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