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, StreamingEditFileTool, SubagentTool,
6 SystemPromptTemplate, Template, Templates, TerminalTool, ThinkingTool, ToolPermissionDecision,
7 WebSearchTool, decide_permission_from_settings,
8};
9use acp_thread::{MentionUri, UserMessageId};
10use action_log::ActionLog;
11use feature_flags::{FeatureFlagAppExt as _, SubagentsFeatureFlag};
12
13use agent_client_protocol as acp;
14use agent_settings::{
15 AgentProfileId, AgentProfileSettings, AgentSettings, SUMMARIZE_THREAD_DETAILED_PROMPT,
16 SUMMARIZE_THREAD_PROMPT,
17};
18use anyhow::{Context as _, Result, anyhow};
19use chrono::{DateTime, Utc};
20use client::UserStore;
21use cloud_api_types::Plan;
22use cloud_llm_client::CompletionIntent;
23use collections::{HashMap, HashSet, IndexMap};
24use fs::Fs;
25use futures::stream;
26use futures::{
27 FutureExt,
28 channel::{mpsc, oneshot},
29 future::Shared,
30 stream::FuturesUnordered,
31};
32use gpui::{
33 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
34};
35use language_model::{
36 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,
37 LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry, LanguageModelRequest,
38 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
39 LanguageModelToolResultContent, LanguageModelToolSchemaFormat, LanguageModelToolUse,
40 LanguageModelToolUseId, Role, SelectedModel, StopReason, TokenUsage, ZED_CLOUD_PROVIDER_ID,
41};
42use project::Project;
43use prompt_store::ProjectContext;
44use schemars::{JsonSchema, Schema};
45use serde::{Deserialize, Serialize};
46use settings::{LanguageModelSelection, Settings, ToolPermissionMode, update_settings_file};
47use smol::stream::StreamExt;
48use std::{
49 collections::BTreeMap,
50 ops::RangeInclusive,
51 path::Path,
52 rc::Rc,
53 sync::Arc,
54 time::{Duration, Instant},
55};
56use std::{fmt::Write, path::PathBuf};
57use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
58use uuid::Uuid;
59
60const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
61pub const MAX_TOOL_NAME_LENGTH: usize = 64;
62pub const MAX_SUBAGENT_DEPTH: u8 = 4;
63pub const MAX_PARALLEL_SUBAGENTS: usize = 8;
64
65/// Context passed to a subagent thread for lifecycle management
66#[derive(Clone)]
67pub struct SubagentContext {
68 /// ID of the parent thread
69 pub parent_thread_id: acp::SessionId,
70
71 /// ID of the tool call that spawned this subagent
72 pub tool_use_id: LanguageModelToolUseId,
73
74 /// Current depth level (0 = root agent, 1 = first-level subagent, etc.)
75 pub depth: u8,
76
77 /// Prompt to send when subagent completes successfully
78 pub summary_prompt: String,
79
80 /// Prompt to send when context is running low (≤25% remaining)
81 pub context_low_prompt: String,
82}
83
84/// The ID of the user prompt that initiated a request.
85///
86/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
87#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
88pub struct PromptId(Arc<str>);
89
90impl PromptId {
91 pub fn new() -> Self {
92 Self(Uuid::new_v4().to_string().into())
93 }
94}
95
96impl std::fmt::Display for PromptId {
97 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
98 write!(f, "{}", self.0)
99 }
100}
101
102pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
103pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
104
105#[derive(Debug, Clone)]
106enum RetryStrategy {
107 ExponentialBackoff {
108 initial_delay: Duration,
109 max_attempts: u8,
110 },
111 Fixed {
112 delay: Duration,
113 max_attempts: u8,
114 },
115}
116
117#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
118pub enum Message {
119 User(UserMessage),
120 Agent(AgentMessage),
121 Resume,
122}
123
124impl Message {
125 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
126 match self {
127 Message::Agent(agent_message) => Some(agent_message),
128 _ => None,
129 }
130 }
131
132 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
133 match self {
134 Message::User(message) => {
135 if message.content.is_empty() {
136 vec![]
137 } else {
138 vec![message.to_request()]
139 }
140 }
141 Message::Agent(message) => message.to_request(),
142 Message::Resume => vec![LanguageModelRequestMessage {
143 role: Role::User,
144 content: vec!["Continue where you left off".into()],
145 cache: false,
146 reasoning_details: None,
147 }],
148 }
149 }
150
151 pub fn to_markdown(&self) -> String {
152 match self {
153 Message::User(message) => message.to_markdown(),
154 Message::Agent(message) => message.to_markdown(),
155 Message::Resume => "[resume]\n".into(),
156 }
157 }
158
159 pub fn role(&self) -> Role {
160 match self {
161 Message::User(_) | Message::Resume => Role::User,
162 Message::Agent(_) => Role::Assistant,
163 }
164 }
165}
166
167#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
168pub struct UserMessage {
169 pub id: UserMessageId,
170 pub content: Vec<UserMessageContent>,
171}
172
173#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
174pub enum UserMessageContent {
175 Text(String),
176 Mention { uri: MentionUri, content: String },
177 Image(LanguageModelImage),
178}
179
180impl UserMessage {
181 pub fn to_markdown(&self) -> String {
182 let mut markdown = String::from("## User\n\n");
183
184 for content in &self.content {
185 match content {
186 UserMessageContent::Text(text) => {
187 markdown.push_str(text);
188 markdown.push('\n');
189 }
190 UserMessageContent::Image(_) => {
191 markdown.push_str("<image />\n");
192 }
193 UserMessageContent::Mention { uri, content } => {
194 if !content.is_empty() {
195 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
196 } else {
197 let _ = writeln!(&mut markdown, "{}", uri.as_link());
198 }
199 }
200 }
201 }
202
203 markdown
204 }
205
206 fn to_request(&self) -> LanguageModelRequestMessage {
207 let mut message = LanguageModelRequestMessage {
208 role: Role::User,
209 content: Vec::with_capacity(self.content.len()),
210 cache: false,
211 reasoning_details: None,
212 };
213
214 const OPEN_CONTEXT: &str = "<context>\n\
215 The following items were attached by the user. \
216 They are up-to-date and don't need to be re-read.\n\n";
217
218 const OPEN_FILES_TAG: &str = "<files>";
219 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
220 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
221 const OPEN_SELECTIONS_TAG: &str = "<selections>";
222 const OPEN_THREADS_TAG: &str = "<threads>";
223 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
224 const OPEN_RULES_TAG: &str =
225 "<rules>\nThe user has specified the following rules that should be applied:\n";
226 const OPEN_DIAGNOSTICS_TAG: &str = "<diagnostics>";
227
228 let mut file_context = OPEN_FILES_TAG.to_string();
229 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
230 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
231 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
232 let mut thread_context = OPEN_THREADS_TAG.to_string();
233 let mut fetch_context = OPEN_FETCH_TAG.to_string();
234 let mut rules_context = OPEN_RULES_TAG.to_string();
235 let mut diagnostics_context = OPEN_DIAGNOSTICS_TAG.to_string();
236
237 for chunk in &self.content {
238 let chunk = match chunk {
239 UserMessageContent::Text(text) => {
240 language_model::MessageContent::Text(text.clone())
241 }
242 UserMessageContent::Image(value) => {
243 language_model::MessageContent::Image(value.clone())
244 }
245 UserMessageContent::Mention { uri, content } => {
246 match uri {
247 MentionUri::File { abs_path } => {
248 write!(
249 &mut file_context,
250 "\n{}",
251 MarkdownCodeBlock {
252 tag: &codeblock_tag(abs_path, None),
253 text: &content.to_string(),
254 }
255 )
256 .ok();
257 }
258 MentionUri::PastedImage => {
259 debug_panic!("pasted image URI should not be used in mention content")
260 }
261 MentionUri::Directory { .. } => {
262 write!(&mut directory_context, "\n{}\n", content).ok();
263 }
264 MentionUri::Symbol {
265 abs_path: path,
266 line_range,
267 ..
268 } => {
269 write!(
270 &mut symbol_context,
271 "\n{}",
272 MarkdownCodeBlock {
273 tag: &codeblock_tag(path, Some(line_range)),
274 text: content
275 }
276 )
277 .ok();
278 }
279 MentionUri::Selection {
280 abs_path: path,
281 line_range,
282 ..
283 } => {
284 write!(
285 &mut selection_context,
286 "\n{}",
287 MarkdownCodeBlock {
288 tag: &codeblock_tag(
289 path.as_deref().unwrap_or("Untitled".as_ref()),
290 Some(line_range)
291 ),
292 text: content
293 }
294 )
295 .ok();
296 }
297 MentionUri::Thread { .. } => {
298 write!(&mut thread_context, "\n{}\n", content).ok();
299 }
300 MentionUri::TextThread { .. } => {
301 write!(&mut thread_context, "\n{}\n", content).ok();
302 }
303 MentionUri::Rule { .. } => {
304 write!(
305 &mut rules_context,
306 "\n{}",
307 MarkdownCodeBlock {
308 tag: "",
309 text: content
310 }
311 )
312 .ok();
313 }
314 MentionUri::Fetch { url } => {
315 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
316 }
317 MentionUri::Diagnostics { .. } => {
318 write!(&mut diagnostics_context, "\n{}\n", content).ok();
319 }
320 MentionUri::TerminalSelection { .. } => {
321 write!(
322 &mut selection_context,
323 "\n{}",
324 MarkdownCodeBlock {
325 tag: "console",
326 text: content
327 }
328 )
329 .ok();
330 }
331 }
332
333 language_model::MessageContent::Text(uri.as_link().to_string())
334 }
335 };
336
337 message.content.push(chunk);
338 }
339
340 let len_before_context = message.content.len();
341
342 if file_context.len() > OPEN_FILES_TAG.len() {
343 file_context.push_str("</files>\n");
344 message
345 .content
346 .push(language_model::MessageContent::Text(file_context));
347 }
348
349 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
350 directory_context.push_str("</directories>\n");
351 message
352 .content
353 .push(language_model::MessageContent::Text(directory_context));
354 }
355
356 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
357 symbol_context.push_str("</symbols>\n");
358 message
359 .content
360 .push(language_model::MessageContent::Text(symbol_context));
361 }
362
363 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
364 selection_context.push_str("</selections>\n");
365 message
366 .content
367 .push(language_model::MessageContent::Text(selection_context));
368 }
369
370 if thread_context.len() > OPEN_THREADS_TAG.len() {
371 thread_context.push_str("</threads>\n");
372 message
373 .content
374 .push(language_model::MessageContent::Text(thread_context));
375 }
376
377 if fetch_context.len() > OPEN_FETCH_TAG.len() {
378 fetch_context.push_str("</fetched_urls>\n");
379 message
380 .content
381 .push(language_model::MessageContent::Text(fetch_context));
382 }
383
384 if rules_context.len() > OPEN_RULES_TAG.len() {
385 rules_context.push_str("</user_rules>\n");
386 message
387 .content
388 .push(language_model::MessageContent::Text(rules_context));
389 }
390
391 if diagnostics_context.len() > OPEN_DIAGNOSTICS_TAG.len() {
392 diagnostics_context.push_str("</diagnostics>\n");
393 message
394 .content
395 .push(language_model::MessageContent::Text(diagnostics_context));
396 }
397
398 if message.content.len() > len_before_context {
399 message.content.insert(
400 len_before_context,
401 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
402 );
403 message
404 .content
405 .push(language_model::MessageContent::Text("</context>".into()));
406 }
407
408 message
409 }
410}
411
412fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
413 let mut result = String::new();
414
415 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
416 let _ = write!(result, "{} ", extension);
417 }
418
419 let _ = write!(result, "{}", full_path.display());
420
421 if let Some(range) = line_range {
422 if range.start() == range.end() {
423 let _ = write!(result, ":{}", range.start() + 1);
424 } else {
425 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
426 }
427 }
428
429 result
430}
431
432impl AgentMessage {
433 pub fn to_markdown(&self) -> String {
434 let mut markdown = String::from("## Assistant\n\n");
435
436 for content in &self.content {
437 match content {
438 AgentMessageContent::Text(text) => {
439 markdown.push_str(text);
440 markdown.push('\n');
441 }
442 AgentMessageContent::Thinking { text, .. } => {
443 markdown.push_str("<think>");
444 markdown.push_str(text);
445 markdown.push_str("</think>\n");
446 }
447 AgentMessageContent::RedactedThinking(_) => {
448 markdown.push_str("<redacted_thinking />\n")
449 }
450 AgentMessageContent::ToolUse(tool_use) => {
451 markdown.push_str(&format!(
452 "**Tool Use**: {} (ID: {})\n",
453 tool_use.name, tool_use.id
454 ));
455 markdown.push_str(&format!(
456 "{}\n",
457 MarkdownCodeBlock {
458 tag: "json",
459 text: &format!("{:#}", tool_use.input)
460 }
461 ));
462 }
463 }
464 }
465
466 for tool_result in self.tool_results.values() {
467 markdown.push_str(&format!(
468 "**Tool Result**: {} (ID: {})\n\n",
469 tool_result.tool_name, tool_result.tool_use_id
470 ));
471 if tool_result.is_error {
472 markdown.push_str("**ERROR:**\n");
473 }
474
475 match &tool_result.content {
476 LanguageModelToolResultContent::Text(text) => {
477 writeln!(markdown, "{text}\n").ok();
478 }
479 LanguageModelToolResultContent::Image(_) => {
480 writeln!(markdown, "<image />\n").ok();
481 }
482 }
483
484 if let Some(output) = tool_result.output.as_ref() {
485 writeln!(
486 markdown,
487 "**Debug Output**:\n\n```json\n{}\n```\n",
488 serde_json::to_string_pretty(output).unwrap()
489 )
490 .unwrap();
491 }
492 }
493
494 markdown
495 }
496
497 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
498 let mut assistant_message = LanguageModelRequestMessage {
499 role: Role::Assistant,
500 content: Vec::with_capacity(self.content.len()),
501 cache: false,
502 reasoning_details: self.reasoning_details.clone(),
503 };
504 for chunk in &self.content {
505 match chunk {
506 AgentMessageContent::Text(text) => {
507 assistant_message
508 .content
509 .push(language_model::MessageContent::Text(text.clone()));
510 }
511 AgentMessageContent::Thinking { text, signature } => {
512 assistant_message
513 .content
514 .push(language_model::MessageContent::Thinking {
515 text: text.clone(),
516 signature: signature.clone(),
517 });
518 }
519 AgentMessageContent::RedactedThinking(value) => {
520 assistant_message.content.push(
521 language_model::MessageContent::RedactedThinking(value.clone()),
522 );
523 }
524 AgentMessageContent::ToolUse(tool_use) => {
525 if self.tool_results.contains_key(&tool_use.id) {
526 assistant_message
527 .content
528 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
529 }
530 }
531 };
532 }
533
534 let mut user_message = LanguageModelRequestMessage {
535 role: Role::User,
536 content: Vec::new(),
537 cache: false,
538 reasoning_details: None,
539 };
540
541 for tool_result in self.tool_results.values() {
542 let mut tool_result = tool_result.clone();
543 // Surprisingly, the API fails if we return an empty string here.
544 // It thinks we are sending a tool use without a tool result.
545 if tool_result.content.is_empty() {
546 tool_result.content = "<Tool returned an empty string>".into();
547 }
548 user_message
549 .content
550 .push(language_model::MessageContent::ToolResult(tool_result));
551 }
552
553 let mut messages = Vec::new();
554 if !assistant_message.content.is_empty() {
555 messages.push(assistant_message);
556 }
557 if !user_message.content.is_empty() {
558 messages.push(user_message);
559 }
560 messages
561 }
562}
563
564#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
565pub struct AgentMessage {
566 pub content: Vec<AgentMessageContent>,
567 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
568 pub reasoning_details: Option<serde_json::Value>,
569}
570
571#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
572pub enum AgentMessageContent {
573 Text(String),
574 Thinking {
575 text: String,
576 signature: Option<String>,
577 },
578 RedactedThinking(String),
579 ToolUse(LanguageModelToolUse),
580}
581
582pub trait TerminalHandle {
583 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
584 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
585 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
586 fn kill(&self, cx: &AsyncApp) -> Result<()>;
587 fn was_stopped_by_user(&self, cx: &AsyncApp) -> Result<bool>;
588}
589
590pub trait ThreadEnvironment {
591 fn create_terminal(
592 &self,
593 command: String,
594 cwd: Option<PathBuf>,
595 output_byte_limit: Option<u64>,
596 cx: &mut AsyncApp,
597 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
598}
599
600#[derive(Debug)]
601pub enum ThreadEvent {
602 UserMessage(UserMessage),
603 AgentText(String),
604 AgentThinking(String),
605 ToolCall(acp::ToolCall),
606 ToolCallUpdate(acp_thread::ToolCallUpdate),
607 ToolCallAuthorization(ToolCallAuthorization),
608 Retry(acp_thread::RetryStatus),
609 Stop(acp::StopReason),
610}
611
612#[derive(Debug)]
613pub struct NewTerminal {
614 pub command: String,
615 pub output_byte_limit: Option<u64>,
616 pub cwd: Option<PathBuf>,
617 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
618}
619
620#[derive(Debug, Clone)]
621pub struct ToolPermissionContext {
622 pub tool_name: String,
623 pub input_value: String,
624}
625
626impl ToolPermissionContext {
627 pub fn new(tool_name: impl Into<String>, input_value: impl Into<String>) -> Self {
628 Self {
629 tool_name: tool_name.into(),
630 input_value: input_value.into(),
631 }
632 }
633
634 /// Builds the permission options for this tool context.
635 ///
636 /// This is the canonical source for permission option generation.
637 /// Tests should use this function rather than manually constructing options.
638 ///
639 /// # Shell Compatibility for Terminal Tool
640 ///
641 /// For the terminal tool, "Always allow" options are only shown when the user's
642 /// shell supports POSIX-like command chaining syntax (`&&`, `||`, `;`, `|`).
643 ///
644 /// **Why this matters:** When a user sets up an "always allow" pattern like `^cargo`,
645 /// we need to parse the command to extract all sub-commands and verify that EVERY
646 /// sub-command matches the pattern. Otherwise, an attacker could craft a command like
647 /// `cargo build && rm -rf /` that would bypass the security check.
648 ///
649 /// **Supported shells:** Posix (sh, bash, dash, zsh), Fish 3.0+, PowerShell 7+/Pwsh,
650 /// Cmd, Xonsh, Csh, Tcsh
651 ///
652 /// **Unsupported shells:** Nushell (uses `and`/`or` keywords), Elvish (uses `and`/`or`
653 /// keywords), Rc (Plan 9 shell - no `&&`/`||` operators)
654 ///
655 /// For unsupported shells, we hide the "Always allow" UI options entirely, and if
656 /// the user has `always_allow` rules configured in settings, `ToolPermissionDecision::from_input`
657 /// will return a `Deny` with an explanatory error message.
658 pub fn build_permission_options(&self) -> acp_thread::PermissionOptions {
659 use crate::pattern_extraction::*;
660 use util::shell::ShellKind;
661
662 let tool_name = &self.tool_name;
663 let input_value = &self.input_value;
664
665 // Check if the user's shell supports POSIX-like command chaining.
666 // See the doc comment above for the full explanation of why this is needed.
667 let shell_supports_always_allow = if tool_name == TerminalTool::name() {
668 ShellKind::system().supports_posix_chaining()
669 } else {
670 true
671 };
672
673 let (pattern, pattern_display) = if tool_name == TerminalTool::name() {
674 (
675 extract_terminal_pattern(input_value),
676 extract_terminal_pattern_display(input_value),
677 )
678 } else if tool_name == EditFileTool::name()
679 || tool_name == DeletePathTool::name()
680 || tool_name == MovePathTool::name()
681 || tool_name == CreateDirectoryTool::name()
682 || tool_name == SaveFileTool::name()
683 {
684 (
685 extract_path_pattern(input_value),
686 extract_path_pattern_display(input_value),
687 )
688 } else if tool_name == FetchTool::name() {
689 (
690 extract_url_pattern(input_value),
691 extract_url_pattern_display(input_value),
692 )
693 } else {
694 (None, None)
695 };
696
697 let mut choices = Vec::new();
698
699 let mut push_choice = |label: String, allow_id, deny_id, allow_kind, deny_kind| {
700 choices.push(acp_thread::PermissionOptionChoice {
701 allow: acp::PermissionOption::new(
702 acp::PermissionOptionId::new(allow_id),
703 label.clone(),
704 allow_kind,
705 ),
706 deny: acp::PermissionOption::new(
707 acp::PermissionOptionId::new(deny_id),
708 label,
709 deny_kind,
710 ),
711 });
712 };
713
714 if shell_supports_always_allow {
715 push_choice(
716 format!("Always for {}", tool_name.replace('_', " ")),
717 format!("always_allow:{}", tool_name),
718 format!("always_deny:{}", tool_name),
719 acp::PermissionOptionKind::AllowAlways,
720 acp::PermissionOptionKind::RejectAlways,
721 );
722
723 if let (Some(pattern), Some(display)) = (pattern, pattern_display) {
724 let button_text = if tool_name == TerminalTool::name() {
725 format!("Always for `{}` commands", display)
726 } else {
727 format!("Always for `{}`", display)
728 };
729 push_choice(
730 button_text,
731 format!("always_allow_pattern:{}:{}", tool_name, pattern),
732 format!("always_deny_pattern:{}:{}", tool_name, pattern),
733 acp::PermissionOptionKind::AllowAlways,
734 acp::PermissionOptionKind::RejectAlways,
735 );
736 }
737 }
738
739 push_choice(
740 "Only this time".to_string(),
741 "allow".to_string(),
742 "deny".to_string(),
743 acp::PermissionOptionKind::AllowOnce,
744 acp::PermissionOptionKind::RejectOnce,
745 );
746
747 acp_thread::PermissionOptions::Dropdown(choices)
748 }
749}
750
751#[derive(Debug)]
752pub struct ToolCallAuthorization {
753 pub tool_call: acp::ToolCallUpdate,
754 pub options: acp_thread::PermissionOptions,
755 pub response: oneshot::Sender<acp::PermissionOptionId>,
756 pub context: Option<ToolPermissionContext>,
757}
758
759#[derive(Debug, thiserror::Error)]
760enum CompletionError {
761 #[error("max tokens")]
762 MaxTokens,
763 #[error("refusal")]
764 Refusal,
765 #[error(transparent)]
766 Other(#[from] anyhow::Error),
767}
768
769pub struct Thread {
770 id: acp::SessionId,
771 prompt_id: PromptId,
772 updated_at: DateTime<Utc>,
773 title: Option<SharedString>,
774 pending_title_generation: Option<Task<()>>,
775 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
776 summary: Option<SharedString>,
777 messages: Vec<Message>,
778 user_store: Entity<UserStore>,
779 /// Holds the task that handles agent interaction until the end of the turn.
780 /// Survives across multiple requests as the model performs tool calls and
781 /// we run tools, report their results.
782 running_turn: Option<RunningTurn>,
783 /// Flag indicating the UI has a queued message waiting to be sent.
784 /// Used to signal that the turn should end at the next message boundary.
785 has_queued_message: bool,
786 pending_message: Option<AgentMessage>,
787 pub(crate) tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
788 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
789 #[allow(unused)]
790 cumulative_token_usage: TokenUsage,
791 #[allow(unused)]
792 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
793 pub(crate) context_server_registry: Entity<ContextServerRegistry>,
794 profile_id: AgentProfileId,
795 project_context: Entity<ProjectContext>,
796 pub(crate) templates: Arc<Templates>,
797 model: Option<Arc<dyn LanguageModel>>,
798 summarization_model: Option<Arc<dyn LanguageModel>>,
799 thinking_enabled: bool,
800 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
801 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
802 pub(crate) project: Entity<Project>,
803 pub(crate) action_log: Entity<ActionLog>,
804 /// Tracks the last time files were read by the agent, to detect external modifications
805 pub(crate) file_read_times: HashMap<PathBuf, fs::MTime>,
806 /// True if this thread was imported from a shared thread and can be synced.
807 imported: bool,
808 /// If this is a subagent thread, contains context about the parent
809 subagent_context: Option<SubagentContext>,
810 /// Weak references to running subagent threads for cancellation propagation
811 running_subagents: Vec<WeakEntity<Thread>>,
812}
813
814impl Thread {
815 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
816 let image = model.map_or(true, |model| model.supports_images());
817 acp::PromptCapabilities::new()
818 .image(image)
819 .embedded_context(true)
820 }
821
822 pub fn new(
823 project: Entity<Project>,
824 project_context: Entity<ProjectContext>,
825 context_server_registry: Entity<ContextServerRegistry>,
826 templates: Arc<Templates>,
827 model: Option<Arc<dyn LanguageModel>>,
828 cx: &mut Context<Self>,
829 ) -> Self {
830 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
831 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
832 let (prompt_capabilities_tx, prompt_capabilities_rx) =
833 watch::channel(Self::prompt_capabilities(model.as_deref()));
834 Self {
835 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
836 prompt_id: PromptId::new(),
837 updated_at: Utc::now(),
838 title: None,
839 pending_title_generation: None,
840 pending_summary_generation: None,
841 summary: None,
842 messages: Vec::new(),
843 user_store: project.read(cx).user_store(),
844 running_turn: None,
845 has_queued_message: false,
846 pending_message: None,
847 tools: BTreeMap::default(),
848 request_token_usage: HashMap::default(),
849 cumulative_token_usage: TokenUsage::default(),
850 initial_project_snapshot: {
851 let project_snapshot = Self::project_snapshot(project.clone(), cx);
852 cx.foreground_executor()
853 .spawn(async move { Some(project_snapshot.await) })
854 .shared()
855 },
856 context_server_registry,
857 profile_id,
858 project_context,
859 templates,
860 model,
861 summarization_model: None,
862 thinking_enabled: true,
863 prompt_capabilities_tx,
864 prompt_capabilities_rx,
865 project,
866 action_log,
867 file_read_times: HashMap::default(),
868 imported: false,
869 subagent_context: None,
870 running_subagents: Vec::new(),
871 }
872 }
873
874 pub fn new_subagent(
875 project: Entity<Project>,
876 project_context: Entity<ProjectContext>,
877 context_server_registry: Entity<ContextServerRegistry>,
878 templates: Arc<Templates>,
879 model: Arc<dyn LanguageModel>,
880 subagent_context: SubagentContext,
881 parent_tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
882 cx: &mut Context<Self>,
883 ) -> Self {
884 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
885 let action_log = cx.new(|_cx| ActionLog::new(project.clone()));
886 let (prompt_capabilities_tx, prompt_capabilities_rx) =
887 watch::channel(Self::prompt_capabilities(Some(model.as_ref())));
888
889 // Rebind tools that hold thread references to use this subagent's thread
890 // instead of the parent's thread. This is critical for tools like EditFileTool
891 // that make model requests using the thread's ID.
892 let weak_self = cx.weak_entity();
893 let tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>> = parent_tools
894 .into_iter()
895 .map(|(name, tool)| {
896 let rebound = tool.rebind_thread(weak_self.clone()).unwrap_or(tool);
897 (name, rebound)
898 })
899 .collect();
900
901 Self {
902 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
903 prompt_id: PromptId::new(),
904 updated_at: Utc::now(),
905 title: None,
906 pending_title_generation: None,
907 pending_summary_generation: None,
908 summary: None,
909 messages: Vec::new(),
910 user_store: project.read(cx).user_store(),
911 running_turn: None,
912 has_queued_message: false,
913 pending_message: None,
914 tools,
915 request_token_usage: HashMap::default(),
916 cumulative_token_usage: TokenUsage::default(),
917 initial_project_snapshot: Task::ready(None).shared(),
918 context_server_registry,
919 profile_id,
920 project_context,
921 templates,
922 model: Some(model),
923 summarization_model: None,
924 thinking_enabled: true,
925 prompt_capabilities_tx,
926 prompt_capabilities_rx,
927 project,
928 action_log,
929 file_read_times: HashMap::default(),
930 imported: false,
931 subagent_context: Some(subagent_context),
932 running_subagents: Vec::new(),
933 }
934 }
935
936 pub fn id(&self) -> &acp::SessionId {
937 &self.id
938 }
939
940 /// Returns true if this thread was imported from a shared thread.
941 pub fn is_imported(&self) -> bool {
942 self.imported
943 }
944
945 pub fn replay(
946 &mut self,
947 cx: &mut Context<Self>,
948 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
949 let (tx, rx) = mpsc::unbounded();
950 let stream = ThreadEventStream(tx);
951 for message in &self.messages {
952 match message {
953 Message::User(user_message) => stream.send_user_message(user_message),
954 Message::Agent(assistant_message) => {
955 for content in &assistant_message.content {
956 match content {
957 AgentMessageContent::Text(text) => stream.send_text(text),
958 AgentMessageContent::Thinking { text, .. } => {
959 stream.send_thinking(text)
960 }
961 AgentMessageContent::RedactedThinking(_) => {}
962 AgentMessageContent::ToolUse(tool_use) => {
963 self.replay_tool_call(
964 tool_use,
965 assistant_message.tool_results.get(&tool_use.id),
966 &stream,
967 cx,
968 );
969 }
970 }
971 }
972 }
973 Message::Resume => {}
974 }
975 }
976 rx
977 }
978
979 fn replay_tool_call(
980 &self,
981 tool_use: &LanguageModelToolUse,
982 tool_result: Option<&LanguageModelToolResult>,
983 stream: &ThreadEventStream,
984 cx: &mut Context<Self>,
985 ) {
986 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
987 self.context_server_registry
988 .read(cx)
989 .servers()
990 .find_map(|(_, tools)| {
991 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
992 Some(tool.clone())
993 } else {
994 None
995 }
996 })
997 });
998
999 let Some(tool) = tool else {
1000 stream
1001 .0
1002 .unbounded_send(Ok(ThreadEvent::ToolCall(
1003 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
1004 .status(acp::ToolCallStatus::Failed)
1005 .raw_input(tool_use.input.clone()),
1006 )))
1007 .ok();
1008 return;
1009 };
1010
1011 let title = tool.initial_title(tool_use.input.clone(), cx);
1012 let kind = tool.kind();
1013 stream.send_tool_call(
1014 &tool_use.id,
1015 &tool_use.name,
1016 title,
1017 kind,
1018 tool_use.input.clone(),
1019 );
1020
1021 let output = tool_result
1022 .as_ref()
1023 .and_then(|result| result.output.clone());
1024 if let Some(output) = output.clone() {
1025 // For replay, we use a dummy cancellation receiver since the tool already completed
1026 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
1027 let tool_event_stream = ToolCallEventStream::new(
1028 tool_use.id.clone(),
1029 stream.clone(),
1030 Some(self.project.read(cx).fs().clone()),
1031 cancellation_rx,
1032 );
1033 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
1034 .log_err();
1035 }
1036
1037 stream.update_tool_call_fields(
1038 &tool_use.id,
1039 acp::ToolCallUpdateFields::new()
1040 .status(
1041 tool_result
1042 .as_ref()
1043 .map_or(acp::ToolCallStatus::Failed, |result| {
1044 if result.is_error {
1045 acp::ToolCallStatus::Failed
1046 } else {
1047 acp::ToolCallStatus::Completed
1048 }
1049 }),
1050 )
1051 .raw_output(output),
1052 );
1053 }
1054
1055 pub fn from_db(
1056 id: acp::SessionId,
1057 db_thread: DbThread,
1058 project: Entity<Project>,
1059 project_context: Entity<ProjectContext>,
1060 context_server_registry: Entity<ContextServerRegistry>,
1061 templates: Arc<Templates>,
1062 cx: &mut Context<Self>,
1063 ) -> Self {
1064 let profile_id = db_thread
1065 .profile
1066 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
1067
1068 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1069 db_thread
1070 .model
1071 .and_then(|model| {
1072 let model = SelectedModel {
1073 provider: model.provider.clone().into(),
1074 model: model.model.into(),
1075 };
1076 registry.select_model(&model, cx)
1077 })
1078 .or_else(|| registry.default_model())
1079 .map(|model| model.model)
1080 });
1081
1082 if model.is_none() {
1083 model = Self::resolve_profile_model(&profile_id, cx);
1084 }
1085 if model.is_none() {
1086 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
1087 registry.default_model().map(|model| model.model)
1088 });
1089 }
1090
1091 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1092 watch::channel(Self::prompt_capabilities(model.as_deref()));
1093
1094 let action_log = cx.new(|_| ActionLog::new(project.clone()));
1095
1096 Self {
1097 id,
1098 prompt_id: PromptId::new(),
1099 title: if db_thread.title.is_empty() {
1100 None
1101 } else {
1102 Some(db_thread.title.clone())
1103 },
1104 pending_title_generation: None,
1105 pending_summary_generation: None,
1106 summary: db_thread.detailed_summary,
1107 messages: db_thread.messages,
1108 user_store: project.read(cx).user_store(),
1109 running_turn: None,
1110 has_queued_message: false,
1111 pending_message: None,
1112 tools: BTreeMap::default(),
1113 request_token_usage: db_thread.request_token_usage.clone(),
1114 cumulative_token_usage: db_thread.cumulative_token_usage,
1115 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
1116 context_server_registry,
1117 profile_id,
1118 project_context,
1119 templates,
1120 model,
1121 summarization_model: None,
1122 // TODO: Persist this on the `DbThread`.
1123 thinking_enabled: true,
1124 project,
1125 action_log,
1126 updated_at: db_thread.updated_at,
1127 prompt_capabilities_tx,
1128 prompt_capabilities_rx,
1129 file_read_times: HashMap::default(),
1130 imported: db_thread.imported,
1131 subagent_context: None,
1132 running_subagents: Vec::new(),
1133 }
1134 }
1135
1136 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
1137 let initial_project_snapshot = self.initial_project_snapshot.clone();
1138 let mut thread = DbThread {
1139 title: self.title(),
1140 messages: self.messages.clone(),
1141 updated_at: self.updated_at,
1142 detailed_summary: self.summary.clone(),
1143 initial_project_snapshot: None,
1144 cumulative_token_usage: self.cumulative_token_usage,
1145 request_token_usage: self.request_token_usage.clone(),
1146 model: self.model.as_ref().map(|model| DbLanguageModel {
1147 provider: model.provider_id().to_string(),
1148 model: model.name().0.to_string(),
1149 }),
1150 profile: Some(self.profile_id.clone()),
1151 imported: self.imported,
1152 };
1153
1154 cx.background_spawn(async move {
1155 let initial_project_snapshot = initial_project_snapshot.await;
1156 thread.initial_project_snapshot = initial_project_snapshot;
1157 thread
1158 })
1159 }
1160
1161 /// Create a snapshot of the current project state including git information and unsaved buffers.
1162 fn project_snapshot(
1163 project: Entity<Project>,
1164 cx: &mut Context<Self>,
1165 ) -> Task<Arc<ProjectSnapshot>> {
1166 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
1167 cx.spawn(async move |_, _| {
1168 let snapshot = task.await;
1169
1170 Arc::new(ProjectSnapshot {
1171 worktree_snapshots: snapshot.worktree_snapshots,
1172 timestamp: Utc::now(),
1173 })
1174 })
1175 }
1176
1177 pub fn project_context(&self) -> &Entity<ProjectContext> {
1178 &self.project_context
1179 }
1180
1181 pub fn project(&self) -> &Entity<Project> {
1182 &self.project
1183 }
1184
1185 pub fn action_log(&self) -> &Entity<ActionLog> {
1186 &self.action_log
1187 }
1188
1189 pub fn is_empty(&self) -> bool {
1190 self.messages.is_empty() && self.title.is_none()
1191 }
1192
1193 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
1194 self.model.as_ref()
1195 }
1196
1197 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
1198 let old_usage = self.latest_token_usage();
1199 self.model = Some(model);
1200 let new_caps = Self::prompt_capabilities(self.model.as_deref());
1201 let new_usage = self.latest_token_usage();
1202 if old_usage != new_usage {
1203 cx.emit(TokenUsageUpdated(new_usage));
1204 }
1205 self.prompt_capabilities_tx.send(new_caps).log_err();
1206 cx.notify()
1207 }
1208
1209 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1210 self.summarization_model.as_ref()
1211 }
1212
1213 pub fn set_summarization_model(
1214 &mut self,
1215 model: Option<Arc<dyn LanguageModel>>,
1216 cx: &mut Context<Self>,
1217 ) {
1218 self.summarization_model = model;
1219 cx.notify()
1220 }
1221
1222 pub fn thinking_enabled(&self) -> bool {
1223 self.thinking_enabled
1224 }
1225
1226 pub fn set_thinking_enabled(&mut self, enabled: bool, cx: &mut Context<Self>) {
1227 self.thinking_enabled = enabled;
1228 cx.notify();
1229 }
1230
1231 pub fn last_message(&self) -> Option<Message> {
1232 if let Some(message) = self.pending_message.clone() {
1233 Some(Message::Agent(message))
1234 } else {
1235 self.messages.last().cloned()
1236 }
1237 }
1238
1239 pub fn add_default_tools(
1240 &mut self,
1241 environment: Rc<dyn ThreadEnvironment>,
1242 cx: &mut Context<Self>,
1243 ) {
1244 let language_registry = self.project.read(cx).languages().clone();
1245 self.add_tool(CopyPathTool::new(self.project.clone()));
1246 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1247 self.add_tool(DeletePathTool::new(
1248 self.project.clone(),
1249 self.action_log.clone(),
1250 ));
1251 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1252 self.add_tool(EditFileTool::new(
1253 self.project.clone(),
1254 cx.weak_entity(),
1255 language_registry.clone(),
1256 Templates::new(),
1257 ));
1258 self.add_tool(StreamingEditFileTool::new(
1259 self.project.clone(),
1260 cx.weak_entity(),
1261 language_registry,
1262 Templates::new(),
1263 ));
1264 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1265 self.add_tool(FindPathTool::new(self.project.clone()));
1266 self.add_tool(GrepTool::new(self.project.clone()));
1267 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1268 self.add_tool(MovePathTool::new(self.project.clone()));
1269 self.add_tool(NowTool);
1270 self.add_tool(OpenTool::new(self.project.clone()));
1271 self.add_tool(ReadFileTool::new(
1272 cx.weak_entity(),
1273 self.project.clone(),
1274 self.action_log.clone(),
1275 ));
1276 self.add_tool(SaveFileTool::new(self.project.clone()));
1277 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1278 self.add_tool(TerminalTool::new(self.project.clone(), environment));
1279 self.add_tool(ThinkingTool);
1280 self.add_tool(WebSearchTool);
1281
1282 if cx.has_flag::<SubagentsFeatureFlag>() && self.depth() < MAX_SUBAGENT_DEPTH {
1283 self.add_tool(SubagentTool::new(cx.weak_entity(), self.depth()));
1284 }
1285 }
1286
1287 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1288 self.tools.insert(T::name().into(), tool.erase());
1289 }
1290
1291 pub fn remove_tool(&mut self, name: &str) -> bool {
1292 self.tools.remove(name).is_some()
1293 }
1294
1295 pub fn restrict_tools(&mut self, allowed: &collections::HashSet<SharedString>) {
1296 self.tools.retain(|name, _| allowed.contains(name));
1297 }
1298
1299 pub fn profile(&self) -> &AgentProfileId {
1300 &self.profile_id
1301 }
1302
1303 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1304 if self.profile_id == profile_id {
1305 return;
1306 }
1307
1308 self.profile_id = profile_id;
1309
1310 // Swap to the profile's preferred model when available.
1311 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1312 self.set_model(model, cx);
1313 }
1314 }
1315
1316 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1317 for subagent in self.running_subagents.drain(..) {
1318 if let Some(subagent) = subagent.upgrade() {
1319 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1320 }
1321 }
1322
1323 let Some(running_turn) = self.running_turn.take() else {
1324 self.flush_pending_message(cx);
1325 return Task::ready(());
1326 };
1327
1328 let turn_task = running_turn.cancel();
1329
1330 cx.spawn(async move |this, cx| {
1331 turn_task.await;
1332 this.update(cx, |this, cx| {
1333 this.flush_pending_message(cx);
1334 })
1335 .ok();
1336 })
1337 }
1338
1339 pub fn set_has_queued_message(&mut self, has_queued: bool) {
1340 self.has_queued_message = has_queued;
1341 }
1342
1343 pub fn has_queued_message(&self) -> bool {
1344 self.has_queued_message
1345 }
1346
1347 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1348 let Some(last_user_message) = self.last_user_message() else {
1349 return;
1350 };
1351
1352 self.request_token_usage
1353 .insert(last_user_message.id.clone(), update);
1354 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1355 cx.notify();
1356 }
1357
1358 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1359 self.cancel(cx).detach();
1360 // Clear pending message since cancel will try to flush it asynchronously,
1361 // and we don't want that content to be added after we truncate
1362 self.pending_message.take();
1363 let Some(position) = self.messages.iter().position(
1364 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1365 ) else {
1366 return Err(anyhow!("Message not found"));
1367 };
1368
1369 for message in self.messages.drain(position..) {
1370 match message {
1371 Message::User(message) => {
1372 self.request_token_usage.remove(&message.id);
1373 }
1374 Message::Agent(_) | Message::Resume => {}
1375 }
1376 }
1377 self.clear_summary();
1378 cx.notify();
1379 Ok(())
1380 }
1381
1382 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1383 let last_user_message = self.last_user_message()?;
1384 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1385 Some(*tokens)
1386 }
1387
1388 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1389 let usage = self.latest_request_token_usage()?;
1390 let model = self.model.clone()?;
1391 Some(acp_thread::TokenUsage {
1392 max_tokens: model.max_token_count(),
1393 used_tokens: usage.total_tokens(),
1394 input_tokens: usage.input_tokens,
1395 output_tokens: usage.output_tokens,
1396 })
1397 }
1398
1399 /// Get the total input token count as of the message before the given message.
1400 ///
1401 /// Returns `None` if:
1402 /// - `target_id` is the first message (no previous message)
1403 /// - The previous message hasn't received a response yet (no usage data)
1404 /// - `target_id` is not found in the messages
1405 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1406 let mut previous_user_message_id: Option<&UserMessageId> = None;
1407
1408 for message in &self.messages {
1409 if let Message::User(user_msg) = message {
1410 if &user_msg.id == target_id {
1411 let prev_id = previous_user_message_id?;
1412 let usage = self.request_token_usage.get(prev_id)?;
1413 return Some(usage.input_tokens);
1414 }
1415 previous_user_message_id = Some(&user_msg.id);
1416 }
1417 }
1418 None
1419 }
1420
1421 /// Look up the active profile and resolve its preferred model if one is configured.
1422 fn resolve_profile_model(
1423 profile_id: &AgentProfileId,
1424 cx: &mut Context<Self>,
1425 ) -> Option<Arc<dyn LanguageModel>> {
1426 let selection = AgentSettings::get_global(cx)
1427 .profiles
1428 .get(profile_id)?
1429 .default_model
1430 .clone()?;
1431 Self::resolve_model_from_selection(&selection, cx)
1432 }
1433
1434 /// Translate a stored model selection into the configured model from the registry.
1435 fn resolve_model_from_selection(
1436 selection: &LanguageModelSelection,
1437 cx: &mut Context<Self>,
1438 ) -> Option<Arc<dyn LanguageModel>> {
1439 let selected = SelectedModel {
1440 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1441 model: LanguageModelId::from(selection.model.clone()),
1442 };
1443 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1444 registry
1445 .select_model(&selected, cx)
1446 .map(|configured| configured.model)
1447 })
1448 }
1449
1450 pub fn resume(
1451 &mut self,
1452 cx: &mut Context<Self>,
1453 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1454 self.messages.push(Message::Resume);
1455 cx.notify();
1456
1457 log::debug!("Total messages in thread: {}", self.messages.len());
1458 self.run_turn(cx)
1459 }
1460
1461 /// Sending a message results in the model streaming a response, which could include tool calls.
1462 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1463 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1464 pub fn send<T>(
1465 &mut self,
1466 id: UserMessageId,
1467 content: impl IntoIterator<Item = T>,
1468 cx: &mut Context<Self>,
1469 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1470 where
1471 T: Into<UserMessageContent>,
1472 {
1473 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1474 log::debug!("Thread::send content: {:?}", content);
1475
1476 self.messages
1477 .push(Message::User(UserMessage { id, content }));
1478 cx.notify();
1479
1480 self.send_existing(cx)
1481 }
1482
1483 pub fn send_existing(
1484 &mut self,
1485 cx: &mut Context<Self>,
1486 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1487 let model = self.model().context("No language model configured")?;
1488
1489 log::info!("Thread::send called with model: {}", model.name().0);
1490 self.advance_prompt_id();
1491
1492 log::debug!("Total messages in thread: {}", self.messages.len());
1493 self.run_turn(cx)
1494 }
1495
1496 pub fn push_acp_user_block(
1497 &mut self,
1498 id: UserMessageId,
1499 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1500 path_style: PathStyle,
1501 cx: &mut Context<Self>,
1502 ) {
1503 let content = blocks
1504 .into_iter()
1505 .map(|block| UserMessageContent::from_content_block(block, path_style))
1506 .collect::<Vec<_>>();
1507 self.messages
1508 .push(Message::User(UserMessage { id, content }));
1509 cx.notify();
1510 }
1511
1512 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1513 let text = match block {
1514 acp::ContentBlock::Text(text_content) => text_content.text,
1515 acp::ContentBlock::Image(_) => "[image]".to_string(),
1516 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1517 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1518 acp::ContentBlock::Resource(resource) => match resource.resource {
1519 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1520 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1521 _ => "[resource]".to_string(),
1522 },
1523 _ => "[unknown]".to_string(),
1524 };
1525
1526 self.messages.push(Message::Agent(AgentMessage {
1527 content: vec![AgentMessageContent::Text(text)],
1528 ..Default::default()
1529 }));
1530 cx.notify();
1531 }
1532
1533 #[cfg(feature = "eval")]
1534 pub fn proceed(
1535 &mut self,
1536 cx: &mut Context<Self>,
1537 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1538 self.run_turn(cx)
1539 }
1540
1541 fn run_turn(
1542 &mut self,
1543 cx: &mut Context<Self>,
1544 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1545 // Flush the old pending message synchronously before cancelling,
1546 // to avoid a race where the detached cancel task might flush the NEW
1547 // turn's pending message instead of the old one.
1548 self.flush_pending_message(cx);
1549 self.cancel(cx).detach();
1550
1551 let model = self.model.clone().context("No language model configured")?;
1552 let profile = AgentSettings::get_global(cx)
1553 .profiles
1554 .get(&self.profile_id)
1555 .context("Profile not found")?;
1556 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1557 let event_stream = ThreadEventStream(events_tx);
1558 let message_ix = self.messages.len().saturating_sub(1);
1559 self.clear_summary();
1560 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1561 self.running_turn = Some(RunningTurn {
1562 event_stream: event_stream.clone(),
1563 tools: self.enabled_tools(profile, &model, cx),
1564 cancellation_tx,
1565 _task: cx.spawn(async move |this, cx| {
1566 log::debug!("Starting agent turn execution");
1567
1568 let turn_result = Self::run_turn_internal(
1569 &this,
1570 model,
1571 &event_stream,
1572 cancellation_rx.clone(),
1573 cx,
1574 )
1575 .await;
1576
1577 // Check if we were cancelled - if so, cancel() already took running_turn
1578 // and we shouldn't touch it (it might be a NEW turn now)
1579 let was_cancelled = *cancellation_rx.borrow();
1580 if was_cancelled {
1581 log::debug!("Turn was cancelled, skipping cleanup");
1582 return;
1583 }
1584
1585 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1586
1587 match turn_result {
1588 Ok(()) => {
1589 log::debug!("Turn execution completed");
1590 event_stream.send_stop(acp::StopReason::EndTurn);
1591 }
1592 Err(error) => {
1593 log::error!("Turn execution failed: {:?}", error);
1594 match error.downcast::<CompletionError>() {
1595 Ok(CompletionError::Refusal) => {
1596 event_stream.send_stop(acp::StopReason::Refusal);
1597 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1598 }
1599 Ok(CompletionError::MaxTokens) => {
1600 event_stream.send_stop(acp::StopReason::MaxTokens);
1601 }
1602 Ok(CompletionError::Other(error)) | Err(error) => {
1603 event_stream.send_error(error);
1604 }
1605 }
1606 }
1607 }
1608
1609 _ = this.update(cx, |this, _| this.running_turn.take());
1610 }),
1611 });
1612 Ok(events_rx)
1613 }
1614
1615 async fn run_turn_internal(
1616 this: &WeakEntity<Self>,
1617 model: Arc<dyn LanguageModel>,
1618 event_stream: &ThreadEventStream,
1619 mut cancellation_rx: watch::Receiver<bool>,
1620 cx: &mut AsyncApp,
1621 ) -> Result<()> {
1622 let mut attempt = 0;
1623 let mut intent = CompletionIntent::UserPrompt;
1624 loop {
1625 let request =
1626 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1627
1628 telemetry::event!(
1629 "Agent Thread Completion",
1630 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1631 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1632 model = model.telemetry_id(),
1633 model_provider = model.provider_id().to_string(),
1634 attempt
1635 );
1636
1637 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1638
1639 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1640 Ok(events) => (events.fuse(), None),
1641 Err(err) => (stream::empty().boxed().fuse(), Some(err)),
1642 };
1643 let mut tool_results = FuturesUnordered::new();
1644 let mut cancelled = false;
1645 loop {
1646 // Race between getting the first event and cancellation
1647 let first_event = futures::select! {
1648 event = events.next().fuse() => event,
1649 _ = cancellation_rx.changed().fuse() => {
1650 if *cancellation_rx.borrow() {
1651 cancelled = true;
1652 break;
1653 }
1654 continue;
1655 }
1656 };
1657 let Some(first_event) = first_event else {
1658 break;
1659 };
1660
1661 // Collect all immediately available events to process as a batch
1662 let mut batch = vec![first_event];
1663 while let Some(event) = events.next().now_or_never().flatten() {
1664 batch.push(event);
1665 }
1666
1667 // Process the batch in a single update
1668 let batch_result = this.update(cx, |this, cx| {
1669 let mut batch_tool_results = Vec::new();
1670 let mut batch_error = None;
1671
1672 for event in batch {
1673 log::trace!("Received completion event: {:?}", event);
1674 match event {
1675 Ok(event) => {
1676 match this.handle_completion_event(
1677 event,
1678 event_stream,
1679 cancellation_rx.clone(),
1680 cx,
1681 ) {
1682 Ok(Some(task)) => batch_tool_results.push(task),
1683 Ok(None) => {}
1684 Err(err) => {
1685 batch_error = Some(err);
1686 break;
1687 }
1688 }
1689 }
1690 Err(err) => {
1691 batch_error = Some(err.into());
1692 break;
1693 }
1694 }
1695 }
1696
1697 cx.notify();
1698 (batch_tool_results, batch_error)
1699 })?;
1700
1701 tool_results.extend(batch_result.0);
1702 if let Some(err) = batch_result.1 {
1703 error = Some(err.downcast()?);
1704 break;
1705 }
1706 }
1707
1708 // Drop the stream to release the rate limit permit before tool execution.
1709 // The stream holds a semaphore guard that limits concurrent requests.
1710 // Without this, the permit would be held during potentially long-running
1711 // tool execution, which could cause deadlocks when tools spawn subagents
1712 // that need their own permits.
1713 drop(events);
1714
1715 let end_turn = tool_results.is_empty();
1716 while let Some(tool_result) = tool_results.next().await {
1717 log::debug!("Tool finished {:?}", tool_result);
1718
1719 event_stream.update_tool_call_fields(
1720 &tool_result.tool_use_id,
1721 acp::ToolCallUpdateFields::new()
1722 .status(if tool_result.is_error {
1723 acp::ToolCallStatus::Failed
1724 } else {
1725 acp::ToolCallStatus::Completed
1726 })
1727 .raw_output(tool_result.output.clone()),
1728 );
1729 this.update(cx, |this, _cx| {
1730 this.pending_message()
1731 .tool_results
1732 .insert(tool_result.tool_use_id.clone(), tool_result);
1733 })?;
1734 }
1735
1736 this.update(cx, |this, cx| {
1737 this.flush_pending_message(cx);
1738 if this.title.is_none() && this.pending_title_generation.is_none() {
1739 this.generate_title(cx);
1740 }
1741 })?;
1742
1743 if cancelled {
1744 log::debug!("Turn cancelled by user, exiting");
1745 return Ok(());
1746 }
1747
1748 if let Some(error) = error {
1749 attempt += 1;
1750 let retry = this.update(cx, |this, cx| {
1751 let user_store = this.user_store.read(cx);
1752 this.handle_completion_error(error, attempt, user_store.plan())
1753 })??;
1754 let timer = cx.background_executor().timer(retry.duration);
1755 event_stream.send_retry(retry);
1756 timer.await;
1757 this.update(cx, |this, _cx| {
1758 if let Some(Message::Agent(message)) = this.messages.last() {
1759 if message.tool_results.is_empty() {
1760 intent = CompletionIntent::UserPrompt;
1761 this.messages.push(Message::Resume);
1762 }
1763 }
1764 })?;
1765 } else if end_turn {
1766 return Ok(());
1767 } else {
1768 let has_queued = this.update(cx, |this, _| this.has_queued_message())?;
1769 if has_queued {
1770 log::debug!("Queued message found, ending turn at message boundary");
1771 return Ok(());
1772 }
1773 intent = CompletionIntent::ToolResults;
1774 attempt = 0;
1775 }
1776 }
1777 }
1778
1779 fn handle_completion_error(
1780 &mut self,
1781 error: LanguageModelCompletionError,
1782 attempt: u8,
1783 plan: Option<Plan>,
1784 ) -> Result<acp_thread::RetryStatus> {
1785 let Some(model) = self.model.as_ref() else {
1786 return Err(anyhow!(error));
1787 };
1788
1789 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
1790 plan.is_some()
1791 } else {
1792 true
1793 };
1794
1795 if !auto_retry {
1796 return Err(anyhow!(error));
1797 }
1798
1799 let Some(strategy) = Self::retry_strategy_for(&error) else {
1800 return Err(anyhow!(error));
1801 };
1802
1803 let max_attempts = match &strategy {
1804 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
1805 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
1806 };
1807
1808 if attempt > max_attempts {
1809 return Err(anyhow!(error));
1810 }
1811
1812 let delay = match &strategy {
1813 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
1814 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
1815 Duration::from_secs(delay_secs)
1816 }
1817 RetryStrategy::Fixed { delay, .. } => *delay,
1818 };
1819 log::debug!("Retry attempt {attempt} with delay {delay:?}");
1820
1821 Ok(acp_thread::RetryStatus {
1822 last_error: error.to_string().into(),
1823 attempt: attempt as usize,
1824 max_attempts: max_attempts as usize,
1825 started_at: Instant::now(),
1826 duration: delay,
1827 })
1828 }
1829
1830 /// A helper method that's called on every streamed completion event.
1831 /// Returns an optional tool result task, which the main agentic loop will
1832 /// send back to the model when it resolves.
1833 fn handle_completion_event(
1834 &mut self,
1835 event: LanguageModelCompletionEvent,
1836 event_stream: &ThreadEventStream,
1837 cancellation_rx: watch::Receiver<bool>,
1838 cx: &mut Context<Self>,
1839 ) -> Result<Option<Task<LanguageModelToolResult>>> {
1840 log::trace!("Handling streamed completion event: {:?}", event);
1841 use LanguageModelCompletionEvent::*;
1842
1843 match event {
1844 StartMessage { .. } => {
1845 self.flush_pending_message(cx);
1846 self.pending_message = Some(AgentMessage::default());
1847 }
1848 Text(new_text) => self.handle_text_event(new_text, event_stream),
1849 Thinking { text, signature } => {
1850 self.handle_thinking_event(text, signature, event_stream)
1851 }
1852 RedactedThinking { data } => self.handle_redacted_thinking_event(data),
1853 ReasoningDetails(details) => {
1854 let last_message = self.pending_message();
1855 // Store the last non-empty reasoning_details (overwrites earlier ones)
1856 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
1857 if let serde_json::Value::Array(ref arr) = details {
1858 if !arr.is_empty() {
1859 last_message.reasoning_details = Some(details);
1860 }
1861 } else {
1862 last_message.reasoning_details = Some(details);
1863 }
1864 }
1865 ToolUse(tool_use) => {
1866 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
1867 }
1868 ToolUseJsonParseError {
1869 id,
1870 tool_name,
1871 raw_input,
1872 json_parse_error,
1873 } => {
1874 return Ok(Some(Task::ready(
1875 self.handle_tool_use_json_parse_error_event(
1876 id,
1877 tool_name,
1878 raw_input,
1879 json_parse_error,
1880 ),
1881 )));
1882 }
1883 UsageUpdate(usage) => {
1884 telemetry::event!(
1885 "Agent Thread Completion Usage Updated",
1886 thread_id = self.id.to_string(),
1887 prompt_id = self.prompt_id.to_string(),
1888 model = self.model.as_ref().map(|m| m.telemetry_id()),
1889 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
1890 input_tokens = usage.input_tokens,
1891 output_tokens = usage.output_tokens,
1892 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1893 cache_read_input_tokens = usage.cache_read_input_tokens,
1894 );
1895 self.update_token_usage(usage, cx);
1896 }
1897 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
1898 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
1899 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
1900 Started | Queued { .. } => {}
1901 }
1902
1903 Ok(None)
1904 }
1905
1906 fn handle_text_event(&mut self, new_text: String, event_stream: &ThreadEventStream) {
1907 event_stream.send_text(&new_text);
1908
1909 let last_message = self.pending_message();
1910 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
1911 text.push_str(&new_text);
1912 } else {
1913 last_message
1914 .content
1915 .push(AgentMessageContent::Text(new_text));
1916 }
1917 }
1918
1919 fn handle_thinking_event(
1920 &mut self,
1921 new_text: String,
1922 new_signature: Option<String>,
1923 event_stream: &ThreadEventStream,
1924 ) {
1925 event_stream.send_thinking(&new_text);
1926
1927 let last_message = self.pending_message();
1928 if let Some(AgentMessageContent::Thinking { text, signature }) =
1929 last_message.content.last_mut()
1930 {
1931 text.push_str(&new_text);
1932 *signature = new_signature.or(signature.take());
1933 } else {
1934 last_message.content.push(AgentMessageContent::Thinking {
1935 text: new_text,
1936 signature: new_signature,
1937 });
1938 }
1939 }
1940
1941 fn handle_redacted_thinking_event(&mut self, data: String) {
1942 let last_message = self.pending_message();
1943 last_message
1944 .content
1945 .push(AgentMessageContent::RedactedThinking(data));
1946 }
1947
1948 fn handle_tool_use_event(
1949 &mut self,
1950 tool_use: LanguageModelToolUse,
1951 event_stream: &ThreadEventStream,
1952 cancellation_rx: watch::Receiver<bool>,
1953 cx: &mut Context<Self>,
1954 ) -> Option<Task<LanguageModelToolResult>> {
1955 cx.notify();
1956
1957 let tool = self.tool(tool_use.name.as_ref());
1958 let mut title = SharedString::from(&tool_use.name);
1959 let mut kind = acp::ToolKind::Other;
1960 if let Some(tool) = tool.as_ref() {
1961 title = tool.initial_title(tool_use.input.clone(), cx);
1962 kind = tool.kind();
1963 }
1964
1965 // Ensure the last message ends in the current tool use
1966 let last_message = self.pending_message();
1967 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
1968 if let AgentMessageContent::ToolUse(last_tool_use) = content {
1969 if last_tool_use.id == tool_use.id {
1970 *last_tool_use = tool_use.clone();
1971 false
1972 } else {
1973 true
1974 }
1975 } else {
1976 true
1977 }
1978 });
1979
1980 if push_new_tool_use {
1981 event_stream.send_tool_call(
1982 &tool_use.id,
1983 &tool_use.name,
1984 title,
1985 kind,
1986 tool_use.input.clone(),
1987 );
1988 last_message
1989 .content
1990 .push(AgentMessageContent::ToolUse(tool_use.clone()));
1991 } else {
1992 event_stream.update_tool_call_fields(
1993 &tool_use.id,
1994 acp::ToolCallUpdateFields::new()
1995 .title(title.as_str())
1996 .kind(kind)
1997 .raw_input(tool_use.input.clone()),
1998 );
1999 }
2000
2001 if !tool_use.is_input_complete {
2002 return None;
2003 }
2004
2005 let Some(tool) = tool else {
2006 let content = format!("No tool named {} exists", tool_use.name);
2007 return Some(Task::ready(LanguageModelToolResult {
2008 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2009 tool_use_id: tool_use.id,
2010 tool_name: tool_use.name,
2011 is_error: true,
2012 output: None,
2013 }));
2014 };
2015
2016 let fs = self.project.read(cx).fs().clone();
2017 let tool_event_stream = ToolCallEventStream::new(
2018 tool_use.id.clone(),
2019 event_stream.clone(),
2020 Some(fs),
2021 cancellation_rx,
2022 );
2023 tool_event_stream.update_fields(
2024 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
2025 );
2026 let supports_images = self.model().is_some_and(|model| model.supports_images());
2027 let tool_result = tool.run(tool_use.input, tool_event_stream, cx);
2028 log::debug!("Running tool {}", tool_use.name);
2029 Some(cx.foreground_executor().spawn(async move {
2030 let tool_result = tool_result.await.and_then(|output| {
2031 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
2032 && !supports_images
2033 {
2034 return Err(anyhow!(
2035 "Attempted to read an image, but this model doesn't support it.",
2036 ));
2037 }
2038 Ok(output)
2039 });
2040
2041 match tool_result {
2042 Ok(output) => LanguageModelToolResult {
2043 tool_use_id: tool_use.id,
2044 tool_name: tool_use.name,
2045 is_error: false,
2046 content: output.llm_output,
2047 output: Some(output.raw_output),
2048 },
2049 Err(error) => LanguageModelToolResult {
2050 tool_use_id: tool_use.id,
2051 tool_name: tool_use.name,
2052 is_error: true,
2053 content: LanguageModelToolResultContent::Text(Arc::from(error.to_string())),
2054 output: Some(error.to_string().into()),
2055 },
2056 }
2057 }))
2058 }
2059
2060 fn handle_tool_use_json_parse_error_event(
2061 &mut self,
2062 tool_use_id: LanguageModelToolUseId,
2063 tool_name: Arc<str>,
2064 raw_input: Arc<str>,
2065 json_parse_error: String,
2066 ) -> LanguageModelToolResult {
2067 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
2068 LanguageModelToolResult {
2069 tool_use_id,
2070 tool_name,
2071 is_error: true,
2072 content: LanguageModelToolResultContent::Text(tool_output.into()),
2073 output: Some(serde_json::Value::String(raw_input.to_string())),
2074 }
2075 }
2076
2077 pub fn title(&self) -> SharedString {
2078 self.title.clone().unwrap_or("New Thread".into())
2079 }
2080
2081 pub fn is_generating_summary(&self) -> bool {
2082 self.pending_summary_generation.is_some()
2083 }
2084
2085 pub fn is_generating_title(&self) -> bool {
2086 self.pending_title_generation.is_some()
2087 }
2088
2089 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
2090 if let Some(summary) = self.summary.as_ref() {
2091 return Task::ready(Some(summary.clone())).shared();
2092 }
2093 if let Some(task) = self.pending_summary_generation.clone() {
2094 return task;
2095 }
2096 let Some(model) = self.summarization_model.clone() else {
2097 log::error!("No summarization model available");
2098 return Task::ready(None).shared();
2099 };
2100 let mut request = LanguageModelRequest {
2101 intent: Some(CompletionIntent::ThreadContextSummarization),
2102 temperature: AgentSettings::temperature_for_model(&model, cx),
2103 ..Default::default()
2104 };
2105
2106 for message in &self.messages {
2107 request.messages.extend(message.to_request());
2108 }
2109
2110 request.messages.push(LanguageModelRequestMessage {
2111 role: Role::User,
2112 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
2113 cache: false,
2114 reasoning_details: None,
2115 });
2116
2117 let task = cx
2118 .spawn(async move |this, cx| {
2119 let mut summary = String::new();
2120 let mut messages = model.stream_completion(request, cx).await.log_err()?;
2121 while let Some(event) = messages.next().await {
2122 let event = event.log_err()?;
2123 let text = match event {
2124 LanguageModelCompletionEvent::Text(text) => text,
2125 _ => continue,
2126 };
2127
2128 let mut lines = text.lines();
2129 summary.extend(lines.next());
2130 }
2131
2132 log::debug!("Setting summary: {}", summary);
2133 let summary = SharedString::from(summary);
2134
2135 this.update(cx, |this, cx| {
2136 this.summary = Some(summary.clone());
2137 this.pending_summary_generation = None;
2138 cx.notify()
2139 })
2140 .ok()?;
2141
2142 Some(summary)
2143 })
2144 .shared();
2145 self.pending_summary_generation = Some(task.clone());
2146 task
2147 }
2148
2149 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
2150 let Some(model) = self.summarization_model.clone() else {
2151 return;
2152 };
2153
2154 log::debug!(
2155 "Generating title with model: {:?}",
2156 self.summarization_model.as_ref().map(|model| model.name())
2157 );
2158 let mut request = LanguageModelRequest {
2159 intent: Some(CompletionIntent::ThreadSummarization),
2160 temperature: AgentSettings::temperature_for_model(&model, cx),
2161 ..Default::default()
2162 };
2163
2164 for message in &self.messages {
2165 request.messages.extend(message.to_request());
2166 }
2167
2168 request.messages.push(LanguageModelRequestMessage {
2169 role: Role::User,
2170 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2171 cache: false,
2172 reasoning_details: None,
2173 });
2174 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2175 let mut title = String::new();
2176
2177 let generate = async {
2178 let mut messages = model.stream_completion(request, cx).await?;
2179 while let Some(event) = messages.next().await {
2180 let event = event?;
2181 let text = match event {
2182 LanguageModelCompletionEvent::Text(text) => text,
2183 _ => continue,
2184 };
2185
2186 let mut lines = text.lines();
2187 title.extend(lines.next());
2188
2189 // Stop if the LLM generated multiple lines.
2190 if lines.next().is_some() {
2191 break;
2192 }
2193 }
2194 anyhow::Ok(())
2195 };
2196
2197 if generate.await.context("failed to generate title").is_ok() {
2198 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
2199 }
2200 _ = this.update(cx, |this, _| this.pending_title_generation = None);
2201 }));
2202 }
2203
2204 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2205 self.pending_title_generation = None;
2206 if Some(&title) != self.title.as_ref() {
2207 self.title = Some(title);
2208 cx.emit(TitleUpdated);
2209 cx.notify();
2210 }
2211 }
2212
2213 fn clear_summary(&mut self) {
2214 self.summary = None;
2215 self.pending_summary_generation = None;
2216 }
2217
2218 fn last_user_message(&self) -> Option<&UserMessage> {
2219 self.messages
2220 .iter()
2221 .rev()
2222 .find_map(|message| match message {
2223 Message::User(user_message) => Some(user_message),
2224 Message::Agent(_) => None,
2225 Message::Resume => None,
2226 })
2227 }
2228
2229 fn pending_message(&mut self) -> &mut AgentMessage {
2230 self.pending_message.get_or_insert_default()
2231 }
2232
2233 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2234 let Some(mut message) = self.pending_message.take() else {
2235 return;
2236 };
2237
2238 if message.content.is_empty() {
2239 return;
2240 }
2241
2242 for content in &message.content {
2243 let AgentMessageContent::ToolUse(tool_use) = content else {
2244 continue;
2245 };
2246
2247 if !message.tool_results.contains_key(&tool_use.id) {
2248 message.tool_results.insert(
2249 tool_use.id.clone(),
2250 LanguageModelToolResult {
2251 tool_use_id: tool_use.id.clone(),
2252 tool_name: tool_use.name.clone(),
2253 is_error: true,
2254 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2255 output: None,
2256 },
2257 );
2258 }
2259 }
2260
2261 self.messages.push(Message::Agent(message));
2262 self.updated_at = Utc::now();
2263 self.clear_summary();
2264 cx.notify()
2265 }
2266
2267 pub(crate) fn build_completion_request(
2268 &self,
2269 completion_intent: CompletionIntent,
2270 cx: &App,
2271 ) -> Result<LanguageModelRequest> {
2272 let model = self.model().context("No language model configured")?;
2273 let tools = if let Some(turn) = self.running_turn.as_ref() {
2274 turn.tools
2275 .iter()
2276 .filter_map(|(tool_name, tool)| {
2277 log::trace!("Including tool: {}", tool_name);
2278 Some(LanguageModelRequestTool {
2279 name: tool_name.to_string(),
2280 description: tool.description().to_string(),
2281 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2282 })
2283 })
2284 .collect::<Vec<_>>()
2285 } else {
2286 Vec::new()
2287 };
2288
2289 log::debug!("Building completion request");
2290 log::debug!("Completion intent: {:?}", completion_intent);
2291
2292 let available_tools: Vec<_> = self
2293 .running_turn
2294 .as_ref()
2295 .map(|turn| turn.tools.keys().cloned().collect())
2296 .unwrap_or_default();
2297
2298 log::debug!("Request includes {} tools", available_tools.len());
2299 let messages = self.build_request_messages(available_tools, cx);
2300 log::debug!("Request will include {} messages", messages.len());
2301
2302 let request = LanguageModelRequest {
2303 thread_id: Some(self.id.to_string()),
2304 prompt_id: Some(self.prompt_id.to_string()),
2305 intent: Some(completion_intent),
2306 messages,
2307 tools,
2308 tool_choice: None,
2309 stop: Vec::new(),
2310 temperature: AgentSettings::temperature_for_model(model, cx),
2311 thinking_allowed: self.thinking_enabled,
2312 };
2313
2314 log::debug!("Completion request built successfully");
2315 Ok(request)
2316 }
2317
2318 fn enabled_tools(
2319 &self,
2320 profile: &AgentProfileSettings,
2321 model: &Arc<dyn LanguageModel>,
2322 cx: &App,
2323 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2324 fn truncate(tool_name: &SharedString) -> SharedString {
2325 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2326 let mut truncated = tool_name.to_string();
2327 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2328 truncated.into()
2329 } else {
2330 tool_name.clone()
2331 }
2332 }
2333
2334 let use_streaming_edit_tool = false;
2335
2336 let mut tools = self
2337 .tools
2338 .iter()
2339 .filter_map(|(tool_name, tool)| {
2340 // For streaming_edit_file, check profile against "edit_file" since that's what users configure
2341 let profile_tool_name = if tool_name == "streaming_edit_file" {
2342 "edit_file"
2343 } else {
2344 tool_name.as_ref()
2345 };
2346
2347 if tool.supports_provider(&model.provider_id())
2348 && profile.is_tool_enabled(profile_tool_name)
2349 {
2350 match (tool_name.as_ref(), use_streaming_edit_tool) {
2351 ("streaming_edit_file", false) | ("edit_file", true) => None,
2352 ("streaming_edit_file", true) => {
2353 // Expose streaming tool as "edit_file"
2354 Some((SharedString::from("edit_file"), tool.clone()))
2355 }
2356 _ => Some((truncate(tool_name), tool.clone())),
2357 }
2358 } else {
2359 None
2360 }
2361 })
2362 .collect::<BTreeMap<_, _>>();
2363
2364 let mut context_server_tools = Vec::new();
2365 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2366 let mut duplicate_tool_names = HashSet::default();
2367 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2368 for (tool_name, tool) in server_tools {
2369 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2370 let tool_name = truncate(tool_name);
2371 if !seen_tools.insert(tool_name.clone()) {
2372 duplicate_tool_names.insert(tool_name.clone());
2373 }
2374 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2375 }
2376 }
2377 }
2378
2379 // When there are duplicate tool names, disambiguate by prefixing them
2380 // with the server ID. In the rare case there isn't enough space for the
2381 // disambiguated tool name, keep only the last tool with this name.
2382 for (server_id, tool_name, tool) in context_server_tools {
2383 if duplicate_tool_names.contains(&tool_name) {
2384 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2385 if available >= 2 {
2386 let mut disambiguated = server_id.0.to_string();
2387 disambiguated.truncate(available - 1);
2388 disambiguated.push('_');
2389 disambiguated.push_str(&tool_name);
2390 tools.insert(disambiguated.into(), tool.clone());
2391 } else {
2392 tools.insert(tool_name, tool.clone());
2393 }
2394 } else {
2395 tools.insert(tool_name, tool.clone());
2396 }
2397 }
2398
2399 tools
2400 }
2401
2402 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2403 self.running_turn.as_ref()?.tools.get(name).cloned()
2404 }
2405
2406 pub fn has_tool(&self, name: &str) -> bool {
2407 self.running_turn
2408 .as_ref()
2409 .is_some_and(|turn| turn.tools.contains_key(name))
2410 }
2411
2412 #[cfg(any(test, feature = "test-support"))]
2413 pub fn has_registered_tool(&self, name: &str) -> bool {
2414 self.tools.contains_key(name)
2415 }
2416
2417 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2418 self.tools.keys().cloned().collect()
2419 }
2420
2421 pub fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2422 self.running_subagents.push(subagent);
2423 }
2424
2425 pub fn unregister_running_subagent(&mut self, subagent: &WeakEntity<Thread>) {
2426 self.running_subagents
2427 .retain(|s| s.entity_id() != subagent.entity_id());
2428 }
2429
2430 pub fn running_subagent_count(&self) -> usize {
2431 self.running_subagents
2432 .iter()
2433 .filter(|s| s.upgrade().is_some())
2434 .count()
2435 }
2436
2437 pub fn is_subagent(&self) -> bool {
2438 self.subagent_context.is_some()
2439 }
2440
2441 pub fn depth(&self) -> u8 {
2442 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2443 }
2444
2445 pub fn is_turn_complete(&self) -> bool {
2446 self.running_turn.is_none()
2447 }
2448
2449 pub fn submit_user_message(
2450 &mut self,
2451 content: impl Into<String>,
2452 cx: &mut Context<Self>,
2453 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2454 let content = content.into();
2455 self.messages.push(Message::User(UserMessage {
2456 id: UserMessageId::new(),
2457 content: vec![UserMessageContent::Text(content)],
2458 }));
2459 cx.notify();
2460 self.send_existing(cx)
2461 }
2462
2463 pub fn interrupt_for_summary(
2464 &mut self,
2465 cx: &mut Context<Self>,
2466 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2467 let context = self
2468 .subagent_context
2469 .as_ref()
2470 .context("Not a subagent thread")?;
2471 let prompt = context.context_low_prompt.clone();
2472 self.cancel(cx).detach();
2473 self.submit_user_message(prompt, cx)
2474 }
2475
2476 pub fn request_final_summary(
2477 &mut self,
2478 cx: &mut Context<Self>,
2479 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
2480 let context = self
2481 .subagent_context
2482 .as_ref()
2483 .context("Not a subagent thread")?;
2484 let prompt = context.summary_prompt.clone();
2485 self.submit_user_message(prompt, cx)
2486 }
2487
2488 fn build_request_messages(
2489 &self,
2490 available_tools: Vec<SharedString>,
2491 cx: &App,
2492 ) -> Vec<LanguageModelRequestMessage> {
2493 log::trace!(
2494 "Building request messages from {} thread messages",
2495 self.messages.len()
2496 );
2497
2498 let system_prompt = SystemPromptTemplate {
2499 project: self.project_context.read(cx),
2500 available_tools,
2501 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2502 }
2503 .render(&self.templates)
2504 .context("failed to build system prompt")
2505 .expect("Invalid template");
2506 let mut messages = vec![LanguageModelRequestMessage {
2507 role: Role::System,
2508 content: vec![system_prompt.into()],
2509 cache: false,
2510 reasoning_details: None,
2511 }];
2512 for message in &self.messages {
2513 messages.extend(message.to_request());
2514 }
2515
2516 if let Some(last_message) = messages.last_mut() {
2517 last_message.cache = true;
2518 }
2519
2520 if let Some(message) = self.pending_message.as_ref() {
2521 messages.extend(message.to_request());
2522 }
2523
2524 messages
2525 }
2526
2527 pub fn to_markdown(&self) -> String {
2528 let mut markdown = String::new();
2529 for (ix, message) in self.messages.iter().enumerate() {
2530 if ix > 0 {
2531 markdown.push('\n');
2532 }
2533 markdown.push_str(&message.to_markdown());
2534 }
2535
2536 if let Some(message) = self.pending_message.as_ref() {
2537 markdown.push('\n');
2538 markdown.push_str(&message.to_markdown());
2539 }
2540
2541 markdown
2542 }
2543
2544 fn advance_prompt_id(&mut self) {
2545 self.prompt_id = PromptId::new();
2546 }
2547
2548 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2549 use LanguageModelCompletionError::*;
2550 use http_client::StatusCode;
2551
2552 // General strategy here:
2553 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2554 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2555 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2556 match error {
2557 HttpResponseError {
2558 status_code: StatusCode::TOO_MANY_REQUESTS,
2559 ..
2560 } => Some(RetryStrategy::ExponentialBackoff {
2561 initial_delay: BASE_RETRY_DELAY,
2562 max_attempts: MAX_RETRY_ATTEMPTS,
2563 }),
2564 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2565 Some(RetryStrategy::Fixed {
2566 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2567 max_attempts: MAX_RETRY_ATTEMPTS,
2568 })
2569 }
2570 UpstreamProviderError {
2571 status,
2572 retry_after,
2573 ..
2574 } => match *status {
2575 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2576 Some(RetryStrategy::Fixed {
2577 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2578 max_attempts: MAX_RETRY_ATTEMPTS,
2579 })
2580 }
2581 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2582 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2583 // Internal Server Error could be anything, retry up to 3 times.
2584 max_attempts: 3,
2585 }),
2586 status => {
2587 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2588 // but we frequently get them in practice. See https://http.dev/529
2589 if status.as_u16() == 529 {
2590 Some(RetryStrategy::Fixed {
2591 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2592 max_attempts: MAX_RETRY_ATTEMPTS,
2593 })
2594 } else {
2595 Some(RetryStrategy::Fixed {
2596 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2597 max_attempts: 2,
2598 })
2599 }
2600 }
2601 },
2602 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2603 delay: BASE_RETRY_DELAY,
2604 max_attempts: 3,
2605 }),
2606 ApiReadResponseError { .. }
2607 | HttpSend { .. }
2608 | DeserializeResponse { .. }
2609 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2610 delay: BASE_RETRY_DELAY,
2611 max_attempts: 3,
2612 }),
2613 // Retrying these errors definitely shouldn't help.
2614 HttpResponseError {
2615 status_code:
2616 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2617 ..
2618 }
2619 | AuthenticationError { .. }
2620 | PermissionError { .. }
2621 | NoApiKey { .. }
2622 | ApiEndpointNotFound { .. }
2623 | PromptTooLarge { .. } => None,
2624 // These errors might be transient, so retry them
2625 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2626 delay: BASE_RETRY_DELAY,
2627 max_attempts: 1,
2628 }),
2629 // Retry all other 4xx and 5xx errors once.
2630 HttpResponseError { status_code, .. }
2631 if status_code.is_client_error() || status_code.is_server_error() =>
2632 {
2633 Some(RetryStrategy::Fixed {
2634 delay: BASE_RETRY_DELAY,
2635 max_attempts: 3,
2636 })
2637 }
2638 Other(err) if err.is::<language_model::PaymentRequiredError>() => {
2639 // Retrying won't help for Payment Required errors.
2640 None
2641 }
2642 // Conservatively assume that any other errors are non-retryable
2643 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2644 delay: BASE_RETRY_DELAY,
2645 max_attempts: 2,
2646 }),
2647 }
2648 }
2649}
2650
2651struct RunningTurn {
2652 /// Holds the task that handles agent interaction until the end of the turn.
2653 /// Survives across multiple requests as the model performs tool calls and
2654 /// we run tools, report their results.
2655 _task: Task<()>,
2656 /// The current event stream for the running turn. Used to report a final
2657 /// cancellation event if we cancel the turn.
2658 event_stream: ThreadEventStream,
2659 /// The tools that were enabled for this turn.
2660 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2661 /// Sender to signal tool cancellation. When cancel is called, this is
2662 /// set to true so all tools can detect user-initiated cancellation.
2663 cancellation_tx: watch::Sender<bool>,
2664}
2665
2666impl RunningTurn {
2667 fn cancel(mut self) -> Task<()> {
2668 log::debug!("Cancelling in progress turn");
2669 self.cancellation_tx.send(true).ok();
2670 self.event_stream.send_canceled();
2671 self._task
2672 }
2673}
2674
2675pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2676
2677impl EventEmitter<TokenUsageUpdated> for Thread {}
2678
2679pub struct TitleUpdated;
2680
2681impl EventEmitter<TitleUpdated> for Thread {}
2682
2683pub trait AgentTool
2684where
2685 Self: 'static + Sized,
2686{
2687 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
2688 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
2689
2690 fn name() -> &'static str;
2691
2692 fn description() -> SharedString {
2693 let schema = schemars::schema_for!(Self::Input);
2694 SharedString::new(
2695 schema
2696 .get("description")
2697 .and_then(|description| description.as_str())
2698 .unwrap_or_default(),
2699 )
2700 }
2701
2702 fn kind() -> acp::ToolKind;
2703
2704 /// The initial tool title to display. Can be updated during the tool run.
2705 fn initial_title(
2706 &self,
2707 input: Result<Self::Input, serde_json::Value>,
2708 cx: &mut App,
2709 ) -> SharedString;
2710
2711 /// Returns the JSON schema that describes the tool's input.
2712 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
2713 language_model::tool_schema::root_schema_for::<Self::Input>(format)
2714 }
2715
2716 /// Some tools rely on a provider for the underlying billing or other reasons.
2717 /// Allow the tool to check if they are compatible, or should be filtered out.
2718 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
2719 true
2720 }
2721
2722 /// Runs the tool with the provided input.
2723 fn run(
2724 self: Arc<Self>,
2725 input: Self::Input,
2726 event_stream: ToolCallEventStream,
2727 cx: &mut App,
2728 ) -> Task<Result<Self::Output>>;
2729
2730 /// Emits events for a previous execution of the tool.
2731 fn replay(
2732 &self,
2733 _input: Self::Input,
2734 _output: Self::Output,
2735 _event_stream: ToolCallEventStream,
2736 _cx: &mut App,
2737 ) -> Result<()> {
2738 Ok(())
2739 }
2740
2741 fn erase(self) -> Arc<dyn AnyAgentTool> {
2742 Arc::new(Erased(Arc::new(self)))
2743 }
2744
2745 /// Create a new instance of this tool bound to a different thread.
2746 /// This is used when creating subagents, so that tools like EditFileTool
2747 /// that hold a thread reference will use the subagent's thread instead
2748 /// of the parent's thread.
2749 /// Returns None if the tool doesn't need rebinding (most tools).
2750 fn rebind_thread(&self, _new_thread: WeakEntity<Thread>) -> Option<Arc<dyn AnyAgentTool>> {
2751 None
2752 }
2753}
2754
2755pub struct Erased<T>(T);
2756
2757pub struct AgentToolOutput {
2758 pub llm_output: LanguageModelToolResultContent,
2759 pub raw_output: serde_json::Value,
2760}
2761
2762pub trait AnyAgentTool {
2763 fn name(&self) -> SharedString;
2764 fn description(&self) -> SharedString;
2765 fn kind(&self) -> acp::ToolKind;
2766 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
2767 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
2768 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
2769 true
2770 }
2771 fn run(
2772 self: Arc<Self>,
2773 input: serde_json::Value,
2774 event_stream: ToolCallEventStream,
2775 cx: &mut App,
2776 ) -> Task<Result<AgentToolOutput>>;
2777 fn replay(
2778 &self,
2779 input: serde_json::Value,
2780 output: serde_json::Value,
2781 event_stream: ToolCallEventStream,
2782 cx: &mut App,
2783 ) -> Result<()>;
2784 /// Create a new instance of this tool bound to a different thread.
2785 /// This is used when creating subagents, so that tools like EditFileTool
2786 /// that hold a thread reference will use the subagent's thread instead
2787 /// of the parent's thread.
2788 /// Returns None if the tool doesn't need rebinding (most tools).
2789 fn rebind_thread(&self, _new_thread: WeakEntity<Thread>) -> Option<Arc<dyn AnyAgentTool>> {
2790 None
2791 }
2792}
2793
2794impl<T> AnyAgentTool for Erased<Arc<T>>
2795where
2796 T: AgentTool,
2797{
2798 fn name(&self) -> SharedString {
2799 T::name().into()
2800 }
2801
2802 fn description(&self) -> SharedString {
2803 T::description()
2804 }
2805
2806 fn kind(&self) -> agent_client_protocol::ToolKind {
2807 T::kind()
2808 }
2809
2810 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
2811 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
2812 self.0.initial_title(parsed_input, _cx)
2813 }
2814
2815 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
2816 let mut json = serde_json::to_value(T::input_schema(format))?;
2817 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
2818 Ok(json)
2819 }
2820
2821 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
2822 T::supports_provider(provider)
2823 }
2824
2825 fn run(
2826 self: Arc<Self>,
2827 input: serde_json::Value,
2828 event_stream: ToolCallEventStream,
2829 cx: &mut App,
2830 ) -> Task<Result<AgentToolOutput>> {
2831 cx.spawn(async move |cx| {
2832 let input = serde_json::from_value(input)?;
2833 let output = cx
2834 .update(|cx| self.0.clone().run(input, event_stream, cx))
2835 .await?;
2836 let raw_output = serde_json::to_value(&output)?;
2837 Ok(AgentToolOutput {
2838 llm_output: output.into(),
2839 raw_output,
2840 })
2841 })
2842 }
2843
2844 fn replay(
2845 &self,
2846 input: serde_json::Value,
2847 output: serde_json::Value,
2848 event_stream: ToolCallEventStream,
2849 cx: &mut App,
2850 ) -> Result<()> {
2851 let input = serde_json::from_value(input)?;
2852 let output = serde_json::from_value(output)?;
2853 self.0.replay(input, output, event_stream, cx)
2854 }
2855
2856 fn rebind_thread(&self, new_thread: WeakEntity<Thread>) -> Option<Arc<dyn AnyAgentTool>> {
2857 self.0.rebind_thread(new_thread)
2858 }
2859}
2860
2861#[derive(Clone)]
2862struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
2863
2864impl ThreadEventStream {
2865 fn send_user_message(&self, message: &UserMessage) {
2866 self.0
2867 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
2868 .ok();
2869 }
2870
2871 fn send_text(&self, text: &str) {
2872 self.0
2873 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
2874 .ok();
2875 }
2876
2877 fn send_thinking(&self, text: &str) {
2878 self.0
2879 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
2880 .ok();
2881 }
2882
2883 fn send_tool_call(
2884 &self,
2885 id: &LanguageModelToolUseId,
2886 tool_name: &str,
2887 title: SharedString,
2888 kind: acp::ToolKind,
2889 input: serde_json::Value,
2890 ) {
2891 self.0
2892 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
2893 id,
2894 tool_name,
2895 title.to_string(),
2896 kind,
2897 input,
2898 ))))
2899 .ok();
2900 }
2901
2902 fn initial_tool_call(
2903 id: &LanguageModelToolUseId,
2904 tool_name: &str,
2905 title: String,
2906 kind: acp::ToolKind,
2907 input: serde_json::Value,
2908 ) -> acp::ToolCall {
2909 acp::ToolCall::new(id.to_string(), title)
2910 .kind(kind)
2911 .raw_input(input)
2912 .meta(acp_thread::meta_with_tool_name(tool_name))
2913 }
2914
2915 fn update_tool_call_fields(
2916 &self,
2917 tool_use_id: &LanguageModelToolUseId,
2918 fields: acp::ToolCallUpdateFields,
2919 ) {
2920 self.0
2921 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
2922 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields).into(),
2923 )))
2924 .ok();
2925 }
2926
2927 fn send_retry(&self, status: acp_thread::RetryStatus) {
2928 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
2929 }
2930
2931 fn send_stop(&self, reason: acp::StopReason) {
2932 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
2933 }
2934
2935 fn send_canceled(&self) {
2936 self.0
2937 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
2938 .ok();
2939 }
2940
2941 fn send_error(&self, error: impl Into<anyhow::Error>) {
2942 self.0.unbounded_send(Err(error.into())).ok();
2943 }
2944}
2945
2946#[derive(Clone)]
2947pub struct ToolCallEventStream {
2948 tool_use_id: LanguageModelToolUseId,
2949 stream: ThreadEventStream,
2950 fs: Option<Arc<dyn Fs>>,
2951 cancellation_rx: watch::Receiver<bool>,
2952}
2953
2954impl ToolCallEventStream {
2955 #[cfg(any(test, feature = "test-support"))]
2956 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
2957 let (stream, receiver, _cancellation_tx) = Self::test_with_cancellation();
2958 (stream, receiver)
2959 }
2960
2961 #[cfg(any(test, feature = "test-support"))]
2962 pub fn test_with_cancellation() -> (Self, ToolCallEventStreamReceiver, watch::Sender<bool>) {
2963 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
2964 let (cancellation_tx, cancellation_rx) = watch::channel(false);
2965
2966 let stream = ToolCallEventStream::new(
2967 "test_id".into(),
2968 ThreadEventStream(events_tx),
2969 None,
2970 cancellation_rx,
2971 );
2972
2973 (
2974 stream,
2975 ToolCallEventStreamReceiver(events_rx),
2976 cancellation_tx,
2977 )
2978 }
2979
2980 /// Signal cancellation for this event stream. Only available in tests.
2981 #[cfg(any(test, feature = "test-support"))]
2982 pub fn signal_cancellation_with_sender(cancellation_tx: &mut watch::Sender<bool>) {
2983 cancellation_tx.send(true).ok();
2984 }
2985
2986 fn new(
2987 tool_use_id: LanguageModelToolUseId,
2988 stream: ThreadEventStream,
2989 fs: Option<Arc<dyn Fs>>,
2990 cancellation_rx: watch::Receiver<bool>,
2991 ) -> Self {
2992 Self {
2993 tool_use_id,
2994 stream,
2995 fs,
2996 cancellation_rx,
2997 }
2998 }
2999
3000 /// Returns a future that resolves when the user cancels the tool call.
3001 /// Tools should select on this alongside their main work to detect user cancellation.
3002 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
3003 let mut rx = self.cancellation_rx.clone();
3004 async move {
3005 loop {
3006 if *rx.borrow() {
3007 return;
3008 }
3009 if rx.changed().await.is_err() {
3010 // Sender dropped, will never be cancelled
3011 std::future::pending::<()>().await;
3012 }
3013 }
3014 }
3015 }
3016
3017 /// Returns true if the user has cancelled this tool call.
3018 /// This is useful for checking cancellation state after an operation completes,
3019 /// to determine if the completion was due to user cancellation.
3020 pub fn was_cancelled_by_user(&self) -> bool {
3021 *self.cancellation_rx.clone().borrow()
3022 }
3023
3024 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
3025 &self.tool_use_id
3026 }
3027
3028 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
3029 self.stream
3030 .update_tool_call_fields(&self.tool_use_id, fields);
3031 }
3032
3033 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
3034 self.stream
3035 .0
3036 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3037 acp_thread::ToolCallUpdateDiff {
3038 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3039 diff,
3040 }
3041 .into(),
3042 )))
3043 .ok();
3044 }
3045
3046 pub fn update_subagent_thread(&self, thread: Entity<acp_thread::AcpThread>) {
3047 self.stream
3048 .0
3049 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3050 acp_thread::ToolCallUpdateSubagentThread {
3051 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3052 thread,
3053 }
3054 .into(),
3055 )))
3056 .ok();
3057 }
3058
3059 /// Authorize a third-party tool (e.g., MCP tool from a context server).
3060 ///
3061 /// Unlike built-in tools, third-party tools don't support pattern-based permissions.
3062 /// They only support `default_mode` (allow/deny/confirm) per tool.
3063 ///
3064 /// Uses the dropdown authorization flow with two granularities:
3065 /// - "Always for <display_name> MCP tool" → sets `tools.<tool_id>.default_mode = "allow"` or "deny"
3066 /// - "Only this time" → allow/deny once
3067 pub fn authorize_third_party_tool(
3068 &self,
3069 title: impl Into<String>,
3070 tool_id: String,
3071 display_name: String,
3072 cx: &mut App,
3073 ) -> Task<Result<()>> {
3074 let settings = agent_settings::AgentSettings::get_global(cx);
3075
3076 let decision = decide_permission_from_settings(&tool_id, "", &settings);
3077
3078 match decision {
3079 ToolPermissionDecision::Allow => return Task::ready(Ok(())),
3080 ToolPermissionDecision::Deny(reason) => return Task::ready(Err(anyhow!(reason))),
3081 ToolPermissionDecision::Confirm => {}
3082 }
3083
3084 let (response_tx, response_rx) = oneshot::channel();
3085 self.stream
3086 .0
3087 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3088 ToolCallAuthorization {
3089 tool_call: acp::ToolCallUpdate::new(
3090 self.tool_use_id.to_string(),
3091 acp::ToolCallUpdateFields::new().title(title.into()),
3092 ),
3093 options: acp_thread::PermissionOptions::Dropdown(vec![
3094 acp_thread::PermissionOptionChoice {
3095 allow: acp::PermissionOption::new(
3096 acp::PermissionOptionId::new(format!(
3097 "always_allow_mcp:{}",
3098 tool_id
3099 )),
3100 format!("Always for {} MCP tool", display_name),
3101 acp::PermissionOptionKind::AllowAlways,
3102 ),
3103 deny: acp::PermissionOption::new(
3104 acp::PermissionOptionId::new(format!(
3105 "always_deny_mcp:{}",
3106 tool_id
3107 )),
3108 format!("Always for {} MCP tool", display_name),
3109 acp::PermissionOptionKind::RejectAlways,
3110 ),
3111 },
3112 acp_thread::PermissionOptionChoice {
3113 allow: acp::PermissionOption::new(
3114 acp::PermissionOptionId::new("allow"),
3115 "Only this time",
3116 acp::PermissionOptionKind::AllowOnce,
3117 ),
3118 deny: acp::PermissionOption::new(
3119 acp::PermissionOptionId::new("deny"),
3120 "Only this time",
3121 acp::PermissionOptionKind::RejectOnce,
3122 ),
3123 },
3124 ]),
3125 response: response_tx,
3126 context: None,
3127 },
3128 )))
3129 .ok();
3130
3131 let fs = self.fs.clone();
3132 cx.spawn(async move |cx| {
3133 let response_str = response_rx.await?.0.to_string();
3134
3135 if response_str == format!("always_allow_mcp:{}", tool_id) {
3136 if let Some(fs) = fs.clone() {
3137 cx.update(|cx| {
3138 update_settings_file(fs, cx, move |settings, _| {
3139 settings
3140 .agent
3141 .get_or_insert_default()
3142 .set_tool_default_mode(&tool_id, ToolPermissionMode::Allow);
3143 });
3144 });
3145 }
3146 return Ok(());
3147 }
3148 if response_str == format!("always_deny_mcp:{}", tool_id) {
3149 if let Some(fs) = fs.clone() {
3150 cx.update(|cx| {
3151 update_settings_file(fs, cx, move |settings, _| {
3152 settings
3153 .agent
3154 .get_or_insert_default()
3155 .set_tool_default_mode(&tool_id, ToolPermissionMode::Deny);
3156 });
3157 });
3158 }
3159 return Err(anyhow!("Permission to run tool denied by user"));
3160 }
3161
3162 if response_str == "allow" {
3163 return Ok(());
3164 }
3165
3166 Err(anyhow!("Permission to run tool denied by user"))
3167 })
3168 }
3169
3170 pub fn authorize(
3171 &self,
3172 title: impl Into<String>,
3173 context: ToolPermissionContext,
3174 cx: &mut App,
3175 ) -> Task<Result<()>> {
3176 use settings::ToolPermissionMode;
3177
3178 let options = context.build_permission_options();
3179
3180 let (response_tx, response_rx) = oneshot::channel();
3181 self.stream
3182 .0
3183 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3184 ToolCallAuthorization {
3185 tool_call: acp::ToolCallUpdate::new(
3186 self.tool_use_id.to_string(),
3187 acp::ToolCallUpdateFields::new().title(title.into()),
3188 ),
3189 options,
3190 response: response_tx,
3191 context: Some(context),
3192 },
3193 )))
3194 .ok();
3195
3196 let fs = self.fs.clone();
3197 cx.spawn(async move |cx| {
3198 let response_str = response_rx.await?.0.to_string();
3199
3200 // Handle "always allow tool" - e.g., "always_allow:terminal"
3201 if let Some(tool) = response_str.strip_prefix("always_allow:") {
3202 if let Some(fs) = fs.clone() {
3203 let tool = tool.to_string();
3204 cx.update(|cx| {
3205 update_settings_file(fs, cx, move |settings, _| {
3206 settings
3207 .agent
3208 .get_or_insert_default()
3209 .set_tool_default_mode(&tool, ToolPermissionMode::Allow);
3210 });
3211 });
3212 }
3213 return Ok(());
3214 }
3215
3216 // Handle "always deny tool" - e.g., "always_deny:terminal"
3217 if let Some(tool) = response_str.strip_prefix("always_deny:") {
3218 if let Some(fs) = fs.clone() {
3219 let tool = tool.to_string();
3220 cx.update(|cx| {
3221 update_settings_file(fs, cx, move |settings, _| {
3222 settings
3223 .agent
3224 .get_or_insert_default()
3225 .set_tool_default_mode(&tool, ToolPermissionMode::Deny);
3226 });
3227 });
3228 }
3229 return Err(anyhow!("Permission to run tool denied by user"));
3230 }
3231
3232 // Handle "always allow pattern" - e.g., "always_allow_pattern:terminal:^cargo\s"
3233 if response_str.starts_with("always_allow_pattern:") {
3234 let parts: Vec<&str> = response_str.splitn(3, ':').collect();
3235 if parts.len() == 3 {
3236 let pattern_tool_name = parts[1].to_string();
3237 let pattern = parts[2].to_string();
3238 if let Some(fs) = fs.clone() {
3239 cx.update(|cx| {
3240 update_settings_file(fs, cx, move |settings, _| {
3241 settings
3242 .agent
3243 .get_or_insert_default()
3244 .add_tool_allow_pattern(&pattern_tool_name, pattern);
3245 });
3246 });
3247 }
3248 }
3249 return Ok(());
3250 }
3251
3252 // Handle "always deny pattern" - e.g., "always_deny_pattern:terminal:^cargo\s"
3253 if response_str.starts_with("always_deny_pattern:") {
3254 let parts: Vec<&str> = response_str.splitn(3, ':').collect();
3255 if parts.len() == 3 {
3256 let pattern_tool_name = parts[1].to_string();
3257 let pattern = parts[2].to_string();
3258 if let Some(fs) = fs.clone() {
3259 cx.update(|cx| {
3260 update_settings_file(fs, cx, move |settings, _| {
3261 settings
3262 .agent
3263 .get_or_insert_default()
3264 .add_tool_deny_pattern(&pattern_tool_name, pattern);
3265 });
3266 });
3267 }
3268 }
3269 return Err(anyhow!("Permission to run tool denied by user"));
3270 }
3271
3272 // Handle simple "allow" (allow once)
3273 if response_str == "allow" {
3274 return Ok(());
3275 }
3276
3277 // Handle simple "deny" (deny once)
3278 Err(anyhow!("Permission to run tool denied by user"))
3279 })
3280 }
3281}
3282
3283#[cfg(any(test, feature = "test-support"))]
3284pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
3285
3286#[cfg(any(test, feature = "test-support"))]
3287impl ToolCallEventStreamReceiver {
3288 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
3289 let event = self.0.next().await;
3290 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
3291 auth
3292 } else {
3293 panic!("Expected ToolCallAuthorization but got: {:?}", event);
3294 }
3295 }
3296
3297 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
3298 let event = self.0.next().await;
3299 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
3300 update,
3301 )))) = event
3302 {
3303 update.fields
3304 } else {
3305 panic!("Expected update fields but got: {:?}", event);
3306 }
3307 }
3308
3309 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
3310 let event = self.0.next().await;
3311 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
3312 update,
3313 )))) = event
3314 {
3315 update.diff
3316 } else {
3317 panic!("Expected diff but got: {:?}", event);
3318 }
3319 }
3320
3321 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
3322 let event = self.0.next().await;
3323 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
3324 update,
3325 )))) = event
3326 {
3327 update.terminal
3328 } else {
3329 panic!("Expected terminal but got: {:?}", event);
3330 }
3331 }
3332}
3333
3334#[cfg(any(test, feature = "test-support"))]
3335impl std::ops::Deref for ToolCallEventStreamReceiver {
3336 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
3337
3338 fn deref(&self) -> &Self::Target {
3339 &self.0
3340 }
3341}
3342
3343#[cfg(any(test, feature = "test-support"))]
3344impl std::ops::DerefMut for ToolCallEventStreamReceiver {
3345 fn deref_mut(&mut self) -> &mut Self::Target {
3346 &mut self.0
3347 }
3348}
3349
3350impl From<&str> for UserMessageContent {
3351 fn from(text: &str) -> Self {
3352 Self::Text(text.into())
3353 }
3354}
3355
3356impl UserMessageContent {
3357 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
3358 match value {
3359 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
3360 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
3361 acp::ContentBlock::Audio(_) => {
3362 // TODO
3363 Self::Text("[audio]".to_string())
3364 }
3365 acp::ContentBlock::ResourceLink(resource_link) => {
3366 match MentionUri::parse(&resource_link.uri, path_style) {
3367 Ok(uri) => Self::Mention {
3368 uri,
3369 content: String::new(),
3370 },
3371 Err(err) => {
3372 log::error!("Failed to parse mention link: {}", err);
3373 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
3374 }
3375 }
3376 }
3377 acp::ContentBlock::Resource(resource) => match resource.resource {
3378 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
3379 match MentionUri::parse(&resource.uri, path_style) {
3380 Ok(uri) => Self::Mention {
3381 uri,
3382 content: resource.text,
3383 },
3384 Err(err) => {
3385 log::error!("Failed to parse mention link: {}", err);
3386 Self::Text(
3387 MarkdownCodeBlock {
3388 tag: &resource.uri,
3389 text: &resource.text,
3390 }
3391 .to_string(),
3392 )
3393 }
3394 }
3395 }
3396 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
3397 // TODO
3398 Self::Text("[blob]".to_string())
3399 }
3400 other => {
3401 log::warn!("Unexpected content type: {:?}", other);
3402 Self::Text("[unknown]".to_string())
3403 }
3404 },
3405 other => {
3406 log::warn!("Unexpected content type: {:?}", other);
3407 Self::Text("[unknown]".to_string())
3408 }
3409 }
3410 }
3411}
3412
3413impl From<UserMessageContent> for acp::ContentBlock {
3414 fn from(content: UserMessageContent) -> Self {
3415 match content {
3416 UserMessageContent::Text(text) => text.into(),
3417 UserMessageContent::Image(image) => {
3418 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
3419 }
3420 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
3421 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
3422 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
3423 )),
3424 ),
3425 }
3426 }
3427}
3428
3429fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
3430 LanguageModelImage {
3431 source: image_content.data.into(),
3432 size: None,
3433 }
3434}