1 files changed, 28 insertions, 1 deletions
diff --git a/makima/src/server/handlers/chat.rs b/makima/src/server/handlers/chat.rs
index 158805b..51f17c1 100644
--- a/makima/src/server/handlers/chat.rs
+++ b/makima/src/server/handlers/chat.rs
@@ -15,7 +15,7 @@ use crate::llm::{
     claude::{self, ClaudeClient, ClaudeError, ClaudeModel},
     execute_tool_call,
     groq::{GroqClient, GroqError, Message, ToolCallResponse},
-    LlmModel, ToolCall, ToolResult, VersionToolRequest, AVAILABLE_TOOLS,
+    LlmModel, ToolCall, ToolResult, UserQuestion, VersionToolRequest, AVAILABLE_TOOLS,
 };
 use crate::server::state::{FileUpdateNotification, SharedState};
 
@@ -66,6 +66,9 @@ pub struct ChatResponse {
     pub updated_body: Vec<BodyElement>,
     /// Updated summary (if changed)
     pub updated_summary: Option<String>,
+    /// Questions pending user answers (pauses conversation)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub pending_questions: Option<Vec<UserQuestion>>,
 }
 
 #[derive(Debug, Serialize, ToSchema)]
@@ -326,6 +329,8 @@ You have access to tools for:
     // Track consecutive failures for agentic retry logic
     let mut consecutive_failures = 0;
     const MAX_CONSECUTIVE_FAILURES: usize = 3;
+    // Track pending user questions (pauses the conversation)
+    let mut pending_questions: Option<Vec<UserQuestion>> = None;
 
     // Multi-turn agentic tool calling loop
     for round in 0..MAX_TOOL_ROUNDS {
@@ -508,6 +513,21 @@ You have access to tools for:
                 );
             }
 
+            // Check for pending user questions (pauses the conversation)
+            if let Some(questions) = execution_result.pending_questions {
+                tracing::info!(
+                    question_count = questions.len(),
+                    "LLM requesting user input, pausing conversation"
+                );
+                pending_questions = Some(questions);
+                // Track this tool call before breaking
+                all_tool_call_infos.push(ToolCallInfo {
+                    name: tool_call.name.clone(),
+                    result: execution_result.result,
+                });
+                break; // Exit inner loop
+            }
+
             // Build tool result message content with enhanced context for agentic reasoning
             let result_content = if let Some(parsed_data) = &execution_result.parsed_data {
                 // Include parsed data in the result for the LLM to use
@@ -559,6 +579,12 @@ You have access to tools for:
             });
         }
 
+        // If user questions are pending, pause the conversation
+        if pending_questions.is_some() {
+            final_response = result.content;
+            break;
+        }
+
         // If finish reason indicates completion, exit loop
         let finish_lower = result.finish_reason.to_lowercase();
         if finish_lower == "stop" || finish_lower == "end_turn" {
@@ -637,6 +663,7 @@ You have access to tools for:
             tool_calls: all_tool_call_infos,
             updated_body: current_body,
             updated_summary: current_summary,
+            pending_questions,
         }),
     )
         .into_response()