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]
1350#[cfg_attr(target_os = "windows", ignore)] // TODO: Fix this test on Windows
1351async fn test_in_progress_send_canceled_by_next_send(cx: &mut TestAppContext) {
1352 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1353 let fake_model = model.as_fake();
1354
1355 let events_1 = thread
1356 .update(cx, |thread, cx| {
1357 thread.send(UserMessageId::new(), ["Hello 1"], cx)
1358 })
1359 .unwrap();
1360 cx.run_until_parked();
1361 fake_model.send_last_completion_stream_text_chunk("Hey 1!");
1362 cx.run_until_parked();
1363
1364 let events_2 = thread
1365 .update(cx, |thread, cx| {
1366 thread.send(UserMessageId::new(), ["Hello 2"], cx)
1367 })
1368 .unwrap();
1369 cx.run_until_parked();
1370 fake_model.send_last_completion_stream_text_chunk("Hey 2!");
1371 fake_model
1372 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1373 fake_model.end_last_completion_stream();
1374
1375 let events_1 = events_1.collect::<Vec<_>>().await;
1376 assert_eq!(stop_events(events_1), vec![acp::StopReason::Cancelled]);
1377 let events_2 = events_2.collect::<Vec<_>>().await;
1378 assert_eq!(stop_events(events_2), vec![acp::StopReason::EndTurn]);
1379}
1380
1381#[gpui::test]
1382async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {
1383 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1384 let fake_model = model.as_fake();
1385
1386 let events_1 = thread
1387 .update(cx, |thread, cx| {
1388 thread.send(UserMessageId::new(), ["Hello 1"], cx)
1389 })
1390 .unwrap();
1391 cx.run_until_parked();
1392 fake_model.send_last_completion_stream_text_chunk("Hey 1!");
1393 fake_model
1394 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1395 fake_model.end_last_completion_stream();
1396 let events_1 = events_1.collect::<Vec<_>>().await;
1397
1398 let events_2 = thread
1399 .update(cx, |thread, cx| {
1400 thread.send(UserMessageId::new(), ["Hello 2"], cx)
1401 })
1402 .unwrap();
1403 cx.run_until_parked();
1404 fake_model.send_last_completion_stream_text_chunk("Hey 2!");
1405 fake_model
1406 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));
1407 fake_model.end_last_completion_stream();
1408 let events_2 = events_2.collect::<Vec<_>>().await;
1409
1410 assert_eq!(stop_events(events_1), vec![acp::StopReason::EndTurn]);
1411 assert_eq!(stop_events(events_2), vec![acp::StopReason::EndTurn]);
1412}
1413
1414#[gpui::test]
1415async fn test_refusal(cx: &mut TestAppContext) {
1416 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1417 let fake_model = model.as_fake();
1418
1419 let events = thread
1420 .update(cx, |thread, cx| {
1421 thread.send(UserMessageId::new(), ["Hello"], cx)
1422 })
1423 .unwrap();
1424 cx.run_until_parked();
1425 thread.read_with(cx, |thread, _| {
1426 assert_eq!(
1427 thread.to_markdown(),
1428 indoc! {"
1429 ## User
1430
1431 Hello
1432 "}
1433 );
1434 });
1435
1436 fake_model.send_last_completion_stream_text_chunk("Hey!");
1437 cx.run_until_parked();
1438 thread.read_with(cx, |thread, _| {
1439 assert_eq!(
1440 thread.to_markdown(),
1441 indoc! {"
1442 ## User
1443
1444 Hello
1445
1446 ## Assistant
1447
1448 Hey!
1449 "}
1450 );
1451 });
1452
1453 // If the model refuses to continue, the thread should remove all the messages after the last user message.
1454 fake_model
1455 .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::Refusal));
1456 let events = events.collect::<Vec<_>>().await;
1457 assert_eq!(stop_events(events), vec![acp::StopReason::Refusal]);
1458 thread.read_with(cx, |thread, _| {
1459 assert_eq!(thread.to_markdown(), "");
1460 });
1461}
1462
1463#[gpui::test]
1464async fn test_truncate_first_message(cx: &mut TestAppContext) {
1465 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1466 let fake_model = model.as_fake();
1467
1468 let message_id = UserMessageId::new();
1469 thread
1470 .update(cx, |thread, cx| {
1471 thread.send(message_id.clone(), ["Hello"], cx)
1472 })
1473 .unwrap();
1474 cx.run_until_parked();
1475 thread.read_with(cx, |thread, _| {
1476 assert_eq!(
1477 thread.to_markdown(),
1478 indoc! {"
1479 ## User
1480
1481 Hello
1482 "}
1483 );
1484 assert_eq!(thread.latest_token_usage(), None);
1485 });
1486
1487 fake_model.send_last_completion_stream_text_chunk("Hey!");
1488 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1489 language_model::TokenUsage {
1490 input_tokens: 32_000,
1491 output_tokens: 16_000,
1492 cache_creation_input_tokens: 0,
1493 cache_read_input_tokens: 0,
1494 },
1495 ));
1496 cx.run_until_parked();
1497 thread.read_with(cx, |thread, _| {
1498 assert_eq!(
1499 thread.to_markdown(),
1500 indoc! {"
1501 ## User
1502
1503 Hello
1504
1505 ## Assistant
1506
1507 Hey!
1508 "}
1509 );
1510 assert_eq!(
1511 thread.latest_token_usage(),
1512 Some(acp_thread::TokenUsage {
1513 used_tokens: 32_000 + 16_000,
1514 max_tokens: 1_000_000,
1515 })
1516 );
1517 });
1518
1519 thread
1520 .update(cx, |thread, cx| thread.truncate(message_id, cx))
1521 .unwrap();
1522 cx.run_until_parked();
1523 thread.read_with(cx, |thread, _| {
1524 assert_eq!(thread.to_markdown(), "");
1525 assert_eq!(thread.latest_token_usage(), None);
1526 });
1527
1528 // Ensure we can still send a new message after truncation.
1529 thread
1530 .update(cx, |thread, cx| {
1531 thread.send(UserMessageId::new(), ["Hi"], cx)
1532 })
1533 .unwrap();
1534 thread.update(cx, |thread, _cx| {
1535 assert_eq!(
1536 thread.to_markdown(),
1537 indoc! {"
1538 ## User
1539
1540 Hi
1541 "}
1542 );
1543 });
1544 cx.run_until_parked();
1545 fake_model.send_last_completion_stream_text_chunk("Ahoy!");
1546 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1547 language_model::TokenUsage {
1548 input_tokens: 40_000,
1549 output_tokens: 20_000,
1550 cache_creation_input_tokens: 0,
1551 cache_read_input_tokens: 0,
1552 },
1553 ));
1554 cx.run_until_parked();
1555 thread.read_with(cx, |thread, _| {
1556 assert_eq!(
1557 thread.to_markdown(),
1558 indoc! {"
1559 ## User
1560
1561 Hi
1562
1563 ## Assistant
1564
1565 Ahoy!
1566 "}
1567 );
1568
1569 assert_eq!(
1570 thread.latest_token_usage(),
1571 Some(acp_thread::TokenUsage {
1572 used_tokens: 40_000 + 20_000,
1573 max_tokens: 1_000_000,
1574 })
1575 );
1576 });
1577}
1578
1579#[gpui::test]
1580async fn test_truncate_second_message(cx: &mut TestAppContext) {
1581 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1582 let fake_model = model.as_fake();
1583
1584 thread
1585 .update(cx, |thread, cx| {
1586 thread.send(UserMessageId::new(), ["Message 1"], cx)
1587 })
1588 .unwrap();
1589 cx.run_until_parked();
1590 fake_model.send_last_completion_stream_text_chunk("Message 1 response");
1591 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1592 language_model::TokenUsage {
1593 input_tokens: 32_000,
1594 output_tokens: 16_000,
1595 cache_creation_input_tokens: 0,
1596 cache_read_input_tokens: 0,
1597 },
1598 ));
1599 fake_model.end_last_completion_stream();
1600 cx.run_until_parked();
1601
1602 let assert_first_message_state = |cx: &mut TestAppContext| {
1603 thread.clone().read_with(cx, |thread, _| {
1604 assert_eq!(
1605 thread.to_markdown(),
1606 indoc! {"
1607 ## User
1608
1609 Message 1
1610
1611 ## Assistant
1612
1613 Message 1 response
1614 "}
1615 );
1616
1617 assert_eq!(
1618 thread.latest_token_usage(),
1619 Some(acp_thread::TokenUsage {
1620 used_tokens: 32_000 + 16_000,
1621 max_tokens: 1_000_000,
1622 })
1623 );
1624 });
1625 };
1626
1627 assert_first_message_state(cx);
1628
1629 let second_message_id = UserMessageId::new();
1630 thread
1631 .update(cx, |thread, cx| {
1632 thread.send(second_message_id.clone(), ["Message 2"], cx)
1633 })
1634 .unwrap();
1635 cx.run_until_parked();
1636
1637 fake_model.send_last_completion_stream_text_chunk("Message 2 response");
1638 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(
1639 language_model::TokenUsage {
1640 input_tokens: 40_000,
1641 output_tokens: 20_000,
1642 cache_creation_input_tokens: 0,
1643 cache_read_input_tokens: 0,
1644 },
1645 ));
1646 fake_model.end_last_completion_stream();
1647 cx.run_until_parked();
1648
1649 thread.read_with(cx, |thread, _| {
1650 assert_eq!(
1651 thread.to_markdown(),
1652 indoc! {"
1653 ## User
1654
1655 Message 1
1656
1657 ## Assistant
1658
1659 Message 1 response
1660
1661 ## User
1662
1663 Message 2
1664
1665 ## Assistant
1666
1667 Message 2 response
1668 "}
1669 );
1670
1671 assert_eq!(
1672 thread.latest_token_usage(),
1673 Some(acp_thread::TokenUsage {
1674 used_tokens: 40_000 + 20_000,
1675 max_tokens: 1_000_000,
1676 })
1677 );
1678 });
1679
1680 thread
1681 .update(cx, |thread, cx| thread.truncate(second_message_id, cx))
1682 .unwrap();
1683 cx.run_until_parked();
1684
1685 assert_first_message_state(cx);
1686}
1687
1688#[gpui::test]
1689#[cfg_attr(target_os = "windows", ignore)] // TODO: Fix this test on Windows
1690async fn test_title_generation(cx: &mut TestAppContext) {
1691 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1692 let fake_model = model.as_fake();
1693
1694 let summary_model = Arc::new(FakeLanguageModel::default());
1695 thread.update(cx, |thread, cx| {
1696 thread.set_summarization_model(Some(summary_model.clone()), cx)
1697 });
1698
1699 let send = thread
1700 .update(cx, |thread, cx| {
1701 thread.send(UserMessageId::new(), ["Hello"], cx)
1702 })
1703 .unwrap();
1704 cx.run_until_parked();
1705
1706 fake_model.send_last_completion_stream_text_chunk("Hey!");
1707 fake_model.end_last_completion_stream();
1708 cx.run_until_parked();
1709 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "New Thread"));
1710
1711 // Ensure the summary model has been invoked to generate a title.
1712 summary_model.send_last_completion_stream_text_chunk("Hello ");
1713 summary_model.send_last_completion_stream_text_chunk("world\nG");
1714 summary_model.send_last_completion_stream_text_chunk("oodnight Moon");
1715 summary_model.end_last_completion_stream();
1716 send.collect::<Vec<_>>().await;
1717 cx.run_until_parked();
1718 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "Hello world"));
1719
1720 // Send another message, ensuring no title is generated this time.
1721 let send = thread
1722 .update(cx, |thread, cx| {
1723 thread.send(UserMessageId::new(), ["Hello again"], cx)
1724 })
1725 .unwrap();
1726 cx.run_until_parked();
1727 fake_model.send_last_completion_stream_text_chunk("Hey again!");
1728 fake_model.end_last_completion_stream();
1729 cx.run_until_parked();
1730 assert_eq!(summary_model.pending_completions(), Vec::new());
1731 send.collect::<Vec<_>>().await;
1732 thread.read_with(cx, |thread, _| assert_eq!(thread.title(), "Hello world"));
1733}
1734
1735#[gpui::test]
1736async fn test_building_request_with_pending_tools(cx: &mut TestAppContext) {
1737 let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
1738 let fake_model = model.as_fake();
1739
1740 let _events = thread
1741 .update(cx, |thread, cx| {
1742 thread.add_tool(ToolRequiringPermission);
1743 thread.add_tool(EchoTool);
1744 thread.send(UserMessageId::new(), ["Hey!"], cx)
1745 })
1746 .unwrap();
1747 cx.run_until_parked();
1748
1749 let permission_tool_use = LanguageModelToolUse {
1750 id: "tool_id_1".into(),
1751 name: ToolRequiringPermission::name().into(),
1752 raw_input: "{}".into(),
1753 input: json!({}),
1754 is_input_complete: true,
1755 };
1756 let echo_tool_use = LanguageModelToolUse {
1757 id: "tool_id_2".into(),
1758 name: EchoTool::name().into(),
1759 raw_input: json!({"text": "test"}).to_string(),
1760 input: json!({"text": "test"}),
1761 is_input_complete: true,
1762 };
1763 fake_model.send_last_completion_stream_text_chunk("Hi!");
1764 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1765 permission_tool_use,
1766 ));
1767 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1768 echo_tool_use.clone(),
1769 ));
1770 fake_model.end_last_completion_stream();
1771 cx.run_until_parked();
1772
1773 // Ensure pending tools are skipped when building a request.
1774 let request = thread
1775 .read_with(cx, |thread, cx| {
1776 thread.build_completion_request(CompletionIntent::EditFile, cx)
1777 })
1778 .unwrap();
1779 assert_eq!(
1780 request.messages[1..],
1781 vec![
1782 LanguageModelRequestMessage {
1783 role: Role::User,
1784 content: vec!["Hey!".into()],
1785 cache: true
1786 },
1787 LanguageModelRequestMessage {
1788 role: Role::Assistant,
1789 content: vec![
1790 MessageContent::Text("Hi!".into()),
1791 MessageContent::ToolUse(echo_tool_use.clone())
1792 ],
1793 cache: false
1794 },
1795 LanguageModelRequestMessage {
1796 role: Role::User,
1797 content: vec![MessageContent::ToolResult(LanguageModelToolResult {
1798 tool_use_id: echo_tool_use.id.clone(),
1799 tool_name: echo_tool_use.name,
1800 is_error: false,
1801 content: "test".into(),
1802 output: Some("test".into())
1803 })],
1804 cache: false
1805 },
1806 ],
1807 );
1808}
1809
1810#[gpui::test]
1811async fn test_agent_connection(cx: &mut TestAppContext) {
1812 cx.update(settings::init);
1813 let templates = Templates::new();
1814
1815 // Initialize language model system with test provider
1816 cx.update(|cx| {
1817 gpui_tokio::init(cx);
1818 client::init_settings(cx);
1819
1820 let http_client = FakeHttpClient::with_404_response();
1821 let clock = Arc::new(clock::FakeSystemClock::new());
1822 let client = Client::new(clock, http_client, cx);
1823 let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
1824 Project::init_settings(cx);
1825 agent_settings::init(cx);
1826 language_model::init(client.clone(), cx);
1827 language_models::init(user_store, client.clone(), cx);
1828 LanguageModelRegistry::test(cx);
1829 });
1830 cx.executor().forbid_parking();
1831
1832 // Create a project for new_thread
1833 let fake_fs = cx.update(|cx| fs::FakeFs::new(cx.background_executor().clone()));
1834 fake_fs.insert_tree(path!("/test"), json!({})).await;
1835 let project = Project::test(fake_fs.clone(), [Path::new("/test")], cx).await;
1836 let cwd = Path::new("/test");
1837 let context_store = cx.new(|cx| assistant_context::ContextStore::fake(project.clone(), cx));
1838 let history_store = cx.new(|cx| HistoryStore::new(context_store, cx));
1839
1840 // Create agent and connection
1841 let agent = NativeAgent::new(
1842 project.clone(),
1843 history_store,
1844 templates.clone(),
1845 None,
1846 fake_fs.clone(),
1847 &mut cx.to_async(),
1848 )
1849 .await
1850 .unwrap();
1851 let connection = NativeAgentConnection(agent.clone());
1852
1853 // Test model_selector returns Some
1854 let selector_opt = connection.model_selector();
1855 assert!(
1856 selector_opt.is_some(),
1857 "agent2 should always support ModelSelector"
1858 );
1859 let selector = selector_opt.unwrap();
1860
1861 // Test list_models
1862 let listed_models = cx
1863 .update(|cx| selector.list_models(cx))
1864 .await
1865 .expect("list_models should succeed");
1866 let AgentModelList::Grouped(listed_models) = listed_models else {
1867 panic!("Unexpected model list type");
1868 };
1869 assert!(!listed_models.is_empty(), "should have at least one model");
1870 assert_eq!(
1871 listed_models[&AgentModelGroupName("Fake".into())][0].id.0,
1872 "fake/fake"
1873 );
1874
1875 // Create a thread using new_thread
1876 let connection_rc = Rc::new(connection.clone());
1877 let acp_thread = cx
1878 .update(|cx| connection_rc.new_thread(project, cwd, cx))
1879 .await
1880 .expect("new_thread should succeed");
1881
1882 // Get the session_id from the AcpThread
1883 let session_id = acp_thread.read_with(cx, |thread, _| thread.session_id().clone());
1884
1885 // Test selected_model returns the default
1886 let model = cx
1887 .update(|cx| selector.selected_model(&session_id, cx))
1888 .await
1889 .expect("selected_model should succeed");
1890 let model = cx
1891 .update(|cx| agent.read(cx).models().model_from_id(&model.id))
1892 .unwrap();
1893 let model = model.as_fake();
1894 assert_eq!(model.id().0, "fake", "should return default model");
1895
1896 let request = acp_thread.update(cx, |thread, cx| thread.send(vec!["abc".into()], cx));
1897 cx.run_until_parked();
1898 model.send_last_completion_stream_text_chunk("def");
1899 cx.run_until_parked();
1900 acp_thread.read_with(cx, |thread, cx| {
1901 assert_eq!(
1902 thread.to_markdown(cx),
1903 indoc! {"
1904 ## User
1905
1906 abc
1907
1908 ## Assistant
1909
1910 def
1911
1912 "}
1913 )
1914 });
1915
1916 // Test cancel
1917 cx.update(|cx| connection.cancel(&session_id, cx));
1918 request.await.expect("prompt should fail gracefully");
1919
1920 // Ensure that dropping the ACP thread causes the native thread to be
1921 // dropped as well.
1922 cx.update(|_| drop(acp_thread));
1923 let result = cx
1924 .update(|cx| {
1925 connection.prompt(
1926 Some(acp_thread::UserMessageId::new()),
1927 acp::PromptRequest {
1928 session_id: session_id.clone(),
1929 prompt: vec!["ghi".into()],
1930 },
1931 cx,
1932 )
1933 })
1934 .await;
1935 assert_eq!(
1936 result.as_ref().unwrap_err().to_string(),
1937 "Session not found",
1938 "unexpected result: {:?}",
1939 result
1940 );
1941}
1942
1943#[gpui::test]
1944async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {
1945 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
1946 thread.update(cx, |thread, _cx| thread.add_tool(ThinkingTool));
1947 let fake_model = model.as_fake();
1948
1949 let mut events = thread
1950 .update(cx, |thread, cx| {
1951 thread.send(UserMessageId::new(), ["Think"], cx)
1952 })
1953 .unwrap();
1954 cx.run_until_parked();
1955
1956 // Simulate streaming partial input.
1957 let input = json!({});
1958 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1959 LanguageModelToolUse {
1960 id: "1".into(),
1961 name: ThinkingTool::name().into(),
1962 raw_input: input.to_string(),
1963 input,
1964 is_input_complete: false,
1965 },
1966 ));
1967
1968 // Input streaming completed
1969 let input = json!({ "content": "Thinking hard!" });
1970 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
1971 LanguageModelToolUse {
1972 id: "1".into(),
1973 name: "thinking".into(),
1974 raw_input: input.to_string(),
1975 input,
1976 is_input_complete: true,
1977 },
1978 ));
1979 fake_model.end_last_completion_stream();
1980 cx.run_until_parked();
1981
1982 let tool_call = expect_tool_call(&mut events).await;
1983 assert_eq!(
1984 tool_call,
1985 acp::ToolCall {
1986 id: acp::ToolCallId("1".into()),
1987 title: "Thinking".into(),
1988 kind: acp::ToolKind::Think,
1989 status: acp::ToolCallStatus::Pending,
1990 content: vec![],
1991 locations: vec![],
1992 raw_input: Some(json!({})),
1993 raw_output: None,
1994 }
1995 );
1996 let update = expect_tool_call_update_fields(&mut events).await;
1997 assert_eq!(
1998 update,
1999 acp::ToolCallUpdate {
2000 id: acp::ToolCallId("1".into()),
2001 fields: acp::ToolCallUpdateFields {
2002 title: Some("Thinking".into()),
2003 kind: Some(acp::ToolKind::Think),
2004 raw_input: Some(json!({ "content": "Thinking hard!" })),
2005 ..Default::default()
2006 },
2007 }
2008 );
2009 let update = expect_tool_call_update_fields(&mut events).await;
2010 assert_eq!(
2011 update,
2012 acp::ToolCallUpdate {
2013 id: acp::ToolCallId("1".into()),
2014 fields: acp::ToolCallUpdateFields {
2015 status: Some(acp::ToolCallStatus::InProgress),
2016 ..Default::default()
2017 },
2018 }
2019 );
2020 let update = expect_tool_call_update_fields(&mut events).await;
2021 assert_eq!(
2022 update,
2023 acp::ToolCallUpdate {
2024 id: acp::ToolCallId("1".into()),
2025 fields: acp::ToolCallUpdateFields {
2026 content: Some(vec!["Thinking hard!".into()]),
2027 ..Default::default()
2028 },
2029 }
2030 );
2031 let update = expect_tool_call_update_fields(&mut events).await;
2032 assert_eq!(
2033 update,
2034 acp::ToolCallUpdate {
2035 id: acp::ToolCallId("1".into()),
2036 fields: acp::ToolCallUpdateFields {
2037 status: Some(acp::ToolCallStatus::Completed),
2038 raw_output: Some("Finished thinking.".into()),
2039 ..Default::default()
2040 },
2041 }
2042 );
2043}
2044
2045#[gpui::test]
2046async fn test_send_no_retry_on_success(cx: &mut TestAppContext) {
2047 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2048 let fake_model = model.as_fake();
2049
2050 let mut events = thread
2051 .update(cx, |thread, cx| {
2052 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2053 thread.send(UserMessageId::new(), ["Hello!"], cx)
2054 })
2055 .unwrap();
2056 cx.run_until_parked();
2057
2058 fake_model.send_last_completion_stream_text_chunk("Hey!");
2059 fake_model.end_last_completion_stream();
2060
2061 let mut retry_events = Vec::new();
2062 while let Some(Ok(event)) = events.next().await {
2063 match event {
2064 ThreadEvent::Retry(retry_status) => {
2065 retry_events.push(retry_status);
2066 }
2067 ThreadEvent::Stop(..) => break,
2068 _ => {}
2069 }
2070 }
2071
2072 assert_eq!(retry_events.len(), 0);
2073 thread.read_with(cx, |thread, _cx| {
2074 assert_eq!(
2075 thread.to_markdown(),
2076 indoc! {"
2077 ## User
2078
2079 Hello!
2080
2081 ## Assistant
2082
2083 Hey!
2084 "}
2085 )
2086 });
2087}
2088
2089#[gpui::test]
2090async fn test_send_retry_on_error(cx: &mut TestAppContext) {
2091 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2092 let fake_model = model.as_fake();
2093
2094 let mut events = thread
2095 .update(cx, |thread, cx| {
2096 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2097 thread.send(UserMessageId::new(), ["Hello!"], cx)
2098 })
2099 .unwrap();
2100 cx.run_until_parked();
2101
2102 fake_model.send_last_completion_stream_text_chunk("Hey,");
2103 fake_model.send_last_completion_stream_error(LanguageModelCompletionError::ServerOverloaded {
2104 provider: LanguageModelProviderName::new("Anthropic"),
2105 retry_after: Some(Duration::from_secs(3)),
2106 });
2107 fake_model.end_last_completion_stream();
2108
2109 cx.executor().advance_clock(Duration::from_secs(3));
2110 cx.run_until_parked();
2111
2112 fake_model.send_last_completion_stream_text_chunk("there!");
2113 fake_model.end_last_completion_stream();
2114 cx.run_until_parked();
2115
2116 let mut retry_events = Vec::new();
2117 while let Some(Ok(event)) = events.next().await {
2118 match event {
2119 ThreadEvent::Retry(retry_status) => {
2120 retry_events.push(retry_status);
2121 }
2122 ThreadEvent::Stop(..) => break,
2123 _ => {}
2124 }
2125 }
2126
2127 assert_eq!(retry_events.len(), 1);
2128 assert!(matches!(
2129 retry_events[0],
2130 acp_thread::RetryStatus { attempt: 1, .. }
2131 ));
2132 thread.read_with(cx, |thread, _cx| {
2133 assert_eq!(
2134 thread.to_markdown(),
2135 indoc! {"
2136 ## User
2137
2138 Hello!
2139
2140 ## Assistant
2141
2142 Hey,
2143
2144 [resume]
2145
2146 ## Assistant
2147
2148 there!
2149 "}
2150 )
2151 });
2152}
2153
2154#[gpui::test]
2155async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {
2156 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2157 let fake_model = model.as_fake();
2158
2159 let events = thread
2160 .update(cx, |thread, cx| {
2161 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2162 thread.add_tool(EchoTool);
2163 thread.send(UserMessageId::new(), ["Call the echo tool!"], cx)
2164 })
2165 .unwrap();
2166 cx.run_until_parked();
2167
2168 let tool_use_1 = LanguageModelToolUse {
2169 id: "tool_1".into(),
2170 name: EchoTool::name().into(),
2171 raw_input: json!({"text": "test"}).to_string(),
2172 input: json!({"text": "test"}),
2173 is_input_complete: true,
2174 };
2175 fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
2176 tool_use_1.clone(),
2177 ));
2178 fake_model.send_last_completion_stream_error(LanguageModelCompletionError::ServerOverloaded {
2179 provider: LanguageModelProviderName::new("Anthropic"),
2180 retry_after: Some(Duration::from_secs(3)),
2181 });
2182 fake_model.end_last_completion_stream();
2183
2184 cx.executor().advance_clock(Duration::from_secs(3));
2185 let completion = fake_model.pending_completions().pop().unwrap();
2186 assert_eq!(
2187 completion.messages[1..],
2188 vec![
2189 LanguageModelRequestMessage {
2190 role: Role::User,
2191 content: vec!["Call the echo tool!".into()],
2192 cache: false
2193 },
2194 LanguageModelRequestMessage {
2195 role: Role::Assistant,
2196 content: vec![language_model::MessageContent::ToolUse(tool_use_1.clone())],
2197 cache: false
2198 },
2199 LanguageModelRequestMessage {
2200 role: Role::User,
2201 content: vec![language_model::MessageContent::ToolResult(
2202 LanguageModelToolResult {
2203 tool_use_id: tool_use_1.id.clone(),
2204 tool_name: tool_use_1.name.clone(),
2205 is_error: false,
2206 content: "test".into(),
2207 output: Some("test".into())
2208 }
2209 )],
2210 cache: true
2211 },
2212 ]
2213 );
2214
2215 fake_model.send_last_completion_stream_text_chunk("Done");
2216 fake_model.end_last_completion_stream();
2217 cx.run_until_parked();
2218 events.collect::<Vec<_>>().await;
2219 thread.read_with(cx, |thread, _cx| {
2220 assert_eq!(
2221 thread.last_message(),
2222 Some(Message::Agent(AgentMessage {
2223 content: vec![AgentMessageContent::Text("Done".into())],
2224 tool_results: IndexMap::default()
2225 }))
2226 );
2227 })
2228}
2229
2230#[gpui::test]
2231async fn test_send_max_retries_exceeded(cx: &mut TestAppContext) {
2232 let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
2233 let fake_model = model.as_fake();
2234
2235 let mut events = thread
2236 .update(cx, |thread, cx| {
2237 thread.set_completion_mode(agent_settings::CompletionMode::Burn, cx);
2238 thread.send(UserMessageId::new(), ["Hello!"], cx)
2239 })
2240 .unwrap();
2241 cx.run_until_parked();
2242
2243 for _ in 0..crate::thread::MAX_RETRY_ATTEMPTS + 1 {
2244 fake_model.send_last_completion_stream_error(
2245 LanguageModelCompletionError::ServerOverloaded {
2246 provider: LanguageModelProviderName::new("Anthropic"),
2247 retry_after: Some(Duration::from_secs(3)),
2248 },
2249 );
2250 fake_model.end_last_completion_stream();
2251 cx.executor().advance_clock(Duration::from_secs(3));
2252 cx.run_until_parked();
2253 }
2254
2255 let mut errors = Vec::new();
2256 let mut retry_events = Vec::new();
2257 while let Some(event) = events.next().await {
2258 match event {
2259 Ok(ThreadEvent::Retry(retry_status)) => {
2260 retry_events.push(retry_status);
2261 }
2262 Ok(ThreadEvent::Stop(..)) => break,
2263 Err(error) => errors.push(error),
2264 _ => {}
2265 }
2266 }
2267
2268 assert_eq!(
2269 retry_events.len(),
2270 crate::thread::MAX_RETRY_ATTEMPTS as usize
2271 );
2272 for i in 0..crate::thread::MAX_RETRY_ATTEMPTS as usize {
2273 assert_eq!(retry_events[i].attempt, i + 1);
2274 }
2275 assert_eq!(errors.len(), 1);
2276 let error = errors[0]
2277 .downcast_ref::<LanguageModelCompletionError>()
2278 .unwrap();
2279 assert!(matches!(
2280 error,
2281 LanguageModelCompletionError::ServerOverloaded { .. }
2282 ));
2283}
2284
2285/// Filters out the stop events for asserting against in tests
2286fn stop_events(result_events: Vec<Result<ThreadEvent>>) -> Vec<acp::StopReason> {
2287 result_events
2288 .into_iter()
2289 .filter_map(|event| match event.unwrap() {
2290 ThreadEvent::Stop(stop_reason) => Some(stop_reason),
2291 _ => None,
2292 })
2293 .collect()
2294}
2295
2296struct ThreadTest {
2297 model: Arc<dyn LanguageModel>,
2298 thread: Entity<Thread>,
2299 project_context: Entity<ProjectContext>,
2300 context_server_store: Entity<ContextServerStore>,
2301 fs: Arc<FakeFs>,
2302}
2303
2304enum TestModel {
2305 Sonnet4,
2306 Fake,
2307}
2308
2309impl TestModel {
2310 fn id(&self) -> LanguageModelId {
2311 match self {
2312 TestModel::Sonnet4 => LanguageModelId("claude-sonnet-4-latest".into()),
2313 TestModel::Fake => unreachable!(),
2314 }
2315 }
2316}
2317
2318async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {
2319 cx.executor().allow_parking();
2320
2321 let fs = FakeFs::new(cx.background_executor.clone());
2322 fs.create_dir(paths::settings_file().parent().unwrap())
2323 .await
2324 .unwrap();
2325 fs.insert_file(
2326 paths::settings_file(),
2327 json!({
2328 "agent": {
2329 "default_profile": "test-profile",
2330 "profiles": {
2331 "test-profile": {
2332 "name": "Test Profile",
2333 "tools": {
2334 EchoTool::name(): true,
2335 DelayTool::name(): true,
2336 WordListTool::name(): true,
2337 ToolRequiringPermission::name(): true,
2338 InfiniteTool::name(): true,
2339 ThinkingTool::name(): true,
2340 }
2341 }
2342 }
2343 }
2344 })
2345 .to_string()
2346 .into_bytes(),
2347 )
2348 .await;
2349
2350 cx.update(|cx| {
2351 settings::init(cx);
2352 Project::init_settings(cx);
2353 agent_settings::init(cx);
2354 gpui_tokio::init(cx);
2355 let http_client = ReqwestClient::user_agent("agent tests").unwrap();
2356 cx.set_http_client(Arc::new(http_client));
2357
2358 client::init_settings(cx);
2359 let client = Client::production(cx);
2360 let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
2361 language_model::init(client.clone(), cx);
2362 language_models::init(user_store, client.clone(), cx);
2363
2364 watch_settings(fs.clone(), cx);
2365 });
2366
2367 let templates = Templates::new();
2368
2369 fs.insert_tree(path!("/test"), json!({})).await;
2370 let project = Project::test(fs.clone(), [path!("/test").as_ref()], cx).await;
2371
2372 let model = cx
2373 .update(|cx| {
2374 if let TestModel::Fake = model {
2375 Task::ready(Arc::new(FakeLanguageModel::default()) as Arc<_>)
2376 } else {
2377 let model_id = model.id();
2378 let models = LanguageModelRegistry::read_global(cx);
2379 let model = models
2380 .available_models(cx)
2381 .find(|model| model.id() == model_id)
2382 .unwrap();
2383
2384 let provider = models.provider(&model.provider_id()).unwrap();
2385 let authenticated = provider.authenticate(cx);
2386
2387 cx.spawn(async move |_cx| {
2388 authenticated.await.unwrap();
2389 model
2390 })
2391 }
2392 })
2393 .await;
2394
2395 let project_context = cx.new(|_cx| ProjectContext::default());
2396 let context_server_store = project.read_with(cx, |project, _| project.context_server_store());
2397 let context_server_registry =
2398 cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));
2399 let thread = cx.new(|cx| {
2400 Thread::new(
2401 project,
2402 project_context.clone(),
2403 context_server_registry,
2404 templates,
2405 Some(model.clone()),
2406 cx,
2407 )
2408 });
2409 ThreadTest {
2410 model,
2411 thread,
2412 project_context,
2413 context_server_store,
2414 fs,
2415 }
2416}
2417
2418#[cfg(test)]
2419#[ctor::ctor]
2420fn init_logger() {
2421 if std::env::var("RUST_LOG").is_ok() {
2422 env_logger::init();
2423 }
2424}
2425
2426fn watch_settings(fs: Arc<dyn Fs>, cx: &mut App) {
2427 let fs = fs.clone();
2428 cx.spawn({
2429 async move |cx| {
2430 let mut new_settings_content_rx = settings::watch_config_file(
2431 cx.background_executor(),
2432 fs,
2433 paths::settings_file().clone(),
2434 );
2435
2436 while let Some(new_settings_content) = new_settings_content_rx.next().await {
2437 cx.update(|cx| {
2438 SettingsStore::update_global(cx, |settings, cx| {
2439 settings.set_user_settings(&new_settings_content, cx)
2440 })
2441 })
2442 .ok();
2443 }
2444 }
2445 })
2446 .detach();
2447}
2448
2449fn tool_names_for_completion(completion: &LanguageModelRequest) -> Vec<String> {
2450 completion
2451 .tools
2452 .iter()
2453 .map(|tool| tool.name.clone())
2454 .collect()
2455}
2456
2457fn setup_context_server(
2458 name: &'static str,
2459 tools: Vec<context_server::types::Tool>,
2460 context_server_store: &Entity<ContextServerStore>,
2461 cx: &mut TestAppContext,
2462) -> mpsc::UnboundedReceiver<(
2463 context_server::types::CallToolParams,
2464 oneshot::Sender<context_server::types::CallToolResponse>,
2465)> {
2466 cx.update(|cx| {
2467 let mut settings = ProjectSettings::get_global(cx).clone();
2468 settings.context_servers.insert(
2469 name.into(),
2470 project::project_settings::ContextServerSettings::Custom {
2471 enabled: true,
2472 command: ContextServerCommand {
2473 path: "somebinary".into(),
2474 args: Vec::new(),
2475 env: None,
2476 },
2477 },
2478 );
2479 ProjectSettings::override_global(settings, cx);
2480 });
2481
2482 let (mcp_tool_calls_tx, mcp_tool_calls_rx) = mpsc::unbounded();
2483 let fake_transport = context_server::test::create_fake_transport(name, cx.executor())
2484 .on_request::<context_server::types::requests::Initialize, _>(move |_params| async move {
2485 context_server::types::InitializeResponse {
2486 protocol_version: context_server::types::ProtocolVersion(
2487 context_server::types::LATEST_PROTOCOL_VERSION.to_string(),
2488 ),
2489 server_info: context_server::types::Implementation {
2490 name: name.into(),
2491 version: "1.0.0".to_string(),
2492 },
2493 capabilities: context_server::types::ServerCapabilities {
2494 tools: Some(context_server::types::ToolsCapabilities {
2495 list_changed: Some(true),
2496 }),
2497 ..Default::default()
2498 },
2499 meta: None,
2500 }
2501 })
2502 .on_request::<context_server::types::requests::ListTools, _>(move |_params| {
2503 let tools = tools.clone();
2504 async move {
2505 context_server::types::ListToolsResponse {
2506 tools,
2507 next_cursor: None,
2508 meta: None,
2509 }
2510 }
2511 })
2512 .on_request::<context_server::types::requests::CallTool, _>(move |params| {
2513 let mcp_tool_calls_tx = mcp_tool_calls_tx.clone();
2514 async move {
2515 let (response_tx, response_rx) = oneshot::channel();
2516 mcp_tool_calls_tx
2517 .unbounded_send((params, response_tx))
2518 .unwrap();
2519 response_rx.await.unwrap()
2520 }
2521 });
2522 context_server_store.update(cx, |store, cx| {
2523 store.start_server(
2524 Arc::new(ContextServer::new(
2525 ContextServerId(name.into()),
2526 Arc::new(fake_transport),
2527 )),
2528 cx,
2529 );
2530 });
2531 cx.run_until_parked();
2532 mcp_tool_calls_rx
2533}