Repository navigation
Conversation
- Allow empty input submission when the last user message can be retried. - Update tests to cover new input bar behavior. - Refactor session transcript handling to prevent duplicate messages during replay. - Introduce trailing user message retry logic to ensure proper message handling. - Improve loading animations and styles for session indicators. - Add new utility scripts for session loading checks. - Update global styles for better visual consistency.
折叠态只取最新 5 条,归档或删除一条后下一条旧会话被拉进来;reconcileSidebarSessions 把所有「上一轮没见过」的会话一律置顶,于是 3 天前的会话在归档别的会话时突然窜到顶部(LIUXIN557 反馈)。现在只有比同工作区已显示会话都新的才置顶,更旧的按时间插回;没有时间戳的保持原行为。
cmd 在解析整行时就展开 %VAR%,同一行里刚 set 的变量还取不到,未定义的 %VAR% 原样留成文字 —— 用户项目里因此出现名为 %T% 的目录(yubo2 反馈)。在 cmd 指引里点明,要求先单独 set 再使用。
App 从 App Translocation 临时位置运行时,自更新要等 38 MB 下载、校验完才在 preflight 里被拒(wuzq 反馈),每次点更新都白下一遍。现在拿到安装的 bundle 后立即用同一个 macos_app_install_path_is_safe 判定,不满足就提示把 ACECode.app 移到 /Applications 后重开再更新。注:该分支在 __APPLE__ 下,本次只在 Windows 构建,未在 macOS 上编译验证。
数据目录迁移铁律、粘贴的文本块(内联/文件两种形态)、文本形式工具调用恢复与纠正、PowerShell 编码前置脚本,以及额度识别、上下文超限识别、压缩占位、搜索索引签名、侧栏补位等小项的改动约束。
用户实测后否掉「每次调用必填 preamble 参数」:强迫模型每次调用都写一句,体验差。
改成用户给的提示词 —— 多步工具任务在第一次调用前与阶段 / 计划变化时用
<text_preamble type="read|write">…</text_preamble> 写恰好一句,最终回答不打标签。
daemon 在正文流里识别标签,闭合的那一刻这句话就进 loading 提示,直到下一段正文 /
下一条前言出现或回合结束;落定后的记录不显示它(用户决定),落盘正文保留标签原文
(模型会模仿自己的历史),type 只解析透传为 kind,读 / 写效果留空待做。
- 纯逻辑:tool_preamble::TextPreambleScanner 流式识别(任意字节切分 / 缺 type /
缺闭合到行尾 / 自闭合跳过 / 大小写 / 1200 字节封顶 / 孤立闭合丢弃 / 标签后空白
吞并),strip_text_preamble_tags 与前端 stripTextPreambleTags 同规则。
- AgentLoop:阶段前言状态 phase_preamble_(tool_preamble_mu_),标签 / 加粗标题
替换,prompt 模式下未加标签的可见正文清除,回合末清空;建立时发
agent_progress{phase:"preamble"}(force 绕节流),有前言期间每帧带
preamble{title,source,kind},model_waiting / reasoning / tool_planning /
tool_running 的 label = 前言(实测批次之间的等待帧退回通用文案会让 Web 实时行
闪动);tool_start 带 preamble / preamble_source / preamble_kind;
metadata.tool_preamble = {source,title,kind} 只作记录。标签总是从 token 流 /
Message 帧 / TUI 行 / 回放里剥掉,只有开启 prompt 模式才发布成前言。
- 删除:sidecar 模式与 tool_preamble_sidecar.{hpp,cpp}、tool_preamble 事件、
inject/strip 参数、ToolCallDelta 参数前缀、TUI 伪行与 tool_call 行前言、
Web pendingToolPreambles / 历史 metadata 传播;配置归一化为 {enabled, mode:
prompt|reasoning},旧 sidecar 键忽略;REST 只剩两字段;设置页二选一。
- 系统提示:「# Progress preamble」段(用户原话),关闭时逐字节不变。
- 测试:C++ 六个套件按新语义重写(扫描器 / 跨步沿用与清除 / 最终回答误打标签 /
写工具进度头 / reasoning / 关闭 / 回放 / REST / 配置 / 系统提示),前端三个
test 重写;docs/daemon-api.md、CLAUDE.md、openspec 同步。
…ender The picker's reset effect depended on the `questions` array, which is rebuilt whenever the request object reference changes. App recreated the visible question object on every render (inline spread, not memoized), so any unrelated re-render while the question was pending reset the answers to their initial state and jumped back to the first question. - Gate the reset on requestId identity via a ref (reset only on new request) - Memoize visibleQuestionReq in App to avoid recreating the object - Add a regression test for same-request re-renders - Update the inline-permission architecture assertion to verify the permission-over-question intent instead of the old syntax
Commit 38b6af5 added workspace-membership checks to the workspaces resume route, so a missing session short-circuits with 404 before reaching SessionClient. The escaped-exception test still expected 500 for that path and failed CI. Assert 404 for the workspace route; the plain /api/sessions/:id/resume route keeps covering the SESSION_RESUME_FAILED 500 path.
…ck-0924 # Conflicts: # src/agent_loop.cpp # src/agent_loop.hpp # web/scripts/i18n-en-overrides.mjs
用户反馈修复(0924):数据目录迁移卡死与必然失败、超长粘贴改为粘贴块(大块落文件引用)、文本形式工具调用恢复与纠正、PowerShell 5.1 编码乱码,以及额度识别、上下文超限识别、压缩占位、长会话索引重建、侧栏补位、macOS 更新位置检查、cmd 变量提示。
grok-4.7 在用户会话 20260925-031619-1f93 里 20 步全部照着「Sharing progress updates」
一节的裸文本示例("Checking the test results.")写进度句、零标签。开启 prompt 模式时
该节的 Good 示例改成 <text_preamble type="read|write">…</text_preamble>,并明说裸文本
进度句是错误;关闭时逐字节不变。
「Sharing progress updates」是 2026-06-14(5cc3123)加入的旧节,教模型在工具调用之间写 10 词以内的裸文本进度句。标签模式开启时整节不出,前言段自带按用户规则写的示例 (第一次调用前写下一步 / 阶段变化时写已验证结果 + 下一步,read / write 各一条)与 裸文本 Bad 示例;「批量调用」「别把结论塞进中途消息」保留;关闭时逐字节不变。
用户决定:不再要求模型输出前言(标签版 grok-4.7 12 个工具步零标签);loading 只说
正在做什么、用现在进行时、不带参数,参数留在工具行;尽量不出现「正在推理」这类
笼统文案;开关从开发者模式挪到「适合日常工作」。
- 文案优先级:本模型步推理摘要的第一对加粗标题(只管本步,删掉推理首句兜底)
> 按原生工具名拼的现在进行时模板(「正在读取 2 个文件并搜索代码」)> 场景文案
(回合开头「正在分析你的请求」、批次后「正在分析文件内容 / 命令输出…」、正文
开始流出「正在撰写回复」)。
- 统一在 emit_agent_progress 替换 model_waiting / reasoning / preamble / responding /
tool_planning / tool_running 的 label 并清空 detail,权限 / 提问 / 压缩 / 重试不换;
TUI 经 on_thinking_title 去重回调;工具行与 TUI 进度头保留参数。
- 系统提示删掉标签段,「Sharing progress updates」逐字恢复,build_system_prompt 去掉
前言开关;历史里的 <text_preamble> 标签只剥不用。
- 配置只剩 tool_preamble.enabled(旧 mode / sidecar_* 忽略),REST 只剩 {enabled};
设置 > 常规 > 工作模式以 daemon 配置为准(lib/workMode.js),删掉开发者模式里的
ToolPreambleSettings。
- 测试:模板 / 场景文案 / kind、AgentLoop 六个端到端场景、配置、REST、前端
toolPreamble / workMode;删除 system_prompt_tool_preamble_test。文档与 openspec 同步。
- Added tests to ensure each workspace icon has distinct open and closed states. - Verified that icons maintain a stroke-only style without fill attributes. - Updated styles for the settings navigation to improve layout and scrolling behavior. - Adjusted padding and margins for workspace folder reordering to ensure proper alignment.
- Add SavedModelReasoningSync class to handle model reasoning updates. - Introduce group_model_reasoning_sync and merge_model_reasoning_sync functions for batch processing and merging model profiles. - Update WebServer to initialize and manage model reasoning synchronization. - Implement API endpoints for refreshing model reasoning and subscribing to updates. - Enhance frontend components to handle model profile updates and refresh logic. - Add tests for model reasoning synchronization functionality and ensure proper integration with the server.
Move the visible question memo before authentication early returns, add a production App hook-order regression, and cover retained drafts. Record PR 72 review findings and baseline-aware validation.
Keep the existing gap-2 layout and update its stale gap-0 assertion. The integrated frontend suite and build pass; record the baseline cause and final review validation without changing sidebar behavior.
…answer-reset fix(question): prevent AskUserQuestion from resetting answers on re-render
On Windows 10 the 1px window border lives in the non-client resize area, and a frameless window must keep its top non-client inset at zero, so the system never draws the top edge. Match Chromium: on Win10 (restored only) move the WebView down by floor(system DPI / 96) pixels and paint that row from the host window with the same colors DWM uses for the other edges (active/inactive gray, accent border when enabled), tracking WM_NCACTIVATE. Windows 11 is unchanged. Offset the Agent Browser child by the WebView origin, and run the original window proc before laying out the WebView on WM_SIZE so the webview-owned fallback window does not stretch it back over the border row.
Gate DTK frame effects and fractional scaling behind the default-off ACECODE_DEEPIN option while retaining GTK ownership and WebKitGTK 4.0. Initialize Xlib threading before GTK to fix the reproduced UOS text-drawing crash caused by Qt's late initialization of existing resource locks. Add native GTK folder selection, stable old-WebKit sidebar gutters, and a centered Linux startup logo/main window. Replace linux-old packaging with Deepin jobs using a pinned DTK SDK and system runtime linkage, with a gate rejecting bundled Qt/DTK libraries and plugins. Validation: generic/Deepin builds, 26 focused native tests, 21 script tests, web tests/build, strict OpenSpec validation, actionlint, GLIBC/dependency checks, and staged/extracted archive verification. The fixed ASan Desktop passed UOS X11 startup, idle, picker cancellation, 12 window transitions, and normal shutdown. Full picker selection and pointer interactions also passed in an isolated X11 session. ARM/Wayland execution remains unverified.
Release v0.9.27 with Deepin packages and shorter pinned label
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
/goal.Verification
npm run buildinweb/.