diff --git a/libs/agent/chat/state.rs b/libs/agent/chat/state.rs
new file mode 100644
index 0000000..3972ce8
--- /dev/null
+++ b/libs/agent/chat/state.rs
@@ -0,0 +1,217 @@
+//! Agent state machine — tracks lifecycle of a single AI agent invocation.
+//!
+//! States: Idle → Thinking → ToolCall → Thinking → ... → Answering | Error
+//! The Thinking ↔ ToolCall cycle repeats until max tool depth or final answer.
+
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+
+/// Current phase of an agent's execution lifecycle.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum AgentState {
+    /// Agent is idle, waiting for input
+    Idle,
+    /// Agent is reasoning/thinking (may produce thinking chunks)
+    Thinking {
+        started_at: DateTime<Utc>,
+        tool_depth: u32,
+    },
+    /// Agent is executing a tool call
+    ToolCall {
+        tool_name: String,
+        started_at: DateTime<Utc>,
+    },
+    /// Agent is returning the final answer
+    Answering {
+        /// Accumulated answer content so far
+        content_chars: u64,
+        started_at: DateTime<Utc>,
+    },
+    /// Agent encountered a non-recoverable error
+    Error {
+        message: String,
+        tool_depth: u32,
+    },
+}
+
+impl AgentState {
+    pub fn is_terminal(&self) -> bool {
+        matches!(self, AgentState::Answering { .. } | AgentState::Error { .. })
+    }
+
+    pub fn is_idle(&self) -> bool {
+        matches!(self, AgentState::Idle)
+    }
+
+    pub fn current_phase(&self) -> &'static str {
+        match self {
+            AgentState::Idle => "idle",
+            AgentState::Thinking { .. } => "thinking",
+            AgentState::ToolCall { .. } => "tool_call",
+            AgentState::Answering { .. } => "answering",
+            AgentState::Error { .. } => "error",
+        }
+    }
+}
+
+/// State machine for agent lifecycle transitions.
+pub struct AgentRuntime {
+    state: AgentState,
+    max_tool_depth: u32,
+    current_depth: u32,
+}
+
+impl AgentRuntime {
+    pub fn new(max_tool_depth: u32) -> Self {
+        Self {
+            state: AgentState::Idle,
+            max_tool_depth,
+            current_depth: 0,
+        }
+    }
+
+    pub fn state(&self) -> &AgentState {
+        &self.state
+    }
+
+    /// Transition from Idle → Thinking
+    pub fn start_thinking(&mut self) {
+        debug_assert!(self.state.is_idle(), "must be Idle to start thinking");
+        self.current_depth = 0;
+        self.state = AgentState::Thinking {
+            started_at: Utc::now(),
+            tool_depth: 0,
+        };
+    }
+
+    /// Transition from Thinking → ToolCall (increments tool depth)
+    pub fn start_tool_call(&mut self, tool_name: String) -> Result<(), &'static str> {
+        if !matches!(self.state, AgentState::Thinking { .. }) {
+            return Err("must be Thinking to start tool call");
+        }
+        if self.current_depth >= self.max_tool_depth {
+            return Err("max tool depth reached");
+        }
+        self.state = AgentState::ToolCall {
+            tool_name,
+            started_at: Utc::now(),
+        };
+        Ok(())
+    }
+
+    /// Transition from ToolCall → Thinking (back after tool result)
+    pub fn complete_tool_call(&mut self) -> Result<(), &'static str> {
+        if !matches!(self.state, AgentState::ToolCall { .. }) {
+            return Err("must be ToolCall to complete");
+        }
+        self.current_depth += 1;
+        self.state = AgentState::Thinking {
+            started_at: Utc::now(),
+            tool_depth: self.current_depth,
+        };
+        Ok(())
+    }
+
+    /// Transition to Answering (terminal)
+    pub fn start_answer(&mut self) {
+        self.state = AgentState::Answering {
+            content_chars: 0,
+            started_at: Utc::now(),
+        };
+    }
+
+    pub fn append_answer(&mut self, content: &str) {
+        if let AgentState::Answering { content_chars, .. } = &mut self.state {
+            *content_chars += content.len() as u64;
+        }
+    }
+
+    /// Transition to Error (terminal)
+    pub fn fail(&mut self, message: String) {
+        self.state = AgentState::Error {
+            message,
+            tool_depth: self.current_depth,
+        };
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_starts_idle() {
+        let rt = AgentRuntime::new(10);
+        assert!(rt.state().is_idle());
+        assert_eq!(rt.state().current_phase(), "idle");
+    }
+
+    #[test]
+    fn test_idle_to_thinking() {
+        let mut rt = AgentRuntime::new(10);
+        rt.start_thinking();
+        assert_eq!(rt.state().current_phase(), "thinking");
+        assert!(!rt.state().is_terminal());
+    }
+
+    #[test]
+    fn test_thinking_to_tool_call_and_back() {
+        let mut rt = AgentRuntime::new(10);
+        rt.start_thinking();
+        rt.start_tool_call("search".into()).unwrap();
+        assert_eq!(rt.state().current_phase(), "tool_call");
+        rt.complete_tool_call().unwrap();
+        assert_eq!(rt.state().current_phase(), "thinking");
+    }
+
+    #[test]
+    fn test_thinking_to_answer() {
+        let mut rt = AgentRuntime::new(10);
+        rt.start_thinking();
+        rt.start_answer();
+        assert_eq!(rt.state().current_phase(), "answering");
+        assert!(rt.state().is_terminal());
+    }
+
+    #[test]
+    fn test_append_answer_tracks_chars() {
+        let mut rt = AgentRuntime::new(10);
+        rt.start_thinking();
+        rt.start_answer();
+        rt.append_answer("hello");
+        if let AgentState::Answering { content_chars, .. } = rt.state() {
+            assert_eq!(*content_chars, 5);
+        } else {
+            panic!("expected Answering state");
+        }
+    }
+
+    #[test]
+    fn test_error_is_terminal() {
+        let mut rt = AgentRuntime::new(10);
+        rt.start_thinking();
+        rt.fail("something broke".into());
+        assert_eq!(rt.state().current_phase(), "error");
+        assert!(rt.state().is_terminal());
+    }
+
+    #[test]
+    fn test_transition_from_wrong_state() {
+        let mut rt = AgentRuntime::new(10);
+        // Can't start tool call from Idle
+        assert!(rt.start_tool_call("tool".into()).is_err());
+        // Can't complete tool call from Idle
+        assert!(rt.complete_tool_call().is_err());
+    }
+
+    #[test]
+    fn test_max_depth_rejected() {
+        let mut rt = AgentRuntime::new(2);
+        rt.start_thinking();
+        rt.start_tool_call("tool1".into()).unwrap();
+        rt.complete_tool_call().unwrap();
+        rt.start_tool_call("tool2".into()).unwrap();
+        rt.complete_tool_call().unwrap();
+        assert!(rt.start_tool_call("tool3".into()).is_err());
+    }
+}
diff --git a/libs/room/src/service/ai_mode_dispatch.rs b/libs/room/src/service/ai_mode_dispatch.rs
new file mode 100644
index 0000000..7a0aebf
--- /dev/null
+++ b/libs/room/src/service/ai_mode_dispatch.rs
@@ -0,0 +1,183 @@
+use std::sync::Arc;
+
+use db::cache::AppCache;
+use db::database::AppDatabase;
+use queue::MessageProducer;
+use uuid::Uuid;
+
+use super::ai_mode_streaming::run_mode_streaming;
+use crate::connection::RoomConnectionManager;
+use agent::chat::{AiRequest, ChatService};
+
+// ── CoT ──────────────────────────────────────────────────────────────────────
+
+pub async fn dispatch_cot(
+    chat_service: Arc<ChatService>,
+    request: AiRequest,
+    room_id: Uuid,
+    project_id: Uuid,
+    model_id: Uuid,
+    lock_guard: crate::room_ai_queue::RoomAiLockGuard,
+    db: AppDatabase,
+    cache: AppCache,
+    queue: MessageProducer,
+    room_manager: Arc<RoomConnectionManager>,
+) {
+    let req = request;
+    run_mode_streaming(
+        chat_service,
+        req,
+        room_id,
+        project_id,
+        model_id,
+        lock_guard,
+        db,
+        cache,
+        queue,
+        room_manager,
+        "cot",
+        Box::new(|chat_svc, ai_req, on_chunk| {
+            Box::pin(async move {
+                let result = chat_svc
+                    .process_cot(&ai_req, |step| {
+                        let on_chunk = on_chunk.clone();
+                        async move {
+                            let (ct, content, is_final) = match step {
+                                agent::modes::cot::CotStep::Thought(t) => {
+                                    ("thinking".to_string(), t, false)
+                                }
+                                agent::modes::cot::CotStep::Action { name, args } => {
+                                    ("tool_call".to_string(),
+                                     serde_json::json!({"name": name, "arguments": args}).to_string(),
+                                     false)
+                                }
+                                agent::modes::cot::CotStep::Observation(o) => {
+                                    ("tool_result".to_string(), o, false)
+                                }
+                                agent::modes::cot::CotStep::Answer(a) => {
+                                    ("answer".to_string(), a, true)
+                                }
+                            };
+                            on_chunk(ct, content, is_final).await;
+                        }
+                    })
+                    .await;
+                result
+            })
+        }),
+    )
+    .await;
+}
+
+// ── ReWOO ────────────────────────────────────────────────────────────────────
+
+pub async fn dispatch_rewoo(
+    chat_service: Arc<ChatService>,
+    request: AiRequest,
+    room_id: Uuid,
+    project_id: Uuid,
+    model_id: Uuid,
+    lock_guard: crate::room_ai_queue::RoomAiLockGuard,
+    db: AppDatabase,
+    cache: AppCache,
+    queue: MessageProducer,
+    room_manager: Arc<RoomConnectionManager>,
+) {
+    let req = request;
+    run_mode_streaming(
+        chat_service.clone(),
+        req,
+        room_id,
+        project_id,
+        model_id,
+        lock_guard,
+        db,
+        cache,
+        queue,
+        room_manager,
+        "rewoo",
+        Box::new(|chat_svc, ai_req, on_chunk| {
+            Box::pin(async move {
+                let result = chat_svc
+                    .process_rewoo(&ai_req, |step| {
+                        let on_chunk = on_chunk.clone();
+                        async move {
+                            let (ct, content, is_final) = match step {
+                                agent::modes::rewoo::ReWooStep::Plan { raw, .. } => {
+                                    ("tool_call".to_string(), raw, false)
+                                }
+                                agent::modes::rewoo::ReWooStep::Execution { tool_name, result } => {
+                                    ("tool_result".to_string(), format!("{}: {}", tool_name, result), false)
+                                }
+                                agent::modes::rewoo::ReWooStep::Synthesis(s) => {
+                                    ("answer".to_string(), s, true)
+                                }
+                            };
+                            on_chunk(ct, content, is_final).await;
+                        }
+                    })
+                    .await;
+                result
+            })
+        }),
+    )
+    .await;
+}
+
+// ── Reflexion ────────────────────────────────────────────────────────────────
+
+pub async fn dispatch_reflexion(
+    chat_service: Arc<ChatService>,
+    request: AiRequest,
+    room_id: Uuid,
+    project_id: Uuid,
+    model_id: Uuid,
+    lock_guard: crate::room_ai_queue::RoomAiLockGuard,
+    db: AppDatabase,
+    cache: AppCache,
+    queue: MessageProducer,
+    room_manager: Arc<RoomConnectionManager>,
+) {
+    let req = request;
+    run_mode_streaming(
+        chat_service.clone(),
+        req,
+        room_id,
+        project_id,
+        model_id,
+        lock_guard,
+        db,
+        cache,
+        queue,
+        room_manager,
+        "reflexion",
+        Box::new(|chat_svc, ai_req, on_chunk| {
+            Box::pin(async move {
+                let result = chat_svc
+                    .process_reflexion(&ai_req, |step| {
+                        let on_chunk = on_chunk.clone();
+                        async move {
+                            let (ct, content, is_final) = match step {
+                                agent::modes::reflexion::ReflexionStep::Generate(s) => {
+                                    ("thinking".to_string(), s, false)
+                                }
+                                agent::modes::reflexion::ReflexionStep::Critique(s) => {
+                                    ("tool_call".to_string(), s, false)
+                                }
+                                agent::modes::reflexion::ReflexionStep::Revise(s) => {
+                                    ("thinking".to_string(), s, false)
+                                }
+                                agent::modes::reflexion::ReflexionStep::Final(s) => {
+                                    ("answer".to_string(), s, true)
+                                }
+                            };
+                            on_chunk(ct, content, is_final).await;
+                        }
+                    }, 3)
+                    .await;
+                result
+            })
+        }),
+    )
+    .await;
+}
diff --git a/libs/room/src/service/ai_mode_streaming.rs b/libs/room/src/service/ai_mode_streaming.rs
new file mode 100644
index 0000000..aeca3f2
--- /dev/null
+++ b/libs/room/src/service/ai_mode_streaming.rs
@@ -0,0 +1,324 @@
+use std::pin::Pin;
+use std::sync::Arc;
+
+use chrono::Utc;
+use db::cache::AppCache;
+use db::database::AppDatabase;
+use models::rooms::room_ai;
+use queue::{MessageProducer, ProjectRoomEvent, RoomMessageEnvelope};
+use sea_orm::{sea_query::Expr, ColumnTrait, EntityTrait, ExprTrait, QueryFilter};
+use uuid::Uuid;
+
+use super::sequence::next_room_message_seq_internal;
+use crate::connection::RoomConnectionManager;
+use agent::chat::{AiRequest, ChatService};
+
+pub type RunModeFn = Box<
+    dyn FnOnce(
+            Arc<ChatService>,
+            AiRequest,
+            Arc<dyn Fn(String, String, bool) -> Pin<Box<dyn std::future::Future<Output = ()> + Send>> + Send + Sync>,
+        ) -> Pin<Box<dyn std::future::Future<Output = Result<(String, i64, i64), agent::AgentError>> + Send>>
+        + Send,
+>;
+
+pub async fn run_mode_streaming(
+    chat_service: Arc<ChatService>,
+    request: AiRequest,
+    room_id: Uuid,
+    project_id: Uuid,
+    model_id: Uuid,
+    lock_guard: crate::room_ai_queue::RoomAiLockGuard,
+    db: AppDatabase,
+    cache: AppCache,
+    queue: MessageProducer,
+    room_manager: Arc<RoomConnectionManager>,
+    mode_name_str: &str,
+    run: RunModeFn,
+) {
+    let mode_name = mode_name_str.to_string();
+    use queue::RoomMessageStreamChunkEvent;
+
+    let streaming_msg_id = Uuid::now_v7();
+    let seq = match next_room_message_seq_internal(room_id, &db, &cache).await {
+        Ok(s) => s,
+        Err(e) => {
+            tracing::error!(error = %e, "Failed to get seq for {} streaming", mode_name);
+            return;
+        }
+    };
+
+    let _ = room_manager
+        .register_stream_channel(streaming_msg_id)
+        .await;
+
+    let initial_event = RoomMessageStreamChunkEvent {
+        message_id: streaming_msg_id,
+        room_id,
+        seq: 0,
+        content: String::new(),
+        done: false,
+        error: None,
+        display_name: Some(request.model.name.clone()),
+        chunk_type: Some("thinking".to_string()),
+    };
+    room_manager.broadcast_stream_chunk(initial_event).await;
+
+    let room_id_inner = room_id;
+    let project_id_inner = project_id;
+    let now = Utc::now();
+    let sender_type = "ai".to_string();
+    let ai_display_name = request.model.name.clone();
+
+    tokio::spawn(async move {
+        let _lock_guard = lock_guard;
+        let cancel = room_manager.register_stream_cancel(room_id_inner).await;
+
+        let ai_typing_id = Uuid::parse_str("00000000-0000-0000-0000-000000000001")
+            .expect("constant UUID should always parse");
+
+        let typing_start = queue::TypingEvent {
+            room_id: room_id_inner,
+            user_id: ai_typing_id,
+            username: ai_display_name.clone(),
+            avatar_url: None,
+            action: "start".to_string(),
+            sender_type: Some("ai".to_string()),
+        };
+        room_manager.broadcast_typing(room_id_inner, typing_start.clone()).await;
+
+        let (typing_cancel_tx, typing_cancel_rx) = tokio::sync::oneshot::channel::<()>();
+        let typing_renew_handle = tokio::spawn({
+            let mut interval = tokio::time::interval(std::time::Duration::from_secs(30));
+            interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+            let mgr = room_manager.clone();
+            let rid = room_id_inner;
+            let evt = typing_start.clone();
+            async move {
+                tokio::select! {
+                    _ = typing_cancel_rx => {}
+                    _ = async {
+                        loop {
+                            interval.tick().await;
+                            mgr.broadcast_typing(rid, evt.clone()).await;
+                        }
+                    } => {}
+                }
+            }
+        });
+
+        let chunk_seq: Arc<std::sync::atomic::AtomicU64> =
+            Arc::new(std::sync::atomic::AtomicU64::new(1));
+        let all_chunks: Arc<std::sync::Mutex<Vec<(String, String)>>> =
+            Arc::new(std::sync::Mutex::new(Vec::new()));
+        let answer_buffer: Arc<std::sync::Mutex<String>> =
+            Arc::new(std::sync::Mutex::new(String::new()));
+
+        fn lock_or_recover<T>(mutex: &std::sync::Mutex<T>) -> std::sync::MutexGuard<'_, T> {
+            mutex.lock().unwrap_or_else(|poisoned| poisoned.into_inner())
+        }
+
+        let on_chunk = {
+            let room_manager = room_manager.clone();
+            let queue = queue.clone();
+            let cancel = cancel.clone();
+            let streaming_msg_id = streaming_msg_id;
+            let room_id = room_id_inner;
+            let chunk_seq = chunk_seq.clone();
+            let ai_display_name = ai_display_name.clone();
+            let all_chunks = all_chunks.clone();
+            let answer_buffer = answer_buffer.clone();
+
+            Arc::new(move |chunk_type: String, content: String, is_answer: bool| {
+                let room_manager = room_manager.clone();
+                let queue = queue.clone();
+                let cancel = cancel.clone();
+                let chunk_seq = chunk_seq.clone();
+                let ai_display_name = ai_display_name.clone();
+                let all_chunks = all_chunks.clone();
+                let answer_buffer = answer_buffer.clone();
+
+                {
+                    let mut chunks = lock_or_recover(&all_chunks);
+                    chunks.push((chunk_type.clone(), content.clone()));
+                }
+                if is_answer {
+                    let mut ab = lock_or_recover(&answer_buffer);
+                    ab.push_str(&content);
+                }
+
+                let current_seq = chunk_seq.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+                let event = RoomMessageStreamChunkEvent {
+                    message_id: streaming_msg_id,
+                    room_id,
+                    seq: current_seq,
+                    content: content.clone(),
+                    done: false,
+                    error: None,
+                    display_name: Some(ai_display_name),
+                    chunk_type: Some(chunk_type),
+                };
+
+                Box::pin(async move {
+                    if cancel.load(std::sync::atomic::Ordering::Acquire) {
+                        return;
+                    }
+                    queue.publish_stream_chunk(&event).await;
+                    room_manager.broadcast_stream_chunk(event).await;
+                }) as Pin<Box<dyn std::future::Future<Output = ()> + Send>>
+            })
+        };
+
+        let result = run(chat_service, request, on_chunk).await;
+
+        let final_stream_content = lock_or_recover(&answer_buffer).clone();
+        let final_event = RoomMessageStreamChunkEvent {
+            message_id: streaming_msg_id,
+            room_id: room_id_inner,
+            seq: chunk_seq.fetch_add(1, std::sync::atomic::Ordering::Relaxed),
+            content: final_stream_content.clone(),
+            done: true,
+            error: None,
+            display_name: Some(ai_display_name.clone()),
+            chunk_type: Some("answer".to_string()),
+        };
+        queue.publish_stream_chunk(&final_event).await;
+        room_manager.broadcast_stream_chunk(final_event).await;
+
+        let (final_content, err_msg) = match result {
+            Ok((content, _input, _output)) => (content, None),
+            Err(e) => {
+                let msg = "[AI 处理发生错误，请稍后再试]".to_string();
+                tracing::error!(error = %e, "{} streaming failed", mode_name);
+                (String::new(), Some(msg))
+            }
+        };
+
+        let all_chunks_data = lock_or_recover(&all_chunks).clone();
+        let reasoning_chain: String = all_chunks_data
+            .iter()
+            .filter(|(t, _)| t != "answer")
+            .map(|(_, c)| c.clone())
+            .collect::<Vec<_>>()
+            .join("\n");
+
+        let content_to_persist = if !final_content.is_empty() {
+            final_content.clone()
+        } else if !reasoning_chain.trim().is_empty() {
+            format!(
+                "[Agent ran through reasoning steps but did not produce a final answer.]\n{}",
+                reasoning_chain.trim_end()
+            )
+        } else {
+            String::from("[No output from reasoning agent]")
+        };
+        let content_to_persist = if let Some(msg) = &err_msg {
+            format!("{}\n[Error: {}]", content_to_persist.trim_end(), msg)
+        } else {
+            content_to_persist
+        };
+
+        let persist_content = content_to_persist.trim().to_string();
+        if persist_content.is_empty() {
+            let _ = typing_cancel_tx.send(());
+            typing_renew_handle.abort();
+            room_manager.unregister_stream_cancel(room_id_inner).await;
+            room_manager.close_stream_channel(streaming_msg_id).await;
+            return;
+        }
+
+        let thinking_content_serialized = {
+            let chunks = lock_or_recover(&all_chunks);
+            if chunks.is_empty() {
+                None
+            } else {
+                let chunks_json = serde_json::json!({
+                    "__chunks__": chunks.iter().map(|(t, c)| serde_json::json!({
+                        "type": t,
+                        "content": c,
+                    })).collect::<Vec<_>>(),
+                });
+                Some(chunks_json.to_string())
+            }
+        };
+        let thinking_content_for_event = thinking_content_serialized.clone();
+
+        let envelope = RoomMessageEnvelope {
+            id: streaming_msg_id,
+            dedup_key: Some(format!("{}:{}", room_id_inner, streaming_msg_id)),
+            room_id: room_id_inner,
+            sender_type: sender_type.clone(),
+            sender_id: None,
+            model_id: Some(model_id),
+            thread_id: None,
+            content: persist_content.clone(),
+            content_type: "text".to_string(),
+            thinking_content: thinking_content_serialized,
+            send_at: now,
+            seq,
+            in_reply_to: None,
+            display_name: Some(ai_display_name.clone()),
+        };
+
+        if let Err(e) = queue.publish(room_id_inner, envelope).await {
+            tracing::error!(error = %e, "Failed to publish {} streaming message", mode_name);
+        } else {
+            let now = Utc::now();
+            if let Err(e) = room_ai::Entity::update_many()
+                .col_expr(room_ai::Column::CallCount, Expr::col(room_ai::Column::CallCount).add(1))
+                .col_expr(room_ai::Column::LastCallAt, Expr::value(Some(now)))
+                .filter(room_ai::Column::Room.eq(room_id_inner))
+                .filter(room_ai::Column::Model.eq(model_id))
+                .exec(&db)
+                .await
+            {
+                tracing::warn!(error = %e, "Failed to update room_ai call stats");
+            }
+
+            let msg_event = queue::RoomMessageEvent {
+                id: streaming_msg_id,
+                room_id: room_id_inner,
+                sender_type: sender_type.clone(),
+                sender_id: None,
+                thread_id: None,
+                content: persist_content,
+                content_type: "text".to_string(),
+                thinking_content: thinking_content_for_event,
+                send_at: now,
+                seq,
+                display_name: Some(ai_display_name.clone()),
+                in_reply_to: None,
+                reactions: None,
+                message_id: None,
+            };
+            room_manager.broadcast(room_id_inner, msg_event).await;
+            room_manager.metrics.messages_sent.increment(1);
+
+            let event = ProjectRoomEvent {
+                event_type: crate::RoomEventType::NewMessage.as_str().into(),
+                project_id: project_id_inner,
+                room_id: Some(room_id_inner),
+                category_id: None,
+                message_id: Some(streaming_msg_id),
+                seq: Some(seq),
+                timestamp: now,
+            };
+            queue.publish_project_room_event(project_id_inner, event).await;
+        }
+
+        let _ = typing_cancel_tx.send(());
+        typing_renew_handle.abort();
+        let typing_stop = queue::TypingEvent {
+            room_id: room_id_inner,
+            user_id: ai_typing_id,
+            username: ai_display_name.clone(),
+            avatar_url: None,
+            action: "stop".to_string(),
+            sender_type: Some("ai".to_string()),
+        };
+        room_manager.broadcast_typing(room_id_inner, typing_stop).await;
+
+        room_manager.unregister_stream_cancel(room_id_inner).await;
+        room_manager.close_stream_channel(streaming_msg_id).await;
+    });
+}
diff --git a/libs/room/src/service/ai_service.rs b/libs/room/src/service/ai_service.rs
new file mode 100644
index 0000000..4742c2d
--- /dev/null
+++ b/libs/room/src/service/ai_service.rs
@@ -0,0 +1,292 @@
+use std::sync::Arc;
+
+use uuid::Uuid;
+
+use db::cache::AppCache;
+use db::database::AppDatabase;
+use models::rooms::room_ai;
+use queue::MessageProducer;
+use sea_orm::{ColumnTrait, EntityTrait, QueryFilter};
+
+use crate::connection::RoomConnectionManager;
+use crate::error::RoomError;
+use crate::service::ai_common::create_and_publish_ai_message;
+use crate::service::ai_mode_dispatch;
+use crate::service::ai_nonstreaming;
+use crate::service::ai_react_nonstreaming;
+use crate::service::ai_react_streaming;
+use crate::service::ai_streaming;
+use crate::service::history;
+use crate::service::patterns::{mention_bracket_re, mention_tag_re};
+use agent::chat::{AiRequest, ChatService};
+
+/// Service responsible for AI message generation orchestration.
+/// Decides which execution path to use (streaming/nonstreaming, ReAct/chat)
+/// and dispatches accordingly.
+#[derive(Clone)]
+pub struct RoomAiService {
+    db: AppDatabase,
+    cache: AppCache,
+    config: config::AppConfig,
+    queue: MessageProducer,
+    room_manager: Arc<RoomConnectionManager>,
+    chat_service: Option<Arc<ChatService>>,
+}
+
+impl RoomAiService {
+    pub fn new(
+        db: AppDatabase,
+        cache: AppCache,
+        config: config::AppConfig,
+        queue: MessageProducer,
+        room_manager: Arc<RoomConnectionManager>,
+        chat_service: Option<Arc<ChatService>>,
+    ) -> Self {
+        Self {
+            db,
+            cache,
+            config,
+            queue,
+            room_manager,
+            chat_service,
+        }
+    }
+
+    /// Check if the message content mentions an AI model configured in this room.
+    pub async fn should_respond(&self, room_id: Uuid, content: &str) -> Result<bool, RoomError> {
+        let ai_configs = history::get_room_ai_configs(&self.db, room_id).await?;
+        if ai_configs.is_empty() {
+            return Ok(false);
+        }
+
+        let model_ids: std::collections::HashSet<String> = ai_configs
+            .iter()
+            .map(|c| c.model.to_string())
+            .collect();
+
+        for cap in mention_bracket_re().captures_iter(content) {
+            if let (Some(type_m), Some(id_m)) = (cap.get(1), cap.get(2)) {
+                if type_m.as_str() == "ai" && model_ids.contains(id_m.as_str().trim()) {
+                    return Ok(true);
+                }
+            }
+        }
+
+        for cap in mention_tag_re().captures_iter(content) {
+            if let (Some(type_m), Some(id_m)) = (cap.get(1), cap.get(2)) {
+                if type_m.as_str() == "ai" && model_ids.contains(id_m.as_str().trim()) {
+                    return Ok(true);
+                }
+            }
+        }
+
+        Ok(false)
+    }
+
+    /// Extract the mentioned AI model ID from message content.
+    fn extract_mentioned_model_id(content: &str) -> Option<Uuid> {
+        for cap in mention_bracket_re().captures_iter(content) {
+            if let (Some(type_m), Some(id_m)) = (cap.get(1), cap.get(2)) {
+                if type_m.as_str() == "ai" {
+                    if let Ok(uuid) = Uuid::parse_str(id_m.as_str().trim()) {
+                        return Some(uuid);
+                    }
+                }
+            }
+        }
+        for cap in mention_tag_re().captures_iter(content) {
+            if let (Some(type_m), Some(id_m)) = (cap.get(1), cap.get(2)) {
+                if type_m.as_str() == "ai" {
+                    if let Ok(uuid) = Uuid::parse_str(id_m.as_str().trim()) {
+                        return Some(uuid);
+                    }
+                }
+            }
+        }
+        None
+    }
+
+    /// Main entry point: process an AI message request.
+    /// Handles model lookup, history building, locking, and dispatch.
+    pub async fn process(
+        &self,
+        room_id: Uuid,
+        sender_id: Uuid,
+        content: &str,
+    ) -> Result<(), RoomError> {
+        let chat_service = match &self.chat_service {
+            Some(cs) => cs.clone(),
+            None => return Ok(()),
+        };
+
+        let model_id = match Self::extract_mentioned_model_id(content) {
+            Some(id) => id,
+            None => return Ok(()),
+        };
+
+        let ai_config = match room_ai::Entity::find()
+            .filter(room_ai::Column::Room.eq(room_id))
+            .filter(room_ai::Column::Model.eq(model_id))
+            .one(&self.db)
+            .await?
+        {
+            Some(c) => c,
+            None => return Ok(()),
+        };
+
+        // Idempotency check: skip if this content already triggered AI within 60s
+        use std::collections::hash_map::DefaultHasher;
+        use std::hash::{Hash, Hasher};
+        let mut hasher = DefaultHasher::new();
+        content.hash(&mut hasher);
+        let idemp_key = format!("ai:idempot:{}:{}", room_id, hasher.finish());
+        {
+            let mut conn = self.cache.conn().await.map_err(|e| {
+                RoomError::Internal(format!("cache conn: {}", e))
+            })?;
+            let exists = redis::cmd("SET")
+                .arg(&idemp_key)
+                .arg("1")
+                .arg("NX")
+                .arg("EX")
+                .arg(60u64)
+                .query_async::<Option<String>>(&mut conn)
+                .await
+                .map_err(|e| RoomError::Internal(format!("idemp SET: {}", e)))?
+                .is_some();
+            if !exists {
+                tracing::debug!(room_id = %room_id, "AI idempotency hit, skipping");
+                return Ok(());
+            }
+        }
+
+        let lock_guard =
+            match crate::room_ai_queue::acquire_room_ai_lock(&self.cache, room_id).await? {
+                Some(g) => g,
+                None => return Ok(()),
+            };
+
+        let room = crate::service::find_room_or_404(&self.db, room_id).await?;
+
+        let project = models::projects::project::Entity::find_by_id(room.project)
+            .one(&self.db)
+            .await?
+            .ok_or_else(|| RoomError::NotFound("Project not found".to_string()))?;
+
+        let model = models::agents::model::Entity::find_by_id(model_id)
+            .one(&self.db)
+            .await?
+            .ok_or_else(|| RoomError::NotFound("AI model not found".to_string()))?;
+
+        let sender = models::users::User::find_by_id(sender_id)
+            .one(&self.db)
+            .await?
+            .ok_or_else(|| RoomError::NotFound("Sender not found".to_string()))?;
+
+        let history_messages = history::get_room_history(&self.db, room_id, 50).await?;
+
+        let user_ids: Vec<Uuid> = history_messages
+            .iter()
+            .filter_map(|m| m.sender_id)
+            .chain(std::iter::once(sender_id))
+            .collect();
+
+        let user_names = history::get_user_names(&self.db, &user_ids).await;
+
+        let mentions =
+            history::extract_mention_context(&self.db, room.project, content).await;
+
+        let request = AiRequest {
+            db: self.db.clone(),
+            cache: self.cache.clone(),
+            config: self.config.clone(),
+            model,
+            project: project.clone(),
+            sender,
+            room: room.clone(),
+            input: content.to_string(),
+            mention: mentions,
+            history: history_messages,
+            user_names,
+            temperature: ai_config.temperature.unwrap_or(0.7),
+            max_tokens: ai_config.max_tokens.unwrap_or(4096) as i32,
+            top_p: 1.0,
+            frequency_penalty: 0.0,
+            presence_penalty: 0.0,
+            think: ai_config.think,
+            tools: Some(chat_service.tools()),
+            max_tool_depth: 1000,
+        };
+
+        let use_streaming = ai_config.stream;
+
+        match ai_config.agent_type.as_deref() {
+            Some("cot") => {
+                if use_streaming {
+                    ai_mode_dispatch::dispatch_cot(
+                        chat_service, request, room_id, room.project, model_id,
+                        lock_guard, self.db.clone(), self.cache.clone(),
+                        self.queue.clone(), self.room_manager.clone(),
+                    ).await;
+                    } else {
+                        if let Ok((content, _in, _out)) = chat_service.process_cot(&request, |_step| async {}).await {
+                        let _ = create_and_publish_ai_message(
+                            &self.db, &self.cache, &self.queue, &self.room_manager,
+                            room_id, room.project, uuid::Uuid::new_v4(), content, model_id,
+                            Some(request.model.name.clone()),
+                        ).await;
+                    }
+                }
+            }
+            Some("rewoo") => {
+                if use_streaming {
+                    ai_mode_dispatch::dispatch_rewoo(
+                        chat_service, request, room_id, room.project, model_id,
+                        lock_guard, self.db.clone(), self.cache.clone(),
+                        self.queue.clone(), self.room_manager.clone(),
+                    ).await;
+                }
+            }
+            Some("reflexion") => {
+                if use_streaming {
+                    ai_mode_dispatch::dispatch_reflexion(
+                        chat_service, request, room_id, room.project, model_id,
+                        lock_guard, self.db.clone(), self.cache.clone(),
+                        self.queue.clone(), self.room_manager.clone(),
+                    ).await;
+                }
+            }
+            Some("react") | _ => {
+                if ai_config.agent_type.as_deref() == Some("react") {
+                    if use_streaming {
+                        ai_react_streaming::process_message_ai_react_streaming(
+                            chat_service, request, room_id, room.project, model_id,
+                            lock_guard, self.db.clone(), self.cache.clone(),
+                            self.queue.clone(), self.room_manager.clone(),
+                        ).await;
+                    } else {
+                        ai_react_nonstreaming::process_message_ai_react_nonstreaming(
+                            chat_service, request, room_id, room.project, model_id,
+                            lock_guard, self.db.clone(), self.cache.clone(),
+                            self.queue.clone(), self.room_manager.clone(),
+                        ).await;
+                    }
+                } else if use_streaming {
+                    ai_streaming::process_message_ai_streaming(
+                        chat_service, request, room_id, room.project, model_id,
+                        lock_guard, self.db.clone(), self.cache.clone(),
+                        self.queue.clone(), self.room_manager.clone(),
+                    ).await;
+                } else {
+                    ai_nonstreaming::process_message_ai_nonstreaming(
+                        chat_service, request, room_id, room.project, model_id,
+                        lock_guard, self.db.clone(), self.cache.clone(),
+                        self.queue.clone(), self.room_manager.clone(),
+                    ).await;
+                }
+            }
+        }
+
+        Ok(())
+    }
+}