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, SpawnAgentTool, StreamingEditFileTool,
6 SystemPromptTemplate, Template, Templates, TerminalTool, ToolPermissionDecision,
7 UpdatePlanTool, WebSearchTool, decide_permission_from_settings,
8};
9use acp_thread::{MentionUri, UserMessageId};
10use action_log::ActionLog;
11use feature_flags::{
12 FeatureFlagAppExt as _, StreamingEditFileToolFeatureFlag, UpdatePlanToolFeatureFlag,
13};
14
15use agent_client_protocol as acp;
16use agent_settings::{
17 AgentProfileId, AgentSettings, SUMMARIZE_THREAD_DETAILED_PROMPT, SUMMARIZE_THREAD_PROMPT,
18};
19use anyhow::{Context as _, Result, anyhow};
20use chrono::{DateTime, Utc};
21use client::UserStore;
22use cloud_api_types::Plan;
23use collections::{HashMap, HashSet, IndexMap};
24use fs::Fs;
25use futures::{
26 FutureExt,
27 channel::{mpsc, oneshot},
28 future::Shared,
29 stream::FuturesUnordered,
30};
31use futures::{StreamExt, stream};
32use gpui::{
33 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
34};
35use heck::ToSnakeCase as _;
36use language_model::{
37 CompletionIntent, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
38 LanguageModelId, LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry,
39 LanguageModelRequest, LanguageModelRequestMessage, LanguageModelRequestTool,
40 LanguageModelToolResult, LanguageModelToolResultContent, LanguageModelToolSchemaFormat,
41 LanguageModelToolUse, LanguageModelToolUseId, Role, SelectedModel, Speed, StopReason,
42 TokenUsage, ZED_CLOUD_PROVIDER_ID,
43};
44use project::Project;
45use prompt_store::ProjectContext;
46use schemars::{JsonSchema, Schema};
47use serde::de::DeserializeOwned;
48use serde::{Deserialize, Serialize};
49use settings::{LanguageModelSelection, Settings, ToolPermissionMode, update_settings_file};
50use std::{
51 collections::BTreeMap,
52 marker::PhantomData,
53 ops::RangeInclusive,
54 path::Path,
55 rc::Rc,
56 sync::Arc,
57 time::{Duration, Instant},
58};
59use std::{fmt::Write, path::PathBuf};
60use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
61use uuid::Uuid;
62
63const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
64pub const MAX_TOOL_NAME_LENGTH: usize = 64;
65pub const MAX_SUBAGENT_DEPTH: u8 = 1;
66
67/// Returned when a turn is attempted but no language model has been selected.
68#[derive(Debug)]
69pub struct NoModelConfiguredError;
70
71impl std::fmt::Display for NoModelConfiguredError {
72 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
73 write!(f, "no language model configured")
74 }
75}
76
77impl std::error::Error for NoModelConfiguredError {}
78
79/// Context passed to a subagent thread for lifecycle management
80#[derive(Clone, Debug, Serialize, Deserialize)]
81pub struct SubagentContext {
82 /// ID of the parent thread
83 pub parent_thread_id: acp::SessionId,
84
85 /// Current depth level (0 = root agent, 1 = first-level subagent, etc.)
86 pub depth: u8,
87}
88
89/// The ID of the user prompt that initiated a request.
90///
91/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
92#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
93pub struct PromptId(Arc<str>);
94
95impl PromptId {
96 pub fn new() -> Self {
97 Self(Uuid::new_v4().to_string().into())
98 }
99}
100
101impl std::fmt::Display for PromptId {
102 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
103 write!(f, "{}", self.0)
104 }
105}
106
107pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
108pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
109
110#[derive(Debug, Clone)]
111enum RetryStrategy {
112 ExponentialBackoff {
113 initial_delay: Duration,
114 max_attempts: u8,
115 },
116 Fixed {
117 delay: Duration,
118 max_attempts: u8,
119 },
120}
121
122#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
123pub enum Message {
124 User(UserMessage),
125 Agent(AgentMessage),
126 Resume,
127}
128
129impl Message {
130 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
131 match self {
132 Message::Agent(agent_message) => Some(agent_message),
133 _ => None,
134 }
135 }
136
137 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
138 match self {
139 Message::User(message) => {
140 if message.content.is_empty() {
141 vec![]
142 } else {
143 vec![message.to_request()]
144 }
145 }
146 Message::Agent(message) => message.to_request(),
147 Message::Resume => vec![LanguageModelRequestMessage {
148 role: Role::User,
149 content: vec!["Continue where you left off".into()],
150 cache: false,
151 reasoning_details: None,
152 }],
153 }
154 }
155
156 pub fn to_markdown(&self) -> String {
157 match self {
158 Message::User(message) => message.to_markdown(),
159 Message::Agent(message) => message.to_markdown(),
160 Message::Resume => "[resume]\n".into(),
161 }
162 }
163
164 pub fn role(&self) -> Role {
165 match self {
166 Message::User(_) | Message::Resume => Role::User,
167 Message::Agent(_) => Role::Assistant,
168 }
169 }
170}
171
172#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
173pub struct UserMessage {
174 pub id: UserMessageId,
175 pub content: Vec<UserMessageContent>,
176}
177
178#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
179pub enum UserMessageContent {
180 Text(String),
181 Mention { uri: MentionUri, content: String },
182 Image(LanguageModelImage),
183}
184
185impl UserMessage {
186 pub fn to_markdown(&self) -> String {
187 let mut markdown = String::new();
188
189 for content in &self.content {
190 match content {
191 UserMessageContent::Text(text) => {
192 markdown.push_str(text);
193 markdown.push('\n');
194 }
195 UserMessageContent::Image(_) => {
196 markdown.push_str("<image />\n");
197 }
198 UserMessageContent::Mention { uri, content } => {
199 if !content.is_empty() {
200 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
201 } else {
202 let _ = writeln!(&mut markdown, "{}", uri.as_link());
203 }
204 }
205 }
206 }
207
208 markdown
209 }
210
211 fn to_request(&self) -> LanguageModelRequestMessage {
212 let mut message = LanguageModelRequestMessage {
213 role: Role::User,
214 content: Vec::with_capacity(self.content.len()),
215 cache: false,
216 reasoning_details: None,
217 };
218
219 const OPEN_CONTEXT: &str = "<context>\n\
220 The following items were attached by the user. \
221 They are up-to-date and don't need to be re-read.\n\n";
222
223 const OPEN_FILES_TAG: &str = "<files>";
224 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
225 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
226 const OPEN_SELECTIONS_TAG: &str = "<selections>";
227 const OPEN_THREADS_TAG: &str = "<threads>";
228 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
229 const OPEN_RULES_TAG: &str =
230 "<rules>\nThe user has specified the following rules that should be applied:\n";
231 const OPEN_DIAGNOSTICS_TAG: &str = "<diagnostics>";
232 const OPEN_DIFFS_TAG: &str = "<diffs>";
233 const MERGE_CONFLICT_TAG: &str = "<merge_conflicts>";
234
235 let mut file_context = OPEN_FILES_TAG.to_string();
236 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
237 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
238 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
239 let mut thread_context = OPEN_THREADS_TAG.to_string();
240 let mut fetch_context = OPEN_FETCH_TAG.to_string();
241 let mut rules_context = OPEN_RULES_TAG.to_string();
242 let mut diagnostics_context = OPEN_DIAGNOSTICS_TAG.to_string();
243 let mut diffs_context = OPEN_DIFFS_TAG.to_string();
244 let mut merge_conflict_context = MERGE_CONFLICT_TAG.to_string();
245
246 for chunk in &self.content {
247 let chunk = match chunk {
248 UserMessageContent::Text(text) => {
249 language_model::MessageContent::Text(text.clone())
250 }
251 UserMessageContent::Image(value) => {
252 language_model::MessageContent::Image(value.clone())
253 }
254 UserMessageContent::Mention { uri, content } => {
255 match uri {
256 MentionUri::File { abs_path } => {
257 write!(
258 &mut file_context,
259 "\n{}",
260 MarkdownCodeBlock {
261 tag: &codeblock_tag(abs_path, None),
262 text: &content.to_string(),
263 }
264 )
265 .ok();
266 }
267 MentionUri::PastedImage { .. } => {
268 debug_panic!("pasted image URI should not be used in mention content")
269 }
270 MentionUri::Directory { .. } => {
271 write!(&mut directory_context, "\n{}\n", content).ok();
272 }
273 MentionUri::Symbol {
274 abs_path: path,
275 line_range,
276 ..
277 } => {
278 write!(
279 &mut symbol_context,
280 "\n{}",
281 MarkdownCodeBlock {
282 tag: &codeblock_tag(path, Some(line_range)),
283 text: content
284 }
285 )
286 .ok();
287 }
288 MentionUri::Selection {
289 abs_path: path,
290 line_range,
291 ..
292 } => {
293 write!(
294 &mut selection_context,
295 "\n{}",
296 MarkdownCodeBlock {
297 tag: &codeblock_tag(
298 path.as_deref().unwrap_or("Untitled".as_ref()),
299 Some(line_range)
300 ),
301 text: content
302 }
303 )
304 .ok();
305 }
306 MentionUri::Thread { .. } => {
307 write!(&mut thread_context, "\n{}\n", content).ok();
308 }
309 MentionUri::Rule { .. } => {
310 write!(
311 &mut rules_context,
312 "\n{}",
313 MarkdownCodeBlock {
314 tag: "",
315 text: content
316 }
317 )
318 .ok();
319 }
320 MentionUri::Fetch { url } => {
321 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
322 }
323 MentionUri::Diagnostics { .. } => {
324 write!(&mut diagnostics_context, "\n{}\n", content).ok();
325 }
326 MentionUri::TerminalSelection { .. } => {
327 write!(
328 &mut selection_context,
329 "\n{}",
330 MarkdownCodeBlock {
331 tag: "console",
332 text: content
333 }
334 )
335 .ok();
336 }
337 MentionUri::GitDiff { base_ref } => {
338 write!(
339 &mut diffs_context,
340 "\nBranch diff against {}:\n{}",
341 base_ref,
342 MarkdownCodeBlock {
343 tag: "diff",
344 text: content
345 }
346 )
347 .ok();
348 }
349 MentionUri::MergeConflict { file_path } => {
350 write!(
351 &mut merge_conflict_context,
352 "\nMerge conflict in {}:\n{}",
353 file_path,
354 MarkdownCodeBlock {
355 tag: "diff",
356 text: content
357 }
358 )
359 .ok();
360 }
361 }
362
363 language_model::MessageContent::Text(uri.as_link().to_string())
364 }
365 };
366
367 message.content.push(chunk);
368 }
369
370 let len_before_context = message.content.len();
371
372 if file_context.len() > OPEN_FILES_TAG.len() {
373 file_context.push_str("</files>\n");
374 message
375 .content
376 .push(language_model::MessageContent::Text(file_context));
377 }
378
379 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
380 directory_context.push_str("</directories>\n");
381 message
382 .content
383 .push(language_model::MessageContent::Text(directory_context));
384 }
385
386 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
387 symbol_context.push_str("</symbols>\n");
388 message
389 .content
390 .push(language_model::MessageContent::Text(symbol_context));
391 }
392
393 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
394 selection_context.push_str("</selections>\n");
395 message
396 .content
397 .push(language_model::MessageContent::Text(selection_context));
398 }
399
400 if diffs_context.len() > OPEN_DIFFS_TAG.len() {
401 diffs_context.push_str("</diffs>\n");
402 message
403 .content
404 .push(language_model::MessageContent::Text(diffs_context));
405 }
406
407 if thread_context.len() > OPEN_THREADS_TAG.len() {
408 thread_context.push_str("</threads>\n");
409 message
410 .content
411 .push(language_model::MessageContent::Text(thread_context));
412 }
413
414 if fetch_context.len() > OPEN_FETCH_TAG.len() {
415 fetch_context.push_str("</fetched_urls>\n");
416 message
417 .content
418 .push(language_model::MessageContent::Text(fetch_context));
419 }
420
421 if rules_context.len() > OPEN_RULES_TAG.len() {
422 rules_context.push_str("</user_rules>\n");
423 message
424 .content
425 .push(language_model::MessageContent::Text(rules_context));
426 }
427
428 if diagnostics_context.len() > OPEN_DIAGNOSTICS_TAG.len() {
429 diagnostics_context.push_str("</diagnostics>\n");
430 message
431 .content
432 .push(language_model::MessageContent::Text(diagnostics_context));
433 }
434
435 if merge_conflict_context.len() > MERGE_CONFLICT_TAG.len() {
436 merge_conflict_context.push_str("</merge_conflicts>\n");
437 message
438 .content
439 .push(language_model::MessageContent::Text(merge_conflict_context));
440 }
441
442 if message.content.len() > len_before_context {
443 message.content.insert(
444 len_before_context,
445 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
446 );
447 message
448 .content
449 .push(language_model::MessageContent::Text("</context>".into()));
450 }
451
452 message
453 }
454}
455
456fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
457 let mut result = String::new();
458
459 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
460 let _ = write!(result, "{} ", extension);
461 }
462
463 let _ = write!(result, "{}", full_path.display());
464
465 if let Some(range) = line_range {
466 if range.start() == range.end() {
467 let _ = write!(result, ":{}", range.start() + 1);
468 } else {
469 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
470 }
471 }
472
473 result
474}
475
476impl AgentMessage {
477 pub fn to_markdown(&self) -> String {
478 let mut markdown = String::new();
479
480 for content in &self.content {
481 match content {
482 AgentMessageContent::Text(text) => {
483 markdown.push_str(text);
484 markdown.push('\n');
485 }
486 AgentMessageContent::Thinking { text, .. } => {
487 markdown.push_str("<think>");
488 markdown.push_str(text);
489 markdown.push_str("</think>\n");
490 }
491 AgentMessageContent::RedactedThinking(_) => {
492 markdown.push_str("<redacted_thinking />\n")
493 }
494 AgentMessageContent::ToolUse(tool_use) => {
495 markdown.push_str(&format!(
496 "**Tool Use**: {} (ID: {})\n",
497 tool_use.name, tool_use.id
498 ));
499 markdown.push_str(&format!(
500 "{}\n",
501 MarkdownCodeBlock {
502 tag: "json",
503 text: &format!("{:#}", tool_use.input)
504 }
505 ));
506 }
507 }
508 }
509
510 for tool_result in self.tool_results.values() {
511 markdown.push_str(&format!(
512 "**Tool Result**: {} (ID: {})\n\n",
513 tool_result.tool_name, tool_result.tool_use_id
514 ));
515 if tool_result.is_error {
516 markdown.push_str("**ERROR:**\n");
517 }
518
519 match &tool_result.content {
520 LanguageModelToolResultContent::Text(text) => {
521 writeln!(markdown, "{text}\n").ok();
522 }
523 LanguageModelToolResultContent::Image(_) => {
524 writeln!(markdown, "<image />\n").ok();
525 }
526 }
527
528 if let Some(output) = tool_result.output.as_ref() {
529 writeln!(
530 markdown,
531 "**Debug Output**:\n\n```json\n{}\n```\n",
532 serde_json::to_string_pretty(output).unwrap()
533 )
534 .unwrap();
535 }
536 }
537
538 markdown
539 }
540
541 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
542 let mut assistant_message = LanguageModelRequestMessage {
543 role: Role::Assistant,
544 content: Vec::with_capacity(self.content.len()),
545 cache: false,
546 reasoning_details: self.reasoning_details.clone(),
547 };
548 for chunk in &self.content {
549 match chunk {
550 AgentMessageContent::Text(text) => {
551 assistant_message
552 .content
553 .push(language_model::MessageContent::Text(text.clone()));
554 }
555 AgentMessageContent::Thinking { text, signature } => {
556 assistant_message
557 .content
558 .push(language_model::MessageContent::Thinking {
559 text: text.clone(),
560 signature: signature.clone(),
561 });
562 }
563 AgentMessageContent::RedactedThinking(value) => {
564 assistant_message.content.push(
565 language_model::MessageContent::RedactedThinking(value.clone()),
566 );
567 }
568 AgentMessageContent::ToolUse(tool_use) => {
569 if self.tool_results.contains_key(&tool_use.id) {
570 assistant_message
571 .content
572 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
573 }
574 }
575 };
576 }
577
578 let mut user_message = LanguageModelRequestMessage {
579 role: Role::User,
580 content: Vec::new(),
581 cache: false,
582 reasoning_details: None,
583 };
584
585 for tool_result in self.tool_results.values() {
586 let mut tool_result = tool_result.clone();
587 // Surprisingly, the API fails if we return an empty string here.
588 // It thinks we are sending a tool use without a tool result.
589 if tool_result.content.is_empty() {
590 tool_result.content = "<Tool returned an empty string>".into();
591 }
592 user_message
593 .content
594 .push(language_model::MessageContent::ToolResult(tool_result));
595 }
596
597 let mut messages = Vec::new();
598 if !assistant_message.content.is_empty() {
599 messages.push(assistant_message);
600 }
601 if !user_message.content.is_empty() {
602 messages.push(user_message);
603 }
604 messages
605 }
606}
607
608#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
609pub struct AgentMessage {
610 pub content: Vec<AgentMessageContent>,
611 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
612 pub reasoning_details: Option<serde_json::Value>,
613}
614
615#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
616pub enum AgentMessageContent {
617 Text(String),
618 Thinking {
619 text: String,
620 signature: Option<String>,
621 },
622 RedactedThinking(String),
623 ToolUse(LanguageModelToolUse),
624}
625
626pub trait TerminalHandle {
627 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
628 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
629 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
630 fn kill(&self, cx: &AsyncApp) -> Result<()>;
631 fn was_stopped_by_user(&self, cx: &AsyncApp) -> Result<bool>;
632}
633
634pub trait SubagentHandle {
635 /// The session ID of this subagent thread
636 fn id(&self) -> acp::SessionId;
637 /// The current number of entries in the thread.
638 /// Useful for knowing where the next turn will begin
639 fn num_entries(&self, cx: &App) -> usize;
640 /// Runs a turn for a given message and returns both the response and the index of that output message.
641 fn send(&self, message: String, cx: &AsyncApp) -> Task<Result<String>>;
642}
643
644pub trait ThreadEnvironment {
645 fn create_terminal(
646 &self,
647 command: String,
648 cwd: Option<PathBuf>,
649 output_byte_limit: Option<u64>,
650 cx: &mut AsyncApp,
651 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
652
653 fn create_subagent(&self, label: String, cx: &mut App) -> Result<Rc<dyn SubagentHandle>>;
654
655 fn resume_subagent(
656 &self,
657 _session_id: acp::SessionId,
658 _cx: &mut App,
659 ) -> Result<Rc<dyn SubagentHandle>> {
660 Err(anyhow::anyhow!(
661 "Resuming subagent sessions is not supported"
662 ))
663 }
664}
665
666#[derive(Debug)]
667pub enum ThreadEvent {
668 UserMessage(UserMessage),
669 AgentText(String),
670 AgentThinking(String),
671 ToolCall(acp::ToolCall),
672 ToolCallUpdate(acp_thread::ToolCallUpdate),
673 Plan(acp::Plan),
674 ToolCallAuthorization(ToolCallAuthorization),
675 SubagentSpawned(acp::SessionId),
676 Retry(acp_thread::RetryStatus),
677 Stop(acp::StopReason),
678}
679
680#[derive(Debug)]
681pub struct NewTerminal {
682 pub command: String,
683 pub output_byte_limit: Option<u64>,
684 pub cwd: Option<PathBuf>,
685 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
686}
687
688#[derive(Debug, Clone)]
689pub struct ToolPermissionContext {
690 pub tool_name: String,
691 pub input_values: Vec<String>,
692 pub scope: ToolPermissionScope,
693}
694
695#[derive(Debug, Clone, Copy, PartialEq, Eq)]
696pub enum ToolPermissionScope {
697 ToolInput,
698 SymlinkTarget,
699}
700
701impl ToolPermissionContext {
702 pub fn new(tool_name: impl Into<String>, input_values: Vec<String>) -> Self {
703 Self {
704 tool_name: tool_name.into(),
705 input_values,
706 scope: ToolPermissionScope::ToolInput,
707 }
708 }
709
710 pub fn symlink_target(tool_name: impl Into<String>, target_paths: Vec<String>) -> Self {
711 Self {
712 tool_name: tool_name.into(),
713 input_values: target_paths,
714 scope: ToolPermissionScope::SymlinkTarget,
715 }
716 }
717
718 /// Builds the permission options for this tool context.
719 ///
720 /// This is the canonical source for permission option generation.
721 /// Tests should use this function rather than manually constructing options.
722 ///
723 /// # Shell Compatibility for Terminal Tool
724 ///
725 /// For the terminal tool, "Always allow" options are only shown when the user's
726 /// shell supports POSIX-like command chaining syntax (`&&`, `||`, `;`, `|`).
727 ///
728 /// **Why this matters:** When a user sets up an "always allow" pattern like `^cargo`,
729 /// we need to parse the command to extract all sub-commands and verify that EVERY
730 /// sub-command matches the pattern. Otherwise, an attacker could craft a command like
731 /// `cargo build && rm -rf /` that would bypass the security check.
732 ///
733 /// **Supported shells:** Posix (sh, bash, dash, zsh), Fish 3.0+, PowerShell 7+/Pwsh,
734 /// Cmd, Xonsh, Csh, Tcsh
735 ///
736 /// **Unsupported shells:** Nushell (uses `and`/`or` keywords), Elvish (uses `and`/`or`
737 /// keywords), Rc (Plan 9 shell - no `&&`/`||` operators)
738 ///
739 /// For unsupported shells, we hide the "Always allow" UI options entirely, and if
740 /// the user has `always_allow` rules configured in settings, `ToolPermissionDecision::from_input`
741 /// will return a `Deny` with an explanatory error message.
742 pub fn build_permission_options(&self) -> acp_thread::PermissionOptions {
743 use crate::pattern_extraction::*;
744 use util::shell::ShellKind;
745
746 let tool_name = &self.tool_name;
747 let input_values = &self.input_values;
748 if self.scope == ToolPermissionScope::SymlinkTarget {
749 return acp_thread::PermissionOptions::Flat(vec![
750 acp::PermissionOption::new(
751 acp::PermissionOptionId::new("allow"),
752 "Yes",
753 acp::PermissionOptionKind::AllowOnce,
754 ),
755 acp::PermissionOption::new(
756 acp::PermissionOptionId::new("deny"),
757 "No",
758 acp::PermissionOptionKind::RejectOnce,
759 ),
760 ]);
761 }
762
763 // Check if the user's shell supports POSIX-like command chaining.
764 // See the doc comment above for the full explanation of why this is needed.
765 let shell_supports_always_allow = if tool_name == TerminalTool::NAME {
766 ShellKind::system().supports_posix_chaining()
767 } else {
768 true
769 };
770
771 // For terminal commands with multiple pipeline commands, use DropdownWithPatterns
772 // to let users individually select which command patterns to always allow.
773 if tool_name == TerminalTool::NAME && shell_supports_always_allow {
774 if let Some(input) = input_values.first() {
775 let all_patterns = extract_all_terminal_patterns(input);
776 if all_patterns.len() > 1 {
777 let mut choices = Vec::new();
778 choices.push(acp_thread::PermissionOptionChoice {
779 allow: acp::PermissionOption::new(
780 acp::PermissionOptionId::new(format!("always_allow:{}", tool_name)),
781 format!("Always for {}", tool_name.replace('_', " ")),
782 acp::PermissionOptionKind::AllowAlways,
783 ),
784 deny: acp::PermissionOption::new(
785 acp::PermissionOptionId::new(format!("always_deny:{}", tool_name)),
786 format!("Always for {}", tool_name.replace('_', " ")),
787 acp::PermissionOptionKind::RejectAlways,
788 ),
789 sub_patterns: vec![],
790 });
791 choices.push(acp_thread::PermissionOptionChoice {
792 allow: acp::PermissionOption::new(
793 acp::PermissionOptionId::new("allow"),
794 "Only this time",
795 acp::PermissionOptionKind::AllowOnce,
796 ),
797 deny: acp::PermissionOption::new(
798 acp::PermissionOptionId::new("deny"),
799 "Only this time",
800 acp::PermissionOptionKind::RejectOnce,
801 ),
802 sub_patterns: vec![],
803 });
804 return acp_thread::PermissionOptions::DropdownWithPatterns {
805 choices,
806 patterns: all_patterns,
807 tool_name: tool_name.clone(),
808 };
809 }
810 }
811 }
812
813 let extract_for_value = |value: &str| -> (Option<String>, Option<String>) {
814 if tool_name == TerminalTool::NAME {
815 (
816 extract_terminal_pattern(value),
817 extract_terminal_pattern_display(value),
818 )
819 } else if tool_name == CopyPathTool::NAME
820 || tool_name == MovePathTool::NAME
821 || tool_name == EditFileTool::NAME
822 || tool_name == DeletePathTool::NAME
823 || tool_name == CreateDirectoryTool::NAME
824 || tool_name == SaveFileTool::NAME
825 {
826 (
827 extract_path_pattern(value),
828 extract_path_pattern_display(value),
829 )
830 } else if tool_name == FetchTool::NAME {
831 (
832 extract_url_pattern(value),
833 extract_url_pattern_display(value),
834 )
835 } else {
836 (None, None)
837 }
838 };
839
840 // Extract patterns from all input values. Only offer a pattern-specific
841 // "always allow/deny" button when every value produces the same pattern.
842 let (pattern, pattern_display) = match input_values.as_slice() {
843 [single] => extract_for_value(single),
844 _ => {
845 let mut iter = input_values.iter().map(|v| extract_for_value(v));
846 match iter.next() {
847 Some(first) => {
848 if iter.all(|pair| pair.0 == first.0) {
849 first
850 } else {
851 (None, None)
852 }
853 }
854 None => (None, None),
855 }
856 }
857 };
858
859 let mut choices = Vec::new();
860
861 let mut push_choice =
862 |label: String, allow_id, deny_id, allow_kind, deny_kind, sub_patterns: Vec<String>| {
863 choices.push(acp_thread::PermissionOptionChoice {
864 allow: acp::PermissionOption::new(
865 acp::PermissionOptionId::new(allow_id),
866 label.clone(),
867 allow_kind,
868 ),
869 deny: acp::PermissionOption::new(
870 acp::PermissionOptionId::new(deny_id),
871 label,
872 deny_kind,
873 ),
874 sub_patterns,
875 });
876 };
877
878 if shell_supports_always_allow {
879 push_choice(
880 format!("Always for {}", tool_name.replace('_', " ")),
881 format!("always_allow:{}", tool_name),
882 format!("always_deny:{}", tool_name),
883 acp::PermissionOptionKind::AllowAlways,
884 acp::PermissionOptionKind::RejectAlways,
885 vec![],
886 );
887
888 if let (Some(pattern), Some(display)) = (pattern, pattern_display) {
889 let button_text = if tool_name == TerminalTool::NAME {
890 format!("Always for `{}` commands", display)
891 } else {
892 format!("Always for `{}`", display)
893 };
894 push_choice(
895 button_text,
896 format!("always_allow:{}", tool_name),
897 format!("always_deny:{}", tool_name),
898 acp::PermissionOptionKind::AllowAlways,
899 acp::PermissionOptionKind::RejectAlways,
900 vec![pattern],
901 );
902 }
903 }
904
905 push_choice(
906 "Only this time".to_string(),
907 "allow".to_string(),
908 "deny".to_string(),
909 acp::PermissionOptionKind::AllowOnce,
910 acp::PermissionOptionKind::RejectOnce,
911 vec![],
912 );
913
914 acp_thread::PermissionOptions::Dropdown(choices)
915 }
916}
917
918#[derive(Debug)]
919pub struct ToolCallAuthorization {
920 pub tool_call: acp::ToolCallUpdate,
921 pub options: acp_thread::PermissionOptions,
922 pub response: oneshot::Sender<acp_thread::SelectedPermissionOutcome>,
923 pub context: Option<ToolPermissionContext>,
924}
925
926#[derive(Debug, thiserror::Error)]
927enum CompletionError {
928 #[error("max tokens")]
929 MaxTokens,
930 #[error("refusal")]
931 Refusal,
932 #[error(transparent)]
933 Other(#[from] anyhow::Error),
934}
935
936pub struct Thread {
937 id: acp::SessionId,
938 prompt_id: PromptId,
939 updated_at: DateTime<Utc>,
940 title: Option<SharedString>,
941 pending_title_generation: Option<Task<()>>,
942 title_generation_failed: bool,
943 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
944 summary: Option<SharedString>,
945 messages: Vec<Message>,
946 user_store: Entity<UserStore>,
947 /// Holds the task that handles agent interaction until the end of the turn.
948 /// Survives across multiple requests as the model performs tool calls and
949 /// we run tools, report their results.
950 running_turn: Option<RunningTurn>,
951 /// Flag indicating the UI has a queued message waiting to be sent.
952 /// Used to signal that the turn should end at the next message boundary.
953 has_queued_message: bool,
954 pending_message: Option<AgentMessage>,
955 pub(crate) tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
956 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
957 #[allow(unused)]
958 cumulative_token_usage: TokenUsage,
959 #[allow(unused)]
960 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
961 pub(crate) context_server_registry: Entity<ContextServerRegistry>,
962 profile_id: AgentProfileId,
963 project_context: Entity<ProjectContext>,
964 pub(crate) templates: Arc<Templates>,
965 model: Option<Arc<dyn LanguageModel>>,
966 summarization_model: Option<Arc<dyn LanguageModel>>,
967 thinking_enabled: bool,
968 thinking_effort: Option<String>,
969 speed: Option<Speed>,
970 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
971 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
972 pub(crate) project: Entity<Project>,
973 pub(crate) action_log: Entity<ActionLog>,
974 /// True if this thread was imported from a shared thread and can be synced.
975 imported: bool,
976 /// If this is a subagent thread, contains context about the parent
977 subagent_context: Option<SubagentContext>,
978 /// The user's unsent prompt text, persisted so it can be restored when reloading the thread.
979 draft_prompt: Option<Vec<acp::ContentBlock>>,
980 ui_scroll_position: Option<gpui::ListOffset>,
981 /// Weak references to running subagent threads for cancellation propagation
982 running_subagents: Vec<WeakEntity<Thread>>,
983}
984
985impl Thread {
986 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
987 let image = model.map_or(true, |model| model.supports_images());
988 acp::PromptCapabilities::new()
989 .image(image)
990 .embedded_context(true)
991 }
992
993 pub fn new_subagent(parent_thread: &Entity<Thread>, cx: &mut Context<Self>) -> Self {
994 let project = parent_thread.read(cx).project.clone();
995 let project_context = parent_thread.read(cx).project_context.clone();
996 let context_server_registry = parent_thread.read(cx).context_server_registry.clone();
997 let templates = parent_thread.read(cx).templates.clone();
998 let model = parent_thread.read(cx).model().cloned();
999 let parent_action_log = parent_thread.read(cx).action_log().clone();
1000 let action_log =
1001 cx.new(|_cx| ActionLog::new(project.clone()).with_linked_action_log(parent_action_log));
1002 let mut thread = Self::new_internal(
1003 project,
1004 project_context,
1005 context_server_registry,
1006 templates,
1007 model,
1008 action_log,
1009 cx,
1010 );
1011 thread.subagent_context = Some(SubagentContext {
1012 parent_thread_id: parent_thread.read(cx).id().clone(),
1013 depth: parent_thread.read(cx).depth() + 1,
1014 });
1015 thread.inherit_parent_settings(parent_thread, cx);
1016 thread
1017 }
1018
1019 pub fn new(
1020 project: Entity<Project>,
1021 project_context: Entity<ProjectContext>,
1022 context_server_registry: Entity<ContextServerRegistry>,
1023 templates: Arc<Templates>,
1024 model: Option<Arc<dyn LanguageModel>>,
1025 cx: &mut Context<Self>,
1026 ) -> Self {
1027 Self::new_internal(
1028 project.clone(),
1029 project_context,
1030 context_server_registry,
1031 templates,
1032 model,
1033 cx.new(|_cx| ActionLog::new(project)),
1034 cx,
1035 )
1036 }
1037
1038 fn new_internal(
1039 project: Entity<Project>,
1040 project_context: Entity<ProjectContext>,
1041 context_server_registry: Entity<ContextServerRegistry>,
1042 templates: Arc<Templates>,
1043 model: Option<Arc<dyn LanguageModel>>,
1044 action_log: Entity<ActionLog>,
1045 cx: &mut Context<Self>,
1046 ) -> Self {
1047 let settings = AgentSettings::get_global(cx);
1048 let profile_id = settings.default_profile.clone();
1049 let enable_thinking = settings
1050 .default_model
1051 .as_ref()
1052 .is_some_and(|model| model.enable_thinking);
1053 let thinking_effort = settings
1054 .default_model
1055 .as_ref()
1056 .and_then(|model| model.effort.clone());
1057 let speed = settings
1058 .default_model
1059 .as_ref()
1060 .and_then(|model| model.speed);
1061 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1062 watch::channel(Self::prompt_capabilities(model.as_deref()));
1063 Self {
1064 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
1065 prompt_id: PromptId::new(),
1066 updated_at: Utc::now(),
1067 title: None,
1068 pending_title_generation: None,
1069 title_generation_failed: false,
1070 pending_summary_generation: None,
1071 summary: None,
1072 messages: Vec::new(),
1073 user_store: project.read(cx).user_store(),
1074 running_turn: None,
1075 has_queued_message: false,
1076 pending_message: None,
1077 tools: BTreeMap::default(),
1078 request_token_usage: HashMap::default(),
1079 cumulative_token_usage: TokenUsage::default(),
1080 initial_project_snapshot: {
1081 let project_snapshot = Self::project_snapshot(project.clone(), cx);
1082 cx.foreground_executor()
1083 .spawn(async move { Some(project_snapshot.await) })
1084 .shared()
1085 },
1086 context_server_registry,
1087 profile_id,
1088 project_context,
1089 templates,
1090 model,
1091 summarization_model: None,
1092 thinking_enabled: enable_thinking,
1093 speed,
1094 thinking_effort,
1095 prompt_capabilities_tx,
1096 prompt_capabilities_rx,
1097 project,
1098 action_log,
1099 imported: false,
1100 subagent_context: None,
1101 draft_prompt: None,
1102 ui_scroll_position: None,
1103 running_subagents: Vec::new(),
1104 }
1105 }
1106
1107 /// Copies runtime-mutable settings from the parent thread so that
1108 /// subagents start with the same configuration the user selected.
1109 /// Every property that `set_*` propagates to `running_subagents`
1110 /// should be inherited here as well.
1111 fn inherit_parent_settings(&mut self, parent_thread: &Entity<Thread>, cx: &mut Context<Self>) {
1112 let parent = parent_thread.read(cx);
1113 self.speed = parent.speed;
1114 self.thinking_enabled = parent.thinking_enabled;
1115 self.thinking_effort = parent.thinking_effort.clone();
1116 self.summarization_model = parent.summarization_model.clone();
1117 self.profile_id = parent.profile_id.clone();
1118 }
1119
1120 pub fn id(&self) -> &acp::SessionId {
1121 &self.id
1122 }
1123
1124 /// Returns true if this thread was imported from a shared thread.
1125 pub fn is_imported(&self) -> bool {
1126 self.imported
1127 }
1128
1129 pub fn replay(
1130 &mut self,
1131 cx: &mut Context<Self>,
1132 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
1133 let (tx, rx) = mpsc::unbounded();
1134 let stream = ThreadEventStream(tx);
1135 for message in &self.messages {
1136 match message {
1137 Message::User(user_message) => stream.send_user_message(user_message),
1138 Message::Agent(assistant_message) => {
1139 for content in &assistant_message.content {
1140 match content {
1141 AgentMessageContent::Text(text) => stream.send_text(text),
1142 AgentMessageContent::Thinking { text, .. } => {
1143 stream.send_thinking(text)
1144 }
1145 AgentMessageContent::RedactedThinking(_) => {}
1146 AgentMessageContent::ToolUse(tool_use) => {
1147 self.replay_tool_call(
1148 tool_use,
1149 assistant_message.tool_results.get(&tool_use.id),
1150 &stream,
1151 cx,
1152 );
1153 }
1154 }
1155 }
1156 }
1157 Message::Resume => {}
1158 }
1159 }
1160 rx
1161 }
1162
1163 fn replay_tool_call(
1164 &self,
1165 tool_use: &LanguageModelToolUse,
1166 tool_result: Option<&LanguageModelToolResult>,
1167 stream: &ThreadEventStream,
1168 cx: &mut Context<Self>,
1169 ) {
1170 // Extract saved output and status first, so they're available even if tool is not found
1171 let output = tool_result
1172 .as_ref()
1173 .and_then(|result| result.output.clone());
1174 let status = tool_result
1175 .as_ref()
1176 .map_or(acp::ToolCallStatus::Failed, |result| {
1177 if result.is_error {
1178 acp::ToolCallStatus::Failed
1179 } else {
1180 acp::ToolCallStatus::Completed
1181 }
1182 });
1183
1184 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
1185 self.context_server_registry
1186 .read(cx)
1187 .servers()
1188 .find_map(|(_, tools)| {
1189 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
1190 Some(tool.clone())
1191 } else {
1192 None
1193 }
1194 })
1195 });
1196
1197 let Some(tool) = tool else {
1198 // Tool not found (e.g., MCP server not connected after restart),
1199 // but still display the saved result if available.
1200 // We need to send both ToolCall and ToolCallUpdate events because the UI
1201 // only converts raw_output to displayable content in update_fields, not from_acp.
1202 stream
1203 .0
1204 .unbounded_send(Ok(ThreadEvent::ToolCall(
1205 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
1206 .status(status)
1207 .raw_input(tool_use.input.clone()),
1208 )))
1209 .ok();
1210 stream.update_tool_call_fields(
1211 &tool_use.id,
1212 acp::ToolCallUpdateFields::new()
1213 .status(status)
1214 .raw_output(output),
1215 None,
1216 );
1217 return;
1218 };
1219
1220 let title = tool.initial_title(tool_use.input.clone(), cx);
1221 let kind = tool.kind();
1222 stream.send_tool_call(
1223 &tool_use.id,
1224 &tool_use.name,
1225 title,
1226 kind,
1227 tool_use.input.clone(),
1228 );
1229
1230 if let Some(output) = output.clone() {
1231 // For replay, we use a dummy cancellation receiver since the tool already completed
1232 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
1233 let tool_event_stream = ToolCallEventStream::new(
1234 tool_use.id.clone(),
1235 stream.clone(),
1236 Some(self.project.read(cx).fs().clone()),
1237 cancellation_rx,
1238 );
1239 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
1240 .log_err();
1241 }
1242
1243 stream.update_tool_call_fields(
1244 &tool_use.id,
1245 acp::ToolCallUpdateFields::new()
1246 .status(status)
1247 .raw_output(output),
1248 None,
1249 );
1250 }
1251
1252 pub fn from_db(
1253 id: acp::SessionId,
1254 db_thread: DbThread,
1255 project: Entity<Project>,
1256 project_context: Entity<ProjectContext>,
1257 context_server_registry: Entity<ContextServerRegistry>,
1258 templates: Arc<Templates>,
1259 cx: &mut Context<Self>,
1260 ) -> Self {
1261 let settings = AgentSettings::get_global(cx);
1262 let profile_id = db_thread
1263 .profile
1264 .unwrap_or_else(|| settings.default_profile.clone());
1265
1266 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1267 db_thread
1268 .model
1269 .and_then(|model| {
1270 let model = SelectedModel {
1271 provider: model.provider.clone().into(),
1272 model: model.model.into(),
1273 };
1274 registry.select_model(&model, cx)
1275 })
1276 .or_else(|| registry.default_model())
1277 .map(|model| model.model)
1278 });
1279
1280 if model.is_none() {
1281 model = Self::resolve_profile_model(&profile_id, cx);
1282 }
1283 if model.is_none() {
1284 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
1285 registry.default_model().map(|model| model.model)
1286 });
1287 }
1288
1289 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1290 watch::channel(Self::prompt_capabilities(model.as_deref()));
1291
1292 let action_log = cx.new(|_| ActionLog::new(project.clone()));
1293
1294 Self {
1295 id,
1296 prompt_id: PromptId::new(),
1297 title: if db_thread.title.is_empty() {
1298 None
1299 } else {
1300 Some(db_thread.title.clone())
1301 },
1302 pending_title_generation: None,
1303 title_generation_failed: false,
1304 pending_summary_generation: None,
1305 summary: db_thread.detailed_summary,
1306 messages: db_thread.messages,
1307 user_store: project.read(cx).user_store(),
1308 running_turn: None,
1309 has_queued_message: false,
1310 pending_message: None,
1311 tools: BTreeMap::default(),
1312 request_token_usage: db_thread.request_token_usage.clone(),
1313 cumulative_token_usage: db_thread.cumulative_token_usage,
1314 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
1315 context_server_registry,
1316 profile_id,
1317 project_context,
1318 templates,
1319 model,
1320 summarization_model: None,
1321 thinking_enabled: db_thread.thinking_enabled,
1322 thinking_effort: db_thread.thinking_effort,
1323 speed: db_thread.speed,
1324 project,
1325 action_log,
1326 updated_at: db_thread.updated_at,
1327 prompt_capabilities_tx,
1328 prompt_capabilities_rx,
1329 imported: db_thread.imported,
1330 subagent_context: db_thread.subagent_context,
1331 draft_prompt: db_thread.draft_prompt,
1332 ui_scroll_position: db_thread.ui_scroll_position.map(|sp| gpui::ListOffset {
1333 item_ix: sp.item_ix,
1334 offset_in_item: gpui::px(sp.offset_in_item),
1335 }),
1336 running_subagents: Vec::new(),
1337 }
1338 }
1339
1340 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
1341 let initial_project_snapshot = self.initial_project_snapshot.clone();
1342 let mut thread = DbThread {
1343 title: self.title().unwrap_or_default(),
1344 messages: self.messages.clone(),
1345 updated_at: self.updated_at,
1346 detailed_summary: self.summary.clone(),
1347 initial_project_snapshot: None,
1348 cumulative_token_usage: self.cumulative_token_usage,
1349 request_token_usage: self.request_token_usage.clone(),
1350 model: self.model.as_ref().map(|model| DbLanguageModel {
1351 provider: model.provider_id().to_string(),
1352 model: model.id().0.to_string(),
1353 }),
1354 profile: Some(self.profile_id.clone()),
1355 imported: self.imported,
1356 subagent_context: self.subagent_context.clone(),
1357 speed: self.speed,
1358 thinking_enabled: self.thinking_enabled,
1359 thinking_effort: self.thinking_effort.clone(),
1360 draft_prompt: self.draft_prompt.clone(),
1361 ui_scroll_position: self.ui_scroll_position.map(|lo| {
1362 crate::db::SerializedScrollPosition {
1363 item_ix: lo.item_ix,
1364 offset_in_item: lo.offset_in_item.as_f32(),
1365 }
1366 }),
1367 };
1368
1369 cx.background_spawn(async move {
1370 let initial_project_snapshot = initial_project_snapshot.await;
1371 thread.initial_project_snapshot = initial_project_snapshot;
1372 thread
1373 })
1374 }
1375
1376 /// Create a snapshot of the current project state including git information and unsaved buffers.
1377 fn project_snapshot(
1378 project: Entity<Project>,
1379 cx: &mut Context<Self>,
1380 ) -> Task<Arc<ProjectSnapshot>> {
1381 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
1382 cx.spawn(async move |_, _| {
1383 let snapshot = task.await;
1384
1385 Arc::new(ProjectSnapshot {
1386 worktree_snapshots: snapshot.worktree_snapshots,
1387 timestamp: Utc::now(),
1388 })
1389 })
1390 }
1391
1392 pub fn project_context(&self) -> &Entity<ProjectContext> {
1393 &self.project_context
1394 }
1395
1396 pub fn project(&self) -> &Entity<Project> {
1397 &self.project
1398 }
1399
1400 pub fn action_log(&self) -> &Entity<ActionLog> {
1401 &self.action_log
1402 }
1403
1404 pub fn is_empty(&self) -> bool {
1405 self.messages.is_empty() && self.title.is_none()
1406 }
1407
1408 pub fn draft_prompt(&self) -> Option<&[acp::ContentBlock]> {
1409 self.draft_prompt.as_deref()
1410 }
1411
1412 pub fn set_draft_prompt(&mut self, prompt: Option<Vec<acp::ContentBlock>>) {
1413 self.draft_prompt = prompt;
1414 }
1415
1416 pub fn ui_scroll_position(&self) -> Option<gpui::ListOffset> {
1417 self.ui_scroll_position
1418 }
1419
1420 pub fn set_ui_scroll_position(&mut self, position: Option<gpui::ListOffset>) {
1421 self.ui_scroll_position = position;
1422 }
1423
1424 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
1425 self.model.as_ref()
1426 }
1427
1428 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
1429 let old_usage = self.latest_token_usage();
1430 self.model = Some(model.clone());
1431 let new_caps = Self::prompt_capabilities(self.model.as_deref());
1432 let new_usage = self.latest_token_usage();
1433 if old_usage != new_usage {
1434 cx.emit(TokenUsageUpdated(new_usage));
1435 }
1436 self.prompt_capabilities_tx.send(new_caps).log_err();
1437
1438 for subagent in &self.running_subagents {
1439 subagent
1440 .update(cx, |thread, cx| thread.set_model(model.clone(), cx))
1441 .ok();
1442 }
1443
1444 cx.notify()
1445 }
1446
1447 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1448 self.summarization_model.as_ref()
1449 }
1450
1451 pub fn set_summarization_model(
1452 &mut self,
1453 model: Option<Arc<dyn LanguageModel>>,
1454 cx: &mut Context<Self>,
1455 ) {
1456 self.summarization_model = model.clone();
1457
1458 for subagent in &self.running_subagents {
1459 subagent
1460 .update(cx, |thread, cx| {
1461 thread.set_summarization_model(model.clone(), cx)
1462 })
1463 .ok();
1464 }
1465 cx.notify()
1466 }
1467
1468 pub fn thinking_enabled(&self) -> bool {
1469 self.thinking_enabled
1470 }
1471
1472 pub fn set_thinking_enabled(&mut self, enabled: bool, cx: &mut Context<Self>) {
1473 self.thinking_enabled = enabled;
1474
1475 for subagent in &self.running_subagents {
1476 subagent
1477 .update(cx, |thread, cx| thread.set_thinking_enabled(enabled, cx))
1478 .ok();
1479 }
1480 cx.notify();
1481 }
1482
1483 pub fn thinking_effort(&self) -> Option<&String> {
1484 self.thinking_effort.as_ref()
1485 }
1486
1487 pub fn set_thinking_effort(&mut self, effort: Option<String>, cx: &mut Context<Self>) {
1488 self.thinking_effort = effort.clone();
1489
1490 for subagent in &self.running_subagents {
1491 subagent
1492 .update(cx, |thread, cx| {
1493 thread.set_thinking_effort(effort.clone(), cx)
1494 })
1495 .ok();
1496 }
1497 cx.notify();
1498 }
1499
1500 pub fn speed(&self) -> Option<Speed> {
1501 self.speed
1502 }
1503
1504 pub fn set_speed(&mut self, speed: Speed, cx: &mut Context<Self>) {
1505 self.speed = Some(speed);
1506
1507 for subagent in &self.running_subagents {
1508 subagent
1509 .update(cx, |thread, cx| thread.set_speed(speed, cx))
1510 .ok();
1511 }
1512 cx.notify();
1513 }
1514
1515 pub fn last_message(&self) -> Option<&Message> {
1516 self.messages.last()
1517 }
1518
1519 #[cfg(any(test, feature = "test-support"))]
1520 pub fn last_received_or_pending_message(&self) -> Option<Message> {
1521 if let Some(message) = self.pending_message.clone() {
1522 Some(Message::Agent(message))
1523 } else {
1524 self.messages.last().cloned()
1525 }
1526 }
1527
1528 pub fn add_default_tools(
1529 &mut self,
1530 environment: Rc<dyn ThreadEnvironment>,
1531 cx: &mut Context<Self>,
1532 ) {
1533 // Only update the agent location for the root thread, not for subagents.
1534 let update_agent_location = self.parent_thread_id().is_none();
1535
1536 let language_registry = self.project.read(cx).languages().clone();
1537 self.add_tool(CopyPathTool::new(self.project.clone()));
1538 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1539 self.add_tool(DeletePathTool::new(
1540 self.project.clone(),
1541 self.action_log.clone(),
1542 ));
1543 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1544 self.add_tool(EditFileTool::new(
1545 self.project.clone(),
1546 cx.weak_entity(),
1547 language_registry.clone(),
1548 Templates::new(),
1549 ));
1550 self.add_tool(StreamingEditFileTool::new(
1551 self.project.clone(),
1552 cx.weak_entity(),
1553 self.action_log.clone(),
1554 language_registry,
1555 ));
1556 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1557 self.add_tool(FindPathTool::new(self.project.clone()));
1558 self.add_tool(GrepTool::new(self.project.clone()));
1559 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1560 self.add_tool(MovePathTool::new(self.project.clone()));
1561 self.add_tool(NowTool);
1562 self.add_tool(OpenTool::new(self.project.clone()));
1563 if cx.has_flag::<UpdatePlanToolFeatureFlag>() {
1564 self.add_tool(UpdatePlanTool);
1565 }
1566 self.add_tool(ReadFileTool::new(
1567 self.project.clone(),
1568 self.action_log.clone(),
1569 update_agent_location,
1570 ));
1571 self.add_tool(SaveFileTool::new(self.project.clone()));
1572 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1573 self.add_tool(TerminalTool::new(self.project.clone(), environment.clone()));
1574 self.add_tool(WebSearchTool);
1575
1576 if self.depth() < MAX_SUBAGENT_DEPTH {
1577 self.add_tool(SpawnAgentTool::new(environment));
1578 }
1579 }
1580
1581 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1582 debug_assert!(
1583 !self.tools.contains_key(T::NAME),
1584 "Duplicate tool name: {}",
1585 T::NAME,
1586 );
1587 self.tools.insert(T::NAME.into(), tool.erase());
1588 }
1589
1590 #[cfg(any(test, feature = "test-support"))]
1591 pub fn remove_tool(&mut self, name: &str) -> bool {
1592 self.tools.remove(name).is_some()
1593 }
1594
1595 pub fn profile(&self) -> &AgentProfileId {
1596 &self.profile_id
1597 }
1598
1599 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1600 if self.profile_id == profile_id {
1601 return;
1602 }
1603
1604 self.profile_id = profile_id.clone();
1605
1606 // Swap to the profile's preferred model when available.
1607 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1608 self.set_model(model, cx);
1609 }
1610
1611 for subagent in &self.running_subagents {
1612 subagent
1613 .update(cx, |thread, cx| thread.set_profile(profile_id.clone(), cx))
1614 .ok();
1615 }
1616 }
1617
1618 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1619 for subagent in self.running_subagents.drain(..) {
1620 if let Some(subagent) = subagent.upgrade() {
1621 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1622 }
1623 }
1624
1625 let Some(running_turn) = self.running_turn.take() else {
1626 self.flush_pending_message(cx);
1627 return Task::ready(());
1628 };
1629
1630 let turn_task = running_turn.cancel();
1631
1632 cx.spawn(async move |this, cx| {
1633 turn_task.await;
1634 this.update(cx, |this, cx| {
1635 this.flush_pending_message(cx);
1636 })
1637 .ok();
1638 })
1639 }
1640
1641 pub fn set_has_queued_message(&mut self, has_queued: bool) {
1642 self.has_queued_message = has_queued;
1643 }
1644
1645 pub fn has_queued_message(&self) -> bool {
1646 self.has_queued_message
1647 }
1648
1649 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1650 let Some(last_user_message) = self.last_user_message() else {
1651 return;
1652 };
1653
1654 self.request_token_usage
1655 .insert(last_user_message.id.clone(), update);
1656 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1657 cx.notify();
1658 }
1659
1660 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1661 self.cancel(cx).detach();
1662 // Clear pending message since cancel will try to flush it asynchronously,
1663 // and we don't want that content to be added after we truncate
1664 self.pending_message.take();
1665 let Some(position) = self.messages.iter().position(
1666 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1667 ) else {
1668 return Err(anyhow!("Message not found"));
1669 };
1670
1671 for message in self.messages.drain(position..) {
1672 match message {
1673 Message::User(message) => {
1674 self.request_token_usage.remove(&message.id);
1675 }
1676 Message::Agent(_) | Message::Resume => {}
1677 }
1678 }
1679 self.clear_summary();
1680 cx.notify();
1681 Ok(())
1682 }
1683
1684 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1685 let last_user_message = self.last_user_message()?;
1686 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1687 Some(*tokens)
1688 }
1689
1690 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1691 let usage = self.latest_request_token_usage()?;
1692 let model = self.model.clone()?;
1693 Some(acp_thread::TokenUsage {
1694 max_tokens: model.max_token_count(),
1695 max_output_tokens: model.max_output_tokens(),
1696 used_tokens: usage.total_tokens(),
1697 input_tokens: usage.input_tokens,
1698 output_tokens: usage.output_tokens,
1699 })
1700 }
1701
1702 /// Get the total input token count as of the message before the given message.
1703 ///
1704 /// Returns `None` if:
1705 /// - `target_id` is the first message (no previous message)
1706 /// - The previous message hasn't received a response yet (no usage data)
1707 /// - `target_id` is not found in the messages
1708 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1709 let mut previous_user_message_id: Option<&UserMessageId> = None;
1710
1711 for message in &self.messages {
1712 if let Message::User(user_msg) = message {
1713 if &user_msg.id == target_id {
1714 let prev_id = previous_user_message_id?;
1715 let usage = self.request_token_usage.get(prev_id)?;
1716 return Some(usage.input_tokens);
1717 }
1718 previous_user_message_id = Some(&user_msg.id);
1719 }
1720 }
1721 None
1722 }
1723
1724 /// Look up the active profile and resolve its preferred model if one is configured.
1725 fn resolve_profile_model(
1726 profile_id: &AgentProfileId,
1727 cx: &mut Context<Self>,
1728 ) -> Option<Arc<dyn LanguageModel>> {
1729 let selection = AgentSettings::get_global(cx)
1730 .profiles
1731 .get(profile_id)?
1732 .default_model
1733 .clone()?;
1734 Self::resolve_model_from_selection(&selection, cx)
1735 }
1736
1737 /// Translate a stored model selection into the configured model from the registry.
1738 fn resolve_model_from_selection(
1739 selection: &LanguageModelSelection,
1740 cx: &mut Context<Self>,
1741 ) -> Option<Arc<dyn LanguageModel>> {
1742 let selected = SelectedModel {
1743 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1744 model: LanguageModelId::from(selection.model.clone()),
1745 };
1746 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1747 registry
1748 .select_model(&selected, cx)
1749 .map(|configured| configured.model)
1750 })
1751 }
1752
1753 pub fn resume(
1754 &mut self,
1755 cx: &mut Context<Self>,
1756 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1757 self.messages.push(Message::Resume);
1758 cx.notify();
1759
1760 log::debug!("Total messages in thread: {}", self.messages.len());
1761 self.run_turn(cx)
1762 }
1763
1764 /// Sending a message results in the model streaming a response, which could include tool calls.
1765 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1766 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1767 pub fn send<T>(
1768 &mut self,
1769 id: UserMessageId,
1770 content: impl IntoIterator<Item = T>,
1771 cx: &mut Context<Self>,
1772 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1773 where
1774 T: Into<UserMessageContent>,
1775 {
1776 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1777 log::debug!("Thread::send content: {:?}", content);
1778
1779 self.messages
1780 .push(Message::User(UserMessage { id, content }));
1781 cx.notify();
1782
1783 self.send_existing(cx)
1784 }
1785
1786 pub fn send_existing(
1787 &mut self,
1788 cx: &mut Context<Self>,
1789 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1790 let model = self
1791 .model()
1792 .ok_or_else(|| anyhow!(NoModelConfiguredError))?;
1793
1794 log::info!("Thread::send called with model: {}", model.name().0);
1795 self.advance_prompt_id();
1796
1797 log::debug!("Total messages in thread: {}", self.messages.len());
1798 self.run_turn(cx)
1799 }
1800
1801 pub fn push_acp_user_block(
1802 &mut self,
1803 id: UserMessageId,
1804 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1805 path_style: PathStyle,
1806 cx: &mut Context<Self>,
1807 ) {
1808 let content = blocks
1809 .into_iter()
1810 .map(|block| UserMessageContent::from_content_block(block, path_style))
1811 .collect::<Vec<_>>();
1812 self.messages
1813 .push(Message::User(UserMessage { id, content }));
1814 cx.notify();
1815 }
1816
1817 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1818 let text = match block {
1819 acp::ContentBlock::Text(text_content) => text_content.text,
1820 acp::ContentBlock::Image(_) => "[image]".to_string(),
1821 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1822 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1823 acp::ContentBlock::Resource(resource) => match resource.resource {
1824 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1825 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1826 _ => "[resource]".to_string(),
1827 },
1828 _ => "[unknown]".to_string(),
1829 };
1830
1831 self.messages.push(Message::Agent(AgentMessage {
1832 content: vec![AgentMessageContent::Text(text)],
1833 ..Default::default()
1834 }));
1835 cx.notify();
1836 }
1837
1838 fn run_turn(
1839 &mut self,
1840 cx: &mut Context<Self>,
1841 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1842 // Flush the old pending message synchronously before cancelling,
1843 // to avoid a race where the detached cancel task might flush the NEW
1844 // turn's pending message instead of the old one.
1845 self.flush_pending_message(cx);
1846 self.cancel(cx).detach();
1847
1848 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1849 let event_stream = ThreadEventStream(events_tx);
1850 let message_ix = self.messages.len().saturating_sub(1);
1851 self.clear_summary();
1852 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1853 self.running_turn = Some(RunningTurn {
1854 event_stream: event_stream.clone(),
1855 tools: self.enabled_tools(cx),
1856 cancellation_tx,
1857 streaming_tool_inputs: HashMap::default(),
1858 _task: cx.spawn(async move |this, cx| {
1859 log::debug!("Starting agent turn execution");
1860
1861 let turn_result =
1862 Self::run_turn_internal(&this, &event_stream, cancellation_rx.clone(), cx)
1863 .await;
1864
1865 // Check if we were cancelled - if so, cancel() already took running_turn
1866 // and we shouldn't touch it (it might be a NEW turn now)
1867 let was_cancelled = *cancellation_rx.borrow();
1868 if was_cancelled {
1869 log::debug!("Turn was cancelled, skipping cleanup");
1870 return;
1871 }
1872
1873 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1874
1875 match turn_result {
1876 Ok(()) => {
1877 log::debug!("Turn execution completed");
1878 event_stream.send_stop(acp::StopReason::EndTurn);
1879 }
1880 Err(error) => {
1881 log::error!("Turn execution failed: {:?}", error);
1882 match error.downcast::<CompletionError>() {
1883 Ok(CompletionError::Refusal) => {
1884 event_stream.send_stop(acp::StopReason::Refusal);
1885 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1886 }
1887 Ok(CompletionError::MaxTokens) => {
1888 event_stream.send_stop(acp::StopReason::MaxTokens);
1889 }
1890 Ok(CompletionError::Other(error)) | Err(error) => {
1891 event_stream.send_error(error);
1892 }
1893 }
1894 }
1895 }
1896
1897 _ = this.update(cx, |this, _| this.running_turn.take());
1898 }),
1899 });
1900 Ok(events_rx)
1901 }
1902
1903 async fn run_turn_internal(
1904 this: &WeakEntity<Self>,
1905 event_stream: &ThreadEventStream,
1906 mut cancellation_rx: watch::Receiver<bool>,
1907 cx: &mut AsyncApp,
1908 ) -> Result<()> {
1909 let mut attempt = 0;
1910 let mut intent = CompletionIntent::UserPrompt;
1911 loop {
1912 // Re-read the model and refresh tools on each iteration so that
1913 // mid-turn changes (e.g. the user switches model, toggles tools,
1914 // or changes profile) take effect between tool-call rounds.
1915 let (model, request) = this.update(cx, |this, cx| {
1916 let model = this
1917 .model
1918 .clone()
1919 .ok_or_else(|| anyhow!(NoModelConfiguredError))?;
1920 this.refresh_turn_tools(cx);
1921 let request = this.build_completion_request(intent, cx)?;
1922 anyhow::Ok((model, request))
1923 })??;
1924
1925 telemetry::event!(
1926 "Agent Thread Completion",
1927 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1928 parent_thread_id = this.read_with(cx, |this, _| this
1929 .parent_thread_id()
1930 .map(|id| id.to_string()))?,
1931 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1932 model = model.telemetry_id(),
1933 model_provider = model.provider_id().to_string(),
1934 attempt
1935 );
1936
1937 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1938
1939 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1940 Ok(events) => (events.fuse(), None),
1941 Err(err) => (stream::empty().boxed().fuse(), Some(err)),
1942 };
1943 let mut tool_results: FuturesUnordered<Task<LanguageModelToolResult>> =
1944 FuturesUnordered::new();
1945 let mut early_tool_results: Vec<LanguageModelToolResult> = Vec::new();
1946 let mut cancelled = false;
1947 loop {
1948 // Race between getting the first event, tool completion, and cancellation.
1949 let first_event = futures::select! {
1950 event = events.next().fuse() => event,
1951 tool_result = futures::StreamExt::select_next_some(&mut tool_results) => {
1952 let is_error = tool_result.is_error;
1953 let is_still_streaming = this
1954 .read_with(cx, |this, _cx| {
1955 this.running_turn
1956 .as_ref()
1957 .and_then(|turn| turn.streaming_tool_inputs.get(&tool_result.tool_use_id))
1958 .map_or(false, |inputs| !inputs.has_received_final())
1959 })
1960 .unwrap_or(false);
1961
1962 early_tool_results.push(tool_result);
1963
1964 // Only break if the tool errored and we are still
1965 // streaming the input of the tool. If the tool errored
1966 // but we are no longer streaming its input (i.e. there
1967 // are parallel tool calls) we want to continue
1968 // processing those tool inputs.
1969 if is_error && is_still_streaming {
1970 break;
1971 }
1972 continue;
1973 }
1974 _ = cancellation_rx.changed().fuse() => {
1975 if *cancellation_rx.borrow() {
1976 cancelled = true;
1977 break;
1978 }
1979 continue;
1980 }
1981 };
1982 let Some(first_event) = first_event else {
1983 break;
1984 };
1985
1986 // Collect all immediately available events to process as a batch
1987 let mut batch = vec![first_event];
1988 while let Some(event) = events.next().now_or_never().flatten() {
1989 batch.push(event);
1990 }
1991
1992 // Process the batch in a single update
1993 let batch_result = this.update(cx, |this, cx| {
1994 let mut batch_tool_results = Vec::new();
1995 let mut batch_error = None;
1996
1997 for event in batch {
1998 log::trace!("Received completion event: {:?}", event);
1999 match event {
2000 Ok(event) => {
2001 match this.handle_completion_event(
2002 event,
2003 event_stream,
2004 cancellation_rx.clone(),
2005 cx,
2006 ) {
2007 Ok(Some(task)) => batch_tool_results.push(task),
2008 Ok(None) => {}
2009 Err(err) => {
2010 batch_error = Some(err);
2011 break;
2012 }
2013 }
2014 }
2015 Err(err) => {
2016 batch_error = Some(err.into());
2017 break;
2018 }
2019 }
2020 }
2021
2022 cx.notify();
2023 (batch_tool_results, batch_error)
2024 })?;
2025
2026 tool_results.extend(batch_result.0);
2027 if let Some(err) = batch_result.1 {
2028 error = Some(err.downcast()?);
2029 break;
2030 }
2031 }
2032
2033 // Drop the stream to release the rate limit permit before tool execution.
2034 // The stream holds a semaphore guard that limits concurrent requests.
2035 // Without this, the permit would be held during potentially long-running
2036 // tool execution, which could cause deadlocks when tools spawn subagents
2037 // that need their own permits.
2038 drop(events);
2039
2040 // Drop streaming tool input senders that never received their final input.
2041 // This prevents deadlock when the LLM stream ends (e.g. because of an error)
2042 // before sending a tool use with `is_input_complete: true`.
2043 this.update(cx, |this, _cx| {
2044 if let Some(running_turn) = this.running_turn.as_mut() {
2045 if running_turn.streaming_tool_inputs.is_empty() {
2046 return;
2047 }
2048 log::warn!("Dropping partial tool inputs because the stream ended");
2049 running_turn.streaming_tool_inputs.drain();
2050 }
2051 })?;
2052
2053 let end_turn = tool_results.is_empty() && early_tool_results.is_empty();
2054
2055 for tool_result in early_tool_results {
2056 Self::process_tool_result(this, event_stream, cx, tool_result)?;
2057 }
2058 while let Some(tool_result) = tool_results.next().await {
2059 Self::process_tool_result(this, event_stream, cx, tool_result)?;
2060 }
2061
2062 this.update(cx, |this, cx| {
2063 this.flush_pending_message(cx);
2064 if this.title.is_none() && this.pending_title_generation.is_none() {
2065 this.generate_title(cx);
2066 }
2067 })?;
2068
2069 if cancelled {
2070 log::debug!("Turn cancelled by user, exiting");
2071 return Ok(());
2072 }
2073
2074 if let Some(error) = error {
2075 attempt += 1;
2076 let retry = this.update(cx, |this, cx| {
2077 let user_store = this.user_store.read(cx);
2078 this.handle_completion_error(error, attempt, user_store.plan())
2079 })??;
2080 let timer = cx.background_executor().timer(retry.duration);
2081 event_stream.send_retry(retry);
2082 futures::select! {
2083 _ = timer.fuse() => {}
2084 _ = cancellation_rx.changed().fuse() => {
2085 if *cancellation_rx.borrow() {
2086 log::debug!("Turn cancelled during retry delay, exiting");
2087 return Ok(());
2088 }
2089 }
2090 }
2091 this.update(cx, |this, _cx| {
2092 if let Some(Message::Agent(message)) = this.messages.last() {
2093 if message.tool_results.is_empty() {
2094 intent = CompletionIntent::UserPrompt;
2095 this.messages.push(Message::Resume);
2096 }
2097 }
2098 })?;
2099 } else if end_turn {
2100 return Ok(());
2101 } else {
2102 let has_queued = this.update(cx, |this, _| this.has_queued_message())?;
2103 if has_queued {
2104 log::debug!("Queued message found, ending turn at message boundary");
2105 return Ok(());
2106 }
2107 intent = CompletionIntent::ToolResults;
2108 attempt = 0;
2109 }
2110 }
2111 }
2112
2113 fn process_tool_result(
2114 this: &WeakEntity<Thread>,
2115 event_stream: &ThreadEventStream,
2116 cx: &mut AsyncApp,
2117 tool_result: LanguageModelToolResult,
2118 ) -> Result<(), anyhow::Error> {
2119 log::debug!("Tool finished {:?}", tool_result);
2120
2121 event_stream.update_tool_call_fields(
2122 &tool_result.tool_use_id,
2123 acp::ToolCallUpdateFields::new()
2124 .status(if tool_result.is_error {
2125 acp::ToolCallStatus::Failed
2126 } else {
2127 acp::ToolCallStatus::Completed
2128 })
2129 .raw_output(tool_result.output.clone()),
2130 None,
2131 );
2132 this.update(cx, |this, _cx| {
2133 this.pending_message()
2134 .tool_results
2135 .insert(tool_result.tool_use_id.clone(), tool_result)
2136 })?;
2137 Ok(())
2138 }
2139
2140 fn handle_completion_error(
2141 &mut self,
2142 error: LanguageModelCompletionError,
2143 attempt: u8,
2144 plan: Option<Plan>,
2145 ) -> Result<acp_thread::RetryStatus> {
2146 let Some(model) = self.model.as_ref() else {
2147 return Err(anyhow!(error));
2148 };
2149
2150 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
2151 plan.is_some()
2152 } else {
2153 true
2154 };
2155
2156 if !auto_retry {
2157 return Err(anyhow!(error));
2158 }
2159
2160 let Some(strategy) = Self::retry_strategy_for(&error) else {
2161 return Err(anyhow!(error));
2162 };
2163
2164 let max_attempts = match &strategy {
2165 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2166 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2167 };
2168
2169 if attempt > max_attempts {
2170 return Err(anyhow!(error));
2171 }
2172
2173 let delay = match &strategy {
2174 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2175 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2176 Duration::from_secs(delay_secs)
2177 }
2178 RetryStrategy::Fixed { delay, .. } => *delay,
2179 };
2180 log::debug!("Retry attempt {attempt} with delay {delay:?}");
2181
2182 Ok(acp_thread::RetryStatus {
2183 last_error: error.to_string().into(),
2184 attempt: attempt as usize,
2185 max_attempts: max_attempts as usize,
2186 started_at: Instant::now(),
2187 duration: delay,
2188 })
2189 }
2190
2191 /// A helper method that's called on every streamed completion event.
2192 /// Returns an optional tool result task, which the main agentic loop will
2193 /// send back to the model when it resolves.
2194 fn handle_completion_event(
2195 &mut self,
2196 event: LanguageModelCompletionEvent,
2197 event_stream: &ThreadEventStream,
2198 cancellation_rx: watch::Receiver<bool>,
2199 cx: &mut Context<Self>,
2200 ) -> Result<Option<Task<LanguageModelToolResult>>> {
2201 log::trace!("Handling streamed completion event: {:?}", event);
2202 use LanguageModelCompletionEvent::*;
2203
2204 match event {
2205 StartMessage { .. } => {
2206 self.flush_pending_message(cx);
2207 self.pending_message = Some(AgentMessage::default());
2208 }
2209 Text(new_text) => self.handle_text_event(new_text, event_stream),
2210 Thinking { text, signature } => {
2211 self.handle_thinking_event(text, signature, event_stream)
2212 }
2213 RedactedThinking { data } => self.handle_redacted_thinking_event(data),
2214 ReasoningDetails(details) => {
2215 let last_message = self.pending_message();
2216 // Store the last non-empty reasoning_details (overwrites earlier ones)
2217 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
2218 if let serde_json::Value::Array(ref arr) = details {
2219 if !arr.is_empty() {
2220 last_message.reasoning_details = Some(details);
2221 }
2222 } else {
2223 last_message.reasoning_details = Some(details);
2224 }
2225 }
2226 ToolUse(tool_use) => {
2227 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
2228 }
2229 ToolUseJsonParseError {
2230 id,
2231 tool_name,
2232 raw_input,
2233 json_parse_error,
2234 } => {
2235 return Ok(self.handle_tool_use_json_parse_error_event(
2236 id,
2237 tool_name,
2238 raw_input,
2239 json_parse_error,
2240 event_stream,
2241 cancellation_rx,
2242 cx,
2243 ));
2244 }
2245 UsageUpdate(usage) => {
2246 telemetry::event!(
2247 "Agent Thread Completion Usage Updated",
2248 thread_id = self.id.to_string(),
2249 parent_thread_id = self.parent_thread_id().map(|id| id.to_string()),
2250 prompt_id = self.prompt_id.to_string(),
2251 model = self.model.as_ref().map(|m| m.telemetry_id()),
2252 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
2253 input_tokens = usage.input_tokens,
2254 output_tokens = usage.output_tokens,
2255 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2256 cache_read_input_tokens = usage.cache_read_input_tokens,
2257 );
2258 self.update_token_usage(usage, cx);
2259 }
2260 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
2261 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
2262 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
2263 Started | Queued { .. } => {}
2264 }
2265
2266 Ok(None)
2267 }
2268
2269 fn handle_text_event(&mut self, new_text: String, event_stream: &ThreadEventStream) {
2270 event_stream.send_text(&new_text);
2271
2272 let last_message = self.pending_message();
2273 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
2274 text.push_str(&new_text);
2275 } else {
2276 last_message
2277 .content
2278 .push(AgentMessageContent::Text(new_text));
2279 }
2280 }
2281
2282 fn handle_thinking_event(
2283 &mut self,
2284 new_text: String,
2285 new_signature: Option<String>,
2286 event_stream: &ThreadEventStream,
2287 ) {
2288 event_stream.send_thinking(&new_text);
2289
2290 let last_message = self.pending_message();
2291 if let Some(AgentMessageContent::Thinking { text, signature }) =
2292 last_message.content.last_mut()
2293 {
2294 text.push_str(&new_text);
2295 *signature = new_signature.or(signature.take());
2296 } else {
2297 last_message.content.push(AgentMessageContent::Thinking {
2298 text: new_text,
2299 signature: new_signature,
2300 });
2301 }
2302 }
2303
2304 fn handle_redacted_thinking_event(&mut self, data: String) {
2305 let last_message = self.pending_message();
2306 last_message
2307 .content
2308 .push(AgentMessageContent::RedactedThinking(data));
2309 }
2310
2311 fn handle_tool_use_event(
2312 &mut self,
2313 tool_use: LanguageModelToolUse,
2314 event_stream: &ThreadEventStream,
2315 cancellation_rx: watch::Receiver<bool>,
2316 cx: &mut Context<Self>,
2317 ) -> Option<Task<LanguageModelToolResult>> {
2318 cx.notify();
2319
2320 let tool = self.tool(tool_use.name.as_ref());
2321 let mut title = SharedString::from(&tool_use.name);
2322 let mut kind = acp::ToolKind::Other;
2323 if let Some(tool) = tool.as_ref() {
2324 title = tool.initial_title(tool_use.input.clone(), cx);
2325 kind = tool.kind();
2326 }
2327
2328 self.send_or_update_tool_use(&tool_use, title, kind, event_stream);
2329
2330 let Some(tool) = tool else {
2331 let content = format!("No tool named {} exists", tool_use.name);
2332 return Some(Task::ready(LanguageModelToolResult {
2333 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2334 tool_use_id: tool_use.id,
2335 tool_name: tool_use.name,
2336 is_error: true,
2337 output: None,
2338 }));
2339 };
2340
2341 if !tool_use.is_input_complete {
2342 if tool.supports_input_streaming() {
2343 let running_turn = self.running_turn.as_mut()?;
2344 if let Some(sender) = running_turn.streaming_tool_inputs.get_mut(&tool_use.id) {
2345 sender.send_partial(tool_use.input);
2346 return None;
2347 }
2348
2349 let (mut sender, tool_input) = ToolInputSender::channel();
2350 sender.send_partial(tool_use.input);
2351 running_turn
2352 .streaming_tool_inputs
2353 .insert(tool_use.id.clone(), sender);
2354
2355 let tool = tool.clone();
2356 log::debug!("Running streaming tool {}", tool_use.name);
2357 return Some(self.run_tool(
2358 tool,
2359 tool_input,
2360 tool_use.id,
2361 tool_use.name,
2362 event_stream,
2363 cancellation_rx,
2364 cx,
2365 ));
2366 } else {
2367 return None;
2368 }
2369 }
2370
2371 if let Some(mut sender) = self
2372 .running_turn
2373 .as_mut()?
2374 .streaming_tool_inputs
2375 .remove(&tool_use.id)
2376 {
2377 sender.send_full(tool_use.input);
2378 return None;
2379 }
2380
2381 log::debug!("Running tool {}", tool_use.name);
2382 let tool_input = ToolInput::ready(tool_use.input);
2383 Some(self.run_tool(
2384 tool,
2385 tool_input,
2386 tool_use.id,
2387 tool_use.name,
2388 event_stream,
2389 cancellation_rx,
2390 cx,
2391 ))
2392 }
2393
2394 fn run_tool(
2395 &self,
2396 tool: Arc<dyn AnyAgentTool>,
2397 tool_input: ToolInput<serde_json::Value>,
2398 tool_use_id: LanguageModelToolUseId,
2399 tool_name: Arc<str>,
2400 event_stream: &ThreadEventStream,
2401 cancellation_rx: watch::Receiver<bool>,
2402 cx: &mut Context<Self>,
2403 ) -> Task<LanguageModelToolResult> {
2404 let fs = self.project.read(cx).fs().clone();
2405 let tool_event_stream = ToolCallEventStream::new(
2406 tool_use_id.clone(),
2407 event_stream.clone(),
2408 Some(fs),
2409 cancellation_rx,
2410 );
2411 tool_event_stream.update_fields(
2412 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
2413 );
2414 let supports_images = self.model().is_some_and(|model| model.supports_images());
2415 let tool_result = tool.run(tool_input, tool_event_stream, cx);
2416 cx.foreground_executor().spawn(async move {
2417 let (is_error, output) = match tool_result.await {
2418 Ok(mut output) => {
2419 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
2420 && !supports_images
2421 {
2422 output = AgentToolOutput::from_error(
2423 "Attempted to read an image, but this model doesn't support it.",
2424 );
2425 (true, output)
2426 } else {
2427 (false, output)
2428 }
2429 }
2430 Err(output) => (true, output),
2431 };
2432
2433 LanguageModelToolResult {
2434 tool_use_id,
2435 tool_name,
2436 is_error,
2437 content: output.llm_output,
2438 output: Some(output.raw_output),
2439 }
2440 })
2441 }
2442
2443 fn handle_tool_use_json_parse_error_event(
2444 &mut self,
2445 tool_use_id: LanguageModelToolUseId,
2446 tool_name: Arc<str>,
2447 raw_input: Arc<str>,
2448 json_parse_error: String,
2449 event_stream: &ThreadEventStream,
2450 cancellation_rx: watch::Receiver<bool>,
2451 cx: &mut Context<Self>,
2452 ) -> Option<Task<LanguageModelToolResult>> {
2453 let tool_use = LanguageModelToolUse {
2454 id: tool_use_id,
2455 name: tool_name,
2456 raw_input: raw_input.to_string(),
2457 input: serde_json::json!({}),
2458 is_input_complete: true,
2459 thought_signature: None,
2460 };
2461 self.send_or_update_tool_use(
2462 &tool_use,
2463 SharedString::from(&tool_use.name),
2464 acp::ToolKind::Other,
2465 event_stream,
2466 );
2467
2468 let tool = self.tool(tool_use.name.as_ref());
2469
2470 let Some(tool) = tool else {
2471 let content = format!("No tool named {} exists", tool_use.name);
2472 return Some(Task::ready(LanguageModelToolResult {
2473 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2474 tool_use_id: tool_use.id,
2475 tool_name: tool_use.name,
2476 is_error: true,
2477 output: None,
2478 }));
2479 };
2480
2481 let error_message = format!("Error parsing input JSON: {json_parse_error}");
2482
2483 if tool.supports_input_streaming()
2484 && let Some(mut sender) = self
2485 .running_turn
2486 .as_mut()?
2487 .streaming_tool_inputs
2488 .remove(&tool_use.id)
2489 {
2490 sender.send_invalid_json(error_message);
2491 return None;
2492 }
2493
2494 log::debug!("Running tool {}. Received invalid JSON", tool_use.name);
2495 let tool_input = ToolInput::invalid_json(error_message);
2496 Some(self.run_tool(
2497 tool,
2498 tool_input,
2499 tool_use.id,
2500 tool_use.name,
2501 event_stream,
2502 cancellation_rx,
2503 cx,
2504 ))
2505 }
2506
2507 fn send_or_update_tool_use(
2508 &mut self,
2509 tool_use: &LanguageModelToolUse,
2510 title: SharedString,
2511 kind: acp::ToolKind,
2512 event_stream: &ThreadEventStream,
2513 ) {
2514 // Ensure the last message ends in the current tool use
2515 let last_message = self.pending_message();
2516
2517 let has_tool_use = last_message.content.iter_mut().rev().any(|content| {
2518 if let AgentMessageContent::ToolUse(last_tool_use) = content {
2519 if last_tool_use.id == tool_use.id {
2520 *last_tool_use = tool_use.clone();
2521 return true;
2522 }
2523 }
2524 false
2525 });
2526
2527 if !has_tool_use {
2528 event_stream.send_tool_call(
2529 &tool_use.id,
2530 &tool_use.name,
2531 title,
2532 kind,
2533 tool_use.input.clone(),
2534 );
2535 last_message
2536 .content
2537 .push(AgentMessageContent::ToolUse(tool_use.clone()));
2538 } else {
2539 event_stream.update_tool_call_fields(
2540 &tool_use.id,
2541 acp::ToolCallUpdateFields::new()
2542 .title(title.as_str())
2543 .kind(kind)
2544 .raw_input(tool_use.input.clone()),
2545 None,
2546 );
2547 }
2548 }
2549
2550 pub fn title(&self) -> Option<SharedString> {
2551 self.title.clone()
2552 }
2553
2554 pub fn is_generating_summary(&self) -> bool {
2555 self.pending_summary_generation.is_some()
2556 }
2557
2558 pub fn is_generating_title(&self) -> bool {
2559 self.pending_title_generation.is_some()
2560 }
2561
2562 pub fn has_failed_title_generation(&self) -> bool {
2563 self.title_generation_failed
2564 }
2565
2566 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
2567 if let Some(summary) = self.summary.as_ref() {
2568 return Task::ready(Some(summary.clone())).shared();
2569 }
2570 if let Some(task) = self.pending_summary_generation.clone() {
2571 return task;
2572 }
2573 let Some(model) = self.summarization_model.clone() else {
2574 log::error!("No summarization model available");
2575 return Task::ready(None).shared();
2576 };
2577 let mut request = LanguageModelRequest {
2578 intent: Some(CompletionIntent::ThreadContextSummarization),
2579 temperature: AgentSettings::temperature_for_model(&model, cx),
2580 ..Default::default()
2581 };
2582
2583 for message in &self.messages {
2584 request.messages.extend(message.to_request());
2585 }
2586
2587 request.messages.push(LanguageModelRequestMessage {
2588 role: Role::User,
2589 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
2590 cache: false,
2591 reasoning_details: None,
2592 });
2593
2594 let task = cx
2595 .spawn(async move |this, cx| {
2596 let mut summary = String::new();
2597 let mut messages = model.stream_completion(request, cx).await.log_err()?;
2598 while let Some(event) = messages.next().await {
2599 let event = event.log_err()?;
2600 let text = match event {
2601 LanguageModelCompletionEvent::Text(text) => text,
2602 _ => continue,
2603 };
2604
2605 let mut lines = text.lines();
2606 summary.extend(lines.next());
2607 }
2608
2609 log::debug!("Setting summary: {}", summary);
2610 let summary = SharedString::from(summary);
2611
2612 this.update(cx, |this, cx| {
2613 this.summary = Some(summary.clone());
2614 this.pending_summary_generation = None;
2615 cx.notify()
2616 })
2617 .ok()?;
2618
2619 Some(summary)
2620 })
2621 .shared();
2622 self.pending_summary_generation = Some(task.clone());
2623 task
2624 }
2625
2626 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
2627 self.title_generation_failed = false;
2628 let Some(model) = self.summarization_model.clone() else {
2629 return;
2630 };
2631
2632 log::debug!(
2633 "Generating title with model: {:?}",
2634 self.summarization_model.as_ref().map(|model| model.name())
2635 );
2636 let mut request = LanguageModelRequest {
2637 intent: Some(CompletionIntent::ThreadSummarization),
2638 temperature: AgentSettings::temperature_for_model(&model, cx),
2639 ..Default::default()
2640 };
2641
2642 for message in &self.messages {
2643 request.messages.extend(message.to_request());
2644 }
2645
2646 request.messages.push(LanguageModelRequestMessage {
2647 role: Role::User,
2648 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2649 cache: false,
2650 reasoning_details: None,
2651 });
2652 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2653 let mut title = String::new();
2654
2655 let generate = async {
2656 let mut messages = model.stream_completion(request, cx).await?;
2657 while let Some(event) = messages.next().await {
2658 let event = event?;
2659 let text = match event {
2660 LanguageModelCompletionEvent::Text(text) => text,
2661 _ => continue,
2662 };
2663
2664 let mut lines = text.lines();
2665 title.extend(lines.next());
2666
2667 // Stop if the LLM generated multiple lines.
2668 if lines.next().is_some() {
2669 break;
2670 }
2671 }
2672 anyhow::Ok(())
2673 };
2674
2675 let succeeded = generate
2676 .await
2677 .context("failed to generate thread title")
2678 .log_err()
2679 .is_some();
2680 _ = this.update(cx, |this, cx| {
2681 this.pending_title_generation = None;
2682 if succeeded {
2683 this.set_title(title.into(), cx);
2684 } else {
2685 this.title_generation_failed = true;
2686 cx.emit(TitleUpdated);
2687 cx.notify();
2688 }
2689 });
2690 }));
2691 }
2692
2693 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2694 self.pending_title_generation = None;
2695 self.title_generation_failed = false;
2696 if Some(&title) != self.title.as_ref() {
2697 self.title = Some(title);
2698 cx.emit(TitleUpdated);
2699 cx.notify();
2700 }
2701 }
2702
2703 fn clear_summary(&mut self) {
2704 self.summary = None;
2705 self.pending_summary_generation = None;
2706 }
2707
2708 fn last_user_message(&self) -> Option<&UserMessage> {
2709 self.messages
2710 .iter()
2711 .rev()
2712 .find_map(|message| match message {
2713 Message::User(user_message) => Some(user_message),
2714 Message::Agent(_) => None,
2715 Message::Resume => None,
2716 })
2717 }
2718
2719 fn pending_message(&mut self) -> &mut AgentMessage {
2720 self.pending_message.get_or_insert_default()
2721 }
2722
2723 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2724 let Some(mut message) = self.pending_message.take() else {
2725 return;
2726 };
2727
2728 if message.content.is_empty() {
2729 return;
2730 }
2731
2732 for content in &message.content {
2733 let AgentMessageContent::ToolUse(tool_use) = content else {
2734 continue;
2735 };
2736
2737 if !message.tool_results.contains_key(&tool_use.id) {
2738 message.tool_results.insert(
2739 tool_use.id.clone(),
2740 LanguageModelToolResult {
2741 tool_use_id: tool_use.id.clone(),
2742 tool_name: tool_use.name.clone(),
2743 is_error: true,
2744 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2745 output: None,
2746 },
2747 );
2748 }
2749 }
2750
2751 self.messages.push(Message::Agent(message));
2752 self.updated_at = Utc::now();
2753 self.clear_summary();
2754 cx.notify()
2755 }
2756
2757 pub(crate) fn build_completion_request(
2758 &self,
2759 completion_intent: CompletionIntent,
2760 cx: &App,
2761 ) -> Result<LanguageModelRequest> {
2762 let completion_intent =
2763 if self.is_subagent() && completion_intent == CompletionIntent::UserPrompt {
2764 CompletionIntent::Subagent
2765 } else {
2766 completion_intent
2767 };
2768
2769 let model = self
2770 .model()
2771 .ok_or_else(|| anyhow!(NoModelConfiguredError))?;
2772 let tools = if let Some(turn) = self.running_turn.as_ref() {
2773 turn.tools
2774 .iter()
2775 .filter_map(|(tool_name, tool)| {
2776 log::trace!("Including tool: {}", tool_name);
2777 Some(LanguageModelRequestTool {
2778 name: tool_name.to_string(),
2779 description: tool.description().to_string(),
2780 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2781 use_input_streaming: tool.supports_input_streaming(),
2782 })
2783 })
2784 .collect::<Vec<_>>()
2785 } else {
2786 Vec::new()
2787 };
2788
2789 log::debug!("Building completion request");
2790 log::debug!("Completion intent: {:?}", completion_intent);
2791
2792 let available_tools: Vec<_> = self
2793 .running_turn
2794 .as_ref()
2795 .map(|turn| turn.tools.keys().cloned().collect())
2796 .unwrap_or_default();
2797
2798 log::debug!("Request includes {} tools", available_tools.len());
2799 let messages = self.build_request_messages(available_tools, cx);
2800 log::debug!("Request will include {} messages", messages.len());
2801
2802 let request = LanguageModelRequest {
2803 thread_id: Some(self.id.to_string()),
2804 prompt_id: Some(self.prompt_id.to_string()),
2805 intent: Some(completion_intent),
2806 messages,
2807 tools,
2808 tool_choice: None,
2809 stop: Vec::new(),
2810 temperature: AgentSettings::temperature_for_model(model, cx),
2811 thinking_allowed: self.thinking_enabled,
2812 thinking_effort: self.thinking_effort.clone(),
2813 speed: self.speed(),
2814 };
2815
2816 log::debug!("Completion request built successfully");
2817 Ok(request)
2818 }
2819
2820 fn enabled_tools(&self, cx: &App) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2821 let Some(model) = self.model.as_ref() else {
2822 return BTreeMap::new();
2823 };
2824 let Some(profile) = AgentSettings::get_global(cx).profiles.get(&self.profile_id) else {
2825 return BTreeMap::new();
2826 };
2827 fn truncate(tool_name: &SharedString) -> SharedString {
2828 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2829 let mut truncated = tool_name.to_string();
2830 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2831 truncated.into()
2832 } else {
2833 tool_name.clone()
2834 }
2835 }
2836
2837 let use_streaming_edit_tool =
2838 cx.has_flag::<StreamingEditFileToolFeatureFlag>() && model.supports_streaming_tools();
2839
2840 let mut tools = self
2841 .tools
2842 .iter()
2843 .filter_map(|(tool_name, tool)| {
2844 // For streaming_edit_file, check profile against "edit_file" since that's what users configure
2845 let profile_tool_name = if tool_name == StreamingEditFileTool::NAME {
2846 EditFileTool::NAME
2847 } else {
2848 tool_name.as_ref()
2849 };
2850
2851 if tool.supports_provider(&model.provider_id())
2852 && profile.is_tool_enabled(profile_tool_name)
2853 {
2854 match (tool_name.as_ref(), use_streaming_edit_tool) {
2855 (StreamingEditFileTool::NAME, false) | (EditFileTool::NAME, true) => None,
2856 (StreamingEditFileTool::NAME, true) => {
2857 // Expose streaming tool as "edit_file"
2858 Some((SharedString::from(EditFileTool::NAME), tool.clone()))
2859 }
2860 _ => Some((truncate(tool_name), tool.clone())),
2861 }
2862 } else {
2863 None
2864 }
2865 })
2866 .collect::<BTreeMap<_, _>>();
2867
2868 let mut context_server_tools = Vec::new();
2869 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2870 let mut duplicate_tool_names = HashSet::default();
2871 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2872 for (tool_name, tool) in server_tools {
2873 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2874 let tool_name = truncate(tool_name);
2875 if !seen_tools.insert(tool_name.clone()) {
2876 duplicate_tool_names.insert(tool_name.clone());
2877 }
2878 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2879 }
2880 }
2881 }
2882
2883 // When there are duplicate tool names, disambiguate by prefixing them
2884 // with the server ID (converted to snake_case for API compatibility).
2885 // In the rare case there isn't enough space for the disambiguated tool
2886 // name, keep only the last tool with this name.
2887 for (server_id, tool_name, tool) in context_server_tools {
2888 if duplicate_tool_names.contains(&tool_name) {
2889 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2890 if available >= 2 {
2891 let mut disambiguated = server_id.0.to_snake_case();
2892 disambiguated.truncate(available - 1);
2893 disambiguated.push('_');
2894 disambiguated.push_str(&tool_name);
2895 tools.insert(disambiguated.into(), tool.clone());
2896 } else {
2897 tools.insert(tool_name, tool.clone());
2898 }
2899 } else {
2900 tools.insert(tool_name, tool.clone());
2901 }
2902 }
2903
2904 tools
2905 }
2906
2907 fn refresh_turn_tools(&mut self, cx: &App) {
2908 let tools = self.enabled_tools(cx);
2909 if let Some(turn) = self.running_turn.as_mut() {
2910 turn.tools = tools;
2911 }
2912 }
2913
2914 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2915 self.running_turn.as_ref()?.tools.get(name).cloned()
2916 }
2917
2918 pub fn has_tool(&self, name: &str) -> bool {
2919 self.running_turn
2920 .as_ref()
2921 .is_some_and(|turn| turn.tools.contains_key(name))
2922 }
2923
2924 #[cfg(any(test, feature = "test-support"))]
2925 pub fn has_registered_tool(&self, name: &str) -> bool {
2926 self.tools.contains_key(name)
2927 }
2928
2929 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2930 self.tools.keys().cloned().collect()
2931 }
2932
2933 pub(crate) fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2934 self.running_subagents.push(subagent);
2935 }
2936
2937 pub(crate) fn unregister_running_subagent(
2938 &mut self,
2939 subagent_session_id: &acp::SessionId,
2940 cx: &App,
2941 ) {
2942 self.running_subagents.retain(|s| {
2943 s.upgrade()
2944 .map_or(false, |s| s.read(cx).id() != subagent_session_id)
2945 });
2946 }
2947
2948 #[cfg(any(test, feature = "test-support"))]
2949 pub fn running_subagent_ids(&self, cx: &App) -> Vec<acp::SessionId> {
2950 self.running_subagents
2951 .iter()
2952 .filter_map(|s| s.upgrade().map(|s| s.read(cx).id().clone()))
2953 .collect()
2954 }
2955
2956 pub fn is_subagent(&self) -> bool {
2957 self.subagent_context.is_some()
2958 }
2959
2960 pub fn parent_thread_id(&self) -> Option<acp::SessionId> {
2961 self.subagent_context
2962 .as_ref()
2963 .map(|c| c.parent_thread_id.clone())
2964 }
2965
2966 pub fn depth(&self) -> u8 {
2967 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2968 }
2969
2970 #[cfg(any(test, feature = "test-support"))]
2971 pub fn set_subagent_context(&mut self, context: SubagentContext) {
2972 self.subagent_context = Some(context);
2973 }
2974
2975 pub fn is_turn_complete(&self) -> bool {
2976 self.running_turn.is_none()
2977 }
2978
2979 fn build_request_messages(
2980 &self,
2981 available_tools: Vec<SharedString>,
2982 cx: &App,
2983 ) -> Vec<LanguageModelRequestMessage> {
2984 log::trace!(
2985 "Building request messages from {} thread messages",
2986 self.messages.len()
2987 );
2988
2989 let system_prompt = SystemPromptTemplate {
2990 project: self.project_context.read(cx),
2991 available_tools,
2992 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2993 }
2994 .render(&self.templates)
2995 .context("failed to build system prompt")
2996 .expect("Invalid template");
2997 let mut messages = vec![LanguageModelRequestMessage {
2998 role: Role::System,
2999 content: vec![system_prompt.into()],
3000 cache: false,
3001 reasoning_details: None,
3002 }];
3003 for message in &self.messages {
3004 messages.extend(message.to_request());
3005 }
3006
3007 if let Some(last_message) = messages.last_mut() {
3008 last_message.cache = true;
3009 }
3010
3011 if let Some(message) = self.pending_message.as_ref() {
3012 messages.extend(message.to_request());
3013 }
3014
3015 messages
3016 }
3017
3018 pub fn to_markdown(&self) -> String {
3019 let mut markdown = String::new();
3020 for (ix, message) in self.messages.iter().enumerate() {
3021 if ix > 0 {
3022 markdown.push('\n');
3023 }
3024 match message {
3025 Message::User(_) => markdown.push_str("## User\n\n"),
3026 Message::Agent(_) => markdown.push_str("## Assistant\n\n"),
3027 Message::Resume => {}
3028 }
3029 markdown.push_str(&message.to_markdown());
3030 }
3031
3032 if let Some(message) = self.pending_message.as_ref() {
3033 markdown.push_str("\n## Assistant\n\n");
3034 markdown.push_str(&message.to_markdown());
3035 }
3036
3037 markdown
3038 }
3039
3040 fn advance_prompt_id(&mut self) {
3041 self.prompt_id = PromptId::new();
3042 }
3043
3044 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
3045 use LanguageModelCompletionError::*;
3046 use http_client::StatusCode;
3047
3048 // General strategy here:
3049 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
3050 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
3051 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
3052 match error {
3053 HttpResponseError {
3054 status_code: StatusCode::TOO_MANY_REQUESTS,
3055 ..
3056 } => Some(RetryStrategy::ExponentialBackoff {
3057 initial_delay: BASE_RETRY_DELAY,
3058 max_attempts: MAX_RETRY_ATTEMPTS,
3059 }),
3060 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
3061 Some(RetryStrategy::Fixed {
3062 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3063 max_attempts: MAX_RETRY_ATTEMPTS,
3064 })
3065 }
3066 UpstreamProviderError {
3067 status,
3068 retry_after,
3069 ..
3070 } => match *status {
3071 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
3072 Some(RetryStrategy::Fixed {
3073 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3074 max_attempts: MAX_RETRY_ATTEMPTS,
3075 })
3076 }
3077 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
3078 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3079 // Internal Server Error could be anything, retry up to 3 times.
3080 max_attempts: 3,
3081 }),
3082 status => {
3083 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
3084 // but we frequently get them in practice. See https://http.dev/529
3085 if status.as_u16() == 529 {
3086 Some(RetryStrategy::Fixed {
3087 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3088 max_attempts: MAX_RETRY_ATTEMPTS,
3089 })
3090 } else {
3091 Some(RetryStrategy::Fixed {
3092 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3093 max_attempts: 2,
3094 })
3095 }
3096 }
3097 },
3098 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
3099 delay: BASE_RETRY_DELAY,
3100 max_attempts: 3,
3101 }),
3102 ApiReadResponseError { .. }
3103 | HttpSend { .. }
3104 | DeserializeResponse { .. }
3105 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
3106 delay: BASE_RETRY_DELAY,
3107 max_attempts: 3,
3108 }),
3109 // Retrying these errors definitely shouldn't help.
3110 HttpResponseError {
3111 status_code:
3112 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
3113 ..
3114 }
3115 | AuthenticationError { .. }
3116 | PermissionError { .. }
3117 | NoApiKey { .. }
3118 | ApiEndpointNotFound { .. }
3119 | PromptTooLarge { .. } => None,
3120 // These errors might be transient, so retry them
3121 SerializeRequest { .. } | BuildRequestBody { .. } | StreamEndedUnexpectedly { .. } => {
3122 Some(RetryStrategy::Fixed {
3123 delay: BASE_RETRY_DELAY,
3124 max_attempts: 1,
3125 })
3126 }
3127 // Retry all other 4xx and 5xx errors once.
3128 HttpResponseError { status_code, .. }
3129 if status_code.is_client_error() || status_code.is_server_error() =>
3130 {
3131 Some(RetryStrategy::Fixed {
3132 delay: BASE_RETRY_DELAY,
3133 max_attempts: 3,
3134 })
3135 }
3136 Other(err) if err.is::<language_model::PaymentRequiredError>() => {
3137 // Retrying won't help for Payment Required errors.
3138 None
3139 }
3140 // Conservatively assume that any other errors are non-retryable
3141 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
3142 delay: BASE_RETRY_DELAY,
3143 max_attempts: 2,
3144 }),
3145 }
3146 }
3147}
3148
3149struct RunningTurn {
3150 /// Holds the task that handles agent interaction until the end of the turn.
3151 /// Survives across multiple requests as the model performs tool calls and
3152 /// we run tools, report their results.
3153 _task: Task<()>,
3154 /// The current event stream for the running turn. Used to report a final
3155 /// cancellation event if we cancel the turn.
3156 event_stream: ThreadEventStream,
3157 /// The tools that are enabled for the current iteration of the turn.
3158 /// Refreshed at the start of each iteration via `refresh_turn_tools`.
3159 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
3160 /// Sender to signal tool cancellation. When cancel is called, this is
3161 /// set to true so all tools can detect user-initiated cancellation.
3162 cancellation_tx: watch::Sender<bool>,
3163 /// Senders for tools that support input streaming and have already been
3164 /// started but are still receiving input from the LLM.
3165 streaming_tool_inputs: HashMap<LanguageModelToolUseId, ToolInputSender>,
3166}
3167
3168impl RunningTurn {
3169 fn cancel(mut self) -> Task<()> {
3170 log::debug!("Cancelling in progress turn");
3171 self.cancellation_tx.send(true).ok();
3172 self.event_stream.send_canceled();
3173 self._task
3174 }
3175}
3176
3177pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
3178
3179impl EventEmitter<TokenUsageUpdated> for Thread {}
3180
3181pub struct TitleUpdated;
3182
3183impl EventEmitter<TitleUpdated> for Thread {}
3184
3185/// A channel-based wrapper that delivers tool input to a running tool.
3186///
3187/// For non-streaming tools, created via `ToolInput::ready()` so `.recv()` resolves immediately.
3188/// For streaming tools, partial JSON snapshots arrive via `.recv_partial()` as the LLM streams
3189/// them, followed by the final complete input available through `.recv()`.
3190pub struct ToolInput<T> {
3191 rx: mpsc::UnboundedReceiver<ToolInputPayload<serde_json::Value>>,
3192 _phantom: PhantomData<T>,
3193}
3194
3195impl<T: DeserializeOwned> ToolInput<T> {
3196 #[cfg(any(test, feature = "test-support"))]
3197 pub fn resolved(input: impl Serialize) -> Self {
3198 let value = serde_json::to_value(input).expect("failed to serialize tool input");
3199 Self::ready(value)
3200 }
3201
3202 pub fn ready(value: serde_json::Value) -> Self {
3203 let (tx, rx) = mpsc::unbounded();
3204 tx.unbounded_send(ToolInputPayload::Full(value)).ok();
3205 Self {
3206 rx,
3207 _phantom: PhantomData,
3208 }
3209 }
3210
3211 pub fn invalid_json(error_message: String) -> Self {
3212 let (tx, rx) = mpsc::unbounded();
3213 tx.unbounded_send(ToolInputPayload::InvalidJson { error_message })
3214 .ok();
3215 Self {
3216 rx,
3217 _phantom: PhantomData,
3218 }
3219 }
3220
3221 #[cfg(any(test, feature = "test-support"))]
3222 pub fn test() -> (ToolInputSender, Self) {
3223 let (sender, input) = ToolInputSender::channel();
3224 (sender, input.cast())
3225 }
3226
3227 /// Wait for the final deserialized input, ignoring all partial updates.
3228 /// Non-streaming tools can use this to wait until the whole input is available.
3229 pub async fn recv(mut self) -> Result<T> {
3230 while let Ok(value) = self.next().await {
3231 match value {
3232 ToolInputPayload::Full(value) => return Ok(value),
3233 ToolInputPayload::Partial(_) => {}
3234 ToolInputPayload::InvalidJson { error_message } => {
3235 return Err(anyhow!(error_message));
3236 }
3237 }
3238 }
3239 Err(anyhow!("tool input was not fully received"))
3240 }
3241
3242 pub async fn next(&mut self) -> Result<ToolInputPayload<T>> {
3243 let value = self
3244 .rx
3245 .next()
3246 .await
3247 .ok_or_else(|| anyhow!("tool input was not fully received"))?;
3248
3249 Ok(match value {
3250 ToolInputPayload::Partial(payload) => ToolInputPayload::Partial(payload),
3251 ToolInputPayload::Full(payload) => {
3252 ToolInputPayload::Full(serde_json::from_value(payload)?)
3253 }
3254 ToolInputPayload::InvalidJson { error_message } => {
3255 ToolInputPayload::InvalidJson { error_message }
3256 }
3257 })
3258 }
3259
3260 fn cast<U: DeserializeOwned>(self) -> ToolInput<U> {
3261 ToolInput {
3262 rx: self.rx,
3263 _phantom: PhantomData,
3264 }
3265 }
3266}
3267
3268pub enum ToolInputPayload<T> {
3269 Partial(serde_json::Value),
3270 Full(T),
3271 InvalidJson { error_message: String },
3272}
3273
3274pub struct ToolInputSender {
3275 has_received_final: bool,
3276 tx: mpsc::UnboundedSender<ToolInputPayload<serde_json::Value>>,
3277}
3278
3279impl ToolInputSender {
3280 pub(crate) fn channel() -> (Self, ToolInput<serde_json::Value>) {
3281 let (tx, rx) = mpsc::unbounded();
3282 let sender = Self {
3283 tx,
3284 has_received_final: false,
3285 };
3286 let input = ToolInput {
3287 rx,
3288 _phantom: PhantomData,
3289 };
3290 (sender, input)
3291 }
3292
3293 pub(crate) fn has_received_final(&self) -> bool {
3294 self.has_received_final
3295 }
3296
3297 pub fn send_partial(&mut self, payload: serde_json::Value) {
3298 self.tx
3299 .unbounded_send(ToolInputPayload::Partial(payload))
3300 .ok();
3301 }
3302
3303 pub fn send_full(&mut self, payload: serde_json::Value) {
3304 self.has_received_final = true;
3305 self.tx.unbounded_send(ToolInputPayload::Full(payload)).ok();
3306 }
3307
3308 pub fn send_invalid_json(&mut self, error_message: String) {
3309 self.has_received_final = true;
3310 self.tx
3311 .unbounded_send(ToolInputPayload::InvalidJson { error_message })
3312 .ok();
3313 }
3314}
3315
3316pub trait AgentTool
3317where
3318 Self: 'static + Sized,
3319{
3320 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
3321 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
3322
3323 const NAME: &'static str;
3324
3325 fn description() -> SharedString {
3326 let schema = schemars::schema_for!(Self::Input);
3327 SharedString::new(
3328 schema
3329 .get("description")
3330 .and_then(|description| description.as_str())
3331 .unwrap_or_default(),
3332 )
3333 }
3334
3335 fn kind() -> acp::ToolKind;
3336
3337 /// The initial tool title to display. Can be updated during the tool run.
3338 fn initial_title(
3339 &self,
3340 input: Result<Self::Input, serde_json::Value>,
3341 cx: &mut App,
3342 ) -> SharedString;
3343
3344 /// Returns the JSON schema that describes the tool's input.
3345 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
3346 language_model::tool_schema::root_schema_for::<Self::Input>(format)
3347 }
3348
3349 /// Returns whether the tool supports streaming of tool use parameters.
3350 fn supports_input_streaming() -> bool {
3351 false
3352 }
3353
3354 /// Some tools rely on a provider for the underlying billing or other reasons.
3355 /// Allow the tool to check if they are compatible, or should be filtered out.
3356 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
3357 true
3358 }
3359
3360 /// Runs the tool with the provided input.
3361 ///
3362 /// Returns `Result<Self::Output, Self::Output>` rather than `Result<Self::Output, anyhow::Error>`
3363 /// because tool errors are sent back to the model as tool results. This means error output must
3364 /// be structured and readable by the agent — not an arbitrary `anyhow::Error`. Returning the
3365 /// same `Output` type for both success and failure lets tools provide structured data while
3366 /// still signaling whether the invocation succeeded or failed.
3367 fn run(
3368 self: Arc<Self>,
3369 input: ToolInput<Self::Input>,
3370 event_stream: ToolCallEventStream,
3371 cx: &mut App,
3372 ) -> Task<Result<Self::Output, Self::Output>>;
3373
3374 /// Emits events for a previous execution of the tool.
3375 fn replay(
3376 &self,
3377 _input: Self::Input,
3378 _output: Self::Output,
3379 _event_stream: ToolCallEventStream,
3380 _cx: &mut App,
3381 ) -> Result<()> {
3382 Ok(())
3383 }
3384
3385 fn erase(self) -> Arc<dyn AnyAgentTool> {
3386 Arc::new(Erased(Arc::new(self)))
3387 }
3388}
3389
3390pub struct Erased<T>(T);
3391
3392pub struct AgentToolOutput {
3393 pub llm_output: LanguageModelToolResultContent,
3394 pub raw_output: serde_json::Value,
3395}
3396
3397impl AgentToolOutput {
3398 pub fn from_error(message: impl Into<String>) -> Self {
3399 let message = message.into();
3400 let llm_output = LanguageModelToolResultContent::Text(Arc::from(message.as_str()));
3401 Self {
3402 raw_output: serde_json::Value::String(message),
3403 llm_output,
3404 }
3405 }
3406}
3407
3408pub trait AnyAgentTool {
3409 fn name(&self) -> SharedString;
3410 fn description(&self) -> SharedString;
3411 fn kind(&self) -> acp::ToolKind;
3412 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
3413 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
3414 fn supports_input_streaming(&self) -> bool {
3415 false
3416 }
3417 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
3418 true
3419 }
3420 /// See [`AgentTool::run`] for why this returns `Result<AgentToolOutput, AgentToolOutput>`.
3421 fn run(
3422 self: Arc<Self>,
3423 input: ToolInput<serde_json::Value>,
3424 event_stream: ToolCallEventStream,
3425 cx: &mut App,
3426 ) -> Task<Result<AgentToolOutput, AgentToolOutput>>;
3427 fn replay(
3428 &self,
3429 input: serde_json::Value,
3430 output: serde_json::Value,
3431 event_stream: ToolCallEventStream,
3432 cx: &mut App,
3433 ) -> Result<()>;
3434}
3435
3436impl<T> AnyAgentTool for Erased<Arc<T>>
3437where
3438 T: AgentTool,
3439{
3440 fn name(&self) -> SharedString {
3441 T::NAME.into()
3442 }
3443
3444 fn description(&self) -> SharedString {
3445 T::description()
3446 }
3447
3448 fn kind(&self) -> agent_client_protocol::ToolKind {
3449 T::kind()
3450 }
3451
3452 fn supports_input_streaming(&self) -> bool {
3453 T::supports_input_streaming()
3454 }
3455
3456 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
3457 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
3458 self.0.initial_title(parsed_input, _cx)
3459 }
3460
3461 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
3462 let mut json = serde_json::to_value(T::input_schema(format))?;
3463 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
3464 Ok(json)
3465 }
3466
3467 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
3468 T::supports_provider(provider)
3469 }
3470
3471 fn run(
3472 self: Arc<Self>,
3473 input: ToolInput<serde_json::Value>,
3474 event_stream: ToolCallEventStream,
3475 cx: &mut App,
3476 ) -> Task<Result<AgentToolOutput, AgentToolOutput>> {
3477 let tool_input: ToolInput<T::Input> = input.cast();
3478 let task = self.0.clone().run(tool_input, event_stream, cx);
3479 cx.spawn(async move |_cx| match task.await {
3480 Ok(output) => {
3481 let raw_output = serde_json::to_value(&output).map_err(|e| {
3482 AgentToolOutput::from_error(format!("Failed to serialize tool output: {e}"))
3483 })?;
3484 Ok(AgentToolOutput {
3485 llm_output: output.into(),
3486 raw_output,
3487 })
3488 }
3489 Err(error_output) => {
3490 let raw_output = serde_json::to_value(&error_output).unwrap_or_else(|e| {
3491 log::error!("Failed to serialize tool error output: {e}");
3492 serde_json::Value::Null
3493 });
3494 Err(AgentToolOutput {
3495 llm_output: error_output.into(),
3496 raw_output,
3497 })
3498 }
3499 })
3500 }
3501
3502 fn replay(
3503 &self,
3504 input: serde_json::Value,
3505 output: serde_json::Value,
3506 event_stream: ToolCallEventStream,
3507 cx: &mut App,
3508 ) -> Result<()> {
3509 let input = serde_json::from_value(input)?;
3510 let output = serde_json::from_value(output)?;
3511 self.0.replay(input, output, event_stream, cx)
3512 }
3513}
3514
3515#[derive(Clone)]
3516struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
3517
3518impl ThreadEventStream {
3519 fn send_user_message(&self, message: &UserMessage) {
3520 self.0
3521 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
3522 .ok();
3523 }
3524
3525 fn send_text(&self, text: &str) {
3526 self.0
3527 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
3528 .ok();
3529 }
3530
3531 fn send_thinking(&self, text: &str) {
3532 self.0
3533 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
3534 .ok();
3535 }
3536
3537 fn send_tool_call(
3538 &self,
3539 id: &LanguageModelToolUseId,
3540 tool_name: &str,
3541 title: SharedString,
3542 kind: acp::ToolKind,
3543 input: serde_json::Value,
3544 ) {
3545 self.0
3546 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
3547 id,
3548 tool_name,
3549 title.to_string(),
3550 kind,
3551 input,
3552 ))))
3553 .ok();
3554 }
3555
3556 fn initial_tool_call(
3557 id: &LanguageModelToolUseId,
3558 tool_name: &str,
3559 title: String,
3560 kind: acp::ToolKind,
3561 input: serde_json::Value,
3562 ) -> acp::ToolCall {
3563 acp::ToolCall::new(id.to_string(), title)
3564 .kind(kind)
3565 .raw_input(input)
3566 .meta(acp_thread::meta_with_tool_name(tool_name))
3567 }
3568
3569 fn update_tool_call_fields(
3570 &self,
3571 tool_use_id: &LanguageModelToolUseId,
3572 fields: acp::ToolCallUpdateFields,
3573 meta: Option<acp::Meta>,
3574 ) {
3575 self.0
3576 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3577 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields)
3578 .meta(meta)
3579 .into(),
3580 )))
3581 .ok();
3582 }
3583
3584 fn send_plan(&self, plan: acp::Plan) {
3585 self.0.unbounded_send(Ok(ThreadEvent::Plan(plan))).ok();
3586 }
3587
3588 fn send_retry(&self, status: acp_thread::RetryStatus) {
3589 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
3590 }
3591
3592 fn send_stop(&self, reason: acp::StopReason) {
3593 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
3594 }
3595
3596 fn send_canceled(&self) {
3597 self.0
3598 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
3599 .ok();
3600 }
3601
3602 fn send_error(&self, error: impl Into<anyhow::Error>) {
3603 self.0.unbounded_send(Err(error.into())).ok();
3604 }
3605}
3606
3607#[derive(Clone)]
3608pub struct ToolCallEventStream {
3609 tool_use_id: LanguageModelToolUseId,
3610 stream: ThreadEventStream,
3611 fs: Option<Arc<dyn Fs>>,
3612 cancellation_rx: watch::Receiver<bool>,
3613}
3614
3615impl ToolCallEventStream {
3616 #[cfg(any(test, feature = "test-support"))]
3617 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
3618 let (stream, receiver, _cancellation_tx) = Self::test_with_cancellation();
3619 (stream, receiver)
3620 }
3621
3622 #[cfg(any(test, feature = "test-support"))]
3623 pub fn test_with_cancellation() -> (Self, ToolCallEventStreamReceiver, watch::Sender<bool>) {
3624 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
3625 let (cancellation_tx, cancellation_rx) = watch::channel(false);
3626
3627 let stream = ToolCallEventStream::new(
3628 "test_id".into(),
3629 ThreadEventStream(events_tx),
3630 None,
3631 cancellation_rx,
3632 );
3633
3634 (
3635 stream,
3636 ToolCallEventStreamReceiver(events_rx),
3637 cancellation_tx,
3638 )
3639 }
3640
3641 /// Signal cancellation for this event stream. Only available in tests.
3642 #[cfg(any(test, feature = "test-support"))]
3643 pub fn signal_cancellation_with_sender(cancellation_tx: &mut watch::Sender<bool>) {
3644 cancellation_tx.send(true).ok();
3645 }
3646
3647 fn new(
3648 tool_use_id: LanguageModelToolUseId,
3649 stream: ThreadEventStream,
3650 fs: Option<Arc<dyn Fs>>,
3651 cancellation_rx: watch::Receiver<bool>,
3652 ) -> Self {
3653 Self {
3654 tool_use_id,
3655 stream,
3656 fs,
3657 cancellation_rx,
3658 }
3659 }
3660
3661 /// Returns a future that resolves when the user cancels the tool call.
3662 /// Tools should select on this alongside their main work to detect user cancellation.
3663 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
3664 let mut rx = self.cancellation_rx.clone();
3665 async move {
3666 loop {
3667 if *rx.borrow() {
3668 return;
3669 }
3670 if rx.changed().await.is_err() {
3671 // Sender dropped, will never be cancelled
3672 std::future::pending::<()>().await;
3673 }
3674 }
3675 }
3676 }
3677
3678 /// Returns true if the user has cancelled this tool call.
3679 /// This is useful for checking cancellation state after an operation completes,
3680 /// to determine if the completion was due to user cancellation.
3681 pub fn was_cancelled_by_user(&self) -> bool {
3682 *self.cancellation_rx.clone().borrow()
3683 }
3684
3685 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
3686 &self.tool_use_id
3687 }
3688
3689 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
3690 self.stream
3691 .update_tool_call_fields(&self.tool_use_id, fields, None);
3692 }
3693
3694 pub fn update_fields_with_meta(
3695 &self,
3696 fields: acp::ToolCallUpdateFields,
3697 meta: Option<acp::Meta>,
3698 ) {
3699 self.stream
3700 .update_tool_call_fields(&self.tool_use_id, fields, meta);
3701 }
3702
3703 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
3704 self.stream
3705 .0
3706 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3707 acp_thread::ToolCallUpdateDiff {
3708 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3709 diff,
3710 }
3711 .into(),
3712 )))
3713 .ok();
3714 }
3715
3716 pub fn subagent_spawned(&self, id: acp::SessionId) {
3717 self.stream
3718 .0
3719 .unbounded_send(Ok(ThreadEvent::SubagentSpawned(id)))
3720 .ok();
3721 }
3722
3723 pub fn update_plan(&self, plan: acp::Plan) {
3724 self.stream.send_plan(plan);
3725 }
3726
3727 /// Authorize a third-party tool (e.g., MCP tool from a context server).
3728 ///
3729 /// Unlike built-in tools, third-party tools don't support pattern-based permissions.
3730 /// They only support `default` (allow/deny/confirm) per tool.
3731 ///
3732 /// Uses the dropdown authorization flow with two granularities:
3733 /// - "Always for <display_name> MCP tool" → sets `tools.<tool_id>.default = "allow"` or "deny"
3734 /// - "Only this time" → allow/deny once
3735 pub fn authorize_third_party_tool(
3736 &self,
3737 title: impl Into<String>,
3738 tool_id: String,
3739 display_name: String,
3740 cx: &mut App,
3741 ) -> Task<Result<()>> {
3742 let settings = agent_settings::AgentSettings::get_global(cx);
3743
3744 let decision = decide_permission_from_settings(&tool_id, &[String::new()], &settings);
3745
3746 match decision {
3747 ToolPermissionDecision::Allow => return Task::ready(Ok(())),
3748 ToolPermissionDecision::Deny(reason) => return Task::ready(Err(anyhow!(reason))),
3749 ToolPermissionDecision::Confirm => {}
3750 }
3751
3752 let (response_tx, response_rx) = oneshot::channel();
3753 if let Err(error) = self
3754 .stream
3755 .0
3756 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3757 ToolCallAuthorization {
3758 tool_call: acp::ToolCallUpdate::new(
3759 self.tool_use_id.to_string(),
3760 acp::ToolCallUpdateFields::new().title(title.into()),
3761 ),
3762 options: acp_thread::PermissionOptions::Dropdown(vec![
3763 acp_thread::PermissionOptionChoice {
3764 allow: acp::PermissionOption::new(
3765 acp::PermissionOptionId::new(format!(
3766 "always_allow_mcp:{}",
3767 tool_id
3768 )),
3769 format!("Always for {} MCP tool", display_name),
3770 acp::PermissionOptionKind::AllowAlways,
3771 ),
3772 deny: acp::PermissionOption::new(
3773 acp::PermissionOptionId::new(format!(
3774 "always_deny_mcp:{}",
3775 tool_id
3776 )),
3777 format!("Always for {} MCP tool", display_name),
3778 acp::PermissionOptionKind::RejectAlways,
3779 ),
3780 sub_patterns: vec![],
3781 },
3782 acp_thread::PermissionOptionChoice {
3783 allow: acp::PermissionOption::new(
3784 acp::PermissionOptionId::new("allow"),
3785 "Only this time",
3786 acp::PermissionOptionKind::AllowOnce,
3787 ),
3788 deny: acp::PermissionOption::new(
3789 acp::PermissionOptionId::new("deny"),
3790 "Only this time",
3791 acp::PermissionOptionKind::RejectOnce,
3792 ),
3793 sub_patterns: vec![],
3794 },
3795 ]),
3796 response: response_tx,
3797 context: None,
3798 },
3799 )))
3800 {
3801 log::error!("Failed to send tool call authorization: {error}");
3802 return Task::ready(Err(anyhow!(
3803 "Failed to send tool call authorization: {error}"
3804 )));
3805 }
3806
3807 let fs = self.fs.clone();
3808 cx.spawn(async move |cx| {
3809 let outcome = response_rx.await?;
3810 let is_allow = Self::persist_permission_outcome(&outcome, fs, &cx);
3811 if is_allow {
3812 Ok(())
3813 } else {
3814 Err(anyhow!("Permission to run tool denied by user"))
3815 }
3816 })
3817 }
3818
3819 pub fn authorize(
3820 &self,
3821 title: impl Into<String>,
3822 context: ToolPermissionContext,
3823 cx: &mut App,
3824 ) -> Task<Result<()>> {
3825 let options = context.build_permission_options();
3826
3827 let (response_tx, response_rx) = oneshot::channel();
3828 if let Err(error) = self
3829 .stream
3830 .0
3831 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3832 ToolCallAuthorization {
3833 tool_call: acp::ToolCallUpdate::new(
3834 self.tool_use_id.to_string(),
3835 acp::ToolCallUpdateFields::new().title(title.into()),
3836 ),
3837 options,
3838 response: response_tx,
3839 context: Some(context),
3840 },
3841 )))
3842 {
3843 log::error!("Failed to send tool call authorization: {error}");
3844 return Task::ready(Err(anyhow!(
3845 "Failed to send tool call authorization: {error}"
3846 )));
3847 }
3848
3849 let fs = self.fs.clone();
3850 cx.spawn(async move |cx| {
3851 let outcome = response_rx.await?;
3852 let is_allow = Self::persist_permission_outcome(&outcome, fs, &cx);
3853 if is_allow {
3854 Ok(())
3855 } else {
3856 Err(anyhow!("Permission to run tool denied by user"))
3857 }
3858 })
3859 }
3860
3861 /// Interprets a `SelectedPermissionOutcome` and persists any settings changes.
3862 /// Returns `true` if the tool call should be allowed, `false` if denied.
3863 fn persist_permission_outcome(
3864 outcome: &acp_thread::SelectedPermissionOutcome,
3865 fs: Option<Arc<dyn Fs>>,
3866 cx: &AsyncApp,
3867 ) -> bool {
3868 let option_id = outcome.option_id.0.as_ref();
3869
3870 let always_permission = option_id
3871 .strip_prefix("always_allow:")
3872 .map(|tool| (tool, ToolPermissionMode::Allow))
3873 .or_else(|| {
3874 option_id
3875 .strip_prefix("always_deny:")
3876 .map(|tool| (tool, ToolPermissionMode::Deny))
3877 })
3878 .or_else(|| {
3879 option_id
3880 .strip_prefix("always_allow_mcp:")
3881 .map(|tool| (tool, ToolPermissionMode::Allow))
3882 })
3883 .or_else(|| {
3884 option_id
3885 .strip_prefix("always_deny_mcp:")
3886 .map(|tool| (tool, ToolPermissionMode::Deny))
3887 });
3888
3889 if let Some((tool, mode)) = always_permission {
3890 let params = outcome.params.as_ref();
3891 Self::persist_always_permission(tool, mode, params, fs, cx);
3892 return mode == ToolPermissionMode::Allow;
3893 }
3894
3895 // Handle simple "allow" / "deny" (once, no persistence)
3896 if option_id == "allow" || option_id == "deny" {
3897 debug_assert!(
3898 outcome.params.is_none(),
3899 "unexpected params for once-only permission"
3900 );
3901 return option_id == "allow";
3902 }
3903
3904 debug_assert!(false, "unexpected permission option_id: {option_id}");
3905 false
3906 }
3907
3908 /// Persists an "always allow" or "always deny" permission, using sub_patterns
3909 /// from params when present.
3910 fn persist_always_permission(
3911 tool: &str,
3912 mode: ToolPermissionMode,
3913 params: Option<&acp_thread::SelectedPermissionParams>,
3914 fs: Option<Arc<dyn Fs>>,
3915 cx: &AsyncApp,
3916 ) {
3917 let Some(fs) = fs else {
3918 return;
3919 };
3920
3921 match params {
3922 Some(acp_thread::SelectedPermissionParams::Terminal {
3923 patterns: sub_patterns,
3924 }) => {
3925 debug_assert!(
3926 !sub_patterns.is_empty(),
3927 "empty sub_patterns for tool {tool} — callers should pass None instead"
3928 );
3929 let tool = tool.to_string();
3930 let sub_patterns = sub_patterns.clone();
3931 cx.update(|cx| {
3932 update_settings_file(fs, cx, move |settings, _| {
3933 let agent = settings.agent.get_or_insert_default();
3934 for pattern in sub_patterns {
3935 match mode {
3936 ToolPermissionMode::Allow => {
3937 agent.add_tool_allow_pattern(&tool, pattern);
3938 }
3939 ToolPermissionMode::Deny => {
3940 agent.add_tool_deny_pattern(&tool, pattern);
3941 }
3942 // If there's no matching pattern this will
3943 // default to confirm, so falling through is
3944 // fine here.
3945 ToolPermissionMode::Confirm => (),
3946 }
3947 }
3948 });
3949 });
3950 }
3951 None => {
3952 let tool = tool.to_string();
3953 cx.update(|cx| {
3954 update_settings_file(fs, cx, move |settings, _| {
3955 settings
3956 .agent
3957 .get_or_insert_default()
3958 .set_tool_default_permission(&tool, mode);
3959 });
3960 });
3961 }
3962 }
3963 }
3964}
3965
3966#[cfg(any(test, feature = "test-support"))]
3967pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
3968
3969#[cfg(any(test, feature = "test-support"))]
3970impl ToolCallEventStreamReceiver {
3971 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
3972 let event = self.0.next().await;
3973 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
3974 auth
3975 } else {
3976 panic!("Expected ToolCallAuthorization but got: {:?}", event);
3977 }
3978 }
3979
3980 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
3981 let event = self.0.next().await;
3982 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
3983 update,
3984 )))) = event
3985 {
3986 update.fields
3987 } else {
3988 panic!("Expected update fields but got: {:?}", event);
3989 }
3990 }
3991
3992 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
3993 let event = self.0.next().await;
3994 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
3995 update,
3996 )))) = event
3997 {
3998 update.diff
3999 } else {
4000 panic!("Expected diff but got: {:?}", event);
4001 }
4002 }
4003
4004 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
4005 let event = self.0.next().await;
4006 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
4007 update,
4008 )))) = event
4009 {
4010 update.terminal
4011 } else {
4012 panic!("Expected terminal but got: {:?}", event);
4013 }
4014 }
4015
4016 pub async fn expect_plan(&mut self) -> acp::Plan {
4017 let event = self.0.next().await;
4018 if let Some(Ok(ThreadEvent::Plan(plan))) = event {
4019 plan
4020 } else {
4021 panic!("Expected plan but got: {:?}", event);
4022 }
4023 }
4024}
4025
4026#[cfg(any(test, feature = "test-support"))]
4027impl std::ops::Deref for ToolCallEventStreamReceiver {
4028 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
4029
4030 fn deref(&self) -> &Self::Target {
4031 &self.0
4032 }
4033}
4034
4035#[cfg(any(test, feature = "test-support"))]
4036impl std::ops::DerefMut for ToolCallEventStreamReceiver {
4037 fn deref_mut(&mut self) -> &mut Self::Target {
4038 &mut self.0
4039 }
4040}
4041
4042impl From<&str> for UserMessageContent {
4043 fn from(text: &str) -> Self {
4044 Self::Text(text.into())
4045 }
4046}
4047
4048impl From<String> for UserMessageContent {
4049 fn from(text: String) -> Self {
4050 Self::Text(text)
4051 }
4052}
4053
4054impl UserMessageContent {
4055 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
4056 match value {
4057 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
4058 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
4059 acp::ContentBlock::Audio(_) => {
4060 // TODO
4061 Self::Text("[audio]".to_string())
4062 }
4063 acp::ContentBlock::ResourceLink(resource_link) => {
4064 match MentionUri::parse(&resource_link.uri, path_style) {
4065 Ok(uri) => Self::Mention {
4066 uri,
4067 content: String::new(),
4068 },
4069 Err(err) => {
4070 log::error!("Failed to parse mention link: {}", err);
4071 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
4072 }
4073 }
4074 }
4075 acp::ContentBlock::Resource(resource) => match resource.resource {
4076 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
4077 match MentionUri::parse(&resource.uri, path_style) {
4078 Ok(uri) => Self::Mention {
4079 uri,
4080 content: resource.text,
4081 },
4082 Err(err) => {
4083 log::error!("Failed to parse mention link: {}", err);
4084 Self::Text(
4085 MarkdownCodeBlock {
4086 tag: &resource.uri,
4087 text: &resource.text,
4088 }
4089 .to_string(),
4090 )
4091 }
4092 }
4093 }
4094 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
4095 // TODO
4096 Self::Text("[blob]".to_string())
4097 }
4098 other => {
4099 log::warn!("Unexpected content type: {:?}", other);
4100 Self::Text("[unknown]".to_string())
4101 }
4102 },
4103 other => {
4104 log::warn!("Unexpected content type: {:?}", other);
4105 Self::Text("[unknown]".to_string())
4106 }
4107 }
4108 }
4109}
4110
4111impl From<UserMessageContent> for acp::ContentBlock {
4112 fn from(content: UserMessageContent) -> Self {
4113 match content {
4114 UserMessageContent::Text(text) => text.into(),
4115 UserMessageContent::Image(image) => {
4116 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
4117 }
4118 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
4119 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
4120 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
4121 )),
4122 ),
4123 }
4124 }
4125}
4126
4127fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
4128 LanguageModelImage {
4129 source: image_content.data.into(),
4130 size: None,
4131 }
4132}
4133
4134#[cfg(test)]
4135mod tests {
4136 use super::*;
4137 use gpui::TestAppContext;
4138 use language_model::LanguageModelToolUseId;
4139 use language_model::fake_provider::FakeLanguageModel;
4140 use serde_json::json;
4141 use std::sync::Arc;
4142
4143 async fn setup_thread_for_test(cx: &mut TestAppContext) -> (Entity<Thread>, ThreadEventStream) {
4144 cx.update(|cx| {
4145 let settings_store = settings::SettingsStore::test(cx);
4146 cx.set_global(settings_store);
4147 });
4148
4149 let fs = fs::FakeFs::new(cx.background_executor.clone());
4150 let templates = Templates::new();
4151 let project = Project::test(fs.clone(), [], cx).await;
4152
4153 cx.update(|cx| {
4154 let project_context = cx.new(|_cx| prompt_store::ProjectContext::default());
4155 let context_server_store = project.read(cx).context_server_store();
4156 let context_server_registry =
4157 cx.new(|cx| ContextServerRegistry::new(context_server_store, cx));
4158
4159 let thread = cx.new(|cx| {
4160 Thread::new(
4161 project,
4162 project_context,
4163 context_server_registry,
4164 templates,
4165 None,
4166 cx,
4167 )
4168 });
4169
4170 let (event_tx, _event_rx) = mpsc::unbounded();
4171 let event_stream = ThreadEventStream(event_tx);
4172
4173 (thread, event_stream)
4174 })
4175 }
4176
4177 fn setup_parent_with_subagents(
4178 cx: &mut TestAppContext,
4179 parent: &Entity<Thread>,
4180 count: usize,
4181 ) -> Vec<Entity<Thread>> {
4182 cx.update(|cx| {
4183 let mut subagents = Vec::new();
4184 for _ in 0..count {
4185 let subagent = cx.new(|cx| Thread::new_subagent(parent, cx));
4186 parent.update(cx, |thread, _cx| {
4187 thread.register_running_subagent(subagent.downgrade());
4188 });
4189 subagents.push(subagent);
4190 }
4191 subagents
4192 })
4193 }
4194
4195 #[gpui::test]
4196 async fn test_set_model_propagates_to_subagents(cx: &mut TestAppContext) {
4197 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4198 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4199
4200 let new_model: Arc<dyn LanguageModel> = Arc::new(FakeLanguageModel::with_id_and_thinking(
4201 "test-provider",
4202 "new-model",
4203 "New Model",
4204 false,
4205 ));
4206
4207 cx.update(|cx| {
4208 parent.update(cx, |thread, cx| {
4209 thread.set_model(new_model, cx);
4210 });
4211
4212 for subagent in &subagents {
4213 let subagent_model_id = subagent.read(cx).model().unwrap().id();
4214 assert_eq!(
4215 subagent_model_id.0.as_ref(),
4216 "new-model",
4217 "Subagent model should match parent model after set_model"
4218 );
4219 }
4220 });
4221 }
4222
4223 #[gpui::test]
4224 async fn test_set_summarization_model_propagates_to_subagents(cx: &mut TestAppContext) {
4225 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4226 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4227
4228 let summary_model: Arc<dyn LanguageModel> =
4229 Arc::new(FakeLanguageModel::with_id_and_thinking(
4230 "test-provider",
4231 "summary-model",
4232 "Summary Model",
4233 false,
4234 ));
4235
4236 cx.update(|cx| {
4237 parent.update(cx, |thread, cx| {
4238 thread.set_summarization_model(Some(summary_model), cx);
4239 });
4240
4241 for subagent in &subagents {
4242 let subagent_summary_id = subagent.read(cx).summarization_model().unwrap().id();
4243 assert_eq!(
4244 subagent_summary_id.0.as_ref(),
4245 "summary-model",
4246 "Subagent summarization model should match parent after set_summarization_model"
4247 );
4248 }
4249 });
4250 }
4251
4252 #[gpui::test]
4253 async fn test_set_thinking_enabled_propagates_to_subagents(cx: &mut TestAppContext) {
4254 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4255 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4256
4257 cx.update(|cx| {
4258 parent.update(cx, |thread, cx| {
4259 thread.set_thinking_enabled(true, cx);
4260 });
4261
4262 for subagent in &subagents {
4263 assert!(
4264 subagent.read(cx).thinking_enabled(),
4265 "Subagent thinking should be enabled after parent enables it"
4266 );
4267 }
4268
4269 parent.update(cx, |thread, cx| {
4270 thread.set_thinking_enabled(false, cx);
4271 });
4272
4273 for subagent in &subagents {
4274 assert!(
4275 !subagent.read(cx).thinking_enabled(),
4276 "Subagent thinking should be disabled after parent disables it"
4277 );
4278 }
4279 });
4280 }
4281
4282 #[gpui::test]
4283 async fn test_set_thinking_effort_propagates_to_subagents(cx: &mut TestAppContext) {
4284 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4285 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4286
4287 cx.update(|cx| {
4288 parent.update(cx, |thread, cx| {
4289 thread.set_thinking_effort(Some("high".to_string()), cx);
4290 });
4291
4292 for subagent in &subagents {
4293 assert_eq!(
4294 subagent.read(cx).thinking_effort().map(|s| s.as_str()),
4295 Some("high"),
4296 "Subagent thinking effort should match parent"
4297 );
4298 }
4299
4300 parent.update(cx, |thread, cx| {
4301 thread.set_thinking_effort(None, cx);
4302 });
4303
4304 for subagent in &subagents {
4305 assert_eq!(
4306 subagent.read(cx).thinking_effort(),
4307 None,
4308 "Subagent thinking effort should be None after parent clears it"
4309 );
4310 }
4311 });
4312 }
4313
4314 #[gpui::test]
4315 async fn test_subagent_inherits_settings_at_creation(cx: &mut TestAppContext) {
4316 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4317
4318 cx.update(|cx| {
4319 parent.update(cx, |thread, cx| {
4320 thread.set_speed(Speed::Fast, cx);
4321 thread.set_thinking_enabled(true, cx);
4322 thread.set_thinking_effort(Some("high".to_string()), cx);
4323 thread.set_profile(AgentProfileId("custom-profile".into()), cx);
4324 });
4325 });
4326
4327 let subagents = setup_parent_with_subagents(cx, &parent, 1);
4328
4329 cx.update(|cx| {
4330 let sub = subagents[0].read(cx);
4331 assert_eq!(sub.speed(), Some(Speed::Fast));
4332 assert!(sub.thinking_enabled());
4333 assert_eq!(sub.thinking_effort().map(|s| s.as_str()), Some("high"));
4334 assert_eq!(sub.profile(), &AgentProfileId("custom-profile".into()));
4335 });
4336 }
4337
4338 #[gpui::test]
4339 async fn test_set_speed_propagates_to_subagents(cx: &mut TestAppContext) {
4340 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4341 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4342
4343 cx.update(|cx| {
4344 parent.update(cx, |thread, cx| {
4345 thread.set_speed(Speed::Fast, cx);
4346 });
4347
4348 for subagent in &subagents {
4349 assert_eq!(
4350 subagent.read(cx).speed(),
4351 Some(Speed::Fast),
4352 "Subagent speed should match parent after set_speed"
4353 );
4354 }
4355 });
4356 }
4357
4358 #[gpui::test]
4359 async fn test_dropped_subagent_does_not_panic(cx: &mut TestAppContext) {
4360 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4361 let subagents = setup_parent_with_subagents(cx, &parent, 1);
4362
4363 // Drop the subagent so the WeakEntity can no longer be upgraded
4364 drop(subagents);
4365
4366 // Should not panic even though the subagent was dropped
4367 cx.update(|cx| {
4368 parent.update(cx, |thread, cx| {
4369 thread.set_thinking_enabled(true, cx);
4370 thread.set_speed(Speed::Fast, cx);
4371 thread.set_thinking_effort(Some("high".to_string()), cx);
4372 });
4373 });
4374 }
4375
4376 #[gpui::test]
4377 async fn test_handle_tool_use_json_parse_error_adds_tool_use_to_content(
4378 cx: &mut TestAppContext,
4379 ) {
4380 let (thread, event_stream) = setup_thread_for_test(cx).await;
4381
4382 let tool_use_id = LanguageModelToolUseId::from("test_tool_id");
4383 let tool_name: Arc<str> = Arc::from("test_tool");
4384 let raw_input: Arc<str> = Arc::from("{invalid json");
4385 let json_parse_error = "expected value at line 1 column 1".to_string();
4386
4387 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
4388
4389 let result = cx
4390 .update(|cx| {
4391 thread.update(cx, |thread, cx| {
4392 // Call the function under test
4393 thread
4394 .handle_tool_use_json_parse_error_event(
4395 tool_use_id.clone(),
4396 tool_name.clone(),
4397 raw_input.clone(),
4398 json_parse_error,
4399 &event_stream,
4400 cancellation_rx,
4401 cx,
4402 )
4403 .unwrap()
4404 })
4405 })
4406 .await;
4407
4408 // Verify the result is an error
4409 assert!(result.is_error);
4410 assert_eq!(result.tool_use_id, tool_use_id);
4411 assert_eq!(result.tool_name, tool_name);
4412 assert!(matches!(
4413 result.content,
4414 LanguageModelToolResultContent::Text(_)
4415 ));
4416
4417 thread.update(cx, |thread, _cx| {
4418 // Verify the tool use was added to the message content
4419 {
4420 let last_message = thread.pending_message();
4421 assert_eq!(
4422 last_message.content.len(),
4423 1,
4424 "Should have one tool_use in content"
4425 );
4426
4427 match &last_message.content[0] {
4428 AgentMessageContent::ToolUse(tool_use) => {
4429 assert_eq!(tool_use.id, tool_use_id);
4430 assert_eq!(tool_use.name, tool_name);
4431 assert_eq!(tool_use.raw_input, raw_input.to_string());
4432 assert!(tool_use.is_input_complete);
4433 // Should fall back to empty object for invalid JSON
4434 assert_eq!(tool_use.input, json!({}));
4435 }
4436 _ => panic!("Expected ToolUse content"),
4437 }
4438 }
4439
4440 // Insert the tool result (simulating what the caller does)
4441 thread
4442 .pending_message()
4443 .tool_results
4444 .insert(result.tool_use_id.clone(), result);
4445
4446 // Verify the tool result was added
4447 let last_message = thread.pending_message();
4448 assert_eq!(
4449 last_message.tool_results.len(),
4450 1,
4451 "Should have one tool_result"
4452 );
4453 assert!(last_message.tool_results.contains_key(&tool_use_id));
4454 })
4455 }
4456}