subagent_tool.rs

  1use acp_thread::{AcpThread, AgentConnection, UserMessageId};
  2use action_log::ActionLog;
  3use agent_client_protocol as acp;
  4use anyhow::{Result, anyhow};
  5use collections::{BTreeMap, HashSet};
  6use futures::{FutureExt, channel::mpsc};
  7use gpui::{App, AppContext, AsyncApp, Entity, SharedString, Task, WeakEntity};
  8use language_model::LanguageModelToolUseId;
  9use project::Project;
 10use schemars::JsonSchema;
 11use serde::{Deserialize, Serialize};
 12use smol::stream::StreamExt;
 13use std::any::Any;
 14use std::path::Path;
 15use std::rc::Rc;
 16use std::sync::Arc;
 17use std::time::Duration;
 18use util::ResultExt;
 19use watch;
 20
 21use crate::{
 22    AgentTool, AnyAgentTool, MAX_PARALLEL_SUBAGENTS, MAX_SUBAGENT_DEPTH, SubagentContext, Thread,
 23    ThreadEvent, ToolCallAuthorization, ToolCallEventStream,
 24};
 25
 26/// When a subagent's remaining context window falls below this fraction (25%),
 27/// the "context running out" prompt is sent to encourage the subagent to wrap up.
 28const CONTEXT_LOW_THRESHOLD: f32 = 0.25;
 29
 30/// Spawns a subagent with its own context window to perform a delegated task.
 31///
 32/// Use this tool when you want to do any of the following:
 33/// - Perform an investigation where all you need to know is the outcome, not the research that led to that outcome.
 34/// - Complete a self-contained task where you need to know if it succeeded or failed (and how), but none of its intermediate output.
 35/// - Run multiple tasks in parallel that would take significantly longer to run sequentially.
 36///
 37/// You control what the subagent does by providing:
 38/// 1. A task prompt describing what the subagent should do
 39/// 2. A summary prompt that tells the subagent how to summarize its work when done
 40/// 3. A "context running out" prompt for when the subagent is low on tokens
 41///
 42/// Each subagent has access to the same tools you do. You can optionally restrict
 43/// which tools each subagent can use.
 44///
 45/// Note:
 46/// - Maximum 8 subagents can run in parallel
 47/// - Subagents cannot use tools you don't have access to
 48/// - If spawning multiple subagents that might write to the filesystem, provide
 49///   guidance on how to avoid conflicts (e.g. assign each to different directories)
 50/// - Instruct subagents to be concise in their summaries to conserve your context
 51#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
 52pub struct SubagentToolInput {
 53    /// Short label displayed in the UI while the subagent runs (e.g., "Researching alternatives")
 54    pub label: String,
 55
 56    /// The initial prompt that tells the subagent what task to perform.
 57    /// Be specific about what you want the subagent to accomplish.
 58    pub task_prompt: String,
 59
 60    /// The prompt sent to the subagent when it completes its task, asking it
 61    /// to summarize what it did and return results. This summary becomes the
 62    /// tool result you receive.
 63    ///
 64    /// Example: "Summarize what you found, listing the top 3 alternatives with pros/cons."
 65    pub summary_prompt: String,
 66
 67    /// The prompt sent if the subagent is running low on context (25% remaining).
 68    /// Should instruct it to stop and summarize progress so far, plus what's left undone.
 69    ///
 70    /// Example: "Context is running low. Stop and summarize your progress so far,
 71    /// and list what remains to be investigated."
 72    pub context_low_prompt: String,
 73
 74    /// Optional: Maximum runtime in milliseconds. If exceeded, the subagent is
 75    /// asked to summarize and return. No timeout by default.
 76    #[serde(default)]
 77    pub timeout_ms: Option<u64>,
 78
 79    /// Optional: List of tool names the subagent is allowed to use.
 80    /// If not provided, the subagent can use all tools available to the parent.
 81    /// Tools listed here must be a subset of the parent's available tools.
 82    #[serde(default)]
 83    pub allowed_tools: Option<Vec<String>>,
 84}
 85
 86/// Tool that spawns a subagent thread to work on a task.
 87pub struct SubagentTool {
 88    parent_thread: WeakEntity<Thread>,
 89    current_depth: u8,
 90}
 91
 92impl SubagentTool {
 93    pub fn new(parent_thread: WeakEntity<Thread>, current_depth: u8) -> Self {
 94        Self {
 95            parent_thread,
 96            current_depth,
 97        }
 98    }
 99
100    pub fn validate_allowed_tools(
101        &self,
102        allowed_tools: &Option<Vec<String>>,
103        cx: &App,
104    ) -> Result<()> {
105        let Some(allowed_tools) = allowed_tools else {
106            return Ok(());
107        };
108
109        let invalid_tools: Vec<_> = self.parent_thread.read_with(cx, |thread, _cx| {
110            allowed_tools
111                .iter()
112                .filter(|tool| !thread.tools.contains_key(tool.as_str()))
113                .map(|s| format!("'{s}'"))
114                .collect()
115        })?;
116
117        if !invalid_tools.is_empty() {
118            return Err(anyhow!(
119                "The following tools do not exist: {}",
120                invalid_tools.join(", ")
121            ));
122        }
123
124        Ok(())
125    }
126}
127
128impl AgentTool for SubagentTool {
129    type Input = SubagentToolInput;
130    type Output = String;
131
132    fn name() -> &'static str {
133        acp_thread::SUBAGENT_TOOL_NAME
134    }
135
136    fn kind() -> acp::ToolKind {
137        acp::ToolKind::Other
138    }
139
140    fn initial_title(
141        &self,
142        input: Result<Self::Input, serde_json::Value>,
143        _cx: &mut App,
144    ) -> SharedString {
145        input
146            .map(|i| i.label.into())
147            .unwrap_or_else(|_| "Subagent".into())
148    }
149
150    fn run(
151        self: Arc<Self>,
152        input: Self::Input,
153        event_stream: ToolCallEventStream,
154        cx: &mut App,
155    ) -> Task<Result<String>> {
156        if self.current_depth >= MAX_SUBAGENT_DEPTH {
157            return Task::ready(Err(anyhow!(
158                "Maximum subagent depth ({}) reached",
159                MAX_SUBAGENT_DEPTH
160            )));
161        }
162
163        if let Err(e) = self.validate_allowed_tools(&input.allowed_tools, cx) {
164            return Task::ready(Err(e));
165        }
166
167        let Some(parent_thread_entity) = self.parent_thread.upgrade() else {
168            return Task::ready(Err(anyhow!(
169                "Parent thread no longer exists (subagent depth={})",
170                self.current_depth + 1
171            )));
172        };
173        let parent_thread = parent_thread_entity.read(cx);
174
175        let running_count = parent_thread.running_subagent_count();
176        if running_count >= MAX_PARALLEL_SUBAGENTS {
177            return Task::ready(Err(anyhow!(
178                "Maximum parallel subagents ({}) reached. Wait for existing subagents to complete.",
179                MAX_PARALLEL_SUBAGENTS
180            )));
181        }
182
183        let parent_model = parent_thread.model().cloned();
184        let Some(model) = parent_model else {
185            return Task::ready(Err(anyhow!("No model configured")));
186        };
187
188        let parent_thread_id = parent_thread.id().clone();
189        let project = parent_thread.project.clone();
190        let project_context = parent_thread.project_context().clone();
191        let context_server_registry = parent_thread.context_server_registry.clone();
192        let templates = parent_thread.templates.clone();
193        let parent_tools = parent_thread.tools.clone();
194        let current_depth = self.current_depth;
195        let parent_thread_weak = self.parent_thread.clone();
196
197        cx.spawn(async move |cx| {
198            let subagent_context = SubagentContext {
199                parent_thread_id: parent_thread_id.clone(),
200                tool_use_id: LanguageModelToolUseId::from(uuid::Uuid::new_v4().to_string()),
201                depth: current_depth + 1,
202                summary_prompt: input.summary_prompt.clone(),
203                context_low_prompt: input.context_low_prompt.clone(),
204            };
205
206            // Determine which tools this subagent gets
207            let subagent_tools: BTreeMap<SharedString, Arc<dyn AnyAgentTool>> =
208                if let Some(ref allowed) = input.allowed_tools {
209                    let allowed_set: HashSet<&str> = allowed.iter().map(|s| s.as_str()).collect();
210                    parent_tools
211                        .iter()
212                        .filter(|(name, _)| allowed_set.contains(name.as_ref()))
213                        .map(|(name, tool)| (name.clone(), tool.clone()))
214                        .collect()
215                } else {
216                    parent_tools.clone()
217                };
218
219            let subagent_thread: Entity<Thread> = cx.new(|cx| {
220                Thread::new_subagent(
221                    project.clone(),
222                    project_context.clone(),
223                    context_server_registry.clone(),
224                    templates.clone(),
225                    model.clone(),
226                    subagent_context,
227                    subagent_tools,
228                    cx,
229                )
230            });
231
232            let subagent_weak = subagent_thread.downgrade();
233
234            let acp_thread: Entity<AcpThread> = cx.new(|cx| {
235                let session_id = subagent_thread.read(cx).id().clone();
236                let action_log: Entity<ActionLog> = cx.new(|_| ActionLog::new(project.clone()));
237                let connection: Rc<dyn AgentConnection> = Rc::new(SubagentDisplayConnection);
238                AcpThread::new(
239                    &input.label,
240                    connection,
241                    project.clone(),
242                    action_log,
243                    session_id,
244                    watch::Receiver::constant(acp::PromptCapabilities::new()),
245                    cx,
246                )
247            });
248
249            event_stream.update_subagent_thread(acp_thread.clone());
250
251            let mut user_stop_rx: watch::Receiver<bool> =
252                acp_thread.update(cx, |thread, _| thread.user_stop_receiver());
253
254            if let Some(parent) = parent_thread_weak.upgrade() {
255                parent.update(cx, |thread, _cx| {
256                    thread.register_running_subagent(subagent_weak.clone());
257                });
258            }
259
260            // Helper to wait for user stop signal on the subagent card
261            let wait_for_user_stop = async {
262                loop {
263                    if *user_stop_rx.borrow() {
264                        return;
265                    }
266                    if user_stop_rx.changed().await.is_err() {
267                        std::future::pending::<()>().await;
268                    }
269                }
270            };
271
272            // Run the subagent, handling cancellation from both:
273            // 1. Parent turn cancellation (event_stream.cancelled_by_user)
274            // 2. Direct user stop on subagent card (user_stop_rx)
275            let result = futures::select! {
276                result = run_subagent(
277                    &subagent_thread,
278                    &acp_thread,
279                    input.task_prompt,
280                    input.timeout_ms,
281                    cx,
282                ).fuse() => result,
283                _ = event_stream.cancelled_by_user().fuse() => {
284                    let _ = subagent_thread.update(cx, |thread, cx| {
285                        thread.cancel(cx).detach();
286                    });
287                    Err(anyhow!("Subagent cancelled by user"))
288                }
289                _ = wait_for_user_stop.fuse() => {
290                    let _ = subagent_thread.update(cx, |thread, cx| {
291                        thread.cancel(cx).detach();
292                    });
293                    Err(anyhow!("Subagent stopped by user"))
294                }
295            };
296
297            if let Some(parent) = parent_thread_weak.upgrade() {
298                let _ = parent.update(cx, |thread, _cx| {
299                    thread.unregister_running_subagent(&subagent_weak);
300                });
301            }
302
303            result
304        })
305    }
306}
307
308async fn run_subagent(
309    subagent_thread: &Entity<Thread>,
310    acp_thread: &Entity<AcpThread>,
311    task_prompt: String,
312    timeout_ms: Option<u64>,
313    cx: &mut AsyncApp,
314) -> Result<String> {
315    let mut events_rx =
316        subagent_thread.update(cx, |thread, cx| thread.submit_user_message(task_prompt, cx))?;
317
318    let acp_thread_weak = acp_thread.downgrade();
319
320    let timed_out = if let Some(timeout) = timeout_ms {
321        forward_events_with_timeout(
322            &mut events_rx,
323            &acp_thread_weak,
324            Duration::from_millis(timeout),
325            cx,
326        )
327        .await
328    } else {
329        forward_events_until_stop(&mut events_rx, &acp_thread_weak, cx).await;
330        false
331    };
332
333    let should_interrupt =
334        timed_out || check_context_low(subagent_thread, CONTEXT_LOW_THRESHOLD, cx);
335
336    if should_interrupt {
337        let mut summary_rx =
338            subagent_thread.update(cx, |thread, cx| thread.interrupt_for_summary(cx))?;
339        forward_events_until_stop(&mut summary_rx, &acp_thread_weak, cx).await;
340    } else {
341        let mut summary_rx =
342            subagent_thread.update(cx, |thread, cx| thread.request_final_summary(cx))?;
343        forward_events_until_stop(&mut summary_rx, &acp_thread_weak, cx).await;
344    }
345
346    Ok(extract_last_message(subagent_thread, cx))
347}
348
349async fn forward_events_until_stop(
350    events_rx: &mut mpsc::UnboundedReceiver<Result<ThreadEvent>>,
351    acp_thread: &WeakEntity<AcpThread>,
352    cx: &mut AsyncApp,
353) {
354    while let Some(event) = events_rx.next().await {
355        match event {
356            Ok(ThreadEvent::Stop(_)) => break,
357            Ok(event) => {
358                forward_event_to_acp_thread(event, acp_thread, cx);
359            }
360            Err(_) => break,
361        }
362    }
363}
364
365async fn forward_events_with_timeout(
366    events_rx: &mut mpsc::UnboundedReceiver<Result<ThreadEvent>>,
367    acp_thread: &WeakEntity<AcpThread>,
368    timeout: Duration,
369    cx: &mut AsyncApp,
370) -> bool {
371    use futures::future::{self, Either};
372
373    let deadline = std::time::Instant::now() + timeout;
374
375    loop {
376        let remaining = deadline.saturating_duration_since(std::time::Instant::now());
377        if remaining.is_zero() {
378            return true;
379        }
380
381        let timeout_future = cx.background_executor().timer(remaining);
382        let event_future = events_rx.next();
383
384        match future::select(event_future, timeout_future).await {
385            Either::Left((event, _)) => match event {
386                Some(Ok(ThreadEvent::Stop(_))) => return false,
387                Some(Ok(event)) => {
388                    forward_event_to_acp_thread(event, acp_thread, cx);
389                }
390                Some(Err(_)) => return false,
391                None => return false,
392            },
393            Either::Right((_, _)) => return true,
394        }
395    }
396}
397
398fn forward_event_to_acp_thread(
399    event: ThreadEvent,
400    acp_thread: &WeakEntity<AcpThread>,
401    cx: &mut AsyncApp,
402) {
403    match event {
404        ThreadEvent::UserMessage(message) => {
405            acp_thread
406                .update(cx, |thread, cx| {
407                    for content in message.content {
408                        thread.push_user_content_block(
409                            Some(message.id.clone()),
410                            content.into(),
411                            cx,
412                        );
413                    }
414                })
415                .log_err();
416        }
417        ThreadEvent::AgentText(text) => {
418            acp_thread
419                .update(cx, |thread, cx| {
420                    thread.push_assistant_content_block(text.into(), false, cx)
421                })
422                .log_err();
423        }
424        ThreadEvent::AgentThinking(text) => {
425            acp_thread
426                .update(cx, |thread, cx| {
427                    thread.push_assistant_content_block(text.into(), true, cx)
428                })
429                .log_err();
430        }
431        ThreadEvent::ToolCallAuthorization(ToolCallAuthorization {
432            tool_call,
433            options,
434            response,
435            ..
436        }) => {
437            let outcome_task = acp_thread.update(cx, |thread, cx| {
438                thread.request_tool_call_authorization(tool_call, options, true, cx)
439            });
440            if let Ok(Ok(task)) = outcome_task {
441                cx.background_spawn(async move {
442                    if let acp::RequestPermissionOutcome::Selected(
443                        acp::SelectedPermissionOutcome { option_id, .. },
444                    ) = task.await
445                    {
446                        response.send(option_id).ok();
447                    }
448                })
449                .detach();
450            }
451        }
452        ThreadEvent::ToolCall(tool_call) => {
453            acp_thread
454                .update(cx, |thread, cx| thread.upsert_tool_call(tool_call, cx))
455                .log_err();
456        }
457        ThreadEvent::ToolCallUpdate(update) => {
458            acp_thread
459                .update(cx, |thread, cx| thread.update_tool_call(update, cx))
460                .log_err();
461        }
462        ThreadEvent::Retry(status) => {
463            acp_thread
464                .update(cx, |thread, cx| thread.update_retry_status(status, cx))
465                .log_err();
466        }
467        ThreadEvent::Stop(_) => {}
468    }
469}
470
471fn check_context_low(thread: &Entity<Thread>, threshold: f32, cx: &mut AsyncApp) -> bool {
472    thread.read_with(cx, |thread, _| {
473        if let Some(usage) = thread.latest_token_usage() {
474            let remaining_ratio = 1.0 - (usage.used_tokens as f32 / usage.max_tokens as f32);
475            remaining_ratio <= threshold
476        } else {
477            false
478        }
479    })
480}
481
482fn extract_last_message(thread: &Entity<Thread>, cx: &mut AsyncApp) -> String {
483    thread.read_with(cx, |thread, _| {
484        thread
485            .last_message()
486            .map(|m| m.to_markdown())
487            .unwrap_or_else(|| "No response from subagent".to_string())
488    })
489}
490
491#[cfg(test)]
492mod tests {
493    use super::*;
494    use language_model::LanguageModelToolSchemaFormat;
495
496    #[test]
497    fn test_subagent_tool_input_json_schema_is_valid() {
498        let schema = SubagentTool::input_schema(LanguageModelToolSchemaFormat::JsonSchema);
499        let schema_json = serde_json::to_value(&schema).expect("schema should serialize to JSON");
500
501        assert!(
502            schema_json.get("properties").is_some(),
503            "schema should have properties"
504        );
505        let properties = schema_json.get("properties").unwrap();
506
507        assert!(properties.get("label").is_some(), "should have label field");
508        assert!(
509            properties.get("task_prompt").is_some(),
510            "should have task_prompt field"
511        );
512        assert!(
513            properties.get("summary_prompt").is_some(),
514            "should have summary_prompt field"
515        );
516        assert!(
517            properties.get("context_low_prompt").is_some(),
518            "should have context_low_prompt field"
519        );
520        assert!(
521            properties.get("timeout_ms").is_some(),
522            "should have timeout_ms field"
523        );
524        assert!(
525            properties.get("allowed_tools").is_some(),
526            "should have allowed_tools field"
527        );
528    }
529
530    #[test]
531    fn test_subagent_tool_name() {
532        assert_eq!(SubagentTool::name(), "subagent");
533    }
534
535    #[test]
536    fn test_subagent_tool_kind() {
537        assert_eq!(SubagentTool::kind(), acp::ToolKind::Other);
538    }
539}
540
541struct SubagentDisplayConnection;
542
543impl AgentConnection for SubagentDisplayConnection {
544    fn telemetry_id(&self) -> SharedString {
545        "subagent".into()
546    }
547
548    fn auth_methods(&self) -> &[acp::AuthMethod] {
549        &[]
550    }
551
552    fn new_thread(
553        self: Rc<Self>,
554        _project: Entity<Project>,
555        _cwd: &Path,
556        _cx: &mut App,
557    ) -> Task<Result<Entity<AcpThread>>> {
558        unimplemented!("SubagentDisplayConnection does not support new_thread")
559    }
560
561    fn authenticate(&self, _method_id: acp::AuthMethodId, _cx: &mut App) -> Task<Result<()>> {
562        unimplemented!("SubagentDisplayConnection does not support authenticate")
563    }
564
565    fn prompt(
566        &self,
567        _id: Option<UserMessageId>,
568        _params: acp::PromptRequest,
569        _cx: &mut App,
570    ) -> Task<Result<acp::PromptResponse>> {
571        unimplemented!("SubagentDisplayConnection does not support prompt")
572    }
573
574    fn cancel(&self, _session_id: &acp::SessionId, _cx: &mut App) {}
575
576    fn into_any(self: Rc<Self>) -> Rc<dyn Any> {
577        self
578    }
579}