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, SubagentTool, SystemPromptTemplate, Template, Templates,
6 TerminalTool, ThinkingTool, WebSearchTool,
7};
8use acp_thread::{MentionUri, UserMessageId};
9use action_log::ActionLog;
10use feature_flags::{FeatureFlagAppExt as _, SubagentsFeatureFlag};
11
12use agent_client_protocol as acp;
13use agent_settings::{
14 AgentProfileId, AgentProfileSettings, AgentSettings, CompletionMode,
15 SUMMARIZE_THREAD_DETAILED_PROMPT, SUMMARIZE_THREAD_PROMPT,
16};
17use anyhow::{Context as _, Result, anyhow};
18use chrono::{DateTime, Utc};
19use client::{ModelRequestUsage, RequestUsage, UserStore};
20use cloud_llm_client::{CompletionIntent, Plan, UsageLimit};
21use collections::{HashMap, HashSet, IndexMap};
22use fs::Fs;
23use futures::stream;
24use futures::{
25 FutureExt,
26 channel::{mpsc, oneshot},
27 future::Shared,
28 stream::FuturesUnordered,
29};
30use gpui::{
31 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
32};
33use language_model::{
34 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelExt,
35 LanguageModelId, LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry,
36 LanguageModelRequest, LanguageModelRequestMessage, LanguageModelRequestTool,
37 LanguageModelToolResult, LanguageModelToolResultContent, LanguageModelToolSchemaFormat,
38 LanguageModelToolUse, LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage,
39 ZED_CLOUD_PROVIDER_ID,
40};
41use project::Project;
42use prompt_store::ProjectContext;
43use schemars::{JsonSchema, Schema};
44use serde::{Deserialize, Serialize};
45use settings::{LanguageModelSelection, Settings, update_settings_file};
46use smol::stream::StreamExt;
47use std::{
48 collections::BTreeMap,
49 ops::RangeInclusive,
50 path::Path,
51 rc::Rc,
52 sync::Arc,
53 time::{Duration, Instant},
54};
55use std::{fmt::Write, path::PathBuf};
56use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
57use uuid::Uuid;
58
59const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
60pub const MAX_TOOL_NAME_LENGTH: usize = 64;
61pub const MAX_SUBAGENT_DEPTH: u8 = 4;
62pub const MAX_PARALLEL_SUBAGENTS: usize = 8;
63
64/// Context passed to a subagent thread for lifecycle management
65#[derive(Clone)]
66pub struct SubagentContext {
67 /// ID of the parent thread
68 pub parent_thread_id: acp::SessionId,
69
70 /// ID of the tool call that spawned this subagent
71 pub tool_use_id: LanguageModelToolUseId,
72
73 /// Current depth level (0 = root agent, 1 = first-level subagent, etc.)
74 pub depth: u8,
75
76 /// Prompt to send when subagent completes successfully
77 pub summary_prompt: String,
78
79 /// Prompt to send when context is running low (≤25% remaining)
80 pub context_low_prompt: String,
81}
82
83/// The ID of the user prompt that initiated a request.
84///
85/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
86#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
87pub struct PromptId(Arc<str>);
88
89impl PromptId {
90 pub fn new() -> Self {
91 Self(Uuid::new_v4().to_string().into())
92 }
93}
94
95impl std::fmt::Display for PromptId {
96 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
97 write!(f, "{}", self.0)
98 }
99}
100
101pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
102pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
103
104#[derive(Debug, Clone)]
105enum RetryStrategy {
106 ExponentialBackoff {
107 initial_delay: Duration,
108 max_attempts: u8,
109 },
110 Fixed {
111 delay: Duration,
112 max_attempts: u8,
113 },
114}
115
116#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
117pub enum Message {
118 User(UserMessage),
119 Agent(AgentMessage),
120 Resume,
121}
122
123impl Message {
124 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
125 match self {
126 Message::Agent(agent_message) => Some(agent_message),
127 _ => None,
128 }
129 }
130
131 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
132 match self {
133 Message::User(message) => {
134 if message.content.is_empty() {
135 vec![]
136 } else {
137 vec![message.to_request()]
138 }
139 }
140 Message::Agent(message) => message.to_request(),
141 Message::Resume => vec![LanguageModelRequestMessage {
142 role: Role::User,
143 content: vec!["Continue where you left off".into()],
144 cache: false,
145 reasoning_details: None,
146 }],
147 }
148 }
149
150 pub fn to_markdown(&self) -> String {
151 match self {
152 Message::User(message) => message.to_markdown(),
153 Message::Agent(message) => message.to_markdown(),
154 Message::Resume => "[resume]\n".into(),
155 }
156 }
157
158 pub fn role(&self) -> Role {
159 match self {
160 Message::User(_) | Message::Resume => Role::User,
161 Message::Agent(_) => Role::Assistant,
162 }
163 }
164}
165
166#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
167pub struct UserMessage {
168 pub id: UserMessageId,
169 pub content: Vec<UserMessageContent>,
170}
171
172#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
173pub enum UserMessageContent {
174 Text(String),
175 Mention { uri: MentionUri, content: String },
176 Image(LanguageModelImage),
177}
178
179impl UserMessage {
180 pub fn to_markdown(&self) -> String {
181 let mut markdown = String::from("## User\n\n");
182
183 for content in &self.content {
184 match content {
185 UserMessageContent::Text(text) => {
186 markdown.push_str(text);
187 markdown.push('\n');
188 }
189 UserMessageContent::Image(_) => {
190 markdown.push_str("<image />\n");
191 }
192 UserMessageContent::Mention { uri, content } => {
193 if !content.is_empty() {
194 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
195 } else {
196 let _ = writeln!(&mut markdown, "{}", uri.as_link());
197 }
198 }
199 }
200 }
201
202 markdown
203 }
204
205 fn to_request(&self) -> LanguageModelRequestMessage {
206 let mut message = LanguageModelRequestMessage {
207 role: Role::User,
208 content: Vec::with_capacity(self.content.len()),
209 cache: false,
210 reasoning_details: None,
211 };
212
213 const OPEN_CONTEXT: &str = "<context>\n\
214 The following items were attached by the user. \
215 They are up-to-date and don't need to be re-read.\n\n";
216
217 const OPEN_FILES_TAG: &str = "<files>";
218 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
219 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
220 const OPEN_SELECTIONS_TAG: &str = "<selections>";
221 const OPEN_THREADS_TAG: &str = "<threads>";
222 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
223 const OPEN_RULES_TAG: &str =
224 "<rules>\nThe user has specified the following rules that should be applied:\n";
225
226 let mut file_context = OPEN_FILES_TAG.to_string();
227 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
228 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
229 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
230 let mut thread_context = OPEN_THREADS_TAG.to_string();
231 let mut fetch_context = OPEN_FETCH_TAG.to_string();
232 let mut rules_context = OPEN_RULES_TAG.to_string();
233
234 for chunk in &self.content {
235 let chunk = match chunk {
236 UserMessageContent::Text(text) => {
237 language_model::MessageContent::Text(text.clone())
238 }
239 UserMessageContent::Image(value) => {
240 language_model::MessageContent::Image(value.clone())
241 }
242 UserMessageContent::Mention { uri, content } => {
243 match uri {
244 MentionUri::File { abs_path } => {
245 write!(
246 &mut file_context,
247 "\n{}",
248 MarkdownCodeBlock {
249 tag: &codeblock_tag(abs_path, None),
250 text: &content.to_string(),
251 }
252 )
253 .ok();
254 }
255 MentionUri::PastedImage => {
256 debug_panic!("pasted image URI should not be used in mention content")
257 }
258 MentionUri::Directory { .. } => {
259 write!(&mut directory_context, "\n{}\n", content).ok();
260 }
261 MentionUri::Symbol {
262 abs_path: path,
263 line_range,
264 ..
265 } => {
266 write!(
267 &mut symbol_context,
268 "\n{}",
269 MarkdownCodeBlock {
270 tag: &codeblock_tag(path, Some(line_range)),
271 text: content
272 }
273 )
274 .ok();
275 }
276 MentionUri::Selection {
277 abs_path: path,
278 line_range,
279 ..
280 } => {
281 write!(
282 &mut selection_context,
283 "\n{}",
284 MarkdownCodeBlock {
285 tag: &codeblock_tag(
286 path.as_deref().unwrap_or("Untitled".as_ref()),
287 Some(line_range)
288 ),
289 text: content
290 }
291 )
292 .ok();
293 }
294 MentionUri::Thread { .. } => {
295 write!(&mut thread_context, "\n{}\n", content).ok();
296 }
297 MentionUri::TextThread { .. } => {
298 write!(&mut thread_context, "\n{}\n", content).ok();
299 }
300 MentionUri::Rule { .. } => {
301 write!(
302 &mut rules_context,
303 "\n{}",
304 MarkdownCodeBlock {
305 tag: "",
306 text: content
307 }
308 )
309 .ok();
310 }
311 MentionUri::Fetch { url } => {
312 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
313 }
314 }
315
316 language_model::MessageContent::Text(uri.as_link().to_string())
317 }
318 };
319
320 message.content.push(chunk);
321 }
322
323 let len_before_context = message.content.len();
324
325 if file_context.len() > OPEN_FILES_TAG.len() {
326 file_context.push_str("</files>\n");
327 message
328 .content
329 .push(language_model::MessageContent::Text(file_context));
330 }
331
332 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
333 directory_context.push_str("</directories>\n");
334 message
335 .content
336 .push(language_model::MessageContent::Text(directory_context));
337 }
338
339 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
340 symbol_context.push_str("</symbols>\n");
341 message
342 .content
343 .push(language_model::MessageContent::Text(symbol_context));
344 }
345
346 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
347 selection_context.push_str("</selections>\n");
348 message
349 .content
350 .push(language_model::MessageContent::Text(selection_context));
351 }
352
353 if thread_context.len() > OPEN_THREADS_TAG.len() {
354 thread_context.push_str("</threads>\n");
355 message
356 .content
357 .push(language_model::MessageContent::Text(thread_context));
358 }
359
360 if fetch_context.len() > OPEN_FETCH_TAG.len() {
361 fetch_context.push_str("</fetched_urls>\n");
362 message
363 .content
364 .push(language_model::MessageContent::Text(fetch_context));
365 }
366
367 if rules_context.len() > OPEN_RULES_TAG.len() {
368 rules_context.push_str("</user_rules>\n");
369 message
370 .content
371 .push(language_model::MessageContent::Text(rules_context));
372 }
373
374 if message.content.len() > len_before_context {
375 message.content.insert(
376 len_before_context,
377 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
378 );
379 message
380 .content
381 .push(language_model::MessageContent::Text("</context>".into()));
382 }
383
384 message
385 }
386}
387
388fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
389 let mut result = String::new();
390
391 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
392 let _ = write!(result, "{} ", extension);
393 }
394
395 let _ = write!(result, "{}", full_path.display());
396
397 if let Some(range) = line_range {
398 if range.start() == range.end() {
399 let _ = write!(result, ":{}", range.start() + 1);
400 } else {
401 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
402 }
403 }
404
405 result
406}
407
408impl AgentMessage {
409 pub fn to_markdown(&self) -> String {
410 let mut markdown = String::from("## Assistant\n\n");
411
412 for content in &self.content {
413 match content {
414 AgentMessageContent::Text(text) => {
415 markdown.push_str(text);
416 markdown.push('\n');
417 }
418 AgentMessageContent::Thinking { text, .. } => {
419 markdown.push_str("<think>");
420 markdown.push_str(text);
421 markdown.push_str("</think>\n");
422 }
423 AgentMessageContent::RedactedThinking(_) => {
424 markdown.push_str("<redacted_thinking />\n")
425 }
426 AgentMessageContent::ToolUse(tool_use) => {
427 markdown.push_str(&format!(
428 "**Tool Use**: {} (ID: {})\n",
429 tool_use.name, tool_use.id
430 ));
431 markdown.push_str(&format!(
432 "{}\n",
433 MarkdownCodeBlock {
434 tag: "json",
435 text: &format!("{:#}", tool_use.input)
436 }
437 ));
438 }
439 }
440 }
441
442 for tool_result in self.tool_results.values() {
443 markdown.push_str(&format!(
444 "**Tool Result**: {} (ID: {})\n\n",
445 tool_result.tool_name, tool_result.tool_use_id
446 ));
447 if tool_result.is_error {
448 markdown.push_str("**ERROR:**\n");
449 }
450
451 match &tool_result.content {
452 LanguageModelToolResultContent::Text(text) => {
453 writeln!(markdown, "{text}\n").ok();
454 }
455 LanguageModelToolResultContent::Image(_) => {
456 writeln!(markdown, "<image />\n").ok();
457 }
458 }
459
460 if let Some(output) = tool_result.output.as_ref() {
461 writeln!(
462 markdown,
463 "**Debug Output**:\n\n```json\n{}\n```\n",
464 serde_json::to_string_pretty(output).unwrap()
465 )
466 .unwrap();
467 }
468 }
469
470 markdown
471 }
472
473 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
474 let mut assistant_message = LanguageModelRequestMessage {
475 role: Role::Assistant,
476 content: Vec::with_capacity(self.content.len()),
477 cache: false,
478 reasoning_details: self.reasoning_details.clone(),
479 };
480 for chunk in &self.content {
481 match chunk {
482 AgentMessageContent::Text(text) => {
483 assistant_message
484 .content
485 .push(language_model::MessageContent::Text(text.clone()));
486 }
487 AgentMessageContent::Thinking { text, signature } => {
488 assistant_message
489 .content
490 .push(language_model::MessageContent::Thinking {
491 text: text.clone(),
492 signature: signature.clone(),
493 });
494 }
495 AgentMessageContent::RedactedThinking(value) => {
496 assistant_message.content.push(
497 language_model::MessageContent::RedactedThinking(value.clone()),
498 );
499 }
500 AgentMessageContent::ToolUse(tool_use) => {
501 if self.tool_results.contains_key(&tool_use.id) {
502 assistant_message
503 .content
504 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
505 }
506 }
507 };
508 }
509
510 let mut user_message = LanguageModelRequestMessage {
511 role: Role::User,
512 content: Vec::new(),
513 cache: false,
514 reasoning_details: None,
515 };
516
517 for tool_result in self.tool_results.values() {
518 let mut tool_result = tool_result.clone();
519 // Surprisingly, the API fails if we return an empty string here.
520 // It thinks we are sending a tool use without a tool result.
521 if tool_result.content.is_empty() {
522 tool_result.content = "<Tool returned an empty string>".into();
523 }
524 user_message
525 .content
526 .push(language_model::MessageContent::ToolResult(tool_result));
527 }
528
529 let mut messages = Vec::new();
530 if !assistant_message.content.is_empty() {
531 messages.push(assistant_message);
532 }
533 if !user_message.content.is_empty() {
534 messages.push(user_message);
535 }
536 messages
537 }
538}
539
540#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
541pub struct AgentMessage {
542 pub content: Vec<AgentMessageContent>,
543 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
544 pub reasoning_details: Option<serde_json::Value>,
545}
546
547#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
548pub enum AgentMessageContent {
549 Text(String),
550 Thinking {
551 text: String,
552 signature: Option<String>,
553 },
554 RedactedThinking(String),
555 ToolUse(LanguageModelToolUse),
556}
557
558pub trait TerminalHandle {
559 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
560 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
561 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
562 fn kill(&self, cx: &AsyncApp) -> Result<()>;
563 fn was_stopped_by_user(&self, cx: &AsyncApp) -> Result<bool>;
564}
565
566pub trait ThreadEnvironment {
567 fn create_terminal(
568 &self,
569 command: String,
570 cwd: Option<PathBuf>,
571 output_byte_limit: Option<u64>,
572 cx: &mut AsyncApp,
573 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
574}
575
576#[derive(Debug)]
577pub enum ThreadEvent {
578 UserMessage(UserMessage),
579 AgentText(String),
580 AgentThinking(String),
581 ToolCall(acp::ToolCall),
582 ToolCallUpdate(acp_thread::ToolCallUpdate),
583 ToolCallAuthorization(ToolCallAuthorization),
584 Retry(acp_thread::RetryStatus),
585 Stop(acp::StopReason),
586}
587
588#[derive(Debug)]
589pub struct NewTerminal {
590 pub command: String,
591 pub output_byte_limit: Option<u64>,
592 pub cwd: Option<PathBuf>,
593 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
594}
595
596#[derive(Debug)]
597pub struct ToolCallAuthorization {
598 pub tool_call: acp::ToolCallUpdate,
599 pub options: Vec<acp::PermissionOption>,
600 pub response: oneshot::Sender<acp::PermissionOptionId>,
601}
602
603#[derive(Debug, thiserror::Error)]
604enum CompletionError {
605 #[error("max tokens")]
606 MaxTokens,
607 #[error("refusal")]
608 Refusal,
609 #[error(transparent)]
610 Other(#[from] anyhow::Error),
611}
612
613pub struct Thread {
614 id: acp::SessionId,
615 prompt_id: PromptId,
616 updated_at: DateTime<Utc>,
617 title: Option<SharedString>,
618 pending_title_generation: Option<Task<()>>,
619 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
620 summary: Option<SharedString>,
621 messages: Vec<Message>,
622 user_store: Entity<UserStore>,
623 completion_mode: CompletionMode,
624 /// Holds the task that handles agent interaction until the end of the turn.
625 /// Survives across multiple requests as the model performs tool calls and
626 /// we run tools, report their results.
627 running_turn: Option<RunningTurn>,
628 pending_message: Option<AgentMessage>,
629 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
630 tool_use_limit_reached: bool,
631 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
632 #[allow(unused)]
633 cumulative_token_usage: TokenUsage,
634 #[allow(unused)]
635 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
636 context_server_registry: Entity<ContextServerRegistry>,
637 profile_id: AgentProfileId,
638 project_context: Entity<ProjectContext>,
639 templates: Arc<Templates>,
640 model: Option<Arc<dyn LanguageModel>>,
641 summarization_model: Option<Arc<dyn LanguageModel>>,
642 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
643 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
644 pub(crate) project: Entity<Project>,
645 pub(crate) action_log: Entity<ActionLog>,
646 /// Tracks the last time files were read by the agent, to detect external modifications
647 pub(crate) file_read_times: HashMap<PathBuf, fs::MTime>,
648 /// True if this thread was imported from a shared thread and can be synced.
649 imported: bool,
650 /// If this is a subagent thread, contains context about the parent
651 subagent_context: Option<SubagentContext>,
652 /// Weak references to running subagent threads for cancellation propagation
653 running_subagents: Vec<WeakEntity<Thread>>,
654}
655
656impl Thread {
657 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
658 let image = model.map_or(true, |model| model.supports_images());
659 acp::PromptCapabilities::new()
660 .image(image)
661 .embedded_context(true)
662 }
663
664 pub fn new(
665 project: Entity<Project>,
666 project_context: Entity<ProjectContext>,
667 context_server_registry: Entity<ContextServerRegistry>,
668 templates: Arc<Templates>,
669 model: Option<Arc<dyn LanguageModel>>,
670 cx: &mut Context<Self>,
671 ) -> Self {
672 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
673 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
674 let (prompt_capabilities_tx, prompt_capabilities_rx) =
675 watch::channel(Self::prompt_capabilities(model.as_deref()));
676 Self {
677 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
678 prompt_id: PromptId::new(),
679 updated_at: Utc::now(),
680 title: None,
681 pending_title_generation: None,
682 pending_summary_generation: None,
683 summary: None,
684 messages: Vec::new(),
685 user_store: project.read(cx).user_store(),
686 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
687 running_turn: None,
688 pending_message: None,
689 tools: BTreeMap::default(),
690 tool_use_limit_reached: false,
691 request_token_usage: HashMap::default(),
692 cumulative_token_usage: TokenUsage::default(),
693 initial_project_snapshot: {
694 let project_snapshot = Self::project_snapshot(project.clone(), cx);
695 cx.foreground_executor()
696 .spawn(async move { Some(project_snapshot.await) })
697 .shared()
698 },
699 context_server_registry,
700 profile_id,
701 project_context,
702 templates,
703 model,
704 summarization_model: None,
705 prompt_capabilities_tx,
706 prompt_capabilities_rx,
707 project,
708 action_log,
709 file_read_times: HashMap::default(),
710 imported: false,
711 subagent_context: None,
712 running_subagents: Vec::new(),
713 }
714 }
715
716 pub fn new_subagent(
717 project: Entity<Project>,
718 project_context: Entity<ProjectContext>,
719 context_server_registry: Entity<ContextServerRegistry>,
720 templates: Arc<Templates>,
721 model: Arc<dyn LanguageModel>,
722 subagent_context: SubagentContext,
723 cx: &mut Context<Self>,
724 ) -> Self {
725 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
726 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
727 let (prompt_capabilities_tx, prompt_capabilities_rx) =
728 watch::channel(Self::prompt_capabilities(Some(model.as_ref())));
729 Self {
730 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
731 prompt_id: PromptId::new(),
732 updated_at: Utc::now(),
733 title: None,
734 pending_title_generation: None,
735 pending_summary_generation: None,
736 summary: None,
737 messages: Vec::new(),
738 user_store: project.read(cx).user_store(),
739 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
740 running_turn: None,
741 pending_message: None,
742 tools: BTreeMap::default(),
743 tool_use_limit_reached: false,
744 request_token_usage: HashMap::default(),
745 cumulative_token_usage: TokenUsage::default(),
746 initial_project_snapshot: Task::ready(None).shared(),
747 context_server_registry,
748 profile_id,
749 project_context,
750 templates,
751 model: Some(model),
752 summarization_model: None,
753 prompt_capabilities_tx,
754 prompt_capabilities_rx,
755 project,
756 action_log,
757 file_read_times: HashMap::default(),
758 imported: false,
759 subagent_context: Some(subagent_context),
760 running_subagents: Vec::new(),
761 }
762 }
763
764 pub fn id(&self) -> &acp::SessionId {
765 &self.id
766 }
767
768 /// Returns true if this thread was imported from a shared thread.
769 pub fn is_imported(&self) -> bool {
770 self.imported
771 }
772
773 pub fn replay(
774 &mut self,
775 cx: &mut Context<Self>,
776 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
777 let (tx, rx) = mpsc::unbounded();
778 let stream = ThreadEventStream(tx);
779 for message in &self.messages {
780 match message {
781 Message::User(user_message) => stream.send_user_message(user_message),
782 Message::Agent(assistant_message) => {
783 for content in &assistant_message.content {
784 match content {
785 AgentMessageContent::Text(text) => stream.send_text(text),
786 AgentMessageContent::Thinking { text, .. } => {
787 stream.send_thinking(text)
788 }
789 AgentMessageContent::RedactedThinking(_) => {}
790 AgentMessageContent::ToolUse(tool_use) => {
791 self.replay_tool_call(
792 tool_use,
793 assistant_message.tool_results.get(&tool_use.id),
794 &stream,
795 cx,
796 );
797 }
798 }
799 }
800 }
801 Message::Resume => {}
802 }
803 }
804 rx
805 }
806
807 fn replay_tool_call(
808 &self,
809 tool_use: &LanguageModelToolUse,
810 tool_result: Option<&LanguageModelToolResult>,
811 stream: &ThreadEventStream,
812 cx: &mut Context<Self>,
813 ) {
814 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
815 self.context_server_registry
816 .read(cx)
817 .servers()
818 .find_map(|(_, tools)| {
819 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
820 Some(tool.clone())
821 } else {
822 None
823 }
824 })
825 });
826
827 let Some(tool) = tool else {
828 stream
829 .0
830 .unbounded_send(Ok(ThreadEvent::ToolCall(
831 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
832 .status(acp::ToolCallStatus::Failed)
833 .raw_input(tool_use.input.clone()),
834 )))
835 .ok();
836 return;
837 };
838
839 let title = tool.initial_title(tool_use.input.clone(), cx);
840 let kind = tool.kind();
841 stream.send_tool_call(
842 &tool_use.id,
843 &tool_use.name,
844 title,
845 kind,
846 tool_use.input.clone(),
847 );
848
849 let output = tool_result
850 .as_ref()
851 .and_then(|result| result.output.clone());
852 if let Some(output) = output.clone() {
853 // For replay, we use a dummy cancellation receiver since the tool already completed
854 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
855 let tool_event_stream = ToolCallEventStream::new(
856 tool_use.id.clone(),
857 stream.clone(),
858 Some(self.project.read(cx).fs().clone()),
859 cancellation_rx,
860 );
861 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
862 .log_err();
863 }
864
865 stream.update_tool_call_fields(
866 &tool_use.id,
867 acp::ToolCallUpdateFields::new()
868 .status(
869 tool_result
870 .as_ref()
871 .map_or(acp::ToolCallStatus::Failed, |result| {
872 if result.is_error {
873 acp::ToolCallStatus::Failed
874 } else {
875 acp::ToolCallStatus::Completed
876 }
877 }),
878 )
879 .raw_output(output),
880 );
881 }
882
883 pub fn from_db(
884 id: acp::SessionId,
885 db_thread: DbThread,
886 project: Entity<Project>,
887 project_context: Entity<ProjectContext>,
888 context_server_registry: Entity<ContextServerRegistry>,
889 templates: Arc<Templates>,
890 cx: &mut Context<Self>,
891 ) -> Self {
892 let profile_id = db_thread
893 .profile
894 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
895
896 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
897 db_thread
898 .model
899 .and_then(|model| {
900 let model = SelectedModel {
901 provider: model.provider.clone().into(),
902 model: model.model.into(),
903 };
904 registry.select_model(&model, cx)
905 })
906 .or_else(|| registry.default_model())
907 .map(|model| model.model)
908 });
909
910 if model.is_none() {
911 model = Self::resolve_profile_model(&profile_id, cx);
912 }
913 if model.is_none() {
914 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
915 registry.default_model().map(|model| model.model)
916 });
917 }
918
919 let (prompt_capabilities_tx, prompt_capabilities_rx) =
920 watch::channel(Self::prompt_capabilities(model.as_deref()));
921
922 let action_log = cx.new(|_| ActionLog::new(project.clone()));
923
924 Self {
925 id,
926 prompt_id: PromptId::new(),
927 title: if db_thread.title.is_empty() {
928 None
929 } else {
930 Some(db_thread.title.clone())
931 },
932 pending_title_generation: None,
933 pending_summary_generation: None,
934 summary: db_thread.detailed_summary,
935 messages: db_thread.messages,
936 user_store: project.read(cx).user_store(),
937 completion_mode: db_thread.completion_mode.unwrap_or_default(),
938 running_turn: None,
939 pending_message: None,
940 tools: BTreeMap::default(),
941 tool_use_limit_reached: false,
942 request_token_usage: db_thread.request_token_usage.clone(),
943 cumulative_token_usage: db_thread.cumulative_token_usage,
944 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
945 context_server_registry,
946 profile_id,
947 project_context,
948 templates,
949 model,
950 summarization_model: None,
951 project,
952 action_log,
953 updated_at: db_thread.updated_at,
954 prompt_capabilities_tx,
955 prompt_capabilities_rx,
956 file_read_times: HashMap::default(),
957 imported: db_thread.imported,
958 subagent_context: None,
959 running_subagents: Vec::new(),
960 }
961 }
962
963 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
964 let initial_project_snapshot = self.initial_project_snapshot.clone();
965 let mut thread = DbThread {
966 title: self.title(),
967 messages: self.messages.clone(),
968 updated_at: self.updated_at,
969 detailed_summary: self.summary.clone(),
970 initial_project_snapshot: None,
971 cumulative_token_usage: self.cumulative_token_usage,
972 request_token_usage: self.request_token_usage.clone(),
973 model: self.model.as_ref().map(|model| DbLanguageModel {
974 provider: model.provider_id().to_string(),
975 model: model.name().0.to_string(),
976 }),
977 completion_mode: Some(self.completion_mode),
978 profile: Some(self.profile_id.clone()),
979 imported: self.imported,
980 };
981
982 cx.background_spawn(async move {
983 let initial_project_snapshot = initial_project_snapshot.await;
984 thread.initial_project_snapshot = initial_project_snapshot;
985 thread
986 })
987 }
988
989 /// Create a snapshot of the current project state including git information and unsaved buffers.
990 fn project_snapshot(
991 project: Entity<Project>,
992 cx: &mut Context<Self>,
993 ) -> Task<Arc<ProjectSnapshot>> {
994 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
995 cx.spawn(async move |_, _| {
996 let snapshot = task.await;
997
998 Arc::new(ProjectSnapshot {
999 worktree_snapshots: snapshot.worktree_snapshots,
1000 timestamp: Utc::now(),
1001 })
1002 })
1003 }
1004
1005 pub fn project_context(&self) -> &Entity<ProjectContext> {
1006 &self.project_context
1007 }
1008
1009 pub fn project(&self) -> &Entity<Project> {
1010 &self.project
1011 }
1012
1013 pub fn action_log(&self) -> &Entity<ActionLog> {
1014 &self.action_log
1015 }
1016
1017 pub fn is_empty(&self) -> bool {
1018 self.messages.is_empty() && self.title.is_none()
1019 }
1020
1021 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
1022 self.model.as_ref()
1023 }
1024
1025 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
1026 let old_usage = self.latest_token_usage();
1027 self.model = Some(model);
1028 let new_caps = Self::prompt_capabilities(self.model.as_deref());
1029 let new_usage = self.latest_token_usage();
1030 if old_usage != new_usage {
1031 cx.emit(TokenUsageUpdated(new_usage));
1032 }
1033 self.prompt_capabilities_tx.send(new_caps).log_err();
1034 cx.notify()
1035 }
1036
1037 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1038 self.summarization_model.as_ref()
1039 }
1040
1041 pub fn set_summarization_model(
1042 &mut self,
1043 model: Option<Arc<dyn LanguageModel>>,
1044 cx: &mut Context<Self>,
1045 ) {
1046 self.summarization_model = model;
1047 cx.notify()
1048 }
1049
1050 pub fn completion_mode(&self) -> CompletionMode {
1051 self.completion_mode
1052 }
1053
1054 pub fn set_completion_mode(&mut self, mode: CompletionMode, cx: &mut Context<Self>) {
1055 let old_usage = self.latest_token_usage();
1056 self.completion_mode = mode;
1057 let new_usage = self.latest_token_usage();
1058 if old_usage != new_usage {
1059 cx.emit(TokenUsageUpdated(new_usage));
1060 }
1061 cx.notify()
1062 }
1063
1064 pub fn last_message(&self) -> Option<Message> {
1065 if let Some(message) = self.pending_message.clone() {
1066 Some(Message::Agent(message))
1067 } else {
1068 self.messages.last().cloned()
1069 }
1070 }
1071
1072 pub fn add_default_tools(
1073 &mut self,
1074 environment: Rc<dyn ThreadEnvironment>,
1075 cx: &mut Context<Self>,
1076 ) {
1077 let language_registry = self.project.read(cx).languages().clone();
1078 self.add_tool(CopyPathTool::new(self.project.clone()));
1079 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1080 self.add_tool(DeletePathTool::new(
1081 self.project.clone(),
1082 self.action_log.clone(),
1083 ));
1084 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1085 self.add_tool(EditFileTool::new(
1086 self.project.clone(),
1087 cx.weak_entity(),
1088 language_registry,
1089 Templates::new(),
1090 ));
1091 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1092 self.add_tool(FindPathTool::new(self.project.clone()));
1093 self.add_tool(GrepTool::new(self.project.clone()));
1094 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1095 self.add_tool(MovePathTool::new(self.project.clone()));
1096 self.add_tool(NowTool);
1097 self.add_tool(OpenTool::new(self.project.clone()));
1098 self.add_tool(ReadFileTool::new(
1099 cx.weak_entity(),
1100 self.project.clone(),
1101 self.action_log.clone(),
1102 ));
1103 self.add_tool(SaveFileTool::new(self.project.clone()));
1104 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1105 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1106 self.add_tool(ThinkingTool);
1107 self.add_tool(WebSearchTool);
1108
1109 if cx.has_flag::<SubagentsFeatureFlag>() && self.depth() < MAX_SUBAGENT_DEPTH {
1110 let tool_names = self.registered_tool_names();
1111 self.add_tool(SubagentTool::new(
1112 cx.weak_entity(),
1113 self.project.clone(),
1114 self.project_context.clone(),
1115 self.context_server_registry.clone(),
1116 self.templates.clone(),
1117 self.depth(),
1118 tool_names,
1119 ));
1120 }
1121 }
1122
1123 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1124 self.tools.insert(T::name().into(), tool.erase());
1125 }
1126
1127 pub fn remove_tool(&mut self, name: &str) -> bool {
1128 self.tools.remove(name).is_some()
1129 }
1130
1131 pub fn restrict_tools(&mut self, allowed: &collections::HashSet<SharedString>) {
1132 self.tools.retain(|name, _| allowed.contains(name));
1133 }
1134
1135 pub fn profile(&self) -> &AgentProfileId {
1136 &self.profile_id
1137 }
1138
1139 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1140 if self.profile_id == profile_id {
1141 return;
1142 }
1143
1144 self.profile_id = profile_id;
1145
1146 // Swap to the profile's preferred model when available.
1147 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1148 self.set_model(model, cx);
1149 }
1150 }
1151
1152 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1153 for subagent in self.running_subagents.drain(..) {
1154 if let Some(subagent) = subagent.upgrade() {
1155 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1156 }
1157 }
1158
1159 let Some(running_turn) = self.running_turn.take() else {
1160 self.flush_pending_message(cx);
1161 return Task::ready(());
1162 };
1163
1164 let turn_task = running_turn.cancel();
1165
1166 cx.spawn(async move |this, cx| {
1167 turn_task.await;
1168 this.update(cx, |this, cx| {
1169 this.flush_pending_message(cx);
1170 })
1171 .ok();
1172 })
1173 }
1174
1175 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1176 let Some(last_user_message) = self.last_user_message() else {
1177 return;
1178 };
1179
1180 self.request_token_usage
1181 .insert(last_user_message.id.clone(), update);
1182 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1183 cx.notify();
1184 }
1185
1186 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1187 self.cancel(cx).detach();
1188 // Clear pending message since cancel will try to flush it asynchronously,
1189 // and we don't want that content to be added after we truncate
1190 self.pending_message.take();
1191 let Some(position) = self.messages.iter().position(
1192 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1193 ) else {
1194 return Err(anyhow!("Message not found"));
1195 };
1196
1197 for message in self.messages.drain(position..) {
1198 match message {
1199 Message::User(message) => {
1200 self.request_token_usage.remove(&message.id);
1201 }
1202 Message::Agent(_) | Message::Resume => {}
1203 }
1204 }
1205 self.clear_summary();
1206 cx.notify();
1207 Ok(())
1208 }
1209
1210 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1211 let last_user_message = self.last_user_message()?;
1212 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1213 Some(*tokens)
1214 }
1215
1216 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1217 let usage = self.latest_request_token_usage()?;
1218 let model = self.model.clone()?;
1219 Some(acp_thread::TokenUsage {
1220 max_tokens: model.max_token_count_for_mode(self.completion_mode.into()),
1221 used_tokens: usage.total_tokens(),
1222 output_tokens: usage.output_tokens,
1223 })
1224 }
1225
1226 /// Get the total input token count as of the message before the given message.
1227 ///
1228 /// Returns `None` if:
1229 /// - `target_id` is the first message (no previous message)
1230 /// - The previous message hasn't received a response yet (no usage data)
1231 /// - `target_id` is not found in the messages
1232 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1233 let mut previous_user_message_id: Option<&UserMessageId> = None;
1234
1235 for message in &self.messages {
1236 if let Message::User(user_msg) = message {
1237 if &user_msg.id == target_id {
1238 let prev_id = previous_user_message_id?;
1239 let usage = self.request_token_usage.get(prev_id)?;
1240 return Some(usage.input_tokens);
1241 }
1242 previous_user_message_id = Some(&user_msg.id);
1243 }
1244 }
1245 None
1246 }
1247
1248 /// Look up the active profile and resolve its preferred model if one is configured.
1249 fn resolve_profile_model(
1250 profile_id: &AgentProfileId,
1251 cx: &mut Context<Self>,
1252 ) -> Option<Arc<dyn LanguageModel>> {
1253 let selection = AgentSettings::get_global(cx)
1254 .profiles
1255 .get(profile_id)?
1256 .default_model
1257 .clone()?;
1258 Self::resolve_model_from_selection(&selection, cx)
1259 }
1260
1261 /// Translate a stored model selection into the configured model from the registry.
1262 fn resolve_model_from_selection(
1263 selection: &LanguageModelSelection,
1264 cx: &mut Context<Self>,
1265 ) -> Option<Arc<dyn LanguageModel>> {
1266 let selected = SelectedModel {
1267 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1268 model: LanguageModelId::from(selection.model.clone()),
1269 };
1270 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1271 registry
1272 .select_model(&selected, cx)
1273 .map(|configured| configured.model)
1274 })
1275 }
1276
1277 pub fn resume(
1278 &mut self,
1279 cx: &mut Context<Self>,
1280 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1281 self.messages.push(Message::Resume);
1282 cx.notify();
1283
1284 log::debug!("Total messages in thread: {}", self.messages.len());
1285 self.run_turn(cx)
1286 }
1287
1288 /// Sending a message results in the model streaming a response, which could include tool calls.
1289 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1290 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1291 pub fn send<T>(
1292 &mut self,
1293 id: UserMessageId,
1294 content: impl IntoIterator<Item = T>,
1295 cx: &mut Context<Self>,
1296 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1297 where
1298 T: Into<UserMessageContent>,
1299 {
1300 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1301 log::debug!("Thread::send content: {:?}", content);
1302
1303 self.messages
1304 .push(Message::User(UserMessage { id, content }));
1305 cx.notify();
1306
1307 self.send_existing(cx)
1308 }
1309
1310 pub fn send_existing(
1311 &mut self,
1312 cx: &mut Context<Self>,
1313 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1314 let model = self.model().context("No language model configured")?;
1315
1316 log::info!("Thread::send called with model: {}", model.name().0);
1317 self.advance_prompt_id();
1318
1319 log::debug!("Total messages in thread: {}", self.messages.len());
1320 self.run_turn(cx)
1321 }
1322
1323 pub fn push_acp_user_block(
1324 &mut self,
1325 id: UserMessageId,
1326 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1327 path_style: PathStyle,
1328 cx: &mut Context<Self>,
1329 ) {
1330 let content = blocks
1331 .into_iter()
1332 .map(|block| UserMessageContent::from_content_block(block, path_style))
1333 .collect::<Vec<_>>();
1334 self.messages
1335 .push(Message::User(UserMessage { id, content }));
1336 cx.notify();
1337 }
1338
1339 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1340 let text = match block {
1341 acp::ContentBlock::Text(text_content) => text_content.text,
1342 acp::ContentBlock::Image(_) => "[image]".to_string(),
1343 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1344 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1345 acp::ContentBlock::Resource(resource) => match resource.resource {
1346 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1347 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1348 _ => "[resource]".to_string(),
1349 },
1350 _ => "[unknown]".to_string(),
1351 };
1352
1353 self.messages.push(Message::Agent(AgentMessage {
1354 content: vec![AgentMessageContent::Text(text)],
1355 ..Default::default()
1356 }));
1357 cx.notify();
1358 }
1359
1360 #[cfg(feature = "eval")]
1361 pub fn proceed(
1362 &mut self,
1363 cx: &mut Context<Self>,
1364 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1365 self.run_turn(cx)
1366 }
1367
1368 fn run_turn(
1369 &mut self,
1370 cx: &mut Context<Self>,
1371 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1372 // Flush the old pending message synchronously before cancelling,
1373 // to avoid a race where the detached cancel task might flush the NEW
1374 // turn's pending message instead of the old one.
1375 self.flush_pending_message(cx);
1376 self.cancel(cx).detach();
1377
1378 let model = self.model.clone().context("No language model configured")?;
1379 let profile = AgentSettings::get_global(cx)
1380 .profiles
1381 .get(&self.profile_id)
1382 .context("Profile not found")?;
1383 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1384 let event_stream = ThreadEventStream(events_tx);
1385 let message_ix = self.messages.len().saturating_sub(1);
1386 self.tool_use_limit_reached = false;
1387 self.clear_summary();
1388 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1389 self.running_turn = Some(RunningTurn {
1390 event_stream: event_stream.clone(),
1391 tools: self.enabled_tools(profile, &model, cx),
1392 cancellation_tx,
1393 _task: cx.spawn(async move |this, cx| {
1394 log::debug!("Starting agent turn execution");
1395
1396 let turn_result = Self::run_turn_internal(
1397 &this,
1398 model,
1399 &event_stream,
1400 cancellation_rx.clone(),
1401 cx,
1402 )
1403 .await;
1404
1405 // Check if we were cancelled - if so, cancel() already took running_turn
1406 // and we shouldn't touch it (it might be a NEW turn now)
1407 let was_cancelled = *cancellation_rx.borrow();
1408 if was_cancelled {
1409 log::debug!("Turn was cancelled, skipping cleanup");
1410 return;
1411 }
1412
1413 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1414
1415 match turn_result {
1416 Ok(()) => {
1417 log::debug!("Turn execution completed");
1418 event_stream.send_stop(acp::StopReason::EndTurn);
1419 }
1420 Err(error) => {
1421 log::error!("Turn execution failed: {:?}", error);
1422 match error.downcast::<CompletionError>() {
1423 Ok(CompletionError::Refusal) => {
1424 event_stream.send_stop(acp::StopReason::Refusal);
1425 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1426 }
1427 Ok(CompletionError::MaxTokens) => {
1428 event_stream.send_stop(acp::StopReason::MaxTokens);
1429 }
1430 Ok(CompletionError::Other(error)) | Err(error) => {
1431 event_stream.send_error(error);
1432 }
1433 }
1434 }
1435 }
1436
1437 _ = this.update(cx, |this, _| this.running_turn.take());
1438 }),
1439 });
1440 Ok(events_rx)
1441 }
1442
1443 async fn run_turn_internal(
1444 this: &WeakEntity<Self>,
1445 model: Arc<dyn LanguageModel>,
1446 event_stream: &ThreadEventStream,
1447 mut cancellation_rx: watch::Receiver<bool>,
1448 cx: &mut AsyncApp,
1449 ) -> Result<()> {
1450 let mut attempt = 0;
1451 let mut intent = CompletionIntent::UserPrompt;
1452 loop {
1453 let request =
1454 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1455
1456 telemetry::event!(
1457 "Agent Thread Completion",
1458 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1459 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1460 model = model.telemetry_id(),
1461 model_provider = model.provider_id().to_string(),
1462 attempt
1463 );
1464
1465 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1466
1467 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1468 Ok(events) => (events, None),
1469 Err(err) => (stream::empty().boxed(), Some(err)),
1470 };
1471 let mut tool_results = FuturesUnordered::new();
1472 let mut cancelled = false;
1473 loop {
1474 // Race between getting the next event and cancellation
1475 let event = futures::select! {
1476 event = events.next().fuse() => event,
1477 _ = cancellation_rx.changed().fuse() => {
1478 if *cancellation_rx.borrow() {
1479 cancelled = true;
1480 break;
1481 }
1482 continue;
1483 }
1484 };
1485 let Some(event) = event else {
1486 break;
1487 };
1488 log::trace!("Received completion event: {:?}", event);
1489 match event {
1490 Ok(event) => {
1491 tool_results.extend(this.update(cx, |this, cx| {
1492 this.handle_completion_event(
1493 event,
1494 event_stream,
1495 cancellation_rx.clone(),
1496 cx,
1497 )
1498 })??);
1499 }
1500 Err(err) => {
1501 error = Some(err);
1502 break;
1503 }
1504 }
1505 }
1506
1507 let end_turn = tool_results.is_empty();
1508 while let Some(tool_result) = tool_results.next().await {
1509 log::debug!("Tool finished {:?}", tool_result);
1510
1511 event_stream.update_tool_call_fields(
1512 &tool_result.tool_use_id,
1513 acp::ToolCallUpdateFields::new()
1514 .status(if tool_result.is_error {
1515 acp::ToolCallStatus::Failed
1516 } else {
1517 acp::ToolCallStatus::Completed
1518 })
1519 .raw_output(tool_result.output.clone()),
1520 );
1521 this.update(cx, |this, _cx| {
1522 this.pending_message()
1523 .tool_results
1524 .insert(tool_result.tool_use_id.clone(), tool_result);
1525 })?;
1526 }
1527
1528 this.update(cx, |this, cx| {
1529 this.flush_pending_message(cx);
1530 if this.title.is_none() && this.pending_title_generation.is_none() {
1531 this.generate_title(cx);
1532 }
1533 })?;
1534
1535 if cancelled {
1536 log::debug!("Turn cancelled by user, exiting");
1537 return Ok(());
1538 }
1539
1540 if let Some(error) = error {
1541 attempt += 1;
1542 let retry = this.update(cx, |this, cx| {
1543 let user_store = this.user_store.read(cx);
1544 this.handle_completion_error(error, attempt, user_store.plan())
1545 })??;
1546 let timer = cx.background_executor().timer(retry.duration);
1547 event_stream.send_retry(retry);
1548 timer.await;
1549 this.update(cx, |this, _cx| {
1550 if let Some(Message::Agent(message)) = this.messages.last() {
1551 if message.tool_results.is_empty() {
1552 intent = CompletionIntent::UserPrompt;
1553 this.messages.push(Message::Resume);
1554 }
1555 }
1556 })?;
1557 } else if this.read_with(cx, |this, _| this.tool_use_limit_reached)? {
1558 return Err(language_model::ToolUseLimitReachedError.into());
1559 } else if end_turn {
1560 return Ok(());
1561 } else {
1562 intent = CompletionIntent::ToolResults;
1563 attempt = 0;
1564 }
1565 }
1566 }
1567
1568 fn handle_completion_error(
1569 &mut self,
1570 error: LanguageModelCompletionError,
1571 attempt: u8,
1572 plan: Option<Plan>,
1573 ) -> Result<acp_thread::RetryStatus> {
1574 let Some(model) = self.model.as_ref() else {
1575 return Err(anyhow!(error));
1576 };
1577
1578 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1579 match plan {
1580 Some(Plan::V2(_)) => true,
1581 Some(Plan::V1(_)) => self.completion_mode == CompletionMode::Burn,
1582 None => false,
1583 }
1584 } else {
1585 true
1586 };
1587
1588 if !auto_retry {
1589 return Err(anyhow!(error));
1590 }
1591
1592 let Some(strategy) = Self::retry_strategy_for(&error) else {
1593 return Err(anyhow!(error));
1594 };
1595
1596 let max_attempts = match &strategy {
1597 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1598 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1599 };
1600
1601 if attempt > max_attempts {
1602 return Err(anyhow!(error));
1603 }
1604
1605 let delay = match &strategy {
1606 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1607 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1608 Duration::from_secs(delay_secs)
1609 }
1610 RetryStrategy::Fixed { delay, .. } => *delay,
1611 };
1612 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1613
1614 Ok(acp_thread::RetryStatus {
1615 last_error: error.to_string().into(),
1616 attempt: attempt as usize,
1617 max_attempts: max_attempts as usize,
1618 started_at: Instant::now(),
1619 duration: delay,
1620 })
1621 }
1622
1623 /// A helper method that's called on every streamed completion event.
1624 /// Returns an optional tool result task, which the main agentic loop will
1625 /// send back to the model when it resolves.
1626 fn handle_completion_event(
1627 &mut self,
1628 event: LanguageModelCompletionEvent,
1629 event_stream: &ThreadEventStream,
1630 cancellation_rx: watch::Receiver<bool>,
1631 cx: &mut Context<Self>,
1632 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1633 log::trace!("Handling streamed completion event: {:?}", event);
1634 use LanguageModelCompletionEvent::*;
1635
1636 match event {
1637 StartMessage { .. } => {
1638 self.flush_pending_message(cx);
1639 self.pending_message = Some(AgentMessage::default());
1640 }
1641 Text(new_text) => self.handle_text_event(new_text, event_stream, cx),
1642 Thinking { text, signature } => {
1643 self.handle_thinking_event(text, signature, event_stream, cx)
1644 }
1645 RedactedThinking { data } => self.handle_redacted_thinking_event(data, cx),
1646 ReasoningDetails(details) => {
1647 let last_message = self.pending_message();
1648 // Store the last non-empty reasoning_details (overwrites earlier ones)
1649 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
1650 if let serde_json::Value::Array(ref arr) = details {
1651 if !arr.is_empty() {
1652 last_message.reasoning_details = Some(details);
1653 }
1654 } else {
1655 last_message.reasoning_details = Some(details);
1656 }
1657 }
1658 ToolUse(tool_use) => {
1659 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
1660 }
1661 ToolUseJsonParseError {
1662 id,
1663 tool_name,
1664 raw_input,
1665 json_parse_error,
1666 } => {
1667 return Ok(Some(Task::ready(
1668 self.handle_tool_use_json_parse_error_event(
1669 id,
1670 tool_name,
1671 raw_input,
1672 json_parse_error,
1673 ),
1674 )));
1675 }
1676 UsageUpdate(usage) => {
1677 telemetry::event!(
1678 "Agent Thread Completion Usage Updated",
1679 thread_id = self.id.to_string(),
1680 prompt_id = self.prompt_id.to_string(),
1681 model = self.model.as_ref().map(|m| m.telemetry_id()),
1682 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1683 input_tokens = usage.input_tokens,
1684 output_tokens = usage.output_tokens,
1685 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1686 cache_read_input_tokens = usage.cache_read_input_tokens,
1687 );
1688 self.update_token_usage(usage, cx);
1689 }
1690 UsageUpdated { amount, limit } => {
1691 self.update_model_request_usage(amount, limit, cx);
1692 }
1693 ToolUseLimitReached => {
1694 self.tool_use_limit_reached = true;
1695 }
1696 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1697 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1698 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1699 Started | Queued { .. } => {}
1700 }
1701
1702 Ok(None)
1703 }
1704
1705 fn handle_text_event(
1706 &mut self,
1707 new_text: String,
1708 event_stream: &ThreadEventStream,
1709 cx: &mut Context<Self>,
1710 ) {
1711 event_stream.send_text(&new_text);
1712
1713 let last_message = self.pending_message();
1714 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1715 text.push_str(&new_text);
1716 } else {
1717 last_message
1718 .content
1719 .push(AgentMessageContent::Text(new_text));
1720 }
1721
1722 cx.notify();
1723 }
1724
1725 fn handle_thinking_event(
1726 &mut self,
1727 new_text: String,
1728 new_signature: Option<String>,
1729 event_stream: &ThreadEventStream,
1730 cx: &mut Context<Self>,
1731 ) {
1732 event_stream.send_thinking(&new_text);
1733
1734 let last_message = self.pending_message();
1735 if let Some(AgentMessageContent::Thinking { text, signature }) =
1736 last_message.content.last_mut()
1737 {
1738 text.push_str(&new_text);
1739 *signature = new_signature.or(signature.take());
1740 } else {
1741 last_message.content.push(AgentMessageContent::Thinking {
1742 text: new_text,
1743 signature: new_signature,
1744 });
1745 }
1746
1747 cx.notify();
1748 }
1749
1750 fn handle_redacted_thinking_event(&mut self, data: String, cx: &mut Context<Self>) {
1751 let last_message = self.pending_message();
1752 last_message
1753 .content
1754 .push(AgentMessageContent::RedactedThinking(data));
1755 cx.notify();
1756 }
1757
1758 fn handle_tool_use_event(
1759 &mut self,
1760 tool_use: LanguageModelToolUse,
1761 event_stream: &ThreadEventStream,
1762 cancellation_rx: watch::Receiver<bool>,
1763 cx: &mut Context<Self>,
1764 ) -> Option<Task<LanguageModelToolResult>> {
1765 cx.notify();
1766
1767 let tool = self.tool(tool_use.name.as_ref());
1768 let mut title = SharedString::from(&tool_use.name);
1769 let mut kind = acp::ToolKind::Other;
1770 if let Some(tool) = tool.as_ref() {
1771 title = tool.initial_title(tool_use.input.clone(), cx);
1772 kind = tool.kind();
1773 }
1774
1775 // Ensure the last message ends in the current tool use
1776 let last_message = self.pending_message();
1777 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1778 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1779 if last_tool_use.id == tool_use.id {
1780 *last_tool_use = tool_use.clone();
1781 false
1782 } else {
1783 true
1784 }
1785 } else {
1786 true
1787 }
1788 });
1789
1790 if push_new_tool_use {
1791 event_stream.send_tool_call(
1792 &tool_use.id,
1793 &tool_use.name,
1794 title,
1795 kind,
1796 tool_use.input.clone(),
1797 );
1798 last_message
1799 .content
1800 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1801 } else {
1802 event_stream.update_tool_call_fields(
1803 &tool_use.id,
1804 acp::ToolCallUpdateFields::new()
1805 .title(title.as_str())
1806 .kind(kind)
1807 .raw_input(tool_use.input.clone()),
1808 );
1809 }
1810
1811 if !tool_use.is_input_complete {
1812 return None;
1813 }
1814
1815 let Some(tool) = tool else {
1816 let content = format!("No tool named {} exists", tool_use.name);
1817 return Some(Task::ready(LanguageModelToolResult {
1818 content: LanguageModelToolResultContent::Text(Arc::from(content)),
1819 tool_use_id: tool_use.id,
1820 tool_name: tool_use.name,
1821 is_error: true,
1822 output: None,
1823 }));
1824 };
1825
1826 let fs = self.project.read(cx).fs().clone();
1827 let tool_event_stream = ToolCallEventStream::new(
1828 tool_use.id.clone(),
1829 event_stream.clone(),
1830 Some(fs),
1831 cancellation_rx,
1832 );
1833 tool_event_stream.update_fields(
1834 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
1835 );
1836 let supports_images = self.model().is_some_and(|model| model.supports_images());
1837 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
1838 log::debug!("Running tool {}", tool_use.name);
1839 Some(cx.foreground_executor().spawn(async move {
1840 let tool_result = tool_result.await.and_then(|output| {
1841 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
1842 && !supports_images
1843 {
1844 return Err(anyhow!(
1845 "Attempted to read an image, but this model doesn't support it.",
1846 ));
1847 }
1848 Ok(output)
1849 });
1850
1851 match tool_result {
1852 Ok(output) => LanguageModelToolResult {
1853 tool_use_id: tool_use.id,
1854 tool_name: tool_use.name,
1855 is_error: false,
1856 content: output.llm_output,
1857 output: Some(output.raw_output),
1858 },
1859 Err(error) => LanguageModelToolResult {
1860 tool_use_id: tool_use.id,
1861 tool_name: tool_use.name,
1862 is_error: true,
1863 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
1864 output: Some(error.to_string().into()),
1865 },
1866 }
1867 }))
1868 }
1869
1870 fn handle_tool_use_json_parse_error_event(
1871 &mut self,
1872 tool_use_id: LanguageModelToolUseId,
1873 tool_name: Arc<str>,
1874 raw_input: Arc<str>,
1875 json_parse_error: String,
1876 ) -> LanguageModelToolResult {
1877 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
1878 LanguageModelToolResult {
1879 tool_use_id,
1880 tool_name,
1881 is_error: true,
1882 content: LanguageModelToolResultContent::Text(tool_output.into()),
1883 output: Some(serde_json::Value::String(raw_input.to_string())),
1884 }
1885 }
1886
1887 fn update_model_request_usage(&self, amount: usize, limit: UsageLimit, cx: &mut Context<Self>) {
1888 self.project
1889 .read(cx)
1890 .user_store()
1891 .update(cx, |user_store, cx| {
1892 user_store.update_model_request_usage(
1893 ModelRequestUsage(RequestUsage {
1894 amount: amount as i32,
1895 limit,
1896 }),
1897 cx,
1898 )
1899 });
1900 }
1901
1902 pub fn title(&self) -> SharedString {
1903 self.title.clone().unwrap_or("New Thread".into())
1904 }
1905
1906 pub fn is_generating_summary(&self) -> bool {
1907 self.pending_summary_generation.is_some()
1908 }
1909
1910 pub fn is_generating_title(&self) -> bool {
1911 self.pending_title_generation.is_some()
1912 }
1913
1914 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
1915 if let Some(summary) = self.summary.as_ref() {
1916 return Task::ready(Some(summary.clone())).shared();
1917 }
1918 if let Some(task) = self.pending_summary_generation.clone() {
1919 return task;
1920 }
1921 let Some(model) = self.summarization_model.clone() else {
1922 log::error!("No summarization model available");
1923 return Task::ready(None).shared();
1924 };
1925 let mut request = LanguageModelRequest {
1926 intent: Some(CompletionIntent::ThreadContextSummarization),
1927 temperature: AgentSettings::temperature_for_model(&model, cx),
1928 ..Default::default()
1929 };
1930
1931 for message in &self.messages {
1932 request.messages.extend(message.to_request());
1933 }
1934
1935 request.messages.push(LanguageModelRequestMessage {
1936 role: Role::User,
1937 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
1938 cache: false,
1939 reasoning_details: None,
1940 });
1941
1942 let task = cx
1943 .spawn(async move |this, cx| {
1944 let mut summary = String::new();
1945 let mut messages = model.stream_completion(request, cx).await.log_err()?;
1946 while let Some(event) = messages.next().await {
1947 let event = event.log_err()?;
1948 let text = match event {
1949 LanguageModelCompletionEvent::Text(text) => text,
1950 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
1951 this.update(cx, |thread, cx| {
1952 thread.update_model_request_usage(amount, limit, cx);
1953 })
1954 .ok()?;
1955 continue;
1956 }
1957 _ => continue,
1958 };
1959
1960 let mut lines = text.lines();
1961 summary.extend(lines.next());
1962 }
1963
1964 log::debug!("Setting summary: {}", summary);
1965 let summary = SharedString::from(summary);
1966
1967 this.update(cx, |this, cx| {
1968 this.summary = Some(summary.clone());
1969 this.pending_summary_generation = None;
1970 cx.notify()
1971 })
1972 .ok()?;
1973
1974 Some(summary)
1975 })
1976 .shared();
1977 self.pending_summary_generation = Some(task.clone());
1978 task
1979 }
1980
1981 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
1982 let Some(model) = self.summarization_model.clone() else {
1983 return;
1984 };
1985
1986 log::debug!(
1987 "Generating title with model: {:?}",
1988 self.summarization_model.as_ref().map(|model| model.name())
1989 );
1990 let mut request = LanguageModelRequest {
1991 intent: Some(CompletionIntent::ThreadSummarization),
1992 temperature: AgentSettings::temperature_for_model(&model, cx),
1993 ..Default::default()
1994 };
1995
1996 for message in &self.messages {
1997 request.messages.extend(message.to_request());
1998 }
1999
2000 request.messages.push(LanguageModelRequestMessage {
2001 role: Role::User,
2002 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2003 cache: false,
2004 reasoning_details: None,
2005 });
2006 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2007 let mut title = String::new();
2008
2009 let generate = async {
2010 let mut messages = model.stream_completion(request, cx).await?;
2011 while let Some(event) = messages.next().await {
2012 let event = event?;
2013 let text = match event {
2014 LanguageModelCompletionEvent::Text(text) => text,
2015 LanguageModelCompletionEvent::UsageUpdated { amount, limit } => {
2016 this.update(cx, |thread, cx| {
2017 thread.update_model_request_usage(amount, limit, cx);
2018 })?;
2019 continue;
2020 }
2021 _ => continue,
2022 };
2023
2024 let mut lines = text.lines();
2025 title.extend(lines.next());
2026
2027 // Stop if the LLM generated multiple lines.
2028 if lines.next().is_some() {
2029 break;
2030 }
2031 }
2032 anyhow::Ok(())
2033 };
2034
2035 if generate.await.context("failed to generate title").is_ok() {
2036 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
2037 }
2038 _ = this.update(cx, |this, _| this.pending_title_generation = None);
2039 }));
2040 }
2041
2042 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2043 self.pending_title_generation = None;
2044 if Some(&title) != self.title.as_ref() {
2045 self.title = Some(title);
2046 cx.emit(TitleUpdated);
2047 cx.notify();
2048 }
2049 }
2050
2051 fn clear_summary(&mut self) {
2052 self.summary = None;
2053 self.pending_summary_generation = None;
2054 }
2055
2056 fn last_user_message(&self) -> Option<&UserMessage> {
2057 self.messages
2058 .iter()
2059 .rev()
2060 .find_map(|message| match message {
2061 Message::User(user_message) => Some(user_message),
2062 Message::Agent(_) => None,
2063 Message::Resume => None,
2064 })
2065 }
2066
2067 fn pending_message(&mut self) -> &mut AgentMessage {
2068 self.pending_message.get_or_insert_default()
2069 }
2070
2071 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2072 let Some(mut message) = self.pending_message.take() else {
2073 return;
2074 };
2075
2076 if message.content.is_empty() {
2077 return;
2078 }
2079
2080 for content in &message.content {
2081 let AgentMessageContent::ToolUse(tool_use) = content else {
2082 continue;
2083 };
2084
2085 if !message.tool_results.contains_key(&tool_use.id) {
2086 message.tool_results.insert(
2087 tool_use.id.clone(),
2088 LanguageModelToolResult {
2089 tool_use_id: tool_use.id.clone(),
2090 tool_name: tool_use.name.clone(),
2091 is_error: true,
2092 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2093 output: None,
2094 },
2095 );
2096 }
2097 }
2098
2099 self.messages.push(Message::Agent(message));
2100 self.updated_at = Utc::now();
2101 self.clear_summary();
2102 cx.notify()
2103 }
2104
2105 pub(crate) fn build_completion_request(
2106 &self,
2107 completion_intent: CompletionIntent,
2108 cx: &App,
2109 ) -> Result<LanguageModelRequest> {
2110 let model = self.model().context("No language model configured")?;
2111 let tools = if let Some(turn) = self.running_turn.as_ref() {
2112 turn.tools
2113 .iter()
2114 .filter_map(|(tool_name, tool)| {
2115 log::trace!("Including tool: {}", tool_name);
2116 Some(LanguageModelRequestTool {
2117 name: tool_name.to_string(),
2118 description: tool.description().to_string(),
2119 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2120 })
2121 })
2122 .collect::<Vec<_>>()
2123 } else {
2124 Vec::new()
2125 };
2126
2127 log::debug!("Building completion request");
2128 log::debug!("Completion intent: {:?}", completion_intent);
2129 log::debug!("Completion mode: {:?}", self.completion_mode);
2130
2131 let available_tools: Vec<_> = self
2132 .running_turn
2133 .as_ref()
2134 .map(|turn| turn.tools.keys().cloned().collect())
2135 .unwrap_or_default();
2136
2137 log::debug!("Request includes {} tools", available_tools.len());
2138 let messages = self.build_request_messages(available_tools, cx);
2139 log::debug!("Request will include {} messages", messages.len());
2140
2141 let request = LanguageModelRequest {
2142 thread_id: Some(self.id.to_string()),
2143 prompt_id: Some(self.prompt_id.to_string()),
2144 intent: Some(completion_intent),
2145 mode: Some(self.completion_mode.into()),
2146 messages,
2147 tools,
2148 tool_choice: None,
2149 stop: Vec::new(),
2150 temperature: AgentSettings::temperature_for_model(model, cx),
2151 thinking_allowed: true,
2152 };
2153
2154 log::debug!("Completion request built successfully");
2155 Ok(request)
2156 }
2157
2158 fn enabled_tools(
2159 &self,
2160 profile: &AgentProfileSettings,
2161 model: &Arc<dyn LanguageModel>,
2162 cx: &App,
2163 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2164 fn truncate(tool_name: &SharedString) -> SharedString {
2165 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2166 let mut truncated = tool_name.to_string();
2167 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2168 truncated.into()
2169 } else {
2170 tool_name.clone()
2171 }
2172 }
2173
2174 let mut tools = self
2175 .tools
2176 .iter()
2177 .filter_map(|(tool_name, tool)| {
2178 if tool.supports_provider(&model.provider_id())
2179 && profile.is_tool_enabled(tool_name)
2180 {
2181 Some((truncate(tool_name), tool.clone()))
2182 } else {
2183 None
2184 }
2185 })
2186 .collect::<BTreeMap<_, _>>();
2187
2188 let mut context_server_tools = Vec::new();
2189 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2190 let mut duplicate_tool_names = HashSet::default();
2191 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2192 for (tool_name, tool) in server_tools {
2193 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2194 let tool_name = truncate(tool_name);
2195 if !seen_tools.insert(tool_name.clone()) {
2196 duplicate_tool_names.insert(tool_name.clone());
2197 }
2198 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2199 }
2200 }
2201 }
2202
2203 // When there are duplicate tool names, disambiguate by prefixing them
2204 // with the server ID. In the rare case there isn't enough space for the
2205 // disambiguated tool name, keep only the last tool with this name.
2206 for (server_id, tool_name, tool) in context_server_tools {
2207 if duplicate_tool_names.contains(&tool_name) {
2208 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2209 if available >= 2 {
2210 let mut disambiguated = server_id.0.to_string();
2211 disambiguated.truncate(available - 1);
2212 disambiguated.push('_');
2213 disambiguated.push_str(&tool_name);
2214 tools.insert(disambiguated.into(), tool.clone());
2215 } else {
2216 tools.insert(tool_name, tool.clone());
2217 }
2218 } else {
2219 tools.insert(tool_name, tool.clone());
2220 }
2221 }
2222
2223 tools
2224 }
2225
2226 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2227 self.running_turn.as_ref()?.tools.get(name).cloned()
2228 }
2229
2230 pub fn has_tool(&self, name: &str) -> bool {
2231 self.running_turn
2232 .as_ref()
2233 .is_some_and(|turn| turn.tools.contains_key(name))
2234 }
2235
2236 #[cfg(any(test, feature = "test-support"))]
2237 pub fn has_registered_tool(&self, name: &str) -> bool {
2238 self.tools.contains_key(name)
2239 }
2240
2241 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2242 self.tools.keys().cloned().collect()
2243 }
2244
2245 pub fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2246 self.running_subagents.push(subagent);
2247 }
2248
2249 pub fn unregister_running_subagent(&mut self, subagent: &WeakEntity<Thread>) {
2250 self.running_subagents
2251 .retain(|s| s.entity_id() != subagent.entity_id());
2252 }
2253
2254 pub fn running_subagent_count(&self) -> usize {
2255 self.running_subagents
2256 .iter()
2257 .filter(|s| s.upgrade().is_some())
2258 .count()
2259 }
2260
2261 pub fn is_subagent(&self) -> bool {
2262 self.subagent_context.is_some()
2263 }
2264
2265 pub fn depth(&self) -> u8 {
2266 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2267 }
2268
2269 pub fn is_turn_complete(&self) -> bool {
2270 self.running_turn.is_none()
2271 }
2272
2273 pub fn submit_user_message(
2274 &mut self,
2275 content: impl Into<String>,
2276 cx: &mut Context<Self>,
2277 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2278 let content = content.into();
2279 self.messages.push(Message::User(UserMessage {
2280 id: UserMessageId::new(),
2281 content: vec![UserMessageContent::Text(content)],
2282 }));
2283 cx.notify();
2284 self.send_existing(cx)
2285 }
2286
2287 pub fn interrupt_for_summary(
2288 &mut self,
2289 cx: &mut Context<Self>,
2290 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2291 let context = self
2292 .subagent_context
2293 .as_ref()
2294 .context("Not a subagent thread")?;
2295 let prompt = context.context_low_prompt.clone();
2296 self.cancel(cx).detach();
2297 self.submit_user_message(prompt, cx)
2298 }
2299
2300 pub fn request_final_summary(
2301 &mut self,
2302 cx: &mut Context<Self>,
2303 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2304 let context = self
2305 .subagent_context
2306 .as_ref()
2307 .context("Not a subagent thread")?;
2308 let prompt = context.summary_prompt.clone();
2309 self.submit_user_message(prompt, cx)
2310 }
2311
2312 fn build_request_messages(
2313 &self,
2314 available_tools: Vec<SharedString>,
2315 cx: &App,
2316 ) -> Vec<LanguageModelRequestMessage> {
2317 log::trace!(
2318 "Building request messages from {} thread messages",
2319 self.messages.len()
2320 );
2321
2322 let system_prompt = SystemPromptTemplate {
2323 project: self.project_context.read(cx),
2324 available_tools,
2325 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2326 }
2327 .render(&self.templates)
2328 .context("failed to build system prompt")
2329 .expect("Invalid template");
2330 let mut messages = vec![LanguageModelRequestMessage {
2331 role: Role::System,
2332 content: vec![system_prompt.into()],
2333 cache: false,
2334 reasoning_details: None,
2335 }];
2336 for message in &self.messages {
2337 messages.extend(message.to_request());
2338 }
2339
2340 if let Some(last_message) = messages.last_mut() {
2341 last_message.cache = true;
2342 }
2343
2344 if let Some(message) = self.pending_message.as_ref() {
2345 messages.extend(message.to_request());
2346 }
2347
2348 messages
2349 }
2350
2351 pub fn to_markdown(&self) -> String {
2352 let mut markdown = String::new();
2353 for (ix, message) in self.messages.iter().enumerate() {
2354 if ix > 0 {
2355 markdown.push('\n');
2356 }
2357 markdown.push_str(&message.to_markdown());
2358 }
2359
2360 if let Some(message) = self.pending_message.as_ref() {
2361 markdown.push('\n');
2362 markdown.push_str(&message.to_markdown());
2363 }
2364
2365 markdown
2366 }
2367
2368 fn advance_prompt_id(&mut self) {
2369 self.prompt_id = PromptId::new();
2370 }
2371
2372 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2373 use LanguageModelCompletionError::*;
2374 use http_client::StatusCode;
2375
2376 // General strategy here:
2377 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2378 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2379 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2380 match error {
2381 HttpResponseError {
2382 status_code: StatusCode::TOO_MANY_REQUESTS,
2383 ..
2384 } => Some(RetryStrategy::ExponentialBackoff {
2385 initial_delay: BASE_RETRY_DELAY,
2386 max_attempts: MAX_RETRY_ATTEMPTS,
2387 }),
2388 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2389 Some(RetryStrategy::Fixed {
2390 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2391 max_attempts: MAX_RETRY_ATTEMPTS,
2392 })
2393 }
2394 UpstreamProviderError {
2395 status,
2396 retry_after,
2397 ..
2398 } => match *status {
2399 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2400 Some(RetryStrategy::Fixed {
2401 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2402 max_attempts: MAX_RETRY_ATTEMPTS,
2403 })
2404 }
2405 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2406 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2407 // Internal Server Error could be anything, retry up to 3 times.
2408 max_attempts: 3,
2409 }),
2410 status => {
2411 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2412 // but we frequently get them in practice. See https://http.dev/529
2413 if status.as_u16() == 529 {
2414 Some(RetryStrategy::Fixed {
2415 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2416 max_attempts: MAX_RETRY_ATTEMPTS,
2417 })
2418 } else {
2419 Some(RetryStrategy::Fixed {
2420 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2421 max_attempts: 2,
2422 })
2423 }
2424 }
2425 },
2426 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2427 delay: BASE_RETRY_DELAY,
2428 max_attempts: 3,
2429 }),
2430 ApiReadResponseError { .. }
2431 | HttpSend { .. }
2432 | DeserializeResponse { .. }
2433 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2434 delay: BASE_RETRY_DELAY,
2435 max_attempts: 3,
2436 }),
2437 // Retrying these errors definitely shouldn't help.
2438 HttpResponseError {
2439 status_code:
2440 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2441 ..
2442 }
2443 | AuthenticationError { .. }
2444 | PermissionError { .. }
2445 | NoApiKey { .. }
2446 | ApiEndpointNotFound { .. }
2447 | PromptTooLarge { .. } => None,
2448 // These errors might be transient, so retry them
2449 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2450 delay: BASE_RETRY_DELAY,
2451 max_attempts: 1,
2452 }),
2453 // Retry all other 4xx and 5xx errors once.
2454 HttpResponseError { status_code, .. }
2455 if status_code.is_client_error() || status_code.is_server_error() =>
2456 {
2457 Some(RetryStrategy::Fixed {
2458 delay: BASE_RETRY_DELAY,
2459 max_attempts: 3,
2460 })
2461 }
2462 Other(err)
2463 if err.is::<language_model::PaymentRequiredError>()
2464 || err.is::<language_model::ModelRequestLimitReachedError>() =>
2465 {
2466 // Retrying won't help for Payment Required or Model Request Limit errors (where
2467 // the user must upgrade to usage-based billing to get more requests, or else wait
2468 // for a significant amount of time for the request limit to reset).
2469 None
2470 }
2471 // Conservatively assume that any other errors are non-retryable
2472 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2473 delay: BASE_RETRY_DELAY,
2474 max_attempts: 2,
2475 }),
2476 }
2477 }
2478}
2479
2480struct RunningTurn {
2481 /// Holds the task that handles agent interaction until the end of the turn.
2482 /// Survives across multiple requests as the model performs tool calls and
2483 /// we run tools, report their results.
2484 _task: Task<()>,
2485 /// The current event stream for the running turn. Used to report a final
2486 /// cancellation event if we cancel the turn.
2487 event_stream: ThreadEventStream,
2488 /// The tools that were enabled for this turn.
2489 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2490 /// Sender to signal tool cancellation. When cancel is called, this is
2491 /// set to true so all tools can detect user-initiated cancellation.
2492 cancellation_tx: watch::Sender<bool>,
2493}
2494
2495impl RunningTurn {
2496 fn cancel(mut self) -> Task<()> {
2497 log::debug!("Cancelling in progress turn");
2498 self.cancellation_tx.send(true).ok();
2499 self.event_stream.send_canceled();
2500 self._task
2501 }
2502}
2503
2504pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2505
2506impl EventEmitter<TokenUsageUpdated> for Thread {}
2507
2508pub struct TitleUpdated;
2509
2510impl EventEmitter<TitleUpdated> for Thread {}
2511
2512pub trait AgentTool
2513where
2514 Self: 'static + Sized,
2515{
2516 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2517 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2518
2519 fn name() -> &'static str;
2520
2521 fn description() -> SharedString {
2522 let schema = schemars::schema_for!(Self::Input);
2523 SharedString::new(
2524 schema
2525 .get("description")
2526 .and_then(|description| description.as_str())
2527 .unwrap_or_default(),
2528 )
2529 }
2530
2531 fn kind() -> acp::ToolKind;
2532
2533 /// The initial tool title to display. Can be updated during the tool run.
2534 fn initial_title(
2535 &self,
2536 input: Result<Self::Input, serde_json::Value>,
2537 cx: &mut App,
2538 ) -> SharedString;
2539
2540 /// Returns the JSON schema that describes the tool's input.
2541 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
2542 language_model::tool_schema::root_schema_for::<Self::Input>(format)
2543 }
2544
2545 /// Some tools rely on a provider for the underlying billing or other reasons.
2546 /// Allow the tool to check if they are compatible, or should be filtered out.
2547 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
2548 true
2549 }
2550
2551 /// Runs the tool with the provided input.
2552 fn run(
2553 self: Arc<Self>,
2554 input: Self::Input,
2555 event_stream: ToolCallEventStream,
2556 cx: &mut App,
2557 ) -> Task<Result<Self::Output>>;
2558
2559 /// Emits events for a previous execution of the tool.
2560 fn replay(
2561 &self,
2562 _input: Self::Input,
2563 _output: Self::Output,
2564 _event_stream: ToolCallEventStream,
2565 _cx: &mut App,
2566 ) -> Result<()> {
2567 Ok(())
2568 }
2569
2570 fn erase(self) -> Arc<dyn AnyAgentTool> {
2571 Arc::new(Erased(Arc::new(self)))
2572 }
2573}
2574
2575pub struct Erased<T>(T);
2576
2577pub struct AgentToolOutput {
2578 pub llm_output: LanguageModelToolResultContent,
2579 pub raw_output: serde_json::Value,
2580}
2581
2582pub trait AnyAgentTool {
2583 fn name(&self) -> SharedString;
2584 fn description(&self) -> SharedString;
2585 fn kind(&self) -> acp::ToolKind;
2586 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2587 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2588 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2589 true
2590 }
2591 fn run(
2592 self: Arc<Self>,
2593 input: serde_json::Value,
2594 event_stream: ToolCallEventStream,
2595 cx: &mut App,
2596 ) -> Task<Result<AgentToolOutput>>;
2597 fn replay(
2598 &self,
2599 input: serde_json::Value,
2600 output: serde_json::Value,
2601 event_stream: ToolCallEventStream,
2602 cx: &mut App,
2603 ) -> Result<()>;
2604}
2605
2606impl<T> AnyAgentTool for Erased<Arc<T>>
2607where
2608 T: AgentTool,
2609{
2610 fn name(&self) -> SharedString {
2611 T::name().into()
2612 }
2613
2614 fn description(&self) -> SharedString {
2615 T::description()
2616 }
2617
2618 fn kind(&self) -> agent_client_protocol::ToolKind {
2619 T::kind()
2620 }
2621
2622 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2623 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2624 self.0.initial_title(parsed_input, _cx)
2625 }
2626
2627 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2628 let mut json = serde_json::to_value(T::input_schema(format))?;
2629 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
2630 Ok(json)
2631 }
2632
2633 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
2634 T::supports_provider(provider)
2635 }
2636
2637 fn run(
2638 self: Arc<Self>,
2639 input: serde_json::Value,
2640 event_stream: ToolCallEventStream,
2641 cx: &mut App,
2642 ) -> Task<Result<AgentToolOutput>> {
2643 cx.spawn(async move |cx| {
2644 let input = serde_json::from_value(input)?;
2645 let output = cx
2646 .update(|cx| self.0.clone().run(input, event_stream, cx))
2647 .await?;
2648 let raw_output = serde_json::to_value(&output)?;
2649 Ok(AgentToolOutput {
2650 llm_output: output.into(),
2651 raw_output,
2652 })
2653 })
2654 }
2655
2656 fn replay(
2657 &self,
2658 input: serde_json::Value,
2659 output: serde_json::Value,
2660 event_stream: ToolCallEventStream,
2661 cx: &mut App,
2662 ) -> Result<()> {
2663 let input = serde_json::from_value(input)?;
2664 let output = serde_json::from_value(output)?;
2665 self.0.replay(input, output, event_stream, cx)
2666 }
2667}
2668
2669#[derive(Clone)]
2670struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2671
2672impl ThreadEventStream {
2673 fn send_user_message(&self, message: &UserMessage) {
2674 self.0
2675 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2676 .ok();
2677 }
2678
2679 fn send_text(&self, text: &str) {
2680 self.0
2681 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2682 .ok();
2683 }
2684
2685 fn send_thinking(&self, text: &str) {
2686 self.0
2687 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2688 .ok();
2689 }
2690
2691 fn send_tool_call(
2692 &self,
2693 id: &LanguageModelToolUseId,
2694 tool_name: &str,
2695 title: SharedString,
2696 kind: acp::ToolKind,
2697 input: serde_json::Value,
2698 ) {
2699 self.0
2700 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2701 id,
2702 tool_name,
2703 title.to_string(),
2704 kind,
2705 input,
2706 ))))
2707 .ok();
2708 }
2709
2710 fn initial_tool_call(
2711 id: &LanguageModelToolUseId,
2712 tool_name: &str,
2713 title: String,
2714 kind: acp::ToolKind,
2715 input: serde_json::Value,
2716 ) -> acp::ToolCall {
2717 acp::ToolCall::new(id.to_string(), title)
2718 .kind(kind)
2719 .raw_input(input)
2720 .meta(acp_thread::meta_with_tool_name(tool_name))
2721 }
2722
2723 fn update_tool_call_fields(
2724 &self,
2725 tool_use_id: &LanguageModelToolUseId,
2726 fields: acp::ToolCallUpdateFields,
2727 ) {
2728 self.0
2729 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2730 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields).into(),
2731 )))
2732 .ok();
2733 }
2734
2735 fn send_retry(&self, status: acp_thread::RetryStatus) {
2736 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2737 }
2738
2739 fn send_stop(&self, reason: acp::StopReason) {
2740 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2741 }
2742
2743 fn send_canceled(&self) {
2744 self.0
2745 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2746 .ok();
2747 }
2748
2749 fn send_error(&self, error: impl Into<anyhow::Error>) {
2750 self.0.unbounded_send(Err(error.into())).ok();
2751 }
2752}
2753
2754#[derive(Clone)]
2755pub struct ToolCallEventStream {
2756 tool_use_id: LanguageModelToolUseId,
2757 stream: ThreadEventStream,
2758 fs: Option<Arc<dyn Fs>>,
2759 cancellation_rx: watch::Receiver<bool>,
2760}
2761
2762impl ToolCallEventStream {
2763 #[cfg(any(test, feature = "test-support"))]
2764 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2765 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2766 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
2767
2768 let stream = ToolCallEventStream::new(
2769 "test_id".into(),
2770 ThreadEventStream(events_tx),
2771 None,
2772 cancellation_rx,
2773 );
2774
2775 (stream, ToolCallEventStreamReceiver(events_rx))
2776 }
2777
2778 fn new(
2779 tool_use_id: LanguageModelToolUseId,
2780 stream: ThreadEventStream,
2781 fs: Option<Arc<dyn Fs>>,
2782 cancellation_rx: watch::Receiver<bool>,
2783 ) -> Self {
2784 Self {
2785 tool_use_id,
2786 stream,
2787 fs,
2788 cancellation_rx,
2789 }
2790 }
2791
2792 /// Returns a future that resolves when the user cancels the tool call.
2793 /// Tools should select on this alongside their main work to detect user cancellation.
2794 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
2795 let mut rx = self.cancellation_rx.clone();
2796 async move {
2797 loop {
2798 if *rx.borrow() {
2799 return;
2800 }
2801 if rx.changed().await.is_err() {
2802 // Sender dropped, will never be cancelled
2803 std::future::pending::<()>().await;
2804 }
2805 }
2806 }
2807 }
2808
2809 /// Returns true if the user has cancelled this tool call.
2810 /// This is useful for checking cancellation state after an operation completes,
2811 /// to determine if the completion was due to user cancellation.
2812 pub fn was_cancelled_by_user(&self) -> bool {
2813 *self.cancellation_rx.clone().borrow()
2814 }
2815
2816 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
2817 &self.tool_use_id
2818 }
2819
2820 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
2821 self.stream
2822 .update_tool_call_fields(&self.tool_use_id, fields);
2823 }
2824
2825 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
2826 self.stream
2827 .0
2828 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2829 acp_thread::ToolCallUpdateDiff {
2830 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
2831 diff,
2832 }
2833 .into(),
2834 )))
2835 .ok();
2836 }
2837
2838 pub fn update_subagent_thread(&self, thread: Entity<acp_thread::AcpThread>) {
2839 self.stream
2840 .0
2841 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2842 acp_thread::ToolCallUpdateSubagentThread {
2843 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
2844 thread,
2845 }
2846 .into(),
2847 )))
2848 .ok();
2849 }
2850
2851 pub fn authorize(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2852 if agent_settings::AgentSettings::get_global(cx).always_allow_tool_actions {
2853 return Task::ready(Ok(()));
2854 }
2855
2856 self.authorize_required(title, cx)
2857 }
2858
2859 /// Like `authorize`, but always prompts for confirmation regardless of
2860 /// the `always_allow_tool_actions` setting. Use this when tool-specific
2861 /// permission rules (like `always_confirm` patterns) have already determined
2862 /// that confirmation is required.
2863 pub fn authorize_required(&self, title: impl Into<String>, cx: &mut App) -> Task<Result<()>> {
2864 let (response_tx, response_rx) = oneshot::channel();
2865 self.stream
2866 .0
2867 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
2868 ToolCallAuthorization {
2869 tool_call: acp::ToolCallUpdate::new(
2870 self.tool_use_id.to_string(),
2871 acp::ToolCallUpdateFields::new().title(title.into()),
2872 ),
2873 options: vec![
2874 acp::PermissionOption::new(
2875 acp::PermissionOptionId::new("always_allow"),
2876 "Always Allow",
2877 acp::PermissionOptionKind::AllowAlways,
2878 ),
2879 acp::PermissionOption::new(
2880 acp::PermissionOptionId::new("allow"),
2881 "Allow",
2882 acp::PermissionOptionKind::AllowOnce,
2883 ),
2884 acp::PermissionOption::new(
2885 acp::PermissionOptionId::new("deny"),
2886 "Deny",
2887 acp::PermissionOptionKind::RejectOnce,
2888 ),
2889 ],
2890 response: response_tx,
2891 },
2892 )))
2893 .ok();
2894 let fs = self.fs.clone();
2895 cx.spawn(async move |cx| match response_rx.await?.0.as_ref() {
2896 "always_allow" => {
2897 if let Some(fs) = fs.clone() {
2898 cx.update(|cx| {
2899 update_settings_file(fs, cx, |settings, _| {
2900 settings
2901 .agent
2902 .get_or_insert_default()
2903 .set_always_allow_tool_actions(true);
2904 });
2905 });
2906 }
2907
2908 Ok(())
2909 }
2910 "allow" => Ok(()),
2911 _ => Err(anyhow!("Permission to run tool denied by user")),
2912 })
2913 }
2914}
2915
2916#[cfg(any(test, feature = "test-support"))]
2917pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
2918
2919#[cfg(any(test, feature = "test-support"))]
2920impl ToolCallEventStreamReceiver {
2921 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
2922 let event = self.0.next().await;
2923 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
2924 auth
2925 } else {
2926 panic!("Expected ToolCallAuthorization but got: {:?}", event);
2927 }
2928 }
2929
2930 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
2931 let event = self.0.next().await;
2932 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
2933 update,
2934 )))) = event
2935 {
2936 update.fields
2937 } else {
2938 panic!("Expected update fields but got: {:?}", event);
2939 }
2940 }
2941
2942 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
2943 let event = self.0.next().await;
2944 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
2945 update,
2946 )))) = event
2947 {
2948 update.diff
2949 } else {
2950 panic!("Expected diff but got: {:?}", event);
2951 }
2952 }
2953
2954 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
2955 let event = self.0.next().await;
2956 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
2957 update,
2958 )))) = event
2959 {
2960 update.terminal
2961 } else {
2962 panic!("Expected terminal but got: {:?}", event);
2963 }
2964 }
2965}
2966
2967#[cfg(any(test, feature = "test-support"))]
2968impl std::ops::Deref for ToolCallEventStreamReceiver {
2969 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
2970
2971 fn deref(&self) -> &Self::Target {
2972 &self.0
2973 }
2974}
2975
2976#[cfg(any(test, feature = "test-support"))]
2977impl std::ops::DerefMut for ToolCallEventStreamReceiver {
2978 fn deref_mut(&mut self) -> &mut Self::Target {
2979 &mut self.0
2980 }
2981}
2982
2983impl From<&str> for UserMessageContent {
2984 fn from(text: &str) -> Self {
2985 Self::Text(text.into())
2986 }
2987}
2988
2989impl UserMessageContent {
2990 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
2991 match value {
2992 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
2993 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
2994 acp::ContentBlock::Audio(_) => {
2995 // TODO
2996 Self::Text("[audio]".to_string())
2997 }
2998 acp::ContentBlock::ResourceLink(resource_link) => {
2999 match MentionUri::parse(&resource_link.uri, path_style) {
3000 Ok(uri) => Self::Mention {
3001 uri,
3002 content: String::new(),
3003 },
3004 Err(err) => {
3005 log::error!("Failed to parse mention link: {}", err);
3006 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
3007 }
3008 }
3009 }
3010 acp::ContentBlock::Resource(resource) => match resource.resource {
3011 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
3012 match MentionUri::parse(&resource.uri, path_style) {
3013 Ok(uri) => Self::Mention {
3014 uri,
3015 content: resource.text,
3016 },
3017 Err(err) => {
3018 log::error!("Failed to parse mention link: {}", err);
3019 Self::Text(
3020 MarkdownCodeBlock {
3021 tag: &resource.uri,
3022 text: &resource.text,
3023 }
3024 .to_string(),
3025 )
3026 }
3027 }
3028 }
3029 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
3030 // TODO
3031 Self::Text("[blob]".to_string())
3032 }
3033 other => {
3034 log::warn!("Unexpected content type: {:?}", other);
3035 Self::Text("[unknown]".to_string())
3036 }
3037 },
3038 other => {
3039 log::warn!("Unexpected content type: {:?}", other);
3040 Self::Text("[unknown]".to_string())
3041 }
3042 }
3043 }
3044}
3045
3046impl From<UserMessageContent> for acp::ContentBlock {
3047 fn from(content: UserMessageContent) -> Self {
3048 match content {
3049 UserMessageContent::Text(text) => text.into(),
3050 UserMessageContent::Image(image) => {
3051 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
3052 }
3053 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
3054 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
3055 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
3056 )),
3057 ),
3058 }
3059 }
3060}
3061
3062fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
3063 LanguageModelImage {
3064 source: image_content.data.into(),
3065 size: None,
3066 }
3067}