aether_core/core/
agent.rs

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

aether_core/core/
agent.rs
