1use crate::{
2 ContextServerRegistry, CopyPathTool, CreateDirectoryTool, DbLanguageModel, DbThread,
3 DeletePathTool, DiagnosticsTool, EditFileTool, FetchTool, FindPathTool, GrepTool,
4 ListDirectoryTool, MovePathTool, NowTool, OpenTool, ProjectSnapshot, ReadFileTool,
5 RestoreFileFromDiskTool, SaveFileTool, SpawnAgentTool, StreamingEditFileTool,
6 SystemPromptTemplate, Template, Templates, TerminalTool, ToolPermissionDecision, WebSearchTool,
7 decide_permission_from_settings,
8};
9use acp_thread::{MentionUri, UserMessageId};
10use action_log::ActionLog;
11use feature_flags::{FeatureFlagAppExt as _, StreamingEditFileToolFeatureFlag};
12
13use agent_client_protocol as acp;
14use agent_settings::{
15 AgentProfileId, AgentProfileSettings, AgentSettings, SUMMARIZE_THREAD_DETAILED_PROMPT,
16 SUMMARIZE_THREAD_PROMPT,
17};
18use anyhow::{Context as _, Result, anyhow};
19use chrono::{DateTime, Utc};
20use client::UserStore;
21use cloud_api_types::Plan;
22use cloud_llm_client::CompletionIntent;
23use collections::{HashMap, HashSet, IndexMap};
24use fs::Fs;
25use futures::stream;
26use futures::{
27 FutureExt,
28 channel::{mpsc, oneshot},
29 future::Shared,
30 stream::FuturesUnordered,
31};
32use gpui::{
33 App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task, WeakEntity,
34};
35use heck::ToSnakeCase as _;
36use language_model::{
37 LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,
38 LanguageModelImage, LanguageModelProviderId, LanguageModelRegistry, LanguageModelRequest,
39 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
40 LanguageModelToolResultContent, LanguageModelToolSchemaFormat, LanguageModelToolUse,
41 LanguageModelToolUseId, Role, SelectedModel, Speed, StopReason, TokenUsage,
42 ZED_CLOUD_PROVIDER_ID,
43};
44use project::Project;
45use prompt_store::ProjectContext;
46use schemars::{JsonSchema, Schema};
47use serde::de::DeserializeOwned;
48use serde::{Deserialize, Serialize};
49use settings::{LanguageModelSelection, Settings, ToolPermissionMode, update_settings_file};
50use smol::stream::StreamExt;
51use std::{
52 collections::BTreeMap,
53 marker::PhantomData,
54 ops::RangeInclusive,
55 path::Path,
56 rc::Rc,
57 sync::Arc,
58 time::{Duration, Instant},
59};
60use std::{fmt::Write, path::PathBuf};
61use util::{ResultExt, debug_panic, markdown::MarkdownCodeBlock, paths::PathStyle};
62use uuid::Uuid;
63
64const TOOL_CANCELED_MESSAGE: &str = "Tool canceled by user";
65pub const MAX_TOOL_NAME_LENGTH: usize = 64;
66pub const MAX_SUBAGENT_DEPTH: u8 = 1;
67
68/// Context passed to a subagent thread for lifecycle management
69#[derive(Clone, Debug, Serialize, Deserialize)]
70pub struct SubagentContext {
71 /// ID of the parent thread
72 pub parent_thread_id: acp::SessionId,
73
74 /// Current depth level (0 = root agent, 1 = first-level subagent, etc.)
75 pub depth: u8,
76}
77
78/// The ID of the user prompt that initiated a request.
79///
80/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
81#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
82pub struct PromptId(Arc<str>);
83
84impl PromptId {
85 pub fn new() -> Self {
86 Self(Uuid::new_v4().to_string().into())
87 }
88}
89
90impl std::fmt::Display for PromptId {
91 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
92 write!(f, "{}", self.0)
93 }
94}
95
96pub(crate) const MAX_RETRY_ATTEMPTS: u8 = 4;
97pub(crate) const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
98
99#[derive(Debug, Clone)]
100enum RetryStrategy {
101 ExponentialBackoff {
102 initial_delay: Duration,
103 max_attempts: u8,
104 },
105 Fixed {
106 delay: Duration,
107 max_attempts: u8,
108 },
109}
110
111#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
112pub enum Message {
113 User(UserMessage),
114 Agent(AgentMessage),
115 Resume,
116}
117
118impl Message {
119 pub fn as_agent_message(&self) -> Option<&AgentMessage> {
120 match self {
121 Message::Agent(agent_message) => Some(agent_message),
122 _ => None,
123 }
124 }
125
126 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
127 match self {
128 Message::User(message) => {
129 if message.content.is_empty() {
130 vec![]
131 } else {
132 vec![message.to_request()]
133 }
134 }
135 Message::Agent(message) => message.to_request(),
136 Message::Resume => vec![LanguageModelRequestMessage {
137 role: Role::User,
138 content: vec!["Continue where you left off".into()],
139 cache: false,
140 reasoning_details: None,
141 }],
142 }
143 }
144
145 pub fn to_markdown(&self) -> String {
146 match self {
147 Message::User(message) => message.to_markdown(),
148 Message::Agent(message) => message.to_markdown(),
149 Message::Resume => "[resume]\n".into(),
150 }
151 }
152
153 pub fn role(&self) -> Role {
154 match self {
155 Message::User(_) | Message::Resume => Role::User,
156 Message::Agent(_) => Role::Assistant,
157 }
158 }
159}
160
161#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
162pub struct UserMessage {
163 pub id: UserMessageId,
164 pub content: Vec<UserMessageContent>,
165}
166
167#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
168pub enum UserMessageContent {
169 Text(String),
170 Mention { uri: MentionUri, content: String },
171 Image(LanguageModelImage),
172}
173
174impl UserMessage {
175 pub fn to_markdown(&self) -> String {
176 let mut markdown = String::new();
177
178 for content in &self.content {
179 match content {
180 UserMessageContent::Text(text) => {
181 markdown.push_str(text);
182 markdown.push('\n');
183 }
184 UserMessageContent::Image(_) => {
185 markdown.push_str("<image />\n");
186 }
187 UserMessageContent::Mention { uri, content } => {
188 if !content.is_empty() {
189 let _ = writeln!(&mut markdown, "{}\n\n{}", uri.as_link(), content);
190 } else {
191 let _ = writeln!(&mut markdown, "{}", uri.as_link());
192 }
193 }
194 }
195 }
196
197 markdown
198 }
199
200 fn to_request(&self) -> LanguageModelRequestMessage {
201 let mut message = LanguageModelRequestMessage {
202 role: Role::User,
203 content: Vec::with_capacity(self.content.len()),
204 cache: false,
205 reasoning_details: None,
206 };
207
208 const OPEN_CONTEXT: &str = "<context>\n\
209 The following items were attached by the user. \
210 They are up-to-date and don't need to be re-read.\n\n";
211
212 const OPEN_FILES_TAG: &str = "<files>";
213 const OPEN_DIRECTORIES_TAG: &str = "<directories>";
214 const OPEN_SYMBOLS_TAG: &str = "<symbols>";
215 const OPEN_SELECTIONS_TAG: &str = "<selections>";
216 const OPEN_THREADS_TAG: &str = "<threads>";
217 const OPEN_FETCH_TAG: &str = "<fetched_urls>";
218 const OPEN_RULES_TAG: &str =
219 "<rules>\nThe user has specified the following rules that should be applied:\n";
220 const OPEN_DIAGNOSTICS_TAG: &str = "<diagnostics>";
221 const OPEN_DIFFS_TAG: &str = "<diffs>";
222
223 let mut file_context = OPEN_FILES_TAG.to_string();
224 let mut directory_context = OPEN_DIRECTORIES_TAG.to_string();
225 let mut symbol_context = OPEN_SYMBOLS_TAG.to_string();
226 let mut selection_context = OPEN_SELECTIONS_TAG.to_string();
227 let mut thread_context = OPEN_THREADS_TAG.to_string();
228 let mut fetch_context = OPEN_FETCH_TAG.to_string();
229 let mut rules_context = OPEN_RULES_TAG.to_string();
230 let mut diagnostics_context = OPEN_DIAGNOSTICS_TAG.to_string();
231 let mut diffs_context = OPEN_DIFFS_TAG.to_string();
232
233 for chunk in &self.content {
234 let chunk = match chunk {
235 UserMessageContent::Text(text) => {
236 language_model::MessageContent::Text(text.clone())
237 }
238 UserMessageContent::Image(value) => {
239 language_model::MessageContent::Image(value.clone())
240 }
241 UserMessageContent::Mention { uri, content } => {
242 match uri {
243 MentionUri::File { abs_path } => {
244 write!(
245 &mut file_context,
246 "\n{}",
247 MarkdownCodeBlock {
248 tag: &codeblock_tag(abs_path, None),
249 text: &content.to_string(),
250 }
251 )
252 .ok();
253 }
254 MentionUri::PastedImage => {
255 debug_panic!("pasted image URI should not be used in mention content")
256 }
257 MentionUri::Directory { .. } => {
258 write!(&mut directory_context, "\n{}\n", content).ok();
259 }
260 MentionUri::Symbol {
261 abs_path: path,
262 line_range,
263 ..
264 } => {
265 write!(
266 &mut symbol_context,
267 "\n{}",
268 MarkdownCodeBlock {
269 tag: &codeblock_tag(path, Some(line_range)),
270 text: content
271 }
272 )
273 .ok();
274 }
275 MentionUri::Selection {
276 abs_path: path,
277 line_range,
278 ..
279 } => {
280 write!(
281 &mut selection_context,
282 "\n{}",
283 MarkdownCodeBlock {
284 tag: &codeblock_tag(
285 path.as_deref().unwrap_or("Untitled".as_ref()),
286 Some(line_range)
287 ),
288 text: content
289 }
290 )
291 .ok();
292 }
293 MentionUri::Thread { .. } => {
294 write!(&mut thread_context, "\n{}\n", content).ok();
295 }
296 MentionUri::TextThread { .. } => {
297 write!(&mut thread_context, "\n{}\n", content).ok();
298 }
299 MentionUri::Rule { .. } => {
300 write!(
301 &mut rules_context,
302 "\n{}",
303 MarkdownCodeBlock {
304 tag: "",
305 text: content
306 }
307 )
308 .ok();
309 }
310 MentionUri::Fetch { url } => {
311 write!(&mut fetch_context, "\nFetch: {}\n\n{}", url, content).ok();
312 }
313 MentionUri::Diagnostics { .. } => {
314 write!(&mut diagnostics_context, "\n{}\n", content).ok();
315 }
316 MentionUri::TerminalSelection { .. } => {
317 write!(
318 &mut selection_context,
319 "\n{}",
320 MarkdownCodeBlock {
321 tag: "console",
322 text: content
323 }
324 )
325 .ok();
326 }
327 MentionUri::GitDiff { base_ref } => {
328 write!(
329 &mut diffs_context,
330 "\nBranch diff against {}:\n{}",
331 base_ref,
332 MarkdownCodeBlock {
333 tag: "diff",
334 text: content
335 }
336 )
337 .ok();
338 }
339 }
340
341 language_model::MessageContent::Text(uri.as_link().to_string())
342 }
343 };
344
345 message.content.push(chunk);
346 }
347
348 let len_before_context = message.content.len();
349
350 if file_context.len() > OPEN_FILES_TAG.len() {
351 file_context.push_str("</files>\n");
352 message
353 .content
354 .push(language_model::MessageContent::Text(file_context));
355 }
356
357 if directory_context.len() > OPEN_DIRECTORIES_TAG.len() {
358 directory_context.push_str("</directories>\n");
359 message
360 .content
361 .push(language_model::MessageContent::Text(directory_context));
362 }
363
364 if symbol_context.len() > OPEN_SYMBOLS_TAG.len() {
365 symbol_context.push_str("</symbols>\n");
366 message
367 .content
368 .push(language_model::MessageContent::Text(symbol_context));
369 }
370
371 if selection_context.len() > OPEN_SELECTIONS_TAG.len() {
372 selection_context.push_str("</selections>\n");
373 message
374 .content
375 .push(language_model::MessageContent::Text(selection_context));
376 }
377
378 if diffs_context.len() > OPEN_DIFFS_TAG.len() {
379 diffs_context.push_str("</diffs>\n");
380 message
381 .content
382 .push(language_model::MessageContent::Text(diffs_context));
383 }
384
385 if thread_context.len() > OPEN_THREADS_TAG.len() {
386 thread_context.push_str("</threads>\n");
387 message
388 .content
389 .push(language_model::MessageContent::Text(thread_context));
390 }
391
392 if fetch_context.len() > OPEN_FETCH_TAG.len() {
393 fetch_context.push_str("</fetched_urls>\n");
394 message
395 .content
396 .push(language_model::MessageContent::Text(fetch_context));
397 }
398
399 if rules_context.len() > OPEN_RULES_TAG.len() {
400 rules_context.push_str("</user_rules>\n");
401 message
402 .content
403 .push(language_model::MessageContent::Text(rules_context));
404 }
405
406 if diagnostics_context.len() > OPEN_DIAGNOSTICS_TAG.len() {
407 diagnostics_context.push_str("</diagnostics>\n");
408 message
409 .content
410 .push(language_model::MessageContent::Text(diagnostics_context));
411 }
412
413 if message.content.len() > len_before_context {
414 message.content.insert(
415 len_before_context,
416 language_model::MessageContent::Text(OPEN_CONTEXT.into()),
417 );
418 message
419 .content
420 .push(language_model::MessageContent::Text("</context>".into()));
421 }
422
423 message
424 }
425}
426
427fn codeblock_tag(full_path: &Path, line_range: Option<&RangeInclusive<u32>>) -> String {
428 let mut result = String::new();
429
430 if let Some(extension) = full_path.extension().and_then(|ext| ext.to_str()) {
431 let _ = write!(result, "{} ", extension);
432 }
433
434 let _ = write!(result, "{}", full_path.display());
435
436 if let Some(range) = line_range {
437 if range.start() == range.end() {
438 let _ = write!(result, ":{}", range.start() + 1);
439 } else {
440 let _ = write!(result, ":{}-{}", range.start() + 1, range.end() + 1);
441 }
442 }
443
444 result
445}
446
447impl AgentMessage {
448 pub fn to_markdown(&self) -> String {
449 let mut markdown = String::new();
450
451 for content in &self.content {
452 match content {
453 AgentMessageContent::Text(text) => {
454 markdown.push_str(text);
455 markdown.push('\n');
456 }
457 AgentMessageContent::Thinking { text, .. } => {
458 markdown.push_str("<think>");
459 markdown.push_str(text);
460 markdown.push_str("</think>\n");
461 }
462 AgentMessageContent::RedactedThinking(_) => {
463 markdown.push_str("<redacted_thinking />\n")
464 }
465 AgentMessageContent::ToolUse(tool_use) => {
466 markdown.push_str(&format!(
467 "**Tool Use**: {} (ID: {})\n",
468 tool_use.name, tool_use.id
469 ));
470 markdown.push_str(&format!(
471 "{}\n",
472 MarkdownCodeBlock {
473 tag: "json",
474 text: &format!("{:#}", tool_use.input)
475 }
476 ));
477 }
478 }
479 }
480
481 for tool_result in self.tool_results.values() {
482 markdown.push_str(&format!(
483 "**Tool Result**: {} (ID: {})\n\n",
484 tool_result.tool_name, tool_result.tool_use_id
485 ));
486 if tool_result.is_error {
487 markdown.push_str("**ERROR:**\n");
488 }
489
490 match &tool_result.content {
491 LanguageModelToolResultContent::Text(text) => {
492 writeln!(markdown, "{text}\n").ok();
493 }
494 LanguageModelToolResultContent::Image(_) => {
495 writeln!(markdown, "<image />\n").ok();
496 }
497 }
498
499 if let Some(output) = tool_result.output.as_ref() {
500 writeln!(
501 markdown,
502 "**Debug Output**:\n\n```json\n{}\n```\n",
503 serde_json::to_string_pretty(output).unwrap()
504 )
505 .unwrap();
506 }
507 }
508
509 markdown
510 }
511
512 pub fn to_request(&self) -> Vec<LanguageModelRequestMessage> {
513 let mut assistant_message = LanguageModelRequestMessage {
514 role: Role::Assistant,
515 content: Vec::with_capacity(self.content.len()),
516 cache: false,
517 reasoning_details: self.reasoning_details.clone(),
518 };
519 for chunk in &self.content {
520 match chunk {
521 AgentMessageContent::Text(text) => {
522 assistant_message
523 .content
524 .push(language_model::MessageContent::Text(text.clone()));
525 }
526 AgentMessageContent::Thinking { text, signature } => {
527 assistant_message
528 .content
529 .push(language_model::MessageContent::Thinking {
530 text: text.clone(),
531 signature: signature.clone(),
532 });
533 }
534 AgentMessageContent::RedactedThinking(value) => {
535 assistant_message.content.push(
536 language_model::MessageContent::RedactedThinking(value.clone()),
537 );
538 }
539 AgentMessageContent::ToolUse(tool_use) => {
540 if self.tool_results.contains_key(&tool_use.id) {
541 assistant_message
542 .content
543 .push(language_model::MessageContent::ToolUse(tool_use.clone()));
544 }
545 }
546 };
547 }
548
549 let mut user_message = LanguageModelRequestMessage {
550 role: Role::User,
551 content: Vec::new(),
552 cache: false,
553 reasoning_details: None,
554 };
555
556 for tool_result in self.tool_results.values() {
557 let mut tool_result = tool_result.clone();
558 // Surprisingly, the API fails if we return an empty string here.
559 // It thinks we are sending a tool use without a tool result.
560 if tool_result.content.is_empty() {
561 tool_result.content = "<Tool returned an empty string>".into();
562 }
563 user_message
564 .content
565 .push(language_model::MessageContent::ToolResult(tool_result));
566 }
567
568 let mut messages = Vec::new();
569 if !assistant_message.content.is_empty() {
570 messages.push(assistant_message);
571 }
572 if !user_message.content.is_empty() {
573 messages.push(user_message);
574 }
575 messages
576 }
577}
578
579#[derive(Default, Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
580pub struct AgentMessage {
581 pub content: Vec<AgentMessageContent>,
582 pub tool_results: IndexMap<LanguageModelToolUseId, LanguageModelToolResult>,
583 pub reasoning_details: Option<serde_json::Value>,
584}
585
586#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
587pub enum AgentMessageContent {
588 Text(String),
589 Thinking {
590 text: String,
591 signature: Option<String>,
592 },
593 RedactedThinking(String),
594 ToolUse(LanguageModelToolUse),
595}
596
597pub trait TerminalHandle {
598 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId>;
599 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse>;
600 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>>;
601 fn kill(&self, cx: &AsyncApp) -> Result<()>;
602 fn was_stopped_by_user(&self, cx: &AsyncApp) -> Result<bool>;
603}
604
605pub trait SubagentHandle {
606 /// The session ID of this subagent thread
607 fn id(&self) -> acp::SessionId;
608 /// The current number of entries in the thread.
609 /// Useful for knowing where the next turn will begin
610 fn num_entries(&self, cx: &App) -> usize;
611 /// Runs a turn for a given message and returns both the response and the index of that output message.
612 fn send(&self, message: String, cx: &AsyncApp) -> Task<Result<String>>;
613}
614
615pub trait ThreadEnvironment {
616 fn create_terminal(
617 &self,
618 command: String,
619 cwd: Option<PathBuf>,
620 output_byte_limit: Option<u64>,
621 cx: &mut AsyncApp,
622 ) -> Task<Result<Rc<dyn TerminalHandle>>>;
623
624 fn create_subagent(&self, label: String, cx: &mut App) -> Result<Rc<dyn SubagentHandle>>;
625
626 fn resume_subagent(
627 &self,
628 _session_id: acp::SessionId,
629 _cx: &mut App,
630 ) -> Result<Rc<dyn SubagentHandle>> {
631 Err(anyhow::anyhow!(
632 "Resuming subagent sessions is not supported"
633 ))
634 }
635}
636
637#[derive(Debug)]
638pub enum ThreadEvent {
639 UserMessage(UserMessage),
640 AgentText(String),
641 AgentThinking(String),
642 ToolCall(acp::ToolCall),
643 ToolCallUpdate(acp_thread::ToolCallUpdate),
644 ToolCallAuthorization(ToolCallAuthorization),
645 SubagentSpawned(acp::SessionId),
646 Retry(acp_thread::RetryStatus),
647 Stop(acp::StopReason),
648}
649
650#[derive(Debug)]
651pub struct NewTerminal {
652 pub command: String,
653 pub output_byte_limit: Option<u64>,
654 pub cwd: Option<PathBuf>,
655 pub response: oneshot::Sender<Result<Entity<acp_thread::Terminal>>>,
656}
657
658#[derive(Debug, Clone)]
659pub struct ToolPermissionContext {
660 pub tool_name: String,
661 pub input_values: Vec<String>,
662 pub scope: ToolPermissionScope,
663}
664
665#[derive(Debug, Clone, Copy, PartialEq, Eq)]
666pub enum ToolPermissionScope {
667 ToolInput,
668 SymlinkTarget,
669}
670
671impl ToolPermissionContext {
672 pub fn new(tool_name: impl Into<String>, input_values: Vec<String>) -> Self {
673 Self {
674 tool_name: tool_name.into(),
675 input_values,
676 scope: ToolPermissionScope::ToolInput,
677 }
678 }
679
680 pub fn symlink_target(tool_name: impl Into<String>, target_paths: Vec<String>) -> Self {
681 Self {
682 tool_name: tool_name.into(),
683 input_values: target_paths,
684 scope: ToolPermissionScope::SymlinkTarget,
685 }
686 }
687
688 /// Builds the permission options for this tool context.
689 ///
690 /// This is the canonical source for permission option generation.
691 /// Tests should use this function rather than manually constructing options.
692 ///
693 /// # Shell Compatibility for Terminal Tool
694 ///
695 /// For the terminal tool, "Always allow" options are only shown when the user's
696 /// shell supports POSIX-like command chaining syntax (`&&`, `||`, `;`, `|`).
697 ///
698 /// **Why this matters:** When a user sets up an "always allow" pattern like `^cargo`,
699 /// we need to parse the command to extract all sub-commands and verify that EVERY
700 /// sub-command matches the pattern. Otherwise, an attacker could craft a command like
701 /// `cargo build && rm -rf /` that would bypass the security check.
702 ///
703 /// **Supported shells:** Posix (sh, bash, dash, zsh), Fish 3.0+, PowerShell 7+/Pwsh,
704 /// Cmd, Xonsh, Csh, Tcsh
705 ///
706 /// **Unsupported shells:** Nushell (uses `and`/`or` keywords), Elvish (uses `and`/`or`
707 /// keywords), Rc (Plan 9 shell - no `&&`/`||` operators)
708 ///
709 /// For unsupported shells, we hide the "Always allow" UI options entirely, and if
710 /// the user has `always_allow` rules configured in settings, `ToolPermissionDecision::from_input`
711 /// will return a `Deny` with an explanatory error message.
712 pub fn build_permission_options(&self) -> acp_thread::PermissionOptions {
713 use crate::pattern_extraction::*;
714 use util::shell::ShellKind;
715
716 let tool_name = &self.tool_name;
717 let input_values = &self.input_values;
718 if self.scope == ToolPermissionScope::SymlinkTarget {
719 return acp_thread::PermissionOptions::Flat(vec![
720 acp::PermissionOption::new(
721 acp::PermissionOptionId::new("allow"),
722 "Yes",
723 acp::PermissionOptionKind::AllowOnce,
724 ),
725 acp::PermissionOption::new(
726 acp::PermissionOptionId::new("deny"),
727 "No",
728 acp::PermissionOptionKind::RejectOnce,
729 ),
730 ]);
731 }
732
733 // Check if the user's shell supports POSIX-like command chaining.
734 // See the doc comment above for the full explanation of why this is needed.
735 let shell_supports_always_allow = if tool_name == TerminalTool::NAME {
736 ShellKind::system().supports_posix_chaining()
737 } else {
738 true
739 };
740
741 let extract_for_value = |value: &str| -> (Option<String>, Option<String>) {
742 if tool_name == TerminalTool::NAME {
743 (
744 extract_terminal_pattern(value),
745 extract_terminal_pattern_display(value),
746 )
747 } else if tool_name == CopyPathTool::NAME
748 || tool_name == MovePathTool::NAME
749 || tool_name == EditFileTool::NAME
750 || tool_name == DeletePathTool::NAME
751 || tool_name == CreateDirectoryTool::NAME
752 || tool_name == SaveFileTool::NAME
753 {
754 (
755 extract_path_pattern(value),
756 extract_path_pattern_display(value),
757 )
758 } else if tool_name == FetchTool::NAME {
759 (
760 extract_url_pattern(value),
761 extract_url_pattern_display(value),
762 )
763 } else {
764 (None, None)
765 }
766 };
767
768 // Extract patterns from all input values. Only offer a pattern-specific
769 // "always allow/deny" button when every value produces the same pattern.
770 let (pattern, pattern_display) = match input_values.as_slice() {
771 [single] => extract_for_value(single),
772 _ => {
773 let mut iter = input_values.iter().map(|v| extract_for_value(v));
774 match iter.next() {
775 Some(first) => {
776 if iter.all(|pair| pair.0 == first.0) {
777 first
778 } else {
779 (None, None)
780 }
781 }
782 None => (None, None),
783 }
784 }
785 };
786
787 let mut choices = Vec::new();
788
789 let mut push_choice = |label: String, allow_id, deny_id, allow_kind, deny_kind| {
790 choices.push(acp_thread::PermissionOptionChoice {
791 allow: acp::PermissionOption::new(
792 acp::PermissionOptionId::new(allow_id),
793 label.clone(),
794 allow_kind,
795 ),
796 deny: acp::PermissionOption::new(
797 acp::PermissionOptionId::new(deny_id),
798 label,
799 deny_kind,
800 ),
801 });
802 };
803
804 if shell_supports_always_allow {
805 push_choice(
806 format!("Always for {}", tool_name.replace('_', " ")),
807 format!("always_allow:{}", tool_name),
808 format!("always_deny:{}", tool_name),
809 acp::PermissionOptionKind::AllowAlways,
810 acp::PermissionOptionKind::RejectAlways,
811 );
812
813 if let (Some(pattern), Some(display)) = (pattern, pattern_display) {
814 let button_text = if tool_name == TerminalTool::NAME {
815 format!("Always for `{}` commands", display)
816 } else {
817 format!("Always for `{}`", display)
818 };
819 push_choice(
820 button_text,
821 format!("always_allow_pattern:{}\n{}", tool_name, pattern),
822 format!("always_deny_pattern:{}\n{}", tool_name, pattern),
823 acp::PermissionOptionKind::AllowAlways,
824 acp::PermissionOptionKind::RejectAlways,
825 );
826 }
827 }
828
829 push_choice(
830 "Only this time".to_string(),
831 "allow".to_string(),
832 "deny".to_string(),
833 acp::PermissionOptionKind::AllowOnce,
834 acp::PermissionOptionKind::RejectOnce,
835 );
836
837 acp_thread::PermissionOptions::Dropdown(choices)
838 }
839}
840
841#[derive(Debug)]
842pub struct ToolCallAuthorization {
843 pub tool_call: acp::ToolCallUpdate,
844 pub options: acp_thread::PermissionOptions,
845 pub response: oneshot::Sender<acp::PermissionOptionId>,
846 pub context: Option<ToolPermissionContext>,
847}
848
849#[derive(Debug, thiserror::Error)]
850enum CompletionError {
851 #[error("max tokens")]
852 MaxTokens,
853 #[error("refusal")]
854 Refusal,
855 #[error(transparent)]
856 Other(#[from] anyhow::Error),
857}
858
859pub struct Thread {
860 id: acp::SessionId,
861 prompt_id: PromptId,
862 updated_at: DateTime<Utc>,
863 title: Option<SharedString>,
864 pending_title_generation: Option<Task<()>>,
865 pending_summary_generation: Option<Shared<Task<Option<SharedString>>>>,
866 summary: Option<SharedString>,
867 messages: Vec<Message>,
868 user_store: Entity<UserStore>,
869 /// Holds the task that handles agent interaction until the end of the turn.
870 /// Survives across multiple requests as the model performs tool calls and
871 /// we run tools, report their results.
872 running_turn: Option<RunningTurn>,
873 /// Flag indicating the UI has a queued message waiting to be sent.
874 /// Used to signal that the turn should end at the next message boundary.
875 has_queued_message: bool,
876 pending_message: Option<AgentMessage>,
877 pub(crate) tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
878 request_token_usage: HashMap<UserMessageId, language_model::TokenUsage>,
879 #[allow(unused)]
880 cumulative_token_usage: TokenUsage,
881 #[allow(unused)]
882 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
883 pub(crate) context_server_registry: Entity<ContextServerRegistry>,
884 profile_id: AgentProfileId,
885 project_context: Entity<ProjectContext>,
886 pub(crate) templates: Arc<Templates>,
887 model: Option<Arc<dyn LanguageModel>>,
888 summarization_model: Option<Arc<dyn LanguageModel>>,
889 thinking_enabled: bool,
890 thinking_effort: Option<String>,
891 speed: Option<Speed>,
892 prompt_capabilities_tx: watch::Sender<acp::PromptCapabilities>,
893 pub(crate) prompt_capabilities_rx: watch::Receiver<acp::PromptCapabilities>,
894 pub(crate) project: Entity<Project>,
895 pub(crate) action_log: Entity<ActionLog>,
896 /// True if this thread was imported from a shared thread and can be synced.
897 imported: bool,
898 /// If this is a subagent thread, contains context about the parent
899 subagent_context: Option<SubagentContext>,
900 /// The user's unsent prompt text, persisted so it can be restored when reloading the thread.
901 draft_prompt: Option<Vec<acp::ContentBlock>>,
902 ui_scroll_position: Option<gpui::ListOffset>,
903 /// Weak references to running subagent threads for cancellation propagation
904 running_subagents: Vec<WeakEntity<Thread>>,
905}
906
907impl Thread {
908 fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {
909 let image = model.map_or(true, |model| model.supports_images());
910 acp::PromptCapabilities::new()
911 .image(image)
912 .embedded_context(true)
913 }
914
915 pub fn new_subagent(parent_thread: &Entity<Thread>, cx: &mut Context<Self>) -> Self {
916 let project = parent_thread.read(cx).project.clone();
917 let project_context = parent_thread.read(cx).project_context.clone();
918 let context_server_registry = parent_thread.read(cx).context_server_registry.clone();
919 let templates = parent_thread.read(cx).templates.clone();
920 let model = parent_thread.read(cx).model().cloned();
921 let parent_action_log = parent_thread.read(cx).action_log().clone();
922 let action_log =
923 cx.new(|_cx| ActionLog::new(project.clone()).with_linked_action_log(parent_action_log));
924 let mut thread = Self::new_internal(
925 project,
926 project_context,
927 context_server_registry,
928 templates,
929 model,
930 action_log,
931 cx,
932 );
933 thread.subagent_context = Some(SubagentContext {
934 parent_thread_id: parent_thread.read(cx).id().clone(),
935 depth: parent_thread.read(cx).depth() + 1,
936 });
937 thread
938 }
939
940 pub fn new(
941 project: Entity<Project>,
942 project_context: Entity<ProjectContext>,
943 context_server_registry: Entity<ContextServerRegistry>,
944 templates: Arc<Templates>,
945 model: Option<Arc<dyn LanguageModel>>,
946 cx: &mut Context<Self>,
947 ) -> Self {
948 Self::new_internal(
949 project.clone(),
950 project_context,
951 context_server_registry,
952 templates,
953 model,
954 cx.new(|_cx| ActionLog::new(project)),
955 cx,
956 )
957 }
958
959 fn new_internal(
960 project: Entity<Project>,
961 project_context: Entity<ProjectContext>,
962 context_server_registry: Entity<ContextServerRegistry>,
963 templates: Arc<Templates>,
964 model: Option<Arc<dyn LanguageModel>>,
965 action_log: Entity<ActionLog>,
966 cx: &mut Context<Self>,
967 ) -> Self {
968 let settings = AgentSettings::get_global(cx);
969 let profile_id = settings.default_profile.clone();
970 let enable_thinking = settings
971 .default_model
972 .as_ref()
973 .is_some_and(|model| model.enable_thinking);
974 let thinking_effort = settings
975 .default_model
976 .as_ref()
977 .and_then(|model| model.effort.clone());
978 let (prompt_capabilities_tx, prompt_capabilities_rx) =
979 watch::channel(Self::prompt_capabilities(model.as_deref()));
980 Self {
981 id: acp::SessionId::new(uuid::Uuid::new_v4().to_string()),
982 prompt_id: PromptId::new(),
983 updated_at: Utc::now(),
984 title: None,
985 pending_title_generation: None,
986 pending_summary_generation: None,
987 summary: None,
988 messages: Vec::new(),
989 user_store: project.read(cx).user_store(),
990 running_turn: None,
991 has_queued_message: false,
992 pending_message: None,
993 tools: BTreeMap::default(),
994 request_token_usage: HashMap::default(),
995 cumulative_token_usage: TokenUsage::default(),
996 initial_project_snapshot: {
997 let project_snapshot = Self::project_snapshot(project.clone(), cx);
998 cx.foreground_executor()
999 .spawn(async move { Some(project_snapshot.await) })
1000 .shared()
1001 },
1002 context_server_registry,
1003 profile_id,
1004 project_context,
1005 templates,
1006 model,
1007 summarization_model: None,
1008 thinking_enabled: enable_thinking,
1009 speed: None,
1010 thinking_effort,
1011 prompt_capabilities_tx,
1012 prompt_capabilities_rx,
1013 project,
1014 action_log,
1015 imported: false,
1016 subagent_context: None,
1017 draft_prompt: None,
1018 ui_scroll_position: None,
1019 running_subagents: Vec::new(),
1020 }
1021 }
1022
1023 pub fn id(&self) -> &acp::SessionId {
1024 &self.id
1025 }
1026
1027 /// Returns true if this thread was imported from a shared thread.
1028 pub fn is_imported(&self) -> bool {
1029 self.imported
1030 }
1031
1032 pub fn replay(
1033 &mut self,
1034 cx: &mut Context<Self>,
1035 ) -> mpsc::UnboundedReceiver<Result<ThreadEvent>> {
1036 let (tx, rx) = mpsc::unbounded();
1037 let stream = ThreadEventStream(tx);
1038 for message in &self.messages {
1039 match message {
1040 Message::User(user_message) => stream.send_user_message(user_message),
1041 Message::Agent(assistant_message) => {
1042 for content in &assistant_message.content {
1043 match content {
1044 AgentMessageContent::Text(text) => stream.send_text(text),
1045 AgentMessageContent::Thinking { text, .. } => {
1046 stream.send_thinking(text)
1047 }
1048 AgentMessageContent::RedactedThinking(_) => {}
1049 AgentMessageContent::ToolUse(tool_use) => {
1050 self.replay_tool_call(
1051 tool_use,
1052 assistant_message.tool_results.get(&tool_use.id),
1053 &stream,
1054 cx,
1055 );
1056 }
1057 }
1058 }
1059 }
1060 Message::Resume => {}
1061 }
1062 }
1063 rx
1064 }
1065
1066 fn replay_tool_call(
1067 &self,
1068 tool_use: &LanguageModelToolUse,
1069 tool_result: Option<&LanguageModelToolResult>,
1070 stream: &ThreadEventStream,
1071 cx: &mut Context<Self>,
1072 ) {
1073 // Extract saved output and status first, so they're available even if tool is not found
1074 let output = tool_result
1075 .as_ref()
1076 .and_then(|result| result.output.clone());
1077 let status = tool_result
1078 .as_ref()
1079 .map_or(acp::ToolCallStatus::Failed, |result| {
1080 if result.is_error {
1081 acp::ToolCallStatus::Failed
1082 } else {
1083 acp::ToolCallStatus::Completed
1084 }
1085 });
1086
1087 let tool = self.tools.get(tool_use.name.as_ref()).cloned().or_else(|| {
1088 self.context_server_registry
1089 .read(cx)
1090 .servers()
1091 .find_map(|(_, tools)| {
1092 if let Some(tool) = tools.get(tool_use.name.as_ref()) {
1093 Some(tool.clone())
1094 } else {
1095 None
1096 }
1097 })
1098 });
1099
1100 let Some(tool) = tool else {
1101 // Tool not found (e.g., MCP server not connected after restart),
1102 // but still display the saved result if available.
1103 // We need to send both ToolCall and ToolCallUpdate events because the UI
1104 // only converts raw_output to displayable content in update_fields, not from_acp.
1105 stream
1106 .0
1107 .unbounded_send(Ok(ThreadEvent::ToolCall(
1108 acp::ToolCall::new(tool_use.id.to_string(), tool_use.name.to_string())
1109 .status(status)
1110 .raw_input(tool_use.input.clone()),
1111 )))
1112 .ok();
1113 stream.update_tool_call_fields(
1114 &tool_use.id,
1115 acp::ToolCallUpdateFields::new()
1116 .status(status)
1117 .raw_output(output),
1118 None,
1119 );
1120 return;
1121 };
1122
1123 let title = tool.initial_title(tool_use.input.clone(), cx);
1124 let kind = tool.kind();
1125 stream.send_tool_call(
1126 &tool_use.id,
1127 &tool_use.name,
1128 title,
1129 kind,
1130 tool_use.input.clone(),
1131 );
1132
1133 if let Some(output) = output.clone() {
1134 // For replay, we use a dummy cancellation receiver since the tool already completed
1135 let (_cancellation_tx, cancellation_rx) = watch::channel(false);
1136 let tool_event_stream = ToolCallEventStream::new(
1137 tool_use.id.clone(),
1138 stream.clone(),
1139 Some(self.project.read(cx).fs().clone()),
1140 cancellation_rx,
1141 );
1142 tool.replay(tool_use.input.clone(), output, tool_event_stream, cx)
1143 .log_err();
1144 }
1145
1146 stream.update_tool_call_fields(
1147 &tool_use.id,
1148 acp::ToolCallUpdateFields::new()
1149 .status(status)
1150 .raw_output(output),
1151 None,
1152 );
1153 }
1154
1155 pub fn from_db(
1156 id: acp::SessionId,
1157 db_thread: DbThread,
1158 project: Entity<Project>,
1159 project_context: Entity<ProjectContext>,
1160 context_server_registry: Entity<ContextServerRegistry>,
1161 templates: Arc<Templates>,
1162 cx: &mut Context<Self>,
1163 ) -> Self {
1164 let settings = AgentSettings::get_global(cx);
1165 let profile_id = db_thread
1166 .profile
1167 .unwrap_or_else(|| settings.default_profile.clone());
1168
1169 let mut model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1170 db_thread
1171 .model
1172 .and_then(|model| {
1173 let model = SelectedModel {
1174 provider: model.provider.clone().into(),
1175 model: model.model.into(),
1176 };
1177 registry.select_model(&model, cx)
1178 })
1179 .or_else(|| registry.default_model())
1180 .map(|model| model.model)
1181 });
1182
1183 if model.is_none() {
1184 model = Self::resolve_profile_model(&profile_id, cx);
1185 }
1186 if model.is_none() {
1187 model = LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {
1188 registry.default_model().map(|model| model.model)
1189 });
1190 }
1191
1192 let (prompt_capabilities_tx, prompt_capabilities_rx) =
1193 watch::channel(Self::prompt_capabilities(model.as_deref()));
1194
1195 let action_log = cx.new(|_| ActionLog::new(project.clone()));
1196
1197 Self {
1198 id,
1199 prompt_id: PromptId::new(),
1200 title: if db_thread.title.is_empty() {
1201 None
1202 } else {
1203 Some(db_thread.title.clone())
1204 },
1205 pending_title_generation: None,
1206 pending_summary_generation: None,
1207 summary: db_thread.detailed_summary,
1208 messages: db_thread.messages,
1209 user_store: project.read(cx).user_store(),
1210 running_turn: None,
1211 has_queued_message: false,
1212 pending_message: None,
1213 tools: BTreeMap::default(),
1214 request_token_usage: db_thread.request_token_usage.clone(),
1215 cumulative_token_usage: db_thread.cumulative_token_usage,
1216 initial_project_snapshot: Task::ready(db_thread.initial_project_snapshot).shared(),
1217 context_server_registry,
1218 profile_id,
1219 project_context,
1220 templates,
1221 model,
1222 summarization_model: None,
1223 thinking_enabled: db_thread.thinking_enabled,
1224 thinking_effort: db_thread.thinking_effort,
1225 speed: db_thread.speed,
1226 project,
1227 action_log,
1228 updated_at: db_thread.updated_at,
1229 prompt_capabilities_tx,
1230 prompt_capabilities_rx,
1231 imported: db_thread.imported,
1232 subagent_context: db_thread.subagent_context,
1233 draft_prompt: db_thread.draft_prompt,
1234 ui_scroll_position: db_thread.ui_scroll_position.map(|sp| gpui::ListOffset {
1235 item_ix: sp.item_ix,
1236 offset_in_item: gpui::px(sp.offset_in_item),
1237 }),
1238 running_subagents: Vec::new(),
1239 }
1240 }
1241
1242 pub fn to_db(&self, cx: &App) -> Task<DbThread> {
1243 let initial_project_snapshot = self.initial_project_snapshot.clone();
1244 let mut thread = DbThread {
1245 title: self.title(),
1246 messages: self.messages.clone(),
1247 updated_at: self.updated_at,
1248 detailed_summary: self.summary.clone(),
1249 initial_project_snapshot: None,
1250 cumulative_token_usage: self.cumulative_token_usage,
1251 request_token_usage: self.request_token_usage.clone(),
1252 model: self.model.as_ref().map(|model| DbLanguageModel {
1253 provider: model.provider_id().to_string(),
1254 model: model.id().0.to_string(),
1255 }),
1256 profile: Some(self.profile_id.clone()),
1257 imported: self.imported,
1258 subagent_context: self.subagent_context.clone(),
1259 speed: self.speed,
1260 thinking_enabled: self.thinking_enabled,
1261 thinking_effort: self.thinking_effort.clone(),
1262 draft_prompt: self.draft_prompt.clone(),
1263 ui_scroll_position: self.ui_scroll_position.map(|lo| {
1264 crate::db::SerializedScrollPosition {
1265 item_ix: lo.item_ix,
1266 offset_in_item: lo.offset_in_item.as_f32(),
1267 }
1268 }),
1269 };
1270
1271 cx.background_spawn(async move {
1272 let initial_project_snapshot = initial_project_snapshot.await;
1273 thread.initial_project_snapshot = initial_project_snapshot;
1274 thread
1275 })
1276 }
1277
1278 /// Create a snapshot of the current project state including git information and unsaved buffers.
1279 fn project_snapshot(
1280 project: Entity<Project>,
1281 cx: &mut Context<Self>,
1282 ) -> Task<Arc<ProjectSnapshot>> {
1283 let task = project::telemetry_snapshot::TelemetrySnapshot::new(&project, cx);
1284 cx.spawn(async move |_, _| {
1285 let snapshot = task.await;
1286
1287 Arc::new(ProjectSnapshot {
1288 worktree_snapshots: snapshot.worktree_snapshots,
1289 timestamp: Utc::now(),
1290 })
1291 })
1292 }
1293
1294 pub fn project_context(&self) -> &Entity<ProjectContext> {
1295 &self.project_context
1296 }
1297
1298 pub fn project(&self) -> &Entity<Project> {
1299 &self.project
1300 }
1301
1302 pub fn action_log(&self) -> &Entity<ActionLog> {
1303 &self.action_log
1304 }
1305
1306 pub fn is_empty(&self) -> bool {
1307 self.messages.is_empty() && self.title.is_none()
1308 }
1309
1310 pub fn draft_prompt(&self) -> Option<&[acp::ContentBlock]> {
1311 self.draft_prompt.as_deref()
1312 }
1313
1314 pub fn set_draft_prompt(&mut self, prompt: Option<Vec<acp::ContentBlock>>) {
1315 self.draft_prompt = prompt;
1316 }
1317
1318 pub fn ui_scroll_position(&self) -> Option<gpui::ListOffset> {
1319 self.ui_scroll_position
1320 }
1321
1322 pub fn set_ui_scroll_position(&mut self, position: Option<gpui::ListOffset>) {
1323 self.ui_scroll_position = position;
1324 }
1325
1326 pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {
1327 self.model.as_ref()
1328 }
1329
1330 pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {
1331 let old_usage = self.latest_token_usage();
1332 self.model = Some(model.clone());
1333 let new_caps = Self::prompt_capabilities(self.model.as_deref());
1334 let new_usage = self.latest_token_usage();
1335 if old_usage != new_usage {
1336 cx.emit(TokenUsageUpdated(new_usage));
1337 }
1338 self.prompt_capabilities_tx.send(new_caps).log_err();
1339
1340 for subagent in &self.running_subagents {
1341 subagent
1342 .update(cx, |thread, cx| thread.set_model(model.clone(), cx))
1343 .ok();
1344 }
1345
1346 cx.notify()
1347 }
1348
1349 pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {
1350 self.summarization_model.as_ref()
1351 }
1352
1353 pub fn set_summarization_model(
1354 &mut self,
1355 model: Option<Arc<dyn LanguageModel>>,
1356 cx: &mut Context<Self>,
1357 ) {
1358 self.summarization_model = model.clone();
1359
1360 for subagent in &self.running_subagents {
1361 subagent
1362 .update(cx, |thread, cx| {
1363 thread.set_summarization_model(model.clone(), cx)
1364 })
1365 .ok();
1366 }
1367 cx.notify()
1368 }
1369
1370 pub fn thinking_enabled(&self) -> bool {
1371 self.thinking_enabled
1372 }
1373
1374 pub fn set_thinking_enabled(&mut self, enabled: bool, cx: &mut Context<Self>) {
1375 self.thinking_enabled = enabled;
1376
1377 for subagent in &self.running_subagents {
1378 subagent
1379 .update(cx, |thread, cx| thread.set_thinking_enabled(enabled, cx))
1380 .ok();
1381 }
1382 cx.notify();
1383 }
1384
1385 pub fn thinking_effort(&self) -> Option<&String> {
1386 self.thinking_effort.as_ref()
1387 }
1388
1389 pub fn set_thinking_effort(&mut self, effort: Option<String>, cx: &mut Context<Self>) {
1390 self.thinking_effort = effort.clone();
1391
1392 for subagent in &self.running_subagents {
1393 subagent
1394 .update(cx, |thread, cx| {
1395 thread.set_thinking_effort(effort.clone(), cx)
1396 })
1397 .ok();
1398 }
1399 cx.notify();
1400 }
1401
1402 pub fn speed(&self) -> Option<Speed> {
1403 self.speed
1404 }
1405
1406 pub fn set_speed(&mut self, speed: Speed, cx: &mut Context<Self>) {
1407 self.speed = Some(speed);
1408
1409 for subagent in &self.running_subagents {
1410 subagent
1411 .update(cx, |thread, cx| thread.set_speed(speed, cx))
1412 .ok();
1413 }
1414 cx.notify();
1415 }
1416
1417 pub fn last_message(&self) -> Option<&Message> {
1418 self.messages.last()
1419 }
1420
1421 #[cfg(any(test, feature = "test-support"))]
1422 pub fn last_received_or_pending_message(&self) -> Option<Message> {
1423 if let Some(message) = self.pending_message.clone() {
1424 Some(Message::Agent(message))
1425 } else {
1426 self.messages.last().cloned()
1427 }
1428 }
1429
1430 pub fn add_default_tools(
1431 &mut self,
1432 environment: Rc<dyn ThreadEnvironment>,
1433 cx: &mut Context<Self>,
1434 ) {
1435 // Only update the agent location for the root thread, not for subagents.
1436 let update_agent_location = self.parent_thread_id().is_none();
1437
1438 let language_registry = self.project.read(cx).languages().clone();
1439 self.add_tool(CopyPathTool::new(self.project.clone()));
1440 self.add_tool(CreateDirectoryTool::new(self.project.clone()));
1441 self.add_tool(DeletePathTool::new(
1442 self.project.clone(),
1443 self.action_log.clone(),
1444 ));
1445 self.add_tool(DiagnosticsTool::new(self.project.clone()));
1446 self.add_tool(EditFileTool::new(
1447 self.project.clone(),
1448 cx.weak_entity(),
1449 language_registry.clone(),
1450 Templates::new(),
1451 ));
1452 self.add_tool(StreamingEditFileTool::new(
1453 self.project.clone(),
1454 cx.weak_entity(),
1455 self.action_log.clone(),
1456 language_registry,
1457 ));
1458 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1459 self.add_tool(FindPathTool::new(self.project.clone()));
1460 self.add_tool(GrepTool::new(self.project.clone()));
1461 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1462 self.add_tool(MovePathTool::new(self.project.clone()));
1463 self.add_tool(NowTool);
1464 self.add_tool(OpenTool::new(self.project.clone()));
1465 self.add_tool(ReadFileTool::new(
1466 self.project.clone(),
1467 self.action_log.clone(),
1468 update_agent_location,
1469 ));
1470 self.add_tool(SaveFileTool::new(self.project.clone()));
1471 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1472 self.add_tool(TerminalTool::new(self.project.clone(), environment.clone()));
1473 self.add_tool(WebSearchTool);
1474
1475 if self.depth() < MAX_SUBAGENT_DEPTH {
1476 self.add_tool(SpawnAgentTool::new(environment));
1477 }
1478 }
1479
1480 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1481 debug_assert!(
1482 !self.tools.contains_key(T::NAME),
1483 "Duplicate tool name: {}",
1484 T::NAME,
1485 );
1486 self.tools.insert(T::NAME.into(), tool.erase());
1487 }
1488
1489 #[cfg(any(test, feature = "test-support"))]
1490 pub fn remove_tool(&mut self, name: &str) -> bool {
1491 self.tools.remove(name).is_some()
1492 }
1493
1494 pub fn profile(&self) -> &AgentProfileId {
1495 &self.profile_id
1496 }
1497
1498 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1499 if self.profile_id == profile_id {
1500 return;
1501 }
1502
1503 self.profile_id = profile_id.clone();
1504
1505 // Swap to the profile's preferred model when available.
1506 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1507 self.set_model(model, cx);
1508 }
1509
1510 for subagent in &self.running_subagents {
1511 subagent
1512 .update(cx, |thread, cx| thread.set_profile(profile_id.clone(), cx))
1513 .ok();
1514 }
1515 }
1516
1517 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1518 for subagent in self.running_subagents.drain(..) {
1519 if let Some(subagent) = subagent.upgrade() {
1520 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1521 }
1522 }
1523
1524 let Some(running_turn) = self.running_turn.take() else {
1525 self.flush_pending_message(cx);
1526 return Task::ready(());
1527 };
1528
1529 let turn_task = running_turn.cancel();
1530
1531 cx.spawn(async move |this, cx| {
1532 turn_task.await;
1533 this.update(cx, |this, cx| {
1534 this.flush_pending_message(cx);
1535 })
1536 .ok();
1537 })
1538 }
1539
1540 pub fn set_has_queued_message(&mut self, has_queued: bool) {
1541 self.has_queued_message = has_queued;
1542 }
1543
1544 pub fn has_queued_message(&self) -> bool {
1545 self.has_queued_message
1546 }
1547
1548 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1549 let Some(last_user_message) = self.last_user_message() else {
1550 return;
1551 };
1552
1553 self.request_token_usage
1554 .insert(last_user_message.id.clone(), update);
1555 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1556 cx.notify();
1557 }
1558
1559 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1560 self.cancel(cx).detach();
1561 // Clear pending message since cancel will try to flush it asynchronously,
1562 // and we don't want that content to be added after we truncate
1563 self.pending_message.take();
1564 let Some(position) = self.messages.iter().position(
1565 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1566 ) else {
1567 return Err(anyhow!("Message not found"));
1568 };
1569
1570 for message in self.messages.drain(position..) {
1571 match message {
1572 Message::User(message) => {
1573 self.request_token_usage.remove(&message.id);
1574 }
1575 Message::Agent(_) | Message::Resume => {}
1576 }
1577 }
1578 self.clear_summary();
1579 cx.notify();
1580 Ok(())
1581 }
1582
1583 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1584 let last_user_message = self.last_user_message()?;
1585 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1586 Some(*tokens)
1587 }
1588
1589 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1590 let usage = self.latest_request_token_usage()?;
1591 let model = self.model.clone()?;
1592 Some(acp_thread::TokenUsage {
1593 max_tokens: model.max_token_count(),
1594 max_output_tokens: model.max_output_tokens(),
1595 used_tokens: usage.total_tokens(),
1596 input_tokens: usage.input_tokens,
1597 output_tokens: usage.output_tokens,
1598 })
1599 }
1600
1601 /// Get the total input token count as of the message before the given message.
1602 ///
1603 /// Returns `None` if:
1604 /// - `target_id` is the first message (no previous message)
1605 /// - The previous message hasn't received a response yet (no usage data)
1606 /// - `target_id` is not found in the messages
1607 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1608 let mut previous_user_message_id: Option<&UserMessageId> = None;
1609
1610 for message in &self.messages {
1611 if let Message::User(user_msg) = message {
1612 if &user_msg.id == target_id {
1613 let prev_id = previous_user_message_id?;
1614 let usage = self.request_token_usage.get(prev_id)?;
1615 return Some(usage.input_tokens);
1616 }
1617 previous_user_message_id = Some(&user_msg.id);
1618 }
1619 }
1620 None
1621 }
1622
1623 /// Look up the active profile and resolve its preferred model if one is configured.
1624 fn resolve_profile_model(
1625 profile_id: &AgentProfileId,
1626 cx: &mut Context<Self>,
1627 ) -> Option<Arc<dyn LanguageModel>> {
1628 let selection = AgentSettings::get_global(cx)
1629 .profiles
1630 .get(profile_id)?
1631 .default_model
1632 .clone()?;
1633 Self::resolve_model_from_selection(&selection, cx)
1634 }
1635
1636 /// Translate a stored model selection into the configured model from the registry.
1637 fn resolve_model_from_selection(
1638 selection: &LanguageModelSelection,
1639 cx: &mut Context<Self>,
1640 ) -> Option<Arc<dyn LanguageModel>> {
1641 let selected = SelectedModel {
1642 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1643 model: LanguageModelId::from(selection.model.clone()),
1644 };
1645 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1646 registry
1647 .select_model(&selected, cx)
1648 .map(|configured| configured.model)
1649 })
1650 }
1651
1652 pub fn resume(
1653 &mut self,
1654 cx: &mut Context<Self>,
1655 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1656 self.messages.push(Message::Resume);
1657 cx.notify();
1658
1659 log::debug!("Total messages in thread: {}", self.messages.len());
1660 self.run_turn(cx)
1661 }
1662
1663 /// Sending a message results in the model streaming a response, which could include tool calls.
1664 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1665 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1666 pub fn send<T>(
1667 &mut self,
1668 id: UserMessageId,
1669 content: impl IntoIterator<Item = T>,
1670 cx: &mut Context<Self>,
1671 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1672 where
1673 T: Into<UserMessageContent>,
1674 {
1675 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1676 log::debug!("Thread::send content: {:?}", content);
1677
1678 self.messages
1679 .push(Message::User(UserMessage { id, content }));
1680 cx.notify();
1681
1682 self.send_existing(cx)
1683 }
1684
1685 pub fn send_existing(
1686 &mut self,
1687 cx: &mut Context<Self>,
1688 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1689 let model = self.model().context("No language model configured")?;
1690
1691 log::info!("Thread::send called with model: {}", model.name().0);
1692 self.advance_prompt_id();
1693
1694 log::debug!("Total messages in thread: {}", self.messages.len());
1695 self.run_turn(cx)
1696 }
1697
1698 pub fn push_acp_user_block(
1699 &mut self,
1700 id: UserMessageId,
1701 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1702 path_style: PathStyle,
1703 cx: &mut Context<Self>,
1704 ) {
1705 let content = blocks
1706 .into_iter()
1707 .map(|block| UserMessageContent::from_content_block(block, path_style))
1708 .collect::<Vec<_>>();
1709 self.messages
1710 .push(Message::User(UserMessage { id, content }));
1711 cx.notify();
1712 }
1713
1714 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1715 let text = match block {
1716 acp::ContentBlock::Text(text_content) => text_content.text,
1717 acp::ContentBlock::Image(_) => "[image]".to_string(),
1718 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1719 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1720 acp::ContentBlock::Resource(resource) => match resource.resource {
1721 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1722 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1723 _ => "[resource]".to_string(),
1724 },
1725 _ => "[unknown]".to_string(),
1726 };
1727
1728 self.messages.push(Message::Agent(AgentMessage {
1729 content: vec![AgentMessageContent::Text(text)],
1730 ..Default::default()
1731 }));
1732 cx.notify();
1733 }
1734
1735 #[cfg(feature = "eval")]
1736 pub fn proceed(
1737 &mut self,
1738 cx: &mut Context<Self>,
1739 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1740 self.run_turn(cx)
1741 }
1742
1743 fn run_turn(
1744 &mut self,
1745 cx: &mut Context<Self>,
1746 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1747 // Flush the old pending message synchronously before cancelling,
1748 // to avoid a race where the detached cancel task might flush the NEW
1749 // turn's pending message instead of the old one.
1750 self.flush_pending_message(cx);
1751 self.cancel(cx).detach();
1752
1753 let model = self.model.clone().context("No language model configured")?;
1754 let profile = AgentSettings::get_global(cx)
1755 .profiles
1756 .get(&self.profile_id)
1757 .context("Profile not found")?;
1758 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1759 let event_stream = ThreadEventStream(events_tx);
1760 let message_ix = self.messages.len().saturating_sub(1);
1761 self.clear_summary();
1762 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1763 self.running_turn = Some(RunningTurn {
1764 event_stream: event_stream.clone(),
1765 tools: self.enabled_tools(profile, &model, cx),
1766 cancellation_tx,
1767 streaming_tool_inputs: HashMap::default(),
1768 _task: cx.spawn(async move |this, cx| {
1769 log::debug!("Starting agent turn execution");
1770
1771 let turn_result = Self::run_turn_internal(
1772 &this,
1773 model,
1774 &event_stream,
1775 cancellation_rx.clone(),
1776 cx,
1777 )
1778 .await;
1779
1780 // Check if we were cancelled - if so, cancel() already took running_turn
1781 // and we shouldn't touch it (it might be a NEW turn now)
1782 let was_cancelled = *cancellation_rx.borrow();
1783 if was_cancelled {
1784 log::debug!("Turn was cancelled, skipping cleanup");
1785 return;
1786 }
1787
1788 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1789
1790 match turn_result {
1791 Ok(()) => {
1792 log::debug!("Turn execution completed");
1793 event_stream.send_stop(acp::StopReason::EndTurn);
1794 }
1795 Err(error) => {
1796 log::error!("Turn execution failed: {:?}", error);
1797 match error.downcast::<CompletionError>() {
1798 Ok(CompletionError::Refusal) => {
1799 event_stream.send_stop(acp::StopReason::Refusal);
1800 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1801 }
1802 Ok(CompletionError::MaxTokens) => {
1803 event_stream.send_stop(acp::StopReason::MaxTokens);
1804 }
1805 Ok(CompletionError::Other(error)) | Err(error) => {
1806 event_stream.send_error(error);
1807 }
1808 }
1809 }
1810 }
1811
1812 _ = this.update(cx, |this, _| this.running_turn.take());
1813 }),
1814 });
1815 Ok(events_rx)
1816 }
1817
1818 async fn run_turn_internal(
1819 this: &WeakEntity<Self>,
1820 model: Arc<dyn LanguageModel>,
1821 event_stream: &ThreadEventStream,
1822 mut cancellation_rx: watch::Receiver<bool>,
1823 cx: &mut AsyncApp,
1824 ) -> Result<()> {
1825 let mut attempt = 0;
1826 let mut intent = CompletionIntent::UserPrompt;
1827 loop {
1828 let request =
1829 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1830
1831 telemetry::event!(
1832 "Agent Thread Completion",
1833 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1834 parent_thread_id = this.read_with(cx, |this, _| this
1835 .parent_thread_id()
1836 .map(|id| id.to_string()))?,
1837 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1838 model = model.telemetry_id(),
1839 model_provider = model.provider_id().to_string(),
1840 attempt
1841 );
1842
1843 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1844
1845 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1846 Ok(events) => (events.fuse(), None),
1847 Err(err) => (stream::empty().boxed().fuse(), Some(err)),
1848 };
1849 let mut tool_results = FuturesUnordered::new();
1850 let mut cancelled = false;
1851 loop {
1852 // Race between getting the first event and cancellation
1853 let first_event = futures::select! {
1854 event = events.next().fuse() => event,
1855 _ = cancellation_rx.changed().fuse() => {
1856 if *cancellation_rx.borrow() {
1857 cancelled = true;
1858 break;
1859 }
1860 continue;
1861 }
1862 };
1863 let Some(first_event) = first_event else {
1864 break;
1865 };
1866
1867 // Collect all immediately available events to process as a batch
1868 let mut batch = vec![first_event];
1869 while let Some(event) = events.next().now_or_never().flatten() {
1870 batch.push(event);
1871 }
1872
1873 // Process the batch in a single update
1874 let batch_result = this.update(cx, |this, cx| {
1875 let mut batch_tool_results = Vec::new();
1876 let mut batch_error = None;
1877
1878 for event in batch {
1879 log::trace!("Received completion event: {:?}", event);
1880 match event {
1881 Ok(event) => {
1882 match this.handle_completion_event(
1883 event,
1884 event_stream,
1885 cancellation_rx.clone(),
1886 cx,
1887 ) {
1888 Ok(Some(task)) => batch_tool_results.push(task),
1889 Ok(None) => {}
1890 Err(err) => {
1891 batch_error = Some(err);
1892 break;
1893 }
1894 }
1895 }
1896 Err(err) => {
1897 batch_error = Some(err.into());
1898 break;
1899 }
1900 }
1901 }
1902
1903 cx.notify();
1904 (batch_tool_results, batch_error)
1905 })?;
1906
1907 tool_results.extend(batch_result.0);
1908 if let Some(err) = batch_result.1 {
1909 error = Some(err.downcast()?);
1910 break;
1911 }
1912 }
1913
1914 // Drop the stream to release the rate limit permit before tool execution.
1915 // The stream holds a semaphore guard that limits concurrent requests.
1916 // Without this, the permit would be held during potentially long-running
1917 // tool execution, which could cause deadlocks when tools spawn subagents
1918 // that need their own permits.
1919 drop(events);
1920
1921 // Drop streaming tool input senders that never received their final input.
1922 // This prevents deadlock when the LLM stream ends (e.g. because of an error)
1923 // before sending a tool use with `is_input_complete: true`.
1924 this.update(cx, |this, _cx| {
1925 if let Some(running_turn) = this.running_turn.as_mut() {
1926 if running_turn.streaming_tool_inputs.is_empty() {
1927 return;
1928 }
1929 log::warn!("Dropping partial tool inputs because the stream ended");
1930 running_turn.streaming_tool_inputs.drain();
1931 }
1932 })?;
1933
1934 let end_turn = tool_results.is_empty();
1935 while let Some(tool_result) = tool_results.next().await {
1936 log::debug!("Tool finished {:?}", tool_result);
1937
1938 event_stream.update_tool_call_fields(
1939 &tool_result.tool_use_id,
1940 acp::ToolCallUpdateFields::new()
1941 .status(if tool_result.is_error {
1942 acp::ToolCallStatus::Failed
1943 } else {
1944 acp::ToolCallStatus::Completed
1945 })
1946 .raw_output(tool_result.output.clone()),
1947 None,
1948 );
1949 this.update(cx, |this, _cx| {
1950 this.pending_message()
1951 .tool_results
1952 .insert(tool_result.tool_use_id.clone(), tool_result);
1953 })?;
1954 }
1955
1956 this.update(cx, |this, cx| {
1957 this.flush_pending_message(cx);
1958 if this.title.is_none() && this.pending_title_generation.is_none() {
1959 this.generate_title(cx);
1960 }
1961 })?;
1962
1963 if cancelled {
1964 log::debug!("Turn cancelled by user, exiting");
1965 return Ok(());
1966 }
1967
1968 if let Some(error) = error {
1969 attempt += 1;
1970 let retry = this.update(cx, |this, cx| {
1971 let user_store = this.user_store.read(cx);
1972 this.handle_completion_error(error, attempt, user_store.plan())
1973 })??;
1974 let timer = cx.background_executor().timer(retry.duration);
1975 event_stream.send_retry(retry);
1976 futures::select! {
1977 _ = timer.fuse() => {}
1978 _ = cancellation_rx.changed().fuse() => {
1979 if *cancellation_rx.borrow() {
1980 log::debug!("Turn cancelled during retry delay, exiting");
1981 return Ok(());
1982 }
1983 }
1984 }
1985 this.update(cx, |this, _cx| {
1986 if let Some(Message::Agent(message)) = this.messages.last() {
1987 if message.tool_results.is_empty() {
1988 intent = CompletionIntent::UserPrompt;
1989 this.messages.push(Message::Resume);
1990 }
1991 }
1992 })?;
1993 } else if end_turn {
1994 return Ok(());
1995 } else {
1996 let has_queued = this.update(cx, |this, _| this.has_queued_message())?;
1997 if has_queued {
1998 log::debug!("Queued message found, ending turn at message boundary");
1999 return Ok(());
2000 }
2001 intent = CompletionIntent::ToolResults;
2002 attempt = 0;
2003 }
2004 }
2005 }
2006
2007 fn handle_completion_error(
2008 &mut self,
2009 error: LanguageModelCompletionError,
2010 attempt: u8,
2011 plan: Option<Plan>,
2012 ) -> Result<acp_thread::RetryStatus> {
2013 let Some(model) = self.model.as_ref() else {
2014 return Err(anyhow!(error));
2015 };
2016
2017 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
2018 plan.is_some()
2019 } else {
2020 true
2021 };
2022
2023 if !auto_retry {
2024 return Err(anyhow!(error));
2025 }
2026
2027 let Some(strategy) = Self::retry_strategy_for(&error) else {
2028 return Err(anyhow!(error));
2029 };
2030
2031 let max_attempts = match &strategy {
2032 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2033 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2034 };
2035
2036 if attempt > max_attempts {
2037 return Err(anyhow!(error));
2038 }
2039
2040 let delay = match &strategy {
2041 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2042 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2043 Duration::from_secs(delay_secs)
2044 }
2045 RetryStrategy::Fixed { delay, .. } => *delay,
2046 };
2047 log::debug!("Retry attempt {attempt} with delay {delay:?}");
2048
2049 Ok(acp_thread::RetryStatus {
2050 last_error: error.to_string().into(),
2051 attempt: attempt as usize,
2052 max_attempts: max_attempts as usize,
2053 started_at: Instant::now(),
2054 duration: delay,
2055 })
2056 }
2057
2058 /// A helper method that's called on every streamed completion event.
2059 /// Returns an optional tool result task, which the main agentic loop will
2060 /// send back to the model when it resolves.
2061 fn handle_completion_event(
2062 &mut self,
2063 event: LanguageModelCompletionEvent,
2064 event_stream: &ThreadEventStream,
2065 cancellation_rx: watch::Receiver<bool>,
2066 cx: &mut Context<Self>,
2067 ) -> Result<Option<Task<LanguageModelToolResult>>> {
2068 log::trace!("Handling streamed completion event: {:?}", event);
2069 use LanguageModelCompletionEvent::*;
2070
2071 match event {
2072 StartMessage { .. } => {
2073 self.flush_pending_message(cx);
2074 self.pending_message = Some(AgentMessage::default());
2075 }
2076 Text(new_text) => self.handle_text_event(new_text, event_stream),
2077 Thinking { text, signature } => {
2078 self.handle_thinking_event(text, signature, event_stream)
2079 }
2080 RedactedThinking { data } => self.handle_redacted_thinking_event(data),
2081 ReasoningDetails(details) => {
2082 let last_message = self.pending_message();
2083 // Store the last non-empty reasoning_details (overwrites earlier ones)
2084 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
2085 if let serde_json::Value::Array(ref arr) = details {
2086 if !arr.is_empty() {
2087 last_message.reasoning_details = Some(details);
2088 }
2089 } else {
2090 last_message.reasoning_details = Some(details);
2091 }
2092 }
2093 ToolUse(tool_use) => {
2094 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
2095 }
2096 ToolUseJsonParseError {
2097 id,
2098 tool_name,
2099 raw_input,
2100 json_parse_error,
2101 } => {
2102 return Ok(Some(Task::ready(
2103 self.handle_tool_use_json_parse_error_event(
2104 id,
2105 tool_name,
2106 raw_input,
2107 json_parse_error,
2108 event_stream,
2109 ),
2110 )));
2111 }
2112 UsageUpdate(usage) => {
2113 telemetry::event!(
2114 "Agent Thread Completion Usage Updated",
2115 thread_id = self.id.to_string(),
2116 parent_thread_id = self.parent_thread_id().map(|id| id.to_string()),
2117 prompt_id = self.prompt_id.to_string(),
2118 model = self.model.as_ref().map(|m| m.telemetry_id()),
2119 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
2120 input_tokens = usage.input_tokens,
2121 output_tokens = usage.output_tokens,
2122 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2123 cache_read_input_tokens = usage.cache_read_input_tokens,
2124 );
2125 self.update_token_usage(usage, cx);
2126 }
2127 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
2128 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
2129 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
2130 Started | Queued { .. } => {}
2131 }
2132
2133 Ok(None)
2134 }
2135
2136 fn handle_text_event(&mut self, new_text: String, event_stream: &ThreadEventStream) {
2137 event_stream.send_text(&new_text);
2138
2139 let last_message = self.pending_message();
2140 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
2141 text.push_str(&new_text);
2142 } else {
2143 last_message
2144 .content
2145 .push(AgentMessageContent::Text(new_text));
2146 }
2147 }
2148
2149 fn handle_thinking_event(
2150 &mut self,
2151 new_text: String,
2152 new_signature: Option<String>,
2153 event_stream: &ThreadEventStream,
2154 ) {
2155 event_stream.send_thinking(&new_text);
2156
2157 let last_message = self.pending_message();
2158 if let Some(AgentMessageContent::Thinking { text, signature }) =
2159 last_message.content.last_mut()
2160 {
2161 text.push_str(&new_text);
2162 *signature = new_signature.or(signature.take());
2163 } else {
2164 last_message.content.push(AgentMessageContent::Thinking {
2165 text: new_text,
2166 signature: new_signature,
2167 });
2168 }
2169 }
2170
2171 fn handle_redacted_thinking_event(&mut self, data: String) {
2172 let last_message = self.pending_message();
2173 last_message
2174 .content
2175 .push(AgentMessageContent::RedactedThinking(data));
2176 }
2177
2178 fn handle_tool_use_event(
2179 &mut self,
2180 tool_use: LanguageModelToolUse,
2181 event_stream: &ThreadEventStream,
2182 cancellation_rx: watch::Receiver<bool>,
2183 cx: &mut Context<Self>,
2184 ) -> Option<Task<LanguageModelToolResult>> {
2185 cx.notify();
2186
2187 let tool = self.tool(tool_use.name.as_ref());
2188 let mut title = SharedString::from(&tool_use.name);
2189 let mut kind = acp::ToolKind::Other;
2190 if let Some(tool) = tool.as_ref() {
2191 title = tool.initial_title(tool_use.input.clone(), cx);
2192 kind = tool.kind();
2193 }
2194
2195 self.send_or_update_tool_use(&tool_use, title, kind, event_stream);
2196
2197 let Some(tool) = tool else {
2198 let content = format!("No tool named {} exists", tool_use.name);
2199 return Some(Task::ready(LanguageModelToolResult {
2200 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2201 tool_use_id: tool_use.id,
2202 tool_name: tool_use.name,
2203 is_error: true,
2204 output: None,
2205 }));
2206 };
2207
2208 if !tool_use.is_input_complete {
2209 if tool.supports_input_streaming() {
2210 let running_turn = self.running_turn.as_mut()?;
2211 if let Some(sender) = running_turn.streaming_tool_inputs.get(&tool_use.id) {
2212 sender.send_partial(tool_use.input);
2213 return None;
2214 }
2215
2216 let (sender, tool_input) = ToolInputSender::channel();
2217 sender.send_partial(tool_use.input);
2218 running_turn
2219 .streaming_tool_inputs
2220 .insert(tool_use.id.clone(), sender);
2221
2222 let tool = tool.clone();
2223 log::debug!("Running streaming tool {}", tool_use.name);
2224 return Some(self.run_tool(
2225 tool,
2226 tool_input,
2227 tool_use.id,
2228 tool_use.name,
2229 event_stream,
2230 cancellation_rx,
2231 cx,
2232 ));
2233 } else {
2234 return None;
2235 }
2236 }
2237
2238 if let Some(sender) = self
2239 .running_turn
2240 .as_mut()?
2241 .streaming_tool_inputs
2242 .remove(&tool_use.id)
2243 {
2244 sender.send_final(tool_use.input);
2245 return None;
2246 }
2247
2248 log::debug!("Running tool {}", tool_use.name);
2249 let tool_input = ToolInput::ready(tool_use.input);
2250 Some(self.run_tool(
2251 tool,
2252 tool_input,
2253 tool_use.id,
2254 tool_use.name,
2255 event_stream,
2256 cancellation_rx,
2257 cx,
2258 ))
2259 }
2260
2261 fn run_tool(
2262 &self,
2263 tool: Arc<dyn AnyAgentTool>,
2264 tool_input: ToolInput<serde_json::Value>,
2265 tool_use_id: LanguageModelToolUseId,
2266 tool_name: Arc<str>,
2267 event_stream: &ThreadEventStream,
2268 cancellation_rx: watch::Receiver<bool>,
2269 cx: &mut Context<Self>,
2270 ) -> Task<LanguageModelToolResult> {
2271 let fs = self.project.read(cx).fs().clone();
2272 let tool_event_stream = ToolCallEventStream::new(
2273 tool_use_id.clone(),
2274 event_stream.clone(),
2275 Some(fs),
2276 cancellation_rx,
2277 );
2278 tool_event_stream.update_fields(
2279 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
2280 );
2281 let supports_images = self.model().is_some_and(|model| model.supports_images());
2282 let tool_result = tool.run(tool_input, tool_event_stream, cx);
2283 cx.foreground_executor().spawn(async move {
2284 let (is_error, output) = match tool_result.await {
2285 Ok(mut output) => {
2286 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
2287 && !supports_images
2288 {
2289 output = AgentToolOutput::from_error(
2290 "Attempted to read an image, but this model doesn't support it.",
2291 );
2292 (true, output)
2293 } else {
2294 (false, output)
2295 }
2296 }
2297 Err(output) => (true, output),
2298 };
2299
2300 LanguageModelToolResult {
2301 tool_use_id,
2302 tool_name,
2303 is_error,
2304 content: output.llm_output,
2305 output: Some(output.raw_output),
2306 }
2307 })
2308 }
2309
2310 fn handle_tool_use_json_parse_error_event(
2311 &mut self,
2312 tool_use_id: LanguageModelToolUseId,
2313 tool_name: Arc<str>,
2314 raw_input: Arc<str>,
2315 json_parse_error: String,
2316 event_stream: &ThreadEventStream,
2317 ) -> LanguageModelToolResult {
2318 let tool_use = LanguageModelToolUse {
2319 id: tool_use_id.clone(),
2320 name: tool_name.clone(),
2321 raw_input: raw_input.to_string(),
2322 input: serde_json::json!({}),
2323 is_input_complete: true,
2324 thought_signature: None,
2325 };
2326 self.send_or_update_tool_use(
2327 &tool_use,
2328 SharedString::from(&tool_use.name),
2329 acp::ToolKind::Other,
2330 event_stream,
2331 );
2332
2333 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
2334 LanguageModelToolResult {
2335 tool_use_id,
2336 tool_name,
2337 is_error: true,
2338 content: LanguageModelToolResultContent::Text(tool_output.into()),
2339 output: Some(serde_json::Value::String(raw_input.to_string())),
2340 }
2341 }
2342
2343 fn send_or_update_tool_use(
2344 &mut self,
2345 tool_use: &LanguageModelToolUse,
2346 title: SharedString,
2347 kind: acp::ToolKind,
2348 event_stream: &ThreadEventStream,
2349 ) {
2350 // Ensure the last message ends in the current tool use
2351 let last_message = self.pending_message();
2352
2353 let has_tool_use = last_message.content.iter_mut().rev().any(|content| {
2354 if let AgentMessageContent::ToolUse(last_tool_use) = content {
2355 if last_tool_use.id == tool_use.id {
2356 *last_tool_use = tool_use.clone();
2357 return true;
2358 }
2359 }
2360 false
2361 });
2362
2363 if !has_tool_use {
2364 event_stream.send_tool_call(
2365 &tool_use.id,
2366 &tool_use.name,
2367 title,
2368 kind,
2369 tool_use.input.clone(),
2370 );
2371 last_message
2372 .content
2373 .push(AgentMessageContent::ToolUse(tool_use.clone()));
2374 } else {
2375 event_stream.update_tool_call_fields(
2376 &tool_use.id,
2377 acp::ToolCallUpdateFields::new()
2378 .title(title.as_str())
2379 .kind(kind)
2380 .raw_input(tool_use.input.clone()),
2381 None,
2382 );
2383 }
2384 }
2385
2386 pub fn title(&self) -> SharedString {
2387 self.title.clone().unwrap_or("New Thread".into())
2388 }
2389
2390 pub fn is_generating_summary(&self) -> bool {
2391 self.pending_summary_generation.is_some()
2392 }
2393
2394 pub fn is_generating_title(&self) -> bool {
2395 self.pending_title_generation.is_some()
2396 }
2397
2398 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
2399 if let Some(summary) = self.summary.as_ref() {
2400 return Task::ready(Some(summary.clone())).shared();
2401 }
2402 if let Some(task) = self.pending_summary_generation.clone() {
2403 return task;
2404 }
2405 let Some(model) = self.summarization_model.clone() else {
2406 log::error!("No summarization model available");
2407 return Task::ready(None).shared();
2408 };
2409 let mut request = LanguageModelRequest {
2410 intent: Some(CompletionIntent::ThreadContextSummarization),
2411 temperature: AgentSettings::temperature_for_model(&model, cx),
2412 ..Default::default()
2413 };
2414
2415 for message in &self.messages {
2416 request.messages.extend(message.to_request());
2417 }
2418
2419 request.messages.push(LanguageModelRequestMessage {
2420 role: Role::User,
2421 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
2422 cache: false,
2423 reasoning_details: None,
2424 });
2425
2426 let task = cx
2427 .spawn(async move |this, cx| {
2428 let mut summary = String::new();
2429 let mut messages = model.stream_completion(request, cx).await.log_err()?;
2430 while let Some(event) = messages.next().await {
2431 let event = event.log_err()?;
2432 let text = match event {
2433 LanguageModelCompletionEvent::Text(text) => text,
2434 _ => continue,
2435 };
2436
2437 let mut lines = text.lines();
2438 summary.extend(lines.next());
2439 }
2440
2441 log::debug!("Setting summary: {}", summary);
2442 let summary = SharedString::from(summary);
2443
2444 this.update(cx, |this, cx| {
2445 this.summary = Some(summary.clone());
2446 this.pending_summary_generation = None;
2447 cx.notify()
2448 })
2449 .ok()?;
2450
2451 Some(summary)
2452 })
2453 .shared();
2454 self.pending_summary_generation = Some(task.clone());
2455 task
2456 }
2457
2458 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
2459 let Some(model) = self.summarization_model.clone() else {
2460 return;
2461 };
2462
2463 log::debug!(
2464 "Generating title with model: {:?}",
2465 self.summarization_model.as_ref().map(|model| model.name())
2466 );
2467 let mut request = LanguageModelRequest {
2468 intent: Some(CompletionIntent::ThreadSummarization),
2469 temperature: AgentSettings::temperature_for_model(&model, cx),
2470 ..Default::default()
2471 };
2472
2473 for message in &self.messages {
2474 request.messages.extend(message.to_request());
2475 }
2476
2477 request.messages.push(LanguageModelRequestMessage {
2478 role: Role::User,
2479 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2480 cache: false,
2481 reasoning_details: None,
2482 });
2483 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2484 let mut title = String::new();
2485
2486 let generate = async {
2487 let mut messages = model.stream_completion(request, cx).await?;
2488 while let Some(event) = messages.next().await {
2489 let event = event?;
2490 let text = match event {
2491 LanguageModelCompletionEvent::Text(text) => text,
2492 _ => continue,
2493 };
2494
2495 let mut lines = text.lines();
2496 title.extend(lines.next());
2497
2498 // Stop if the LLM generated multiple lines.
2499 if lines.next().is_some() {
2500 break;
2501 }
2502 }
2503 anyhow::Ok(())
2504 };
2505
2506 if generate
2507 .await
2508 .context("failed to generate thread title")
2509 .log_err()
2510 .is_some()
2511 {
2512 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
2513 }
2514 _ = this.update(cx, |this, _| this.pending_title_generation = None);
2515 }));
2516 }
2517
2518 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2519 self.pending_title_generation = None;
2520 if Some(&title) != self.title.as_ref() {
2521 self.title = Some(title);
2522 cx.emit(TitleUpdated);
2523 cx.notify();
2524 }
2525 }
2526
2527 fn clear_summary(&mut self) {
2528 self.summary = None;
2529 self.pending_summary_generation = None;
2530 }
2531
2532 fn last_user_message(&self) -> Option<&UserMessage> {
2533 self.messages
2534 .iter()
2535 .rev()
2536 .find_map(|message| match message {
2537 Message::User(user_message) => Some(user_message),
2538 Message::Agent(_) => None,
2539 Message::Resume => None,
2540 })
2541 }
2542
2543 fn pending_message(&mut self) -> &mut AgentMessage {
2544 self.pending_message.get_or_insert_default()
2545 }
2546
2547 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2548 let Some(mut message) = self.pending_message.take() else {
2549 return;
2550 };
2551
2552 if message.content.is_empty() {
2553 return;
2554 }
2555
2556 for content in &message.content {
2557 let AgentMessageContent::ToolUse(tool_use) = content else {
2558 continue;
2559 };
2560
2561 if !message.tool_results.contains_key(&tool_use.id) {
2562 message.tool_results.insert(
2563 tool_use.id.clone(),
2564 LanguageModelToolResult {
2565 tool_use_id: tool_use.id.clone(),
2566 tool_name: tool_use.name.clone(),
2567 is_error: true,
2568 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2569 output: None,
2570 },
2571 );
2572 }
2573 }
2574
2575 self.messages.push(Message::Agent(message));
2576 self.updated_at = Utc::now();
2577 self.clear_summary();
2578 cx.notify()
2579 }
2580
2581 pub(crate) fn build_completion_request(
2582 &self,
2583 completion_intent: CompletionIntent,
2584 cx: &App,
2585 ) -> Result<LanguageModelRequest> {
2586 let model = self.model().context("No language model configured")?;
2587 let tools = if let Some(turn) = self.running_turn.as_ref() {
2588 turn.tools
2589 .iter()
2590 .filter_map(|(tool_name, tool)| {
2591 log::trace!("Including tool: {}", tool_name);
2592 Some(LanguageModelRequestTool {
2593 name: tool_name.to_string(),
2594 description: tool.description().to_string(),
2595 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2596 use_input_streaming: tool.supports_input_streaming(),
2597 })
2598 })
2599 .collect::<Vec<_>>()
2600 } else {
2601 Vec::new()
2602 };
2603
2604 log::debug!("Building completion request");
2605 log::debug!("Completion intent: {:?}", completion_intent);
2606
2607 let available_tools: Vec<_> = self
2608 .running_turn
2609 .as_ref()
2610 .map(|turn| turn.tools.keys().cloned().collect())
2611 .unwrap_or_default();
2612
2613 log::debug!("Request includes {} tools", available_tools.len());
2614 let messages = self.build_request_messages(available_tools, cx);
2615 log::debug!("Request will include {} messages", messages.len());
2616
2617 let request = LanguageModelRequest {
2618 thread_id: Some(self.id.to_string()),
2619 prompt_id: Some(self.prompt_id.to_string()),
2620 intent: Some(completion_intent),
2621 messages,
2622 tools,
2623 tool_choice: None,
2624 stop: Vec::new(),
2625 temperature: AgentSettings::temperature_for_model(model, cx),
2626 thinking_allowed: self.thinking_enabled,
2627 thinking_effort: self.thinking_effort.clone(),
2628 speed: self.speed(),
2629 };
2630
2631 log::debug!("Completion request built successfully");
2632 Ok(request)
2633 }
2634
2635 fn enabled_tools(
2636 &self,
2637 profile: &AgentProfileSettings,
2638 model: &Arc<dyn LanguageModel>,
2639 cx: &App,
2640 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2641 fn truncate(tool_name: &SharedString) -> SharedString {
2642 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2643 let mut truncated = tool_name.to_string();
2644 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2645 truncated.into()
2646 } else {
2647 tool_name.clone()
2648 }
2649 }
2650
2651 let use_streaming_edit_tool =
2652 cx.has_flag::<StreamingEditFileToolFeatureFlag>() && model.supports_streaming_tools();
2653
2654 let mut tools = self
2655 .tools
2656 .iter()
2657 .filter_map(|(tool_name, tool)| {
2658 // For streaming_edit_file, check profile against "edit_file" since that's what users configure
2659 let profile_tool_name = if tool_name == StreamingEditFileTool::NAME {
2660 EditFileTool::NAME
2661 } else {
2662 tool_name.as_ref()
2663 };
2664
2665 if tool.supports_provider(&model.provider_id())
2666 && profile.is_tool_enabled(profile_tool_name)
2667 {
2668 match (tool_name.as_ref(), use_streaming_edit_tool) {
2669 (StreamingEditFileTool::NAME, false) | (EditFileTool::NAME, true) => None,
2670 (StreamingEditFileTool::NAME, true) => {
2671 // Expose streaming tool as "edit_file"
2672 Some((SharedString::from(EditFileTool::NAME), tool.clone()))
2673 }
2674 _ => Some((truncate(tool_name), tool.clone())),
2675 }
2676 } else {
2677 None
2678 }
2679 })
2680 .collect::<BTreeMap<_, _>>();
2681
2682 let mut context_server_tools = Vec::new();
2683 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2684 let mut duplicate_tool_names = HashSet::default();
2685 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2686 for (tool_name, tool) in server_tools {
2687 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2688 let tool_name = truncate(tool_name);
2689 if !seen_tools.insert(tool_name.clone()) {
2690 duplicate_tool_names.insert(tool_name.clone());
2691 }
2692 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2693 }
2694 }
2695 }
2696
2697 // When there are duplicate tool names, disambiguate by prefixing them
2698 // with the server ID (converted to snake_case for API compatibility).
2699 // In the rare case there isn't enough space for the disambiguated tool
2700 // name, keep only the last tool with this name.
2701 for (server_id, tool_name, tool) in context_server_tools {
2702 if duplicate_tool_names.contains(&tool_name) {
2703 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2704 if available >= 2 {
2705 let mut disambiguated = server_id.0.to_snake_case();
2706 disambiguated.truncate(available - 1);
2707 disambiguated.push('_');
2708 disambiguated.push_str(&tool_name);
2709 tools.insert(disambiguated.into(), tool.clone());
2710 } else {
2711 tools.insert(tool_name, tool.clone());
2712 }
2713 } else {
2714 tools.insert(tool_name, tool.clone());
2715 }
2716 }
2717
2718 tools
2719 }
2720
2721 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2722 self.running_turn.as_ref()?.tools.get(name).cloned()
2723 }
2724
2725 pub fn has_tool(&self, name: &str) -> bool {
2726 self.running_turn
2727 .as_ref()
2728 .is_some_and(|turn| turn.tools.contains_key(name))
2729 }
2730
2731 #[cfg(any(test, feature = "test-support"))]
2732 pub fn has_registered_tool(&self, name: &str) -> bool {
2733 self.tools.contains_key(name)
2734 }
2735
2736 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2737 self.tools.keys().cloned().collect()
2738 }
2739
2740 pub(crate) fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2741 self.running_subagents.push(subagent);
2742 }
2743
2744 pub(crate) fn unregister_running_subagent(
2745 &mut self,
2746 subagent_session_id: &acp::SessionId,
2747 cx: &App,
2748 ) {
2749 self.running_subagents.retain(|s| {
2750 s.upgrade()
2751 .map_or(false, |s| s.read(cx).id() != subagent_session_id)
2752 });
2753 }
2754
2755 #[cfg(any(test, feature = "test-support"))]
2756 pub fn running_subagent_ids(&self, cx: &App) -> Vec<acp::SessionId> {
2757 self.running_subagents
2758 .iter()
2759 .filter_map(|s| s.upgrade().map(|s| s.read(cx).id().clone()))
2760 .collect()
2761 }
2762
2763 pub fn is_subagent(&self) -> bool {
2764 self.subagent_context.is_some()
2765 }
2766
2767 pub fn parent_thread_id(&self) -> Option<acp::SessionId> {
2768 self.subagent_context
2769 .as_ref()
2770 .map(|c| c.parent_thread_id.clone())
2771 }
2772
2773 pub fn depth(&self) -> u8 {
2774 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2775 }
2776
2777 #[cfg(any(test, feature = "test-support"))]
2778 pub fn set_subagent_context(&mut self, context: SubagentContext) {
2779 self.subagent_context = Some(context);
2780 }
2781
2782 pub fn is_turn_complete(&self) -> bool {
2783 self.running_turn.is_none()
2784 }
2785
2786 fn build_request_messages(
2787 &self,
2788 available_tools: Vec<SharedString>,
2789 cx: &App,
2790 ) -> Vec<LanguageModelRequestMessage> {
2791 log::trace!(
2792 "Building request messages from {} thread messages",
2793 self.messages.len()
2794 );
2795
2796 let system_prompt = SystemPromptTemplate {
2797 project: self.project_context.read(cx),
2798 available_tools,
2799 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2800 }
2801 .render(&self.templates)
2802 .context("failed to build system prompt")
2803 .expect("Invalid template");
2804 let mut messages = vec![LanguageModelRequestMessage {
2805 role: Role::System,
2806 content: vec![system_prompt.into()],
2807 cache: false,
2808 reasoning_details: None,
2809 }];
2810 for message in &self.messages {
2811 messages.extend(message.to_request());
2812 }
2813
2814 if let Some(last_message) = messages.last_mut() {
2815 last_message.cache = true;
2816 }
2817
2818 if let Some(message) = self.pending_message.as_ref() {
2819 messages.extend(message.to_request());
2820 }
2821
2822 messages
2823 }
2824
2825 pub fn to_markdown(&self) -> String {
2826 let mut markdown = String::new();
2827 for (ix, message) in self.messages.iter().enumerate() {
2828 if ix > 0 {
2829 markdown.push('\n');
2830 }
2831 match message {
2832 Message::User(_) => markdown.push_str("## User\n\n"),
2833 Message::Agent(_) => markdown.push_str("## Assistant\n\n"),
2834 Message::Resume => {}
2835 }
2836 markdown.push_str(&message.to_markdown());
2837 }
2838
2839 if let Some(message) = self.pending_message.as_ref() {
2840 markdown.push_str("\n## Assistant\n\n");
2841 markdown.push_str(&message.to_markdown());
2842 }
2843
2844 markdown
2845 }
2846
2847 fn advance_prompt_id(&mut self) {
2848 self.prompt_id = PromptId::new();
2849 }
2850
2851 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2852 use LanguageModelCompletionError::*;
2853 use http_client::StatusCode;
2854
2855 // General strategy here:
2856 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2857 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2858 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2859 match error {
2860 HttpResponseError {
2861 status_code: StatusCode::TOO_MANY_REQUESTS,
2862 ..
2863 } => Some(RetryStrategy::ExponentialBackoff {
2864 initial_delay: BASE_RETRY_DELAY,
2865 max_attempts: MAX_RETRY_ATTEMPTS,
2866 }),
2867 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2868 Some(RetryStrategy::Fixed {
2869 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2870 max_attempts: MAX_RETRY_ATTEMPTS,
2871 })
2872 }
2873 UpstreamProviderError {
2874 status,
2875 retry_after,
2876 ..
2877 } => match *status {
2878 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2879 Some(RetryStrategy::Fixed {
2880 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2881 max_attempts: MAX_RETRY_ATTEMPTS,
2882 })
2883 }
2884 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2885 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2886 // Internal Server Error could be anything, retry up to 3 times.
2887 max_attempts: 3,
2888 }),
2889 status => {
2890 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2891 // but we frequently get them in practice. See https://http.dev/529
2892 if status.as_u16() == 529 {
2893 Some(RetryStrategy::Fixed {
2894 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2895 max_attempts: MAX_RETRY_ATTEMPTS,
2896 })
2897 } else {
2898 Some(RetryStrategy::Fixed {
2899 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2900 max_attempts: 2,
2901 })
2902 }
2903 }
2904 },
2905 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2906 delay: BASE_RETRY_DELAY,
2907 max_attempts: 3,
2908 }),
2909 ApiReadResponseError { .. }
2910 | HttpSend { .. }
2911 | DeserializeResponse { .. }
2912 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2913 delay: BASE_RETRY_DELAY,
2914 max_attempts: 3,
2915 }),
2916 // Retrying these errors definitely shouldn't help.
2917 HttpResponseError {
2918 status_code:
2919 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2920 ..
2921 }
2922 | AuthenticationError { .. }
2923 | PermissionError { .. }
2924 | NoApiKey { .. }
2925 | ApiEndpointNotFound { .. }
2926 | PromptTooLarge { .. } => None,
2927 // These errors might be transient, so retry them
2928 SerializeRequest { .. } | BuildRequestBody { .. } | StreamEndedUnexpectedly { .. } => {
2929 Some(RetryStrategy::Fixed {
2930 delay: BASE_RETRY_DELAY,
2931 max_attempts: 1,
2932 })
2933 }
2934 // Retry all other 4xx and 5xx errors once.
2935 HttpResponseError { status_code, .. }
2936 if status_code.is_client_error() || status_code.is_server_error() =>
2937 {
2938 Some(RetryStrategy::Fixed {
2939 delay: BASE_RETRY_DELAY,
2940 max_attempts: 3,
2941 })
2942 }
2943 Other(err) if err.is::<language_model::PaymentRequiredError>() => {
2944 // Retrying won't help for Payment Required errors.
2945 None
2946 }
2947 // Conservatively assume that any other errors are non-retryable
2948 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2949 delay: BASE_RETRY_DELAY,
2950 max_attempts: 2,
2951 }),
2952 }
2953 }
2954}
2955
2956struct RunningTurn {
2957 /// Holds the task that handles agent interaction until the end of the turn.
2958 /// Survives across multiple requests as the model performs tool calls and
2959 /// we run tools, report their results.
2960 _task: Task<()>,
2961 /// The current event stream for the running turn. Used to report a final
2962 /// cancellation event if we cancel the turn.
2963 event_stream: ThreadEventStream,
2964 /// The tools that were enabled for this turn.
2965 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2966 /// Sender to signal tool cancellation. When cancel is called, this is
2967 /// set to true so all tools can detect user-initiated cancellation.
2968 cancellation_tx: watch::Sender<bool>,
2969 /// Senders for tools that support input streaming and have already been
2970 /// started but are still receiving input from the LLM.
2971 streaming_tool_inputs: HashMap<LanguageModelToolUseId, ToolInputSender>,
2972}
2973
2974impl RunningTurn {
2975 fn cancel(mut self) -> Task<()> {
2976 log::debug!("Cancelling in progress turn");
2977 self.cancellation_tx.send(true).ok();
2978 self.event_stream.send_canceled();
2979 self._task
2980 }
2981}
2982
2983pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2984
2985impl EventEmitter<TokenUsageUpdated> for Thread {}
2986
2987pub struct TitleUpdated;
2988
2989impl EventEmitter<TitleUpdated> for Thread {}
2990
2991/// A channel-based wrapper that delivers tool input to a running tool.
2992///
2993/// For non-streaming tools, created via `ToolInput::ready()` so `.recv()` resolves immediately.
2994/// For streaming tools, partial JSON snapshots arrive via `.recv_partial()` as the LLM streams
2995/// them, followed by the final complete input available through `.recv()`.
2996pub struct ToolInput<T> {
2997 partial_rx: mpsc::UnboundedReceiver<serde_json::Value>,
2998 final_rx: oneshot::Receiver<serde_json::Value>,
2999 _phantom: PhantomData<T>,
3000}
3001
3002impl<T: DeserializeOwned> ToolInput<T> {
3003 #[cfg(any(test, feature = "test-support"))]
3004 pub fn resolved(input: impl Serialize) -> Self {
3005 let value = serde_json::to_value(input).expect("failed to serialize tool input");
3006 Self::ready(value)
3007 }
3008
3009 pub fn ready(value: serde_json::Value) -> Self {
3010 let (partial_tx, partial_rx) = mpsc::unbounded();
3011 drop(partial_tx);
3012 let (final_tx, final_rx) = oneshot::channel();
3013 final_tx.send(value).ok();
3014 Self {
3015 partial_rx,
3016 final_rx,
3017 _phantom: PhantomData,
3018 }
3019 }
3020
3021 #[cfg(any(test, feature = "test-support"))]
3022 pub fn test() -> (ToolInputSender, Self) {
3023 let (sender, input) = ToolInputSender::channel();
3024 (sender, input.cast())
3025 }
3026
3027 /// Wait for the final deserialized input, ignoring all partial updates.
3028 /// Non-streaming tools can use this to wait until the whole input is available.
3029 pub async fn recv(mut self) -> Result<T> {
3030 // Drain any remaining partials
3031 while self.partial_rx.next().await.is_some() {}
3032 let value = self
3033 .final_rx
3034 .await
3035 .map_err(|_| anyhow!("tool input was not fully received"))?;
3036 serde_json::from_value(value).map_err(Into::into)
3037 }
3038
3039 /// Returns the next partial JSON snapshot, or `None` when input is complete.
3040 /// Once this returns `None`, call `recv()` to get the final input.
3041 pub async fn recv_partial(&mut self) -> Option<serde_json::Value> {
3042 self.partial_rx.next().await
3043 }
3044
3045 fn cast<U: DeserializeOwned>(self) -> ToolInput<U> {
3046 ToolInput {
3047 partial_rx: self.partial_rx,
3048 final_rx: self.final_rx,
3049 _phantom: PhantomData,
3050 }
3051 }
3052}
3053
3054pub struct ToolInputSender {
3055 partial_tx: mpsc::UnboundedSender<serde_json::Value>,
3056 final_tx: Option<oneshot::Sender<serde_json::Value>>,
3057}
3058
3059impl ToolInputSender {
3060 pub(crate) fn channel() -> (Self, ToolInput<serde_json::Value>) {
3061 let (partial_tx, partial_rx) = mpsc::unbounded();
3062 let (final_tx, final_rx) = oneshot::channel();
3063 let sender = Self {
3064 partial_tx,
3065 final_tx: Some(final_tx),
3066 };
3067 let input = ToolInput {
3068 partial_rx,
3069 final_rx,
3070 _phantom: PhantomData,
3071 };
3072 (sender, input)
3073 }
3074
3075 pub(crate) fn send_partial(&self, value: serde_json::Value) {
3076 self.partial_tx.unbounded_send(value).ok();
3077 }
3078
3079 pub(crate) fn send_final(mut self, value: serde_json::Value) {
3080 // Close the partial channel so recv_partial() returns None
3081 self.partial_tx.close_channel();
3082 if let Some(final_tx) = self.final_tx.take() {
3083 final_tx.send(value).ok();
3084 }
3085 }
3086}
3087
3088pub trait AgentTool
3089where
3090 Self: 'static + Sized,
3091{
3092 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
3093 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
3094
3095 const NAME: &'static str;
3096
3097 fn description() -> SharedString {
3098 let schema = schemars::schema_for!(Self::Input);
3099 SharedString::new(
3100 schema
3101 .get("description")
3102 .and_then(|description| description.as_str())
3103 .unwrap_or_default(),
3104 )
3105 }
3106
3107 fn kind() -> acp::ToolKind;
3108
3109 /// The initial tool title to display. Can be updated during the tool run.
3110 fn initial_title(
3111 &self,
3112 input: Result<Self::Input, serde_json::Value>,
3113 cx: &mut App,
3114 ) -> SharedString;
3115
3116 /// Returns the JSON schema that describes the tool's input.
3117 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
3118 language_model::tool_schema::root_schema_for::<Self::Input>(format)
3119 }
3120
3121 /// Returns whether the tool supports streaming of tool use parameters.
3122 fn supports_input_streaming() -> bool {
3123 false
3124 }
3125
3126 /// Some tools rely on a provider for the underlying billing or other reasons.
3127 /// Allow the tool to check if they are compatible, or should be filtered out.
3128 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
3129 true
3130 }
3131
3132 /// Runs the tool with the provided input.
3133 ///
3134 /// Returns `Result<Self::Output, Self::Output>` rather than `Result<Self::Output, anyhow::Error>`
3135 /// because tool errors are sent back to the model as tool results. This means error output must
3136 /// be structured and readable by the agent — not an arbitrary `anyhow::Error`. Returning the
3137 /// same `Output` type for both success and failure lets tools provide structured data while
3138 /// still signaling whether the invocation succeeded or failed.
3139 fn run(
3140 self: Arc<Self>,
3141 input: ToolInput<Self::Input>,
3142 event_stream: ToolCallEventStream,
3143 cx: &mut App,
3144 ) -> Task<Result<Self::Output, Self::Output>>;
3145
3146 /// Emits events for a previous execution of the tool.
3147 fn replay(
3148 &self,
3149 _input: Self::Input,
3150 _output: Self::Output,
3151 _event_stream: ToolCallEventStream,
3152 _cx: &mut App,
3153 ) -> Result<()> {
3154 Ok(())
3155 }
3156
3157 fn erase(self) -> Arc<dyn AnyAgentTool> {
3158 Arc::new(Erased(Arc::new(self)))
3159 }
3160}
3161
3162pub struct Erased<T>(T);
3163
3164pub struct AgentToolOutput {
3165 pub llm_output: LanguageModelToolResultContent,
3166 pub raw_output: serde_json::Value,
3167}
3168
3169impl AgentToolOutput {
3170 pub fn from_error(message: impl Into<String>) -> Self {
3171 let message = message.into();
3172 let llm_output = LanguageModelToolResultContent::Text(Arc::from(message.as_str()));
3173 Self {
3174 raw_output: serde_json::Value::String(message),
3175 llm_output,
3176 }
3177 }
3178}
3179
3180pub trait AnyAgentTool {
3181 fn name(&self) -> SharedString;
3182 fn description(&self) -> SharedString;
3183 fn kind(&self) -> acp::ToolKind;
3184 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
3185 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
3186 fn supports_input_streaming(&self) -> bool {
3187 false
3188 }
3189 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
3190 true
3191 }
3192 /// See [`AgentTool::run`] for why this returns `Result<AgentToolOutput, AgentToolOutput>`.
3193 fn run(
3194 self: Arc<Self>,
3195 input: ToolInput<serde_json::Value>,
3196 event_stream: ToolCallEventStream,
3197 cx: &mut App,
3198 ) -> Task<Result<AgentToolOutput, AgentToolOutput>>;
3199 fn replay(
3200 &self,
3201 input: serde_json::Value,
3202 output: serde_json::Value,
3203 event_stream: ToolCallEventStream,
3204 cx: &mut App,
3205 ) -> Result<()>;
3206}
3207
3208impl<T> AnyAgentTool for Erased<Arc<T>>
3209where
3210 T: AgentTool,
3211{
3212 fn name(&self) -> SharedString {
3213 T::NAME.into()
3214 }
3215
3216 fn description(&self) -> SharedString {
3217 T::description()
3218 }
3219
3220 fn kind(&self) -> agent_client_protocol::ToolKind {
3221 T::kind()
3222 }
3223
3224 fn supports_input_streaming(&self) -> bool {
3225 T::supports_input_streaming()
3226 }
3227
3228 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
3229 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
3230 self.0.initial_title(parsed_input, _cx)
3231 }
3232
3233 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
3234 let mut json = serde_json::to_value(T::input_schema(format))?;
3235 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
3236 Ok(json)
3237 }
3238
3239 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
3240 T::supports_provider(provider)
3241 }
3242
3243 fn run(
3244 self: Arc<Self>,
3245 input: ToolInput<serde_json::Value>,
3246 event_stream: ToolCallEventStream,
3247 cx: &mut App,
3248 ) -> Task<Result<AgentToolOutput, AgentToolOutput>> {
3249 let tool_input: ToolInput<T::Input> = input.cast();
3250 let task = self.0.clone().run(tool_input, event_stream, cx);
3251 cx.spawn(async move |_cx| match task.await {
3252 Ok(output) => {
3253 let raw_output = serde_json::to_value(&output).map_err(|e| {
3254 AgentToolOutput::from_error(format!("Failed to serialize tool output: {e}"))
3255 })?;
3256 Ok(AgentToolOutput {
3257 llm_output: output.into(),
3258 raw_output,
3259 })
3260 }
3261 Err(error_output) => {
3262 let raw_output = serde_json::to_value(&error_output).unwrap_or_else(|e| {
3263 log::error!("Failed to serialize tool error output: {e}");
3264 serde_json::Value::Null
3265 });
3266 Err(AgentToolOutput {
3267 llm_output: error_output.into(),
3268 raw_output,
3269 })
3270 }
3271 })
3272 }
3273
3274 fn replay(
3275 &self,
3276 input: serde_json::Value,
3277 output: serde_json::Value,
3278 event_stream: ToolCallEventStream,
3279 cx: &mut App,
3280 ) -> Result<()> {
3281 let input = serde_json::from_value(input)?;
3282 let output = serde_json::from_value(output)?;
3283 self.0.replay(input, output, event_stream, cx)
3284 }
3285}
3286
3287#[derive(Clone)]
3288struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
3289
3290impl ThreadEventStream {
3291 fn send_user_message(&self, message: &UserMessage) {
3292 self.0
3293 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
3294 .ok();
3295 }
3296
3297 fn send_text(&self, text: &str) {
3298 self.0
3299 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
3300 .ok();
3301 }
3302
3303 fn send_thinking(&self, text: &str) {
3304 self.0
3305 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
3306 .ok();
3307 }
3308
3309 fn send_tool_call(
3310 &self,
3311 id: &LanguageModelToolUseId,
3312 tool_name: &str,
3313 title: SharedString,
3314 kind: acp::ToolKind,
3315 input: serde_json::Value,
3316 ) {
3317 self.0
3318 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
3319 id,
3320 tool_name,
3321 title.to_string(),
3322 kind,
3323 input,
3324 ))))
3325 .ok();
3326 }
3327
3328 fn initial_tool_call(
3329 id: &LanguageModelToolUseId,
3330 tool_name: &str,
3331 title: String,
3332 kind: acp::ToolKind,
3333 input: serde_json::Value,
3334 ) -> acp::ToolCall {
3335 acp::ToolCall::new(id.to_string(), title)
3336 .kind(kind)
3337 .raw_input(input)
3338 .meta(acp_thread::meta_with_tool_name(tool_name))
3339 }
3340
3341 fn update_tool_call_fields(
3342 &self,
3343 tool_use_id: &LanguageModelToolUseId,
3344 fields: acp::ToolCallUpdateFields,
3345 meta: Option<acp::Meta>,
3346 ) {
3347 self.0
3348 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3349 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields)
3350 .meta(meta)
3351 .into(),
3352 )))
3353 .ok();
3354 }
3355
3356 fn send_retry(&self, status: acp_thread::RetryStatus) {
3357 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
3358 }
3359
3360 fn send_stop(&self, reason: acp::StopReason) {
3361 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
3362 }
3363
3364 fn send_canceled(&self) {
3365 self.0
3366 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
3367 .ok();
3368 }
3369
3370 fn send_error(&self, error: impl Into<anyhow::Error>) {
3371 self.0.unbounded_send(Err(error.into())).ok();
3372 }
3373}
3374
3375#[derive(Clone)]
3376pub struct ToolCallEventStream {
3377 tool_use_id: LanguageModelToolUseId,
3378 stream: ThreadEventStream,
3379 fs: Option<Arc<dyn Fs>>,
3380 cancellation_rx: watch::Receiver<bool>,
3381}
3382
3383impl ToolCallEventStream {
3384 #[cfg(any(test, feature = "test-support"))]
3385 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
3386 let (stream, receiver, _cancellation_tx) = Self::test_with_cancellation();
3387 (stream, receiver)
3388 }
3389
3390 #[cfg(any(test, feature = "test-support"))]
3391 pub fn test_with_cancellation() -> (Self, ToolCallEventStreamReceiver, watch::Sender<bool>) {
3392 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
3393 let (cancellation_tx, cancellation_rx) = watch::channel(false);
3394
3395 let stream = ToolCallEventStream::new(
3396 "test_id".into(),
3397 ThreadEventStream(events_tx),
3398 None,
3399 cancellation_rx,
3400 );
3401
3402 (
3403 stream,
3404 ToolCallEventStreamReceiver(events_rx),
3405 cancellation_tx,
3406 )
3407 }
3408
3409 /// Signal cancellation for this event stream. Only available in tests.
3410 #[cfg(any(test, feature = "test-support"))]
3411 pub fn signal_cancellation_with_sender(cancellation_tx: &mut watch::Sender<bool>) {
3412 cancellation_tx.send(true).ok();
3413 }
3414
3415 fn new(
3416 tool_use_id: LanguageModelToolUseId,
3417 stream: ThreadEventStream,
3418 fs: Option<Arc<dyn Fs>>,
3419 cancellation_rx: watch::Receiver<bool>,
3420 ) -> Self {
3421 Self {
3422 tool_use_id,
3423 stream,
3424 fs,
3425 cancellation_rx,
3426 }
3427 }
3428
3429 /// Returns a future that resolves when the user cancels the tool call.
3430 /// Tools should select on this alongside their main work to detect user cancellation.
3431 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
3432 let mut rx = self.cancellation_rx.clone();
3433 async move {
3434 loop {
3435 if *rx.borrow() {
3436 return;
3437 }
3438 if rx.changed().await.is_err() {
3439 // Sender dropped, will never be cancelled
3440 std::future::pending::<()>().await;
3441 }
3442 }
3443 }
3444 }
3445
3446 /// Returns true if the user has cancelled this tool call.
3447 /// This is useful for checking cancellation state after an operation completes,
3448 /// to determine if the completion was due to user cancellation.
3449 pub fn was_cancelled_by_user(&self) -> bool {
3450 *self.cancellation_rx.clone().borrow()
3451 }
3452
3453 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
3454 &self.tool_use_id
3455 }
3456
3457 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
3458 self.stream
3459 .update_tool_call_fields(&self.tool_use_id, fields, None);
3460 }
3461
3462 pub fn update_fields_with_meta(
3463 &self,
3464 fields: acp::ToolCallUpdateFields,
3465 meta: Option<acp::Meta>,
3466 ) {
3467 self.stream
3468 .update_tool_call_fields(&self.tool_use_id, fields, meta);
3469 }
3470
3471 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
3472 self.stream
3473 .0
3474 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3475 acp_thread::ToolCallUpdateDiff {
3476 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3477 diff,
3478 }
3479 .into(),
3480 )))
3481 .ok();
3482 }
3483
3484 pub fn subagent_spawned(&self, id: acp::SessionId) {
3485 self.stream
3486 .0
3487 .unbounded_send(Ok(ThreadEvent::SubagentSpawned(id)))
3488 .ok();
3489 }
3490
3491 /// Authorize a third-party tool (e.g., MCP tool from a context server).
3492 ///
3493 /// Unlike built-in tools, third-party tools don't support pattern-based permissions.
3494 /// They only support `default` (allow/deny/confirm) per tool.
3495 ///
3496 /// Uses the dropdown authorization flow with two granularities:
3497 /// - "Always for <display_name> MCP tool" → sets `tools.<tool_id>.default = "allow"` or "deny"
3498 /// - "Only this time" → allow/deny once
3499 pub fn authorize_third_party_tool(
3500 &self,
3501 title: impl Into<String>,
3502 tool_id: String,
3503 display_name: String,
3504 cx: &mut App,
3505 ) -> Task<Result<()>> {
3506 let settings = agent_settings::AgentSettings::get_global(cx);
3507
3508 let decision = decide_permission_from_settings(&tool_id, &[String::new()], &settings);
3509
3510 match decision {
3511 ToolPermissionDecision::Allow => return Task::ready(Ok(())),
3512 ToolPermissionDecision::Deny(reason) => return Task::ready(Err(anyhow!(reason))),
3513 ToolPermissionDecision::Confirm => {}
3514 }
3515
3516 let (response_tx, response_rx) = oneshot::channel();
3517 if let Err(error) = self
3518 .stream
3519 .0
3520 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3521 ToolCallAuthorization {
3522 tool_call: acp::ToolCallUpdate::new(
3523 self.tool_use_id.to_string(),
3524 acp::ToolCallUpdateFields::new().title(title.into()),
3525 ),
3526 options: acp_thread::PermissionOptions::Dropdown(vec![
3527 acp_thread::PermissionOptionChoice {
3528 allow: acp::PermissionOption::new(
3529 acp::PermissionOptionId::new(format!(
3530 "always_allow_mcp:{}",
3531 tool_id
3532 )),
3533 format!("Always for {} MCP tool", display_name),
3534 acp::PermissionOptionKind::AllowAlways,
3535 ),
3536 deny: acp::PermissionOption::new(
3537 acp::PermissionOptionId::new(format!(
3538 "always_deny_mcp:{}",
3539 tool_id
3540 )),
3541 format!("Always for {} MCP tool", display_name),
3542 acp::PermissionOptionKind::RejectAlways,
3543 ),
3544 },
3545 acp_thread::PermissionOptionChoice {
3546 allow: acp::PermissionOption::new(
3547 acp::PermissionOptionId::new("allow"),
3548 "Only this time",
3549 acp::PermissionOptionKind::AllowOnce,
3550 ),
3551 deny: acp::PermissionOption::new(
3552 acp::PermissionOptionId::new("deny"),
3553 "Only this time",
3554 acp::PermissionOptionKind::RejectOnce,
3555 ),
3556 },
3557 ]),
3558 response: response_tx,
3559 context: None,
3560 },
3561 )))
3562 {
3563 log::error!("Failed to send tool call authorization: {error}");
3564 return Task::ready(Err(anyhow!(
3565 "Failed to send tool call authorization: {error}"
3566 )));
3567 }
3568
3569 let fs = self.fs.clone();
3570 cx.spawn(async move |cx| {
3571 let response_str = response_rx.await?.0.to_string();
3572
3573 if response_str == format!("always_allow_mcp:{}", tool_id) {
3574 if let Some(fs) = fs.clone() {
3575 cx.update(|cx| {
3576 update_settings_file(fs, cx, move |settings, _| {
3577 settings
3578 .agent
3579 .get_or_insert_default()
3580 .set_tool_default_permission(&tool_id, ToolPermissionMode::Allow);
3581 });
3582 });
3583 }
3584 return Ok(());
3585 }
3586 if response_str == format!("always_deny_mcp:{}", tool_id) {
3587 if let Some(fs) = fs.clone() {
3588 cx.update(|cx| {
3589 update_settings_file(fs, cx, move |settings, _| {
3590 settings
3591 .agent
3592 .get_or_insert_default()
3593 .set_tool_default_permission(&tool_id, ToolPermissionMode::Deny);
3594 });
3595 });
3596 }
3597 return Err(anyhow!("Permission to run tool denied by user"));
3598 }
3599
3600 if response_str == "allow" {
3601 return Ok(());
3602 }
3603
3604 Err(anyhow!("Permission to run tool denied by user"))
3605 })
3606 }
3607
3608 pub fn authorize(
3609 &self,
3610 title: impl Into<String>,
3611 context: ToolPermissionContext,
3612 cx: &mut App,
3613 ) -> Task<Result<()>> {
3614 use settings::ToolPermissionMode;
3615
3616 let options = context.build_permission_options();
3617
3618 let (response_tx, response_rx) = oneshot::channel();
3619 if let Err(error) = self
3620 .stream
3621 .0
3622 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3623 ToolCallAuthorization {
3624 tool_call: acp::ToolCallUpdate::new(
3625 self.tool_use_id.to_string(),
3626 acp::ToolCallUpdateFields::new().title(title.into()),
3627 ),
3628 options,
3629 response: response_tx,
3630 context: Some(context),
3631 },
3632 )))
3633 {
3634 log::error!("Failed to send tool call authorization: {error}");
3635 return Task::ready(Err(anyhow!(
3636 "Failed to send tool call authorization: {error}"
3637 )));
3638 }
3639
3640 let fs = self.fs.clone();
3641 cx.spawn(async move |cx| {
3642 let response_str = response_rx.await?.0.to_string();
3643
3644 // Handle "always allow tool" - e.g., "always_allow:terminal"
3645 if let Some(tool) = response_str.strip_prefix("always_allow:") {
3646 if let Some(fs) = fs.clone() {
3647 let tool = tool.to_string();
3648 cx.update(|cx| {
3649 update_settings_file(fs, cx, move |settings, _| {
3650 settings
3651 .agent
3652 .get_or_insert_default()
3653 .set_tool_default_permission(&tool, ToolPermissionMode::Allow);
3654 });
3655 });
3656 }
3657 return Ok(());
3658 }
3659
3660 // Handle "always deny tool" - e.g., "always_deny:terminal"
3661 if let Some(tool) = response_str.strip_prefix("always_deny:") {
3662 if let Some(fs) = fs.clone() {
3663 let tool = tool.to_string();
3664 cx.update(|cx| {
3665 update_settings_file(fs, cx, move |settings, _| {
3666 settings
3667 .agent
3668 .get_or_insert_default()
3669 .set_tool_default_permission(&tool, ToolPermissionMode::Deny);
3670 });
3671 });
3672 }
3673 return Err(anyhow!("Permission to run tool denied by user"));
3674 }
3675
3676 // Handle "always allow pattern" - e.g., "always_allow_pattern:mcp:server:tool\n^cargo\s"
3677 if let Some(rest) = response_str.strip_prefix("always_allow_pattern:") {
3678 if let Some((pattern_tool_name, pattern)) = rest.split_once('\n') {
3679 let pattern_tool_name = pattern_tool_name.to_string();
3680 let pattern = pattern.to_string();
3681 if let Some(fs) = fs.clone() {
3682 cx.update(|cx| {
3683 update_settings_file(fs, cx, move |settings, _| {
3684 settings
3685 .agent
3686 .get_or_insert_default()
3687 .add_tool_allow_pattern(&pattern_tool_name, pattern);
3688 });
3689 });
3690 }
3691 } else {
3692 log::error!("Failed to parse always allow pattern: missing newline separator in '{rest}'");
3693 }
3694 return Ok(());
3695 }
3696
3697 // Handle "always deny pattern" - e.g., "always_deny_pattern:mcp:server:tool\n^cargo\s"
3698 if let Some(rest) = response_str.strip_prefix("always_deny_pattern:") {
3699 if let Some((pattern_tool_name, pattern)) = rest.split_once('\n') {
3700 let pattern_tool_name = pattern_tool_name.to_string();
3701 let pattern = pattern.to_string();
3702 if let Some(fs) = fs.clone() {
3703 cx.update(|cx| {
3704 update_settings_file(fs, cx, move |settings, _| {
3705 settings
3706 .agent
3707 .get_or_insert_default()
3708 .add_tool_deny_pattern(&pattern_tool_name, pattern);
3709 });
3710 });
3711 }
3712 } else {
3713 log::error!("Failed to parse always deny pattern: missing newline separator in '{rest}'");
3714 }
3715 return Err(anyhow!("Permission to run tool denied by user"));
3716 }
3717
3718 // Handle simple "allow" (allow once)
3719 if response_str == "allow" {
3720 return Ok(());
3721 }
3722
3723 // Handle simple "deny" (deny once)
3724 Err(anyhow!("Permission to run tool denied by user"))
3725 })
3726 }
3727}
3728
3729#[cfg(any(test, feature = "test-support"))]
3730pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
3731
3732#[cfg(any(test, feature = "test-support"))]
3733impl ToolCallEventStreamReceiver {
3734 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
3735 let event = self.0.next().await;
3736 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
3737 auth
3738 } else {
3739 panic!("Expected ToolCallAuthorization but got: {:?}", event);
3740 }
3741 }
3742
3743 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
3744 let event = self.0.next().await;
3745 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
3746 update,
3747 )))) = event
3748 {
3749 update.fields
3750 } else {
3751 panic!("Expected update fields but got: {:?}", event);
3752 }
3753 }
3754
3755 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
3756 let event = self.0.next().await;
3757 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
3758 update,
3759 )))) = event
3760 {
3761 update.diff
3762 } else {
3763 panic!("Expected diff but got: {:?}", event);
3764 }
3765 }
3766
3767 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
3768 let event = self.0.next().await;
3769 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
3770 update,
3771 )))) = event
3772 {
3773 update.terminal
3774 } else {
3775 panic!("Expected terminal but got: {:?}", event);
3776 }
3777 }
3778}
3779
3780#[cfg(any(test, feature = "test-support"))]
3781impl std::ops::Deref for ToolCallEventStreamReceiver {
3782 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
3783
3784 fn deref(&self) -> &Self::Target {
3785 &self.0
3786 }
3787}
3788
3789#[cfg(any(test, feature = "test-support"))]
3790impl std::ops::DerefMut for ToolCallEventStreamReceiver {
3791 fn deref_mut(&mut self) -> &mut Self::Target {
3792 &mut self.0
3793 }
3794}
3795
3796impl From<&str> for UserMessageContent {
3797 fn from(text: &str) -> Self {
3798 Self::Text(text.into())
3799 }
3800}
3801
3802impl From<String> for UserMessageContent {
3803 fn from(text: String) -> Self {
3804 Self::Text(text)
3805 }
3806}
3807
3808impl UserMessageContent {
3809 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
3810 match value {
3811 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
3812 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
3813 acp::ContentBlock::Audio(_) => {
3814 // TODO
3815 Self::Text("[audio]".to_string())
3816 }
3817 acp::ContentBlock::ResourceLink(resource_link) => {
3818 match MentionUri::parse(&resource_link.uri, path_style) {
3819 Ok(uri) => Self::Mention {
3820 uri,
3821 content: String::new(),
3822 },
3823 Err(err) => {
3824 log::error!("Failed to parse mention link: {}", err);
3825 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
3826 }
3827 }
3828 }
3829 acp::ContentBlock::Resource(resource) => match resource.resource {
3830 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
3831 match MentionUri::parse(&resource.uri, path_style) {
3832 Ok(uri) => Self::Mention {
3833 uri,
3834 content: resource.text,
3835 },
3836 Err(err) => {
3837 log::error!("Failed to parse mention link: {}", err);
3838 Self::Text(
3839 MarkdownCodeBlock {
3840 tag: &resource.uri,
3841 text: &resource.text,
3842 }
3843 .to_string(),
3844 )
3845 }
3846 }
3847 }
3848 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
3849 // TODO
3850 Self::Text("[blob]".to_string())
3851 }
3852 other => {
3853 log::warn!("Unexpected content type: {:?}", other);
3854 Self::Text("[unknown]".to_string())
3855 }
3856 },
3857 other => {
3858 log::warn!("Unexpected content type: {:?}", other);
3859 Self::Text("[unknown]".to_string())
3860 }
3861 }
3862 }
3863}
3864
3865impl From<UserMessageContent> for acp::ContentBlock {
3866 fn from(content: UserMessageContent) -> Self {
3867 match content {
3868 UserMessageContent::Text(text) => text.into(),
3869 UserMessageContent::Image(image) => {
3870 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
3871 }
3872 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
3873 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
3874 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
3875 )),
3876 ),
3877 }
3878 }
3879}
3880
3881fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
3882 LanguageModelImage {
3883 source: image_content.data.into(),
3884 size: None,
3885 }
3886}
3887
3888#[cfg(test)]
3889mod tests {
3890 use super::*;
3891 use gpui::TestAppContext;
3892 use language_model::LanguageModelToolUseId;
3893 use language_model::fake_provider::FakeLanguageModel;
3894 use serde_json::json;
3895 use std::sync::Arc;
3896
3897 async fn setup_thread_for_test(cx: &mut TestAppContext) -> (Entity<Thread>, ThreadEventStream) {
3898 cx.update(|cx| {
3899 let settings_store = settings::SettingsStore::test(cx);
3900 cx.set_global(settings_store);
3901 });
3902
3903 let fs = fs::FakeFs::new(cx.background_executor.clone());
3904 let templates = Templates::new();
3905 let project = Project::test(fs.clone(), [], cx).await;
3906
3907 cx.update(|cx| {
3908 let project_context = cx.new(|_cx| prompt_store::ProjectContext::default());
3909 let context_server_store = project.read(cx).context_server_store();
3910 let context_server_registry =
3911 cx.new(|cx| ContextServerRegistry::new(context_server_store, cx));
3912
3913 let thread = cx.new(|cx| {
3914 Thread::new(
3915 project,
3916 project_context,
3917 context_server_registry,
3918 templates,
3919 None,
3920 cx,
3921 )
3922 });
3923
3924 let (event_tx, _event_rx) = mpsc::unbounded();
3925 let event_stream = ThreadEventStream(event_tx);
3926
3927 (thread, event_stream)
3928 })
3929 }
3930
3931 fn setup_parent_with_subagents(
3932 cx: &mut TestAppContext,
3933 parent: &Entity<Thread>,
3934 count: usize,
3935 ) -> Vec<Entity<Thread>> {
3936 cx.update(|cx| {
3937 let mut subagents = Vec::new();
3938 for _ in 0..count {
3939 let subagent = cx.new(|cx| Thread::new_subagent(parent, cx));
3940 parent.update(cx, |thread, _cx| {
3941 thread.register_running_subagent(subagent.downgrade());
3942 });
3943 subagents.push(subagent);
3944 }
3945 subagents
3946 })
3947 }
3948
3949 #[gpui::test]
3950 async fn test_set_model_propagates_to_subagents(cx: &mut TestAppContext) {
3951 let (parent, _event_stream) = setup_thread_for_test(cx).await;
3952 let subagents = setup_parent_with_subagents(cx, &parent, 2);
3953
3954 let new_model: Arc<dyn LanguageModel> = Arc::new(FakeLanguageModel::with_id_and_thinking(
3955 "test-provider",
3956 "new-model",
3957 "New Model",
3958 false,
3959 ));
3960
3961 cx.update(|cx| {
3962 parent.update(cx, |thread, cx| {
3963 thread.set_model(new_model, cx);
3964 });
3965
3966 for subagent in &subagents {
3967 let subagent_model_id = subagent.read(cx).model().unwrap().id();
3968 assert_eq!(
3969 subagent_model_id.0.as_ref(),
3970 "new-model",
3971 "Subagent model should match parent model after set_model"
3972 );
3973 }
3974 });
3975 }
3976
3977 #[gpui::test]
3978 async fn test_set_summarization_model_propagates_to_subagents(cx: &mut TestAppContext) {
3979 let (parent, _event_stream) = setup_thread_for_test(cx).await;
3980 let subagents = setup_parent_with_subagents(cx, &parent, 2);
3981
3982 let summary_model: Arc<dyn LanguageModel> =
3983 Arc::new(FakeLanguageModel::with_id_and_thinking(
3984 "test-provider",
3985 "summary-model",
3986 "Summary Model",
3987 false,
3988 ));
3989
3990 cx.update(|cx| {
3991 parent.update(cx, |thread, cx| {
3992 thread.set_summarization_model(Some(summary_model), cx);
3993 });
3994
3995 for subagent in &subagents {
3996 let subagent_summary_id = subagent.read(cx).summarization_model().unwrap().id();
3997 assert_eq!(
3998 subagent_summary_id.0.as_ref(),
3999 "summary-model",
4000 "Subagent summarization model should match parent after set_summarization_model"
4001 );
4002 }
4003 });
4004 }
4005
4006 #[gpui::test]
4007 async fn test_set_thinking_enabled_propagates_to_subagents(cx: &mut TestAppContext) {
4008 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4009 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4010
4011 cx.update(|cx| {
4012 parent.update(cx, |thread, cx| {
4013 thread.set_thinking_enabled(true, cx);
4014 });
4015
4016 for subagent in &subagents {
4017 assert!(
4018 subagent.read(cx).thinking_enabled(),
4019 "Subagent thinking should be enabled after parent enables it"
4020 );
4021 }
4022
4023 parent.update(cx, |thread, cx| {
4024 thread.set_thinking_enabled(false, cx);
4025 });
4026
4027 for subagent in &subagents {
4028 assert!(
4029 !subagent.read(cx).thinking_enabled(),
4030 "Subagent thinking should be disabled after parent disables it"
4031 );
4032 }
4033 });
4034 }
4035
4036 #[gpui::test]
4037 async fn test_set_thinking_effort_propagates_to_subagents(cx: &mut TestAppContext) {
4038 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4039 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4040
4041 cx.update(|cx| {
4042 parent.update(cx, |thread, cx| {
4043 thread.set_thinking_effort(Some("high".to_string()), cx);
4044 });
4045
4046 for subagent in &subagents {
4047 assert_eq!(
4048 subagent.read(cx).thinking_effort().map(|s| s.as_str()),
4049 Some("high"),
4050 "Subagent thinking effort should match parent"
4051 );
4052 }
4053
4054 parent.update(cx, |thread, cx| {
4055 thread.set_thinking_effort(None, cx);
4056 });
4057
4058 for subagent in &subagents {
4059 assert_eq!(
4060 subagent.read(cx).thinking_effort(),
4061 None,
4062 "Subagent thinking effort should be None after parent clears it"
4063 );
4064 }
4065 });
4066 }
4067
4068 #[gpui::test]
4069 async fn test_set_speed_propagates_to_subagents(cx: &mut TestAppContext) {
4070 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4071 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4072
4073 cx.update(|cx| {
4074 parent.update(cx, |thread, cx| {
4075 thread.set_speed(Speed::Fast, cx);
4076 });
4077
4078 for subagent in &subagents {
4079 assert_eq!(
4080 subagent.read(cx).speed(),
4081 Some(Speed::Fast),
4082 "Subagent speed should match parent after set_speed"
4083 );
4084 }
4085 });
4086 }
4087
4088 #[gpui::test]
4089 async fn test_dropped_subagent_does_not_panic(cx: &mut TestAppContext) {
4090 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4091 let subagents = setup_parent_with_subagents(cx, &parent, 1);
4092
4093 // Drop the subagent so the WeakEntity can no longer be upgraded
4094 drop(subagents);
4095
4096 // Should not panic even though the subagent was dropped
4097 cx.update(|cx| {
4098 parent.update(cx, |thread, cx| {
4099 thread.set_thinking_enabled(true, cx);
4100 thread.set_speed(Speed::Fast, cx);
4101 thread.set_thinking_effort(Some("high".to_string()), cx);
4102 });
4103 });
4104 }
4105
4106 #[gpui::test]
4107 async fn test_handle_tool_use_json_parse_error_adds_tool_use_to_content(
4108 cx: &mut TestAppContext,
4109 ) {
4110 let (thread, event_stream) = setup_thread_for_test(cx).await;
4111
4112 cx.update(|cx| {
4113 thread.update(cx, |thread, _cx| {
4114 let tool_use_id = LanguageModelToolUseId::from("test_tool_id");
4115 let tool_name: Arc<str> = Arc::from("test_tool");
4116 let raw_input: Arc<str> = Arc::from("{invalid json");
4117 let json_parse_error = "expected value at line 1 column 1".to_string();
4118
4119 // Call the function under test
4120 let result = thread.handle_tool_use_json_parse_error_event(
4121 tool_use_id.clone(),
4122 tool_name.clone(),
4123 raw_input.clone(),
4124 json_parse_error,
4125 &event_stream,
4126 );
4127
4128 // Verify the result is an error
4129 assert!(result.is_error);
4130 assert_eq!(result.tool_use_id, tool_use_id);
4131 assert_eq!(result.tool_name, tool_name);
4132 assert!(matches!(
4133 result.content,
4134 LanguageModelToolResultContent::Text(_)
4135 ));
4136
4137 // Verify the tool use was added to the message content
4138 {
4139 let last_message = thread.pending_message();
4140 assert_eq!(
4141 last_message.content.len(),
4142 1,
4143 "Should have one tool_use in content"
4144 );
4145
4146 match &last_message.content[0] {
4147 AgentMessageContent::ToolUse(tool_use) => {
4148 assert_eq!(tool_use.id, tool_use_id);
4149 assert_eq!(tool_use.name, tool_name);
4150 assert_eq!(tool_use.raw_input, raw_input.to_string());
4151 assert!(tool_use.is_input_complete);
4152 // Should fall back to empty object for invalid JSON
4153 assert_eq!(tool_use.input, json!({}));
4154 }
4155 _ => panic!("Expected ToolUse content"),
4156 }
4157 }
4158
4159 // Insert the tool result (simulating what the caller does)
4160 thread
4161 .pending_message()
4162 .tool_results
4163 .insert(result.tool_use_id.clone(), result);
4164
4165 // Verify the tool result was added
4166 let last_message = thread.pending_message();
4167 assert_eq!(
4168 last_message.tool_results.len(),
4169 1,
4170 "Should have one tool_result"
4171 );
4172 assert!(last_message.tool_results.contains_key(&tool_use_id));
4173 });
4174 });
4175 }
4176}