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