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