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