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
1003 }
1004
1005 pub fn new(
1006 project: Entity<Project>,
1007 project_context: Entity<ProjectContext>,
1008 context_server_registry: Entity<ContextServerRegistry>,
1009 templates: Arc<Templates>,
1010 model: Option<Arc<dyn LanguageModel>>,
1011 cx: &mut Context<Self>,
1012 ) -> Self {
1013 Self::new_internal(
1014 project.clone(),
1015 project_context,
1016 context_server_registry,
1017 templates,
1018 model,
1019 cx.new(|_cx| ActionLog::new(project)),
1020 cx,
1021 )
1022 }
1023
1024 fn new_internal(
1025 project: Entity<Project>,
1026 project_context: Entity<ProjectContext>,
1027 context_server_registry: Entity<ContextServerRegistry>,
1028 templates: Arc<Templates>,
1029 model: Option<Arc<dyn LanguageModel>>,
1030 action_log: Entity<ActionLog>,
1031 cx: &mut Context<Self>,
1032 ) -> Self {
1033 let settings = AgentSettings::get_global(cx);
1034 let profile_id = settings.default_profile.clone();
1035 let enable_thinking = settings
1036 .default_model
1037 .as_ref()
1038 .is_some_and(|model| model.enable_thinking);
1039 let thinking_effort = settings
1040 .default_model
1041 .as_ref()
1042 .and_then(|model| model.effort.clone());
1043 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1044 watch::channel(Self::prompt_capabilities(model.as_deref()));
1045 Self {
1046 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
1047 prompt_id: PromptId::new(),
1048 updated_at: Utc::now(),
1049 title: None,
1050 pending_title_generation: None,
1051 pending_summary_generation: None,
1052 summary: None,
1053 messages: Vec::new(),
1054 user_store: project.read(cx).user_store(),
1055 running_turn: None,
1056 has_queued_message: false,
1057 pending_message: None,
1058 tools: BTreeMap::default(),
1059 request_token_usage: HashMap::default(),
1060 cumulative_token_usage: TokenUsage::default(),
1061 initial_project_snapshot: {
1062 let project_snapshot = Self::project_snapshot(project.clone(), cx);
1063 cx.foreground_executor()
1064 .spawn(async move { Some(project_snapshot.await) })
1065 .shared()
1066 },
1067 context_server_registry,
1068 profile_id,
1069 project_context,
1070 templates,
1071 model,
1072 summarization_model: None,
1073 thinking_enabled: enable_thinking,
1074 speed: None,
1075 thinking_effort,
1076 prompt_capabilities_tx,
1077 prompt_capabilities_rx,
1078 project,
1079 action_log,
1080 imported: false,
1081 subagent_context: None,
1082 draft_prompt: None,
1083 ui_scroll_position: None,
1084 running_subagents: Vec::new(),
1085 }
1086 }
1087
1088 pub fn id(&self) -> &acp::SessionId {
1089 &self.id
1090 }
1091
1092 /// Returns true if this thread was imported from a shared thread.
1093 pub fn is_imported(&self) -> bool {
1094 self.imported
1095 }
1096
1097 pub fn replay(
1098 &mut self,
1099 cx: &mut Context<Self>,
1100 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
1101 let (tx, rx) = mpsc::unbounded();
1102 let stream = ThreadEventStream(tx);
1103 for message in &self.messages {
1104 match message {
1105 Message::User(user_message) => stream.send_user_message(user_message),
1106 Message::Agent(assistant_message) => {
1107 for content in &assistant_message.content {
1108 match content {
1109 AgentMessageContent::Text(text) => stream.send_text(text),
1110 AgentMessageContent::Thinking { text, .. } => {
1111 stream.send_thinking(text)
1112 }
1113 AgentMessageContent::RedactedThinking(_) => {}
1114 AgentMessageContent::ToolUse(tool_use) => {
1115 self.replay_tool_call(
1116 tool_use,
1117 assistant_message.tool_results.get(&tool_use.id),
1118 &stream,
1119 cx,
1120 );
1121 }
1122 }
1123 }
1124 }
1125 Message::Resume => {}
1126 }
1127 }
1128 rx
1129 }
1130
1131 fn replay_tool_call(
1132 &self,
1133 tool_use: &LanguageModelToolUse,
1134 tool_result: Option<&LanguageModelToolResult>,
1135 stream: &ThreadEventStream,
1136 cx: &mut Context<Self>,
1137 ) {
1138 // Extract saved output and status first, so they're available even if tool is not found
1139 let output = tool_result
1140 .as_ref()
1141 .and_then(|result| result.output.clone());
1142 let status = tool_result
1143 .as_ref()
1144 .map_or(acp::ToolCallStatus::Failed, |result| {
1145 if result.is_error {
1146 acp::ToolCallStatus::Failed
1147 } else {
1148 acp::ToolCallStatus::Completed
1149 }
1150 });
1151
1152 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
1153 self.context_server_registry
1154 .read(cx)
1155 .servers()
1156 .find_map(|(_, tools)| {
1157 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
1158 Some(tool.clone())
1159 } else {
1160 None
1161 }
1162 })
1163 });
1164
1165 let Some(tool) = tool else {
1166 // Tool not found (e.g., MCP server not connected after restart),
1167 // but still display the saved result if available.
1168 // We need to send both ToolCall and ToolCallUpdate events because the UI
1169 // only converts raw_output to displayable content in update_fields, not from_acp.
1170 stream
1171 .0
1172 .unbounded_send(Ok(ThreadEvent::ToolCall(
1173 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
1174 .status(status)
1175 .raw_input(tool_use.input.clone()),
1176 )))
1177 .ok();
1178 stream.update_tool_call_fields(
1179 &tool_use.id,
1180 acp::ToolCallUpdateFields::new()
1181 .status(status)
1182 .raw_output(output),
1183 None,
1184 );
1185 return;
1186 };
1187
1188 let title = tool.initial_title(tool_use.input.clone(), cx);
1189 let kind = tool.kind();
1190 stream.send_tool_call(
1191 &tool_use.id,
1192 &tool_use.name,
1193 title,
1194 kind,
1195 tool_use.input.clone(),
1196 );
1197
1198 if let Some(output) = output.clone() {
1199 // For replay, we use a dummy cancellation receiver since the tool already completed
1200 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
1201 let tool_event_stream = ToolCallEventStream::new(
1202 tool_use.id.clone(),
1203 stream.clone(),
1204 Some(self.project.read(cx).fs().clone()),
1205 cancellation_rx,
1206 );
1207 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
1208 .log_err();
1209 }
1210
1211 stream.update_tool_call_fields(
1212 &tool_use.id,
1213 acp::ToolCallUpdateFields::new()
1214 .status(status)
1215 .raw_output(output),
1216 None,
1217 );
1218 }
1219
1220 pub fn from_db(
1221 id: acp::SessionId,
1222 db_thread: DbThread,
1223 project: Entity<Project>,
1224 project_context: Entity<ProjectContext>,
1225 context_server_registry: Entity<ContextServerRegistry>,
1226 templates: Arc<Templates>,
1227 cx: &mut Context<Self>,
1228 ) -> Self {
1229 let settings = AgentSettings::get_global(cx);
1230 let profile_id = db_thread
1231 .profile
1232 .unwrap_or_else(|| settings.default_profile.clone());
1233
1234 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1235 db_thread
1236 .model
1237 .and_then(|model| {
1238 let model = SelectedModel {
1239 provider: model.provider.clone().into(),
1240 model: model.model.into(),
1241 };
1242 registry.select_model(&model, cx)
1243 })
1244 .or_else(|| registry.default_model())
1245 .map(|model| model.model)
1246 });
1247
1248 if model.is_none() {
1249 model = Self::resolve_profile_model(&profile_id, cx);
1250 }
1251 if model.is_none() {
1252 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
1253 registry.default_model().map(|model| model.model)
1254 });
1255 }
1256
1257 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1258 watch::channel(Self::prompt_capabilities(model.as_deref()));
1259
1260 let action_log = cx.new(|_| ActionLog::new(project.clone()));
1261
1262 Self {
1263 id,
1264 prompt_id: PromptId::new(),
1265 title: if db_thread.title.is_empty() {
1266 None
1267 } else {
1268 Some(db_thread.title.clone())
1269 },
1270 pending_title_generation: None,
1271 pending_summary_generation: None,
1272 summary: db_thread.detailed_summary,
1273 messages: db_thread.messages,
1274 user_store: project.read(cx).user_store(),
1275 running_turn: None,
1276 has_queued_message: false,
1277 pending_message: None,
1278 tools: BTreeMap::default(),
1279 request_token_usage: db_thread.request_token_usage.clone(),
1280 cumulative_token_usage: db_thread.cumulative_token_usage,
1281 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
1282 context_server_registry,
1283 profile_id,
1284 project_context,
1285 templates,
1286 model,
1287 summarization_model: None,
1288 thinking_enabled: db_thread.thinking_enabled,
1289 thinking_effort: db_thread.thinking_effort,
1290 speed: db_thread.speed,
1291 project,
1292 action_log,
1293 updated_at: db_thread.updated_at,
1294 prompt_capabilities_tx,
1295 prompt_capabilities_rx,
1296 imported: db_thread.imported,
1297 subagent_context: db_thread.subagent_context,
1298 draft_prompt: db_thread.draft_prompt,
1299 ui_scroll_position: db_thread.ui_scroll_position.map(|sp| gpui::ListOffset {
1300 item_ix: sp.item_ix,
1301 offset_in_item: gpui::px(sp.offset_in_item),
1302 }),
1303 running_subagents: Vec::new(),
1304 }
1305 }
1306
1307 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
1308 let initial_project_snapshot = self.initial_project_snapshot.clone();
1309 let mut thread = DbThread {
1310 title: self.title().unwrap_or_default(),
1311 messages: self.messages.clone(),
1312 updated_at: self.updated_at,
1313 detailed_summary: self.summary.clone(),
1314 initial_project_snapshot: None,
1315 cumulative_token_usage: self.cumulative_token_usage,
1316 request_token_usage: self.request_token_usage.clone(),
1317 model: self.model.as_ref().map(|model| DbLanguageModel {
1318 provider: model.provider_id().to_string(),
1319 model: model.id().0.to_string(),
1320 }),
1321 profile: Some(self.profile_id.clone()),
1322 imported: self.imported,
1323 subagent_context: self.subagent_context.clone(),
1324 speed: self.speed,
1325 thinking_enabled: self.thinking_enabled,
1326 thinking_effort: self.thinking_effort.clone(),
1327 draft_prompt: self.draft_prompt.clone(),
1328 ui_scroll_position: self.ui_scroll_position.map(|lo| {
1329 crate::db::SerializedScrollPosition {
1330 item_ix: lo.item_ix,
1331 offset_in_item: lo.offset_in_item.as_f32(),
1332 }
1333 }),
1334 };
1335
1336 cx.background_spawn(async move {
1337 let initial_project_snapshot = initial_project_snapshot.await;
1338 thread.initial_project_snapshot = initial_project_snapshot;
1339 thread
1340 })
1341 }
1342
1343 /// Create a snapshot of the current project state including git information and unsaved buffers.
1344 fn project_snapshot(
1345 project: Entity<Project>,
1346 cx: &mut Context<Self>,
1347 ) -> Task<Arc<ProjectSnapshot>> {
1348 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
1349 cx.spawn(async move |_, _| {
1350 let snapshot = task.await;
1351
1352 Arc::new(ProjectSnapshot {
1353 worktree_snapshots: snapshot.worktree_snapshots,
1354 timestamp: Utc::now(),
1355 })
1356 })
1357 }
1358
1359 pub fn project_context(&self) -> &Entity<ProjectContext> {
1360 &self.project_context
1361 }
1362
1363 pub fn project(&self) -> &Entity<Project> {
1364 &self.project
1365 }
1366
1367 pub fn action_log(&self) -> &Entity<ActionLog> {
1368 &self.action_log
1369 }
1370
1371 pub fn is_empty(&self) -> bool {
1372 self.messages.is_empty() && self.title.is_none()
1373 }
1374
1375 pub fn draft_prompt(&self) -> Option<&[acp::ContentBlock]> {
1376 self.draft_prompt.as_deref()
1377 }
1378
1379 pub fn set_draft_prompt(&mut self, prompt: Option<Vec<acp::ContentBlock>>) {
1380 self.draft_prompt = prompt;
1381 }
1382
1383 pub fn ui_scroll_position(&self) -> Option<gpui::ListOffset> {
1384 self.ui_scroll_position
1385 }
1386
1387 pub fn set_ui_scroll_position(&mut self, position: Option<gpui::ListOffset>) {
1388 self.ui_scroll_position = position;
1389 }
1390
1391 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
1392 self.model.as_ref()
1393 }
1394
1395 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
1396 let old_usage = self.latest_token_usage();
1397 self.model = Some(model.clone());
1398 let new_caps = Self::prompt_capabilities(self.model.as_deref());
1399 let new_usage = self.latest_token_usage();
1400 if old_usage != new_usage {
1401 cx.emit(TokenUsageUpdated(new_usage));
1402 }
1403 self.prompt_capabilities_tx.send(new_caps).log_err();
1404
1405 for subagent in &self.running_subagents {
1406 subagent
1407 .update(cx, |thread, cx| thread.set_model(model.clone(), cx))
1408 .ok();
1409 }
1410
1411 cx.notify()
1412 }
1413
1414 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1415 self.summarization_model.as_ref()
1416 }
1417
1418 pub fn set_summarization_model(
1419 &mut self,
1420 model: Option<Arc<dyn LanguageModel>>,
1421 cx: &mut Context<Self>,
1422 ) {
1423 self.summarization_model = model.clone();
1424
1425 for subagent in &self.running_subagents {
1426 subagent
1427 .update(cx, |thread, cx| {
1428 thread.set_summarization_model(model.clone(), cx)
1429 })
1430 .ok();
1431 }
1432 cx.notify()
1433 }
1434
1435 pub fn thinking_enabled(&self) -> bool {
1436 self.thinking_enabled
1437 }
1438
1439 pub fn set_thinking_enabled(&mut self, enabled: bool, cx: &mut Context<Self>) {
1440 self.thinking_enabled = enabled;
1441
1442 for subagent in &self.running_subagents {
1443 subagent
1444 .update(cx, |thread, cx| thread.set_thinking_enabled(enabled, cx))
1445 .ok();
1446 }
1447 cx.notify();
1448 }
1449
1450 pub fn thinking_effort(&self) -> Option<&String> {
1451 self.thinking_effort.as_ref()
1452 }
1453
1454 pub fn set_thinking_effort(&mut self, effort: Option<String>, cx: &mut Context<Self>) {
1455 self.thinking_effort = effort.clone();
1456
1457 for subagent in &self.running_subagents {
1458 subagent
1459 .update(cx, |thread, cx| {
1460 thread.set_thinking_effort(effort.clone(), cx)
1461 })
1462 .ok();
1463 }
1464 cx.notify();
1465 }
1466
1467 pub fn speed(&self) -> Option<Speed> {
1468 self.speed
1469 }
1470
1471 pub fn set_speed(&mut self, speed: Speed, cx: &mut Context<Self>) {
1472 self.speed = Some(speed);
1473
1474 for subagent in &self.running_subagents {
1475 subagent
1476 .update(cx, |thread, cx| thread.set_speed(speed, cx))
1477 .ok();
1478 }
1479 cx.notify();
1480 }
1481
1482 pub fn last_message(&self) -> Option<&Message> {
1483 self.messages.last()
1484 }
1485
1486 #[cfg(any(test, feature = "test-support"))]
1487 pub fn last_received_or_pending_message(&self) -> Option<Message> {
1488 if let Some(message) = self.pending_message.clone() {
1489 Some(Message::Agent(message))
1490 } else {
1491 self.messages.last().cloned()
1492 }
1493 }
1494
1495 pub fn add_default_tools(
1496 &mut self,
1497 environment: Rc<dyn ThreadEnvironment>,
1498 cx: &mut Context<Self>,
1499 ) {
1500 // Only update the agent location for the root thread, not for subagents.
1501 let update_agent_location = self.parent_thread_id().is_none();
1502
1503 let language_registry = self.project.read(cx).languages().clone();
1504 self.add_tool(CopyPathTool::new(self.project.clone()));
1505 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1506 self.add_tool(DeletePathTool::new(
1507 self.project.clone(),
1508 self.action_log.clone(),
1509 ));
1510 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1511 self.add_tool(EditFileTool::new(
1512 self.project.clone(),
1513 cx.weak_entity(),
1514 language_registry.clone(),
1515 Templates::new(),
1516 ));
1517 self.add_tool(StreamingEditFileTool::new(
1518 self.project.clone(),
1519 cx.weak_entity(),
1520 self.action_log.clone(),
1521 language_registry,
1522 ));
1523 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1524 self.add_tool(FindPathTool::new(self.project.clone()));
1525 self.add_tool(GrepTool::new(self.project.clone()));
1526 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1527 self.add_tool(MovePathTool::new(self.project.clone()));
1528 self.add_tool(NowTool);
1529 self.add_tool(OpenTool::new(self.project.clone()));
1530 if cx.has_flag::<UpdatePlanToolFeatureFlag>() {
1531 self.add_tool(UpdatePlanTool);
1532 }
1533 self.add_tool(ReadFileTool::new(
1534 self.project.clone(),
1535 self.action_log.clone(),
1536 update_agent_location,
1537 ));
1538 self.add_tool(SaveFileTool::new(self.project.clone()));
1539 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1540 self.add_tool(TerminalTool::new(self.project.clone(), environment.clone()));
1541 self.add_tool(WebSearchTool);
1542
1543 if self.depth() < MAX_SUBAGENT_DEPTH {
1544 self.add_tool(SpawnAgentTool::new(environment));
1545 }
1546 }
1547
1548 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1549 debug_assert!(
1550 !self.tools.contains_key(T::NAME),
1551 "Duplicate tool name: {}",
1552 T::NAME,
1553 );
1554 self.tools.insert(T::NAME.into(), tool.erase());
1555 }
1556
1557 #[cfg(any(test, feature = "test-support"))]
1558 pub fn remove_tool(&mut self, name: &str) -> bool {
1559 self.tools.remove(name).is_some()
1560 }
1561
1562 pub fn profile(&self) -> &AgentProfileId {
1563 &self.profile_id
1564 }
1565
1566 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1567 if self.profile_id == profile_id {
1568 return;
1569 }
1570
1571 self.profile_id = profile_id.clone();
1572
1573 // Swap to the profile's preferred model when available.
1574 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1575 self.set_model(model, cx);
1576 }
1577
1578 for subagent in &self.running_subagents {
1579 subagent
1580 .update(cx, |thread, cx| thread.set_profile(profile_id.clone(), cx))
1581 .ok();
1582 }
1583 }
1584
1585 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1586 for subagent in self.running_subagents.drain(..) {
1587 if let Some(subagent) = subagent.upgrade() {
1588 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1589 }
1590 }
1591
1592 let Some(running_turn) = self.running_turn.take() else {
1593 self.flush_pending_message(cx);
1594 return Task::ready(());
1595 };
1596
1597 let turn_task = running_turn.cancel();
1598
1599 cx.spawn(async move |this, cx| {
1600 turn_task.await;
1601 this.update(cx, |this, cx| {
1602 this.flush_pending_message(cx);
1603 })
1604 .ok();
1605 })
1606 }
1607
1608 pub fn set_has_queued_message(&mut self, has_queued: bool) {
1609 self.has_queued_message = has_queued;
1610 }
1611
1612 pub fn has_queued_message(&self) -> bool {
1613 self.has_queued_message
1614 }
1615
1616 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1617 let Some(last_user_message) = self.last_user_message() else {
1618 return;
1619 };
1620
1621 self.request_token_usage
1622 .insert(last_user_message.id.clone(), update);
1623 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1624 cx.notify();
1625 }
1626
1627 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1628 self.cancel(cx).detach();
1629 // Clear pending message since cancel will try to flush it asynchronously,
1630 // and we don't want that content to be added after we truncate
1631 self.pending_message.take();
1632 let Some(position) = self.messages.iter().position(
1633 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1634 ) else {
1635 return Err(anyhow!("Message not found"));
1636 };
1637
1638 for message in self.messages.drain(position..) {
1639 match message {
1640 Message::User(message) => {
1641 self.request_token_usage.remove(&message.id);
1642 }
1643 Message::Agent(_) | Message::Resume => {}
1644 }
1645 }
1646 self.clear_summary();
1647 cx.notify();
1648 Ok(())
1649 }
1650
1651 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1652 let last_user_message = self.last_user_message()?;
1653 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1654 Some(*tokens)
1655 }
1656
1657 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1658 let usage = self.latest_request_token_usage()?;
1659 let model = self.model.clone()?;
1660 Some(acp_thread::TokenUsage {
1661 max_tokens: model.max_token_count(),
1662 max_output_tokens: model.max_output_tokens(),
1663 used_tokens: usage.total_tokens(),
1664 input_tokens: usage.input_tokens,
1665 output_tokens: usage.output_tokens,
1666 })
1667 }
1668
1669 /// Get the total input token count as of the message before the given message.
1670 ///
1671 /// Returns `None` if:
1672 /// - `target_id` is the first message (no previous message)
1673 /// - The previous message hasn't received a response yet (no usage data)
1674 /// - `target_id` is not found in the messages
1675 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1676 let mut previous_user_message_id: Option<&UserMessageId> = None;
1677
1678 for message in &self.messages {
1679 if let Message::User(user_msg) = message {
1680 if &user_msg.id == target_id {
1681 let prev_id = previous_user_message_id?;
1682 let usage = self.request_token_usage.get(prev_id)?;
1683 return Some(usage.input_tokens);
1684 }
1685 previous_user_message_id = Some(&user_msg.id);
1686 }
1687 }
1688 None
1689 }
1690
1691 /// Look up the active profile and resolve its preferred model if one is configured.
1692 fn resolve_profile_model(
1693 profile_id: &AgentProfileId,
1694 cx: &mut Context<Self>,
1695 ) -> Option<Arc<dyn LanguageModel>> {
1696 let selection = AgentSettings::get_global(cx)
1697 .profiles
1698 .get(profile_id)?
1699 .default_model
1700 .clone()?;
1701 Self::resolve_model_from_selection(&selection, cx)
1702 }
1703
1704 /// Translate a stored model selection into the configured model from the registry.
1705 fn resolve_model_from_selection(
1706 selection: &LanguageModelSelection,
1707 cx: &mut Context<Self>,
1708 ) -> Option<Arc<dyn LanguageModel>> {
1709 let selected = SelectedModel {
1710 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1711 model: LanguageModelId::from(selection.model.clone()),
1712 };
1713 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1714 registry
1715 .select_model(&selected, cx)
1716 .map(|configured| configured.model)
1717 })
1718 }
1719
1720 pub fn resume(
1721 &mut self,
1722 cx: &mut Context<Self>,
1723 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1724 self.messages.push(Message::Resume);
1725 cx.notify();
1726
1727 log::debug!("Total messages in thread: {}", self.messages.len());
1728 self.run_turn(cx)
1729 }
1730
1731 /// Sending a message results in the model streaming a response, which could include tool calls.
1732 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1733 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1734 pub fn send<T>(
1735 &mut self,
1736 id: UserMessageId,
1737 content: impl IntoIterator<Item = T>,
1738 cx: &mut Context<Self>,
1739 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1740 where
1741 T: Into<UserMessageContent>,
1742 {
1743 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1744 log::debug!("Thread::send content: {:?}", content);
1745
1746 self.messages
1747 .push(Message::User(UserMessage { id, content }));
1748 cx.notify();
1749
1750 self.send_existing(cx)
1751 }
1752
1753 pub fn send_existing(
1754 &mut self,
1755 cx: &mut Context<Self>,
1756 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1757 let model = self.model().context("No language model configured")?;
1758
1759 log::info!("Thread::send called with model: {}", model.name().0);
1760 self.advance_prompt_id();
1761
1762 log::debug!("Total messages in thread: {}", self.messages.len());
1763 self.run_turn(cx)
1764 }
1765
1766 pub fn push_acp_user_block(
1767 &mut self,
1768 id: UserMessageId,
1769 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1770 path_style: PathStyle,
1771 cx: &mut Context<Self>,
1772 ) {
1773 let content = blocks
1774 .into_iter()
1775 .map(|block| UserMessageContent::from_content_block(block, path_style))
1776 .collect::<Vec<_>>();
1777 self.messages
1778 .push(Message::User(UserMessage { id, content }));
1779 cx.notify();
1780 }
1781
1782 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1783 let text = match block {
1784 acp::ContentBlock::Text(text_content) => text_content.text,
1785 acp::ContentBlock::Image(_) => "[image]".to_string(),
1786 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1787 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1788 acp::ContentBlock::Resource(resource) => match resource.resource {
1789 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1790 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1791 _ => "[resource]".to_string(),
1792 },
1793 _ => "[unknown]".to_string(),
1794 };
1795
1796 self.messages.push(Message::Agent(AgentMessage {
1797 content: vec![AgentMessageContent::Text(text)],
1798 ..Default::default()
1799 }));
1800 cx.notify();
1801 }
1802
1803 fn run_turn(
1804 &mut self,
1805 cx: &mut Context<Self>,
1806 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1807 // Flush the old pending message synchronously before cancelling,
1808 // to avoid a race where the detached cancel task might flush the NEW
1809 // turn's pending message instead of the old one.
1810 self.flush_pending_message(cx);
1811 self.cancel(cx).detach();
1812
1813 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1814 let event_stream = ThreadEventStream(events_tx);
1815 let message_ix = self.messages.len().saturating_sub(1);
1816 self.clear_summary();
1817 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1818 self.running_turn = Some(RunningTurn {
1819 event_stream: event_stream.clone(),
1820 tools: self.enabled_tools(cx),
1821 cancellation_tx,
1822 streaming_tool_inputs: HashMap::default(),
1823 _task: cx.spawn(async move |this, cx| {
1824 log::debug!("Starting agent turn execution");
1825
1826 let turn_result =
1827 Self::run_turn_internal(&this, &event_stream, cancellation_rx.clone(), cx)
1828 .await;
1829
1830 // Check if we were cancelled - if so, cancel() already took running_turn
1831 // and we shouldn't touch it (it might be a NEW turn now)
1832 let was_cancelled = *cancellation_rx.borrow();
1833 if was_cancelled {
1834 log::debug!("Turn was cancelled, skipping cleanup");
1835 return;
1836 }
1837
1838 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1839
1840 match turn_result {
1841 Ok(()) => {
1842 log::debug!("Turn execution completed");
1843 event_stream.send_stop(acp::StopReason::EndTurn);
1844 }
1845 Err(error) => {
1846 log::error!("Turn execution failed: {:?}", error);
1847 match error.downcast::<CompletionError>() {
1848 Ok(CompletionError::Refusal) => {
1849 event_stream.send_stop(acp::StopReason::Refusal);
1850 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1851 }
1852 Ok(CompletionError::MaxTokens) => {
1853 event_stream.send_stop(acp::StopReason::MaxTokens);
1854 }
1855 Ok(CompletionError::Other(error)) | Err(error) => {
1856 event_stream.send_error(error);
1857 }
1858 }
1859 }
1860 }
1861
1862 _ = this.update(cx, |this, _| this.running_turn.take());
1863 }),
1864 });
1865 Ok(events_rx)
1866 }
1867
1868 async fn run_turn_internal(
1869 this: &WeakEntity<Self>,
1870 event_stream: &ThreadEventStream,
1871 mut cancellation_rx: watch::Receiver<bool>,
1872 cx: &mut AsyncApp,
1873 ) -> Result<()> {
1874 let mut attempt = 0;
1875 let mut intent = CompletionIntent::UserPrompt;
1876 loop {
1877 // Re-read the model and refresh tools on each iteration so that
1878 // mid-turn changes (e.g. the user switches model, toggles tools,
1879 // or changes profile) take effect between tool-call rounds.
1880 let (model, request) = this.update(cx, |this, cx| {
1881 let model = this.model.clone().context("No language model configured")?;
1882 this.refresh_turn_tools(cx);
1883 let request = this.build_completion_request(intent, cx)?;
1884 anyhow::Ok((model, request))
1885 })??;
1886
1887 telemetry::event!(
1888 "Agent Thread Completion",
1889 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1890 parent_thread_id = this.read_with(cx, |this, _| this
1891 .parent_thread_id()
1892 .map(|id| id.to_string()))?,
1893 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1894 model = model.telemetry_id(),
1895 model_provider = model.provider_id().to_string(),
1896 attempt
1897 );
1898
1899 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1900
1901 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1902 Ok(events) => (events.fuse(), None),
1903 Err(err) => (stream::empty().boxed().fuse(), Some(err)),
1904 };
1905 let mut tool_results: FuturesUnordered<Task<LanguageModelToolResult>> =
1906 FuturesUnordered::new();
1907 let mut early_tool_results: Vec<LanguageModelToolResult> = Vec::new();
1908 let mut cancelled = false;
1909 loop {
1910 // Race between getting the first event, tool completion, and cancellation.
1911 let first_event = futures::select! {
1912 event = events.next().fuse() => event,
1913 tool_result = futures::StreamExt::select_next_some(&mut tool_results) => {
1914 let is_error = tool_result.is_error;
1915 let is_still_streaming = this
1916 .read_with(cx, |this, _cx| {
1917 this.running_turn
1918 .as_ref()
1919 .and_then(|turn| turn.streaming_tool_inputs.get(&tool_result.tool_use_id))
1920 .map_or(false, |inputs| !inputs.has_received_final())
1921 })
1922 .unwrap_or(false);
1923
1924 early_tool_results.push(tool_result);
1925
1926 // Only break if the tool errored and we are still
1927 // streaming the input of the tool. If the tool errored
1928 // but we are no longer streaming its input (i.e. there
1929 // are parallel tool calls) we want to continue
1930 // processing those tool inputs.
1931 if is_error && is_still_streaming {
1932 break;
1933 }
1934 continue;
1935 }
1936 _ = cancellation_rx.changed().fuse() => {
1937 if *cancellation_rx.borrow() {
1938 cancelled = true;
1939 break;
1940 }
1941 continue;
1942 }
1943 };
1944 let Some(first_event) = first_event else {
1945 break;
1946 };
1947
1948 // Collect all immediately available events to process as a batch
1949 let mut batch = vec![first_event];
1950 while let Some(event) = events.next().now_or_never().flatten() {
1951 batch.push(event);
1952 }
1953
1954 // Process the batch in a single update
1955 let batch_result = this.update(cx, |this, cx| {
1956 let mut batch_tool_results = Vec::new();
1957 let mut batch_error = None;
1958
1959 for event in batch {
1960 log::trace!("Received completion event: {:?}", event);
1961 match event {
1962 Ok(event) => {
1963 match this.handle_completion_event(
1964 event,
1965 event_stream,
1966 cancellation_rx.clone(),
1967 cx,
1968 ) {
1969 Ok(Some(task)) => batch_tool_results.push(task),
1970 Ok(None) => {}
1971 Err(err) => {
1972 batch_error = Some(err);
1973 break;
1974 }
1975 }
1976 }
1977 Err(err) => {
1978 batch_error = Some(err.into());
1979 break;
1980 }
1981 }
1982 }
1983
1984 cx.notify();
1985 (batch_tool_results, batch_error)
1986 })?;
1987
1988 tool_results.extend(batch_result.0);
1989 if let Some(err) = batch_result.1 {
1990 error = Some(err.downcast()?);
1991 break;
1992 }
1993 }
1994
1995 // Drop the stream to release the rate limit permit before tool execution.
1996 // The stream holds a semaphore guard that limits concurrent requests.
1997 // Without this, the permit would be held during potentially long-running
1998 // tool execution, which could cause deadlocks when tools spawn subagents
1999 // that need their own permits.
2000 drop(events);
2001
2002 // Drop streaming tool input senders that never received their final input.
2003 // This prevents deadlock when the LLM stream ends (e.g. because of an error)
2004 // before sending a tool use with `is_input_complete: true`.
2005 this.update(cx, |this, _cx| {
2006 if let Some(running_turn) = this.running_turn.as_mut() {
2007 if running_turn.streaming_tool_inputs.is_empty() {
2008 return;
2009 }
2010 log::warn!("Dropping partial tool inputs because the stream ended");
2011 running_turn.streaming_tool_inputs.drain();
2012 }
2013 })?;
2014
2015 let end_turn = tool_results.is_empty() && early_tool_results.is_empty();
2016
2017 for tool_result in early_tool_results {
2018 Self::process_tool_result(this, event_stream, cx, tool_result)?;
2019 }
2020 while let Some(tool_result) = tool_results.next().await {
2021 Self::process_tool_result(this, event_stream, cx, tool_result)?;
2022 }
2023
2024 this.update(cx, |this, cx| {
2025 this.flush_pending_message(cx);
2026 if this.title.is_none() && this.pending_title_generation.is_none() {
2027 this.generate_title(cx);
2028 }
2029 })?;
2030
2031 if cancelled {
2032 log::debug!("Turn cancelled by user, exiting");
2033 return Ok(());
2034 }
2035
2036 if let Some(error) = error {
2037 attempt += 1;
2038 let retry = this.update(cx, |this, cx| {
2039 let user_store = this.user_store.read(cx);
2040 this.handle_completion_error(error, attempt, user_store.plan())
2041 })??;
2042 let timer = cx.background_executor().timer(retry.duration);
2043 event_stream.send_retry(retry);
2044 futures::select! {
2045 _ = timer.fuse() => {}
2046 _ = cancellation_rx.changed().fuse() => {
2047 if *cancellation_rx.borrow() {
2048 log::debug!("Turn cancelled during retry delay, exiting");
2049 return Ok(());
2050 }
2051 }
2052 }
2053 this.update(cx, |this, _cx| {
2054 if let Some(Message::Agent(message)) = this.messages.last() {
2055 if message.tool_results.is_empty() {
2056 intent = CompletionIntent::UserPrompt;
2057 this.messages.push(Message::Resume);
2058 }
2059 }
2060 })?;
2061 } else if end_turn {
2062 return Ok(());
2063 } else {
2064 let has_queued = this.update(cx, |this, _| this.has_queued_message())?;
2065 if has_queued {
2066 log::debug!("Queued message found, ending turn at message boundary");
2067 return Ok(());
2068 }
2069 intent = CompletionIntent::ToolResults;
2070 attempt = 0;
2071 }
2072 }
2073 }
2074
2075 fn process_tool_result(
2076 this: &WeakEntity<Thread>,
2077 event_stream: &ThreadEventStream,
2078 cx: &mut AsyncApp,
2079 tool_result: LanguageModelToolResult,
2080 ) -> Result<(), anyhow::Error> {
2081 log::debug!("Tool finished {:?}", tool_result);
2082
2083 event_stream.update_tool_call_fields(
2084 &tool_result.tool_use_id,
2085 acp::ToolCallUpdateFields::new()
2086 .status(if tool_result.is_error {
2087 acp::ToolCallStatus::Failed
2088 } else {
2089 acp::ToolCallStatus::Completed
2090 })
2091 .raw_output(tool_result.output.clone()),
2092 None,
2093 );
2094 this.update(cx, |this, _cx| {
2095 this.pending_message()
2096 .tool_results
2097 .insert(tool_result.tool_use_id.clone(), tool_result)
2098 })?;
2099 Ok(())
2100 }
2101
2102 fn handle_completion_error(
2103 &mut self,
2104 error: LanguageModelCompletionError,
2105 attempt: u8,
2106 plan: Option<Plan>,
2107 ) -> Result<acp_thread::RetryStatus> {
2108 let Some(model) = self.model.as_ref() else {
2109 return Err(anyhow!(error));
2110 };
2111
2112 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
2113 plan.is_some()
2114 } else {
2115 true
2116 };
2117
2118 if !auto_retry {
2119 return Err(anyhow!(error));
2120 }
2121
2122 let Some(strategy) = Self::retry_strategy_for(&error) else {
2123 return Err(anyhow!(error));
2124 };
2125
2126 let max_attempts = match &strategy {
2127 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2128 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2129 };
2130
2131 if attempt > max_attempts {
2132 return Err(anyhow!(error));
2133 }
2134
2135 let delay = match &strategy {
2136 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2137 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2138 Duration::from_secs(delay_secs)
2139 }
2140 RetryStrategy::Fixed { delay, .. } => *delay,
2141 };
2142 log::debug!("Retry attempt {attempt} with delay {delay:?}");
2143
2144 Ok(acp_thread::RetryStatus {
2145 last_error: error.to_string().into(),
2146 attempt: attempt as usize,
2147 max_attempts: max_attempts as usize,
2148 started_at: Instant::now(),
2149 duration: delay,
2150 })
2151 }
2152
2153 /// A helper method that's called on every streamed completion event.
2154 /// Returns an optional tool result task, which the main agentic loop will
2155 /// send back to the model when it resolves.
2156 fn handle_completion_event(
2157 &mut self,
2158 event: LanguageModelCompletionEvent,
2159 event_stream: &ThreadEventStream,
2160 cancellation_rx: watch::Receiver<bool>,
2161 cx: &mut Context<Self>,
2162 ) -> Result<Option<Task<LanguageModelToolResult>>> {
2163 log::trace!("Handling streamed completion event: {:?}", event);
2164 use LanguageModelCompletionEvent::*;
2165
2166 match event {
2167 StartMessage { .. } => {
2168 self.flush_pending_message(cx);
2169 self.pending_message = Some(AgentMessage::default());
2170 }
2171 Text(new_text) => self.handle_text_event(new_text, event_stream),
2172 Thinking { text, signature } => {
2173 self.handle_thinking_event(text, signature, event_stream)
2174 }
2175 RedactedThinking { data } => self.handle_redacted_thinking_event(data),
2176 ReasoningDetails(details) => {
2177 let last_message = self.pending_message();
2178 // Store the last non-empty reasoning_details (overwrites earlier ones)
2179 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
2180 if let serde_json::Value::Array(ref arr) = details {
2181 if !arr.is_empty() {
2182 last_message.reasoning_details = Some(details);
2183 }
2184 } else {
2185 last_message.reasoning_details = Some(details);
2186 }
2187 }
2188 ToolUse(tool_use) => {
2189 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
2190 }
2191 ToolUseJsonParseError {
2192 id,
2193 tool_name,
2194 raw_input,
2195 json_parse_error,
2196 } => {
2197 return Ok(self.handle_tool_use_json_parse_error_event(
2198 id,
2199 tool_name,
2200 raw_input,
2201 json_parse_error,
2202 event_stream,
2203 cancellation_rx,
2204 cx,
2205 ));
2206 }
2207 UsageUpdate(usage) => {
2208 telemetry::event!(
2209 "Agent Thread Completion Usage Updated",
2210 thread_id = self.id.to_string(),
2211 parent_thread_id = self.parent_thread_id().map(|id| id.to_string()),
2212 prompt_id = self.prompt_id.to_string(),
2213 model = self.model.as_ref().map(|m| m.telemetry_id()),
2214 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
2215 input_tokens = usage.input_tokens,
2216 output_tokens = usage.output_tokens,
2217 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2218 cache_read_input_tokens = usage.cache_read_input_tokens,
2219 );
2220 self.update_token_usage(usage, cx);
2221 }
2222 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
2223 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
2224 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
2225 Started | Queued { .. } => {}
2226 }
2227
2228 Ok(None)
2229 }
2230
2231 fn handle_text_event(&mut self, new_text: String, event_stream: &ThreadEventStream) {
2232 event_stream.send_text(&new_text);
2233
2234 let last_message = self.pending_message();
2235 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
2236 text.push_str(&new_text);
2237 } else {
2238 last_message
2239 .content
2240 .push(AgentMessageContent::Text(new_text));
2241 }
2242 }
2243
2244 fn handle_thinking_event(
2245 &mut self,
2246 new_text: String,
2247 new_signature: Option<String>,
2248 event_stream: &ThreadEventStream,
2249 ) {
2250 event_stream.send_thinking(&new_text);
2251
2252 let last_message = self.pending_message();
2253 if let Some(AgentMessageContent::Thinking { text, signature }) =
2254 last_message.content.last_mut()
2255 {
2256 text.push_str(&new_text);
2257 *signature = new_signature.or(signature.take());
2258 } else {
2259 last_message.content.push(AgentMessageContent::Thinking {
2260 text: new_text,
2261 signature: new_signature,
2262 });
2263 }
2264 }
2265
2266 fn handle_redacted_thinking_event(&mut self, data: String) {
2267 let last_message = self.pending_message();
2268 last_message
2269 .content
2270 .push(AgentMessageContent::RedactedThinking(data));
2271 }
2272
2273 fn handle_tool_use_event(
2274 &mut self,
2275 tool_use: LanguageModelToolUse,
2276 event_stream: &ThreadEventStream,
2277 cancellation_rx: watch::Receiver<bool>,
2278 cx: &mut Context<Self>,
2279 ) -> Option<Task<LanguageModelToolResult>> {
2280 cx.notify();
2281
2282 let tool = self.tool(tool_use.name.as_ref());
2283 let mut title = SharedString::from(&tool_use.name);
2284 let mut kind = acp::ToolKind::Other;
2285 if let Some(tool) = tool.as_ref() {
2286 title = tool.initial_title(tool_use.input.clone(), cx);
2287 kind = tool.kind();
2288 }
2289
2290 self.send_or_update_tool_use(&tool_use, title, kind, event_stream);
2291
2292 let Some(tool) = tool else {
2293 let content = format!("No tool named {} exists", tool_use.name);
2294 return Some(Task::ready(LanguageModelToolResult {
2295 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2296 tool_use_id: tool_use.id,
2297 tool_name: tool_use.name,
2298 is_error: true,
2299 output: None,
2300 }));
2301 };
2302
2303 if !tool_use.is_input_complete {
2304 if tool.supports_input_streaming() {
2305 let running_turn = self.running_turn.as_mut()?;
2306 if let Some(sender) = running_turn.streaming_tool_inputs.get_mut(&tool_use.id) {
2307 sender.send_partial(tool_use.input);
2308 return None;
2309 }
2310
2311 let (mut sender, tool_input) = ToolInputSender::channel();
2312 sender.send_partial(tool_use.input);
2313 running_turn
2314 .streaming_tool_inputs
2315 .insert(tool_use.id.clone(), sender);
2316
2317 let tool = tool.clone();
2318 log::debug!("Running streaming tool {}", tool_use.name);
2319 return Some(self.run_tool(
2320 tool,
2321 tool_input,
2322 tool_use.id,
2323 tool_use.name,
2324 event_stream,
2325 cancellation_rx,
2326 cx,
2327 ));
2328 } else {
2329 return None;
2330 }
2331 }
2332
2333 if let Some(mut sender) = self
2334 .running_turn
2335 .as_mut()?
2336 .streaming_tool_inputs
2337 .remove(&tool_use.id)
2338 {
2339 sender.send_full(tool_use.input);
2340 return None;
2341 }
2342
2343 log::debug!("Running tool {}", tool_use.name);
2344 let tool_input = ToolInput::ready(tool_use.input);
2345 Some(self.run_tool(
2346 tool,
2347 tool_input,
2348 tool_use.id,
2349 tool_use.name,
2350 event_stream,
2351 cancellation_rx,
2352 cx,
2353 ))
2354 }
2355
2356 fn run_tool(
2357 &self,
2358 tool: Arc<dyn AnyAgentTool>,
2359 tool_input: ToolInput<serde_json::Value>,
2360 tool_use_id: LanguageModelToolUseId,
2361 tool_name: Arc<str>,
2362 event_stream: &ThreadEventStream,
2363 cancellation_rx: watch::Receiver<bool>,
2364 cx: &mut Context<Self>,
2365 ) -> Task<LanguageModelToolResult> {
2366 let fs = self.project.read(cx).fs().clone();
2367 let tool_event_stream = ToolCallEventStream::new(
2368 tool_use_id.clone(),
2369 event_stream.clone(),
2370 Some(fs),
2371 cancellation_rx,
2372 );
2373 tool_event_stream.update_fields(
2374 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
2375 );
2376 let supports_images = self.model().is_some_and(|model| model.supports_images());
2377 let tool_result = tool.run(tool_input, tool_event_stream, cx);
2378 cx.foreground_executor().spawn(async move {
2379 let (is_error, output) = match tool_result.await {
2380 Ok(mut output) => {
2381 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
2382 && !supports_images
2383 {
2384 output = AgentToolOutput::from_error(
2385 "Attempted to read an image, but this model doesn't support it.",
2386 );
2387 (true, output)
2388 } else {
2389 (false, output)
2390 }
2391 }
2392 Err(output) => (true, output),
2393 };
2394
2395 LanguageModelToolResult {
2396 tool_use_id,
2397 tool_name,
2398 is_error,
2399 content: output.llm_output,
2400 output: Some(output.raw_output),
2401 }
2402 })
2403 }
2404
2405 fn handle_tool_use_json_parse_error_event(
2406 &mut self,
2407 tool_use_id: LanguageModelToolUseId,
2408 tool_name: Arc<str>,
2409 raw_input: Arc<str>,
2410 json_parse_error: String,
2411 event_stream: &ThreadEventStream,
2412 cancellation_rx: watch::Receiver<bool>,
2413 cx: &mut Context<Self>,
2414 ) -> Option<Task<LanguageModelToolResult>> {
2415 let tool_use = LanguageModelToolUse {
2416 id: tool_use_id,
2417 name: tool_name,
2418 raw_input: raw_input.to_string(),
2419 input: serde_json::json!({}),
2420 is_input_complete: true,
2421 thought_signature: None,
2422 };
2423 self.send_or_update_tool_use(
2424 &tool_use,
2425 SharedString::from(&tool_use.name),
2426 acp::ToolKind::Other,
2427 event_stream,
2428 );
2429
2430 let tool = self.tool(tool_use.name.as_ref());
2431
2432 let Some(tool) = tool else {
2433 let content = format!("No tool named {} exists", tool_use.name);
2434 return Some(Task::ready(LanguageModelToolResult {
2435 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2436 tool_use_id: tool_use.id,
2437 tool_name: tool_use.name,
2438 is_error: true,
2439 output: None,
2440 }));
2441 };
2442
2443 let error_message = format!("Error parsing input JSON: {json_parse_error}");
2444
2445 if tool.supports_input_streaming()
2446 && let Some(mut sender) = self
2447 .running_turn
2448 .as_mut()?
2449 .streaming_tool_inputs
2450 .remove(&tool_use.id)
2451 {
2452 sender.send_invalid_json(error_message);
2453 return None;
2454 }
2455
2456 log::debug!("Running tool {}. Received invalid JSON", tool_use.name);
2457 let tool_input = ToolInput::invalid_json(error_message);
2458 Some(self.run_tool(
2459 tool,
2460 tool_input,
2461 tool_use.id,
2462 tool_use.name,
2463 event_stream,
2464 cancellation_rx,
2465 cx,
2466 ))
2467 }
2468
2469 fn send_or_update_tool_use(
2470 &mut self,
2471 tool_use: &LanguageModelToolUse,
2472 title: SharedString,
2473 kind: acp::ToolKind,
2474 event_stream: &ThreadEventStream,
2475 ) {
2476 // Ensure the last message ends in the current tool use
2477 let last_message = self.pending_message();
2478
2479 let has_tool_use = last_message.content.iter_mut().rev().any(|content| {
2480 if let AgentMessageContent::ToolUse(last_tool_use) = content {
2481 if last_tool_use.id == tool_use.id {
2482 *last_tool_use = tool_use.clone();
2483 return true;
2484 }
2485 }
2486 false
2487 });
2488
2489 if !has_tool_use {
2490 event_stream.send_tool_call(
2491 &tool_use.id,
2492 &tool_use.name,
2493 title,
2494 kind,
2495 tool_use.input.clone(),
2496 );
2497 last_message
2498 .content
2499 .push(AgentMessageContent::ToolUse(tool_use.clone()));
2500 } else {
2501 event_stream.update_tool_call_fields(
2502 &tool_use.id,
2503 acp::ToolCallUpdateFields::new()
2504 .title(title.as_str())
2505 .kind(kind)
2506 .raw_input(tool_use.input.clone()),
2507 None,
2508 );
2509 }
2510 }
2511
2512 pub fn title(&self) -> Option<SharedString> {
2513 self.title.clone()
2514 }
2515
2516 pub fn is_generating_summary(&self) -> bool {
2517 self.pending_summary_generation.is_some()
2518 }
2519
2520 pub fn is_generating_title(&self) -> bool {
2521 self.pending_title_generation.is_some()
2522 }
2523
2524 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
2525 if let Some(summary) = self.summary.as_ref() {
2526 return Task::ready(Some(summary.clone())).shared();
2527 }
2528 if let Some(task) = self.pending_summary_generation.clone() {
2529 return task;
2530 }
2531 let Some(model) = self.summarization_model.clone() else {
2532 log::error!("No summarization model available");
2533 return Task::ready(None).shared();
2534 };
2535 let mut request = LanguageModelRequest {
2536 intent: Some(CompletionIntent::ThreadContextSummarization),
2537 temperature: AgentSettings::temperature_for_model(&model, cx),
2538 ..Default::default()
2539 };
2540
2541 for message in &self.messages {
2542 request.messages.extend(message.to_request());
2543 }
2544
2545 request.messages.push(LanguageModelRequestMessage {
2546 role: Role::User,
2547 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
2548 cache: false,
2549 reasoning_details: None,
2550 });
2551
2552 let task = cx
2553 .spawn(async move |this, cx| {
2554 let mut summary = String::new();
2555 let mut messages = model.stream_completion(request, cx).await.log_err()?;
2556 while let Some(event) = messages.next().await {
2557 let event = event.log_err()?;
2558 let text = match event {
2559 LanguageModelCompletionEvent::Text(text) => text,
2560 _ => continue,
2561 };
2562
2563 let mut lines = text.lines();
2564 summary.extend(lines.next());
2565 }
2566
2567 log::debug!("Setting summary: {}", summary);
2568 let summary = SharedString::from(summary);
2569
2570 this.update(cx, |this, cx| {
2571 this.summary = Some(summary.clone());
2572 this.pending_summary_generation = None;
2573 cx.notify()
2574 })
2575 .ok()?;
2576
2577 Some(summary)
2578 })
2579 .shared();
2580 self.pending_summary_generation = Some(task.clone());
2581 task
2582 }
2583
2584 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
2585 let Some(model) = self.summarization_model.clone() else {
2586 return;
2587 };
2588
2589 log::debug!(
2590 "Generating title with model: {:?}",
2591 self.summarization_model.as_ref().map(|model| model.name())
2592 );
2593 let mut request = LanguageModelRequest {
2594 intent: Some(CompletionIntent::ThreadSummarization),
2595 temperature: AgentSettings::temperature_for_model(&model, cx),
2596 ..Default::default()
2597 };
2598
2599 for message in &self.messages {
2600 request.messages.extend(message.to_request());
2601 }
2602
2603 request.messages.push(LanguageModelRequestMessage {
2604 role: Role::User,
2605 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2606 cache: false,
2607 reasoning_details: None,
2608 });
2609 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2610 let mut title = String::new();
2611
2612 let generate = async {
2613 let mut messages = model.stream_completion(request, cx).await?;
2614 while let Some(event) = messages.next().await {
2615 let event = event?;
2616 let text = match event {
2617 LanguageModelCompletionEvent::Text(text) => text,
2618 _ => continue,
2619 };
2620
2621 let mut lines = text.lines();
2622 title.extend(lines.next());
2623
2624 // Stop if the LLM generated multiple lines.
2625 if lines.next().is_some() {
2626 break;
2627 }
2628 }
2629 anyhow::Ok(())
2630 };
2631
2632 if generate
2633 .await
2634 .context("failed to generate thread title")
2635 .log_err()
2636 .is_some()
2637 {
2638 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
2639 } else {
2640 // Emit TitleUpdated even on failure so that the propagation
2641 // chain (agent::Thread → NativeAgent → AcpThread) fires and
2642 // clears any provisional title that was set before the turn.
2643 _ = this.update(cx, |_, cx| {
2644 cx.emit(TitleUpdated);
2645 cx.notify();
2646 });
2647 }
2648 _ = this.update(cx, |this, _| this.pending_title_generation = None);
2649 }));
2650 }
2651
2652 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2653 self.pending_title_generation = None;
2654 if Some(&title) != self.title.as_ref() {
2655 self.title = Some(title);
2656 cx.emit(TitleUpdated);
2657 cx.notify();
2658 }
2659 }
2660
2661 fn clear_summary(&mut self) {
2662 self.summary = None;
2663 self.pending_summary_generation = None;
2664 }
2665
2666 fn last_user_message(&self) -> Option<&UserMessage> {
2667 self.messages
2668 .iter()
2669 .rev()
2670 .find_map(|message| match message {
2671 Message::User(user_message) => Some(user_message),
2672 Message::Agent(_) => None,
2673 Message::Resume => None,
2674 })
2675 }
2676
2677 fn pending_message(&mut self) -> &mut AgentMessage {
2678 self.pending_message.get_or_insert_default()
2679 }
2680
2681 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2682 let Some(mut message) = self.pending_message.take() else {
2683 return;
2684 };
2685
2686 if message.content.is_empty() {
2687 return;
2688 }
2689
2690 for content in &message.content {
2691 let AgentMessageContent::ToolUse(tool_use) = content else {
2692 continue;
2693 };
2694
2695 if !message.tool_results.contains_key(&tool_use.id) {
2696 message.tool_results.insert(
2697 tool_use.id.clone(),
2698 LanguageModelToolResult {
2699 tool_use_id: tool_use.id.clone(),
2700 tool_name: tool_use.name.clone(),
2701 is_error: true,
2702 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2703 output: None,
2704 },
2705 );
2706 }
2707 }
2708
2709 self.messages.push(Message::Agent(message));
2710 self.updated_at = Utc::now();
2711 self.clear_summary();
2712 cx.notify()
2713 }
2714
2715 pub(crate) fn build_completion_request(
2716 &self,
2717 completion_intent: CompletionIntent,
2718 cx: &App,
2719 ) -> Result<LanguageModelRequest> {
2720 let completion_intent =
2721 if self.is_subagent() && completion_intent == CompletionIntent::UserPrompt {
2722 CompletionIntent::Subagent
2723 } else {
2724 completion_intent
2725 };
2726
2727 let model = self.model().context("No language model configured")?;
2728 let tools = if let Some(turn) = self.running_turn.as_ref() {
2729 turn.tools
2730 .iter()
2731 .filter_map(|(tool_name, tool)| {
2732 log::trace!("Including tool: {}", tool_name);
2733 Some(LanguageModelRequestTool {
2734 name: tool_name.to_string(),
2735 description: tool.description().to_string(),
2736 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2737 use_input_streaming: tool.supports_input_streaming(),
2738 })
2739 })
2740 .collect::<Vec<_>>()
2741 } else {
2742 Vec::new()
2743 };
2744
2745 log::debug!("Building completion request");
2746 log::debug!("Completion intent: {:?}", completion_intent);
2747
2748 let available_tools: Vec<_> = self
2749 .running_turn
2750 .as_ref()
2751 .map(|turn| turn.tools.keys().cloned().collect())
2752 .unwrap_or_default();
2753
2754 log::debug!("Request includes {} tools", available_tools.len());
2755 let messages = self.build_request_messages(available_tools, cx);
2756 log::debug!("Request will include {} messages", messages.len());
2757
2758 let request = LanguageModelRequest {
2759 thread_id: Some(self.id.to_string()),
2760 prompt_id: Some(self.prompt_id.to_string()),
2761 intent: Some(completion_intent),
2762 messages,
2763 tools,
2764 tool_choice: None,
2765 stop: Vec::new(),
2766 temperature: AgentSettings::temperature_for_model(model, cx),
2767 thinking_allowed: self.thinking_enabled,
2768 thinking_effort: self.thinking_effort.clone(),
2769 speed: self.speed(),
2770 };
2771
2772 log::debug!("Completion request built successfully");
2773 Ok(request)
2774 }
2775
2776 fn enabled_tools(&self, cx: &App) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2777 let Some(model) = self.model.as_ref() else {
2778 return BTreeMap::new();
2779 };
2780 let Some(profile) = AgentSettings::get_global(cx).profiles.get(&self.profile_id) else {
2781 return BTreeMap::new();
2782 };
2783 fn truncate(tool_name: &SharedString) -> SharedString {
2784 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2785 let mut truncated = tool_name.to_string();
2786 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2787 truncated.into()
2788 } else {
2789 tool_name.clone()
2790 }
2791 }
2792
2793 let use_streaming_edit_tool =
2794 cx.has_flag::<StreamingEditFileToolFeatureFlag>() && model.supports_streaming_tools();
2795
2796 let mut tools = self
2797 .tools
2798 .iter()
2799 .filter_map(|(tool_name, tool)| {
2800 // For streaming_edit_file, check profile against "edit_file" since that's what users configure
2801 let profile_tool_name = if tool_name == StreamingEditFileTool::NAME {
2802 EditFileTool::NAME
2803 } else {
2804 tool_name.as_ref()
2805 };
2806
2807 if tool.supports_provider(&model.provider_id())
2808 && profile.is_tool_enabled(profile_tool_name)
2809 {
2810 match (tool_name.as_ref(), use_streaming_edit_tool) {
2811 (StreamingEditFileTool::NAME, false) | (EditFileTool::NAME, true) => None,
2812 (StreamingEditFileTool::NAME, true) => {
2813 // Expose streaming tool as "edit_file"
2814 Some((SharedString::from(EditFileTool::NAME), tool.clone()))
2815 }
2816 _ => Some((truncate(tool_name), tool.clone())),
2817 }
2818 } else {
2819 None
2820 }
2821 })
2822 .collect::<BTreeMap<_, _>>();
2823
2824 let mut context_server_tools = Vec::new();
2825 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2826 let mut duplicate_tool_names = HashSet::default();
2827 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2828 for (tool_name, tool) in server_tools {
2829 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2830 let tool_name = truncate(tool_name);
2831 if !seen_tools.insert(tool_name.clone()) {
2832 duplicate_tool_names.insert(tool_name.clone());
2833 }
2834 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2835 }
2836 }
2837 }
2838
2839 // When there are duplicate tool names, disambiguate by prefixing them
2840 // with the server ID (converted to snake_case for API compatibility).
2841 // In the rare case there isn't enough space for the disambiguated tool
2842 // name, keep only the last tool with this name.
2843 for (server_id, tool_name, tool) in context_server_tools {
2844 if duplicate_tool_names.contains(&tool_name) {
2845 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2846 if available >= 2 {
2847 let mut disambiguated = server_id.0.to_snake_case();
2848 disambiguated.truncate(available - 1);
2849 disambiguated.push('_');
2850 disambiguated.push_str(&tool_name);
2851 tools.insert(disambiguated.into(), tool.clone());
2852 } else {
2853 tools.insert(tool_name, tool.clone());
2854 }
2855 } else {
2856 tools.insert(tool_name, tool.clone());
2857 }
2858 }
2859
2860 tools
2861 }
2862
2863 fn refresh_turn_tools(&mut self, cx: &App) {
2864 let tools = self.enabled_tools(cx);
2865 if let Some(turn) = self.running_turn.as_mut() {
2866 turn.tools = tools;
2867 }
2868 }
2869
2870 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2871 self.running_turn.as_ref()?.tools.get(name).cloned()
2872 }
2873
2874 pub fn has_tool(&self, name: &str) -> bool {
2875 self.running_turn
2876 .as_ref()
2877 .is_some_and(|turn| turn.tools.contains_key(name))
2878 }
2879
2880 #[cfg(any(test, feature = "test-support"))]
2881 pub fn has_registered_tool(&self, name: &str) -> bool {
2882 self.tools.contains_key(name)
2883 }
2884
2885 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2886 self.tools.keys().cloned().collect()
2887 }
2888
2889 pub(crate) fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2890 self.running_subagents.push(subagent);
2891 }
2892
2893 pub(crate) fn unregister_running_subagent(
2894 &mut self,
2895 subagent_session_id: &acp::SessionId,
2896 cx: &App,
2897 ) {
2898 self.running_subagents.retain(|s| {
2899 s.upgrade()
2900 .map_or(false, |s| s.read(cx).id() != subagent_session_id)
2901 });
2902 }
2903
2904 #[cfg(any(test, feature = "test-support"))]
2905 pub fn running_subagent_ids(&self, cx: &App) -> Vec<acp::SessionId> {
2906 self.running_subagents
2907 .iter()
2908 .filter_map(|s| s.upgrade().map(|s| s.read(cx).id().clone()))
2909 .collect()
2910 }
2911
2912 pub fn is_subagent(&self) -> bool {
2913 self.subagent_context.is_some()
2914 }
2915
2916 pub fn parent_thread_id(&self) -> Option<acp::SessionId> {
2917 self.subagent_context
2918 .as_ref()
2919 .map(|c| c.parent_thread_id.clone())
2920 }
2921
2922 pub fn depth(&self) -> u8 {
2923 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2924 }
2925
2926 #[cfg(any(test, feature = "test-support"))]
2927 pub fn set_subagent_context(&mut self, context: SubagentContext) {
2928 self.subagent_context = Some(context);
2929 }
2930
2931 pub fn is_turn_complete(&self) -> bool {
2932 self.running_turn.is_none()
2933 }
2934
2935 fn build_request_messages(
2936 &self,
2937 available_tools: Vec<SharedString>,
2938 cx: &App,
2939 ) -> Vec<LanguageModelRequestMessage> {
2940 log::trace!(
2941 "Building request messages from {} thread messages",
2942 self.messages.len()
2943 );
2944
2945 let system_prompt = SystemPromptTemplate {
2946 project: self.project_context.read(cx),
2947 available_tools,
2948 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2949 }
2950 .render(&self.templates)
2951 .context("failed to build system prompt")
2952 .expect("Invalid template");
2953 let mut messages = vec![LanguageModelRequestMessage {
2954 role: Role::System,
2955 content: vec![system_prompt.into()],
2956 cache: false,
2957 reasoning_details: None,
2958 }];
2959 for message in &self.messages {
2960 messages.extend(message.to_request());
2961 }
2962
2963 if let Some(last_message) = messages.last_mut() {
2964 last_message.cache = true;
2965 }
2966
2967 if let Some(message) = self.pending_message.as_ref() {
2968 messages.extend(message.to_request());
2969 }
2970
2971 messages
2972 }
2973
2974 pub fn to_markdown(&self) -> String {
2975 let mut markdown = String::new();
2976 for (ix, message) in self.messages.iter().enumerate() {
2977 if ix > 0 {
2978 markdown.push('\n');
2979 }
2980 match message {
2981 Message::User(_) => markdown.push_str("## User\n\n"),
2982 Message::Agent(_) => markdown.push_str("## Assistant\n\n"),
2983 Message::Resume => {}
2984 }
2985 markdown.push_str(&message.to_markdown());
2986 }
2987
2988 if let Some(message) = self.pending_message.as_ref() {
2989 markdown.push_str("\n## Assistant\n\n");
2990 markdown.push_str(&message.to_markdown());
2991 }
2992
2993 markdown
2994 }
2995
2996 fn advance_prompt_id(&mut self) {
2997 self.prompt_id = PromptId::new();
2998 }
2999
3000 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
3001 use LanguageModelCompletionError::*;
3002 use http_client::StatusCode;
3003
3004 // General strategy here:
3005 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
3006 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
3007 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
3008 match error {
3009 HttpResponseError {
3010 status_code: StatusCode::TOO_MANY_REQUESTS,
3011 ..
3012 } => Some(RetryStrategy::ExponentialBackoff {
3013 initial_delay: BASE_RETRY_DELAY,
3014 max_attempts: MAX_RETRY_ATTEMPTS,
3015 }),
3016 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
3017 Some(RetryStrategy::Fixed {
3018 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3019 max_attempts: MAX_RETRY_ATTEMPTS,
3020 })
3021 }
3022 UpstreamProviderError {
3023 status,
3024 retry_after,
3025 ..
3026 } => match *status {
3027 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
3028 Some(RetryStrategy::Fixed {
3029 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3030 max_attempts: MAX_RETRY_ATTEMPTS,
3031 })
3032 }
3033 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
3034 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3035 // Internal Server Error could be anything, retry up to 3 times.
3036 max_attempts: 3,
3037 }),
3038 status => {
3039 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
3040 // but we frequently get them in practice. See https://http.dev/529
3041 if status.as_u16() == 529 {
3042 Some(RetryStrategy::Fixed {
3043 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3044 max_attempts: MAX_RETRY_ATTEMPTS,
3045 })
3046 } else {
3047 Some(RetryStrategy::Fixed {
3048 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
3049 max_attempts: 2,
3050 })
3051 }
3052 }
3053 },
3054 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
3055 delay: BASE_RETRY_DELAY,
3056 max_attempts: 3,
3057 }),
3058 ApiReadResponseError { .. }
3059 | HttpSend { .. }
3060 | DeserializeResponse { .. }
3061 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
3062 delay: BASE_RETRY_DELAY,
3063 max_attempts: 3,
3064 }),
3065 // Retrying these errors definitely shouldn't help.
3066 HttpResponseError {
3067 status_code:
3068 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
3069 ..
3070 }
3071 | AuthenticationError { .. }
3072 | PermissionError { .. }
3073 | NoApiKey { .. }
3074 | ApiEndpointNotFound { .. }
3075 | PromptTooLarge { .. } => None,
3076 // These errors might be transient, so retry them
3077 SerializeRequest { .. } | BuildRequestBody { .. } | StreamEndedUnexpectedly { .. } => {
3078 Some(RetryStrategy::Fixed {
3079 delay: BASE_RETRY_DELAY,
3080 max_attempts: 1,
3081 })
3082 }
3083 // Retry all other 4xx and 5xx errors once.
3084 HttpResponseError { status_code, .. }
3085 if status_code.is_client_error() || status_code.is_server_error() =>
3086 {
3087 Some(RetryStrategy::Fixed {
3088 delay: BASE_RETRY_DELAY,
3089 max_attempts: 3,
3090 })
3091 }
3092 Other(err) if err.is::<language_model::PaymentRequiredError>() => {
3093 // Retrying won't help for Payment Required errors.
3094 None
3095 }
3096 // Conservatively assume that any other errors are non-retryable
3097 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
3098 delay: BASE_RETRY_DELAY,
3099 max_attempts: 2,
3100 }),
3101 }
3102 }
3103}
3104
3105struct RunningTurn {
3106 /// Holds the task that handles agent interaction until the end of the turn.
3107 /// Survives across multiple requests as the model performs tool calls and
3108 /// we run tools, report their results.
3109 _task: Task<()>,
3110 /// The current event stream for the running turn. Used to report a final
3111 /// cancellation event if we cancel the turn.
3112 event_stream: ThreadEventStream,
3113 /// The tools that are enabled for the current iteration of the turn.
3114 /// Refreshed at the start of each iteration via `refresh_turn_tools`.
3115 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
3116 /// Sender to signal tool cancellation. When cancel is called, this is
3117 /// set to true so all tools can detect user-initiated cancellation.
3118 cancellation_tx: watch::Sender<bool>,
3119 /// Senders for tools that support input streaming and have already been
3120 /// started but are still receiving input from the LLM.
3121 streaming_tool_inputs: HashMap<LanguageModelToolUseId, ToolInputSender>,
3122}
3123
3124impl RunningTurn {
3125 fn cancel(mut self) -> Task<()> {
3126 log::debug!("Cancelling in progress turn");
3127 self.cancellation_tx.send(true).ok();
3128 self.event_stream.send_canceled();
3129 self._task
3130 }
3131}
3132
3133pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
3134
3135impl EventEmitter<TokenUsageUpdated> for Thread {}
3136
3137pub struct TitleUpdated;
3138
3139impl EventEmitter<TitleUpdated> for Thread {}
3140
3141/// A channel-based wrapper that delivers tool input to a running tool.
3142///
3143/// For non-streaming tools, created via `ToolInput::ready()` so `.recv()` resolves immediately.
3144/// For streaming tools, partial JSON snapshots arrive via `.recv_partial()` as the LLM streams
3145/// them, followed by the final complete input available through `.recv()`.
3146pub struct ToolInput<T> {
3147 rx: mpsc::UnboundedReceiver<ToolInputPayload<serde_json::Value>>,
3148 _phantom: PhantomData<T>,
3149}
3150
3151impl<T: DeserializeOwned> ToolInput<T> {
3152 #[cfg(any(test, feature = "test-support"))]
3153 pub fn resolved(input: impl Serialize) -> Self {
3154 let value = serde_json::to_value(input).expect("failed to serialize tool input");
3155 Self::ready(value)
3156 }
3157
3158 pub fn ready(value: serde_json::Value) -> Self {
3159 let (tx, rx) = mpsc::unbounded();
3160 tx.unbounded_send(ToolInputPayload::Full(value)).ok();
3161 Self {
3162 rx,
3163 _phantom: PhantomData,
3164 }
3165 }
3166
3167 pub fn invalid_json(error_message: String) -> Self {
3168 let (tx, rx) = mpsc::unbounded();
3169 tx.unbounded_send(ToolInputPayload::InvalidJson { error_message })
3170 .ok();
3171 Self {
3172 rx,
3173 _phantom: PhantomData,
3174 }
3175 }
3176
3177 #[cfg(any(test, feature = "test-support"))]
3178 pub fn test() -> (ToolInputSender, Self) {
3179 let (sender, input) = ToolInputSender::channel();
3180 (sender, input.cast())
3181 }
3182
3183 /// Wait for the final deserialized input, ignoring all partial updates.
3184 /// Non-streaming tools can use this to wait until the whole input is available.
3185 pub async fn recv(mut self) -> Result<T> {
3186 while let Ok(value) = self.next().await {
3187 match value {
3188 ToolInputPayload::Full(value) => return Ok(value),
3189 ToolInputPayload::Partial(_) => {}
3190 ToolInputPayload::InvalidJson { error_message } => {
3191 return Err(anyhow!(error_message));
3192 }
3193 }
3194 }
3195 Err(anyhow!("tool input was not fully received"))
3196 }
3197
3198 pub async fn next(&mut self) -> Result<ToolInputPayload<T>> {
3199 let value = self
3200 .rx
3201 .next()
3202 .await
3203 .ok_or_else(|| anyhow!("tool input was not fully received"))?;
3204
3205 Ok(match value {
3206 ToolInputPayload::Partial(payload) => ToolInputPayload::Partial(payload),
3207 ToolInputPayload::Full(payload) => {
3208 ToolInputPayload::Full(serde_json::from_value(payload)?)
3209 }
3210 ToolInputPayload::InvalidJson { error_message } => {
3211 ToolInputPayload::InvalidJson { error_message }
3212 }
3213 })
3214 }
3215
3216 fn cast<U: DeserializeOwned>(self) -> ToolInput<U> {
3217 ToolInput {
3218 rx: self.rx,
3219 _phantom: PhantomData,
3220 }
3221 }
3222}
3223
3224pub enum ToolInputPayload<T> {
3225 Partial(serde_json::Value),
3226 Full(T),
3227 InvalidJson { error_message: String },
3228}
3229
3230pub struct ToolInputSender {
3231 has_received_final: bool,
3232 tx: mpsc::UnboundedSender<ToolInputPayload<serde_json::Value>>,
3233}
3234
3235impl ToolInputSender {
3236 pub(crate) fn channel() -> (Self, ToolInput<serde_json::Value>) {
3237 let (tx, rx) = mpsc::unbounded();
3238 let sender = Self {
3239 tx,
3240 has_received_final: false,
3241 };
3242 let input = ToolInput {
3243 rx,
3244 _phantom: PhantomData,
3245 };
3246 (sender, input)
3247 }
3248
3249 pub(crate) fn has_received_final(&self) -> bool {
3250 self.has_received_final
3251 }
3252
3253 pub fn send_partial(&mut self, payload: serde_json::Value) {
3254 self.tx
3255 .unbounded_send(ToolInputPayload::Partial(payload))
3256 .ok();
3257 }
3258
3259 pub fn send_full(&mut self, payload: serde_json::Value) {
3260 self.has_received_final = true;
3261 self.tx.unbounded_send(ToolInputPayload::Full(payload)).ok();
3262 }
3263
3264 pub fn send_invalid_json(&mut self, error_message: String) {
3265 self.has_received_final = true;
3266 self.tx
3267 .unbounded_send(ToolInputPayload::InvalidJson { error_message })
3268 .ok();
3269 }
3270}
3271
3272pub trait AgentTool
3273where
3274 Self: 'static + Sized,
3275{
3276 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
3277 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
3278
3279 const NAME: &'static str;
3280
3281 fn description() -> SharedString {
3282 let schema = schemars::schema_for!(Self::Input);
3283 SharedString::new(
3284 schema
3285 .get("description")
3286 .and_then(|description| description.as_str())
3287 .unwrap_or_default(),
3288 )
3289 }
3290
3291 fn kind() -> acp::ToolKind;
3292
3293 /// The initial tool title to display. Can be updated during the tool run.
3294 fn initial_title(
3295 &self,
3296 input: Result<Self::Input, serde_json::Value>,
3297 cx: &mut App,
3298 ) -> SharedString;
3299
3300 /// Returns the JSON schema that describes the tool's input.
3301 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
3302 language_model::tool_schema::root_schema_for::<Self::Input>(format)
3303 }
3304
3305 /// Returns whether the tool supports streaming of tool use parameters.
3306 fn supports_input_streaming() -> bool {
3307 false
3308 }
3309
3310 /// Some tools rely on a provider for the underlying billing or other reasons.
3311 /// Allow the tool to check if they are compatible, or should be filtered out.
3312 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
3313 true
3314 }
3315
3316 /// Runs the tool with the provided input.
3317 ///
3318 /// Returns `Result<Self::Output, Self::Output>` rather than `Result<Self::Output, anyhow::Error>`
3319 /// because tool errors are sent back to the model as tool results. This means error output must
3320 /// be structured and readable by the agent — not an arbitrary `anyhow::Error`. Returning the
3321 /// same `Output` type for both success and failure lets tools provide structured data while
3322 /// still signaling whether the invocation succeeded or failed.
3323 fn run(
3324 self: Arc<Self>,
3325 input: ToolInput<Self::Input>,
3326 event_stream: ToolCallEventStream,
3327 cx: &mut App,
3328 ) -> Task<Result<Self::Output, Self::Output>>;
3329
3330 /// Emits events for a previous execution of the tool.
3331 fn replay(
3332 &self,
3333 _input: Self::Input,
3334 _output: Self::Output,
3335 _event_stream: ToolCallEventStream,
3336 _cx: &mut App,
3337 ) -> Result<()> {
3338 Ok(())
3339 }
3340
3341 fn erase(self) -> Arc<dyn AnyAgentTool> {
3342 Arc::new(Erased(Arc::new(self)))
3343 }
3344}
3345
3346pub struct Erased<T>(T);
3347
3348pub struct AgentToolOutput {
3349 pub llm_output: LanguageModelToolResultContent,
3350 pub raw_output: serde_json::Value,
3351}
3352
3353impl AgentToolOutput {
3354 pub fn from_error(message: impl Into<String>) -> Self {
3355 let message = message.into();
3356 let llm_output = LanguageModelToolResultContent::Text(Arc::from(message.as_str()));
3357 Self {
3358 raw_output: serde_json::Value::String(message),
3359 llm_output,
3360 }
3361 }
3362}
3363
3364pub trait AnyAgentTool {
3365 fn name(&self) -> SharedString;
3366 fn description(&self) -> SharedString;
3367 fn kind(&self) -> acp::ToolKind;
3368 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
3369 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
3370 fn supports_input_streaming(&self) -> bool {
3371 false
3372 }
3373 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
3374 true
3375 }
3376 /// See [`AgentTool::run`] for why this returns `Result<AgentToolOutput, AgentToolOutput>`.
3377 fn run(
3378 self: Arc<Self>,
3379 input: ToolInput<serde_json::Value>,
3380 event_stream: ToolCallEventStream,
3381 cx: &mut App,
3382 ) -> Task<Result<AgentToolOutput, AgentToolOutput>>;
3383 fn replay(
3384 &self,
3385 input: serde_json::Value,
3386 output: serde_json::Value,
3387 event_stream: ToolCallEventStream,
3388 cx: &mut App,
3389 ) -> Result<()>;
3390}
3391
3392impl<T> AnyAgentTool for Erased<Arc<T>>
3393where
3394 T: AgentTool,
3395{
3396 fn name(&self) -> SharedString {
3397 T::NAME.into()
3398 }
3399
3400 fn description(&self) -> SharedString {
3401 T::description()
3402 }
3403
3404 fn kind(&self) -> agent_client_protocol::ToolKind {
3405 T::kind()
3406 }
3407
3408 fn supports_input_streaming(&self) -> bool {
3409 T::supports_input_streaming()
3410 }
3411
3412 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
3413 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
3414 self.0.initial_title(parsed_input, _cx)
3415 }
3416
3417 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
3418 let mut json = serde_json::to_value(T::input_schema(format))?;
3419 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
3420 Ok(json)
3421 }
3422
3423 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
3424 T::supports_provider(provider)
3425 }
3426
3427 fn run(
3428 self: Arc<Self>,
3429 input: ToolInput<serde_json::Value>,
3430 event_stream: ToolCallEventStream,
3431 cx: &mut App,
3432 ) -> Task<Result<AgentToolOutput, AgentToolOutput>> {
3433 let tool_input: ToolInput<T::Input> = input.cast();
3434 let task = self.0.clone().run(tool_input, event_stream, cx);
3435 cx.spawn(async move |_cx| match task.await {
3436 Ok(output) => {
3437 let raw_output = serde_json::to_value(&output).map_err(|e| {
3438 AgentToolOutput::from_error(format!("Failed to serialize tool output: {e}"))
3439 })?;
3440 Ok(AgentToolOutput {
3441 llm_output: output.into(),
3442 raw_output,
3443 })
3444 }
3445 Err(error_output) => {
3446 let raw_output = serde_json::to_value(&error_output).unwrap_or_else(|e| {
3447 log::error!("Failed to serialize tool error output: {e}");
3448 serde_json::Value::Null
3449 });
3450 Err(AgentToolOutput {
3451 llm_output: error_output.into(),
3452 raw_output,
3453 })
3454 }
3455 })
3456 }
3457
3458 fn replay(
3459 &self,
3460 input: serde_json::Value,
3461 output: serde_json::Value,
3462 event_stream: ToolCallEventStream,
3463 cx: &mut App,
3464 ) -> Result<()> {
3465 let input = serde_json::from_value(input)?;
3466 let output = serde_json::from_value(output)?;
3467 self.0.replay(input, output, event_stream, cx)
3468 }
3469}
3470
3471#[derive(Clone)]
3472struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
3473
3474impl ThreadEventStream {
3475 fn send_user_message(&self, message: &UserMessage) {
3476 self.0
3477 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
3478 .ok();
3479 }
3480
3481 fn send_text(&self, text: &str) {
3482 self.0
3483 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
3484 .ok();
3485 }
3486
3487 fn send_thinking(&self, text: &str) {
3488 self.0
3489 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
3490 .ok();
3491 }
3492
3493 fn send_tool_call(
3494 &self,
3495 id: &LanguageModelToolUseId,
3496 tool_name: &str,
3497 title: SharedString,
3498 kind: acp::ToolKind,
3499 input: serde_json::Value,
3500 ) {
3501 self.0
3502 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
3503 id,
3504 tool_name,
3505 title.to_string(),
3506 kind,
3507 input,
3508 ))))
3509 .ok();
3510 }
3511
3512 fn initial_tool_call(
3513 id: &LanguageModelToolUseId,
3514 tool_name: &str,
3515 title: String,
3516 kind: acp::ToolKind,
3517 input: serde_json::Value,
3518 ) -> acp::ToolCall {
3519 acp::ToolCall::new(id.to_string(), title)
3520 .kind(kind)
3521 .raw_input(input)
3522 .meta(acp_thread::meta_with_tool_name(tool_name))
3523 }
3524
3525 fn update_tool_call_fields(
3526 &self,
3527 tool_use_id: &LanguageModelToolUseId,
3528 fields: acp::ToolCallUpdateFields,
3529 meta: Option<acp::Meta>,
3530 ) {
3531 self.0
3532 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3533 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields)
3534 .meta(meta)
3535 .into(),
3536 )))
3537 .ok();
3538 }
3539
3540 fn send_plan(&self, plan: acp::Plan) {
3541 self.0.unbounded_send(Ok(ThreadEvent::Plan(plan))).ok();
3542 }
3543
3544 fn send_retry(&self, status: acp_thread::RetryStatus) {
3545 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
3546 }
3547
3548 fn send_stop(&self, reason: acp::StopReason) {
3549 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
3550 }
3551
3552 fn send_canceled(&self) {
3553 self.0
3554 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
3555 .ok();
3556 }
3557
3558 fn send_error(&self, error: impl Into<anyhow::Error>) {
3559 self.0.unbounded_send(Err(error.into())).ok();
3560 }
3561}
3562
3563#[derive(Clone)]
3564pub struct ToolCallEventStream {
3565 tool_use_id: LanguageModelToolUseId,
3566 stream: ThreadEventStream,
3567 fs: Option<Arc<dyn Fs>>,
3568 cancellation_rx: watch::Receiver<bool>,
3569}
3570
3571impl ToolCallEventStream {
3572 #[cfg(any(test, feature = "test-support"))]
3573 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
3574 let (stream, receiver, _cancellation_tx) = Self::test_with_cancellation();
3575 (stream, receiver)
3576 }
3577
3578 #[cfg(any(test, feature = "test-support"))]
3579 pub fn test_with_cancellation() -> (Self, ToolCallEventStreamReceiver, watch::Sender<bool>) {
3580 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
3581 let (cancellation_tx, cancellation_rx) = watch::channel(false);
3582
3583 let stream = ToolCallEventStream::new(
3584 "test_id".into(),
3585 ThreadEventStream(events_tx),
3586 None,
3587 cancellation_rx,
3588 );
3589
3590 (
3591 stream,
3592 ToolCallEventStreamReceiver(events_rx),
3593 cancellation_tx,
3594 )
3595 }
3596
3597 /// Signal cancellation for this event stream. Only available in tests.
3598 #[cfg(any(test, feature = "test-support"))]
3599 pub fn signal_cancellation_with_sender(cancellation_tx: &mut watch::Sender<bool>) {
3600 cancellation_tx.send(true).ok();
3601 }
3602
3603 fn new(
3604 tool_use_id: LanguageModelToolUseId,
3605 stream: ThreadEventStream,
3606 fs: Option<Arc<dyn Fs>>,
3607 cancellation_rx: watch::Receiver<bool>,
3608 ) -> Self {
3609 Self {
3610 tool_use_id,
3611 stream,
3612 fs,
3613 cancellation_rx,
3614 }
3615 }
3616
3617 /// Returns a future that resolves when the user cancels the tool call.
3618 /// Tools should select on this alongside their main work to detect user cancellation.
3619 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
3620 let mut rx = self.cancellation_rx.clone();
3621 async move {
3622 loop {
3623 if *rx.borrow() {
3624 return;
3625 }
3626 if rx.changed().await.is_err() {
3627 // Sender dropped, will never be cancelled
3628 std::future::pending::<()>().await;
3629 }
3630 }
3631 }
3632 }
3633
3634 /// Returns true if the user has cancelled this tool call.
3635 /// This is useful for checking cancellation state after an operation completes,
3636 /// to determine if the completion was due to user cancellation.
3637 pub fn was_cancelled_by_user(&self) -> bool {
3638 *self.cancellation_rx.clone().borrow()
3639 }
3640
3641 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
3642 &self.tool_use_id
3643 }
3644
3645 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
3646 self.stream
3647 .update_tool_call_fields(&self.tool_use_id, fields, None);
3648 }
3649
3650 pub fn update_fields_with_meta(
3651 &self,
3652 fields: acp::ToolCallUpdateFields,
3653 meta: Option<acp::Meta>,
3654 ) {
3655 self.stream
3656 .update_tool_call_fields(&self.tool_use_id, fields, meta);
3657 }
3658
3659 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
3660 self.stream
3661 .0
3662 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3663 acp_thread::ToolCallUpdateDiff {
3664 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3665 diff,
3666 }
3667 .into(),
3668 )))
3669 .ok();
3670 }
3671
3672 pub fn subagent_spawned(&self, id: acp::SessionId) {
3673 self.stream
3674 .0
3675 .unbounded_send(Ok(ThreadEvent::SubagentSpawned(id)))
3676 .ok();
3677 }
3678
3679 pub fn update_plan(&self, plan: acp::Plan) {
3680 self.stream.send_plan(plan);
3681 }
3682
3683 /// Authorize a third-party tool (e.g., MCP tool from a context server).
3684 ///
3685 /// Unlike built-in tools, third-party tools don't support pattern-based permissions.
3686 /// They only support `default` (allow/deny/confirm) per tool.
3687 ///
3688 /// Uses the dropdown authorization flow with two granularities:
3689 /// - "Always for <display_name> MCP tool" → sets `tools.<tool_id>.default = "allow"` or "deny"
3690 /// - "Only this time" → allow/deny once
3691 pub fn authorize_third_party_tool(
3692 &self,
3693 title: impl Into<String>,
3694 tool_id: String,
3695 display_name: String,
3696 cx: &mut App,
3697 ) -> Task<Result<()>> {
3698 let settings = agent_settings::AgentSettings::get_global(cx);
3699
3700 let decision = decide_permission_from_settings(&tool_id, &[String::new()], &settings);
3701
3702 match decision {
3703 ToolPermissionDecision::Allow => return Task::ready(Ok(())),
3704 ToolPermissionDecision::Deny(reason) => return Task::ready(Err(anyhow!(reason))),
3705 ToolPermissionDecision::Confirm => {}
3706 }
3707
3708 let (response_tx, response_rx) = oneshot::channel();
3709 if let Err(error) = self
3710 .stream
3711 .0
3712 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3713 ToolCallAuthorization {
3714 tool_call: acp::ToolCallUpdate::new(
3715 self.tool_use_id.to_string(),
3716 acp::ToolCallUpdateFields::new().title(title.into()),
3717 ),
3718 options: acp_thread::PermissionOptions::Dropdown(vec![
3719 acp_thread::PermissionOptionChoice {
3720 allow: acp::PermissionOption::new(
3721 acp::PermissionOptionId::new(format!(
3722 "always_allow_mcp:{}",
3723 tool_id
3724 )),
3725 format!("Always for {} MCP tool", display_name),
3726 acp::PermissionOptionKind::AllowAlways,
3727 ),
3728 deny: acp::PermissionOption::new(
3729 acp::PermissionOptionId::new(format!(
3730 "always_deny_mcp:{}",
3731 tool_id
3732 )),
3733 format!("Always for {} MCP tool", display_name),
3734 acp::PermissionOptionKind::RejectAlways,
3735 ),
3736 sub_patterns: vec![],
3737 },
3738 acp_thread::PermissionOptionChoice {
3739 allow: acp::PermissionOption::new(
3740 acp::PermissionOptionId::new("allow"),
3741 "Only this time",
3742 acp::PermissionOptionKind::AllowOnce,
3743 ),
3744 deny: acp::PermissionOption::new(
3745 acp::PermissionOptionId::new("deny"),
3746 "Only this time",
3747 acp::PermissionOptionKind::RejectOnce,
3748 ),
3749 sub_patterns: vec![],
3750 },
3751 ]),
3752 response: response_tx,
3753 context: None,
3754 },
3755 )))
3756 {
3757 log::error!("Failed to send tool call authorization: {error}");
3758 return Task::ready(Err(anyhow!(
3759 "Failed to send tool call authorization: {error}"
3760 )));
3761 }
3762
3763 let fs = self.fs.clone();
3764 cx.spawn(async move |cx| {
3765 let outcome = response_rx.await?;
3766 let is_allow = Self::persist_permission_outcome(&outcome, fs, &cx);
3767 if is_allow {
3768 Ok(())
3769 } else {
3770 Err(anyhow!("Permission to run tool denied by user"))
3771 }
3772 })
3773 }
3774
3775 pub fn authorize(
3776 &self,
3777 title: impl Into<String>,
3778 context: ToolPermissionContext,
3779 cx: &mut App,
3780 ) -> Task<Result<()>> {
3781 let options = context.build_permission_options();
3782
3783 let (response_tx, response_rx) = oneshot::channel();
3784 if let Err(error) = self
3785 .stream
3786 .0
3787 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3788 ToolCallAuthorization {
3789 tool_call: acp::ToolCallUpdate::new(
3790 self.tool_use_id.to_string(),
3791 acp::ToolCallUpdateFields::new().title(title.into()),
3792 ),
3793 options,
3794 response: response_tx,
3795 context: Some(context),
3796 },
3797 )))
3798 {
3799 log::error!("Failed to send tool call authorization: {error}");
3800 return Task::ready(Err(anyhow!(
3801 "Failed to send tool call authorization: {error}"
3802 )));
3803 }
3804
3805 let fs = self.fs.clone();
3806 cx.spawn(async move |cx| {
3807 let outcome = response_rx.await?;
3808 let is_allow = Self::persist_permission_outcome(&outcome, fs, &cx);
3809 if is_allow {
3810 Ok(())
3811 } else {
3812 Err(anyhow!("Permission to run tool denied by user"))
3813 }
3814 })
3815 }
3816
3817 /// Interprets a `SelectedPermissionOutcome` and persists any settings changes.
3818 /// Returns `true` if the tool call should be allowed, `false` if denied.
3819 fn persist_permission_outcome(
3820 outcome: &acp_thread::SelectedPermissionOutcome,
3821 fs: Option<Arc<dyn Fs>>,
3822 cx: &AsyncApp,
3823 ) -> bool {
3824 let option_id = outcome.option_id.0.as_ref();
3825
3826 let always_permission = option_id
3827 .strip_prefix("always_allow:")
3828 .map(|tool| (tool, ToolPermissionMode::Allow))
3829 .or_else(|| {
3830 option_id
3831 .strip_prefix("always_deny:")
3832 .map(|tool| (tool, ToolPermissionMode::Deny))
3833 })
3834 .or_else(|| {
3835 option_id
3836 .strip_prefix("always_allow_mcp:")
3837 .map(|tool| (tool, ToolPermissionMode::Allow))
3838 })
3839 .or_else(|| {
3840 option_id
3841 .strip_prefix("always_deny_mcp:")
3842 .map(|tool| (tool, ToolPermissionMode::Deny))
3843 });
3844
3845 if let Some((tool, mode)) = always_permission {
3846 let params = outcome.params.as_ref();
3847 Self::persist_always_permission(tool, mode, params, fs, cx);
3848 return mode == ToolPermissionMode::Allow;
3849 }
3850
3851 // Handle simple "allow" / "deny" (once, no persistence)
3852 if option_id == "allow" || option_id == "deny" {
3853 debug_assert!(
3854 outcome.params.is_none(),
3855 "unexpected params for once-only permission"
3856 );
3857 return option_id == "allow";
3858 }
3859
3860 debug_assert!(false, "unexpected permission option_id: {option_id}");
3861 false
3862 }
3863
3864 /// Persists an "always allow" or "always deny" permission, using sub_patterns
3865 /// from params when present.
3866 fn persist_always_permission(
3867 tool: &str,
3868 mode: ToolPermissionMode,
3869 params: Option<&acp_thread::SelectedPermissionParams>,
3870 fs: Option<Arc<dyn Fs>>,
3871 cx: &AsyncApp,
3872 ) {
3873 let Some(fs) = fs else {
3874 return;
3875 };
3876
3877 match params {
3878 Some(acp_thread::SelectedPermissionParams::Terminal {
3879 patterns: sub_patterns,
3880 }) => {
3881 debug_assert!(
3882 !sub_patterns.is_empty(),
3883 "empty sub_patterns for tool {tool} — callers should pass None instead"
3884 );
3885 let tool = tool.to_string();
3886 let sub_patterns = sub_patterns.clone();
3887 cx.update(|cx| {
3888 update_settings_file(fs, cx, move |settings, _| {
3889 let agent = settings.agent.get_or_insert_default();
3890 for pattern in sub_patterns {
3891 match mode {
3892 ToolPermissionMode::Allow => {
3893 agent.add_tool_allow_pattern(&tool, pattern);
3894 }
3895 ToolPermissionMode::Deny => {
3896 agent.add_tool_deny_pattern(&tool, pattern);
3897 }
3898 // If there's no matching pattern this will
3899 // default to confirm, so falling through is
3900 // fine here.
3901 ToolPermissionMode::Confirm => (),
3902 }
3903 }
3904 });
3905 });
3906 }
3907 None => {
3908 let tool = tool.to_string();
3909 cx.update(|cx| {
3910 update_settings_file(fs, cx, move |settings, _| {
3911 settings
3912 .agent
3913 .get_or_insert_default()
3914 .set_tool_default_permission(&tool, mode);
3915 });
3916 });
3917 }
3918 }
3919 }
3920}
3921
3922#[cfg(any(test, feature = "test-support"))]
3923pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
3924
3925#[cfg(any(test, feature = "test-support"))]
3926impl ToolCallEventStreamReceiver {
3927 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
3928 let event = self.0.next().await;
3929 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
3930 auth
3931 } else {
3932 panic!("Expected ToolCallAuthorization but got: {:?}", event);
3933 }
3934 }
3935
3936 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
3937 let event = self.0.next().await;
3938 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
3939 update,
3940 )))) = event
3941 {
3942 update.fields
3943 } else {
3944 panic!("Expected update fields but got: {:?}", event);
3945 }
3946 }
3947
3948 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
3949 let event = self.0.next().await;
3950 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
3951 update,
3952 )))) = event
3953 {
3954 update.diff
3955 } else {
3956 panic!("Expected diff but got: {:?}", event);
3957 }
3958 }
3959
3960 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
3961 let event = self.0.next().await;
3962 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
3963 update,
3964 )))) = event
3965 {
3966 update.terminal
3967 } else {
3968 panic!("Expected terminal but got: {:?}", event);
3969 }
3970 }
3971
3972 pub async fn expect_plan(&mut self) -> acp::Plan {
3973 let event = self.0.next().await;
3974 if let Some(Ok(ThreadEvent::Plan(plan))) = event {
3975 plan
3976 } else {
3977 panic!("Expected plan but got: {:?}", event);
3978 }
3979 }
3980}
3981
3982#[cfg(any(test, feature = "test-support"))]
3983impl std::ops::Deref for ToolCallEventStreamReceiver {
3984 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
3985
3986 fn deref(&self) -> &Self::Target {
3987 &self.0
3988 }
3989}
3990
3991#[cfg(any(test, feature = "test-support"))]
3992impl std::ops::DerefMut for ToolCallEventStreamReceiver {
3993 fn deref_mut(&mut self) -> &mut Self::Target {
3994 &mut self.0
3995 }
3996}
3997
3998impl From<&str> for UserMessageContent {
3999 fn from(text: &str) -> Self {
4000 Self::Text(text.into())
4001 }
4002}
4003
4004impl From<String> for UserMessageContent {
4005 fn from(text: String) -> Self {
4006 Self::Text(text)
4007 }
4008}
4009
4010impl UserMessageContent {
4011 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
4012 match value {
4013 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
4014 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
4015 acp::ContentBlock::Audio(_) => {
4016 // TODO
4017 Self::Text("[audio]".to_string())
4018 }
4019 acp::ContentBlock::ResourceLink(resource_link) => {
4020 match MentionUri::parse(&resource_link.uri, path_style) {
4021 Ok(uri) => Self::Mention {
4022 uri,
4023 content: String::new(),
4024 },
4025 Err(err) => {
4026 log::error!("Failed to parse mention link: {}", err);
4027 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
4028 }
4029 }
4030 }
4031 acp::ContentBlock::Resource(resource) => match resource.resource {
4032 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
4033 match MentionUri::parse(&resource.uri, path_style) {
4034 Ok(uri) => Self::Mention {
4035 uri,
4036 content: resource.text,
4037 },
4038 Err(err) => {
4039 log::error!("Failed to parse mention link: {}", err);
4040 Self::Text(
4041 MarkdownCodeBlock {
4042 tag: &resource.uri,
4043 text: &resource.text,
4044 }
4045 .to_string(),
4046 )
4047 }
4048 }
4049 }
4050 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
4051 // TODO
4052 Self::Text("[blob]".to_string())
4053 }
4054 other => {
4055 log::warn!("Unexpected content type: {:?}", other);
4056 Self::Text("[unknown]".to_string())
4057 }
4058 },
4059 other => {
4060 log::warn!("Unexpected content type: {:?}", other);
4061 Self::Text("[unknown]".to_string())
4062 }
4063 }
4064 }
4065}
4066
4067impl From<UserMessageContent> for acp::ContentBlock {
4068 fn from(content: UserMessageContent) -> Self {
4069 match content {
4070 UserMessageContent::Text(text) => text.into(),
4071 UserMessageContent::Image(image) => {
4072 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
4073 }
4074 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
4075 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
4076 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
4077 )),
4078 ),
4079 }
4080 }
4081}
4082
4083fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
4084 LanguageModelImage {
4085 source: image_content.data.into(),
4086 size: None,
4087 }
4088}
4089
4090#[cfg(test)]
4091mod tests {
4092 use super::*;
4093 use gpui::TestAppContext;
4094 use language_model::LanguageModelToolUseId;
4095 use language_model::fake_provider::FakeLanguageModel;
4096 use serde_json::json;
4097 use std::sync::Arc;
4098
4099 async fn setup_thread_for_test(cx: &mut TestAppContext) -> (Entity<Thread>, ThreadEventStream) {
4100 cx.update(|cx| {
4101 let settings_store = settings::SettingsStore::test(cx);
4102 cx.set_global(settings_store);
4103 });
4104
4105 let fs = fs::FakeFs::new(cx.background_executor.clone());
4106 let templates = Templates::new();
4107 let project = Project::test(fs.clone(), [], cx).await;
4108
4109 cx.update(|cx| {
4110 let project_context = cx.new(|_cx| prompt_store::ProjectContext::default());
4111 let context_server_store = project.read(cx).context_server_store();
4112 let context_server_registry =
4113 cx.new(|cx| ContextServerRegistry::new(context_server_store, cx));
4114
4115 let thread = cx.new(|cx| {
4116 Thread::new(
4117 project,
4118 project_context,
4119 context_server_registry,
4120 templates,
4121 None,
4122 cx,
4123 )
4124 });
4125
4126 let (event_tx, _event_rx) = mpsc::unbounded();
4127 let event_stream = ThreadEventStream(event_tx);
4128
4129 (thread, event_stream)
4130 })
4131 }
4132
4133 fn setup_parent_with_subagents(
4134 cx: &mut TestAppContext,
4135 parent: &Entity<Thread>,
4136 count: usize,
4137 ) -> Vec<Entity<Thread>> {
4138 cx.update(|cx| {
4139 let mut subagents = Vec::new();
4140 for _ in 0..count {
4141 let subagent = cx.new(|cx| Thread::new_subagent(parent, cx));
4142 parent.update(cx, |thread, _cx| {
4143 thread.register_running_subagent(subagent.downgrade());
4144 });
4145 subagents.push(subagent);
4146 }
4147 subagents
4148 })
4149 }
4150
4151 #[gpui::test]
4152 async fn test_set_model_propagates_to_subagents(cx: &mut TestAppContext) {
4153 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4154 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4155
4156 let new_model: Arc<dyn LanguageModel> = Arc::new(FakeLanguageModel::with_id_and_thinking(
4157 "test-provider",
4158 "new-model",
4159 "New Model",
4160 false,
4161 ));
4162
4163 cx.update(|cx| {
4164 parent.update(cx, |thread, cx| {
4165 thread.set_model(new_model, cx);
4166 });
4167
4168 for subagent in &subagents {
4169 let subagent_model_id = subagent.read(cx).model().unwrap().id();
4170 assert_eq!(
4171 subagent_model_id.0.as_ref(),
4172 "new-model",
4173 "Subagent model should match parent model after set_model"
4174 );
4175 }
4176 });
4177 }
4178
4179 #[gpui::test]
4180 async fn test_set_summarization_model_propagates_to_subagents(cx: &mut TestAppContext) {
4181 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4182 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4183
4184 let summary_model: Arc<dyn LanguageModel> =
4185 Arc::new(FakeLanguageModel::with_id_and_thinking(
4186 "test-provider",
4187 "summary-model",
4188 "Summary Model",
4189 false,
4190 ));
4191
4192 cx.update(|cx| {
4193 parent.update(cx, |thread, cx| {
4194 thread.set_summarization_model(Some(summary_model), cx);
4195 });
4196
4197 for subagent in &subagents {
4198 let subagent_summary_id = subagent.read(cx).summarization_model().unwrap().id();
4199 assert_eq!(
4200 subagent_summary_id.0.as_ref(),
4201 "summary-model",
4202 "Subagent summarization model should match parent after set_summarization_model"
4203 );
4204 }
4205 });
4206 }
4207
4208 #[gpui::test]
4209 async fn test_set_thinking_enabled_propagates_to_subagents(cx: &mut TestAppContext) {
4210 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4211 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4212
4213 cx.update(|cx| {
4214 parent.update(cx, |thread, cx| {
4215 thread.set_thinking_enabled(true, cx);
4216 });
4217
4218 for subagent in &subagents {
4219 assert!(
4220 subagent.read(cx).thinking_enabled(),
4221 "Subagent thinking should be enabled after parent enables it"
4222 );
4223 }
4224
4225 parent.update(cx, |thread, cx| {
4226 thread.set_thinking_enabled(false, cx);
4227 });
4228
4229 for subagent in &subagents {
4230 assert!(
4231 !subagent.read(cx).thinking_enabled(),
4232 "Subagent thinking should be disabled after parent disables it"
4233 );
4234 }
4235 });
4236 }
4237
4238 #[gpui::test]
4239 async fn test_set_thinking_effort_propagates_to_subagents(cx: &mut TestAppContext) {
4240 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4241 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4242
4243 cx.update(|cx| {
4244 parent.update(cx, |thread, cx| {
4245 thread.set_thinking_effort(Some("high".to_string()), cx);
4246 });
4247
4248 for subagent in &subagents {
4249 assert_eq!(
4250 subagent.read(cx).thinking_effort().map(|s| s.as_str()),
4251 Some("high"),
4252 "Subagent thinking effort should match parent"
4253 );
4254 }
4255
4256 parent.update(cx, |thread, cx| {
4257 thread.set_thinking_effort(None, cx);
4258 });
4259
4260 for subagent in &subagents {
4261 assert_eq!(
4262 subagent.read(cx).thinking_effort(),
4263 None,
4264 "Subagent thinking effort should be None after parent clears it"
4265 );
4266 }
4267 });
4268 }
4269
4270 #[gpui::test]
4271 async fn test_set_speed_propagates_to_subagents(cx: &mut TestAppContext) {
4272 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4273 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4274
4275 cx.update(|cx| {
4276 parent.update(cx, |thread, cx| {
4277 thread.set_speed(Speed::Fast, cx);
4278 });
4279
4280 for subagent in &subagents {
4281 assert_eq!(
4282 subagent.read(cx).speed(),
4283 Some(Speed::Fast),
4284 "Subagent speed should match parent after set_speed"
4285 );
4286 }
4287 });
4288 }
4289
4290 #[gpui::test]
4291 async fn test_dropped_subagent_does_not_panic(cx: &mut TestAppContext) {
4292 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4293 let subagents = setup_parent_with_subagents(cx, &parent, 1);
4294
4295 // Drop the subagent so the WeakEntity can no longer be upgraded
4296 drop(subagents);
4297
4298 // Should not panic even though the subagent was dropped
4299 cx.update(|cx| {
4300 parent.update(cx, |thread, cx| {
4301 thread.set_thinking_enabled(true, cx);
4302 thread.set_speed(Speed::Fast, cx);
4303 thread.set_thinking_effort(Some("high".to_string()), cx);
4304 });
4305 });
4306 }
4307
4308 #[gpui::test]
4309 async fn test_handle_tool_use_json_parse_error_adds_tool_use_to_content(
4310 cx: &mut TestAppContext,
4311 ) {
4312 let (thread, event_stream) = setup_thread_for_test(cx).await;
4313
4314 let tool_use_id = LanguageModelToolUseId::from("test_tool_id");
4315 let tool_name: Arc<str> = Arc::from("test_tool");
4316 let raw_input: Arc<str> = Arc::from("{invalid json");
4317 let json_parse_error = "expected value at line 1 column 1".to_string();
4318
4319 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
4320
4321 let result = cx
4322 .update(|cx| {
4323 thread.update(cx, |thread, cx| {
4324 // Call the function under test
4325 thread
4326 .handle_tool_use_json_parse_error_event(
4327 tool_use_id.clone(),
4328 tool_name.clone(),
4329 raw_input.clone(),
4330 json_parse_error,
4331 &event_stream,
4332 cancellation_rx,
4333 cx,
4334 )
4335 .unwrap()
4336 })
4337 })
4338 .await;
4339
4340 // Verify the result is an error
4341 assert!(result.is_error);
4342 assert_eq!(result.tool_use_id, tool_use_id);
4343 assert_eq!(result.tool_name, tool_name);
4344 assert!(matches!(
4345 result.content,
4346 LanguageModelToolResultContent::Text(_)
4347 ));
4348
4349 thread.update(cx, |thread, _cx| {
4350 // Verify the tool use was added to the message content
4351 {
4352 let last_message = thread.pending_message();
4353 assert_eq!(
4354 last_message.content.len(),
4355 1,
4356 "Should have one tool_use in content"
4357 );
4358
4359 match &last_message.content[0] {
4360 AgentMessageContent::ToolUse(tool_use) => {
4361 assert_eq!(tool_use.id, tool_use_id);
4362 assert_eq!(tool_use.name, tool_name);
4363 assert_eq!(tool_use.raw_input, raw_input.to_string());
4364 assert!(tool_use.is_input_complete);
4365 // Should fall back to empty object for invalid JSON
4366 assert_eq!(tool_use.input, json!({}));
4367 }
4368 _ => panic!("Expected ToolUse content"),
4369 }
4370 }
4371
4372 // Insert the tool result (simulating what the caller does)
4373 thread
4374 .pending_message()
4375 .tool_results
4376 .insert(result.tool_use_id.clone(), result);
4377
4378 // Verify the tool result was added
4379 let last_message = thread.pending_message();
4380 assert_eq!(
4381 last_message.tool_results.len(),
4382 1,
4383 "Should have one tool_result"
4384 );
4385 assert!(last_message.tool_results.contains_key(&tool_use_id));
4386 })
4387 }
4388}