Move composition of systemMessage to a single place and enforce empty…

… system message if given explicitly by user

core/config/load.ts

@@ -369,7 +369,7 @@ async function intermediateToFinalConfig(
               ideSettings,
               writeLog,
               config.completionOptions,
-              config.systemMessage,
+              config.systemMessage
             );
 
             if (llm?.providerName === "free-trial") {
@@ -397,8 +397,8 @@ async function intermediateToFinalConfig(
       (config.contextProviders || [])
         .filter(isContextProviderWithParams)
         .find((cp) => cp.name === "codebase") as
-        | ContextProviderWithParams
-        | undefined
+      | ContextProviderWithParams
+      | undefined
     )?.params || {};
 
   const DEFAULT_CONTEXT_PROVIDERS = [
@@ -991,5 +991,6 @@ export {
   finalToBrowserConfig,
   intermediateToFinalConfig,
   loadContinueConfigFromJson,
-  type BrowserSerializedContinueConfig,
+  type BrowserSerializedContinueConfig
 };
+

core/llm/constructMessages.ts

@@ -1,4 +1,5 @@
 import {
+  BrowserSerializedContinueConfig,
   ChatHistoryItem,
   ChatMessage,
   MessagePart,
@@ -14,12 +15,31 @@ const TOOL_USE_RULES = `When using tools, follow the following guidelines:
 function constructSystemPrompt(
   modelDescription: ModelDescription,
   useTools: boolean,
+  continueConfig: BrowserSerializedContinueConfig
 ): string | null {
   let systemMessage =
     "Always include the language and file name in the info string when you write code blocks, for example '```python file.py'.";
   if (useTools && modelSupportsTools(modelDescription)) {
     systemMessage += "\n\n" + TOOL_USE_RULES;
   }
+
+  // We we have no access to the LLM class, we final systemMessage have to be the same as in core/llm/index.ts
+  const userSystemMessage = modelDescription.systemMessage ?? continueConfig.systemMessage;
+
+  // logic moved from core/llm/countTokens.ts
+  if (userSystemMessage && userSystemMessage.trim() !== "") {
+    const shouldAddNewLines = systemMessage !== "";
+    if (shouldAddNewLines) {
+      systemMessage += "\n\n";
+    }
+    systemMessage += userSystemMessage;
+  }
+
+  if (userSystemMessage === "") {
+    // Used defined explicit empty system message will be forced
+    systemMessage = "";
+  }
+
   return systemMessage;
 }
 
@@ -30,13 +50,14 @@ export function constructMessages(
   history: ChatHistoryItem[],
   modelDescription: ModelDescription,
   useTools: boolean,
+  continueConfig: BrowserSerializedContinueConfig,
 ): ChatMessage[] {
   const filteredHistory = history.filter(
     (item) => item.message.role !== "system",
   );
   const msgs: ChatMessage[] = [];
 
-  const systemMessage = constructSystemPrompt(modelDescription, useTools);
+  const systemMessage = constructSystemPrompt(modelDescription, useTools, continueConfig);
   if (systemMessage) {
     msgs.push({
       role: "system",

core/llm/countTokens.ts

@@ -27,9 +27,9 @@ class LlamaEncoding implements Encoding {
 }
 
 class NonWorkerAsyncEncoder implements AsyncEncoder {
-  constructor(private readonly encoding: Encoding) {}
+  constructor(private readonly encoding: Encoding) { }
 
-  async close(): Promise<void> {}
+  async close(): Promise<void> { }
 
   async encode(text: string): Promise<number[]> {
     return this.encoding.encode(text);
@@ -383,8 +383,8 @@ function compileChatMessages(
 ): ChatMessage[] {
   let msgsCopy = msgs
     ? msgs
-        .map((msg) => ({ ...msg }))
-        .filter((msg) => !chatMessageIsEmpty(msg) && msg.role !== "system")
+      .map((msg) => ({ ...msg }))
+      .filter((msg) => !chatMessageIsEmpty(msg) && msg.role !== "system")
     : [];
 
   msgsCopy = addSpaceToAnyEmptyMessages(msgsCopy);
@@ -397,6 +397,7 @@ function compileChatMessages(
     msgsCopy.push(promptMsg);
   }
 
+  /* Original logic, moved to core/llm/constructMessages.ts
   if (
     (systemMessage && systemMessage.trim() !== "") ||
     msgs?.[0]?.role === "system"
@@ -419,6 +420,29 @@ function compileChatMessages(
     // Insert as second to last
     // Later moved to top, but want second-priority to last user message
     msgsCopy.splice(-1, 0, systemChatMsg);
+  }*/
+
+  if (
+    msgs?.[0]?.role === "system"
+  ) {
+    let content = "";
+
+    content = renderChatMessage(msgs?.[0]);
+
+    const systemChatMsg: ChatMessage = {
+      role: "system",
+      content,
+    };
+    // Insert as second to last
+    // Later moved to top, but want second-priority to last user message
+    msgsCopy.splice(-1, 0, systemChatMsg);
+  } else if (systemMessage && systemMessage.trim() !== "") {
+    // In case core/llm/constructMessages.ts constructSystemPrompt() is not called
+    const systemChatMsg: ChatMessage = {
+      role: "system",
+      content: systemMessage,
+    };
+    msgsCopy.splice(-1, 0, systemChatMsg);
   }
 
   let functionTokens = 0;
@@ -469,5 +493,6 @@ export {
   pruneLinesFromTop,
   pruneRawPromptFromTop,
   pruneStringFromBottom,
-  pruneStringFromTop,
+  pruneStringFromTop
 };
+

core/llm/llms/index.ts

@@ -3,7 +3,7 @@ import {
   IdeSettings,
   ILLM,
   LLMOptions,
-  ModelDescription,
+  ModelDescription
 } from "../..";
 import { renderTemplatedString } from "../../promptFiles/v1/renderTemplatedString";
 import { BaseLLM } from "../index";
@@ -19,7 +19,6 @@ import Cohere from "./Cohere";
 import DeepInfra from "./DeepInfra";
 import Deepseek from "./Deepseek";
 import Fireworks from "./Fireworks";
-import NCompass from "./NCompass";
 import Flowise from "./Flowise";
 import FreeTrial from "./FreeTrial";
 import FunctionNetwork from "./FunctionNetwork";
@@ -35,7 +34,9 @@ import Mistral from "./Mistral";
 import MockLLM from "./Mock";
 import Moonshot from "./Moonshot";
 import Msty from "./Msty";
+import NCompass from "./NCompass";
 import Nebius from "./Nebius";
+import Novita from "./Novita";
 import Nvidia from "./Nvidia";
 import Ollama from "./Ollama";
 import OpenAI from "./OpenAI";
@@ -49,7 +50,6 @@ import ContinueProxy from "./stubs/ContinueProxy";
 import TestLLM from "./Test";
 import TextGenWebUI from "./TextGenWebUI";
 import Together from "./Together";
-import Novita from "./Novita";
 import VertexAI from "./VertexAI";
 import Vllm from "./Vllm";
 import WatsonX from "./WatsonX";

gui/src/redux/thunks/streamResponse.ts

@@ -34,10 +34,10 @@ const getSlashCommandForInput = (
     typeof input === "string"
       ? input
       : (
-          input.filter((part) => part.type === "text").slice(-1)[0] as
-            | TextMessagePart
-            | undefined
-        )?.text || "";
+        input.filter((part) => part.type === "text").slice(-1)[0] as
+        | TextMessagePart
+        | undefined
+      )?.text || "";
 
   if (lastText.startsWith("/")) {
     slashCommandName = lastText.split(" ")[0].substring(1);
@@ -124,6 +124,7 @@ export const streamResponseThunk = createAsyncThunk<
           [...updatedHistory],
           defaultModel,
           useTools,
+          state.config.config
         );
 
         posthog.capture("step run", {

gui/src/redux/thunks/streamResponseAfterToolCall.ts

@@ -64,6 +64,7 @@ export const streamResponseAfterToolCall = createAsyncThunk<
           [...updatedHistory],
           defaultModel,
           useTools,
+          state.config.config
         );
         const output = await dispatch(streamNormalInput(messages));
         unwrapResult(output);