Use recent messages in prompt and tool prompts

messagesToPrompt now preserves the system message and includes up to the last 10 non-system conversation turns, formatting them as System/User/Assistant blocks and handling content arrays. buildPromptWithTools was changed to construct a prompt consisting of tool instructions, any existing system context, and the last user message (extracted safely from array content), instead of mutating the messages array. The chat route streaming logic was adjusted to suppress immediate text chunk streaming when tools are present so tool call responses can be emitted as proper tool_calls on completion.

Author: foxy1402

src/helpers/format.js

@@ -214,32 +214,55 @@ const getModelMapping = (requestedModel) => {
 // ---------------------------------------------------------------------------
 
 /**
- * Convert an OpenAI messages array into a single prompt string.
+ * Convert an OpenAI messages array into a single prompt string for qodercli.
  *
- * For better qodercli compatibility, we only send the latest user message
- * instead of full conversation history, as qodercli responds better to
- * individual prompts than conversation threads.
+ * Includes conversation history (up to last 10 messages) so the model has
+ * context for follow-up questions and multi-turn edits. Older messages are
+ * dropped to avoid exceeding qodercli's context limits.
+ *
+ * Format:
+ *   System: <system message if present>
+ *   User: <message>
+ *   Assistant: <message>
+ *   User: <latest message>
  */
 const messagesToPrompt = (messages) => {
-  // Find the last user message
-  const lastUserMessage = messages
-    .slice()
-    .reverse()
-    .find((msg) => msg.role === "user");
-
-  if (!lastUserMessage) {
-    return "Hello";
-  }
+  if (!messages || messages.length === 0) return "Hello";
+
+  // Separate system message from conversation
+  const systemMsg = messages.find((m) => m.role === "system");
+  const conversation = messages.filter((m) => m.role !== "system");
 
-  // Extract content
-  const content = Array.isArray(lastUserMessage.content)
-    ? lastUserMessage.content
+  // Keep last 10 conversation turns to avoid context overflow
+  const recent = conversation.slice(-10);
+
+  const extractContent = (msg) => {
+    if (Array.isArray(msg.content)) {
+      return msg.content
         .filter((p) => p.type === "text")
         .map((p) => p.text)
-        .join("")
-    : lastUserMessage.content || "";
+        .join("");
+    }
+    return msg.content || "";
+  };
+
+  const parts = [];
+
+  // Include system message if present
+  if (systemMsg) {
+    const sysContent = extractContent(systemMsg);
+    if (sysContent.trim()) parts.push(`System: ${sysContent.trim()}`);
+  }
+
+  // Include conversation history
+  for (const msg of recent) {
+    const content = extractContent(msg).trim();
+    if (!content) continue;
+    if (msg.role === "user") parts.push(`User: ${content}`);
+    else if (msg.role === "assistant") parts.push(`Assistant: ${content}`);
+  }
 
-  return `User: ${content.trim()}`;
+  return parts.join("\n\n") || "Hello";
 };
 
 // ---------------------------------------------------------------------------

src/helpers/toolPrompt.js

@@ -61,22 +61,34 @@ const buildPromptWithTools = (messages, tools, messagesToPromptFn) => {
 
   const toolSystem = buildToolSystemPrompt(tools);
 
-  // Find existing system message and prepend tool instructions to it,
-  // or insert a new system message at the start
-  const existingSystemIdx = messages.findIndex((m) => m.role === 'system');
-  let augmented;
-
-  if (existingSystemIdx !== -1) {
-    augmented = messages.map((m, i) =>
-      i === existingSystemIdx
-        ? { ...m, content: `${toolSystem}\n\n${m.content}` }
-        : m
-    );
-  } else {
-    augmented = [{ role: 'system', content: toolSystem }, ...messages];
-  }
-
-  return messagesToPromptFn(augmented);
+  // Extract the last user message directly — bypasses messagesToPrompt entirely
+  // because messagesToPrompt drops system messages and we need to include tool instructions.
+  const lastUser = messages
+    .slice()
+    .reverse()
+    .find((m) => m.role === 'user');
+
+  const userContent = Array.isArray(lastUser?.content)
+    ? lastUser.content
+        .filter((p) => p.type === 'text')
+        .map((p) => p.text)
+        .join('')
+    : lastUser?.content || '';
+
+  // Also extract any existing system message to include alongside tool instructions
+  const existingSystem = messages.find((m) => m.role === 'system');
+  const systemContent = existingSystem
+    ? Array.isArray(existingSystem.content)
+      ? existingSystem.content.filter((p) => p.type === 'text').map((p) => p.text).join('')
+      : existingSystem.content || ''
+    : '';
+
+  // Build final prompt: tool instructions + optional system context + user message
+  const parts = [toolSystem];
+  if (systemContent) parts.push(`System context: ${systemContent.trim()}`);
+  parts.push(`User: ${userContent.trim()}`);
+
+  return parts.join('\n\n');
 };
 
 // ── Response parsing ──────────────────────────────────────────────────────────

src/routes/chat.js

@@ -155,15 +155,19 @@ router.post("/", (req, res) => {
         if (finishReason) lastFinishReason = finishReason;
         if (content) {
           fullStreamText += content;
-          // Stream content chunks normally — tool call detection happens onDone
-          const chunk = {
-            id,
-            object: "chat.completion.chunk",
-            created: Math.floor(Date.now() / 1000),
-            model,
-            choices: [{ index: 0, delta: { content }, finish_reason: null }],
-          };
-          res.write(`data: ${JSON.stringify(chunk)}\n\n`);
+          // When tools are present: accumulate silently, don't stream raw text chunks.
+          // The response will be emitted as a proper tool_calls chunk in onDone.
+          // When no tools: stream content immediately as normal.
+          if (!hasTools) {
+            const chunk = {
+              id,
+              object: "chat.completion.chunk",
+              created: Math.floor(Date.now() / 1000),
+              model,
+              choices: [{ index: 0, delta: { content }, finish_reason: null }],
+            };
+            res.write(`data: ${JSON.stringify(chunk)}\n\n`);
+          }
         }
       },
       onDone: (code, stderr) => {