a3s-code-core 8.0.3

use super::completion_runtime::CompletionFlow;
use super::execution_state::ExecutionLoopState;
use super::llm_turn::LlmTurnRequest;
use super::queue_forwarder::QueueEventForwarder;
use super::{AgentEvent, AgentLoop, AgentResult};
use crate::llm::{ContentBlock, Message};
use crate::prompts::AgentStyle;
use anyhow::Result;
use tokio::sync::mpsc;

const TOOL_BUDGET_FINALIZATION: &str = "Tool-use budget reached. Stop gathering evidence and return the best complete final answer now using only the tool results already present. Do not call any tool.";

impl AgentLoop {
    /// Core execution loop (without planning routing).
    ///
    /// This is the inner loop that runs LLM calls and tool executions.
    /// Called directly by `execute_with_session` (after planning check)
    /// and by `execute_plan` (for individual steps, bypassing planning).
    #[allow(clippy::too_many_arguments)]
    pub(super) async fn execute_loop(
        &self,
        history: &[Message],
        prompt: &str,
        effective_style: AgentStyle,
        session_id: Option<&str>,
        event_tx: Option<mpsc::Sender<AgentEvent>>,
        cancel_token: &tokio_util::sync::CancellationToken,
        emit_end: bool,
    ) -> Result<AgentResult> {
        // When called via execute_loop, the prompt is used for both
        // message-adding and hook/memory/event purposes.
        self.execute_loop_inner(
            history,
            prompt,
            prompt,
            Some(effective_style),
            session_id,
            event_tx,
            cancel_token,
            emit_end,
            None,
        )
        .await
    }

    /// Inner execution loop.
    ///
    /// `msg_prompt` controls whether a user message is appended (empty = skip).
    /// `effective_prompt` is used for hooks, memory recall, taint tracking, and events.
    /// `effective_style` pre-computed style to skip redundant LLM-based intent detection.
    /// `emit_end` controls whether to send `AgentEvent::End` when the loop completes
    /// (should be false when called from `execute_plan` to avoid duplicate End events).
    #[allow(clippy::too_many_arguments)]
    pub(super) async fn execute_loop_inner(
        &self,
        history: &[Message],
        msg_prompt: &str,
        effective_prompt: &str,
        effective_style: Option<AgentStyle>,
        session_id: Option<&str>,
        event_tx: Option<mpsc::Sender<AgentEvent>>,
        cancel_token: &tokio_util::sync::CancellationToken,
        emit_end: bool,
        seed: Option<super::execution_state::ExecutionSeed>,
    ) -> Result<AgentResult> {
        let mut state = ExecutionLoopState::new_seeded(history, seed);

        let style_prompt = if effective_prompt.is_empty() {
            msg_prompt
        } else {
            effective_prompt
        };
        let prompt_mode = self
            .resolve_prompt_mode(effective_style, style_prompt, &event_tx)
            .await;
        let effective_system_prompt = prompt_mode.system_prompt;

        // Send start event
        if let Some(tx) = &event_tx {
            tx.send(AgentEvent::Start {
                prompt: effective_prompt.to_string(),
            })
            .await
            .ok();
        }

        let _queue_forwarder = QueueEventForwarder::start(
            self.command_queue.as_ref(),
            event_tx.as_ref(),
            cancel_token,
        );

        let prompt_before_hooks = effective_prompt;
        let turn_context = self
            .prepare_turn_context(
                &effective_system_prompt,
                effective_prompt,
                state.messages.len(),
                session_id,
                &event_tx,
            )
            .await?;
        let effective_prompt = turn_context.effective_prompt.as_str();
        let augmented_system = turn_context.augmented_system;

        self.config.rl_trajectory_recorder.record_execution_start(
            crate::rl_trajectory::ExecutionStartRecord {
                session_id: session_id.unwrap_or(""),
                workspace: &self.tool_context.workspace,
                prompt: effective_prompt,
                history,
                system_prompt: augmented_system.as_deref(),
                max_tool_rounds: self.config.max_tool_rounds,
                planning_mode: &format!("{:?}", self.config.planning_mode),
            },
        );

        // Put the hook-effective prompt on the wire. Previously the rewritten
        // value was used only for context lookup and telemetry while the LLM
        // still received the original user message, making prompt rewrites and
        // additionalContext observational instead of authoritative.
        if !msg_prompt.is_empty() {
            state.messages.push(Message::user(effective_prompt));
        } else if effective_prompt != prompt_before_hooks {
            rewrite_latest_user_prompt(&mut state.messages, prompt_before_hooks, effective_prompt);
        }

        loop {
            // `max_tool_rounds` bounds evidence gathering, not the agent's
            // ability to return the evidence it already collected. Reserve one
            // provider turn with an empty tool set so bounded child runs
            // converge instead of discarding all work at the limit.
            let force_finalization = state.current_turn() >= self.config.max_tool_rounds;
            if force_finalization {
                state.messages.push(Message::user(TOOL_BUDGET_FINALIZATION));
            }
            let turn = state.next_turn();
            let capability_turn = self
                .capability_runtime
                .as_ref()
                .map(|runtime| runtime.begin_turn(turn))
                .transpose()?;
            let scoped_cancellation = capability_turn
                .as_ref()
                .map(crate::capability::AgentCapabilityTurn::cancellation)
                .unwrap_or_else(|| cancel_token.clone());
            let scoped_tool_context = capability_turn.as_ref().map_or_else(
                || self.tool_context.clone(),
                |scope| {
                    self.tool_context
                        .clone()
                        .with_capability_context(scope.tool_context())
                },
            );
            let llm_turn = match self
                .execute_llm_turn(
                    &mut state,
                    LlmTurnRequest {
                        turn,
                        augmented_system: &augmented_system,
                        effective_prompt,
                        session_id,
                        event_tx: &event_tx,
                        cancel_token: &scoped_cancellation,
                        force_no_tools: force_finalization,
                    },
                )
                .await
            {
                Ok(turn) => turn,
                // Interrupted mid-generation (Esc / cancel): keep the conversation
                // accumulated so far — above all the user's message — and return it
                // as the result so it is committed to history. Without this the
                // whole turn is dropped and the agent "forgets" what was just asked
                // when the user continues.
                Err(_) if scoped_cancellation.is_cancelled() => {
                    close_capability_turn(capability_turn.as_ref()).await?;
                    return Ok(state.finish_interrupted());
                }
                Err(error) => {
                    if let Err(close_error) = close_capability_turn(capability_turn.as_ref()).await
                    {
                        tracing::warn!(
                            error = %close_error,
                            "Capability Turn close also failed after provider failure"
                        );
                    }
                    return Err(error);
                }
            };
            debug_assert_eq!(llm_turn.turn, turn);
            let response = llm_turn.response;
            let tool_calls = llm_turn.tool_calls;

            if force_finalization && !tool_calls.is_empty() {
                let error = format!(
                    "Max tool rounds ({}) exceeded; the reserved finalization turn attempted another tool call",
                    self.config.max_tool_rounds
                );
                self.emit_error(&event_tx, error.clone()).await;
                close_capability_turn(capability_turn.as_ref()).await?;
                anyhow::bail!(error);
            }

            if tool_calls.is_empty() {
                match self
                    .complete_no_tool_response(
                        &mut state,
                        turn,
                        &response,
                        effective_prompt,
                        session_id,
                        &event_tx,
                        emit_end,
                        &scoped_cancellation,
                        &scoped_tool_context,
                        force_finalization,
                    )
                    .await
                {
                    CompletionFlow::Continue => {
                        close_capability_turn(capability_turn.as_ref()).await?;
                        continue;
                    }
                    CompletionFlow::Finished(final_text) => {
                        close_capability_turn(capability_turn.as_ref()).await?;
                        return Ok(state.finish(final_text));
                    }
                }
            }

            if let Err(e) = self
                .execute_tool_turn(
                    tool_calls,
                    &mut state,
                    &event_tx,
                    session_id,
                    &scoped_cancellation,
                    &scoped_tool_context,
                )
                .await
            {
                // Same as above: a cancelled tool round commits its partial
                // history rather than being dropped.
                if scoped_cancellation.is_cancelled() {
                    close_capability_turn(capability_turn.as_ref()).await?;
                    return Ok(state.finish_interrupted());
                }
                if let Err(close_error) = close_capability_turn(capability_turn.as_ref()).await {
                    tracing::warn!(
                        error = %close_error,
                        "Capability Turn close also failed after Tool failure"
                    );
                }
                return Err(e);
            }

            close_capability_turn(capability_turn.as_ref()).await?;
            // Quiescent boundary: all tools and capability-owned effects have
            // settled, and `state.messages` is consistent. The runtime sink
            // drains every preceding event before acknowledging persistence.
            self.persist_loop_checkpoint(turn, &state, session_id).await;
        }
    }

    /// Persist a `LoopCheckpoint` if both a sink and a bound run id are
    /// configured. Failures are swallowed (the sink already logs them)
    /// so an unavailable store cannot halt a live run.
    async fn persist_loop_checkpoint(
        &self,
        turn: usize,
        state: &super::execution_state::ExecutionLoopState,
        session_id: Option<&str>,
    ) {
        let Some(sink) = self.checkpoint_sink.as_ref() else {
            return;
        };
        let Some(run_id) = self.checkpoint_run_id.as_ref() else {
            return;
        };
        let checkpoint = crate::loop_checkpoint::LoopCheckpoint {
            schema_version: crate::loop_checkpoint::LOOP_CHECKPOINT_SCHEMA_VERSION,
            run_id: run_id.clone(),
            session_id: session_id.unwrap_or("").to_string(),
            capability_binding: self.checkpoint_capability_binding.clone(),
            turn,
            messages: state.messages.clone(),
            total_usage: state.total_usage.clone(),
            tool_calls_count: state.tool_calls_count,
            verification_reports: state.verification_reports.clone(),
            convergence: state.convergence_checkpoint(),
            checkpoint_ms: self.config.host_env.now_ms(),
        };
        sink.save_checkpoint(&checkpoint).await;
    }
}

async fn close_capability_turn(
    turn: Option<&crate::capability::AgentCapabilityTurn>,
) -> anyhow::Result<()> {
    let Some(turn) = turn else {
        return Ok(());
    };
    let report = turn.close().await?;
    if !report.is_clean() {
        anyhow::bail!(
            "Capability Turn close was incomplete (tasks failed: {}, tasks timed out: {}, child scopes failed: {}, child scopes timed out: {}, effects failed: {}, effects timed out: {})",
            report.tasks_failed,
            report.tasks_timed_out,
            report.child_scopes_failed,
            report.child_scopes_timed_out,
            report.effects_failed,
            report.effects_timed_out,
        );
    }
    Ok(())
}

fn rewrite_latest_user_prompt(messages: &mut [Message], original: &str, replacement: &str) {
    let candidate = messages
        .iter()
        .rposition(|message| message.role == "user" && message.text() == original)
        .or_else(|| {
            messages
                .iter()
                .rposition(|message| message.role == "user" && !message.text().is_empty())
        });
    let Some(index) = candidate else {
        tracing::warn!("PrePrompt modified input but no user message could be rewritten");
        return;
    };
    let message = &mut messages[index];

    let mut wrote_text = false;
    message.content.retain_mut(|block| match block {
        ContentBlock::Text { text } if !wrote_text => {
            *text = replacement.to_string();
            wrote_text = true;
            true
        }
        ContentBlock::Text { .. } => false,
        _ => true,
    });
    if !wrote_text {
        message.content.push(ContentBlock::Text {
            text: replacement.to_string(),
        });
    }
}