aether_core/core/
agent.rs

1use crate::context::{
2    CompactionConfig, CompactionError, CompactionResult, Compactor, SessionUsageTracker, TokenTracker,
3};
4use crate::core::PromptCache;
5use crate::core::prompt_cache_key::derive_prompt_cache_key;
6use crate::core::queued_input::QueuedInput;
7pub use crate::core::retry_config::RetryConfig;
8use crate::core::tool_execution::{ToolAbortPolicy, ToolExecutionUpdate, ToolExecutions};
9use crate::events::{
10    AgentCommand, AgentEvent, AgentObserver, Command, CompactionId, CompactionOutcome, ContextEvent, LlmCallOutcome,
11    ModelEvent, StreamState, TaskOutcome, ToolEvent, TraceContext, TurnEvent, TurnOutcome, UserCommand,
12};
13use crate::mcp::McpHandle;
14use futures::Stream;
15use llm::{
16    AssistantReasoning, ChatMessage, Context, EncryptedReasoningContent, LlmCallPurpose, LlmError, LlmModel,
17    LlmResponse, MessageId, ModelIdentity, StopReason, StreamingModelProvider, TokenUsage, ToolCallError,
18    ToolCallRequest, ToolCallResult,
19};
20use mcp_utils::client::{CallToolError, CallToolOptions, ToolCallEvent};
21use std::collections::VecDeque;
22use std::pin::Pin;
23use std::sync::Arc;
24use std::time::Duration;
25use tokio::sync::mpsc;
26use tokio::time::sleep;
27use tokio_stream::StreamExt;
28use tokio_stream::StreamMap;
29use tokio_stream::wrappers::ReceiverStream;
30
31/// Internal event type for merging LLM and tool result streams
32#[derive(Debug)]
33#[allow(clippy::large_enum_variant)]
34enum StreamEvent {
35    LlmRequestStarted { attempt: u32 },
36    Llm(Result<LlmResponse, LlmError>),
37    ToolExecution(ToolCallEvent),
38    Command(Command),
39    InputClosed,
40    Compaction(Result<CompactionResult, CompactionError>),
41}
42
43type EventStream = Pin<Box<dyn Stream<Item = StreamEvent> + Send>>;
44
45/// Keys for the merged stream map. Tool-call IDs come from providers, so the
46/// typed key keeps them from colliding with reserved streams.
47#[derive(Debug, Clone, PartialEq, Eq, Hash)]
48enum StreamKey {
49    Input,
50    Llm,
51    Compaction,
52    Tool(String),
53}
54
55pub(crate) struct AgentConfig {
56    pub llm: Arc<dyn StreamingModelProvider>,
57    pub context: Context,
58    pub mcp: Option<McpHandle>,
59    pub tool_timeout: Duration,
60    pub compaction_config: Option<CompactionConfig>,
61    pub auto_continue: AutoContinue,
62    pub retry_config: RetryConfig,
63    pub context_window: Option<u32>,
64    pub prompt_cache: PromptCache,
65    pub observers: Vec<Box<dyn AgentObserver>>,
66    pub session_usage: SessionUsageTracker,
67}
68
69pub struct Agent {
70    llm: Arc<dyn StreamingModelProvider>,
71    context: Context,
72    mcp: Option<McpHandle>,
73    message_tx: mpsc::Sender<AgentEvent>,
74    observers: Vec<Box<dyn AgentObserver>>,
75    streams: StreamMap<StreamKey, EventStream>,
76    tool_timeout: Duration,
77    token_tracker: TokenTracker,
78    compaction_config: Option<CompactionConfig>,
79    auto_continue: AutoContinue,
80    retry_config: RetryConfig,
81    tool_executions: ToolExecutions,
82    pending_inputs: VecDeque<QueuedInput>,
83    queued_inputs: VecDeque<QueuedInput>,
84    context_window: Option<u32>,
85    prompt_cache: PromptCache,
86    turn_active: bool,
87    llm_call_active: bool,
88    active_compaction: Option<CompactionId>,
89    active_model: Option<LlmModel>,
90    session_usage: SessionUsageTracker,
91}
92
93impl Agent {
94    pub(crate) fn new(
95        config: AgentConfig,
96        command_rx: mpsc::Receiver<Command>,
97        message_tx: mpsc::Sender<AgentEvent>,
98    ) -> Self {
99        let mut streams: StreamMap<StreamKey, EventStream> = StreamMap::new();
100        let input_stream = ReceiverStream::new(command_rx)
101            .map(StreamEvent::Command)
102            .chain(futures::stream::once(async { StreamEvent::InputClosed }));
103        streams.insert(StreamKey::Input, Box::pin(input_stream));
104
105        let context_limit = config.context_window.or_else(|| config.llm.context_window());
106
107        Self {
108            llm: config.llm,
109            context: config.context,
110            mcp: config.mcp,
111            message_tx,
112            observers: config.observers,
113            streams,
114            tool_timeout: config.tool_timeout,
115            token_tracker: TokenTracker::new(context_limit),
116            compaction_config: config.compaction_config,
117            auto_continue: config.auto_continue,
118            retry_config: config.retry_config,
119            tool_executions: ToolExecutions::default(),
120            pending_inputs: VecDeque::new(),
121            queued_inputs: VecDeque::new(),
122            context_window: config.context_window,
123            prompt_cache: config.prompt_cache,
124            turn_active: false,
125            llm_call_active: false,
126            active_compaction: None,
127            active_model: None,
128            session_usage: config.session_usage,
129        }
130    }
131
132    pub fn current_model_display_name(&self) -> String {
133        self.llm.display_name()
134    }
135
136    /// Get a reference to the token tracker
137    pub fn token_tracker(&self) -> &TokenTracker {
138        &self.token_tracker
139    }
140
141    pub async fn run(mut self) {
142        let mut state = IterationState::default();
143        let mut input_closed = false;
144        self.emit_tool_definitions().await;
145
146        while let Some((stream_key, event)) = self.streams.next().await {
147            match event {
148                StreamEvent::Command(Command::UserCommand(UserCommand::Cancel)) => {
149                    self.on_user_cancel(&mut state).await;
150                }
151
152                StreamEvent::Command(Command::UserCommand(UserCommand::ClearContext)) => {
153                    self.on_user_clear_context(&mut state).await;
154                }
155
156                StreamEvent::Command(Command::UserCommand(UserCommand::Text { message_id, content })) => {
157                    if self.is_busy() {
158                        self.queued_inputs.push_back(QueuedInput::User { message_id, content });
159                    } else {
160                        self.begin_turn(QueuedInput::User { message_id, content }, &mut state).await;
161                    }
162                }
163
164                StreamEvent::Command(Command::AgentCommand(AgentCommand::SwitchModel(new_provider))) => {
165                    self.on_switch_model(new_provider).await;
166                }
167
168                StreamEvent::Command(Command::AgentCommand(AgentCommand::UpdateTools(tools))) => {
169                    self.context.set_tools(tools);
170                    self.emit_tool_definitions().await;
171                }
172
173                StreamEvent::Command(Command::AgentCommand(AgentCommand::UpdateMcpInstructions { server, body })) => {
174                    self.on_update_instruction(server, body).await;
175                }
176
177                StreamEvent::Command(Command::AgentCommand(AgentCommand::SetReasoningEffort(effort))) => {
178                    self.context.set_reasoning_effort(effort.unwrap_or_default());
179                }
180
181                StreamEvent::Command(Command::AgentCommand(AgentCommand::ReplaceConversation(messages))) => {
182                    self.on_replace_conversation(messages, &mut state).await;
183                }
184
185                StreamEvent::InputClosed => {
186                    input_closed = true;
187                }
188
189                StreamEvent::LlmRequestStarted { attempt } => {
190                    self.begin_chat_call(attempt).await;
191                }
192
193                StreamEvent::Llm(llm_event) => {
194                    self.on_llm_event(llm_event, &mut state).await;
195                }
196
197                StreamEvent::ToolExecution(tool_event) => {
198                    let StreamKey::Tool(tool_id) = stream_key else {
199                        unreachable!("tool events must come from a tool stream")
200                    };
201                    self.on_tool_execution_event(tool_id, tool_event, &mut state).await;
202                }
203
204                StreamEvent::Compaction(result) => {
205                    self.on_compaction_complete(result).await;
206                }
207            }
208
209            if state.is_complete(self.tool_executions.has_foreground())
210                && let Some(id) = state.current_message_id.take()
211            {
212                let iteration = std::mem::take(&mut state);
213                self.on_iteration_complete(id, iteration).await;
214            }
215
216            if input_closed && !self.turn_active && !self.is_busy() {
217                if self.tool_executions.is_empty() {
218                    break;
219                }
220                self.abort_in_flight_work(ToolAbortPolicy::CancelAll).await;
221            }
222        }
223
224        tracing::debug!("Agent task shutting down - input channel closed");
225    }
226
227    async fn on_iteration_complete(&mut self, id: MessageId, iteration: IterationState) {
228        let IterationState {
229            message_content,
230            reasoning_summary_text,
231            encrypted_reasoning,
232            completed_tool_calls,
233            stop_reason,
234            ..
235        } = iteration;
236        let has_tool_calls = !completed_tool_calls.is_empty();
237        let has_content = !message_content.is_empty() || !reasoning_summary_text.is_empty() || has_tool_calls;
238        let should_auto_continue = self.auto_continue.should_continue(stop_reason.as_ref());
239
240        if has_content {
241            let reasoning = AssistantReasoning::from_parts(reasoning_summary_text.clone(), encrypted_reasoning);
242            self.context.push_assistant_turn(id.clone(), &message_content, reasoning, completed_tool_calls);
243
244            self.emit(AgentEvent::text(&id, &message_content, StreamState::Complete)).await;
245
246            if !reasoning_summary_text.is_empty() {
247                self.emit(AgentEvent::thought(&id, &reasoning_summary_text, StreamState::Complete)).await;
248            }
249        }
250
251        let has_queued_input = !self.queued_inputs.is_empty();
252        if has_queued_input || has_tool_calls {
253            self.auto_continue.reset();
254            self.start_next_turn().await;
255        } else if should_auto_continue {
256            self.auto_continue.advance();
257            tracing::info!(
258                "LLM stopped with {:?}, auto-continuing (attempt {}/{})",
259                stop_reason,
260                self.auto_continue.count,
261                self.auto_continue.max
262            );
263
264            self.inject_continuation_prompt(stop_reason.as_ref()).await;
265            self.start_next_turn().await;
266        } else {
267            tracing::debug!("LLM completed turn with stop reason: {:?}", stop_reason);
268            self.auto_continue.reset();
269            self.finish_turn(TurnOutcome::Completed).await;
270        }
271    }
272
273    async fn start_next_turn(&mut self) {
274        debug_assert!(self.pending_inputs.is_empty());
275        self.pending_inputs.append(&mut self.queued_inputs);
276        if self.compaction_needed() {
277            self.begin_compaction().await;
278        } else {
279            self.start_chat_turn().await;
280        }
281    }
282
283    async fn start_chat_turn(&mut self) {
284        self.commit_pending_inputs().await;
285        self.start_llm_stream(None, 0).await;
286    }
287
288    async fn on_user_cancel(&mut self, state: &mut IterationState) {
289        self.abort_in_flight_work(ToolAbortPolicy::PreserveBackgroundAcknowledgements).await;
290        self.commit_pending_inputs().await;
291        self.queued_inputs.retain(|input| matches!(input, QueuedInput::TaskOutcome(_)));
292        self.commit_queued_inputs().await;
293        *state = IterationState::default();
294        self.finish_turn(TurnOutcome::Cancelled).await;
295    }
296
297    async fn discard_in_flight_work(&mut self, state: &mut IterationState) {
298        self.abort_in_flight_work(ToolAbortPolicy::CancelAll).await;
299        self.pending_inputs.clear();
300        self.queued_inputs.clear();
301        self.auto_continue.reset();
302        *state = IterationState::default();
303    }
304
305    async fn on_user_clear_context(&mut self, state: &mut IterationState) {
306        self.discard_in_flight_work(state).await;
307        self.context.clear_conversation();
308        self.token_tracker.reset_current_usage();
309        self.emit(AgentEvent::Context(ContextEvent::Cleared)).await;
310        self.finish_turn(TurnOutcome::Cancelled).await;
311    }
312
313    async fn on_replace_conversation(&mut self, messages: Vec<ChatMessage>, state: &mut IterationState) {
314        self.discard_in_flight_work(state).await;
315        self.context.replace_conversation(messages);
316        self.emit(self.context_usage_message()).await;
317        self.finish_turn(TurnOutcome::Cancelled).await;
318    }
319
320    async fn begin_turn(&mut self, input: QueuedInput, state: &mut IterationState) {
321        *state = IterationState::default();
322        self.auto_continue.reset();
323        self.turn_active = true;
324        let content = input.content_blocks();
325        self.emit(AgentEvent::Turn(TurnEvent::Started { content })).await;
326        self.queued_inputs.push_back(input);
327        self.start_next_turn().await;
328    }
329
330    async fn enqueue_task_outcome(&mut self, outcome: TaskOutcome, state: &mut IterationState) {
331        let input = QueuedInput::TaskOutcome(Box::new(outcome));
332        if self.is_busy() {
333            self.queued_inputs.push_back(input);
334        } else {
335            self.begin_turn(input, state).await;
336        }
337    }
338
339    async fn on_update_instruction(&mut self, server: String, body: Option<String>) {
340        self.prompt_cache.update_mcp_instruction(server, body);
341        match self.prompt_cache.render().await {
342            Ok(content) => self.context.set_system_content(content),
343            Err(e) => tracing::warn!("Failed to rebuild system prompt after instructions update: {e}"),
344        }
345    }
346
347    async fn on_switch_model(&mut self, new_provider: Box<dyn StreamingModelProvider>) {
348        let previous = self.llm.display_name();
349        let new_context_limit = self.context_window.or_else(|| new_provider.context_window());
350        self.llm = Arc::from(new_provider);
351        self.token_tracker.reset_current_usage();
352        self.token_tracker.set_context_limit(new_context_limit);
353        let new = self.llm.display_name();
354        self.emit(AgentEvent::Model(ModelEvent::Switched { previous, new })).await;
355
356        self.emit(self.context_usage_message()).await;
357    }
358
359    async fn start_llm_stream(&mut self, delay: Option<Duration>, attempt: u32) {
360        self.refresh_prompt_cache_key();
361        self.streams.remove(&StreamKey::Llm);
362        let stream: EventStream = match delay {
363            None => {
364                self.begin_chat_call(attempt).await;
365                Box::pin(self.llm.stream_response(&self.context).map(StreamEvent::Llm))
366            }
367            Some(delay) => {
368                self.emit(AgentEvent::Turn(TurnEvent::RetryScheduled {
369                    purpose: LlmCallPurpose::Chat,
370                    attempt,
371                    max_attempts: self.retry_config.max_attempts,
372                    delay_ms: u64::try_from(delay.as_millis()).unwrap_or(u64::MAX),
373                }))
374                .await;
375                let llm = Arc::clone(&self.llm);
376                let context = self.context.clone();
377                Box::pin(async_stream::stream! {
378                    sleep(delay).await;
379                    yield StreamEvent::LlmRequestStarted { attempt };
380                    let mut inner = llm.stream_response(&context);
381                    while let Some(item) = inner.next().await {
382                        yield StreamEvent::Llm(item);
383                    }
384                })
385            }
386        };
387        self.streams.insert(StreamKey::Llm, stream);
388    }
389
390    async fn on_llm_error(&mut self, error: LlmError, state: &mut IterationState) {
391        let will_retry = error.is_retryable() && state.retry_attempt < self.retry_config.max_attempts;
392        let outcome = LlmCallOutcome::from_llm_error(&error, will_retry);
393        let error_message = error.to_string();
394        self.finish_chat_call(outcome).await;
395
396        if !will_retry {
397            self.finish_turn(TurnOutcome::failed(error_message)).await;
398            return;
399        }
400
401        state.retry_attempt += 1;
402        let delay = self.retry_config.compute_delay(state.retry_attempt);
403
404        tracing::warn!(
405            attempt = state.retry_attempt,
406            max_attempts = self.retry_config.max_attempts,
407            delay_ms = u64::try_from(delay.as_millis()).unwrap_or(u64::MAX),
408            error = %error,
409            "Retrying LLM request after transient failure"
410        );
411
412        self.tool_executions.retire_foreground();
413        self.start_llm_stream(Some(delay), state.retry_attempt).await;
414    }
415
416    fn is_busy(&self) -> bool {
417        self.streams.contains_key(&StreamKey::Llm)
418            || self.streams.contains_key(&StreamKey::Compaction)
419            || self.tool_executions.has_foreground()
420    }
421
422    async fn abort_in_flight_work(&mut self, tool_policy: ToolAbortPolicy) {
423        if self.llm_call_active {
424            self.finish_chat_call(LlmCallOutcome::Cancelled).await;
425        }
426        if self.streams.remove(&StreamKey::Compaction).is_some() {
427            let compaction_id = self.active_compaction.take().expect("active compaction stream has an identity");
428            self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded {
429                purpose: LlmCallPurpose::Compaction,
430                outcome: LlmCallOutcome::Cancelled,
431            }))
432            .await;
433            self.emit(AgentEvent::Context(ContextEvent::CompactionEnded {
434                compaction_id,
435                outcome: CompactionOutcome::Cancelled,
436            }))
437            .await;
438        }
439        self.streams.remove(&StreamKey::Llm);
440        for tool_id in self.tool_executions.abort(&tool_policy) {
441            self.streams.remove(&StreamKey::Tool(tool_id));
442        }
443    }
444
445    /// Inject a continuation prompt when the LLM stops due to a resumable reason.
446    async fn inject_continuation_prompt(&mut self, stop_reason: Option<&StopReason>) {
447        let reason = stop_reason.map_or_else(|| "Unknown".to_string(), |reason| format!("{reason:?}"));
448        let message_id = MessageId::new();
449        let content = vec![llm::ContentBlock::text(format!(
450            "<system-notification>The LLM API stopped with reason '{reason}'. Continue from where you left off and finish your task.</system-notification>"
451        ))];
452        self.context.add_message(ChatMessage::user_with_id(message_id.clone(), content.clone()));
453        self.emit(AgentEvent::Turn(TurnEvent::AutoContinue {
454            attempt: self.auto_continue.count,
455            max_attempts: self.auto_continue.max,
456            message_id,
457            content,
458        }))
459        .await;
460    }
461
462    async fn on_llm_event(&mut self, result: Result<LlmResponse, LlmError>, state: &mut IterationState) {
463        use LlmResponse::{
464            Done, EncryptedReasoning, Error, Reasoning, Start, Text, ToolRequestArg, ToolRequestComplete,
465            ToolRequestStart, Usage,
466        };
467
468        let response = match result {
469            Ok(response) => response,
470            Err(e) => {
471                self.on_llm_error(e, state).await;
472                return;
473            }
474        };
475
476        match response {
477            Start => state.on_llm_start(MessageId::new()),
478
479            Text { chunk } => {
480                self.handle_llm_text(chunk, state).await;
481            }
482
483            Reasoning { chunk } => {
484                state.reasoning_summary_text.push_str(&chunk);
485                if let Some(id) = state.current_message_id.clone() {
486                    self.emit(AgentEvent::thought(&id, &chunk, StreamState::Partial)).await;
487                }
488            }
489
490            EncryptedReasoning { id, content } => {
491                if let Some(model) = self.active_model.clone() {
492                    state.encrypted_reasoning = Some(EncryptedReasoningContent { id, model, content });
493                }
494            }
495
496            ToolRequestStart { id, name } => {
497                self.emit(AgentEvent::Tool(ToolEvent::InputStarted { id, name })).await;
498            }
499
500            ToolRequestArg { id, chunk } => {
501                self.emit(AgentEvent::Tool(ToolEvent::InputDelta { id, chunk })).await;
502            }
503
504            ToolRequestComplete { tool_call } => {
505                self.handle_tool_completion(tool_call).await;
506            }
507
508            Done { stop_reason } => {
509                state.llm_done = true;
510                state.stop_reason = stop_reason;
511                self.finish_chat_call(LlmCallOutcome::Completed {
512                    stop_reason: state.stop_reason.clone(),
513                    usage: state.call_usage.take(),
514                })
515                .await;
516            }
517
518            Error { message } => {
519                self.finish_chat_call(LlmCallOutcome::failed(message.clone(), false)).await;
520                self.finish_turn(TurnOutcome::failed(message)).await;
521            }
522
523            Usage { tokens: sample } => {
524                self.handle_llm_usage(sample, state).await;
525            }
526        }
527    }
528
529    async fn handle_llm_text(&mut self, chunk: String, state: &mut IterationState) {
530        state.message_content.push_str(&chunk);
531
532        if let Some(id) = state.current_message_id.clone() {
533            self.emit(AgentEvent::text(&id, &chunk, StreamState::Partial)).await;
534        }
535    }
536
537    async fn handle_tool_completion(&mut self, tool_call: ToolCallRequest) {
538        let cancel = self.tool_executions.start(tool_call.clone());
539
540        let tool_id = tool_call.id.clone();
541        tracing::debug!("Tool execution started: {} ({})", tool_call.name, tool_id);
542        self.emit(AgentEvent::Tool(ToolEvent::Call { request: tool_call.clone() })).await;
543
544        let Some(mcp) = self.mcp.clone() else {
545            let stream = futures::stream::once(async {
546                StreamEvent::ToolExecution(ToolCallEvent::Complete(Err(CallToolError::Unavailable {
547                    message: "MCP runtime is not available".to_string(),
548                })))
549            });
550            self.streams.insert(StreamKey::Tool(tool_id), Box::pin(stream));
551            return;
552        };
553
554        let trace_context = self.observers.iter().find_map(|observer| observer.tool_trace_context(&tool_id));
555        let options = CallToolOptions {
556            timeout: self.tool_timeout,
557            meta: trace_context.as_ref().map(TraceContext::to_meta),
558            cancel,
559        };
560        let stream =
561            mcp.call_model_visible(tool_call.name, &tool_call.arguments, options).map(StreamEvent::ToolExecution);
562        self.streams.insert(StreamKey::Tool(tool_id), Box::pin(stream));
563    }
564
565    async fn handle_llm_usage(&mut self, sample: TokenUsage, state: &mut IterationState) {
566        state.call_usage = Some(sample);
567        self.token_tracker.record_usage(sample);
568        let ratio_pct = self.token_tracker.usage_ratio().map(|r| r * 100.0);
569        let remaining = self.token_tracker.tokens_remaining();
570        tracing::debug!(?sample, ?ratio_pct, ?remaining, "Token usage");
571
572        self.emit(self.context_usage_message()).await;
573        self.emit_session_usage(LlmCallPurpose::Chat, sample).await;
574    }
575
576    async fn emit_session_usage(&mut self, purpose: LlmCallPurpose, tokens: TokenUsage) {
577        let model = ModelIdentity::of(self.active_model.as_ref());
578        let event = self.session_usage.record(purpose, model, tokens);
579        self.emit(AgentEvent::SessionUsage(event)).await;
580    }
581
582    fn context_usage_message(&self) -> AgentEvent {
583        AgentEvent::Context(ContextEvent::UsageUpdated { usage: self.token_tracker.snapshot().clone() })
584    }
585
586    fn compaction_needed(&self) -> bool {
587        self.compaction_config.as_ref().is_some_and(|config| {
588            self.token_tracker.needs_compaction(self.context.estimated_token_count(), config.threshold)
589        })
590    }
591
592    async fn begin_compaction(&mut self) {
593        tracing::info!("Starting context compaction - {} messages", self.context.message_count());
594        let compaction_id = CompactionId::new();
595        self.active_compaction = Some(compaction_id.clone());
596        self.emit(AgentEvent::Context(ContextEvent::CompactionStarted {
597            compaction_id,
598            message_count: self.context.message_count(),
599        }))
600        .await;
601        let started = self.begin_llm_call(LlmCallPurpose::Compaction, 0);
602        self.emit(started).await;
603
604        let compactor = Compactor::new(self.llm.clone());
605        let context = self.context.clone();
606        let stream: EventStream =
607            Box::pin(futures::stream::once(async move { StreamEvent::Compaction(compactor.compact(context).await) }));
608        self.streams.insert(StreamKey::Compaction, stream);
609    }
610
611    async fn on_compaction_complete(&mut self, result: Result<CompactionResult, CompactionError>) {
612        let compaction_id = self.active_compaction.take().expect("completed compaction has an identity");
613        if let Ok(result) = &result
614            && let Some(usage) = result.usage
615        {
616            self.emit_session_usage(LlmCallPurpose::Compaction, usage).await;
617        }
618        let outcome = match &result {
619            Ok(result) => LlmCallOutcome::Completed { stop_reason: None, usage: result.usage },
620            Err(e) => LlmCallOutcome::failed(e.to_string(), false),
621        };
622        self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded { purpose: LlmCallPurpose::Compaction, outcome })).await;
623
624        match result {
625            Ok(result) => {
626                tracing::info!("Context compacted: {} messages removed", result.messages_removed);
627                let message_id = MessageId::new();
628                self.context = self.context.with_compacted_summary(message_id.clone(), &result.summary);
629                self.token_tracker.reset_current_usage();
630                self.emit(AgentEvent::Context(ContextEvent::CompactionResult {
631                    compaction_id: compaction_id.clone(),
632                    message_id,
633                    summary: result.summary,
634                    messages_removed: result.messages_removed,
635                }))
636                .await;
637                self.emit(AgentEvent::Context(ContextEvent::CompactionEnded {
638                    compaction_id,
639                    outcome: CompactionOutcome::Completed,
640                }))
641                .await;
642            }
643            Err(e) => {
644                tracing::warn!("Context compaction failed: {e}");
645                self.emit(AgentEvent::Context(ContextEvent::CompactionEnded {
646                    compaction_id,
647                    outcome: CompactionOutcome::Failed { error: e.to_string() },
648                }))
649                .await;
650            }
651        }
652
653        self.start_chat_turn().await;
654    }
655
656    async fn on_tool_execution_event(&mut self, tool_id: String, event: ToolCallEvent, state: &mut IterationState) {
657        match self.tool_executions.on_event(&tool_id, event) {
658            ToolExecutionUpdate::Event(event) => {
659                if let ToolEvent::SubAgentProgress { payload, .. } = &event
660                    && let AgentEvent::SessionUsage(child) = &payload.event
661                {
662                    let folded = self.session_usage.record_child(&payload.task_id, child.clone());
663                    self.emit(AgentEvent::SessionUsage(folded)).await;
664                }
665                self.emit(AgentEvent::Tool(event)).await;
666            }
667            ToolExecutionUpdate::Completed { result, event } => {
668                self.streams.remove(&StreamKey::Tool(tool_id));
669                state.completed_tool_calls.push(result);
670                self.emit(AgentEvent::Tool(event)).await;
671            }
672            ToolExecutionUpdate::TaskCreated { result, event } => {
673                state.completed_tool_calls.push(Ok(result));
674                self.emit(AgentEvent::Tool(event)).await;
675            }
676            ToolExecutionUpdate::TaskCompleted(outcome) => {
677                self.streams.remove(&StreamKey::Tool(tool_id));
678                self.enqueue_task_outcome(outcome, state).await;
679            }
680            ToolExecutionUpdate::TaskCancelled(outcome) => {
681                self.streams.remove(&StreamKey::Tool(tool_id));
682                self.record_task_outcome(outcome).await;
683            }
684            ToolExecutionUpdate::Retired => {
685                self.streams.remove(&StreamKey::Tool(tool_id));
686            }
687            ToolExecutionUpdate::Ignored => {
688                tracing::debug!(%tool_id, "Ignoring unexpected tool execution event");
689            }
690        }
691    }
692
693    async fn record_task_outcome(&mut self, outcome: TaskOutcome) {
694        self.context.add_message(outcome.context_message());
695        self.emit(AgentEvent::Tool(outcome.into())).await;
696    }
697
698    fn refresh_prompt_cache_key(&mut self) {
699        let key = derive_prompt_cache_key(self.llm.as_ref(), &self.context);
700        self.context.set_prompt_cache_key(Some(key));
701    }
702
703    async fn commit_pending_inputs(&mut self) {
704        let inputs = std::mem::take(&mut self.pending_inputs);
705        self.commit_inputs(inputs).await;
706    }
707
708    async fn commit_queued_inputs(&mut self) {
709        let inputs = std::mem::take(&mut self.queued_inputs);
710        self.commit_inputs(inputs).await;
711    }
712
713    async fn commit_inputs(&mut self, inputs: VecDeque<QueuedInput>) {
714        for input in inputs {
715            match input {
716                QueuedInput::User { message_id, content } => {
717                    self.context.add_message(ChatMessage::user_with_id(message_id, content));
718                }
719                QueuedInput::TaskOutcome(outcome) => self.record_task_outcome(*outcome).await,
720            }
721        }
722    }
723
724    async fn emit_tool_definitions(&mut self) {
725        let tools = self.context.tools().clone();
726        if !tools.is_empty() {
727            self.emit(AgentEvent::Tool(ToolEvent::DefinitionsUpdated { tools })).await;
728        }
729    }
730
731    async fn emit(&mut self, message: AgentEvent) {
732        for observer in &mut self.observers {
733            observer.on_event(&message);
734        }
735
736        if let Err(e) = self.message_tx.send(message).await {
737            tracing::warn!("Failed to send agent message: {e:?}");
738        }
739    }
740
741    async fn finish_turn(&mut self, outcome: TurnOutcome) {
742        if std::mem::take(&mut self.turn_active) {
743            self.emit(AgentEvent::turn_ended(outcome)).await;
744        }
745    }
746
747    async fn begin_chat_call(&mut self, attempt: u32) {
748        self.llm_call_active = true;
749        let started = self.begin_llm_call(LlmCallPurpose::Chat, attempt);
750        if let Some(system_prompt) = self.context.system_content() {
751            for observer in &mut self.observers {
752                observer.on_system_prompt(system_prompt);
753            }
754        }
755        self.emit(started).await;
756    }
757
758    async fn finish_chat_call(&mut self, outcome: LlmCallOutcome) {
759        if std::mem::take(&mut self.llm_call_active) {
760            self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded { purpose: LlmCallPurpose::Chat, outcome })).await;
761        }
762    }
763
764    fn begin_llm_call(&mut self, purpose: LlmCallPurpose, attempt: u32) -> AgentEvent {
765        self.active_model = self.llm.model();
766        AgentEvent::Turn(TurnEvent::LlmCallStarted {
767            purpose,
768            model: ModelIdentity::of(self.active_model.as_ref()),
769            display_name: self.llm.display_name(),
770            attempt,
771            max_attempts: self.retry_config.max_attempts,
772        })
773    }
774}
775
776pub(crate) struct AutoContinue {
777    max: u32,
778    count: u32,
779}
780
781impl AutoContinue {
782    pub(crate) fn new(max: u32) -> Self {
783        Self { max, count: 0 }
784    }
785
786    fn reset(&mut self) {
787        self.count = 0;
788    }
789
790    fn should_continue(&self, stop_reason: Option<&StopReason>) -> bool {
791        matches!(stop_reason, Some(StopReason::Length)) && self.count < self.max
792    }
793
794    fn advance(&mut self) {
795        self.count += 1;
796    }
797}
798
799#[derive(Debug, Default)]
800struct IterationState {
801    current_message_id: Option<MessageId>,
802    message_content: String,
803    reasoning_summary_text: String,
804    encrypted_reasoning: Option<EncryptedReasoningContent>,
805    completed_tool_calls: Vec<Result<ToolCallResult, ToolCallError>>,
806    llm_done: bool,
807    stop_reason: Option<StopReason>,
808    retry_attempt: u32,
809    call_usage: Option<TokenUsage>,
810}
811
812impl IterationState {
813    fn on_llm_start(&mut self, message_id: MessageId) {
814        self.current_message_id = Some(message_id);
815        self.message_content.clear();
816        self.reasoning_summary_text.clear();
817        self.encrypted_reasoning = None;
818        self.stop_reason = None;
819        self.call_usage = None;
820    }
821
822    fn is_complete(&self, has_foreground_tools: bool) -> bool {
823        self.llm_done && !has_foreground_tools
824    }
825}
aether_core/core/agent.rs

aether_core/core/
agent.rs
