Additional cleanup of caching

app/src/ai/agent/api.rs

@@ -134,6 +134,9 @@ pub struct RequestParams {
     /// Full Bedrock conversation history for direct Bedrock calls.
     /// When present, the Bedrock path uses this instead of extracting from task_context.
     pub bedrock_message_history: Vec<crate::ai::bedrock::convert::ConversationMessage>,
+    /// Compacted summary from a prior summarization pass. Injected into the system
+    /// prompt so it benefits from system-level caching.
+    pub bedrock_compact_summary: Option<String>,
     /// Populated by the Bedrock path after building the message list.
     /// Contains the full messages sent (old history + new input) so the controller
     /// can store them back into the conversation for the next request cycle.
@@ -328,6 +331,7 @@ impl RequestParams {
             parent_agent_id: None,
             agent_name: None,
             bedrock_message_history: Vec::new(),
+            bedrock_compact_summary: None,
             bedrock_messages_sent: std::sync::Arc::new(std::sync::Mutex::new(Vec::new())),
             is_summarization: false,
         }

app/src/ai/agent/api/impl.rs

@@ -153,6 +153,7 @@ pub async fn generate_multi_agent_output(
         model_id,
         root_task_id: params.root_task_id.clone(),
         bedrock_message_history: params.bedrock_message_history.clone(),
+        bedrock_compact_summary: params.bedrock_compact_summary.clone(),
         bedrock_messages_sent: params.bedrock_messages_sent.clone(),
         is_summarization: params.is_summarization,
     };

app/src/ai/agent/api/impl_tests.rs

@@ -41,6 +41,7 @@ fn request_params_with_ask_user_question_enabled(ask_user_question_enabled: bool
         parent_agent_id: None,
         agent_name: None,
         bedrock_message_history: Vec::new(),
+        bedrock_compact_summary: None,
         bedrock_messages_sent: std::sync::Arc::new(std::sync::Mutex::new(Vec::new())),
         is_summarization: false,
     }

app/src/ai/agent/conversation.rs

@@ -241,6 +241,11 @@ pub struct AIConversation {
     /// context window size, NOT a cumulative total.
     current_context_tokens: u32,
 
+    /// When set, contains the compacted summary of prior conversation history.
+    /// Injected into the system prompt (not the messages array) so it benefits
+    /// from system-level caching at the 1-hour TTL.
+    compact_summary: Option<String>,
+
     /// Guards against repeated auto-compact triggers within the same high-usage window.
     /// Set to true when auto-compact fires; reset when summarization completes.
     has_pending_auto_compact: bool,
@@ -298,6 +303,7 @@ impl AIConversation {
             is_remote_child: false,
             last_event_sequence: None,
             bedrock_message_history: Vec::new(),
+            compact_summary: None,
             current_context_tokens: 0,
             has_pending_auto_compact: false,
             subagent_retry_count: 0,
@@ -498,6 +504,7 @@ impl AIConversation {
             is_remote_child: false,
             last_event_sequence,
             bedrock_message_history,
+            compact_summary: None,
             current_context_tokens: 0,
             has_pending_auto_compact: false,
             subagent_retry_count: 0,
@@ -525,6 +532,14 @@ impl AIConversation {
         self.bedrock_message_history.extend(messages);
     }
 
+    pub fn compact_summary(&self) -> Option<&str> {
+        self.compact_summary.as_deref()
+    }
+
+    pub fn set_compact_summary(&mut self, summary: Option<String>) {
+        self.compact_summary = summary;
+    }
+
     /// Assigns fresh exchange IDs to all exchanges in this conversation.
     /// Used when forking conversations to avoid ID collisions with persisted blocks.
     pub fn reassign_exchange_ids(&mut self) {