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, CompactionOutcome, ContextEvent, LlmCallOutcome, ModelEvent,
11    StreamState, TaskOutcome, ToolEvent, TraceContext, TurnEvent, TurnOutcome, UserCommand,
12};
13use crate::mcp::McpHandle;
14use futures::Stream;
15use llm::types::IsoString;
16use llm::{
17    AssistantReasoning, ChatMessage, Context, EncryptedReasoningContent, LlmCallPurpose, LlmError, LlmModel,
18    LlmResponse, ModelIdentity, StopReason, StreamingModelProvider, TokenUsage, ToolCallError, ToolCallRequest,
19    ToolCallResult,
20};
21use mcp_utils::client::{CallToolError, CallToolOptions, ToolCallEvent};
22use std::collections::VecDeque;
23use std::pin::Pin;
24use std::sync::Arc;
25use std::time::Duration;
26use tokio::sync::mpsc;
27use tokio::time::sleep;
28use tokio_stream::StreamExt;
29use tokio_stream::StreamMap;
30use tokio_stream::wrappers::ReceiverStream;
31
32/// Internal event type for merging LLM and tool result streams
33#[derive(Debug)]
34#[allow(clippy::large_enum_variant)]
35enum StreamEvent {
36    LlmRequestStarted { attempt: u32 },
37    Llm(Result<LlmResponse, LlmError>),
38    ToolExecution(ToolCallEvent),
39    Command(Command),
40    InputClosed,
41    Compaction(Result<CompactionResult, CompactionError>),
42}
43
44type EventStream = Pin<Box<dyn Stream<Item = StreamEvent> + Send>>;
45
46/// Keys for the merged stream map. Tool-call IDs come from providers, so the
47/// typed key keeps them from colliding with reserved streams.
48#[derive(Debug, Clone, PartialEq, Eq, Hash)]
49enum StreamKey {
50    Input,
51    Llm,
52    Compaction,
53    Tool(String),
54}
55
56pub(crate) struct AgentConfig {
57    pub llm: Arc<dyn StreamingModelProvider>,
58    pub context: Context,
59    pub mcp: Option<McpHandle>,
60    pub tool_timeout: Duration,
61    pub compaction_config: Option<CompactionConfig>,
62    pub auto_continue: AutoContinue,
63    pub retry_config: RetryConfig,
64    pub context_window: Option<u32>,
65    pub prompt_cache: PromptCache,
66    pub observers: Vec<Box<dyn AgentObserver>>,
67    pub session_usage: SessionUsageTracker,
68}
69
70pub struct Agent {
71    llm: Arc<dyn StreamingModelProvider>,
72    context: Context,
73    mcp: Option<McpHandle>,
74    message_tx: mpsc::Sender<AgentEvent>,
75    observers: Vec<Box<dyn AgentObserver>>,
76    streams: StreamMap<StreamKey, EventStream>,
77    tool_timeout: Duration,
78    token_tracker: TokenTracker,
79    compaction_config: Option<CompactionConfig>,
80    auto_continue: AutoContinue,
81    retry_config: RetryConfig,
82    tool_executions: ToolExecutions,
83    pending_inputs: VecDeque<QueuedInput>,
84    queued_inputs: VecDeque<QueuedInput>,
85    context_window: Option<u32>,
86    prompt_cache: PromptCache,
87    turn_active: bool,
88    llm_call_active: bool,
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_model: None,
127            session_usage: config.session_usage,
128        }
129    }
130
131    pub fn current_model_display_name(&self) -> String {
132        self.llm.display_name()
133    }
134
135    /// Get a reference to the token tracker
136    pub fn token_tracker(&self) -> &TokenTracker {
137        &self.token_tracker
138    }
139
140    pub async fn run(mut self) {
141        let mut state = IterationState::default();
142        let mut input_closed = false;
143        self.emit_tool_definitions().await;
144
145        while let Some((stream_key, event)) = self.streams.next().await {
146            match event {
147                StreamEvent::Command(Command::UserCommand(UserCommand::Cancel)) => {
148                    self.on_user_cancel(&mut state).await;
149                }
150
151                StreamEvent::Command(Command::UserCommand(UserCommand::ClearContext)) => {
152                    self.on_user_clear_context(&mut state).await;
153                }
154
155                StreamEvent::Command(Command::UserCommand(UserCommand::Text { content })) => {
156                    if self.is_busy() {
157                        self.queued_inputs.push_back(QueuedInput::User(content));
158                    } else {
159                        self.begin_turn(QueuedInput::User(content), &mut state).await;
160                    }
161                }
162
163                StreamEvent::Command(Command::AgentCommand(AgentCommand::SwitchModel(new_provider))) => {
164                    self.on_switch_model(new_provider).await;
165                }
166
167                StreamEvent::Command(Command::AgentCommand(AgentCommand::UpdateTools(tools))) => {
168                    self.context.set_tools(tools);
169                    self.emit_tool_definitions().await;
170                }
171
172                StreamEvent::Command(Command::AgentCommand(AgentCommand::UpdateMcpInstructions { server, body })) => {
173                    self.on_update_instruction(server, body).await;
174                }
175
176                StreamEvent::Command(Command::AgentCommand(AgentCommand::SetReasoningEffort(effort))) => {
177                    self.context.set_reasoning_effort(effort);
178                }
179
180                StreamEvent::Command(Command::AgentCommand(AgentCommand::ReplaceConversation(messages))) => {
181                    self.on_replace_conversation(messages, &mut state).await;
182                }
183
184                StreamEvent::InputClosed => {
185                    input_closed = true;
186                }
187
188                StreamEvent::LlmRequestStarted { attempt } => {
189                    self.begin_chat_call(attempt).await;
190                }
191
192                StreamEvent::Llm(llm_event) => {
193                    self.on_llm_event(llm_event, &mut state).await;
194                }
195
196                StreamEvent::ToolExecution(tool_event) => {
197                    let StreamKey::Tool(tool_id) = stream_key else {
198                        unreachable!("tool events must come from a tool stream")
199                    };
200                    self.on_tool_execution_event(tool_id, tool_event, &mut state).await;
201                }
202
203                StreamEvent::Compaction(result) => {
204                    self.on_compaction_complete(result).await;
205                }
206            }
207
208            if state.is_complete(self.tool_executions.has_foreground())
209                && let Some(id) = state.current_message_id.take()
210            {
211                let iteration = std::mem::take(&mut state);
212                self.on_iteration_complete(id, iteration).await;
213            }
214
215            if input_closed && !self.turn_active && !self.is_busy() {
216                if self.tool_executions.is_empty() {
217                    break;
218                }
219                self.abort_in_flight_work(ToolAbortPolicy::CancelAll).await;
220            }
221        }
222
223        tracing::debug!("Agent task shutting down - input channel closed");
224    }
225
226    async fn on_iteration_complete(&mut self, id: String, iteration: IterationState) {
227        let IterationState {
228            message_content,
229            reasoning_summary_text,
230            encrypted_reasoning,
231            completed_tool_calls,
232            stop_reason,
233            ..
234        } = iteration;
235        let has_tool_calls = !completed_tool_calls.is_empty();
236        let has_content = !message_content.is_empty() || has_tool_calls;
237        let should_auto_continue = self.auto_continue.should_continue(stop_reason.as_ref());
238
239        if has_content {
240            let reasoning = AssistantReasoning::from_parts(reasoning_summary_text.clone(), encrypted_reasoning);
241            self.context.push_assistant_turn(&message_content, reasoning, completed_tool_calls);
242
243            self.emit(AgentEvent::text(&id, &message_content, StreamState::Complete)).await;
244
245            if !reasoning_summary_text.is_empty() {
246                self.emit(AgentEvent::thought(&id, &reasoning_summary_text, StreamState::Complete)).await;
247            }
248        }
249
250        let has_queued_input = !self.queued_inputs.is_empty();
251        if has_queued_input || has_tool_calls {
252            self.auto_continue.reset();
253            self.start_next_turn().await;
254        } else if should_auto_continue {
255            self.auto_continue.advance();
256            tracing::info!(
257                "LLM stopped with {:?}, auto-continuing (attempt {}/{})",
258                stop_reason,
259                self.auto_continue.count,
260                self.auto_continue.max
261            );
262
263            self.emit(AgentEvent::Turn(TurnEvent::AutoContinue {
264                attempt: self.auto_continue.count,
265                max_attempts: self.auto_continue.max,
266            }))
267            .await;
268
269            self.inject_continuation_prompt(&message_content, stop_reason.as_ref());
270            self.start_next_turn().await;
271        } else {
272            tracing::debug!("LLM completed turn with stop reason: {:?}", stop_reason);
273            self.auto_continue.reset();
274            self.finish_turn(TurnOutcome::Completed).await;
275        }
276    }
277
278    async fn start_next_turn(&mut self) {
279        debug_assert!(self.pending_inputs.is_empty());
280        self.pending_inputs.append(&mut self.queued_inputs);
281        if self.compaction_needed() {
282            self.begin_compaction().await;
283        } else {
284            self.start_chat_turn().await;
285        }
286    }
287
288    async fn start_chat_turn(&mut self) {
289        self.commit_pending_inputs().await;
290        self.start_llm_stream(None, 0).await;
291    }
292
293    async fn on_user_cancel(&mut self, state: &mut IterationState) {
294        self.abort_in_flight_work(ToolAbortPolicy::PreserveBackgroundAcknowledgements).await;
295        self.commit_pending_inputs().await;
296        self.queued_inputs.retain(|input| matches!(input, QueuedInput::TaskOutcome(_)));
297        self.commit_queued_inputs().await;
298        *state = IterationState::default();
299        self.finish_turn(TurnOutcome::Cancelled).await;
300    }
301
302    async fn discard_in_flight_work(&mut self, state: &mut IterationState) {
303        self.abort_in_flight_work(ToolAbortPolicy::CancelAll).await;
304        self.pending_inputs.clear();
305        self.queued_inputs.clear();
306        self.auto_continue.reset();
307        *state = IterationState::default();
308    }
309
310    async fn on_user_clear_context(&mut self, state: &mut IterationState) {
311        self.discard_in_flight_work(state).await;
312        self.context.clear_conversation();
313        self.token_tracker.reset_current_usage();
314        self.emit(AgentEvent::Context(ContextEvent::Cleared)).await;
315        self.finish_turn(TurnOutcome::Cancelled).await;
316    }
317
318    async fn on_replace_conversation(&mut self, messages: Vec<ChatMessage>, state: &mut IterationState) {
319        self.discard_in_flight_work(state).await;
320        self.context.replace_conversation(messages);
321        self.emit(self.context_usage_message()).await;
322        self.finish_turn(TurnOutcome::Cancelled).await;
323    }
324
325    async fn begin_turn(&mut self, input: QueuedInput, state: &mut IterationState) {
326        *state = IterationState::default();
327        self.auto_continue.reset();
328        self.turn_active = true;
329        let content = input.content_blocks();
330        self.emit(AgentEvent::Turn(TurnEvent::Started { content })).await;
331        self.queued_inputs.push_back(input);
332        self.start_next_turn().await;
333    }
334
335    async fn enqueue_task_outcome(&mut self, outcome: TaskOutcome, state: &mut IterationState) {
336        let input = QueuedInput::TaskOutcome(Box::new(outcome));
337        if self.is_busy() {
338            self.queued_inputs.push_back(input);
339        } else {
340            self.begin_turn(input, state).await;
341        }
342    }
343
344    async fn on_update_instruction(&mut self, server: String, body: Option<String>) {
345        self.prompt_cache.update_mcp_instruction(server, body);
346        match self.prompt_cache.render().await {
347            Ok(content) => self.context.set_system_content(content),
348            Err(e) => tracing::warn!("Failed to rebuild system prompt after instructions update: {e}"),
349        }
350    }
351
352    async fn on_switch_model(&mut self, new_provider: Box<dyn StreamingModelProvider>) {
353        let previous = self.llm.display_name();
354        let new_context_limit = self.context_window.or_else(|| new_provider.context_window());
355        self.llm = Arc::from(new_provider);
356        self.token_tracker.reset_current_usage();
357        self.token_tracker.set_context_limit(new_context_limit);
358        let new = self.llm.display_name();
359        self.emit(AgentEvent::Model(ModelEvent::Switched { previous, new })).await;
360
361        self.emit(self.context_usage_message()).await;
362    }
363
364    async fn start_llm_stream(&mut self, delay: Option<Duration>, attempt: u32) {
365        self.refresh_prompt_cache_key();
366        self.streams.remove(&StreamKey::Llm);
367        let stream: EventStream = match delay {
368            None => {
369                self.begin_chat_call(attempt).await;
370                Box::pin(self.llm.stream_response(&self.context).map(StreamEvent::Llm))
371            }
372            Some(delay) => {
373                self.emit(AgentEvent::Turn(TurnEvent::RetryScheduled {
374                    purpose: LlmCallPurpose::Chat,
375                    attempt,
376                    max_attempts: self.retry_config.max_attempts,
377                    delay_ms: u64::try_from(delay.as_millis()).unwrap_or(u64::MAX),
378                }))
379                .await;
380                let llm = Arc::clone(&self.llm);
381                let context = self.context.clone();
382                Box::pin(async_stream::stream! {
383                    sleep(delay).await;
384                    yield StreamEvent::LlmRequestStarted { attempt };
385                    let mut inner = llm.stream_response(&context);
386                    while let Some(item) = inner.next().await {
387                        yield StreamEvent::Llm(item);
388                    }
389                })
390            }
391        };
392        self.streams.insert(StreamKey::Llm, stream);
393    }
394
395    async fn on_llm_error(&mut self, error: LlmError, state: &mut IterationState) {
396        let will_retry = error.is_retryable() && state.retry_attempt < self.retry_config.max_attempts;
397        let outcome = LlmCallOutcome::from_llm_error(&error, will_retry);
398        let error_message = error.to_string();
399        self.finish_chat_call(outcome).await;
400
401        if !will_retry {
402            self.finish_turn(TurnOutcome::Failed { error: error_message }).await;
403            return;
404        }
405
406        state.retry_attempt += 1;
407        let delay = self.retry_config.compute_delay(state.retry_attempt);
408
409        tracing::warn!(
410            attempt = state.retry_attempt,
411            max_attempts = self.retry_config.max_attempts,
412            delay_ms = u64::try_from(delay.as_millis()).unwrap_or(u64::MAX),
413            error = %error,
414            "Retrying LLM request after transient failure"
415        );
416
417        self.tool_executions.retire_foreground();
418        self.start_llm_stream(Some(delay), state.retry_attempt).await;
419    }
420
421    fn is_busy(&self) -> bool {
422        self.streams.contains_key(&StreamKey::Llm)
423            || self.streams.contains_key(&StreamKey::Compaction)
424            || self.tool_executions.has_foreground()
425    }
426
427    async fn abort_in_flight_work(&mut self, tool_policy: ToolAbortPolicy) {
428        if self.llm_call_active {
429            self.finish_chat_call(LlmCallOutcome::Cancelled).await;
430        }
431        if self.streams.remove(&StreamKey::Compaction).is_some() {
432            self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded {
433                purpose: LlmCallPurpose::Compaction,
434                outcome: LlmCallOutcome::Cancelled,
435            }))
436            .await;
437            self.emit(AgentEvent::Context(ContextEvent::CompactionEnded { outcome: CompactionOutcome::Cancelled }))
438                .await;
439        }
440        self.streams.remove(&StreamKey::Llm);
441        for tool_id in self.tool_executions.abort(&tool_policy) {
442            self.streams.remove(&StreamKey::Tool(tool_id));
443        }
444    }
445
446    /// Inject a continuation prompt when the LLM stops due to a resumable reason.
447    fn inject_continuation_prompt(&mut self, previous_response: &str, stop_reason: Option<&StopReason>) {
448        if !previous_response.is_empty() {
449            self.context.add_message(ChatMessage::Assistant {
450                content: previous_response.to_string(),
451                reasoning: AssistantReasoning::default(),
452                timestamp: IsoString::now(),
453                tool_calls: Vec::new(),
454            });
455        }
456
457        let reason = stop_reason.map_or_else(|| "Unknown".to_string(), |reason| format!("{reason:?}"));
458
459        self.context.add_message(ChatMessage::User {
460            content: vec![llm::ContentBlock::text(format!(
461                "<system-notification>The LLM API stopped with reason '{reason}'. Continue from where you left off and finish your task.</system-notification>"
462            ))],
463            timestamp: IsoString::now(),
464        });
465    }
466
467    async fn on_llm_event(&mut self, result: Result<LlmResponse, LlmError>, state: &mut IterationState) {
468        use LlmResponse::{
469            Done, EncryptedReasoning, Error, Reasoning, Start, Text, ToolRequestArg, ToolRequestComplete,
470            ToolRequestStart, Usage,
471        };
472
473        let response = match result {
474            Ok(response) => response,
475            Err(e) => {
476                self.on_llm_error(e, state).await;
477                return;
478            }
479        };
480
481        match response {
482            Start { message_id } => {
483                state.on_llm_start(message_id);
484            }
485
486            Text { chunk } => {
487                self.handle_llm_text(chunk, state).await;
488            }
489
490            Reasoning { chunk } => {
491                state.reasoning_summary_text.push_str(&chunk);
492                if let Some(id) = state.current_message_id.clone() {
493                    self.emit(AgentEvent::thought(&id, &chunk, StreamState::Partial)).await;
494                }
495            }
496
497            EncryptedReasoning { id, content } => {
498                if let Some(model) = self.active_model.clone() {
499                    state.encrypted_reasoning = Some(EncryptedReasoningContent { id, model, content });
500                }
501            }
502
503            ToolRequestStart { id, name } => {
504                let request = ToolCallRequest { id, name, arguments: String::new() };
505                self.emit(AgentEvent::Tool(ToolEvent::Call { request })).await;
506            }
507
508            ToolRequestArg { id, chunk } => {
509                self.emit(AgentEvent::Tool(ToolEvent::CallUpdate { tool_call_id: id, chunk })).await;
510            }
511
512            ToolRequestComplete { tool_call } => {
513                self.handle_tool_completion(tool_call).await;
514            }
515
516            Done { stop_reason } => {
517                state.llm_done = true;
518                state.stop_reason = stop_reason;
519                self.finish_chat_call(LlmCallOutcome::Completed {
520                    stop_reason: state.stop_reason.clone(),
521                    usage: state.call_usage.take(),
522                })
523                .await;
524            }
525
526            Error { message } => {
527                self.finish_chat_call(LlmCallOutcome::failed(message.clone(), false)).await;
528                self.finish_turn(TurnOutcome::Failed { error: message }).await;
529            }
530
531            Usage { tokens: sample } => {
532                self.handle_llm_usage(sample, state).await;
533            }
534        }
535    }
536
537    async fn handle_llm_text(&mut self, chunk: String, state: &mut IterationState) {
538        state.message_content.push_str(&chunk);
539
540        if let Some(id) = state.current_message_id.clone() {
541            self.emit(AgentEvent::text(&id, &chunk, StreamState::Partial)).await;
542        }
543    }
544
545    async fn handle_tool_completion(&mut self, tool_call: ToolCallRequest) {
546        let cancel = self.tool_executions.start(tool_call.clone());
547
548        let tool_id = tool_call.id.clone();
549        tracing::debug!("Tool execution started: {} ({})", tool_call.name, tool_id);
550        self.emit(AgentEvent::Tool(ToolEvent::ExecutionStarted {
551            tool_id: tool_id.clone(),
552            tool_name: tool_call.name.clone(),
553        }))
554        .await;
555
556        let Some(mcp) = self.mcp.clone() else {
557            let stream = futures::stream::once(async {
558                StreamEvent::ToolExecution(ToolCallEvent::Complete(Err(CallToolError::Unavailable {
559                    message: "MCP runtime is not available".to_string(),
560                })))
561            });
562            self.streams.insert(StreamKey::Tool(tool_id), Box::pin(stream));
563            return;
564        };
565
566        let trace_context = self.observers.iter().find_map(|observer| observer.tool_trace_context(&tool_id));
567        let options = CallToolOptions {
568            timeout: self.tool_timeout,
569            meta: trace_context.as_ref().map(TraceContext::to_meta),
570            cancel,
571        };
572        let stream =
573            mcp.call_model_visible(tool_call.name, &tool_call.arguments, options).map(StreamEvent::ToolExecution);
574        self.streams.insert(StreamKey::Tool(tool_id), Box::pin(stream));
575    }
576
577    async fn handle_llm_usage(&mut self, sample: TokenUsage, state: &mut IterationState) {
578        state.call_usage = Some(sample);
579        self.token_tracker.record_usage(sample);
580        let ratio_pct = self.token_tracker.usage_ratio().map(|r| r * 100.0);
581        let remaining = self.token_tracker.tokens_remaining();
582        tracing::debug!(?sample, ?ratio_pct, ?remaining, "Token usage");
583
584        self.emit(self.context_usage_message()).await;
585        self.emit_session_usage(LlmCallPurpose::Chat, sample).await;
586    }
587
588    async fn emit_session_usage(&mut self, purpose: LlmCallPurpose, tokens: TokenUsage) {
589        let model = ModelIdentity::of(self.active_model.as_ref());
590        let event = self.session_usage.record(purpose, model, tokens);
591        self.emit(AgentEvent::SessionUsage(event)).await;
592    }
593
594    fn context_usage_message(&self) -> AgentEvent {
595        AgentEvent::Context(ContextEvent::UsageUpdated { usage: self.token_tracker.snapshot().clone() })
596    }
597
598    fn compaction_needed(&self) -> bool {
599        self.compaction_config.as_ref().is_some_and(|config| {
600            self.token_tracker.needs_compaction(self.context.estimated_token_count(), config.threshold)
601        })
602    }
603
604    async fn begin_compaction(&mut self) {
605        tracing::info!("Starting context compaction - {} messages", self.context.message_count());
606        self.emit(AgentEvent::Context(ContextEvent::CompactionStarted { message_count: self.context.message_count() }))
607            .await;
608        let started = self.begin_llm_call(LlmCallPurpose::Compaction, 0);
609        self.emit(started).await;
610
611        let compactor = Compactor::new(self.llm.clone());
612        let context = self.context.clone();
613        let stream: EventStream =
614            Box::pin(futures::stream::once(async move { StreamEvent::Compaction(compactor.compact(context).await) }));
615        self.streams.insert(StreamKey::Compaction, stream);
616    }
617
618    async fn on_compaction_complete(&mut self, result: Result<CompactionResult, CompactionError>) {
619        if let Ok(result) = &result
620            && let Some(usage) = result.usage
621        {
622            self.emit_session_usage(LlmCallPurpose::Compaction, usage).await;
623        }
624        let outcome = match &result {
625            Ok(result) => LlmCallOutcome::Completed { stop_reason: None, usage: result.usage },
626            Err(e) => LlmCallOutcome::failed(e.to_string(), false),
627        };
628        self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded { purpose: LlmCallPurpose::Compaction, outcome })).await;
629
630        match result {
631            Ok(result) => {
632                tracing::info!("Context compacted: {} messages removed", result.messages_removed);
633                self.context = self.context.with_compacted_summary(&result.summary);
634                self.token_tracker.reset_current_usage();
635                self.emit(AgentEvent::Context(ContextEvent::CompactionResult {
636                    summary: result.summary,
637                    messages_removed: result.messages_removed,
638                }))
639                .await;
640                self.emit(AgentEvent::Context(ContextEvent::CompactionEnded { outcome: CompactionOutcome::Completed }))
641                    .await;
642            }
643            Err(e) => {
644                tracing::warn!("Context compaction failed: {e}");
645                self.emit(AgentEvent::Context(ContextEvent::CompactionEnded {
646                    outcome: CompactionOutcome::Failed { error: e.to_string() },
647                }))
648                .await;
649            }
650        }
651
652        self.start_chat_turn().await;
653    }
654
655    async fn on_tool_execution_event(&mut self, tool_id: String, event: ToolCallEvent, state: &mut IterationState) {
656        match self.tool_executions.on_event(&tool_id, event) {
657            ToolExecutionUpdate::Event(event) => {
658                if let ToolEvent::SubAgentProgress { payload, .. } = &event
659                    && let AgentEvent::SessionUsage(child) = &payload.event
660                {
661                    let folded = self.session_usage.record_child(&payload.task_id, child.clone());
662                    self.emit(AgentEvent::SessionUsage(folded)).await;
663                }
664                self.emit(AgentEvent::Tool(event)).await;
665            }
666            ToolExecutionUpdate::Completed { result, event } => {
667                self.streams.remove(&StreamKey::Tool(tool_id));
668                state.completed_tool_calls.push(result);
669                self.emit(AgentEvent::Tool(event)).await;
670            }
671            ToolExecutionUpdate::TaskCreated { result, event } => {
672                state.completed_tool_calls.push(Ok(result));
673                self.emit(AgentEvent::Tool(event)).await;
674            }
675            ToolExecutionUpdate::TaskCompleted(outcome) => {
676                self.streams.remove(&StreamKey::Tool(tool_id));
677                self.enqueue_task_outcome(outcome, state).await;
678            }
679            ToolExecutionUpdate::TaskCancelled(outcome) => {
680                self.streams.remove(&StreamKey::Tool(tool_id));
681                self.record_task_outcome(outcome).await;
682            }
683            ToolExecutionUpdate::Retired => {
684                self.streams.remove(&StreamKey::Tool(tool_id));
685            }
686            ToolExecutionUpdate::Ignored => {
687                tracing::debug!(%tool_id, "Ignoring unexpected tool execution event");
688            }
689        }
690    }
691
692    async fn record_task_outcome(&mut self, outcome: TaskOutcome) {
693        self.context.add_message(outcome.context_message());
694        self.emit(AgentEvent::Tool(outcome.into())).await;
695    }
696
697    fn refresh_prompt_cache_key(&mut self) {
698        let key = derive_prompt_cache_key(self.llm.as_ref(), &self.context);
699        self.context.set_prompt_cache_key(Some(key));
700    }
701
702    async fn commit_pending_inputs(&mut self) {
703        let inputs = std::mem::take(&mut self.pending_inputs);
704        self.commit_inputs(inputs).await;
705    }
706
707    async fn commit_queued_inputs(&mut self) {
708        let inputs = std::mem::take(&mut self.queued_inputs);
709        self.commit_inputs(inputs).await;
710    }
711
712    async fn commit_inputs(&mut self, inputs: VecDeque<QueuedInput>) {
713        let mut user_content = Vec::new();
714        for input in inputs {
715            match input {
716                QueuedInput::User(content) => user_content.extend(content),
717                QueuedInput::TaskOutcome(outcome) => {
718                    self.commit_user_content(&mut user_content);
719                    self.record_task_outcome(*outcome).await;
720                }
721            }
722        }
723        self.commit_user_content(&mut user_content);
724    }
725
726    fn commit_user_content(&mut self, content: &mut Vec<llm::ContentBlock>) {
727        if !content.is_empty() {
728            self.context
729                .add_message(ChatMessage::User { content: std::mem::take(content), timestamp: IsoString::now() });
730        }
731    }
732
733    async fn emit_tool_definitions(&mut self) {
734        let tools = self.context.tools().clone();
735        if !tools.is_empty() {
736            self.emit(AgentEvent::Tool(ToolEvent::DefinitionsUpdated { tools })).await;
737        }
738    }
739
740    async fn emit(&mut self, message: AgentEvent) {
741        for observer in &mut self.observers {
742            observer.on_event(&message);
743        }
744
745        if let Err(e) = self.message_tx.send(message).await {
746            tracing::warn!("Failed to send agent message: {e:?}");
747        }
748    }
749
750    async fn finish_turn(&mut self, outcome: TurnOutcome) {
751        if std::mem::take(&mut self.turn_active) {
752            self.emit(AgentEvent::turn_ended(outcome)).await;
753        }
754    }
755
756    async fn begin_chat_call(&mut self, attempt: u32) {
757        self.llm_call_active = true;
758        let started = self.begin_llm_call(LlmCallPurpose::Chat, attempt);
759        if let Some(system_prompt) = self.context.system_content() {
760            for observer in &mut self.observers {
761                observer.on_system_prompt(system_prompt);
762            }
763        }
764        self.emit(started).await;
765    }
766
767    async fn finish_chat_call(&mut self, outcome: LlmCallOutcome) {
768        if std::mem::take(&mut self.llm_call_active) {
769            self.emit(AgentEvent::Turn(TurnEvent::LlmCallEnded { purpose: LlmCallPurpose::Chat, outcome })).await;
770        }
771    }
772
773    fn begin_llm_call(&mut self, purpose: LlmCallPurpose, attempt: u32) -> AgentEvent {
774        self.active_model = self.llm.model();
775        AgentEvent::Turn(TurnEvent::LlmCallStarted {
776            purpose,
777            model: ModelIdentity::of(self.active_model.as_ref()),
778            display_name: self.llm.display_name(),
779            attempt,
780            max_attempts: self.retry_config.max_attempts,
781        })
782    }
783}
784
785pub(crate) struct AutoContinue {
786    max: u32,
787    count: u32,
788}
789
790impl AutoContinue {
791    pub(crate) fn new(max: u32) -> Self {
792        Self { max, count: 0 }
793    }
794
795    fn reset(&mut self) {
796        self.count = 0;
797    }
798
799    fn should_continue(&self, stop_reason: Option<&StopReason>) -> bool {
800        matches!(stop_reason, Some(StopReason::Length)) && self.count < self.max
801    }
802
803    fn advance(&mut self) {
804        self.count += 1;
805    }
806}
807
808#[derive(Debug, Default)]
809struct IterationState {
810    current_message_id: Option<String>,
811    message_content: String,
812    reasoning_summary_text: String,
813    encrypted_reasoning: Option<EncryptedReasoningContent>,
814    completed_tool_calls: Vec<Result<ToolCallResult, ToolCallError>>,
815    llm_done: bool,
816    stop_reason: Option<StopReason>,
817    retry_attempt: u32,
818    call_usage: Option<TokenUsage>,
819}
820
821impl IterationState {
822    fn on_llm_start(&mut self, message_id: String) {
823        self.current_message_id = Some(message_id);
824        self.message_content.clear();
825        self.reasoning_summary_text.clear();
826        self.encrypted_reasoning = None;
827        self.stop_reason = None;
828        self.call_usage = None;
829    }
830
831    fn is_complete(&self, has_foreground_tools: bool) -> bool {
832        self.llm_done && !has_foreground_tools
833    }
834}
aether_core/core/agent.rs

aether_core/core/
agent.rs
