gitdataai/libs/agent/chat/service.rs

use futures::StreamExt;
use models::projects::project_skill;
use models::rooms::room_ai;
use rig::agent::{AgentBuilder, MultiTurnStreamItem};
use rig::client::CompletionClient;
use rig::completion::{CompletionModel, GetTokenUsage, Prompt};
use rig::streaming::{StreamedAssistantContent, StreamingPrompt};
use sea_orm::*;
use std::pin::Pin;
use std::sync::Arc;
use uuid::Uuid;

use super::context::RoomMessageContext;
use super::{AiChunkType, AiRequest, AiStreamChunk, Mention, StreamCallback};
use crate::billing;
use crate::client::AiClientConfig;

use crate::client::types::{ChatRequestMessage, ToolCall};
use crate::client::{
    StreamChunk, StreamChunkType, StreamedToolCall, call_stream, call_with_params,
};
use crate::compact::{CompactConfig, CompactService};
use crate::embed::EmbedService;
use crate::error::{AgentError, Result};
use crate::perception::{PerceptionService, SkillEntry, ToolCallEvent};
use crate::react::{DEFAULT_SYSTEM_PROMPT, ReactStep};
use crate::react::types::Action as ReactAction;
use crate::tool::{
    RecordingTool, ToolCall as AgentToolCall, ToolContext, ToolExecutor,
    registry::ToolRegistry,
};

/// Result from streaming AI response.
pub struct StreamResult {
    pub content: String,
    pub reasoning_content: String,
    pub input_tokens: i64,
    pub output_tokens: i64,
    /// All chunks in arrival order — preserves ReAct multi-cycle ordering.
    pub chunks: Vec<StreamChunk>,
}

/// Result from non-streaming AI response.
pub struct ProcessResult {
    pub content: String,
    pub input_tokens: i64,
    pub output_tokens: i64,
}

/// Record an AI session with cost calculation.
async fn record_ai_session(
    cache: &db::cache::AppCache,
    db: &db::database::AppDatabase,
    project_id: Uuid,
    session_id: Uuid,
    room_id: Uuid,
    model_id: Uuid,
    version_id: Uuid,
    input_tokens: i64,
    output_tokens: i64,
    latency_ms: i64,
) {
    metrics::histogram!("ai_call_latency_ms", "model" => model_id.to_string()).record(latency_ms as f64);

    let session = models::ai::ai_session::ActiveModel {
        id: Set(session_id),
        room: Set(room_id),
        model: Set(model_id),
        version: Set(version_id),
        token_input: Set(input_tokens),
        token_output: Set(output_tokens),
        latency_ms: Set(Some(latency_ms)),
        cost: Set(None),
        currency: Set(None),
        error_message: Set(None),
        error_code: Set(None),
        created_at: Set(chrono::Utc::now()),
    };

    if let Err(e) = session.insert(db).await {
        tracing::error!(error = %e, session_id = %session_id, "failed to insert ai session record");
        return;
    }

    let (cost, currency, error_msg) = match billing::record_ai_usage(
        db,
        project_id,
        version_id,
        input_tokens,
        output_tokens,
    )
    .await
    {
        Ok(billing::BillingResult::Success(record)) => {
            (Some(record.cost), Some(record.currency), None)
        }
        Ok(billing::BillingResult::InsufficientBalance { message }) => {
            create_system_message(cache, db, room_id, &message).await;
            (None, None, Some(message))
        }
        Err(e) => (None, None, Some(e.to_string())),
    };

    use sea_orm::sea_query::Expr;
    let _ = models::ai::ai_session::Entity::update_many()
        .col_expr(models::ai::ai_session::Column::Cost, Expr::value(cost))
        .col_expr(models::ai::ai_session::Column::Currency, Expr::value(currency))
        .col_expr(models::ai::ai_session::Column::ErrorMessage, Expr::value(error_msg))
        .filter(models::ai::ai_session::Column::Id.eq(session_id))
        .exec(db)
        .await;
}

/// Create a system message in the room for billing errors.
async fn create_system_message(
    cache: &db::cache::AppCache,
    db: &db::database::AppDatabase,
    room_id: Uuid,
    message: &str,
) {
    use models::rooms::{room_message, MessageSenderType, MessageContentType};
    use sea_orm::Set;

    let seq_key = format!("room:seq:{}", room_id);
    let seq = match cache.conn().await {
        Ok(mut conn) => {
            match redis::cmd("INCR").arg(&seq_key).query_async::<i64>(&mut conn).await {
                Ok(s) => s,
                Err(e) => {
                    tracing::warn!(error = %e, "cache INCR failed for system message seq, falling back to DB");
                    let last_seq = match room_message::Entity::find()
                        .filter(room_message::Column::Room.eq(room_id))
                        .order_by_desc(room_message::Column::Seq)
                        .one(db)
                        .await
                    {
                        Ok(Some(m)) => m.seq,
                        Ok(None) => 0,
                        Err(e) => {
                            tracing::warn!(error = %e, "Failed to get last seq for system message");
                            return;
                        }
                    };
                    last_seq + 1
                }
            }
        }
        Err(e) => {
            tracing::warn!(error = %e, "Failed to get Redis connection for system message seq");
            return;
        }
    };

    let now = chrono::Utc::now();

    let result = room_message::ActiveModel {
        id: Set(Uuid::now_v7()),
        seq: Set(seq),
        room: Set(room_id),
        sender_type: Set(MessageSenderType::System),
        sender_id: Set(None),
        model_id: Set(None),
        thread: Set(None),
        in_reply_to: Set(None),
        content: Set(message.to_string()),
        content_type: Set(MessageContentType::Text),
        thinking_content: Set(None),
        edited_at: Set(None),
        send_at: Set(now),
        revoked: Set(None),
        revoked_by: Set(None),
    }
    .insert(db)
    .await;

    match result {
        Ok(_) => {
            tracing::info!(
                room_id = %room_id,
                message = %message,
                "system_message_created_for_billing_error"
            );
        }
        Err(e) => {
            tracing::warn!(
                error = %e,
                room_id = %room_id,
                "Failed to create system message for billing error"
            );
        }
    }
}

/// Service for handling AI chat requests in rooms.
pub struct ChatService {
    ai_base_url: Option<String>,
    ai_api_key: Option<String>,
    compact_service: Option<CompactService>,
    embed_service: Option<EmbedService>,
    perception_service: PerceptionService,
    tool_registry: Option<ToolRegistry>,
}

impl ChatService {
    pub fn new() -> Self {
        Self {
            ai_base_url: None,
            ai_api_key: None,
            compact_service: None,
            embed_service: None,
            perception_service: PerceptionService::default(),
            tool_registry: None,
        }
    }

    pub fn with_ai_client_config(mut self, config: AiClientConfig) -> Self {
        self.ai_base_url = config.base_url.clone();
        self.ai_api_key = Some(config.api_key.clone());
        self
    }

    pub fn with_compact_service(mut self, compact_service: CompactService) -> Self {
        self.compact_service = Some(compact_service);
        self
    }

    pub fn with_embed_service(mut self, embed_service: EmbedService) -> Self {
        self.embed_service = Some(embed_service);
        self
    }

    pub fn with_perception_service(mut self, perception_service: PerceptionService) -> Self {
        self.perception_service = perception_service;
        self
    }

    pub fn with_tool_registry(mut self, registry: ToolRegistry) -> Self {
        self.tool_registry = Some(registry);
        self
    }

    /// Returns all registered tools as JSON tool definitions.
    pub fn tools(&self) -> Vec<serde_json::Value> {
        self.tool_registry
            .as_ref()
            .map(|r| r.to_openai_tools())
            .unwrap_or_default()
    }

    /// Build a RigToolSet from the registered tool registry.
    ///
    /// This enables using the same tools with `RigAgentService` via rig's native Agent.
    /// The context (db, cache, config, room_id, sender_id) is passed through to each
    /// tool handler at creation time.
    #[cfg(feature = "rig")]
    pub fn rig_toolset(
        &self,
        db: db::database::AppDatabase,
        cache: db::cache::AppCache,
        config: config::AppConfig,
        room_id: uuid::Uuid,
        sender_id: Option<uuid::Uuid>,
        project_id: uuid::Uuid,
    ) -> Option<crate::RigToolSet> {
        self.tool_registry.as_ref().map(|registry| {
            crate::RigToolSet::from_registry(
                registry, db, cache, config, room_id, sender_id, project_id,
            )
        })
    }

    /// Get a reference to the underlying ToolRegistry.
    pub fn tool_registry(&self) -> Option<&ToolRegistry> {
        self.tool_registry.as_ref()
    }

    pub async fn process(&self, request: AiRequest) -> Result<ProcessResult> {
        let tools: Vec<serde_json::Value> = request.tools.clone().unwrap_or_default();
        let tools_enabled = !tools.is_empty();
        let max_tool_depth = request.max_tool_depth;

        let mut messages = self.build_messages(&request).await?;

        let room_ai = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await?;

        let model_name = request.model.name.clone();
        let temperature = room_ai
            .as_ref()
            .and_then(|r| r.temperature.map(|v| v as f32))
            .unwrap_or(request.temperature as f32);
        let max_tokens = room_ai
            .as_ref()
            .and_then(|r| r.max_tokens.map(|v| v as u32))
            .unwrap_or(request.max_tokens as u32);
        let mut tool_depth = 0;
        let mut input_tokens = 0i64;
        let mut output_tokens = 0i64;
        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();
        let version_id = room_ai.as_ref().and_then(|r| r.version);

        let config = AiClientConfig::new(self.ai_api_key.clone().unwrap_or_default())
            .with_base_url(
                self.ai_base_url
                    .clone()
                    .unwrap_or_else(|| "https://api.openai.com".into()),
            );

        loop {
            let response = call_with_params(
                &messages,
                &model_name,
                &config,
                temperature,
                max_tokens,
                None,
                if tools_enabled { Some(&tools) } else { None },
                if tools_enabled { None } else { Some("none") },
            )
            .await?;

            let text = response.content.clone();
            input_tokens += response.input_tokens;
            output_tokens += response.output_tokens;

            if tools_enabled && !response.tool_calls_finished.is_empty() {
                // Build assistant message with tool_calls
                let tool_call_messages: Vec<_> = response
                    .tool_calls_finished
                    .iter()
                    .map(|name| {
                        // We need ID and arguments — for non-streaming we reconstruct from content
                        // The model returns tool_calls in its content; for now we create a placeholder
                        // that will be replaced by actual tool results
                        ToolCall {
                            id: Uuid::new_v4().to_string(),
                            type_: "function".into(),
                            function: crate::client::types::ToolCallFunction {
                                name: name.clone(),
                                arguments: "{}".into(),
                            },
                        }
                    })
                    .collect();

                messages.push(ChatRequestMessage::assistant(
                    Some(text.clone()),
                    Some(tool_call_messages.clone()),
                ));

                // Create ToolCall list for executor (we need real IDs and args)
                // Since we can't get args from streaming, use name matching from the text
                let calls: Vec<AgentToolCall> = tool_call_messages
                    .into_iter()
                    .map(|tc| AgentToolCall {
                        id: tc.id.clone(),
                        name: tc.function.name.clone(),
                        arguments: tc.function.arguments.clone(),
                    })
                    .collect();

                let tool_messages = {
                    let mut ctx = ToolContext::new(
                        request.db.clone(),
                        request.cache.clone(),
                        request.config.clone(),
                        request.room.id,
                        Some(request.sender.uid),
                    )
                    .with_project(request.project.id);
                    if let Some(ref es) = self.embed_service {
                        ctx = ctx.with_embed_service(es.clone());
                    }
                    if let Some(ref registry) = self.tool_registry {
                        ctx.registry_mut().merge(registry.clone());
                    }

                    let recorder = crate::tool::recorder::ToolCallRecorder::with_session(
                        request.db.clone(),
                        session_id,
                    );
                    let start = std::time::Instant::now();

                    let executor = ToolExecutor::new();
                    match executor.execute_batch(calls, &mut ctx).await {
                        Ok(results) => {
                            for (call, result) in
                                response.tool_calls_finished.iter().zip(results.iter())
                            {
                                let elapsed = start.elapsed().as_millis() as i64;
                                let is_error =
                                    matches!(result.result, crate::tool::ToolResult::Error(_));
                                let error_msg = match &result.result {
                                    crate::tool::ToolResult::Error(msg) => Some(msg.clone()),
                                    _ => None,
                                };
                                recorder.record(crate::tool::recorder::ToolCallRecord {
                                    tool_call_id: call.clone(),
                                    session_id: recorder.session_id(),
                                    tool_name: call.clone(),
                                    caller: request.sender.uid,
                                    arguments: serde_json::Value::Null,
                                    status: if is_error {
                                        models::ai::ToolCallStatus::Failed
                                    } else {
                                        models::ai::ToolCallStatus::Success
                                    },
                                    execution_time_ms: Some(elapsed),
                                    error_message: error_msg,
                                    error_stack: None,
                                    retry_count: 0,
                                });
                            }
                            crate::tool::ToolExecutor::to_tool_messages(&results)
                        }
                        Err(e) => {
                            let elapsed = start.elapsed().as_millis() as i64;
                            for call_name in &response.tool_calls_finished {
                                recorder.record(crate::tool::recorder::ToolCallRecord {
                                    tool_call_id: Uuid::new_v4().to_string(),
                                    session_id: recorder.session_id(),
                                    tool_name: call_name.clone(),
                                    caller: request.sender.uid,
                                    arguments: serde_json::Value::Null,
                                    status: models::ai::ToolCallStatus::Failed,
                                    execution_time_ms: Some(elapsed),
                                    error_message: Some(e.to_string()),
                                    error_stack: None,
                                    retry_count: 0,
                                });
                            }

                            let err_msg = format!("[Tool call failed: {}]", e);
                            response
                                .tool_calls_finished
                                .iter()
                                .map(|_| {
                                    ChatRequestMessage::tool(Uuid::new_v4().to_string(), &err_msg)
                                })
                                .collect()
                        }
                    }
                };
                messages.extend(tool_messages);

                // Inject passive-detected skills based on tool calls
                if let Ok(skills) = project_skill::Entity::find()
                    .filter(project_skill::Column::ProjectUuid.eq(request.project.id))
                    .filter(project_skill::Column::Enabled.eq(true))
                    .all(&request.db)
                    .await
                {
                    let skill_entries: Vec<SkillEntry> = skills
                        .into_iter()
                        .map(|s| SkillEntry {
                            slug: s.slug,
                            name: s.name,
                            description: s.description,
                            content: s.content,
                        })
                        .collect();
                    let tool_events: Vec<ToolCallEvent> = response
                        .tool_calls_finished
                        .iter()
                        .map(|name| ToolCallEvent {
                            tool_name: name.clone(),
                            arguments: String::new(),
                        })
                        .collect();
                    for event in &tool_events {
                        if let Some(ctx) = self
                            .perception_service
                            .passive
                            .detect(event, &skill_entries)
                        {
                            messages.push(ctx.to_system_message());
                        }
                    }
                }

                tool_depth += 1;
                if tool_depth >= max_tool_depth {
                    let content = if text.is_empty() {
                        format!(
                            "[AI reached maximum tool depth ({}) — no final answer produced]",
                            max_tool_depth
                        )
                    } else {
                        text
                    };
                    // Record session
                    record_ai_session(
                        &request.cache,
                        &request.db,
                        request.project.id,
                        session_id,
                        request.room.id,
                        request.model.id,
                        version_id.unwrap_or_default(),
                        input_tokens,
                        output_tokens,
                        session_start.elapsed().as_millis() as i64,
                    )
                    .await;
                    return Ok(ProcessResult {
                        content,
                        input_tokens,
                        output_tokens,
                    });
                }
                continue;
            }

            // Record session
            record_ai_session(
                &request.cache,
                &request.db,
                request.project.id,
                session_id,
                request.room.id,
                request.model.id,
                version_id.unwrap_or_default(),
                input_tokens,
                output_tokens,
                session_start.elapsed().as_millis() as i64,
            )
            .await;
            return Ok(ProcessResult {
                content: text,
                input_tokens,
                output_tokens,
            });
        }
    }

    pub async fn process_stream(
        &self,
        request: AiRequest,
        on_chunk: StreamCallback,
    ) -> Result<StreamResult> {
        // Wrap on_chunk in Arc so it can be shared across loop iterations
        let on_chunk = Arc::new(on_chunk);
        let tools: Vec<serde_json::Value> = request.tools.clone().unwrap_or_default();
        let tools_enabled = !tools.is_empty();
        let max_tool_depth = request.max_tool_depth;

        let mut messages = self.build_messages(&request).await?;

        let room_ai = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await?;

        let model_name = request.model.name.clone();
        let temperature = room_ai
            .as_ref()
            .and_then(|r| r.temperature.map(|v| v as f32))
            .unwrap_or(request.temperature as f32);
        let max_tokens = room_ai
            .as_ref()
            .and_then(|r| r.max_tokens.map(|v| v as u32))
            .unwrap_or(request.max_tokens as u32);
        let mut tool_depth = 0;
        let mut total_input_tokens = 0i64;
        let mut total_output_tokens = 0i64;
        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();

        let version_id = room_ai.as_ref().and_then(|r| r.version);

        let config = AiClientConfig::new(self.ai_api_key.clone().unwrap_or_default())
            .with_base_url(
                self.ai_base_url
                    .clone()
                    .unwrap_or_else(|| "https://api.openai.com".into()),
            );

        let mut full_content = String::new();
        let mut all_chunks: Vec<StreamChunk> = Vec::new();
        // Collect tool calls during streaming, push them incrementally after.
        let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<StreamedToolCall>();

        loop {
            let on_chunk_cb = on_chunk.clone();
            let on_chunk_cb2 = on_chunk_cb.clone();
            let tx_arc = Arc::new(tx.clone());
            let tx_arc2 = tx_arc.clone();
            let response = call_stream(
                &messages,
                &model_name,
                &config,
                temperature,
                max_tokens,
                if tools_enabled { Some(&tools) } else { None },
                None, // tool_choice — auto (let model decide)
                Arc::new(move |delta| {
                    let fut = on_chunk_cb(AiStreamChunk {
                        content: delta.to_string(),
                        done: false,
                        chunk_type: AiChunkType::Answer,
                    });
                    fut
                }),
                Arc::new(move |delta| {
                    let fut = on_chunk_cb2(AiStreamChunk {
                        content: delta.to_string(),
                        done: false,
                        chunk_type: AiChunkType::Thinking,
                    });
                    fut
                }),
                Arc::new(move |tc: &StreamedToolCall| {
                    let tx = tx_arc2.clone();
                    let tc_owned = tc.clone();
                    Box::pin(async move {
                        let _ = tx.send(tc_owned);
                    }) as Pin<Box<dyn std::future::Future<Output = ()> + Send>>
                }),
            )
            .await?;

            total_input_tokens += response.input_tokens;
            total_output_tokens += response.output_tokens;

            // Collect chunks from this streaming iteration in order.
            all_chunks.extend(response.chunks);

            let has_tool_calls = tools_enabled && !response.tool_calls.is_empty();

            if has_tool_calls {
                // Accumulate the assistant's text before tool calls
                full_content.push_str(&response.content);
                full_content.push('\n');

                // Build assistant message with tool_calls from streaming response
                let tool_calls: Vec<ToolCall> = response
                    .tool_calls
                    .iter()
                    .map(|tc| ToolCall {
                        id: tc.id.clone(),
                        type_: "function".into(),
                        function: crate::client::types::ToolCallFunction {
                            name: tc.name.clone(),
                            arguments: tc.arguments.clone(),
                        },
                    })
                    .collect();

                messages.push(ChatRequestMessage::assistant(
                    Some(response.content.clone()),
                    Some(tool_calls.clone()),
                ));

                // Push each tool call incrementally to frontend.
                // Use try_recv() — tx is never dropped so recv() would deadlock.
                loop {
                    match rx.try_recv() {
                        Ok(tc) => {
                            let args_display = if tc.arguments.len() > 100 {
                                format!("{}...", &tc.arguments[..100])
                            } else {
                                tc.arguments.clone()
                            };
                            let tool_display = format!("🔧 {}({})", tc.name, args_display);
                            on_chunk(AiStreamChunk {
                                content: tool_display.clone(),
                                done: false,
                                chunk_type: AiChunkType::ToolCall,
                            })
                            .await;
                            all_chunks.push(StreamChunk {
                                chunk_type: StreamChunkType::ToolCall,
                                content: tool_display,
                            });
                        }
                        Err(tokio::sync::mpsc::error::TryRecvError::Empty) => break,
                        Err(tokio::sync::mpsc::error::TryRecvError::Disconnected) => break,
                    }
                }

                // Execute tools one at a time, push each result incrementally
                let calls: Vec<AgentToolCall> = response
                    .tool_calls
                    .iter()
                    .map(|tc| AgentToolCall {
                        id: tc.id.clone(),
                        name: tc.name.clone(),
                        arguments: tc.arguments.clone(),
                    })
                    .collect();

                let mut tool_messages = Vec::new();
                let mut ctx = crate::tool::ToolContext::new(
                    request.db.clone(),
                    request.cache.clone(),
                    request.config.clone(),
                    request.room.id,
                    Some(request.sender.uid),
                )
                .with_project(request.project.id);
                if let Some(ref es) = self.embed_service {
                    ctx = ctx.with_embed_service(es.clone());
                }
                if let Some(ref registry) = self.tool_registry {
                    ctx.registry_mut().merge(registry.clone());
                }

                let recorder = crate::tool::recorder::ToolCallRecorder::with_session(
                    request.db.clone(),
                    session_id,
                );

                for call in &calls {
                    let start = std::time::Instant::now();

                    // Spawn tool execution in a separate task to avoid blocking the
                    // tokio worker thread (git2 operations are synchronous).
                    // This allows the heartbeat timer to fire independently.
                    let call_clone = call.clone();
                    let mut ctx_clone = ctx.clone();
                    let (result_tx, mut result_rx) = tokio::sync::oneshot::channel();
                    tokio::spawn(async move {
                        let executor = crate::tool::ToolExecutor::new();
                        let res = executor.execute_batch(vec![call_clone], &mut ctx_clone).await;
                        let _ = result_tx.send(res);
                    });

                    // Send heartbeats every 10s until tool execution completes
                    let heartbeat_dur = std::time::Duration::from_secs(10);
                    let results = loop {
                        tokio::select! {
                            res = &mut result_rx => {
                                match res {
                                    Ok(inner) => break inner,
                                    Err(_) => break Err(crate::tool::ToolError::ExecutionError("tool task cancelled".into())),
                                }
                            },
                            _ = tokio::time::sleep(heartbeat_dur) => {
                                on_chunk(AiStreamChunk {
                                    content: String::new(),
                                    done: false,
                                    chunk_type: AiChunkType::ToolCall,
                                }).await;
                            }
                        }
                    };

                    let results = match results {
                        Ok(r) => r,
                        Err(e) => {
                            let elapsed = start.elapsed().as_millis() as i64;
                            recorder.record(crate::tool::recorder::ToolCallRecord {
                                tool_call_id: call.id.clone(),
                                session_id: recorder.session_id(),
                                tool_name: call.name.clone(),
                                caller: request.sender.uid,
                                arguments: call.arguments_json().unwrap_or_default(),
                                status: models::ai::ToolCallStatus::Failed,
                                execution_time_ms: Some(elapsed),
                                error_message: Some(e.to_string()),
                                error_stack: None,
                                retry_count: 0,
                            });

                            let err_text = format!("[Tool call failed: {}]", e);
                            tracing::warn!(tool = %call.name, args = %call.arguments, error = %e, "tool_call_failed");
                            let err_display = format!("❌ {} (failed)", call.name);
                            on_chunk(AiStreamChunk {
                                content: err_display.clone(),
                                done: false,
                                chunk_type: AiChunkType::ToolResult,
                            })
                            .await;
                            all_chunks.push(StreamChunk {
                                chunk_type: StreamChunkType::ToolCall,
                                content: err_display,
                            });
                            tool_messages.push(ChatRequestMessage::tool(&call.id, &err_text));
                            continue;
                        }
                    };

                    for result in &results {
                        let text = match &result.result {
                            crate::tool::ToolResult::Ok(v) => v.to_string(),
                            crate::tool::ToolResult::Error(msg) => msg.clone(),
                        };
                        let preview = if text.len() > 300 {
                            format!("{}...", &text[..300])
                        } else {
                            text.clone()
                        };
                        tracing::debug!("tool_result: {} — {}", call.name, preview);

                        let elapsed = start.elapsed().as_millis() as i64;
                        let is_error = matches!(result.result, crate::tool::ToolResult::Error(_));
                        let error_msg = match &result.result {
                            crate::tool::ToolResult::Error(msg) => Some(msg.clone()),
                            _ => None,
                        };
                        recorder.record(crate::tool::recorder::ToolCallRecord {
                            tool_call_id: call.id.clone(),
                            session_id: recorder.session_id(),
                            tool_name: call.name.clone(),
                            caller: request.sender.uid,
                            arguments: call.arguments_json().unwrap_or_default(),
                            status: if is_error {
                                models::ai::ToolCallStatus::Failed
                            } else {
                                models::ai::ToolCallStatus::Success
                            },
                            execution_time_ms: Some(elapsed),
                            error_message: error_msg,
                            error_stack: None,
                            retry_count: 0,
                        });
                        // Do NOT emit tool_result chunks to frontend — raw output may contain sensitive data.
                        // Log server-side only; frontend sees tool_call status via on_chunk below.
                    }
                    let success_display = format!("✅ {}", call.name);
                    on_chunk(AiStreamChunk {
                        content: success_display.clone(),
                        done: false,
                        chunk_type: AiChunkType::ToolResult,
                    })
                    .await;
                    all_chunks.push(StreamChunk {
                        chunk_type: StreamChunkType::ToolCall,
                        content: success_display,
                    });

                    let msgs = crate::tool::ToolExecutor::to_tool_messages(&results);
                    tool_messages.extend(msgs);
                }
                messages.extend(tool_messages);

                // Inject passive-detected skills based on tool calls
                if let Ok(skills) = project_skill::Entity::find()
                    .filter(project_skill::Column::ProjectUuid.eq(request.project.id))
                    .filter(project_skill::Column::Enabled.eq(true))
                    .all(&request.db)
                    .await
                {
                    let skill_entries: Vec<SkillEntry> = skills
                        .into_iter()
                        .map(|s| SkillEntry {
                            slug: s.slug,
                            name: s.name,
                            description: s.description,
                            content: s.content,
                        })
                        .collect();
                    let tool_events: Vec<ToolCallEvent> = response
                        .tool_calls
                        .iter()
                        .map(|tc| ToolCallEvent {
                            tool_name: tc.name.clone(),
                            arguments: tc.arguments.clone(),
                        })
                        .collect();
                    for event in &tool_events {
                        if let Some(ctx) = self
                            .perception_service
                            .passive
                            .detect(event, &skill_entries)
                        {
                            messages.push(ctx.to_system_message());
                        }
                    }
                }

                tool_depth += 1;
                if tool_depth >= max_tool_depth {
                    let max_depth_text = format!(
                        "[AI reached maximum tool depth ({}) — no final answer produced]",
                        max_tool_depth
                    );
                    on_chunk(AiStreamChunk {
                        content: max_depth_text.clone(),
                        done: true,
                        chunk_type: AiChunkType::Answer,
                    })
                    .await;
                    all_chunks.push(StreamChunk {
                        chunk_type: StreamChunkType::Answer,
                        content: max_depth_text,
                    });
                    // Record session
                    record_ai_session(
                        &request.cache,
                        &request.db,
                        request.project.id,
                        session_id,
                        request.room.id,
                        request.model.id,
                        version_id.unwrap_or_default(),
                        total_input_tokens,
                        total_output_tokens,
                        session_start.elapsed().as_millis() as i64,
                    )
                    .await;
                    return Ok(StreamResult {
                        content: full_content,
                        reasoning_content: String::new(),
                        input_tokens: 0,
                        output_tokens: 0,
                        chunks: all_chunks,
                    });
                }
                continue;
            }

            // Final answer — accumulate and return
            full_content.push_str(&response.content);

            on_chunk(AiStreamChunk {
                content: response.content.clone(),
                done: true,
                chunk_type: AiChunkType::Answer,
            })
            .await;
            all_chunks.push(StreamChunk {
                chunk_type: StreamChunkType::Answer,
                content: response.content.clone(),
            });
            // Record session
            record_ai_session(
                &request.cache,
                &request.db,
                request.project.id,
                session_id,
                request.room.id,
                request.model.id,
                version_id.unwrap_or_default(),
                total_input_tokens,
                total_output_tokens,
                session_start.elapsed().as_millis() as i64,
            )
            .await;
            return Ok(StreamResult {
                content: full_content,
                reasoning_content: response.reasoning_content,
                input_tokens: response.input_tokens,
                output_tokens: response.output_tokens,
                chunks: all_chunks,
            });
        }
    }

    async fn build_messages(&self, request: &AiRequest) -> Result<Vec<ChatRequestMessage>> {
        let mut messages = Vec::new();

        // Core reasoning instruction — prioritize analysis before answering.
        messages.push(ChatRequestMessage::system(
            "When receiving a question or problem, follow this reasoning process:\n\
             1. ANALYZE: Break down the question. Identify what is being asked, what context is available, and what information is missing.\n\
             2. GATHER: Use available tools (repository search, file reading, etc.) to collect relevant information before answering.\n\
             3. REASON: Synthesize the gathered information. Consider edge cases and potential issues.\n\
             4. ANSWER: Provide a clear, actionable answer based on your analysis.\n\
             \n\
             Do NOT guess or assume when tools can provide concrete answers. Always verify claims against actual code or documentation.".to_string()
        ));

        let mut processed_history = Vec::new();
        if let Some(compact_service) = &self.compact_service {
            let compact_cache_key = format!("ai:compact:{}", request.room.id);
            let compact_config = CompactConfig::default();

            // Try cached compaction summary (avoids re-compacting same history)
            let cached_summary: Option<String> = {
                let conn_result = request.cache.conn().await;
                match conn_result {
                    Ok(mut conn) => {
                        redis::cmd("GET")
                            .arg(&compact_cache_key)
                            .query_async::<Option<String>>(&mut conn)
                            .await
                            .unwrap_or(None)
                    }
                    Err(e) => {
                        tracing::warn!(error = %e, "compact cache: conn failed");
                        None
                    }
                }
            };

            if let Some(cached_json) = cached_summary {
                if let Ok(summary) = serde_json::from_str::<crate::compact::CompactSummary>(&cached_json) {
                    if !summary.summary.is_empty() {
                        messages.push(ChatRequestMessage::system(format!(
                            "Conversation summary:\n{}",
                            summary.summary
                        )));
                    }
                    processed_history = summary.retained;
                }
            }

            if processed_history.is_empty() {
                match compact_service
                    .compact_room_auto(request.room.id, Some(request.user_names.clone()), compact_config)
                    .await
                {
                    Ok(compact_summary) => {
                        if !compact_summary.summary.is_empty() {
                            messages.push(ChatRequestMessage::system(format!(
                                "Conversation summary:\n{}",
                                compact_summary.summary
                            )));
                        }
                        // Cache for subsequent calls (5 min TTL)
                        if let Ok(json) = serde_json::to_string(&compact_summary) {
                            if let Ok(mut conn) = request.cache.conn().await {
                                let _ = redis::cmd("SETEX")
                                    .arg(&compact_cache_key)
                                    .arg(300u64)
                                    .arg(&json)
                                    .query_async::<()>(&mut conn)
                                    .await
                                    .inspect_err(|e| {
                                        tracing::warn!(error = %e, "compact cache: SETEX failed");
                                    });
                            }
                        }
                        processed_history = compact_summary.retained;
                    }
                    Err(e) => {
                        tracing::warn!(error = %e, "conversation compaction failed, using full history");
                    }
                }
            }
        }

        if !processed_history.is_empty() {
            for msg_summary in processed_history {
                let ctx = RoomMessageContext::from(msg_summary);
                messages.push(ctx.to_message());
            }
        } else {
            for msg in &request.history {
                let ctx = RoomMessageContext::from_model_with_names(msg, &request.user_names);
                messages.push(ctx.to_message());
            }
        }

        for mention in &request.mention {
            match mention {
                Mention::Repo(repo) => {
                    // Inject repo details into system prompt so AI knows the repo context
                    let mut parts = vec![
                        format!("Name: {}", repo.repo_name),
                        format!("ID: {}", repo.id),
                    ];
                    if let Some(ref desc) = repo.description {
                        parts.push(format!("Description: {}", desc));
                    }
                    parts.push(format!("Default branch: {}", repo.default_branch));
                    parts.push(format!(
                        "Private: {}",
                        if repo.is_private { "yes" } else { "no" }
                    ));
                    parts.push(format!("Created: {}", repo.created_at.format("%Y-%m-%d")));
                    messages.push(ChatRequestMessage::system(format!(
                        "Mentioned repository:\n{}",
                        parts.join("\n")
                    )));

                    // Vector search for related issues and repos (enhancement, optional)
                    if let Some(embed_service) = &self.embed_service {
                        let query = format!(
                            "{} {}",
                            repo.repo_name,
                            repo.description.as_deref().unwrap_or_default()
                        );
                        if let Ok(issues) = embed_service.search_issues(&query, 5).await {
                            if !issues.is_empty() {
                                let context = format!(
                                    "Related issues for repo {}:\n{}",
                                    repo.repo_name,
                                    issues
                                        .iter()
                                        .map(|i| format!("- {}", i.payload.text))
                                        .collect::<Vec<_>>()
                                        .join("\n")
                                );
                                messages.push(ChatRequestMessage::system(context));
                            }
                        }
                        if let Ok(repos) = embed_service.search_repos(&query, 3).await {
                            if !repos.is_empty() {
                                let context = format!(
                                    "Similar repositories:\n{}",
                                    repos
                                        .iter()
                                        .map(|r| format!("- {}", r.payload.text))
                                        .collect::<Vec<_>>()
                                        .join("\n")
                                );
                                messages.push(ChatRequestMessage::system(context));
                            }
                        }
                    }
                }
                Mention::User(user) => {
                    let mut profile_parts = vec![format!("Username: {}", user.username)];
                    if let Some(ref display_name) = user.display_name {
                        profile_parts.push(format!("Display name: {}", display_name));
                    }
                    if let Some(ref org) = user.organization {
                        profile_parts.push(format!("Organization: {}", org));
                    }
                    if let Some(ref website) = user.website_url {
                        profile_parts.push(format!("Website: {}", website));
                    }
                    messages.push(ChatRequestMessage::system(format!(
                        "Mentioned user profile:\n{}",
                        profile_parts.join("\n")
                    )));
                }
            }
        }

        let skill_contexts = self.build_skill_context(request).await;
        for ctx in skill_contexts {
            messages.push(ctx.to_system_message());
        }

        let memories = self.build_memory_context(request).await;
        for mem in memories {
            messages.push(mem.to_system_message());
        }

        messages.push(ChatRequestMessage::system(format!(
            "Current Project:\n{}\nDescription: {}\nPublic: {}",
            request.project.display_name,
            request.project.description.as_deref().unwrap_or("(none)"),
            if request.project.is_public {
                "yes"
            } else {
                "no"
            }
        )));

        let mut sender_parts = vec![format!("**Sender:** {}", request.sender.username)];
        if let Some(ref display_name) = request.sender.display_name {
            sender_parts.push(display_name.clone());
        }
        if let Some(ref org) = request.sender.organization {
            sender_parts.push(format!("({})", org));
        }
        let sender_display = sender_parts.join(" ");
        messages.push(ChatRequestMessage::system(format!(
            "The person sending the next message:\n{}",
            sender_display
        )));

        messages.push(ChatRequestMessage::user(&request.input));

        Ok(messages)
    }

    async fn build_skill_context(
        &self,
        request: &AiRequest,
    ) -> Vec<crate::perception::SkillContext> {
        // Load database skills for this project
        let db_skills: Vec<SkillEntry> = match project_skill::Entity::find()
            .filter(project_skill::Column::ProjectUuid.eq(request.project.id))
            .filter(project_skill::Column::Enabled.eq(true))
            .all(&request.db)
            .await
        {
            Ok(models) => models
                .into_iter()
                .map(|s| SkillEntry {
                    slug: s.slug,
                    name: s.name,
                    description: s.description,
                    content: s.content,
                })
                .collect(),
            Err(_) => Vec::new(),
        };

        // Load built-in skills and merge with database skills
        // Built-in skills override database skills with the same slug
        let mut all_skills: Vec<SkillEntry> = db_skills;
        for built_in in crate::skills::all_skills() {
            // Skip if a database skill with the same slug already exists
            if !all_skills.iter().any(|s| s.slug == built_in.slug) {
                all_skills.push(SkillEntry {
                    slug: built_in.slug.to_string(),
                    name: built_in.name.to_string(),
                    description: Some(built_in.description.to_string()),
                    content: built_in.content.clone(),
                });
            }
        }

        if all_skills.is_empty() {
            return Vec::new();
        }

        let history_texts: Vec<String> = request
            .history
            .iter()
            .rev()
            .take(10)
            .map(|msg| msg.content.clone())
            .collect();

        let tool_events: Vec<ToolCallEvent> = Vec::new();
        let keyword_skills = self
            .perception_service
            .inject_skills(&request.input, &history_texts, &tool_events, &all_skills)
            .await;

        let mut vector_skills = Vec::new();
        if let Some(embed_service) = &self.embed_service {
            let awareness = crate::perception::VectorActiveAwareness::default();
            vector_skills = awareness
                .detect(
                    embed_service,
                    &request.input,
                    &request.project.id.to_string(),
                )
                .await;
        }

        let mut seen = std::collections::HashSet::new();
        let mut result = Vec::new();
        for ctx in vector_skills {
            if seen.insert(ctx.label.clone()) {
                result.push(ctx);
            }
        }
        for ctx in keyword_skills {
            if seen.insert(ctx.label.clone()) {
                result.push(ctx);
            }
        }

        result
    }

    async fn build_memory_context(
        &self,
        request: &AiRequest,
    ) -> Vec<crate::perception::vector::MemoryContext> {
        let embed_service = match &self.embed_service {
            Some(s) => s,
            None => return Vec::new(),
        };

        let awareness = crate::perception::VectorPassiveAwareness::default();
        awareness
            .detect(
                embed_service,
                &request.input,
                &request.project.display_name,
                &request.room.id.to_string(),
            )
            .await
    }

    pub async fn process_react<C, Fut>(&self, request: &AiRequest, mut on_chunk: C) -> Result<(String, i64, i64)>
    where
        C: FnMut(crate::react::ReactStep) -> Fut + Send,
        Fut: std::future::Future<Output = ()> + Send,
    {
        let base_url = self
            .ai_base_url
            .clone()
            .unwrap_or_else(|| "https://api.openai.com".into());
        let api_key = self.ai_api_key.clone().unwrap_or_default();
        let client_config = AiClientConfig::new(api_key).with_base_url(base_url);

        let Some(registry) = &self.tool_registry else {
            return Err(AgentError::Internal("no tool registry registered".into()));
        };

        let db = request.db.clone();
        let cache = request.cache.clone();
        let cfg = request.config.clone();
        let room_id = request.room.id;
        let sender_uid = request.sender.uid;
        let project_id = request.project.id;
        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();
        let version_id = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await
            .ok()
            .flatten()
            .and_then(|r| r.version);

        // Build rig tools with recording wrapper directly from registry
        let mut tools: Vec<Box<dyn rig::tool::ToolDyn + 'static>> = Vec::new();
        for def in registry.definitions() {
            let name = def.name.clone();
            if let Some(handler) = registry.get(&name) {
                let adapter = crate::tool::RigToolAdapter::new(
                    handler.clone(),
                    def.clone(),
                    db.clone(),
                    cache.clone(),
                    cfg.clone(),
                    room_id,
                    Some(sender_uid),
                    project_id,
                );
                tools.push(Box::new(RecordingTool::new(
                    Box::new(adapter),
                    db.clone(),
                    session_id,
                    sender_uid,
                )));
            }
        }

        // Build rig agent (handles multi-turn tool calls natively)
        let rig_client = client_config.build_rig_client();
        let model = rig_client.completion_model(&request.model.name);
        let agent = AgentBuilder::new(model)
            .preamble(DEFAULT_SYSTEM_PROMPT)
            .tools(tools)
            .default_max_turns(request.max_tool_depth)
            .build();

        let stream = agent
            .stream_prompt(&request.input)
            .with_history(Vec::new())
            .multi_turn(request.max_tool_depth)
            .await;

        tokio::pin!(stream);

        let mut step_count = 0usize;
        let mut final_content = String::new();
        let mut total_input_tokens: i64 = 0;
        let mut total_output_tokens: i64 = 0;

        while let Some(item) = stream.next().await {
            match item {
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::Text(text),
                )) => {
                    step_count += 1;
                    let t = text.text;
                    on_chunk(ReactStep::Answer {
                        step: step_count,
                        answer: t.clone(),
                    })
                    .await;
                    final_content.push_str(&t);
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::Reasoning(reasoning),
                )) => {
                    let reasoning_text = reasoning.reasoning.join("");
                    if !reasoning_text.is_empty() {
                        step_count += 1;
                        on_chunk(ReactStep::Thought {
                            step: step_count,
                            thought: reasoning_text,
                        })
                        .await;
                    }
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::ReasoningDelta { reasoning, .. },
                )) => {
                    if !reasoning.is_empty() {
                        step_count += 1;
                        on_chunk(ReactStep::Thought {
                            step: step_count,
                            thought: reasoning,
                        })
                        .await;
                    }
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::ToolCall { tool_call, .. },
                )) => {
                    step_count += 1;
                    let args: serde_json::Value = match &tool_call.function.arguments {
                        serde_json::Value::String(s) => {
                            serde_json::from_str(s).unwrap_or(serde_json::Value::Null)
                        }
                        v => v.clone(),
                    };
                    on_chunk(ReactStep::Action {
                        step: step_count,
                        action: ReactAction::new(&tool_call.function.name, args),
                    })
                    .await;
                }
                Ok(MultiTurnStreamItem::StreamUserItem(
                    rig::streaming::StreamedUserContent::ToolResult { tool_result, .. },
                )) => {
                    step_count += 1;
                    let obs = tool_result_content_to_string(&tool_result.content);
                    on_chunk(ReactStep::Observation {
                        step: step_count,
                        observation: obs,
                    })
                    .await;
                }
                Ok(MultiTurnStreamItem::FinalResponse(resp)) => {
                    let usage = resp.usage();
                    total_input_tokens = usage.input_tokens as i64;
                    total_output_tokens = usage.output_tokens as i64;
                    // Text was already streamed incrementally via Answer events.
                }
                Err(e) => {
                    let err_msg = format!("rig agent stream error: {}", e);
                    return Err(AgentError::OpenAi(err_msg));
                }
                _ => {}
            }
        }

        let elapsed_ms = session_start.elapsed().as_millis() as i64;
        record_ai_session(
            &request.cache,
            &request.db,
            request.project.id,
            session_id,
            request.room.id,
            request.model.id,
            version_id.unwrap_or_default(),
            total_input_tokens,
            total_output_tokens,
            elapsed_ms,
        )
        .await;

        Ok((final_content, total_input_tokens, total_output_tokens))
    }

    // ── CoT (Chain-of-Thought) ────────────────────────────────────────────

    /// Run a CoT (Chain-of-Thought) reasoning cycle — step-by-step reasoning with optional tool use.
    pub async fn process_cot<C, Fut>(&self, request: &AiRequest, mut on_chunk: C) -> Result<(String, i64, i64)>
    where
        C: FnMut(crate::modes::cot::CotStep) -> Fut + Send,
        Fut: std::future::Future<Output = ()> + Send,
    {
        let client_config = AiClientConfig::new(
            self.ai_api_key.clone().unwrap_or_default(),
        )
        .with_base_url(
            self.ai_base_url
                .clone()
                .unwrap_or_else(|| "https://api.openai.com".into()),
        );
        let rig_client = client_config.build_rig_client();

        let Some(registry) = &self.tool_registry else {
            return Err(AgentError::Internal("no tool registry registered".into()));
        };

        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();
        let version_id = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await
            .ok()
            .flatten()
            .and_then(|r| r.version);

        let db = request.db.clone();
        let cache = request.cache.clone();
        let cfg = request.config.clone();
        let room_id = request.room.id;
        let sender_uid = request.sender.uid;
        let project_id = request.project.id;

        let mut tools: Vec<Box<dyn rig::tool::ToolDyn + 'static>> = Vec::new();
        for def in registry.definitions() {
            let name = def.name.clone();
            if let Some(handler) = registry.get(&name) {
                let adapter = crate::tool::RigToolAdapter::new(
                    handler.clone(), def.clone(),
                    db.clone(), cache.clone(), cfg.clone(),
                    room_id, Some(sender_uid), project_id,
                );
                tools.push(Box::new(crate::tool::RecordingTool::new(
                    Box::new(adapter), db.clone(), session_id, sender_uid,
                )));
            }
        }

        let model = rig_client.completion_model(&request.model.name);
        let agent = AgentBuilder::new(model)
            .preamble(crate::modes::cot::COT_SYSTEM_PROMPT)
            .tools(tools)
            .default_max_turns(request.max_tool_depth)
            .build();

        let stream = agent
            .stream_prompt(&request.input)
            .with_history(Vec::new())
            .multi_turn(request.max_tool_depth)
            .await;

        tokio::pin!(stream);

        let mut final_content = String::new();
        let mut total_input_tokens: i64 = 0;
        let mut total_output_tokens: i64 = 0;

        while let Some(item) = stream.next().await {
            match item {
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::Text(text),
                )) => {
                    let t = text.text;
                    on_chunk(crate::modes::cot::CotStep::Answer(t.clone())).await;
                    final_content.push_str(&t);
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::Reasoning(reasoning),
                )) => {
                    let r = reasoning.reasoning.join("");
                    if !r.is_empty() {
                        on_chunk(crate::modes::cot::CotStep::Thought(r)).await;
                    }
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::ReasoningDelta { reasoning, .. },
                )) => {
                    if !reasoning.is_empty() {
                        on_chunk(crate::modes::cot::CotStep::Thought(reasoning)).await;
                    }
                }
                Ok(MultiTurnStreamItem::StreamAssistantItem(
                    StreamedAssistantContent::ToolCall { tool_call, .. },
                )) => {
                    let args: serde_json::Value = match &tool_call.function.arguments {
                        serde_json::Value::String(s) => {
                            serde_json::from_str(s).unwrap_or(serde_json::Value::Null)
                        }
                        v => v.clone(),
                    };
                    on_chunk(crate::modes::cot::CotStep::Action {
                        name: tool_call.function.name.clone(),
                        args,
                    }).await;
                }
                Ok(MultiTurnStreamItem::StreamUserItem(
                    rig::streaming::StreamedUserContent::ToolResult { tool_result, .. },
                )) => {
                    let obs = tool_result_content_to_string(&tool_result.content);
                    on_chunk(crate::modes::cot::CotStep::Observation(obs)).await;
                }
                Ok(MultiTurnStreamItem::FinalResponse(resp)) => {
                    let usage = resp.usage();
                    total_input_tokens = usage.input_tokens as i64;
                    total_output_tokens = usage.output_tokens as i64;
                }
                Err(e) => {
                    return Err(AgentError::OpenAi(e.to_string()));
                }
                _ => {}
            }
        }

        let elapsed_ms = session_start.elapsed().as_millis() as i64;
        record_ai_session(
            &request.cache, &request.db, request.project.id,
            session_id, request.room.id, request.model.id,
            version_id.unwrap_or_default(),
            total_input_tokens, total_output_tokens, elapsed_ms,
        ).await;

        Ok((final_content, total_input_tokens, total_output_tokens))
    }

    // ── ReWOO (Plan → Execute → Synthesize) ───────────────────────────────

    /// Run a ReWOO reasoning cycle: model plans tool calls, they are executed,
    /// then the model synthesises the final answer.
    pub async fn process_rewoo<C, Fut>(&self, request: &AiRequest, mut on_chunk: C) -> Result<(String, i64, i64)>
    where
        C: FnMut(crate::modes::rewoo::ReWooStep) -> Fut + Send,
        Fut: std::future::Future<Output = ()> + Send,
    {
        let client_config = AiClientConfig::new(
            self.ai_api_key.clone().unwrap_or_default(),
        )
        .with_base_url(
            self.ai_base_url
                .clone()
                .unwrap_or_else(|| "https://api.openai.com".into()),
        );
        let rig_client = client_config.build_rig_client();

        let Some(registry) = &self.tool_registry else {
            return Err(AgentError::Internal("no tool registry registered".into()));
        };

        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();
        let version_id = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await
            .ok()
            .flatten()
            .and_then(|r| r.version);

        let mut total_input_tokens: i64 = 0;
        let mut total_output_tokens: i64 = 0;

        let mut messages = self.build_messages(request).await?;
        messages.insert(0, crate::client::types::ChatRequestMessage::system(
            crate::modes::rewoo::REWOO_SYSTEM_PROMPT.to_string(),
        ));
        let model = rig_client.completion_model(&request.model.name);

        let plan_tools = {
            let db = request.db.clone();
            let cache = request.cache.clone();
            let cfg = request.config.clone();
            let room_id = request.room.id;
            let sender_uid = request.sender.uid;
            let project_id = request.project.id;

            let mut tools: Vec<Box<dyn rig::tool::ToolDyn + 'static>> = Vec::new();
            for def in registry.definitions() {
                let name = def.name.clone();
                if let Some(handler) = registry.get(&name) {
                    let adapter = crate::tool::RigToolAdapter::new(
                        handler.clone(),
                        def.clone(),
                        db.clone(),
                        cache.clone(),
                        cfg.clone(),
                        room_id,
                        Some(sender_uid),
                        project_id,
                    );
                    tools.push(Box::new(crate::tool::RecordingTool::new(
                        Box::new(adapter),
                        db.clone(),
                        session_id,
                        sender_uid,
                    )));
                }
            }
            tools
        };

        let plan_agent = rig::agent::AgentBuilder::new(model)
            .preamble(crate::modes::rewoo::REWOO_SYSTEM_PROMPT)
            .tools(plan_tools)
            .default_max_turns(1)
            .build();

        let plan_response = plan_agent
            .prompt(&request.input)
            .extended_details()
            .await
            .map_err(|e| AgentError::OpenAi(e.to_string()))?;

        total_input_tokens += plan_response.total_usage.input_tokens as i64;
        total_output_tokens += plan_response.total_usage.output_tokens as i64;

        let plan = crate::modes::rewoo::extract_plan(&plan_response.output)
            .unwrap_or_default();

        if plan.calls.is_empty() {
            on_chunk(crate::modes::rewoo::ReWooStep::Synthesis(plan_response.output.clone())).await;
            let elapsed_ms = session_start.elapsed().as_millis() as i64;
            record_ai_session(
                &request.cache, &request.db, request.project.id,
                session_id, request.room.id, request.model.id,
                version_id.unwrap_or_default(),
                total_input_tokens, total_output_tokens, elapsed_ms,
            ).await;
            return Ok((plan_response.output, total_input_tokens, total_output_tokens));
        }

        on_chunk(crate::modes::rewoo::ReWooStep::Plan {
            calls: plan.calls.clone(),
            raw: plan.raw_text,
        }).await;

        // ── Phase 2: Execute all tool calls in parallel ───────────────────
        let mut tool_results: Vec<(String, String)> = Vec::new();
        let mut handles = Vec::new();

        for call in &plan.calls {
            let ctx = crate::tool::ToolContext::new(
                request.db.clone(),
                request.cache.clone(),
                request.config.clone(),
                request.room.id,
                Some(request.sender.uid),
            )
            .with_project(request.project.id);
            if let Some(ref es) = self.embed_service {
                // ctx = ctx.with_embed_service(es.clone()); -- not clone-able via pattern, skip
                let _ = es;
            }

            let call_id = call.step.to_string();
            let tool_name = call.tool.clone();
            let args = call.args.clone();
            let ctx_clone = ctx.clone();

            let handle = tokio::spawn(async move {
                let executor = crate::tool::ToolExecutor::new();
                let agent_call = crate::tool::ToolCall {
                    id: call_id,
                    name: tool_name.clone(),
                    arguments: args.to_string(),
                };
                let mut local_ctx = ctx_clone;
                let result = executor.execute_batch(vec![agent_call], &mut local_ctx).await;
                match result {
                    Ok(results) => {
                        for r in &results {
                            match &r.result {
                                crate::tool::ToolResult::Ok(v) => {
                                    return (tool_name, v.to_string());
                                }
                                crate::tool::ToolResult::Error(e) => {
                                    return (tool_name, format!("[Error: {}]", e));
                                }
                            }
                        }
                        (tool_name, "[No result]".to_string())
                    }
                    Err(e) => (tool_name, format!("[Execution error: {}]", e)),
                }
            });
            handles.push(handle);
        }

        for handle in handles {
            match handle.await {
                Ok((name, result)) => {
                    on_chunk(crate::modes::rewoo::ReWooStep::Execution {
                        tool_name: name.clone(),
                        result: result.clone(),
                    }).await;
                    tool_results.push((name, result));
                }
                Err(e) => {
                    let msg = format!("[Task panicked: {}]", e);
                    on_chunk(crate::modes::rewoo::ReWooStep::Execution {
                        tool_name: "unknown".into(),
                        result: msg.clone(),
                    }).await;
                    tool_results.push(("unknown".into(), msg));
                }
            }
        }

        // ── Phase 3: Synthesize ───────────────────────────────────────────
        let mut synth_messages = self.build_messages(request).await?;
        synth_messages.insert(0, crate::client::types::ChatRequestMessage::system(
            crate::modes::rewoo::REWOO_SYSTEM_PROMPT.to_string(),
        ));

        let results_summary: String = tool_results
            .iter()
            .map(|(name, res)| format!("- {}:\n{}", name, res))
            .collect::<Vec<_>>()
            .join("\n");

        synth_messages.push(crate::client::types::ChatRequestMessage::system(format!(
            "## Tool Execution Results\n\nThe following tool calls were executed:\n\n{}\n\nNow synthesize your final answer based on these results.",
            results_summary
        )));
        synth_messages.push(crate::client::types::ChatRequestMessage::user(&request.input));

        let preamble = synth_messages
            .iter()
            .find(|m| m.role == "system")
            .and_then(|m| m.content.as_deref())
            .unwrap_or("")
            .to_string();
        let non_system: Vec<_> = synth_messages
            .iter()
            .filter(|m| m.role != "system")
            .map(|m| crate::client::to_rig_message(m))
            .collect();

        let synth_model = rig_client.completion_model(&request.model.name);
        let synth_stream = synth_model
            .completion_request("")
            .preamble(preamble)
            .messages(non_system)
            .temperature(request.temperature as f64)
            .max_tokens(request.max_tokens as u64)
            .stream()
            .await
            .map_err(|e| AgentError::OpenAi(e.to_string()))?;

        use rig::streaming::StreamedAssistantContent;
        tokio::pin!(synth_stream);

        let mut synthesis = String::new();
        while let Some(item) = synth_stream.next().await {
            match item {
                Ok(StreamedAssistantContent::Text(text)) => {
                    let t = text.text;
                    on_chunk(crate::modes::rewoo::ReWooStep::Synthesis(t.clone())).await;
                    synthesis.push_str(&t);
                }
                Ok(StreamedAssistantContent::Final(response)) => {
                    if let Some(usage) = response.token_usage() {
                        total_input_tokens += usage.input_tokens as i64;
                        total_output_tokens += usage.output_tokens as i64;
                    }
                }
                Err(e) => return Err(AgentError::OpenAi(e.to_string())),
                _ => {}
            }
        }

        let elapsed_ms = session_start.elapsed().as_millis() as i64;
        record_ai_session(
            &request.cache, &request.db, request.project.id,
            session_id, request.room.id, request.model.id,
            version_id.unwrap_or_default(),
            total_input_tokens, total_output_tokens, elapsed_ms,
        ).await;

        Ok((synthesis, total_input_tokens, total_output_tokens))
    }

    // ── Reflexion (Generate → Critique → Revise) ──────────────────────────

    /// Run a Reflexion reasoning cycle: generate → critique → revise (up to 3 rounds).
    pub async fn process_reflexion<C, Fut>(
        &self,
        request: &AiRequest,
        mut on_chunk: C,
        max_cycles: usize,
    ) -> Result<(String, i64, i64)>
    where
        C: FnMut(crate::modes::reflexion::ReflexionStep) -> Fut + Send,
        Fut: std::future::Future<Output = ()> + Send,
    {
        let client_config = AiClientConfig::new(
            self.ai_api_key.clone().unwrap_or_default(),
        )
        .with_base_url(
            self.ai_base_url
                .clone()
                .unwrap_or_else(|| "https://api.openai.com".into()),
        );
        let rig_client = client_config.build_rig_client();
        let Some(registry) = &self.tool_registry else {
            return Err(AgentError::Internal("no tool registry registered".into()));
        };

        let session_id = Uuid::now_v7();
        let session_start = std::time::Instant::now();
        let version_id = room_ai::Entity::find()
            .filter(room_ai::Column::Room.eq(request.room.id))
            .filter(room_ai::Column::Model.eq(request.model.id))
            .one(&request.db)
            .await
            .ok()
            .flatten()
            .and_then(|r| r.version);

        let max_cycles = max_cycles.min(3);

        let mut total_input_tokens: i64 = 0;
        let mut total_output_tokens: i64 = 0;
        let mut best_answer = String::new();

        for cycle in 0..max_cycles {
            let mut messages = self.build_messages(request).await?;
            messages.insert(0, crate::client::types::ChatRequestMessage::system(
                crate::modes::reflexion::REFLEXION_SYSTEM_PROMPT.to_string(),
            ));

            if cycle > 0 {
                messages.push(crate::client::types::ChatRequestMessage::system(format!(
                    "This is cycle {} of the reflexion process. Your previous answer was:\n\n{}\n\nPlease critique and improve upon it.",
                    cycle + 1,
                    best_answer
                )));
            }

            // Build tools for this cycle (not cloneable, so rebuild each iteration)
            let cycle_tools = build_rig_tools(
                registry, &request.db, &request.cache, &request.config,
                request.room.id, request.sender.uid, request.project.id, session_id,
            );

            // ── Generate ──────────────────────────────────────────────
            let model = rig_client.completion_model(&request.model.name);
            let agent = rig::agent::AgentBuilder::new(model)
                .preamble(crate::modes::reflexion::REFLEXION_SYSTEM_PROMPT)
                .tools(cycle_tools)
                .default_max_turns(request.max_tool_depth)
                .build();

            let stream = agent
                .stream_prompt(&request.input)
                .with_history(Vec::new())
                .multi_turn(request.max_tool_depth)
                .await;

            tokio::pin!(stream);
            let mut generated = String::new();

            while let Some(item) = stream.next().await {
                match item {
                    Ok(rig::agent::MultiTurnStreamItem::StreamAssistantItem(
                        rig::streaming::StreamedAssistantContent::Text(text),
                    )) => {
                        generated.push_str(&text.text);
                    }
                    Ok(rig::agent::MultiTurnStreamItem::FinalResponse(resp)) => {
                        let usage = resp.usage();
                        total_input_tokens += usage.input_tokens as i64;
                        total_output_tokens += usage.output_tokens as i64;
                    }
                    Err(e) => return Err(AgentError::OpenAi(e.to_string())),
                    _ => {}
                }
            }

            best_answer = generated.clone();
            on_chunk(crate::modes::reflexion::ReflexionStep::Generate(generated.clone())).await;

            // If only 1 cycle, emit final and exit
            if max_cycles == 1 || cycle + 1 >= max_cycles {
                on_chunk(crate::modes::reflexion::ReflexionStep::Final(generated.clone())).await;
                break;
            }

            // ── Self-critique ─────────────────────────────────────────
            let critique_messages = vec![
                crate::client::types::ChatRequestMessage::system(crate::modes::reflexion::REFLEXION_SYSTEM_PROMPT),
                crate::client::types::ChatRequestMessage::system(format!(
                    "Your previous answer was:\n\n{}", generated
                )),
                crate::client::types::ChatRequestMessage::user(crate::modes::reflexion::REFLEXION_CRITIQUE_PROMPT),
            ];

            let critique_result = crate::client::call_with_params(
                &critique_messages,
                &request.model.name,
                &client_config,
                request.temperature as f32,
                request.max_tokens as u32,
                None,
                None,
                Some("none"),
            ).await?;

            total_input_tokens += critique_result.input_tokens;
            total_output_tokens += critique_result.output_tokens;
            let critique = critique_result.content;
            on_chunk(crate::modes::reflexion::ReflexionStep::Critique(critique.clone())).await;

            // ── Revise ───────────────────────────────────────────────
            let revise_messages = vec![
                crate::client::types::ChatRequestMessage::user(format!(
                    "Your previous answer:\n\n{}\n\nYour self-critique:\n\n{}",
                    generated, critique
                )),
                crate::client::types::ChatRequestMessage::user(crate::modes::reflexion::REFLEXION_REVISE_PROMPT),
            ];

            let revise_model = rig_client.completion_model(&request.model.name);
            let revise_stream = revise_model
                .completion_request("")
                .preamble(crate::modes::reflexion::REFLEXION_SYSTEM_PROMPT.to_string())
                .messages(revise_messages.iter().map(|m| {
                    crate::client::to_rig_message(m)
                }).collect::<Vec<_>>())
                .temperature(request.temperature as f64)
                .max_tokens(request.max_tokens as u64)
                .stream()
                .await
                .map_err(|e| AgentError::OpenAi(e.to_string()))?;

            tokio::pin!(revise_stream);
            let mut revised = String::new();

            while let Some(item) = revise_stream.next().await {
                match item {
                    Ok(rig::streaming::StreamedAssistantContent::Text(text)) => {
                        revised.push_str(&text.text);
                    }
                    Ok(rig::streaming::StreamedAssistantContent::Final(response)) => {
                        if let Some(usage) = response.token_usage() {
                            total_input_tokens += usage.input_tokens as i64;
                            total_output_tokens += usage.output_tokens as i64;
                        }
                    }
                    Err(e) => return Err(AgentError::OpenAi(e.to_string())),
                    _ => {}
                }
            }

            best_answer = revised.clone();
            on_chunk(crate::modes::reflexion::ReflexionStep::Revise(revised.clone())).await;

            // If last cycle, emit final
            if cycle + 1 >= max_cycles {
                on_chunk(crate::modes::reflexion::ReflexionStep::Final(revised.clone())).await;
            }
        }

        let elapsed_ms = session_start.elapsed().as_millis() as i64;
        record_ai_session(
            &request.cache, &request.db, request.project.id,
            session_id, request.room.id, request.model.id,
            version_id.unwrap_or_default(),
            total_input_tokens, total_output_tokens, elapsed_ms,
        ).await;

        Ok((best_answer, total_input_tokens, total_output_tokens))
    }
}

fn build_rig_tools(
    registry: &crate::tool::ToolRegistry,
    db: &db::database::AppDatabase,
    cache: &db::cache::AppCache,
    cfg: &config::AppConfig,
    room_id: uuid::Uuid,
    sender_uid: uuid::Uuid,
    project_id: uuid::Uuid,
    session_id: uuid::Uuid,
) -> Vec<Box<dyn rig::tool::ToolDyn + 'static>> {
    let mut tools: Vec<Box<dyn rig::tool::ToolDyn + 'static>> = Vec::new();
    for def in registry.definitions() {
        let name = def.name.clone();
        if let Some(handler) = registry.get(&name) {
            let adapter = crate::tool::RigToolAdapter::new(
                handler.clone(), def.clone(),
                db.clone(), cache.clone(), cfg.clone(),
                room_id, Some(sender_uid), project_id,
            );
            tools.push(Box::new(crate::tool::RecordingTool::new(
                Box::new(adapter), db.clone(), session_id, sender_uid,
            )));
        }
    }
    tools
}

/// Extract text from rig's ToolResultContent, ignoring images.
fn tool_result_content_to_string(content: &rig::one_or_many::OneOrMany<rig::completion::message::ToolResultContent>) -> String {
    use rig::completion::message::ToolResultContent;
    content
        .iter()
        .filter_map(|item| {
            if let ToolResultContent::Text(t) = item {
                Some(t.text.clone())
            } else {
                None
            }
        })
        .collect::<Vec<_>>()
        .join("\n")
}