fix(bedrock): cache trailing message for stable prefix across agent turns

crates/goose/src/providers/bedrock.rs

@@ -232,15 +232,14 @@ impl BedrockProvider {
         let visible_messages: Vec<&Message> =
             messages.iter().filter(|m| m.is_agent_visible()).collect();
 
-        // Cache the earliest messages (not most recent) because prompt caching
-        // requires exact prefix matching — caching recent messages would shift
-        // positions each turn, causing misses.
-        const MESSAGE_CACHE_BUDGET: usize = 3;
-        let cache_count = if enable_caching {
-            visible_messages.len().min(MESSAGE_CACHE_BUDGET)
-        } else {
-            0
-        };
+        // Place the cache point on the trailing message. Cache entries are keyed by
+        // the hash of the prefix ending at the breakpoint; on each new turn, the
+        // lookback walks backward from the new trailing position and finds the
+        // previous turn's entry, so fresh processing is bounded to the content added
+        // since the last request. See
+        // https://platform.claude.com/docs/en/build-with-claude/prompt-caching
+        let last_idx = visible_messages.len().checked_sub(1);
+        let cache_last = enable_caching && last_idx.is_some();
 
         let mut request = self
             .client
@@ -251,7 +250,9 @@ impl BedrockProvider {
                 visible_messages
                     .iter()
                     .enumerate()
-                    .map(|(idx, m)| to_bedrock_message_with_caching(m, idx < cache_count))
+                    .map(|(idx, m)| {
+                        to_bedrock_message_with_caching(m, cache_last && Some(idx) == last_idx)
+                    })
                     .collect::<Result<_>>()?,
             ));