 @@ export async function processStreamingResponse(
     chatHistory,
     systemMessage,
   ) as ChatCompletionMessageParam[];
+
+  // Normalize messages for model-specific compatibility (GitHub Issue #9249)
+  // Fixes: Mistral "Unexpected role 'system' after role 'tool'"
+  //        Gemma "Invalid 'tool_calls': unknown variant 'index'"
+  const normalizedMessages = normalizeMessagesForModel(
+    openaiChatHistory,
+    model.model,
+  );
+
   const requestStartTime = Date.now();
 
   const streamFactory = async (retryAbortSignal: AbortSignal) => {
@@ -269,7 +279,7 @@ export async function processStreamingResponse(
       llmApi,
       {
         model: model.model,
-        messages: openaiChatHistory,
+        messages: normalizedMessages,
         stream: true,
         tools,
         ...getDefaultCompletionOptions(model.defaultCompletionOptions),

extensions/cli/src/util/messageNormalizer.ts

@@ -0,0 +1,106 @@
+/**
+ * Message Normalization for Model-Specific Compatibility
+ *
+ * Handles model-specific message formatting quirks to ensure compatibility
+ * across different LLM providers when using Ollama's OpenAI endpoint.
+ *
+ * Issues addressed:
+ * 1. Mistral/Ministral: "Unexpected role 'system' after role 'tool'"
+ * 2. Gemma3: "Invalid 'tool_calls': unknown variant 'index'"
+ *
+ * GitHub Issue: https://github.com/continuedev/continue/issues/9249
+ */
+
+import type { ChatCompletionMessageParam } from "openai/resources";
+
+/**
+ * Normalize message list for model-specific requirements.
+ *
+ * @param messages - List of OpenAI-format messages
+ * @param modelName - Model identifier (e.g., 'mistral-large-3:675b-cloud')
+ * @returns Normalized message list safe for the target model
+ */
+export function normalizeMessagesForModel(
+  messages: ChatCompletionMessageParam[],
+  modelName: string,
+): ChatCompletionMessageParam[] {
+  const modelLower = modelName.toLowerCase();
+
+  // Detect model family and apply appropriate normalization
+  if (modelLower.includes("mistral") || modelLower.includes("ministral")) {
+    return normalizeForMistral(messages);
+  } else if (modelLower.includes("gemma")) {
+    return normalizeForGemma(messages);
+  }
+
+  // No normalization needed for other models
+  return messages;
+}
+
+/**
+ * Fix Mistral's "Unexpected role 'system' after role 'tool'" error.
+ *
+ * Strategy: Move system messages before any tool interactions.
+ * If system message appears after tool, convert to user message.
+ */
+function normalizeForMistral(
+  messages: ChatCompletionMessageParam[],
+): ChatCompletionMessageParam[] {
+  const normalized: ChatCompletionMessageParam[] = [];
+  const systemMessages: ChatCompletionMessageParam[] = [];
+  let hasToolInteraction = false;
+
+  for (const msg of messages) {
+    const role = msg.role;
+
+    if (role === "system") {
+      if (hasToolInteraction) {
+        // System after tool - convert to user message
+        normalized.push({
+          role: "user",
+          content: `[System instruction]: ${typeof msg.content === "string" ? msg.content : JSON.stringify(msg.content)}`,
+        });
+      } else {
+        // System before tool - keep as system
+        systemMessages.push(msg);
+      }
+    } else if (role === "tool") {
+      hasToolInteraction = true;
+      normalized.push(msg);
+    } else {
+      normalized.push(msg);
+    }
+  }
+
+  // Prepend system messages at the start
+  return [...systemMessages, ...normalized];
+}
+
+/**
+ * Fix Gemma's "Invalid 'tool_calls': unknown variant 'index'" error.
+ *
+ * Strategy: Remove 'index' field from tool_calls if present.
+ * Gemma expects only: id, type, function (name, arguments)
+ */
+function normalizeForGemma(
+  messages: ChatCompletionMessageParam[],
+): ChatCompletionMessageParam[] {
+  return messages.map((msg) => {
+    // Only process assistant messages with tool_calls
+    if (msg.role !== "assistant" || !("tool_calls" in msg) || !msg.tool_calls) {
+      return msg;
+    }
+
+    // Remove 'index' field from each tool call
+    const cleanedToolCalls = msg.tool_calls.map((call: any) => {
+      // Create a new object without the 'index' field
+      const { index: _index, ...cleanedCall } = call;
+      return cleanedCall;
+    });
+
+    return {
+      ...msg,
+      tool_calls: cleanedToolCalls,
+    };
+  });
+}