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 language_registry,
1456 ));
1457 self.add_tool(FetchTool::new(self.project.read(cx).client().http_client()));
1458 self.add_tool(FindPathTool::new(self.project.clone()));
1459 self.add_tool(GrepTool::new(self.project.clone()));
1460 self.add_tool(ListDirectoryTool::new(self.project.clone()));
1461 self.add_tool(MovePathTool::new(self.project.clone()));
1462 self.add_tool(NowTool);
1463 self.add_tool(OpenTool::new(self.project.clone()));
1464 self.add_tool(ReadFileTool::new(
1465 self.project.clone(),
1466 self.action_log.clone(),
1467 update_agent_location,
1468 ));
1469 self.add_tool(SaveFileTool::new(self.project.clone()));
1470 self.add_tool(RestoreFileFromDiskTool::new(self.project.clone()));
1471 self.add_tool(TerminalTool::new(self.project.clone(), environment.clone()));
1472 self.add_tool(WebSearchTool);
1473
1474 if self.depth() < MAX_SUBAGENT_DEPTH {
1475 self.add_tool(SpawnAgentTool::new(environment));
1476 }
1477 }
1478
1479 pub fn add_tool<T: AgentTool>(&mut self, tool: T) {
1480 debug_assert!(
1481 !self.tools.contains_key(T::NAME),
1482 "Duplicate tool name: {}",
1483 T::NAME,
1484 );
1485 self.tools.insert(T::NAME.into(), tool.erase());
1486 }
1487
1488 #[cfg(any(test, feature = "test-support"))]
1489 pub fn remove_tool(&mut self, name: &str) -> bool {
1490 self.tools.remove(name).is_some()
1491 }
1492
1493 pub fn profile(&self) -> &AgentProfileId {
1494 &self.profile_id
1495 }
1496
1497 pub fn set_profile(&mut self, profile_id: AgentProfileId, cx: &mut Context<Self>) {
1498 if self.profile_id == profile_id {
1499 return;
1500 }
1501
1502 self.profile_id = profile_id.clone();
1503
1504 // Swap to the profile's preferred model when available.
1505 if let Some(model) = Self::resolve_profile_model(&self.profile_id, cx) {
1506 self.set_model(model, cx);
1507 }
1508
1509 for subagent in &self.running_subagents {
1510 subagent
1511 .update(cx, |thread, cx| thread.set_profile(profile_id.clone(), cx))
1512 .ok();
1513 }
1514 }
1515
1516 pub fn cancel(&mut self, cx: &mut Context<Self>) -> Task<()> {
1517 for subagent in self.running_subagents.drain(..) {
1518 if let Some(subagent) = subagent.upgrade() {
1519 subagent.update(cx, |thread, cx| thread.cancel(cx)).detach();
1520 }
1521 }
1522
1523 let Some(running_turn) = self.running_turn.take() else {
1524 self.flush_pending_message(cx);
1525 return Task::ready(());
1526 };
1527
1528 let turn_task = running_turn.cancel();
1529
1530 cx.spawn(async move |this, cx| {
1531 turn_task.await;
1532 this.update(cx, |this, cx| {
1533 this.flush_pending_message(cx);
1534 })
1535 .ok();
1536 })
1537 }
1538
1539 pub fn set_has_queued_message(&mut self, has_queued: bool) {
1540 self.has_queued_message = has_queued;
1541 }
1542
1543 pub fn has_queued_message(&self) -> bool {
1544 self.has_queued_message
1545 }
1546
1547 fn update_token_usage(&mut self, update: language_model::TokenUsage, cx: &mut Context<Self>) {
1548 let Some(last_user_message) = self.last_user_message() else {
1549 return;
1550 };
1551
1552 self.request_token_usage
1553 .insert(last_user_message.id.clone(), update);
1554 cx.emit(TokenUsageUpdated(self.latest_token_usage()));
1555 cx.notify();
1556 }
1557
1558 pub fn truncate(&mut self, message_id: UserMessageId, cx: &mut Context<Self>) -> Result<()> {
1559 self.cancel(cx).detach();
1560 // Clear pending message since cancel will try to flush it asynchronously,
1561 // and we don't want that content to be added after we truncate
1562 self.pending_message.take();
1563 let Some(position) = self.messages.iter().position(
1564 |msg| matches!(msg, Message::User(UserMessage { id, .. }) if id == &message_id),
1565 ) else {
1566 return Err(anyhow!("Message not found"));
1567 };
1568
1569 for message in self.messages.drain(position..) {
1570 match message {
1571 Message::User(message) => {
1572 self.request_token_usage.remove(&message.id);
1573 }
1574 Message::Agent(_) | Message::Resume => {}
1575 }
1576 }
1577 self.clear_summary();
1578 cx.notify();
1579 Ok(())
1580 }
1581
1582 pub fn latest_request_token_usage(&self) -> Option<language_model::TokenUsage> {
1583 let last_user_message = self.last_user_message()?;
1584 let tokens = self.request_token_usage.get(&last_user_message.id)?;
1585 Some(*tokens)
1586 }
1587
1588 pub fn latest_token_usage(&self) -> Option<acp_thread::TokenUsage> {
1589 let usage = self.latest_request_token_usage()?;
1590 let model = self.model.clone()?;
1591 Some(acp_thread::TokenUsage {
1592 max_tokens: model.max_token_count(),
1593 max_output_tokens: model.max_output_tokens(),
1594 used_tokens: usage.total_tokens(),
1595 input_tokens: usage.input_tokens,
1596 output_tokens: usage.output_tokens,
1597 })
1598 }
1599
1600 /// Get the total input token count as of the message before the given message.
1601 ///
1602 /// Returns `None` if:
1603 /// - `target_id` is the first message (no previous message)
1604 /// - The previous message hasn't received a response yet (no usage data)
1605 /// - `target_id` is not found in the messages
1606 pub fn tokens_before_message(&self, target_id: &UserMessageId) -> Option<u64> {
1607 let mut previous_user_message_id: Option<&UserMessageId> = None;
1608
1609 for message in &self.messages {
1610 if let Message::User(user_msg) = message {
1611 if &user_msg.id == target_id {
1612 let prev_id = previous_user_message_id?;
1613 let usage = self.request_token_usage.get(prev_id)?;
1614 return Some(usage.input_tokens);
1615 }
1616 previous_user_message_id = Some(&user_msg.id);
1617 }
1618 }
1619 None
1620 }
1621
1622 /// Look up the active profile and resolve its preferred model if one is configured.
1623 fn resolve_profile_model(
1624 profile_id: &AgentProfileId,
1625 cx: &mut Context<Self>,
1626 ) -> Option<Arc<dyn LanguageModel>> {
1627 let selection = AgentSettings::get_global(cx)
1628 .profiles
1629 .get(profile_id)?
1630 .default_model
1631 .clone()?;
1632 Self::resolve_model_from_selection(&selection, cx)
1633 }
1634
1635 /// Translate a stored model selection into the configured model from the registry.
1636 fn resolve_model_from_selection(
1637 selection: &LanguageModelSelection,
1638 cx: &mut Context<Self>,
1639 ) -> Option<Arc<dyn LanguageModel>> {
1640 let selected = SelectedModel {
1641 provider: LanguageModelProviderId::from(selection.provider.0.clone()),
1642 model: LanguageModelId::from(selection.model.clone()),
1643 };
1644 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
1645 registry
1646 .select_model(&selected, cx)
1647 .map(|configured| configured.model)
1648 })
1649 }
1650
1651 pub fn resume(
1652 &mut self,
1653 cx: &mut Context<Self>,
1654 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1655 self.messages.push(Message::Resume);
1656 cx.notify();
1657
1658 log::debug!("Total messages in thread: {}", self.messages.len());
1659 self.run_turn(cx)
1660 }
1661
1662 /// Sending a message results in the model streaming a response, which could include tool calls.
1663 /// After calling tools, the model will stops and waits for any outstanding tool calls to be completed and their results sent.
1664 /// The returned channel will report all the occurrences in which the model stops before erroring or ending its turn.
1665 pub fn send<T>(
1666 &mut self,
1667 id: UserMessageId,
1668 content: impl IntoIterator<Item = T>,
1669 cx: &mut Context<Self>,
1670 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>>
1671 where
1672 T: Into<UserMessageContent>,
1673 {
1674 let content = content.into_iter().map(Into::into).collect::<Vec<_>>();
1675 log::debug!("Thread::send content: {:?}", content);
1676
1677 self.messages
1678 .push(Message::User(UserMessage { id, content }));
1679 cx.notify();
1680
1681 self.send_existing(cx)
1682 }
1683
1684 pub fn send_existing(
1685 &mut self,
1686 cx: &mut Context<Self>,
1687 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1688 let model = self.model().context("No language model configured")?;
1689
1690 log::info!("Thread::send called with model: {}", model.name().0);
1691 self.advance_prompt_id();
1692
1693 log::debug!("Total messages in thread: {}", self.messages.len());
1694 self.run_turn(cx)
1695 }
1696
1697 pub fn push_acp_user_block(
1698 &mut self,
1699 id: UserMessageId,
1700 blocks: impl IntoIterator<Item = acp::ContentBlock>,
1701 path_style: PathStyle,
1702 cx: &mut Context<Self>,
1703 ) {
1704 let content = blocks
1705 .into_iter()
1706 .map(|block| UserMessageContent::from_content_block(block, path_style))
1707 .collect::<Vec<_>>();
1708 self.messages
1709 .push(Message::User(UserMessage { id, content }));
1710 cx.notify();
1711 }
1712
1713 pub fn push_acp_agent_block(&mut self, block: acp::ContentBlock, cx: &mut Context<Self>) {
1714 let text = match block {
1715 acp::ContentBlock::Text(text_content) => text_content.text,
1716 acp::ContentBlock::Image(_) => "[image]".to_string(),
1717 acp::ContentBlock::Audio(_) => "[audio]".to_string(),
1718 acp::ContentBlock::ResourceLink(resource_link) => resource_link.uri,
1719 acp::ContentBlock::Resource(resource) => match resource.resource {
1720 acp::EmbeddedResourceResource::TextResourceContents(resource) => resource.uri,
1721 acp::EmbeddedResourceResource::BlobResourceContents(resource) => resource.uri,
1722 _ => "[resource]".to_string(),
1723 },
1724 _ => "[unknown]".to_string(),
1725 };
1726
1727 self.messages.push(Message::Agent(AgentMessage {
1728 content: vec![AgentMessageContent::Text(text)],
1729 ..Default::default()
1730 }));
1731 cx.notify();
1732 }
1733
1734 #[cfg(feature = "eval")]
1735 pub fn proceed(
1736 &mut self,
1737 cx: &mut Context<Self>,
1738 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1739 self.run_turn(cx)
1740 }
1741
1742 fn run_turn(
1743 &mut self,
1744 cx: &mut Context<Self>,
1745 ) -> Result<mpsc::UnboundedReceiver<Result<ThreadEvent>>> {
1746 // Flush the old pending message synchronously before cancelling,
1747 // to avoid a race where the detached cancel task might flush the NEW
1748 // turn's pending message instead of the old one.
1749 self.flush_pending_message(cx);
1750 self.cancel(cx).detach();
1751
1752 let model = self.model.clone().context("No language model configured")?;
1753 let profile = AgentSettings::get_global(cx)
1754 .profiles
1755 .get(&self.profile_id)
1756 .context("Profile not found")?;
1757 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
1758 let event_stream = ThreadEventStream(events_tx);
1759 let message_ix = self.messages.len().saturating_sub(1);
1760 self.clear_summary();
1761 let (cancellation_tx, mut cancellation_rx) = watch::channel(false);
1762 self.running_turn = Some(RunningTurn {
1763 event_stream: event_stream.clone(),
1764 tools: self.enabled_tools(profile, &model, cx),
1765 cancellation_tx,
1766 streaming_tool_inputs: HashMap::default(),
1767 _task: cx.spawn(async move |this, cx| {
1768 log::debug!("Starting agent turn execution");
1769
1770 let turn_result = Self::run_turn_internal(
1771 &this,
1772 model,
1773 &event_stream,
1774 cancellation_rx.clone(),
1775 cx,
1776 )
1777 .await;
1778
1779 // Check if we were cancelled - if so, cancel() already took running_turn
1780 // and we shouldn't touch it (it might be a NEW turn now)
1781 let was_cancelled = *cancellation_rx.borrow();
1782 if was_cancelled {
1783 log::debug!("Turn was cancelled, skipping cleanup");
1784 return;
1785 }
1786
1787 _ = this.update(cx, |this, cx| this.flush_pending_message(cx));
1788
1789 match turn_result {
1790 Ok(()) => {
1791 log::debug!("Turn execution completed");
1792 event_stream.send_stop(acp::StopReason::EndTurn);
1793 }
1794 Err(error) => {
1795 log::error!("Turn execution failed: {:?}", error);
1796 match error.downcast::<CompletionError>() {
1797 Ok(CompletionError::Refusal) => {
1798 event_stream.send_stop(acp::StopReason::Refusal);
1799 _ = this.update(cx, |this, _| this.messages.truncate(message_ix));
1800 }
1801 Ok(CompletionError::MaxTokens) => {
1802 event_stream.send_stop(acp::StopReason::MaxTokens);
1803 }
1804 Ok(CompletionError::Other(error)) | Err(error) => {
1805 event_stream.send_error(error);
1806 }
1807 }
1808 }
1809 }
1810
1811 _ = this.update(cx, |this, _| this.running_turn.take());
1812 }),
1813 });
1814 Ok(events_rx)
1815 }
1816
1817 async fn run_turn_internal(
1818 this: &WeakEntity<Self>,
1819 model: Arc<dyn LanguageModel>,
1820 event_stream: &ThreadEventStream,
1821 mut cancellation_rx: watch::Receiver<bool>,
1822 cx: &mut AsyncApp,
1823 ) -> Result<()> {
1824 let mut attempt = 0;
1825 let mut intent = CompletionIntent::UserPrompt;
1826 loop {
1827 let request =
1828 this.update(cx, |this, cx| this.build_completion_request(intent, cx))??;
1829
1830 telemetry::event!(
1831 "Agent Thread Completion",
1832 thread_id = this.read_with(cx, |this, _| this.id.to_string())?,
1833 parent_thread_id = this.read_with(cx, |this, _| this
1834 .parent_thread_id()
1835 .map(|id| id.to_string()))?,
1836 prompt_id = this.read_with(cx, |this, _| this.prompt_id.to_string())?,
1837 model = model.telemetry_id(),
1838 model_provider = model.provider_id().to_string(),
1839 attempt
1840 );
1841
1842 log::debug!("Calling model.stream_completion, attempt {}", attempt);
1843
1844 let (mut events, mut error) = match model.stream_completion(request, cx).await {
1845 Ok(events) => (events.fuse(), None),
1846 Err(err) => (stream::empty().boxed().fuse(), Some(err)),
1847 };
1848 let mut tool_results = FuturesUnordered::new();
1849 let mut cancelled = false;
1850 loop {
1851 // Race between getting the first event and cancellation
1852 let first_event = futures::select! {
1853 event = events.next().fuse() => event,
1854 _ = cancellation_rx.changed().fuse() => {
1855 if *cancellation_rx.borrow() {
1856 cancelled = true;
1857 break;
1858 }
1859 continue;
1860 }
1861 };
1862 let Some(first_event) = first_event else {
1863 break;
1864 };
1865
1866 // Collect all immediately available events to process as a batch
1867 let mut batch = vec![first_event];
1868 while let Some(event) = events.next().now_or_never().flatten() {
1869 batch.push(event);
1870 }
1871
1872 // Process the batch in a single update
1873 let batch_result = this.update(cx, |this, cx| {
1874 let mut batch_tool_results = Vec::new();
1875 let mut batch_error = None;
1876
1877 for event in batch {
1878 log::trace!("Received completion event: {:?}", event);
1879 match event {
1880 Ok(event) => {
1881 match this.handle_completion_event(
1882 event,
1883 event_stream,
1884 cancellation_rx.clone(),
1885 cx,
1886 ) {
1887 Ok(Some(task)) => batch_tool_results.push(task),
1888 Ok(None) => {}
1889 Err(err) => {
1890 batch_error = Some(err);
1891 break;
1892 }
1893 }
1894 }
1895 Err(err) => {
1896 batch_error = Some(err.into());
1897 break;
1898 }
1899 }
1900 }
1901
1902 cx.notify();
1903 (batch_tool_results, batch_error)
1904 })?;
1905
1906 tool_results.extend(batch_result.0);
1907 if let Some(err) = batch_result.1 {
1908 error = Some(err.downcast()?);
1909 break;
1910 }
1911 }
1912
1913 // Drop the stream to release the rate limit permit before tool execution.
1914 // The stream holds a semaphore guard that limits concurrent requests.
1915 // Without this, the permit would be held during potentially long-running
1916 // tool execution, which could cause deadlocks when tools spawn subagents
1917 // that need their own permits.
1918 drop(events);
1919
1920 let end_turn = tool_results.is_empty();
1921 while let Some(tool_result) = tool_results.next().await {
1922 log::debug!("Tool finished {:?}", tool_result);
1923
1924 event_stream.update_tool_call_fields(
1925 &tool_result.tool_use_id,
1926 acp::ToolCallUpdateFields::new()
1927 .status(if tool_result.is_error {
1928 acp::ToolCallStatus::Failed
1929 } else {
1930 acp::ToolCallStatus::Completed
1931 })
1932 .raw_output(tool_result.output.clone()),
1933 None,
1934 );
1935 this.update(cx, |this, _cx| {
1936 this.pending_message()
1937 .tool_results
1938 .insert(tool_result.tool_use_id.clone(), tool_result);
1939 })?;
1940 }
1941
1942 this.update(cx, |this, cx| {
1943 this.flush_pending_message(cx);
1944 if this.title.is_none() && this.pending_title_generation.is_none() {
1945 this.generate_title(cx);
1946 }
1947 })?;
1948
1949 if cancelled {
1950 log::debug!("Turn cancelled by user, exiting");
1951 return Ok(());
1952 }
1953
1954 if let Some(error) = error {
1955 attempt += 1;
1956 let retry = this.update(cx, |this, cx| {
1957 let user_store = this.user_store.read(cx);
1958 this.handle_completion_error(error, attempt, user_store.plan())
1959 })??;
1960 let timer = cx.background_executor().timer(retry.duration);
1961 event_stream.send_retry(retry);
1962 futures::select! {
1963 _ = timer.fuse() => {}
1964 _ = cancellation_rx.changed().fuse() => {
1965 if *cancellation_rx.borrow() {
1966 log::debug!("Turn cancelled during retry delay, exiting");
1967 return Ok(());
1968 }
1969 }
1970 }
1971 this.update(cx, |this, _cx| {
1972 if let Some(Message::Agent(message)) = this.messages.last() {
1973 if message.tool_results.is_empty() {
1974 intent = CompletionIntent::UserPrompt;
1975 this.messages.push(Message::Resume);
1976 }
1977 }
1978 })?;
1979 } else if end_turn {
1980 return Ok(());
1981 } else {
1982 let has_queued = this.update(cx, |this, _| this.has_queued_message())?;
1983 if has_queued {
1984 log::debug!("Queued message found, ending turn at message boundary");
1985 return Ok(());
1986 }
1987 intent = CompletionIntent::ToolResults;
1988 attempt = 0;
1989 }
1990 }
1991 }
1992
1993 fn handle_completion_error(
1994 &mut self,
1995 error: LanguageModelCompletionError,
1996 attempt: u8,
1997 plan: Option<Plan>,
1998 ) -> Result<acp_thread::RetryStatus> {
1999 let Some(model) = self.model.as_ref() else {
2000 return Err(anyhow!(error));
2001 };
2002
2003 let auto_retry = if model.provider_id() == ZED_CLOUD_PROVIDER_ID {
2004 plan.is_some()
2005 } else {
2006 true
2007 };
2008
2009 if !auto_retry {
2010 return Err(anyhow!(error));
2011 }
2012
2013 let Some(strategy) = Self::retry_strategy_for(&error) else {
2014 return Err(anyhow!(error));
2015 };
2016
2017 let max_attempts = match &strategy {
2018 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2019 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2020 };
2021
2022 if attempt > max_attempts {
2023 return Err(anyhow!(error));
2024 }
2025
2026 let delay = match &strategy {
2027 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2028 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2029 Duration::from_secs(delay_secs)
2030 }
2031 RetryStrategy::Fixed { delay, .. } => *delay,
2032 };
2033 log::debug!("Retry attempt {attempt} with delay {delay:?}");
2034
2035 Ok(acp_thread::RetryStatus {
2036 last_error: error.to_string().into(),
2037 attempt: attempt as usize,
2038 max_attempts: max_attempts as usize,
2039 started_at: Instant::now(),
2040 duration: delay,
2041 })
2042 }
2043
2044 /// A helper method that's called on every streamed completion event.
2045 /// Returns an optional tool result task, which the main agentic loop will
2046 /// send back to the model when it resolves.
2047 fn handle_completion_event(
2048 &mut self,
2049 event: LanguageModelCompletionEvent,
2050 event_stream: &ThreadEventStream,
2051 cancellation_rx: watch::Receiver<bool>,
2052 cx: &mut Context<Self>,
2053 ) -> Result<Option<Task<LanguageModelToolResult>>> {
2054 log::trace!("Handling streamed completion event: {:?}", event);
2055 use LanguageModelCompletionEvent::*;
2056
2057 match event {
2058 StartMessage { .. } => {
2059 self.flush_pending_message(cx);
2060 self.pending_message = Some(AgentMessage::default());
2061 }
2062 Text(new_text) => self.handle_text_event(new_text, event_stream),
2063 Thinking { text, signature } => {
2064 self.handle_thinking_event(text, signature, event_stream)
2065 }
2066 RedactedThinking { data } => self.handle_redacted_thinking_event(data),
2067 ReasoningDetails(details) => {
2068 let last_message = self.pending_message();
2069 // Store the last non-empty reasoning_details (overwrites earlier ones)
2070 // This ensures we keep the encrypted reasoning with signatures, not the early text reasoning
2071 if let serde_json::Value::Array(ref arr) = details {
2072 if !arr.is_empty() {
2073 last_message.reasoning_details = Some(details);
2074 }
2075 } else {
2076 last_message.reasoning_details = Some(details);
2077 }
2078 }
2079 ToolUse(tool_use) => {
2080 return Ok(self.handle_tool_use_event(tool_use, event_stream, cancellation_rx, cx));
2081 }
2082 ToolUseJsonParseError {
2083 id,
2084 tool_name,
2085 raw_input,
2086 json_parse_error,
2087 } => {
2088 return Ok(Some(Task::ready(
2089 self.handle_tool_use_json_parse_error_event(
2090 id,
2091 tool_name,
2092 raw_input,
2093 json_parse_error,
2094 event_stream,
2095 ),
2096 )));
2097 }
2098 UsageUpdate(usage) => {
2099 telemetry::event!(
2100 "Agent Thread Completion Usage Updated",
2101 thread_id = self.id.to_string(),
2102 parent_thread_id = self.parent_thread_id().map(|id| id.to_string()),
2103 prompt_id = self.prompt_id.to_string(),
2104 model = self.model.as_ref().map(|m| m.telemetry_id()),
2105 model_provider = self.model.as_ref().map(|m| m.provider_id().to_string()),
2106 input_tokens = usage.input_tokens,
2107 output_tokens = usage.output_tokens,
2108 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2109 cache_read_input_tokens = usage.cache_read_input_tokens,
2110 );
2111 self.update_token_usage(usage, cx);
2112 }
2113 Stop(StopReason::Refusal) => return Err(CompletionError::Refusal.into()),
2114 Stop(StopReason::MaxTokens) => return Err(CompletionError::MaxTokens.into()),
2115 Stop(StopReason::ToolUse | StopReason::EndTurn) => {}
2116 Started | Queued { .. } => {}
2117 }
2118
2119 Ok(None)
2120 }
2121
2122 fn handle_text_event(&mut self, new_text: String, event_stream: &ThreadEventStream) {
2123 event_stream.send_text(&new_text);
2124
2125 let last_message = self.pending_message();
2126 if let Some(AgentMessageContent::Text(text)) = last_message.content.last_mut() {
2127 text.push_str(&new_text);
2128 } else {
2129 last_message
2130 .content
2131 .push(AgentMessageContent::Text(new_text));
2132 }
2133 }
2134
2135 fn handle_thinking_event(
2136 &mut self,
2137 new_text: String,
2138 new_signature: Option<String>,
2139 event_stream: &ThreadEventStream,
2140 ) {
2141 event_stream.send_thinking(&new_text);
2142
2143 let last_message = self.pending_message();
2144 if let Some(AgentMessageContent::Thinking { text, signature }) =
2145 last_message.content.last_mut()
2146 {
2147 text.push_str(&new_text);
2148 *signature = new_signature.or(signature.take());
2149 } else {
2150 last_message.content.push(AgentMessageContent::Thinking {
2151 text: new_text,
2152 signature: new_signature,
2153 });
2154 }
2155 }
2156
2157 fn handle_redacted_thinking_event(&mut self, data: String) {
2158 let last_message = self.pending_message();
2159 last_message
2160 .content
2161 .push(AgentMessageContent::RedactedThinking(data));
2162 }
2163
2164 fn handle_tool_use_event(
2165 &mut self,
2166 tool_use: LanguageModelToolUse,
2167 event_stream: &ThreadEventStream,
2168 cancellation_rx: watch::Receiver<bool>,
2169 cx: &mut Context<Self>,
2170 ) -> Option<Task<LanguageModelToolResult>> {
2171 cx.notify();
2172
2173 let tool = self.tool(tool_use.name.as_ref());
2174 let mut title = SharedString::from(&tool_use.name);
2175 let mut kind = acp::ToolKind::Other;
2176 if let Some(tool) = tool.as_ref() {
2177 title = tool.initial_title(tool_use.input.clone(), cx);
2178 kind = tool.kind();
2179 }
2180
2181 self.send_or_update_tool_use(&tool_use, title, kind, event_stream);
2182
2183 let Some(tool) = tool else {
2184 let content = format!("No tool named {} exists", tool_use.name);
2185 return Some(Task::ready(LanguageModelToolResult {
2186 content: LanguageModelToolResultContent::Text(Arc::from(content)),
2187 tool_use_id: tool_use.id,
2188 tool_name: tool_use.name,
2189 is_error: true,
2190 output: None,
2191 }));
2192 };
2193
2194 if !tool_use.is_input_complete {
2195 if tool.supports_input_streaming() {
2196 let running_turn = self.running_turn.as_mut()?;
2197 if let Some(sender) = running_turn.streaming_tool_inputs.get(&tool_use.id) {
2198 sender.send_partial(tool_use.input);
2199 return None;
2200 }
2201
2202 let (sender, tool_input) = ToolInputSender::channel();
2203 sender.send_partial(tool_use.input);
2204 running_turn
2205 .streaming_tool_inputs
2206 .insert(tool_use.id.clone(), sender);
2207
2208 let tool = tool.clone();
2209 log::debug!("Running streaming tool {}", tool_use.name);
2210 return Some(self.run_tool(
2211 tool,
2212 tool_input,
2213 tool_use.id,
2214 tool_use.name,
2215 event_stream,
2216 cancellation_rx,
2217 cx,
2218 ));
2219 } else {
2220 return None;
2221 }
2222 }
2223
2224 if let Some(sender) = self
2225 .running_turn
2226 .as_mut()?
2227 .streaming_tool_inputs
2228 .remove(&tool_use.id)
2229 {
2230 sender.send_final(tool_use.input);
2231 return None;
2232 }
2233
2234 log::debug!("Running tool {}", tool_use.name);
2235 let tool_input = ToolInput::ready(tool_use.input);
2236 Some(self.run_tool(
2237 tool,
2238 tool_input,
2239 tool_use.id,
2240 tool_use.name,
2241 event_stream,
2242 cancellation_rx,
2243 cx,
2244 ))
2245 }
2246
2247 fn run_tool(
2248 &self,
2249 tool: Arc<dyn AnyAgentTool>,
2250 tool_input: ToolInput<serde_json::Value>,
2251 tool_use_id: LanguageModelToolUseId,
2252 tool_name: Arc<str>,
2253 event_stream: &ThreadEventStream,
2254 cancellation_rx: watch::Receiver<bool>,
2255 cx: &mut Context<Self>,
2256 ) -> Task<LanguageModelToolResult> {
2257 let fs = self.project.read(cx).fs().clone();
2258 let tool_event_stream = ToolCallEventStream::new(
2259 tool_use_id.clone(),
2260 event_stream.clone(),
2261 Some(fs),
2262 cancellation_rx,
2263 );
2264 tool_event_stream.update_fields(
2265 acp::ToolCallUpdateFields::new().status(acp::ToolCallStatus::InProgress),
2266 );
2267 let supports_images = self.model().is_some_and(|model| model.supports_images());
2268 let tool_result = tool.run(tool_input, tool_event_stream, cx);
2269 cx.foreground_executor().spawn(async move {
2270 let (is_error, output) = match tool_result.await {
2271 Ok(mut output) => {
2272 if let LanguageModelToolResultContent::Image(_) = &output.llm_output
2273 && !supports_images
2274 {
2275 output = AgentToolOutput::from_error(
2276 "Attempted to read an image, but this model doesn't support it.",
2277 );
2278 (true, output)
2279 } else {
2280 (false, output)
2281 }
2282 }
2283 Err(output) => (true, output),
2284 };
2285
2286 LanguageModelToolResult {
2287 tool_use_id,
2288 tool_name,
2289 is_error,
2290 content: output.llm_output,
2291 output: Some(output.raw_output),
2292 }
2293 })
2294 }
2295
2296 fn handle_tool_use_json_parse_error_event(
2297 &mut self,
2298 tool_use_id: LanguageModelToolUseId,
2299 tool_name: Arc<str>,
2300 raw_input: Arc<str>,
2301 json_parse_error: String,
2302 event_stream: &ThreadEventStream,
2303 ) -> LanguageModelToolResult {
2304 let tool_use = LanguageModelToolUse {
2305 id: tool_use_id.clone(),
2306 name: tool_name.clone(),
2307 raw_input: raw_input.to_string(),
2308 input: serde_json::json!({}),
2309 is_input_complete: true,
2310 thought_signature: None,
2311 };
2312 self.send_or_update_tool_use(
2313 &tool_use,
2314 SharedString::from(&tool_use.name),
2315 acp::ToolKind::Other,
2316 event_stream,
2317 );
2318
2319 let tool_output = format!("Error parsing input JSON: {json_parse_error}");
2320 LanguageModelToolResult {
2321 tool_use_id,
2322 tool_name,
2323 is_error: true,
2324 content: LanguageModelToolResultContent::Text(tool_output.into()),
2325 output: Some(serde_json::Value::String(raw_input.to_string())),
2326 }
2327 }
2328
2329 fn send_or_update_tool_use(
2330 &mut self,
2331 tool_use: &LanguageModelToolUse,
2332 title: SharedString,
2333 kind: acp::ToolKind,
2334 event_stream: &ThreadEventStream,
2335 ) {
2336 // Ensure the last message ends in the current tool use
2337 let last_message = self.pending_message();
2338 let push_new_tool_use = last_message.content.last_mut().is_none_or(|content| {
2339 if let AgentMessageContent::ToolUse(last_tool_use) = content {
2340 if last_tool_use.id == tool_use.id {
2341 *last_tool_use = tool_use.clone();
2342 false
2343 } else {
2344 true
2345 }
2346 } else {
2347 true
2348 }
2349 });
2350
2351 if push_new_tool_use {
2352 event_stream.send_tool_call(
2353 &tool_use.id,
2354 &tool_use.name,
2355 title,
2356 kind,
2357 tool_use.input.clone(),
2358 );
2359 last_message
2360 .content
2361 .push(AgentMessageContent::ToolUse(tool_use.clone()));
2362 } else {
2363 event_stream.update_tool_call_fields(
2364 &tool_use.id,
2365 acp::ToolCallUpdateFields::new()
2366 .title(title.as_str())
2367 .kind(kind)
2368 .raw_input(tool_use.input.clone()),
2369 None,
2370 );
2371 }
2372 }
2373
2374 pub fn title(&self) -> SharedString {
2375 self.title.clone().unwrap_or("New Thread".into())
2376 }
2377
2378 pub fn is_generating_summary(&self) -> bool {
2379 self.pending_summary_generation.is_some()
2380 }
2381
2382 pub fn is_generating_title(&self) -> bool {
2383 self.pending_title_generation.is_some()
2384 }
2385
2386 pub fn summary(&mut self, cx: &mut Context<Self>) -> Shared<Task<Option<SharedString>>> {
2387 if let Some(summary) = self.summary.as_ref() {
2388 return Task::ready(Some(summary.clone())).shared();
2389 }
2390 if let Some(task) = self.pending_summary_generation.clone() {
2391 return task;
2392 }
2393 let Some(model) = self.summarization_model.clone() else {
2394 log::error!("No summarization model available");
2395 return Task::ready(None).shared();
2396 };
2397 let mut request = LanguageModelRequest {
2398 intent: Some(CompletionIntent::ThreadContextSummarization),
2399 temperature: AgentSettings::temperature_for_model(&model, cx),
2400 ..Default::default()
2401 };
2402
2403 for message in &self.messages {
2404 request.messages.extend(message.to_request());
2405 }
2406
2407 request.messages.push(LanguageModelRequestMessage {
2408 role: Role::User,
2409 content: vec![SUMMARIZE_THREAD_DETAILED_PROMPT.into()],
2410 cache: false,
2411 reasoning_details: None,
2412 });
2413
2414 let task = cx
2415 .spawn(async move |this, cx| {
2416 let mut summary = String::new();
2417 let mut messages = model.stream_completion(request, cx).await.log_err()?;
2418 while let Some(event) = messages.next().await {
2419 let event = event.log_err()?;
2420 let text = match event {
2421 LanguageModelCompletionEvent::Text(text) => text,
2422 _ => continue,
2423 };
2424
2425 let mut lines = text.lines();
2426 summary.extend(lines.next());
2427 }
2428
2429 log::debug!("Setting summary: {}", summary);
2430 let summary = SharedString::from(summary);
2431
2432 this.update(cx, |this, cx| {
2433 this.summary = Some(summary.clone());
2434 this.pending_summary_generation = None;
2435 cx.notify()
2436 })
2437 .ok()?;
2438
2439 Some(summary)
2440 })
2441 .shared();
2442 self.pending_summary_generation = Some(task.clone());
2443 task
2444 }
2445
2446 pub fn generate_title(&mut self, cx: &mut Context<Self>) {
2447 let Some(model) = self.summarization_model.clone() else {
2448 return;
2449 };
2450
2451 log::debug!(
2452 "Generating title with model: {:?}",
2453 self.summarization_model.as_ref().map(|model| model.name())
2454 );
2455 let mut request = LanguageModelRequest {
2456 intent: Some(CompletionIntent::ThreadSummarization),
2457 temperature: AgentSettings::temperature_for_model(&model, cx),
2458 ..Default::default()
2459 };
2460
2461 for message in &self.messages {
2462 request.messages.extend(message.to_request());
2463 }
2464
2465 request.messages.push(LanguageModelRequestMessage {
2466 role: Role::User,
2467 content: vec![SUMMARIZE_THREAD_PROMPT.into()],
2468 cache: false,
2469 reasoning_details: None,
2470 });
2471 self.pending_title_generation = Some(cx.spawn(async move |this, cx| {
2472 let mut title = String::new();
2473
2474 let generate = async {
2475 let mut messages = model.stream_completion(request, cx).await?;
2476 while let Some(event) = messages.next().await {
2477 let event = event?;
2478 let text = match event {
2479 LanguageModelCompletionEvent::Text(text) => text,
2480 _ => continue,
2481 };
2482
2483 let mut lines = text.lines();
2484 title.extend(lines.next());
2485
2486 // Stop if the LLM generated multiple lines.
2487 if lines.next().is_some() {
2488 break;
2489 }
2490 }
2491 anyhow::Ok(())
2492 };
2493
2494 if generate
2495 .await
2496 .context("failed to generate thread title")
2497 .log_err()
2498 .is_some()
2499 {
2500 _ = this.update(cx, |this, cx| this.set_title(title.into(), cx));
2501 }
2502 _ = this.update(cx, |this, _| this.pending_title_generation = None);
2503 }));
2504 }
2505
2506 pub fn set_title(&mut self, title: SharedString, cx: &mut Context<Self>) {
2507 self.pending_title_generation = None;
2508 if Some(&title) != self.title.as_ref() {
2509 self.title = Some(title);
2510 cx.emit(TitleUpdated);
2511 cx.notify();
2512 }
2513 }
2514
2515 fn clear_summary(&mut self) {
2516 self.summary = None;
2517 self.pending_summary_generation = None;
2518 }
2519
2520 fn last_user_message(&self) -> Option<&UserMessage> {
2521 self.messages
2522 .iter()
2523 .rev()
2524 .find_map(|message| match message {
2525 Message::User(user_message) => Some(user_message),
2526 Message::Agent(_) => None,
2527 Message::Resume => None,
2528 })
2529 }
2530
2531 fn pending_message(&mut self) -> &mut AgentMessage {
2532 self.pending_message.get_or_insert_default()
2533 }
2534
2535 fn flush_pending_message(&mut self, cx: &mut Context<Self>) {
2536 let Some(mut message) = self.pending_message.take() else {
2537 return;
2538 };
2539
2540 if message.content.is_empty() {
2541 return;
2542 }
2543
2544 for content in &message.content {
2545 let AgentMessageContent::ToolUse(tool_use) = content else {
2546 continue;
2547 };
2548
2549 if !message.tool_results.contains_key(&tool_use.id) {
2550 message.tool_results.insert(
2551 tool_use.id.clone(),
2552 LanguageModelToolResult {
2553 tool_use_id: tool_use.id.clone(),
2554 tool_name: tool_use.name.clone(),
2555 is_error: true,
2556 content: LanguageModelToolResultContent::Text(TOOL_CANCELED_MESSAGE.into()),
2557 output: None,
2558 },
2559 );
2560 }
2561 }
2562
2563 self.messages.push(Message::Agent(message));
2564 self.updated_at = Utc::now();
2565 self.clear_summary();
2566 cx.notify()
2567 }
2568
2569 pub(crate) fn build_completion_request(
2570 &self,
2571 completion_intent: CompletionIntent,
2572 cx: &App,
2573 ) -> Result<LanguageModelRequest> {
2574 let model = self.model().context("No language model configured")?;
2575 let tools = if let Some(turn) = self.running_turn.as_ref() {
2576 turn.tools
2577 .iter()
2578 .filter_map(|(tool_name, tool)| {
2579 log::trace!("Including tool: {}", tool_name);
2580 Some(LanguageModelRequestTool {
2581 name: tool_name.to_string(),
2582 description: tool.description().to_string(),
2583 input_schema: tool.input_schema(model.tool_input_format()).log_err()?,
2584 use_input_streaming: tool.supports_input_streaming(),
2585 })
2586 })
2587 .collect::<Vec<_>>()
2588 } else {
2589 Vec::new()
2590 };
2591
2592 log::debug!("Building completion request");
2593 log::debug!("Completion intent: {:?}", completion_intent);
2594
2595 let available_tools: Vec<_> = self
2596 .running_turn
2597 .as_ref()
2598 .map(|turn| turn.tools.keys().cloned().collect())
2599 .unwrap_or_default();
2600
2601 log::debug!("Request includes {} tools", available_tools.len());
2602 let messages = self.build_request_messages(available_tools, cx);
2603 log::debug!("Request will include {} messages", messages.len());
2604
2605 let request = LanguageModelRequest {
2606 thread_id: Some(self.id.to_string()),
2607 prompt_id: Some(self.prompt_id.to_string()),
2608 intent: Some(completion_intent),
2609 messages,
2610 tools,
2611 tool_choice: None,
2612 stop: Vec::new(),
2613 temperature: AgentSettings::temperature_for_model(model, cx),
2614 thinking_allowed: self.thinking_enabled,
2615 thinking_effort: self.thinking_effort.clone(),
2616 speed: self.speed(),
2617 };
2618
2619 log::debug!("Completion request built successfully");
2620 Ok(request)
2621 }
2622
2623 fn enabled_tools(
2624 &self,
2625 profile: &AgentProfileSettings,
2626 model: &Arc<dyn LanguageModel>,
2627 cx: &App,
2628 ) -> BTreeMap<SharedString, Arc<dyn AnyAgentTool>> {
2629 fn truncate(tool_name: &SharedString) -> SharedString {
2630 if tool_name.len() > MAX_TOOL_NAME_LENGTH {
2631 let mut truncated = tool_name.to_string();
2632 truncated.truncate(MAX_TOOL_NAME_LENGTH);
2633 truncated.into()
2634 } else {
2635 tool_name.clone()
2636 }
2637 }
2638
2639 let use_streaming_edit_tool =
2640 cx.has_flag::<StreamingEditFileToolFeatureFlag>() && model.supports_streaming_tools();
2641
2642 let mut tools = self
2643 .tools
2644 .iter()
2645 .filter_map(|(tool_name, tool)| {
2646 // For streaming_edit_file, check profile against "edit_file" since that's what users configure
2647 let profile_tool_name = if tool_name == StreamingEditFileTool::NAME {
2648 EditFileTool::NAME
2649 } else {
2650 tool_name.as_ref()
2651 };
2652
2653 if tool.supports_provider(&model.provider_id())
2654 && profile.is_tool_enabled(profile_tool_name)
2655 {
2656 match (tool_name.as_ref(), use_streaming_edit_tool) {
2657 (StreamingEditFileTool::NAME, false) | (EditFileTool::NAME, true) => None,
2658 (StreamingEditFileTool::NAME, true) => {
2659 // Expose streaming tool as "edit_file"
2660 Some((SharedString::from(EditFileTool::NAME), tool.clone()))
2661 }
2662 _ => Some((truncate(tool_name), tool.clone())),
2663 }
2664 } else {
2665 None
2666 }
2667 })
2668 .collect::<BTreeMap<_, _>>();
2669
2670 let mut context_server_tools = Vec::new();
2671 let mut seen_tools = tools.keys().cloned().collect::<HashSet<_>>();
2672 let mut duplicate_tool_names = HashSet::default();
2673 for (server_id, server_tools) in self.context_server_registry.read(cx).servers() {
2674 for (tool_name, tool) in server_tools {
2675 if profile.is_context_server_tool_enabled(&server_id.0, &tool_name) {
2676 let tool_name = truncate(tool_name);
2677 if !seen_tools.insert(tool_name.clone()) {
2678 duplicate_tool_names.insert(tool_name.clone());
2679 }
2680 context_server_tools.push((server_id.clone(), tool_name, tool.clone()));
2681 }
2682 }
2683 }
2684
2685 // When there are duplicate tool names, disambiguate by prefixing them
2686 // with the server ID (converted to snake_case for API compatibility).
2687 // In the rare case there isn't enough space for the disambiguated tool
2688 // name, keep only the last tool with this name.
2689 for (server_id, tool_name, tool) in context_server_tools {
2690 if duplicate_tool_names.contains(&tool_name) {
2691 let available = MAX_TOOL_NAME_LENGTH.saturating_sub(tool_name.len());
2692 if available >= 2 {
2693 let mut disambiguated = server_id.0.to_snake_case();
2694 disambiguated.truncate(available - 1);
2695 disambiguated.push('_');
2696 disambiguated.push_str(&tool_name);
2697 tools.insert(disambiguated.into(), tool.clone());
2698 } else {
2699 tools.insert(tool_name, tool.clone());
2700 }
2701 } else {
2702 tools.insert(tool_name, tool.clone());
2703 }
2704 }
2705
2706 tools
2707 }
2708
2709 fn tool(&self, name: &str) -> Option<Arc<dyn AnyAgentTool>> {
2710 self.running_turn.as_ref()?.tools.get(name).cloned()
2711 }
2712
2713 pub fn has_tool(&self, name: &str) -> bool {
2714 self.running_turn
2715 .as_ref()
2716 .is_some_and(|turn| turn.tools.contains_key(name))
2717 }
2718
2719 #[cfg(any(test, feature = "test-support"))]
2720 pub fn has_registered_tool(&self, name: &str) -> bool {
2721 self.tools.contains_key(name)
2722 }
2723
2724 pub fn registered_tool_names(&self) -> Vec<SharedString> {
2725 self.tools.keys().cloned().collect()
2726 }
2727
2728 pub(crate) fn register_running_subagent(&mut self, subagent: WeakEntity<Thread>) {
2729 self.running_subagents.push(subagent);
2730 }
2731
2732 pub(crate) fn unregister_running_subagent(
2733 &mut self,
2734 subagent_session_id: &acp::SessionId,
2735 cx: &App,
2736 ) {
2737 self.running_subagents.retain(|s| {
2738 s.upgrade()
2739 .map_or(false, |s| s.read(cx).id() != subagent_session_id)
2740 });
2741 }
2742
2743 #[cfg(any(test, feature = "test-support"))]
2744 pub fn running_subagent_ids(&self, cx: &App) -> Vec<acp::SessionId> {
2745 self.running_subagents
2746 .iter()
2747 .filter_map(|s| s.upgrade().map(|s| s.read(cx).id().clone()))
2748 .collect()
2749 }
2750
2751 pub fn is_subagent(&self) -> bool {
2752 self.subagent_context.is_some()
2753 }
2754
2755 pub fn parent_thread_id(&self) -> Option<acp::SessionId> {
2756 self.subagent_context
2757 .as_ref()
2758 .map(|c| c.parent_thread_id.clone())
2759 }
2760
2761 pub fn depth(&self) -> u8 {
2762 self.subagent_context.as_ref().map(|c| c.depth).unwrap_or(0)
2763 }
2764
2765 #[cfg(any(test, feature = "test-support"))]
2766 pub fn set_subagent_context(&mut self, context: SubagentContext) {
2767 self.subagent_context = Some(context);
2768 }
2769
2770 pub fn is_turn_complete(&self) -> bool {
2771 self.running_turn.is_none()
2772 }
2773
2774 fn build_request_messages(
2775 &self,
2776 available_tools: Vec<SharedString>,
2777 cx: &App,
2778 ) -> Vec<LanguageModelRequestMessage> {
2779 log::trace!(
2780 "Building request messages from {} thread messages",
2781 self.messages.len()
2782 );
2783
2784 let system_prompt = SystemPromptTemplate {
2785 project: self.project_context.read(cx),
2786 available_tools,
2787 model_name: self.model.as_ref().map(|m| m.name().0.to_string()),
2788 }
2789 .render(&self.templates)
2790 .context("failed to build system prompt")
2791 .expect("Invalid template");
2792 let mut messages = vec![LanguageModelRequestMessage {
2793 role: Role::System,
2794 content: vec![system_prompt.into()],
2795 cache: false,
2796 reasoning_details: None,
2797 }];
2798 for message in &self.messages {
2799 messages.extend(message.to_request());
2800 }
2801
2802 if let Some(last_message) = messages.last_mut() {
2803 last_message.cache = true;
2804 }
2805
2806 if let Some(message) = self.pending_message.as_ref() {
2807 messages.extend(message.to_request());
2808 }
2809
2810 messages
2811 }
2812
2813 pub fn to_markdown(&self) -> String {
2814 let mut markdown = String::new();
2815 for (ix, message) in self.messages.iter().enumerate() {
2816 if ix > 0 {
2817 markdown.push('\n');
2818 }
2819 match message {
2820 Message::User(_) => markdown.push_str("## User\n\n"),
2821 Message::Agent(_) => markdown.push_str("## Assistant\n\n"),
2822 Message::Resume => {}
2823 }
2824 markdown.push_str(&message.to_markdown());
2825 }
2826
2827 if let Some(message) = self.pending_message.as_ref() {
2828 markdown.push_str("\n## Assistant\n\n");
2829 markdown.push_str(&message.to_markdown());
2830 }
2831
2832 markdown
2833 }
2834
2835 fn advance_prompt_id(&mut self) {
2836 self.prompt_id = PromptId::new();
2837 }
2838
2839 fn retry_strategy_for(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2840 use LanguageModelCompletionError::*;
2841 use http_client::StatusCode;
2842
2843 // General strategy here:
2844 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2845 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2846 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2847 match error {
2848 HttpResponseError {
2849 status_code: StatusCode::TOO_MANY_REQUESTS,
2850 ..
2851 } => Some(RetryStrategy::ExponentialBackoff {
2852 initial_delay: BASE_RETRY_DELAY,
2853 max_attempts: MAX_RETRY_ATTEMPTS,
2854 }),
2855 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2856 Some(RetryStrategy::Fixed {
2857 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2858 max_attempts: MAX_RETRY_ATTEMPTS,
2859 })
2860 }
2861 UpstreamProviderError {
2862 status,
2863 retry_after,
2864 ..
2865 } => match *status {
2866 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2867 Some(RetryStrategy::Fixed {
2868 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2869 max_attempts: MAX_RETRY_ATTEMPTS,
2870 })
2871 }
2872 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2873 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2874 // Internal Server Error could be anything, retry up to 3 times.
2875 max_attempts: 3,
2876 }),
2877 status => {
2878 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2879 // but we frequently get them in practice. See https://http.dev/529
2880 if status.as_u16() == 529 {
2881 Some(RetryStrategy::Fixed {
2882 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2883 max_attempts: MAX_RETRY_ATTEMPTS,
2884 })
2885 } else {
2886 Some(RetryStrategy::Fixed {
2887 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2888 max_attempts: 2,
2889 })
2890 }
2891 }
2892 },
2893 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2894 delay: BASE_RETRY_DELAY,
2895 max_attempts: 3,
2896 }),
2897 ApiReadResponseError { .. }
2898 | HttpSend { .. }
2899 | DeserializeResponse { .. }
2900 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2901 delay: BASE_RETRY_DELAY,
2902 max_attempts: 3,
2903 }),
2904 // Retrying these errors definitely shouldn't help.
2905 HttpResponseError {
2906 status_code:
2907 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2908 ..
2909 }
2910 | AuthenticationError { .. }
2911 | PermissionError { .. }
2912 | NoApiKey { .. }
2913 | ApiEndpointNotFound { .. }
2914 | PromptTooLarge { .. } => None,
2915 // These errors might be transient, so retry them
2916 SerializeRequest { .. } | BuildRequestBody { .. } | StreamEndedUnexpectedly { .. } => {
2917 Some(RetryStrategy::Fixed {
2918 delay: BASE_RETRY_DELAY,
2919 max_attempts: 1,
2920 })
2921 }
2922 // Retry all other 4xx and 5xx errors once.
2923 HttpResponseError { status_code, .. }
2924 if status_code.is_client_error() || status_code.is_server_error() =>
2925 {
2926 Some(RetryStrategy::Fixed {
2927 delay: BASE_RETRY_DELAY,
2928 max_attempts: 3,
2929 })
2930 }
2931 Other(err) if err.is::<language_model::PaymentRequiredError>() => {
2932 // Retrying won't help for Payment Required errors.
2933 None
2934 }
2935 // Conservatively assume that any other errors are non-retryable
2936 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2937 delay: BASE_RETRY_DELAY,
2938 max_attempts: 2,
2939 }),
2940 }
2941 }
2942}
2943
2944struct RunningTurn {
2945 /// Holds the task that handles agent interaction until the end of the turn.
2946 /// Survives across multiple requests as the model performs tool calls and
2947 /// we run tools, report their results.
2948 _task: Task<()>,
2949 /// The current event stream for the running turn. Used to report a final
2950 /// cancellation event if we cancel the turn.
2951 event_stream: ThreadEventStream,
2952 /// The tools that were enabled for this turn.
2953 tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>>,
2954 /// Sender to signal tool cancellation. When cancel is called, this is
2955 /// set to true so all tools can detect user-initiated cancellation.
2956 cancellation_tx: watch::Sender<bool>,
2957 /// Senders for tools that support input streaming and have already been
2958 /// started but are still receiving input from the LLM.
2959 streaming_tool_inputs: HashMap<LanguageModelToolUseId, ToolInputSender>,
2960}
2961
2962impl RunningTurn {
2963 fn cancel(mut self) -> Task<()> {
2964 log::debug!("Cancelling in progress turn");
2965 self.cancellation_tx.send(true).ok();
2966 self.event_stream.send_canceled();
2967 self._task
2968 }
2969}
2970
2971pub struct TokenUsageUpdated(pub Option<acp_thread::TokenUsage>);
2972
2973impl EventEmitter<TokenUsageUpdated> for Thread {}
2974
2975pub struct TitleUpdated;
2976
2977impl EventEmitter<TitleUpdated> for Thread {}
2978
2979/// A channel-based wrapper that delivers tool input to a running tool.
2980///
2981/// For non-streaming tools, created via `ToolInput::ready()` so `.recv()` resolves immediately.
2982/// For streaming tools, partial JSON snapshots arrive via `.recv_partial()` as the LLM streams
2983/// them, followed by the final complete input available through `.recv()`.
2984pub struct ToolInput<T> {
2985 partial_rx: mpsc::UnboundedReceiver<serde_json::Value>,
2986 final_rx: oneshot::Receiver<serde_json::Value>,
2987 _phantom: PhantomData<T>,
2988}
2989
2990impl<T: DeserializeOwned> ToolInput<T> {
2991 #[cfg(any(test, feature = "test-support"))]
2992 pub fn resolved(input: impl Serialize) -> Self {
2993 let value = serde_json::to_value(input).expect("failed to serialize tool input");
2994 Self::ready(value)
2995 }
2996
2997 pub fn ready(value: serde_json::Value) -> Self {
2998 let (partial_tx, partial_rx) = mpsc::unbounded();
2999 drop(partial_tx);
3000 let (final_tx, final_rx) = oneshot::channel();
3001 final_tx.send(value).ok();
3002 Self {
3003 partial_rx,
3004 final_rx,
3005 _phantom: PhantomData,
3006 }
3007 }
3008
3009 #[cfg(any(test, feature = "test-support"))]
3010 pub fn test() -> (ToolInputSender, Self) {
3011 let (sender, input) = ToolInputSender::channel();
3012 (sender, input.cast())
3013 }
3014
3015 /// Wait for the final deserialized input, ignoring all partial updates.
3016 /// Non-streaming tools can use this to wait until the whole input is available.
3017 pub async fn recv(mut self) -> Result<T> {
3018 // Drain any remaining partials
3019 while self.partial_rx.next().await.is_some() {}
3020 let value = self
3021 .final_rx
3022 .await
3023 .map_err(|_| anyhow!("tool input sender was dropped before sending final input"))?;
3024 serde_json::from_value(value).map_err(Into::into)
3025 }
3026
3027 /// Returns the next partial JSON snapshot, or `None` when input is complete.
3028 /// Once this returns `None`, call `recv()` to get the final input.
3029 pub async fn recv_partial(&mut self) -> Option<serde_json::Value> {
3030 self.partial_rx.next().await
3031 }
3032
3033 fn cast<U: DeserializeOwned>(self) -> ToolInput<U> {
3034 ToolInput {
3035 partial_rx: self.partial_rx,
3036 final_rx: self.final_rx,
3037 _phantom: PhantomData,
3038 }
3039 }
3040}
3041
3042pub struct ToolInputSender {
3043 partial_tx: mpsc::UnboundedSender<serde_json::Value>,
3044 final_tx: Option<oneshot::Sender<serde_json::Value>>,
3045}
3046
3047impl ToolInputSender {
3048 pub(crate) fn channel() -> (Self, ToolInput<serde_json::Value>) {
3049 let (partial_tx, partial_rx) = mpsc::unbounded();
3050 let (final_tx, final_rx) = oneshot::channel();
3051 let sender = Self {
3052 partial_tx,
3053 final_tx: Some(final_tx),
3054 };
3055 let input = ToolInput {
3056 partial_rx,
3057 final_rx,
3058 _phantom: PhantomData,
3059 };
3060 (sender, input)
3061 }
3062
3063 pub(crate) fn send_partial(&self, value: serde_json::Value) {
3064 self.partial_tx.unbounded_send(value).ok();
3065 }
3066
3067 pub(crate) fn send_final(mut self, value: serde_json::Value) {
3068 // Close the partial channel so recv_partial() returns None
3069 self.partial_tx.close_channel();
3070 if let Some(final_tx) = self.final_tx.take() {
3071 final_tx.send(value).ok();
3072 }
3073 }
3074}
3075
3076pub trait AgentTool
3077where
3078 Self: 'static + Sized,
3079{
3080 type Input: for<'de> Deserialize<'de> + Serialize + JsonSchema;
3081 type Output: for<'de> Deserialize<'de> + Serialize + Into<LanguageModelToolResultContent>;
3082
3083 const NAME: &'static str;
3084
3085 fn description() -> SharedString {
3086 let schema = schemars::schema_for!(Self::Input);
3087 SharedString::new(
3088 schema
3089 .get("description")
3090 .and_then(|description| description.as_str())
3091 .unwrap_or_default(),
3092 )
3093 }
3094
3095 fn kind() -> acp::ToolKind;
3096
3097 /// The initial tool title to display. Can be updated during the tool run.
3098 fn initial_title(
3099 &self,
3100 input: Result<Self::Input, serde_json::Value>,
3101 cx: &mut App,
3102 ) -> SharedString;
3103
3104 /// Returns the JSON schema that describes the tool's input.
3105 fn input_schema(format: LanguageModelToolSchemaFormat) -> Schema {
3106 language_model::tool_schema::root_schema_for::<Self::Input>(format)
3107 }
3108
3109 /// Returns whether the tool supports streaming of tool use parameters.
3110 fn supports_input_streaming() -> bool {
3111 false
3112 }
3113
3114 /// Some tools rely on a provider for the underlying billing or other reasons.
3115 /// Allow the tool to check if they are compatible, or should be filtered out.
3116 fn supports_provider(_provider: &LanguageModelProviderId) -> bool {
3117 true
3118 }
3119
3120 /// Runs the tool with the provided input.
3121 ///
3122 /// Returns `Result<Self::Output, Self::Output>` rather than `Result<Self::Output, anyhow::Error>`
3123 /// because tool errors are sent back to the model as tool results. This means error output must
3124 /// be structured and readable by the agent — not an arbitrary `anyhow::Error`. Returning the
3125 /// same `Output` type for both success and failure lets tools provide structured data while
3126 /// still signaling whether the invocation succeeded or failed.
3127 fn run(
3128 self: Arc<Self>,
3129 input: ToolInput<Self::Input>,
3130 event_stream: ToolCallEventStream,
3131 cx: &mut App,
3132 ) -> Task<Result<Self::Output, Self::Output>>;
3133
3134 /// Emits events for a previous execution of the tool.
3135 fn replay(
3136 &self,
3137 _input: Self::Input,
3138 _output: Self::Output,
3139 _event_stream: ToolCallEventStream,
3140 _cx: &mut App,
3141 ) -> Result<()> {
3142 Ok(())
3143 }
3144
3145 fn erase(self) -> Arc<dyn AnyAgentTool> {
3146 Arc::new(Erased(Arc::new(self)))
3147 }
3148}
3149
3150pub struct Erased<T>(T);
3151
3152pub struct AgentToolOutput {
3153 pub llm_output: LanguageModelToolResultContent,
3154 pub raw_output: serde_json::Value,
3155}
3156
3157impl AgentToolOutput {
3158 pub fn from_error(message: impl Into<String>) -> Self {
3159 let message = message.into();
3160 let llm_output = LanguageModelToolResultContent::Text(Arc::from(message.as_str()));
3161 Self {
3162 raw_output: serde_json::Value::String(message),
3163 llm_output,
3164 }
3165 }
3166}
3167
3168pub trait AnyAgentTool {
3169 fn name(&self) -> SharedString;
3170 fn description(&self) -> SharedString;
3171 fn kind(&self) -> acp::ToolKind;
3172 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString;
3173 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value>;
3174 fn supports_input_streaming(&self) -> bool {
3175 false
3176 }
3177 fn supports_provider(&self, _provider: &LanguageModelProviderId) -> bool {
3178 true
3179 }
3180 /// See [`AgentTool::run`] for why this returns `Result<AgentToolOutput, AgentToolOutput>`.
3181 fn run(
3182 self: Arc<Self>,
3183 input: ToolInput<serde_json::Value>,
3184 event_stream: ToolCallEventStream,
3185 cx: &mut App,
3186 ) -> Task<Result<AgentToolOutput, AgentToolOutput>>;
3187 fn replay(
3188 &self,
3189 input: serde_json::Value,
3190 output: serde_json::Value,
3191 event_stream: ToolCallEventStream,
3192 cx: &mut App,
3193 ) -> Result<()>;
3194}
3195
3196impl<T> AnyAgentTool for Erased<Arc<T>>
3197where
3198 T: AgentTool,
3199{
3200 fn name(&self) -> SharedString {
3201 T::NAME.into()
3202 }
3203
3204 fn description(&self) -> SharedString {
3205 T::description()
3206 }
3207
3208 fn kind(&self) -> agent_client_protocol::ToolKind {
3209 T::kind()
3210 }
3211
3212 fn supports_input_streaming(&self) -> bool {
3213 T::supports_input_streaming()
3214 }
3215
3216 fn initial_title(&self, input: serde_json::Value, _cx: &mut App) -> SharedString {
3217 let parsed_input = serde_json::from_value(input.clone()).map_err(|_| input);
3218 self.0.initial_title(parsed_input, _cx)
3219 }
3220
3221 fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
3222 let mut json = serde_json::to_value(T::input_schema(format))?;
3223 language_model::tool_schema::adapt_schema_to_format(&mut json, format)?;
3224 Ok(json)
3225 }
3226
3227 fn supports_provider(&self, provider: &LanguageModelProviderId) -> bool {
3228 T::supports_provider(provider)
3229 }
3230
3231 fn run(
3232 self: Arc<Self>,
3233 input: ToolInput<serde_json::Value>,
3234 event_stream: ToolCallEventStream,
3235 cx: &mut App,
3236 ) -> Task<Result<AgentToolOutput, AgentToolOutput>> {
3237 let tool_input: ToolInput<T::Input> = input.cast();
3238 let task = self.0.clone().run(tool_input, event_stream, cx);
3239 cx.spawn(async move |_cx| match task.await {
3240 Ok(output) => {
3241 let raw_output = serde_json::to_value(&output).map_err(|e| {
3242 AgentToolOutput::from_error(format!("Failed to serialize tool output: {e}"))
3243 })?;
3244 Ok(AgentToolOutput {
3245 llm_output: output.into(),
3246 raw_output,
3247 })
3248 }
3249 Err(error_output) => {
3250 let raw_output = serde_json::to_value(&error_output).unwrap_or_else(|e| {
3251 log::error!("Failed to serialize tool error output: {e}");
3252 serde_json::Value::Null
3253 });
3254 Err(AgentToolOutput {
3255 llm_output: error_output.into(),
3256 raw_output,
3257 })
3258 }
3259 })
3260 }
3261
3262 fn replay(
3263 &self,
3264 input: serde_json::Value,
3265 output: serde_json::Value,
3266 event_stream: ToolCallEventStream,
3267 cx: &mut App,
3268 ) -> Result<()> {
3269 let input = serde_json::from_value(input)?;
3270 let output = serde_json::from_value(output)?;
3271 self.0.replay(input, output, event_stream, cx)
3272 }
3273}
3274
3275#[derive(Clone)]
3276struct ThreadEventStream(mpsc::UnboundedSender<Result<ThreadEvent>>);
3277
3278impl ThreadEventStream {
3279 fn send_user_message(&self, message: &UserMessage) {
3280 self.0
3281 .unbounded_send(Ok(ThreadEvent::UserMessage(message.clone())))
3282 .ok();
3283 }
3284
3285 fn send_text(&self, text: &str) {
3286 self.0
3287 .unbounded_send(Ok(ThreadEvent::AgentText(text.to_string())))
3288 .ok();
3289 }
3290
3291 fn send_thinking(&self, text: &str) {
3292 self.0
3293 .unbounded_send(Ok(ThreadEvent::AgentThinking(text.to_string())))
3294 .ok();
3295 }
3296
3297 fn send_tool_call(
3298 &self,
3299 id: &LanguageModelToolUseId,
3300 tool_name: &str,
3301 title: SharedString,
3302 kind: acp::ToolKind,
3303 input: serde_json::Value,
3304 ) {
3305 self.0
3306 .unbounded_send(Ok(ThreadEvent::ToolCall(Self::initial_tool_call(
3307 id,
3308 tool_name,
3309 title.to_string(),
3310 kind,
3311 input,
3312 ))))
3313 .ok();
3314 }
3315
3316 fn initial_tool_call(
3317 id: &LanguageModelToolUseId,
3318 tool_name: &str,
3319 title: String,
3320 kind: acp::ToolKind,
3321 input: serde_json::Value,
3322 ) -> acp::ToolCall {
3323 acp::ToolCall::new(id.to_string(), title)
3324 .kind(kind)
3325 .raw_input(input)
3326 .meta(acp_thread::meta_with_tool_name(tool_name))
3327 }
3328
3329 fn update_tool_call_fields(
3330 &self,
3331 tool_use_id: &LanguageModelToolUseId,
3332 fields: acp::ToolCallUpdateFields,
3333 meta: Option<acp::Meta>,
3334 ) {
3335 self.0
3336 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3337 acp::ToolCallUpdate::new(tool_use_id.to_string(), fields)
3338 .meta(meta)
3339 .into(),
3340 )))
3341 .ok();
3342 }
3343
3344 fn send_retry(&self, status: acp_thread::RetryStatus) {
3345 self.0.unbounded_send(Ok(ThreadEvent::Retry(status))).ok();
3346 }
3347
3348 fn send_stop(&self, reason: acp::StopReason) {
3349 self.0.unbounded_send(Ok(ThreadEvent::Stop(reason))).ok();
3350 }
3351
3352 fn send_canceled(&self) {
3353 self.0
3354 .unbounded_send(Ok(ThreadEvent::Stop(acp::StopReason::Cancelled)))
3355 .ok();
3356 }
3357
3358 fn send_error(&self, error: impl Into<anyhow::Error>) {
3359 self.0.unbounded_send(Err(error.into())).ok();
3360 }
3361}
3362
3363#[derive(Clone)]
3364pub struct ToolCallEventStream {
3365 tool_use_id: LanguageModelToolUseId,
3366 stream: ThreadEventStream,
3367 fs: Option<Arc<dyn Fs>>,
3368 cancellation_rx: watch::Receiver<bool>,
3369}
3370
3371impl ToolCallEventStream {
3372 #[cfg(any(test, feature = "test-support"))]
3373 pub fn test() -> (Self, ToolCallEventStreamReceiver) {
3374 let (stream, receiver, _cancellation_tx) = Self::test_with_cancellation();
3375 (stream, receiver)
3376 }
3377
3378 #[cfg(any(test, feature = "test-support"))]
3379 pub fn test_with_cancellation() -> (Self, ToolCallEventStreamReceiver, watch::Sender<bool>) {
3380 let (events_tx, events_rx) = mpsc::unbounded::<Result<ThreadEvent>>();
3381 let (cancellation_tx, cancellation_rx) = watch::channel(false);
3382
3383 let stream = ToolCallEventStream::new(
3384 "test_id".into(),
3385 ThreadEventStream(events_tx),
3386 None,
3387 cancellation_rx,
3388 );
3389
3390 (
3391 stream,
3392 ToolCallEventStreamReceiver(events_rx),
3393 cancellation_tx,
3394 )
3395 }
3396
3397 /// Signal cancellation for this event stream. Only available in tests.
3398 #[cfg(any(test, feature = "test-support"))]
3399 pub fn signal_cancellation_with_sender(cancellation_tx: &mut watch::Sender<bool>) {
3400 cancellation_tx.send(true).ok();
3401 }
3402
3403 fn new(
3404 tool_use_id: LanguageModelToolUseId,
3405 stream: ThreadEventStream,
3406 fs: Option<Arc<dyn Fs>>,
3407 cancellation_rx: watch::Receiver<bool>,
3408 ) -> Self {
3409 Self {
3410 tool_use_id,
3411 stream,
3412 fs,
3413 cancellation_rx,
3414 }
3415 }
3416
3417 /// Returns a future that resolves when the user cancels the tool call.
3418 /// Tools should select on this alongside their main work to detect user cancellation.
3419 pub fn cancelled_by_user(&self) -> impl std::future::Future<Output = ()> + '_ {
3420 let mut rx = self.cancellation_rx.clone();
3421 async move {
3422 loop {
3423 if *rx.borrow() {
3424 return;
3425 }
3426 if rx.changed().await.is_err() {
3427 // Sender dropped, will never be cancelled
3428 std::future::pending::<()>().await;
3429 }
3430 }
3431 }
3432 }
3433
3434 /// Returns true if the user has cancelled this tool call.
3435 /// This is useful for checking cancellation state after an operation completes,
3436 /// to determine if the completion was due to user cancellation.
3437 pub fn was_cancelled_by_user(&self) -> bool {
3438 *self.cancellation_rx.clone().borrow()
3439 }
3440
3441 pub fn tool_use_id(&self) -> &LanguageModelToolUseId {
3442 &self.tool_use_id
3443 }
3444
3445 pub fn update_fields(&self, fields: acp::ToolCallUpdateFields) {
3446 self.stream
3447 .update_tool_call_fields(&self.tool_use_id, fields, None);
3448 }
3449
3450 pub fn update_fields_with_meta(
3451 &self,
3452 fields: acp::ToolCallUpdateFields,
3453 meta: Option<acp::Meta>,
3454 ) {
3455 self.stream
3456 .update_tool_call_fields(&self.tool_use_id, fields, meta);
3457 }
3458
3459 pub fn update_diff(&self, diff: Entity<acp_thread::Diff>) {
3460 self.stream
3461 .0
3462 .unbounded_send(Ok(ThreadEvent::ToolCallUpdate(
3463 acp_thread::ToolCallUpdateDiff {
3464 id: acp::ToolCallId::new(self.tool_use_id.to_string()),
3465 diff,
3466 }
3467 .into(),
3468 )))
3469 .ok();
3470 }
3471
3472 pub fn subagent_spawned(&self, id: acp::SessionId) {
3473 self.stream
3474 .0
3475 .unbounded_send(Ok(ThreadEvent::SubagentSpawned(id)))
3476 .ok();
3477 }
3478
3479 /// Authorize a third-party tool (e.g., MCP tool from a context server).
3480 ///
3481 /// Unlike built-in tools, third-party tools don't support pattern-based permissions.
3482 /// They only support `default` (allow/deny/confirm) per tool.
3483 ///
3484 /// Uses the dropdown authorization flow with two granularities:
3485 /// - "Always for <display_name> MCP tool" → sets `tools.<tool_id>.default = "allow"` or "deny"
3486 /// - "Only this time" → allow/deny once
3487 pub fn authorize_third_party_tool(
3488 &self,
3489 title: impl Into<String>,
3490 tool_id: String,
3491 display_name: String,
3492 cx: &mut App,
3493 ) -> Task<Result<()>> {
3494 let settings = agent_settings::AgentSettings::get_global(cx);
3495
3496 let decision = decide_permission_from_settings(&tool_id, &[String::new()], &settings);
3497
3498 match decision {
3499 ToolPermissionDecision::Allow => return Task::ready(Ok(())),
3500 ToolPermissionDecision::Deny(reason) => return Task::ready(Err(anyhow!(reason))),
3501 ToolPermissionDecision::Confirm => {}
3502 }
3503
3504 let (response_tx, response_rx) = oneshot::channel();
3505 if let Err(error) = self
3506 .stream
3507 .0
3508 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3509 ToolCallAuthorization {
3510 tool_call: acp::ToolCallUpdate::new(
3511 self.tool_use_id.to_string(),
3512 acp::ToolCallUpdateFields::new().title(title.into()),
3513 ),
3514 options: acp_thread::PermissionOptions::Dropdown(vec![
3515 acp_thread::PermissionOptionChoice {
3516 allow: acp::PermissionOption::new(
3517 acp::PermissionOptionId::new(format!(
3518 "always_allow_mcp:{}",
3519 tool_id
3520 )),
3521 format!("Always for {} MCP tool", display_name),
3522 acp::PermissionOptionKind::AllowAlways,
3523 ),
3524 deny: acp::PermissionOption::new(
3525 acp::PermissionOptionId::new(format!(
3526 "always_deny_mcp:{}",
3527 tool_id
3528 )),
3529 format!("Always for {} MCP tool", display_name),
3530 acp::PermissionOptionKind::RejectAlways,
3531 ),
3532 },
3533 acp_thread::PermissionOptionChoice {
3534 allow: acp::PermissionOption::new(
3535 acp::PermissionOptionId::new("allow"),
3536 "Only this time",
3537 acp::PermissionOptionKind::AllowOnce,
3538 ),
3539 deny: acp::PermissionOption::new(
3540 acp::PermissionOptionId::new("deny"),
3541 "Only this time",
3542 acp::PermissionOptionKind::RejectOnce,
3543 ),
3544 },
3545 ]),
3546 response: response_tx,
3547 context: None,
3548 },
3549 )))
3550 {
3551 log::error!("Failed to send tool call authorization: {error}");
3552 return Task::ready(Err(anyhow!(
3553 "Failed to send tool call authorization: {error}"
3554 )));
3555 }
3556
3557 let fs = self.fs.clone();
3558 cx.spawn(async move |cx| {
3559 let response_str = response_rx.await?.0.to_string();
3560
3561 if response_str == format!("always_allow_mcp:{}", tool_id) {
3562 if let Some(fs) = fs.clone() {
3563 cx.update(|cx| {
3564 update_settings_file(fs, cx, move |settings, _| {
3565 settings
3566 .agent
3567 .get_or_insert_default()
3568 .set_tool_default_permission(&tool_id, ToolPermissionMode::Allow);
3569 });
3570 });
3571 }
3572 return Ok(());
3573 }
3574 if response_str == format!("always_deny_mcp:{}", tool_id) {
3575 if let Some(fs) = fs.clone() {
3576 cx.update(|cx| {
3577 update_settings_file(fs, cx, move |settings, _| {
3578 settings
3579 .agent
3580 .get_or_insert_default()
3581 .set_tool_default_permission(&tool_id, ToolPermissionMode::Deny);
3582 });
3583 });
3584 }
3585 return Err(anyhow!("Permission to run tool denied by user"));
3586 }
3587
3588 if response_str == "allow" {
3589 return Ok(());
3590 }
3591
3592 Err(anyhow!("Permission to run tool denied by user"))
3593 })
3594 }
3595
3596 pub fn authorize(
3597 &self,
3598 title: impl Into<String>,
3599 context: ToolPermissionContext,
3600 cx: &mut App,
3601 ) -> Task<Result<()>> {
3602 use settings::ToolPermissionMode;
3603
3604 let options = context.build_permission_options();
3605
3606 let (response_tx, response_rx) = oneshot::channel();
3607 if let Err(error) = self
3608 .stream
3609 .0
3610 .unbounded_send(Ok(ThreadEvent::ToolCallAuthorization(
3611 ToolCallAuthorization {
3612 tool_call: acp::ToolCallUpdate::new(
3613 self.tool_use_id.to_string(),
3614 acp::ToolCallUpdateFields::new().title(title.into()),
3615 ),
3616 options,
3617 response: response_tx,
3618 context: Some(context),
3619 },
3620 )))
3621 {
3622 log::error!("Failed to send tool call authorization: {error}");
3623 return Task::ready(Err(anyhow!(
3624 "Failed to send tool call authorization: {error}"
3625 )));
3626 }
3627
3628 let fs = self.fs.clone();
3629 cx.spawn(async move |cx| {
3630 let response_str = response_rx.await?.0.to_string();
3631
3632 // Handle "always allow tool" - e.g., "always_allow:terminal"
3633 if let Some(tool) = response_str.strip_prefix("always_allow:") {
3634 if let Some(fs) = fs.clone() {
3635 let tool = tool.to_string();
3636 cx.update(|cx| {
3637 update_settings_file(fs, cx, move |settings, _| {
3638 settings
3639 .agent
3640 .get_or_insert_default()
3641 .set_tool_default_permission(&tool, ToolPermissionMode::Allow);
3642 });
3643 });
3644 }
3645 return Ok(());
3646 }
3647
3648 // Handle "always deny tool" - e.g., "always_deny:terminal"
3649 if let Some(tool) = response_str.strip_prefix("always_deny:") {
3650 if let Some(fs) = fs.clone() {
3651 let tool = tool.to_string();
3652 cx.update(|cx| {
3653 update_settings_file(fs, cx, move |settings, _| {
3654 settings
3655 .agent
3656 .get_or_insert_default()
3657 .set_tool_default_permission(&tool, ToolPermissionMode::Deny);
3658 });
3659 });
3660 }
3661 return Err(anyhow!("Permission to run tool denied by user"));
3662 }
3663
3664 // Handle "always allow pattern" - e.g., "always_allow_pattern:mcp:server:tool\n^cargo\s"
3665 if let Some(rest) = response_str.strip_prefix("always_allow_pattern:") {
3666 if let Some((pattern_tool_name, pattern)) = rest.split_once('\n') {
3667 let pattern_tool_name = pattern_tool_name.to_string();
3668 let pattern = pattern.to_string();
3669 if let Some(fs) = fs.clone() {
3670 cx.update(|cx| {
3671 update_settings_file(fs, cx, move |settings, _| {
3672 settings
3673 .agent
3674 .get_or_insert_default()
3675 .add_tool_allow_pattern(&pattern_tool_name, pattern);
3676 });
3677 });
3678 }
3679 } else {
3680 log::error!("Failed to parse always allow pattern: missing newline separator in '{rest}'");
3681 }
3682 return Ok(());
3683 }
3684
3685 // Handle "always deny pattern" - e.g., "always_deny_pattern:mcp:server:tool\n^cargo\s"
3686 if let Some(rest) = response_str.strip_prefix("always_deny_pattern:") {
3687 if let Some((pattern_tool_name, pattern)) = rest.split_once('\n') {
3688 let pattern_tool_name = pattern_tool_name.to_string();
3689 let pattern = pattern.to_string();
3690 if let Some(fs) = fs.clone() {
3691 cx.update(|cx| {
3692 update_settings_file(fs, cx, move |settings, _| {
3693 settings
3694 .agent
3695 .get_or_insert_default()
3696 .add_tool_deny_pattern(&pattern_tool_name, pattern);
3697 });
3698 });
3699 }
3700 } else {
3701 log::error!("Failed to parse always deny pattern: missing newline separator in '{rest}'");
3702 }
3703 return Err(anyhow!("Permission to run tool denied by user"));
3704 }
3705
3706 // Handle simple "allow" (allow once)
3707 if response_str == "allow" {
3708 return Ok(());
3709 }
3710
3711 // Handle simple "deny" (deny once)
3712 Err(anyhow!("Permission to run tool denied by user"))
3713 })
3714 }
3715}
3716
3717#[cfg(any(test, feature = "test-support"))]
3718pub struct ToolCallEventStreamReceiver(mpsc::UnboundedReceiver<Result<ThreadEvent>>);
3719
3720#[cfg(any(test, feature = "test-support"))]
3721impl ToolCallEventStreamReceiver {
3722 pub async fn expect_authorization(&mut self) -> ToolCallAuthorization {
3723 let event = self.0.next().await;
3724 if let Some(Ok(ThreadEvent::ToolCallAuthorization(auth))) = event {
3725 auth
3726 } else {
3727 panic!("Expected ToolCallAuthorization but got: {:?}", event);
3728 }
3729 }
3730
3731 pub async fn expect_update_fields(&mut self) -> acp::ToolCallUpdateFields {
3732 let event = self.0.next().await;
3733 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateFields(
3734 update,
3735 )))) = event
3736 {
3737 update.fields
3738 } else {
3739 panic!("Expected update fields but got: {:?}", event);
3740 }
3741 }
3742
3743 pub async fn expect_diff(&mut self) -> Entity<acp_thread::Diff> {
3744 let event = self.0.next().await;
3745 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateDiff(
3746 update,
3747 )))) = event
3748 {
3749 update.diff
3750 } else {
3751 panic!("Expected diff but got: {:?}", event);
3752 }
3753 }
3754
3755 pub async fn expect_terminal(&mut self) -> Entity<acp_thread::Terminal> {
3756 let event = self.0.next().await;
3757 if let Some(Ok(ThreadEvent::ToolCallUpdate(acp_thread::ToolCallUpdate::UpdateTerminal(
3758 update,
3759 )))) = event
3760 {
3761 update.terminal
3762 } else {
3763 panic!("Expected terminal but got: {:?}", event);
3764 }
3765 }
3766}
3767
3768#[cfg(any(test, feature = "test-support"))]
3769impl std::ops::Deref for ToolCallEventStreamReceiver {
3770 type Target = mpsc::UnboundedReceiver<Result<ThreadEvent>>;
3771
3772 fn deref(&self) -> &Self::Target {
3773 &self.0
3774 }
3775}
3776
3777#[cfg(any(test, feature = "test-support"))]
3778impl std::ops::DerefMut for ToolCallEventStreamReceiver {
3779 fn deref_mut(&mut self) -> &mut Self::Target {
3780 &mut self.0
3781 }
3782}
3783
3784impl From<&str> for UserMessageContent {
3785 fn from(text: &str) -> Self {
3786 Self::Text(text.into())
3787 }
3788}
3789
3790impl From<String> for UserMessageContent {
3791 fn from(text: String) -> Self {
3792 Self::Text(text)
3793 }
3794}
3795
3796impl UserMessageContent {
3797 pub fn from_content_block(value: acp::ContentBlock, path_style: PathStyle) -> Self {
3798 match value {
3799 acp::ContentBlock::Text(text_content) => Self::Text(text_content.text),
3800 acp::ContentBlock::Image(image_content) => Self::Image(convert_image(image_content)),
3801 acp::ContentBlock::Audio(_) => {
3802 // TODO
3803 Self::Text("[audio]".to_string())
3804 }
3805 acp::ContentBlock::ResourceLink(resource_link) => {
3806 match MentionUri::parse(&resource_link.uri, path_style) {
3807 Ok(uri) => Self::Mention {
3808 uri,
3809 content: String::new(),
3810 },
3811 Err(err) => {
3812 log::error!("Failed to parse mention link: {}", err);
3813 Self::Text(format!("[{}]({})", resource_link.name, resource_link.uri))
3814 }
3815 }
3816 }
3817 acp::ContentBlock::Resource(resource) => match resource.resource {
3818 acp::EmbeddedResourceResource::TextResourceContents(resource) => {
3819 match MentionUri::parse(&resource.uri, path_style) {
3820 Ok(uri) => Self::Mention {
3821 uri,
3822 content: resource.text,
3823 },
3824 Err(err) => {
3825 log::error!("Failed to parse mention link: {}", err);
3826 Self::Text(
3827 MarkdownCodeBlock {
3828 tag: &resource.uri,
3829 text: &resource.text,
3830 }
3831 .to_string(),
3832 )
3833 }
3834 }
3835 }
3836 acp::EmbeddedResourceResource::BlobResourceContents(_) => {
3837 // TODO
3838 Self::Text("[blob]".to_string())
3839 }
3840 other => {
3841 log::warn!("Unexpected content type: {:?}", other);
3842 Self::Text("[unknown]".to_string())
3843 }
3844 },
3845 other => {
3846 log::warn!("Unexpected content type: {:?}", other);
3847 Self::Text("[unknown]".to_string())
3848 }
3849 }
3850 }
3851}
3852
3853impl From<UserMessageContent> for acp::ContentBlock {
3854 fn from(content: UserMessageContent) -> Self {
3855 match content {
3856 UserMessageContent::Text(text) => text.into(),
3857 UserMessageContent::Image(image) => {
3858 acp::ContentBlock::Image(acp::ImageContent::new(image.source, "image/png"))
3859 }
3860 UserMessageContent::Mention { uri, content } => acp::ContentBlock::Resource(
3861 acp::EmbeddedResource::new(acp::EmbeddedResourceResource::TextResourceContents(
3862 acp::TextResourceContents::new(content, uri.to_uri().to_string()),
3863 )),
3864 ),
3865 }
3866 }
3867}
3868
3869fn convert_image(image_content: acp::ImageContent) -> LanguageModelImage {
3870 LanguageModelImage {
3871 source: image_content.data.into(),
3872 size: None,
3873 }
3874}
3875
3876#[cfg(test)]
3877mod tests {
3878 use super::*;
3879 use gpui::TestAppContext;
3880 use language_model::LanguageModelToolUseId;
3881 use language_model::fake_provider::FakeLanguageModel;
3882 use serde_json::json;
3883 use std::sync::Arc;
3884
3885 async fn setup_thread_for_test(cx: &mut TestAppContext) -> (Entity<Thread>, ThreadEventStream) {
3886 cx.update(|cx| {
3887 let settings_store = settings::SettingsStore::test(cx);
3888 cx.set_global(settings_store);
3889 });
3890
3891 let fs = fs::FakeFs::new(cx.background_executor.clone());
3892 let templates = Templates::new();
3893 let project = Project::test(fs.clone(), [], cx).await;
3894
3895 cx.update(|cx| {
3896 let project_context = cx.new(|_cx| prompt_store::ProjectContext::default());
3897 let context_server_store = project.read(cx).context_server_store();
3898 let context_server_registry =
3899 cx.new(|cx| ContextServerRegistry::new(context_server_store, cx));
3900
3901 let thread = cx.new(|cx| {
3902 Thread::new(
3903 project,
3904 project_context,
3905 context_server_registry,
3906 templates,
3907 None,
3908 cx,
3909 )
3910 });
3911
3912 let (event_tx, _event_rx) = mpsc::unbounded();
3913 let event_stream = ThreadEventStream(event_tx);
3914
3915 (thread, event_stream)
3916 })
3917 }
3918
3919 fn setup_parent_with_subagents(
3920 cx: &mut TestAppContext,
3921 parent: &Entity<Thread>,
3922 count: usize,
3923 ) -> Vec<Entity<Thread>> {
3924 cx.update(|cx| {
3925 let mut subagents = Vec::new();
3926 for _ in 0..count {
3927 let subagent = cx.new(|cx| Thread::new_subagent(parent, cx));
3928 parent.update(cx, |thread, _cx| {
3929 thread.register_running_subagent(subagent.downgrade());
3930 });
3931 subagents.push(subagent);
3932 }
3933 subagents
3934 })
3935 }
3936
3937 #[gpui::test]
3938 async fn test_set_model_propagates_to_subagents(cx: &mut TestAppContext) {
3939 let (parent, _event_stream) = setup_thread_for_test(cx).await;
3940 let subagents = setup_parent_with_subagents(cx, &parent, 2);
3941
3942 let new_model: Arc<dyn LanguageModel> = Arc::new(FakeLanguageModel::with_id_and_thinking(
3943 "test-provider",
3944 "new-model",
3945 "New Model",
3946 false,
3947 ));
3948
3949 cx.update(|cx| {
3950 parent.update(cx, |thread, cx| {
3951 thread.set_model(new_model, cx);
3952 });
3953
3954 for subagent in &subagents {
3955 let subagent_model_id = subagent.read(cx).model().unwrap().id();
3956 assert_eq!(
3957 subagent_model_id.0.as_ref(),
3958 "new-model",
3959 "Subagent model should match parent model after set_model"
3960 );
3961 }
3962 });
3963 }
3964
3965 #[gpui::test]
3966 async fn test_set_summarization_model_propagates_to_subagents(cx: &mut TestAppContext) {
3967 let (parent, _event_stream) = setup_thread_for_test(cx).await;
3968 let subagents = setup_parent_with_subagents(cx, &parent, 2);
3969
3970 let summary_model: Arc<dyn LanguageModel> =
3971 Arc::new(FakeLanguageModel::with_id_and_thinking(
3972 "test-provider",
3973 "summary-model",
3974 "Summary Model",
3975 false,
3976 ));
3977
3978 cx.update(|cx| {
3979 parent.update(cx, |thread, cx| {
3980 thread.set_summarization_model(Some(summary_model), cx);
3981 });
3982
3983 for subagent in &subagents {
3984 let subagent_summary_id = subagent.read(cx).summarization_model().unwrap().id();
3985 assert_eq!(
3986 subagent_summary_id.0.as_ref(),
3987 "summary-model",
3988 "Subagent summarization model should match parent after set_summarization_model"
3989 );
3990 }
3991 });
3992 }
3993
3994 #[gpui::test]
3995 async fn test_set_thinking_enabled_propagates_to_subagents(cx: &mut TestAppContext) {
3996 let (parent, _event_stream) = setup_thread_for_test(cx).await;
3997 let subagents = setup_parent_with_subagents(cx, &parent, 2);
3998
3999 cx.update(|cx| {
4000 parent.update(cx, |thread, cx| {
4001 thread.set_thinking_enabled(true, cx);
4002 });
4003
4004 for subagent in &subagents {
4005 assert!(
4006 subagent.read(cx).thinking_enabled(),
4007 "Subagent thinking should be enabled after parent enables it"
4008 );
4009 }
4010
4011 parent.update(cx, |thread, cx| {
4012 thread.set_thinking_enabled(false, cx);
4013 });
4014
4015 for subagent in &subagents {
4016 assert!(
4017 !subagent.read(cx).thinking_enabled(),
4018 "Subagent thinking should be disabled after parent disables it"
4019 );
4020 }
4021 });
4022 }
4023
4024 #[gpui::test]
4025 async fn test_set_thinking_effort_propagates_to_subagents(cx: &mut TestAppContext) {
4026 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4027 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4028
4029 cx.update(|cx| {
4030 parent.update(cx, |thread, cx| {
4031 thread.set_thinking_effort(Some("high".to_string()), cx);
4032 });
4033
4034 for subagent in &subagents {
4035 assert_eq!(
4036 subagent.read(cx).thinking_effort().map(|s| s.as_str()),
4037 Some("high"),
4038 "Subagent thinking effort should match parent"
4039 );
4040 }
4041
4042 parent.update(cx, |thread, cx| {
4043 thread.set_thinking_effort(None, cx);
4044 });
4045
4046 for subagent in &subagents {
4047 assert_eq!(
4048 subagent.read(cx).thinking_effort(),
4049 None,
4050 "Subagent thinking effort should be None after parent clears it"
4051 );
4052 }
4053 });
4054 }
4055
4056 #[gpui::test]
4057 async fn test_set_speed_propagates_to_subagents(cx: &mut TestAppContext) {
4058 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4059 let subagents = setup_parent_with_subagents(cx, &parent, 2);
4060
4061 cx.update(|cx| {
4062 parent.update(cx, |thread, cx| {
4063 thread.set_speed(Speed::Fast, cx);
4064 });
4065
4066 for subagent in &subagents {
4067 assert_eq!(
4068 subagent.read(cx).speed(),
4069 Some(Speed::Fast),
4070 "Subagent speed should match parent after set_speed"
4071 );
4072 }
4073 });
4074 }
4075
4076 #[gpui::test]
4077 async fn test_dropped_subagent_does_not_panic(cx: &mut TestAppContext) {
4078 let (parent, _event_stream) = setup_thread_for_test(cx).await;
4079 let subagents = setup_parent_with_subagents(cx, &parent, 1);
4080
4081 // Drop the subagent so the WeakEntity can no longer be upgraded
4082 drop(subagents);
4083
4084 // Should not panic even though the subagent was dropped
4085 cx.update(|cx| {
4086 parent.update(cx, |thread, cx| {
4087 thread.set_thinking_enabled(true, cx);
4088 thread.set_speed(Speed::Fast, cx);
4089 thread.set_thinking_effort(Some("high".to_string()), cx);
4090 });
4091 });
4092 }
4093
4094 #[gpui::test]
4095 async fn test_handle_tool_use_json_parse_error_adds_tool_use_to_content(
4096 cx: &mut TestAppContext,
4097 ) {
4098 let (thread, event_stream) = setup_thread_for_test(cx).await;
4099
4100 cx.update(|cx| {
4101 thread.update(cx, |thread, _cx| {
4102 let tool_use_id = LanguageModelToolUseId::from("test_tool_id");
4103 let tool_name: Arc<str> = Arc::from("test_tool");
4104 let raw_input: Arc<str> = Arc::from("{invalid json");
4105 let json_parse_error = "expected value at line 1 column 1".to_string();
4106
4107 // Call the function under test
4108 let result = thread.handle_tool_use_json_parse_error_event(
4109 tool_use_id.clone(),
4110 tool_name.clone(),
4111 raw_input.clone(),
4112 json_parse_error,
4113 &event_stream,
4114 );
4115
4116 // Verify the result is an error
4117 assert!(result.is_error);
4118 assert_eq!(result.tool_use_id, tool_use_id);
4119 assert_eq!(result.tool_name, tool_name);
4120 assert!(matches!(
4121 result.content,
4122 LanguageModelToolResultContent::Text(_)
4123 ));
4124
4125 // Verify the tool use was added to the message content
4126 {
4127 let last_message = thread.pending_message();
4128 assert_eq!(
4129 last_message.content.len(),
4130 1,
4131 "Should have one tool_use in content"
4132 );
4133
4134 match &last_message.content[0] {
4135 AgentMessageContent::ToolUse(tool_use) => {
4136 assert_eq!(tool_use.id, tool_use_id);
4137 assert_eq!(tool_use.name, tool_name);
4138 assert_eq!(tool_use.raw_input, raw_input.to_string());
4139 assert!(tool_use.is_input_complete);
4140 // Should fall back to empty object for invalid JSON
4141 assert_eq!(tool_use.input, json!({}));
4142 }
4143 _ => panic!("Expected ToolUse content"),
4144 }
4145 }
4146
4147 // Insert the tool result (simulating what the caller does)
4148 thread
4149 .pending_message()
4150 .tool_results
4151 .insert(result.tool_use_id.clone(), result);
4152
4153 // Verify the tool result was added
4154 let last_message = thread.pending_message();
4155 assert_eq!(
4156 last_message.tool_results.len(),
4157 1,
4158 "Should have one tool_result"
4159 );
4160 assert!(last_message.tool_results.contains_key(&tool_use_id));
4161 });
4162 });
4163 }
4164}