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, UserStore};
19use cloud_llm_client::{CompletionIntent, CompletionRequestStatus, Plan, UsageLimit};
20use collections::{HashMap, HashSet, IndexMap};
21use fs::Fs;
22use futures::stream;
23use futures::{
24 FutureExt,
25 channel::{mpsc, oneshot},
26 future::Shared,
27 stream::FuturesUnordered,
28};
29use git::repository::DiffType;
30use gpui::{
31 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
32};
33use language_model::{
34 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelExt,
35 LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry, LanguageModelRequest,
36 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
37 LanguageModelToolResultContent, LanguageModelToolSchemaFormat, LanguageModelToolUse,
38 LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage, ZED_CLOUD_PROVIDER_ID,
39};
40use project::{
41 Project,
42 git_store::{GitStore, RepositoryState},
43};
44use prompt_store::ProjectContext;
45use schemars::{JsonSchema, Schema};
46use serde::{Deserialize, Serialize};
47use settings::{Settings, update_settings_file};
48use smol::stream::StreamExt;
49use std::{
50 collections::BTreeMap,
51 ops::RangeInclusive,
52 path::Path,
53 rc::Rc,
54 sync::Arc,
55 time::{Duration, Instant},
56};
57use std::{fmt::Write, path::PathBuf};
58use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock};
59use uuid::Uuid;
60
61const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
62pub const MAX_TOOL_NAME_LENGTH: usize = 64;
63
64/// The ID of the user prompt that initiated a request.
65///
66/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
67#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
68pub struct PromptId(Arc<str>);
69
70impl PromptId {
71 pub fn new() -> Self {
72 Self(Uuid::new_v4().to_string().into())
73 }
74}
75
76impl std::fmt::Display for PromptId {
77 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
78 write!(f, "{}", self.0)
79 }
80}
81
82pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
83pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
84
85#[derive(Debug, Clone)]
86enum RetryStrategy {
87 ExponentialBackoff {
88 initial_delay: Duration,
89 max_attempts: u8,
90 },
91 Fixed {
92 delay: Duration,
93 max_attempts: u8,
94 },
95}
96
97#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
98pub enum Message {
99 User(UserMessage),
100 Agent(AgentMessage),
101 Resume,
102}
103
104impl Message {
105 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
106 match self {
107 Message::Agent(agent_message) => Some(agent_message),
108 _ => None,
109 }
110 }
111
112 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
113 match self {
114 Message::User(message) => vec![message.to_request()],
115 Message::Agent(message) => message.to_request(),
116 Message::Resume => vec![LanguageModelRequestMessage {
117 role: Role::User,
118 content: vec!["Continue where you left off".into()],
119 cache: false,
120 }],
121 }
122 }
123
124 pub fn to_markdown(&self) -> String {
125 match self {
126 Message::User(message) => message.to_markdown(),
127 Message::Agent(message) => message.to_markdown(),
128 Message::Resume => "[resume]\n".into(),
129 }
130 }
131
132 pub fn role(&self) -> Role {
133 match self {
134 Message::User(_) | Message::Resume => Role::User,
135 Message::Agent(_) => Role::Assistant,
136 }
137 }
138}
139
140#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
141pub struct UserMessage {
142 pub id: UserMessageId,
143 pub content: Vec<UserMessageContent>,
144}
145
146#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
147pub enum UserMessageContent {
148 Text(String),
149 Mention { uri: MentionUri, content: String },
150 Image(LanguageModelImage),
151}
152
153impl UserMessage {
154 pub fn to_markdown(&self) -> String {
155 let mut markdown = String::from("## User\n\n");
156
157 for content in &self.content {
158 match content {
159 UserMessageContent::Text(text) => {
160 markdown.push_str(text);
161 markdown.push('\n');
162 }
163 UserMessageContent::Image(_) => {
164 markdown.push_str("<image />\n");
165 }
166 UserMessageContent::Mention { uri, content } => {
167 if !content.is_empty() {
168 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
169 } else {
170 let _ = writeln!(&mut markdown, "{}", uri.as_link());
171 }
172 }
173 }
174 }
175
176 markdown
177 }
178
179 fn to_request(&self) -> LanguageModelRequestMessage {
180 let mut message = LanguageModelRequestMessage {
181 role: Role::User,
182 content: Vec::with_capacity(self.content.len()),
183 cache: false,
184 };
185
186 const OPEN_CONTEXT: &str = "<context>\n\
187 The following items were attached by the user. \
188 They are up-to-date and don't need to be re-read.\n\n";
189
190 const OPEN_FILES_TAG: &str = "<files>";
191 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
192 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
193 const OPEN_SELECTIONS_TAG: &str = "<selections>";
194 const OPEN_THREADS_TAG: &str = "<threads>";
195 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
196 const OPEN_RULES_TAG: &str =
197 "<rules>\nThe user has specified the following rules that should be applied:\n";
198
199 let mut file_context = OPEN_FILES_TAG.to_string();
200 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
201 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
202 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
203 let mut thread_context = OPEN_THREADS_TAG.to_string();
204 let mut fetch_context = OPEN_FETCH_TAG.to_string();
205 let mut rules_context = OPEN_RULES_TAG.to_string();
206
207 for chunk in &self.content {
208 let chunk = match chunk {
209 UserMessageContent::Text(text) => {
210 language_model::MessageContent::Text(text.clone())
211 }
212 UserMessageContent::Image(value) => {
213 language_model::MessageContent::Image(value.clone())
214 }
215 UserMessageContent::Mention { uri, content } => {
216 match uri {
217 MentionUri::File { abs_path } => {
218 write!(
219 &mut file_context,
220 "\n{}",
221 MarkdownCodeBlock {
222 tag: &codeblock_tag(abs_path, None),
223 text: &content.to_string(),
224 }
225 )
226 .ok();
227 }
228 MentionUri::PastedImage => {
229 debug_panic!("pasted image URI should not be used in mention content")
230 }
231 MentionUri::Directory { .. } => {
232 write!(&mut directory_context, "\n{}\n", content).ok();
233 }
234 MentionUri::Symbol {
235 abs_path: path,
236 line_range,
237 ..
238 } => {
239 write!(
240 &mut symbol_context,
241 "\n{}",
242 MarkdownCodeBlock {
243 tag: &codeblock_tag(path, Some(line_range)),
244 text: content
245 }
246 )
247 .ok();
248 }
249 MentionUri::Selection {
250 abs_path: path,
251 line_range,
252 ..
253 } => {
254 write!(
255 &mut selection_context,
256 "\n{}",
257 MarkdownCodeBlock {
258 tag: &codeblock_tag(
259 path.as_deref().unwrap_or("Untitled".as_ref()),
260 Some(line_range)
261 ),
262 text: content
263 }
264 )
265 .ok();
266 }
267 MentionUri::Thread { .. } => {
268 write!(&mut thread_context, "\n{}\n", content).ok();
269 }
270 MentionUri::TextThread { .. } => {
271 write!(&mut thread_context, "\n{}\n", content).ok();
272 }
273 MentionUri::Rule { .. } => {
274 write!(
275 &mut rules_context,
276 "\n{}",
277 MarkdownCodeBlock {
278 tag: "",
279 text: content
280 }
281 )
282 .ok();
283 }
284 MentionUri::Fetch { url } => {
285 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
286 }
287 }
288
289 language_model::MessageContent::Text(uri.as_link().to_string())
290 }
291 };
292
293 message.content.push(chunk);
294 }
295
296 let len_before_context = message.content.len();
297
298 if file_context.len() > OPEN_FILES_TAG.len() {
299 file_context.push_str("</files>\n");
300 message
301 .content
302 .push(language_model::MessageContent::Text(file_context));
303 }
304
305 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
306 directory_context.push_str("</directories>\n");
307 message
308 .content
309 .push(language_model::MessageContent::Text(directory_context));
310 }
311
312 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
313 symbol_context.push_str("</symbols>\n");
314 message
315 .content
316 .push(language_model::MessageContent::Text(symbol_context));
317 }
318
319 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
320 selection_context.push_str("</selections>\n");
321 message
322 .content
323 .push(language_model::MessageContent::Text(selection_context));
324 }
325
326 if thread_context.len() > OPEN_THREADS_TAG.len() {
327 thread_context.push_str("</threads>\n");
328 message
329 .content
330 .push(language_model::MessageContent::Text(thread_context));
331 }
332
333 if fetch_context.len() > OPEN_FETCH_TAG.len() {
334 fetch_context.push_str("</fetched_urls>\n");
335 message
336 .content
337 .push(language_model::MessageContent::Text(fetch_context));
338 }
339
340 if rules_context.len() > OPEN_RULES_TAG.len() {
341 rules_context.push_str("</user_rules>\n");
342 message
343 .content
344 .push(language_model::MessageContent::Text(rules_context));
345 }
346
347 if message.content.len() > len_before_context {
348 message.content.insert(
349 len_before_context,
350 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
351 );
352 message
353 .content
354 .push(language_model::MessageContent::Text("</context>".into()));
355 }
356
357 message
358 }
359}
360
361fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
362 let mut result = String::new();
363
364 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
365 let _ = write!(result, "{} ", extension);
366 }
367
368 let _ = write!(result, "{}", full_path.display());
369
370 if let Some(range) = line_range {
371 if range.start() == range.end() {
372 let _ = write!(result, ":{}", range.start() + 1);
373 } else {
374 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
375 }
376 }
377
378 result
379}
380
381impl AgentMessage {
382 pub fn to_markdown(&self) -> String {
383 let mut markdown = String::from("## Assistant\n\n");
384
385 for content in &self.content {
386 match content {
387 AgentMessageContent::Text(text) => {
388 markdown.push_str(text);
389 markdown.push('\n');
390 }
391 AgentMessageContent::Thinking { text, .. } => {
392 markdown.push_str("<think>");
393 markdown.push_str(text);
394 markdown.push_str("</think>\n");
395 }
396 AgentMessageContent::RedactedThinking(_) => {
397 markdown.push_str("<redacted_thinking />\n")
398 }
399 AgentMessageContent::ToolUse(tool_use) => {
400 markdown.push_str(&format!(
401 "**Tool Use**: {} (ID: {})\n",
402 tool_use.name, tool_use.id
403 ));
404 markdown.push_str(&format!(
405 "{}\n",
406 MarkdownCodeBlock {
407 tag: "json",
408 text: &format!("{:#}", tool_use.input)
409 }
410 ));
411 }
412 }
413 }
414
415 for tool_result in self.tool_results.values() {
416 markdown.push_str(&format!(
417 "**Tool Result**: {} (ID: {})\n\n",
418 tool_result.tool_name, tool_result.tool_use_id
419 ));
420 if tool_result.is_error {
421 markdown.push_str("**ERROR:**\n");
422 }
423
424 match &tool_result.content {
425 LanguageModelToolResultContent::Text(text) => {
426 writeln!(markdown, "{text}\n").ok();
427 }
428 LanguageModelToolResultContent::Image(_) => {
429 writeln!(markdown, "<image />\n").ok();
430 }
431 }
432
433 if let Some(output) = tool_result.output.as_ref() {
434 writeln!(
435 markdown,
436 "**Debug Output**:\n\n```json\n{}\n```\n",
437 serde_json::to_string_pretty(output).unwrap()
438 )
439 .unwrap();
440 }
441 }
442
443 markdown
444 }
445
446 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
447 let mut assistant_message = LanguageModelRequestMessage {
448 role: Role::Assistant,
449 content: Vec::with_capacity(self.content.len()),
450 cache: false,
451 };
452 for chunk in &self.content {
453 match chunk {
454 AgentMessageContent::Text(text) => {
455 assistant_message
456 .content
457 .push(language_model::MessageContent::Text(text.clone()));
458 }
459 AgentMessageContent::Thinking { text, signature } => {
460 assistant_message
461 .content
462 .push(language_model::MessageContent::Thinking {
463 text: text.clone(),
464 signature: signature.clone(),
465 });
466 }
467 AgentMessageContent::RedactedThinking(value) => {
468 assistant_message.content.push(
469 language_model::MessageContent::RedactedThinking(value.clone()),
470 );
471 }
472 AgentMessageContent::ToolUse(tool_use) => {
473 if self.tool_results.contains_key(&tool_use.id) {
474 assistant_message
475 .content
476 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
477 }
478 }
479 };
480 }
481
482 let mut user_message = LanguageModelRequestMessage {
483 role: Role::User,
484 content: Vec::new(),
485 cache: false,
486 };
487
488 for tool_result in self.tool_results.values() {
489 let mut tool_result = tool_result.clone();
490 // Surprisingly, the API fails if we return an empty string here.
491 // It thinks we are sending a tool use without a tool result.
492 if tool_result.content.is_empty() {
493 tool_result.content = "<Tool returned an empty string>".into();
494 }
495 user_message
496 .content
497 .push(language_model::MessageContent::ToolResult(tool_result));
498 }
499
500 let mut messages = Vec::new();
501 if !assistant_message.content.is_empty() {
502 messages.push(assistant_message);
503 }
504 if !user_message.content.is_empty() {
505 messages.push(user_message);
506 }
507 messages
508 }
509}
510
511#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
512pub struct AgentMessage {
513 pub content: Vec<AgentMessageContent>,
514 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
515}
516
517#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
518pub enum AgentMessageContent {
519 Text(String),
520 Thinking {
521 text: String,
522 signature: Option<String>,
523 },
524 RedactedThinking(String),
525 ToolUse(LanguageModelToolUse),
526}
527
528pub trait TerminalHandle {
529 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
530 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
531 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
532}
533
534pub trait ThreadEnvironment {
535 fn create_terminal(
536 &self,
537 command: String,
538 cwd: Option<PathBuf>,
539 output_byte_limit: Option<u64>,
540 cx: &mut AsyncApp,
541 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
542}
543
544#[derive(Debug)]
545pub enum ThreadEvent {
546 UserMessage(UserMessage),
547 AgentText(String),
548 AgentThinking(String),
549 ToolCall(acp::ToolCall),
550 ToolCallUpdate(acp_thread::ToolCallUpdate),
551 ToolCallAuthorization(ToolCallAuthorization),
552 Retry(acp_thread::RetryStatus),
553 Stop(acp::StopReason),
554}
555
556#[derive(Debug)]
557pub struct NewTerminal {
558 pub command: String,
559 pub output_byte_limit: Option<u64>,
560 pub cwd: Option<PathBuf>,
561 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
562}
563
564#[derive(Debug)]
565pub struct ToolCallAuthorization {
566 pub tool_call: acp::ToolCallUpdate,
567 pub options: Vec<acp::PermissionOption>,
568 pub response: oneshot::Sender<acp::PermissionOptionId>,
569}
570
571#[derive(Debug, thiserror::Error)]
572enum CompletionError {
573 #[error("max tokens")]
574 MaxTokens,
575 #[error("refusal")]
576 Refusal,
577 #[error(transparent)]
578 Other(#[from] anyhow::Error),
579}
580
581pub struct Thread {
582 id: acp::SessionId,
583 prompt_id: PromptId,
584 updated_at: DateTime<Utc>,
585 title: Option<SharedString>,
586 pending_title_generation: Option<Task<()>>,
587 summary: Option<SharedString>,
588 messages: Vec<Message>,
589 user_store: Entity<UserStore>,
590 completion_mode: CompletionMode,
591 /// Holds the task that handles agent interaction until the end of the turn.
592 /// Survives across multiple requests as the model performs tool calls and
593 /// we run tools, report their results.
594 running_turn: Option<RunningTurn>,
595 pending_message: Option<AgentMessage>,
596 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
597 tool_use_limit_reached: bool,
598 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
599 #[allow(unused)]
600 cumulative_token_usage: TokenUsage,
601 #[allow(unused)]
602 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
603 context_server_registry: Entity<ContextServerRegistry>,
604 profile_id: AgentProfileId,
605 project_context: Entity<ProjectContext>,
606 templates: Arc<Templates>,
607 model: Option<Arc<dyn LanguageModel>>,
608 summarization_model: Option<Arc<dyn LanguageModel>>,
609 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
610 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
611 pub(crate) project: Entity<Project>,
612 pub(crate) action_log: Entity<ActionLog>,
613}
614
615impl Thread {
616 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
617 let image = model.map_or(true, |model| model.supports_images());
618 acp::PromptCapabilities {
619 meta: None,
620 image,
621 audio: false,
622 embedded_context: true,
623 }
624 }
625
626 pub fn new(
627 project: Entity<Project>,
628 project_context: Entity<ProjectContext>,
629 context_server_registry: Entity<ContextServerRegistry>,
630 templates: Arc<Templates>,
631 model: Option<Arc<dyn LanguageModel>>,
632 cx: &mut Context<Self>,
633 ) -> Self {
634 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
635 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
636 let (prompt_capabilities_tx, prompt_capabilities_rx) =
637 watch::channel(Self::prompt_capabilities(model.as_deref()));
638 Self {
639 id: acp::SessionId(uuid::Uuid::new_v4().to_string().into()),
640 prompt_id: PromptId::new(),
641 updated_at: Utc::now(),
642 title: None,
643 pending_title_generation: None,
644 summary: None,
645 messages: Vec::new(),
646 user_store: project.read(cx).user_store(),
647 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
648 running_turn: None,
649 pending_message: None,
650 tools: BTreeMap::default(),
651 tool_use_limit_reached: false,
652 request_token_usage: HashMap::default(),
653 cumulative_token_usage: TokenUsage::default(),
654 initial_project_snapshot: {
655 let project_snapshot = Self::project_snapshot(project.clone(), cx);
656 cx.foreground_executor()
657 .spawn(async move { Some(project_snapshot.await) })
658 .shared()
659 },
660 context_server_registry,
661 profile_id,
662 project_context,
663 templates,
664 model,
665 summarization_model: None,
666 prompt_capabilities_tx,
667 prompt_capabilities_rx,
668 project,
669 action_log,
670 }
671 }
672
673 pub fn id(&self) -> &acp::SessionId {
674 &self.id
675 }
676
677 pub fn replay(
678 &mut self,
679 cx: &mut Context<Self>,
680 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
681 let (tx, rx) = mpsc::unbounded();
682 let stream = ThreadEventStream(tx);
683 for message in &self.messages {
684 match message {
685 Message::User(user_message) => stream.send_user_message(user_message),
686 Message::Agent(assistant_message) => {
687 for content in &assistant_message.content {
688 match content {
689 AgentMessageContent::Text(text) => stream.send_text(text),
690 AgentMessageContent::Thinking { text, .. } => {
691 stream.send_thinking(text)
692 }
693 AgentMessageContent::RedactedThinking(_) => {}
694 AgentMessageContent::ToolUse(tool_use) => {
695 self.replay_tool_call(
696 tool_use,
697 assistant_message.tool_results.get(&tool_use.id),
698 &stream,
699 cx,
700 );
701 }
702 }
703 }
704 }
705 Message::Resume => {}
706 }
707 }
708 rx
709 }
710
711 fn replay_tool_call(
712 &self,
713 tool_use: &LanguageModelToolUse,
714 tool_result: Option<&LanguageModelToolResult>,
715 stream: &ThreadEventStream,
716 cx: &mut Context<Self>,
717 ) {
718 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
719 self.context_server_registry
720 .read(cx)
721 .servers()
722 .find_map(|(_, tools)| {
723 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
724 Some(tool.clone())
725 } else {
726 None
727 }
728 })
729 });
730
731 let Some(tool) = tool else {
732 stream
733 .0
734 .unbounded_send(Ok(ThreadEvent::ToolCall(acp::ToolCall {
735 meta: None,
736 id: acp::ToolCallId(tool_use.id.to_string().into()),
737 title: tool_use.name.to_string(),
738 kind: acp::ToolKind::Other,
739 status: acp::ToolCallStatus::Failed,
740 content: Vec::new(),
741 locations: Vec::new(),
742 raw_input: Some(tool_use.input.clone()),
743 raw_output: None,
744 })))
745 .ok();
746 return;
747 };
748
749 let title = tool.initial_title(tool_use.input.clone(), cx);
750 let kind = tool.kind();
751 stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
752
753 let output = tool_result
754 .as_ref()
755 .and_then(|result| result.output.clone());
756 if let Some(output) = output.clone() {
757 let tool_event_stream = ToolCallEventStream::new(
758 tool_use.id.clone(),
759 stream.clone(),
760 Some(self.project.read(cx).fs().clone()),
761 );
762 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
763 .log_err();
764 }
765
766 stream.update_tool_call_fields(
767 &tool_use.id,
768 acp::ToolCallUpdateFields {
769 status: Some(
770 tool_result
771 .as_ref()
772 .map_or(acp::ToolCallStatus::Failed, |result| {
773 if result.is_error {
774 acp::ToolCallStatus::Failed
775 } else {
776 acp::ToolCallStatus::Completed
777 }
778 }),
779 ),
780 raw_output: output,
781 ..Default::default()
782 },
783 );
784 }
785
786 pub fn from_db(
787 id: acp::SessionId,
788 db_thread: DbThread,
789 project: Entity<Project>,
790 project_context: Entity<ProjectContext>,
791 context_server_registry: Entity<ContextServerRegistry>,
792 action_log: Entity<ActionLog>,
793 templates: Arc<Templates>,
794 cx: &mut Context<Self>,
795 ) -> Self {
796 let profile_id = db_thread
797 .profile
798 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
799 let model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
800 db_thread
801 .model
802 .and_then(|model| {
803 let model = SelectedModel {
804 provider: model.provider.clone().into(),
805 model: model.model.into(),
806 };
807 registry.select_model(&model, cx)
808 })
809 .or_else(|| registry.default_model())
810 .map(|model| model.model)
811 });
812 let (prompt_capabilities_tx, prompt_capabilities_rx) =
813 watch::channel(Self::prompt_capabilities(model.as_deref()));
814
815 Self {
816 id,
817 prompt_id: PromptId::new(),
818 title: if db_thread.title.is_empty() {
819 None
820 } else {
821 Some(db_thread.title.clone())
822 },
823 pending_title_generation: None,
824 summary: db_thread.detailed_summary,
825 messages: db_thread.messages,
826 user_store: project.read(cx).user_store(),
827 completion_mode: db_thread.completion_mode.unwrap_or_default(),
828 running_turn: None,
829 pending_message: None,
830 tools: BTreeMap::default(),
831 tool_use_limit_reached: false,
832 request_token_usage: db_thread.request_token_usage.clone(),
833 cumulative_token_usage: db_thread.cumulative_token_usage,
834 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
835 context_server_registry,
836 profile_id,
837 project_context,
838 templates,
839 model,
840 summarization_model: None,
841 project,
842 action_log,
843 updated_at: db_thread.updated_at,
844 prompt_capabilities_tx,
845 prompt_capabilities_rx,
846 }
847 }
848
849 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
850 let initial_project_snapshot = self.initial_project_snapshot.clone();
851 let mut thread = DbThread {
852 title: self.title(),
853 messages: self.messages.clone(),
854 updated_at: self.updated_at,
855 detailed_summary: self.summary.clone(),
856 initial_project_snapshot: None,
857 cumulative_token_usage: self.cumulative_token_usage,
858 request_token_usage: self.request_token_usage.clone(),
859 model: self.model.as_ref().map(|model| DbLanguageModel {
860 provider: model.provider_id().to_string(),
861 model: model.name().0.to_string(),
862 }),
863 completion_mode: Some(self.completion_mode),
864 profile: Some(self.profile_id.clone()),
865 };
866
867 cx.background_spawn(async move {
868 let initial_project_snapshot = initial_project_snapshot.await;
869 thread.initial_project_snapshot = initial_project_snapshot;
870 thread
871 })
872 }
873
874 /// Create a snapshot of the current project state including git information and unsaved buffers.
875 fn project_snapshot(
876 project: Entity<Project>,
877 cx: &mut Context<Self>,
878 ) -> Task<Arc<agent::thread::ProjectSnapshot>> {
879 let git_store = project.read(cx).git_store().clone();
880 let worktree_snapshots: Vec<_> = project
881 .read(cx)
882 .visible_worktrees(cx)
883 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
884 .collect();
885
886 cx.spawn(async move |_, _| {
887 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
888
889 Arc::new(ProjectSnapshot {
890 worktree_snapshots,
891 timestamp: Utc::now(),
892 })
893 })
894 }
895
896 fn worktree_snapshot(
897 worktree: Entity<project::Worktree>,
898 git_store: Entity<GitStore>,
899 cx: &App,
900 ) -> Task<agent::thread::WorktreeSnapshot> {
901 cx.spawn(async move |cx| {
902 // Get worktree path and snapshot
903 let worktree_info = cx.update(|app_cx| {
904 let worktree = worktree.read(app_cx);
905 let path = worktree.abs_path().to_string_lossy().into_owned();
906 let snapshot = worktree.snapshot();
907 (path, snapshot)
908 });
909
910 let Ok((worktree_path, _snapshot)) = worktree_info else {
911 return WorktreeSnapshot {
912 worktree_path: String::new(),
913 git_state: None,
914 };
915 };
916
917 let git_state = git_store
918 .update(cx, |git_store, cx| {
919 git_store
920 .repositories()
921 .values()
922 .find(|repo| {
923 repo.read(cx)
924 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
925 .is_some()
926 })
927 .cloned()
928 })
929 .ok()
930 .flatten()
931 .map(|repo| {
932 repo.update(cx, |repo, _| {
933 let current_branch =
934 repo.branch.as_ref().map(|branch| branch.name().to_owned());
935 repo.send_job(None, |state, _| async move {
936 let RepositoryState::Local { backend, .. } = state else {
937 return GitState {
938 remote_url: None,
939 head_sha: None,
940 current_branch,
941 diff: None,
942 };
943 };
944
945 let remote_url = backend.remote_url("origin");
946 let head_sha = backend.head_sha().await;
947 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
948
949 GitState {
950 remote_url,
951 head_sha,
952 current_branch,
953 diff,
954 }
955 })
956 })
957 });
958
959 let git_state = match git_state {
960 Some(git_state) => match git_state.ok() {
961 Some(git_state) => git_state.await.ok(),
962 None => None,
963 },
964 None => None,
965 };
966
967 WorktreeSnapshot {
968 worktree_path,
969 git_state,
970 }
971 })
972 }
973
974 pub fn project_context(&self) -> &Entity<ProjectContext> {
975 &self.project_context
976 }
977
978 pub fn project(&self) -> &Entity<Project> {
979 &self.project
980 }
981
982 pub fn action_log(&self) -> &Entity<ActionLog> {
983 &self.action_log
984 }
985
986 pub fn is_empty(&self) -> bool {
987 self.messages.is_empty() && self.title.is_none()
988 }
989
990 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
991 self.model.as_ref()
992 }
993
994 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
995 let old_usage = self.latest_token_usage();
996 self.model = Some(model);
997 let new_caps = Self::prompt_capabilities(self.model.as_deref());
998 let new_usage = self.latest_token_usage();
999 if old_usage != new_usage {
1000 cx.emit(TokenUsageUpdated(new_usage));
1001 }
1002 self.prompt_capabilities_tx.send(new_caps).log_err();
1003 cx.notify()
1004 }
1005
1006 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1007 self.summarization_model.as_ref()
1008 }
1009
1010 pub fn set_summarization_model(
1011 &mut self,
1012 model: Option<Arc<dyn LanguageModel>>,
1013 cx: &mut Context<Self>,
1014 ) {
1015 self.summarization_model = model;
1016 cx.notify()
1017 }
1018
1019 pub fn completion_mode(&self) -> CompletionMode {
1020 self.completion_mode
1021 }
1022
1023 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
1024 let old_usage = self.latest_token_usage();
1025 self.completion_mode = mode;
1026 let new_usage = self.latest_token_usage();
1027 if old_usage != new_usage {
1028 cx.emit(TokenUsageUpdated(new_usage));
1029 }
1030 cx.notify()
1031 }
1032
1033 #[cfg(any(test, feature = "test-support"))]
1034 pub fn last_message(&self) -> Option<Message> {
1035 if let Some(message) = self.pending_message.clone() {
1036 Some(Message::Agent(message))
1037 } else {
1038 self.messages.last().cloned()
1039 }
1040 }
1041
1042 pub fn add_default_tools(
1043 &mut self,
1044 environment: Rc<dyn ThreadEnvironment>,
1045 cx: &mut Context<Self>,
1046 ) {
1047 let language_registry = self.project.read(cx).languages().clone();
1048 self.add_tool(CopyPathTool::new(self.project.clone()));
1049 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1050 self.add_tool(DeletePathTool::new(
1051 self.project.clone(),
1052 self.action_log.clone(),
1053 ));
1054 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1055 self.add_tool(EditFileTool::new(
1056 self.project.clone(),
1057 cx.weak_entity(),
1058 language_registry,
1059 ));
1060 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1061 self.add_tool(FindPathTool::new(self.project.clone()));
1062 self.add_tool(GrepTool::new(self.project.clone()));
1063 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1064 self.add_tool(MovePathTool::new(self.project.clone()));
1065 self.add_tool(NowTool);
1066 self.add_tool(OpenTool::new(self.project.clone()));
1067 self.add_tool(ReadFileTool::new(
1068 self.project.clone(),
1069 self.action_log.clone(),
1070 ));
1071 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1072 self.add_tool(ThinkingTool);
1073 self.add_tool(WebSearchTool);
1074 }
1075
1076 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1077 self.tools.insert(T::name().into(), tool.erase());
1078 }
1079
1080 pub fn remove_tool(&mut self, name: &str) -> bool {
1081 self.tools.remove(name).is_some()
1082 }
1083
1084 pub fn profile(&self) -> &AgentProfileId {
1085 &self.profile_id
1086 }
1087
1088 pub fn set_profile(&mut self, profile_id: AgentProfileId) {
1089 self.profile_id = profile_id;
1090 }
1091
1092 pub fn cancel(&mut self, cx: &mut Context<Self>) {
1093 if let Some(running_turn) = self.running_turn.take() {
1094 running_turn.cancel();
1095 }
1096 self.flush_pending_message(cx);
1097 }
1098
1099 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1100 let Some(last_user_message) = self.last_user_message() else {
1101 return;
1102 };
1103
1104 self.request_token_usage
1105 .insert(last_user_message.id.clone(), update);
1106 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1107 cx.notify();
1108 }
1109
1110 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1111 self.cancel(cx);
1112 let Some(position) = self.messages.iter().position(
1113 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1114 ) else {
1115 return Err(anyhow!("Message not found"));
1116 };
1117
1118 for message in self.messages.drain(position..) {
1119 match message {
1120 Message::User(message) => {
1121 self.request_token_usage.remove(&message.id);
1122 }
1123 Message::Agent(_) | Message::Resume => {}
1124 }
1125 }
1126 self.summary = None;
1127 cx.notify();
1128 Ok(())
1129 }
1130
1131 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1132 let last_user_message = self.last_user_message()?;
1133 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1134 let model = self.model.clone()?;
1135
1136 Some(acp_thread::TokenUsage {
1137 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1138 used_tokens: tokens.total_tokens(),
1139 })
1140 }
1141
1142 pub fn resume(
1143 &mut self,
1144 cx: &mut Context<Self>,
1145 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1146 self.messages.push(Message::Resume);
1147 cx.notify();
1148
1149 log::debug!("Total messages in thread: {}", self.messages.len());
1150 self.run_turn(cx)
1151 }
1152
1153 /// Sending a message results in the model streaming a response, which could include tool calls.
1154 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1155 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1156 pub fn send<T>(
1157 &mut self,
1158 id: UserMessageId,
1159 content: impl IntoIterator<Item = T>,
1160 cx: &mut Context<Self>,
1161 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1162 where
1163 T: Into<UserMessageContent>,
1164 {
1165 let model = self.model().context("No language model configured")?;
1166
1167 log::info!("Thread::send called with model: {}", model.name().0);
1168 self.advance_prompt_id();
1169
1170 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1171 log::debug!("Thread::send content: {:?}", content);
1172
1173 self.messages
1174 .push(Message::User(UserMessage { id, content }));
1175 cx.notify();
1176
1177 log::debug!("Total messages in thread: {}", self.messages.len());
1178 self.run_turn(cx)
1179 }
1180
1181 fn run_turn(
1182 &mut self,
1183 cx: &mut Context<Self>,
1184 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1185 self.cancel(cx);
1186
1187 let model = self.model.clone().context("No language model configured")?;
1188 let profile = AgentSettings::get_global(cx)
1189 .profiles
1190 .get(&self.profile_id)
1191 .context("Profile not found")?;
1192 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1193 let event_stream = ThreadEventStream(events_tx);
1194 let message_ix = self.messages.len().saturating_sub(1);
1195 self.tool_use_limit_reached = false;
1196 self.summary = None;
1197 self.running_turn = Some(RunningTurn {
1198 event_stream: event_stream.clone(),
1199 tools: self.enabled_tools(profile, &model, cx),
1200 _task: cx.spawn(async move |this, cx| {
1201 log::debug!("Starting agent turn execution");
1202
1203 let turn_result = Self::run_turn_internal(&this, model, &event_stream, cx).await;
1204 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1205
1206 match turn_result {
1207 Ok(()) => {
1208 log::debug!("Turn execution completed");
1209 event_stream.send_stop(acp::StopReason::EndTurn);
1210 }
1211 Err(error) => {
1212 log::error!("Turn execution failed: {:?}", error);
1213 match error.downcast::<CompletionError>() {
1214 Ok(CompletionError::Refusal) => {
1215 event_stream.send_stop(acp::StopReason::Refusal);
1216 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1217 }
1218 Ok(CompletionError::MaxTokens) => {
1219 event_stream.send_stop(acp::StopReason::MaxTokens);
1220 }
1221 Ok(CompletionError::Other(error)) | Err(error) => {
1222 event_stream.send_error(error);
1223 }
1224 }
1225 }
1226 }
1227
1228 _ = this.update(cx, |this, _| this.running_turn.take());
1229 }),
1230 });
1231 Ok(events_rx)
1232 }
1233
1234 async fn run_turn_internal(
1235 this: &WeakEntity<Self>,
1236 model: Arc<dyn LanguageModel>,
1237 event_stream: &ThreadEventStream,
1238 cx: &mut AsyncApp,
1239 ) -> Result<()> {
1240 let mut attempt = 0;
1241 let mut intent = CompletionIntent::UserPrompt;
1242 loop {
1243 let request =
1244 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1245
1246 telemetry::event!(
1247 "Agent Thread Completion",
1248 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1249 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1250 model = model.telemetry_id(),
1251 model_provider = model.provider_id().to_string(),
1252 attempt
1253 );
1254
1255 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1256
1257 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1258 Ok(events) => (events, None),
1259 Err(err) => (stream::empty().boxed(), Some(err)),
1260 };
1261 let mut tool_results = FuturesUnordered::new();
1262 while let Some(event) = events.next().await {
1263 log::trace!("Received completion event: {:?}", event);
1264 match event {
1265 Ok(event) => {
1266 tool_results.extend(this.update(cx, |this, cx| {
1267 this.handle_completion_event(event, event_stream, cx)
1268 })??);
1269 }
1270 Err(err) => {
1271 error = Some(err);
1272 break;
1273 }
1274 }
1275 }
1276
1277 let end_turn = tool_results.is_empty();
1278 while let Some(tool_result) = tool_results.next().await {
1279 log::debug!("Tool finished {:?}", tool_result);
1280
1281 event_stream.update_tool_call_fields(
1282 &tool_result.tool_use_id,
1283 acp::ToolCallUpdateFields {
1284 status: Some(if tool_result.is_error {
1285 acp::ToolCallStatus::Failed
1286 } else {
1287 acp::ToolCallStatus::Completed
1288 }),
1289 raw_output: tool_result.output.clone(),
1290 ..Default::default()
1291 },
1292 );
1293 this.update(cx, |this, _cx| {
1294 this.pending_message()
1295 .tool_results
1296 .insert(tool_result.tool_use_id.clone(), tool_result);
1297 })?;
1298 }
1299
1300 this.update(cx, |this, cx| {
1301 this.flush_pending_message(cx);
1302 if this.title.is_none() && this.pending_title_generation.is_none() {
1303 this.generate_title(cx);
1304 }
1305 })?;
1306
1307 if let Some(error) = error {
1308 attempt += 1;
1309 let retry = this.update(cx, |this, cx| {
1310 let user_store = this.user_store.read(cx);
1311 this.handle_completion_error(error, attempt, user_store.plan())
1312 })??;
1313 let timer = cx.background_executor().timer(retry.duration);
1314 event_stream.send_retry(retry);
1315 timer.await;
1316 this.update(cx, |this, _cx| {
1317 if let Some(Message::Agent(message)) = this.messages.last() {
1318 if message.tool_results.is_empty() {
1319 intent = CompletionIntent::UserPrompt;
1320 this.messages.push(Message::Resume);
1321 }
1322 }
1323 })?;
1324 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1325 return Err(language_model::ToolUseLimitReachedError.into());
1326 } else if end_turn {
1327 return Ok(());
1328 } else {
1329 intent = CompletionIntent::ToolResults;
1330 attempt = 0;
1331 }
1332 }
1333 }
1334
1335 fn handle_completion_error(
1336 &mut self,
1337 error: LanguageModelCompletionError,
1338 attempt: u8,
1339 plan: Option<Plan>,
1340 ) -> Result<acp_thread::RetryStatus> {
1341 let Some(model) = self.model.as_ref() else {
1342 return Err(anyhow!(error));
1343 };
1344
1345 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1346 match plan {
1347 Some(Plan::V2(_)) => true,
1348 Some(Plan::V1(_)) => self.completion_mode == CompletionMode::Burn,
1349 None => false,
1350 }
1351 } else {
1352 true
1353 };
1354
1355 if !auto_retry {
1356 return Err(anyhow!(error));
1357 }
1358
1359 let Some(strategy) = Self::retry_strategy_for(&error) else {
1360 return Err(anyhow!(error));
1361 };
1362
1363 let max_attempts = match &strategy {
1364 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1365 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1366 };
1367
1368 if attempt > max_attempts {
1369 return Err(anyhow!(error));
1370 }
1371
1372 let delay = match &strategy {
1373 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1374 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1375 Duration::from_secs(delay_secs)
1376 }
1377 RetryStrategy::Fixed { delay, .. } => *delay,
1378 };
1379 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1380
1381 Ok(acp_thread::RetryStatus {
1382 last_error: error.to_string().into(),
1383 attempt: attempt as usize,
1384 max_attempts: max_attempts as usize,
1385 started_at: Instant::now(),
1386 duration: delay,
1387 })
1388 }
1389
1390 /// A helper method that's called on every streamed completion event.
1391 /// Returns an optional tool result task, which the main agentic loop will
1392 /// send back to the model when it resolves.
1393 fn handle_completion_event(
1394 &mut self,
1395 event: LanguageModelCompletionEvent,
1396 event_stream: &ThreadEventStream,
1397 cx: &mut Context<Self>,
1398 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1399 log::trace!("Handling streamed completion event: {:?}", event);
1400 use LanguageModelCompletionEvent::*;
1401
1402 match event {
1403 StartMessage { .. } => {
1404 self.flush_pending_message(cx);
1405 self.pending_message = Some(AgentMessage::default());
1406 }
1407 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1408 Thinking { text, signature } => {
1409 self.handle_thinking_event(text, signature, event_stream, cx)
1410 }
1411 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1412 ToolUse(tool_use) => {
1413 return Ok(self.handle_tool_use_event(tool_use, event_stream, cx));
1414 }
1415 ToolUseJsonParseError {
1416 id,
1417 tool_name,
1418 raw_input,
1419 json_parse_error,
1420 } => {
1421 return Ok(Some(Task::ready(
1422 self.handle_tool_use_json_parse_error_event(
1423 id,
1424 tool_name,
1425 raw_input,
1426 json_parse_error,
1427 ),
1428 )));
1429 }
1430 UsageUpdate(usage) => {
1431 telemetry::event!(
1432 "Agent Thread Completion Usage Updated",
1433 thread_id = self.id.to_string(),
1434 prompt_id = self.prompt_id.to_string(),
1435 model = self.model.as_ref().map(|m| m.telemetry_id()),
1436 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1437 input_tokens = usage.input_tokens,
1438 output_tokens = usage.output_tokens,
1439 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1440 cache_read_input_tokens = usage.cache_read_input_tokens,
1441 );
1442 self.update_token_usage(usage, cx);
1443 }
1444 StatusUpdate(CompletionRequestStatus::UsageUpdated { amount, limit }) => {
1445 self.update_model_request_usage(amount, limit, cx);
1446 }
1447 StatusUpdate(
1448 CompletionRequestStatus::Started
1449 | CompletionRequestStatus::Queued { .. }
1450 | CompletionRequestStatus::Failed { .. },
1451 ) => {}
1452 StatusUpdate(CompletionRequestStatus::ToolUseLimitReached) => {
1453 self.tool_use_limit_reached = true;
1454 }
1455 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1456 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1457 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1458 }
1459
1460 Ok(None)
1461 }
1462
1463 fn handle_text_event(
1464 &mut self,
1465 new_text: String,
1466 event_stream: &ThreadEventStream,
1467 cx: &mut Context<Self>,
1468 ) {
1469 event_stream.send_text(&new_text);
1470
1471 let last_message = self.pending_message();
1472 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1473 text.push_str(&new_text);
1474 } else {
1475 last_message
1476 .content
1477 .push(AgentMessageContent::Text(new_text));
1478 }
1479
1480 cx.notify();
1481 }
1482
1483 fn handle_thinking_event(
1484 &mut self,
1485 new_text: String,
1486 new_signature: Option<String>,
1487 event_stream: &ThreadEventStream,
1488 cx: &mut Context<Self>,
1489 ) {
1490 event_stream.send_thinking(&new_text);
1491
1492 let last_message = self.pending_message();
1493 if let Some(AgentMessageContent::Thinking { text, signature }) =
1494 last_message.content.last_mut()
1495 {
1496 text.push_str(&new_text);
1497 *signature = new_signature.or(signature.take());
1498 } else {
1499 last_message.content.push(AgentMessageContent::Thinking {
1500 text: new_text,
1501 signature: new_signature,
1502 });
1503 }
1504
1505 cx.notify();
1506 }
1507
1508 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1509 let last_message = self.pending_message();
1510 last_message
1511 .content
1512 .push(AgentMessageContent::RedactedThinking(data));
1513 cx.notify();
1514 }
1515
1516 fn handle_tool_use_event(
1517 &mut self,
1518 tool_use: LanguageModelToolUse,
1519 event_stream: &ThreadEventStream,
1520 cx: &mut Context<Self>,
1521 ) -> Option<Task<LanguageModelToolResult>> {
1522 cx.notify();
1523
1524 let tool = self.tool(tool_use.name.as_ref());
1525 let mut title = SharedString::from(&tool_use.name);
1526 let mut kind = acp::ToolKind::Other;
1527 if let Some(tool) = tool.as_ref() {
1528 title = tool.initial_title(tool_use.input.clone(), cx);
1529 kind = tool.kind();
1530 }
1531
1532 // Ensure the last message ends in the current tool use
1533 let last_message = self.pending_message();
1534 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1535 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1536 if last_tool_use.id == tool_use.id {
1537 *last_tool_use = tool_use.clone();
1538 false
1539 } else {
1540 true
1541 }
1542 } else {
1543 true
1544 }
1545 });
1546
1547 if push_new_tool_use {
1548 event_stream.send_tool_call(&tool_use.id, title, kind, tool_use.input.clone());
1549 last_message
1550 .content
1551 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1552 } else {
1553 event_stream.update_tool_call_fields(
1554 &tool_use.id,
1555 acp::ToolCallUpdateFields {
1556 title: Some(title.into()),
1557 kind: Some(kind),
1558 raw_input: Some(tool_use.input.clone()),
1559 ..Default::default()
1560 },
1561 );
1562 }
1563
1564 if !tool_use.is_input_complete {
1565 return None;
1566 }
1567
1568 let Some(tool) = tool else {
1569 let content = format!("No tool named {} exists", tool_use.name);
1570 return Some(Task::ready(LanguageModelToolResult {
1571 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1572 tool_use_id: tool_use.id,
1573 tool_name: tool_use.name,
1574 is_error: true,
1575 output: None,
1576 }));
1577 };
1578
1579 let fs = self.project.read(cx).fs().clone();
1580 let tool_event_stream =
1581 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1582 tool_event_stream.update_fields(acp::ToolCallUpdateFields {
1583 status: Some(acp::ToolCallStatus::InProgress),
1584 ..Default::default()
1585 });
1586 let supports_images = self.model().is_some_and(|model| model.supports_images());
1587 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1588 log::debug!("Running tool {}", tool_use.name);
1589 Some(cx.foreground_executor().spawn(async move {
1590 let tool_result = tool_result.await.and_then(|output| {
1591 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1592 && !supports_images
1593 {
1594 return Err(anyhow!(
1595 "Attempted to read an image, but this model doesn't support it.",
1596 ));
1597 }
1598 Ok(output)
1599 });
1600
1601 match tool_result {
1602 Ok(output) => LanguageModelToolResult {
1603 tool_use_id: tool_use.id,
1604 tool_name: tool_use.name,
1605 is_error: false,
1606 content: output.llm_output,
1607 output: Some(output.raw_output),
1608 },
1609 Err(error) => LanguageModelToolResult {
1610 tool_use_id: tool_use.id,
1611 tool_name: tool_use.name,
1612 is_error: true,
1613 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1614 output: Some(error.to_string().into()),
1615 },
1616 }
1617 }))
1618 }
1619
1620 fn handle_tool_use_json_parse_error_event(
1621 &mut self,
1622 tool_use_id: LanguageModelToolUseId,
1623 tool_name: Arc<str>,
1624 raw_input: Arc<str>,
1625 json_parse_error: String,
1626 ) -> LanguageModelToolResult {
1627 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1628 LanguageModelToolResult {
1629 tool_use_id,
1630 tool_name,
1631 is_error: true,
1632 content: LanguageModelToolResultContent::Text(tool_output.into()),
1633 output: Some(serde_json::Value::String(raw_input.to_string())),
1634 }
1635 }
1636
1637 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1638 self.project
1639 .read(cx)
1640 .user_store()
1641 .update(cx, |user_store, cx| {
1642 user_store.update_model_request_usage(
1643 ModelRequestUsage(RequestUsage {
1644 amount: amount as i32,
1645 limit,
1646 }),
1647 cx,
1648 )
1649 });
1650 }
1651
1652 pub fn title(&self) -> SharedString {
1653 self.title.clone().unwrap_or("New Thread".into())
1654 }
1655
1656 pub fn summary(&mut self, cx: &mut Context<Self>) -> Task<Result<SharedString>> {
1657 if let Some(summary) = self.summary.as_ref() {
1658 return Task::ready(Ok(summary.clone()));
1659 }
1660 let Some(model) = self.summarization_model.clone() else {
1661 return Task::ready(Err(anyhow!("No summarization model available")));
1662 };
1663 let mut request = LanguageModelRequest {
1664 intent: Some(CompletionIntent::ThreadContextSummarization),
1665 temperature: AgentSettings::temperature_for_model(&model, cx),
1666 ..Default::default()
1667 };
1668
1669 for message in &self.messages {
1670 request.messages.extend(message.to_request());
1671 }
1672
1673 request.messages.push(LanguageModelRequestMessage {
1674 role: Role::User,
1675 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1676 cache: false,
1677 });
1678 cx.spawn(async move |this, cx| {
1679 let mut summary = String::new();
1680 let mut messages = model.stream_completion(request, cx).await?;
1681 while let Some(event) = messages.next().await {
1682 let event = event?;
1683 let text = match event {
1684 LanguageModelCompletionEvent::Text(text) => text,
1685 LanguageModelCompletionEvent::StatusUpdate(
1686 CompletionRequestStatus::UsageUpdated { amount, limit },
1687 ) => {
1688 this.update(cx, |thread, cx| {
1689 thread.update_model_request_usage(amount, limit, cx);
1690 })?;
1691 continue;
1692 }
1693 _ => continue,
1694 };
1695
1696 let mut lines = text.lines();
1697 summary.extend(lines.next());
1698 }
1699
1700 log::debug!("Setting summary: {}", summary);
1701 let summary = SharedString::from(summary);
1702
1703 this.update(cx, |this, cx| {
1704 this.summary = Some(summary.clone());
1705 cx.notify()
1706 })?;
1707
1708 Ok(summary)
1709 })
1710 }
1711
1712 fn generate_title(&mut self, cx: &mut Context<Self>) {
1713 let Some(model) = self.summarization_model.clone() else {
1714 return;
1715 };
1716
1717 log::debug!(
1718 "Generating title with model: {:?}",
1719 self.summarization_model.as_ref().map(|model| model.name())
1720 );
1721 let mut request = LanguageModelRequest {
1722 intent: Some(CompletionIntent::ThreadSummarization),
1723 temperature: AgentSettings::temperature_for_model(&model, cx),
1724 ..Default::default()
1725 };
1726
1727 for message in &self.messages {
1728 request.messages.extend(message.to_request());
1729 }
1730
1731 request.messages.push(LanguageModelRequestMessage {
1732 role: Role::User,
1733 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1734 cache: false,
1735 });
1736 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
1737 let mut title = String::new();
1738
1739 let generate = async {
1740 let mut messages = model.stream_completion(request, cx).await?;
1741 while let Some(event) = messages.next().await {
1742 let event = event?;
1743 let text = match event {
1744 LanguageModelCompletionEvent::Text(text) => text,
1745 LanguageModelCompletionEvent::StatusUpdate(
1746 CompletionRequestStatus::UsageUpdated { amount, limit },
1747 ) => {
1748 this.update(cx, |thread, cx| {
1749 thread.update_model_request_usage(amount, limit, cx);
1750 })?;
1751 continue;
1752 }
1753 _ => continue,
1754 };
1755
1756 let mut lines = text.lines();
1757 title.extend(lines.next());
1758
1759 // Stop if the LLM generated multiple lines.
1760 if lines.next().is_some() {
1761 break;
1762 }
1763 }
1764 anyhow::Ok(())
1765 };
1766
1767 if generate.await.context("failed to generate title").is_ok() {
1768 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
1769 }
1770 _ = this.update(cx, |this, _| this.pending_title_generation = None);
1771 }));
1772 }
1773
1774 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
1775 self.pending_title_generation = None;
1776 if Some(&title) != self.title.as_ref() {
1777 self.title = Some(title);
1778 cx.emit(TitleUpdated);
1779 cx.notify();
1780 }
1781 }
1782
1783 fn last_user_message(&self) -> Option<&UserMessage> {
1784 self.messages
1785 .iter()
1786 .rev()
1787 .find_map(|message| match message {
1788 Message::User(user_message) => Some(user_message),
1789 Message::Agent(_) => None,
1790 Message::Resume => None,
1791 })
1792 }
1793
1794 fn pending_message(&mut self) -> &mut AgentMessage {
1795 self.pending_message.get_or_insert_default()
1796 }
1797
1798 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1799 let Some(mut message) = self.pending_message.take() else {
1800 return;
1801 };
1802
1803 if message.content.is_empty() {
1804 return;
1805 }
1806
1807 for content in &message.content {
1808 let AgentMessageContent::ToolUse(tool_use) = content else {
1809 continue;
1810 };
1811
1812 if !message.tool_results.contains_key(&tool_use.id) {
1813 message.tool_results.insert(
1814 tool_use.id.clone(),
1815 LanguageModelToolResult {
1816 tool_use_id: tool_use.id.clone(),
1817 tool_name: tool_use.name.clone(),
1818 is_error: true,
1819 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1820 output: None,
1821 },
1822 );
1823 }
1824 }
1825
1826 self.messages.push(Message::Agent(message));
1827 self.updated_at = Utc::now();
1828 self.summary = None;
1829 cx.notify()
1830 }
1831
1832 pub(crate) fn build_completion_request(
1833 &self,
1834 completion_intent: CompletionIntent,
1835 cx: &App,
1836 ) -> Result<LanguageModelRequest> {
1837 let model = self.model().context("No language model configured")?;
1838 let tools = if let Some(turn) = self.running_turn.as_ref() {
1839 turn.tools
1840 .iter()
1841 .filter_map(|(tool_name, tool)| {
1842 log::trace!("Including tool: {}", tool_name);
1843 Some(LanguageModelRequestTool {
1844 name: tool_name.to_string(),
1845 description: tool.description().to_string(),
1846 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1847 })
1848 })
1849 .collect::<Vec<_>>()
1850 } else {
1851 Vec::new()
1852 };
1853
1854 log::debug!("Building completion request");
1855 log::debug!("Completion intent: {:?}", completion_intent);
1856 log::debug!("Completion mode: {:?}", self.completion_mode);
1857
1858 let messages = self.build_request_messages(cx);
1859 log::debug!("Request will include {} messages", messages.len());
1860 log::debug!("Request includes {} tools", tools.len());
1861
1862 let request = LanguageModelRequest {
1863 thread_id: Some(self.id.to_string()),
1864 prompt_id: Some(self.prompt_id.to_string()),
1865 intent: Some(completion_intent),
1866 mode: Some(self.completion_mode.into()),
1867 messages,
1868 tools,
1869 tool_choice: None,
1870 stop: Vec::new(),
1871 temperature: AgentSettings::temperature_for_model(model, cx),
1872 thinking_allowed: true,
1873 };
1874
1875 log::debug!("Completion request built successfully");
1876 Ok(request)
1877 }
1878
1879 fn enabled_tools(
1880 &self,
1881 profile: &AgentProfileSettings,
1882 model: &Arc<dyn LanguageModel>,
1883 cx: &App,
1884 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
1885 fn truncate(tool_name: &SharedString) -> SharedString {
1886 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
1887 let mut truncated = tool_name.to_string();
1888 truncated.truncate(MAX_TOOL_NAME_LENGTH);
1889 truncated.into()
1890 } else {
1891 tool_name.clone()
1892 }
1893 }
1894
1895 let mut tools = self
1896 .tools
1897 .iter()
1898 .filter_map(|(tool_name, tool)| {
1899 if tool.supported_provider(&model.provider_id())
1900 && profile.is_tool_enabled(tool_name)
1901 {
1902 Some((truncate(tool_name), tool.clone()))
1903 } else {
1904 None
1905 }
1906 })
1907 .collect::<BTreeMap<_, _>>();
1908
1909 let mut context_server_tools = Vec::new();
1910 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
1911 let mut duplicate_tool_names = HashSet::default();
1912 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
1913 for (tool_name, tool) in server_tools {
1914 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
1915 let tool_name = truncate(tool_name);
1916 if !seen_tools.insert(tool_name.clone()) {
1917 duplicate_tool_names.insert(tool_name.clone());
1918 }
1919 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
1920 }
1921 }
1922 }
1923
1924 // When there are duplicate tool names, disambiguate by prefixing them
1925 // with the server ID. In the rare case there isn't enough space for the
1926 // disambiguated tool name, keep only the last tool with this name.
1927 for (server_id, tool_name, tool) in context_server_tools {
1928 if duplicate_tool_names.contains(&tool_name) {
1929 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
1930 if available >= 2 {
1931 let mut disambiguated = server_id.0.to_string();
1932 disambiguated.truncate(available - 1);
1933 disambiguated.push('_');
1934 disambiguated.push_str(&tool_name);
1935 tools.insert(disambiguated.into(), tool.clone());
1936 } else {
1937 tools.insert(tool_name, tool.clone());
1938 }
1939 } else {
1940 tools.insert(tool_name, tool.clone());
1941 }
1942 }
1943
1944 tools
1945 }
1946
1947 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
1948 self.running_turn.as_ref()?.tools.get(name).cloned()
1949 }
1950
1951 fn build_request_messages(&self, cx: &App) -> Vec<LanguageModelRequestMessage> {
1952 log::trace!(
1953 "Building request messages from {} thread messages",
1954 self.messages.len()
1955 );
1956
1957 let system_prompt = SystemPromptTemplate {
1958 project: self.project_context.read(cx),
1959 available_tools: self.tools.keys().cloned().collect(),
1960 }
1961 .render(&self.templates)
1962 .context("failed to build system prompt")
1963 .expect("Invalid template");
1964 let mut messages = vec![LanguageModelRequestMessage {
1965 role: Role::System,
1966 content: vec![system_prompt.into()],
1967 cache: false,
1968 }];
1969 for message in &self.messages {
1970 messages.extend(message.to_request());
1971 }
1972
1973 if let Some(last_message) = messages.last_mut() {
1974 last_message.cache = true;
1975 }
1976
1977 if let Some(message) = self.pending_message.as_ref() {
1978 messages.extend(message.to_request());
1979 }
1980
1981 messages
1982 }
1983
1984 pub fn to_markdown(&self) -> String {
1985 let mut markdown = String::new();
1986 for (ix, message) in self.messages.iter().enumerate() {
1987 if ix > 0 {
1988 markdown.push('\n');
1989 }
1990 markdown.push_str(&message.to_markdown());
1991 }
1992
1993 if let Some(message) = self.pending_message.as_ref() {
1994 markdown.push('\n');
1995 markdown.push_str(&message.to_markdown());
1996 }
1997
1998 markdown
1999 }
2000
2001 fn advance_prompt_id(&mut self) {
2002 self.prompt_id = PromptId::new();
2003 }
2004
2005 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2006 use LanguageModelCompletionError::*;
2007 use http_client::StatusCode;
2008
2009 // General strategy here:
2010 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2011 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2012 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2013 match error {
2014 HttpResponseError {
2015 status_code: StatusCode::TOO_MANY_REQUESTS,
2016 ..
2017 } => Some(RetryStrategy::ExponentialBackoff {
2018 initial_delay: BASE_RETRY_DELAY,
2019 max_attempts: MAX_RETRY_ATTEMPTS,
2020 }),
2021 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2022 Some(RetryStrategy::Fixed {
2023 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2024 max_attempts: MAX_RETRY_ATTEMPTS,
2025 })
2026 }
2027 UpstreamProviderError {
2028 status,
2029 retry_after,
2030 ..
2031 } => match *status {
2032 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2033 Some(RetryStrategy::Fixed {
2034 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2035 max_attempts: MAX_RETRY_ATTEMPTS,
2036 })
2037 }
2038 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2039 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2040 // Internal Server Error could be anything, retry up to 3 times.
2041 max_attempts: 3,
2042 }),
2043 status => {
2044 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2045 // but we frequently get them in practice. See https://http.dev/529
2046 if status.as_u16() == 529 {
2047 Some(RetryStrategy::Fixed {
2048 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2049 max_attempts: MAX_RETRY_ATTEMPTS,
2050 })
2051 } else {
2052 Some(RetryStrategy::Fixed {
2053 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2054 max_attempts: 2,
2055 })
2056 }
2057 }
2058 },
2059 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2060 delay: BASE_RETRY_DELAY,
2061 max_attempts: 3,
2062 }),
2063 ApiReadResponseError { .. }
2064 | HttpSend { .. }
2065 | DeserializeResponse { .. }
2066 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2067 delay: BASE_RETRY_DELAY,
2068 max_attempts: 3,
2069 }),
2070 // Retrying these errors definitely shouldn't help.
2071 HttpResponseError {
2072 status_code:
2073 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2074 ..
2075 }
2076 | AuthenticationError { .. }
2077 | PermissionError { .. }
2078 | NoApiKey { .. }
2079 | ApiEndpointNotFound { .. }
2080 | PromptTooLarge { .. } => None,
2081 // These errors might be transient, so retry them
2082 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2083 delay: BASE_RETRY_DELAY,
2084 max_attempts: 1,
2085 }),
2086 // Retry all other 4xx and 5xx errors once.
2087 HttpResponseError { status_code, .. }
2088 if status_code.is_client_error() || status_code.is_server_error() =>
2089 {
2090 Some(RetryStrategy::Fixed {
2091 delay: BASE_RETRY_DELAY,
2092 max_attempts: 3,
2093 })
2094 }
2095 Other(err)
2096 if err.is::<language_model::PaymentRequiredError>()
2097 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2098 {
2099 // Retrying won't help for Payment Required or Model Request Limit errors (where
2100 // the user must upgrade to usage-based billing to get more requests, or else wait
2101 // for a significant amount of time for the request limit to reset).
2102 None
2103 }
2104 // Conservatively assume that any other errors are non-retryable
2105 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2106 delay: BASE_RETRY_DELAY,
2107 max_attempts: 2,
2108 }),
2109 }
2110 }
2111}
2112
2113struct RunningTurn {
2114 /// Holds the task that handles agent interaction until the end of the turn.
2115 /// Survives across multiple requests as the model performs tool calls and
2116 /// we run tools, report their results.
2117 _task: Task<()>,
2118 /// The current event stream for the running turn. Used to report a final
2119 /// cancellation event if we cancel the turn.
2120 event_stream: ThreadEventStream,
2121 /// The tools that were enabled for this turn.
2122 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2123}
2124
2125impl RunningTurn {
2126 fn cancel(self) {
2127 log::debug!("Cancelling in progress turn");
2128 self.event_stream.send_canceled();
2129 }
2130}
2131
2132pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2133
2134impl EventEmitter<TokenUsageUpdated> for Thread {}
2135
2136pub struct TitleUpdated;
2137
2138impl EventEmitter<TitleUpdated> for Thread {}
2139
2140pub trait AgentTool
2141where
2142 Self: 'static + Sized,
2143{
2144 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2145 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2146
2147 fn name() -> &'static str;
2148
2149 fn description(&self) -> SharedString {
2150 let schema = schemars::schema_for!(Self::Input);
2151 SharedString::new(
2152 schema
2153 .get("description")
2154 .and_then(|description| description.as_str())
2155 .unwrap_or_default(),
2156 )
2157 }
2158
2159 fn kind() -> acp::ToolKind;
2160
2161 /// The initial tool title to display. Can be updated during the tool run.
2162 fn initial_title(
2163 &self,
2164 input: Result<Self::Input, serde_json::Value>,
2165 cx: &mut App,
2166 ) -> SharedString;
2167
2168 /// Returns the JSON schema that describes the tool's input.
2169 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Schema {
2170 crate::tool_schema::root_schema_for::<Self::Input>(format)
2171 }
2172
2173 /// Some tools rely on a provider for the underlying billing or other reasons.
2174 /// Allow the tool to check if they are compatible, or should be filtered out.
2175 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2176 true
2177 }
2178
2179 /// Runs the tool with the provided input.
2180 fn run(
2181 self: Arc<Self>,
2182 input: Self::Input,
2183 event_stream: ToolCallEventStream,
2184 cx: &mut App,
2185 ) -> Task<Result<Self::Output>>;
2186
2187 /// Emits events for a previous execution of the tool.
2188 fn replay(
2189 &self,
2190 _input: Self::Input,
2191 _output: Self::Output,
2192 _event_stream: ToolCallEventStream,
2193 _cx: &mut App,
2194 ) -> Result<()> {
2195 Ok(())
2196 }
2197
2198 fn erase(self) -> Arc<dyn AnyAgentTool> {
2199 Arc::new(Erased(Arc::new(self)))
2200 }
2201}
2202
2203pub struct Erased<T>(T);
2204
2205pub struct AgentToolOutput {
2206 pub llm_output: LanguageModelToolResultContent,
2207 pub raw_output: serde_json::Value,
2208}
2209
2210pub trait AnyAgentTool {
2211 fn name(&self) -> SharedString;
2212 fn description(&self) -> SharedString;
2213 fn kind(&self) -> acp::ToolKind;
2214 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2215 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2216 fn supported_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2217 true
2218 }
2219 fn run(
2220 self: Arc<Self>,
2221 input: serde_json::Value,
2222 event_stream: ToolCallEventStream,
2223 cx: &mut App,
2224 ) -> Task<Result<AgentToolOutput>>;
2225 fn replay(
2226 &self,
2227 input: serde_json::Value,
2228 output: serde_json::Value,
2229 event_stream: ToolCallEventStream,
2230 cx: &mut App,
2231 ) -> Result<()>;
2232}
2233
2234impl<T> AnyAgentTool for Erased<Arc<T>>
2235where
2236 T: AgentTool,
2237{
2238 fn name(&self) -> SharedString {
2239 T::name().into()
2240 }
2241
2242 fn description(&self) -> SharedString {
2243 self.0.description()
2244 }
2245
2246 fn kind(&self) -> agent_client_protocol::ToolKind {
2247 T::kind()
2248 }
2249
2250 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2251 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2252 self.0.initial_title(parsed_input, _cx)
2253 }
2254
2255 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2256 let mut json = serde_json::to_value(self.0.input_schema(format))?;
2257 adapt_schema_to_format(&mut json, format)?;
2258 Ok(json)
2259 }
2260
2261 fn supported_provider(&self, provider: &LanguageModelProviderId) -> bool {
2262 self.0.supported_provider(provider)
2263 }
2264
2265 fn run(
2266 self: Arc<Self>,
2267 input: serde_json::Value,
2268 event_stream: ToolCallEventStream,
2269 cx: &mut App,
2270 ) -> Task<Result<AgentToolOutput>> {
2271 cx.spawn(async move |cx| {
2272 let input = serde_json::from_value(input)?;
2273 let output = cx
2274 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2275 .await?;
2276 let raw_output = serde_json::to_value(&output)?;
2277 Ok(AgentToolOutput {
2278 llm_output: output.into(),
2279 raw_output,
2280 })
2281 })
2282 }
2283
2284 fn replay(
2285 &self,
2286 input: serde_json::Value,
2287 output: serde_json::Value,
2288 event_stream: ToolCallEventStream,
2289 cx: &mut App,
2290 ) -> Result<()> {
2291 let input = serde_json::from_value(input)?;
2292 let output = serde_json::from_value(output)?;
2293 self.0.replay(input, output, event_stream, cx)
2294 }
2295}
2296
2297#[derive(Clone)]
2298struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2299
2300impl ThreadEventStream {
2301 fn send_user_message(&self, message: &UserMessage) {
2302 self.0
2303 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2304 .ok();
2305 }
2306
2307 fn send_text(&self, text: &str) {
2308 self.0
2309 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2310 .ok();
2311 }
2312
2313 fn send_thinking(&self, text: &str) {
2314 self.0
2315 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2316 .ok();
2317 }
2318
2319 fn send_tool_call(
2320 &self,
2321 id: &LanguageModelToolUseId,
2322 title: SharedString,
2323 kind: acp::ToolKind,
2324 input: serde_json::Value,
2325 ) {
2326 self.0
2327 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2328 id,
2329 title.to_string(),
2330 kind,
2331 input,
2332 ))))
2333 .ok();
2334 }
2335
2336 fn initial_tool_call(
2337 id: &LanguageModelToolUseId,
2338 title: String,
2339 kind: acp::ToolKind,
2340 input: serde_json::Value,
2341 ) -> acp::ToolCall {
2342 acp::ToolCall {
2343 meta: None,
2344 id: acp::ToolCallId(id.to_string().into()),
2345 title,
2346 kind,
2347 status: acp::ToolCallStatus::Pending,
2348 content: vec![],
2349 locations: vec![],
2350 raw_input: Some(input),
2351 raw_output: None,
2352 }
2353 }
2354
2355 fn update_tool_call_fields(
2356 &self,
2357 tool_use_id: &LanguageModelToolUseId,
2358 fields: acp::ToolCallUpdateFields,
2359 ) {
2360 self.0
2361 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2362 acp::ToolCallUpdate {
2363 meta: None,
2364 id: acp::ToolCallId(tool_use_id.to_string().into()),
2365 fields,
2366 }
2367 .into(),
2368 )))
2369 .ok();
2370 }
2371
2372 fn send_retry(&self, status: acp_thread::RetryStatus) {
2373 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2374 }
2375
2376 fn send_stop(&self, reason: acp::StopReason) {
2377 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2378 }
2379
2380 fn send_canceled(&self) {
2381 self.0
2382 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2383 .ok();
2384 }
2385
2386 fn send_error(&self, error: impl Into<anyhow::Error>) {
2387 self.0.unbounded_send(Err(error.into())).ok();
2388 }
2389}
2390
2391#[derive(Clone)]
2392pub struct ToolCallEventStream {
2393 tool_use_id: LanguageModelToolUseId,
2394 stream: ThreadEventStream,
2395 fs: Option<Arc<dyn Fs>>,
2396}
2397
2398impl ToolCallEventStream {
2399 #[cfg(test)]
2400 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2401 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2402
2403 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2404
2405 (stream, ToolCallEventStreamReceiver(events_rx))
2406 }
2407
2408 fn new(
2409 tool_use_id: LanguageModelToolUseId,
2410 stream: ThreadEventStream,
2411 fs: Option<Arc<dyn Fs>>,
2412 ) -> Self {
2413 Self {
2414 tool_use_id,
2415 stream,
2416 fs,
2417 }
2418 }
2419
2420 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2421 self.stream
2422 .update_tool_call_fields(&self.tool_use_id, fields);
2423 }
2424
2425 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2426 self.stream
2427 .0
2428 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2429 acp_thread::ToolCallUpdateDiff {
2430 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2431 diff,
2432 }
2433 .into(),
2434 )))
2435 .ok();
2436 }
2437
2438 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2439 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2440 return Task::ready(Ok(()));
2441 }
2442
2443 let (response_tx, response_rx) = oneshot::channel();
2444 self.stream
2445 .0
2446 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2447 ToolCallAuthorization {
2448 tool_call: acp::ToolCallUpdate {
2449 meta: None,
2450 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2451 fields: acp::ToolCallUpdateFields {
2452 title: Some(title.into()),
2453 ..Default::default()
2454 },
2455 },
2456 options: vec![
2457 acp::PermissionOption {
2458 id: acp::PermissionOptionId("always_allow".into()),
2459 name: "Always Allow".into(),
2460 kind: acp::PermissionOptionKind::AllowAlways,
2461 meta: None,
2462 },
2463 acp::PermissionOption {
2464 id: acp::PermissionOptionId("allow".into()),
2465 name: "Allow".into(),
2466 kind: acp::PermissionOptionKind::AllowOnce,
2467 meta: None,
2468 },
2469 acp::PermissionOption {
2470 id: acp::PermissionOptionId("deny".into()),
2471 name: "Deny".into(),
2472 kind: acp::PermissionOptionKind::RejectOnce,
2473 meta: None,
2474 },
2475 ],
2476 response: response_tx,
2477 },
2478 )))
2479 .ok();
2480 let fs = self.fs.clone();
2481 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2482 "always_allow" => {
2483 if let Some(fs) = fs.clone() {
2484 cx.update(|cx| {
2485 update_settings_file(fs, cx, |settings, _| {
2486 settings
2487 .agent
2488 .get_or_insert_default()
2489 .set_always_allow_tool_actions(true);
2490 });
2491 })?;
2492 }
2493
2494 Ok(())
2495 }
2496 "allow" => Ok(()),
2497 _ => Err(anyhow!("Permission to run tool denied by user")),
2498 })
2499 }
2500}
2501
2502#[cfg(test)]
2503pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2504
2505#[cfg(test)]
2506impl ToolCallEventStreamReceiver {
2507 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2508 let event = self.0.next().await;
2509 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2510 auth
2511 } else {
2512 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2513 }
2514 }
2515
2516 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
2517 let event = self.0.next().await;
2518 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
2519 update,
2520 )))) = event
2521 {
2522 update.fields
2523 } else {
2524 panic!("Expected update fields but got: {:?}", event);
2525 }
2526 }
2527
2528 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
2529 let event = self.0.next().await;
2530 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
2531 update,
2532 )))) = event
2533 {
2534 update.diff
2535 } else {
2536 panic!("Expected diff but got: {:?}", event);
2537 }
2538 }
2539
2540 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2541 let event = self.0.next().await;
2542 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2543 update,
2544 )))) = event
2545 {
2546 update.terminal
2547 } else {
2548 panic!("Expected terminal but got: {:?}", event);
2549 }
2550 }
2551}
2552
2553#[cfg(test)]
2554impl std::ops::Deref for ToolCallEventStreamReceiver {
2555 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2556
2557 fn deref(&self) -> &Self::Target {
2558 &self.0
2559 }
2560}
2561
2562#[cfg(test)]
2563impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2564 fn deref_mut(&mut self) -> &mut Self::Target {
2565 &mut self.0
2566 }
2567}
2568
2569impl From<&str> for UserMessageContent {
2570 fn from(text: &str) -> Self {
2571 Self::Text(text.into())
2572 }
2573}
2574
2575impl From<acp::ContentBlock> for UserMessageContent {
2576 fn from(value: acp::ContentBlock) -> Self {
2577 match value {
2578 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2579 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2580 acp::ContentBlock::Audio(_) => {
2581 // TODO
2582 Self::Text("[audio]".to_string())
2583 }
2584 acp::ContentBlock::ResourceLink(resource_link) => {
2585 match MentionUri::parse(&resource_link.uri) {
2586 Ok(uri) => Self::Mention {
2587 uri,
2588 content: String::new(),
2589 },
2590 Err(err) => {
2591 log::error!("Failed to parse mention link: {}", err);
2592 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2593 }
2594 }
2595 }
2596 acp::ContentBlock::Resource(resource) => match resource.resource {
2597 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2598 match MentionUri::parse(&resource.uri) {
2599 Ok(uri) => Self::Mention {
2600 uri,
2601 content: resource.text,
2602 },
2603 Err(err) => {
2604 log::error!("Failed to parse mention link: {}", err);
2605 Self::Text(
2606 MarkdownCodeBlock {
2607 tag: &resource.uri,
2608 text: &resource.text,
2609 }
2610 .to_string(),
2611 )
2612 }
2613 }
2614 }
2615 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2616 // TODO
2617 Self::Text("[blob]".to_string())
2618 }
2619 },
2620 }
2621 }
2622}
2623
2624impl From<UserMessageContent> for acp::ContentBlock {
2625 fn from(content: UserMessageContent) -> Self {
2626 match content {
2627 UserMessageContent::Text(text) => acp::ContentBlock::Text(acp::TextContent {
2628 text,
2629 annotations: None,
2630 meta: None,
2631 }),
2632 UserMessageContent::Image(image) => acp::ContentBlock::Image(acp::ImageContent {
2633 data: image.source.to_string(),
2634 mime_type: "image/png".to_string(),
2635 meta: None,
2636 annotations: None,
2637 uri: None,
2638 }),
2639 UserMessageContent::Mention { uri, content } => {
2640 acp::ContentBlock::Resource(acp::EmbeddedResource {
2641 meta: None,
2642 resource: acp::EmbeddedResourceResource::TextResourceContents(
2643 acp::TextResourceContents {
2644 meta: None,
2645 mime_type: None,
2646 text: content,
2647 uri: uri.to_uri().to_string(),
2648 },
2649 ),
2650 annotations: None,
2651 })
2652 }
2653 }
2654 }
2655}
2656
2657fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2658 LanguageModelImage {
2659 source: image_content.data.into(),
2660 // TODO: make this optional?
2661 size: gpui::Size::new(0.into(), 0.into()),
2662 }
2663}