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, AgentProfileSettings, AgentSettings, CompletionMode,
13 SUMMARIZE_THREAD_DETAILED_PROMPT, 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, HashSet, 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::fmt::Write;
49use std::{
50 collections::BTreeMap,
51 ops::RangeInclusive,
52 path::Path,
53 sync::Arc,
54 time::{Duration, Instant},
55};
56use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock};
57use uuid::Uuid;
58
59const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
60pub const MAX_TOOL_NAME_LENGTH: usize = 64;
61
62/// The ID of the user prompt that initiated a request.
63///
64/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
65#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
66pub struct PromptId(Arc<str>);
67
68impl PromptId {
69 pub fn new() -> Self {
70 Self(Uuid::new_v4().to_string().into())
71 }
72}
73
74impl std::fmt::Display for PromptId {
75 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
76 write!(f, "{}", self.0)
77 }
78}
79
80pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
81pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
82
83#[derive(Debug, Clone)]
84enum RetryStrategy {
85 ExponentialBackoff {
86 initial_delay: Duration,
87 max_attempts: u8,
88 },
89 Fixed {
90 delay: Duration,
91 max_attempts: u8,
92 },
93}
94
95#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
96pub enum Message {
97 User(UserMessage),
98 Agent(AgentMessage),
99 Resume,
100}
101
102impl Message {
103 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
104 match self {
105 Message::Agent(agent_message) => Some(agent_message),
106 _ => None,
107 }
108 }
109
110 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
111 match self {
112 Message::User(message) => vec![message.to_request()],
113 Message::Agent(message) => message.to_request(),
114 Message::Resume => vec![LanguageModelRequestMessage {
115 role: Role::User,
116 content: vec!["Continue where you left off".into()],
117 cache: false,
118 }],
119 }
120 }
121
122 pub fn to_markdown(&self) -> String {
123 match self {
124 Message::User(message) => message.to_markdown(),
125 Message::Agent(message) => message.to_markdown(),
126 Message::Resume => "[resume]\n".into(),
127 }
128 }
129
130 pub fn role(&self) -> Role {
131 match self {
132 Message::User(_) | Message::Resume => Role::User,
133 Message::Agent(_) => Role::Assistant,
134 }
135 }
136}
137
138#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
139pub struct UserMessage {
140 pub id: UserMessageId,
141 pub content: Vec<UserMessageContent>,
142}
143
144#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
145pub enum UserMessageContent {
146 Text(String),
147 Mention { uri: MentionUri, content: String },
148 Image(LanguageModelImage),
149}
150
151impl UserMessage {
152 pub fn to_markdown(&self) -> String {
153 let mut markdown = String::from("## User\n\n");
154
155 for content in &self.content {
156 match content {
157 UserMessageContent::Text(text) => {
158 markdown.push_str(text);
159 markdown.push('\n');
160 }
161 UserMessageContent::Image(_) => {
162 markdown.push_str("<image />\n");
163 }
164 UserMessageContent::Mention { uri, content } => {
165 if !content.is_empty() {
166 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
167 } else {
168 let _ = writeln!(&mut markdown, "{}", uri.as_link());
169 }
170 }
171 }
172 }
173
174 markdown
175 }
176
177 fn to_request(&self) -> LanguageModelRequestMessage {
178 let mut message = LanguageModelRequestMessage {
179 role: Role::User,
180 content: Vec::with_capacity(self.content.len()),
181 cache: false,
182 };
183
184 const OPEN_CONTEXT: &str = "<context>\n\
185 The following items were attached by the user. \
186 They are up-to-date and don't need to be re-read.\n\n";
187
188 const OPEN_FILES_TAG: &str = "<files>";
189 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
190 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
191 const OPEN_SELECTIONS_TAG: &str = "<selections>";
192 const OPEN_THREADS_TAG: &str = "<threads>";
193 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
194 const OPEN_RULES_TAG: &str =
195 "<rules>\nThe user has specified the following rules that should be applied:\n";
196
197 let mut file_context = OPEN_FILES_TAG.to_string();
198 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
199 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
200 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
201 let mut thread_context = OPEN_THREADS_TAG.to_string();
202 let mut fetch_context = OPEN_FETCH_TAG.to_string();
203 let mut rules_context = OPEN_RULES_TAG.to_string();
204
205 for chunk in &self.content {
206 let chunk = match chunk {
207 UserMessageContent::Text(text) => {
208 language_model::MessageContent::Text(text.clone())
209 }
210 UserMessageContent::Image(value) => {
211 language_model::MessageContent::Image(value.clone())
212 }
213 UserMessageContent::Mention { uri, content } => {
214 match uri {
215 MentionUri::File { abs_path } => {
216 write!(
217 &mut file_context,
218 "\n{}",
219 MarkdownCodeBlock {
220 tag: &codeblock_tag(abs_path, None),
221 text: &content.to_string(),
222 }
223 )
224 .ok();
225 }
226 MentionUri::PastedImage => {
227 debug_panic!("pasted image URI should not be used in mention content")
228 }
229 MentionUri::Directory { .. } => {
230 write!(&mut directory_context, "\n{}\n", content).ok();
231 }
232 MentionUri::Symbol {
233 abs_path: path,
234 line_range,
235 ..
236 } => {
237 write!(
238 &mut symbol_context,
239 "\n{}",
240 MarkdownCodeBlock {
241 tag: &codeblock_tag(path, Some(line_range)),
242 text: content
243 }
244 )
245 .ok();
246 }
247 MentionUri::Selection {
248 abs_path: path,
249 line_range,
250 ..
251 } => {
252 write!(
253 &mut selection_context,
254 "\n{}",
255 MarkdownCodeBlock {
256 tag: &codeblock_tag(
257 path.as_deref().unwrap_or("Untitled".as_ref()),
258 Some(line_range)
259 ),
260 text: content
261 }
262 )
263 .ok();
264 }
265 MentionUri::Thread { .. } => {
266 write!(&mut thread_context, "\n{}\n", content).ok();
267 }
268 MentionUri::TextThread { .. } => {
269 write!(&mut thread_context, "\n{}\n", content).ok();
270 }
271 MentionUri::Rule { .. } => {
272 write!(
273 &mut rules_context,
274 "\n{}",
275 MarkdownCodeBlock {
276 tag: "",
277 text: content
278 }
279 )
280 .ok();
281 }
282 MentionUri::Fetch { url } => {
283 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
284 }
285 }
286
287 language_model::MessageContent::Text(uri.as_link().to_string())
288 }
289 };
290
291 message.content.push(chunk);
292 }
293
294 let len_before_context = message.content.len();
295
296 if file_context.len() > OPEN_FILES_TAG.len() {
297 file_context.push_str("</files>\n");
298 message
299 .content
300 .push(language_model::MessageContent::Text(file_context));
301 }
302
303 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
304 directory_context.push_str("</directories>\n");
305 message
306 .content
307 .push(language_model::MessageContent::Text(directory_context));
308 }
309
310 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
311 symbol_context.push_str("</symbols>\n");
312 message
313 .content
314 .push(language_model::MessageContent::Text(symbol_context));
315 }
316
317 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
318 selection_context.push_str("</selections>\n");
319 message
320 .content
321 .push(language_model::MessageContent::Text(selection_context));
322 }
323
324 if thread_context.len() > OPEN_THREADS_TAG.len() {
325 thread_context.push_str("</threads>\n");
326 message
327 .content
328 .push(language_model::MessageContent::Text(thread_context));
329 }
330
331 if fetch_context.len() > OPEN_FETCH_TAG.len() {
332 fetch_context.push_str("</fetched_urls>\n");
333 message
334 .content
335 .push(language_model::MessageContent::Text(fetch_context));
336 }
337
338 if rules_context.len() > OPEN_RULES_TAG.len() {
339 rules_context.push_str("</user_rules>\n");
340 message
341 .content
342 .push(language_model::MessageContent::Text(rules_context));
343 }
344
345 if message.content.len() > len_before_context {
346 message.content.insert(
347 len_before_context,
348 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
349 );
350 message
351 .content
352 .push(language_model::MessageContent::Text("</context>".into()));
353 }
354
355 message
356 }
357}
358
359fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
360 let mut result = String::new();
361
362 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
363 let _ = write!(result, "{} ", extension);
364 }
365
366 let _ = write!(result, "{}", full_path.display());
367
368 if let Some(range) = line_range {
369 if range.start() == range.end() {
370 let _ = write!(result, ":{}", range.start() + 1);
371 } else {
372 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
373 }
374 }
375
376 result
377}
378
379impl AgentMessage {
380 pub fn to_markdown(&self) -> String {
381 let mut markdown = String::from("## Assistant\n\n");
382
383 for content in &self.content {
384 match content {
385 AgentMessageContent::Text(text) => {
386 markdown.push_str(text);
387 markdown.push('\n');
388 }
389 AgentMessageContent::Thinking { text, .. } => {
390 markdown.push_str("<think>");
391 markdown.push_str(text);
392 markdown.push_str("</think>\n");
393 }
394 AgentMessageContent::RedactedThinking(_) => {
395 markdown.push_str("<redacted_thinking />\n")
396 }
397 AgentMessageContent::ToolUse(tool_use) => {
398 markdown.push_str(&format!(
399 "**Tool Use**: {} (ID: {})\n",
400 tool_use.name, tool_use.id
401 ));
402 markdown.push_str(&format!(
403 "{}\n",
404 MarkdownCodeBlock {
405 tag: "json",
406 text: &format!("{:#}", tool_use.input)
407 }
408 ));
409 }
410 }
411 }
412
413 for tool_result in self.tool_results.values() {
414 markdown.push_str(&format!(
415 "**Tool Result**: {} (ID: {})\n\n",
416 tool_result.tool_name, tool_result.tool_use_id
417 ));
418 if tool_result.is_error {
419 markdown.push_str("**ERROR:**\n");
420 }
421
422 match &tool_result.content {
423 LanguageModelToolResultContent::Text(text) => {
424 writeln!(markdown, "{text}\n").ok();
425 }
426 LanguageModelToolResultContent::Image(_) => {
427 writeln!(markdown, "<image />\n").ok();
428 }
429 }
430
431 if let Some(output) = tool_result.output.as_ref() {
432 writeln!(
433 markdown,
434 "**Debug Output**:\n\n```json\n{}\n```\n",
435 serde_json::to_string_pretty(output).unwrap()
436 )
437 .unwrap();
438 }
439 }
440
441 markdown
442 }
443
444 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
445 let mut assistant_message = LanguageModelRequestMessage {
446 role: Role::Assistant,
447 content: Vec::with_capacity(self.content.len()),
448 cache: false,
449 };
450 for chunk in &self.content {
451 match chunk {
452 AgentMessageContent::Text(text) => {
453 assistant_message
454 .content
455 .push(language_model::MessageContent::Text(text.clone()));
456 }
457 AgentMessageContent::Thinking { text, signature } => {
458 assistant_message
459 .content
460 .push(language_model::MessageContent::Thinking {
461 text: text.clone(),
462 signature: signature.clone(),
463 });
464 }
465 AgentMessageContent::RedactedThinking(value) => {
466 assistant_message.content.push(
467 language_model::MessageContent::RedactedThinking(value.clone()),
468 );
469 }
470 AgentMessageContent::ToolUse(tool_use) => {
471 if self.tool_results.contains_key(&tool_use.id) {
472 assistant_message
473 .content
474 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
475 }
476 }
477 };
478 }
479
480 let mut user_message = LanguageModelRequestMessage {
481 role: Role::User,
482 content: Vec::new(),
483 cache: false,
484 };
485
486 for tool_result in self.tool_results.values() {
487 user_message
488 .content
489 .push(language_model::MessageContent::ToolResult(
490 tool_result.clone(),
491 ));
492 }
493
494 let mut messages = Vec::new();
495 if !assistant_message.content.is_empty() {
496 messages.push(assistant_message);
497 }
498 if !user_message.content.is_empty() {
499 messages.push(user_message);
500 }
501 messages
502 }
503}
504
505#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
506pub struct AgentMessage {
507 pub content: Vec<AgentMessageContent>,
508 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
509}
510
511#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
512pub enum AgentMessageContent {
513 Text(String),
514 Thinking {
515 text: String,
516 signature: Option<String>,
517 },
518 RedactedThinking(String),
519 ToolUse(LanguageModelToolUse),
520}
521
522#[derive(Debug)]
523pub enum ThreadEvent {
524 UserMessage(UserMessage),
525 AgentText(String),
526 AgentThinking(String),
527 ToolCall(acp::ToolCall),
528 ToolCallUpdate(acp_thread::ToolCallUpdate),
529 ToolCallAuthorization(ToolCallAuthorization),
530 Retry(acp_thread::RetryStatus),
531 Stop(acp::StopReason),
532}
533
534#[derive(Debug)]
535pub struct ToolCallAuthorization {
536 pub tool_call: acp::ToolCallUpdate,
537 pub options: Vec<acp::PermissionOption>,
538 pub response: oneshot::Sender<acp::PermissionOptionId>,
539}
540
541#[derive(Debug, thiserror::Error)]
542enum CompletionError {
543 #[error("max tokens")]
544 MaxTokens,
545 #[error("refusal")]
546 Refusal,
547 #[error(transparent)]
548 Other(#[from] anyhow::Error),
549}
550
551pub struct Thread {
552 id: acp::SessionId,
553 prompt_id: PromptId,
554 updated_at: DateTime<Utc>,
555 title: Option<SharedString>,
556 pending_title_generation: Option<Task<()>>,
557 summary: Option<SharedString>,
558 messages: Vec<Message>,
559 completion_mode: CompletionMode,
560 /// Holds the task that handles agent interaction until the end of the turn.
561 /// Survives across multiple requests as the model performs tool calls and
562 /// we run tools, report their results.
563 running_turn: Option<RunningTurn>,
564 pending_message: Option<AgentMessage>,
565 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
566 tool_use_limit_reached: bool,
567 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
568 #[allow(unused)]
569 cumulative_token_usage: TokenUsage,
570 #[allow(unused)]
571 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
572 context_server_registry: Entity<ContextServerRegistry>,
573 profile_id: AgentProfileId,
574 project_context: Entity<ProjectContext>,
575 templates: Arc<Templates>,
576 model: Option<Arc<dyn LanguageModel>>,
577 summarization_model: Option<Arc<dyn LanguageModel>>,
578 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
579 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
580 pub(crate) project: Entity<Project>,
581 pub(crate) action_log: Entity<ActionLog>,
582}
583
584impl Thread {
585 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
586 let image = model.map_or(true, |model| model.supports_images());
587 acp::PromptCapabilities {
588 image,
589 audio: false,
590 embedded_context: true,
591 }
592 }
593
594 pub fn new(
595 project: Entity<Project>,
596 project_context: Entity<ProjectContext>,
597 context_server_registry: Entity<ContextServerRegistry>,
598 templates: Arc<Templates>,
599 model: Option<Arc<dyn LanguageModel>>,
600 cx: &mut Context<Self>,
601 ) -> Self {
602 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
603 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
604 let (prompt_capabilities_tx, prompt_capabilities_rx) =
605 watch::channel(Self::prompt_capabilities(model.as_deref()));
606 Self {
607 id: acp::SessionId(uuid::Uuid::new_v4().to_string().into()),
608 prompt_id: PromptId::new(),
609 updated_at: Utc::now(),
610 title: None,
611 pending_title_generation: None,
612 summary: None,
613 messages: Vec::new(),
614 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
615 running_turn: None,
616 pending_message: None,
617 tools: BTreeMap::default(),
618 tool_use_limit_reached: false,
619 request_token_usage: HashMap::default(),
620 cumulative_token_usage: TokenUsage::default(),
621 initial_project_snapshot: {
622 let project_snapshot = Self::project_snapshot(project.clone(), cx);
623 cx.foreground_executor()
624 .spawn(async move { Some(project_snapshot.await) })
625 .shared()
626 },
627 context_server_registry,
628 profile_id,
629 project_context,
630 templates,
631 model,
632 summarization_model: None,
633 prompt_capabilities_tx,
634 prompt_capabilities_rx,
635 project,
636 action_log,
637 }
638 }
639
640 pub fn id(&self) -> &acp::SessionId {
641 &self.id
642 }
643
644 pub fn replay(
645 &mut self,
646 cx: &mut Context<Self>,
647 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
648 let (tx, rx) = mpsc::unbounded();
649 let stream = ThreadEventStream(tx);
650 for message in &self.messages {
651 match message {
652 Message::User(user_message) => stream.send_user_message(user_message),
653 Message::Agent(assistant_message) => {
654 for content in &assistant_message.content {
655 match content {
656 AgentMessageContent::Text(text) => stream.send_text(text),
657 AgentMessageContent::Thinking { text, .. } => {
658 stream.send_thinking(text)
659 }
660 AgentMessageContent::RedactedThinking(_) => {}
661 AgentMessageContent::ToolUse(tool_use) => {
662 self.replay_tool_call(
663 tool_use,
664 assistant_message.tool_results.get(&tool_use.id),
665 &stream,
666 cx,
667 );
668 }
669 }
670 }
671 }
672 Message::Resume => {}
673 }
674 }
675 rx
676 }
677
678 fn replay_tool_call(
679 &self,
680 tool_use: &LanguageModelToolUse,
681 tool_result: Option<&LanguageModelToolResult>,
682 stream: &ThreadEventStream,
683 cx: &mut Context<Self>,
684 ) {
685 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
686 self.context_server_registry
687 .read(cx)
688 .servers()
689 .find_map(|(_, tools)| {
690 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
691 Some(tool.clone())
692 } else {
693 None
694 }
695 })
696 });
697
698 let Some(tool) = tool else {
699 stream
700 .0
701 .unbounded_send(Ok(ThreadEvent::ToolCall(acp::ToolCall {
702 id: acp::ToolCallId(tool_use.id.to_string().into()),
703 title: tool_use.name.to_string(),
704 kind: acp::ToolKind::Other,
705 status: acp::ToolCallStatus::Failed,
706 content: Vec::new(),
707 locations: Vec::new(),
708 raw_input: Some(tool_use.input.clone()),
709 raw_output: None,
710 })))
711 .ok();
712 return;
713 };
714
715 let title = tool.initial_title(tool_use.input.clone());
716 let kind = tool.kind();
717 stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
718
719 let output = tool_result
720 .as_ref()
721 .and_then(|result| result.output.clone());
722 if let Some(output) = output.clone() {
723 let tool_event_stream = ToolCallEventStream::new(
724 tool_use.id.clone(),
725 stream.clone(),
726 Some(self.project.read(cx).fs().clone()),
727 );
728 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
729 .log_err();
730 }
731
732 stream.update_tool_call_fields(
733 &tool_use.id,
734 acp::ToolCallUpdateFields {
735 status: Some(acp::ToolCallStatus::Completed),
736 raw_output: output,
737 ..Default::default()
738 },
739 );
740 }
741
742 pub fn from_db(
743 id: acp::SessionId,
744 db_thread: DbThread,
745 project: Entity<Project>,
746 project_context: Entity<ProjectContext>,
747 context_server_registry: Entity<ContextServerRegistry>,
748 action_log: Entity<ActionLog>,
749 templates: Arc<Templates>,
750 cx: &mut Context<Self>,
751 ) -> Self {
752 let profile_id = db_thread
753 .profile
754 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
755 let model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
756 db_thread
757 .model
758 .and_then(|model| {
759 let model = SelectedModel {
760 provider: model.provider.clone().into(),
761 model: model.model.into(),
762 };
763 registry.select_model(&model, cx)
764 })
765 .or_else(|| registry.default_model())
766 .map(|model| model.model)
767 });
768 let (prompt_capabilities_tx, prompt_capabilities_rx) =
769 watch::channel(Self::prompt_capabilities(model.as_deref()));
770
771 Self {
772 id,
773 prompt_id: PromptId::new(),
774 title: if db_thread.title.is_empty() {
775 None
776 } else {
777 Some(db_thread.title.clone())
778 },
779 pending_title_generation: None,
780 summary: db_thread.detailed_summary,
781 messages: db_thread.messages,
782 completion_mode: db_thread.completion_mode.unwrap_or_default(),
783 running_turn: None,
784 pending_message: None,
785 tools: BTreeMap::default(),
786 tool_use_limit_reached: false,
787 request_token_usage: db_thread.request_token_usage.clone(),
788 cumulative_token_usage: db_thread.cumulative_token_usage,
789 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
790 context_server_registry,
791 profile_id,
792 project_context,
793 templates,
794 model,
795 summarization_model: None,
796 project,
797 action_log,
798 updated_at: db_thread.updated_at,
799 prompt_capabilities_tx,
800 prompt_capabilities_rx,
801 }
802 }
803
804 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
805 let initial_project_snapshot = self.initial_project_snapshot.clone();
806 let mut thread = DbThread {
807 title: self.title(),
808 messages: self.messages.clone(),
809 updated_at: self.updated_at,
810 detailed_summary: self.summary.clone(),
811 initial_project_snapshot: None,
812 cumulative_token_usage: self.cumulative_token_usage,
813 request_token_usage: self.request_token_usage.clone(),
814 model: self.model.as_ref().map(|model| DbLanguageModel {
815 provider: model.provider_id().to_string(),
816 model: model.name().0.to_string(),
817 }),
818 completion_mode: Some(self.completion_mode),
819 profile: Some(self.profile_id.clone()),
820 };
821
822 cx.background_spawn(async move {
823 let initial_project_snapshot = initial_project_snapshot.await;
824 thread.initial_project_snapshot = initial_project_snapshot;
825 thread
826 })
827 }
828
829 /// Create a snapshot of the current project state including git information and unsaved buffers.
830 fn project_snapshot(
831 project: Entity<Project>,
832 cx: &mut Context<Self>,
833 ) -> Task<Arc<agent::thread::ProjectSnapshot>> {
834 let git_store = project.read(cx).git_store().clone();
835 let worktree_snapshots: Vec<_> = project
836 .read(cx)
837 .visible_worktrees(cx)
838 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
839 .collect();
840
841 cx.spawn(async move |_, cx| {
842 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
843
844 let mut unsaved_buffers = Vec::new();
845 cx.update(|app_cx| {
846 let buffer_store = project.read(app_cx).buffer_store();
847 for buffer_handle in buffer_store.read(app_cx).buffers() {
848 let buffer = buffer_handle.read(app_cx);
849 if buffer.is_dirty()
850 && let Some(file) = buffer.file()
851 {
852 let path = file.path().to_string_lossy().to_string();
853 unsaved_buffers.push(path);
854 }
855 }
856 })
857 .ok();
858
859 Arc::new(ProjectSnapshot {
860 worktree_snapshots,
861 unsaved_buffer_paths: unsaved_buffers,
862 timestamp: Utc::now(),
863 })
864 })
865 }
866
867 fn worktree_snapshot(
868 worktree: Entity<project::Worktree>,
869 git_store: Entity<GitStore>,
870 cx: &App,
871 ) -> Task<agent::thread::WorktreeSnapshot> {
872 cx.spawn(async move |cx| {
873 // Get worktree path and snapshot
874 let worktree_info = cx.update(|app_cx| {
875 let worktree = worktree.read(app_cx);
876 let path = worktree.abs_path().to_string_lossy().to_string();
877 let snapshot = worktree.snapshot();
878 (path, snapshot)
879 });
880
881 let Ok((worktree_path, _snapshot)) = worktree_info else {
882 return WorktreeSnapshot {
883 worktree_path: String::new(),
884 git_state: None,
885 };
886 };
887
888 let git_state = git_store
889 .update(cx, |git_store, cx| {
890 git_store
891 .repositories()
892 .values()
893 .find(|repo| {
894 repo.read(cx)
895 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
896 .is_some()
897 })
898 .cloned()
899 })
900 .ok()
901 .flatten()
902 .map(|repo| {
903 repo.update(cx, |repo, _| {
904 let current_branch =
905 repo.branch.as_ref().map(|branch| branch.name().to_owned());
906 repo.send_job(None, |state, _| async move {
907 let RepositoryState::Local { backend, .. } = state else {
908 return GitState {
909 remote_url: None,
910 head_sha: None,
911 current_branch,
912 diff: None,
913 };
914 };
915
916 let remote_url = backend.remote_url("origin");
917 let head_sha = backend.head_sha().await;
918 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
919
920 GitState {
921 remote_url,
922 head_sha,
923 current_branch,
924 diff,
925 }
926 })
927 })
928 });
929
930 let git_state = match git_state {
931 Some(git_state) => match git_state.ok() {
932 Some(git_state) => git_state.await.ok(),
933 None => None,
934 },
935 None => None,
936 };
937
938 WorktreeSnapshot {
939 worktree_path,
940 git_state,
941 }
942 })
943 }
944
945 pub fn project_context(&self) -> &Entity<ProjectContext> {
946 &self.project_context
947 }
948
949 pub fn project(&self) -> &Entity<Project> {
950 &self.project
951 }
952
953 pub fn action_log(&self) -> &Entity<ActionLog> {
954 &self.action_log
955 }
956
957 pub fn is_empty(&self) -> bool {
958 self.messages.is_empty() && self.title.is_none()
959 }
960
961 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
962 self.model.as_ref()
963 }
964
965 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
966 let old_usage = self.latest_token_usage();
967 self.model = Some(model);
968 let new_caps = Self::prompt_capabilities(self.model.as_deref());
969 let new_usage = self.latest_token_usage();
970 if old_usage != new_usage {
971 cx.emit(TokenUsageUpdated(new_usage));
972 }
973 self.prompt_capabilities_tx.send(new_caps).log_err();
974 cx.notify()
975 }
976
977 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
978 self.summarization_model.as_ref()
979 }
980
981 pub fn set_summarization_model(
982 &mut self,
983 model: Option<Arc<dyn LanguageModel>>,
984 cx: &mut Context<Self>,
985 ) {
986 self.summarization_model = model;
987 cx.notify()
988 }
989
990 pub fn completion_mode(&self) -> CompletionMode {
991 self.completion_mode
992 }
993
994 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
995 let old_usage = self.latest_token_usage();
996 self.completion_mode = mode;
997 let new_usage = self.latest_token_usage();
998 if old_usage != new_usage {
999 cx.emit(TokenUsageUpdated(new_usage));
1000 }
1001 cx.notify()
1002 }
1003
1004 #[cfg(any(test, feature = "test-support"))]
1005 pub fn last_message(&self) -> Option<Message> {
1006 if let Some(message) = self.pending_message.clone() {
1007 Some(Message::Agent(message))
1008 } else {
1009 self.messages.last().cloned()
1010 }
1011 }
1012
1013 pub fn add_default_tools(&mut self, cx: &mut Context<Self>) {
1014 let language_registry = self.project.read(cx).languages().clone();
1015 self.add_tool(CopyPathTool::new(self.project.clone()));
1016 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1017 self.add_tool(DeletePathTool::new(
1018 self.project.clone(),
1019 self.action_log.clone(),
1020 ));
1021 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1022 self.add_tool(EditFileTool::new(cx.weak_entity(), language_registry));
1023 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1024 self.add_tool(FindPathTool::new(self.project.clone()));
1025 self.add_tool(GrepTool::new(self.project.clone()));
1026 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1027 self.add_tool(MovePathTool::new(self.project.clone()));
1028 self.add_tool(NowTool);
1029 self.add_tool(OpenTool::new(self.project.clone()));
1030 self.add_tool(ReadFileTool::new(
1031 self.project.clone(),
1032 self.action_log.clone(),
1033 ));
1034 self.add_tool(TerminalTool::new(self.project.clone(), cx));
1035 self.add_tool(ThinkingTool);
1036 self.add_tool(WebSearchTool);
1037 }
1038
1039 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1040 self.tools.insert(T::name().into(), tool.erase());
1041 }
1042
1043 pub fn remove_tool(&mut self, name: &str) -> bool {
1044 self.tools.remove(name).is_some()
1045 }
1046
1047 pub fn profile(&self) -> &AgentProfileId {
1048 &self.profile_id
1049 }
1050
1051 pub fn set_profile(&mut self, profile_id: AgentProfileId) {
1052 self.profile_id = profile_id;
1053 }
1054
1055 pub fn cancel(&mut self, cx: &mut Context<Self>) {
1056 if let Some(running_turn) = self.running_turn.take() {
1057 running_turn.cancel();
1058 }
1059 self.flush_pending_message(cx);
1060 }
1061
1062 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1063 let Some(last_user_message) = self.last_user_message() else {
1064 return;
1065 };
1066
1067 self.request_token_usage
1068 .insert(last_user_message.id.clone(), update);
1069 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1070 cx.notify();
1071 }
1072
1073 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1074 self.cancel(cx);
1075 let Some(position) = self.messages.iter().position(
1076 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1077 ) else {
1078 return Err(anyhow!("Message not found"));
1079 };
1080
1081 for message in self.messages.drain(position..) {
1082 match message {
1083 Message::User(message) => {
1084 self.request_token_usage.remove(&message.id);
1085 }
1086 Message::Agent(_) | Message::Resume => {}
1087 }
1088 }
1089 self.summary = None;
1090 cx.notify();
1091 Ok(())
1092 }
1093
1094 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1095 let last_user_message = self.last_user_message()?;
1096 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1097 let model = self.model.clone()?;
1098
1099 Some(acp_thread::TokenUsage {
1100 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1101 used_tokens: tokens.total_tokens(),
1102 })
1103 }
1104
1105 pub fn resume(
1106 &mut self,
1107 cx: &mut Context<Self>,
1108 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1109 self.messages.push(Message::Resume);
1110 cx.notify();
1111
1112 log::debug!("Total messages in thread: {}", self.messages.len());
1113 self.run_turn(cx)
1114 }
1115
1116 /// Sending a message results in the model streaming a response, which could include tool calls.
1117 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1118 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1119 pub fn send<T>(
1120 &mut self,
1121 id: UserMessageId,
1122 content: impl IntoIterator<Item = T>,
1123 cx: &mut Context<Self>,
1124 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1125 where
1126 T: Into<UserMessageContent>,
1127 {
1128 let model = self.model().context("No language model configured")?;
1129
1130 log::info!("Thread::send called with model: {}", model.name().0);
1131 self.advance_prompt_id();
1132
1133 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1134 log::debug!("Thread::send content: {:?}", content);
1135
1136 self.messages
1137 .push(Message::User(UserMessage { id, content }));
1138 cx.notify();
1139
1140 log::debug!("Total messages in thread: {}", self.messages.len());
1141 self.run_turn(cx)
1142 }
1143
1144 fn run_turn(
1145 &mut self,
1146 cx: &mut Context<Self>,
1147 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1148 self.cancel(cx);
1149
1150 let model = self.model.clone().context("No language model configured")?;
1151 let profile = AgentSettings::get_global(cx)
1152 .profiles
1153 .get(&self.profile_id)
1154 .context("Profile not found")?;
1155 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1156 let event_stream = ThreadEventStream(events_tx);
1157 let message_ix = self.messages.len().saturating_sub(1);
1158 self.tool_use_limit_reached = false;
1159 self.summary = None;
1160 self.running_turn = Some(RunningTurn {
1161 event_stream: event_stream.clone(),
1162 tools: self.enabled_tools(profile, &model, cx),
1163 _task: cx.spawn(async move |this, cx| {
1164 log::debug!("Starting agent turn execution");
1165
1166 let turn_result = Self::run_turn_internal(&this, model, &event_stream, cx).await;
1167 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1168
1169 match turn_result {
1170 Ok(()) => {
1171 log::debug!("Turn execution completed");
1172 event_stream.send_stop(acp::StopReason::EndTurn);
1173 }
1174 Err(error) => {
1175 log::error!("Turn execution failed: {:?}", error);
1176 match error.downcast::<CompletionError>() {
1177 Ok(CompletionError::Refusal) => {
1178 event_stream.send_stop(acp::StopReason::Refusal);
1179 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1180 }
1181 Ok(CompletionError::MaxTokens) => {
1182 event_stream.send_stop(acp::StopReason::MaxTokens);
1183 }
1184 Ok(CompletionError::Other(error)) | Err(error) => {
1185 event_stream.send_error(error);
1186 }
1187 }
1188 }
1189 }
1190
1191 _ = this.update(cx, |this, _| this.running_turn.take());
1192 }),
1193 });
1194 Ok(events_rx)
1195 }
1196
1197 async fn run_turn_internal(
1198 this: &WeakEntity<Self>,
1199 model: Arc<dyn LanguageModel>,
1200 event_stream: &ThreadEventStream,
1201 cx: &mut AsyncApp,
1202 ) -> Result<()> {
1203 let mut attempt = 0;
1204 let mut intent = CompletionIntent::UserPrompt;
1205 loop {
1206 let request =
1207 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1208
1209 telemetry::event!(
1210 "Agent Thread Completion",
1211 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1212 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1213 model = model.telemetry_id(),
1214 model_provider = model.provider_id().to_string(),
1215 attempt
1216 );
1217
1218 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1219 let mut events = model
1220 .stream_completion(request, cx)
1221 .await
1222 .map_err(|error| anyhow!(error))?;
1223 let mut tool_results = FuturesUnordered::new();
1224 let mut error = None;
1225 while let Some(event) = events.next().await {
1226 log::trace!("Received completion event: {:?}", event);
1227 match event {
1228 Ok(event) => {
1229 tool_results.extend(this.update(cx, |this, cx| {
1230 this.handle_completion_event(event, event_stream, cx)
1231 })??);
1232 }
1233 Err(err) => {
1234 error = Some(err);
1235 break;
1236 }
1237 }
1238 }
1239
1240 let end_turn = tool_results.is_empty();
1241 while let Some(tool_result) = tool_results.next().await {
1242 log::debug!("Tool finished {:?}", tool_result);
1243
1244 event_stream.update_tool_call_fields(
1245 &tool_result.tool_use_id,
1246 acp::ToolCallUpdateFields {
1247 status: Some(if tool_result.is_error {
1248 acp::ToolCallStatus::Failed
1249 } else {
1250 acp::ToolCallStatus::Completed
1251 }),
1252 raw_output: tool_result.output.clone(),
1253 ..Default::default()
1254 },
1255 );
1256 this.update(cx, |this, _cx| {
1257 this.pending_message()
1258 .tool_results
1259 .insert(tool_result.tool_use_id.clone(), tool_result);
1260 })?;
1261 }
1262
1263 this.update(cx, |this, cx| {
1264 this.flush_pending_message(cx);
1265 if this.title.is_none() && this.pending_title_generation.is_none() {
1266 this.generate_title(cx);
1267 }
1268 })?;
1269
1270 if let Some(error) = error {
1271 attempt += 1;
1272 let retry =
1273 this.update(cx, |this, _| this.handle_completion_error(error, attempt))??;
1274 let timer = cx.background_executor().timer(retry.duration);
1275 event_stream.send_retry(retry);
1276 timer.await;
1277 this.update(cx, |this, _cx| {
1278 if let Some(Message::Agent(message)) = this.messages.last() {
1279 if message.tool_results.is_empty() {
1280 intent = CompletionIntent::UserPrompt;
1281 this.messages.push(Message::Resume);
1282 }
1283 }
1284 })?;
1285 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1286 return Err(language_model::ToolUseLimitReachedError.into());
1287 } else if end_turn {
1288 return Ok(());
1289 } else {
1290 intent = CompletionIntent::ToolResults;
1291 attempt = 0;
1292 }
1293 }
1294 }
1295
1296 fn handle_completion_error(
1297 &mut self,
1298 error: LanguageModelCompletionError,
1299 attempt: u8,
1300 ) -> Result<acp_thread::RetryStatus> {
1301 if self.completion_mode == CompletionMode::Normal {
1302 return Err(anyhow!(error));
1303 }
1304
1305 let Some(strategy) = Self::retry_strategy_for(&error) else {
1306 return Err(anyhow!(error));
1307 };
1308
1309 let max_attempts = match &strategy {
1310 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1311 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1312 };
1313
1314 if attempt > max_attempts {
1315 return Err(anyhow!(error));
1316 }
1317
1318 let delay = match &strategy {
1319 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1320 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1321 Duration::from_secs(delay_secs)
1322 }
1323 RetryStrategy::Fixed { delay, .. } => *delay,
1324 };
1325 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1326
1327 Ok(acp_thread::RetryStatus {
1328 last_error: error.to_string().into(),
1329 attempt: attempt as usize,
1330 max_attempts: max_attempts as usize,
1331 started_at: Instant::now(),
1332 duration: delay,
1333 })
1334 }
1335
1336 /// A helper method that's called on every streamed completion event.
1337 /// Returns an optional tool result task, which the main agentic loop will
1338 /// send back to the model when it resolves.
1339 fn handle_completion_event(
1340 &mut self,
1341 event: LanguageModelCompletionEvent,
1342 event_stream: &ThreadEventStream,
1343 cx: &mut Context<Self>,
1344 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1345 log::trace!("Handling streamed completion event: {:?}", event);
1346 use LanguageModelCompletionEvent::*;
1347
1348 match event {
1349 StartMessage { .. } => {
1350 self.flush_pending_message(cx);
1351 self.pending_message = Some(AgentMessage::default());
1352 }
1353 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1354 Thinking { text, signature } => {
1355 self.handle_thinking_event(text, signature, event_stream, cx)
1356 }
1357 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1358 ToolUse(tool_use) => {
1359 return Ok(self.handle_tool_use_event(tool_use, event_stream, cx));
1360 }
1361 ToolUseJsonParseError {
1362 id,
1363 tool_name,
1364 raw_input,
1365 json_parse_error,
1366 } => {
1367 return Ok(Some(Task::ready(
1368 self.handle_tool_use_json_parse_error_event(
1369 id,
1370 tool_name,
1371 raw_input,
1372 json_parse_error,
1373 ),
1374 )));
1375 }
1376 UsageUpdate(usage) => {
1377 telemetry::event!(
1378 "Agent Thread Completion Usage Updated",
1379 thread_id = self.id.to_string(),
1380 prompt_id = self.prompt_id.to_string(),
1381 model = self.model.as_ref().map(|m| m.telemetry_id()),
1382 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1383 input_tokens = usage.input_tokens,
1384 output_tokens = usage.output_tokens,
1385 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1386 cache_read_input_tokens = usage.cache_read_input_tokens,
1387 );
1388 self.update_token_usage(usage, cx);
1389 }
1390 StatusUpdate(CompletionRequestStatus::UsageUpdated { amount, limit }) => {
1391 self.update_model_request_usage(amount, limit, cx);
1392 }
1393 StatusUpdate(
1394 CompletionRequestStatus::Started
1395 | CompletionRequestStatus::Queued { .. }
1396 | CompletionRequestStatus::Failed { .. },
1397 ) => {}
1398 StatusUpdate(CompletionRequestStatus::ToolUseLimitReached) => {
1399 self.tool_use_limit_reached = true;
1400 }
1401 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1402 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1403 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1404 }
1405
1406 Ok(None)
1407 }
1408
1409 fn handle_text_event(
1410 &mut self,
1411 new_text: String,
1412 event_stream: &ThreadEventStream,
1413 cx: &mut Context<Self>,
1414 ) {
1415 event_stream.send_text(&new_text);
1416
1417 let last_message = self.pending_message();
1418 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1419 text.push_str(&new_text);
1420 } else {
1421 last_message
1422 .content
1423 .push(AgentMessageContent::Text(new_text));
1424 }
1425
1426 cx.notify();
1427 }
1428
1429 fn handle_thinking_event(
1430 &mut self,
1431 new_text: String,
1432 new_signature: Option<String>,
1433 event_stream: &ThreadEventStream,
1434 cx: &mut Context<Self>,
1435 ) {
1436 event_stream.send_thinking(&new_text);
1437
1438 let last_message = self.pending_message();
1439 if let Some(AgentMessageContent::Thinking { text, signature }) =
1440 last_message.content.last_mut()
1441 {
1442 text.push_str(&new_text);
1443 *signature = new_signature.or(signature.take());
1444 } else {
1445 last_message.content.push(AgentMessageContent::Thinking {
1446 text: new_text,
1447 signature: new_signature,
1448 });
1449 }
1450
1451 cx.notify();
1452 }
1453
1454 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1455 let last_message = self.pending_message();
1456 last_message
1457 .content
1458 .push(AgentMessageContent::RedactedThinking(data));
1459 cx.notify();
1460 }
1461
1462 fn handle_tool_use_event(
1463 &mut self,
1464 tool_use: LanguageModelToolUse,
1465 event_stream: &ThreadEventStream,
1466 cx: &mut Context<Self>,
1467 ) -> Option<Task<LanguageModelToolResult>> {
1468 cx.notify();
1469
1470 let tool = self.tool(tool_use.name.as_ref());
1471 let mut title = SharedString::from(&tool_use.name);
1472 let mut kind = acp::ToolKind::Other;
1473 if let Some(tool) = tool.as_ref() {
1474 title = tool.initial_title(tool_use.input.clone());
1475 kind = tool.kind();
1476 }
1477
1478 // Ensure the last message ends in the current tool use
1479 let last_message = self.pending_message();
1480 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1481 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1482 if last_tool_use.id == tool_use.id {
1483 *last_tool_use = tool_use.clone();
1484 false
1485 } else {
1486 true
1487 }
1488 } else {
1489 true
1490 }
1491 });
1492
1493 if push_new_tool_use {
1494 event_stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
1495 last_message
1496 .content
1497 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1498 } else {
1499 event_stream.update_tool_call_fields(
1500 &tool_use.id,
1501 acp::ToolCallUpdateFields {
1502 title: Some(title.into()),
1503 kind: Some(kind),
1504 raw_input: Some(tool_use.input.clone()),
1505 ..Default::default()
1506 },
1507 );
1508 }
1509
1510 if !tool_use.is_input_complete {
1511 return None;
1512 }
1513
1514 let Some(tool) = tool else {
1515 let content = format!("No tool named {} exists", tool_use.name);
1516 return Some(Task::ready(LanguageModelToolResult {
1517 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1518 tool_use_id: tool_use.id,
1519 tool_name: tool_use.name,
1520 is_error: true,
1521 output: None,
1522 }));
1523 };
1524
1525 let fs = self.project.read(cx).fs().clone();
1526 let tool_event_stream =
1527 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1528 tool_event_stream.update_fields(acp::ToolCallUpdateFields {
1529 status: Some(acp::ToolCallStatus::InProgress),
1530 ..Default::default()
1531 });
1532 let supports_images = self.model().is_some_and(|model| model.supports_images());
1533 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1534 log::debug!("Running tool {}", tool_use.name);
1535 Some(cx.foreground_executor().spawn(async move {
1536 let tool_result = tool_result.await.and_then(|output| {
1537 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1538 && !supports_images
1539 {
1540 return Err(anyhow!(
1541 "Attempted to read an image, but this model doesn't support it.",
1542 ));
1543 }
1544 Ok(output)
1545 });
1546
1547 match tool_result {
1548 Ok(output) => LanguageModelToolResult {
1549 tool_use_id: tool_use.id,
1550 tool_name: tool_use.name,
1551 is_error: false,
1552 content: output.llm_output,
1553 output: Some(output.raw_output),
1554 },
1555 Err(error) => LanguageModelToolResult {
1556 tool_use_id: tool_use.id,
1557 tool_name: tool_use.name,
1558 is_error: true,
1559 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1560 output: None,
1561 },
1562 }
1563 }))
1564 }
1565
1566 fn handle_tool_use_json_parse_error_event(
1567 &mut self,
1568 tool_use_id: LanguageModelToolUseId,
1569 tool_name: Arc<str>,
1570 raw_input: Arc<str>,
1571 json_parse_error: String,
1572 ) -> LanguageModelToolResult {
1573 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1574 LanguageModelToolResult {
1575 tool_use_id,
1576 tool_name,
1577 is_error: true,
1578 content: LanguageModelToolResultContent::Text(tool_output.into()),
1579 output: Some(serde_json::Value::String(raw_input.to_string())),
1580 }
1581 }
1582
1583 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1584 self.project
1585 .read(cx)
1586 .user_store()
1587 .update(cx, |user_store, cx| {
1588 user_store.update_model_request_usage(
1589 ModelRequestUsage(RequestUsage {
1590 amount: amount as i32,
1591 limit,
1592 }),
1593 cx,
1594 )
1595 });
1596 }
1597
1598 pub fn title(&self) -> SharedString {
1599 self.title.clone().unwrap_or("New Thread".into())
1600 }
1601
1602 pub fn summary(&mut self, cx: &mut Context<Self>) -> Task<Result<SharedString>> {
1603 if let Some(summary) = self.summary.as_ref() {
1604 return Task::ready(Ok(summary.clone()));
1605 }
1606 let Some(model) = self.summarization_model.clone() else {
1607 return Task::ready(Err(anyhow!("No summarization model available")));
1608 };
1609 let mut request = LanguageModelRequest {
1610 intent: Some(CompletionIntent::ThreadContextSummarization),
1611 temperature: AgentSettings::temperature_for_model(&model, cx),
1612 ..Default::default()
1613 };
1614
1615 for message in &self.messages {
1616 request.messages.extend(message.to_request());
1617 }
1618
1619 request.messages.push(LanguageModelRequestMessage {
1620 role: Role::User,
1621 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1622 cache: false,
1623 });
1624 cx.spawn(async move |this, cx| {
1625 let mut summary = String::new();
1626 let mut messages = model.stream_completion(request, cx).await?;
1627 while let Some(event) = messages.next().await {
1628 let event = event?;
1629 let text = match event {
1630 LanguageModelCompletionEvent::Text(text) => text,
1631 LanguageModelCompletionEvent::StatusUpdate(
1632 CompletionRequestStatus::UsageUpdated { amount, limit },
1633 ) => {
1634 this.update(cx, |thread, cx| {
1635 thread.update_model_request_usage(amount, limit, cx);
1636 })?;
1637 continue;
1638 }
1639 _ => continue,
1640 };
1641
1642 let mut lines = text.lines();
1643 summary.extend(lines.next());
1644 }
1645
1646 log::debug!("Setting summary: {}", summary);
1647 let summary = SharedString::from(summary);
1648
1649 this.update(cx, |this, cx| {
1650 this.summary = Some(summary.clone());
1651 cx.notify()
1652 })?;
1653
1654 Ok(summary)
1655 })
1656 }
1657
1658 fn generate_title(&mut self, cx: &mut Context<Self>) {
1659 let Some(model) = self.summarization_model.clone() else {
1660 return;
1661 };
1662
1663 log::debug!(
1664 "Generating title with model: {:?}",
1665 self.summarization_model.as_ref().map(|model| model.name())
1666 );
1667 let mut request = LanguageModelRequest {
1668 intent: Some(CompletionIntent::ThreadSummarization),
1669 temperature: AgentSettings::temperature_for_model(&model, cx),
1670 ..Default::default()
1671 };
1672
1673 for message in &self.messages {
1674 request.messages.extend(message.to_request());
1675 }
1676
1677 request.messages.push(LanguageModelRequestMessage {
1678 role: Role::User,
1679 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1680 cache: false,
1681 });
1682 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
1683 let mut title = String::new();
1684
1685 let generate = async {
1686 let mut messages = model.stream_completion(request, cx).await?;
1687 while let Some(event) = messages.next().await {
1688 let event = event?;
1689 let text = match event {
1690 LanguageModelCompletionEvent::Text(text) => text,
1691 LanguageModelCompletionEvent::StatusUpdate(
1692 CompletionRequestStatus::UsageUpdated { amount, limit },
1693 ) => {
1694 this.update(cx, |thread, cx| {
1695 thread.update_model_request_usage(amount, limit, cx);
1696 })?;
1697 continue;
1698 }
1699 _ => continue,
1700 };
1701
1702 let mut lines = text.lines();
1703 title.extend(lines.next());
1704
1705 // Stop if the LLM generated multiple lines.
1706 if lines.next().is_some() {
1707 break;
1708 }
1709 }
1710 anyhow::Ok(())
1711 };
1712
1713 if generate.await.context("failed to generate title").is_ok() {
1714 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
1715 }
1716 _ = this.update(cx, |this, _| this.pending_title_generation = None);
1717 }));
1718 }
1719
1720 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
1721 self.pending_title_generation = None;
1722 if Some(&title) != self.title.as_ref() {
1723 self.title = Some(title);
1724 cx.emit(TitleUpdated);
1725 cx.notify();
1726 }
1727 }
1728
1729 fn last_user_message(&self) -> Option<&UserMessage> {
1730 self.messages
1731 .iter()
1732 .rev()
1733 .find_map(|message| match message {
1734 Message::User(user_message) => Some(user_message),
1735 Message::Agent(_) => None,
1736 Message::Resume => None,
1737 })
1738 }
1739
1740 fn pending_message(&mut self) -> &mut AgentMessage {
1741 self.pending_message.get_or_insert_default()
1742 }
1743
1744 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1745 let Some(mut message) = self.pending_message.take() else {
1746 return;
1747 };
1748
1749 if message.content.is_empty() {
1750 return;
1751 }
1752
1753 for content in &message.content {
1754 let AgentMessageContent::ToolUse(tool_use) = content else {
1755 continue;
1756 };
1757
1758 if !message.tool_results.contains_key(&tool_use.id) {
1759 message.tool_results.insert(
1760 tool_use.id.clone(),
1761 LanguageModelToolResult {
1762 tool_use_id: tool_use.id.clone(),
1763 tool_name: tool_use.name.clone(),
1764 is_error: true,
1765 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1766 output: None,
1767 },
1768 );
1769 }
1770 }
1771
1772 self.messages.push(Message::Agent(message));
1773 self.updated_at = Utc::now();
1774 self.summary = None;
1775 cx.notify()
1776 }
1777
1778 pub(crate) fn build_completion_request(
1779 &self,
1780 completion_intent: CompletionIntent,
1781 cx: &App,
1782 ) -> Result<LanguageModelRequest> {
1783 let model = self.model().context("No language model configured")?;
1784 let tools = if let Some(turn) = self.running_turn.as_ref() {
1785 turn.tools
1786 .iter()
1787 .filter_map(|(tool_name, tool)| {
1788 log::trace!("Including tool: {}", tool_name);
1789 Some(LanguageModelRequestTool {
1790 name: tool_name.to_string(),
1791 description: tool.description().to_string(),
1792 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1793 })
1794 })
1795 .collect::<Vec<_>>()
1796 } else {
1797 Vec::new()
1798 };
1799
1800 log::debug!("Building completion request");
1801 log::debug!("Completion intent: {:?}", completion_intent);
1802 log::debug!("Completion mode: {:?}", self.completion_mode);
1803
1804 let messages = self.build_request_messages(cx);
1805 log::debug!("Request will include {} messages", messages.len());
1806 log::debug!("Request includes {} tools", tools.len());
1807
1808 let request = LanguageModelRequest {
1809 thread_id: Some(self.id.to_string()),
1810 prompt_id: Some(self.prompt_id.to_string()),
1811 intent: Some(completion_intent),
1812 mode: Some(self.completion_mode.into()),
1813 messages,
1814 tools,
1815 tool_choice: None,
1816 stop: Vec::new(),
1817 temperature: AgentSettings::temperature_for_model(model, cx),
1818 thinking_allowed: true,
1819 };
1820
1821 log::debug!("Completion request built successfully");
1822 Ok(request)
1823 }
1824
1825 fn enabled_tools(
1826 &self,
1827 profile: &AgentProfileSettings,
1828 model: &Arc<dyn LanguageModel>,
1829 cx: &App,
1830 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
1831 fn truncate(tool_name: &SharedString) -> SharedString {
1832 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
1833 let mut truncated = tool_name.to_string();
1834 truncated.truncate(MAX_TOOL_NAME_LENGTH);
1835 truncated.into()
1836 } else {
1837 tool_name.clone()
1838 }
1839 }
1840
1841 let mut tools = self
1842 .tools
1843 .iter()
1844 .filter_map(|(tool_name, tool)| {
1845 if tool.supported_provider(&model.provider_id())
1846 && profile.is_tool_enabled(tool_name)
1847 {
1848 Some((truncate(tool_name), tool.clone()))
1849 } else {
1850 None
1851 }
1852 })
1853 .collect::<BTreeMap<_, _>>();
1854
1855 let mut context_server_tools = Vec::new();
1856 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
1857 let mut duplicate_tool_names = HashSet::default();
1858 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
1859 for (tool_name, tool) in server_tools {
1860 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
1861 let tool_name = truncate(tool_name);
1862 if !seen_tools.insert(tool_name.clone()) {
1863 duplicate_tool_names.insert(tool_name.clone());
1864 }
1865 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
1866 }
1867 }
1868 }
1869
1870 // When there are duplicate tool names, disambiguate by prefixing them
1871 // with the server ID. In the rare case there isn't enough space for the
1872 // disambiguated tool name, keep only the last tool with this name.
1873 for (server_id, tool_name, tool) in context_server_tools {
1874 if duplicate_tool_names.contains(&tool_name) {
1875 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
1876 if available >= 2 {
1877 let mut disambiguated = server_id.0.to_string();
1878 disambiguated.truncate(available - 1);
1879 disambiguated.push('_');
1880 disambiguated.push_str(&tool_name);
1881 tools.insert(disambiguated.into(), tool.clone());
1882 } else {
1883 tools.insert(tool_name, tool.clone());
1884 }
1885 } else {
1886 tools.insert(tool_name, tool.clone());
1887 }
1888 }
1889
1890 tools
1891 }
1892
1893 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
1894 self.running_turn.as_ref()?.tools.get(name).cloned()
1895 }
1896
1897 fn build_request_messages(&self, cx: &App) -> Vec<LanguageModelRequestMessage> {
1898 log::trace!(
1899 "Building request messages from {} thread messages",
1900 self.messages.len()
1901 );
1902
1903 let system_prompt = SystemPromptTemplate {
1904 project: self.project_context.read(cx),
1905 available_tools: self.tools.keys().cloned().collect(),
1906 }
1907 .render(&self.templates)
1908 .context("failed to build system prompt")
1909 .expect("Invalid template");
1910 let mut messages = vec![LanguageModelRequestMessage {
1911 role: Role::System,
1912 content: vec![system_prompt.into()],
1913 cache: false,
1914 }];
1915 for message in &self.messages {
1916 messages.extend(message.to_request());
1917 }
1918
1919 if let Some(last_message) = messages.last_mut() {
1920 last_message.cache = true;
1921 }
1922
1923 if let Some(message) = self.pending_message.as_ref() {
1924 messages.extend(message.to_request());
1925 }
1926
1927 messages
1928 }
1929
1930 pub fn to_markdown(&self) -> String {
1931 let mut markdown = String::new();
1932 for (ix, message) in self.messages.iter().enumerate() {
1933 if ix > 0 {
1934 markdown.push('\n');
1935 }
1936 markdown.push_str(&message.to_markdown());
1937 }
1938
1939 if let Some(message) = self.pending_message.as_ref() {
1940 markdown.push('\n');
1941 markdown.push_str(&message.to_markdown());
1942 }
1943
1944 markdown
1945 }
1946
1947 fn advance_prompt_id(&mut self) {
1948 self.prompt_id = PromptId::new();
1949 }
1950
1951 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
1952 use LanguageModelCompletionError::*;
1953 use http_client::StatusCode;
1954
1955 // General strategy here:
1956 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
1957 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
1958 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
1959 match error {
1960 HttpResponseError {
1961 status_code: StatusCode::TOO_MANY_REQUESTS,
1962 ..
1963 } => Some(RetryStrategy::ExponentialBackoff {
1964 initial_delay: BASE_RETRY_DELAY,
1965 max_attempts: MAX_RETRY_ATTEMPTS,
1966 }),
1967 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
1968 Some(RetryStrategy::Fixed {
1969 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1970 max_attempts: MAX_RETRY_ATTEMPTS,
1971 })
1972 }
1973 UpstreamProviderError {
1974 status,
1975 retry_after,
1976 ..
1977 } => match *status {
1978 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
1979 Some(RetryStrategy::Fixed {
1980 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1981 max_attempts: MAX_RETRY_ATTEMPTS,
1982 })
1983 }
1984 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
1985 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1986 // Internal Server Error could be anything, retry up to 3 times.
1987 max_attempts: 3,
1988 }),
1989 status => {
1990 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
1991 // but we frequently get them in practice. See https://http.dev/529
1992 if status.as_u16() == 529 {
1993 Some(RetryStrategy::Fixed {
1994 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
1995 max_attempts: MAX_RETRY_ATTEMPTS,
1996 })
1997 } else {
1998 Some(RetryStrategy::Fixed {
1999 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2000 max_attempts: 2,
2001 })
2002 }
2003 }
2004 },
2005 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2006 delay: BASE_RETRY_DELAY,
2007 max_attempts: 3,
2008 }),
2009 ApiReadResponseError { .. }
2010 | HttpSend { .. }
2011 | DeserializeResponse { .. }
2012 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2013 delay: BASE_RETRY_DELAY,
2014 max_attempts: 3,
2015 }),
2016 // Retrying these errors definitely shouldn't help.
2017 HttpResponseError {
2018 status_code:
2019 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2020 ..
2021 }
2022 | AuthenticationError { .. }
2023 | PermissionError { .. }
2024 | NoApiKey { .. }
2025 | ApiEndpointNotFound { .. }
2026 | PromptTooLarge { .. } => None,
2027 // These errors might be transient, so retry them
2028 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2029 delay: BASE_RETRY_DELAY,
2030 max_attempts: 1,
2031 }),
2032 // Retry all other 4xx and 5xx errors once.
2033 HttpResponseError { status_code, .. }
2034 if status_code.is_client_error() || status_code.is_server_error() =>
2035 {
2036 Some(RetryStrategy::Fixed {
2037 delay: BASE_RETRY_DELAY,
2038 max_attempts: 3,
2039 })
2040 }
2041 Other(err)
2042 if err.is::<language_model::PaymentRequiredError>()
2043 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2044 {
2045 // Retrying won't help for Payment Required or Model Request Limit errors (where
2046 // the user must upgrade to usage-based billing to get more requests, or else wait
2047 // for a significant amount of time for the request limit to reset).
2048 None
2049 }
2050 // Conservatively assume that any other errors are non-retryable
2051 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2052 delay: BASE_RETRY_DELAY,
2053 max_attempts: 2,
2054 }),
2055 }
2056 }
2057}
2058
2059struct RunningTurn {
2060 /// Holds the task that handles agent interaction until the end of the turn.
2061 /// Survives across multiple requests as the model performs tool calls and
2062 /// we run tools, report their results.
2063 _task: Task<()>,
2064 /// The current event stream for the running turn. Used to report a final
2065 /// cancellation event if we cancel the turn.
2066 event_stream: ThreadEventStream,
2067 /// The tools that were enabled for this turn.
2068 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2069}
2070
2071impl RunningTurn {
2072 fn cancel(self) {
2073 log::debug!("Cancelling in progress turn");
2074 self.event_stream.send_canceled();
2075 }
2076}
2077
2078pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2079
2080impl EventEmitter<TokenUsageUpdated> for Thread {}
2081
2082pub struct TitleUpdated;
2083
2084impl EventEmitter<TitleUpdated> for Thread {}
2085
2086pub trait AgentTool
2087where
2088 Self: 'static + Sized,
2089{
2090 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2091 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2092
2093 fn name() -> &'static str;
2094
2095 fn description(&self) -> SharedString {
2096 let schema = schemars::schema_for!(Self::Input);
2097 SharedString::new(
2098 schema
2099 .get("description")
2100 .and_then(|description| description.as_str())
2101 .unwrap_or_default(),
2102 )
2103 }
2104
2105 fn kind() -> acp::ToolKind;
2106
2107 /// The initial tool title to display. Can be updated during the tool run.
2108 fn initial_title(&self, input: Result<Self::Input, serde_json::Value>) -> SharedString;
2109
2110 /// Returns the JSON schema that describes the tool's input.
2111 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Schema {
2112 crate::tool_schema::root_schema_for::<Self::Input>(format)
2113 }
2114
2115 /// Some tools rely on a provider for the underlying billing or other reasons.
2116 /// Allow the tool to check if they are compatible, or should be filtered out.
2117 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2118 true
2119 }
2120
2121 /// Runs the tool with the provided input.
2122 fn run(
2123 self: Arc<Self>,
2124 input: Self::Input,
2125 event_stream: ToolCallEventStream,
2126 cx: &mut App,
2127 ) -> Task<Result<Self::Output>>;
2128
2129 /// Emits events for a previous execution of the tool.
2130 fn replay(
2131 &self,
2132 _input: Self::Input,
2133 _output: Self::Output,
2134 _event_stream: ToolCallEventStream,
2135 _cx: &mut App,
2136 ) -> Result<()> {
2137 Ok(())
2138 }
2139
2140 fn erase(self) -> Arc<dyn AnyAgentTool> {
2141 Arc::new(Erased(Arc::new(self)))
2142 }
2143}
2144
2145pub struct Erased<T>(T);
2146
2147pub struct AgentToolOutput {
2148 pub llm_output: LanguageModelToolResultContent,
2149 pub raw_output: serde_json::Value,
2150}
2151
2152pub trait AnyAgentTool {
2153 fn name(&self) -> SharedString;
2154 fn description(&self) -> SharedString;
2155 fn kind(&self) -> acp::ToolKind;
2156 fn initial_title(&self, input: serde_json::Value) -> SharedString;
2157 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2158 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2159 true
2160 }
2161 fn run(
2162 self: Arc<Self>,
2163 input: serde_json::Value,
2164 event_stream: ToolCallEventStream,
2165 cx: &mut App,
2166 ) -> Task<Result<AgentToolOutput>>;
2167 fn replay(
2168 &self,
2169 input: serde_json::Value,
2170 output: serde_json::Value,
2171 event_stream: ToolCallEventStream,
2172 cx: &mut App,
2173 ) -> Result<()>;
2174}
2175
2176impl<T> AnyAgentTool for Erased<Arc<T>>
2177where
2178 T: AgentTool,
2179{
2180 fn name(&self) -> SharedString {
2181 T::name().into()
2182 }
2183
2184 fn description(&self) -> SharedString {
2185 self.0.description()
2186 }
2187
2188 fn kind(&self) -> agent_client_protocol::ToolKind {
2189 T::kind()
2190 }
2191
2192 fn initial_title(&self, input: serde_json::Value) -> SharedString {
2193 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2194 self.0.initial_title(parsed_input)
2195 }
2196
2197 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2198 let mut json = serde_json::to_value(self.0.input_schema(format))?;
2199 adapt_schema_to_format(&mut json, format)?;
2200 Ok(json)
2201 }
2202
2203 fn supported_provider(&self, provider: &LanguageModelProviderId) -> bool {
2204 self.0.supported_provider(provider)
2205 }
2206
2207 fn run(
2208 self: Arc<Self>,
2209 input: serde_json::Value,
2210 event_stream: ToolCallEventStream,
2211 cx: &mut App,
2212 ) -> Task<Result<AgentToolOutput>> {
2213 cx.spawn(async move |cx| {
2214 let input = serde_json::from_value(input)?;
2215 let output = cx
2216 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2217 .await?;
2218 let raw_output = serde_json::to_value(&output)?;
2219 Ok(AgentToolOutput {
2220 llm_output: output.into(),
2221 raw_output,
2222 })
2223 })
2224 }
2225
2226 fn replay(
2227 &self,
2228 input: serde_json::Value,
2229 output: serde_json::Value,
2230 event_stream: ToolCallEventStream,
2231 cx: &mut App,
2232 ) -> Result<()> {
2233 let input = serde_json::from_value(input)?;
2234 let output = serde_json::from_value(output)?;
2235 self.0.replay(input, output, event_stream, cx)
2236 }
2237}
2238
2239#[derive(Clone)]
2240struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2241
2242impl ThreadEventStream {
2243 fn send_user_message(&self, message: &UserMessage) {
2244 self.0
2245 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2246 .ok();
2247 }
2248
2249 fn send_text(&self, text: &str) {
2250 self.0
2251 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2252 .ok();
2253 }
2254
2255 fn send_thinking(&self, text: &str) {
2256 self.0
2257 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2258 .ok();
2259 }
2260
2261 fn send_tool_call(
2262 &self,
2263 id: &LanguageModelToolUseId,
2264 title: SharedString,
2265 kind: acp::ToolKind,
2266 input: serde_json::Value,
2267 ) {
2268 self.0
2269 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2270 id,
2271 title.to_string(),
2272 kind,
2273 input,
2274 ))))
2275 .ok();
2276 }
2277
2278 fn initial_tool_call(
2279 id: &LanguageModelToolUseId,
2280 title: String,
2281 kind: acp::ToolKind,
2282 input: serde_json::Value,
2283 ) -> acp::ToolCall {
2284 acp::ToolCall {
2285 id: acp::ToolCallId(id.to_string().into()),
2286 title,
2287 kind,
2288 status: acp::ToolCallStatus::Pending,
2289 content: vec![],
2290 locations: vec![],
2291 raw_input: Some(input),
2292 raw_output: None,
2293 }
2294 }
2295
2296 fn update_tool_call_fields(
2297 &self,
2298 tool_use_id: &LanguageModelToolUseId,
2299 fields: acp::ToolCallUpdateFields,
2300 ) {
2301 self.0
2302 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2303 acp::ToolCallUpdate {
2304 id: acp::ToolCallId(tool_use_id.to_string().into()),
2305 fields,
2306 }
2307 .into(),
2308 )))
2309 .ok();
2310 }
2311
2312 fn send_retry(&self, status: acp_thread::RetryStatus) {
2313 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2314 }
2315
2316 fn send_stop(&self, reason: acp::StopReason) {
2317 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2318 }
2319
2320 fn send_canceled(&self) {
2321 self.0
2322 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2323 .ok();
2324 }
2325
2326 fn send_error(&self, error: impl Into<anyhow::Error>) {
2327 self.0.unbounded_send(Err(error.into())).ok();
2328 }
2329}
2330
2331#[derive(Clone)]
2332pub struct ToolCallEventStream {
2333 tool_use_id: LanguageModelToolUseId,
2334 stream: ThreadEventStream,
2335 fs: Option<Arc<dyn Fs>>,
2336}
2337
2338impl ToolCallEventStream {
2339 #[cfg(test)]
2340 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2341 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2342
2343 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2344
2345 (stream, ToolCallEventStreamReceiver(events_rx))
2346 }
2347
2348 fn new(
2349 tool_use_id: LanguageModelToolUseId,
2350 stream: ThreadEventStream,
2351 fs: Option<Arc<dyn Fs>>,
2352 ) -> Self {
2353 Self {
2354 tool_use_id,
2355 stream,
2356 fs,
2357 }
2358 }
2359
2360 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2361 self.stream
2362 .update_tool_call_fields(&self.tool_use_id, fields);
2363 }
2364
2365 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2366 self.stream
2367 .0
2368 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2369 acp_thread::ToolCallUpdateDiff {
2370 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2371 diff,
2372 }
2373 .into(),
2374 )))
2375 .ok();
2376 }
2377
2378 pub fn update_terminal(&self, terminal: Entity<acp_thread::Terminal>) {
2379 self.stream
2380 .0
2381 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2382 acp_thread::ToolCallUpdateTerminal {
2383 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2384 terminal,
2385 }
2386 .into(),
2387 )))
2388 .ok();
2389 }
2390
2391 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2392 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2393 return Task::ready(Ok(()));
2394 }
2395
2396 let (response_tx, response_rx) = oneshot::channel();
2397 self.stream
2398 .0
2399 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2400 ToolCallAuthorization {
2401 tool_call: acp::ToolCallUpdate {
2402 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2403 fields: acp::ToolCallUpdateFields {
2404 title: Some(title.into()),
2405 ..Default::default()
2406 },
2407 },
2408 options: vec![
2409 acp::PermissionOption {
2410 id: acp::PermissionOptionId("always_allow".into()),
2411 name: "Always Allow".into(),
2412 kind: acp::PermissionOptionKind::AllowAlways,
2413 },
2414 acp::PermissionOption {
2415 id: acp::PermissionOptionId("allow".into()),
2416 name: "Allow".into(),
2417 kind: acp::PermissionOptionKind::AllowOnce,
2418 },
2419 acp::PermissionOption {
2420 id: acp::PermissionOptionId("deny".into()),
2421 name: "Deny".into(),
2422 kind: acp::PermissionOptionKind::RejectOnce,
2423 },
2424 ],
2425 response: response_tx,
2426 },
2427 )))
2428 .ok();
2429 let fs = self.fs.clone();
2430 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2431 "always_allow" => {
2432 if let Some(fs) = fs.clone() {
2433 cx.update(|cx| {
2434 update_settings_file::<AgentSettings>(fs, cx, |settings, _| {
2435 settings.set_always_allow_tool_actions(true);
2436 });
2437 })?;
2438 }
2439
2440 Ok(())
2441 }
2442 "allow" => Ok(()),
2443 _ => Err(anyhow!("Permission to run tool denied by user")),
2444 })
2445 }
2446}
2447
2448#[cfg(test)]
2449pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2450
2451#[cfg(test)]
2452impl ToolCallEventStreamReceiver {
2453 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2454 let event = self.0.next().await;
2455 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2456 auth
2457 } else {
2458 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2459 }
2460 }
2461
2462 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2463 let event = self.0.next().await;
2464 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2465 update,
2466 )))) = event
2467 {
2468 update.terminal
2469 } else {
2470 panic!("Expected terminal but got: {:?}", event);
2471 }
2472 }
2473}
2474
2475#[cfg(test)]
2476impl std::ops::Deref for ToolCallEventStreamReceiver {
2477 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2478
2479 fn deref(&self) -> &Self::Target {
2480 &self.0
2481 }
2482}
2483
2484#[cfg(test)]
2485impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2486 fn deref_mut(&mut self) -> &mut Self::Target {
2487 &mut self.0
2488 }
2489}
2490
2491impl From<&str> for UserMessageContent {
2492 fn from(text: &str) -> Self {
2493 Self::Text(text.into())
2494 }
2495}
2496
2497impl From<acp::ContentBlock> for UserMessageContent {
2498 fn from(value: acp::ContentBlock) -> Self {
2499 match value {
2500 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2501 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2502 acp::ContentBlock::Audio(_) => {
2503 // TODO
2504 Self::Text("[audio]".to_string())
2505 }
2506 acp::ContentBlock::ResourceLink(resource_link) => {
2507 match MentionUri::parse(&resource_link.uri) {
2508 Ok(uri) => Self::Mention {
2509 uri,
2510 content: String::new(),
2511 },
2512 Err(err) => {
2513 log::error!("Failed to parse mention link: {}", err);
2514 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2515 }
2516 }
2517 }
2518 acp::ContentBlock::Resource(resource) => match resource.resource {
2519 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2520 match MentionUri::parse(&resource.uri) {
2521 Ok(uri) => Self::Mention {
2522 uri,
2523 content: resource.text,
2524 },
2525 Err(err) => {
2526 log::error!("Failed to parse mention link: {}", err);
2527 Self::Text(
2528 MarkdownCodeBlock {
2529 tag: &resource.uri,
2530 text: &resource.text,
2531 }
2532 .to_string(),
2533 )
2534 }
2535 }
2536 }
2537 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2538 // TODO
2539 Self::Text("[blob]".to_string())
2540 }
2541 },
2542 }
2543 }
2544}
2545
2546impl From<UserMessageContent> for acp::ContentBlock {
2547 fn from(content: UserMessageContent) -> Self {
2548 match content {
2549 UserMessageContent::Text(text) => acp::ContentBlock::Text(acp::TextContent {
2550 text,
2551 annotations: None,
2552 }),
2553 UserMessageContent::Image(image) => acp::ContentBlock::Image(acp::ImageContent {
2554 data: image.source.to_string(),
2555 mime_type: "image/png".to_string(),
2556 annotations: None,
2557 uri: None,
2558 }),
2559 UserMessageContent::Mention { uri, content } => {
2560 acp::ContentBlock::Resource(acp::EmbeddedResource {
2561 resource: acp::EmbeddedResourceResource::TextResourceContents(
2562 acp::TextResourceContents {
2563 mime_type: None,
2564 text: content,
2565 uri: uri.to_uri().to_string(),
2566 },
2567 ),
2568 annotations: None,
2569 })
2570 }
2571 }
2572 }
2573}
2574
2575fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2576 LanguageModelImage {
2577 source: image_content.data.into(),
2578 // TODO: make this optional?
2579 size: gpui::Size::new(0.into(), 0.into()),
2580 }
2581}