refactor(cli): route Peer Host through the assembled runtime

docs/architecture/agent-runtime-services-design.md

@@ -693,9 +693,9 @@ pub struct HarnessExecutionContext {
 通过 `assemble(ProductAssemblyInput)` 产生 `ProductRuntimeParts`，本文件不再为它定义第二套目标接口。
 
 当前 CLI 入口已使用类型化 `RuntimeServices` 构造 `ProductRuntimeParts`，并通过一个调用级上下文把
-Agent Runtime SDK、Harness、能力注册、调用级权限和 Agentic 事件广播交给 TUI、Exec、Session 与 Usage。
-SDK 已承接会话创建/列举/删除、轮次提交和取消；SDK v1 尚未覆盖的恢复视图、消息、分支、用量和
-工具确认由 `assembly/core` 的单一兼容门面转发。`doctor` 与 `health` 校验真实组装结果及必需注册完整性；
+Agent Runtime SDK、Harness、能力注册、调用级权限和 Agentic 事件广播交给 TUI、Exec、Session、Usage 与
+交互模式下的 Peer Host。SDK 已承接会话创建/列举/删除、轮次提交和取消；SDK v1 尚未覆盖的恢复视图、
+消息、分支、用量、快照和工具确认由 `assembly/core` 的单一兼容门面转发。`doctor` 与 `health` 校验真实组装结果及必需注册完整性；
 Core 的 Network、Git 和 MCP Catalog 当前仍含兼容 marker，因此该诊断不等于对这些外部服务做实时探活。
 
 该切换仍是 `product-full` 兼容组装，不是 owner 迁移。协调器、调度器、持久化、工具管线和 Agentic Event Queue
@@ -713,7 +713,7 @@ ping 路由。未接入入口的 profile、枚举分支和单元测试仍不能
 
 | 阶段 | 约束 |
 |---|---|
-| 当前 | CLI 消费真实 Runtime Parts 与 SDK，Core 兼容门面只承接 SDK v1 缺口；不扩张字段或再造描述符 |
+| 当前 | CLI 与 Peer Host 消费真实 Runtime Parts 与 SDK，Core 兼容门面只承接 SDK v1 缺口；不扩张字段或再造描述符 |
 | 迁移 | 迁移执行 owner 或继续接入 ACP、Desktop 前，必须分别证明行为等价；`assembly/core -> apps/relay-server` 反向依赖仍需消除 |
 | 完成 | 每个声称支持的 profile 都由生产入口消费组装结果，并有最小入口验证；无消费方的 profile 不对外宣称可用 |
 
@@ -998,9 +998,10 @@ Product 测试：
 - `bitfun-runtime-services` 提供类型化服务注入；工具 contracts、provider groups 与 execution 已分层。
 - `bitfun-harness` 已提供类型化工作流描述与注册能力。
 - `bitfun-core` 可继续作为 `product-full` 兼容门面，避免迁移期间一次性重写入口。
-- CLI 已以 `DeliveryProfile::Cli` 构造真实 Runtime Parts 和 SDK runtime；本地 Agent 入口、会话和用量
-  共用一个调用级上下文与广播事件源，审批策略不再写回全局配置。Peer Host 仍保留既有 Core 兼容路径，
-  其协议与生命周期切换由独立变更处理。
+- CLI 已以 `DeliveryProfile::Cli` 构造真实 Runtime Parts 和 SDK runtime；本地 Agent 入口、会话、用量和
+  Peer Host 共用一个调用级上下文与广播事件源，审批策略不再写回全局配置。Peer Host 通过 SDK 提交/取消
+  turn，通过单一 Core 兼容门面处理会话与快照缺口，不再构造独立调度器、持久化 manager 或事件队列；
+  wire schema、Relay ACK/重放和重连协议未在该切换中扩张。
 - CLI 通过 SDK 处理已覆盖的 session/turn/cancel 操作，并通过一个 Core 兼容门面处理 SDK v1 缺口；
   该门面复用现有 owner，不建立第二套状态或事件 schema。
 

docs/architecture/cli-product-line-design.md

@@ -88,9 +88,10 @@ BitFun CLI 应成为可独立安装和发布的 Agent 产品，而不是 Desktop
 - BitFun 原生插件目录的发现、内容校验、来源确认，以及 OpenCode custom tool 静态名称预览。
 - CLI 本地 Agent 入口以类型化 `RuntimeServices` 调用 `ProductAssembler`，选择 `DeliveryProfile::Cli`，
   并把 `ProductRuntimeParts`、Agent Runtime SDK、事件源和调用级审批策略保存在一个 `CliRuntimeContext` 中。
-- TUI、`exec`、会话和用量复用同一上下文。SDK 已承接会话创建/列举/删除、轮次提交和取消；
-  SDK v1 尚未覆盖的固定 ID、恢复视图、消息、分支、用量和工具确认由一个 Core 兼容门面转发给原 owner。
-- Agentic Event Queue 仍是唯一事件 owner；TUI 与 `exec` 使用独立广播订阅，不互相消费事件。
+- TUI、`exec`、会话、用量和交互模式下的 Peer Host 复用同一上下文。SDK 已承接会话创建/列举/删除、
+  轮次提交和取消；SDK v1 尚未覆盖的固定 ID、恢复视图、消息、分支、用量、快照和工具确认由一个 Core
+  兼容门面转发给原 owner。
+- Agentic Event Queue 仍是唯一事件 owner；TUI、`exec` 与 Peer Host 使用独立广播订阅，不互相消费事件。
 - 有界旧队列只承担兼容存储；达到容量时不得抑制广播。CLI 保持一个后台 drain，订阅方一旦报告 lag/closed，
   必须取消活动 turn 并显式失败，不能在状态不完整时继续报告成功。
 - 会话 ID 在进入存储路径前统一校验；运行时索引同时绑定 ID 与规范化存储路径，并以待提交 claim 计数保护
@@ -101,14 +102,24 @@ BitFun CLI 应成为可独立安装和发布的 Agent 产品，而不是 Desktop
   关闭输入捕获、关闭 raw mode 并显示光标。真实 PTY/ConPTY 故障注入仍需独立验收。
 - 初始化按入口分级：交互模式启动 Peer Host 与 MCP，`exec` 只启动 MCP；本地 session 管理和 usage 查询不启动
   Peer Host/MCP。该分级不改变 Agentic/Terminal owner，也不等同于管理命令已有独立轻量 Runtime。
-- Peer Host 的 HostInvoke、Relay、控制器身份、确认和重连协议仍走既有兼容路径，不属于本次本地 Runtime 切换。
+- Peer Host 保持既有 HostInvoke / DeviceEvent wire schema 与 Relay 路由，但执行已接入上述调用级上下文：
+  对话提交和精确取消走 SDK，会话与快照缺口走单一 Core 兼容门面。Peer Host 只跟踪由 Peer 提交的根 turn、
+  其子 turn 与待确认工具；可确认工具始终由控制器确认，即使宿主全局策略跳过确认，Agent 也会暂停等待控制器。
+  该 Peer 专属确认要求会沿精确后台结果 follow-up 保留。后台结果按 Core 内部元数据中的精确父 turn 与来源子 turn
+  继承 ownership；仅在父 turn 仍运行时注入，否则排在
+  无关 turn 之后并保留 Peer ownership。来源 turn 完成而结果仍等待会话串行化时，仅保留有界、一次性内部 tombstone；
+  会话清理或事件流中断会移除它。最后一个控制器离线
+  或分离、事件订阅 lag/closed 时取消这些 turn；事件失步同时投递既有失败 terminal event，终态在实际发送尝试前
+  不提前清理，队列关闭时改走同一直接投递路径。事件以入队时的控制器快照为上界，每个目标发送前再确认仍连接；
+  单目标投递租约将分离或离线移除与本地 Relay 入队尝试串行化。显式断开无法确认取消时，本地界面仍安全退出并
+  显示警告。不承诺本次变更范围外的 ACK、重放或重连恢复。
 - `doctor` 与 `health` 构造并校验真实 Runtime Parts，区分 assembly-ready、Core compatibility owner 和不可用扩展。
   它们证明必需能力已注册，不把 Core 的 Network/Git/MCP compatibility marker 描述为外部服务实时可用。
 - 独立 CLI 测试与打包工作流；主 CI 的三平台 workspace check 同时覆盖 `bitfun-cli` 编译。
 
 上述切换不等于运行时 owner 已迁移，也不表示 CLI-P0 全部完成。CLI crate 仍以 `bitfun-core/product-full`
-承载协调器、调度器、持久化、工具管线和部分 SDK v1 缺口；ACP stdio 仍走原入口，插件命令仍以来源管理和
-静态预览为主。兼容门面只转发，不重新计算或写入同一事实。
+承载协调器、调度器、持久化、工具管线和部分 SDK v1 缺口，但 Peer Host 不再自行构造这些 owner；ACP stdio
+仍走原入口，插件命令仍以来源管理和静态预览为主。兼容门面只转发，不重新计算或写入同一事实。
 
 目标态仍存在以下结构缺口：
 

docs/architecture/peer-device-mode.md

@@ -55,12 +55,37 @@ FS) and must not be mixed with Peer Device Mode.
 - Peer: decrypt → allow/deny → execute on the peer host:
   - Desktop: webview bridge `peer-host-invoke://request` → same Tauri handlers
     as local UI → `peer_host_invoke_complete`
-  - CLI: Core HostInvoke registry (`WorkspaceService`, FS, config, session,
-    git, `DialogScheduler`) — no webview. Desktop-only surfaces (MiniApp /
-    cron / ACP list) return empty or no-op so hydrate does not fail.
+  - CLI: the invocation-scoped CLI product runtime handles dialog submit/cancel
+    through the Agent Runtime SDK and session/snapshot gaps through one Core
+    compatibility facade — no webview and no second scheduler, persistence
+    manager, or event queue. Desktop-only surfaces (MiniApp / cron / ACP list)
+    return empty or no-op so hydrate does not fail.
 - Events: peer agentic projection (and other product events such as terminal /
   FS / MCP interaction) fan-out as `RemoteCommand::DeviceEvent` to attached
   controllers; controller re-emits the same event names locally.
+- CLI Peer Host forwards only turns submitted through Peer Host and linked
+  child turns. A background-result follow-up inherits ownership only when its
+  Core-internal metadata identifies the exact tracked parent and source child
+  turns; if an unrelated turn is running in the same session, the result queues
+  behind it without losing Peer ownership. Completed source lineage uses a
+  bounded, one-shot tombstone while delivery waits on session serialization;
+  session drain or event-stream interruption clears it. Peer Host
+  requires an attached controller before submit and binds tool confirmation to
+  the exact observed tool and turn. Confirmable Peer tools always wait for the
+  controller even when the host's global policy skips confirmation, so an Agent
+  pauses until the controller responds; exact background-result follow-ups
+  retain this Peer-only confirmation requirement. The host cancels tracked turns when the
+  last controller detaches/goes offline or the agent-event subscription
+  lags/closes; continuity loss also projects the existing dialog-turn-failed
+  terminal event. Terminal ownership remains tracked until the event reaches the
+  delivery attempt, and a closed local delivery queue uses the same direct
+  DeviceEvent path. Delivery targets are captured when an event is queued and
+  rechecked against the currently attached set before each send. A per-target
+  delivery lease serializes detach or offline removal with the local Relay
+  enqueue attempt. An explicit disconnect still restores the local controller
+  UI, but reports a warning when host cancellation was not confirmed. This
+  boundary does not change the Relay envelope or add ACK, replay, or reconnect
+  recovery.
 - Relay `POST /api/devices/:id/rpc` waits up to **120s** for the peer response;
   reverse proxies in front of the relay must use a matching (or higher) read
   timeout or they will return 504 first.

docs/architecture/product-architecture.md

@@ -236,7 +236,7 @@ flowchart LR
 | 产品形态 | 当前 P0 插件能力 | 入口行为 |
 |---|---|---|
 | Desktop / product-full | 生产入口仍直接依赖 `bitfun-core/product-full`；当前没有 managed-plugin 管理或 OpenCode 静态预览的生产 UI/调用方 | 共享代码可编译不等于 Desktop 已消费插件能力 |
-| CLI | 入口仍以 `bitfun-core/product-full` 作为执行兼容 owner；只为 BitFun 原生包提供来源审核、启用预览、精确内容确认和停用 | 本地 Agent 路径选择 `DeliveryProfile::Cli`，校验必需 Runtime Service 注册并消费 Runtime Parts/SDK；SDK 缺口由单一 Core 兼容门面转发。部分注册仍是 compatibility marker，不代表实时探活；插件 binding 明确禁用，不执行 OpenCode 插件代码 |
+| CLI | 入口仍以 `bitfun-core/product-full` 作为执行兼容 owner；只为 BitFun 原生包提供来源审核、启用预览、精确内容确认和停用 | 本地 Agent 与 Peer Host 路径选择 `DeliveryProfile::Cli`，校验必需 Runtime Service 注册并消费同一 Runtime Parts/SDK；SDK 缺口由单一 Core 兼容门面转发。Peer Host 不再构造第二套调度、持久化或事件 owner。部分注册仍是 compatibility marker，不代表实时探活；插件 binding 明确禁用，不执行 OpenCode 插件代码 |
 | ACP | 生产入口仍直接依赖 `bitfun-core/product-full` | `DeliveryProfile::Acp` 尚未进入入口组装；不得把测试中的 profile 解释为生产隔离 |
 | Server / Remote | 当前生产路由没有插件状态消费闭环；Remote 插件执行未实现 | 不在本地替远端项目发现、准备或执行插件；未接入时返回明确不支持 |
 | Web / Mobile Web | 依赖现有后端入口，不持有插件执行单元 | 对应 profile 当前为空计划或未接入生产，不能据枚举值宣称独立产品能力 |

src/apps/cli/README.md

@@ -6,7 +6,9 @@ The local Agent paths build the CLI product profile once per invocation. Interac
 session commands, and usage reports use that invocation-scoped runtime context and event source.
 Local management queries do not start Peer Host or MCP; `exec` starts MCP but not Peer Host.
 Core remains the compatibility owner for execution and persistence operations not yet covered by the
-Agent Runtime SDK; plugin execution is not enabled by this assembly path.
+Agent Runtime SDK. When interactive mode enables Peer Host, Peer dialog submission, cancellation,
+and agent-event fan-out reuse the same runtime context; Peer Host does not construct another
+scheduler, persistence manager, or event queue. Plugin execution is not enabled by this assembly path.
 
 ## Common commands
 

src/apps/cli/src/account.rs

@@ -241,6 +241,9 @@ async fn spawn_device_routing(relay_url: &str, device_name: &str) -> Result<()>
         while let Some(event) = event_rx.recv().await {
             handle_relay_event(event, &session_arc, &relay_client_arc).await;
         }
+        if is_current_routing_client(&relay_client_arc).await {
+            crate::peer_host::update_controller_presence(Vec::new()).await;
+        }
         tracing::info!("Device routing event loop exited");
     });
 
@@ -249,9 +252,19 @@ async fn spawn_device_routing(relay_url: &str, device_name: &str) -> Result<()>
 
 /// Disconnect the device-routing connection (if any).
 pub(crate) async fn stop_device_routing() {
-    if let Some(client) = device_relay_client().write().await.take() {
+    let client = { device_relay_client().write().await.take() };
+    if let Some(client) = client {
         client.disconnect().await;
     }
+    crate::peer_host::update_controller_presence(Vec::new()).await;
+}
+
+async fn is_current_routing_client(client: &Arc<RelayClient>) -> bool {
+    device_relay_client()
+        .read()
+        .await
+        .as_ref()
+        .is_some_and(|current| Arc::ptr_eq(current, client))
 }
 
 /// Log out: tear down routing, revoke the token (best-effort), clear state.
@@ -277,16 +290,25 @@ async fn handle_relay_event(
     session_arc: &Arc<RwLock<Option<AccountSession>>>,
     relay_client: &Arc<RelayClient>,
 ) {
+    if !is_current_routing_client(relay_client).await {
+        tracing::debug!("Ignoring event from a stale device routing client");
+        return;
+    }
     match event {
         RelayEvent::AuthOk { user_id, device_id } => {
             tracing::info!("Device routing auth ok: user={user_id} device={device_id}");
         }
         RelayEvent::AuthError { message } => {
             tracing::warn!("Device routing auth error: {message}");
             TOKEN_EXPIRED.store(true, Ordering::Relaxed);
+            crate::peer_host::update_controller_presence(Vec::new()).await;
         }
         RelayEvent::DevicePresence { devices } => {
             tracing::info!("Device presence updated: {} online", devices.len());
+            crate::peer_host::update_controller_presence(
+                devices.into_iter().map(|device| device.device_id).collect(),
+            )
+            .await;
         }
         RelayEvent::DeviceMessageReceived {
             source_device_id,
@@ -365,6 +387,7 @@ async fn handle_relay_event(
         }
         RelayEvent::Disconnected => {
             tracing::info!("Device routing disconnected");
+            crate::peer_host::update_controller_presence(Vec::new()).await;
         }
         RelayEvent::Reconnected => {
             tracing::info!("Device routing reconnected");

src/apps/cli/src/main.rs

@@ -498,7 +498,7 @@ async fn initialize_core_services(
     );
 
     if bootstrap_profile.starts_peer_host() {
-        if let Err(e) = peer_host::ensure_peer_host_ready(runtime.agentic_system()).await {
+        if let Err(e) = peer_host::ensure_peer_host_ready(runtime.as_ref()).await {
             tracing::warn!("Failed to initialize CLI peer host services: {e}");
         } else {
             tracing::info!("CLI peer host services initialized");

src/apps/cli/src/peer_host/bootstrap.rs

@@ -3,42 +3,20 @@
 use std::sync::Arc;
 
 use anyhow::{Context, Result};
-use bitfun_core::agentic::coordination::{self, DialogScheduler};
-use bitfun_core::agentic::system::AgenticSystem;
-use bitfun_core::infrastructure::try_get_path_manager_arc;
 use bitfun_core::service::filesystem::FileSystemServiceFactory;
 use bitfun_core::service::workspace::{self, WorkspaceService};
 
+use crate::runtime::CliRuntimeContext;
+
 use super::fanout::start_peer_event_fanout;
-use super::state::{set_peer_host_state, try_peer_host_state, PeerHostState};
+use super::state::{set_peer_host_state, try_peer_host_state, PeerHostState, PeerTurnTracker};
 
 /// Ensure Peer Host services are ready. Idempotent.
-pub(crate) async fn ensure_peer_host_ready(agentic: &AgenticSystem) -> Result<()> {
+pub(crate) async fn ensure_peer_host_ready(runtime: &CliRuntimeContext) -> Result<()> {
     if try_peer_host_state().is_some() {
         return Ok(());
     }
 
-    let path_manager = try_get_path_manager_arc().context("path manager")?;
-    let persistence = Arc::new(
-        bitfun_core::agentic::persistence::PersistenceManager::new(path_manager)
-            .context("persistence manager")?,
-    );
-
-    let scheduler = if let Some(existing) = coordination::get_global_scheduler() {
-        existing
-    } else {
-        let session_manager = agentic.coordinator.get_session_manager().clone();
-        let scheduler = DialogScheduler::new(agentic.coordinator.clone(), session_manager);
-        agentic
-            .coordinator
-            .set_scheduler_notifier(scheduler.outcome_sender());
-        agentic
-            .coordinator
-            .set_round_injection_source(scheduler.round_injection_monitor());
-        coordination::set_global_scheduler(scheduler.clone());
-        scheduler
-    };
-
     let workspace_service = if let Some(existing) = workspace::get_global_workspace_service() {
         existing
     } else {
@@ -54,20 +32,20 @@ pub(crate) async fn ensure_peer_host_ready(agentic: &AgenticSystem) -> Result<()
     let filesystem_service = Arc::new(FileSystemServiceFactory::create_default());
 
     let state = PeerHostState {
-        coordinator: agentic.coordinator.clone(),
-        scheduler,
-        event_queue: agentic.event_queue.clone(),
+        agent_runtime: runtime.agent_runtime().clone(),
+        compatibility: runtime.compatibility().clone(),
+        agent_events: runtime.agent_events().clone(),
+        turns: PeerTurnTracker::new(),
         workspace_service,
         filesystem_service,
-        persistence,
     };
 
     if set_peer_host_state(state.clone()).is_err() {
         // Another task won the race; treat as success.
         return Ok(());
     }
 
-    start_peer_event_fanout(state.event_queue.clone());
+    start_peer_event_fanout(state);
     tracing::info!("CLI peer host services ready");
     Ok(())
 }