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