diff --git a/src/atlas.ts b/src/atlas.ts
index 1871ca9..8016904 100644
--- a/src/atlas.ts
+++ b/src/atlas.ts
@@ -88,14 +88,23 @@ export async function atlasChat(
     const oaiTools = toOAITools(ATLAS_TOOLS);
     const systemPrompt = resolvePrompt('atlas');
 
-    // For init triggers, don't add the synthetic prompt as a user turn
-    if (!opts?.isInit) {
-        session.history.push({ role: 'user', content: userMessage });
-    }
+    // Always push the user message so Gemini gets a valid conversation (requires at least one user turn).
+    // For init triggers, we mark it so we can strip it from the returned history — it's an internal
+    // prompt, not a real user message, and shouldn't appear in the conversation UI or DB.
+    const INIT_MARKER = '__atlas_init_marker__';
+    session.history.push({
+        role: 'user',
+        content: opts?.isInit ? INIT_MARKER + userMessage : userMessage
+    });
 
     const buildMessages = (): LLMMessage[] => [
         { role: 'system', content: systemPrompt },
-        ...session.history.slice(-60)
+        ...session.history.slice(-60).map(m =>
+            // Strip the init marker before sending to the LLM
+            m.role === 'user' && typeof m.content === 'string' && m.content.startsWith(INIT_MARKER)
+                ? { ...m, content: m.content.slice(INIT_MARKER.length) }
+                : m
+        )
     ];
 
     let turn = 0;
@@ -156,6 +165,8 @@ export async function atlasChat(
         reply: finalReply,
         sessionId,
         history: session.history
+            // Drop the internal init user turn — it's not a real user message
+            .filter(m => !(m.role === 'user' && typeof m.content === 'string' && m.content.startsWith(INIT_MARKER)))
             .filter(m => m.role !== 'assistant' || m.content || m.tool_calls?.length)
             .slice(-60),
         prdContent,