1use crate::{
2 ContextServerRegistry, CopyPathTool, CreateDirectoryTool, DbLanguageModel, DbThread,
3 DeletePathTool, DiagnosticsTool, EditFileTool, FetchTool, FindPathTool, GrepTool,
4 ListDirectoryTool, MovePathTool, NowTool, OpenTool, ProjectSnapshot, ReadFileTool,
5 SystemPromptTemplate, Template, Templates, TerminalTool, ThinkingTool, WebSearchTool,
6};
7use acp_thread::{MentionUri, UserMessageId};
8use action_log::ActionLog;
9
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 chrono::{DateTime, Utc};
17use client::{ModelRequestUsage, RequestUsage, UserStore};
18use cloud_llm_client::{CompletionIntent, CompletionRequestStatus, Plan, UsageLimit};
19use collections::{HashMap, HashSet, IndexMap};
20use fs::Fs;
21use futures::stream;
22use futures::{
23 FutureExt,
24 channel::{mpsc, oneshot},
25 future::Shared,
26 stream::FuturesUnordered,
27};
28use gpui::{
29 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
30};
31use language_model::{
32 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelExt,
33 LanguageModelId, LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry,
34 LanguageModelRequest, LanguageModelRequestMessage, LanguageModelRequestTool,
35 LanguageModelToolResult, LanguageModelToolResultContent, LanguageModelToolSchemaFormat,
36 LanguageModelToolUse, LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage,
37 ZED_CLOUD_PROVIDER_ID,
38};
39use project::Project;
40use prompt_store::ProjectContext;
41use schemars::{JsonSchema, Schema};
42use serde::{Deserialize, Serialize};
43use settings::{LanguageModelSelection, Settings, update_settings_file};
44use smol::stream::StreamExt;
45use std::{
46 collections::BTreeMap,
47 ops::RangeInclusive,
48 path::Path,
49 rc::Rc,
50 sync::Arc,
51 time::{Duration, Instant},
52};
53use std::{fmt::Write, path::PathBuf};
54use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
55use uuid::Uuid;
56
57const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
58pub const MAX_TOOL_NAME_LENGTH: usize = 64;
59
60/// The ID of the user prompt that initiated a request.
61///
62/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
63#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
64pub struct PromptId(Arc<str>);
65
66impl PromptId {
67 pub fn new() -> Self {
68 Self(Uuid::new_v4().to_string().into())
69 }
70}
71
72impl std::fmt::Display for PromptId {
73 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
74 write!(f, "{}", self.0)
75 }
76}
77
78pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
79pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
80
81#[derive(Debug, Clone)]
82enum RetryStrategy {
83 ExponentialBackoff {
84 initial_delay: Duration,
85 max_attempts: u8,
86 },
87 Fixed {
88 delay: Duration,
89 max_attempts: u8,
90 },
91}
92
93#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
94pub enum Message {
95 User(UserMessage),
96 Agent(AgentMessage),
97 Resume,
98}
99
100impl Message {
101 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
102 match self {
103 Message::Agent(agent_message) => Some(agent_message),
104 _ => None,
105 }
106 }
107
108 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
109 match self {
110 Message::User(message) => vec![message.to_request()],
111 Message::Agent(message) => message.to_request(),
112 Message::Resume => vec![LanguageModelRequestMessage {
113 role: Role::User,
114 content: vec!["Continue where you left off".into()],
115 cache: false,
116 reasoning_details: None,
117 }],
118 }
119 }
120
121 pub fn to_markdown(&self) -> String {
122 match self {
123 Message::User(message) => message.to_markdown(),
124 Message::Agent(message) => message.to_markdown(),
125 Message::Resume => "[resume]\n".into(),
126 }
127 }
128
129 pub fn role(&self) -> Role {
130 match self {
131 Message::User(_) | Message::Resume => Role::User,
132 Message::Agent(_) => Role::Assistant,
133 }
134 }
135}
136
137#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
138pub struct UserMessage {
139 pub id: UserMessageId,
140 pub content: Vec<UserMessageContent>,
141}
142
143#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
144pub enum UserMessageContent {
145 Text(String),
146 Mention { uri: MentionUri, content: String },
147 Image(LanguageModelImage),
148}
149
150impl UserMessage {
151 pub fn to_markdown(&self) -> String {
152 let mut markdown = String::from("## User\n\n");
153
154 for content in &self.content {
155 match content {
156 UserMessageContent::Text(text) => {
157 markdown.push_str(text);
158 markdown.push('\n');
159 }
160 UserMessageContent::Image(_) => {
161 markdown.push_str("<image />\n");
162 }
163 UserMessageContent::Mention { uri, content } => {
164 if !content.is_empty() {
165 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
166 } else {
167 let _ = writeln!(&mut markdown, "{}", uri.as_link());
168 }
169 }
170 }
171 }
172
173 markdown
174 }
175
176 fn to_request(&self) -> LanguageModelRequestMessage {
177 let mut message = LanguageModelRequestMessage {
178 role: Role::User,
179 content: Vec::with_capacity(self.content.len()),
180 cache: false,
181 reasoning_details: None,
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 reasoning_details: self.reasoning_details.clone(),
450 };
451 for chunk in &self.content {
452 match chunk {
453 AgentMessageContent::Text(text) => {
454 assistant_message
455 .content
456 .push(language_model::MessageContent::Text(text.clone()));
457 }
458 AgentMessageContent::Thinking { text, signature } => {
459 assistant_message
460 .content
461 .push(language_model::MessageContent::Thinking {
462 text: text.clone(),
463 signature: signature.clone(),
464 });
465 }
466 AgentMessageContent::RedactedThinking(value) => {
467 assistant_message.content.push(
468 language_model::MessageContent::RedactedThinking(value.clone()),
469 );
470 }
471 AgentMessageContent::ToolUse(tool_use) => {
472 if self.tool_results.contains_key(&tool_use.id) {
473 assistant_message
474 .content
475 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
476 }
477 }
478 };
479 }
480
481 let mut user_message = LanguageModelRequestMessage {
482 role: Role::User,
483 content: Vec::new(),
484 cache: false,
485 reasoning_details: None,
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 pub reasoning_details: Option<serde_json::Value>,
516}
517
518#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
519pub enum AgentMessageContent {
520 Text(String),
521 Thinking {
522 text: String,
523 signature: Option<String>,
524 },
525 RedactedThinking(String),
526 ToolUse(LanguageModelToolUse),
527}
528
529pub trait TerminalHandle {
530 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
531 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
532 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
533}
534
535pub trait ThreadEnvironment {
536 fn create_terminal(
537 &self,
538 command: String,
539 cwd: Option<PathBuf>,
540 output_byte_limit: Option<u64>,
541 cx: &mut AsyncApp,
542 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
543}
544
545#[derive(Debug)]
546pub enum ThreadEvent {
547 UserMessage(UserMessage),
548 AgentText(String),
549 AgentThinking(String),
550 ToolCall(acp::ToolCall),
551 ToolCallUpdate(acp_thread::ToolCallUpdate),
552 ToolCallAuthorization(ToolCallAuthorization),
553 Retry(acp_thread::RetryStatus),
554 Stop(acp::StopReason),
555}
556
557#[derive(Debug)]
558pub struct NewTerminal {
559 pub command: String,
560 pub output_byte_limit: Option<u64>,
561 pub cwd: Option<PathBuf>,
562 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
563}
564
565#[derive(Debug)]
566pub struct ToolCallAuthorization {
567 pub tool_call: acp::ToolCallUpdate,
568 pub options: Vec<acp::PermissionOption>,
569 pub response: oneshot::Sender<acp::PermissionOptionId>,
570}
571
572#[derive(Debug, thiserror::Error)]
573enum CompletionError {
574 #[error("max tokens")]
575 MaxTokens,
576 #[error("refusal")]
577 Refusal,
578 #[error(transparent)]
579 Other(#[from] anyhow::Error),
580}
581
582pub struct Thread {
583 id: acp::SessionId,
584 prompt_id: PromptId,
585 updated_at: DateTime<Utc>,
586 title: Option<SharedString>,
587 pending_title_generation: Option<Task<()>>,
588 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
589 summary: Option<SharedString>,
590 messages: Vec<Message>,
591 user_store: Entity<UserStore>,
592 completion_mode: CompletionMode,
593 /// Holds the task that handles agent interaction until the end of the turn.
594 /// Survives across multiple requests as the model performs tool calls and
595 /// we run tools, report their results.
596 running_turn: Option<RunningTurn>,
597 pending_message: Option<AgentMessage>,
598 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
599 tool_use_limit_reached: bool,
600 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
601 #[allow(unused)]
602 cumulative_token_usage: TokenUsage,
603 #[allow(unused)]
604 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
605 context_server_registry: Entity<ContextServerRegistry>,
606 profile_id: AgentProfileId,
607 project_context: Entity<ProjectContext>,
608 templates: Arc<Templates>,
609 model: Option<Arc<dyn LanguageModel>>,
610 summarization_model: Option<Arc<dyn LanguageModel>>,
611 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
612 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
613 pub(crate) project: Entity<Project>,
614 pub(crate) action_log: Entity<ActionLog>,
615}
616
617impl Thread {
618 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
619 let image = model.map_or(true, |model| model.supports_images());
620 acp::PromptCapabilities {
621 meta: None,
622 image,
623 audio: false,
624 embedded_context: true,
625 }
626 }
627
628 pub fn new(
629 project: Entity<Project>,
630 project_context: Entity<ProjectContext>,
631 context_server_registry: Entity<ContextServerRegistry>,
632 templates: Arc<Templates>,
633 model: Option<Arc<dyn LanguageModel>>,
634 cx: &mut Context<Self>,
635 ) -> Self {
636 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
637 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
638 let (prompt_capabilities_tx, prompt_capabilities_rx) =
639 watch::channel(Self::prompt_capabilities(model.as_deref()));
640 Self {
641 id: acp::SessionId(uuid::Uuid::new_v4().to_string().into()),
642 prompt_id: PromptId::new(),
643 updated_at: Utc::now(),
644 title: None,
645 pending_title_generation: None,
646 pending_summary_generation: None,
647 summary: None,
648 messages: Vec::new(),
649 user_store: project.read(cx).user_store(),
650 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
651 running_turn: None,
652 pending_message: None,
653 tools: BTreeMap::default(),
654 tool_use_limit_reached: false,
655 request_token_usage: HashMap::default(),
656 cumulative_token_usage: TokenUsage::default(),
657 initial_project_snapshot: {
658 let project_snapshot = Self::project_snapshot(project.clone(), cx);
659 cx.foreground_executor()
660 .spawn(async move { Some(project_snapshot.await) })
661 .shared()
662 },
663 context_server_registry,
664 profile_id,
665 project_context,
666 templates,
667 model,
668 summarization_model: None,
669 prompt_capabilities_tx,
670 prompt_capabilities_rx,
671 project,
672 action_log,
673 }
674 }
675
676 pub fn id(&self) -> &acp::SessionId {
677 &self.id
678 }
679
680 pub fn replay(
681 &mut self,
682 cx: &mut Context<Self>,
683 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
684 let (tx, rx) = mpsc::unbounded();
685 let stream = ThreadEventStream(tx);
686 for message in &self.messages {
687 match message {
688 Message::User(user_message) => stream.send_user_message(user_message),
689 Message::Agent(assistant_message) => {
690 for content in &assistant_message.content {
691 match content {
692 AgentMessageContent::Text(text) => stream.send_text(text),
693 AgentMessageContent::Thinking { text, .. } => {
694 stream.send_thinking(text)
695 }
696 AgentMessageContent::RedactedThinking(_) => {}
697 AgentMessageContent::ToolUse(tool_use) => {
698 self.replay_tool_call(
699 tool_use,
700 assistant_message.tool_results.get(&tool_use.id),
701 &stream,
702 cx,
703 );
704 }
705 }
706 }
707 }
708 Message::Resume => {}
709 }
710 }
711 rx
712 }
713
714 fn replay_tool_call(
715 &self,
716 tool_use: &LanguageModelToolUse,
717 tool_result: Option<&LanguageModelToolResult>,
718 stream: &ThreadEventStream,
719 cx: &mut Context<Self>,
720 ) {
721 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
722 self.context_server_registry
723 .read(cx)
724 .servers()
725 .find_map(|(_, tools)| {
726 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
727 Some(tool.clone())
728 } else {
729 None
730 }
731 })
732 });
733
734 let Some(tool) = tool else {
735 stream
736 .0
737 .unbounded_send(Ok(ThreadEvent::ToolCall(acp::ToolCall {
738 meta: None,
739 id: acp::ToolCallId(tool_use.id.to_string().into()),
740 title: tool_use.name.to_string(),
741 kind: acp::ToolKind::Other,
742 status: acp::ToolCallStatus::Failed,
743 content: Vec::new(),
744 locations: Vec::new(),
745 raw_input: Some(tool_use.input.clone()),
746 raw_output: None,
747 })))
748 .ok();
749 return;
750 };
751
752 let title = tool.initial_title(tool_use.input.clone(), cx);
753 let kind = tool.kind();
754 stream.send_tool_call(
755 &tool_use.id,
756 &tool_use.name,
757 title,
758 kind,
759 tool_use.input.clone(),
760 );
761
762 let output = tool_result
763 .as_ref()
764 .and_then(|result| result.output.clone());
765 if let Some(output) = output.clone() {
766 let tool_event_stream = ToolCallEventStream::new(
767 tool_use.id.clone(),
768 stream.clone(),
769 Some(self.project.read(cx).fs().clone()),
770 );
771 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
772 .log_err();
773 }
774
775 stream.update_tool_call_fields(
776 &tool_use.id,
777 acp::ToolCallUpdateFields {
778 status: Some(
779 tool_result
780 .as_ref()
781 .map_or(acp::ToolCallStatus::Failed, |result| {
782 if result.is_error {
783 acp::ToolCallStatus::Failed
784 } else {
785 acp::ToolCallStatus::Completed
786 }
787 }),
788 ),
789 raw_output: output,
790 ..Default::default()
791 },
792 );
793 }
794
795 pub fn from_db(
796 id: acp::SessionId,
797 db_thread: DbThread,
798 project: Entity<Project>,
799 project_context: Entity<ProjectContext>,
800 context_server_registry: Entity<ContextServerRegistry>,
801 templates: Arc<Templates>,
802 cx: &mut Context<Self>,
803 ) -> Self {
804 let profile_id = db_thread
805 .profile
806 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
807
808 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
809 db_thread
810 .model
811 .and_then(|model| {
812 let model = SelectedModel {
813 provider: model.provider.clone().into(),
814 model: model.model.into(),
815 };
816 registry.select_model(&model, cx)
817 })
818 .or_else(|| registry.default_model())
819 .map(|model| model.model)
820 });
821
822 if model.is_none() {
823 model = Self::resolve_profile_model(&profile_id, cx);
824 }
825 if model.is_none() {
826 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
827 registry.default_model().map(|model| model.model)
828 });
829 }
830
831 let (prompt_capabilities_tx, prompt_capabilities_rx) =
832 watch::channel(Self::prompt_capabilities(model.as_deref()));
833
834 let action_log = cx.new(|_| ActionLog::new(project.clone()));
835
836 Self {
837 id,
838 prompt_id: PromptId::new(),
839 title: if db_thread.title.is_empty() {
840 None
841 } else {
842 Some(db_thread.title.clone())
843 },
844 pending_title_generation: None,
845 pending_summary_generation: None,
846 summary: db_thread.detailed_summary,
847 messages: db_thread.messages,
848 user_store: project.read(cx).user_store(),
849 completion_mode: db_thread.completion_mode.unwrap_or_default(),
850 running_turn: None,
851 pending_message: None,
852 tools: BTreeMap::default(),
853 tool_use_limit_reached: false,
854 request_token_usage: db_thread.request_token_usage.clone(),
855 cumulative_token_usage: db_thread.cumulative_token_usage,
856 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
857 context_server_registry,
858 profile_id,
859 project_context,
860 templates,
861 model,
862 summarization_model: None,
863 project,
864 action_log,
865 updated_at: db_thread.updated_at,
866 prompt_capabilities_tx,
867 prompt_capabilities_rx,
868 }
869 }
870
871 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
872 let initial_project_snapshot = self.initial_project_snapshot.clone();
873 let mut thread = DbThread {
874 title: self.title(),
875 messages: self.messages.clone(),
876 updated_at: self.updated_at,
877 detailed_summary: self.summary.clone(),
878 initial_project_snapshot: None,
879 cumulative_token_usage: self.cumulative_token_usage,
880 request_token_usage: self.request_token_usage.clone(),
881 model: self.model.as_ref().map(|model| DbLanguageModel {
882 provider: model.provider_id().to_string(),
883 model: model.name().0.to_string(),
884 }),
885 completion_mode: Some(self.completion_mode),
886 profile: Some(self.profile_id.clone()),
887 };
888
889 cx.background_spawn(async move {
890 let initial_project_snapshot = initial_project_snapshot.await;
891 thread.initial_project_snapshot = initial_project_snapshot;
892 thread
893 })
894 }
895
896 /// Create a snapshot of the current project state including git information and unsaved buffers.
897 fn project_snapshot(
898 project: Entity<Project>,
899 cx: &mut Context<Self>,
900 ) -> Task<Arc<ProjectSnapshot>> {
901 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
902 cx.spawn(async move |_, _| {
903 let snapshot = task.await;
904
905 Arc::new(ProjectSnapshot {
906 worktree_snapshots: snapshot.worktree_snapshots,
907 timestamp: Utc::now(),
908 })
909 })
910 }
911
912 pub fn project_context(&self) -> &Entity<ProjectContext> {
913 &self.project_context
914 }
915
916 pub fn project(&self) -> &Entity<Project> {
917 &self.project
918 }
919
920 pub fn action_log(&self) -> &Entity<ActionLog> {
921 &self.action_log
922 }
923
924 pub fn is_empty(&self) -> bool {
925 self.messages.is_empty() && self.title.is_none()
926 }
927
928 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
929 self.model.as_ref()
930 }
931
932 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
933 let old_usage = self.latest_token_usage();
934 self.model = Some(model);
935 let new_caps = Self::prompt_capabilities(self.model.as_deref());
936 let new_usage = self.latest_token_usage();
937 if old_usage != new_usage {
938 cx.emit(TokenUsageUpdated(new_usage));
939 }
940 self.prompt_capabilities_tx.send(new_caps).log_err();
941 cx.notify()
942 }
943
944 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
945 self.summarization_model.as_ref()
946 }
947
948 pub fn set_summarization_model(
949 &mut self,
950 model: Option<Arc<dyn LanguageModel>>,
951 cx: &mut Context<Self>,
952 ) {
953 self.summarization_model = model;
954 cx.notify()
955 }
956
957 pub fn completion_mode(&self) -> CompletionMode {
958 self.completion_mode
959 }
960
961 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
962 let old_usage = self.latest_token_usage();
963 self.completion_mode = mode;
964 let new_usage = self.latest_token_usage();
965 if old_usage != new_usage {
966 cx.emit(TokenUsageUpdated(new_usage));
967 }
968 cx.notify()
969 }
970
971 #[cfg(any(test, feature = "test-support"))]
972 pub fn last_message(&self) -> Option<Message> {
973 if let Some(message) = self.pending_message.clone() {
974 Some(Message::Agent(message))
975 } else {
976 self.messages.last().cloned()
977 }
978 }
979
980 pub fn add_default_tools(
981 &mut self,
982 environment: Rc<dyn ThreadEnvironment>,
983 cx: &mut Context<Self>,
984 ) {
985 let language_registry = self.project.read(cx).languages().clone();
986 self.add_tool(CopyPathTool::new(self.project.clone()));
987 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
988 self.add_tool(DeletePathTool::new(
989 self.project.clone(),
990 self.action_log.clone(),
991 ));
992 self.add_tool(DiagnosticsTool::new(self.project.clone()));
993 self.add_tool(EditFileTool::new(
994 self.project.clone(),
995 cx.weak_entity(),
996 language_registry,
997 Templates::new(),
998 ));
999 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1000 self.add_tool(FindPathTool::new(self.project.clone()));
1001 self.add_tool(GrepTool::new(self.project.clone()));
1002 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1003 self.add_tool(MovePathTool::new(self.project.clone()));
1004 self.add_tool(NowTool);
1005 self.add_tool(OpenTool::new(self.project.clone()));
1006 self.add_tool(ReadFileTool::new(
1007 self.project.clone(),
1008 self.action_log.clone(),
1009 ));
1010 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1011 self.add_tool(ThinkingTool);
1012 self.add_tool(WebSearchTool);
1013 }
1014
1015 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1016 self.tools.insert(T::name().into(), tool.erase());
1017 }
1018
1019 pub fn remove_tool(&mut self, name: &str) -> bool {
1020 self.tools.remove(name).is_some()
1021 }
1022
1023 pub fn profile(&self) -> &AgentProfileId {
1024 &self.profile_id
1025 }
1026
1027 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1028 if self.profile_id == profile_id {
1029 return;
1030 }
1031
1032 self.profile_id = profile_id;
1033
1034 // Swap to the profile's preferred model when available.
1035 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1036 self.set_model(model, cx);
1037 }
1038 }
1039
1040 pub fn cancel(&mut self, cx: &mut Context<Self>) {
1041 if let Some(running_turn) = self.running_turn.take() {
1042 running_turn.cancel();
1043 }
1044 self.flush_pending_message(cx);
1045 }
1046
1047 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1048 let Some(last_user_message) = self.last_user_message() else {
1049 return;
1050 };
1051
1052 self.request_token_usage
1053 .insert(last_user_message.id.clone(), update);
1054 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1055 cx.notify();
1056 }
1057
1058 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1059 self.cancel(cx);
1060 let Some(position) = self.messages.iter().position(
1061 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1062 ) else {
1063 return Err(anyhow!("Message not found"));
1064 };
1065
1066 for message in self.messages.drain(position..) {
1067 match message {
1068 Message::User(message) => {
1069 self.request_token_usage.remove(&message.id);
1070 }
1071 Message::Agent(_) | Message::Resume => {}
1072 }
1073 }
1074 self.clear_summary();
1075 cx.notify();
1076 Ok(())
1077 }
1078
1079 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1080 let last_user_message = self.last_user_message()?;
1081 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1082 Some(*tokens)
1083 }
1084
1085 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1086 let usage = self.latest_request_token_usage()?;
1087 let model = self.model.clone()?;
1088 Some(acp_thread::TokenUsage {
1089 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1090 used_tokens: usage.total_tokens(),
1091 })
1092 }
1093
1094 /// Look up the active profile and resolve its preferred model if one is configured.
1095 fn resolve_profile_model(
1096 profile_id: &AgentProfileId,
1097 cx: &mut Context<Self>,
1098 ) -> Option<Arc<dyn LanguageModel>> {
1099 let selection = AgentSettings::get_global(cx)
1100 .profiles
1101 .get(profile_id)?
1102 .default_model
1103 .clone()?;
1104 Self::resolve_model_from_selection(&selection, cx)
1105 }
1106
1107 /// Translate a stored model selection into the configured model from the registry.
1108 fn resolve_model_from_selection(
1109 selection: &LanguageModelSelection,
1110 cx: &mut Context<Self>,
1111 ) -> Option<Arc<dyn LanguageModel>> {
1112 let selected = SelectedModel {
1113 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1114 model: LanguageModelId::from(selection.model.clone()),
1115 };
1116 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1117 registry
1118 .select_model(&selected, cx)
1119 .map(|configured| configured.model)
1120 })
1121 }
1122
1123 pub fn resume(
1124 &mut self,
1125 cx: &mut Context<Self>,
1126 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1127 self.messages.push(Message::Resume);
1128 cx.notify();
1129
1130 log::debug!("Total messages in thread: {}", self.messages.len());
1131 self.run_turn(cx)
1132 }
1133
1134 /// Sending a message results in the model streaming a response, which could include tool calls.
1135 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1136 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1137 pub fn send<T>(
1138 &mut self,
1139 id: UserMessageId,
1140 content: impl IntoIterator<Item = T>,
1141 cx: &mut Context<Self>,
1142 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1143 where
1144 T: Into<UserMessageContent>,
1145 {
1146 let model = self.model().context("No language model configured")?;
1147
1148 log::info!("Thread::send called with model: {}", model.name().0);
1149 self.advance_prompt_id();
1150
1151 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1152 log::debug!("Thread::send content: {:?}", content);
1153
1154 self.messages
1155 .push(Message::User(UserMessage { id, content }));
1156 cx.notify();
1157
1158 log::debug!("Total messages in thread: {}", self.messages.len());
1159 self.run_turn(cx)
1160 }
1161
1162 #[cfg(feature = "eval")]
1163 pub fn proceed(
1164 &mut self,
1165 cx: &mut Context<Self>,
1166 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1167 self.run_turn(cx)
1168 }
1169
1170 fn run_turn(
1171 &mut self,
1172 cx: &mut Context<Self>,
1173 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1174 self.cancel(cx);
1175
1176 let model = self.model.clone().context("No language model configured")?;
1177 let profile = AgentSettings::get_global(cx)
1178 .profiles
1179 .get(&self.profile_id)
1180 .context("Profile not found")?;
1181 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1182 let event_stream = ThreadEventStream(events_tx);
1183 let message_ix = self.messages.len().saturating_sub(1);
1184 self.tool_use_limit_reached = false;
1185 self.clear_summary();
1186 self.running_turn = Some(RunningTurn {
1187 event_stream: event_stream.clone(),
1188 tools: self.enabled_tools(profile, &model, cx),
1189 _task: cx.spawn(async move |this, cx| {
1190 log::debug!("Starting agent turn execution");
1191
1192 let turn_result = Self::run_turn_internal(&this, model, &event_stream, cx).await;
1193 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1194
1195 match turn_result {
1196 Ok(()) => {
1197 log::debug!("Turn execution completed");
1198 event_stream.send_stop(acp::StopReason::EndTurn);
1199 }
1200 Err(error) => {
1201 log::error!("Turn execution failed: {:?}", error);
1202 match error.downcast::<CompletionError>() {
1203 Ok(CompletionError::Refusal) => {
1204 event_stream.send_stop(acp::StopReason::Refusal);
1205 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1206 }
1207 Ok(CompletionError::MaxTokens) => {
1208 event_stream.send_stop(acp::StopReason::MaxTokens);
1209 }
1210 Ok(CompletionError::Other(error)) | Err(error) => {
1211 event_stream.send_error(error);
1212 }
1213 }
1214 }
1215 }
1216
1217 _ = this.update(cx, |this, _| this.running_turn.take());
1218 }),
1219 });
1220 Ok(events_rx)
1221 }
1222
1223 async fn run_turn_internal(
1224 this: &WeakEntity<Self>,
1225 model: Arc<dyn LanguageModel>,
1226 event_stream: &ThreadEventStream,
1227 cx: &mut AsyncApp,
1228 ) -> Result<()> {
1229 let mut attempt = 0;
1230 let mut intent = CompletionIntent::UserPrompt;
1231 loop {
1232 let request =
1233 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1234
1235 telemetry::event!(
1236 "Agent Thread Completion",
1237 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1238 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1239 model = model.telemetry_id(),
1240 model_provider = model.provider_id().to_string(),
1241 attempt
1242 );
1243
1244 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1245
1246 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1247 Ok(events) => (events, None),
1248 Err(err) => (stream::empty().boxed(), Some(err)),
1249 };
1250 let mut tool_results = FuturesUnordered::new();
1251 while let Some(event) = events.next().await {
1252 log::trace!("Received completion event: {:?}", event);
1253 match event {
1254 Ok(event) => {
1255 tool_results.extend(this.update(cx, |this, cx| {
1256 this.handle_completion_event(event, event_stream, cx)
1257 })??);
1258 }
1259 Err(err) => {
1260 error = Some(err);
1261 break;
1262 }
1263 }
1264 }
1265
1266 let end_turn = tool_results.is_empty();
1267 while let Some(tool_result) = tool_results.next().await {
1268 log::debug!("Tool finished {:?}", tool_result);
1269
1270 event_stream.update_tool_call_fields(
1271 &tool_result.tool_use_id,
1272 acp::ToolCallUpdateFields {
1273 status: Some(if tool_result.is_error {
1274 acp::ToolCallStatus::Failed
1275 } else {
1276 acp::ToolCallStatus::Completed
1277 }),
1278 raw_output: tool_result.output.clone(),
1279 ..Default::default()
1280 },
1281 );
1282 this.update(cx, |this, _cx| {
1283 this.pending_message()
1284 .tool_results
1285 .insert(tool_result.tool_use_id.clone(), tool_result);
1286 })?;
1287 }
1288
1289 this.update(cx, |this, cx| {
1290 this.flush_pending_message(cx);
1291 if this.title.is_none() && this.pending_title_generation.is_none() {
1292 this.generate_title(cx);
1293 }
1294 })?;
1295
1296 if let Some(error) = error {
1297 attempt += 1;
1298 let retry = this.update(cx, |this, cx| {
1299 let user_store = this.user_store.read(cx);
1300 this.handle_completion_error(error, attempt, user_store.plan())
1301 })??;
1302 let timer = cx.background_executor().timer(retry.duration);
1303 event_stream.send_retry(retry);
1304 timer.await;
1305 this.update(cx, |this, _cx| {
1306 if let Some(Message::Agent(message)) = this.messages.last() {
1307 if message.tool_results.is_empty() {
1308 intent = CompletionIntent::UserPrompt;
1309 this.messages.push(Message::Resume);
1310 }
1311 }
1312 })?;
1313 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1314 return Err(language_model::ToolUseLimitReachedError.into());
1315 } else if end_turn {
1316 return Ok(());
1317 } else {
1318 intent = CompletionIntent::ToolResults;
1319 attempt = 0;
1320 }
1321 }
1322 }
1323
1324 fn handle_completion_error(
1325 &mut self,
1326 error: LanguageModelCompletionError,
1327 attempt: u8,
1328 plan: Option<Plan>,
1329 ) -> Result<acp_thread::RetryStatus> {
1330 let Some(model) = self.model.as_ref() else {
1331 return Err(anyhow!(error));
1332 };
1333
1334 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1335 match plan {
1336 Some(Plan::V2(_)) => true,
1337 Some(Plan::V1(_)) => self.completion_mode == CompletionMode::Burn,
1338 None => false,
1339 }
1340 } else {
1341 true
1342 };
1343
1344 if !auto_retry {
1345 return Err(anyhow!(error));
1346 }
1347
1348 let Some(strategy) = Self::retry_strategy_for(&error) else {
1349 return Err(anyhow!(error));
1350 };
1351
1352 let max_attempts = match &strategy {
1353 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1354 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1355 };
1356
1357 if attempt > max_attempts {
1358 return Err(anyhow!(error));
1359 }
1360
1361 let delay = match &strategy {
1362 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1363 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1364 Duration::from_secs(delay_secs)
1365 }
1366 RetryStrategy::Fixed { delay, .. } => *delay,
1367 };
1368 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1369
1370 Ok(acp_thread::RetryStatus {
1371 last_error: error.to_string().into(),
1372 attempt: attempt as usize,
1373 max_attempts: max_attempts as usize,
1374 started_at: Instant::now(),
1375 duration: delay,
1376 })
1377 }
1378
1379 /// A helper method that's called on every streamed completion event.
1380 /// Returns an optional tool result task, which the main agentic loop will
1381 /// send back to the model when it resolves.
1382 fn handle_completion_event(
1383 &mut self,
1384 event: LanguageModelCompletionEvent,
1385 event_stream: &ThreadEventStream,
1386 cx: &mut Context<Self>,
1387 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1388 log::trace!("Handling streamed completion event: {:?}", event);
1389 use LanguageModelCompletionEvent::*;
1390
1391 match event {
1392 StartMessage { .. } => {
1393 self.flush_pending_message(cx);
1394 self.pending_message = Some(AgentMessage::default());
1395 }
1396 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1397 Thinking { text, signature } => {
1398 self.handle_thinking_event(text, signature, event_stream, cx)
1399 }
1400 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1401 ReasoningDetails(details) => {
1402 let last_message = self.pending_message();
1403 // Store the last non-empty reasoning_details (overwrites earlier ones)
1404 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
1405 if let serde_json::Value::Array(ref arr) = details {
1406 if !arr.is_empty() {
1407 last_message.reasoning_details = Some(details);
1408 }
1409 } else {
1410 last_message.reasoning_details = Some(details);
1411 }
1412 }
1413 ToolUse(tool_use) => {
1414 return Ok(self.handle_tool_use_event(tool_use, event_stream, cx));
1415 }
1416 ToolUseJsonParseError {
1417 id,
1418 tool_name,
1419 raw_input,
1420 json_parse_error,
1421 } => {
1422 return Ok(Some(Task::ready(
1423 self.handle_tool_use_json_parse_error_event(
1424 id,
1425 tool_name,
1426 raw_input,
1427 json_parse_error,
1428 ),
1429 )));
1430 }
1431 UsageUpdate(usage) => {
1432 telemetry::event!(
1433 "Agent Thread Completion Usage Updated",
1434 thread_id = self.id.to_string(),
1435 prompt_id = self.prompt_id.to_string(),
1436 model = self.model.as_ref().map(|m| m.telemetry_id()),
1437 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1438 input_tokens = usage.input_tokens,
1439 output_tokens = usage.output_tokens,
1440 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1441 cache_read_input_tokens = usage.cache_read_input_tokens,
1442 );
1443 self.update_token_usage(usage, cx);
1444 }
1445 StatusUpdate(CompletionRequestStatus::UsageUpdated { amount, limit }) => {
1446 self.update_model_request_usage(amount, limit, cx);
1447 }
1448 StatusUpdate(
1449 CompletionRequestStatus::Started
1450 | CompletionRequestStatus::Queued { .. }
1451 | CompletionRequestStatus::Failed { .. },
1452 ) => {}
1453 StatusUpdate(CompletionRequestStatus::ToolUseLimitReached) => {
1454 self.tool_use_limit_reached = true;
1455 }
1456 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1457 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1458 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1459 }
1460
1461 Ok(None)
1462 }
1463
1464 fn handle_text_event(
1465 &mut self,
1466 new_text: String,
1467 event_stream: &ThreadEventStream,
1468 cx: &mut Context<Self>,
1469 ) {
1470 event_stream.send_text(&new_text);
1471
1472 let last_message = self.pending_message();
1473 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1474 text.push_str(&new_text);
1475 } else {
1476 last_message
1477 .content
1478 .push(AgentMessageContent::Text(new_text));
1479 }
1480
1481 cx.notify();
1482 }
1483
1484 fn handle_thinking_event(
1485 &mut self,
1486 new_text: String,
1487 new_signature: Option<String>,
1488 event_stream: &ThreadEventStream,
1489 cx: &mut Context<Self>,
1490 ) {
1491 event_stream.send_thinking(&new_text);
1492
1493 let last_message = self.pending_message();
1494 if let Some(AgentMessageContent::Thinking { text, signature }) =
1495 last_message.content.last_mut()
1496 {
1497 text.push_str(&new_text);
1498 *signature = new_signature.or(signature.take());
1499 } else {
1500 last_message.content.push(AgentMessageContent::Thinking {
1501 text: new_text,
1502 signature: new_signature,
1503 });
1504 }
1505
1506 cx.notify();
1507 }
1508
1509 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1510 let last_message = self.pending_message();
1511 last_message
1512 .content
1513 .push(AgentMessageContent::RedactedThinking(data));
1514 cx.notify();
1515 }
1516
1517 fn handle_tool_use_event(
1518 &mut self,
1519 tool_use: LanguageModelToolUse,
1520 event_stream: &ThreadEventStream,
1521 cx: &mut Context<Self>,
1522 ) -> Option<Task<LanguageModelToolResult>> {
1523 cx.notify();
1524
1525 let tool = self.tool(tool_use.name.as_ref());
1526 let mut title = SharedString::from(&tool_use.name);
1527 let mut kind = acp::ToolKind::Other;
1528 if let Some(tool) = tool.as_ref() {
1529 title = tool.initial_title(tool_use.input.clone(), cx);
1530 kind = tool.kind();
1531 }
1532
1533 // Ensure the last message ends in the current tool use
1534 let last_message = self.pending_message();
1535 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1536 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1537 if last_tool_use.id == tool_use.id {
1538 *last_tool_use = tool_use.clone();
1539 false
1540 } else {
1541 true
1542 }
1543 } else {
1544 true
1545 }
1546 });
1547
1548 if push_new_tool_use {
1549 event_stream.send_tool_call(
1550 &tool_use.id,
1551 &tool_use.name,
1552 title,
1553 kind,
1554 tool_use.input.clone(),
1555 );
1556 last_message
1557 .content
1558 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1559 } else {
1560 event_stream.update_tool_call_fields(
1561 &tool_use.id,
1562 acp::ToolCallUpdateFields {
1563 title: Some(title.into()),
1564 kind: Some(kind),
1565 raw_input: Some(tool_use.input.clone()),
1566 ..Default::default()
1567 },
1568 );
1569 }
1570
1571 if !tool_use.is_input_complete {
1572 return None;
1573 }
1574
1575 let Some(tool) = tool else {
1576 let content = format!("No tool named {} exists", tool_use.name);
1577 return Some(Task::ready(LanguageModelToolResult {
1578 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1579 tool_use_id: tool_use.id,
1580 tool_name: tool_use.name,
1581 is_error: true,
1582 output: None,
1583 }));
1584 };
1585
1586 let fs = self.project.read(cx).fs().clone();
1587 let tool_event_stream =
1588 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1589 tool_event_stream.update_fields(acp::ToolCallUpdateFields {
1590 status: Some(acp::ToolCallStatus::InProgress),
1591 ..Default::default()
1592 });
1593 let supports_images = self.model().is_some_and(|model| model.supports_images());
1594 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1595 log::debug!("Running tool {}", tool_use.name);
1596 Some(cx.foreground_executor().spawn(async move {
1597 let tool_result = tool_result.await.and_then(|output| {
1598 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1599 && !supports_images
1600 {
1601 return Err(anyhow!(
1602 "Attempted to read an image, but this model doesn't support it.",
1603 ));
1604 }
1605 Ok(output)
1606 });
1607
1608 match tool_result {
1609 Ok(output) => LanguageModelToolResult {
1610 tool_use_id: tool_use.id,
1611 tool_name: tool_use.name,
1612 is_error: false,
1613 content: output.llm_output,
1614 output: Some(output.raw_output),
1615 },
1616 Err(error) => LanguageModelToolResult {
1617 tool_use_id: tool_use.id,
1618 tool_name: tool_use.name,
1619 is_error: true,
1620 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1621 output: Some(error.to_string().into()),
1622 },
1623 }
1624 }))
1625 }
1626
1627 fn handle_tool_use_json_parse_error_event(
1628 &mut self,
1629 tool_use_id: LanguageModelToolUseId,
1630 tool_name: Arc<str>,
1631 raw_input: Arc<str>,
1632 json_parse_error: String,
1633 ) -> LanguageModelToolResult {
1634 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1635 LanguageModelToolResult {
1636 tool_use_id,
1637 tool_name,
1638 is_error: true,
1639 content: LanguageModelToolResultContent::Text(tool_output.into()),
1640 output: Some(serde_json::Value::String(raw_input.to_string())),
1641 }
1642 }
1643
1644 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1645 self.project
1646 .read(cx)
1647 .user_store()
1648 .update(cx, |user_store, cx| {
1649 user_store.update_model_request_usage(
1650 ModelRequestUsage(RequestUsage {
1651 amount: amount as i32,
1652 limit,
1653 }),
1654 cx,
1655 )
1656 });
1657 }
1658
1659 pub fn title(&self) -> SharedString {
1660 self.title.clone().unwrap_or("New Thread".into())
1661 }
1662
1663 pub fn is_generating_summary(&self) -> bool {
1664 self.pending_summary_generation.is_some()
1665 }
1666
1667 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
1668 if let Some(summary) = self.summary.as_ref() {
1669 return Task::ready(Some(summary.clone())).shared();
1670 }
1671 if let Some(task) = self.pending_summary_generation.clone() {
1672 return task;
1673 }
1674 let Some(model) = self.summarization_model.clone() else {
1675 log::error!("No summarization model available");
1676 return Task::ready(None).shared();
1677 };
1678 let mut request = LanguageModelRequest {
1679 intent: Some(CompletionIntent::ThreadContextSummarization),
1680 temperature: AgentSettings::temperature_for_model(&model, cx),
1681 ..Default::default()
1682 };
1683
1684 for message in &self.messages {
1685 request.messages.extend(message.to_request());
1686 }
1687
1688 request.messages.push(LanguageModelRequestMessage {
1689 role: Role::User,
1690 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1691 cache: false,
1692 reasoning_details: None,
1693 });
1694
1695 let task = cx
1696 .spawn(async move |this, cx| {
1697 let mut summary = String::new();
1698 let mut messages = model.stream_completion(request, cx).await.log_err()?;
1699 while let Some(event) = messages.next().await {
1700 let event = event.log_err()?;
1701 let text = match event {
1702 LanguageModelCompletionEvent::Text(text) => text,
1703 LanguageModelCompletionEvent::StatusUpdate(
1704 CompletionRequestStatus::UsageUpdated { amount, limit },
1705 ) => {
1706 this.update(cx, |thread, cx| {
1707 thread.update_model_request_usage(amount, limit, cx);
1708 })
1709 .ok()?;
1710 continue;
1711 }
1712 _ => continue,
1713 };
1714
1715 let mut lines = text.lines();
1716 summary.extend(lines.next());
1717 }
1718
1719 log::debug!("Setting summary: {}", summary);
1720 let summary = SharedString::from(summary);
1721
1722 this.update(cx, |this, cx| {
1723 this.summary = Some(summary.clone());
1724 this.pending_summary_generation = None;
1725 cx.notify()
1726 })
1727 .ok()?;
1728
1729 Some(summary)
1730 })
1731 .shared();
1732 self.pending_summary_generation = Some(task.clone());
1733 task
1734 }
1735
1736 fn generate_title(&mut self, cx: &mut Context<Self>) {
1737 let Some(model) = self.summarization_model.clone() else {
1738 return;
1739 };
1740
1741 log::debug!(
1742 "Generating title with model: {:?}",
1743 self.summarization_model.as_ref().map(|model| model.name())
1744 );
1745 let mut request = LanguageModelRequest {
1746 intent: Some(CompletionIntent::ThreadSummarization),
1747 temperature: AgentSettings::temperature_for_model(&model, cx),
1748 ..Default::default()
1749 };
1750
1751 for message in &self.messages {
1752 request.messages.extend(message.to_request());
1753 }
1754
1755 request.messages.push(LanguageModelRequestMessage {
1756 role: Role::User,
1757 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1758 cache: false,
1759 reasoning_details: None,
1760 });
1761 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
1762 let mut title = String::new();
1763
1764 let generate = async {
1765 let mut messages = model.stream_completion(request, cx).await?;
1766 while let Some(event) = messages.next().await {
1767 let event = event?;
1768 let text = match event {
1769 LanguageModelCompletionEvent::Text(text) => text,
1770 LanguageModelCompletionEvent::StatusUpdate(
1771 CompletionRequestStatus::UsageUpdated { amount, limit },
1772 ) => {
1773 this.update(cx, |thread, cx| {
1774 thread.update_model_request_usage(amount, limit, cx);
1775 })?;
1776 continue;
1777 }
1778 _ => continue,
1779 };
1780
1781 let mut lines = text.lines();
1782 title.extend(lines.next());
1783
1784 // Stop if the LLM generated multiple lines.
1785 if lines.next().is_some() {
1786 break;
1787 }
1788 }
1789 anyhow::Ok(())
1790 };
1791
1792 if generate.await.context("failed to generate title").is_ok() {
1793 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
1794 }
1795 _ = this.update(cx, |this, _| this.pending_title_generation = None);
1796 }));
1797 }
1798
1799 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
1800 self.pending_title_generation = None;
1801 if Some(&title) != self.title.as_ref() {
1802 self.title = Some(title);
1803 cx.emit(TitleUpdated);
1804 cx.notify();
1805 }
1806 }
1807
1808 fn clear_summary(&mut self) {
1809 self.summary = None;
1810 self.pending_summary_generation = None;
1811 }
1812
1813 fn last_user_message(&self) -> Option<&UserMessage> {
1814 self.messages
1815 .iter()
1816 .rev()
1817 .find_map(|message| match message {
1818 Message::User(user_message) => Some(user_message),
1819 Message::Agent(_) => None,
1820 Message::Resume => None,
1821 })
1822 }
1823
1824 fn pending_message(&mut self) -> &mut AgentMessage {
1825 self.pending_message.get_or_insert_default()
1826 }
1827
1828 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1829 let Some(mut message) = self.pending_message.take() else {
1830 return;
1831 };
1832
1833 if message.content.is_empty() {
1834 return;
1835 }
1836
1837 for content in &message.content {
1838 let AgentMessageContent::ToolUse(tool_use) = content else {
1839 continue;
1840 };
1841
1842 if !message.tool_results.contains_key(&tool_use.id) {
1843 message.tool_results.insert(
1844 tool_use.id.clone(),
1845 LanguageModelToolResult {
1846 tool_use_id: tool_use.id.clone(),
1847 tool_name: tool_use.name.clone(),
1848 is_error: true,
1849 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1850 output: None,
1851 },
1852 );
1853 }
1854 }
1855
1856 self.messages.push(Message::Agent(message));
1857 self.updated_at = Utc::now();
1858 self.clear_summary();
1859 cx.notify()
1860 }
1861
1862 pub(crate) fn build_completion_request(
1863 &self,
1864 completion_intent: CompletionIntent,
1865 cx: &App,
1866 ) -> Result<LanguageModelRequest> {
1867 let model = self.model().context("No language model configured")?;
1868 let tools = if let Some(turn) = self.running_turn.as_ref() {
1869 turn.tools
1870 .iter()
1871 .filter_map(|(tool_name, tool)| {
1872 log::trace!("Including tool: {}", tool_name);
1873 Some(LanguageModelRequestTool {
1874 name: tool_name.to_string(),
1875 description: tool.description().to_string(),
1876 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1877 })
1878 })
1879 .collect::<Vec<_>>()
1880 } else {
1881 Vec::new()
1882 };
1883
1884 log::debug!("Building completion request");
1885 log::debug!("Completion intent: {:?}", completion_intent);
1886 log::debug!("Completion mode: {:?}", self.completion_mode);
1887
1888 let available_tools: Vec<_> = self
1889 .running_turn
1890 .as_ref()
1891 .map(|turn| turn.tools.keys().cloned().collect())
1892 .unwrap_or_default();
1893
1894 log::debug!("Request includes {} tools", available_tools.len());
1895 let messages = self.build_request_messages(available_tools, cx);
1896 log::debug!("Request will include {} messages", messages.len());
1897
1898 let request = LanguageModelRequest {
1899 thread_id: Some(self.id.to_string()),
1900 prompt_id: Some(self.prompt_id.to_string()),
1901 intent: Some(completion_intent),
1902 mode: Some(self.completion_mode.into()),
1903 messages,
1904 tools,
1905 tool_choice: None,
1906 stop: Vec::new(),
1907 temperature: AgentSettings::temperature_for_model(model, cx),
1908 thinking_allowed: true,
1909 };
1910
1911 log::debug!("Completion request built successfully");
1912 Ok(request)
1913 }
1914
1915 fn enabled_tools(
1916 &self,
1917 profile: &AgentProfileSettings,
1918 model: &Arc<dyn LanguageModel>,
1919 cx: &App,
1920 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
1921 fn truncate(tool_name: &SharedString) -> SharedString {
1922 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
1923 let mut truncated = tool_name.to_string();
1924 truncated.truncate(MAX_TOOL_NAME_LENGTH);
1925 truncated.into()
1926 } else {
1927 tool_name.clone()
1928 }
1929 }
1930
1931 let mut tools = self
1932 .tools
1933 .iter()
1934 .filter_map(|(tool_name, tool)| {
1935 if tool.supports_provider(&model.provider_id())
1936 && profile.is_tool_enabled(tool_name)
1937 {
1938 Some((truncate(tool_name), tool.clone()))
1939 } else {
1940 None
1941 }
1942 })
1943 .collect::<BTreeMap<_, _>>();
1944
1945 let mut context_server_tools = Vec::new();
1946 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
1947 let mut duplicate_tool_names = HashSet::default();
1948 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
1949 for (tool_name, tool) in server_tools {
1950 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
1951 let tool_name = truncate(tool_name);
1952 if !seen_tools.insert(tool_name.clone()) {
1953 duplicate_tool_names.insert(tool_name.clone());
1954 }
1955 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
1956 }
1957 }
1958 }
1959
1960 // When there are duplicate tool names, disambiguate by prefixing them
1961 // with the server ID. In the rare case there isn't enough space for the
1962 // disambiguated tool name, keep only the last tool with this name.
1963 for (server_id, tool_name, tool) in context_server_tools {
1964 if duplicate_tool_names.contains(&tool_name) {
1965 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
1966 if available >= 2 {
1967 let mut disambiguated = server_id.0.to_string();
1968 disambiguated.truncate(available - 1);
1969 disambiguated.push('_');
1970 disambiguated.push_str(&tool_name);
1971 tools.insert(disambiguated.into(), tool.clone());
1972 } else {
1973 tools.insert(tool_name, tool.clone());
1974 }
1975 } else {
1976 tools.insert(tool_name, tool.clone());
1977 }
1978 }
1979
1980 tools
1981 }
1982
1983 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
1984 self.running_turn.as_ref()?.tools.get(name).cloned()
1985 }
1986
1987 fn build_request_messages(
1988 &self,
1989 available_tools: Vec<SharedString>,
1990 cx: &App,
1991 ) -> Vec<LanguageModelRequestMessage> {
1992 log::trace!(
1993 "Building request messages from {} thread messages",
1994 self.messages.len()
1995 );
1996
1997 let system_prompt = SystemPromptTemplate {
1998 project: self.project_context.read(cx),
1999 available_tools,
2000 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2001 }
2002 .render(&self.templates)
2003 .context("failed to build system prompt")
2004 .expect("Invalid template");
2005 let mut messages = vec![LanguageModelRequestMessage {
2006 role: Role::System,
2007 content: vec![system_prompt.into()],
2008 cache: false,
2009 reasoning_details: None,
2010 }];
2011 for message in &self.messages {
2012 messages.extend(message.to_request());
2013 }
2014
2015 if let Some(last_message) = messages.last_mut() {
2016 last_message.cache = true;
2017 }
2018
2019 if let Some(message) = self.pending_message.as_ref() {
2020 messages.extend(message.to_request());
2021 }
2022
2023 messages
2024 }
2025
2026 pub fn to_markdown(&self) -> String {
2027 let mut markdown = String::new();
2028 for (ix, message) in self.messages.iter().enumerate() {
2029 if ix > 0 {
2030 markdown.push('\n');
2031 }
2032 markdown.push_str(&message.to_markdown());
2033 }
2034
2035 if let Some(message) = self.pending_message.as_ref() {
2036 markdown.push('\n');
2037 markdown.push_str(&message.to_markdown());
2038 }
2039
2040 markdown
2041 }
2042
2043 fn advance_prompt_id(&mut self) {
2044 self.prompt_id = PromptId::new();
2045 }
2046
2047 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2048 use LanguageModelCompletionError::*;
2049 use http_client::StatusCode;
2050
2051 // General strategy here:
2052 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2053 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2054 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2055 match error {
2056 HttpResponseError {
2057 status_code: StatusCode::TOO_MANY_REQUESTS,
2058 ..
2059 } => Some(RetryStrategy::ExponentialBackoff {
2060 initial_delay: BASE_RETRY_DELAY,
2061 max_attempts: MAX_RETRY_ATTEMPTS,
2062 }),
2063 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2064 Some(RetryStrategy::Fixed {
2065 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2066 max_attempts: MAX_RETRY_ATTEMPTS,
2067 })
2068 }
2069 UpstreamProviderError {
2070 status,
2071 retry_after,
2072 ..
2073 } => match *status {
2074 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2075 Some(RetryStrategy::Fixed {
2076 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2077 max_attempts: MAX_RETRY_ATTEMPTS,
2078 })
2079 }
2080 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2081 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2082 // Internal Server Error could be anything, retry up to 3 times.
2083 max_attempts: 3,
2084 }),
2085 status => {
2086 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2087 // but we frequently get them in practice. See https://http.dev/529
2088 if status.as_u16() == 529 {
2089 Some(RetryStrategy::Fixed {
2090 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2091 max_attempts: MAX_RETRY_ATTEMPTS,
2092 })
2093 } else {
2094 Some(RetryStrategy::Fixed {
2095 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2096 max_attempts: 2,
2097 })
2098 }
2099 }
2100 },
2101 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2102 delay: BASE_RETRY_DELAY,
2103 max_attempts: 3,
2104 }),
2105 ApiReadResponseError { .. }
2106 | HttpSend { .. }
2107 | DeserializeResponse { .. }
2108 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2109 delay: BASE_RETRY_DELAY,
2110 max_attempts: 3,
2111 }),
2112 // Retrying these errors definitely shouldn't help.
2113 HttpResponseError {
2114 status_code:
2115 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2116 ..
2117 }
2118 | AuthenticationError { .. }
2119 | PermissionError { .. }
2120 | NoApiKey { .. }
2121 | ApiEndpointNotFound { .. }
2122 | PromptTooLarge { .. } => None,
2123 // These errors might be transient, so retry them
2124 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2125 delay: BASE_RETRY_DELAY,
2126 max_attempts: 1,
2127 }),
2128 // Retry all other 4xx and 5xx errors once.
2129 HttpResponseError { status_code, .. }
2130 if status_code.is_client_error() || status_code.is_server_error() =>
2131 {
2132 Some(RetryStrategy::Fixed {
2133 delay: BASE_RETRY_DELAY,
2134 max_attempts: 3,
2135 })
2136 }
2137 Other(err)
2138 if err.is::<language_model::PaymentRequiredError>()
2139 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2140 {
2141 // Retrying won't help for Payment Required or Model Request Limit errors (where
2142 // the user must upgrade to usage-based billing to get more requests, or else wait
2143 // for a significant amount of time for the request limit to reset).
2144 None
2145 }
2146 // Conservatively assume that any other errors are non-retryable
2147 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2148 delay: BASE_RETRY_DELAY,
2149 max_attempts: 2,
2150 }),
2151 }
2152 }
2153}
2154
2155struct RunningTurn {
2156 /// Holds the task that handles agent interaction until the end of the turn.
2157 /// Survives across multiple requests as the model performs tool calls and
2158 /// we run tools, report their results.
2159 _task: Task<()>,
2160 /// The current event stream for the running turn. Used to report a final
2161 /// cancellation event if we cancel the turn.
2162 event_stream: ThreadEventStream,
2163 /// The tools that were enabled for this turn.
2164 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2165}
2166
2167impl RunningTurn {
2168 fn cancel(self) {
2169 log::debug!("Cancelling in progress turn");
2170 self.event_stream.send_canceled();
2171 }
2172}
2173
2174pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2175
2176impl EventEmitter<TokenUsageUpdated> for Thread {}
2177
2178pub struct TitleUpdated;
2179
2180impl EventEmitter<TitleUpdated> for Thread {}
2181
2182pub trait AgentTool
2183where
2184 Self: 'static + Sized,
2185{
2186 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2187 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2188
2189 fn name() -> &'static str;
2190
2191 fn description() -> SharedString {
2192 let schema = schemars::schema_for!(Self::Input);
2193 SharedString::new(
2194 schema
2195 .get("description")
2196 .and_then(|description| description.as_str())
2197 .unwrap_or_default(),
2198 )
2199 }
2200
2201 fn kind() -> acp::ToolKind;
2202
2203 /// The initial tool title to display. Can be updated during the tool run.
2204 fn initial_title(
2205 &self,
2206 input: Result<Self::Input, serde_json::Value>,
2207 cx: &mut App,
2208 ) -> SharedString;
2209
2210 /// Returns the JSON schema that describes the tool's input.
2211 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
2212 language_model::tool_schema::root_schema_for::<Self::Input>(format)
2213 }
2214
2215 /// Some tools rely on a provider for the underlying billing or other reasons.
2216 /// Allow the tool to check if they are compatible, or should be filtered out.
2217 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
2218 true
2219 }
2220
2221 /// Runs the tool with the provided input.
2222 fn run(
2223 self: Arc<Self>,
2224 input: Self::Input,
2225 event_stream: ToolCallEventStream,
2226 cx: &mut App,
2227 ) -> Task<Result<Self::Output>>;
2228
2229 /// Emits events for a previous execution of the tool.
2230 fn replay(
2231 &self,
2232 _input: Self::Input,
2233 _output: Self::Output,
2234 _event_stream: ToolCallEventStream,
2235 _cx: &mut App,
2236 ) -> Result<()> {
2237 Ok(())
2238 }
2239
2240 fn erase(self) -> Arc<dyn AnyAgentTool> {
2241 Arc::new(Erased(Arc::new(self)))
2242 }
2243}
2244
2245pub struct Erased<T>(T);
2246
2247pub struct AgentToolOutput {
2248 pub llm_output: LanguageModelToolResultContent,
2249 pub raw_output: serde_json::Value,
2250}
2251
2252pub trait AnyAgentTool {
2253 fn name(&self) -> SharedString;
2254 fn description(&self) -> SharedString;
2255 fn kind(&self) -> acp::ToolKind;
2256 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2257 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2258 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2259 true
2260 }
2261 fn run(
2262 self: Arc<Self>,
2263 input: serde_json::Value,
2264 event_stream: ToolCallEventStream,
2265 cx: &mut App,
2266 ) -> Task<Result<AgentToolOutput>>;
2267 fn replay(
2268 &self,
2269 input: serde_json::Value,
2270 output: serde_json::Value,
2271 event_stream: ToolCallEventStream,
2272 cx: &mut App,
2273 ) -> Result<()>;
2274}
2275
2276impl<T> AnyAgentTool for Erased<Arc<T>>
2277where
2278 T: AgentTool,
2279{
2280 fn name(&self) -> SharedString {
2281 T::name().into()
2282 }
2283
2284 fn description(&self) -> SharedString {
2285 T::description()
2286 }
2287
2288 fn kind(&self) -> agent_client_protocol::ToolKind {
2289 T::kind()
2290 }
2291
2292 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2293 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2294 self.0.initial_title(parsed_input, _cx)
2295 }
2296
2297 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2298 let mut json = serde_json::to_value(T::input_schema(format))?;
2299 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
2300 Ok(json)
2301 }
2302
2303 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
2304 T::supports_provider(provider)
2305 }
2306
2307 fn run(
2308 self: Arc<Self>,
2309 input: serde_json::Value,
2310 event_stream: ToolCallEventStream,
2311 cx: &mut App,
2312 ) -> Task<Result<AgentToolOutput>> {
2313 cx.spawn(async move |cx| {
2314 let input = serde_json::from_value(input)?;
2315 let output = cx
2316 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2317 .await?;
2318 let raw_output = serde_json::to_value(&output)?;
2319 Ok(AgentToolOutput {
2320 llm_output: output.into(),
2321 raw_output,
2322 })
2323 })
2324 }
2325
2326 fn replay(
2327 &self,
2328 input: serde_json::Value,
2329 output: serde_json::Value,
2330 event_stream: ToolCallEventStream,
2331 cx: &mut App,
2332 ) -> Result<()> {
2333 let input = serde_json::from_value(input)?;
2334 let output = serde_json::from_value(output)?;
2335 self.0.replay(input, output, event_stream, cx)
2336 }
2337}
2338
2339#[derive(Clone)]
2340struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2341
2342impl ThreadEventStream {
2343 fn send_user_message(&self, message: &UserMessage) {
2344 self.0
2345 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2346 .ok();
2347 }
2348
2349 fn send_text(&self, text: &str) {
2350 self.0
2351 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2352 .ok();
2353 }
2354
2355 fn send_thinking(&self, text: &str) {
2356 self.0
2357 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2358 .ok();
2359 }
2360
2361 fn send_tool_call(
2362 &self,
2363 id: &LanguageModelToolUseId,
2364 tool_name: &str,
2365 title: SharedString,
2366 kind: acp::ToolKind,
2367 input: serde_json::Value,
2368 ) {
2369 self.0
2370 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2371 id,
2372 tool_name,
2373 title.to_string(),
2374 kind,
2375 input,
2376 ))))
2377 .ok();
2378 }
2379
2380 fn initial_tool_call(
2381 id: &LanguageModelToolUseId,
2382 tool_name: &str,
2383 title: String,
2384 kind: acp::ToolKind,
2385 input: serde_json::Value,
2386 ) -> acp::ToolCall {
2387 acp::ToolCall {
2388 meta: Some(serde_json::json!({
2389 "tool_name": tool_name
2390 })),
2391 id: acp::ToolCallId(id.to_string().into()),
2392 title,
2393 kind,
2394 status: acp::ToolCallStatus::Pending,
2395 content: vec![],
2396 locations: vec![],
2397 raw_input: Some(input),
2398 raw_output: None,
2399 }
2400 }
2401
2402 fn update_tool_call_fields(
2403 &self,
2404 tool_use_id: &LanguageModelToolUseId,
2405 fields: acp::ToolCallUpdateFields,
2406 ) {
2407 self.0
2408 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2409 acp::ToolCallUpdate {
2410 meta: None,
2411 id: acp::ToolCallId(tool_use_id.to_string().into()),
2412 fields,
2413 }
2414 .into(),
2415 )))
2416 .ok();
2417 }
2418
2419 fn send_retry(&self, status: acp_thread::RetryStatus) {
2420 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2421 }
2422
2423 fn send_stop(&self, reason: acp::StopReason) {
2424 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2425 }
2426
2427 fn send_canceled(&self) {
2428 self.0
2429 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2430 .ok();
2431 }
2432
2433 fn send_error(&self, error: impl Into<anyhow::Error>) {
2434 self.0.unbounded_send(Err(error.into())).ok();
2435 }
2436}
2437
2438#[derive(Clone)]
2439pub struct ToolCallEventStream {
2440 tool_use_id: LanguageModelToolUseId,
2441 stream: ThreadEventStream,
2442 fs: Option<Arc<dyn Fs>>,
2443}
2444
2445impl ToolCallEventStream {
2446 #[cfg(any(test, feature = "test-support"))]
2447 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2448 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2449
2450 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2451
2452 (stream, ToolCallEventStreamReceiver(events_rx))
2453 }
2454
2455 fn new(
2456 tool_use_id: LanguageModelToolUseId,
2457 stream: ThreadEventStream,
2458 fs: Option<Arc<dyn Fs>>,
2459 ) -> Self {
2460 Self {
2461 tool_use_id,
2462 stream,
2463 fs,
2464 }
2465 }
2466
2467 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2468 self.stream
2469 .update_tool_call_fields(&self.tool_use_id, fields);
2470 }
2471
2472 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2473 self.stream
2474 .0
2475 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2476 acp_thread::ToolCallUpdateDiff {
2477 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2478 diff,
2479 }
2480 .into(),
2481 )))
2482 .ok();
2483 }
2484
2485 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2486 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2487 return Task::ready(Ok(()));
2488 }
2489
2490 let (response_tx, response_rx) = oneshot::channel();
2491 self.stream
2492 .0
2493 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2494 ToolCallAuthorization {
2495 tool_call: acp::ToolCallUpdate {
2496 meta: None,
2497 id: acp::ToolCallId(self.tool_use_id.to_string().into()),
2498 fields: acp::ToolCallUpdateFields {
2499 title: Some(title.into()),
2500 ..Default::default()
2501 },
2502 },
2503 options: vec![
2504 acp::PermissionOption {
2505 id: acp::PermissionOptionId("always_allow".into()),
2506 name: "Always Allow".into(),
2507 kind: acp::PermissionOptionKind::AllowAlways,
2508 meta: None,
2509 },
2510 acp::PermissionOption {
2511 id: acp::PermissionOptionId("allow".into()),
2512 name: "Allow".into(),
2513 kind: acp::PermissionOptionKind::AllowOnce,
2514 meta: None,
2515 },
2516 acp::PermissionOption {
2517 id: acp::PermissionOptionId("deny".into()),
2518 name: "Deny".into(),
2519 kind: acp::PermissionOptionKind::RejectOnce,
2520 meta: None,
2521 },
2522 ],
2523 response: response_tx,
2524 },
2525 )))
2526 .ok();
2527 let fs = self.fs.clone();
2528 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2529 "always_allow" => {
2530 if let Some(fs) = fs.clone() {
2531 cx.update(|cx| {
2532 update_settings_file(fs, cx, |settings, _| {
2533 settings
2534 .agent
2535 .get_or_insert_default()
2536 .set_always_allow_tool_actions(true);
2537 });
2538 })?;
2539 }
2540
2541 Ok(())
2542 }
2543 "allow" => Ok(()),
2544 _ => Err(anyhow!("Permission to run tool denied by user")),
2545 })
2546 }
2547}
2548
2549#[cfg(any(test, feature = "test-support"))]
2550pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2551
2552#[cfg(any(test, feature = "test-support"))]
2553impl ToolCallEventStreamReceiver {
2554 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2555 let event = self.0.next().await;
2556 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2557 auth
2558 } else {
2559 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2560 }
2561 }
2562
2563 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
2564 let event = self.0.next().await;
2565 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
2566 update,
2567 )))) = event
2568 {
2569 update.fields
2570 } else {
2571 panic!("Expected update fields but got: {:?}", event);
2572 }
2573 }
2574
2575 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
2576 let event = self.0.next().await;
2577 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
2578 update,
2579 )))) = event
2580 {
2581 update.diff
2582 } else {
2583 panic!("Expected diff but got: {:?}", event);
2584 }
2585 }
2586
2587 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2588 let event = self.0.next().await;
2589 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2590 update,
2591 )))) = event
2592 {
2593 update.terminal
2594 } else {
2595 panic!("Expected terminal but got: {:?}", event);
2596 }
2597 }
2598}
2599
2600#[cfg(any(test, feature = "test-support"))]
2601impl std::ops::Deref for ToolCallEventStreamReceiver {
2602 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2603
2604 fn deref(&self) -> &Self::Target {
2605 &self.0
2606 }
2607}
2608
2609#[cfg(any(test, feature = "test-support"))]
2610impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2611 fn deref_mut(&mut self) -> &mut Self::Target {
2612 &mut self.0
2613 }
2614}
2615
2616impl From<&str> for UserMessageContent {
2617 fn from(text: &str) -> Self {
2618 Self::Text(text.into())
2619 }
2620}
2621
2622impl UserMessageContent {
2623 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
2624 match value {
2625 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2626 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2627 acp::ContentBlock::Audio(_) => {
2628 // TODO
2629 Self::Text("[audio]".to_string())
2630 }
2631 acp::ContentBlock::ResourceLink(resource_link) => {
2632 match MentionUri::parse(&resource_link.uri, path_style) {
2633 Ok(uri) => Self::Mention {
2634 uri,
2635 content: String::new(),
2636 },
2637 Err(err) => {
2638 log::error!("Failed to parse mention link: {}", err);
2639 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2640 }
2641 }
2642 }
2643 acp::ContentBlock::Resource(resource) => match resource.resource {
2644 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2645 match MentionUri::parse(&resource.uri, path_style) {
2646 Ok(uri) => Self::Mention {
2647 uri,
2648 content: resource.text,
2649 },
2650 Err(err) => {
2651 log::error!("Failed to parse mention link: {}", err);
2652 Self::Text(
2653 MarkdownCodeBlock {
2654 tag: &resource.uri,
2655 text: &resource.text,
2656 }
2657 .to_string(),
2658 )
2659 }
2660 }
2661 }
2662 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2663 // TODO
2664 Self::Text("[blob]".to_string())
2665 }
2666 },
2667 }
2668 }
2669}
2670
2671impl From<UserMessageContent> for acp::ContentBlock {
2672 fn from(content: UserMessageContent) -> Self {
2673 match content {
2674 UserMessageContent::Text(text) => acp::ContentBlock::Text(acp::TextContent {
2675 text,
2676 annotations: None,
2677 meta: None,
2678 }),
2679 UserMessageContent::Image(image) => acp::ContentBlock::Image(acp::ImageContent {
2680 data: image.source.to_string(),
2681 mime_type: "image/png".to_string(),
2682 meta: None,
2683 annotations: None,
2684 uri: None,
2685 }),
2686 UserMessageContent::Mention { uri, content } => {
2687 acp::ContentBlock::Resource(acp::EmbeddedResource {
2688 meta: None,
2689 resource: acp::EmbeddedResourceResource::TextResourceContents(
2690 acp::TextResourceContents {
2691 meta: None,
2692 mime_type: None,
2693 text: content,
2694 uri: uri.to_uri().to_string(),
2695 },
2696 ),
2697 annotations: None,
2698 })
2699 }
2700 }
2701 }
2702}
2703
2704fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2705 LanguageModelImage {
2706 source: image_content.data.into(),
2707 // TODO: make this optional?
2708 size: gpui::Size::new(0.into(), 0.into()),
2709 }
2710}