1use super::*;
2use acp_thread::{AgentConnection, AgentModelGroupName, AgentModelList, UserMessageId};
3use agent_client_protocol::{self as acp};
4use agent_settings::AgentProfileId;
5use anyhow::Result;
6use client::{Client, UserStore};
7use cloud_llm_client::CompletionIntent;
8use collections::IndexMap;
9use context_server::{ContextServer, ContextServerCommand, ContextServerId};
10use fs::{FakeFs, Fs};
11use futures::{
12 StreamExt,
13 channel::{
14 mpsc::{self, UnboundedReceiver},
15 oneshot,
16 },
17};
18use gpui::{
19 App, AppContext, Entity, Task, TestAppContext, UpdateGlobal, http_client::FakeHttpClient,
20};
21use indoc::indoc;
22use language_model::{
23 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,
24 LanguageModelProviderName, LanguageModelRegistry, LanguageModelRequest,
25 LanguageModelRequestMessage, LanguageModelToolResult, LanguageModelToolSchemaFormat,
26 LanguageModelToolUse, MessageContent, Role, StopReason, fake_provider::FakeLanguageModel,
27};
28use pretty_assertions::assert_eq;
29use project::{
30 Project, context_server_store::ContextServerStore, project_settings::ProjectSettings,
31};
32use prompt_store::ProjectContext;
33use reqwest_client::ReqwestClient;
34use schemars::JsonSchema;
35use serde::{Deserialize, Serialize};
36use serde_json::json;
37use settings::{Settings, SettingsStore};
38use std::{path::Path, rc::Rc, sync::Arc, time::Duration};
39use util::path;
40
41mod test_tools;
42use test_tools::*;
43
44#[gpui::test]
45async fn test_echo(cx: &mut TestAppContext) {
46 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
47 let fake_model = model.as_fake();
48
49 let events = thread
50 .update(cx, |thread, cx| {
51 thread.send(UserMessageId::new(), ["Testing: Reply with 'Hello'"], cx)
52 })
53 .unwrap();
54 cx.run_until_parked();
55 fake_model.send_last_completion_stream_text_chunk("Hello");
56 fake_model
57 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
58 fake_model.end_last_completion_stream();
59
60 let events = events.collect().await;
61 thread.update(cx, |thread, _cx| {
62 assert_eq!(
63 thread.last_message().unwrap().to_markdown(),
64 indoc! {"
65 ## Assistant
66
67 Hello
68 "}
69 )
70 });
71 assert_eq!(stop_events(events), vec![acp::StopReason::EndTurn]);
72}
73
74#[gpui::test]
75async fn test_thinking(cx: &mut TestAppContext) {
76 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
77 let fake_model = model.as_fake();
78
79 let events = thread
80 .update(cx, |thread, cx| {
81 thread.send(
82 UserMessageId::new(),
83 [indoc! {"
84 Testing:
85
86 Generate a thinking step where you just think the word 'Think',
87 and have your final answer be 'Hello'
88 "}],
89 cx,
90 )
91 })
92 .unwrap();
93 cx.run_until_parked();
94 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Thinking {
95 text: "Think".to_string(),
96 signature: None,
97 });
98 fake_model.send_last_completion_stream_text_chunk("Hello");
99 fake_model
100 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
101 fake_model.end_last_completion_stream();
102
103 let events = events.collect().await;
104 thread.update(cx, |thread, _cx| {
105 assert_eq!(
106 thread.last_message().unwrap().to_markdown(),
107 indoc! {"
108 ## Assistant
109
110 <think>Think</think>
111 Hello
112 "}
113 )
114 });
115 assert_eq!(stop_events(events), vec![acp::StopReason::EndTurn]);
116}
117
118#[gpui::test]
119async fn test_system_prompt(cx: &mut TestAppContext) {
120 let ThreadTest {
121 model,
122 thread,
123 project_context,
124 ..
125 } = setup(cx, TestModel::Fake).await;
126 let fake_model = model.as_fake();
127
128 project_context.update(cx, |project_context, _cx| {
129 project_context.shell = "test-shell".into()
130 });
131 thread.update(cx, |thread, _| thread.add_tool(EchoTool));
132 thread
133 .update(cx, |thread, cx| {
134 thread.send(UserMessageId::new(), ["abc"], cx)
135 })
136 .unwrap();
137 cx.run_until_parked();
138 let mut pending_completions = fake_model.pending_completions();
139 assert_eq!(
140 pending_completions.len(),
141 1,
142 "unexpected pending completions: {:?}",
143 pending_completions
144 );
145
146 let pending_completion = pending_completions.pop().unwrap();
147 assert_eq!(pending_completion.messages[0].role, Role::System);
148
149 let system_message = &pending_completion.messages[0];
150 let system_prompt = system_message.content[0].to_str().unwrap();
151 assert!(
152 system_prompt.contains("test-shell"),
153 "unexpected system message: {:?}",
154 system_message
155 );
156 assert!(
157 system_prompt.contains("## Fixing Diagnostics"),
158 "unexpected system message: {:?}",
159 system_message
160 );
161}
162
163#[gpui::test]
164async fn test_prompt_caching(cx: &mut TestAppContext) {
165 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
166 let fake_model = model.as_fake();
167
168 // Send initial user message and verify it's cached
169 thread
170 .update(cx, |thread, cx| {
171 thread.send(UserMessageId::new(), ["Message 1"], cx)
172 })
173 .unwrap();
174 cx.run_until_parked();
175
176 let completion = fake_model.pending_completions().pop().unwrap();
177 assert_eq!(
178 completion.messages[1..],
179 vec![LanguageModelRequestMessage {
180 role: Role::User,
181 content: vec!["Message 1".into()],
182 cache: true
183 }]
184 );
185 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Text(
186 "Response to Message 1".into(),
187 ));
188 fake_model.end_last_completion_stream();
189 cx.run_until_parked();
190
191 // Send another user message and verify only the latest is cached
192 thread
193 .update(cx, |thread, cx| {
194 thread.send(UserMessageId::new(), ["Message 2"], cx)
195 })
196 .unwrap();
197 cx.run_until_parked();
198
199 let completion = fake_model.pending_completions().pop().unwrap();
200 assert_eq!(
201 completion.messages[1..],
202 vec![
203 LanguageModelRequestMessage {
204 role: Role::User,
205 content: vec!["Message 1".into()],
206 cache: false
207 },
208 LanguageModelRequestMessage {
209 role: Role::Assistant,
210 content: vec!["Response to Message 1".into()],
211 cache: false
212 },
213 LanguageModelRequestMessage {
214 role: Role::User,
215 content: vec!["Message 2".into()],
216 cache: true
217 }
218 ]
219 );
220 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Text(
221 "Response to Message 2".into(),
222 ));
223 fake_model.end_last_completion_stream();
224 cx.run_until_parked();
225
226 // Simulate a tool call and verify that the latest tool result is cached
227 thread.update(cx, |thread, _| thread.add_tool(EchoTool));
228 thread
229 .update(cx, |thread, cx| {
230 thread.send(UserMessageId::new(), ["Use the echo tool"], cx)
231 })
232 .unwrap();
233 cx.run_until_parked();
234
235 let tool_use = LanguageModelToolUse {
236 id: "tool_1".into(),
237 name: EchoTool::name().into(),
238 raw_input: json!({"text": "test"}).to_string(),
239 input: json!({"text": "test"}),
240 is_input_complete: true,
241 };
242 fake_model
243 .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));
244 fake_model.end_last_completion_stream();
245 cx.run_until_parked();
246
247 let completion = fake_model.pending_completions().pop().unwrap();
248 let tool_result = LanguageModelToolResult {
249 tool_use_id: "tool_1".into(),
250 tool_name: EchoTool::name().into(),
251 is_error: false,
252 content: "test".into(),
253 output: Some("test".into()),
254 };
255 assert_eq!(
256 completion.messages[1..],
257 vec![
258 LanguageModelRequestMessage {
259 role: Role::User,
260 content: vec!["Message 1".into()],
261 cache: false
262 },
263 LanguageModelRequestMessage {
264 role: Role::Assistant,
265 content: vec!["Response to Message 1".into()],
266 cache: false
267 },
268 LanguageModelRequestMessage {
269 role: Role::User,
270 content: vec!["Message 2".into()],
271 cache: false
272 },
273 LanguageModelRequestMessage {
274 role: Role::Assistant,
275 content: vec!["Response to Message 2".into()],
276 cache: false
277 },
278 LanguageModelRequestMessage {
279 role: Role::User,
280 content: vec!["Use the echo tool".into()],
281 cache: false
282 },
283 LanguageModelRequestMessage {
284 role: Role::Assistant,
285 content: vec![MessageContent::ToolUse(tool_use)],
286 cache: false
287 },
288 LanguageModelRequestMessage {
289 role: Role::User,
290 content: vec![MessageContent::ToolResult(tool_result)],
291 cache: true
292 }
293 ]
294 );
295}
296
297#[gpui::test]
298#[cfg_attr(not(feature = "e2e"), ignore)]
299async fn test_basic_tool_calls(cx: &mut TestAppContext) {
300 let ThreadTest { thread, .. } = setup(cx, TestModel::Sonnet4).await;
301
302 // Test a tool call that's likely to complete *before* streaming stops.
303 let events = thread
304 .update(cx, |thread, cx| {
305 thread.add_tool(EchoTool);
306 thread.send(
307 UserMessageId::new(),
308 ["Now test the echo tool with 'Hello'. Does it work? Say 'Yes' or 'No'."],
309 cx,
310 )
311 })
312 .unwrap()
313 .collect()
314 .await;
315 assert_eq!(stop_events(events), vec![acp::StopReason::EndTurn]);
316
317 // Test a tool calls that's likely to complete *after* streaming stops.
318 let events = thread
319 .update(cx, |thread, cx| {
320 thread.remove_tool(&EchoTool::name());
321 thread.add_tool(DelayTool);
322 thread.send(
323 UserMessageId::new(),
324 [
325 "Now call the delay tool with 200ms.",
326 "When the timer goes off, then you echo the output of the tool.",
327 ],
328 cx,
329 )
330 })
331 .unwrap()
332 .collect()
333 .await;
334 assert_eq!(stop_events(events), vec![acp::StopReason::EndTurn]);
335 thread.update(cx, |thread, _cx| {
336 assert!(
337 thread
338 .last_message()
339 .unwrap()
340 .as_agent_message()
341 .unwrap()
342 .content
343 .iter()
344 .any(|content| {
345 if let AgentMessageContent::Text(text) = content {
346 text.contains("Ding")
347 } else {
348 false
349 }
350 }),
351 "{}",
352 thread.to_markdown()
353 );
354 });
355}
356
357#[gpui::test]
358#[cfg_attr(not(feature = "e2e"), ignore)]
359async fn test_streaming_tool_calls(cx: &mut TestAppContext) {
360 let ThreadTest { thread, .. } = setup(cx, TestModel::Sonnet4).await;
361
362 // Test a tool call that's likely to complete *before* streaming stops.
363 let mut events = thread
364 .update(cx, |thread, cx| {
365 thread.add_tool(WordListTool);
366 thread.send(UserMessageId::new(), ["Test the word_list tool."], cx)
367 })
368 .unwrap();
369
370 let mut saw_partial_tool_use = false;
371 while let Some(event) = events.next().await {
372 if let Ok(ThreadEvent::ToolCall(tool_call)) = event {
373 thread.update(cx, |thread, _cx| {
374 // Look for a tool use in the thread's last message
375 let message = thread.last_message().unwrap();
376 let agent_message = message.as_agent_message().unwrap();
377 let last_content = agent_message.content.last().unwrap();
378 if let AgentMessageContent::ToolUse(last_tool_use) = last_content {
379 assert_eq!(last_tool_use.name.as_ref(), "word_list");
380 if tool_call.status == acp::ToolCallStatus::Pending {
381 if !last_tool_use.is_input_complete
382 && last_tool_use.input.get("g").is_none()
383 {
384 saw_partial_tool_use = true;
385 }
386 } else {
387 last_tool_use
388 .input
389 .get("a")
390 .expect("'a' has streamed because input is now complete");
391 last_tool_use
392 .input
393 .get("g")
394 .expect("'g' has streamed because input is now complete");
395 }
396 } else {
397 panic!("last content should be a tool use");
398 }
399 });
400 }
401 }
402
403 assert!(
404 saw_partial_tool_use,
405 "should see at least one partially streamed tool use in the history"
406 );
407}
408
409#[gpui::test]
410async fn test_tool_authorization(cx: &mut TestAppContext) {
411 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
412 let fake_model = model.as_fake();
413
414 let mut events = thread
415 .update(cx, |thread, cx| {
416 thread.add_tool(ToolRequiringPermission);
417 thread.send(UserMessageId::new(), ["abc"], cx)
418 })
419 .unwrap();
420 cx.run_until_parked();
421 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
422 LanguageModelToolUse {
423 id: "tool_id_1".into(),
424 name: ToolRequiringPermission::name().into(),
425 raw_input: "{}".into(),
426 input: json!({}),
427 is_input_complete: true,
428 },
429 ));
430 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
431 LanguageModelToolUse {
432 id: "tool_id_2".into(),
433 name: ToolRequiringPermission::name().into(),
434 raw_input: "{}".into(),
435 input: json!({}),
436 is_input_complete: true,
437 },
438 ));
439 fake_model.end_last_completion_stream();
440 let tool_call_auth_1 = next_tool_call_authorization(&mut events).await;
441 let tool_call_auth_2 = next_tool_call_authorization(&mut events).await;
442
443 // Approve the first
444 tool_call_auth_1
445 .response
446 .send(tool_call_auth_1.options[1].id.clone())
447 .unwrap();
448 cx.run_until_parked();
449
450 // Reject the second
451 tool_call_auth_2
452 .response
453 .send(tool_call_auth_1.options[2].id.clone())
454 .unwrap();
455 cx.run_until_parked();
456
457 let completion = fake_model.pending_completions().pop().unwrap();
458 let message = completion.messages.last().unwrap();
459 assert_eq!(
460 message.content,
461 vec![
462 language_model::MessageContent::ToolResult(LanguageModelToolResult {
463 tool_use_id: tool_call_auth_1.tool_call.id.0.to_string().into(),
464 tool_name: ToolRequiringPermission::name().into(),
465 is_error: false,
466 content: "Allowed".into(),
467 output: Some("Allowed".into())
468 }),
469 language_model::MessageContent::ToolResult(LanguageModelToolResult {
470 tool_use_id: tool_call_auth_2.tool_call.id.0.to_string().into(),
471 tool_name: ToolRequiringPermission::name().into(),
472 is_error: true,
473 content: "Permission to run tool denied by user".into(),
474 output: None
475 })
476 ]
477 );
478
479 // Simulate yet another tool call.
480 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
481 LanguageModelToolUse {
482 id: "tool_id_3".into(),
483 name: ToolRequiringPermission::name().into(),
484 raw_input: "{}".into(),
485 input: json!({}),
486 is_input_complete: true,
487 },
488 ));
489 fake_model.end_last_completion_stream();
490
491 // Respond by always allowing tools.
492 let tool_call_auth_3 = next_tool_call_authorization(&mut events).await;
493 tool_call_auth_3
494 .response
495 .send(tool_call_auth_3.options[0].id.clone())
496 .unwrap();
497 cx.run_until_parked();
498 let completion = fake_model.pending_completions().pop().unwrap();
499 let message = completion.messages.last().unwrap();
500 assert_eq!(
501 message.content,
502 vec![language_model::MessageContent::ToolResult(
503 LanguageModelToolResult {
504 tool_use_id: tool_call_auth_3.tool_call.id.0.to_string().into(),
505 tool_name: ToolRequiringPermission::name().into(),
506 is_error: false,
507 content: "Allowed".into(),
508 output: Some("Allowed".into())
509 }
510 )]
511 );
512
513 // Simulate a final tool call, ensuring we don't trigger authorization.
514 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
515 LanguageModelToolUse {
516 id: "tool_id_4".into(),
517 name: ToolRequiringPermission::name().into(),
518 raw_input: "{}".into(),
519 input: json!({}),
520 is_input_complete: true,
521 },
522 ));
523 fake_model.end_last_completion_stream();
524 cx.run_until_parked();
525 let completion = fake_model.pending_completions().pop().unwrap();
526 let message = completion.messages.last().unwrap();
527 assert_eq!(
528 message.content,
529 vec![language_model::MessageContent::ToolResult(
530 LanguageModelToolResult {
531 tool_use_id: "tool_id_4".into(),
532 tool_name: ToolRequiringPermission::name().into(),
533 is_error: false,
534 content: "Allowed".into(),
535 output: Some("Allowed".into())
536 }
537 )]
538 );
539}
540
541#[gpui::test]
542async fn test_tool_hallucination(cx: &mut TestAppContext) {
543 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
544 let fake_model = model.as_fake();
545
546 let mut events = thread
547 .update(cx, |thread, cx| {
548 thread.send(UserMessageId::new(), ["abc"], cx)
549 })
550 .unwrap();
551 cx.run_until_parked();
552 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
553 LanguageModelToolUse {
554 id: "tool_id_1".into(),
555 name: "nonexistent_tool".into(),
556 raw_input: "{}".into(),
557 input: json!({}),
558 is_input_complete: true,
559 },
560 ));
561 fake_model.end_last_completion_stream();
562
563 let tool_call = expect_tool_call(&mut events).await;
564 assert_eq!(tool_call.title, "nonexistent_tool");
565 assert_eq!(tool_call.status, acp::ToolCallStatus::Pending);
566 let update = expect_tool_call_update_fields(&mut events).await;
567 assert_eq!(update.fields.status, Some(acp::ToolCallStatus::Failed));
568}
569
570#[gpui::test]
571async fn test_resume_after_tool_use_limit(cx: &mut TestAppContext) {
572 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
573 let fake_model = model.as_fake();
574
575 let events = thread
576 .update(cx, |thread, cx| {
577 thread.add_tool(EchoTool);
578 thread.send(UserMessageId::new(), ["abc"], cx)
579 })
580 .unwrap();
581 cx.run_until_parked();
582 let tool_use = LanguageModelToolUse {
583 id: "tool_id_1".into(),
584 name: EchoTool::name().into(),
585 raw_input: "{}".into(),
586 input: serde_json::to_value(&EchoToolInput { text: "def".into() }).unwrap(),
587 is_input_complete: true,
588 };
589 fake_model
590 .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));
591 fake_model.end_last_completion_stream();
592
593 cx.run_until_parked();
594 let completion = fake_model.pending_completions().pop().unwrap();
595 let tool_result = LanguageModelToolResult {
596 tool_use_id: "tool_id_1".into(),
597 tool_name: EchoTool::name().into(),
598 is_error: false,
599 content: "def".into(),
600 output: Some("def".into()),
601 };
602 assert_eq!(
603 completion.messages[1..],
604 vec![
605 LanguageModelRequestMessage {
606 role: Role::User,
607 content: vec!["abc".into()],
608 cache: false
609 },
610 LanguageModelRequestMessage {
611 role: Role::Assistant,
612 content: vec![MessageContent::ToolUse(tool_use.clone())],
613 cache: false
614 },
615 LanguageModelRequestMessage {
616 role: Role::User,
617 content: vec![MessageContent::ToolResult(tool_result.clone())],
618 cache: true
619 },
620 ]
621 );
622
623 // Simulate reaching tool use limit.
624 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::StatusUpdate(
625 cloud_llm_client::CompletionRequestStatus::ToolUseLimitReached,
626 ));
627 fake_model.end_last_completion_stream();
628 let last_event = events.collect::<Vec<_>>().await.pop().unwrap();
629 assert!(
630 last_event
631 .unwrap_err()
632 .is::<language_model::ToolUseLimitReachedError>()
633 );
634
635 let events = thread.update(cx, |thread, cx| thread.resume(cx)).unwrap();
636 cx.run_until_parked();
637 let completion = fake_model.pending_completions().pop().unwrap();
638 assert_eq!(
639 completion.messages[1..],
640 vec![
641 LanguageModelRequestMessage {
642 role: Role::User,
643 content: vec!["abc".into()],
644 cache: false
645 },
646 LanguageModelRequestMessage {
647 role: Role::Assistant,
648 content: vec![MessageContent::ToolUse(tool_use)],
649 cache: false
650 },
651 LanguageModelRequestMessage {
652 role: Role::User,
653 content: vec![MessageContent::ToolResult(tool_result)],
654 cache: false
655 },
656 LanguageModelRequestMessage {
657 role: Role::User,
658 content: vec!["Continue where you left off".into()],
659 cache: true
660 }
661 ]
662 );
663
664 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Text("Done".into()));
665 fake_model.end_last_completion_stream();
666 events.collect::<Vec<_>>().await;
667 thread.read_with(cx, |thread, _cx| {
668 assert_eq!(
669 thread.last_message().unwrap().to_markdown(),
670 indoc! {"
671 ## Assistant
672
673 Done
674 "}
675 )
676 });
677}
678
679#[gpui::test]
680async fn test_send_after_tool_use_limit(cx: &mut TestAppContext) {
681 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
682 let fake_model = model.as_fake();
683
684 let events = thread
685 .update(cx, |thread, cx| {
686 thread.add_tool(EchoTool);
687 thread.send(UserMessageId::new(), ["abc"], cx)
688 })
689 .unwrap();
690 cx.run_until_parked();
691
692 let tool_use = LanguageModelToolUse {
693 id: "tool_id_1".into(),
694 name: EchoTool::name().into(),
695 raw_input: "{}".into(),
696 input: serde_json::to_value(&EchoToolInput { text: "def".into() }).unwrap(),
697 is_input_complete: true,
698 };
699 let tool_result = LanguageModelToolResult {
700 tool_use_id: "tool_id_1".into(),
701 tool_name: EchoTool::name().into(),
702 is_error: false,
703 content: "def".into(),
704 output: Some("def".into()),
705 };
706 fake_model
707 .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));
708 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::StatusUpdate(
709 cloud_llm_client::CompletionRequestStatus::ToolUseLimitReached,
710 ));
711 fake_model.end_last_completion_stream();
712 let last_event = events.collect::<Vec<_>>().await.pop().unwrap();
713 assert!(
714 last_event
715 .unwrap_err()
716 .is::<language_model::ToolUseLimitReachedError>()
717 );
718
719 thread
720 .update(cx, |thread, cx| {
721 thread.send(UserMessageId::new(), vec!["ghi"], cx)
722 })
723 .unwrap();
724 cx.run_until_parked();
725 let completion = fake_model.pending_completions().pop().unwrap();
726 assert_eq!(
727 completion.messages[1..],
728 vec![
729 LanguageModelRequestMessage {
730 role: Role::User,
731 content: vec!["abc".into()],
732 cache: false
733 },
734 LanguageModelRequestMessage {
735 role: Role::Assistant,
736 content: vec![MessageContent::ToolUse(tool_use)],
737 cache: false
738 },
739 LanguageModelRequestMessage {
740 role: Role::User,
741 content: vec![MessageContent::ToolResult(tool_result)],
742 cache: false
743 },
744 LanguageModelRequestMessage {
745 role: Role::User,
746 content: vec!["ghi".into()],
747 cache: true
748 }
749 ]
750 );
751}
752
753async fn expect_tool_call(events: &mut UnboundedReceiver<Result<ThreadEvent>>) -> acp::ToolCall {
754 let event = events
755 .next()
756 .await
757 .expect("no tool call authorization event received")
758 .unwrap();
759 match event {
760 ThreadEvent::ToolCall(tool_call) => tool_call,
761 event => {
762 panic!("Unexpected event {event:?}");
763 }
764 }
765}
766
767async fn expect_tool_call_update_fields(
768 events: &mut UnboundedReceiver<Result<ThreadEvent>>,
769) -> acp::ToolCallUpdate {
770 let event = events
771 .next()
772 .await
773 .expect("no tool call authorization event received")
774 .unwrap();
775 match event {
776 ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(update)) => update,
777 event => {
778 panic!("Unexpected event {event:?}");
779 }
780 }
781}
782
783async fn next_tool_call_authorization(
784 events: &mut UnboundedReceiver<Result<ThreadEvent>>,
785) -> ToolCallAuthorization {
786 loop {
787 let event = events
788 .next()
789 .await
790 .expect("no tool call authorization event received")
791 .unwrap();
792 if let ThreadEvent::ToolCallAuthorization(tool_call_authorization) = event {
793 let permission_kinds = tool_call_authorization
794 .options
795 .iter()
796 .map(|o| o.kind)
797 .collect::<Vec<_>>();
798 assert_eq!(
799 permission_kinds,
800 vec![
801 acp::PermissionOptionKind::AllowAlways,
802 acp::PermissionOptionKind::AllowOnce,
803 acp::PermissionOptionKind::RejectOnce,
804 ]
805 );
806 return tool_call_authorization;
807 }
808 }
809}
810
811#[gpui::test]
812#[cfg_attr(not(feature = "e2e"), ignore)]
813async fn test_concurrent_tool_calls(cx: &mut TestAppContext) {
814 let ThreadTest { thread, .. } = setup(cx, TestModel::Sonnet4).await;
815
816 // Test concurrent tool calls with different delay times
817 let events = thread
818 .update(cx, |thread, cx| {
819 thread.add_tool(DelayTool);
820 thread.send(
821 UserMessageId::new(),
822 [
823 "Call the delay tool twice in the same message.",
824 "Once with 100ms. Once with 300ms.",
825 "When both timers are complete, describe the outputs.",
826 ],
827 cx,
828 )
829 })
830 .unwrap()
831 .collect()
832 .await;
833
834 let stop_reasons = stop_events(events);
835 assert_eq!(stop_reasons, vec![acp::StopReason::EndTurn]);
836
837 thread.update(cx, |thread, _cx| {
838 let last_message = thread.last_message().unwrap();
839 let agent_message = last_message.as_agent_message().unwrap();
840 let text = agent_message
841 .content
842 .iter()
843 .filter_map(|content| {
844 if let AgentMessageContent::Text(text) = content {
845 Some(text.as_str())
846 } else {
847 None
848 }
849 })
850 .collect::<String>();
851
852 assert!(text.contains("Ding"));
853 });
854}
855
856#[gpui::test]
857async fn test_profiles(cx: &mut TestAppContext) {
858 let ThreadTest {
859 model, thread, fs, ..
860 } = setup(cx, TestModel::Fake).await;
861 let fake_model = model.as_fake();
862
863 thread.update(cx, |thread, _cx| {
864 thread.add_tool(DelayTool);
865 thread.add_tool(EchoTool);
866 thread.add_tool(InfiniteTool);
867 });
868
869 // Override profiles and wait for settings to be loaded.
870 fs.insert_file(
871 paths::settings_file(),
872 json!({
873 "agent": {
874 "profiles": {
875 "test-1": {
876 "name": "Test Profile 1",
877 "tools": {
878 EchoTool::name(): true,
879 DelayTool::name(): true,
880 }
881 },
882 "test-2": {
883 "name": "Test Profile 2",
884 "tools": {
885 InfiniteTool::name(): true,
886 }
887 }
888 }
889 }
890 })
891 .to_string()
892 .into_bytes(),
893 )
894 .await;
895 cx.run_until_parked();
896
897 // Test that test-1 profile (default) has echo and delay tools
898 thread
899 .update(cx, |thread, cx| {
900 thread.set_profile(AgentProfileId("test-1".into()));
901 thread.send(UserMessageId::new(), ["test"], cx)
902 })
903 .unwrap();
904 cx.run_until_parked();
905
906 let mut pending_completions = fake_model.pending_completions();
907 assert_eq!(pending_completions.len(), 1);
908 let completion = pending_completions.pop().unwrap();
909 let tool_names: Vec<String> = completion
910 .tools
911 .iter()
912 .map(|tool| tool.name.clone())
913 .collect();
914 assert_eq!(tool_names, vec![DelayTool::name(), EchoTool::name()]);
915 fake_model.end_last_completion_stream();
916
917 // Switch to test-2 profile, and verify that it has only the infinite tool.
918 thread
919 .update(cx, |thread, cx| {
920 thread.set_profile(AgentProfileId("test-2".into()));
921 thread.send(UserMessageId::new(), ["test2"], cx)
922 })
923 .unwrap();
924 cx.run_until_parked();
925 let mut pending_completions = fake_model.pending_completions();
926 assert_eq!(pending_completions.len(), 1);
927 let completion = pending_completions.pop().unwrap();
928 let tool_names: Vec<String> = completion
929 .tools
930 .iter()
931 .map(|tool| tool.name.clone())
932 .collect();
933 assert_eq!(tool_names, vec![InfiniteTool::name()]);
934}
935
936#[gpui::test]
937async fn test_mcp_tools(cx: &mut TestAppContext) {
938 let ThreadTest {
939 model,
940 thread,
941 context_server_store,
942 fs,
943 ..
944 } = setup(cx, TestModel::Fake).await;
945 let fake_model = model.as_fake();
946
947 // Override profiles and wait for settings to be loaded.
948 fs.insert_file(
949 paths::settings_file(),
950 json!({
951 "agent": {
952 "profiles": {
953 "test": {
954 "name": "Test Profile",
955 "enable_all_context_servers": true,
956 "tools": {
957 EchoTool::name(): true,
958 }
959 },
960 }
961 }
962 })
963 .to_string()
964 .into_bytes(),
965 )
966 .await;
967 cx.run_until_parked();
968 thread.update(cx, |thread, _| {
969 thread.set_profile(AgentProfileId("test".into()))
970 });
971
972 let mut mcp_tool_calls = setup_context_server(
973 "test_server",
974 vec![context_server::types::Tool {
975 name: "echo".into(),
976 description: None,
977 input_schema: serde_json::to_value(
978 EchoTool.input_schema(LanguageModelToolSchemaFormat::JsonSchema),
979 )
980 .unwrap(),
981 output_schema: None,
982 annotations: None,
983 }],
984 &context_server_store,
985 cx,
986 );
987
988 let events = thread.update(cx, |thread, cx| {
989 thread.send(UserMessageId::new(), ["Hey"], cx).unwrap()
990 });
991 cx.run_until_parked();
992
993 // Simulate the model calling the MCP tool.
994 let completion = fake_model.pending_completions().pop().unwrap();
995 assert_eq!(tool_names_for_completion(&completion), vec!["echo"]);
996 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
997 LanguageModelToolUse {
998 id: "tool_1".into(),
999 name: "echo".into(),
1000 raw_input: json!({"text": "test"}).to_string(),
1001 input: json!({"text": "test"}),
1002 is_input_complete: true,
1003 },
1004 ));
1005 fake_model.end_last_completion_stream();
1006 cx.run_until_parked();
1007
1008 let (tool_call_params, tool_call_response) = mcp_tool_calls.next().await.unwrap();
1009 assert_eq!(tool_call_params.name, "echo");
1010 assert_eq!(tool_call_params.arguments, Some(json!({"text": "test"})));
1011 tool_call_response
1012 .send(context_server::types::CallToolResponse {
1013 content: vec![context_server::types::ToolResponseContent::Text {
1014 text: "test".into(),
1015 }],
1016 is_error: None,
1017 meta: None,
1018 structured_content: None,
1019 })
1020 .unwrap();
1021 cx.run_until_parked();
1022
1023 assert_eq!(tool_names_for_completion(&completion), vec!["echo"]);
1024 fake_model.send_last_completion_stream_text_chunk("Done!");
1025 fake_model.end_last_completion_stream();
1026 events.collect::<Vec<_>>().await;
1027
1028 // Send again after adding the echo tool, ensuring the name collision is resolved.
1029 let events = thread.update(cx, |thread, cx| {
1030 thread.add_tool(EchoTool);
1031 thread.send(UserMessageId::new(), ["Go"], cx).unwrap()
1032 });
1033 cx.run_until_parked();
1034 let completion = fake_model.pending_completions().pop().unwrap();
1035 assert_eq!(
1036 tool_names_for_completion(&completion),
1037 vec!["echo", "test_server_echo"]
1038 );
1039 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1040 LanguageModelToolUse {
1041 id: "tool_2".into(),
1042 name: "test_server_echo".into(),
1043 raw_input: json!({"text": "mcp"}).to_string(),
1044 input: json!({"text": "mcp"}),
1045 is_input_complete: true,
1046 },
1047 ));
1048 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1049 LanguageModelToolUse {
1050 id: "tool_3".into(),
1051 name: "echo".into(),
1052 raw_input: json!({"text": "native"}).to_string(),
1053 input: json!({"text": "native"}),
1054 is_input_complete: true,
1055 },
1056 ));
1057 fake_model.end_last_completion_stream();
1058 cx.run_until_parked();
1059
1060 let (tool_call_params, tool_call_response) = mcp_tool_calls.next().await.unwrap();
1061 assert_eq!(tool_call_params.name, "echo");
1062 assert_eq!(tool_call_params.arguments, Some(json!({"text": "mcp"})));
1063 tool_call_response
1064 .send(context_server::types::CallToolResponse {
1065 content: vec![context_server::types::ToolResponseContent::Text { text: "mcp".into() }],
1066 is_error: None,
1067 meta: None,
1068 structured_content: None,
1069 })
1070 .unwrap();
1071 cx.run_until_parked();
1072
1073 // Ensure the tool results were inserted with the correct names.
1074 let completion = fake_model.pending_completions().pop().unwrap();
1075 assert_eq!(
1076 completion.messages.last().unwrap().content,
1077 vec![
1078 MessageContent::ToolResult(LanguageModelToolResult {
1079 tool_use_id: "tool_3".into(),
1080 tool_name: "echo".into(),
1081 is_error: false,
1082 content: "native".into(),
1083 output: Some("native".into()),
1084 },),
1085 MessageContent::ToolResult(LanguageModelToolResult {
1086 tool_use_id: "tool_2".into(),
1087 tool_name: "test_server_echo".into(),
1088 is_error: false,
1089 content: "mcp".into(),
1090 output: Some("mcp".into()),
1091 },),
1092 ]
1093 );
1094 fake_model.end_last_completion_stream();
1095 events.collect::<Vec<_>>().await;
1096}
1097
1098#[gpui::test]
1099async fn test_mcp_tool_truncation(cx: &mut TestAppContext) {
1100 let ThreadTest {
1101 model,
1102 thread,
1103 context_server_store,
1104 fs,
1105 ..
1106 } = setup(cx, TestModel::Fake).await;
1107 let fake_model = model.as_fake();
1108
1109 // Set up a profile with all tools enabled
1110 fs.insert_file(
1111 paths::settings_file(),
1112 json!({
1113 "agent": {
1114 "profiles": {
1115 "test": {
1116 "name": "Test Profile",
1117 "enable_all_context_servers": true,
1118 "tools": {
1119 EchoTool::name(): true,
1120 DelayTool::name(): true,
1121 WordListTool::name(): true,
1122 ToolRequiringPermission::name(): true,
1123 InfiniteTool::name(): true,
1124 }
1125 },
1126 }
1127 }
1128 })
1129 .to_string()
1130 .into_bytes(),
1131 )
1132 .await;
1133 cx.run_until_parked();
1134
1135 thread.update(cx, |thread, _| {
1136 thread.set_profile(AgentProfileId("test".into()));
1137 thread.add_tool(EchoTool);
1138 thread.add_tool(DelayTool);
1139 thread.add_tool(WordListTool);
1140 thread.add_tool(ToolRequiringPermission);
1141 thread.add_tool(InfiniteTool);
1142 });
1143
1144 // Set up multiple context servers with some overlapping tool names
1145 let _server1_calls = setup_context_server(
1146 "xxx",
1147 vec![
1148 context_server::types::Tool {
1149 name: "echo".into(), // Conflicts with native EchoTool
1150 description: None,
1151 input_schema: serde_json::to_value(
1152 EchoTool.input_schema(LanguageModelToolSchemaFormat::JsonSchema),
1153 )
1154 .unwrap(),
1155 output_schema: None,
1156 annotations: None,
1157 },
1158 context_server::types::Tool {
1159 name: "unique_tool_1".into(),
1160 description: None,
1161 input_schema: json!({"type": "object", "properties": {}}),
1162 output_schema: None,
1163 annotations: None,
1164 },
1165 ],
1166 &context_server_store,
1167 cx,
1168 );
1169
1170 let _server2_calls = setup_context_server(
1171 "yyy",
1172 vec![
1173 context_server::types::Tool {
1174 name: "echo".into(), // Also conflicts with native EchoTool
1175 description: None,
1176 input_schema: serde_json::to_value(
1177 EchoTool.input_schema(LanguageModelToolSchemaFormat::JsonSchema),
1178 )
1179 .unwrap(),
1180 output_schema: None,
1181 annotations: None,
1182 },
1183 context_server::types::Tool {
1184 name: "unique_tool_2".into(),
1185 description: None,
1186 input_schema: json!({"type": "object", "properties": {}}),
1187 output_schema: None,
1188 annotations: None,
1189 },
1190 context_server::types::Tool {
1191 name: "a".repeat(MAX_TOOL_NAME_LENGTH - 2),
1192 description: None,
1193 input_schema: json!({"type": "object", "properties": {}}),
1194 output_schema: None,
1195 annotations: None,
1196 },
1197 context_server::types::Tool {
1198 name: "b".repeat(MAX_TOOL_NAME_LENGTH - 1),
1199 description: None,
1200 input_schema: json!({"type": "object", "properties": {}}),
1201 output_schema: None,
1202 annotations: None,
1203 },
1204 ],
1205 &context_server_store,
1206 cx,
1207 );
1208 let _server3_calls = setup_context_server(
1209 "zzz",
1210 vec![
1211 context_server::types::Tool {
1212 name: "a".repeat(MAX_TOOL_NAME_LENGTH - 2),
1213 description: None,
1214 input_schema: json!({"type": "object", "properties": {}}),
1215 output_schema: None,
1216 annotations: None,
1217 },
1218 context_server::types::Tool {
1219 name: "b".repeat(MAX_TOOL_NAME_LENGTH - 1),
1220 description: None,
1221 input_schema: json!({"type": "object", "properties": {}}),
1222 output_schema: None,
1223 annotations: None,
1224 },
1225 context_server::types::Tool {
1226 name: "c".repeat(MAX_TOOL_NAME_LENGTH + 1),
1227 description: None,
1228 input_schema: json!({"type": "object", "properties": {}}),
1229 output_schema: None,
1230 annotations: None,
1231 },
1232 ],
1233 &context_server_store,
1234 cx,
1235 );
1236
1237 thread
1238 .update(cx, |thread, cx| {
1239 thread.send(UserMessageId::new(), ["Go"], cx)
1240 })
1241 .unwrap();
1242 cx.run_until_parked();
1243 let completion = fake_model.pending_completions().pop().unwrap();
1244 assert_eq!(
1245 tool_names_for_completion(&completion),
1246 vec![
1247 "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb",
1248 "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc",
1249 "delay",
1250 "echo",
1251 "infinite",
1252 "tool_requiring_permission",
1253 "unique_tool_1",
1254 "unique_tool_2",
1255 "word_list",
1256 "xxx_echo",
1257 "y_aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
1258 "yyy_echo",
1259 "z_aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa",
1260 ]
1261 );
1262}
1263
1264#[gpui::test]
1265#[cfg_attr(not(feature = "e2e"), ignore)]
1266async fn test_cancellation(cx: &mut TestAppContext) {
1267 let ThreadTest { thread, .. } = setup(cx, TestModel::Sonnet4).await;
1268
1269 let mut events = thread
1270 .update(cx, |thread, cx| {
1271 thread.add_tool(InfiniteTool);
1272 thread.add_tool(EchoTool);
1273 thread.send(
1274 UserMessageId::new(),
1275 ["Call the echo tool, then call the infinite tool, then explain their output"],
1276 cx,
1277 )
1278 })
1279 .unwrap();
1280
1281 // Wait until both tools are called.
1282 let mut expected_tools = vec!["Echo", "Infinite Tool"];
1283 let mut echo_id = None;
1284 let mut echo_completed = false;
1285 while let Some(event) = events.next().await {
1286 match event.unwrap() {
1287 ThreadEvent::ToolCall(tool_call) => {
1288 assert_eq!(tool_call.title, expected_tools.remove(0));
1289 if tool_call.title == "Echo" {
1290 echo_id = Some(tool_call.id);
1291 }
1292 }
1293 ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
1294 acp::ToolCallUpdate {
1295 id,
1296 fields:
1297 acp::ToolCallUpdateFields {
1298 status: Some(acp::ToolCallStatus::Completed),
1299 ..
1300 },
1301 },
1302 )) if Some(&id) == echo_id.as_ref() => {
1303 echo_completed = true;
1304 }
1305 _ => {}
1306 }
1307
1308 if expected_tools.is_empty() && echo_completed {
1309 break;
1310 }
1311 }
1312
1313 // Cancel the current send and ensure that the event stream is closed, even
1314 // if one of the tools is still running.
1315 thread.update(cx, |thread, cx| thread.cancel(cx));
1316 let events = events.collect::<Vec<_>>().await;
1317 let last_event = events.last();
1318 assert!(
1319 matches!(
1320 last_event,
1321 Some(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
1322 ),
1323 "unexpected event {last_event:?}"
1324 );
1325
1326 // Ensure we can still send a new message after cancellation.
1327 let events = thread
1328 .update(cx, |thread, cx| {
1329 thread.send(
1330 UserMessageId::new(),
1331 ["Testing: reply with 'Hello' then stop."],
1332 cx,
1333 )
1334 })
1335 .unwrap()
1336 .collect::<Vec<_>>()
1337 .await;
1338 thread.update(cx, |thread, _cx| {
1339 let message = thread.last_message().unwrap();
1340 let agent_message = message.as_agent_message().unwrap();
1341 assert_eq!(
1342 agent_message.content,
1343 vec![AgentMessageContent::Text("Hello".to_string())]
1344 );
1345 });
1346 assert_eq!(stop_events(events), vec![acp::StopReason::EndTurn]);
1347}
1348
1349#[gpui::test]
1350async fn test_in_progress_send_canceled_by_next_send(cx: &mut TestAppContext) {
1351 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1352 let fake_model = model.as_fake();
1353
1354 let events_1 = thread
1355 .update(cx, |thread, cx| {
1356 thread.send(UserMessageId::new(), ["Hello 1"], cx)
1357 })
1358 .unwrap();
1359 cx.run_until_parked();
1360 fake_model.send_last_completion_stream_text_chunk("Hey 1!");
1361 cx.run_until_parked();
1362
1363 let events_2 = thread
1364 .update(cx, |thread, cx| {
1365 thread.send(UserMessageId::new(), ["Hello 2"], cx)
1366 })
1367 .unwrap();
1368 cx.run_until_parked();
1369 fake_model.send_last_completion_stream_text_chunk("Hey 2!");
1370 fake_model
1371 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1372 fake_model.end_last_completion_stream();
1373
1374 let events_1 = events_1.collect::<Vec<_>>().await;
1375 assert_eq!(stop_events(events_1), vec![acp::StopReason::Cancelled]);
1376 let events_2 = events_2.collect::<Vec<_>>().await;
1377 assert_eq!(stop_events(events_2), vec![acp::StopReason::EndTurn]);
1378}
1379
1380#[gpui::test]
1381async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {
1382 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1383 let fake_model = model.as_fake();
1384
1385 let events_1 = thread
1386 .update(cx, |thread, cx| {
1387 thread.send(UserMessageId::new(), ["Hello 1"], cx)
1388 })
1389 .unwrap();
1390 cx.run_until_parked();
1391 fake_model.send_last_completion_stream_text_chunk("Hey 1!");
1392 fake_model
1393 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1394 fake_model.end_last_completion_stream();
1395 let events_1 = events_1.collect::<Vec<_>>().await;
1396
1397 let events_2 = thread
1398 .update(cx, |thread, cx| {
1399 thread.send(UserMessageId::new(), ["Hello 2"], cx)
1400 })
1401 .unwrap();
1402 cx.run_until_parked();
1403 fake_model.send_last_completion_stream_text_chunk("Hey 2!");
1404 fake_model
1405 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1406 fake_model.end_last_completion_stream();
1407 let events_2 = events_2.collect::<Vec<_>>().await;
1408
1409 assert_eq!(stop_events(events_1), vec![acp::StopReason::EndTurn]);
1410 assert_eq!(stop_events(events_2), vec![acp::StopReason::EndTurn]);
1411}
1412
1413#[gpui::test]
1414async fn test_refusal(cx: &mut TestAppContext) {
1415 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1416 let fake_model = model.as_fake();
1417
1418 let events = thread
1419 .update(cx, |thread, cx| {
1420 thread.send(UserMessageId::new(), ["Hello"], cx)
1421 })
1422 .unwrap();
1423 cx.run_until_parked();
1424 thread.read_with(cx, |thread, _| {
1425 assert_eq!(
1426 thread.to_markdown(),
1427 indoc! {"
1428 ## User
1429
1430 Hello
1431 "}
1432 );
1433 });
1434
1435 fake_model.send_last_completion_stream_text_chunk("Hey!");
1436 cx.run_until_parked();
1437 thread.read_with(cx, |thread, _| {
1438 assert_eq!(
1439 thread.to_markdown(),
1440 indoc! {"
1441 ## User
1442
1443 Hello
1444
1445 ## Assistant
1446
1447 Hey!
1448 "}
1449 );
1450 });
1451
1452 // If the model refuses to continue, the thread should remove all the messages after the last user message.
1453 fake_model
1454 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::Refusal));
1455 let events = events.collect::<Vec<_>>().await;
1456 assert_eq!(stop_events(events), vec![acp::StopReason::Refusal]);
1457 thread.read_with(cx, |thread, _| {
1458 assert_eq!(thread.to_markdown(), "");
1459 });
1460}
1461
1462#[gpui::test]
1463async fn test_truncate_first_message(cx: &mut TestAppContext) {
1464 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1465 let fake_model = model.as_fake();
1466
1467 let message_id = UserMessageId::new();
1468 thread
1469 .update(cx, |thread, cx| {
1470 thread.send(message_id.clone(), ["Hello"], cx)
1471 })
1472 .unwrap();
1473 cx.run_until_parked();
1474 thread.read_with(cx, |thread, _| {
1475 assert_eq!(
1476 thread.to_markdown(),
1477 indoc! {"
1478 ## User
1479
1480 Hello
1481 "}
1482 );
1483 assert_eq!(thread.latest_token_usage(), None);
1484 });
1485
1486 fake_model.send_last_completion_stream_text_chunk("Hey!");
1487 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1488 language_model::TokenUsage {
1489 input_tokens: 32_000,
1490 output_tokens: 16_000,
1491 cache_creation_input_tokens: 0,
1492 cache_read_input_tokens: 0,
1493 },
1494 ));
1495 cx.run_until_parked();
1496 thread.read_with(cx, |thread, _| {
1497 assert_eq!(
1498 thread.to_markdown(),
1499 indoc! {"
1500 ## User
1501
1502 Hello
1503
1504 ## Assistant
1505
1506 Hey!
1507 "}
1508 );
1509 assert_eq!(
1510 thread.latest_token_usage(),
1511 Some(acp_thread::TokenUsage {
1512 used_tokens: 32_000 + 16_000,
1513 max_tokens: 1_000_000,
1514 })
1515 );
1516 });
1517
1518 thread
1519 .update(cx, |thread, cx| thread.truncate(message_id, cx))
1520 .unwrap();
1521 cx.run_until_parked();
1522 thread.read_with(cx, |thread, _| {
1523 assert_eq!(thread.to_markdown(), "");
1524 assert_eq!(thread.latest_token_usage(), None);
1525 });
1526
1527 // Ensure we can still send a new message after truncation.
1528 thread
1529 .update(cx, |thread, cx| {
1530 thread.send(UserMessageId::new(), ["Hi"], cx)
1531 })
1532 .unwrap();
1533 thread.update(cx, |thread, _cx| {
1534 assert_eq!(
1535 thread.to_markdown(),
1536 indoc! {"
1537 ## User
1538
1539 Hi
1540 "}
1541 );
1542 });
1543 cx.run_until_parked();
1544 fake_model.send_last_completion_stream_text_chunk("Ahoy!");
1545 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1546 language_model::TokenUsage {
1547 input_tokens: 40_000,
1548 output_tokens: 20_000,
1549 cache_creation_input_tokens: 0,
1550 cache_read_input_tokens: 0,
1551 },
1552 ));
1553 cx.run_until_parked();
1554 thread.read_with(cx, |thread, _| {
1555 assert_eq!(
1556 thread.to_markdown(),
1557 indoc! {"
1558 ## User
1559
1560 Hi
1561
1562 ## Assistant
1563
1564 Ahoy!
1565 "}
1566 );
1567
1568 assert_eq!(
1569 thread.latest_token_usage(),
1570 Some(acp_thread::TokenUsage {
1571 used_tokens: 40_000 + 20_000,
1572 max_tokens: 1_000_000,
1573 })
1574 );
1575 });
1576}
1577
1578#[gpui::test]
1579async fn test_truncate_second_message(cx: &mut TestAppContext) {
1580 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1581 let fake_model = model.as_fake();
1582
1583 thread
1584 .update(cx, |thread, cx| {
1585 thread.send(UserMessageId::new(), ["Message 1"], cx)
1586 })
1587 .unwrap();
1588 cx.run_until_parked();
1589 fake_model.send_last_completion_stream_text_chunk("Message 1 response");
1590 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1591 language_model::TokenUsage {
1592 input_tokens: 32_000,
1593 output_tokens: 16_000,
1594 cache_creation_input_tokens: 0,
1595 cache_read_input_tokens: 0,
1596 },
1597 ));
1598 fake_model.end_last_completion_stream();
1599 cx.run_until_parked();
1600
1601 let assert_first_message_state = |cx: &mut TestAppContext| {
1602 thread.clone().read_with(cx, |thread, _| {
1603 assert_eq!(
1604 thread.to_markdown(),
1605 indoc! {"
1606 ## User
1607
1608 Message 1
1609
1610 ## Assistant
1611
1612 Message 1 response
1613 "}
1614 );
1615
1616 assert_eq!(
1617 thread.latest_token_usage(),
1618 Some(acp_thread::TokenUsage {
1619 used_tokens: 32_000 + 16_000,
1620 max_tokens: 1_000_000,
1621 })
1622 );
1623 });
1624 };
1625
1626 assert_first_message_state(cx);
1627
1628 let second_message_id = UserMessageId::new();
1629 thread
1630 .update(cx, |thread, cx| {
1631 thread.send(second_message_id.clone(), ["Message 2"], cx)
1632 })
1633 .unwrap();
1634 cx.run_until_parked();
1635
1636 fake_model.send_last_completion_stream_text_chunk("Message 2 response");
1637 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1638 language_model::TokenUsage {
1639 input_tokens: 40_000,
1640 output_tokens: 20_000,
1641 cache_creation_input_tokens: 0,
1642 cache_read_input_tokens: 0,
1643 },
1644 ));
1645 fake_model.end_last_completion_stream();
1646 cx.run_until_parked();
1647
1648 thread.read_with(cx, |thread, _| {
1649 assert_eq!(
1650 thread.to_markdown(),
1651 indoc! {"
1652 ## User
1653
1654 Message 1
1655
1656 ## Assistant
1657
1658 Message 1 response
1659
1660 ## User
1661
1662 Message 2
1663
1664 ## Assistant
1665
1666 Message 2 response
1667 "}
1668 );
1669
1670 assert_eq!(
1671 thread.latest_token_usage(),
1672 Some(acp_thread::TokenUsage {
1673 used_tokens: 40_000 + 20_000,
1674 max_tokens: 1_000_000,
1675 })
1676 );
1677 });
1678
1679 thread
1680 .update(cx, |thread, cx| thread.truncate(second_message_id, cx))
1681 .unwrap();
1682 cx.run_until_parked();
1683
1684 assert_first_message_state(cx);
1685}
1686
1687#[gpui::test]
1688async fn test_title_generation(cx: &mut TestAppContext) {
1689 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1690 let fake_model = model.as_fake();
1691
1692 let summary_model = Arc::new(FakeLanguageModel::default());
1693 thread.update(cx, |thread, cx| {
1694 thread.set_summarization_model(Some(summary_model.clone()), cx)
1695 });
1696
1697 let send = thread
1698 .update(cx, |thread, cx| {
1699 thread.send(UserMessageId::new(), ["Hello"], cx)
1700 })
1701 .unwrap();
1702 cx.run_until_parked();
1703
1704 fake_model.send_last_completion_stream_text_chunk("Hey!");
1705 fake_model.end_last_completion_stream();
1706 cx.run_until_parked();
1707 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "New Thread"));
1708
1709 // Ensure the summary model has been invoked to generate a title.
1710 summary_model.send_last_completion_stream_text_chunk("Hello ");
1711 summary_model.send_last_completion_stream_text_chunk("world\nG");
1712 summary_model.send_last_completion_stream_text_chunk("oodnight Moon");
1713 summary_model.end_last_completion_stream();
1714 send.collect::<Vec<_>>().await;
1715 cx.run_until_parked();
1716 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "Hello world"));
1717
1718 // Send another message, ensuring no title is generated this time.
1719 let send = thread
1720 .update(cx, |thread, cx| {
1721 thread.send(UserMessageId::new(), ["Hello again"], cx)
1722 })
1723 .unwrap();
1724 cx.run_until_parked();
1725 fake_model.send_last_completion_stream_text_chunk("Hey again!");
1726 fake_model.end_last_completion_stream();
1727 cx.run_until_parked();
1728 assert_eq!(summary_model.pending_completions(), Vec::new());
1729 send.collect::<Vec<_>>().await;
1730 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "Hello world"));
1731}
1732
1733#[gpui::test]
1734async fn test_building_request_with_pending_tools(cx: &mut TestAppContext) {
1735 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1736 let fake_model = model.as_fake();
1737
1738 let _events = thread
1739 .update(cx, |thread, cx| {
1740 thread.add_tool(ToolRequiringPermission);
1741 thread.add_tool(EchoTool);
1742 thread.send(UserMessageId::new(), ["Hey!"], cx)
1743 })
1744 .unwrap();
1745 cx.run_until_parked();
1746
1747 let permission_tool_use = LanguageModelToolUse {
1748 id: "tool_id_1".into(),
1749 name: ToolRequiringPermission::name().into(),
1750 raw_input: "{}".into(),
1751 input: json!({}),
1752 is_input_complete: true,
1753 };
1754 let echo_tool_use = LanguageModelToolUse {
1755 id: "tool_id_2".into(),
1756 name: EchoTool::name().into(),
1757 raw_input: json!({"text": "test"}).to_string(),
1758 input: json!({"text": "test"}),
1759 is_input_complete: true,
1760 };
1761 fake_model.send_last_completion_stream_text_chunk("Hi!");
1762 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1763 permission_tool_use,
1764 ));
1765 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1766 echo_tool_use.clone(),
1767 ));
1768 fake_model.end_last_completion_stream();
1769 cx.run_until_parked();
1770
1771 // Ensure pending tools are skipped when building a request.
1772 let request = thread
1773 .read_with(cx, |thread, cx| {
1774 thread.build_completion_request(CompletionIntent::EditFile, cx)
1775 })
1776 .unwrap();
1777 assert_eq!(
1778 request.messages[1..],
1779 vec![
1780 LanguageModelRequestMessage {
1781 role: Role::User,
1782 content: vec!["Hey!".into()],
1783 cache: true
1784 },
1785 LanguageModelRequestMessage {
1786 role: Role::Assistant,
1787 content: vec![
1788 MessageContent::Text("Hi!".into()),
1789 MessageContent::ToolUse(echo_tool_use.clone())
1790 ],
1791 cache: false
1792 },
1793 LanguageModelRequestMessage {
1794 role: Role::User,
1795 content: vec![MessageContent::ToolResult(LanguageModelToolResult {
1796 tool_use_id: echo_tool_use.id.clone(),
1797 tool_name: echo_tool_use.name,
1798 is_error: false,
1799 content: "test".into(),
1800 output: Some("test".into())
1801 })],
1802 cache: false
1803 },
1804 ],
1805 );
1806}
1807
1808#[gpui::test]
1809async fn test_agent_connection(cx: &mut TestAppContext) {
1810 cx.update(settings::init);
1811 let templates = Templates::new();
1812
1813 // Initialize language model system with test provider
1814 cx.update(|cx| {
1815 gpui_tokio::init(cx);
1816 client::init_settings(cx);
1817
1818 let http_client = FakeHttpClient::with_404_response();
1819 let clock = Arc::new(clock::FakeSystemClock::new());
1820 let client = Client::new(clock, http_client, cx);
1821 let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
1822 Project::init_settings(cx);
1823 agent_settings::init(cx);
1824 language_model::init(client.clone(), cx);
1825 language_models::init(user_store, client.clone(), cx);
1826 LanguageModelRegistry::test(cx);
1827 });
1828 cx.executor().forbid_parking();
1829
1830 // Create a project for new_thread
1831 let fake_fs = cx.update(|cx| fs::FakeFs::new(cx.background_executor().clone()));
1832 fake_fs.insert_tree(path!("/test"), json!({})).await;
1833 let project = Project::test(fake_fs.clone(), [Path::new("/test")], cx).await;
1834 let cwd = Path::new("/test");
1835 let context_store = cx.new(|cx| assistant_context::ContextStore::fake(project.clone(), cx));
1836 let history_store = cx.new(|cx| HistoryStore::new(context_store, cx));
1837
1838 // Create agent and connection
1839 let agent = NativeAgent::new(
1840 project.clone(),
1841 history_store,
1842 templates.clone(),
1843 None,
1844 fake_fs.clone(),
1845 &mut cx.to_async(),
1846 )
1847 .await
1848 .unwrap();
1849 let connection = NativeAgentConnection(agent.clone());
1850
1851 // Test model_selector returns Some
1852 let selector_opt = connection.model_selector();
1853 assert!(
1854 selector_opt.is_some(),
1855 "agent2 should always support ModelSelector"
1856 );
1857 let selector = selector_opt.unwrap();
1858
1859 // Test list_models
1860 let listed_models = cx
1861 .update(|cx| selector.list_models(cx))
1862 .await
1863 .expect("list_models should succeed");
1864 let AgentModelList::Grouped(listed_models) = listed_models else {
1865 panic!("Unexpected model list type");
1866 };
1867 assert!(!listed_models.is_empty(), "should have at least one model");
1868 assert_eq!(
1869 listed_models[&AgentModelGroupName("Fake".into())][0].id.0,
1870 "fake/fake"
1871 );
1872
1873 // Create a thread using new_thread
1874 let connection_rc = Rc::new(connection.clone());
1875 let acp_thread = cx
1876 .update(|cx| connection_rc.new_thread(project, cwd, cx))
1877 .await
1878 .expect("new_thread should succeed");
1879
1880 // Get the session_id from the AcpThread
1881 let session_id = acp_thread.read_with(cx, |thread, _| thread.session_id().clone());
1882
1883 // Test selected_model returns the default
1884 let model = cx
1885 .update(|cx| selector.selected_model(&session_id, cx))
1886 .await
1887 .expect("selected_model should succeed");
1888 let model = cx
1889 .update(|cx| agent.read(cx).models().model_from_id(&model.id))
1890 .unwrap();
1891 let model = model.as_fake();
1892 assert_eq!(model.id().0, "fake", "should return default model");
1893
1894 let request = acp_thread.update(cx, |thread, cx| thread.send(vec!["abc".into()], cx));
1895 cx.run_until_parked();
1896 model.send_last_completion_stream_text_chunk("def");
1897 cx.run_until_parked();
1898 acp_thread.read_with(cx, |thread, cx| {
1899 assert_eq!(
1900 thread.to_markdown(cx),
1901 indoc! {"
1902 ## User
1903
1904 abc
1905
1906 ## Assistant
1907
1908 def
1909
1910 "}
1911 )
1912 });
1913
1914 // Test cancel
1915 cx.update(|cx| connection.cancel(&session_id, cx));
1916 request.await.expect("prompt should fail gracefully");
1917
1918 // Ensure that dropping the ACP thread causes the native thread to be
1919 // dropped as well.
1920 cx.update(|_| drop(acp_thread));
1921 let result = cx
1922 .update(|cx| {
1923 connection.prompt(
1924 Some(acp_thread::UserMessageId::new()),
1925 acp::PromptRequest {
1926 session_id: session_id.clone(),
1927 prompt: vec!["ghi".into()],
1928 },
1929 cx,
1930 )
1931 })
1932 .await;
1933 assert_eq!(
1934 result.as_ref().unwrap_err().to_string(),
1935 "Session not found",
1936 "unexpected result: {:?}",
1937 result
1938 );
1939}
1940
1941#[gpui::test]
1942async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {
1943 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
1944 thread.update(cx, |thread, _cx| thread.add_tool(ThinkingTool));
1945 let fake_model = model.as_fake();
1946
1947 let mut events = thread
1948 .update(cx, |thread, cx| {
1949 thread.send(UserMessageId::new(), ["Think"], cx)
1950 })
1951 .unwrap();
1952 cx.run_until_parked();
1953
1954 // Simulate streaming partial input.
1955 let input = json!({});
1956 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1957 LanguageModelToolUse {
1958 id: "1".into(),
1959 name: ThinkingTool::name().into(),
1960 raw_input: input.to_string(),
1961 input,
1962 is_input_complete: false,
1963 },
1964 ));
1965
1966 // Input streaming completed
1967 let input = json!({ "content": "Thinking hard!" });
1968 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1969 LanguageModelToolUse {
1970 id: "1".into(),
1971 name: "thinking".into(),
1972 raw_input: input.to_string(),
1973 input,
1974 is_input_complete: true,
1975 },
1976 ));
1977 fake_model.end_last_completion_stream();
1978 cx.run_until_parked();
1979
1980 let tool_call = expect_tool_call(&mut events).await;
1981 assert_eq!(
1982 tool_call,
1983 acp::ToolCall {
1984 id: acp::ToolCallId("1".into()),
1985 title: "Thinking".into(),
1986 kind: acp::ToolKind::Think,
1987 status: acp::ToolCallStatus::Pending,
1988 content: vec![],
1989 locations: vec![],
1990 raw_input: Some(json!({})),
1991 raw_output: None,
1992 }
1993 );
1994 let update = expect_tool_call_update_fields(&mut events).await;
1995 assert_eq!(
1996 update,
1997 acp::ToolCallUpdate {
1998 id: acp::ToolCallId("1".into()),
1999 fields: acp::ToolCallUpdateFields {
2000 title: Some("Thinking".into()),
2001 kind: Some(acp::ToolKind::Think),
2002 raw_input: Some(json!({ "content": "Thinking hard!" })),
2003 ..Default::default()
2004 },
2005 }
2006 );
2007 let update = expect_tool_call_update_fields(&mut events).await;
2008 assert_eq!(
2009 update,
2010 acp::ToolCallUpdate {
2011 id: acp::ToolCallId("1".into()),
2012 fields: acp::ToolCallUpdateFields {
2013 status: Some(acp::ToolCallStatus::InProgress),
2014 ..Default::default()
2015 },
2016 }
2017 );
2018 let update = expect_tool_call_update_fields(&mut events).await;
2019 assert_eq!(
2020 update,
2021 acp::ToolCallUpdate {
2022 id: acp::ToolCallId("1".into()),
2023 fields: acp::ToolCallUpdateFields {
2024 content: Some(vec!["Thinking hard!".into()]),
2025 ..Default::default()
2026 },
2027 }
2028 );
2029 let update = expect_tool_call_update_fields(&mut events).await;
2030 assert_eq!(
2031 update,
2032 acp::ToolCallUpdate {
2033 id: acp::ToolCallId("1".into()),
2034 fields: acp::ToolCallUpdateFields {
2035 status: Some(acp::ToolCallStatus::Completed),
2036 raw_output: Some("Finished thinking.".into()),
2037 ..Default::default()
2038 },
2039 }
2040 );
2041}
2042
2043#[gpui::test]
2044async fn test_send_no_retry_on_success(cx: &mut TestAppContext) {
2045 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2046 let fake_model = model.as_fake();
2047
2048 let mut events = thread
2049 .update(cx, |thread, cx| {
2050 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2051 thread.send(UserMessageId::new(), ["Hello!"], cx)
2052 })
2053 .unwrap();
2054 cx.run_until_parked();
2055
2056 fake_model.send_last_completion_stream_text_chunk("Hey!");
2057 fake_model.end_last_completion_stream();
2058
2059 let mut retry_events = Vec::new();
2060 while let Some(Ok(event)) = events.next().await {
2061 match event {
2062 ThreadEvent::Retry(retry_status) => {
2063 retry_events.push(retry_status);
2064 }
2065 ThreadEvent::Stop(..) => break,
2066 _ => {}
2067 }
2068 }
2069
2070 assert_eq!(retry_events.len(), 0);
2071 thread.read_with(cx, |thread, _cx| {
2072 assert_eq!(
2073 thread.to_markdown(),
2074 indoc! {"
2075 ## User
2076
2077 Hello!
2078
2079 ## Assistant
2080
2081 Hey!
2082 "}
2083 )
2084 });
2085}
2086
2087#[gpui::test]
2088async fn test_send_retry_on_error(cx: &mut TestAppContext) {
2089 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2090 let fake_model = model.as_fake();
2091
2092 let mut events = thread
2093 .update(cx, |thread, cx| {
2094 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2095 thread.send(UserMessageId::new(), ["Hello!"], cx)
2096 })
2097 .unwrap();
2098 cx.run_until_parked();
2099
2100 fake_model.send_last_completion_stream_text_chunk("Hey,");
2101 fake_model.send_last_completion_stream_error(LanguageModelCompletionError::ServerOverloaded {
2102 provider: LanguageModelProviderName::new("Anthropic"),
2103 retry_after: Some(Duration::from_secs(3)),
2104 });
2105 fake_model.end_last_completion_stream();
2106
2107 cx.executor().advance_clock(Duration::from_secs(3));
2108 cx.run_until_parked();
2109
2110 fake_model.send_last_completion_stream_text_chunk("there!");
2111 fake_model.end_last_completion_stream();
2112 cx.run_until_parked();
2113
2114 let mut retry_events = Vec::new();
2115 while let Some(Ok(event)) = events.next().await {
2116 match event {
2117 ThreadEvent::Retry(retry_status) => {
2118 retry_events.push(retry_status);
2119 }
2120 ThreadEvent::Stop(..) => break,
2121 _ => {}
2122 }
2123 }
2124
2125 assert_eq!(retry_events.len(), 1);
2126 assert!(matches!(
2127 retry_events[0],
2128 acp_thread::RetryStatus { attempt: 1, .. }
2129 ));
2130 thread.read_with(cx, |thread, _cx| {
2131 assert_eq!(
2132 thread.to_markdown(),
2133 indoc! {"
2134 ## User
2135
2136 Hello!
2137
2138 ## Assistant
2139
2140 Hey,
2141
2142 [resume]
2143
2144 ## Assistant
2145
2146 there!
2147 "}
2148 )
2149 });
2150}
2151
2152#[gpui::test]
2153async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {
2154 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2155 let fake_model = model.as_fake();
2156
2157 let events = thread
2158 .update(cx, |thread, cx| {
2159 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2160 thread.add_tool(EchoTool);
2161 thread.send(UserMessageId::new(), ["Call the echo tool!"], cx)
2162 })
2163 .unwrap();
2164 cx.run_until_parked();
2165
2166 let tool_use_1 = LanguageModelToolUse {
2167 id: "tool_1".into(),
2168 name: EchoTool::name().into(),
2169 raw_input: json!({"text": "test"}).to_string(),
2170 input: json!({"text": "test"}),
2171 is_input_complete: true,
2172 };
2173 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
2174 tool_use_1.clone(),
2175 ));
2176 fake_model.send_last_completion_stream_error(LanguageModelCompletionError::ServerOverloaded {
2177 provider: LanguageModelProviderName::new("Anthropic"),
2178 retry_after: Some(Duration::from_secs(3)),
2179 });
2180 fake_model.end_last_completion_stream();
2181
2182 cx.executor().advance_clock(Duration::from_secs(3));
2183 let completion = fake_model.pending_completions().pop().unwrap();
2184 assert_eq!(
2185 completion.messages[1..],
2186 vec![
2187 LanguageModelRequestMessage {
2188 role: Role::User,
2189 content: vec!["Call the echo tool!".into()],
2190 cache: false
2191 },
2192 LanguageModelRequestMessage {
2193 role: Role::Assistant,
2194 content: vec![language_model::MessageContent::ToolUse(tool_use_1.clone())],
2195 cache: false
2196 },
2197 LanguageModelRequestMessage {
2198 role: Role::User,
2199 content: vec![language_model::MessageContent::ToolResult(
2200 LanguageModelToolResult {
2201 tool_use_id: tool_use_1.id.clone(),
2202 tool_name: tool_use_1.name.clone(),
2203 is_error: false,
2204 content: "test".into(),
2205 output: Some("test".into())
2206 }
2207 )],
2208 cache: true
2209 },
2210 ]
2211 );
2212
2213 fake_model.send_last_completion_stream_text_chunk("Done");
2214 fake_model.end_last_completion_stream();
2215 cx.run_until_parked();
2216 events.collect::<Vec<_>>().await;
2217 thread.read_with(cx, |thread, _cx| {
2218 assert_eq!(
2219 thread.last_message(),
2220 Some(Message::Agent(AgentMessage {
2221 content: vec![AgentMessageContent::Text("Done".into())],
2222 tool_results: IndexMap::default()
2223 }))
2224 );
2225 })
2226}
2227
2228#[gpui::test]
2229async fn test_send_max_retries_exceeded(cx: &mut TestAppContext) {
2230 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2231 let fake_model = model.as_fake();
2232
2233 let mut events = thread
2234 .update(cx, |thread, cx| {
2235 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2236 thread.send(UserMessageId::new(), ["Hello!"], cx)
2237 })
2238 .unwrap();
2239 cx.run_until_parked();
2240
2241 for _ in 0..crate::thread::MAX_RETRY_ATTEMPTS + 1 {
2242 fake_model.send_last_completion_stream_error(
2243 LanguageModelCompletionError::ServerOverloaded {
2244 provider: LanguageModelProviderName::new("Anthropic"),
2245 retry_after: Some(Duration::from_secs(3)),
2246 },
2247 );
2248 fake_model.end_last_completion_stream();
2249 cx.executor().advance_clock(Duration::from_secs(3));
2250 cx.run_until_parked();
2251 }
2252
2253 let mut errors = Vec::new();
2254 let mut retry_events = Vec::new();
2255 while let Some(event) = events.next().await {
2256 match event {
2257 Ok(ThreadEvent::Retry(retry_status)) => {
2258 retry_events.push(retry_status);
2259 }
2260 Ok(ThreadEvent::Stop(..)) => break,
2261 Err(error) => errors.push(error),
2262 _ => {}
2263 }
2264 }
2265
2266 assert_eq!(
2267 retry_events.len(),
2268 crate::thread::MAX_RETRY_ATTEMPTS as usize
2269 );
2270 for i in 0..crate::thread::MAX_RETRY_ATTEMPTS as usize {
2271 assert_eq!(retry_events[i].attempt, i + 1);
2272 }
2273 assert_eq!(errors.len(), 1);
2274 let error = errors[0]
2275 .downcast_ref::<LanguageModelCompletionError>()
2276 .unwrap();
2277 assert!(matches!(
2278 error,
2279 LanguageModelCompletionError::ServerOverloaded { .. }
2280 ));
2281}
2282
2283/// Filters out the stop events for asserting against in tests
2284fn stop_events(result_events: Vec<Result<ThreadEvent>>) -> Vec<acp::StopReason> {
2285 result_events
2286 .into_iter()
2287 .filter_map(|event| match event.unwrap() {
2288 ThreadEvent::Stop(stop_reason) => Some(stop_reason),
2289 _ => None,
2290 })
2291 .collect()
2292}
2293
2294struct ThreadTest {
2295 model: Arc<dyn LanguageModel>,
2296 thread: Entity<Thread>,
2297 project_context: Entity<ProjectContext>,
2298 context_server_store: Entity<ContextServerStore>,
2299 fs: Arc<FakeFs>,
2300}
2301
2302enum TestModel {
2303 Sonnet4,
2304 Fake,
2305}
2306
2307impl TestModel {
2308 fn id(&self) -> LanguageModelId {
2309 match self {
2310 TestModel::Sonnet4 => LanguageModelId("claude-sonnet-4-latest".into()),
2311 TestModel::Fake => unreachable!(),
2312 }
2313 }
2314}
2315
2316async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {
2317 cx.executor().allow_parking();
2318
2319 let fs = FakeFs::new(cx.background_executor.clone());
2320 fs.create_dir(paths::settings_file().parent().unwrap())
2321 .await
2322 .unwrap();
2323 fs.insert_file(
2324 paths::settings_file(),
2325 json!({
2326 "agent": {
2327 "default_profile": "test-profile",
2328 "profiles": {
2329 "test-profile": {
2330 "name": "Test Profile",
2331 "tools": {
2332 EchoTool::name(): true,
2333 DelayTool::name(): true,
2334 WordListTool::name(): true,
2335 ToolRequiringPermission::name(): true,
2336 InfiniteTool::name(): true,
2337 ThinkingTool::name(): true,
2338 }
2339 }
2340 }
2341 }
2342 })
2343 .to_string()
2344 .into_bytes(),
2345 )
2346 .await;
2347
2348 cx.update(|cx| {
2349 settings::init(cx);
2350 Project::init_settings(cx);
2351 agent_settings::init(cx);
2352 gpui_tokio::init(cx);
2353 let http_client = ReqwestClient::user_agent("agent tests").unwrap();
2354 cx.set_http_client(Arc::new(http_client));
2355
2356 client::init_settings(cx);
2357 let client = Client::production(cx);
2358 let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
2359 language_model::init(client.clone(), cx);
2360 language_models::init(user_store, client.clone(), cx);
2361
2362 watch_settings(fs.clone(), cx);
2363 });
2364
2365 let templates = Templates::new();
2366
2367 fs.insert_tree(path!("/test"), json!({})).await;
2368 let project = Project::test(fs.clone(), [path!("/test").as_ref()], cx).await;
2369
2370 let model = cx
2371 .update(|cx| {
2372 if let TestModel::Fake = model {
2373 Task::ready(Arc::new(FakeLanguageModel::default()) as Arc<_>)
2374 } else {
2375 let model_id = model.id();
2376 let models = LanguageModelRegistry::read_global(cx);
2377 let model = models
2378 .available_models(cx)
2379 .find(|model| model.id() == model_id)
2380 .unwrap();
2381
2382 let provider = models.provider(&model.provider_id()).unwrap();
2383 let authenticated = provider.authenticate(cx);
2384
2385 cx.spawn(async move |_cx| {
2386 authenticated.await.unwrap();
2387 model
2388 })
2389 }
2390 })
2391 .await;
2392
2393 let project_context = cx.new(|_cx| ProjectContext::default());
2394 let context_server_store = project.read_with(cx, |project, _| project.context_server_store());
2395 let context_server_registry =
2396 cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));
2397 let thread = cx.new(|cx| {
2398 Thread::new(
2399 project,
2400 project_context.clone(),
2401 context_server_registry,
2402 templates,
2403 Some(model.clone()),
2404 cx,
2405 )
2406 });
2407 ThreadTest {
2408 model,
2409 thread,
2410 project_context,
2411 context_server_store,
2412 fs,
2413 }
2414}
2415
2416#[cfg(test)]
2417#[ctor::ctor]
2418fn init_logger() {
2419 if std::env::var("RUST_LOG").is_ok() {
2420 env_logger::init();
2421 }
2422}
2423
2424fn watch_settings(fs: Arc<dyn Fs>, cx: &mut App) {
2425 let fs = fs.clone();
2426 cx.spawn({
2427 async move |cx| {
2428 let mut new_settings_content_rx = settings::watch_config_file(
2429 cx.background_executor(),
2430 fs,
2431 paths::settings_file().clone(),
2432 );
2433
2434 while let Some(new_settings_content) = new_settings_content_rx.next().await {
2435 cx.update(|cx| {
2436 SettingsStore::update_global(cx, |settings, cx| {
2437 settings.set_user_settings(&new_settings_content, cx)
2438 })
2439 })
2440 .ok();
2441 }
2442 }
2443 })
2444 .detach();
2445}
2446
2447fn tool_names_for_completion(completion: &LanguageModelRequest) -> Vec<String> {
2448 completion
2449 .tools
2450 .iter()
2451 .map(|tool| tool.name.clone())
2452 .collect()
2453}
2454
2455fn setup_context_server(
2456 name: &'static str,
2457 tools: Vec<context_server::types::Tool>,
2458 context_server_store: &Entity<ContextServerStore>,
2459 cx: &mut TestAppContext,
2460) -> mpsc::UnboundedReceiver<(
2461 context_server::types::CallToolParams,
2462 oneshot::Sender<context_server::types::CallToolResponse>,
2463)> {
2464 cx.update(|cx| {
2465 let mut settings = ProjectSettings::get_global(cx).clone();
2466 settings.context_servers.insert(
2467 name.into(),
2468 project::project_settings::ContextServerSettings::Custom {
2469 enabled: true,
2470 command: ContextServerCommand {
2471 path: "somebinary".into(),
2472 args: Vec::new(),
2473 env: None,
2474 },
2475 },
2476 );
2477 ProjectSettings::override_global(settings, cx);
2478 });
2479
2480 let (mcp_tool_calls_tx, mcp_tool_calls_rx) = mpsc::unbounded();
2481 let fake_transport = context_server::test::create_fake_transport(name, cx.executor())
2482 .on_request::<context_server::types::requests::Initialize, _>(move |_params| async move {
2483 context_server::types::InitializeResponse {
2484 protocol_version: context_server::types::ProtocolVersion(
2485 context_server::types::LATEST_PROTOCOL_VERSION.to_string(),
2486 ),
2487 server_info: context_server::types::Implementation {
2488 name: name.into(),
2489 version: "1.0.0".to_string(),
2490 },
2491 capabilities: context_server::types::ServerCapabilities {
2492 tools: Some(context_server::types::ToolsCapabilities {
2493 list_changed: Some(true),
2494 }),
2495 ..Default::default()
2496 },
2497 meta: None,
2498 }
2499 })
2500 .on_request::<context_server::types::requests::ListTools, _>(move |_params| {
2501 let tools = tools.clone();
2502 async move {
2503 context_server::types::ListToolsResponse {
2504 tools,
2505 next_cursor: None,
2506 meta: None,
2507 }
2508 }
2509 })
2510 .on_request::<context_server::types::requests::CallTool, _>(move |params| {
2511 let mcp_tool_calls_tx = mcp_tool_calls_tx.clone();
2512 async move {
2513 let (response_tx, response_rx) = oneshot::channel();
2514 mcp_tool_calls_tx
2515 .unbounded_send((params, response_tx))
2516 .unwrap();
2517 response_rx.await.unwrap()
2518 }
2519 });
2520 context_server_store.update(cx, |store, cx| {
2521 store.start_server(
2522 Arc::new(ContextServer::new(
2523 ContextServerId(name.into()),
2524 Arc::new(fake_transport),
2525 )),
2526 cx,
2527 );
2528 });
2529 cx.run_until_parked();
2530 mcp_tool_calls_rx
2531}