1use super::server::{AcpServer, ServerRunError};
2use super::session::actor::SessionActorInit;
3use super::session::agent_key::AgentKey;
4use super::session::agents::SessionAgents;
5use super::session::config::SessionConfigState;
6use super::session::error::SessionError;
7use super::session::hooks::SessionHooks;
8use super::session::model::{Modes, ValidatedMode};
9use super::session::runtime::{AgentRuntime, RuntimeFactory};
10use super::state::{AcpState, AcpStateConfig};
11use crate::resolve::InitialSessionSelection;
12use crate::runtime::{Runtime, RuntimeBuilder};
13use crate::settings_args::SettingsSourceArgs;
14use crate::workspace::WorkspaceManager;
15use crate::workspace::testing::StdCopyCloner;
16use acp_utils::notifications::{AuthMethodsUpdatedParams, McpNotification};
17use acp_utils::testing::{TestPeer, initialize_request};
18use aether_auth::OAuthCredentialStorage;
19use aether_core::agent_spec::{AgentSpec, AgentSpecExposure};
20use aether_core::core::{AgentBuilder, AgentHandle, Prompt};
21use aether_core::events::{AgentEvent, Command, MessageEvent, ToolEvent, TurnEvent, TurnOutcome, UserCommand};
22use aether_core::mcp::{McpBuilder, McpCatalog, mcp};
23use aether_core::testing::ready_catalog;
24use aether_project::AgentCatalog;
25use aether_sessions::SessionStore;
26use aether_sessions::{SessionControlEvent, SessionEvent, SessionMeta, UserEvent, last_agent_from_events};
27use agent_client_protocol::schema::v2::{
28 AbsolutePath, AvailableCommand, InitializeResponse, ReplayFrom, ReplayFromStart, ResumeSessionRequest, SessionId,
29 SessionUpdate, StateUpdate, StopReason,
30};
31use agent_client_protocol::{Agent, Channel, Client, ConnectionTo, on_receive_notification};
32use llm::testing::FakeLlmProvider;
33use llm::{ChatMessage, Context, LlmModel, LlmResponse, SessionUsageEvent, StreamingModelProvider};
34use llm::{MessageId, ProviderConnectionOverrides, ToolCallRequest, ToolCallResult};
35use mcp_utils::client::{Elicitation, ElicitationSink};
36use mcp_utils::gateway::ServerSpec;
37use mcp_utils::model::{CallToolResult, CreateTaskResult, DetailedTask, ElicitResult, Task, TaskPayload, TaskStatus};
38use mcp_utils::testing::{FakeMcpServer, FakeTool, FakeToolResponse, Gate, completed_task_payload, elicitation_sink};
39use std::collections::HashMap;
40use std::future::Future;
41use std::net::SocketAddr;
42use std::path::PathBuf;
43use std::sync::{Arc, Mutex};
44use tokio::sync::{mpsc, oneshot, watch};
45use tokio::task::{JoinHandle, JoinSet, LocalSet};
46use tokio_util::sync::CancellationToken;
47use utils::mcp_status::McpServerStatusEntry;
48
49const PLANNER_REPLY: &str = "planner reply";
50const CODER_REPLY: &str = "coder reply";
51
52pub struct AcpTestHarness {
53 pub client_cx: ConnectionTo<Agent>,
54 pub peer: TestPeer,
55 pub initialize_response: InitializeResponse,
56 connection: Option<HarnessConnectionTasks>,
57 pub auth_updates: mpsc::UnboundedReceiver<acp_utils::notifications::AuthMethodsUpdatedParams>,
58 resume_agent: FakeAcpAgent,
59 runtime_control: Arc<Mutex<FakeRuntimeControl>>,
60 pub oauth_store: Arc<aether_auth::FakeOAuthCredentialStore>,
61 agent_cx: ConnectionTo<Client>,
62 state: Arc<AcpState>,
63 session_store: Arc<SessionStore>,
64 _tmp: tempfile::TempDir,
65}
66
67pub struct AcpWebSocketTestServer {
69 pub address: SocketAddr,
70 detached: watch::Receiver<u64>,
71 shutdown: Option<oneshot::Sender<()>>,
72 task: JoinHandle<Result<(), ServerRunError>>,
73}
74
75impl AcpWebSocketTestServer {
76 pub fn url(&self) -> String {
77 format!("ws://{}", self.address)
78 }
79
80 pub fn subscribe(&self) -> watch::Receiver<u64> {
81 self.detached.clone()
82 }
83
84 pub async fn shutdown(mut self) {
86 let _ = self.shutdown.take().expect("server is running").send(());
87 (&mut self.task).await.expect("listener task joins").expect("listener shuts down");
88 }
89
90 pub async fn abort(mut self) {
92 self.task.abort();
93 assert!((&mut self.task).await.expect_err("listener task is aborted").is_cancelled());
94 }
95}
96
97impl Drop for AcpWebSocketTestServer {
98 fn drop(&mut self) {
99 self.task.abort();
100 }
101}
102
103pub struct FakeAgentSwitchingSession {
104 session_id: SessionId,
105 planner: FakeAcpAgent,
106 coder: FakeAcpAgent,
107}
108
109#[derive(Clone)]
110pub struct FakeAcpAgent {
111 name: String,
112 captured_contexts: Arc<Mutex<Vec<Context>>>,
113}
114
115pub struct FakeBackgroundTask {
116 server: FakeMcpServer,
117}
118
119pub struct PendingRuntime {
120 started: oneshot::Receiver<()>,
121 release: oneshot::Sender<bool>,
122}
123
124impl PendingRuntime {
125 pub async fn wait_until_started(&mut self) {
126 (&mut self.started).await.expect("runtime startup reached");
127 }
128
129 pub fn finish(self, succeed: bool) {
130 self.release.send(succeed).expect("runtime startup is waiting");
131 }
132}
133
134impl AcpTestHarness {
135 pub fn pause_next_runtime(&self) -> PendingRuntime {
136 let (started, observed) = oneshot::channel();
137 let (release, proceed) = oneshot::channel();
138 self.runtime_control.lock().unwrap().pending = Some((started, proceed));
139 PendingRuntime { started: observed, release }
140 }
141
142 pub fn pause_prompt_expansion(&self) -> Gate {
143 let gate = Gate::new();
144 self.runtime_control.lock().unwrap().prompt_gate = Some(gate.clone());
145 gate
146 }
147
148 pub fn add_mcp_server(&self, server: ServerSpec) {
149 self.runtime_control.lock().unwrap().extra_servers.push(server);
150 }
151
152 pub fn add_stalled_mcp_server(&self) {
153 self.runtime_control.lock().unwrap().stalled_server = true;
154 }
155
156 pub fn watch_agent_prompts(&self) -> watch::Receiver<usize> {
157 self.runtime_control.lock().unwrap().agent_prompts.subscribe()
158 }
159
160 pub fn elicit_during_prompt_expansion(&self) -> mpsc::UnboundedReceiver<ElicitResult> {
161 let (sender, receiver) = mpsc::unbounded_channel();
162 self.runtime_control.lock().unwrap().elicitation_results = Some(sender);
163 receiver
164 }
165
166 pub fn live_runtime_count(&self) -> usize {
167 self.runtime_control.lock().unwrap().agents.iter().filter(|sender| !sender.is_closed()).count()
168 }
169
170 pub async fn disconnect(&mut self) {
171 if let Some(connection) = self.connection.take() {
172 connection.shutdown().await;
173 }
174 }
175
176 pub async fn start_unattached_session(&mut self, prompt: impl Into<String>) -> SessionId {
177 self.disconnect().await;
178 self.state.start_session(prompt.into()).await.expect("unattached session starts")
179 }
180
181 pub async fn reconnect(&mut self) {
183 assert!(self.connection.is_none(), "disconnect the current client before reconnecting");
184 assert!(!self.state.stop_token().is_cancelled(), "host is running");
185 let connection = connect_client(self.state.clone()).await;
186 self.client_cx = connection.client_cx;
187 self.agent_cx = connection.agent_cx;
188 self.peer = connection.peer;
189 self.initialize_response = connection.initialize_response;
190 self.auth_updates = connection.auth_updates;
191 self.connection = Some(connection.tasks);
192 }
193
194 pub async fn resume(&self, session_id: &SessionId) {
195 self.client_cx
196 .send_request(ResumeSessionRequest::new(session_id.clone(), AbsolutePath::new("/tmp")))
197 .block_task()
198 .await
199 .expect("session resumes");
200 }
201
202 pub async fn resume_with_replay(&self, session_id: &SessionId) {
203 self.client_cx
204 .send_request(
205 ResumeSessionRequest::new(session_id.clone(), AbsolutePath::new("/tmp"))
206 .replay_from(ReplayFrom::Start(ReplayFromStart::new())),
207 )
208 .block_task()
209 .await
210 .expect("session resumes with replay");
211 }
212
213 pub async fn serve_websocket(&mut self) -> AcpWebSocketTestServer {
215 self.disconnect().await;
216 self.listen_websocket().await
217 }
218
219 pub async fn listen_websocket(&self) -> AcpWebSocketTestServer {
221 assert!(!self.state.stop_token().is_cancelled(), "host is running");
222 let server =
223 AcpServer::bind("127.0.0.1:0".parse().unwrap(), self.state.clone()).await.expect("bind test server");
224 let address = server.local_addr().expect("listener address");
225 let detached = self.state.client_slot().subscribe();
226 let (shutdown, stopped) = oneshot::channel();
227 let task = tokio::spawn(server.run_until(async move {
228 let _ = stopped.await;
229 Ok(())
230 }));
231 AcpWebSocketTestServer { address, detached, shutdown: Some(shutdown), task }
232 }
233
234 pub async fn shutdown(&mut self) {
236 self.disconnect().await;
237 self.state.shutdown_all().await;
238 }
239
240 pub fn stored_events(&self, session_id: &SessionId) -> Vec<SessionEvent> {
241 self.session_store.load(session_id.0.as_ref()).expect("stored session loads").1
242 }
243
244 pub async fn run<F, Fut, T>(body: F) -> T
248 where
249 F: FnOnce(Self) -> Fut,
250 Fut: Future<Output = T>,
251 {
252 LocalSet::new().run_until(Box::pin(async move { body(Self::start().await).await })).await
253 }
254
255 pub async fn run_with_hooks<F, Fut>(hooks: SessionHooks, body: F)
256 where
257 F: FnOnce(Self) -> Fut,
258 Fut: Future<Output = ()>,
259 {
260 LocalSet::new().run_until(Box::pin(async move { body(Self::start_with(hooks).await).await })).await;
261 }
262
263 pub async fn start() -> Self {
264 Self::start_with(SessionHooks::default()).await
265 }
266
267 async fn start_with(hooks: SessionHooks) -> Self {
268 let tmp = tempfile::tempdir().expect("tempdir for session store");
269 let session_store = Arc::new(SessionStore::from_path(tmp.path().to_path_buf()));
270 let workspace_manager = Arc::new(WorkspaceManager::from_registry_path_with_cloner(
271 tmp.path().join("workspaces.json"),
272 Arc::new(StdCopyCloner),
273 ));
274 let (resume_def, resume_agent) = fake_agent("Resume", "resume-mcp", "resume", "resumed reply");
275 let mut resume_agents = HashMap::new();
276 resume_agents.insert(resume_def.spec.name.clone(), resume_def);
277 let runtime_control = Arc::new(Mutex::new(FakeRuntimeControl::default()));
278 let runtime_factory = Arc::new(FakeRuntimeFactory {
279 cwd: PathBuf::from("/tmp"),
280 agents: resume_agents,
281 control: runtime_control.clone(),
282 });
283 let oauth_store = Arc::new(aether_auth::FakeOAuthCredentialStore::new());
284 let state = Arc::new(AcpState::with_login(
285 AcpStateConfig {
286 session_store: session_store.clone(),
287 workspace_manager,
288 oauth_credential_store: oauth_store.clone(),
289 initial_selection: InitialSessionSelection::default(),
290 settings_source: SettingsSourceArgs {
292 settings_json: Some(r#"{"agents":[]}"#.to_string()),
293 settings_file: None,
294 },
295 provider_connections: ProviderConnectionOverrides::default(),
296 telemetry: None,
297 runtime_factory: Some(runtime_factory),
298 cwd: PathBuf::from("/tmp"),
299 hooks,
300 },
301 Arc::new(FakeProviderLogin),
302 ));
303
304 let connection = connect_client(state.clone()).await;
305 Self {
306 client_cx: connection.client_cx,
307 peer: connection.peer,
308 initialize_response: connection.initialize_response,
309 connection: Some(connection.tasks),
310 auth_updates: connection.auth_updates,
311 resume_agent,
312 runtime_control,
313 oauth_store,
314 agent_cx: connection.agent_cx,
315 state,
316 session_store,
317 _tmp: tmp,
318 }
319 }
320
321 pub fn resume_agent(&self) -> &FakeAcpAgent {
322 &self.resume_agent
323 }
324
325 pub async fn insert_agent_switching_session(&self) -> FakeAgentSwitchingSession {
326 self.insert_switching_session(
327 SessionId::new("agent-switching-session"),
328 Vec::new(),
329 Some("Planner".to_string()),
330 false,
331 )
332 .await
333 }
334
335 pub async fn insert_agent_switching_session_with_serverless_coder(&self) -> FakeAgentSwitchingSession {
336 self.insert_switching_session(
337 SessionId::new("agent-switching-serverless-session"),
338 Vec::new(),
339 Some("Planner".to_string()),
340 true,
341 )
342 .await
343 }
344
345 pub async fn insert_loaded_agent_switching_session(&self, session_id: &str) -> FakeAgentSwitchingSession {
346 let events = self.session_store.load(session_id).map(|(_, events)| events).unwrap_or_default();
347 let selected_mode = last_agent_from_events(Some("Planner".to_string()), &events);
348 self.insert_switching_session(SessionId::new(session_id), events, selected_mode, false).await
349 }
350
351 pub async fn expect_idle(&mut self, session_id: &SessionId, expected: StopReason) {
352 loop {
353 let notification = self.peer.next_session_notification().await;
354 if notification.session_id == *session_id
355 && let SessionUpdate::StateUpdate(StateUpdate::Idle(idle)) = notification.update
356 {
357 assert_eq!(idle.stop_reason, Some(expected));
358 return;
359 }
360 }
361 }
362
363 pub async fn updates_until(
364 &mut self,
365 session_id: &SessionId,
366 last: impl Fn(&SessionUpdate) -> bool,
367 ) -> Vec<SessionUpdate> {
368 let mut updates = Vec::new();
369 loop {
370 let notification = self.peer.next_session_notification().await;
371 if notification.session_id != *session_id {
372 continue;
373 }
374 let done = last(¬ification.update);
375 updates.push(notification.update);
376 if done {
377 return updates;
378 }
379 }
380 }
381
382 pub async fn expect_mcp_server_status(&mut self, expected: &[&str]) {
383 let statuses = self.next_mcp_server_statuses().await;
384 let names = statuses.iter().map(|server| server.name.as_str()).collect::<Vec<_>>();
385 for server_name in expected {
386 assert!(names.contains(server_name), "expected server {server_name} in {names:?}");
387 }
388 }
389
390 pub async fn expect_mcp_server_status_exact(&mut self, expected: &[&str]) {
391 let statuses = self.next_mcp_server_statuses().await;
392 let names = statuses.iter().map(|server| server.name.as_str()).collect::<Vec<_>>();
393 assert_eq!(names, expected);
394 }
395
396 pub async fn next_mcp_server_statuses(&mut self) -> Vec<McpServerStatusEntry> {
397 let McpNotification::ServerStatus { servers } = self.peer.next_mcp_notification().await;
398 servers
399 }
400
401 pub async fn expect_available_commands(&mut self, expected: &[&str], unexpected: &[&str]) {
402 let commands = self.next_available_commands().await;
403 let names = commands.iter().map(|command| command.name.as_str()).collect::<Vec<_>>();
404 for name in expected {
405 assert!(names.contains(name), "expected command /{name} in {names:?}");
406 }
407 for name in unexpected {
408 assert!(!names.contains(name), "did not expect command /{name} in {names:?}");
409 }
410 }
411
412 pub async fn next_available_commands(&mut self) -> Vec<AvailableCommand> {
413 loop {
414 if let SessionUpdate::AvailableCommandsUpdate(update) = self.peer.next_session_notification().await.update {
415 return update.available_commands;
416 }
417 }
418 }
419
420 pub fn append_agent_switch(&self, session_id: &str, from: Option<&str>, to: Option<&str>) {
421 self.append_stored_event(
422 session_id,
423 &SessionEvent::Control(SessionControlEvent::AgentSwitched {
424 from: from.map(str::to_string),
425 to: to.map(str::to_string),
426 }),
427 );
428 }
429
430 pub async fn insert_stub_session(
436 &self,
437 agent_tx: mpsc::Sender<Command>,
438 agent_rx: mpsc::Receiver<AgentEvent>,
439 agent_handle: AgentHandle,
440 id: SessionId,
441 model: &str,
442 ) {
443 let (mcp, elicitations) = spawn_ready(mcp("/tmp")).await;
444 let runtime = Runtime { agent_tx, agent_rx, agent_handle, mcp };
445 self.register_stub_session(runtime, elicitations, id, model).await;
446 }
447
448 pub async fn insert_background_task_session(
449 &self,
450 provider: FakeLlmProvider,
451 task: &FakeBackgroundTask,
452 id: SessionId,
453 ) -> McpCatalog {
454 let server = task.server.clone();
455 let prompts = self.runtime_control.lock().unwrap().prompt_server("plan");
456 let servers = vec![ServerSpec::in_process("tasks", server), ServerSpec::in_process("prompts", prompts)];
457 let (mcp, elicitations) = spawn_ready(mcp("/tmp").with_servers(servers)).await;
458 let (agent_tx, agent_rx, agent_handle) = AgentBuilder::new(Arc::new(provider))
459 .max_auto_continues(0)
460 .mcp(mcp.clone())
461 .spawn()
462 .await
463 .expect("background task agent spawns");
464 let runtime = Runtime { agent_tx, agent_rx, agent_handle, mcp: mcp.clone() };
465 self.register_stub_session(runtime, elicitations, id, "fake:fake").await;
466 mcp
467 }
468
469 pub fn append_stored_session(&self, session_id: &str, created_at: &str) {
470 self.append_stored_session_in(session_id, created_at, std::path::Path::new("/tmp"));
471 }
472
473 pub fn append_stored_session_in(&self, session_id: &str, created_at: &str, cwd: &std::path::Path) {
474 let meta = SessionMeta {
475 session_id: session_id.to_string(),
476 cwd: cwd.to_path_buf(),
477 model: "anthropic:claude-sonnet-4-5".to_string(),
478 selected_mode: None,
479 created_at: created_at.to_string(),
480 };
481
482 self.session_store.append_meta(session_id, &meta).expect("stored session meta appends");
483 }
484
485 pub fn append_stored_prompt(&self, session_id: &str, prompt: &str) {
486 self.append_stored_event(
487 session_id,
488 &SessionEvent::User(UserEvent::Message {
489 message_id: llm::MessageId::new(),
490 content: vec![llm::ContentBlock::text(prompt)],
491 display_content: None,
492 }),
493 );
494 }
495
496 pub fn append_stored_user_blocks(&self, session_id: &str, blocks: Vec<llm::ContentBlock>) {
497 self.append_stored_event(
498 session_id,
499 &SessionEvent::User(UserEvent::Message {
500 message_id: llm::MessageId::new(),
501 content: blocks,
502 display_content: None,
503 }),
504 );
505 }
506
507 pub fn append_stored_agent_turn(&self, session_id: &str, text: &str) {
508 self.append_stored_agent_text(session_id, text);
509 self.append_stored_event(
510 session_id,
511 &SessionEvent::Agent(AgentEvent::Turn(TurnEvent::Ended { outcome: TurnOutcome::Completed })),
512 );
513 }
514
515 pub fn append_stored_agent_text(&self, session_id: &str, text: &str) {
516 self.append_stored_event(
517 session_id,
518 &SessionEvent::Agent(AgentEvent::Message(MessageEvent::Text {
519 message_id: MessageId::new(),
520 chunk: text.to_string(),
521 is_complete: true,
522 })),
523 );
524 }
525
526 pub fn append_stored_tool_call(&self, session_id: &str, id: &str, name: &str, arguments: &str) {
527 self.append_stored_event(
528 session_id,
529 &SessionEvent::Agent(AgentEvent::Tool(ToolEvent::Call {
530 request: ToolCallRequest {
531 id: id.to_string(),
532 name: name.to_string(),
533 arguments: arguments.to_string(),
534 },
535 })),
536 );
537 }
538
539 pub fn append_stored_tool_result(&self, session_id: &str, id: &str, name: &str, result: &str) {
540 self.append_stored_event(
541 session_id,
542 &SessionEvent::Agent(AgentEvent::Tool(ToolEvent::Result {
543 result: ToolCallResult {
544 id: id.to_string(),
545 name: name.to_string(),
546 arguments: "{}".to_string(),
547 result: result.to_string(),
548 },
549 result_meta: None,
550 })),
551 );
552 }
553
554 async fn insert_switching_session(
555 &self,
556 acp_session_id: SessionId,
557 events: Vec<SessionEvent>,
558 selected_mode: Option<String>,
559 serverless_coder: bool,
560 ) -> FakeAgentSwitchingSession {
561 let (planner_def, planner) = fake_agent("Planner", "planner-mcp", "plan", PLANNER_REPLY);
562 let (mut coder_def, coder) = fake_agent("Coder", "coder-mcp", "edit", CODER_REPLY);
563 if serverless_coder {
564 coder_def.mcp = None;
565 }
566
567 let mut catalog_specs = Vec::new();
568 let mut agents = HashMap::new();
569 for def in [planner_def, coder_def] {
570 catalog_specs.push(def.spec.clone());
571 agents.insert(def.spec.name.clone(), def);
572 }
573 let specs = SessionAgents::new(AgentCatalog::new(PathBuf::from("/tmp"), catalog_specs, None));
574
575 let factory =
576 Arc::new(FakeRuntimeFactory { cwd: PathBuf::from("/tmp"), agents, control: self.runtime_control.clone() });
577 let initial_agent = selected_mode.clone().unwrap_or_else(|| "Planner".to_string());
578
579 self.state
580 .register_session(SessionActorInit {
581 cwd: PathBuf::from("/tmp"),
582 mcp_servers: Vec::new(),
583 session_id: acp_session_id.clone(),
584 connection: Some(self.agent_cx.clone()),
585 repository: self.session_store.clone(),
586 oauth_credential_store: self.oauth_store.clone(),
587 active_agent: AgentKey::Named(initial_agent),
588 specs,
589 runtime_factory: factory,
590 transcript: events,
591 replay: false,
592 modes: switching_modes(),
593 config: SessionConfigState::with_selection(
594 "anthropic:claude-sonnet-4-5".to_string(),
595 selected_mode,
596 None,
597 ),
598 hooks: SessionHooks::default(),
599 })
600 .await;
601 FakeAgentSwitchingSession { session_id: acp_session_id, planner, coder }
602 }
603
604 pub fn append_stored_event(&self, session_id: &str, event: &SessionEvent) {
605 self.session_store.append_event(session_id, event).expect("stored session event appends");
606 }
607
608 async fn register_stub_session(
609 &self,
610 mut runtime: Runtime,
611 elicitations: mpsc::Receiver<Elicitation>,
612 id: SessionId,
613 model: &str,
614 ) {
615 let model_spec: llm::catalog::LlmModel = "anthropic:claude-sonnet-4-5".parse().expect("test model parses");
616 let mut specs = SessionAgents::new(AgentCatalog::empty(PathBuf::from("/tmp")));
617 specs.set_default(AgentSpec::bare(&model_spec, None, Vec::new()));
618 let agent_prompts = self.runtime_control.lock().unwrap().agent_prompts.clone();
619 self.runtime_control.lock().unwrap().agents.push(runtime.agent_tx.clone());
620 runtime.agent_tx = count_prompts(runtime.agent_tx, agent_prompts);
621 let factory = Arc::new(StubRuntimeFactory { runtime: Mutex::new(Some((runtime, elicitations))) });
622
623 self.state
624 .register_session(SessionActorInit {
625 cwd: PathBuf::from("/tmp"),
626 mcp_servers: Vec::new(),
627 session_id: id.clone(),
628 connection: Some(self.agent_cx.clone()),
629 repository: self.session_store.clone(),
630 oauth_credential_store: self.oauth_store.clone(),
631 active_agent: AgentKey::Default,
632 specs,
633 runtime_factory: factory,
634 transcript: Vec::new(),
635 replay: false,
636 modes: Modes::default(),
637 config: SessionConfigState::with_selection(model.to_string(), None, None),
638 hooks: SessionHooks::default(),
639 })
640 .await;
641 }
642}
643
644impl FakeAgentSwitchingSession {
645 pub fn session_id(&self) -> &SessionId {
646 &self.session_id
647 }
648
649 pub fn planner(&self) -> &FakeAcpAgent {
650 &self.planner
651 }
652
653 pub fn coder(&self) -> &FakeAcpAgent {
654 &self.coder
655 }
656}
657
658impl FakeAcpAgent {
659 pub fn assert_saw(&self, expected: &[&str]) {
662 let seen = self.latest_conversation();
663 for text in expected {
664 assert!(seen.iter().any(|m| m == text), "{} should have seen {text:?}; saw {seen:?}", self.name);
665 }
666 }
667
668 pub fn assert_saw_exactly(&self, expected: &[&str]) {
672 let seen = self.latest_conversation();
673 let expected: Vec<String> = expected.iter().map(|t| (*t).to_string()).collect();
674 assert_eq!(seen, expected, "{} conversation mismatch", self.name);
675 }
676
677 pub fn assert_saw_user_content(&self, expected: &[llm::ContentBlock]) {
678 let contexts = self.captured_contexts.lock().expect("captured contexts lock is healthy");
679 let latest = contexts.last().expect("agent should have run a turn");
680 assert!(
681 latest
682 .messages()
683 .iter()
684 .any(|message| { matches!(message, ChatMessage::User { content, .. } if content == expected) }),
685 "{} should have seen user content {expected:?}",
686 self.name
687 );
688 }
689
690 pub fn assert_never_ran(&self) {
692 let contexts = self.captured_contexts.lock().expect("captured contexts lock is healthy");
693 assert!(contexts.is_empty(), "{} should not have run; captured {} context(s)", self.name, contexts.len());
694 }
695
696 fn latest_conversation(&self) -> Vec<String> {
697 let contexts = self.captured_contexts.lock().expect("captured contexts lock is healthy");
698 let latest = contexts.last().unwrap_or_else(|| panic!("{} should have run a turn", self.name));
699 conversation_texts(latest)
700 }
701}
702
703impl FakeBackgroundTask {
704 pub const TOOL: &str = "tasks__start";
705 const ID: &str = "background-task";
706
707 pub fn new() -> Self {
708 let server = FakeMcpServer::new()
709 .with_tool(Self::start_tool())
710 .with_task(Self::ID, [DetailedTask::new(background_task(TaskStatus::Working), TaskPayload::Working)]);
711 Self { server }
712 }
713
714 pub fn complete(&self, output: &str) {
715 let result = CallToolResult::success(vec![mcp_utils::model::ContentBlock::text(output)]);
716 let completed = DetailedTask::new(background_task(TaskStatus::Completed), completed_task_payload(result));
717 self.server.state().script_task(Self::ID, [completed]);
718 }
719
720 pub async fn announce_unchanged_tools(&self) {
721 let state = self.server.state();
722 let listed = state.tool_list_requests();
723 state.notify_tools_changed().await;
724 state.wait_until(|state| state.tool_list_requests() > listed).await;
725 }
726
727 pub async fn add_tool(&self, name: &str) {
728 self.server.state().add_tool_and_notify(FakeTool::new(name)).await;
729 }
730
731 fn start_tool() -> FakeTool {
732 let working = background_task(TaskStatus::Working);
733 FakeTool::new("start").responds(FakeToolResponse::task(CreateTaskResult::new(working)))
734 }
735}
736
737impl Default for FakeBackgroundTask {
738 fn default() -> Self {
739 Self::new()
740 }
741}
742
743struct HarnessConnection {
744 client_cx: ConnectionTo<Agent>,
745 agent_cx: ConnectionTo<Client>,
746 peer: TestPeer,
747 initialize_response: InitializeResponse,
748 auth_updates: mpsc::UnboundedReceiver<acp_utils::notifications::AuthMethodsUpdatedParams>,
749 tasks: HarnessConnectionTasks,
750}
751
752struct HarnessConnectionTasks {
753 stop: CancellationToken,
754 tasks: JoinSet<Result<(), agent_client_protocol::Error>>,
755}
756
757impl HarnessConnectionTasks {
758 async fn shutdown(mut self) {
759 self.stop.cancel();
760 while let Some(result) = self.tasks.join_next().await {
761 result.expect("connection task joins").expect("connection cleanup completes");
762 }
763 }
764}
765
766async fn connect_client(state: Arc<AcpState>) -> HarnessConnection {
767 let (peer, client_builder) = TestPeer::new();
768 let (auth_tx, auth_updates) = mpsc::unbounded_channel();
769 let client_builder = client_builder.on_receive_notification(
770 async move |notification: AuthMethodsUpdatedParams, _cx| {
771 let _ = auth_tx.send(notification);
772 Ok(())
773 },
774 on_receive_notification!(),
775 );
776 let (agent_transport, client_transport) = Channel::duplex();
777 let (send_agent_connection, agent_ready) = oneshot::channel();
778 let (send_client_connection, client_ready) = oneshot::channel();
779 let stop = state.stop_token().child_token();
780 let mut tasks = JoinSet::new();
781 let agent_stop = stop.clone();
782 tasks.spawn_local(async move { state.serve_with(agent_transport, agent_stop, send_agent_connection).await });
783 tasks.spawn_local(async move {
784 client_builder
785 .with_runner(acp_utils::testing::CaptureConnection(send_client_connection))
786 .connect_to(client_transport)
787 .await
788 });
789 let agent_cx = agent_ready.await.expect("agent connection");
790 let client_cx = client_ready.await.expect("client connection");
791 let initialize_response =
792 client_cx.send_request(initialize_request()).block_task().await.expect("initialize harness");
793 HarnessConnection {
794 client_cx,
795 agent_cx,
796 peer,
797 initialize_response,
798 auth_updates,
799 tasks: HarnessConnectionTasks { stop, tasks },
800 }
801}
802
803struct FakeRuntimeFactory {
807 cwd: PathBuf,
808 agents: HashMap<String, FakeAgentDef>,
809 control: Arc<Mutex<FakeRuntimeControl>>,
810}
811
812#[derive(Default)]
813struct FakeRuntimeControl {
814 agents: Vec<mpsc::Sender<Command>>,
815 pending: Option<(oneshot::Sender<()>, oneshot::Receiver<bool>)>,
816 prompt_gate: Option<Gate>,
817 elicitation_results: Option<mpsc::UnboundedSender<ElicitResult>>,
818 agent_prompts: watch::Sender<usize>,
819 extra_servers: Vec<ServerSpec>,
820 stalled_server: bool,
821}
822
823impl FakeRuntimeControl {
824 fn servers(&self, def: &FakeAgentDef) -> Vec<ServerSpec> {
825 let mut servers = def
826 .mcp
827 .iter()
828 .map(|(server_name, prompt_name)| ServerSpec::in_process(server_name, self.prompt_server(prompt_name)))
829 .collect::<Vec<_>>();
830 servers.extend(self.extra_servers.iter().cloned());
831 if self.stalled_server {
832 servers.push(ServerSpec::in_process("stalled-mcp", FakeMcpServer::new().with_tool_list_gate(Gate::new())));
833 }
834 servers
835 }
836
837 fn prompt_server(&self, prompt_name: &str) -> FakeMcpServer {
838 let mut server = FakeMcpServer::new().with_prompt(prompt_name, format!("expanded {prompt_name}"));
839 if let Some(gate) = self.prompt_gate.clone() {
840 server = server.with_prompt_gate(gate);
841 }
842 if let Some(results) = self.elicitation_results.clone() {
843 server = server.with_prompt_elicitation(results);
844 }
845 server
846 }
847}
848
849struct FakeAgentDef {
850 spec: AgentSpec,
851 provider: Arc<dyn StreamingModelProvider>,
852 mcp: Option<(String, String)>,
853}
854
855#[async_trait::async_trait]
856impl RuntimeFactory for FakeRuntimeFactory {
857 async fn spawn(
858 &self,
859 _agent: AgentKey,
860 spec: &AgentSpec,
861 initial_messages: Vec<ChatMessage>,
862 usage_seed: Option<SessionUsageEvent>,
863 elicitations: mpsc::Sender<Elicitation>,
864 ) -> Result<AgentRuntime, SessionError> {
865 let pending = self.control.lock().unwrap().pending.take();
866 if let Some((started, proceed)) = pending {
867 let _ = started.send(());
868 if !proceed.await.unwrap_or(false) {
869 return Err(SessionError::AgentNotFound("injected startup failure".to_string()));
870 }
871 }
872 let def = self
873 .agents
874 .get(&spec.name)
875 .or_else(|| self.agents.values().next())
876 .ok_or_else(|| SessionError::AgentNotFound(spec.name.clone()))?;
877 let (servers, stalled_server) = {
878 let control = self.control.lock().unwrap();
879 (control.servers(def), control.stalled_server)
880 };
881 let mut builder = RuntimeBuilder::from_spec(self.cwd.clone(), spec.clone())
882 .provider(def.provider.clone())
883 .extra_servers(servers)
884 .elicitation(ElicitationSink::form_and_url(elicitations))
885 .resume_usage(usage_seed);
886
887 if !stalled_server {
888 builder = builder.wait_for_mcp();
889 }
890
891 let runtime = builder.build(initial_messages).await?;
892 self.control.lock().unwrap().agents.push(runtime.agent_tx.clone());
893
894 Ok(AgentRuntime::new(runtime))
895 }
896}
897
898struct StubRuntimeFactory {
899 runtime: Mutex<Option<(Runtime, mpsc::Receiver<Elicitation>)>>,
900}
901
902#[async_trait::async_trait]
903impl RuntimeFactory for StubRuntimeFactory {
904 async fn spawn(
905 &self,
906 _agent: AgentKey,
907 _spec: &AgentSpec,
908 _initial_messages: Vec<ChatMessage>,
909 _usage_seed: Option<SessionUsageEvent>,
910 elicitations: mpsc::Sender<Elicitation>,
911 ) -> Result<AgentRuntime, SessionError> {
912 let (runtime, mut stub_elicitations) =
913 self.runtime.lock().expect("stub runtime lock is healthy").take().expect("stub runtime spawned once");
914 tokio::spawn(async move {
915 while let Some(elicitation) = stub_elicitations.recv().await {
916 if elicitations.send(elicitation).await.is_err() {
917 return;
918 }
919 }
920 });
921 Ok(AgentRuntime::new(runtime))
922 }
923}
924
925async fn spawn_ready(builder: McpBuilder) -> (McpCatalog, mpsc::Receiver<Elicitation>) {
926 let (sink, elicitations) = elicitation_sink();
927 (ready_catalog(builder.with_elicitation(sink)).await, elicitations)
928}
929
930fn count_prompts(agent_tx: mpsc::Sender<Command>, prompts: watch::Sender<usize>) -> mpsc::Sender<Command> {
931 let (tx, mut rx) = mpsc::channel(32);
932 tokio::spawn(async move {
933 while let Some(command) = rx.recv().await {
934 let is_prompt = matches!(command, Command::UserCommand(UserCommand::Text { .. }));
935 if agent_tx.send(command).await.is_err() {
936 return;
937 }
938 if is_prompt {
939 prompts.send_modify(|count| *count += 1);
940 }
941 }
942 });
943 tx
944}
945
946fn background_task(status: TaskStatus) -> Task {
947 const CREATED_AT: &str = "2026-01-01T00:00:00Z";
948 Task::new(FakeBackgroundTask::ID, status, CREATED_AT.to_string(), CREATED_AT.to_string()).with_poll_interval_ms(10)
949}
950
951fn fake_agent(name: &str, server_name: &str, prompt_name: &str, reply: &str) -> (FakeAgentDef, FakeAcpAgent) {
952 const TURNS_BEFORE_AND_AFTER_REATTACH: usize = 2;
953 let provider = FakeLlmProvider::new(vec![
954 vec![LlmResponse::Start, LlmResponse::text(reply), LlmResponse::done()];
955 TURNS_BEFORE_AND_AFTER_REATTACH
956 ])
957 .with_display_name(name);
958 let captured_contexts = provider.captured_contexts();
959 let def = FakeAgentDef {
960 spec: fake_agent_spec(name),
961 provider: Arc::new(provider),
962 mcp: Some((server_name.to_string(), prompt_name.to_string())),
963 };
964 let observer = FakeAcpAgent { name: name.to_string(), captured_contexts };
965 (def, observer)
966}
967
968struct FakeProviderLogin;
969
970#[async_trait::async_trait]
971impl super::state::ProviderLogin for FakeProviderLogin {
972 async fn login(&self, store: &dyn OAuthCredentialStorage) -> Result<(), llm::LlmError> {
973 store
974 .save("codex", serde_json::json!({"access_token": "fake-access", "refresh_token": "fake-refresh"}))
975 .await?;
976 Ok(())
977 }
978}
979
980fn switching_modes() -> Modes {
981 Modes::new(vec![
982 ValidatedMode {
983 name: "Planner".to_string(),
984 model: "anthropic:claude-sonnet-4-5".to_string(),
985 reasoning_effort: None,
986 },
987 ValidatedMode { name: "Coder".to_string(), model: text_only_model(), reasoning_effort: None },
988 ])
989}
990
991fn text_only_model() -> String {
992 LlmModel::all()
993 .iter()
994 .find(|model| !model.supports_image())
995 .expect("catalog contains a text-only model")
996 .to_string()
997}
998
999fn fake_agent_spec(name: &str) -> AgentSpec {
1000 let model: llm::catalog::LlmModel = "anthropic:claude-sonnet-4-5".parse().expect("test model parses");
1001 let mut spec = AgentSpec::bare(&model, None, vec![Prompt::text(&format!("{name} system prompt"))]);
1002 spec.name = name.to_string();
1003 spec.description = format!("{name} test agent");
1004 spec.exposure = AgentSpecExposure::user_only();
1005 spec
1006}
1007
1008fn conversation_texts(context: &Context) -> Vec<String> {
1009 context
1010 .messages()
1011 .iter()
1012 .filter_map(|message| match message {
1013 ChatMessage::User { content, .. } => llm::ContentBlock::first_text(content).map(str::to_string),
1014 ChatMessage::Assistant { content, .. } if !content.is_empty() => Some(content.clone()),
1015 _ => None,
1016 })
1017 .collect()
1018}