Wire up find_path tool in agent2 (#35799)

Release Notes:

- N/A

---------

Co-authored-by: Antonio Scandurra <me@as-cii.com>

crates/agent2/src/tests/mod.rs

@@ -270,14 +270,14 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {
         vec![
             MessageContent::ToolResult(LanguageModelToolResult {
                 tool_use_id: tool_call_auth_1.tool_call.id.0.to_string().into(),
-                tool_name: tool_call_auth_1.tool_call.title.into(),
+                tool_name: ToolRequiringPermission.name().into(),
                 is_error: false,
                 content: "Allowed".into(),
                 output: None
             }),
             MessageContent::ToolResult(LanguageModelToolResult {
                 tool_use_id: tool_call_auth_2.tool_call.id.0.to_string().into(),
-                tool_name: tool_call_auth_2.tool_call.title.into(),
+                tool_name: ToolRequiringPermission.name().into(),
                 is_error: true,
                 content: "Permission to run tool denied by user".into(),
                 output: None
@@ -286,6 +286,63 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {
     );
 }
 
+#[gpui::test]
+async fn test_tool_hallucination(cx: &mut TestAppContext) {
+    let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;
+    let fake_model = model.as_fake();
+
+    let mut events = thread.update(cx, |thread, cx| thread.send(model.clone(), "abc", cx));
+    cx.run_until_parked();
+    fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
+        LanguageModelToolUse {
+            id: "tool_id_1".into(),
+            name: "nonexistent_tool".into(),
+            raw_input: "{}".into(),
+            input: json!({}),
+            is_input_complete: true,
+        },
+    ));
+    fake_model.end_last_completion_stream();
+
+    let tool_call = expect_tool_call(&mut events).await;
+    assert_eq!(tool_call.title, "nonexistent_tool");
+    assert_eq!(tool_call.status, acp::ToolCallStatus::Pending);
+    let update = expect_tool_call_update(&mut events).await;
+    assert_eq!(update.fields.status, Some(acp::ToolCallStatus::Failed));
+}
+
+async fn expect_tool_call(
+    events: &mut UnboundedReceiver<Result<AgentResponseEvent, LanguageModelCompletionError>>,
+) -> acp::ToolCall {
+    let event = events
+        .next()
+        .await
+        .expect("no tool call authorization event received")
+        .unwrap();
+    match event {
+        AgentResponseEvent::ToolCall(tool_call) => return tool_call,
+        event => {
+            panic!("Unexpected event {event:?}");
+        }
+    }
+}
+
+async fn expect_tool_call_update(
+    events: &mut UnboundedReceiver<Result<AgentResponseEvent, LanguageModelCompletionError>>,
+) -> acp::ToolCallUpdate {
+    let event = events
+        .next()
+        .await
+        .expect("no tool call authorization event received")
+        .unwrap();
+    match event {
+        AgentResponseEvent::ToolCallUpdate(tool_call_update) => return tool_call_update,
+        event => {
+            panic!("Unexpected event {event:?}");
+        }
+    }
+}
+
 async fn next_tool_call_authorization(
     events: &mut UnboundedReceiver<Result<AgentResponseEvent, LanguageModelCompletionError>>,
 ) -> ToolCallAuthorization {
@@ -582,6 +639,77 @@ async fn test_agent_connection(cx: &mut TestAppContext) {
     );
 }
 
+#[gpui::test]
+async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {
+    let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;
+    thread.update(cx, |thread, _cx| thread.add_tool(ThinkingTool));
+    let fake_model = model.as_fake();
+
+    let mut events = thread.update(cx, |thread, cx| thread.send(model.clone(), "Think", cx));
+    cx.run_until_parked();
+
+    let input = json!({ "content": "Thinking hard!" });
+    fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(
+        LanguageModelToolUse {
+            id: "1".into(),
+            name: ThinkingTool.name().into(),
+            raw_input: input.to_string(),
+            input,
+            is_input_complete: true,
+        },
+    ));
+    fake_model.end_last_completion_stream();
+    cx.run_until_parked();
+
+    let tool_call = expect_tool_call(&mut events).await;
+    assert_eq!(
+        tool_call,
+        acp::ToolCall {
+            id: acp::ToolCallId("1".into()),
+            title: "Thinking".into(),
+            kind: acp::ToolKind::Think,
+            status: acp::ToolCallStatus::Pending,
+            content: vec![],
+            locations: vec![],
+            raw_input: Some(json!({ "content": "Thinking hard!" })),
+            raw_output: None,
+        }
+    );
+    let update = expect_tool_call_update(&mut events).await;
+    assert_eq!(
+        update,
+        acp::ToolCallUpdate {
+            id: acp::ToolCallId("1".into()),
+            fields: acp::ToolCallUpdateFields {
+                status: Some(acp::ToolCallStatus::InProgress,),
+                ..Default::default()
+            },
+        }
+    );
+    let update = expect_tool_call_update(&mut events).await;
+    assert_eq!(
+        update,
+        acp::ToolCallUpdate {
+            id: acp::ToolCallId("1".into()),
+            fields: acp::ToolCallUpdateFields {
+                content: Some(vec!["Thinking hard!".into()]),
+                ..Default::default()
+            },
+        }
+    );
+    let update = expect_tool_call_update(&mut events).await;
+    assert_eq!(
+        update,
+        acp::ToolCallUpdate {
+            id: acp::ToolCallId("1".into()),
+            fields: acp::ToolCallUpdateFields {
+                status: Some(acp::ToolCallStatus::Completed),
+                ..Default::default()
+            },
+        }
+    );
+}
+
 /// Filters out the stop events for asserting against in tests
 fn stop_events(
     result_events: Vec<Result<AgentResponseEvent, LanguageModelCompletionError>>,

crates/agent2/src/tests/test_tools.rs

@@ -19,11 +19,20 @@ impl AgentTool for EchoTool {
         "echo".into()
     }
 
-    fn needs_authorization(&self, _input: Self::Input, _cx: &App) -> bool {
-        false
+    fn kind(&self) -> acp::ToolKind {
+        acp::ToolKind::Other
     }
 
-    fn run(self: Arc<Self>, input: Self::Input, _cx: &mut App) -> Task<Result<String>> {
+    fn initial_title(&self, _: Self::Input) -> SharedString {
+        "Echo".into()
+    }
+
+    fn run(
+        self: Arc<Self>,
+        input: Self::Input,
+        _event_stream: ToolCallEventStream,
+        _cx: &mut App,
+    ) -> Task<Result<String>> {
         Task::ready(Ok(input.text))
     }
 }
@@ -44,11 +53,20 @@ impl AgentTool for DelayTool {
         "delay".into()
     }
 
-    fn needs_authorization(&self, _input: Self::Input, _cx: &App) -> bool {
-        false
+    fn initial_title(&self, input: Self::Input) -> SharedString {
+        format!("Delay {}ms", input.ms).into()
     }
 
-    fn run(self: Arc<Self>, input: Self::Input, cx: &mut App) -> Task<Result<String>>
+    fn kind(&self) -> acp::ToolKind {
+        acp::ToolKind::Other
+    }
+
+    fn run(
+        self: Arc<Self>,
+        input: Self::Input,
+        _event_stream: ToolCallEventStream,
+        cx: &mut App,
+    ) -> Task<Result<String>>
     where
         Self: Sized,
     {
@@ -71,16 +89,28 @@ impl AgentTool for ToolRequiringPermission {
         "tool_requiring_permission".into()
     }
 
-    fn needs_authorization(&self, _input: Self::Input, _cx: &App) -> bool {
-        true
+    fn kind(&self) -> acp::ToolKind {
+        acp::ToolKind::Other
     }
 
-    fn run(self: Arc<Self>, _input: Self::Input, cx: &mut App) -> Task<Result<String>>
+    fn initial_title(&self, _input: Self::Input) -> SharedString {
+        "This tool requires permission".into()
+    }
+
+    fn run(
+        self: Arc<Self>,
+        input: Self::Input,
+        event_stream: ToolCallEventStream,
+        cx: &mut App,
+    ) -> Task<Result<String>>
     where
         Self: Sized,
     {
-        cx.foreground_executor()
-            .spawn(async move { Ok("Allowed".to_string()) })
+        let auth_check = self.authorize(input, event_stream);
+        cx.foreground_executor().spawn(async move {
+            auth_check.await?;
+            Ok("Allowed".to_string())
+        })
     }
 }
 
@@ -96,11 +126,20 @@ impl AgentTool for InfiniteTool {
         "infinite".into()
     }
 
-    fn needs_authorization(&self, _input: Self::Input, _cx: &App) -> bool {
-        false
+    fn kind(&self) -> acp::ToolKind {
+        acp::ToolKind::Other
     }
 
-    fn run(self: Arc<Self>, _input: Self::Input, cx: &mut App) -> Task<Result<String>> {
+    fn initial_title(&self, _input: Self::Input) -> SharedString {
+        "This is the tool that never ends... it just goes on and on my friends!".into()
+    }
+
+    fn run(
+        self: Arc<Self>,
+        _input: Self::Input,
+        _event_stream: ToolCallEventStream,
+        cx: &mut App,
+    ) -> Task<Result<String>> {
         cx.foreground_executor().spawn(async move {
             future::pending::<()>().await;
             unreachable!()
@@ -137,11 +176,20 @@ impl AgentTool for WordListTool {
         "word_list".into()
     }
 
-    fn needs_authorization(&self, _input: Self::Input, _cx: &App) -> bool {
-        false
+    fn initial_title(&self, _input: Self::Input) -> SharedString {
+        "List of random words".into()
     }
 
-    fn run(self: Arc<Self>, _input: Self::Input, _cx: &mut App) -> Task<Result<String>> {
+    fn kind(&self) -> acp::ToolKind {
+        acp::ToolKind::Other
+    }
+
+    fn run(
+        self: Arc<Self>,
+        _input: Self::Input,
+        _event_stream: ToolCallEventStream,
+        _cx: &mut App,
+    ) -> Task<Result<String>> {
         Task::ready(Ok("ok".to_string()))
     }
 }