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 fn was_stopped_by_user(&self, cx: &AsyncApp) -> Result<bool>;
542}
543
544pub trait ThreadEnvironment {
545 fn create_terminal(
546 &self,
547 command: String,
548 cwd: Option<PathBuf>,
549 output_byte_limit: Option<u64>,
550 cx: &mut AsyncApp,
551 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
552}
553
554#[derive(Debug)]
555pub enum ThreadEvent {
556 UserMessage(UserMessage),
557 AgentText(String),
558 AgentThinking(String),
559 ToolCall(acp::ToolCall),
560 ToolCallUpdate(acp_thread::ToolCallUpdate),
561 ToolCallAuthorization(ToolCallAuthorization),
562 Retry(acp_thread::RetryStatus),
563 Stop(acp::StopReason),
564}
565
566#[derive(Debug)]
567pub struct NewTerminal {
568 pub command: String,
569 pub output_byte_limit: Option<u64>,
570 pub cwd: Option<PathBuf>,
571 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
572}
573
574#[derive(Debug)]
575pub struct ToolCallAuthorization {
576 pub tool_call: acp::ToolCallUpdate,
577 pub options: Vec<acp::PermissionOption>,
578 pub response: oneshot::Sender<acp::PermissionOptionId>,
579}
580
581#[derive(Debug, thiserror::Error)]
582enum CompletionError {
583 #[error("max tokens")]
584 MaxTokens,
585 #[error("refusal")]
586 Refusal,
587 #[error(transparent)]
588 Other(#[from] anyhow::Error),
589}
590
591pub struct Thread {
592 id: acp::SessionId,
593 prompt_id: PromptId,
594 updated_at: DateTime<Utc>,
595 title: Option<SharedString>,
596 pending_title_generation: Option<Task<()>>,
597 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
598 summary: Option<SharedString>,
599 messages: Vec<Message>,
600 user_store: Entity<UserStore>,
601 completion_mode: CompletionMode,
602 /// Holds the task that handles agent interaction until the end of the turn.
603 /// Survives across multiple requests as the model performs tool calls and
604 /// we run tools, report their results.
605 running_turn: Option<RunningTurn>,
606 pending_message: Option<AgentMessage>,
607 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
608 tool_use_limit_reached: bool,
609 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
610 #[allow(unused)]
611 cumulative_token_usage: TokenUsage,
612 #[allow(unused)]
613 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
614 context_server_registry: Entity<ContextServerRegistry>,
615 profile_id: AgentProfileId,
616 project_context: Entity<ProjectContext>,
617 templates: Arc<Templates>,
618 model: Option<Arc<dyn LanguageModel>>,
619 summarization_model: Option<Arc<dyn LanguageModel>>,
620 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
621 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
622 pub(crate) project: Entity<Project>,
623 pub(crate) action_log: Entity<ActionLog>,
624 /// Tracks the last time files were read by the agent, to detect external modifications
625 pub(crate) file_read_times: HashMap<PathBuf, fs::MTime>,
626 /// True if this thread was imported from a shared thread and can be synced.
627 imported: bool,
628}
629
630impl Thread {
631 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
632 let image = model.map_or(true, |model| model.supports_images());
633 acp::PromptCapabilities::new()
634 .image(image)
635 .embedded_context(true)
636 }
637
638 pub fn new(
639 project: Entity<Project>,
640 project_context: Entity<ProjectContext>,
641 context_server_registry: Entity<ContextServerRegistry>,
642 templates: Arc<Templates>,
643 model: Option<Arc<dyn LanguageModel>>,
644 cx: &mut Context<Self>,
645 ) -> Self {
646 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
647 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
648 let (prompt_capabilities_tx, prompt_capabilities_rx) =
649 watch::channel(Self::prompt_capabilities(model.as_deref()));
650 Self {
651 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
652 prompt_id: PromptId::new(),
653 updated_at: Utc::now(),
654 title: None,
655 pending_title_generation: None,
656 pending_summary_generation: None,
657 summary: None,
658 messages: Vec::new(),
659 user_store: project.read(cx).user_store(),
660 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
661 running_turn: None,
662 pending_message: None,
663 tools: BTreeMap::default(),
664 tool_use_limit_reached: false,
665 request_token_usage: HashMap::default(),
666 cumulative_token_usage: TokenUsage::default(),
667 initial_project_snapshot: {
668 let project_snapshot = Self::project_snapshot(project.clone(), cx);
669 cx.foreground_executor()
670 .spawn(async move { Some(project_snapshot.await) })
671 .shared()
672 },
673 context_server_registry,
674 profile_id,
675 project_context,
676 templates,
677 model,
678 summarization_model: None,
679 prompt_capabilities_tx,
680 prompt_capabilities_rx,
681 project,
682 action_log,
683 file_read_times: HashMap::default(),
684 imported: false,
685 }
686 }
687
688 pub fn id(&self) -> &acp::SessionId {
689 &self.id
690 }
691
692 /// Returns true if this thread was imported from a shared thread.
693 pub fn is_imported(&self) -> bool {
694 self.imported
695 }
696
697 pub fn replay(
698 &mut self,
699 cx: &mut Context<Self>,
700 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
701 let (tx, rx) = mpsc::unbounded();
702 let stream = ThreadEventStream(tx);
703 for message in &self.messages {
704 match message {
705 Message::User(user_message) => stream.send_user_message(user_message),
706 Message::Agent(assistant_message) => {
707 for content in &assistant_message.content {
708 match content {
709 AgentMessageContent::Text(text) => stream.send_text(text),
710 AgentMessageContent::Thinking { text, .. } => {
711 stream.send_thinking(text)
712 }
713 AgentMessageContent::RedactedThinking(_) => {}
714 AgentMessageContent::ToolUse(tool_use) => {
715 self.replay_tool_call(
716 tool_use,
717 assistant_message.tool_results.get(&tool_use.id),
718 &stream,
719 cx,
720 );
721 }
722 }
723 }
724 }
725 Message::Resume => {}
726 }
727 }
728 rx
729 }
730
731 fn replay_tool_call(
732 &self,
733 tool_use: &LanguageModelToolUse,
734 tool_result: Option<&LanguageModelToolResult>,
735 stream: &ThreadEventStream,
736 cx: &mut Context<Self>,
737 ) {
738 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
739 self.context_server_registry
740 .read(cx)
741 .servers()
742 .find_map(|(_, tools)| {
743 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
744 Some(tool.clone())
745 } else {
746 None
747 }
748 })
749 });
750
751 let Some(tool) = tool else {
752 stream
753 .0
754 .unbounded_send(Ok(ThreadEvent::ToolCall(
755 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
756 .status(acp::ToolCallStatus::Failed)
757 .raw_input(tool_use.input.clone()),
758 )))
759 .ok();
760 return;
761 };
762
763 let title = tool.initial_title(tool_use.input.clone(), cx);
764 let kind = tool.kind();
765 stream.send_tool_call(
766 &tool_use.id,
767 &tool_use.name,
768 title,
769 kind,
770 tool_use.input.clone(),
771 );
772
773 let output = tool_result
774 .as_ref()
775 .and_then(|result| result.output.clone());
776 if let Some(output) = output.clone() {
777 // For replay, we use a dummy cancellation receiver since the tool already completed
778 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
779 let tool_event_stream = ToolCallEventStream::new(
780 tool_use.id.clone(),
781 stream.clone(),
782 Some(self.project.read(cx).fs().clone()),
783 cancellation_rx,
784 );
785 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
786 .log_err();
787 }
788
789 stream.update_tool_call_fields(
790 &tool_use.id,
791 acp::ToolCallUpdateFields::new()
792 .status(
793 tool_result
794 .as_ref()
795 .map_or(acp::ToolCallStatus::Failed, |result| {
796 if result.is_error {
797 acp::ToolCallStatus::Failed
798 } else {
799 acp::ToolCallStatus::Completed
800 }
801 }),
802 )
803 .raw_output(output),
804 );
805 }
806
807 pub fn from_db(
808 id: acp::SessionId,
809 db_thread: DbThread,
810 project: Entity<Project>,
811 project_context: Entity<ProjectContext>,
812 context_server_registry: Entity<ContextServerRegistry>,
813 templates: Arc<Templates>,
814 cx: &mut Context<Self>,
815 ) -> Self {
816 let profile_id = db_thread
817 .profile
818 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
819
820 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
821 db_thread
822 .model
823 .and_then(|model| {
824 let model = SelectedModel {
825 provider: model.provider.clone().into(),
826 model: model.model.into(),
827 };
828 registry.select_model(&model, cx)
829 })
830 .or_else(|| registry.default_model())
831 .map(|model| model.model)
832 });
833
834 if model.is_none() {
835 model = Self::resolve_profile_model(&profile_id, cx);
836 }
837 if model.is_none() {
838 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
839 registry.default_model().map(|model| model.model)
840 });
841 }
842
843 let (prompt_capabilities_tx, prompt_capabilities_rx) =
844 watch::channel(Self::prompt_capabilities(model.as_deref()));
845
846 let action_log = cx.new(|_| ActionLog::new(project.clone()));
847
848 Self {
849 id,
850 prompt_id: PromptId::new(),
851 title: if db_thread.title.is_empty() {
852 None
853 } else {
854 Some(db_thread.title.clone())
855 },
856 pending_title_generation: None,
857 pending_summary_generation: None,
858 summary: db_thread.detailed_summary,
859 messages: db_thread.messages,
860 user_store: project.read(cx).user_store(),
861 completion_mode: db_thread.completion_mode.unwrap_or_default(),
862 running_turn: None,
863 pending_message: None,
864 tools: BTreeMap::default(),
865 tool_use_limit_reached: false,
866 request_token_usage: db_thread.request_token_usage.clone(),
867 cumulative_token_usage: db_thread.cumulative_token_usage,
868 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
869 context_server_registry,
870 profile_id,
871 project_context,
872 templates,
873 model,
874 summarization_model: None,
875 project,
876 action_log,
877 updated_at: db_thread.updated_at,
878 prompt_capabilities_tx,
879 prompt_capabilities_rx,
880 file_read_times: HashMap::default(),
881 imported: db_thread.imported,
882 }
883 }
884
885 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
886 let initial_project_snapshot = self.initial_project_snapshot.clone();
887 let mut thread = DbThread {
888 title: self.title(),
889 messages: self.messages.clone(),
890 updated_at: self.updated_at,
891 detailed_summary: self.summary.clone(),
892 initial_project_snapshot: None,
893 cumulative_token_usage: self.cumulative_token_usage,
894 request_token_usage: self.request_token_usage.clone(),
895 model: self.model.as_ref().map(|model| DbLanguageModel {
896 provider: model.provider_id().to_string(),
897 model: model.name().0.to_string(),
898 }),
899 completion_mode: Some(self.completion_mode),
900 profile: Some(self.profile_id.clone()),
901 imported: self.imported,
902 };
903
904 cx.background_spawn(async move {
905 let initial_project_snapshot = initial_project_snapshot.await;
906 thread.initial_project_snapshot = initial_project_snapshot;
907 thread
908 })
909 }
910
911 /// Create a snapshot of the current project state including git information and unsaved buffers.
912 fn project_snapshot(
913 project: Entity<Project>,
914 cx: &mut Context<Self>,
915 ) -> Task<Arc<ProjectSnapshot>> {
916 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
917 cx.spawn(async move |_, _| {
918 let snapshot = task.await;
919
920 Arc::new(ProjectSnapshot {
921 worktree_snapshots: snapshot.worktree_snapshots,
922 timestamp: Utc::now(),
923 })
924 })
925 }
926
927 pub fn project_context(&self) -> &Entity<ProjectContext> {
928 &self.project_context
929 }
930
931 pub fn project(&self) -> &Entity<Project> {
932 &self.project
933 }
934
935 pub fn action_log(&self) -> &Entity<ActionLog> {
936 &self.action_log
937 }
938
939 pub fn is_empty(&self) -> bool {
940 self.messages.is_empty() && self.title.is_none()
941 }
942
943 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
944 self.model.as_ref()
945 }
946
947 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
948 let old_usage = self.latest_token_usage();
949 self.model = Some(model);
950 let new_caps = Self::prompt_capabilities(self.model.as_deref());
951 let new_usage = self.latest_token_usage();
952 if old_usage != new_usage {
953 cx.emit(TokenUsageUpdated(new_usage));
954 }
955 self.prompt_capabilities_tx.send(new_caps).log_err();
956 cx.notify()
957 }
958
959 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
960 self.summarization_model.as_ref()
961 }
962
963 pub fn set_summarization_model(
964 &mut self,
965 model: Option<Arc<dyn LanguageModel>>,
966 cx: &mut Context<Self>,
967 ) {
968 self.summarization_model = model;
969 cx.notify()
970 }
971
972 pub fn completion_mode(&self) -> CompletionMode {
973 self.completion_mode
974 }
975
976 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
977 let old_usage = self.latest_token_usage();
978 self.completion_mode = mode;
979 let new_usage = self.latest_token_usage();
980 if old_usage != new_usage {
981 cx.emit(TokenUsageUpdated(new_usage));
982 }
983 cx.notify()
984 }
985
986 #[cfg(any(test, feature = "test-support"))]
987 pub fn last_message(&self) -> Option<Message> {
988 if let Some(message) = self.pending_message.clone() {
989 Some(Message::Agent(message))
990 } else {
991 self.messages.last().cloned()
992 }
993 }
994
995 pub fn add_default_tools(
996 &mut self,
997 environment: Rc<dyn ThreadEnvironment>,
998 cx: &mut Context<Self>,
999 ) {
1000 let language_registry = self.project.read(cx).languages().clone();
1001 self.add_tool(CopyPathTool::new(self.project.clone()));
1002 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1003 self.add_tool(DeletePathTool::new(
1004 self.project.clone(),
1005 self.action_log.clone(),
1006 ));
1007 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1008 self.add_tool(EditFileTool::new(
1009 self.project.clone(),
1010 cx.weak_entity(),
1011 language_registry,
1012 Templates::new(),
1013 ));
1014 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1015 self.add_tool(FindPathTool::new(self.project.clone()));
1016 self.add_tool(GrepTool::new(self.project.clone()));
1017 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1018 self.add_tool(MovePathTool::new(self.project.clone()));
1019 self.add_tool(NowTool);
1020 self.add_tool(OpenTool::new(self.project.clone()));
1021 self.add_tool(ReadFileTool::new(
1022 cx.weak_entity(),
1023 self.project.clone(),
1024 self.action_log.clone(),
1025 ));
1026 self.add_tool(SaveFileTool::new(self.project.clone()));
1027 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1028 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1029 self.add_tool(ThinkingTool);
1030 self.add_tool(WebSearchTool);
1031 }
1032
1033 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1034 self.tools.insert(T::name().into(), tool.erase());
1035 }
1036
1037 pub fn remove_tool(&mut self, name: &str) -> bool {
1038 self.tools.remove(name).is_some()
1039 }
1040
1041 pub fn profile(&self) -> &AgentProfileId {
1042 &self.profile_id
1043 }
1044
1045 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1046 if self.profile_id == profile_id {
1047 return;
1048 }
1049
1050 self.profile_id = profile_id;
1051
1052 // Swap to the profile's preferred model when available.
1053 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1054 self.set_model(model, cx);
1055 }
1056 }
1057
1058 pub fn cancel(&mut self, cx: &mut Context<Self>) {
1059 if let Some(running_turn) = self.running_turn.take() {
1060 running_turn.cancel();
1061 }
1062 self.flush_pending_message(cx);
1063 }
1064
1065 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1066 let Some(last_user_message) = self.last_user_message() else {
1067 return;
1068 };
1069
1070 self.request_token_usage
1071 .insert(last_user_message.id.clone(), update);
1072 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1073 cx.notify();
1074 }
1075
1076 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1077 self.cancel(cx);
1078 let Some(position) = self.messages.iter().position(
1079 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1080 ) else {
1081 return Err(anyhow!("Message not found"));
1082 };
1083
1084 for message in self.messages.drain(position..) {
1085 match message {
1086 Message::User(message) => {
1087 self.request_token_usage.remove(&message.id);
1088 }
1089 Message::Agent(_) | Message::Resume => {}
1090 }
1091 }
1092 self.clear_summary();
1093 cx.notify();
1094 Ok(())
1095 }
1096
1097 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1098 let last_user_message = self.last_user_message()?;
1099 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1100 Some(*tokens)
1101 }
1102
1103 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1104 let usage = self.latest_request_token_usage()?;
1105 let model = self.model.clone()?;
1106 Some(acp_thread::TokenUsage {
1107 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1108 used_tokens: usage.total_tokens(),
1109 })
1110 }
1111
1112 /// Get the total input token count as of the message before the given message.
1113 ///
1114 /// Returns `None` if:
1115 /// - `target_id` is the first message (no previous message)
1116 /// - The previous message hasn't received a response yet (no usage data)
1117 /// - `target_id` is not found in the messages
1118 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1119 let mut previous_user_message_id: Option<&UserMessageId> = None;
1120
1121 for message in &self.messages {
1122 if let Message::User(user_msg) = message {
1123 if &user_msg.id == target_id {
1124 let prev_id = previous_user_message_id?;
1125 let usage = self.request_token_usage.get(prev_id)?;
1126 return Some(usage.input_tokens);
1127 }
1128 previous_user_message_id = Some(&user_msg.id);
1129 }
1130 }
1131 None
1132 }
1133
1134 /// Look up the active profile and resolve its preferred model if one is configured.
1135 fn resolve_profile_model(
1136 profile_id: &AgentProfileId,
1137 cx: &mut Context<Self>,
1138 ) -> Option<Arc<dyn LanguageModel>> {
1139 let selection = AgentSettings::get_global(cx)
1140 .profiles
1141 .get(profile_id)?
1142 .default_model
1143 .clone()?;
1144 Self::resolve_model_from_selection(&selection, cx)
1145 }
1146
1147 /// Translate a stored model selection into the configured model from the registry.
1148 fn resolve_model_from_selection(
1149 selection: &LanguageModelSelection,
1150 cx: &mut Context<Self>,
1151 ) -> Option<Arc<dyn LanguageModel>> {
1152 let selected = SelectedModel {
1153 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1154 model: LanguageModelId::from(selection.model.clone()),
1155 };
1156 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1157 registry
1158 .select_model(&selected, cx)
1159 .map(|configured| configured.model)
1160 })
1161 }
1162
1163 pub fn resume(
1164 &mut self,
1165 cx: &mut Context<Self>,
1166 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1167 self.messages.push(Message::Resume);
1168 cx.notify();
1169
1170 log::debug!("Total messages in thread: {}", self.messages.len());
1171 self.run_turn(cx)
1172 }
1173
1174 /// Sending a message results in the model streaming a response, which could include tool calls.
1175 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1176 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1177 pub fn send<T>(
1178 &mut self,
1179 id: UserMessageId,
1180 content: impl IntoIterator<Item = T>,
1181 cx: &mut Context<Self>,
1182 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1183 where
1184 T: Into<UserMessageContent>,
1185 {
1186 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1187 log::debug!("Thread::send content: {:?}", content);
1188
1189 self.messages
1190 .push(Message::User(UserMessage { id, content }));
1191 cx.notify();
1192
1193 self.send_existing(cx)
1194 }
1195
1196 pub fn send_existing(
1197 &mut self,
1198 cx: &mut Context<Self>,
1199 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1200 let model = self.model().context("No language model configured")?;
1201
1202 log::info!("Thread::send called with model: {}", model.name().0);
1203 self.advance_prompt_id();
1204
1205 log::debug!("Total messages in thread: {}", self.messages.len());
1206 self.run_turn(cx)
1207 }
1208
1209 pub fn push_acp_user_block(
1210 &mut self,
1211 id: UserMessageId,
1212 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1213 path_style: PathStyle,
1214 cx: &mut Context<Self>,
1215 ) {
1216 let content = blocks
1217 .into_iter()
1218 .map(|block| UserMessageContent::from_content_block(block, path_style))
1219 .collect::<Vec<_>>();
1220 self.messages
1221 .push(Message::User(UserMessage { id, content }));
1222 cx.notify();
1223 }
1224
1225 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1226 let text = match block {
1227 acp::ContentBlock::Text(text_content) => text_content.text,
1228 acp::ContentBlock::Image(_) => "[image]".to_string(),
1229 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1230 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1231 acp::ContentBlock::Resource(resource) => match resource.resource {
1232 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1233 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1234 _ => "[resource]".to_string(),
1235 },
1236 _ => "[unknown]".to_string(),
1237 };
1238
1239 self.messages.push(Message::Agent(AgentMessage {
1240 content: vec![AgentMessageContent::Text(text)],
1241 ..Default::default()
1242 }));
1243 cx.notify();
1244 }
1245
1246 #[cfg(feature = "eval")]
1247 pub fn proceed(
1248 &mut self,
1249 cx: &mut Context<Self>,
1250 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1251 self.run_turn(cx)
1252 }
1253
1254 fn run_turn(
1255 &mut self,
1256 cx: &mut Context<Self>,
1257 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1258 self.cancel(cx);
1259
1260 let model = self.model.clone().context("No language model configured")?;
1261 let profile = AgentSettings::get_global(cx)
1262 .profiles
1263 .get(&self.profile_id)
1264 .context("Profile not found")?;
1265 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1266 let event_stream = ThreadEventStream(events_tx);
1267 let message_ix = self.messages.len().saturating_sub(1);
1268 self.tool_use_limit_reached = false;
1269 self.clear_summary();
1270 let (cancellation_tx, cancellation_rx) = watch::channel(false);
1271 self.running_turn = Some(RunningTurn {
1272 event_stream: event_stream.clone(),
1273 tools: self.enabled_tools(profile, &model, cx),
1274 cancellation_tx,
1275 _task: cx.spawn(async move |this, cx| {
1276 log::debug!("Starting agent turn execution");
1277
1278 let turn_result =
1279 Self::run_turn_internal(&this, model, &event_stream, cancellation_rx, cx).await;
1280 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1281
1282 match turn_result {
1283 Ok(()) => {
1284 log::debug!("Turn execution completed");
1285 event_stream.send_stop(acp::StopReason::EndTurn);
1286 }
1287 Err(error) => {
1288 log::error!("Turn execution failed: {:?}", error);
1289 match error.downcast::<CompletionError>() {
1290 Ok(CompletionError::Refusal) => {
1291 event_stream.send_stop(acp::StopReason::Refusal);
1292 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1293 }
1294 Ok(CompletionError::MaxTokens) => {
1295 event_stream.send_stop(acp::StopReason::MaxTokens);
1296 }
1297 Ok(CompletionError::Other(error)) | Err(error) => {
1298 event_stream.send_error(error);
1299 }
1300 }
1301 }
1302 }
1303
1304 _ = this.update(cx, |this, _| this.running_turn.take());
1305 }),
1306 });
1307 Ok(events_rx)
1308 }
1309
1310 async fn run_turn_internal(
1311 this: &WeakEntity<Self>,
1312 model: Arc<dyn LanguageModel>,
1313 event_stream: &ThreadEventStream,
1314 cancellation_rx: watch::Receiver<bool>,
1315 cx: &mut AsyncApp,
1316 ) -> Result<()> {
1317 let mut attempt = 0;
1318 let mut intent = CompletionIntent::UserPrompt;
1319 loop {
1320 let request =
1321 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1322
1323 telemetry::event!(
1324 "Agent Thread Completion",
1325 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1326 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1327 model = model.telemetry_id(),
1328 model_provider = model.provider_id().to_string(),
1329 attempt
1330 );
1331
1332 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1333
1334 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1335 Ok(events) => (events, None),
1336 Err(err) => (stream::empty().boxed(), Some(err)),
1337 };
1338 let mut tool_results = FuturesUnordered::new();
1339 while let Some(event) = events.next().await {
1340 log::trace!("Received completion event: {:?}", event);
1341 match event {
1342 Ok(event) => {
1343 tool_results.extend(this.update(cx, |this, cx| {
1344 this.handle_completion_event(
1345 event,
1346 event_stream,
1347 cancellation_rx.clone(),
1348 cx,
1349 )
1350 })??);
1351 }
1352 Err(err) => {
1353 error = Some(err);
1354 break;
1355 }
1356 }
1357 }
1358
1359 let end_turn = tool_results.is_empty();
1360 while let Some(tool_result) = tool_results.next().await {
1361 log::debug!("Tool finished {:?}", tool_result);
1362
1363 event_stream.update_tool_call_fields(
1364 &tool_result.tool_use_id,
1365 acp::ToolCallUpdateFields::new()
1366 .status(if tool_result.is_error {
1367 acp::ToolCallStatus::Failed
1368 } else {
1369 acp::ToolCallStatus::Completed
1370 })
1371 .raw_output(tool_result.output.clone()),
1372 );
1373 this.update(cx, |this, _cx| {
1374 this.pending_message()
1375 .tool_results
1376 .insert(tool_result.tool_use_id.clone(), tool_result);
1377 })?;
1378 }
1379
1380 this.update(cx, |this, cx| {
1381 this.flush_pending_message(cx);
1382 if this.title.is_none() && this.pending_title_generation.is_none() {
1383 this.generate_title(cx);
1384 }
1385 })?;
1386
1387 if let Some(error) = error {
1388 attempt += 1;
1389 let retry = this.update(cx, |this, cx| {
1390 let user_store = this.user_store.read(cx);
1391 this.handle_completion_error(error, attempt, user_store.plan())
1392 })??;
1393 let timer = cx.background_executor().timer(retry.duration);
1394 event_stream.send_retry(retry);
1395 timer.await;
1396 this.update(cx, |this, _cx| {
1397 if let Some(Message::Agent(message)) = this.messages.last() {
1398 if message.tool_results.is_empty() {
1399 intent = CompletionIntent::UserPrompt;
1400 this.messages.push(Message::Resume);
1401 }
1402 }
1403 })?;
1404 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1405 return Err(language_model::ToolUseLimitReachedError.into());
1406 } else if end_turn {
1407 return Ok(());
1408 } else {
1409 intent = CompletionIntent::ToolResults;
1410 attempt = 0;
1411 }
1412 }
1413 }
1414
1415 fn handle_completion_error(
1416 &mut self,
1417 error: LanguageModelCompletionError,
1418 attempt: u8,
1419 plan: Option<Plan>,
1420 ) -> Result<acp_thread::RetryStatus> {
1421 let Some(model) = self.model.as_ref() else {
1422 return Err(anyhow!(error));
1423 };
1424
1425 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1426 match plan {
1427 Some(Plan::V2(_)) => true,
1428 Some(Plan::V1(_)) => self.completion_mode == CompletionMode::Burn,
1429 None => false,
1430 }
1431 } else {
1432 true
1433 };
1434
1435 if !auto_retry {
1436 return Err(anyhow!(error));
1437 }
1438
1439 let Some(strategy) = Self::retry_strategy_for(&error) else {
1440 return Err(anyhow!(error));
1441 };
1442
1443 let max_attempts = match &strategy {
1444 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1445 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1446 };
1447
1448 if attempt > max_attempts {
1449 return Err(anyhow!(error));
1450 }
1451
1452 let delay = match &strategy {
1453 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1454 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1455 Duration::from_secs(delay_secs)
1456 }
1457 RetryStrategy::Fixed { delay, .. } => *delay,
1458 };
1459 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1460
1461 Ok(acp_thread::RetryStatus {
1462 last_error: error.to_string().into(),
1463 attempt: attempt as usize,
1464 max_attempts: max_attempts as usize,
1465 started_at: Instant::now(),
1466 duration: delay,
1467 })
1468 }
1469
1470 /// A helper method that's called on every streamed completion event.
1471 /// Returns an optional tool result task, which the main agentic loop will
1472 /// send back to the model when it resolves.
1473 fn handle_completion_event(
1474 &mut self,
1475 event: LanguageModelCompletionEvent,
1476 event_stream: &ThreadEventStream,
1477 cancellation_rx: watch::Receiver<bool>,
1478 cx: &mut Context<Self>,
1479 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1480 log::trace!("Handling streamed completion event: {:?}", event);
1481 use LanguageModelCompletionEvent::*;
1482
1483 match event {
1484 StartMessage { .. } => {
1485 self.flush_pending_message(cx);
1486 self.pending_message = Some(AgentMessage::default());
1487 }
1488 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1489 Thinking { text, signature } => {
1490 self.handle_thinking_event(text, signature, event_stream, cx)
1491 }
1492 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1493 ReasoningDetails(details) => {
1494 let last_message = self.pending_message();
1495 // Store the last non-empty reasoning_details (overwrites earlier ones)
1496 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
1497 if let serde_json::Value::Array(ref arr) = details {
1498 if !arr.is_empty() {
1499 last_message.reasoning_details = Some(details);
1500 }
1501 } else {
1502 last_message.reasoning_details = Some(details);
1503 }
1504 }
1505 ToolUse(tool_use) => {
1506 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
1507 }
1508 ToolUseJsonParseError {
1509 id,
1510 tool_name,
1511 raw_input,
1512 json_parse_error,
1513 } => {
1514 return Ok(Some(Task::ready(
1515 self.handle_tool_use_json_parse_error_event(
1516 id,
1517 tool_name,
1518 raw_input,
1519 json_parse_error,
1520 ),
1521 )));
1522 }
1523 UsageUpdate(usage) => {
1524 telemetry::event!(
1525 "Agent Thread Completion Usage Updated",
1526 thread_id = self.id.to_string(),
1527 prompt_id = self.prompt_id.to_string(),
1528 model = self.model.as_ref().map(|m| m.telemetry_id()),
1529 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1530 input_tokens = usage.input_tokens,
1531 output_tokens = usage.output_tokens,
1532 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1533 cache_read_input_tokens = usage.cache_read_input_tokens,
1534 );
1535 self.update_token_usage(usage, cx);
1536 }
1537 UsageUpdated { amount, limit } => {
1538 self.update_model_request_usage(amount, limit, cx);
1539 }
1540 ToolUseLimitReached => {
1541 self.tool_use_limit_reached = true;
1542 }
1543 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1544 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1545 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1546 Started | Queued { .. } => {}
1547 }
1548
1549 Ok(None)
1550 }
1551
1552 fn handle_text_event(
1553 &mut self,
1554 new_text: String,
1555 event_stream: &ThreadEventStream,
1556 cx: &mut Context<Self>,
1557 ) {
1558 event_stream.send_text(&new_text);
1559
1560 let last_message = self.pending_message();
1561 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1562 text.push_str(&new_text);
1563 } else {
1564 last_message
1565 .content
1566 .push(AgentMessageContent::Text(new_text));
1567 }
1568
1569 cx.notify();
1570 }
1571
1572 fn handle_thinking_event(
1573 &mut self,
1574 new_text: String,
1575 new_signature: Option<String>,
1576 event_stream: &ThreadEventStream,
1577 cx: &mut Context<Self>,
1578 ) {
1579 event_stream.send_thinking(&new_text);
1580
1581 let last_message = self.pending_message();
1582 if let Some(AgentMessageContent::Thinking { text, signature }) =
1583 last_message.content.last_mut()
1584 {
1585 text.push_str(&new_text);
1586 *signature = new_signature.or(signature.take());
1587 } else {
1588 last_message.content.push(AgentMessageContent::Thinking {
1589 text: new_text,
1590 signature: new_signature,
1591 });
1592 }
1593
1594 cx.notify();
1595 }
1596
1597 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1598 let last_message = self.pending_message();
1599 last_message
1600 .content
1601 .push(AgentMessageContent::RedactedThinking(data));
1602 cx.notify();
1603 }
1604
1605 fn handle_tool_use_event(
1606 &mut self,
1607 tool_use: LanguageModelToolUse,
1608 event_stream: &ThreadEventStream,
1609 cancellation_rx: watch::Receiver<bool>,
1610 cx: &mut Context<Self>,
1611 ) -> Option<Task<LanguageModelToolResult>> {
1612 cx.notify();
1613
1614 let tool = self.tool(tool_use.name.as_ref());
1615 let mut title = SharedString::from(&tool_use.name);
1616 let mut kind = acp::ToolKind::Other;
1617 if let Some(tool) = tool.as_ref() {
1618 title = tool.initial_title(tool_use.input.clone(), cx);
1619 kind = tool.kind();
1620 }
1621
1622 // Ensure the last message ends in the current tool use
1623 let last_message = self.pending_message();
1624 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1625 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1626 if last_tool_use.id == tool_use.id {
1627 *last_tool_use = tool_use.clone();
1628 false
1629 } else {
1630 true
1631 }
1632 } else {
1633 true
1634 }
1635 });
1636
1637 if push_new_tool_use {
1638 event_stream.send_tool_call(
1639 &tool_use.id,
1640 &tool_use.name,
1641 title,
1642 kind,
1643 tool_use.input.clone(),
1644 );
1645 last_message
1646 .content
1647 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1648 } else {
1649 event_stream.update_tool_call_fields(
1650 &tool_use.id,
1651 acp::ToolCallUpdateFields::new()
1652 .title(title.as_str())
1653 .kind(kind)
1654 .raw_input(tool_use.input.clone()),
1655 );
1656 }
1657
1658 if !tool_use.is_input_complete {
1659 return None;
1660 }
1661
1662 let Some(tool) = tool else {
1663 let content = format!("No tool named {} exists", tool_use.name);
1664 return Some(Task::ready(LanguageModelToolResult {
1665 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1666 tool_use_id: tool_use.id,
1667 tool_name: tool_use.name,
1668 is_error: true,
1669 output: None,
1670 }));
1671 };
1672
1673 let fs = self.project.read(cx).fs().clone();
1674 let tool_event_stream = ToolCallEventStream::new(
1675 tool_use.id.clone(),
1676 event_stream.clone(),
1677 Some(fs),
1678 cancellation_rx,
1679 );
1680 tool_event_stream.update_fields(
1681 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
1682 );
1683 let supports_images = self.model().is_some_and(|model| model.supports_images());
1684 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1685 log::debug!("Running tool {}", tool_use.name);
1686 Some(cx.foreground_executor().spawn(async move {
1687 let tool_result = tool_result.await.and_then(|output| {
1688 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1689 && !supports_images
1690 {
1691 return Err(anyhow!(
1692 "Attempted to read an image, but this model doesn't support it.",
1693 ));
1694 }
1695 Ok(output)
1696 });
1697
1698 match tool_result {
1699 Ok(output) => LanguageModelToolResult {
1700 tool_use_id: tool_use.id,
1701 tool_name: tool_use.name,
1702 is_error: false,
1703 content: output.llm_output,
1704 output: Some(output.raw_output),
1705 },
1706 Err(error) => LanguageModelToolResult {
1707 tool_use_id: tool_use.id,
1708 tool_name: tool_use.name,
1709 is_error: true,
1710 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1711 output: Some(error.to_string().into()),
1712 },
1713 }
1714 }))
1715 }
1716
1717 fn handle_tool_use_json_parse_error_event(
1718 &mut self,
1719 tool_use_id: LanguageModelToolUseId,
1720 tool_name: Arc<str>,
1721 raw_input: Arc<str>,
1722 json_parse_error: String,
1723 ) -> LanguageModelToolResult {
1724 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1725 LanguageModelToolResult {
1726 tool_use_id,
1727 tool_name,
1728 is_error: true,
1729 content: LanguageModelToolResultContent::Text(tool_output.into()),
1730 output: Some(serde_json::Value::String(raw_input.to_string())),
1731 }
1732 }
1733
1734 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1735 self.project
1736 .read(cx)
1737 .user_store()
1738 .update(cx, |user_store, cx| {
1739 user_store.update_model_request_usage(
1740 ModelRequestUsage(RequestUsage {
1741 amount: amount as i32,
1742 limit,
1743 }),
1744 cx,
1745 )
1746 });
1747 }
1748
1749 pub fn title(&self) -> SharedString {
1750 self.title.clone().unwrap_or("New Thread".into())
1751 }
1752
1753 pub fn is_generating_summary(&self) -> bool {
1754 self.pending_summary_generation.is_some()
1755 }
1756
1757 pub fn is_generating_title(&self) -> bool {
1758 self.pending_title_generation.is_some()
1759 }
1760
1761 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
1762 if let Some(summary) = self.summary.as_ref() {
1763 return Task::ready(Some(summary.clone())).shared();
1764 }
1765 if let Some(task) = self.pending_summary_generation.clone() {
1766 return task;
1767 }
1768 let Some(model) = self.summarization_model.clone() else {
1769 log::error!("No summarization model available");
1770 return Task::ready(None).shared();
1771 };
1772 let mut request = LanguageModelRequest {
1773 intent: Some(CompletionIntent::ThreadContextSummarization),
1774 temperature: AgentSettings::temperature_for_model(&model, cx),
1775 ..Default::default()
1776 };
1777
1778 for message in &self.messages {
1779 request.messages.extend(message.to_request());
1780 }
1781
1782 request.messages.push(LanguageModelRequestMessage {
1783 role: Role::User,
1784 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1785 cache: false,
1786 reasoning_details: None,
1787 });
1788
1789 let task = cx
1790 .spawn(async move |this, cx| {
1791 let mut summary = String::new();
1792 let mut messages = model.stream_completion(request, cx).await.log_err()?;
1793 while let Some(event) = messages.next().await {
1794 let event = event.log_err()?;
1795 let text = match event {
1796 LanguageModelCompletionEvent::Text(text) => text,
1797 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
1798 this.update(cx, |thread, cx| {
1799 thread.update_model_request_usage(amount, limit, cx);
1800 })
1801 .ok()?;
1802 continue;
1803 }
1804 _ => continue,
1805 };
1806
1807 let mut lines = text.lines();
1808 summary.extend(lines.next());
1809 }
1810
1811 log::debug!("Setting summary: {}", summary);
1812 let summary = SharedString::from(summary);
1813
1814 this.update(cx, |this, cx| {
1815 this.summary = Some(summary.clone());
1816 this.pending_summary_generation = None;
1817 cx.notify()
1818 })
1819 .ok()?;
1820
1821 Some(summary)
1822 })
1823 .shared();
1824 self.pending_summary_generation = Some(task.clone());
1825 task
1826 }
1827
1828 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
1829 let Some(model) = self.summarization_model.clone() else {
1830 return;
1831 };
1832
1833 log::debug!(
1834 "Generating title with model: {:?}",
1835 self.summarization_model.as_ref().map(|model| model.name())
1836 );
1837 let mut request = LanguageModelRequest {
1838 intent: Some(CompletionIntent::ThreadSummarization),
1839 temperature: AgentSettings::temperature_for_model(&model, cx),
1840 ..Default::default()
1841 };
1842
1843 for message in &self.messages {
1844 request.messages.extend(message.to_request());
1845 }
1846
1847 request.messages.push(LanguageModelRequestMessage {
1848 role: Role::User,
1849 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
1850 cache: false,
1851 reasoning_details: None,
1852 });
1853 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
1854 let mut title = String::new();
1855
1856 let generate = async {
1857 let mut messages = model.stream_completion(request, cx).await?;
1858 while let Some(event) = messages.next().await {
1859 let event = event?;
1860 let text = match event {
1861 LanguageModelCompletionEvent::Text(text) => text,
1862 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
1863 this.update(cx, |thread, cx| {
1864 thread.update_model_request_usage(amount, limit, cx);
1865 })?;
1866 continue;
1867 }
1868 _ => continue,
1869 };
1870
1871 let mut lines = text.lines();
1872 title.extend(lines.next());
1873
1874 // Stop if the LLM generated multiple lines.
1875 if lines.next().is_some() {
1876 break;
1877 }
1878 }
1879 anyhow::Ok(())
1880 };
1881
1882 if generate.await.context("failed to generate title").is_ok() {
1883 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
1884 }
1885 _ = this.update(cx, |this, _| this.pending_title_generation = None);
1886 }));
1887 }
1888
1889 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
1890 self.pending_title_generation = None;
1891 if Some(&title) != self.title.as_ref() {
1892 self.title = Some(title);
1893 cx.emit(TitleUpdated);
1894 cx.notify();
1895 }
1896 }
1897
1898 fn clear_summary(&mut self) {
1899 self.summary = None;
1900 self.pending_summary_generation = None;
1901 }
1902
1903 fn last_user_message(&self) -> Option<&UserMessage> {
1904 self.messages
1905 .iter()
1906 .rev()
1907 .find_map(|message| match message {
1908 Message::User(user_message) => Some(user_message),
1909 Message::Agent(_) => None,
1910 Message::Resume => None,
1911 })
1912 }
1913
1914 fn pending_message(&mut self) -> &mut AgentMessage {
1915 self.pending_message.get_or_insert_default()
1916 }
1917
1918 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
1919 let Some(mut message) = self.pending_message.take() else {
1920 return;
1921 };
1922
1923 if message.content.is_empty() {
1924 return;
1925 }
1926
1927 for content in &message.content {
1928 let AgentMessageContent::ToolUse(tool_use) = content else {
1929 continue;
1930 };
1931
1932 if !message.tool_results.contains_key(&tool_use.id) {
1933 message.tool_results.insert(
1934 tool_use.id.clone(),
1935 LanguageModelToolResult {
1936 tool_use_id: tool_use.id.clone(),
1937 tool_name: tool_use.name.clone(),
1938 is_error: true,
1939 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
1940 output: None,
1941 },
1942 );
1943 }
1944 }
1945
1946 self.messages.push(Message::Agent(message));
1947 self.updated_at = Utc::now();
1948 self.clear_summary();
1949 cx.notify()
1950 }
1951
1952 pub(crate) fn build_completion_request(
1953 &self,
1954 completion_intent: CompletionIntent,
1955 cx: &App,
1956 ) -> Result<LanguageModelRequest> {
1957 let model = self.model().context("No language model configured")?;
1958 let tools = if let Some(turn) = self.running_turn.as_ref() {
1959 turn.tools
1960 .iter()
1961 .filter_map(|(tool_name, tool)| {
1962 log::trace!("Including tool: {}", tool_name);
1963 Some(LanguageModelRequestTool {
1964 name: tool_name.to_string(),
1965 description: tool.description().to_string(),
1966 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
1967 })
1968 })
1969 .collect::<Vec<_>>()
1970 } else {
1971 Vec::new()
1972 };
1973
1974 log::debug!("Building completion request");
1975 log::debug!("Completion intent: {:?}", completion_intent);
1976 log::debug!("Completion mode: {:?}", self.completion_mode);
1977
1978 let available_tools: Vec<_> = self
1979 .running_turn
1980 .as_ref()
1981 .map(|turn| turn.tools.keys().cloned().collect())
1982 .unwrap_or_default();
1983
1984 log::debug!("Request includes {} tools", available_tools.len());
1985 let messages = self.build_request_messages(available_tools, cx);
1986 log::debug!("Request will include {} messages", messages.len());
1987
1988 let request = LanguageModelRequest {
1989 thread_id: Some(self.id.to_string()),
1990 prompt_id: Some(self.prompt_id.to_string()),
1991 intent: Some(completion_intent),
1992 mode: Some(self.completion_mode.into()),
1993 messages,
1994 tools,
1995 tool_choice: None,
1996 stop: Vec::new(),
1997 temperature: AgentSettings::temperature_for_model(model, cx),
1998 thinking_allowed: true,
1999 };
2000
2001 log::debug!("Completion request built successfully");
2002 Ok(request)
2003 }
2004
2005 fn enabled_tools(
2006 &self,
2007 profile: &AgentProfileSettings,
2008 model: &Arc<dyn LanguageModel>,
2009 cx: &App,
2010 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2011 fn truncate(tool_name: &SharedString) -> SharedString {
2012 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2013 let mut truncated = tool_name.to_string();
2014 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2015 truncated.into()
2016 } else {
2017 tool_name.clone()
2018 }
2019 }
2020
2021 let mut tools = self
2022 .tools
2023 .iter()
2024 .filter_map(|(tool_name, tool)| {
2025 if tool.supports_provider(&model.provider_id())
2026 && profile.is_tool_enabled(tool_name)
2027 {
2028 Some((truncate(tool_name), tool.clone()))
2029 } else {
2030 None
2031 }
2032 })
2033 .collect::<BTreeMap<_, _>>();
2034
2035 let mut context_server_tools = Vec::new();
2036 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2037 let mut duplicate_tool_names = HashSet::default();
2038 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2039 for (tool_name, tool) in server_tools {
2040 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2041 let tool_name = truncate(tool_name);
2042 if !seen_tools.insert(tool_name.clone()) {
2043 duplicate_tool_names.insert(tool_name.clone());
2044 }
2045 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2046 }
2047 }
2048 }
2049
2050 // When there are duplicate tool names, disambiguate by prefixing them
2051 // with the server ID. In the rare case there isn't enough space for the
2052 // disambiguated tool name, keep only the last tool with this name.
2053 for (server_id, tool_name, tool) in context_server_tools {
2054 if duplicate_tool_names.contains(&tool_name) {
2055 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2056 if available >= 2 {
2057 let mut disambiguated = server_id.0.to_string();
2058 disambiguated.truncate(available - 1);
2059 disambiguated.push('_');
2060 disambiguated.push_str(&tool_name);
2061 tools.insert(disambiguated.into(), tool.clone());
2062 } else {
2063 tools.insert(tool_name, tool.clone());
2064 }
2065 } else {
2066 tools.insert(tool_name, tool.clone());
2067 }
2068 }
2069
2070 tools
2071 }
2072
2073 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2074 self.running_turn.as_ref()?.tools.get(name).cloned()
2075 }
2076
2077 pub fn has_tool(&self, name: &str) -> bool {
2078 self.running_turn
2079 .as_ref()
2080 .is_some_and(|turn| turn.tools.contains_key(name))
2081 }
2082
2083 fn build_request_messages(
2084 &self,
2085 available_tools: Vec<SharedString>,
2086 cx: &App,
2087 ) -> Vec<LanguageModelRequestMessage> {
2088 log::trace!(
2089 "Building request messages from {} thread messages",
2090 self.messages.len()
2091 );
2092
2093 let system_prompt = SystemPromptTemplate {
2094 project: self.project_context.read(cx),
2095 available_tools,
2096 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2097 }
2098 .render(&self.templates)
2099 .context("failed to build system prompt")
2100 .expect("Invalid template");
2101 let mut messages = vec![LanguageModelRequestMessage {
2102 role: Role::System,
2103 content: vec![system_prompt.into()],
2104 cache: false,
2105 reasoning_details: None,
2106 }];
2107 for message in &self.messages {
2108 messages.extend(message.to_request());
2109 }
2110
2111 if let Some(last_message) = messages.last_mut() {
2112 last_message.cache = true;
2113 }
2114
2115 if let Some(message) = self.pending_message.as_ref() {
2116 messages.extend(message.to_request());
2117 }
2118
2119 messages
2120 }
2121
2122 pub fn to_markdown(&self) -> String {
2123 let mut markdown = String::new();
2124 for (ix, message) in self.messages.iter().enumerate() {
2125 if ix > 0 {
2126 markdown.push('\n');
2127 }
2128 markdown.push_str(&message.to_markdown());
2129 }
2130
2131 if let Some(message) = self.pending_message.as_ref() {
2132 markdown.push('\n');
2133 markdown.push_str(&message.to_markdown());
2134 }
2135
2136 markdown
2137 }
2138
2139 fn advance_prompt_id(&mut self) {
2140 self.prompt_id = PromptId::new();
2141 }
2142
2143 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2144 use LanguageModelCompletionError::*;
2145 use http_client::StatusCode;
2146
2147 // General strategy here:
2148 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2149 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2150 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2151 match error {
2152 HttpResponseError {
2153 status_code: StatusCode::TOO_MANY_REQUESTS,
2154 ..
2155 } => Some(RetryStrategy::ExponentialBackoff {
2156 initial_delay: BASE_RETRY_DELAY,
2157 max_attempts: MAX_RETRY_ATTEMPTS,
2158 }),
2159 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2160 Some(RetryStrategy::Fixed {
2161 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2162 max_attempts: MAX_RETRY_ATTEMPTS,
2163 })
2164 }
2165 UpstreamProviderError {
2166 status,
2167 retry_after,
2168 ..
2169 } => match *status {
2170 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2171 Some(RetryStrategy::Fixed {
2172 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2173 max_attempts: MAX_RETRY_ATTEMPTS,
2174 })
2175 }
2176 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2177 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2178 // Internal Server Error could be anything, retry up to 3 times.
2179 max_attempts: 3,
2180 }),
2181 status => {
2182 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2183 // but we frequently get them in practice. See https://http.dev/529
2184 if status.as_u16() == 529 {
2185 Some(RetryStrategy::Fixed {
2186 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2187 max_attempts: MAX_RETRY_ATTEMPTS,
2188 })
2189 } else {
2190 Some(RetryStrategy::Fixed {
2191 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2192 max_attempts: 2,
2193 })
2194 }
2195 }
2196 },
2197 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2198 delay: BASE_RETRY_DELAY,
2199 max_attempts: 3,
2200 }),
2201 ApiReadResponseError { .. }
2202 | HttpSend { .. }
2203 | DeserializeResponse { .. }
2204 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2205 delay: BASE_RETRY_DELAY,
2206 max_attempts: 3,
2207 }),
2208 // Retrying these errors definitely shouldn't help.
2209 HttpResponseError {
2210 status_code:
2211 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2212 ..
2213 }
2214 | AuthenticationError { .. }
2215 | PermissionError { .. }
2216 | NoApiKey { .. }
2217 | ApiEndpointNotFound { .. }
2218 | PromptTooLarge { .. } => None,
2219 // These errors might be transient, so retry them
2220 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2221 delay: BASE_RETRY_DELAY,
2222 max_attempts: 1,
2223 }),
2224 // Retry all other 4xx and 5xx errors once.
2225 HttpResponseError { status_code, .. }
2226 if status_code.is_client_error() || status_code.is_server_error() =>
2227 {
2228 Some(RetryStrategy::Fixed {
2229 delay: BASE_RETRY_DELAY,
2230 max_attempts: 3,
2231 })
2232 }
2233 Other(err)
2234 if err.is::<language_model::PaymentRequiredError>()
2235 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2236 {
2237 // Retrying won't help for Payment Required or Model Request Limit errors (where
2238 // the user must upgrade to usage-based billing to get more requests, or else wait
2239 // for a significant amount of time for the request limit to reset).
2240 None
2241 }
2242 // Conservatively assume that any other errors are non-retryable
2243 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2244 delay: BASE_RETRY_DELAY,
2245 max_attempts: 2,
2246 }),
2247 }
2248 }
2249}
2250
2251struct RunningTurn {
2252 /// Holds the task that handles agent interaction until the end of the turn.
2253 /// Survives across multiple requests as the model performs tool calls and
2254 /// we run tools, report their results.
2255 _task: Task<()>,
2256 /// The current event stream for the running turn. Used to report a final
2257 /// cancellation event if we cancel the turn.
2258 event_stream: ThreadEventStream,
2259 /// The tools that were enabled for this turn.
2260 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2261 /// Sender to signal tool cancellation. When cancel is called, this is
2262 /// set to true so all tools can detect user-initiated cancellation.
2263 cancellation_tx: watch::Sender<bool>,
2264}
2265
2266impl RunningTurn {
2267 fn cancel(mut self) {
2268 log::debug!("Cancelling in progress turn");
2269 self.cancellation_tx.send(true).ok();
2270 self.event_stream.send_canceled();
2271 }
2272}
2273
2274pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2275
2276impl EventEmitter<TokenUsageUpdated> for Thread {}
2277
2278pub struct TitleUpdated;
2279
2280impl EventEmitter<TitleUpdated> for Thread {}
2281
2282pub trait AgentTool
2283where
2284 Self: 'static + Sized,
2285{
2286 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2287 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2288
2289 fn name() -> &'static str;
2290
2291 fn description() -> SharedString {
2292 let schema = schemars::schema_for!(Self::Input);
2293 SharedString::new(
2294 schema
2295 .get("description")
2296 .and_then(|description| description.as_str())
2297 .unwrap_or_default(),
2298 )
2299 }
2300
2301 fn kind() -> acp::ToolKind;
2302
2303 /// The initial tool title to display. Can be updated during the tool run.
2304 fn initial_title(
2305 &self,
2306 input: Result<Self::Input, serde_json::Value>,
2307 cx: &mut App,
2308 ) -> SharedString;
2309
2310 /// Returns the JSON schema that describes the tool's input.
2311 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
2312 language_model::tool_schema::root_schema_for::<Self::Input>(format)
2313 }
2314
2315 /// Some tools rely on a provider for the underlying billing or other reasons.
2316 /// Allow the tool to check if they are compatible, or should be filtered out.
2317 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
2318 true
2319 }
2320
2321 /// Runs the tool with the provided input.
2322 fn run(
2323 self: Arc<Self>,
2324 input: Self::Input,
2325 event_stream: ToolCallEventStream,
2326 cx: &mut App,
2327 ) -> Task<Result<Self::Output>>;
2328
2329 /// Emits events for a previous execution of the tool.
2330 fn replay(
2331 &self,
2332 _input: Self::Input,
2333 _output: Self::Output,
2334 _event_stream: ToolCallEventStream,
2335 _cx: &mut App,
2336 ) -> Result<()> {
2337 Ok(())
2338 }
2339
2340 fn erase(self) -> Arc<dyn AnyAgentTool> {
2341 Arc::new(Erased(Arc::new(self)))
2342 }
2343}
2344
2345pub struct Erased<T>(T);
2346
2347pub struct AgentToolOutput {
2348 pub llm_output: LanguageModelToolResultContent,
2349 pub raw_output: serde_json::Value,
2350}
2351
2352pub trait AnyAgentTool {
2353 fn name(&self) -> SharedString;
2354 fn description(&self) -> SharedString;
2355 fn kind(&self) -> acp::ToolKind;
2356 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2357 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2358 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2359 true
2360 }
2361 fn run(
2362 self: Arc<Self>,
2363 input: serde_json::Value,
2364 event_stream: ToolCallEventStream,
2365 cx: &mut App,
2366 ) -> Task<Result<AgentToolOutput>>;
2367 fn replay(
2368 &self,
2369 input: serde_json::Value,
2370 output: serde_json::Value,
2371 event_stream: ToolCallEventStream,
2372 cx: &mut App,
2373 ) -> Result<()>;
2374}
2375
2376impl<T> AnyAgentTool for Erased<Arc<T>>
2377where
2378 T: AgentTool,
2379{
2380 fn name(&self) -> SharedString {
2381 T::name().into()
2382 }
2383
2384 fn description(&self) -> SharedString {
2385 T::description()
2386 }
2387
2388 fn kind(&self) -> agent_client_protocol::ToolKind {
2389 T::kind()
2390 }
2391
2392 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2393 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2394 self.0.initial_title(parsed_input, _cx)
2395 }
2396
2397 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2398 let mut json = serde_json::to_value(T::input_schema(format))?;
2399 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
2400 Ok(json)
2401 }
2402
2403 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
2404 T::supports_provider(provider)
2405 }
2406
2407 fn run(
2408 self: Arc<Self>,
2409 input: serde_json::Value,
2410 event_stream: ToolCallEventStream,
2411 cx: &mut App,
2412 ) -> Task<Result<AgentToolOutput>> {
2413 cx.spawn(async move |cx| {
2414 let input = serde_json::from_value(input)?;
2415 let output = cx
2416 .update(|cx| self.0.clone().run(input, event_stream, cx))
2417 .await?;
2418 let raw_output = serde_json::to_value(&output)?;
2419 Ok(AgentToolOutput {
2420 llm_output: output.into(),
2421 raw_output,
2422 })
2423 })
2424 }
2425
2426 fn replay(
2427 &self,
2428 input: serde_json::Value,
2429 output: serde_json::Value,
2430 event_stream: ToolCallEventStream,
2431 cx: &mut App,
2432 ) -> Result<()> {
2433 let input = serde_json::from_value(input)?;
2434 let output = serde_json::from_value(output)?;
2435 self.0.replay(input, output, event_stream, cx)
2436 }
2437}
2438
2439#[derive(Clone)]
2440struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2441
2442impl ThreadEventStream {
2443 fn send_user_message(&self, message: &UserMessage) {
2444 self.0
2445 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2446 .ok();
2447 }
2448
2449 fn send_text(&self, text: &str) {
2450 self.0
2451 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2452 .ok();
2453 }
2454
2455 fn send_thinking(&self, text: &str) {
2456 self.0
2457 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2458 .ok();
2459 }
2460
2461 fn send_tool_call(
2462 &self,
2463 id: &LanguageModelToolUseId,
2464 tool_name: &str,
2465 title: SharedString,
2466 kind: acp::ToolKind,
2467 input: serde_json::Value,
2468 ) {
2469 self.0
2470 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2471 id,
2472 tool_name,
2473 title.to_string(),
2474 kind,
2475 input,
2476 ))))
2477 .ok();
2478 }
2479
2480 fn initial_tool_call(
2481 id: &LanguageModelToolUseId,
2482 tool_name: &str,
2483 title: String,
2484 kind: acp::ToolKind,
2485 input: serde_json::Value,
2486 ) -> acp::ToolCall {
2487 acp::ToolCall::new(id.to_string(), title)
2488 .kind(kind)
2489 .raw_input(input)
2490 .meta(acp::Meta::from_iter([(
2491 "tool_name".into(),
2492 tool_name.into(),
2493 )]))
2494 }
2495
2496 fn update_tool_call_fields(
2497 &self,
2498 tool_use_id: &LanguageModelToolUseId,
2499 fields: acp::ToolCallUpdateFields,
2500 ) {
2501 self.0
2502 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2503 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields).into(),
2504 )))
2505 .ok();
2506 }
2507
2508 fn send_retry(&self, status: acp_thread::RetryStatus) {
2509 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2510 }
2511
2512 fn send_stop(&self, reason: acp::StopReason) {
2513 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2514 }
2515
2516 fn send_canceled(&self) {
2517 self.0
2518 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2519 .ok();
2520 }
2521
2522 fn send_error(&self, error: impl Into<anyhow::Error>) {
2523 self.0.unbounded_send(Err(error.into())).ok();
2524 }
2525}
2526
2527#[derive(Clone)]
2528pub struct ToolCallEventStream {
2529 tool_use_id: LanguageModelToolUseId,
2530 stream: ThreadEventStream,
2531 fs: Option<Arc<dyn Fs>>,
2532 cancellation_rx: watch::Receiver<bool>,
2533}
2534
2535impl ToolCallEventStream {
2536 #[cfg(any(test, feature = "test-support"))]
2537 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2538 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2539 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
2540
2541 let stream = ToolCallEventStream::new(
2542 "test_id".into(),
2543 ThreadEventStream(events_tx),
2544 None,
2545 cancellation_rx,
2546 );
2547
2548 (stream, ToolCallEventStreamReceiver(events_rx))
2549 }
2550
2551 fn new(
2552 tool_use_id: LanguageModelToolUseId,
2553 stream: ThreadEventStream,
2554 fs: Option<Arc<dyn Fs>>,
2555 cancellation_rx: watch::Receiver<bool>,
2556 ) -> Self {
2557 Self {
2558 tool_use_id,
2559 stream,
2560 fs,
2561 cancellation_rx,
2562 }
2563 }
2564
2565 /// Returns a future that resolves when the user cancels the tool call.
2566 /// Tools should select on this alongside their main work to detect user cancellation.
2567 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
2568 let mut rx = self.cancellation_rx.clone();
2569 async move {
2570 loop {
2571 if *rx.borrow() {
2572 return;
2573 }
2574 if rx.changed().await.is_err() {
2575 // Sender dropped, will never be cancelled
2576 std::future::pending::<()>().await;
2577 }
2578 }
2579 }
2580 }
2581
2582 /// Returns true if the user has cancelled this tool call.
2583 /// This is useful for checking cancellation state after an operation completes,
2584 /// to determine if the completion was due to user cancellation.
2585 pub fn was_cancelled_by_user(&self) -> bool {
2586 *self.cancellation_rx.clone().borrow()
2587 }
2588
2589 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2590 self.stream
2591 .update_tool_call_fields(&self.tool_use_id, fields);
2592 }
2593
2594 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2595 self.stream
2596 .0
2597 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2598 acp_thread::ToolCallUpdateDiff {
2599 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
2600 diff,
2601 }
2602 .into(),
2603 )))
2604 .ok();
2605 }
2606
2607 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2608 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2609 return Task::ready(Ok(()));
2610 }
2611
2612 let (response_tx, response_rx) = oneshot::channel();
2613 self.stream
2614 .0
2615 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2616 ToolCallAuthorization {
2617 tool_call: acp::ToolCallUpdate::new(
2618 self.tool_use_id.to_string(),
2619 acp::ToolCallUpdateFields::new().title(title.into()),
2620 ),
2621 options: vec![
2622 acp::PermissionOption::new(
2623 acp::PermissionOptionId::new("always_allow"),
2624 "Always Allow",
2625 acp::PermissionOptionKind::AllowAlways,
2626 ),
2627 acp::PermissionOption::new(
2628 acp::PermissionOptionId::new("allow"),
2629 "Allow",
2630 acp::PermissionOptionKind::AllowOnce,
2631 ),
2632 acp::PermissionOption::new(
2633 acp::PermissionOptionId::new("deny"),
2634 "Deny",
2635 acp::PermissionOptionKind::RejectOnce,
2636 ),
2637 ],
2638 response: response_tx,
2639 },
2640 )))
2641 .ok();
2642 let fs = self.fs.clone();
2643 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2644 "always_allow" => {
2645 if let Some(fs) = fs.clone() {
2646 cx.update(|cx| {
2647 update_settings_file(fs, cx, |settings, _| {
2648 settings
2649 .agent
2650 .get_or_insert_default()
2651 .set_always_allow_tool_actions(true);
2652 });
2653 });
2654 }
2655
2656 Ok(())
2657 }
2658 "allow" => Ok(()),
2659 _ => Err(anyhow!("Permission to run tool denied by user")),
2660 })
2661 }
2662}
2663
2664#[cfg(any(test, feature = "test-support"))]
2665pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2666
2667#[cfg(any(test, feature = "test-support"))]
2668impl ToolCallEventStreamReceiver {
2669 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2670 let event = self.0.next().await;
2671 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2672 auth
2673 } else {
2674 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2675 }
2676 }
2677
2678 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
2679 let event = self.0.next().await;
2680 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
2681 update,
2682 )))) = event
2683 {
2684 update.fields
2685 } else {
2686 panic!("Expected update fields but got: {:?}", event);
2687 }
2688 }
2689
2690 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
2691 let event = self.0.next().await;
2692 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
2693 update,
2694 )))) = event
2695 {
2696 update.diff
2697 } else {
2698 panic!("Expected diff but got: {:?}", event);
2699 }
2700 }
2701
2702 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2703 let event = self.0.next().await;
2704 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2705 update,
2706 )))) = event
2707 {
2708 update.terminal
2709 } else {
2710 panic!("Expected terminal but got: {:?}", event);
2711 }
2712 }
2713}
2714
2715#[cfg(any(test, feature = "test-support"))]
2716impl std::ops::Deref for ToolCallEventStreamReceiver {
2717 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2718
2719 fn deref(&self) -> &Self::Target {
2720 &self.0
2721 }
2722}
2723
2724#[cfg(any(test, feature = "test-support"))]
2725impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2726 fn deref_mut(&mut self) -> &mut Self::Target {
2727 &mut self.0
2728 }
2729}
2730
2731impl From<&str> for UserMessageContent {
2732 fn from(text: &str) -> Self {
2733 Self::Text(text.into())
2734 }
2735}
2736
2737impl UserMessageContent {
2738 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
2739 match value {
2740 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2741 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2742 acp::ContentBlock::Audio(_) => {
2743 // TODO
2744 Self::Text("[audio]".to_string())
2745 }
2746 acp::ContentBlock::ResourceLink(resource_link) => {
2747 match MentionUri::parse(&resource_link.uri, path_style) {
2748 Ok(uri) => Self::Mention {
2749 uri,
2750 content: String::new(),
2751 },
2752 Err(err) => {
2753 log::error!("Failed to parse mention link: {}", err);
2754 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
2755 }
2756 }
2757 }
2758 acp::ContentBlock::Resource(resource) => match resource.resource {
2759 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
2760 match MentionUri::parse(&resource.uri, path_style) {
2761 Ok(uri) => Self::Mention {
2762 uri,
2763 content: resource.text,
2764 },
2765 Err(err) => {
2766 log::error!("Failed to parse mention link: {}", err);
2767 Self::Text(
2768 MarkdownCodeBlock {
2769 tag: &resource.uri,
2770 text: &resource.text,
2771 }
2772 .to_string(),
2773 )
2774 }
2775 }
2776 }
2777 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
2778 // TODO
2779 Self::Text("[blob]".to_string())
2780 }
2781 other => {
2782 log::warn!("Unexpected content type: {:?}", other);
2783 Self::Text("[unknown]".to_string())
2784 }
2785 },
2786 other => {
2787 log::warn!("Unexpected content type: {:?}", other);
2788 Self::Text("[unknown]".to_string())
2789 }
2790 }
2791 }
2792}
2793
2794impl From<UserMessageContent> for acp::ContentBlock {
2795 fn from(content: UserMessageContent) -> Self {
2796 match content {
2797 UserMessageContent::Text(text) => text.into(),
2798 UserMessageContent::Image(image) => {
2799 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
2800 }
2801 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
2802 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
2803 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
2804 )),
2805 ),
2806 }
2807 }
2808}
2809
2810fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
2811 LanguageModelImage {
2812 source: image_content.data.into(),
2813 size: None,
2814 }
2815}