Goose Simple Compact UX (#4202)

Co-authored-by: David Katz <dkatz@squareup.com>

crates/goose-cli/src/session/export.rs

@@ -347,6 +347,9 @@ pub fn message_to_markdown(message: &Message, export_all_content: bool) -> Strin
                 md.push_str("**Thinking:**\n");
                 md.push_str("> *Thinking was redacted*\n\n");
             }
+            MessageContent::SummarizationRequested(summarization) => {
+                md.push_str(&format!("*{}*\n\n", summarization.msg));
+            }
             _ => {
                 md.push_str(
                     "`WARNING: Message content type could not be rendered to Markdown`\n\n",

crates/goose-cli/src/session/output.rs

@@ -185,6 +185,9 @@ pub fn render_message(message: &Message, debug: bool) {
                 println!("\n{}", style("Thinking:").dim().italic());
                 print_markdown("Thinking was redacted", theme);
             }
+            MessageContent::SummarizationRequested(summarization) => {
+                println!("\n{}", style(&summarization.msg).yellow());
+            }
             _ => {
                 println!("WARNING: Message content type could not be rendered");
             }

crates/goose/src/agents/agent.rs

@@ -946,7 +946,7 @@ impl Agent {
         // If we compacted, yield the compaction message and history replacement event
         if let Some(compaction_msg) = compaction_msg {
             return Ok(Box::pin(async_stream::try_stream! {
-                yield AgentEvent::Message(Message::assistant().with_text(compaction_msg));
+                yield AgentEvent::Message(Message::assistant().with_summarization_requested(compaction_msg));
                 yield AgentEvent::HistoryReplaced(messages.messages().clone());
 
                 // Continue with normal reply processing using compacted messages

crates/goose/src/agents/context.rs

@@ -84,10 +84,21 @@ impl Agent {
 
         // Add an assistant message to the summarized messages to ensure the assistant's response is included in the context.
         if new_messages.len() == 1 {
-            let assistant_message = Message::assistant().with_text(
-                "I ran into a context length exceeded error so I summarized our conversation.",
-            );
-            let assistant_message_tokens: usize = 14;
+            let compaction_marker = Message::assistant()
+                .with_summarization_requested("Conversation compacted and summarized");
+            let compaction_marker_tokens: usize = 8;
+
+            // Insert the marker before the summary message
+            new_messages.insert(0, compaction_marker);
+            new_token_counts.insert(0, compaction_marker_tokens);
+
+            // Add an assistant message to continue the conversation
+            let assistant_message = Message::assistant().with_text("
+                The previous message contains a summary that was prepared because a context limit was reached.
+                Do not mention that you read a summary or that conversation summarization occurred
+                Just continue the conversation naturally based on the summarized context
+            ");
+            let assistant_message_tokens: usize = 41;
             new_messages.push(assistant_message);
             new_token_counts.push(assistant_message_tokens);
         }

crates/goose/src/context_mgmt/auto_compact.rs

@@ -595,6 +595,8 @@ mod tests {
             create_test_message("First message"),
             create_test_message("Second message"),
             create_test_message("Third message"),
+            create_test_message("Fourth message"),
+            create_test_message("Fifth message"),
         ];
 
         // Create session metadata with high token count to trigger compaction
@@ -617,6 +619,7 @@ mod tests {
 
         // Verify the compacted messages are returned
         assert!(!result.messages.is_empty());
+
         // Should have fewer messages after compaction
         assert!(result.messages.len() <= messages.len());
     }