Skip to content

armgit / zed

Public Imported from GitHub
0

Code at the speed of thought – Zed is a high-performance, multiplayer code editor from the creators of Atom and Tree-sitter.

agent: Fix compaction limit (#58431)

Bennet Bo Fenner0357c4a8d2c7
diff --git a/crates/agent/src/thread.rs b/crates/agent/src/thread.rs
index 3f990345363..d742750b226 100644
--- a/crates/agent/src/thread.rs
+++ b/crates/agent/src/thread.rs
@@ -3657,7 +3657,10 @@ impl Thread {
.and_then(|v| v.parse().ok())
.unwrap_or(AGENT_COMPACTION_REMAINING_TOKEN_BUDGET);
- let compaction_threshold = model.max_token_count().saturating_sub(remaining_budget);
+ let compaction_threshold = model
+ .max_token_count()
+ .saturating_sub(model.max_output_tokens().unwrap_or_default())
+ .saturating_sub(remaining_budget);
if active_tokens < compaction_threshold {
return None;
}
@@ -5509,6 +5512,34 @@ mod tests {
});
}
+ #[gpui::test]
+ async fn test_compaction_threshold_reserves_max_output_tokens(cx: &mut TestAppContext) {
+ let (thread, _event_stream) = setup_thread_for_test(cx).await;
+ let model = Arc::new(FakeLanguageModel::default());
+ model.set_max_token_count(272_000);
+ model.set_max_output_tokens(Some(128_000));
+ let user_message_id = UserMessageId::new();
+
+ cx.update(|cx| {
+ thread.update(cx, |thread, cx| {
+ thread.set_model(model, cx);
+ thread.messages.push(user_text_message(
+ user_message_id.clone(),
+ "near output-reserved limit",
+ ));
+ thread.request_token_usage.insert(
+ user_message_id.clone(),
+ language_model::TokenUsage {
+ input_tokens: 105_000,
+ ..Default::default()
+ },
+ );
+
+ assert_eq!(thread.compaction_message_target_ix(), Some(1));
+ });
+ });
+ }
+
#[gpui::test]
async fn test_compaction_unavailable_for_small_context_window(cx: &mut TestAppContext) {
let (thread, _event_stream) = setup_thread_for_test(cx).await;
diff --git a/crates/language_model/src/fake_provider.rs b/crates/language_model/src/fake_provider.rs
index d0527d495d7..e9f4fba4430 100644
--- a/crates/language_model/src/fake_provider.rs
+++ b/crates/language_model/src/fake_provider.rs
@@ -127,6 +127,7 @@ pub struct FakeLanguageModel {
supports_streaming_tools: AtomicBool,
supports_images: AtomicBool,
max_token_count: AtomicU64,
+ max_output_tokens: AtomicU64,
}
impl Default for FakeLanguageModel {
@@ -142,6 +143,7 @@ impl Default for FakeLanguageModel {
supports_streaming_tools: AtomicBool::new(false),
supports_images: AtomicBool::new(false),
max_token_count: AtomicU64::new(1_000_000),
+ max_output_tokens: AtomicU64::new(0),
}
}
}
@@ -186,6 +188,11 @@ impl FakeLanguageModel {
self.max_token_count.store(count, SeqCst);
}
+ pub fn set_max_output_tokens(&self, count: Option<u64>) {
+ self.max_output_tokens
+ .store(count.unwrap_or_default(), SeqCst);
+ }
+
pub fn pending_completions(&self) -> Vec<LanguageModelRequest> {
self.current_completion_txs
.lock()
@@ -311,6 +318,15 @@ impl LanguageModel for FakeLanguageModel {
self.max_token_count.load(SeqCst)
}
+ fn max_output_tokens(&self) -> Option<u64> {
+ let max_output_tokens = self.max_output_tokens.load(SeqCst);
+ if max_output_tokens == 0 {
+ None
+ } else {
+ Some(max_output_tokens)
+ }
+ }
+
fn stream_completion(
&self,
request: LanguageModelRequest,