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 _ = write!(&mut markdown, "{}\n\n{}\n", uri.as_link(), content);
165 } else {
166 let _ = write!(&mut markdown, "{}\n", 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.clone().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.clone().unwrap_or_default(),
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 model(&self) -> Option<&Arc<dyn LanguageModel>> {
874 self.model.as_ref()
875 }
876
877 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
878 let old_usage = self.latest_token_usage();
879 self.model = Some(model);
880 let new_usage = self.latest_token_usage();
881 if old_usage != new_usage {
882 cx.emit(TokenUsageUpdated(new_usage));
883 }
884 cx.notify()
885 }
886
887 pub fn set_summarization_model(
888 &mut self,
889 model: Option<Arc<dyn LanguageModel>>,
890 cx: &mut Context<Self>,
891 ) {
892 self.summarization_model = model;
893 cx.notify()
894 }
895
896 pub fn completion_mode(&self) -> CompletionMode {
897 self.completion_mode
898 }
899
900 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
901 let old_usage = self.latest_token_usage();
902 self.completion_mode = mode;
903 let new_usage = self.latest_token_usage();
904 if old_usage != new_usage {
905 cx.emit(TokenUsageUpdated(new_usage));
906 }
907 cx.notify()
908 }
909
910 #[cfg(any(test, feature = "test-support"))]
911 pub fn last_message(&self) -> Option<Message> {
912 if let Some(message) = self.pending_message.clone() {
913 Some(Message::Agent(message))
914 } else {
915 self.messages.last().cloned()
916 }
917 }
918
919 pub fn add_default_tools(&mut self, cx: &mut Context<Self>) {
920 let language_registry = self.project.read(cx).languages().clone();
921 self.add_tool(CopyPathTool::new(self.project.clone()));
922 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
923 self.add_tool(DeletePathTool::new(
924 self.project.clone(),
925 self.action_log.clone(),
926 ));
927 self.add_tool(DiagnosticsTool::new(self.project.clone()));
928 self.add_tool(EditFileTool::new(cx.weak_entity(), language_registry));
929 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
930 self.add_tool(FindPathTool::new(self.project.clone()));
931 self.add_tool(GrepTool::new(self.project.clone()));
932 self.add_tool(ListDirectoryTool::new(self.project.clone()));
933 self.add_tool(MovePathTool::new(self.project.clone()));
934 self.add_tool(NowTool);
935 self.add_tool(OpenTool::new(self.project.clone()));
936 self.add_tool(ReadFileTool::new(
937 self.project.clone(),
938 self.action_log.clone(),
939 ));
940 self.add_tool(TerminalTool::new(self.project.clone(), cx));
941 self.add_tool(ThinkingTool);
942 self.add_tool(WebSearchTool); // TODO: Enable this only if it's a zed model.
943 }
944
945 pub fn add_tool(&mut self, tool: impl AgentTool) {
946 self.tools.insert(tool.name(), tool.erase());
947 }
948
949 pub fn remove_tool(&mut self, name: &str) -> bool {
950 self.tools.remove(name).is_some()
951 }
952
953 pub fn profile(&self) -> &AgentProfileId {
954 &self.profile_id
955 }
956
957 pub fn set_profile(&mut self, profile_id: AgentProfileId) {
958 self.profile_id = profile_id;
959 }
960
961 pub fn cancel(&mut self, cx: &mut Context<Self>) {
962 if let Some(running_turn) = self.running_turn.take() {
963 running_turn.cancel();
964 }
965 self.flush_pending_message(cx);
966 }
967
968 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
969 let Some(last_user_message) = self.last_user_message() else {
970 return;
971 };
972
973 self.request_token_usage
974 .insert(last_user_message.id.clone(), update);
975 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
976 cx.notify();
977 }
978
979 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
980 self.cancel(cx);
981 let Some(position) = self.messages.iter().position(
982 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
983 ) else {
984 return Err(anyhow!("Message not found"));
985 };
986
987 for message in self.messages.drain(position..) {
988 match message {
989 Message::User(message) => {
990 self.request_token_usage.remove(&message.id);
991 }
992 Message::Agent(_) | Message::Resume => {}
993 }
994 }
995 self.summary = None;
996 cx.notify();
997 Ok(())
998 }
999
1000 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1001 let last_user_message = self.last_user_message()?;
1002 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1003 let model = self.model.clone()?;
1004
1005 Some(acp_thread::TokenUsage {
1006 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1007 used_tokens: tokens.total_tokens(),
1008 })
1009 }
1010
1011 pub fn resume(
1012 &mut self,
1013 cx: &mut Context<Self>,
1014 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1015 anyhow::ensure!(
1016 self.tool_use_limit_reached,
1017 "can only resume after tool use limit is reached"
1018 );
1019
1020 self.messages.push(Message::Resume);
1021 cx.notify();
1022
1023 log::info!("Total messages in thread: {}", self.messages.len());
1024 self.run_turn(cx)
1025 }
1026
1027 /// Sending a message results in the model streaming a response, which could include tool calls.
1028 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1029 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1030 pub fn send<T>(
1031 &mut self,
1032 id: UserMessageId,
1033 content: impl IntoIterator<Item = T>,
1034 cx: &mut Context<Self>,
1035 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1036 where
1037 T: Into<UserMessageContent>,
1038 {
1039 let model = self.model().context("No language model configured")?;
1040
1041 log::info!("Thread::send called with model: {:?}", model.name());
1042 self.advance_prompt_id();
1043
1044 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1045 log::debug!("Thread::send content: {:?}", content);
1046
1047 self.messages
1048 .push(Message::User(UserMessage { id, content }));
1049 cx.notify();
1050
1051 log::info!("Total messages in thread: {}", self.messages.len());
1052 self.run_turn(cx)
1053 }
1054
1055 fn run_turn(
1056 &mut self,
1057 cx: &mut Context<Self>,
1058 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1059 self.cancel(cx);
1060
1061 let model = self.model.clone().context("No language model configured")?;
1062 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1063 let event_stream = ThreadEventStream(events_tx);
1064 let message_ix = self.messages.len().saturating_sub(1);
1065 self.tool_use_limit_reached = false;
1066 self.summary = None;
1067 self.running_turn = Some(RunningTurn {
1068 event_stream: event_stream.clone(),
1069 _task: cx.spawn(async move |this, cx| {
1070 log::info!("Starting agent turn execution");
1071 let turn_result: Result<StopReason> = async {
1072 let mut completion_intent = CompletionIntent::UserPrompt;
1073 loop {
1074 log::debug!(
1075 "Building completion request with intent: {:?}",
1076 completion_intent
1077 );
1078 let request = this.update(cx, |this, cx| {
1079 this.build_completion_request(completion_intent, cx)
1080 })??;
1081
1082 log::info!("Calling model.stream_completion");
1083
1084 let mut tool_use_limit_reached = false;
1085 let mut refused = false;
1086 let mut reached_max_tokens = false;
1087 let mut tool_uses = Self::stream_completion_with_retries(
1088 this.clone(),
1089 model.clone(),
1090 request,
1091 &event_stream,
1092 &mut tool_use_limit_reached,
1093 &mut refused,
1094 &mut reached_max_tokens,
1095 cx,
1096 )
1097 .await?;
1098
1099 if refused {
1100 return Ok(StopReason::Refusal);
1101 } else if reached_max_tokens {
1102 return Ok(StopReason::MaxTokens);
1103 }
1104
1105 let end_turn = tool_uses.is_empty();
1106 while let Some(tool_result) = tool_uses.next().await {
1107 log::info!("Tool finished {:?}", tool_result);
1108
1109 event_stream.update_tool_call_fields(
1110 &tool_result.tool_use_id,
1111 acp::ToolCallUpdateFields {
1112 status: Some(if tool_result.is_error {
1113 acp::ToolCallStatus::Failed
1114 } else {
1115 acp::ToolCallStatus::Completed
1116 }),
1117 raw_output: tool_result.output.clone(),
1118 ..Default::default()
1119 },
1120 );
1121 this.update(cx, |this, _cx| {
1122 this.pending_message()
1123 .tool_results
1124 .insert(tool_result.tool_use_id.clone(), tool_result);
1125 })
1126 .ok();
1127 }
1128
1129 if tool_use_limit_reached {
1130 log::info!("Tool use limit reached, completing turn");
1131 this.update(cx, |this, _cx| this.tool_use_limit_reached = true)?;
1132 return Err(language_model::ToolUseLimitReachedError.into());
1133 } else if end_turn {
1134 log::info!("No tool uses found, completing turn");
1135 return Ok(StopReason::EndTurn);
1136 } else {
1137 this.update(cx, |this, cx| this.flush_pending_message(cx))?;
1138 completion_intent = CompletionIntent::ToolResults;
1139 }
1140 }
1141 }
1142 .await;
1143 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1144
1145 match turn_result {
1146 Ok(reason) => {
1147 log::info!("Turn execution completed: {:?}", reason);
1148
1149 let update_title = this
1150 .update(cx, |this, cx| this.update_title(&event_stream, cx))
1151 .ok()
1152 .flatten();
1153 if let Some(update_title) = update_title {
1154 update_title.await.context("update title failed").log_err();
1155 }
1156
1157 event_stream.send_stop(reason);
1158 if reason == StopReason::Refusal {
1159 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1160 }
1161 }
1162 Err(error) => {
1163 log::error!("Turn execution failed: {:?}", error);
1164 event_stream.send_error(error);
1165 }
1166 }
1167
1168 _ = this.update(cx, |this, _| this.running_turn.take());
1169 }),
1170 });
1171 Ok(events_rx)
1172 }
1173
1174 async fn stream_completion_with_retries(
1175 this: WeakEntity<Self>,
1176 model: Arc<dyn LanguageModel>,
1177 request: LanguageModelRequest,
1178 event_stream: &ThreadEventStream,
1179 tool_use_limit_reached: &mut bool,
1180 refusal: &mut bool,
1181 max_tokens_reached: &mut bool,
1182 cx: &mut AsyncApp,
1183 ) -> Result<FuturesUnordered<Task<LanguageModelToolResult>>> {
1184 log::debug!("Stream completion started successfully");
1185
1186 let mut attempt = None;
1187 'retry: loop {
1188 let mut events = model.stream_completion(request.clone(), cx).await?;
1189 let mut tool_uses = FuturesUnordered::new();
1190 while let Some(event) = events.next().await {
1191 match event {
1192 Ok(LanguageModelCompletionEvent::StatusUpdate(
1193 CompletionRequestStatus::ToolUseLimitReached,
1194 )) => {
1195 *tool_use_limit_reached = true;
1196 }
1197 Ok(LanguageModelCompletionEvent::StatusUpdate(
1198 CompletionRequestStatus::UsageUpdated { amount, limit },
1199 )) => {
1200 this.update(cx, |this, cx| {
1201 this.update_model_request_usage(amount, limit, cx)
1202 })?;
1203 }
1204 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {
1205 this.update(cx, |this, cx| this.update_token_usage(token_usage, cx))?;
1206 }
1207 Ok(LanguageModelCompletionEvent::Stop(StopReason::Refusal)) => {
1208 *refusal = true;
1209 return Ok(FuturesUnordered::default());
1210 }
1211 Ok(LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)) => {
1212 *max_tokens_reached = true;
1213 return Ok(FuturesUnordered::default());
1214 }
1215 Ok(LanguageModelCompletionEvent::Stop(
1216 StopReason::ToolUse | StopReason::EndTurn,
1217 )) => break,
1218 Ok(event) => {
1219 log::trace!("Received completion event: {:?}", event);
1220 this.update(cx, |this, cx| {
1221 tool_uses.extend(this.handle_streamed_completion_event(
1222 event,
1223 event_stream,
1224 cx,
1225 ));
1226 })?;
1227 }
1228 Err(error) => {
1229 let completion_mode =
1230 this.read_with(cx, |thread, _cx| thread.completion_mode())?;
1231 if completion_mode == CompletionMode::Normal {
1232 return Err(error.into());
1233 }
1234
1235 let Some(strategy) = Self::retry_strategy_for(&error) else {
1236 return Err(error.into());
1237 };
1238
1239 let max_attempts = match &strategy {
1240 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1241 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1242 };
1243
1244 let attempt = attempt.get_or_insert(0u8);
1245
1246 *attempt += 1;
1247
1248 let attempt = *attempt;
1249 if attempt > max_attempts {
1250 return Err(error.into());
1251 }
1252
1253 let delay = match &strategy {
1254 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1255 let delay_secs =
1256 initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1257 Duration::from_secs(delay_secs)
1258 }
1259 RetryStrategy::Fixed { delay, .. } => *delay,
1260 };
1261 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1262
1263 event_stream.send_retry(acp_thread::RetryStatus {
1264 last_error: error.to_string().into(),
1265 attempt: attempt as usize,
1266 max_attempts: max_attempts as usize,
1267 started_at: Instant::now(),
1268 duration: delay,
1269 });
1270
1271 cx.background_executor().timer(delay).await;
1272 continue 'retry;
1273 }
1274 }
1275 }
1276
1277 return Ok(tool_uses);
1278 }
1279 }
1280
1281 pub fn build_system_message(&self, cx: &App) -> LanguageModelRequestMessage {
1282 log::debug!("Building system message");
1283 let prompt = SystemPromptTemplate {
1284 project: self.project_context.read(cx),
1285 available_tools: self.tools.keys().cloned().collect(),
1286 }
1287 .render(&self.templates)
1288 .context("failed to build system prompt")
1289 .expect("Invalid template");
1290 log::debug!("System message built");
1291 LanguageModelRequestMessage {
1292 role: Role::System,
1293 content: vec![prompt.into()],
1294 cache: true,
1295 }
1296 }
1297
1298 /// A helper method that's called on every streamed completion event.
1299 /// Returns an optional tool result task, which the main agentic loop in
1300 /// send will send back to the model when it resolves.
1301 fn handle_streamed_completion_event(
1302 &mut self,
1303 event: LanguageModelCompletionEvent,
1304 event_stream: &ThreadEventStream,
1305 cx: &mut Context<Self>,
1306 ) -> Option<Task<LanguageModelToolResult>> {
1307 log::trace!("Handling streamed completion event: {:?}", event);
1308 use LanguageModelCompletionEvent::*;
1309
1310 match event {
1311 StartMessage { .. } => {
1312 self.flush_pending_message(cx);
1313 self.pending_message = Some(AgentMessage::default());
1314 }
1315 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1316 Thinking { text, signature } => {
1317 self.handle_thinking_event(text, signature, event_stream, cx)
1318 }
1319 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1320 ToolUse(tool_use) => {
1321 return self.handle_tool_use_event(tool_use, event_stream, cx);
1322 }
1323 ToolUseJsonParseError {
1324 id,
1325 tool_name,
1326 raw_input,
1327 json_parse_error,
1328 } => {
1329 return Some(Task::ready(self.handle_tool_use_json_parse_error_event(
1330 id,
1331 tool_name,
1332 raw_input,
1333 json_parse_error,
1334 )));
1335 }
1336 StatusUpdate(_) => {}
1337 UsageUpdate(_) | Stop(_) => unreachable!(),
1338 }
1339
1340 None
1341 }
1342
1343 fn handle_text_event(
1344 &mut self,
1345 new_text: String,
1346 event_stream: &ThreadEventStream,
1347 cx: &mut Context<Self>,
1348 ) {
1349 event_stream.send_text(&new_text);
1350
1351 let last_message = self.pending_message();
1352 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1353 text.push_str(&new_text);
1354 } else {
1355 last_message
1356 .content
1357 .push(AgentMessageContent::Text(new_text));
1358 }
1359
1360 cx.notify();
1361 }
1362
1363 fn handle_thinking_event(
1364 &mut self,
1365 new_text: String,
1366 new_signature: Option<String>,
1367 event_stream: &ThreadEventStream,
1368 cx: &mut Context<Self>,
1369 ) {
1370 event_stream.send_thinking(&new_text);
1371
1372 let last_message = self.pending_message();
1373 if let Some(AgentMessageContent::Thinking { text, signature }) =
1374 last_message.content.last_mut()
1375 {
1376 text.push_str(&new_text);
1377 *signature = new_signature.or(signature.take());
1378 } else {
1379 last_message.content.push(AgentMessageContent::Thinking {
1380 text: new_text,
1381 signature: new_signature,
1382 });
1383 }
1384
1385 cx.notify();
1386 }
1387
1388 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1389 let last_message = self.pending_message();
1390 last_message
1391 .content
1392 .push(AgentMessageContent::RedactedThinking(data));
1393 cx.notify();
1394 }
1395
1396 fn handle_tool_use_event(
1397 &mut self,
1398 tool_use: LanguageModelToolUse,
1399 event_stream: &ThreadEventStream,
1400 cx: &mut Context<Self>,
1401 ) -> Option<Task<LanguageModelToolResult>> {
1402 cx.notify();
1403
1404 let tool = self.tools.get(tool_use.name.as_ref()).cloned();
1405 let mut title = SharedString::from(&tool_use.name);
1406 let mut kind = acp::ToolKind::Other;
1407 if let Some(tool) = tool.as_ref() {
1408 title = tool.initial_title(tool_use.input.clone());
1409 kind = tool.kind();
1410 }
1411
1412 // Ensure the last message ends in the current tool use
1413 let last_message = self.pending_message();
1414 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1415 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1416 if last_tool_use.id == tool_use.id {
1417 *last_tool_use = tool_use.clone();
1418 false
1419 } else {
1420 true
1421 }
1422 } else {
1423 true
1424 }
1425 });
1426
1427 if push_new_tool_use {
1428 event_stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
1429 last_message
1430 .content
1431 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1432 } else {
1433 event_stream.update_tool_call_fields(
1434 &tool_use.id,
1435 acp::ToolCallUpdateFields {
1436 title: Some(title.into()),
1437 kind: Some(kind),
1438 raw_input: Some(tool_use.input.clone()),
1439 ..Default::default()
1440 },
1441 );
1442 }
1443
1444 if !tool_use.is_input_complete {
1445 return None;
1446 }
1447
1448 let Some(tool) = tool else {
1449 let content = format!("No tool named {} exists", tool_use.name);
1450 return Some(Task::ready(LanguageModelToolResult {
1451 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1452 tool_use_id: tool_use.id,
1453 tool_name: tool_use.name,
1454 is_error: true,
1455 output: None,
1456 }));
1457 };
1458
1459 let fs = self.project.read(cx).fs().clone();
1460 let tool_event_stream =
1461 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1462 tool_event_stream.update_fields(acp::ToolCallUpdateFields {
1463 status: Some(acp::ToolCallStatus::InProgress),
1464 ..Default::default()
1465 });
1466 let supports_images = self.model().is_some_and(|model| model.supports_images());
1467 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1468 log::info!("Running tool {}", tool_use.name);
1469 Some(cx.foreground_executor().spawn(async move {
1470 let tool_result = tool_result.await.and_then(|output| {
1471 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1472 && !supports_images
1473 {
1474 return Err(anyhow!(
1475 "Attempted to read an image, but this model doesn't support it.",
1476 ));
1477 }
1478 Ok(output)
1479 });
1480
1481 match tool_result {
1482 Ok(output) => LanguageModelToolResult {
1483 tool_use_id: tool_use.id,
1484 tool_name: tool_use.name,
1485 is_error: false,
1486 content: output.llm_output,
1487 output: Some(output.raw_output),
1488 },
1489 Err(error) => LanguageModelToolResult {
1490 tool_use_id: tool_use.id,
1491 tool_name: tool_use.name,
1492 is_error: true,
1493 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1494 output: None,
1495 },
1496 }
1497 }))
1498 }
1499
1500 fn handle_tool_use_json_parse_error_event(
1501 &mut self,
1502 tool_use_id: LanguageModelToolUseId,
1503 tool_name: Arc<str>,
1504 raw_input: Arc<str>,
1505 json_parse_error: String,
1506 ) -> LanguageModelToolResult {
1507 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1508 LanguageModelToolResult {
1509 tool_use_id,
1510 tool_name,
1511 is_error: true,
1512 content: LanguageModelToolResultContent::Text(tool_output.into()),
1513 output: Some(serde_json::Value::String(raw_input.to_string())),
1514 }
1515 }
1516
1517 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1518 self.project
1519 .read(cx)
1520 .user_store()
1521 .update(cx, |user_store, cx| {
1522 user_store.update_model_request_usage(
1523 ModelRequestUsage(RequestUsage {
1524 amount: amount as i32,
1525 limit,
1526 }),
1527 cx,
1528 )
1529 });
1530 }
1531
1532 pub fn title(&self) -> SharedString {
1533 self.title.clone().unwrap_or("New Thread".into())
1534 }
1535
1536 pub fn summary(&mut self, cx: &mut Context<Self>) -> Task<Result<SharedString>> {
1537 if let Some(summary) = self.summary.as_ref() {
1538 return Task::ready(Ok(summary.clone()));
1539 }
1540 let Some(model) = self.summarization_model.clone() else {
1541 return Task::ready(Err(anyhow!("No summarization model available")));
1542 };
1543 let mut request = LanguageModelRequest {
1544 intent: Some(CompletionIntent::ThreadContextSummarization),
1545 temperature: AgentSettings::temperature_for_model(&model, cx),
1546 ..Default::default()
1547 };
1548
1549 for message in &self.messages {
1550 request.messages.extend(message.to_request());
1551 }
1552
1553 request.messages.push(LanguageModelRequestMessage {
1554 role: Role::User,
1555 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1556 cache: false,
1557 });
1558 cx.spawn(async move |this, cx| {
1559 let mut summary = String::new();
1560 let mut messages = model.stream_completion(request, cx).await?;
1561 while let Some(event) = messages.next().await {
1562 let event = event?;
1563 let text = match event {
1564 LanguageModelCompletionEvent::Text(text) => text,
1565 LanguageModelCompletionEvent::StatusUpdate(
1566 CompletionRequestStatus::UsageUpdated { amount, limit },
1567 ) => {
1568 this.update(cx, |thread, cx| {
1569 thread.update_model_request_usage(amount, limit, cx);
1570 })?;
1571 continue;
1572 }
1573 _ => continue,
1574 };
1575
1576 let mut lines = text.lines();
1577 summary.extend(lines.next());
1578 }
1579
1580 log::info!("Setting summary: {}", summary);
1581 let summary = SharedString::from(summary);
1582
1583 this.update(cx, |this, cx| {
1584 this.summary = Some(summary.clone());
1585 cx.notify()
1586 })?;
1587
1588 Ok(summary)
1589 })
1590 }
1591
1592 fn update_title(
1593 &mut self,
1594 event_stream: &ThreadEventStream,
1595 cx: &mut Context<Self>,
1596 ) -> Option<Task<Result<()>>> {
1597 if self.title.is_some() {
1598 log::debug!("Skipping title generation because we already have one.");
1599 return None;
1600 }
1601
1602 log::info!(
1603 "Generating title with model: {:?}",
1604 self.summarization_model.as_ref().map(|model| model.name())
1605 );
1606 let model = self.summarization_model.clone()?;
1607 let event_stream = event_stream.clone();
1608 let mut request = LanguageModelRequest {
1609 intent: Some(CompletionIntent::ThreadSummarization),
1610 temperature: AgentSettings::temperature_for_model(&model, cx),
1611 ..Default::default()
1612 };
1613
1614 for message in &self.messages {
1615 request.messages.extend(message.to_request());
1616 }
1617
1618 request.messages.push(LanguageModelRequestMessage {
1619 role: Role::User,
1620 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1621 cache: false,
1622 });
1623 Some(cx.spawn(async move |this, cx| {
1624 let mut title = String::new();
1625 let mut messages = model.stream_completion(request, cx).await?;
1626 while let Some(event) = messages.next().await {
1627 let event = event?;
1628 let text = match event {
1629 LanguageModelCompletionEvent::Text(text) => text,
1630 LanguageModelCompletionEvent::StatusUpdate(
1631 CompletionRequestStatus::UsageUpdated { amount, limit },
1632 ) => {
1633 this.update(cx, |thread, cx| {
1634 thread.update_model_request_usage(amount, limit, cx);
1635 })?;
1636 continue;
1637 }
1638 _ => continue,
1639 };
1640
1641 let mut lines = text.lines();
1642 title.extend(lines.next());
1643
1644 // Stop if the LLM generated multiple lines.
1645 if lines.next().is_some() {
1646 break;
1647 }
1648 }
1649
1650 log::info!("Setting title: {}", title);
1651
1652 this.update(cx, |this, cx| {
1653 let title = SharedString::from(title);
1654 event_stream.send_title_update(title.clone());
1655 this.title = Some(title);
1656 cx.notify();
1657 })
1658 }))
1659 }
1660
1661 fn last_user_message(&self) -> Option<&UserMessage> {
1662 self.messages
1663 .iter()
1664 .rev()
1665 .find_map(|message| match message {
1666 Message::User(user_message) => Some(user_message),
1667 Message::Agent(_) => None,
1668 Message::Resume => None,
1669 })
1670 }
1671
1672 fn pending_message(&mut self) -> &mut AgentMessage {
1673 self.pending_message.get_or_insert_default()
1674 }
1675
1676 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1677 let Some(mut message) = self.pending_message.take() else {
1678 return;
1679 };
1680
1681 for content in &message.content {
1682 let AgentMessageContent::ToolUse(tool_use) = content else {
1683 continue;
1684 };
1685
1686 if !message.tool_results.contains_key(&tool_use.id) {
1687 message.tool_results.insert(
1688 tool_use.id.clone(),
1689 LanguageModelToolResult {
1690 tool_use_id: tool_use.id.clone(),
1691 tool_name: tool_use.name.clone(),
1692 is_error: true,
1693 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1694 output: None,
1695 },
1696 );
1697 }
1698 }
1699
1700 self.messages.push(Message::Agent(message));
1701 self.updated_at = Utc::now();
1702 self.summary = None;
1703 cx.notify()
1704 }
1705
1706 pub(crate) fn build_completion_request(
1707 &self,
1708 completion_intent: CompletionIntent,
1709 cx: &mut App,
1710 ) -> Result<LanguageModelRequest> {
1711 let model = self.model().context("No language model configured")?;
1712
1713 log::debug!("Building completion request");
1714 log::debug!("Completion intent: {:?}", completion_intent);
1715 log::debug!("Completion mode: {:?}", self.completion_mode);
1716
1717 let messages = self.build_request_messages(cx);
1718 log::info!("Request will include {} messages", messages.len());
1719
1720 let tools = if let Some(tools) = self.tools(cx).log_err() {
1721 tools
1722 .filter_map(|tool| {
1723 let tool_name = tool.name().to_string();
1724 log::trace!("Including tool: {}", tool_name);
1725 Some(LanguageModelRequestTool {
1726 name: tool_name,
1727 description: tool.description().to_string(),
1728 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1729 })
1730 })
1731 .collect()
1732 } else {
1733 Vec::new()
1734 };
1735
1736 log::info!("Request includes {} tools", tools.len());
1737
1738 let request = LanguageModelRequest {
1739 thread_id: Some(self.id.to_string()),
1740 prompt_id: Some(self.prompt_id.to_string()),
1741 intent: Some(completion_intent),
1742 mode: Some(self.completion_mode.into()),
1743 messages,
1744 tools,
1745 tool_choice: None,
1746 stop: Vec::new(),
1747 temperature: AgentSettings::temperature_for_model(model, cx),
1748 thinking_allowed: true,
1749 };
1750
1751 log::debug!("Completion request built successfully");
1752 Ok(request)
1753 }
1754
1755 fn tools<'a>(&'a self, cx: &'a App) -> Result<impl Iterator<Item = &'a Arc<dyn AnyAgentTool>>> {
1756 let model = self.model().context("No language model configured")?;
1757
1758 let profile = AgentSettings::get_global(cx)
1759 .profiles
1760 .get(&self.profile_id)
1761 .context("profile not found")?;
1762 let provider_id = model.provider_id();
1763
1764 Ok(self
1765 .tools
1766 .iter()
1767 .filter(move |(_, tool)| tool.supported_provider(&provider_id))
1768 .filter_map(|(tool_name, tool)| {
1769 if profile.is_tool_enabled(tool_name) {
1770 Some(tool)
1771 } else {
1772 None
1773 }
1774 })
1775 .chain(self.context_server_registry.read(cx).servers().flat_map(
1776 |(server_id, tools)| {
1777 tools.iter().filter_map(|(tool_name, tool)| {
1778 if profile.is_context_server_tool_enabled(&server_id.0, tool_name) {
1779 Some(tool)
1780 } else {
1781 None
1782 }
1783 })
1784 },
1785 )))
1786 }
1787
1788 fn build_request_messages(&self, cx: &App) -> Vec<LanguageModelRequestMessage> {
1789 log::trace!(
1790 "Building request messages from {} thread messages",
1791 self.messages.len()
1792 );
1793 let mut messages = vec![self.build_system_message(cx)];
1794 for message in &self.messages {
1795 messages.extend(message.to_request());
1796 }
1797
1798 if let Some(message) = self.pending_message.as_ref() {
1799 messages.extend(message.to_request());
1800 }
1801
1802 if let Some(last_user_message) = messages
1803 .iter_mut()
1804 .rev()
1805 .find(|message| message.role == Role::User)
1806 {
1807 last_user_message.cache = true;
1808 }
1809
1810 messages
1811 }
1812
1813 pub fn to_markdown(&self) -> String {
1814 let mut markdown = String::new();
1815 for (ix, message) in self.messages.iter().enumerate() {
1816 if ix > 0 {
1817 markdown.push('\n');
1818 }
1819 markdown.push_str(&message.to_markdown());
1820 }
1821
1822 if let Some(message) = self.pending_message.as_ref() {
1823 markdown.push('\n');
1824 markdown.push_str(&message.to_markdown());
1825 }
1826
1827 markdown
1828 }
1829
1830 fn advance_prompt_id(&mut self) {
1831 self.prompt_id = PromptId::new();
1832 }
1833
1834 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
1835 use LanguageModelCompletionError::*;
1836 use http_client::StatusCode;
1837
1838 // General strategy here:
1839 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
1840 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
1841 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
1842 match error {
1843 HttpResponseError {
1844 status_code: StatusCode::TOO_MANY_REQUESTS,
1845 ..
1846 } => Some(RetryStrategy::ExponentialBackoff {
1847 initial_delay: BASE_RETRY_DELAY,
1848 max_attempts: MAX_RETRY_ATTEMPTS,
1849 }),
1850 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
1851 Some(RetryStrategy::Fixed {
1852 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1853 max_attempts: MAX_RETRY_ATTEMPTS,
1854 })
1855 }
1856 UpstreamProviderError {
1857 status,
1858 retry_after,
1859 ..
1860 } => match *status {
1861 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
1862 Some(RetryStrategy::Fixed {
1863 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1864 max_attempts: MAX_RETRY_ATTEMPTS,
1865 })
1866 }
1867 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
1868 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1869 // Internal Server Error could be anything, retry up to 3 times.
1870 max_attempts: 3,
1871 }),
1872 status => {
1873 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
1874 // but we frequently get them in practice. See https://http.dev/529
1875 if status.as_u16() == 529 {
1876 Some(RetryStrategy::Fixed {
1877 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1878 max_attempts: MAX_RETRY_ATTEMPTS,
1879 })
1880 } else {
1881 Some(RetryStrategy::Fixed {
1882 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1883 max_attempts: 2,
1884 })
1885 }
1886 }
1887 },
1888 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
1889 delay: BASE_RETRY_DELAY,
1890 max_attempts: 3,
1891 }),
1892 ApiReadResponseError { .. }
1893 | HttpSend { .. }
1894 | DeserializeResponse { .. }
1895 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
1896 delay: BASE_RETRY_DELAY,
1897 max_attempts: 3,
1898 }),
1899 // Retrying these errors definitely shouldn't help.
1900 HttpResponseError {
1901 status_code:
1902 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
1903 ..
1904 }
1905 | AuthenticationError { .. }
1906 | PermissionError { .. }
1907 | NoApiKey { .. }
1908 | ApiEndpointNotFound { .. }
1909 | PromptTooLarge { .. } => None,
1910 // These errors might be transient, so retry them
1911 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
1912 delay: BASE_RETRY_DELAY,
1913 max_attempts: 1,
1914 }),
1915 // Retry all other 4xx and 5xx errors once.
1916 HttpResponseError { status_code, .. }
1917 if status_code.is_client_error() || status_code.is_server_error() =>
1918 {
1919 Some(RetryStrategy::Fixed {
1920 delay: BASE_RETRY_DELAY,
1921 max_attempts: 3,
1922 })
1923 }
1924 Other(err)
1925 if err.is::<language_model::PaymentRequiredError>()
1926 || err.is::<language_model::ModelRequestLimitReachedError>() =>
1927 {
1928 // Retrying won't help for Payment Required or Model Request Limit errors (where
1929 // the user must upgrade to usage-based billing to get more requests, or else wait
1930 // for a significant amount of time for the request limit to reset).
1931 None
1932 }
1933 // Conservatively assume that any other errors are non-retryable
1934 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
1935 delay: BASE_RETRY_DELAY,
1936 max_attempts: 2,
1937 }),
1938 }
1939 }
1940}
1941
1942struct RunningTurn {
1943 /// Holds the task that handles agent interaction until the end of the turn.
1944 /// Survives across multiple requests as the model performs tool calls and
1945 /// we run tools, report their results.
1946 _task: Task<()>,
1947 /// The current event stream for the running turn. Used to report a final
1948 /// cancellation event if we cancel the turn.
1949 event_stream: ThreadEventStream,
1950}
1951
1952impl RunningTurn {
1953 fn cancel(self) {
1954 log::debug!("Cancelling in progress turn");
1955 self.event_stream.send_canceled();
1956 }
1957}
1958
1959pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
1960
1961impl EventEmitter<TokenUsageUpdated> for Thread {}
1962
1963pub trait AgentTool
1964where
1965 Self: 'static + Sized,
1966{
1967 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
1968 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
1969
1970 fn name(&self) -> SharedString;
1971
1972 fn description(&self) -> SharedString {
1973 let schema = schemars::schema_for!(Self::Input);
1974 SharedString::new(
1975 schema
1976 .get("description")
1977 .and_then(|description| description.as_str())
1978 .unwrap_or_default(),
1979 )
1980 }
1981
1982 fn kind(&self) -> acp::ToolKind;
1983
1984 /// The initial tool title to display. Can be updated during the tool run.
1985 fn initial_title(&self, input: Result<Self::Input, serde_json::Value>) -> SharedString;
1986
1987 /// Returns the JSON schema that describes the tool's input.
1988 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Schema {
1989 crate::tool_schema::root_schema_for::<Self::Input>(format)
1990 }
1991
1992 /// Some tools rely on a provider for the underlying billing or other reasons.
1993 /// Allow the tool to check if they are compatible, or should be filtered out.
1994 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
1995 true
1996 }
1997
1998 /// Runs the tool with the provided input.
1999 fn run(
2000 self: Arc<Self>,
2001 input: Self::Input,
2002 event_stream: ToolCallEventStream,
2003 cx: &mut App,
2004 ) -> Task<Result<Self::Output>>;
2005
2006 /// Emits events for a previous execution of the tool.
2007 fn replay(
2008 &self,
2009 _input: Self::Input,
2010 _output: Self::Output,
2011 _event_stream: ToolCallEventStream,
2012 _cx: &mut App,
2013 ) -> Result<()> {
2014 Ok(())
2015 }
2016
2017 fn erase(self) -> Arc<dyn AnyAgentTool> {
2018 Arc::new(Erased(Arc::new(self)))
2019 }
2020}
2021
2022pub struct Erased<T>(T);
2023
2024pub struct AgentToolOutput {
2025 pub llm_output: LanguageModelToolResultContent,
2026 pub raw_output: serde_json::Value,
2027}
2028
2029pub trait AnyAgentTool {
2030 fn name(&self) -> SharedString;
2031 fn description(&self) -> SharedString;
2032 fn kind(&self) -> acp::ToolKind;
2033 fn initial_title(&self, input: serde_json::Value) -> SharedString;
2034 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2035 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2036 true
2037 }
2038 fn run(
2039 self: Arc<Self>,
2040 input: serde_json::Value,
2041 event_stream: ToolCallEventStream,
2042 cx: &mut App,
2043 ) -> Task<Result<AgentToolOutput>>;
2044 fn replay(
2045 &self,
2046 input: serde_json::Value,
2047 output: serde_json::Value,
2048 event_stream: ToolCallEventStream,
2049 cx: &mut App,
2050 ) -> Result<()>;
2051}
2052
2053impl<T> AnyAgentTool for Erased<Arc<T>>
2054where
2055 T: AgentTool,
2056{
2057 fn name(&self) -> SharedString {
2058 self.0.name()
2059 }
2060
2061 fn description(&self) -> SharedString {
2062 self.0.description()
2063 }
2064
2065 fn kind(&self) -> agent_client_protocol::ToolKind {
2066 self.0.kind()
2067 }
2068
2069 fn initial_title(&self, input: serde_json::Value) -> SharedString {
2070 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2071 self.0.initial_title(parsed_input)
2072 }
2073
2074 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2075 let mut json = serde_json::to_value(self.0.input_schema(format))?;
2076 adapt_schema_to_format(&mut json, format)?;
2077 Ok(json)
2078 }
2079
2080 fn supported_provider(&self, provider: &LanguageModelProviderId) -> bool {
2081 self.0.supported_provider(provider)
2082 }
2083
2084 fn run(
2085 self: Arc<Self>,
2086 input: serde_json::Value,
2087 event_stream: ToolCallEventStream,
2088 cx: &mut App,
2089 ) -> Task<Result<AgentToolOutput>> {
2090 cx.spawn(async move |cx| {
2091 let input = serde_json::from_value(input)?;
2092 let output = cx
2093 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2094 .await?;
2095 let raw_output = serde_json::to_value(&output)?;
2096 Ok(AgentToolOutput {
2097 llm_output: output.into(),
2098 raw_output,
2099 })
2100 })
2101 }
2102
2103 fn replay(
2104 &self,
2105 input: serde_json::Value,
2106 output: serde_json::Value,
2107 event_stream: ToolCallEventStream,
2108 cx: &mut App,
2109 ) -> Result<()> {
2110 let input = serde_json::from_value(input)?;
2111 let output = serde_json::from_value(output)?;
2112 self.0.replay(input, output, event_stream, cx)
2113 }
2114}
2115
2116#[derive(Clone)]
2117struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2118
2119impl ThreadEventStream {
2120 fn send_title_update(&self, text: SharedString) {
2121 self.0
2122 .unbounded_send(Ok(ThreadEvent::TitleUpdate(text)))
2123 .ok();
2124 }
2125
2126 fn send_user_message(&self, message: &UserMessage) {
2127 self.0
2128 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2129 .ok();
2130 }
2131
2132 fn send_text(&self, text: &str) {
2133 self.0
2134 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2135 .ok();
2136 }
2137
2138 fn send_thinking(&self, text: &str) {
2139 self.0
2140 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2141 .ok();
2142 }
2143
2144 fn send_tool_call(
2145 &self,
2146 id: &LanguageModelToolUseId,
2147 title: SharedString,
2148 kind: acp::ToolKind,
2149 input: serde_json::Value,
2150 ) {
2151 self.0
2152 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2153 id,
2154 title.to_string(),
2155 kind,
2156 input,
2157 ))))
2158 .ok();
2159 }
2160
2161 fn initial_tool_call(
2162 id: &LanguageModelToolUseId,
2163 title: String,
2164 kind: acp::ToolKind,
2165 input: serde_json::Value,
2166 ) -> acp::ToolCall {
2167 acp::ToolCall {
2168 id: acp::ToolCallId(id.to_string().into()),
2169 title,
2170 kind,
2171 status: acp::ToolCallStatus::Pending,
2172 content: vec![],
2173 locations: vec![],
2174 raw_input: Some(input),
2175 raw_output: None,
2176 }
2177 }
2178
2179 fn update_tool_call_fields(
2180 &self,
2181 tool_use_id: &LanguageModelToolUseId,
2182 fields: acp::ToolCallUpdateFields,
2183 ) {
2184 self.0
2185 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2186 acp::ToolCallUpdate {
2187 id: acp::ToolCallId(tool_use_id.to_string().into()),
2188 fields,
2189 }
2190 .into(),
2191 )))
2192 .ok();
2193 }
2194
2195 fn send_retry(&self, status: acp_thread::RetryStatus) {
2196 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2197 }
2198
2199 fn send_stop(&self, reason: StopReason) {
2200 match reason {
2201 StopReason::EndTurn => {
2202 self.0
2203 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::EndTurn)))
2204 .ok();
2205 }
2206 StopReason::MaxTokens => {
2207 self.0
2208 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::MaxTokens)))
2209 .ok();
2210 }
2211 StopReason::Refusal => {
2212 self.0
2213 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Refusal)))
2214 .ok();
2215 }
2216 StopReason::ToolUse => {}
2217 }
2218 }
2219
2220 fn send_canceled(&self) {
2221 self.0
2222 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Canceled)))
2223 .ok();
2224 }
2225
2226 fn send_error(&self, error: impl Into<anyhow::Error>) {
2227 self.0.unbounded_send(Err(error.into())).ok();
2228 }
2229}
2230
2231#[derive(Clone)]
2232pub struct ToolCallEventStream {
2233 tool_use_id: LanguageModelToolUseId,
2234 stream: ThreadEventStream,
2235 fs: Option<Arc<dyn Fs>>,
2236}
2237
2238impl ToolCallEventStream {
2239 #[cfg(test)]
2240 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2241 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2242
2243 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2244
2245 (stream, ToolCallEventStreamReceiver(events_rx))
2246 }
2247
2248 fn new(
2249 tool_use_id: LanguageModelToolUseId,
2250 stream: ThreadEventStream,
2251 fs: Option<Arc<dyn Fs>>,
2252 ) -> Self {
2253 Self {
2254 tool_use_id,
2255 stream,
2256 fs,
2257 }
2258 }
2259
2260 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2261 self.stream
2262 .update_tool_call_fields(&self.tool_use_id, fields);
2263 }
2264
2265 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2266 self.stream
2267 .0
2268 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2269 acp_thread::ToolCallUpdateDiff {
2270 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2271 diff,
2272 }
2273 .into(),
2274 )))
2275 .ok();
2276 }
2277
2278 pub fn update_terminal(&self, terminal: Entity<acp_thread::Terminal>) {
2279 self.stream
2280 .0
2281 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2282 acp_thread::ToolCallUpdateTerminal {
2283 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2284 terminal,
2285 }
2286 .into(),
2287 )))
2288 .ok();
2289 }
2290
2291 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2292 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2293 return Task::ready(Ok(()));
2294 }
2295
2296 let (response_tx, response_rx) = oneshot::channel();
2297 self.stream
2298 .0
2299 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2300 ToolCallAuthorization {
2301 tool_call: acp::ToolCallUpdate {
2302 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2303 fields: acp::ToolCallUpdateFields {
2304 title: Some(title.into()),
2305 ..Default::default()
2306 },
2307 },
2308 options: vec![
2309 acp::PermissionOption {
2310 id: acp::PermissionOptionId("always_allow".into()),
2311 name: "Always Allow".into(),
2312 kind: acp::PermissionOptionKind::AllowAlways,
2313 },
2314 acp::PermissionOption {
2315 id: acp::PermissionOptionId("allow".into()),
2316 name: "Allow".into(),
2317 kind: acp::PermissionOptionKind::AllowOnce,
2318 },
2319 acp::PermissionOption {
2320 id: acp::PermissionOptionId("deny".into()),
2321 name: "Deny".into(),
2322 kind: acp::PermissionOptionKind::RejectOnce,
2323 },
2324 ],
2325 response: response_tx,
2326 },
2327 )))
2328 .ok();
2329 let fs = self.fs.clone();
2330 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2331 "always_allow" => {
2332 if let Some(fs) = fs.clone() {
2333 cx.update(|cx| {
2334 update_settings_file::<AgentSettings>(fs, cx, |settings, _| {
2335 settings.set_always_allow_tool_actions(true);
2336 });
2337 })?;
2338 }
2339
2340 Ok(())
2341 }
2342 "allow" => Ok(()),
2343 _ => Err(anyhow!("Permission to run tool denied by user")),
2344 })
2345 }
2346}
2347
2348#[cfg(test)]
2349pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2350
2351#[cfg(test)]
2352impl ToolCallEventStreamReceiver {
2353 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2354 let event = self.0.next().await;
2355 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2356 auth
2357 } else {
2358 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2359 }
2360 }
2361
2362 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2363 let event = self.0.next().await;
2364 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2365 update,
2366 )))) = event
2367 {
2368 update.terminal
2369 } else {
2370 panic!("Expected terminal but got: {:?}", event);
2371 }
2372 }
2373}
2374
2375#[cfg(test)]
2376impl std::ops::Deref for ToolCallEventStreamReceiver {
2377 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2378
2379 fn deref(&self) -> &Self::Target {
2380 &self.0
2381 }
2382}
2383
2384#[cfg(test)]
2385impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2386 fn deref_mut(&mut self) -> &mut Self::Target {
2387 &mut self.0
2388 }
2389}
2390
2391impl From<&str> for UserMessageContent {
2392 fn from(text: &str) -> Self {
2393 Self::Text(text.into())
2394 }
2395}
2396
2397impl From<acp::ContentBlock> for UserMessageContent {
2398 fn from(value: acp::ContentBlock) -> Self {
2399 match value {
2400 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2401 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2402 acp::ContentBlock::Audio(_) => {
2403 // TODO
2404 Self::Text("[audio]".to_string())
2405 }
2406 acp::ContentBlock::ResourceLink(resource_link) => {
2407 match MentionUri::parse(&resource_link.uri) {
2408 Ok(uri) => Self::Mention {
2409 uri,
2410 content: String::new(),
2411 },
2412 Err(err) => {
2413 log::error!("Failed to parse mention link: {}", err);
2414 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2415 }
2416 }
2417 }
2418 acp::ContentBlock::Resource(resource) => match resource.resource {
2419 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2420 match MentionUri::parse(&resource.uri) {
2421 Ok(uri) => Self::Mention {
2422 uri,
2423 content: resource.text,
2424 },
2425 Err(err) => {
2426 log::error!("Failed to parse mention link: {}", err);
2427 Self::Text(
2428 MarkdownCodeBlock {
2429 tag: &resource.uri,
2430 text: &resource.text,
2431 }
2432 .to_string(),
2433 )
2434 }
2435 }
2436 }
2437 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2438 // TODO
2439 Self::Text("[blob]".to_string())
2440 }
2441 },
2442 }
2443 }
2444}
2445
2446impl From<UserMessageContent> for acp::ContentBlock {
2447 fn from(content: UserMessageContent) -> Self {
2448 match content {
2449 UserMessageContent::Text(text) => acp::ContentBlock::Text(acp::TextContent {
2450 text,
2451 annotations: None,
2452 }),
2453 UserMessageContent::Image(image) => acp::ContentBlock::Image(acp::ImageContent {
2454 data: image.source.to_string(),
2455 mime_type: "image/png".to_string(),
2456 annotations: None,
2457 uri: None,
2458 }),
2459 UserMessageContent::Mention { uri, content } => {
2460 acp::ContentBlock::ResourceLink(acp::ResourceLink {
2461 uri: uri.to_uri().to_string(),
2462 name: uri.name(),
2463 annotations: None,
2464 description: if content.is_empty() {
2465 None
2466 } else {
2467 Some(content)
2468 },
2469 mime_type: None,
2470 size: None,
2471 title: None,
2472 })
2473 }
2474 }
2475 }
2476}
2477
2478fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2479 LanguageModelImage {
2480 source: image_content.data.into(),
2481 // TODO: make this optional?
2482 size: gpui::Size::new(0.into(), 0.into()),
2483 }
2484}