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