1use crate::{
2 ContextServerRegistry, CopyPathTool, CreateDirectoryTool, DbLanguageModel, DbThread,
3 DeletePathTool, DiagnosticsTool, EditFileTool, FetchTool, FindPathTool, GrepTool,
4 ListDirectoryTool, MovePathTool, NowTool, OpenTool, ProjectSnapshot, ReadFileTool,
5 RestoreFileFromDiskTool, SaveFileTool, SystemPromptTemplate, Template, Templates, TerminalTool,
6 ThinkingTool, WebSearchTool,
7};
8use acp_thread::{MentionUri, UserMessageId};
9use action_log::ActionLog;
10
11use agent_client_protocol as acp;
12use agent_settings::{
13 AgentProfileId, AgentProfileSettings, AgentSettings, CompletionMode,
14 SUMMARIZE_THREAD_DETAILED_PROMPT, SUMMARIZE_THREAD_PROMPT,
15};
16use anyhow::{Context as _, Result, anyhow};
17use chrono::{DateTime, Utc};
18use client::{ModelRequestUsage, RequestUsage, UserStore};
19use cloud_llm_client::{CompletionIntent, 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 gpui::{
30 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
31};
32use language_model::{
33 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelExt,
34 LanguageModelId, LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry,
35 LanguageModelRequest, LanguageModelRequestMessage, LanguageModelRequestTool,
36 LanguageModelToolResult, LanguageModelToolResultContent, LanguageModelToolSchemaFormat,
37 LanguageModelToolUse, LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage,
38 ZED_CLOUD_PROVIDER_ID,
39};
40use project::Project;
41use prompt_store::ProjectContext;
42use schemars::{JsonSchema, Schema};
43use serde::{Deserialize, Serialize};
44use settings::{LanguageModelSelection, Settings, update_settings_file};
45use smol::stream::StreamExt;
46use std::{
47 collections::BTreeMap,
48 ops::RangeInclusive,
49 path::Path,
50 rc::Rc,
51 sync::Arc,
52 time::{Duration, Instant},
53};
54use std::{fmt::Write, path::PathBuf};
55use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
56use uuid::Uuid;
57
58const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
59pub const MAX_TOOL_NAME_LENGTH: usize = 64;
60
61/// The ID of the user prompt that initiated a request.
62///
63/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
64#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
65pub struct PromptId(Arc<str>);
66
67impl PromptId {
68 pub fn new() -> Self {
69 Self(Uuid::new_v4().to_string().into())
70 }
71}
72
73impl std::fmt::Display for PromptId {
74 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
75 write!(f, "{}", self.0)
76 }
77}
78
79pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
80pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
81
82#[derive(Debug, Clone)]
83enum RetryStrategy {
84 ExponentialBackoff {
85 initial_delay: Duration,
86 max_attempts: u8,
87 },
88 Fixed {
89 delay: Duration,
90 max_attempts: u8,
91 },
92}
93
94#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
95pub enum Message {
96 User(UserMessage),
97 Agent(AgentMessage),
98 Resume,
99}
100
101impl Message {
102 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
103 match self {
104 Message::Agent(agent_message) => Some(agent_message),
105 _ => None,
106 }
107 }
108
109 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
110 match self {
111 Message::User(message) => {
112 if message.content.is_empty() {
113 vec![]
114 } else {
115 vec![message.to_request()]
116 }
117 }
118 Message::Agent(message) => message.to_request(),
119 Message::Resume => vec![LanguageModelRequestMessage {
120 role: Role::User,
121 content: vec!["Continue where you left off".into()],
122 cache: false,
123 reasoning_details: None,
124 }],
125 }
126 }
127
128 pub fn to_markdown(&self) -> String {
129 match self {
130 Message::User(message) => message.to_markdown(),
131 Message::Agent(message) => message.to_markdown(),
132 Message::Resume => "[resume]\n".into(),
133 }
134 }
135
136 pub fn role(&self) -> Role {
137 match self {
138 Message::User(_) | Message::Resume => Role::User,
139 Message::Agent(_) => Role::Assistant,
140 }
141 }
142}
143
144#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
145pub struct UserMessage {
146 pub id: UserMessageId,
147 pub content: Vec<UserMessageContent>,
148}
149
150#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
151pub enum UserMessageContent {
152 Text(String),
153 Mention { uri: MentionUri, content: String },
154 Image(LanguageModelImage),
155}
156
157impl UserMessage {
158 pub fn to_markdown(&self) -> String {
159 let mut markdown = String::from("## User\n\n");
160
161 for content in &self.content {
162 match content {
163 UserMessageContent::Text(text) => {
164 markdown.push_str(text);
165 markdown.push('\n');
166 }
167 UserMessageContent::Image(_) => {
168 markdown.push_str("<image />\n");
169 }
170 UserMessageContent::Mention { uri, content } => {
171 if !content.is_empty() {
172 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
173 } else {
174 let _ = writeln!(&mut markdown, "{}", uri.as_link());
175 }
176 }
177 }
178 }
179
180 markdown
181 }
182
183 fn to_request(&self) -> LanguageModelRequestMessage {
184 let mut message = LanguageModelRequestMessage {
185 role: Role::User,
186 content: Vec::with_capacity(self.content.len()),
187 cache: false,
188 reasoning_details: None,
189 };
190
191 const OPEN_CONTEXT: &str = "<context>\n\
192 The following items were attached by the user. \
193 They are up-to-date and don't need to be re-read.\n\n";
194
195 const OPEN_FILES_TAG: &str = "<files>";
196 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
197 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
198 const OPEN_SELECTIONS_TAG: &str = "<selections>";
199 const OPEN_THREADS_TAG: &str = "<threads>";
200 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
201 const OPEN_RULES_TAG: &str =
202 "<rules>\nThe user has specified the following rules that should be applied:\n";
203
204 let mut file_context = OPEN_FILES_TAG.to_string();
205 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
206 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
207 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
208 let mut thread_context = OPEN_THREADS_TAG.to_string();
209 let mut fetch_context = OPEN_FETCH_TAG.to_string();
210 let mut rules_context = OPEN_RULES_TAG.to_string();
211
212 for chunk in &self.content {
213 let chunk = match chunk {
214 UserMessageContent::Text(text) => {
215 language_model::MessageContent::Text(text.clone())
216 }
217 UserMessageContent::Image(value) => {
218 language_model::MessageContent::Image(value.clone())
219 }
220 UserMessageContent::Mention { uri, content } => {
221 match uri {
222 MentionUri::File { abs_path } => {
223 write!(
224 &mut file_context,
225 "\n{}",
226 MarkdownCodeBlock {
227 tag: &codeblock_tag(abs_path, None),
228 text: &content.to_string(),
229 }
230 )
231 .ok();
232 }
233 MentionUri::PastedImage => {
234 debug_panic!("pasted image URI should not be used in mention content")
235 }
236 MentionUri::Directory { .. } => {
237 write!(&mut directory_context, "\n{}\n", content).ok();
238 }
239 MentionUri::Symbol {
240 abs_path: path,
241 line_range,
242 ..
243 } => {
244 write!(
245 &mut symbol_context,
246 "\n{}",
247 MarkdownCodeBlock {
248 tag: &codeblock_tag(path, Some(line_range)),
249 text: content
250 }
251 )
252 .ok();
253 }
254 MentionUri::Selection {
255 abs_path: path,
256 line_range,
257 ..
258 } => {
259 write!(
260 &mut selection_context,
261 "\n{}",
262 MarkdownCodeBlock {
263 tag: &codeblock_tag(
264 path.as_deref().unwrap_or("Untitled".as_ref()),
265 Some(line_range)
266 ),
267 text: content
268 }
269 )
270 .ok();
271 }
272 MentionUri::Thread { .. } => {
273 write!(&mut thread_context, "\n{}\n", content).ok();
274 }
275 MentionUri::TextThread { .. } => {
276 write!(&mut thread_context, "\n{}\n", content).ok();
277 }
278 MentionUri::Rule { .. } => {
279 write!(
280 &mut rules_context,
281 "\n{}",
282 MarkdownCodeBlock {
283 tag: "",
284 text: content
285 }
286 )
287 .ok();
288 }
289 MentionUri::Fetch { url } => {
290 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
291 }
292 }
293
294 language_model::MessageContent::Text(uri.as_link().to_string())
295 }
296 };
297
298 message.content.push(chunk);
299 }
300
301 let len_before_context = message.content.len();
302
303 if file_context.len() > OPEN_FILES_TAG.len() {
304 file_context.push_str("</files>\n");
305 message
306 .content
307 .push(language_model::MessageContent::Text(file_context));
308 }
309
310 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
311 directory_context.push_str("</directories>\n");
312 message
313 .content
314 .push(language_model::MessageContent::Text(directory_context));
315 }
316
317 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
318 symbol_context.push_str("</symbols>\n");
319 message
320 .content
321 .push(language_model::MessageContent::Text(symbol_context));
322 }
323
324 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
325 selection_context.push_str("</selections>\n");
326 message
327 .content
328 .push(language_model::MessageContent::Text(selection_context));
329 }
330
331 if thread_context.len() > OPEN_THREADS_TAG.len() {
332 thread_context.push_str("</threads>\n");
333 message
334 .content
335 .push(language_model::MessageContent::Text(thread_context));
336 }
337
338 if fetch_context.len() > OPEN_FETCH_TAG.len() {
339 fetch_context.push_str("</fetched_urls>\n");
340 message
341 .content
342 .push(language_model::MessageContent::Text(fetch_context));
343 }
344
345 if rules_context.len() > OPEN_RULES_TAG.len() {
346 rules_context.push_str("</user_rules>\n");
347 message
348 .content
349 .push(language_model::MessageContent::Text(rules_context));
350 }
351
352 if message.content.len() > len_before_context {
353 message.content.insert(
354 len_before_context,
355 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
356 );
357 message
358 .content
359 .push(language_model::MessageContent::Text("</context>".into()));
360 }
361
362 message
363 }
364}
365
366fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
367 let mut result = String::new();
368
369 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
370 let _ = write!(result, "{} ", extension);
371 }
372
373 let _ = write!(result, "{}", full_path.display());
374
375 if let Some(range) = line_range {
376 if range.start() == range.end() {
377 let _ = write!(result, ":{}", range.start() + 1);
378 } else {
379 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
380 }
381 }
382
383 result
384}
385
386impl AgentMessage {
387 pub fn to_markdown(&self) -> String {
388 let mut markdown = String::from("## Assistant\n\n");
389
390 for content in &self.content {
391 match content {
392 AgentMessageContent::Text(text) => {
393 markdown.push_str(text);
394 markdown.push('\n');
395 }
396 AgentMessageContent::Thinking { text, .. } => {
397 markdown.push_str("<think>");
398 markdown.push_str(text);
399 markdown.push_str("</think>\n");
400 }
401 AgentMessageContent::RedactedThinking(_) => {
402 markdown.push_str("<redacted_thinking />\n")
403 }
404 AgentMessageContent::ToolUse(tool_use) => {
405 markdown.push_str(&format!(
406 "**Tool Use**: {} (ID: {})\n",
407 tool_use.name, tool_use.id
408 ));
409 markdown.push_str(&format!(
410 "{}\n",
411 MarkdownCodeBlock {
412 tag: "json",
413 text: &format!("{:#}", tool_use.input)
414 }
415 ));
416 }
417 }
418 }
419
420 for tool_result in self.tool_results.values() {
421 markdown.push_str(&format!(
422 "**Tool Result**: {} (ID: {})\n\n",
423 tool_result.tool_name, tool_result.tool_use_id
424 ));
425 if tool_result.is_error {
426 markdown.push_str("**ERROR:**\n");
427 }
428
429 match &tool_result.content {
430 LanguageModelToolResultContent::Text(text) => {
431 writeln!(markdown, "{text}\n").ok();
432 }
433 LanguageModelToolResultContent::Image(_) => {
434 writeln!(markdown, "<image />\n").ok();
435 }
436 }
437
438 if let Some(output) = tool_result.output.as_ref() {
439 writeln!(
440 markdown,
441 "**Debug Output**:\n\n```json\n{}\n```\n",
442 serde_json::to_string_pretty(output).unwrap()
443 )
444 .unwrap();
445 }
446 }
447
448 markdown
449 }
450
451 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
452 let mut assistant_message = LanguageModelRequestMessage {
453 role: Role::Assistant,
454 content: Vec::with_capacity(self.content.len()),
455 cache: false,
456 reasoning_details: self.reasoning_details.clone(),
457 };
458 for chunk in &self.content {
459 match chunk {
460 AgentMessageContent::Text(text) => {
461 assistant_message
462 .content
463 .push(language_model::MessageContent::Text(text.clone()));
464 }
465 AgentMessageContent::Thinking { text, signature } => {
466 assistant_message
467 .content
468 .push(language_model::MessageContent::Thinking {
469 text: text.clone(),
470 signature: signature.clone(),
471 });
472 }
473 AgentMessageContent::RedactedThinking(value) => {
474 assistant_message.content.push(
475 language_model::MessageContent::RedactedThinking(value.clone()),
476 );
477 }
478 AgentMessageContent::ToolUse(tool_use) => {
479 if self.tool_results.contains_key(&tool_use.id) {
480 assistant_message
481 .content
482 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
483 }
484 }
485 };
486 }
487
488 let mut user_message = LanguageModelRequestMessage {
489 role: Role::User,
490 content: Vec::new(),
491 cache: false,
492 reasoning_details: None,
493 };
494
495 for tool_result in self.tool_results.values() {
496 let mut tool_result = tool_result.clone();
497 // Surprisingly, the API fails if we return an empty string here.
498 // It thinks we are sending a tool use without a tool result.
499 if tool_result.content.is_empty() {
500 tool_result.content = "<Tool returned an empty string>".into();
501 }
502 user_message
503 .content
504 .push(language_model::MessageContent::ToolResult(tool_result));
505 }
506
507 let mut messages = Vec::new();
508 if !assistant_message.content.is_empty() {
509 messages.push(assistant_message);
510 }
511 if !user_message.content.is_empty() {
512 messages.push(user_message);
513 }
514 messages
515 }
516}
517
518#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
519pub struct AgentMessage {
520 pub content: Vec<AgentMessageContent>,
521 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
522 pub reasoning_details: Option<serde_json::Value>,
523}
524
525#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
526pub enum AgentMessageContent {
527 Text(String),
528 Thinking {
529 text: String,
530 signature: Option<String>,
531 },
532 RedactedThinking(String),
533 ToolUse(LanguageModelToolUse),
534}
535
536pub trait TerminalHandle {
537 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
538 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
539 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
540 fn kill(&self, cx: &AsyncApp) -> Result<()>;
541}
542
543pub trait ThreadEnvironment {
544 fn create_terminal(
545 &self,
546 command: String,
547 cwd: Option<PathBuf>,
548 output_byte_limit: Option<u64>,
549 cx: &mut AsyncApp,
550 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
551}
552
553#[derive(Debug)]
554pub enum ThreadEvent {
555 UserMessage(UserMessage),
556 AgentText(String),
557 AgentThinking(String),
558 ToolCall(acp::ToolCall),
559 ToolCallUpdate(acp_thread::ToolCallUpdate),
560 ToolCallAuthorization(ToolCallAuthorization),
561 Retry(acp_thread::RetryStatus),
562 Stop(acp::StopReason),
563}
564
565#[derive(Debug)]
566pub struct NewTerminal {
567 pub command: String,
568 pub output_byte_limit: Option<u64>,
569 pub cwd: Option<PathBuf>,
570 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
571}
572
573#[derive(Debug)]
574pub struct ToolCallAuthorization {
575 pub tool_call: acp::ToolCallUpdate,
576 pub options: Vec<acp::PermissionOption>,
577 pub response: oneshot::Sender<acp::PermissionOptionId>,
578}
579
580#[derive(Debug, thiserror::Error)]
581enum CompletionError {
582 #[error("max tokens")]
583 MaxTokens,
584 #[error("refusal")]
585 Refusal,
586 #[error(transparent)]
587 Other(#[from] anyhow::Error),
588}
589
590pub struct Thread {
591 id: acp::SessionId,
592 prompt_id: PromptId,
593 updated_at: DateTime<Utc>,
594 title: Option<SharedString>,
595 pending_title_generation: Option<Task<()>>,
596 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
597 summary: Option<SharedString>,
598 messages: Vec<Message>,
599 user_store: Entity<UserStore>,
600 completion_mode: CompletionMode,
601 /// Holds the task that handles agent interaction until the end of the turn.
602 /// Survives across multiple requests as the model performs tool calls and
603 /// we run tools, report their results.
604 running_turn: Option<RunningTurn>,
605 pending_message: Option<AgentMessage>,
606 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
607 tool_use_limit_reached: bool,
608 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
609 #[allow(unused)]
610 cumulative_token_usage: TokenUsage,
611 #[allow(unused)]
612 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
613 context_server_registry: Entity<ContextServerRegistry>,
614 profile_id: AgentProfileId,
615 project_context: Entity<ProjectContext>,
616 templates: Arc<Templates>,
617 model: Option<Arc<dyn LanguageModel>>,
618 summarization_model: Option<Arc<dyn LanguageModel>>,
619 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
620 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
621 pub(crate) project: Entity<Project>,
622 pub(crate) action_log: Entity<ActionLog>,
623 /// Tracks the last time files were read by the agent, to detect external modifications
624 pub(crate) file_read_times: HashMap<PathBuf, fs::MTime>,
625}
626
627impl Thread {
628 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
629 let image = model.map_or(true, |model| model.supports_images());
630 acp::PromptCapabilities::new()
631 .image(image)
632 .embedded_context(true)
633 }
634
635 pub fn new(
636 project: Entity<Project>,
637 project_context: Entity<ProjectContext>,
638 context_server_registry: Entity<ContextServerRegistry>,
639 templates: Arc<Templates>,
640 model: Option<Arc<dyn LanguageModel>>,
641 cx: &mut Context<Self>,
642 ) -> Self {
643 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
644 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
645 let (prompt_capabilities_tx, prompt_capabilities_rx) =
646 watch::channel(Self::prompt_capabilities(model.as_deref()));
647 Self {
648 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
649 prompt_id: PromptId::new(),
650 updated_at: Utc::now(),
651 title: None,
652 pending_title_generation: None,
653 pending_summary_generation: None,
654 summary: None,
655 messages: Vec::new(),
656 user_store: project.read(cx).user_store(),
657 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
658 running_turn: None,
659 pending_message: None,
660 tools: BTreeMap::default(),
661 tool_use_limit_reached: false,
662 request_token_usage: HashMap::default(),
663 cumulative_token_usage: TokenUsage::default(),
664 initial_project_snapshot: {
665 let project_snapshot = Self::project_snapshot(project.clone(), cx);
666 cx.foreground_executor()
667 .spawn(async move { Some(project_snapshot.await) })
668 .shared()
669 },
670 context_server_registry,
671 profile_id,
672 project_context,
673 templates,
674 model,
675 summarization_model: None,
676 prompt_capabilities_tx,
677 prompt_capabilities_rx,
678 project,
679 action_log,
680 file_read_times: HashMap::default(),
681 }
682 }
683
684 pub fn id(&self) -> &acp::SessionId {
685 &self.id
686 }
687
688 pub fn replay(
689 &mut self,
690 cx: &mut Context<Self>,
691 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
692 let (tx, rx) = mpsc::unbounded();
693 let stream = ThreadEventStream(tx);
694 for message in &self.messages {
695 match message {
696 Message::User(user_message) => stream.send_user_message(user_message),
697 Message::Agent(assistant_message) => {
698 for content in &assistant_message.content {
699 match content {
700 AgentMessageContent::Text(text) => stream.send_text(text),
701 AgentMessageContent::Thinking { text, .. } => {
702 stream.send_thinking(text)
703 }
704 AgentMessageContent::RedactedThinking(_) => {}
705 AgentMessageContent::ToolUse(tool_use) => {
706 self.replay_tool_call(
707 tool_use,
708 assistant_message.tool_results.get(&tool_use.id),
709 &stream,
710 cx,
711 );
712 }
713 }
714 }
715 }
716 Message::Resume => {}
717 }
718 }
719 rx
720 }
721
722 fn replay_tool_call(
723 &self,
724 tool_use: &LanguageModelToolUse,
725 tool_result: Option<&LanguageModelToolResult>,
726 stream: &ThreadEventStream,
727 cx: &mut Context<Self>,
728 ) {
729 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
730 self.context_server_registry
731 .read(cx)
732 .servers()
733 .find_map(|(_, tools)| {
734 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
735 Some(tool.clone())
736 } else {
737 None
738 }
739 })
740 });
741
742 let Some(tool) = tool else {
743 stream
744 .0
745 .unbounded_send(Ok(ThreadEvent::ToolCall(
746 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
747 .status(acp::ToolCallStatus::Failed)
748 .raw_input(tool_use.input.clone()),
749 )))
750 .ok();
751 return;
752 };
753
754 let title = tool.initial_title(tool_use.input.clone(), cx);
755 let kind = tool.kind();
756 stream.send_tool_call(
757 &tool_use.id,
758 &tool_use.name,
759 title,
760 kind,
761 tool_use.input.clone(),
762 );
763
764 let output = tool_result
765 .as_ref()
766 .and_then(|result| result.output.clone());
767 if let Some(output) = output.clone() {
768 let tool_event_stream = ToolCallEventStream::new(
769 tool_use.id.clone(),
770 stream.clone(),
771 Some(self.project.read(cx).fs().clone()),
772 );
773 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
774 .log_err();
775 }
776
777 stream.update_tool_call_fields(
778 &tool_use.id,
779 acp::ToolCallUpdateFields::new()
780 .status(
781 tool_result
782 .as_ref()
783 .map_or(acp::ToolCallStatus::Failed, |result| {
784 if result.is_error {
785 acp::ToolCallStatus::Failed
786 } else {
787 acp::ToolCallStatus::Completed
788 }
789 }),
790 )
791 .raw_output(output),
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 file_read_times: HashMap::default(),
869 }
870 }
871
872 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
873 let initial_project_snapshot = self.initial_project_snapshot.clone();
874 let mut thread = DbThread {
875 title: self.title(),
876 messages: self.messages.clone(),
877 updated_at: self.updated_at,
878 detailed_summary: self.summary.clone(),
879 initial_project_snapshot: None,
880 cumulative_token_usage: self.cumulative_token_usage,
881 request_token_usage: self.request_token_usage.clone(),
882 model: self.model.as_ref().map(|model| DbLanguageModel {
883 provider: model.provider_id().to_string(),
884 model: model.name().0.to_string(),
885 }),
886 completion_mode: Some(self.completion_mode),
887 profile: Some(self.profile_id.clone()),
888 };
889
890 cx.background_spawn(async move {
891 let initial_project_snapshot = initial_project_snapshot.await;
892 thread.initial_project_snapshot = initial_project_snapshot;
893 thread
894 })
895 }
896
897 /// Create a snapshot of the current project state including git information and unsaved buffers.
898 fn project_snapshot(
899 project: Entity<Project>,
900 cx: &mut Context<Self>,
901 ) -> Task<Arc<ProjectSnapshot>> {
902 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
903 cx.spawn(async move |_, _| {
904 let snapshot = task.await;
905
906 Arc::new(ProjectSnapshot {
907 worktree_snapshots: snapshot.worktree_snapshots,
908 timestamp: Utc::now(),
909 })
910 })
911 }
912
913 pub fn project_context(&self) -> &Entity<ProjectContext> {
914 &self.project_context
915 }
916
917 pub fn project(&self) -> &Entity<Project> {
918 &self.project
919 }
920
921 pub fn action_log(&self) -> &Entity<ActionLog> {
922 &self.action_log
923 }
924
925 pub fn is_empty(&self) -> bool {
926 self.messages.is_empty() && self.title.is_none()
927 }
928
929 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
930 self.model.as_ref()
931 }
932
933 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
934 let old_usage = self.latest_token_usage();
935 self.model = Some(model);
936 let new_caps = Self::prompt_capabilities(self.model.as_deref());
937 let new_usage = self.latest_token_usage();
938 if old_usage != new_usage {
939 cx.emit(TokenUsageUpdated(new_usage));
940 }
941 self.prompt_capabilities_tx.send(new_caps).log_err();
942 cx.notify()
943 }
944
945 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
946 self.summarization_model.as_ref()
947 }
948
949 pub fn set_summarization_model(
950 &mut self,
951 model: Option<Arc<dyn LanguageModel>>,
952 cx: &mut Context<Self>,
953 ) {
954 self.summarization_model = model;
955 cx.notify()
956 }
957
958 pub fn completion_mode(&self) -> CompletionMode {
959 self.completion_mode
960 }
961
962 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
963 let old_usage = self.latest_token_usage();
964 self.completion_mode = mode;
965 let new_usage = self.latest_token_usage();
966 if old_usage != new_usage {
967 cx.emit(TokenUsageUpdated(new_usage));
968 }
969 cx.notify()
970 }
971
972 #[cfg(any(test, feature = "test-support"))]
973 pub fn last_message(&self) -> Option<Message> {
974 if let Some(message) = self.pending_message.clone() {
975 Some(Message::Agent(message))
976 } else {
977 self.messages.last().cloned()
978 }
979 }
980
981 pub fn add_default_tools(
982 &mut self,
983 environment: Rc<dyn ThreadEnvironment>,
984 cx: &mut Context<Self>,
985 ) {
986 let language_registry = self.project.read(cx).languages().clone();
987 self.add_tool(CopyPathTool::new(self.project.clone()));
988 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
989 self.add_tool(DeletePathTool::new(
990 self.project.clone(),
991 self.action_log.clone(),
992 ));
993 self.add_tool(DiagnosticsTool::new(self.project.clone()));
994 self.add_tool(EditFileTool::new(
995 self.project.clone(),
996 cx.weak_entity(),
997 language_registry,
998 Templates::new(),
999 ));
1000 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1001 self.add_tool(FindPathTool::new(self.project.clone()));
1002 self.add_tool(GrepTool::new(self.project.clone()));
1003 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1004 self.add_tool(MovePathTool::new(self.project.clone()));
1005 self.add_tool(NowTool);
1006 self.add_tool(OpenTool::new(self.project.clone()));
1007 self.add_tool(ReadFileTool::new(
1008 cx.weak_entity(),
1009 self.project.clone(),
1010 self.action_log.clone(),
1011 ));
1012 self.add_tool(SaveFileTool::new(self.project.clone()));
1013 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1014 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1015 self.add_tool(ThinkingTool);
1016 self.add_tool(WebSearchTool);
1017 }
1018
1019 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1020 self.tools.insert(T::name().into(), tool.erase());
1021 }
1022
1023 pub fn remove_tool(&mut self, name: &str) -> bool {
1024 self.tools.remove(name).is_some()
1025 }
1026
1027 pub fn profile(&self) -> &AgentProfileId {
1028 &self.profile_id
1029 }
1030
1031 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1032 if self.profile_id == profile_id {
1033 return;
1034 }
1035
1036 self.profile_id = profile_id;
1037
1038 // Swap to the profile's preferred model when available.
1039 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1040 self.set_model(model, cx);
1041 }
1042 }
1043
1044 pub fn cancel(&mut self, cx: &mut Context<Self>) {
1045 if let Some(running_turn) = self.running_turn.take() {
1046 running_turn.cancel();
1047 }
1048 self.flush_pending_message(cx);
1049 }
1050
1051 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1052 let Some(last_user_message) = self.last_user_message() else {
1053 return;
1054 };
1055
1056 self.request_token_usage
1057 .insert(last_user_message.id.clone(), update);
1058 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1059 cx.notify();
1060 }
1061
1062 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1063 self.cancel(cx);
1064 let Some(position) = self.messages.iter().position(
1065 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1066 ) else {
1067 return Err(anyhow!("Message not found"));
1068 };
1069
1070 for message in self.messages.drain(position..) {
1071 match message {
1072 Message::User(message) => {
1073 self.request_token_usage.remove(&message.id);
1074 }
1075 Message::Agent(_) | Message::Resume => {}
1076 }
1077 }
1078 self.clear_summary();
1079 cx.notify();
1080 Ok(())
1081 }
1082
1083 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1084 let last_user_message = self.last_user_message()?;
1085 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1086 Some(*tokens)
1087 }
1088
1089 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1090 let usage = self.latest_request_token_usage()?;
1091 let model = self.model.clone()?;
1092 Some(acp_thread::TokenUsage {
1093 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1094 used_tokens: usage.total_tokens(),
1095 })
1096 }
1097
1098 /// Get the total input token count as of the message before the given message.
1099 ///
1100 /// Returns `None` if:
1101 /// - `target_id` is the first message (no previous message)
1102 /// - The previous message hasn't received a response yet (no usage data)
1103 /// - `target_id` is not found in the messages
1104 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1105 let mut previous_user_message_id: Option<&UserMessageId> = None;
1106
1107 for message in &self.messages {
1108 if let Message::User(user_msg) = message {
1109 if &user_msg.id == target_id {
1110 let prev_id = previous_user_message_id?;
1111 let usage = self.request_token_usage.get(prev_id)?;
1112 return Some(usage.input_tokens);
1113 }
1114 previous_user_message_id = Some(&user_msg.id);
1115 }
1116 }
1117 None
1118 }
1119
1120 /// Look up the active profile and resolve its preferred model if one is configured.
1121 fn resolve_profile_model(
1122 profile_id: &AgentProfileId,
1123 cx: &mut Context<Self>,
1124 ) -> Option<Arc<dyn LanguageModel>> {
1125 let selection = AgentSettings::get_global(cx)
1126 .profiles
1127 .get(profile_id)?
1128 .default_model
1129 .clone()?;
1130 Self::resolve_model_from_selection(&selection, cx)
1131 }
1132
1133 /// Translate a stored model selection into the configured model from the registry.
1134 fn resolve_model_from_selection(
1135 selection: &LanguageModelSelection,
1136 cx: &mut Context<Self>,
1137 ) -> Option<Arc<dyn LanguageModel>> {
1138 let selected = SelectedModel {
1139 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1140 model: LanguageModelId::from(selection.model.clone()),
1141 };
1142 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1143 registry
1144 .select_model(&selected, cx)
1145 .map(|configured| configured.model)
1146 })
1147 }
1148
1149 pub fn resume(
1150 &mut self,
1151 cx: &mut Context<Self>,
1152 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1153 self.messages.push(Message::Resume);
1154 cx.notify();
1155
1156 log::debug!("Total messages in thread: {}", self.messages.len());
1157 self.run_turn(cx)
1158 }
1159
1160 /// Sending a message results in the model streaming a response, which could include tool calls.
1161 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1162 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1163 pub fn send<T>(
1164 &mut self,
1165 id: UserMessageId,
1166 content: impl IntoIterator<Item = T>,
1167 cx: &mut Context<Self>,
1168 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1169 where
1170 T: Into<UserMessageContent>,
1171 {
1172 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1173 log::debug!("Thread::send content: {:?}", content);
1174
1175 self.messages
1176 .push(Message::User(UserMessage { id, content }));
1177 cx.notify();
1178
1179 self.send_existing(cx)
1180 }
1181
1182 pub fn send_existing(
1183 &mut self,
1184 cx: &mut Context<Self>,
1185 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1186 let model = self.model().context("No language model configured")?;
1187
1188 log::info!("Thread::send called with model: {}", model.name().0);
1189 self.advance_prompt_id();
1190
1191 log::debug!("Total messages in thread: {}", self.messages.len());
1192 self.run_turn(cx)
1193 }
1194
1195 pub fn push_acp_user_block(
1196 &mut self,
1197 id: UserMessageId,
1198 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1199 path_style: PathStyle,
1200 cx: &mut Context<Self>,
1201 ) {
1202 let content = blocks
1203 .into_iter()
1204 .map(|block| UserMessageContent::from_content_block(block, path_style))
1205 .collect::<Vec<_>>();
1206 self.messages
1207 .push(Message::User(UserMessage { id, content }));
1208 cx.notify();
1209 }
1210
1211 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1212 let text = match block {
1213 acp::ContentBlock::Text(text_content) => text_content.text,
1214 acp::ContentBlock::Image(_) => "[image]".to_string(),
1215 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1216 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1217 acp::ContentBlock::Resource(resource) => match resource.resource {
1218 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1219 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1220 _ => "[resource]".to_string(),
1221 },
1222 _ => "[unknown]".to_string(),
1223 };
1224
1225 self.messages.push(Message::Agent(AgentMessage {
1226 content: vec![AgentMessageContent::Text(text)],
1227 ..Default::default()
1228 }));
1229 cx.notify();
1230 }
1231
1232 #[cfg(feature = "eval")]
1233 pub fn proceed(
1234 &mut self,
1235 cx: &mut Context<Self>,
1236 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1237 self.run_turn(cx)
1238 }
1239
1240 fn run_turn(
1241 &mut self,
1242 cx: &mut Context<Self>,
1243 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1244 self.cancel(cx);
1245
1246 let model = self.model.clone().context("No language model configured")?;
1247 let profile = AgentSettings::get_global(cx)
1248 .profiles
1249 .get(&self.profile_id)
1250 .context("Profile not found")?;
1251 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1252 let event_stream = ThreadEventStream(events_tx);
1253 let message_ix = self.messages.len().saturating_sub(1);
1254 self.tool_use_limit_reached = false;
1255 self.clear_summary();
1256 self.running_turn = Some(RunningTurn {
1257 event_stream: event_stream.clone(),
1258 tools: self.enabled_tools(profile, &model, cx),
1259 _task: cx.spawn(async move |this, cx| {
1260 log::debug!("Starting agent turn execution");
1261
1262 let turn_result = Self::run_turn_internal(&this, model, &event_stream, cx).await;
1263 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1264
1265 match turn_result {
1266 Ok(()) => {
1267 log::debug!("Turn execution completed");
1268 event_stream.send_stop(acp::StopReason::EndTurn);
1269 }
1270 Err(error) => {
1271 log::error!("Turn execution failed: {:?}", error);
1272 match error.downcast::<CompletionError>() {
1273 Ok(CompletionError::Refusal) => {
1274 event_stream.send_stop(acp::StopReason::Refusal);
1275 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1276 }
1277 Ok(CompletionError::MaxTokens) => {
1278 event_stream.send_stop(acp::StopReason::MaxTokens);
1279 }
1280 Ok(CompletionError::Other(error)) | Err(error) => {
1281 event_stream.send_error(error);
1282 }
1283 }
1284 }
1285 }
1286
1287 _ = this.update(cx, |this, _| this.running_turn.take());
1288 }),
1289 });
1290 Ok(events_rx)
1291 }
1292
1293 async fn run_turn_internal(
1294 this: &WeakEntity<Self>,
1295 model: Arc<dyn LanguageModel>,
1296 event_stream: &ThreadEventStream,
1297 cx: &mut AsyncApp,
1298 ) -> Result<()> {
1299 let mut attempt = 0;
1300 let mut intent = CompletionIntent::UserPrompt;
1301 loop {
1302 let request =
1303 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1304
1305 telemetry::event!(
1306 "Agent Thread Completion",
1307 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1308 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1309 model = model.telemetry_id(),
1310 model_provider = model.provider_id().to_string(),
1311 attempt
1312 );
1313
1314 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1315
1316 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1317 Ok(events) => (events, None),
1318 Err(err) => (stream::empty().boxed(), Some(err)),
1319 };
1320 let mut tool_results = FuturesUnordered::new();
1321 while let Some(event) = events.next().await {
1322 log::trace!("Received completion event: {:?}", event);
1323 match event {
1324 Ok(event) => {
1325 tool_results.extend(this.update(cx, |this, cx| {
1326 this.handle_completion_event(event, event_stream, cx)
1327 })??);
1328 }
1329 Err(err) => {
1330 error = Some(err);
1331 break;
1332 }
1333 }
1334 }
1335
1336 let end_turn = tool_results.is_empty();
1337 while let Some(tool_result) = tool_results.next().await {
1338 log::debug!("Tool finished {:?}", tool_result);
1339
1340 event_stream.update_tool_call_fields(
1341 &tool_result.tool_use_id,
1342 acp::ToolCallUpdateFields::new()
1343 .status(if tool_result.is_error {
1344 acp::ToolCallStatus::Failed
1345 } else {
1346 acp::ToolCallStatus::Completed
1347 })
1348 .raw_output(tool_result.output.clone()),
1349 );
1350 this.update(cx, |this, _cx| {
1351 this.pending_message()
1352 .tool_results
1353 .insert(tool_result.tool_use_id.clone(), tool_result);
1354 })?;
1355 }
1356
1357 this.update(cx, |this, cx| {
1358 this.flush_pending_message(cx);
1359 if this.title.is_none() && this.pending_title_generation.is_none() {
1360 this.generate_title(cx);
1361 }
1362 })?;
1363
1364 if let Some(error) = error {
1365 attempt += 1;
1366 let retry = this.update(cx, |this, cx| {
1367 let user_store = this.user_store.read(cx);
1368 this.handle_completion_error(error, attempt, user_store.plan())
1369 })??;
1370 let timer = cx.background_executor().timer(retry.duration);
1371 event_stream.send_retry(retry);
1372 timer.await;
1373 this.update(cx, |this, _cx| {
1374 if let Some(Message::Agent(message)) = this.messages.last() {
1375 if message.tool_results.is_empty() {
1376 intent = CompletionIntent::UserPrompt;
1377 this.messages.push(Message::Resume);
1378 }
1379 }
1380 })?;
1381 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1382 return Err(language_model::ToolUseLimitReachedError.into());
1383 } else if end_turn {
1384 return Ok(());
1385 } else {
1386 intent = CompletionIntent::ToolResults;
1387 attempt = 0;
1388 }
1389 }
1390 }
1391
1392 fn handle_completion_error(
1393 &mut self,
1394 error: LanguageModelCompletionError,
1395 attempt: u8,
1396 plan: Option<Plan>,
1397 ) -> Result<acp_thread::RetryStatus> {
1398 let Some(model) = self.model.as_ref() else {
1399 return Err(anyhow!(error));
1400 };
1401
1402 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1403 match plan {
1404 Some(Plan::V2(_)) => true,
1405 Some(Plan::V1(_)) => self.completion_mode == CompletionMode::Burn,
1406 None => false,
1407 }
1408 } else {
1409 true
1410 };
1411
1412 if !auto_retry {
1413 return Err(anyhow!(error));
1414 }
1415
1416 let Some(strategy) = Self::retry_strategy_for(&error) else {
1417 return Err(anyhow!(error));
1418 };
1419
1420 let max_attempts = match &strategy {
1421 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1422 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1423 };
1424
1425 if attempt > max_attempts {
1426 return Err(anyhow!(error));
1427 }
1428
1429 let delay = match &strategy {
1430 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1431 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1432 Duration::from_secs(delay_secs)
1433 }
1434 RetryStrategy::Fixed { delay, .. } => *delay,
1435 };
1436 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1437
1438 Ok(acp_thread::RetryStatus {
1439 last_error: error.to_string().into(),
1440 attempt: attempt as usize,
1441 max_attempts: max_attempts as usize,
1442 started_at: Instant::now(),
1443 duration: delay,
1444 })
1445 }
1446
1447 /// A helper method that's called on every streamed completion event.
1448 /// Returns an optional tool result task, which the main agentic loop will
1449 /// send back to the model when it resolves.
1450 fn handle_completion_event(
1451 &mut self,
1452 event: LanguageModelCompletionEvent,
1453 event_stream: &ThreadEventStream,
1454 cx: &mut Context<Self>,
1455 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1456 log::trace!("Handling streamed completion event: {:?}", event);
1457 use LanguageModelCompletionEvent::*;
1458
1459 match event {
1460 StartMessage { .. } => {
1461 self.flush_pending_message(cx);
1462 self.pending_message = Some(AgentMessage::default());
1463 }
1464 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1465 Thinking { text, signature } => {
1466 self.handle_thinking_event(text, signature, event_stream, cx)
1467 }
1468 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1469 ReasoningDetails(details) => {
1470 let last_message = self.pending_message();
1471 // Store the last non-empty reasoning_details (overwrites earlier ones)
1472 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
1473 if let serde_json::Value::Array(ref arr) = details {
1474 if !arr.is_empty() {
1475 last_message.reasoning_details = Some(details);
1476 }
1477 } else {
1478 last_message.reasoning_details = Some(details);
1479 }
1480 }
1481 ToolUse(tool_use) => {
1482 return Ok(self.handle_tool_use_event(tool_use, event_stream, cx));
1483 }
1484 ToolUseJsonParseError {
1485 id,
1486 tool_name,
1487 raw_input,
1488 json_parse_error,
1489 } => {
1490 return Ok(Some(Task::ready(
1491 self.handle_tool_use_json_parse_error_event(
1492 id,
1493 tool_name,
1494 raw_input,
1495 json_parse_error,
1496 ),
1497 )));
1498 }
1499 UsageUpdate(usage) => {
1500 telemetry::event!(
1501 "Agent Thread Completion Usage Updated",
1502 thread_id = self.id.to_string(),
1503 prompt_id = self.prompt_id.to_string(),
1504 model = self.model.as_ref().map(|m| m.telemetry_id()),
1505 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1506 input_tokens = usage.input_tokens,
1507 output_tokens = usage.output_tokens,
1508 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1509 cache_read_input_tokens = usage.cache_read_input_tokens,
1510 );
1511 self.update_token_usage(usage, cx);
1512 }
1513 UsageUpdated { amount, limit } => {
1514 self.update_model_request_usage(amount, limit, cx);
1515 }
1516 ToolUseLimitReached => {
1517 self.tool_use_limit_reached = true;
1518 }
1519 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1520 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1521 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1522 Started | Queued { .. } => {}
1523 }
1524
1525 Ok(None)
1526 }
1527
1528 fn handle_text_event(
1529 &mut self,
1530 new_text: String,
1531 event_stream: &ThreadEventStream,
1532 cx: &mut Context<Self>,
1533 ) {
1534 event_stream.send_text(&new_text);
1535
1536 let last_message = self.pending_message();
1537 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1538 text.push_str(&new_text);
1539 } else {
1540 last_message
1541 .content
1542 .push(AgentMessageContent::Text(new_text));
1543 }
1544
1545 cx.notify();
1546 }
1547
1548 fn handle_thinking_event(
1549 &mut self,
1550 new_text: String,
1551 new_signature: Option<String>,
1552 event_stream: &ThreadEventStream,
1553 cx: &mut Context<Self>,
1554 ) {
1555 event_stream.send_thinking(&new_text);
1556
1557 let last_message = self.pending_message();
1558 if let Some(AgentMessageContent::Thinking { text, signature }) =
1559 last_message.content.last_mut()
1560 {
1561 text.push_str(&new_text);
1562 *signature = new_signature.or(signature.take());
1563 } else {
1564 last_message.content.push(AgentMessageContent::Thinking {
1565 text: new_text,
1566 signature: new_signature,
1567 });
1568 }
1569
1570 cx.notify();
1571 }
1572
1573 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1574 let last_message = self.pending_message();
1575 last_message
1576 .content
1577 .push(AgentMessageContent::RedactedThinking(data));
1578 cx.notify();
1579 }
1580
1581 fn handle_tool_use_event(
1582 &mut self,
1583 tool_use: LanguageModelToolUse,
1584 event_stream: &ThreadEventStream,
1585 cx: &mut Context<Self>,
1586 ) -> Option<Task<LanguageModelToolResult>> {
1587 cx.notify();
1588
1589 let tool = self.tool(tool_use.name.as_ref());
1590 let mut title = SharedString::from(&tool_use.name);
1591 let mut kind = acp::ToolKind::Other;
1592 if let Some(tool) = tool.as_ref() {
1593 title = tool.initial_title(tool_use.input.clone(), cx);
1594 kind = tool.kind();
1595 }
1596
1597 // Ensure the last message ends in the current tool use
1598 let last_message = self.pending_message();
1599 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1600 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1601 if last_tool_use.id == tool_use.id {
1602 *last_tool_use = tool_use.clone();
1603 false
1604 } else {
1605 true
1606 }
1607 } else {
1608 true
1609 }
1610 });
1611
1612 if push_new_tool_use {
1613 event_stream.send_tool_call(
1614 &tool_use.id,
1615 &tool_use.name,
1616 title,
1617 kind,
1618 tool_use.input.clone(),
1619 );
1620 last_message
1621 .content
1622 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1623 } else {
1624 event_stream.update_tool_call_fields(
1625 &tool_use.id,
1626 acp::ToolCallUpdateFields::new()
1627 .title(title.as_str())
1628 .kind(kind)
1629 .raw_input(tool_use.input.clone()),
1630 );
1631 }
1632
1633 if !tool_use.is_input_complete {
1634 return None;
1635 }
1636
1637 let Some(tool) = tool else {
1638 let content = format!("No tool named {} exists", tool_use.name);
1639 return Some(Task::ready(LanguageModelToolResult {
1640 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1641 tool_use_id: tool_use.id,
1642 tool_name: tool_use.name,
1643 is_error: true,
1644 output: None,
1645 }));
1646 };
1647
1648 let fs = self.project.read(cx).fs().clone();
1649 let tool_event_stream =
1650 ToolCallEventStream::new(tool_use.id.clone(), event_stream.clone(), Some(fs));
1651 tool_event_stream.update_fields(
1652 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
1653 );
1654 let supports_images = self.model().is_some_and(|model| model.supports_images());
1655 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1656 log::debug!("Running tool {}", tool_use.name);
1657 Some(cx.foreground_executor().spawn(async move {
1658 let tool_result = tool_result.await.and_then(|output| {
1659 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1660 && !supports_images
1661 {
1662 return Err(anyhow!(
1663 "Attempted to read an image, but this model doesn't support it.",
1664 ));
1665 }
1666 Ok(output)
1667 });
1668
1669 match tool_result {
1670 Ok(output) => LanguageModelToolResult {
1671 tool_use_id: tool_use.id,
1672 tool_name: tool_use.name,
1673 is_error: false,
1674 content: output.llm_output,
1675 output: Some(output.raw_output),
1676 },
1677 Err(error) => LanguageModelToolResult {
1678 tool_use_id: tool_use.id,
1679 tool_name: tool_use.name,
1680 is_error: true,
1681 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1682 output: Some(error.to_string().into()),
1683 },
1684 }
1685 }))
1686 }
1687
1688 fn handle_tool_use_json_parse_error_event(
1689 &mut self,
1690 tool_use_id: LanguageModelToolUseId,
1691 tool_name: Arc<str>,
1692 raw_input: Arc<str>,
1693 json_parse_error: String,
1694 ) -> LanguageModelToolResult {
1695 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1696 LanguageModelToolResult {
1697 tool_use_id,
1698 tool_name,
1699 is_error: true,
1700 content: LanguageModelToolResultContent::Text(tool_output.into()),
1701 output: Some(serde_json::Value::String(raw_input.to_string())),
1702 }
1703 }
1704
1705 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1706 self.project
1707 .read(cx)
1708 .user_store()
1709 .update(cx, |user_store, cx| {
1710 user_store.update_model_request_usage(
1711 ModelRequestUsage(RequestUsage {
1712 amount: amount as i32,
1713 limit,
1714 }),
1715 cx,
1716 )
1717 });
1718 }
1719
1720 pub fn title(&self) -> SharedString {
1721 self.title.clone().unwrap_or("New Thread".into())
1722 }
1723
1724 pub fn is_generating_summary(&self) -> bool {
1725 self.pending_summary_generation.is_some()
1726 }
1727
1728 pub fn is_generating_title(&self) -> bool {
1729 self.pending_title_generation.is_some()
1730 }
1731
1732 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
1733 if let Some(summary) = self.summary.as_ref() {
1734 return Task::ready(Some(summary.clone())).shared();
1735 }
1736 if let Some(task) = self.pending_summary_generation.clone() {
1737 return task;
1738 }
1739 let Some(model) = self.summarization_model.clone() else {
1740 log::error!("No summarization model available");
1741 return Task::ready(None).shared();
1742 };
1743 let mut request = LanguageModelRequest {
1744 intent: Some(CompletionIntent::ThreadContextSummarization),
1745 temperature: AgentSettings::temperature_for_model(&model, cx),
1746 ..Default::default()
1747 };
1748
1749 for message in &self.messages {
1750 request.messages.extend(message.to_request());
1751 }
1752
1753 request.messages.push(LanguageModelRequestMessage {
1754 role: Role::User,
1755 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1756 cache: false,
1757 reasoning_details: None,
1758 });
1759
1760 let task = cx
1761 .spawn(async move |this, cx| {
1762 let mut summary = String::new();
1763 let mut messages = model.stream_completion(request, cx).await.log_err()?;
1764 while let Some(event) = messages.next().await {
1765 let event = event.log_err()?;
1766 let text = match event {
1767 LanguageModelCompletionEvent::Text(text) => text,
1768 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
1769 this.update(cx, |thread, cx| {
1770 thread.update_model_request_usage(amount, limit, cx);
1771 })
1772 .ok()?;
1773 continue;
1774 }
1775 _ => continue,
1776 };
1777
1778 let mut lines = text.lines();
1779 summary.extend(lines.next());
1780 }
1781
1782 log::debug!("Setting summary: {}", summary);
1783 let summary = SharedString::from(summary);
1784
1785 this.update(cx, |this, cx| {
1786 this.summary = Some(summary.clone());
1787 this.pending_summary_generation = None;
1788 cx.notify()
1789 })
1790 .ok()?;
1791
1792 Some(summary)
1793 })
1794 .shared();
1795 self.pending_summary_generation = Some(task.clone());
1796 task
1797 }
1798
1799 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
1800 let Some(model) = self.summarization_model.clone() else {
1801 return;
1802 };
1803
1804 log::debug!(
1805 "Generating title with model: {:?}",
1806 self.summarization_model.as_ref().map(|model| model.name())
1807 );
1808 let mut request = LanguageModelRequest {
1809 intent: Some(CompletionIntent::ThreadSummarization),
1810 temperature: AgentSettings::temperature_for_model(&model, cx),
1811 ..Default::default()
1812 };
1813
1814 for message in &self.messages {
1815 request.messages.extend(message.to_request());
1816 }
1817
1818 request.messages.push(LanguageModelRequestMessage {
1819 role: Role::User,
1820 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1821 cache: false,
1822 reasoning_details: None,
1823 });
1824 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
1825 let mut title = String::new();
1826
1827 let generate = async {
1828 let mut messages = model.stream_completion(request, cx).await?;
1829 while let Some(event) = messages.next().await {
1830 let event = event?;
1831 let text = match event {
1832 LanguageModelCompletionEvent::Text(text) => text,
1833 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
1834 this.update(cx, |thread, cx| {
1835 thread.update_model_request_usage(amount, limit, cx);
1836 })?;
1837 continue;
1838 }
1839 _ => continue,
1840 };
1841
1842 let mut lines = text.lines();
1843 title.extend(lines.next());
1844
1845 // Stop if the LLM generated multiple lines.
1846 if lines.next().is_some() {
1847 break;
1848 }
1849 }
1850 anyhow::Ok(())
1851 };
1852
1853 if generate.await.context("failed to generate title").is_ok() {
1854 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
1855 }
1856 _ = this.update(cx, |this, _| this.pending_title_generation = None);
1857 }));
1858 }
1859
1860 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
1861 self.pending_title_generation = None;
1862 if Some(&title) != self.title.as_ref() {
1863 self.title = Some(title);
1864 cx.emit(TitleUpdated);
1865 cx.notify();
1866 }
1867 }
1868
1869 fn clear_summary(&mut self) {
1870 self.summary = None;
1871 self.pending_summary_generation = None;
1872 }
1873
1874 fn last_user_message(&self) -> Option<&UserMessage> {
1875 self.messages
1876 .iter()
1877 .rev()
1878 .find_map(|message| match message {
1879 Message::User(user_message) => Some(user_message),
1880 Message::Agent(_) => None,
1881 Message::Resume => None,
1882 })
1883 }
1884
1885 fn pending_message(&mut self) -> &mut AgentMessage {
1886 self.pending_message.get_or_insert_default()
1887 }
1888
1889 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1890 let Some(mut message) = self.pending_message.take() else {
1891 return;
1892 };
1893
1894 if message.content.is_empty() {
1895 return;
1896 }
1897
1898 for content in &message.content {
1899 let AgentMessageContent::ToolUse(tool_use) = content else {
1900 continue;
1901 };
1902
1903 if !message.tool_results.contains_key(&tool_use.id) {
1904 message.tool_results.insert(
1905 tool_use.id.clone(),
1906 LanguageModelToolResult {
1907 tool_use_id: tool_use.id.clone(),
1908 tool_name: tool_use.name.clone(),
1909 is_error: true,
1910 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1911 output: None,
1912 },
1913 );
1914 }
1915 }
1916
1917 self.messages.push(Message::Agent(message));
1918 self.updated_at = Utc::now();
1919 self.clear_summary();
1920 cx.notify()
1921 }
1922
1923 pub(crate) fn build_completion_request(
1924 &self,
1925 completion_intent: CompletionIntent,
1926 cx: &App,
1927 ) -> Result<LanguageModelRequest> {
1928 let model = self.model().context("No language model configured")?;
1929 let tools = if let Some(turn) = self.running_turn.as_ref() {
1930 turn.tools
1931 .iter()
1932 .filter_map(|(tool_name, tool)| {
1933 log::trace!("Including tool: {}", tool_name);
1934 Some(LanguageModelRequestTool {
1935 name: tool_name.to_string(),
1936 description: tool.description().to_string(),
1937 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1938 })
1939 })
1940 .collect::<Vec<_>>()
1941 } else {
1942 Vec::new()
1943 };
1944
1945 log::debug!("Building completion request");
1946 log::debug!("Completion intent: {:?}", completion_intent);
1947 log::debug!("Completion mode: {:?}", self.completion_mode);
1948
1949 let available_tools: Vec<_> = self
1950 .running_turn
1951 .as_ref()
1952 .map(|turn| turn.tools.keys().cloned().collect())
1953 .unwrap_or_default();
1954
1955 log::debug!("Request includes {} tools", available_tools.len());
1956 let messages = self.build_request_messages(available_tools, cx);
1957 log::debug!("Request will include {} messages", messages.len());
1958
1959 let request = LanguageModelRequest {
1960 thread_id: Some(self.id.to_string()),
1961 prompt_id: Some(self.prompt_id.to_string()),
1962 intent: Some(completion_intent),
1963 mode: Some(self.completion_mode.into()),
1964 messages,
1965 tools,
1966 tool_choice: None,
1967 stop: Vec::new(),
1968 temperature: AgentSettings::temperature_for_model(model, cx),
1969 thinking_allowed: true,
1970 };
1971
1972 log::debug!("Completion request built successfully");
1973 Ok(request)
1974 }
1975
1976 fn enabled_tools(
1977 &self,
1978 profile: &AgentProfileSettings,
1979 model: &Arc<dyn LanguageModel>,
1980 cx: &App,
1981 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
1982 fn truncate(tool_name: &SharedString) -> SharedString {
1983 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
1984 let mut truncated = tool_name.to_string();
1985 truncated.truncate(MAX_TOOL_NAME_LENGTH);
1986 truncated.into()
1987 } else {
1988 tool_name.clone()
1989 }
1990 }
1991
1992 let mut tools = self
1993 .tools
1994 .iter()
1995 .filter_map(|(tool_name, tool)| {
1996 if tool.supports_provider(&model.provider_id())
1997 && profile.is_tool_enabled(tool_name)
1998 {
1999 Some((truncate(tool_name), tool.clone()))
2000 } else {
2001 None
2002 }
2003 })
2004 .collect::<BTreeMap<_, _>>();
2005
2006 let mut context_server_tools = Vec::new();
2007 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2008 let mut duplicate_tool_names = HashSet::default();
2009 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2010 for (tool_name, tool) in server_tools {
2011 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2012 let tool_name = truncate(tool_name);
2013 if !seen_tools.insert(tool_name.clone()) {
2014 duplicate_tool_names.insert(tool_name.clone());
2015 }
2016 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2017 }
2018 }
2019 }
2020
2021 // When there are duplicate tool names, disambiguate by prefixing them
2022 // with the server ID. In the rare case there isn't enough space for the
2023 // disambiguated tool name, keep only the last tool with this name.
2024 for (server_id, tool_name, tool) in context_server_tools {
2025 if duplicate_tool_names.contains(&tool_name) {
2026 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2027 if available >= 2 {
2028 let mut disambiguated = server_id.0.to_string();
2029 disambiguated.truncate(available - 1);
2030 disambiguated.push('_');
2031 disambiguated.push_str(&tool_name);
2032 tools.insert(disambiguated.into(), tool.clone());
2033 } else {
2034 tools.insert(tool_name, tool.clone());
2035 }
2036 } else {
2037 tools.insert(tool_name, tool.clone());
2038 }
2039 }
2040
2041 tools
2042 }
2043
2044 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2045 self.running_turn.as_ref()?.tools.get(name).cloned()
2046 }
2047
2048 pub fn has_tool(&self, name: &str) -> bool {
2049 self.running_turn
2050 .as_ref()
2051 .is_some_and(|turn| turn.tools.contains_key(name))
2052 }
2053
2054 fn build_request_messages(
2055 &self,
2056 available_tools: Vec<SharedString>,
2057 cx: &App,
2058 ) -> Vec<LanguageModelRequestMessage> {
2059 log::trace!(
2060 "Building request messages from {} thread messages",
2061 self.messages.len()
2062 );
2063
2064 let system_prompt = SystemPromptTemplate {
2065 project: self.project_context.read(cx),
2066 available_tools,
2067 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2068 }
2069 .render(&self.templates)
2070 .context("failed to build system prompt")
2071 .expect("Invalid template");
2072 let mut messages = vec![LanguageModelRequestMessage {
2073 role: Role::System,
2074 content: vec![system_prompt.into()],
2075 cache: false,
2076 reasoning_details: None,
2077 }];
2078 for message in &self.messages {
2079 messages.extend(message.to_request());
2080 }
2081
2082 if let Some(last_message) = messages.last_mut() {
2083 last_message.cache = true;
2084 }
2085
2086 if let Some(message) = self.pending_message.as_ref() {
2087 messages.extend(message.to_request());
2088 }
2089
2090 messages
2091 }
2092
2093 pub fn to_markdown(&self) -> String {
2094 let mut markdown = String::new();
2095 for (ix, message) in self.messages.iter().enumerate() {
2096 if ix > 0 {
2097 markdown.push('\n');
2098 }
2099 markdown.push_str(&message.to_markdown());
2100 }
2101
2102 if let Some(message) = self.pending_message.as_ref() {
2103 markdown.push('\n');
2104 markdown.push_str(&message.to_markdown());
2105 }
2106
2107 markdown
2108 }
2109
2110 fn advance_prompt_id(&mut self) {
2111 self.prompt_id = PromptId::new();
2112 }
2113
2114 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2115 use LanguageModelCompletionError::*;
2116 use http_client::StatusCode;
2117
2118 // General strategy here:
2119 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2120 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2121 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2122 match error {
2123 HttpResponseError {
2124 status_code: StatusCode::TOO_MANY_REQUESTS,
2125 ..
2126 } => Some(RetryStrategy::ExponentialBackoff {
2127 initial_delay: BASE_RETRY_DELAY,
2128 max_attempts: MAX_RETRY_ATTEMPTS,
2129 }),
2130 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2131 Some(RetryStrategy::Fixed {
2132 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2133 max_attempts: MAX_RETRY_ATTEMPTS,
2134 })
2135 }
2136 UpstreamProviderError {
2137 status,
2138 retry_after,
2139 ..
2140 } => match *status {
2141 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2142 Some(RetryStrategy::Fixed {
2143 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2144 max_attempts: MAX_RETRY_ATTEMPTS,
2145 })
2146 }
2147 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2148 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2149 // Internal Server Error could be anything, retry up to 3 times.
2150 max_attempts: 3,
2151 }),
2152 status => {
2153 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2154 // but we frequently get them in practice. See https://http.dev/529
2155 if status.as_u16() == 529 {
2156 Some(RetryStrategy::Fixed {
2157 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2158 max_attempts: MAX_RETRY_ATTEMPTS,
2159 })
2160 } else {
2161 Some(RetryStrategy::Fixed {
2162 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2163 max_attempts: 2,
2164 })
2165 }
2166 }
2167 },
2168 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2169 delay: BASE_RETRY_DELAY,
2170 max_attempts: 3,
2171 }),
2172 ApiReadResponseError { .. }
2173 | HttpSend { .. }
2174 | DeserializeResponse { .. }
2175 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2176 delay: BASE_RETRY_DELAY,
2177 max_attempts: 3,
2178 }),
2179 // Retrying these errors definitely shouldn't help.
2180 HttpResponseError {
2181 status_code:
2182 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2183 ..
2184 }
2185 | AuthenticationError { .. }
2186 | PermissionError { .. }
2187 | NoApiKey { .. }
2188 | ApiEndpointNotFound { .. }
2189 | PromptTooLarge { .. } => None,
2190 // These errors might be transient, so retry them
2191 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2192 delay: BASE_RETRY_DELAY,
2193 max_attempts: 1,
2194 }),
2195 // Retry all other 4xx and 5xx errors once.
2196 HttpResponseError { status_code, .. }
2197 if status_code.is_client_error() || status_code.is_server_error() =>
2198 {
2199 Some(RetryStrategy::Fixed {
2200 delay: BASE_RETRY_DELAY,
2201 max_attempts: 3,
2202 })
2203 }
2204 Other(err)
2205 if err.is::<language_model::PaymentRequiredError>()
2206 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2207 {
2208 // Retrying won't help for Payment Required or Model Request Limit errors (where
2209 // the user must upgrade to usage-based billing to get more requests, or else wait
2210 // for a significant amount of time for the request limit to reset).
2211 None
2212 }
2213 // Conservatively assume that any other errors are non-retryable
2214 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2215 delay: BASE_RETRY_DELAY,
2216 max_attempts: 2,
2217 }),
2218 }
2219 }
2220}
2221
2222struct RunningTurn {
2223 /// Holds the task that handles agent interaction until the end of the turn.
2224 /// Survives across multiple requests as the model performs tool calls and
2225 /// we run tools, report their results.
2226 _task: Task<()>,
2227 /// The current event stream for the running turn. Used to report a final
2228 /// cancellation event if we cancel the turn.
2229 event_stream: ThreadEventStream,
2230 /// The tools that were enabled for this turn.
2231 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2232}
2233
2234impl RunningTurn {
2235 fn cancel(self) {
2236 log::debug!("Cancelling in progress turn");
2237 self.event_stream.send_canceled();
2238 }
2239}
2240
2241pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2242
2243impl EventEmitter<TokenUsageUpdated> for Thread {}
2244
2245pub struct TitleUpdated;
2246
2247impl EventEmitter<TitleUpdated> for Thread {}
2248
2249pub trait AgentTool
2250where
2251 Self: 'static + Sized,
2252{
2253 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2254 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2255
2256 fn name() -> &'static str;
2257
2258 fn description() -> SharedString {
2259 let schema = schemars::schema_for!(Self::Input);
2260 SharedString::new(
2261 schema
2262 .get("description")
2263 .and_then(|description| description.as_str())
2264 .unwrap_or_default(),
2265 )
2266 }
2267
2268 fn kind() -> acp::ToolKind;
2269
2270 /// The initial tool title to display. Can be updated during the tool run.
2271 fn initial_title(
2272 &self,
2273 input: Result<Self::Input, serde_json::Value>,
2274 cx: &mut App,
2275 ) -> SharedString;
2276
2277 /// Returns the JSON schema that describes the tool's input.
2278 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
2279 language_model::tool_schema::root_schema_for::<Self::Input>(format)
2280 }
2281
2282 /// Some tools rely on a provider for the underlying billing or other reasons.
2283 /// Allow the tool to check if they are compatible, or should be filtered out.
2284 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
2285 true
2286 }
2287
2288 /// Runs the tool with the provided input.
2289 fn run(
2290 self: Arc<Self>,
2291 input: Self::Input,
2292 event_stream: ToolCallEventStream,
2293 cx: &mut App,
2294 ) -> Task<Result<Self::Output>>;
2295
2296 /// Emits events for a previous execution of the tool.
2297 fn replay(
2298 &self,
2299 _input: Self::Input,
2300 _output: Self::Output,
2301 _event_stream: ToolCallEventStream,
2302 _cx: &mut App,
2303 ) -> Result<()> {
2304 Ok(())
2305 }
2306
2307 fn erase(self) -> Arc<dyn AnyAgentTool> {
2308 Arc::new(Erased(Arc::new(self)))
2309 }
2310}
2311
2312pub struct Erased<T>(T);
2313
2314pub struct AgentToolOutput {
2315 pub llm_output: LanguageModelToolResultContent,
2316 pub raw_output: serde_json::Value,
2317}
2318
2319pub trait AnyAgentTool {
2320 fn name(&self) -> SharedString;
2321 fn description(&self) -> SharedString;
2322 fn kind(&self) -> acp::ToolKind;
2323 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2324 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2325 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2326 true
2327 }
2328 fn run(
2329 self: Arc<Self>,
2330 input: serde_json::Value,
2331 event_stream: ToolCallEventStream,
2332 cx: &mut App,
2333 ) -> Task<Result<AgentToolOutput>>;
2334 fn replay(
2335 &self,
2336 input: serde_json::Value,
2337 output: serde_json::Value,
2338 event_stream: ToolCallEventStream,
2339 cx: &mut App,
2340 ) -> Result<()>;
2341}
2342
2343impl<T> AnyAgentTool for Erased<Arc<T>>
2344where
2345 T: AgentTool,
2346{
2347 fn name(&self) -> SharedString {
2348 T::name().into()
2349 }
2350
2351 fn description(&self) -> SharedString {
2352 T::description()
2353 }
2354
2355 fn kind(&self) -> agent_client_protocol::ToolKind {
2356 T::kind()
2357 }
2358
2359 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2360 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2361 self.0.initial_title(parsed_input, _cx)
2362 }
2363
2364 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2365 let mut json = serde_json::to_value(T::input_schema(format))?;
2366 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
2367 Ok(json)
2368 }
2369
2370 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
2371 T::supports_provider(provider)
2372 }
2373
2374 fn run(
2375 self: Arc<Self>,
2376 input: serde_json::Value,
2377 event_stream: ToolCallEventStream,
2378 cx: &mut App,
2379 ) -> Task<Result<AgentToolOutput>> {
2380 cx.spawn(async move |cx| {
2381 let input = serde_json::from_value(input)?;
2382 let output = cx
2383 .update(|cx| self.0.clone().run(input, event_stream, cx))?
2384 .await?;
2385 let raw_output = serde_json::to_value(&output)?;
2386 Ok(AgentToolOutput {
2387 llm_output: output.into(),
2388 raw_output,
2389 })
2390 })
2391 }
2392
2393 fn replay(
2394 &self,
2395 input: serde_json::Value,
2396 output: serde_json::Value,
2397 event_stream: ToolCallEventStream,
2398 cx: &mut App,
2399 ) -> Result<()> {
2400 let input = serde_json::from_value(input)?;
2401 let output = serde_json::from_value(output)?;
2402 self.0.replay(input, output, event_stream, cx)
2403 }
2404}
2405
2406#[derive(Clone)]
2407struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2408
2409impl ThreadEventStream {
2410 fn send_user_message(&self, message: &UserMessage) {
2411 self.0
2412 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2413 .ok();
2414 }
2415
2416 fn send_text(&self, text: &str) {
2417 self.0
2418 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2419 .ok();
2420 }
2421
2422 fn send_thinking(&self, text: &str) {
2423 self.0
2424 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2425 .ok();
2426 }
2427
2428 fn send_tool_call(
2429 &self,
2430 id: &LanguageModelToolUseId,
2431 tool_name: &str,
2432 title: SharedString,
2433 kind: acp::ToolKind,
2434 input: serde_json::Value,
2435 ) {
2436 self.0
2437 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2438 id,
2439 tool_name,
2440 title.to_string(),
2441 kind,
2442 input,
2443 ))))
2444 .ok();
2445 }
2446
2447 fn initial_tool_call(
2448 id: &LanguageModelToolUseId,
2449 tool_name: &str,
2450 title: String,
2451 kind: acp::ToolKind,
2452 input: serde_json::Value,
2453 ) -> acp::ToolCall {
2454 acp::ToolCall::new(id.to_string(), title)
2455 .kind(kind)
2456 .raw_input(input)
2457 .meta(acp::Meta::from_iter([(
2458 "tool_name".into(),
2459 tool_name.into(),
2460 )]))
2461 }
2462
2463 fn update_tool_call_fields(
2464 &self,
2465 tool_use_id: &LanguageModelToolUseId,
2466 fields: acp::ToolCallUpdateFields,
2467 ) {
2468 self.0
2469 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2470 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields).into(),
2471 )))
2472 .ok();
2473 }
2474
2475 fn send_retry(&self, status: acp_thread::RetryStatus) {
2476 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2477 }
2478
2479 fn send_stop(&self, reason: acp::StopReason) {
2480 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2481 }
2482
2483 fn send_canceled(&self) {
2484 self.0
2485 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2486 .ok();
2487 }
2488
2489 fn send_error(&self, error: impl Into<anyhow::Error>) {
2490 self.0.unbounded_send(Err(error.into())).ok();
2491 }
2492}
2493
2494#[derive(Clone)]
2495pub struct ToolCallEventStream {
2496 tool_use_id: LanguageModelToolUseId,
2497 stream: ThreadEventStream,
2498 fs: Option<Arc<dyn Fs>>,
2499}
2500
2501impl ToolCallEventStream {
2502 #[cfg(any(test, feature = "test-support"))]
2503 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2504 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2505
2506 let stream = ToolCallEventStream::new("test_id".into(), ThreadEventStream(events_tx), None);
2507
2508 (stream, ToolCallEventStreamReceiver(events_rx))
2509 }
2510
2511 fn new(
2512 tool_use_id: LanguageModelToolUseId,
2513 stream: ThreadEventStream,
2514 fs: Option<Arc<dyn Fs>>,
2515 ) -> Self {
2516 Self {
2517 tool_use_id,
2518 stream,
2519 fs,
2520 }
2521 }
2522
2523 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2524 self.stream
2525 .update_tool_call_fields(&self.tool_use_id, fields);
2526 }
2527
2528 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2529 self.stream
2530 .0
2531 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2532 acp_thread::ToolCallUpdateDiff {
2533 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
2534 diff,
2535 }
2536 .into(),
2537 )))
2538 .ok();
2539 }
2540
2541 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2542 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2543 return Task::ready(Ok(()));
2544 }
2545
2546 let (response_tx, response_rx) = oneshot::channel();
2547 self.stream
2548 .0
2549 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2550 ToolCallAuthorization {
2551 tool_call: acp::ToolCallUpdate::new(
2552 self.tool_use_id.to_string(),
2553 acp::ToolCallUpdateFields::new().title(title.into()),
2554 ),
2555 options: vec![
2556 acp::PermissionOption::new(
2557 acp::PermissionOptionId::new("always_allow"),
2558 "Always Allow",
2559 acp::PermissionOptionKind::AllowAlways,
2560 ),
2561 acp::PermissionOption::new(
2562 acp::PermissionOptionId::new("allow"),
2563 "Allow",
2564 acp::PermissionOptionKind::AllowOnce,
2565 ),
2566 acp::PermissionOption::new(
2567 acp::PermissionOptionId::new("deny"),
2568 "Deny",
2569 acp::PermissionOptionKind::RejectOnce,
2570 ),
2571 ],
2572 response: response_tx,
2573 },
2574 )))
2575 .ok();
2576 let fs = self.fs.clone();
2577 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2578 "always_allow" => {
2579 if let Some(fs) = fs.clone() {
2580 cx.update(|cx| {
2581 update_settings_file(fs, cx, |settings, _| {
2582 settings
2583 .agent
2584 .get_or_insert_default()
2585 .set_always_allow_tool_actions(true);
2586 });
2587 })?;
2588 }
2589
2590 Ok(())
2591 }
2592 "allow" => Ok(()),
2593 _ => Err(anyhow!("Permission to run tool denied by user")),
2594 })
2595 }
2596}
2597
2598#[cfg(any(test, feature = "test-support"))]
2599pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2600
2601#[cfg(any(test, feature = "test-support"))]
2602impl ToolCallEventStreamReceiver {
2603 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2604 let event = self.0.next().await;
2605 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2606 auth
2607 } else {
2608 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2609 }
2610 }
2611
2612 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
2613 let event = self.0.next().await;
2614 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
2615 update,
2616 )))) = event
2617 {
2618 update.fields
2619 } else {
2620 panic!("Expected update fields but got: {:?}", event);
2621 }
2622 }
2623
2624 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
2625 let event = self.0.next().await;
2626 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
2627 update,
2628 )))) = event
2629 {
2630 update.diff
2631 } else {
2632 panic!("Expected diff but got: {:?}", event);
2633 }
2634 }
2635
2636 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2637 let event = self.0.next().await;
2638 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2639 update,
2640 )))) = event
2641 {
2642 update.terminal
2643 } else {
2644 panic!("Expected terminal but got: {:?}", event);
2645 }
2646 }
2647}
2648
2649#[cfg(any(test, feature = "test-support"))]
2650impl std::ops::Deref for ToolCallEventStreamReceiver {
2651 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2652
2653 fn deref(&self) -> &Self::Target {
2654 &self.0
2655 }
2656}
2657
2658#[cfg(any(test, feature = "test-support"))]
2659impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2660 fn deref_mut(&mut self) -> &mut Self::Target {
2661 &mut self.0
2662 }
2663}
2664
2665impl From<&str> for UserMessageContent {
2666 fn from(text: &str) -> Self {
2667 Self::Text(text.into())
2668 }
2669}
2670
2671impl UserMessageContent {
2672 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
2673 match value {
2674 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2675 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2676 acp::ContentBlock::Audio(_) => {
2677 // TODO
2678 Self::Text("[audio]".to_string())
2679 }
2680 acp::ContentBlock::ResourceLink(resource_link) => {
2681 match MentionUri::parse(&resource_link.uri, path_style) {
2682 Ok(uri) => Self::Mention {
2683 uri,
2684 content: String::new(),
2685 },
2686 Err(err) => {
2687 log::error!("Failed to parse mention link: {}", err);
2688 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2689 }
2690 }
2691 }
2692 acp::ContentBlock::Resource(resource) => match resource.resource {
2693 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2694 match MentionUri::parse(&resource.uri, path_style) {
2695 Ok(uri) => Self::Mention {
2696 uri,
2697 content: resource.text,
2698 },
2699 Err(err) => {
2700 log::error!("Failed to parse mention link: {}", err);
2701 Self::Text(
2702 MarkdownCodeBlock {
2703 tag: &resource.uri,
2704 text: &resource.text,
2705 }
2706 .to_string(),
2707 )
2708 }
2709 }
2710 }
2711 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2712 // TODO
2713 Self::Text("[blob]".to_string())
2714 }
2715 other => {
2716 log::warn!("Unexpected content type: {:?}", other);
2717 Self::Text("[unknown]".to_string())
2718 }
2719 },
2720 other => {
2721 log::warn!("Unexpected content type: {:?}", other);
2722 Self::Text("[unknown]".to_string())
2723 }
2724 }
2725 }
2726}
2727
2728impl From<UserMessageContent> for acp::ContentBlock {
2729 fn from(content: UserMessageContent) -> Self {
2730 match content {
2731 UserMessageContent::Text(text) => text.into(),
2732 UserMessageContent::Image(image) => {
2733 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
2734 }
2735 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
2736 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
2737 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
2738 )),
2739 ),
2740 }
2741 }
2742}
2743
2744fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2745 LanguageModelImage {
2746 source: image_content.data.into(),
2747 size: None,
2748 }
2749}