aether_core/testing/
utils.rs

1use mcp_utils::client::ElicitationSink;
2use mcp_utils::gateway::ServerSpec;
3use mcp_utils::model::ElicitResult;
4use mcp_utils::server::McpServer;
5use mcp_utils::testing::{CapturedElicitation, ElicitationScriptBuilder, FakeMcpServer};
6use std::future::Future;
7use std::sync::Arc;
8use std::time::Duration;
9use tokio::sync::{Notify, mpsc};
10
11use crate::context::CompactionConfig;
12use crate::core::{AgentBuilder, AgentError, McpInstructions, Prompt, RetryConfig, agent};
13use crate::events::{
14    AgentCommand, AgentEvent, AgentObserver, Command, ContextEvent, ToolEvent, TurnEvent, UserCommand,
15};
16use crate::mcp::{McpBuilder, McpCatalog, mcp};
17use crate::testing::{AgentTrace, FakeAgentObserver};
18use llm::{ChatMessage, Context, LlmError, LlmModel, LlmResponse, ModelSettings, StreamingModelProvider};
19
20use llm::testing::FakeLlmProvider;
21
22pub async fn drain_until(
23    receiver: &mut mpsc::Receiver<AgentEvent>,
24    predicate: impl Fn(&AgentEvent) -> bool,
25) -> Vec<AgentEvent> {
26    let mut events = Vec::new();
27    while let Some(event) = receiver.recv().await {
28        let matched = predicate(&event);
29        events.push(event);
30        if matched {
31            return events;
32        }
33    }
34    panic!("agent event channel closed before predicate matched");
35}
36
37pub fn content_events(events: Vec<AgentEvent>) -> Vec<AgentEvent> {
38    events
39        .into_iter()
40        .filter(|event| {
41            !matches!(
42                event,
43                AgentEvent::Turn(
44                    TurnEvent::Started { .. }
45                        | TurnEvent::UserMessageInserted { .. }
46                        | TurnEvent::UserMessageDiscarded { .. }
47                        | TurnEvent::LlmCallStarted { .. }
48                        | TurnEvent::LlmCallEnded { .. }
49                ) | AgentEvent::Tool(ToolEvent::DefinitionsUpdated { .. })
50            )
51        })
52        .collect()
53}
54
55pub async fn ready_catalog(builder: McpBuilder) -> McpCatalog {
56    let catalog = builder.spawn().expect("the MCP catalog spawns");
57    catalog.ready().await;
58    catalog
59}
60
61pub fn mcp_instructions(entries: &[(&str, &str)]) -> McpInstructions {
62    let servers = entries.iter().map(|(k, v)| ((*k).to_string(), (*v).to_string())).collect();
63    McpInstructions { servers, progressive_discovery: false }
64}
65
66/// Millisecond-scale retry delays so retry tests run fast under virtual time.
67pub fn fast_retry(max_attempts: u32) -> RetryConfig {
68    RetryConfig { max_attempts, base_delay: Duration::from_millis(1), max_delay: Duration::from_millis(5) }
69}
70
71pub fn test_agent() -> TestAgentBuilder {
72    TestAgentBuilder::default()
73}
74
75/// A fluent sequence of commands and synchronization points for a test agent.
76#[derive(Default)]
77pub struct TestScenario {
78    steps: Vec<Step>,
79}
80
81impl TestScenario {
82    pub fn new() -> Self {
83        Self::default()
84    }
85
86    pub fn send(self, command: Command) -> Self {
87        self.step(Step::Send(command))
88    }
89
90    pub fn user_text(self, text: impl Into<String>) -> Self {
91        self.send(Command::text(text))
92    }
93
94    pub fn cancel(self) -> Self {
95        self.send(Command::UserCommand(UserCommand::Cancel))
96    }
97
98    pub fn switch_model(self, provider: impl StreamingModelProvider + 'static) -> Self {
99        self.send(Command::AgentCommand(AgentCommand::SwitchModel(Box::new(provider))))
100    }
101
102    pub fn replace_conversation(self, messages: Vec<ChatMessage>) -> Self {
103        self.send(Command::AgentCommand(AgentCommand::ReplaceConversation(messages)))
104    }
105
106    pub fn wait_for(self, predicate: impl Fn(&AgentEvent) -> bool + Send + 'static) -> Self {
107        self.step(Step::WaitFor(Box::new(predicate)))
108    }
109
110    pub fn wait_for_turn_end(self) -> Self {
111        self.wait_for(|event| matches!(event, AgentEvent::Turn(TurnEvent::Ended { .. })))
112    }
113
114    pub fn wait_for_compaction_start(self) -> Self {
115        self.wait_for(|event| matches!(event, AgentEvent::Context(ContextEvent::CompactionStarted { .. })))
116    }
117
118    pub fn wait_for_retry(self, attempt: u32) -> Self {
119        self.wait_for(
120            move |event| matches!(event, AgentEvent::Turn(TurnEvent::RetryScheduled { attempt: actual, .. }) if *actual == attempt),
121        )
122    }
123
124    /// Run an arbitrary side effect between scenario steps, e.g. releasing a
125    /// paused LLM stream so a queued message can be injected mid-turn.
126    pub fn perform(self, action: impl FnOnce() + Send + 'static) -> Self {
127        self.step(Step::Perform(Box::new(action)))
128    }
129
130    fn step(mut self, step: Step) -> Self {
131        self.steps.push(step);
132        self
133    }
134}
135
136/// Result of running a test agent, including messages, captured contexts, and the MCP elicitations it answered.
137pub struct TestAgentResult {
138    pub messages: Vec<AgentEvent>,
139    pub contexts: Vec<Context>,
140    pub elicitations: Vec<CapturedElicitation>,
141}
142
143/// Error returned by [`TestAgentBuilder::run`], [`TestAgentBuilder::run_trace`], and
144/// [`TestAgentBuilder::run_with_context`].
145#[derive(Debug, thiserror::Error)]
146pub enum TestAgentError {
147    #[error(transparent)]
148    Agent(#[from] AgentError),
149    #[error("failed to send command to the test agent: {0}")]
150    SendCommand(#[from] mpsc::error::SendError<Command>),
151}
152
153pub type TestResult<T> = std::result::Result<T, TestAgentError>;
154
155#[derive(Default)]
156pub struct TestAgentBuilder {
157    responses: Vec<Vec<Result<LlmResponse, LlmError>>>,
158    llm_config: Vec<Configure<FakeLlmProvider>>,
159    agent_config: Vec<Configure<AgentBuilder>>,
160    /// `None` connects a single fake server named `test`.
161    mcp_servers: Option<Vec<ServerSpec>>,
162    elicitations: Option<ElicitationScriptBuilder>,
163    scenario: Option<TestScenario>,
164}
165
166impl TestAgentBuilder {
167    pub fn commands(self, commands: Vec<Command>) -> Self {
168        assert!(!commands.is_empty(), "commands() requires at least one command");
169        let scenario = commands.into_iter().fold(TestScenario::new(), TestScenario::send);
170        self.scenario(scenario.wait_for_turn_end())
171    }
172
173    pub fn scenario(mut self, scenario: TestScenario) -> Self {
174        assert!(!scenario.steps.is_empty(), "scenario() requires at least one step");
175        assert!(self.scenario.is_none(), "commands(), user_text(), and scenario() are mutually exclusive");
176        self.scenario = Some(scenario);
177        self
178    }
179
180    pub fn user_text(self, text: &str) -> Self {
181        self.commands(vec![Command::text(text)])
182    }
183
184    pub fn llm_responses(mut self, llm_responses: &[Vec<LlmResponse>]) -> Self {
185        self.responses = llm_responses.iter().map(|turn| turn.iter().cloned().map(Ok).collect()).collect();
186        self
187    }
188
189    pub fn llm_result_responses(mut self, llm_responses: &[Vec<Result<LlmResponse, LlmError>>]) -> Self {
190        self.responses = Vec::from(llm_responses);
191        self
192    }
193
194    pub fn model(self, model: LlmModel) -> Self {
195        self.configure_llm(|llm| llm.with_model(model))
196    }
197
198    pub fn provider_context_window(self, window: Option<u32>) -> Self {
199        self.configure_llm(move |llm| llm.with_context_window(window))
200    }
201
202    pub fn context_window_override(self, window: u32) -> Self {
203        self.configure(move |agent| agent.context_window(Some(window)))
204    }
205
206    pub fn tool_timeout(self, timeout: Duration) -> Self {
207        self.configure(move |agent| agent.tool_timeout(timeout))
208    }
209
210    pub fn max_auto_continues(self, max: u32) -> Self {
211        self.configure(move |agent| agent.max_auto_continues(max))
212    }
213
214    pub fn retry_config(self, config: RetryConfig) -> Self {
215        self.configure(move |agent| agent.retry(config))
216    }
217
218    /// Run without the default fake MCP server when the scenario does not exercise tools.
219    pub fn without_mcp(mut self) -> Self {
220        self.mcp_servers = Some(Vec::new());
221        self
222    }
223
224    pub fn mcp_server(mut self, name: &str, server: impl Into<McpServer>) -> Self {
225        self.mcp_servers.get_or_insert_with(Vec::new).push(ServerSpec::in_process(name, server));
226        self
227    }
228
229    pub fn elicitation_response(mut self, response: ElicitResult) -> Self {
230        self.elicitations = Some(self.elicitations.take().unwrap_or_default().response(response));
231        self
232    }
233
234    pub fn on_url_elicitation<F, Fut>(mut self, handler: F) -> Self
235    where
236        F: Fn(String, String) -> Fut + Send + Sync + 'static,
237        Fut: Future<Output = ()> + Send + 'static,
238    {
239        self.elicitations = Some(self.elicitations.take().unwrap_or_default().on_url(handler));
240        self
241    }
242
243    /// Pre-populate the context with conversation history.
244    pub fn messages(self, messages: Vec<ChatMessage>) -> Self {
245        self.configure(|agent| agent.messages(messages))
246    }
247
248    /// Set the system prompt. Multiple prompts are concatenated with double newlines.
249    pub fn system_prompt(self, prompt: Prompt) -> Self {
250        self.configure(|agent| agent.system_prompt(prompt))
251    }
252
253    /// Set the session affinity key attached to every LLM call; defaults to a
254    /// fresh random key per agent.
255    pub fn session_affinity_key(self, key: impl Into<String>) -> Self {
256        let key = key.into();
257        self.configure(|agent| agent.session_affinity_key(key))
258    }
259
260    /// Configure context compaction settings.
261    pub fn compaction_config(self, config: CompactionConfig) -> Self {
262        self.configure(|agent| agent.compaction(config))
263    }
264
265    /// Set the model settings applied to every LLM call.
266    pub fn model_settings(self, settings: ModelSettings) -> Self {
267        self.configure(|agent| agent.model_settings(settings))
268    }
269
270    /// Pause the fake LLM stream at `turn_index` / `chunk_index` until
271    /// `release.notify_one()` is called. Used for deterministic timing tests.
272    pub fn pause_turn_after(self, turn_index: usize, chunk_index: usize, release: Arc<Notify>) -> Self {
273        self.configure_llm(move |llm| llm.pause_turn_after(turn_index, chunk_index, release))
274    }
275
276    /// Attach an observer of the test agent's event stream.
277    pub fn observer(self, observer: Box<dyn AgentObserver>) -> Self {
278        self.configure(|agent| agent.observer(observer))
279    }
280
281    pub async fn run(self) -> TestResult<Vec<AgentEvent>> {
282        let result = self.run_with_context().await?;
283        Ok(result.messages)
284    }
285
286    /// Runs the test agent with a recording observer attached and returns the
287    /// full event trace, including internal events.
288    pub async fn run_trace(self) -> TestResult<AgentTrace> {
289        let observer = FakeAgentObserver::new();
290        let events = observer.events();
291        self.observer(Box::new(observer)).run().await?;
292        Ok(AgentTrace::from_observer_events(&events))
293    }
294
295    /// Runs the test agent and returns both messages and captured contexts.
296    ///
297    /// Use this when you need to verify what context was passed to the LLM,
298    /// for example when testing that file attachments are properly formatted.
299    pub async fn run_with_context(self) -> TestResult<TestAgentResult> {
300        let Self { responses, llm_config, agent_config, mcp_servers, elicitations, scenario } = self;
301        let scenario = scenario.expect("test agent requires commands(), user_text(), or scenario()");
302        let llm =
303            llm_config.into_iter().fold(FakeLlmProvider::from_results(responses), |llm, configure| configure(llm));
304        let captured_contexts = llm.captured_contexts();
305
306        let servers = mcp_servers.unwrap_or_else(|| vec![ServerSpec::in_process("test", FakeMcpServer::math())]);
307        let mut mcp = mcp("/workspace").with_servers(servers);
308        let mut script = None;
309        if let Some(elicitations) = elicitations {
310            let (sink, events) = mpsc::channel(32);
311            mcp = mcp.with_elicitation(ElicitationSink::form_and_url(sink));
312            script = Some(elicitations.spawn(events));
313        }
314
315        let builder = agent(llm).mcp(ready_catalog(mcp).await).retry(RetryConfig::disabled());
316        let builder = agent_config.into_iter().fold(builder, |builder, configure| configure(builder));
317        let (tx, mut rx, handle) = builder.spawn().await?;
318        let mut messages = Vec::new();
319
320        for step in scenario.steps {
321            match step {
322                Step::Send(command) => tx.send(command).await?,
323                Step::WaitFor(predicate) => loop {
324                    let message = rx.recv().await.expect("agent event channel closed before scenario step matched");
325                    let matched = predicate(&message);
326                    messages.push(message);
327                    if matched {
328                        break;
329                    }
330                },
331                Step::Perform(action) => action(),
332            }
333        }
334        drop(tx);
335
336        handle.await_completion().await;
337
338        let contexts = captured_contexts.lock().expect("captured contexts lock poisoned").clone();
339        let elicitations = script.map(|script| script.captured()).unwrap_or_default();
340        Ok(TestAgentResult { messages, contexts, elicitations })
341    }
342
343    fn configure(mut self, configure: impl FnOnce(AgentBuilder) -> AgentBuilder + Send + 'static) -> Self {
344        self.agent_config.push(Box::new(configure));
345        self
346    }
347
348    fn configure_llm(mut self, configure: impl FnOnce(FakeLlmProvider) -> FakeLlmProvider + Send + 'static) -> Self {
349        self.llm_config.push(Box::new(configure));
350        self
351    }
352}
353
354type Configure<T> = Box<dyn FnOnce(T) -> T + Send>;
355
356enum Step {
357    Send(Command),
358    WaitFor(Box<dyn Fn(&AgentEvent) -> bool + Send>),
359    Perform(Box<dyn FnOnce() + Send>),
360}
aether_core/testing/utils.rs

aether_core/testing/
utils.rs
