Fix nameless live tool rows and prepare v1.0.13

CLAUDE.md

@@ -615,6 +615,8 @@ web/
 
 **实时状态卡住的兜底(用户 1.0.9 反馈:会话一直「正在推理」,切走再切回才看到已完成)。** 本质是页面没处理这一回合的 `busy_changed(false)` / `done`,而既有「完成回合自愈」只在**看到了** busy→false 之后才触发。三层兜底,不追究具体丢帧原因:(1) **WS 心跳**(`connection.js` + `lib/liveConnectionHealth.js`):静默 20 秒发 `ping`,ping 后 10 秒没有任何帧就摘掉旧 socket 直接重连(不等 onclose,假活连接的关闭握手要很久),页面重新可见时立即补验;`pong` 在连接层消化,**不能**派发成会话事件(否则每次心跳一次 transcript 提交)。(2) **事件游标带实例号**:`EventDispatcher::instance_id()` 构造时随机生成,`subscribe_ack` 的 payload、messages 快照与新的 `GET /api/sessions/:id/runtime` 都带 `event_seq` / `event_epoch`;daemon 重启或会话运行实体重建后 seq 从 1 重新数,前端发现实例号变了或本地 seq 比服务端还大就清零游标、广播 `session_seq_reset`,transcript 手里的 seq 确实作废时才对账 —— 否则新实例的事件全被 `isStaleSequencedEvent` 当成过期帧静默丢弃。(3) **卡住检测**(`lib/transcriptLiveWatchdog.js` 纯逻辑,`useSessionTranscript` 里只在「实时 + 在跑」时挂计时器):运行中 20 秒没收到本会话事件(页面重新可见时 4 秒)就调 runtime 探针;服务端空闲、或服务端 seq 超过页面、或实例号变化 → `resyncWithServer` 重拉尾页 + `reconcileTranscriptWithServer` + 从服务端当前 seq `replay_from_start` 重订阅,等同自动「来回切一下」;服务端仍忙且 seq 一致(长时间无输出的工具)只探不改。对账**不能**复用 `reloadHistoryTail`:它套的 `preserveLiveRuntimeOnLoad` 会在实时 seq 更大时把卡住的运行态原样保留。同一服务端状态只对账一次(`lastResyncKey`,提交成功后才记)。每次触发都经 `POST /api/diagnostics/transcript-stall`(白名单 `handlers/transcript_stall_report.cpp`)在 daemon 日志留一行 `[transcript-stall]`;`[event_dispatcher] listener failed` 现在带事件类型、seq 与异常文本。回归:前端 `liveConnectionHealth` / `connectionHeartbeat` / `transcriptLiveWatchdog` / `transcriptStallResync` 测试,C++ `transcript_stall_report_test.cpp`、`event_dispatcher_test.cpp::InstanceIdIsStableWithinAndDistinctAcrossInstances`、`web_server_smoke_test.cpp` 的 `TranscriptLiveSyncSmoke.*` 与 `TranscriptStallReportAcceptsAllowlistAndRejectsInjection`。
 
+**工具伪行按 call id 配对(fix: 运行中偶现一串只显示「Tool」、展开只有返回的工具行,回合结束自愈后恢复)。** 实时线上每次工具调用除 `tool_start` / `tool_end` 外还有两条只为显示、不落盘的 `message` 伪行:`tool_call`(`[Tool: 名] 参数`)与 `tool_result`(输出)。它们曾经不带 call id、id 只是 `sha1(role+content)`,于是:(1) Web 只能用全局的「按名认领 + 按序号配结果」去猜伪行属于哪个工具行;(2) 内容相同的两次调用(同一条 `git status`、同样的输出)撞 id,实时 reducer 把后一次原位并进前一次;(3) 切进运行中会话时 REST 历史 + 事件环回放,伪 `tool_call` 按内容 key 被历史的 tool_call 包装拦下,伪 `tool_result` 与历史的 role:tool 不同名拦不住,一串无名结果追加到末尾。现在 `tool_row_metadata`(`tool_result_presenter.cpp`)给两种伪行带 `tool_call_id` / `tool_index` / `tool_name`,`TranscriptWriter::dispatch_message` 把 call id 并进伪行 id;前端实时伪行不按消息 id 合并、回放按 call id 判重(`transcriptHasToolRow`)。兜底:当前回合仍有只有返回的结果(请求在重连缺口 / 订阅晚于快照时丢了)时,`useSessionTranscript` 等 600ms 拉一次 REST 尾页,`repairOrphanToolRequests` 用落盘的 `assistant.tool_calls` 在结果前补一条与历史同形的 tool_call 包装;每个 call id 只试一次,计时器不随 items 重置(流式 token 每帧都改 items)。回归:`web/src/lib/sessionTranscriptToolRows.test.js`、`agent_loop_tool_lifecycle_events_test.cpp::PseudoToolRowsCarryCallIdentity`。
+
 **Transcript 尾部窗口(渐进虚拟化)**:大会话切入时只渲染最近 `INITIAL_TAIL_ITEMS`(120)条投影行,更早的行折成顶部「显示更早的 N 条消息 / 显示全部」入口(`web/src/lib/transcriptWindow.js` 纯函数 + Node 测试)。窗口按**首个可见行的 user 行 id** 锚定(user 行不参与活动折叠,id 跨投影稳定;流式追加在尾部,锚不动、窗口自然增长;锚丢失 fail-open 全量)。渲染 map 走 `windowedItems`,但 `itemsRef`/scrubber 回合列表/turnFileListPlacement 仍吃全量 `renderedItems` —— scrubber 跳转与搜索深链(`searchJumpOrdinal`)目标不在 DOM 时先 `expandTranscriptWindow()` 再重试;Ctrl+F 查找是 DOM 文本搜索,`GlobalFindOverlay.openFind` 广播 `acecode:conversation-find-open` 事件让 ChatView 先全量展开。「显示更早」向上补条目需手动滚动补偿(`windowRevealScrollRef` 按 scrollHeight 差值回补,原生锚定已被 overflow-anchor:none 关掉)。窗口状态是 render-phase 调整(React 文档模式)—— 用 useEffect 会先全量渲染一帧,虚拟化就白做了;transcript 重置(条目清零)回到未初始化重定锚,因为 item id 会被重载复用。
 
 `lib/usePreference.js` 是 UI 偏好读写的统一入口(`usePreference(key, defaults, validator?)` 返回 `[value, setValue]`,setter 接受 partial 浅合并 / 函数式更新,localStorage 写失败静默吞)。当前订阅:`ace.theme`(由 `theme.jsx::ThemeProvider`)、`acecode.singleLayoutWidths.v1`(由 `App.jsx`,`{sidebar, sidePanel}` 像素宽)、`acecode.uiPrefs.v1`(由 `App.jsx`,`{view: 'single'|'grid4'|'grid9', sidePanelCollapsed: boolean}` — view 模式与 SidePanel 折叠态跨刷新持久化)。同一 key 在整个 App 内只允许一个 hook 实例订阅,避免多 useState 之间互覆盖。

CMakeLists.txt

@@ -7,7 +7,7 @@
     set(VCPKG_TARGET_TRIPLET "x64-windows-static" CACHE STRING "Vcpkg target triplet" FORCE)
 endif()
 
-project(acecode VERSION 1.0.12 LANGUAGES C CXX)
+project(acecode VERSION 1.0.13 LANGUAGES C CXX)
 
 include(${CMAKE_SOURCE_DIR}/cmake/acecode_source_guards.cmake)
 
@@ -105,7 +105,7 @@
 set(ACECODE_WEB_DIST "${CMAKE_SOURCE_DIR}/web/dist")
 set(ACECODE_WEB_EMBED_DIR "${ACECODE_WEB_DIST}")
 if(NOT EXISTS "${ACECODE_WEB_DIST}")
     message(WARNING
         "[acecode] web/dist/ not found — embedding a minimal fallback page. "
         "Run `pnpm install && pnpm build` inside web/ before re-configuring for the full UI.")
     set(ACECODE_WEB_EMBED_DIR "${CMAKE_BINARY_DIR}/generated/web-dist-fallback")

docs/daemon-api.md

@@ -5200,6 +5200,16 @@ the messages snapshot carry the current `swarm_mode` too, plus `agent_path`
 `[Error]`) and `summary` otherwise, in both the session list and the chat
 header; they must not derive a title from message bodies.
 
+Each tool call is also mirrored by two display-only `message` events that are
+never persisted: role `tool_call` (content `[Tool: <name>] <arguments>`) before
+the result is collected, and role `tool_result` (the displayed output) after it.
+Both carry `metadata.tool_call_id`, `metadata.tool_index` and
+`metadata.tool_name`, matching the `tool_start` / `tool_end` payloads of the same
+call, and their `id` is distinct per call even when two calls have identical
+content. Clients pair these rows with the structured tool by `tool_call_id`, and
+de-duplicate replayed rows against history (the persisted `assistant.tool_calls`
+and role `tool` result) by the same id rather than by content.
+
 For a successful `task_complete` call, the `tool_end` payload also includes
 `message_id`, the canonical id of the persisted tool-role result. Live and
 trajectory/replay records use the same id so clients can attach copy, fork, and

src/engine/agent/tool_exec/tool_batch_scheduler.cpp

@@ -110,7 +110,8 @@ ToolBatchOutcome ToolBatchScheduler::execute(
                 // 的调用不再显示伪行,canonical 的 [Interrupted] 由 Phase 3 落盘。
                 transcript_.dispatch_message("tool_call",
                     "[Tool: " + item.call.function_name + "] " +
-                        item.call.function_arguments, true, nlohmann::json::object(), nlohmann::json::array());
+                        item.call.function_arguments, true,
+                    tool_row_metadata(item.call, idx), nlohmann::json::array());
                 try {
                     slots[idx].outcome = futures.get(item.future_index);
                 } catch (const std::exception& e) {
@@ -124,7 +125,7 @@ ToolBatchOutcome ToolBatchScheduler::execute(
                 }
                 batch.doom_guard.record_result(item.call, slots[idx].outcome->result);
                 goal_.account_usage(session_manager_, 0, false);
-                presenter_.display(item.call, slots[idx].outcome->result);
+                presenter_.display(item.call, slots[idx].outcome->result, idx);
             }
 
             futures.join();
@@ -141,14 +142,15 @@ ToolBatchOutcome ToolBatchScheduler::execute(
         LOG_INFO("Tool call (write): " + tc.function_name + " id=" + tc.id);
 
         transcript_.dispatch_message("tool_call",
-                "[Tool: " + tc.function_name + "] " + tc.function_arguments, true, nlohmann::json::object(), nlohmann::json::array());
+                "[Tool: " + tc.function_name + "] " + tc.function_arguments, true,
+                tool_row_metadata(tc, slot.original_index), nlohmann::json::array());
 
         slot.outcome = lifecycle_.run(batch, tc, slot.original_index, true);
         batch.doom_guard.record_result(tc, slot.outcome->result);
         goal_.account_usage(session_manager_, 0, false);
         // 结果行紧跟派发。调用行在执行前已显示(权限确认弹窗需要上下文),
         // 写工具串行执行,顺序天然成对。
-        presenter_.display(tc, slot.outcome->result);
+        presenter_.display(tc, slot.outcome->result, slot.original_index);
         if (slot.outcome->result.terminate_session_after_turn) {
             LOG_INFO("Stopping remaining write tools after terminal session action");
             break;

src/engine/agent/tool_exec/tool_result_presenter.cpp

@@ -55,7 +55,16 @@ void ToolResultPresenter::materialize_attachments(ToolResult& result) {
         materialized.warnings.end());
 }
 
-void ToolResultPresenter::display(const ToolCall& tc, const ToolResult& result) {
+nlohmann::json tool_row_metadata(const ToolCall& call, std::size_t tool_index) {
+    return {
+        {"tool_call_id", call.id},
+        {"tool_index", tool_index},
+        {"tool_name", call.function_name},
+    };
+}
+
+void ToolResultPresenter::display(
+    const ToolCall& tc, const ToolResult& result, std::size_t tool_index) {
     const auto callbacks = callbacks_.snapshot();
     std::string display_output = result.output;
     std::string ask_display =
@@ -71,7 +80,8 @@ void ToolResultPresenter::display(const ToolCall& tc, const ToolResult& result)
         }
         display_output += attachment_fallback;
     }
-    transcript_.dispatch_message("tool_result", display_output, true, nlohmann::json::object(), nlohmann::json::array());
+    transcript_.dispatch_message("tool_result", display_output, true,
+        tool_row_metadata(tc, tool_index), nlohmann::json::array());
     if (callbacks.on_tool_result) {
         ChatMessage call_msg;
         call_msg.role = "tool_call";

src/engine/agent/tool_exec/tool_result_presenter.hpp

@@ -9,14 +9,19 @@ class WorkspaceBoundary;
 class ToolSessionHost;
 class TranscriptWriter;
 
+// 工具调用 / 结果伪行(tool_call / tool_result message 事件)的身份:与
+// tool_start / tool_end 同一个 call id 与序号。Web 据此把伪行配到结构化工具行上;
+// 没有它只能按全局顺序猜,内容相同的两次调用还会被当成同一条消息合并。
+nlohmann::json tool_row_metadata(const ToolCall& call, std::size_t tool_index);
+
 class ToolResultPresenter {
 public:
     ToolResultPresenter(WorkspaceBoundary& boundary, ToolSessionHost& host,
         TranscriptWriter& transcript, CallbacksSlot& callbacks, SessionManager* session)
         : boundary_(boundary), host_(host), transcript_(transcript),
           callbacks_(callbacks), session_manager_(session) {}
     void materialize_attachments(ToolResult& result);
-    void display(const ToolCall& call, const ToolResult& result);
+    void display(const ToolCall& call, const ToolResult& result, std::size_t tool_index);
 private:
     WorkspaceBoundary& boundary_;
     ToolSessionHost& host_;

src/engine/agent/transcript/transcript_writer.cpp

@@ -56,6 +56,16 @@ void TranscriptWriter::dispatch_message(const std::string& role,
     ChatMessage tmp;
     tmp.role    = role;
     tmp.content = content;
+    // 工具调用 / 结果伪行不落盘,id 只用于前端去重。带 tool_call_id 时把它并进
+    // id 依据:同样的命令或同样的输出(git status、"OK")若共用内容哈希,前端会把
+    // 后一次调用当成已有消息原位合并,工具行就少了一半、只剩无名的返回。
+    if (metadata.is_object()) {
+        const auto call_id = metadata.find("tool_call_id");
+        if (call_id != metadata.end() && call_id->is_string() &&
+            !call_id->get_ref<const std::string&>().empty()) {
+            tmp.timestamp = "tool_call_id:" + call_id->get<std::string>();
+        }
+    }
     if (content_parts.is_array() && !content_parts.empty()) {
         tmp.content_parts = content_parts;
     }

tests/agent/agent_loop_tool_lifecycle_events_test.cpp

@@ -605,6 +605,56 @@ TEST(AgentLoopToolLifecycleEvents, ReadOnlySameNameParallelToolsEmitKeyedStartAn
     fs::remove_all(cwd);
 }
 
+// 回归: 工具调用 / 结果伪行(tool_call / tool_result Message 事件)必须带上与
+// tool_start 相同的 call id,且内容相同的两次调用 id 不同。修复前伪行只有内容
+// 哈希 id、没有 call id:Web 只能按全局顺序猜它属于哪个工具行,同样的命令或
+// 输出还会被当成同一条消息合并,进行中的回合就出现一串只有返回的「Tool」行。
+TEST(AgentLoopToolLifecycleEvents, PseudoToolRowsCarryCallIdentity) {
+    auto cwd = make_temp_dir("acecode_lifecycle_pseudo_rows");
+    std::atomic<int> reads{0};
+    std::atomic<int> writes{0};
+    ToolLifecycleHarness h(cwd.string(), PermissionMode::Yolo);
+    h.tools().register_tool(make_probe_tool("ro_probe", true, &reads));
+    h.tools().register_tool(make_fake_bash_tool(&writes));
+
+    ScriptedResponse tools_turn;
+    tools_turn.tool_calls.push_back({"call-r1", "ro_probe", "{}"});
+    tools_turn.tool_calls.push_back({"call-r2", "ro_probe", "{}"});
+    tools_turn.tool_calls.push_back({"call-w1", "bash", R"({"command":"git status"})"});
+    tools_turn.tool_calls.push_back({"call-w2", "bash", R"({"command":"git status"})"});
+    h.provider().push_response(std::move(tools_turn));
+    h.provider().push_text("done");
+
+    ASSERT_TRUE(h.submit_and_wait());
+    EXPECT_EQ(reads.load(), 2);
+    EXPECT_EQ(writes.load(), 2);
+
+    const std::vector<std::string> expected_ids{"call-r1", "call-r2", "call-w1", "call-w2"};
+    for (const std::string role : {"tool_call", "tool_result"}) {
+        std::vector<std::string> call_ids;
+        std::vector<std::string> message_ids;
+        for (const auto& e : h.events_of(SessionEventKind::Message)) {
+            if (e.payload.value("role", std::string{}) != role) continue;
+            ASSERT_TRUE(e.payload.contains("metadata")) << role;
+            const auto& metadata = e.payload["metadata"];
+            const std::string call_id = metadata.value("tool_call_id", std::string{});
+            call_ids.push_back(call_id);
+            message_ids.push_back(e.payload.value("id", std::string{}));
+            const bool is_read = call_id.rfind("call-r", 0) == 0;
+            EXPECT_EQ(metadata.value("tool_name", std::string{}), is_read ? "ro_probe" : "bash");
+            EXPECT_EQ(metadata.value("tool_index", -1),
+                      static_cast<int>(std::find(expected_ids.begin(), expected_ids.end(), call_id) -
+                                       expected_ids.begin()));
+        }
+        EXPECT_EQ(call_ids, expected_ids) << role;
+        ASSERT_EQ(message_ids.size(), 4u) << role;
+        EXPECT_NE(message_ids[0], message_ids[1]) << role << " identical read rows share an id";
+        EXPECT_NE(message_ids[2], message_ids[3]) << role << " identical bash rows share an id";
+    }
+
+    fs::remove_all(cwd);
+}
+
 // 回归: 工具调用回合的 assistant 文本必须作为 Message 事件下发,给前端一个
 // 权威帧整体替换流式草稿。
 //

vcpkg.json

@@ -1,6 +1,6 @@
 {
   "name": "acecode",
-  "version-semver": "1.0.12",
+  "version-semver": "1.0.13",
   "dependencies": [
     "cpr",
     {

web/src/lib/runTests.js

@@ -139,6 +139,7 @@ import './queueCardItem.test.js';
 import './sessionSubscription.test.js';
 import './connectionStatusSubscriptions.test.js';
 import './sessionTranscript.test.js';
+import './sessionTranscriptToolRows.test.js';
 import './sessionHistoryPaging.test.js';
 import './sessionOpenDiagnostics.test.js';
 import { execFileSync as runHookTest } from 'node:child_process';