1use crate::{
2 ContextServerRegistry, CopyPathTool, CreateDirectoryTool, DbLanguageModel, DbThread,
3 DeletePathTool, DiagnosticsTool, EditFileTool, FetchTool, FindPathTool, GrepTool,
4 ListDirectoryTool, MovePathTool, NowTool, OpenTool, ReadFileTool, SystemPromptTemplate,
5 Template, Templates, TerminalTool, ThinkingTool, WebSearchTool,
6};
7use acp_thread::{MentionUri, UserMessageId};
8use action_log::ActionLog;
9use agent::thread::{GitState, ProjectSnapshot, WorktreeSnapshot};
10use agent_client_protocol as acp;
11use agent_settings::{
12 AgentProfileId, AgentSettings, CompletionMode, SUMMARIZE_THREAD_DETAILED_PROMPT,
13 SUMMARIZE_THREAD_PROMPT,
14};
15use anyhow::{Context as _, Result, anyhow};
16use assistant_tool::adapt_schema_to_format;
17use chrono::{DateTime, Utc};
18use client::{ModelRequestUsage, RequestUsage};
19use cloud_llm_client::{CompletionIntent, CompletionRequestStatus, UsageLimit};
20use collections::{HashMap, IndexMap};
21use fs::Fs;
22use futures::{
23 FutureExt,
24 channel::{mpsc, oneshot},
25 future::Shared,
26 stream::FuturesUnordered,
27};
28use git::repository::DiffType;
29use gpui::{
30 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
31};
32use language_model::{
33 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelExt,
34 LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry, LanguageModelRequest,
35 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
36 LanguageModelToolResultContent, LanguageModelToolSchemaFormat, LanguageModelToolUse,
37 LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage,
38};
39use project::{
40 Project,
41 git_store::{GitStore, RepositoryState},
42};
43use prompt_store::ProjectContext;
44use schemars::{JsonSchema, Schema};
45use serde::{Deserialize, Serialize};
46use settings::{Settings, update_settings_file};
47use smol::stream::StreamExt;
48use std::{
49 collections::BTreeMap,
50 path::Path,
51 sync::Arc,
52 time::{Duration, Instant},
53};
54use std::{fmt::Write, ops::Range};
55use util::{ResultExt, markdown::MarkdownCodeBlock};
56use uuid::Uuid;
57
58const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
59
60/// The ID of the user prompt that initiated a request.
61///
62/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
63#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
64pub struct PromptId(Arc<str>);
65
66impl PromptId {
67 pub fn new() -> Self {
68 Self(Uuid::new_v4().to_string().into())
69 }
70}
71
72impl std::fmt::Display for PromptId {
73 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
74 write!(f, "{}", self.0)
75 }
76}
77
78pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
79pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
80
81#[derive(Debug, Clone)]
82enum RetryStrategy {
83 ExponentialBackoff {
84 initial_delay: Duration,
85 max_attempts: u8,
86 },
87 Fixed {
88 delay: Duration,
89 max_attempts: u8,
90 },
91}
92
93#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
94pub enum Message {
95 User(UserMessage),
96 Agent(AgentMessage),
97 Resume,
98}
99
100impl Message {
101 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
102 match self {
103 Message::Agent(agent_message) => Some(agent_message),
104 _ => None,
105 }
106 }
107
108 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
109 match self {
110 Message::User(message) => vec![message.to_request()],
111 Message::Agent(message) => message.to_request(),
112 Message::Resume => vec![LanguageModelRequestMessage {
113 role: Role::User,
114 content: vec!["Continue where you left off".into()],
115 cache: false,
116 }],
117 }
118 }
119
120 pub fn to_markdown(&self) -> String {
121 match self {
122 Message::User(message) => message.to_markdown(),
123 Message::Agent(message) => message.to_markdown(),
124 Message::Resume => "[resumed after tool use limit was reached]".into(),
125 }
126 }
127
128 pub fn role(&self) -> Role {
129 match self {
130 Message::User(_) | Message::Resume => Role::User,
131 Message::Agent(_) => Role::Assistant,
132 }
133 }
134}
135
136#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
137pub struct UserMessage {
138 pub id: UserMessageId,
139 pub content: Vec<UserMessageContent>,
140}
141
142#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
143pub enum UserMessageContent {
144 Text(String),
145 Mention { uri: MentionUri, content: String },
146 Image(LanguageModelImage),
147}
148
149impl UserMessage {
150 pub fn to_markdown(&self) -> String {
151 let mut markdown = String::from("## User\n\n");
152
153 for content in &self.content {
154 match content {
155 UserMessageContent::Text(text) => {
156 markdown.push_str(text);
157 markdown.push('\n');
158 }
159 UserMessageContent::Image(_) => {
160 markdown.push_str("<image />\n");
161 }
162 UserMessageContent::Mention { uri, content } => {
163 if !content.is_empty() {
164 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
165 } else {
166 let _ = writeln!(&mut markdown, "{}", uri.as_link());
167 }
168 }
169 }
170 }
171
172 markdown
173 }
174
175 fn to_request(&self) -> LanguageModelRequestMessage {
176 let mut message = LanguageModelRequestMessage {
177 role: Role::User,
178 content: Vec::with_capacity(self.content.len()),
179 cache: false,
180 };
181
182 const OPEN_CONTEXT: &str = "<context>\n\
183 The following items were attached by the user. \
184 They are up-to-date and don't need to be re-read.\n\n";
185
186 const OPEN_FILES_TAG: &str = "<files>";
187 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
188 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
189 const OPEN_THREADS_TAG: &str = "<threads>";
190 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
191 const OPEN_RULES_TAG: &str =
192 "<rules>\nThe user has specified the following rules that should be applied:\n";
193
194 let mut file_context = OPEN_FILES_TAG.to_string();
195 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
196 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
197 let mut thread_context = OPEN_THREADS_TAG.to_string();
198 let mut fetch_context = OPEN_FETCH_TAG.to_string();
199 let mut rules_context = OPEN_RULES_TAG.to_string();
200
201 for chunk in &self.content {
202 let chunk = match chunk {
203 UserMessageContent::Text(text) => {
204 language_model::MessageContent::Text(text.clone())
205 }
206 UserMessageContent::Image(value) => {
207 language_model::MessageContent::Image(value.clone())
208 }
209 UserMessageContent::Mention { uri, content } => {
210 match uri {
211 MentionUri::File { abs_path } => {
212 write!(
213 &mut symbol_context,
214 "\n{}",
215 MarkdownCodeBlock {
216 tag: &codeblock_tag(abs_path, None),
217 text: &content.to_string(),
218 }
219 )
220 .ok();
221 }
222 MentionUri::Directory { .. } => {
223 write!(&mut directory_context, "\n{}\n", content).ok();
224 }
225 MentionUri::Symbol {
226 path, line_range, ..
227 }
228 | MentionUri::Selection {
229 path, line_range, ..
230 } => {
231 write!(
232 &mut rules_context,
233 "\n{}",
234 MarkdownCodeBlock {
235 tag: &codeblock_tag(path, Some(line_range)),
236 text: content
237 }
238 )
239 .ok();
240 }
241 MentionUri::Thread { .. } => {
242 write!(&mut thread_context, "\n{}\n", content).ok();
243 }
244 MentionUri::TextThread { .. } => {
245 write!(&mut thread_context, "\n{}\n", content).ok();
246 }
247 MentionUri::Rule { .. } => {
248 write!(
249 &mut rules_context,
250 "\n{}",
251 MarkdownCodeBlock {
252 tag: "",
253 text: content
254 }
255 )
256 .ok();
257 }
258 MentionUri::Fetch { url } => {
259 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
260 }
261 }
262
263 language_model::MessageContent::Text(uri.as_link().to_string())
264 }
265 };
266
267 message.content.push(chunk);
268 }
269
270 let len_before_context = message.content.len();
271
272 if file_context.len() > OPEN_FILES_TAG.len() {
273 file_context.push_str("</files>\n");
274 message
275 .content
276 .push(language_model::MessageContent::Text(file_context));
277 }
278
279 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
280 directory_context.push_str("</directories>\n");
281 message
282 .content
283 .push(language_model::MessageContent::Text(directory_context));
284 }
285
286 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
287 symbol_context.push_str("</symbols>\n");
288 message
289 .content
290 .push(language_model::MessageContent::Text(symbol_context));
291 }
292
293 if thread_context.len() > OPEN_THREADS_TAG.len() {
294 thread_context.push_str("</threads>\n");
295 message
296 .content
297 .push(language_model::MessageContent::Text(thread_context));
298 }
299
300 if fetch_context.len() > OPEN_FETCH_TAG.len() {
301 fetch_context.push_str("</fetched_urls>\n");
302 message
303 .content
304 .push(language_model::MessageContent::Text(fetch_context));
305 }
306
307 if rules_context.len() > OPEN_RULES_TAG.len() {
308 rules_context.push_str("</user_rules>\n");
309 message
310 .content
311 .push(language_model::MessageContent::Text(rules_context));
312 }
313
314 if message.content.len() > len_before_context {
315 message.content.insert(
316 len_before_context,
317 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
318 );
319 message
320 .content
321 .push(language_model::MessageContent::Text("</context>".into()));
322 }
323
324 message
325 }
326}
327
328fn codeblock_tag(full_path: &Path, line_range: Option<&Range<u32>>) -> String {
329 let mut result = String::new();
330
331 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
332 let _ = write!(result, "{} ", extension);
333 }
334
335 let _ = write!(result, "{}", full_path.display());
336
337 if let Some(range) = line_range {
338 if range.start == range.end {
339 let _ = write!(result, ":{}", range.start + 1);
340 } else {
341 let _ = write!(result, ":{}-{}", range.start + 1, range.end + 1);
342 }
343 }
344
345 result
346}
347
348impl AgentMessage {
349 pub fn to_markdown(&self) -> String {
350 let mut markdown = String::from("## Assistant\n\n");
351
352 for content in &self.content {
353 match content {
354 AgentMessageContent::Text(text) => {
355 markdown.push_str(text);
356 markdown.push('\n');
357 }
358 AgentMessageContent::Thinking { text, .. } => {
359 markdown.push_str("<think>");
360 markdown.push_str(text);
361 markdown.push_str("</think>\n");
362 }
363 AgentMessageContent::RedactedThinking(_) => {
364 markdown.push_str("<redacted_thinking />\n")
365 }
366 AgentMessageContent::ToolUse(tool_use) => {
367 markdown.push_str(&format!(
368 "**Tool Use**: {} (ID: {})\n",
369 tool_use.name, tool_use.id
370 ));
371 markdown.push_str(&format!(
372 "{}\n",
373 MarkdownCodeBlock {
374 tag: "json",
375 text: &format!("{:#}", tool_use.input)
376 }
377 ));
378 }
379 }
380 }
381
382 for tool_result in self.tool_results.values() {
383 markdown.push_str(&format!(
384 "**Tool Result**: {} (ID: {})\n\n",
385 tool_result.tool_name, tool_result.tool_use_id
386 ));
387 if tool_result.is_error {
388 markdown.push_str("**ERROR:**\n");
389 }
390
391 match &tool_result.content {
392 LanguageModelToolResultContent::Text(text) => {
393 writeln!(markdown, "{text}\n").ok();
394 }
395 LanguageModelToolResultContent::Image(_) => {
396 writeln!(markdown, "<image />\n").ok();
397 }
398 }
399
400 if let Some(output) = tool_result.output.as_ref() {
401 writeln!(
402 markdown,
403 "**Debug Output**:\n\n```json\n{}\n```\n",
404 serde_json::to_string_pretty(output).unwrap()
405 )
406 .unwrap();
407 }
408 }
409
410 markdown
411 }
412
413 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
414 let mut assistant_message = LanguageModelRequestMessage {
415 role: Role::Assistant,
416 content: Vec::with_capacity(self.content.len()),
417 cache: false,
418 };
419 for chunk in &self.content {
420 let chunk = match chunk {
421 AgentMessageContent::Text(text) => {
422 language_model::MessageContent::Text(text.clone())
423 }
424 AgentMessageContent::Thinking { text, signature } => {
425 language_model::MessageContent::Thinking {
426 text: text.clone(),
427 signature: signature.clone(),
428 }
429 }
430 AgentMessageContent::RedactedThinking(value) => {
431 language_model::MessageContent::RedactedThinking(value.clone())
432 }
433 AgentMessageContent::ToolUse(value) => {
434 language_model::MessageContent::ToolUse(value.clone())
435 }
436 };
437 assistant_message.content.push(chunk);
438 }
439
440 let mut user_message = LanguageModelRequestMessage {
441 role: Role::User,
442 content: Vec::new(),
443 cache: false,
444 };
445
446 for tool_result in self.tool_results.values() {
447 user_message
448 .content
449 .push(language_model::MessageContent::ToolResult(
450 tool_result.clone(),
451 ));
452 }
453
454 let mut messages = Vec::new();
455 if !assistant_message.content.is_empty() {
456 messages.push(assistant_message);
457 }
458 if !user_message.content.is_empty() {
459 messages.push(user_message);
460 }
461 messages
462 }
463}
464
465#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
466pub struct AgentMessage {
467 pub content: Vec<AgentMessageContent>,
468 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
469}
470
471#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
472pub enum AgentMessageContent {
473 Text(String),
474 Thinking {
475 text: String,
476 signature: Option<String>,
477 },
478 RedactedThinking(String),
479 ToolUse(LanguageModelToolUse),
480}
481
482#[derive(Debug)]
483pub enum ThreadEvent {
484 UserMessage(UserMessage),
485 AgentText(String),
486 AgentThinking(String),
487 ToolCall(acp::ToolCall),
488 ToolCallUpdate(acp_thread::ToolCallUpdate),
489 ToolCallAuthorization(ToolCallAuthorization),
490 TitleUpdate(SharedString),
491 Retry(acp_thread::RetryStatus),
492 Stop(acp::StopReason),
493}
494
495#[derive(Debug)]
496pub struct ToolCallAuthorization {
497 pub tool_call: acp::ToolCallUpdate,
498 pub options: Vec<acp::PermissionOption>,
499 pub response: oneshot::Sender<acp::PermissionOptionId>,
500}
501
502pub struct Thread {
503 id: acp::SessionId,
504 prompt_id: PromptId,
505 updated_at: DateTime<Utc>,
506 title: Option<SharedString>,
507 summary: Option<SharedString>,
508 messages: Vec<Message>,
509 completion_mode: CompletionMode,
510 /// Holds the task that handles agent interaction until the end of the turn.
511 /// Survives across multiple requests as the model performs tool calls and
512 /// we run tools, report their results.
513 running_turn: Option<RunningTurn>,
514 pending_message: Option<AgentMessage>,
515 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
516 tool_use_limit_reached: bool,
517 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
518 #[allow(unused)]
519 cumulative_token_usage: TokenUsage,
520 #[allow(unused)]
521 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
522 context_server_registry: Entity<ContextServerRegistry>,
523 profile_id: AgentProfileId,
524 project_context: Entity<ProjectContext>,
525 templates: Arc<Templates>,
526 model: Option<Arc<dyn LanguageModel>>,
527 summarization_model: Option<Arc<dyn LanguageModel>>,
528 pub(crate) project: Entity<Project>,
529 pub(crate) action_log: Entity<ActionLog>,
530}
531
532impl Thread {
533 pub fn new(
534 project: Entity<Project>,
535 project_context: Entity<ProjectContext>,
536 context_server_registry: Entity<ContextServerRegistry>,
537 action_log: Entity<ActionLog>,
538 templates: Arc<Templates>,
539 model: Option<Arc<dyn LanguageModel>>,
540 cx: &mut Context<Self>,
541 ) -> Self {
542 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
543 Self {
544 id: acp::SessionId(uuid::Uuid::new_v4().to_string().into()),
545 prompt_id: PromptId::new(),
546 updated_at: Utc::now(),
547 title: None,
548 summary: None,
549 messages: Vec::new(),
550 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
551 running_turn: None,
552 pending_message: None,
553 tools: BTreeMap::default(),
554 tool_use_limit_reached: false,
555 request_token_usage: HashMap::default(),
556 cumulative_token_usage: TokenUsage::default(),
557 initial_project_snapshot: {
558 let project_snapshot = Self::project_snapshot(project.clone(), cx);
559 cx.foreground_executor()
560 .spawn(async move { Some(project_snapshot.await) })
561 .shared()
562 },
563 context_server_registry,
564 profile_id,
565 project_context,
566 templates,
567 model,
568 summarization_model: None,
569 project,
570 action_log,
571 }
572 }
573
574 pub fn id(&self) -> &acp::SessionId {
575 &self.id
576 }
577
578 pub fn replay(
579 &mut self,
580 cx: &mut Context<Self>,
581 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
582 let (tx, rx) = mpsc::unbounded();
583 let stream = ThreadEventStream(tx);
584 for message in &self.messages {
585 match message {
586 Message::User(user_message) => stream.send_user_message(user_message),
587 Message::Agent(assistant_message) => {
588 for content in &assistant_message.content {
589 match content {
590 AgentMessageContent::Text(text) => stream.send_text(text),
591 AgentMessageContent::Thinking { text, .. } => {
592 stream.send_thinking(text)
593 }
594 AgentMessageContent::RedactedThinking(_) => {}
595 AgentMessageContent::ToolUse(tool_use) => {
596 self.replay_tool_call(
597 tool_use,
598 assistant_message.tool_results.get(&tool_use.id),
599 &stream,
600 cx,
601 );
602 }
603 }
604 }
605 }
606 Message::Resume => {}
607 }
608 }
609 rx
610 }
611
612 fn replay_tool_call(
613 &self,
614 tool_use: &LanguageModelToolUse,
615 tool_result: Option<&LanguageModelToolResult>,
616 stream: &ThreadEventStream,
617 cx: &mut Context<Self>,
618 ) {
619 let Some(tool) = self.tools.get(tool_use.name.as_ref()) else {
620 stream
621 .0
622 .unbounded_send(Ok(ThreadEvent::ToolCall(acp::ToolCall {
623 id: acp::ToolCallId(tool_use.id.to_string().into()),
624 title: tool_use.name.to_string(),
625 kind: acp::ToolKind::Other,
626 status: acp::ToolCallStatus::Failed,
627 content: Vec::new(),
628 locations: Vec::new(),
629 raw_input: Some(tool_use.input.clone()),
630 raw_output: None,
631 })))
632 .ok();
633 return;
634 };
635
636 let title = tool.initial_title(tool_use.input.clone());
637 let kind = tool.kind();
638 stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
639
640 let output = tool_result
641 .as_ref()
642 .and_then(|result| result.output.clone());
643 if let Some(output) = output.clone() {
644 let tool_event_stream = ToolCallEventStream::new(
645 tool_use.id.clone(),
646 stream.clone(),
647 Some(self.project.read(cx).fs().clone()),
648 );
649 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
650 .log_err();
651 }
652
653 stream.update_tool_call_fields(
654 &tool_use.id,
655 acp::ToolCallUpdateFields {
656 status: Some(acp::ToolCallStatus::Completed),
657 raw_output: output,
658 ..Default::default()
659 },
660 );
661 }
662
663 pub fn from_db(
664 id: acp::SessionId,
665 db_thread: DbThread,
666 project: Entity<Project>,
667 project_context: Entity<ProjectContext>,
668 context_server_registry: Entity<ContextServerRegistry>,
669 action_log: Entity<ActionLog>,
670 templates: Arc<Templates>,
671 cx: &mut Context<Self>,
672 ) -> Self {
673 let profile_id = db_thread
674 .profile
675 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
676 let model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
677 db_thread
678 .model
679 .and_then(|model| {
680 let model = SelectedModel {
681 provider: model.provider.clone().into(),
682 model: model.model.into(),
683 };
684 registry.select_model(&model, cx)
685 })
686 .or_else(|| registry.default_model())
687 .map(|model| model.model)
688 });
689
690 Self {
691 id,
692 prompt_id: PromptId::new(),
693 title: if db_thread.title.is_empty() {
694 None
695 } else {
696 Some(db_thread.title.clone())
697 },
698 summary: db_thread.detailed_summary,
699 messages: db_thread.messages,
700 completion_mode: db_thread.completion_mode.unwrap_or_default(),
701 running_turn: None,
702 pending_message: None,
703 tools: BTreeMap::default(),
704 tool_use_limit_reached: false,
705 request_token_usage: db_thread.request_token_usage.clone(),
706 cumulative_token_usage: db_thread.cumulative_token_usage,
707 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
708 context_server_registry,
709 profile_id,
710 project_context,
711 templates,
712 model,
713 summarization_model: None,
714 project,
715 action_log,
716 updated_at: db_thread.updated_at,
717 }
718 }
719
720 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
721 let initial_project_snapshot = self.initial_project_snapshot.clone();
722 let mut thread = DbThread {
723 title: self.title(),
724 messages: self.messages.clone(),
725 updated_at: self.updated_at,
726 detailed_summary: self.summary.clone(),
727 initial_project_snapshot: None,
728 cumulative_token_usage: self.cumulative_token_usage,
729 request_token_usage: self.request_token_usage.clone(),
730 model: self.model.as_ref().map(|model| DbLanguageModel {
731 provider: model.provider_id().to_string(),
732 model: model.name().0.to_string(),
733 }),
734 completion_mode: Some(self.completion_mode),
735 profile: Some(self.profile_id.clone()),
736 };
737
738 cx.background_spawn(async move {
739 let initial_project_snapshot = initial_project_snapshot.await;
740 thread.initial_project_snapshot = initial_project_snapshot;
741 thread
742 })
743 }
744
745 /// Create a snapshot of the current project state including git information and unsaved buffers.
746 fn project_snapshot(
747 project: Entity<Project>,
748 cx: &mut Context<Self>,
749 ) -> Task<Arc<agent::thread::ProjectSnapshot>> {
750 let git_store = project.read(cx).git_store().clone();
751 let worktree_snapshots: Vec<_> = project
752 .read(cx)
753 .visible_worktrees(cx)
754 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
755 .collect();
756
757 cx.spawn(async move |_, cx| {
758 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
759
760 let mut unsaved_buffers = Vec::new();
761 cx.update(|app_cx| {
762 let buffer_store = project.read(app_cx).buffer_store();
763 for buffer_handle in buffer_store.read(app_cx).buffers() {
764 let buffer = buffer_handle.read(app_cx);
765 if buffer.is_dirty()
766 && let Some(file) = buffer.file()
767 {
768 let path = file.path().to_string_lossy().to_string();
769 unsaved_buffers.push(path);
770 }
771 }
772 })
773 .ok();
774
775 Arc::new(ProjectSnapshot {
776 worktree_snapshots,
777 unsaved_buffer_paths: unsaved_buffers,
778 timestamp: Utc::now(),
779 })
780 })
781 }
782
783 fn worktree_snapshot(
784 worktree: Entity<project::Worktree>,
785 git_store: Entity<GitStore>,
786 cx: &App,
787 ) -> Task<agent::thread::WorktreeSnapshot> {
788 cx.spawn(async move |cx| {
789 // Get worktree path and snapshot
790 let worktree_info = cx.update(|app_cx| {
791 let worktree = worktree.read(app_cx);
792 let path = worktree.abs_path().to_string_lossy().to_string();
793 let snapshot = worktree.snapshot();
794 (path, snapshot)
795 });
796
797 let Ok((worktree_path, _snapshot)) = worktree_info else {
798 return WorktreeSnapshot {
799 worktree_path: String::new(),
800 git_state: None,
801 };
802 };
803
804 let git_state = git_store
805 .update(cx, |git_store, cx| {
806 git_store
807 .repositories()
808 .values()
809 .find(|repo| {
810 repo.read(cx)
811 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
812 .is_some()
813 })
814 .cloned()
815 })
816 .ok()
817 .flatten()
818 .map(|repo| {
819 repo.update(cx, |repo, _| {
820 let current_branch =
821 repo.branch.as_ref().map(|branch| branch.name().to_owned());
822 repo.send_job(None, |state, _| async move {
823 let RepositoryState::Local { backend, .. } = state else {
824 return GitState {
825 remote_url: None,
826 head_sha: None,
827 current_branch,
828 diff: None,
829 };
830 };
831
832 let remote_url = backend.remote_url("origin");
833 let head_sha = backend.head_sha().await;
834 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
835
836 GitState {
837 remote_url,
838 head_sha,
839 current_branch,
840 diff,
841 }
842 })
843 })
844 });
845
846 let git_state = match git_state {
847 Some(git_state) => match git_state.ok() {
848 Some(git_state) => git_state.await.ok(),
849 None => None,
850 },
851 None => None,
852 };
853
854 WorktreeSnapshot {
855 worktree_path,
856 git_state,
857 }
858 })
859 }
860
861 pub fn project_context(&self) -> &Entity<ProjectContext> {
862 &self.project_context
863 }
864
865 pub fn project(&self) -> &Entity<Project> {
866 &self.project
867 }
868
869 pub fn action_log(&self) -> &Entity<ActionLog> {
870 &self.action_log
871 }
872
873 pub fn is_empty(&self) -> bool {
874 self.messages.is_empty() && self.title.is_none()
875 }
876
877 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
878 self.model.as_ref()
879 }
880
881 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
882 let old_usage = self.latest_token_usage();
883 self.model = Some(model);
884 let new_usage = self.latest_token_usage();
885 if old_usage != new_usage {
886 cx.emit(TokenUsageUpdated(new_usage));
887 }
888 cx.notify()
889 }
890
891 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
892 self.summarization_model.as_ref()
893 }
894
895 pub fn set_summarization_model(
896 &mut self,
897 model: Option<Arc<dyn LanguageModel>>,
898 cx: &mut Context<Self>,
899 ) {
900 self.summarization_model = model;
901 cx.notify()
902 }
903
904 pub fn completion_mode(&self) -> CompletionMode {
905 self.completion_mode
906 }
907
908 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
909 let old_usage = self.latest_token_usage();
910 self.completion_mode = mode;
911 let new_usage = self.latest_token_usage();
912 if old_usage != new_usage {
913 cx.emit(TokenUsageUpdated(new_usage));
914 }
915 cx.notify()
916 }
917
918 #[cfg(any(test, feature = "test-support"))]
919 pub fn last_message(&self) -> Option<Message> {
920 if let Some(message) = self.pending_message.clone() {
921 Some(Message::Agent(message))
922 } else {
923 self.messages.last().cloned()
924 }
925 }
926
927 pub fn add_default_tools(&mut self, cx: &mut Context<Self>) {
928 let language_registry = self.project.read(cx).languages().clone();
929 self.add_tool(CopyPathTool::new(self.project.clone()));
930 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
931 self.add_tool(DeletePathTool::new(
932 self.project.clone(),
933 self.action_log.clone(),
934 ));
935 self.add_tool(DiagnosticsTool::new(self.project.clone()));
936 self.add_tool(EditFileTool::new(cx.weak_entity(), language_registry));
937 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
938 self.add_tool(FindPathTool::new(self.project.clone()));
939 self.add_tool(GrepTool::new(self.project.clone()));
940 self.add_tool(ListDirectoryTool::new(self.project.clone()));
941 self.add_tool(MovePathTool::new(self.project.clone()));
942 self.add_tool(NowTool);
943 self.add_tool(OpenTool::new(self.project.clone()));
944 self.add_tool(ReadFileTool::new(
945 self.project.clone(),
946 self.action_log.clone(),
947 ));
948 self.add_tool(TerminalTool::new(self.project.clone(), cx));
949 self.add_tool(ThinkingTool);
950 self.add_tool(WebSearchTool); // TODO: Enable this only if it's a zed model.
951 }
952
953 pub fn add_tool(&mut self, tool: impl AgentTool) {
954 self.tools.insert(tool.name(), tool.erase());
955 }
956
957 pub fn remove_tool(&mut self, name: &str) -> bool {
958 self.tools.remove(name).is_some()
959 }
960
961 pub fn profile(&self) -> &AgentProfileId {
962 &self.profile_id
963 }
964
965 pub fn set_profile(&mut self, profile_id: AgentProfileId) {
966 self.profile_id = profile_id;
967 }
968
969 pub fn cancel(&mut self, cx: &mut Context<Self>) {
970 if let Some(running_turn) = self.running_turn.take() {
971 running_turn.cancel();
972 }
973 self.flush_pending_message(cx);
974 }
975
976 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
977 let Some(last_user_message) = self.last_user_message() else {
978 return;
979 };
980
981 self.request_token_usage
982 .insert(last_user_message.id.clone(), update);
983 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
984 cx.notify();
985 }
986
987 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
988 self.cancel(cx);
989 let Some(position) = self.messages.iter().position(
990 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
991 ) else {
992 return Err(anyhow!("Message not found"));
993 };
994
995 for message in self.messages.drain(position..) {
996 match message {
997 Message::User(message) => {
998 self.request_token_usage.remove(&message.id);
999 }
1000 Message::Agent(_) | Message::Resume => {}
1001 }
1002 }
1003 self.summary = None;
1004 cx.notify();
1005 Ok(())
1006 }
1007
1008 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1009 let last_user_message = self.last_user_message()?;
1010 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1011 let model = self.model.clone()?;
1012
1013 Some(acp_thread::TokenUsage {
1014 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1015 used_tokens: tokens.total_tokens(),
1016 })
1017 }
1018
1019 pub fn resume(
1020 &mut self,
1021 cx: &mut Context<Self>,
1022 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1023 anyhow::ensure!(
1024 self.tool_use_limit_reached,
1025 "can only resume after tool use limit is reached"
1026 );
1027
1028 self.messages.push(Message::Resume);
1029 cx.notify();
1030
1031 log::info!("Total messages in thread: {}", self.messages.len());
1032 self.run_turn(cx)
1033 }
1034
1035 /// Sending a message results in the model streaming a response, which could include tool calls.
1036 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1037 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1038 pub fn send<T>(
1039 &mut self,
1040 id: UserMessageId,
1041 content: impl IntoIterator<Item = T>,
1042 cx: &mut Context<Self>,
1043 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1044 where
1045 T: Into<UserMessageContent>,
1046 {
1047 let model = self.model().context("No language model configured")?;
1048
1049 log::info!("Thread::send called with model: {:?}", model.name());
1050 self.advance_prompt_id();
1051
1052 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1053 log::debug!("Thread::send content: {:?}", content);
1054
1055 self.messages
1056 .push(Message::User(UserMessage { id, content }));
1057 cx.notify();
1058
1059 log::info!("Total messages in thread: {}", self.messages.len());
1060 self.run_turn(cx)
1061 }
1062
1063 fn run_turn(
1064 &mut self,
1065 cx: &mut Context<Self>,
1066 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1067 self.cancel(cx);
1068
1069 let model = self.model.clone().context("No language model configured")?;
1070 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1071 let event_stream = ThreadEventStream(events_tx);
1072 let message_ix = self.messages.len().saturating_sub(1);
1073 self.tool_use_limit_reached = false;
1074 self.summary = None;
1075 self.running_turn = Some(RunningTurn {
1076 event_stream: event_stream.clone(),
1077 _task: cx.spawn(async move |this, cx| {
1078 log::info!("Starting agent turn execution");
1079 let mut update_title = None;
1080 let turn_result: Result<StopReason> = async {
1081 let mut completion_intent = CompletionIntent::UserPrompt;
1082 loop {
1083 log::debug!(
1084 "Building completion request with intent: {:?}",
1085 completion_intent
1086 );
1087 let request = this.update(cx, |this, cx| {
1088 this.build_completion_request(completion_intent, cx)
1089 })??;
1090
1091 log::info!("Calling model.stream_completion");
1092
1093 let mut tool_use_limit_reached = false;
1094 let mut refused = false;
1095 let mut reached_max_tokens = false;
1096 let mut tool_uses = Self::stream_completion_with_retries(
1097 this.clone(),
1098 model.clone(),
1099 request,
1100 &event_stream,
1101 &mut tool_use_limit_reached,
1102 &mut refused,
1103 &mut reached_max_tokens,
1104 cx,
1105 )
1106 .await?;
1107
1108 if refused {
1109 return Ok(StopReason::Refusal);
1110 } else if reached_max_tokens {
1111 return Ok(StopReason::MaxTokens);
1112 }
1113
1114 let end_turn = tool_uses.is_empty();
1115 while let Some(tool_result) = tool_uses.next().await {
1116 log::info!("Tool finished {:?}", tool_result);
1117
1118 event_stream.update_tool_call_fields(
1119 &tool_result.tool_use_id,
1120 acp::ToolCallUpdateFields {
1121 status: Some(if tool_result.is_error {
1122 acp::ToolCallStatus::Failed
1123 } else {
1124 acp::ToolCallStatus::Completed
1125 }),
1126 raw_output: tool_result.output.clone(),
1127 ..Default::default()
1128 },
1129 );
1130 this.update(cx, |this, _cx| {
1131 this.pending_message()
1132 .tool_results
1133 .insert(tool_result.tool_use_id.clone(), tool_result);
1134 })?;
1135 }
1136
1137 this.update(cx, |this, cx| {
1138 if this.title.is_none() && update_title.is_none() {
1139 update_title = Some(this.update_title(&event_stream, cx));
1140 }
1141 })?;
1142
1143 if tool_use_limit_reached {
1144 log::info!("Tool use limit reached, completing turn");
1145 this.update(cx, |this, _cx| this.tool_use_limit_reached = true)?;
1146 return Err(language_model::ToolUseLimitReachedError.into());
1147 } else if end_turn {
1148 log::info!("No tool uses found, completing turn");
1149 return Ok(StopReason::EndTurn);
1150 } else {
1151 this.update(cx, |this, cx| this.flush_pending_message(cx))?;
1152 completion_intent = CompletionIntent::ToolResults;
1153 }
1154 }
1155 }
1156 .await;
1157 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1158
1159 match turn_result {
1160 Ok(reason) => {
1161 log::info!("Turn execution completed: {:?}", reason);
1162
1163 if let Some(update_title) = update_title {
1164 update_title.await.context("update title failed").log_err();
1165 }
1166
1167 event_stream.send_stop(reason);
1168 if reason == StopReason::Refusal {
1169 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1170 }
1171 }
1172 Err(error) => {
1173 log::error!("Turn execution failed: {:?}", error);
1174 event_stream.send_error(error);
1175 }
1176 }
1177
1178 _ = this.update(cx, |this, _| this.running_turn.take());
1179 }),
1180 });
1181 Ok(events_rx)
1182 }
1183
1184 async fn stream_completion_with_retries(
1185 this: WeakEntity<Self>,
1186 model: Arc<dyn LanguageModel>,
1187 request: LanguageModelRequest,
1188 event_stream: &ThreadEventStream,
1189 tool_use_limit_reached: &mut bool,
1190 refusal: &mut bool,
1191 max_tokens_reached: &mut bool,
1192 cx: &mut AsyncApp,
1193 ) -> Result<FuturesUnordered<Task<LanguageModelToolResult>>> {
1194 log::debug!("Stream completion started successfully");
1195
1196 let mut attempt = None;
1197 'retry: loop {
1198 telemetry::event!(
1199 "Agent Thread Completion",
1200 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1201 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1202 model = model.telemetry_id(),
1203 model_provider = model.provider_id().to_string(),
1204 attempt
1205 );
1206
1207 let mut events = model.stream_completion(request.clone(), cx).await?;
1208 let mut tool_uses = FuturesUnordered::new();
1209 while let Some(event) = events.next().await {
1210 match event {
1211 Ok(LanguageModelCompletionEvent::StatusUpdate(
1212 CompletionRequestStatus::ToolUseLimitReached,
1213 )) => {
1214 *tool_use_limit_reached = true;
1215 }
1216 Ok(LanguageModelCompletionEvent::StatusUpdate(
1217 CompletionRequestStatus::UsageUpdated { amount, limit },
1218 )) => {
1219 this.update(cx, |this, cx| {
1220 this.update_model_request_usage(amount, limit, cx)
1221 })?;
1222 }
1223 Ok(LanguageModelCompletionEvent::UsageUpdate(usage)) => {
1224 telemetry::event!(
1225 "Agent Thread Completion Usage Updated",
1226 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1227 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1228 model = model.telemetry_id(),
1229 model_provider = model.provider_id().to_string(),
1230 attempt,
1231 input_tokens = usage.input_tokens,
1232 output_tokens = usage.output_tokens,
1233 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1234 cache_read_input_tokens = usage.cache_read_input_tokens,
1235 );
1236
1237 this.update(cx, |this, cx| this.update_token_usage(usage, cx))?;
1238 }
1239 Ok(LanguageModelCompletionEvent::Stop(StopReason::Refusal)) => {
1240 *refusal = true;
1241 return Ok(FuturesUnordered::default());
1242 }
1243 Ok(LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)) => {
1244 *max_tokens_reached = true;
1245 return Ok(FuturesUnordered::default());
1246 }
1247 Ok(LanguageModelCompletionEvent::Stop(
1248 StopReason::ToolUse | StopReason::EndTurn,
1249 )) => break,
1250 Ok(event) => {
1251 log::trace!("Received completion event: {:?}", event);
1252 this.update(cx, |this, cx| {
1253 tool_uses.extend(this.handle_streamed_completion_event(
1254 event,
1255 event_stream,
1256 cx,
1257 ));
1258 })?;
1259 }
1260 Err(error) => {
1261 let completion_mode =
1262 this.read_with(cx, |thread, _cx| thread.completion_mode())?;
1263 if completion_mode == CompletionMode::Normal {
1264 return Err(error.into());
1265 }
1266
1267 let Some(strategy) = Self::retry_strategy_for(&error) else {
1268 return Err(error.into());
1269 };
1270
1271 let max_attempts = match &strategy {
1272 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1273 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1274 };
1275
1276 let attempt = attempt.get_or_insert(0u8);
1277
1278 *attempt += 1;
1279
1280 let attempt = *attempt;
1281 if attempt > max_attempts {
1282 return Err(error.into());
1283 }
1284
1285 let delay = match &strategy {
1286 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1287 let delay_secs =
1288 initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1289 Duration::from_secs(delay_secs)
1290 }
1291 RetryStrategy::Fixed { delay, .. } => *delay,
1292 };
1293 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1294
1295 event_stream.send_retry(acp_thread::RetryStatus {
1296 last_error: error.to_string().into(),
1297 attempt: attempt as usize,
1298 max_attempts: max_attempts as usize,
1299 started_at: Instant::now(),
1300 duration: delay,
1301 });
1302
1303 cx.background_executor().timer(delay).await;
1304 continue 'retry;
1305 }
1306 }
1307 }
1308
1309 return Ok(tool_uses);
1310 }
1311 }
1312
1313 pub fn build_system_message(&self, cx: &App) -> LanguageModelRequestMessage {
1314 log::debug!("Building system message");
1315 let prompt = SystemPromptTemplate {
1316 project: self.project_context.read(cx),
1317 available_tools: self.tools.keys().cloned().collect(),
1318 }
1319 .render(&self.templates)
1320 .context("failed to build system prompt")
1321 .expect("Invalid template");
1322 log::debug!("System message built");
1323 LanguageModelRequestMessage {
1324 role: Role::System,
1325 content: vec![prompt.into()],
1326 cache: true,
1327 }
1328 }
1329
1330 /// A helper method that's called on every streamed completion event.
1331 /// Returns an optional tool result task, which the main agentic loop in
1332 /// send will send back to the model when it resolves.
1333 fn handle_streamed_completion_event(
1334 &mut self,
1335 event: LanguageModelCompletionEvent,
1336 event_stream: &ThreadEventStream,
1337 cx: &mut Context<Self>,
1338 ) -> Option<Task<LanguageModelToolResult>> {
1339 log::trace!("Handling streamed completion event: {:?}", event);
1340 use LanguageModelCompletionEvent::*;
1341
1342 match event {
1343 StartMessage { .. } => {
1344 self.flush_pending_message(cx);
1345 self.pending_message = Some(AgentMessage::default());
1346 }
1347 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1348 Thinking { text, signature } => {
1349 self.handle_thinking_event(text, signature, event_stream, cx)
1350 }
1351 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1352 ToolUse(tool_use) => {
1353 return self.handle_tool_use_event(tool_use, event_stream, cx);
1354 }
1355 ToolUseJsonParseError {
1356 id,
1357 tool_name,
1358 raw_input,
1359 json_parse_error,
1360 } => {
1361 return Some(Task::ready(self.handle_tool_use_json_parse_error_event(
1362 id,
1363 tool_name,
1364 raw_input,
1365 json_parse_error,
1366 )));
1367 }
1368 StatusUpdate(_) => {}
1369 UsageUpdate(_) | Stop(_) => unreachable!(),
1370 }
1371
1372 None
1373 }
1374
1375 fn handle_text_event(
1376 &mut self,
1377 new_text: String,
1378 event_stream: &ThreadEventStream,
1379 cx: &mut Context<Self>,
1380 ) {
1381 event_stream.send_text(&new_text);
1382
1383 let last_message = self.pending_message();
1384 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1385 text.push_str(&new_text);
1386 } else {
1387 last_message
1388 .content
1389 .push(AgentMessageContent::Text(new_text));
1390 }
1391
1392 cx.notify();
1393 }
1394
1395 fn handle_thinking_event(
1396 &mut self,
1397 new_text: String,
1398 new_signature: Option<String>,
1399 event_stream: &ThreadEventStream,
1400 cx: &mut Context<Self>,
1401 ) {
1402 event_stream.send_thinking(&new_text);
1403
1404 let last_message = self.pending_message();
1405 if let Some(AgentMessageContent::Thinking { text, signature }) =
1406 last_message.content.last_mut()
1407 {
1408 text.push_str(&new_text);
1409 *signature = new_signature.or(signature.take());
1410 } else {
1411 last_message.content.push(AgentMessageContent::Thinking {
1412 text: new_text,
1413 signature: new_signature,
1414 });
1415 }
1416
1417 cx.notify();
1418 }
1419
1420 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1421 let last_message = self.pending_message();
1422 last_message
1423 .content
1424 .push(AgentMessageContent::RedactedThinking(data));
1425 cx.notify();
1426 }
1427
1428 fn handle_tool_use_event(
1429 &mut self,
1430 tool_use: LanguageModelToolUse,
1431 event_stream: &ThreadEventStream,
1432 cx: &mut Context<Self>,
1433 ) -> Option<Task<LanguageModelToolResult>> {
1434 cx.notify();
1435
1436 let tool = self.tools.get(tool_use.name.as_ref()).cloned();
1437 let mut title = SharedString::from(&tool_use.name);
1438 let mut kind = acp::ToolKind::Other;
1439 if let Some(tool) = tool.as_ref() {
1440 title = tool.initial_title(tool_use.input.clone());
1441 kind = tool.kind();
1442 }
1443
1444 // Ensure the last message ends in the current tool use
1445 let last_message = self.pending_message();
1446 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1447 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1448 if last_tool_use.id == tool_use.id {
1449 *last_tool_use = tool_use.clone();
1450 false
1451 } else {
1452 true
1453 }
1454 } else {
1455 true
1456 }
1457 });
1458
1459 if push_new_tool_use {
1460 event_stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
1461 last_message
1462 .content
1463 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1464 } else {
1465 event_stream.update_tool_call_fields(
1466 &tool_use.id,
1467 acp::ToolCallUpdateFields {
1468 title: Some(title.into()),
1469 kind: Some(kind),
1470 raw_input: Some(tool_use.input.clone()),
1471 ..Default::default()
1472 },
1473 );
1474 }
1475
1476 if !tool_use.is_input_complete {
1477 return None;
1478 }
1479
1480 let Some(tool) = tool else {
1481 let content = format!("No tool named {} exists", tool_use.name);
1482 return Some(Task::ready(LanguageModelToolResult {
1483 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1484 tool_use_id: tool_use.id,
1485 tool_name: tool_use.name,
1486 is_error: true,
1487 output: None,
1488 }));
1489 };
1490
1491 let fs = self.project.read(cx).fs().clone();
1492 let tool_event_stream =
1493 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1494 tool_event_stream.update_fields(acp::ToolCallUpdateFields {
1495 status: Some(acp::ToolCallStatus::InProgress),
1496 ..Default::default()
1497 });
1498 let supports_images = self.model().is_some_and(|model| model.supports_images());
1499 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1500 log::info!("Running tool {}", tool_use.name);
1501 Some(cx.foreground_executor().spawn(async move {
1502 let tool_result = tool_result.await.and_then(|output| {
1503 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1504 && !supports_images
1505 {
1506 return Err(anyhow!(
1507 "Attempted to read an image, but this model doesn't support it.",
1508 ));
1509 }
1510 Ok(output)
1511 });
1512
1513 match tool_result {
1514 Ok(output) => LanguageModelToolResult {
1515 tool_use_id: tool_use.id,
1516 tool_name: tool_use.name,
1517 is_error: false,
1518 content: output.llm_output,
1519 output: Some(output.raw_output),
1520 },
1521 Err(error) => LanguageModelToolResult {
1522 tool_use_id: tool_use.id,
1523 tool_name: tool_use.name,
1524 is_error: true,
1525 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1526 output: None,
1527 },
1528 }
1529 }))
1530 }
1531
1532 fn handle_tool_use_json_parse_error_event(
1533 &mut self,
1534 tool_use_id: LanguageModelToolUseId,
1535 tool_name: Arc<str>,
1536 raw_input: Arc<str>,
1537 json_parse_error: String,
1538 ) -> LanguageModelToolResult {
1539 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1540 LanguageModelToolResult {
1541 tool_use_id,
1542 tool_name,
1543 is_error: true,
1544 content: LanguageModelToolResultContent::Text(tool_output.into()),
1545 output: Some(serde_json::Value::String(raw_input.to_string())),
1546 }
1547 }
1548
1549 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1550 self.project
1551 .read(cx)
1552 .user_store()
1553 .update(cx, |user_store, cx| {
1554 user_store.update_model_request_usage(
1555 ModelRequestUsage(RequestUsage {
1556 amount: amount as i32,
1557 limit,
1558 }),
1559 cx,
1560 )
1561 });
1562 }
1563
1564 pub fn title(&self) -> SharedString {
1565 self.title.clone().unwrap_or("New Thread".into())
1566 }
1567
1568 pub fn summary(&mut self, cx: &mut Context<Self>) -> Task<Result<SharedString>> {
1569 if let Some(summary) = self.summary.as_ref() {
1570 return Task::ready(Ok(summary.clone()));
1571 }
1572 let Some(model) = self.summarization_model.clone() else {
1573 return Task::ready(Err(anyhow!("No summarization model available")));
1574 };
1575 let mut request = LanguageModelRequest {
1576 intent: Some(CompletionIntent::ThreadContextSummarization),
1577 temperature: AgentSettings::temperature_for_model(&model, cx),
1578 ..Default::default()
1579 };
1580
1581 for message in &self.messages {
1582 request.messages.extend(message.to_request());
1583 }
1584
1585 request.messages.push(LanguageModelRequestMessage {
1586 role: Role::User,
1587 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1588 cache: false,
1589 });
1590 cx.spawn(async move |this, cx| {
1591 let mut summary = String::new();
1592 let mut messages = model.stream_completion(request, cx).await?;
1593 while let Some(event) = messages.next().await {
1594 let event = event?;
1595 let text = match event {
1596 LanguageModelCompletionEvent::Text(text) => text,
1597 LanguageModelCompletionEvent::StatusUpdate(
1598 CompletionRequestStatus::UsageUpdated { amount, limit },
1599 ) => {
1600 this.update(cx, |thread, cx| {
1601 thread.update_model_request_usage(amount, limit, cx);
1602 })?;
1603 continue;
1604 }
1605 _ => continue,
1606 };
1607
1608 let mut lines = text.lines();
1609 summary.extend(lines.next());
1610 }
1611
1612 log::info!("Setting summary: {}", summary);
1613 let summary = SharedString::from(summary);
1614
1615 this.update(cx, |this, cx| {
1616 this.summary = Some(summary.clone());
1617 cx.notify()
1618 })?;
1619
1620 Ok(summary)
1621 })
1622 }
1623
1624 fn update_title(
1625 &mut self,
1626 event_stream: &ThreadEventStream,
1627 cx: &mut Context<Self>,
1628 ) -> Task<Result<()>> {
1629 log::info!(
1630 "Generating title with model: {:?}",
1631 self.summarization_model.as_ref().map(|model| model.name())
1632 );
1633 let Some(model) = self.summarization_model.clone() else {
1634 return Task::ready(Ok(()));
1635 };
1636 let event_stream = event_stream.clone();
1637 let mut request = LanguageModelRequest {
1638 intent: Some(CompletionIntent::ThreadSummarization),
1639 temperature: AgentSettings::temperature_for_model(&model, cx),
1640 ..Default::default()
1641 };
1642
1643 for message in &self.messages {
1644 request.messages.extend(message.to_request());
1645 }
1646
1647 request.messages.push(LanguageModelRequestMessage {
1648 role: Role::User,
1649 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1650 cache: false,
1651 });
1652 cx.spawn(async move |this, cx| {
1653 let mut title = String::new();
1654 let mut messages = model.stream_completion(request, cx).await?;
1655 while let Some(event) = messages.next().await {
1656 let event = event?;
1657 let text = match event {
1658 LanguageModelCompletionEvent::Text(text) => text,
1659 LanguageModelCompletionEvent::StatusUpdate(
1660 CompletionRequestStatus::UsageUpdated { amount, limit },
1661 ) => {
1662 this.update(cx, |thread, cx| {
1663 thread.update_model_request_usage(amount, limit, cx);
1664 })?;
1665 continue;
1666 }
1667 _ => continue,
1668 };
1669
1670 let mut lines = text.lines();
1671 title.extend(lines.next());
1672
1673 // Stop if the LLM generated multiple lines.
1674 if lines.next().is_some() {
1675 break;
1676 }
1677 }
1678
1679 log::info!("Setting title: {}", title);
1680
1681 this.update(cx, |this, cx| {
1682 let title = SharedString::from(title);
1683 event_stream.send_title_update(title.clone());
1684 this.title = Some(title);
1685 cx.notify();
1686 })
1687 })
1688 }
1689
1690 fn last_user_message(&self) -> Option<&UserMessage> {
1691 self.messages
1692 .iter()
1693 .rev()
1694 .find_map(|message| match message {
1695 Message::User(user_message) => Some(user_message),
1696 Message::Agent(_) => None,
1697 Message::Resume => None,
1698 })
1699 }
1700
1701 fn pending_message(&mut self) -> &mut AgentMessage {
1702 self.pending_message.get_or_insert_default()
1703 }
1704
1705 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1706 let Some(mut message) = self.pending_message.take() else {
1707 return;
1708 };
1709
1710 for content in &message.content {
1711 let AgentMessageContent::ToolUse(tool_use) = content else {
1712 continue;
1713 };
1714
1715 if !message.tool_results.contains_key(&tool_use.id) {
1716 message.tool_results.insert(
1717 tool_use.id.clone(),
1718 LanguageModelToolResult {
1719 tool_use_id: tool_use.id.clone(),
1720 tool_name: tool_use.name.clone(),
1721 is_error: true,
1722 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1723 output: None,
1724 },
1725 );
1726 }
1727 }
1728
1729 self.messages.push(Message::Agent(message));
1730 self.updated_at = Utc::now();
1731 self.summary = None;
1732 cx.notify()
1733 }
1734
1735 pub(crate) fn build_completion_request(
1736 &self,
1737 completion_intent: CompletionIntent,
1738 cx: &mut App,
1739 ) -> Result<LanguageModelRequest> {
1740 let model = self.model().context("No language model configured")?;
1741
1742 log::debug!("Building completion request");
1743 log::debug!("Completion intent: {:?}", completion_intent);
1744 log::debug!("Completion mode: {:?}", self.completion_mode);
1745
1746 let messages = self.build_request_messages(cx);
1747 log::info!("Request will include {} messages", messages.len());
1748
1749 let tools = if let Some(tools) = self.tools(cx).log_err() {
1750 tools
1751 .filter_map(|tool| {
1752 let tool_name = tool.name().to_string();
1753 log::trace!("Including tool: {}", tool_name);
1754 Some(LanguageModelRequestTool {
1755 name: tool_name,
1756 description: tool.description().to_string(),
1757 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1758 })
1759 })
1760 .collect()
1761 } else {
1762 Vec::new()
1763 };
1764
1765 log::info!("Request includes {} tools", tools.len());
1766
1767 let request = LanguageModelRequest {
1768 thread_id: Some(self.id.to_string()),
1769 prompt_id: Some(self.prompt_id.to_string()),
1770 intent: Some(completion_intent),
1771 mode: Some(self.completion_mode.into()),
1772 messages,
1773 tools,
1774 tool_choice: None,
1775 stop: Vec::new(),
1776 temperature: AgentSettings::temperature_for_model(model, cx),
1777 thinking_allowed: true,
1778 };
1779
1780 log::debug!("Completion request built successfully");
1781 Ok(request)
1782 }
1783
1784 fn tools<'a>(&'a self, cx: &'a App) -> Result<impl Iterator<Item = &'a Arc<dyn AnyAgentTool>>> {
1785 let model = self.model().context("No language model configured")?;
1786
1787 let profile = AgentSettings::get_global(cx)
1788 .profiles
1789 .get(&self.profile_id)
1790 .context("profile not found")?;
1791 let provider_id = model.provider_id();
1792
1793 Ok(self
1794 .tools
1795 .iter()
1796 .filter(move |(_, tool)| tool.supported_provider(&provider_id))
1797 .filter_map(|(tool_name, tool)| {
1798 if profile.is_tool_enabled(tool_name) {
1799 Some(tool)
1800 } else {
1801 None
1802 }
1803 })
1804 .chain(self.context_server_registry.read(cx).servers().flat_map(
1805 |(server_id, tools)| {
1806 tools.iter().filter_map(|(tool_name, tool)| {
1807 if profile.is_context_server_tool_enabled(&server_id.0, tool_name) {
1808 Some(tool)
1809 } else {
1810 None
1811 }
1812 })
1813 },
1814 )))
1815 }
1816
1817 fn build_request_messages(&self, cx: &App) -> Vec<LanguageModelRequestMessage> {
1818 log::trace!(
1819 "Building request messages from {} thread messages",
1820 self.messages.len()
1821 );
1822 let mut messages = vec![self.build_system_message(cx)];
1823 for message in &self.messages {
1824 messages.extend(message.to_request());
1825 }
1826
1827 if let Some(message) = self.pending_message.as_ref() {
1828 messages.extend(message.to_request());
1829 }
1830
1831 if let Some(last_user_message) = messages
1832 .iter_mut()
1833 .rev()
1834 .find(|message| message.role == Role::User)
1835 {
1836 last_user_message.cache = true;
1837 }
1838
1839 messages
1840 }
1841
1842 pub fn to_markdown(&self) -> String {
1843 let mut markdown = String::new();
1844 for (ix, message) in self.messages.iter().enumerate() {
1845 if ix > 0 {
1846 markdown.push('\n');
1847 }
1848 markdown.push_str(&message.to_markdown());
1849 }
1850
1851 if let Some(message) = self.pending_message.as_ref() {
1852 markdown.push('\n');
1853 markdown.push_str(&message.to_markdown());
1854 }
1855
1856 markdown
1857 }
1858
1859 fn advance_prompt_id(&mut self) {
1860 self.prompt_id = PromptId::new();
1861 }
1862
1863 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
1864 use LanguageModelCompletionError::*;
1865 use http_client::StatusCode;
1866
1867 // General strategy here:
1868 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
1869 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
1870 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
1871 match error {
1872 HttpResponseError {
1873 status_code: StatusCode::TOO_MANY_REQUESTS,
1874 ..
1875 } => Some(RetryStrategy::ExponentialBackoff {
1876 initial_delay: BASE_RETRY_DELAY,
1877 max_attempts: MAX_RETRY_ATTEMPTS,
1878 }),
1879 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
1880 Some(RetryStrategy::Fixed {
1881 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1882 max_attempts: MAX_RETRY_ATTEMPTS,
1883 })
1884 }
1885 UpstreamProviderError {
1886 status,
1887 retry_after,
1888 ..
1889 } => match *status {
1890 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
1891 Some(RetryStrategy::Fixed {
1892 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1893 max_attempts: MAX_RETRY_ATTEMPTS,
1894 })
1895 }
1896 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
1897 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1898 // Internal Server Error could be anything, retry up to 3 times.
1899 max_attempts: 3,
1900 }),
1901 status => {
1902 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
1903 // but we frequently get them in practice. See https://http.dev/529
1904 if status.as_u16() == 529 {
1905 Some(RetryStrategy::Fixed {
1906 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1907 max_attempts: MAX_RETRY_ATTEMPTS,
1908 })
1909 } else {
1910 Some(RetryStrategy::Fixed {
1911 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1912 max_attempts: 2,
1913 })
1914 }
1915 }
1916 },
1917 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
1918 delay: BASE_RETRY_DELAY,
1919 max_attempts: 3,
1920 }),
1921 ApiReadResponseError { .. }
1922 | HttpSend { .. }
1923 | DeserializeResponse { .. }
1924 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
1925 delay: BASE_RETRY_DELAY,
1926 max_attempts: 3,
1927 }),
1928 // Retrying these errors definitely shouldn't help.
1929 HttpResponseError {
1930 status_code:
1931 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
1932 ..
1933 }
1934 | AuthenticationError { .. }
1935 | PermissionError { .. }
1936 | NoApiKey { .. }
1937 | ApiEndpointNotFound { .. }
1938 | PromptTooLarge { .. } => None,
1939 // These errors might be transient, so retry them
1940 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
1941 delay: BASE_RETRY_DELAY,
1942 max_attempts: 1,
1943 }),
1944 // Retry all other 4xx and 5xx errors once.
1945 HttpResponseError { status_code, .. }
1946 if status_code.is_client_error() || status_code.is_server_error() =>
1947 {
1948 Some(RetryStrategy::Fixed {
1949 delay: BASE_RETRY_DELAY,
1950 max_attempts: 3,
1951 })
1952 }
1953 Other(err)
1954 if err.is::<language_model::PaymentRequiredError>()
1955 || err.is::<language_model::ModelRequestLimitReachedError>() =>
1956 {
1957 // Retrying won't help for Payment Required or Model Request Limit errors (where
1958 // the user must upgrade to usage-based billing to get more requests, or else wait
1959 // for a significant amount of time for the request limit to reset).
1960 None
1961 }
1962 // Conservatively assume that any other errors are non-retryable
1963 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
1964 delay: BASE_RETRY_DELAY,
1965 max_attempts: 2,
1966 }),
1967 }
1968 }
1969}
1970
1971struct RunningTurn {
1972 /// Holds the task that handles agent interaction until the end of the turn.
1973 /// Survives across multiple requests as the model performs tool calls and
1974 /// we run tools, report their results.
1975 _task: Task<()>,
1976 /// The current event stream for the running turn. Used to report a final
1977 /// cancellation event if we cancel the turn.
1978 event_stream: ThreadEventStream,
1979}
1980
1981impl RunningTurn {
1982 fn cancel(self) {
1983 log::debug!("Cancelling in progress turn");
1984 self.event_stream.send_canceled();
1985 }
1986}
1987
1988pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
1989
1990impl EventEmitter<TokenUsageUpdated> for Thread {}
1991
1992pub trait AgentTool
1993where
1994 Self: 'static + Sized,
1995{
1996 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
1997 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
1998
1999 fn name(&self) -> SharedString;
2000
2001 fn description(&self) -> SharedString {
2002 let schema = schemars::schema_for!(Self::Input);
2003 SharedString::new(
2004 schema
2005 .get("description")
2006 .and_then(|description| description.as_str())
2007 .unwrap_or_default(),
2008 )
2009 }
2010
2011 fn kind(&self) -> acp::ToolKind;
2012
2013 /// The initial tool title to display. Can be updated during the tool run.
2014 fn initial_title(&self, input: Result<Self::Input, serde_json::Value>) -> SharedString;
2015
2016 /// Returns the JSON schema that describes the tool's input.
2017 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Schema {
2018 crate::tool_schema::root_schema_for::<Self::Input>(format)
2019 }
2020
2021 /// Some tools rely on a provider for the underlying billing or other reasons.
2022 /// Allow the tool to check if they are compatible, or should be filtered out.
2023 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2024 true
2025 }
2026
2027 /// Runs the tool with the provided input.
2028 fn run(
2029 self: Arc<Self>,
2030 input: Self::Input,
2031 event_stream: ToolCallEventStream,
2032 cx: &mut App,
2033 ) -> Task<Result<Self::Output>>;
2034
2035 /// Emits events for a previous execution of the tool.
2036 fn replay(
2037 &self,
2038 _input: Self::Input,
2039 _output: Self::Output,
2040 _event_stream: ToolCallEventStream,
2041 _cx: &mut App,
2042 ) -> Result<()> {
2043 Ok(())
2044 }
2045
2046 fn erase(self) -> Arc<dyn AnyAgentTool> {
2047 Arc::new(Erased(Arc::new(self)))
2048 }
2049}
2050
2051pub struct Erased<T>(T);
2052
2053pub struct AgentToolOutput {
2054 pub llm_output: LanguageModelToolResultContent,
2055 pub raw_output: serde_json::Value,
2056}
2057
2058pub trait AnyAgentTool {
2059 fn name(&self) -> SharedString;
2060 fn description(&self) -> SharedString;
2061 fn kind(&self) -> acp::ToolKind;
2062 fn initial_title(&self, input: serde_json::Value) -> SharedString;
2063 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2064 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2065 true
2066 }
2067 fn run(
2068 self: Arc<Self>,
2069 input: serde_json::Value,
2070 event_stream: ToolCallEventStream,
2071 cx: &mut App,
2072 ) -> Task<Result<AgentToolOutput>>;
2073 fn replay(
2074 &self,
2075 input: serde_json::Value,
2076 output: serde_json::Value,
2077 event_stream: ToolCallEventStream,
2078 cx: &mut App,
2079 ) -> Result<()>;
2080}
2081
2082impl<T> AnyAgentTool for Erased<Arc<T>>
2083where
2084 T: AgentTool,
2085{
2086 fn name(&self) -> SharedString {
2087 self.0.name()
2088 }
2089
2090 fn description(&self) -> SharedString {
2091 self.0.description()
2092 }
2093
2094 fn kind(&self) -> agent_client_protocol::ToolKind {
2095 self.0.kind()
2096 }
2097
2098 fn initial_title(&self, input: serde_json::Value) -> SharedString {
2099 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2100 self.0.initial_title(parsed_input)
2101 }
2102
2103 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2104 let mut json = serde_json::to_value(self.0.input_schema(format))?;
2105 adapt_schema_to_format(&mut json, format)?;
2106 Ok(json)
2107 }
2108
2109 fn supported_provider(&self, provider: &LanguageModelProviderId) -> bool {
2110 self.0.supported_provider(provider)
2111 }
2112
2113 fn run(
2114 self: Arc<Self>,
2115 input: serde_json::Value,
2116 event_stream: ToolCallEventStream,
2117 cx: &mut App,
2118 ) -> Task<Result<AgentToolOutput>> {
2119 cx.spawn(async move |cx| {
2120 let input = serde_json::from_value(input)?;
2121 let output = cx
2122 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2123 .await?;
2124 let raw_output = serde_json::to_value(&output)?;
2125 Ok(AgentToolOutput {
2126 llm_output: output.into(),
2127 raw_output,
2128 })
2129 })
2130 }
2131
2132 fn replay(
2133 &self,
2134 input: serde_json::Value,
2135 output: serde_json::Value,
2136 event_stream: ToolCallEventStream,
2137 cx: &mut App,
2138 ) -> Result<()> {
2139 let input = serde_json::from_value(input)?;
2140 let output = serde_json::from_value(output)?;
2141 self.0.replay(input, output, event_stream, cx)
2142 }
2143}
2144
2145#[derive(Clone)]
2146struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2147
2148impl ThreadEventStream {
2149 fn send_title_update(&self, text: SharedString) {
2150 self.0
2151 .unbounded_send(Ok(ThreadEvent::TitleUpdate(text)))
2152 .ok();
2153 }
2154
2155 fn send_user_message(&self, message: &UserMessage) {
2156 self.0
2157 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2158 .ok();
2159 }
2160
2161 fn send_text(&self, text: &str) {
2162 self.0
2163 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2164 .ok();
2165 }
2166
2167 fn send_thinking(&self, text: &str) {
2168 self.0
2169 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2170 .ok();
2171 }
2172
2173 fn send_tool_call(
2174 &self,
2175 id: &LanguageModelToolUseId,
2176 title: SharedString,
2177 kind: acp::ToolKind,
2178 input: serde_json::Value,
2179 ) {
2180 self.0
2181 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2182 id,
2183 title.to_string(),
2184 kind,
2185 input,
2186 ))))
2187 .ok();
2188 }
2189
2190 fn initial_tool_call(
2191 id: &LanguageModelToolUseId,
2192 title: String,
2193 kind: acp::ToolKind,
2194 input: serde_json::Value,
2195 ) -> acp::ToolCall {
2196 acp::ToolCall {
2197 id: acp::ToolCallId(id.to_string().into()),
2198 title,
2199 kind,
2200 status: acp::ToolCallStatus::Pending,
2201 content: vec![],
2202 locations: vec![],
2203 raw_input: Some(input),
2204 raw_output: None,
2205 }
2206 }
2207
2208 fn update_tool_call_fields(
2209 &self,
2210 tool_use_id: &LanguageModelToolUseId,
2211 fields: acp::ToolCallUpdateFields,
2212 ) {
2213 self.0
2214 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2215 acp::ToolCallUpdate {
2216 id: acp::ToolCallId(tool_use_id.to_string().into()),
2217 fields,
2218 }
2219 .into(),
2220 )))
2221 .ok();
2222 }
2223
2224 fn send_retry(&self, status: acp_thread::RetryStatus) {
2225 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2226 }
2227
2228 fn send_stop(&self, reason: StopReason) {
2229 match reason {
2230 StopReason::EndTurn => {
2231 self.0
2232 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::EndTurn)))
2233 .ok();
2234 }
2235 StopReason::MaxTokens => {
2236 self.0
2237 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::MaxTokens)))
2238 .ok();
2239 }
2240 StopReason::Refusal => {
2241 self.0
2242 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Refusal)))
2243 .ok();
2244 }
2245 StopReason::ToolUse => {}
2246 }
2247 }
2248
2249 fn send_canceled(&self) {
2250 self.0
2251 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Canceled)))
2252 .ok();
2253 }
2254
2255 fn send_error(&self, error: impl Into<anyhow::Error>) {
2256 self.0.unbounded_send(Err(error.into())).ok();
2257 }
2258}
2259
2260#[derive(Clone)]
2261pub struct ToolCallEventStream {
2262 tool_use_id: LanguageModelToolUseId,
2263 stream: ThreadEventStream,
2264 fs: Option<Arc<dyn Fs>>,
2265}
2266
2267impl ToolCallEventStream {
2268 #[cfg(test)]
2269 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2270 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2271
2272 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2273
2274 (stream, ToolCallEventStreamReceiver(events_rx))
2275 }
2276
2277 fn new(
2278 tool_use_id: LanguageModelToolUseId,
2279 stream: ThreadEventStream,
2280 fs: Option<Arc<dyn Fs>>,
2281 ) -> Self {
2282 Self {
2283 tool_use_id,
2284 stream,
2285 fs,
2286 }
2287 }
2288
2289 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2290 self.stream
2291 .update_tool_call_fields(&self.tool_use_id, fields);
2292 }
2293
2294 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2295 self.stream
2296 .0
2297 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2298 acp_thread::ToolCallUpdateDiff {
2299 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2300 diff,
2301 }
2302 .into(),
2303 )))
2304 .ok();
2305 }
2306
2307 pub fn update_terminal(&self, terminal: Entity<acp_thread::Terminal>) {
2308 self.stream
2309 .0
2310 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2311 acp_thread::ToolCallUpdateTerminal {
2312 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2313 terminal,
2314 }
2315 .into(),
2316 )))
2317 .ok();
2318 }
2319
2320 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2321 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2322 return Task::ready(Ok(()));
2323 }
2324
2325 let (response_tx, response_rx) = oneshot::channel();
2326 self.stream
2327 .0
2328 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2329 ToolCallAuthorization {
2330 tool_call: acp::ToolCallUpdate {
2331 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2332 fields: acp::ToolCallUpdateFields {
2333 title: Some(title.into()),
2334 ..Default::default()
2335 },
2336 },
2337 options: vec![
2338 acp::PermissionOption {
2339 id: acp::PermissionOptionId("always_allow".into()),
2340 name: "Always Allow".into(),
2341 kind: acp::PermissionOptionKind::AllowAlways,
2342 },
2343 acp::PermissionOption {
2344 id: acp::PermissionOptionId("allow".into()),
2345 name: "Allow".into(),
2346 kind: acp::PermissionOptionKind::AllowOnce,
2347 },
2348 acp::PermissionOption {
2349 id: acp::PermissionOptionId("deny".into()),
2350 name: "Deny".into(),
2351 kind: acp::PermissionOptionKind::RejectOnce,
2352 },
2353 ],
2354 response: response_tx,
2355 },
2356 )))
2357 .ok();
2358 let fs = self.fs.clone();
2359 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2360 "always_allow" => {
2361 if let Some(fs) = fs.clone() {
2362 cx.update(|cx| {
2363 update_settings_file::<AgentSettings>(fs, cx, |settings, _| {
2364 settings.set_always_allow_tool_actions(true);
2365 });
2366 })?;
2367 }
2368
2369 Ok(())
2370 }
2371 "allow" => Ok(()),
2372 _ => Err(anyhow!("Permission to run tool denied by user")),
2373 })
2374 }
2375}
2376
2377#[cfg(test)]
2378pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2379
2380#[cfg(test)]
2381impl ToolCallEventStreamReceiver {
2382 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2383 let event = self.0.next().await;
2384 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2385 auth
2386 } else {
2387 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2388 }
2389 }
2390
2391 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2392 let event = self.0.next().await;
2393 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2394 update,
2395 )))) = event
2396 {
2397 update.terminal
2398 } else {
2399 panic!("Expected terminal but got: {:?}", event);
2400 }
2401 }
2402}
2403
2404#[cfg(test)]
2405impl std::ops::Deref for ToolCallEventStreamReceiver {
2406 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2407
2408 fn deref(&self) -> &Self::Target {
2409 &self.0
2410 }
2411}
2412
2413#[cfg(test)]
2414impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2415 fn deref_mut(&mut self) -> &mut Self::Target {
2416 &mut self.0
2417 }
2418}
2419
2420impl From<&str> for UserMessageContent {
2421 fn from(text: &str) -> Self {
2422 Self::Text(text.into())
2423 }
2424}
2425
2426impl From<acp::ContentBlock> for UserMessageContent {
2427 fn from(value: acp::ContentBlock) -> Self {
2428 match value {
2429 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2430 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2431 acp::ContentBlock::Audio(_) => {
2432 // TODO
2433 Self::Text("[audio]".to_string())
2434 }
2435 acp::ContentBlock::ResourceLink(resource_link) => {
2436 match MentionUri::parse(&resource_link.uri) {
2437 Ok(uri) => Self::Mention {
2438 uri,
2439 content: String::new(),
2440 },
2441 Err(err) => {
2442 log::error!("Failed to parse mention link: {}", err);
2443 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2444 }
2445 }
2446 }
2447 acp::ContentBlock::Resource(resource) => match resource.resource {
2448 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2449 match MentionUri::parse(&resource.uri) {
2450 Ok(uri) => Self::Mention {
2451 uri,
2452 content: resource.text,
2453 },
2454 Err(err) => {
2455 log::error!("Failed to parse mention link: {}", err);
2456 Self::Text(
2457 MarkdownCodeBlock {
2458 tag: &resource.uri,
2459 text: &resource.text,
2460 }
2461 .to_string(),
2462 )
2463 }
2464 }
2465 }
2466 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2467 // TODO
2468 Self::Text("[blob]".to_string())
2469 }
2470 },
2471 }
2472 }
2473}
2474
2475impl From<UserMessageContent> for acp::ContentBlock {
2476 fn from(content: UserMessageContent) -> Self {
2477 match content {
2478 UserMessageContent::Text(text) => acp::ContentBlock::Text(acp::TextContent {
2479 text,
2480 annotations: None,
2481 }),
2482 UserMessageContent::Image(image) => acp::ContentBlock::Image(acp::ImageContent {
2483 data: image.source.to_string(),
2484 mime_type: "image/png".to_string(),
2485 annotations: None,
2486 uri: None,
2487 }),
2488 UserMessageContent::Mention { uri, content } => {
2489 acp::ContentBlock::Resource(acp::EmbeddedResource {
2490 resource: acp::EmbeddedResourceResource::TextResourceContents(
2491 acp::TextResourceContents {
2492 mime_type: None,
2493 text: content,
2494 uri: uri.to_uri().to_string(),
2495 },
2496 ),
2497 annotations: None,
2498 })
2499 }
2500 }
2501 }
2502}
2503
2504fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2505 LanguageModelImage {
2506 source: image_content.data.into(),
2507 // TODO: make this optional?
2508 size: gpui::Size::new(0.into(), 0.into()),
2509 }
2510}