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