Changelog
Every stable release, newest first. Synced automatically from GitHub Releases; prereleases are not listed. Release notes are bilingual (Chinese and English), synced from each release changelog.
ChengYoung 1.34.0
Released 2026-11-14中文:ACP 集成能力扩展与可选的 MCP 工具延迟加载——通过 ACP 集成的客户端可为会话打标签并按标签过滤;新增可选的 MCP 工具延迟加载模式(默认关闭),适合挂载了大量 MCP 工具的高级场景;本地会话数据库随升级自动迁移,无需手动操作。
新增
- ACP 会话标签:通过 ACP 集成的客户端可在创建会话时附加标签(最多 16 组),并按标签精确过滤会话列表。
- MCP 工具延迟加载(可选,默认关闭):设置
GOOSE_TOOL_DEFERRAL=lookup后,系统提示只列出 MCP 工具名,模型可按需加载完整工具定义——适合挂载大量 MCP 工具的场景;GOOSE_TOOL_DEFERRAL_ALWAYS_LOAD可指定始终完整发布的工具。
English: ACP integration surface extended, plus optional MCP tool deferral — clients integrating over ACP can now tag sessions and filter by tag; an optional MCP tool deferral mode (off by default) suits setups with many MCP tools; the local session database upgrades automatically.
Added
- ACP session tags: clients integrating over ACP can attach up to 16 tags when creating a session and filter the session list by tags.
- MCP tool deferral (optional, off by default): with
GOOSE_TOOL_DEFERRAL=lookup, the system prompt lists MCP tool names and the model loads full schemas on demand viaenable_tools— useful with many MCP tools;GOOSE_TOOL_DEFERRAL_ALWAYS_LOADkeeps selected tools eager.
ChengYoung 1.34.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
- Authenticode: the installer, main exe, kernel sidecar and gateway are code-signed (Certum cloud HSM, RFC3161 timestamp via time.certum.pl); NSIS plugin DLLs unsigned by design (ADR-071 design C)
Verify the download
certutil -hashfile ChengYoung_1.34.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.33.0
Released 2026-11-13中文:上游基线同步(v1.59.0)+ 代码模式运行时换代 + 依赖安全清零——代码模式改用内置 JavaScript 运行时,启动更轻、依赖面收敛;修复取消/超时的 shell 命令残留子进程、多家提供商的请求细节(Databricks Gemini 思考等级、OpenRouter 变体、LiteLLM 视觉、429 重试等待、Bedrock 回落)与压缩/回合错误报告;MCP 运行时与 Markdown 渲染依赖升级,未登记安全通告归零。
新增
- 三个内置提供商模板:Lemonade、RouterPlex、AIHubMix(选择即可接入,无需手工配置)。
- 代码模式提供更完整的工具发现与批量调用(
list_functions/get_function_details/execute_javascript)。
- 子智能体会话一键跳转:工具调用卡片新增「查看子智能体会话」入口,可直达对应子任务的会话;子智能体并发数可用
GOOSE_MAX_CONCURRENT_SUBAGENTS环境变量配置。(上游 #12946 / #12998)
变更
- 代码模式改用内置 JavaScript 运行时:不再依赖外部 JS 运行时组件,安装包与依赖面更轻;代码块标记与工具名统一为
execute_javascript。 - 自定义系统提示词模板出错时会明确报错:此前语法错误会静默改用默认提示词,容易让人误以为自定义已生效。
- 工具调用标题简化:对话中的工具调用标题改为「工具名 + 首参」,更简洁(首参过长时截断至 80 字符)。
修复
- 取消或超时的 shell 命令会连同其子进程一起结束:此前可能残留后台进程继续读写输出。
- 多家提供商请求细节修正:Databricks(Gemini 思考等级与 Gemini 3 Pro 档位)、OpenRouter(
:free/:nitro变体解析)、LiteLLM(按模型信息判断视觉能力)、限流 429 的 Retry-After 等待、Bedrock(不支持文档块的模型回落默认推理)。 - 错误不再被吞:压缩失败、以错误结束的回合、无效的系统提示词模板现在都会明确报错;调度任务使用配方自身的设置。
- 安全阈值说明方向修正:阈值说明改为「值越低越严格」,此前方向表述有误。
安全
- MCP 运行时升级至 rmcp 3.9.0(修复 3 项 RustSec 通告,含 1 项 High)。
- Markdown 渲染依赖升级至 pulldown-cmark 0.13.6(修复 RUSTSEC-2026-0433,High)。
- 解压库升级至 zip 9(我方代码路径不再使用 zip 8;其余 zip 8 通告经不可达论证登记)。
English: Upstream baseline sync (v1.59.0), a new built-in JavaScript runtime for code mode, and a clean dependency audit — code mode now runs on a built-in JavaScript runtime (lighter startup, smaller dependency surface); fixed shell commands leaving child processes behind on cancel/timeout, several provider request details (Databricks Gemini thinking level, OpenRouter variant suffixes, LiteLLM vision, 429 Retry-After, Bedrock fallback) and silent error reporting; MCP runtime and Markdown rendering dependencies upgraded, no unregistered advisories remain.
Added
- Three built-in provider templates: Lemonade, RouterPlex, and AIHubMix (pick and connect, no manual configuration).
- Code mode offers fuller tool discovery and batching (
list_functions,get_function_details,execute_javascript).
- One-click jump to subagent sessions: tool-call cards now offer a "View subagent session" entry to open the corresponding subagent's session; the subagent concurrency limit is configurable via the
GOOSE_MAX_CONCURRENT_SUBAGENTSenvironment variable. (upstream #12946 / #12998)
Changed
- Code mode now uses a built-in JavaScript runtime: no external JS runtime component, lighter install and dependency surface; the code fence tag and tool name are unified as
execute_javascript. - Invalid custom system-prompt templates now fail loudly: a syntax error used to silently fall back to the default prompt, which made custom prompts look applied when they were not.
- Simplified tool-call titles: tool-call titles in chats now show "tool name + first argument" (truncated at 80 characters).
Fixed
- Cancelled or timed-out shell commands now terminate their child processes too: previously a background process could survive and keep writing output.
- Provider request details corrected: Databricks (Gemini thinking level and the Gemini 3 Pro tier), OpenRouter (
:free/:nitrovariant suffixes), LiteLLM (vision support from model info), 429 Retry-After honoring, Bedrock (models without document-block support fall back to default inference). - Errors are no longer swallowed: failed compaction, turns ending on an error, and invalid system-prompt templates now report an error; scheduled runs use the recipe's own settings.
- Security threshold wording corrected: the explanation now reads "lower values are stricter"; the direction was previously stated incorrectly.
Security
- MCP runtime upgraded to rmcp 3.9.0 (fixes 3 RustSec advisories, 1 High).
- Markdown rendering dependency upgraded to pulldown-cmark 0.13.6 (RUSTSEC-2026-0433, High).
- Extraction library upgraded to zip 9 (our code paths no longer use zip 8; remaining zip 8 advisories are documented as unreachable).
ChengYoung 1.33.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
- Authenticode: the installer, main exe, kernel sidecar and gateway are code-signed (Certum cloud HSM, RFC3161 timestamp via time.certum.pl); NSIS plugin DLLs unsigned by design (ADR-071 design C)
Verify the download
certutil -hashfile ChengYoung_1.33.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.32.1
Released 2026-11-101.32.1 — 2026-11-10
中文:Windows 安装包与全部程序文件现已代码签名——安装包、主程序、内核与网关均带 Authenticode 数字签名(RFC3161 时间戳),浏览器下载时的 SmartScreen「未知发布者」拦截将随签名信誉逐步消失;应用内自动更新链路不变。
安全
- 安装包与全部程序文件启用 Authenticode 代码签名(Certum 云端证书;RFC3161 时间戳服务 time.certum.pl)。
English: The Windows installer and all bundled executables are now code-signed — the installer, main program, bundled kernel and gateway all carry Authenticode signatures (RFC3161 timestamped); browser-download SmartScreen "unknown publisher" warnings will fade as signature reputation builds; the in-app auto-update flow is unchanged.
Security
- Installer and all bundled executables are now Authenticode-signed (Certum cloud certificate; timestamp service time.certum.pl).
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
- Authenticode: the installer and all bundled executables are code-signed (Certum cloud HSM certificate, RFC3161 timestamp via time.certum.pl)
Verify the download
certutil -hashfile ChengYoung_1.32.1_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.32.0
Released 2026-11-03中文:界面语言切换即时生效 + 空回复重试与提醒——切换界面语言不再需要重启(16 种语言、含跟随系统);模型偶发空回复时消息内一键重试、通知中心留档提醒。
新增
- 空回复一键重试与提醒:模型偶发返回空白内容时,消息中出现占位提示与重试按钮,一键重发原问题;通知中心同步留档提醒。
变更
- 切换界面语言立即生效:不再需要重启应用,偏好自动保存(16 种语言,含跟随系统)。
- 截屏全局热键默认改为 Ctrl+Alt+X:原默认 Ctrl+Alt+A 与微信/QQ 等常见截图工具冲突,导致截屏热键时好时坏;已自定义过键位的不受影响。
English: Instant UI language switching + empty-reply retry & reminder — switching the interface language no longer requires a restart (16 languages, including "follow system"); when the model occasionally returns an empty reply, an in-message retry button resends the original prompt and the notification center keeps a reminder.
Added
- Empty-reply retry & reminder: when the model occasionally returns an empty reply, an in-message placeholder with a Retry button resends the original prompt, and the notification center keeps an archived reminder.
Changed
- UI language switches take effect immediately: no restart needed, and the preference is saved automatically (16 languages, including "follow system").
- Screen-capture global hotkey default changed to Ctrl+Alt+X: the previous default (Ctrl+Alt+A) clashed with WeChat/QQ-style screenshot hotkeys, making the capture hotkey unreliable; custom key bindings are unaffected.
ChengYoung 1.32.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.32.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.31.0
Released 2026-11-02中文:依赖安全更新 + 企业扩展白名单 + 对话修复——修复 13 项 RustSec 安全通告(zstd 系 4 项 + TLS/加密与 Markdown 渲染依赖 9 项),并移除 jsonwebtoken 及其依赖链;企业部署可限定只允许安装清单内的扩展;修复「继续对话」把上一条模型回复当提问重发、以及分叉或切换语言后助手被启动两次。
新增
- 企业扩展安装白名单:设置
GOOSE_ALLOWLIST后只允许清单内的命令安装扩展;策略拉取失败时提示「扩展策略不可用」并阻止安装,不再静默放行。
安全
- 升级 zstd 0.14.3 / zstd-safe 8.2.0 / zstd-sys 2.1.2(修复 RUSTSEC-2026-0391/-0398/-0403/-0410,对齐上游 v1.57.0);
- 升级 rustls 0.23.46 / rustls-webpki 0.103.16 / webpki-roots 1.0.10 / webpki-root-certs 1.0.10 / ctutils 0.4.6 / pulldown-cmark 0.13.5(修复 RUSTSEC-2026-0346/-0347/-0348/-0349/-0381/-0382/-0383/-0385/-0414;其中 -0414 上游尚未修复,我方先行);
- 移除 jsonwebtoken 及其 rsa 依赖链(消除对应通告的可达路径)。
修复
- 「继续对话」不再重发上一条模型回复:此前会把上一条回答当成用户提问再发一次;
- 分叉或切换语言后助手不再被启动两次:此前会重复触发一次自动启动;
- 本地小模型不再偶发输出超长命令:此前
tree之类命令可能吐出数万行、撑爆上下文并触发压缩。
English: Dependency security updates, an enterprise extension allowlist, and chat fixes — 13 RustSec advisories fixed (4 in the zstd family and 9 in TLS/crypto/markdown-rendering dependencies), jsonwebtoken and its rsa chain removed; enterprise deployments can restrict extension installs to an allowlist; "continue" no longer re-sends the previous model reply, and the assistant no longer starts twice after a fork or a language switch.
Added
- Enterprise extension allowlist: with
GOOSE_ALLOWLISTset, only listed commands may install extensions; if the policy cannot be fetched, installation is blocked with an "extension policy unavailable" message instead of being silently allowed.
Security
- zstd 0.14.3 / zstd-safe 8.2.0 / zstd-sys 2.1.2 (RUSTSEC-2026-0391/-0398/-0403/-0410, aligned with upstream v1.57.0);
- rustls 0.23.46 / rustls-webpki 0.103.16 / webpki-roots 1.0.10 / webpki-root-certs 1.0.10 / ctutils 0.4.6 / pulldown-cmark 0.13.5 (RUSTSEC-2026-0346/-0347/-0348/-0349/-0381/-0382/-0383/-0385/-0414; -0414 is fixed here before upstream);
- jsonwebtoken and its rsa dependency chain removed (the corresponding advisory is no longer reachable).
Fixed
- "Continue" no longer re-sends the previous model reply — it used to send the last answer back as if it were a new question;
- The assistant no longer starts twice after a fork or a language switch;
- Small local models no longer emit oversized commands — output like
treecould previously produce tens of thousands of lines, blowing up the context and forcing a compaction.
ChengYoung 1.31.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.31.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.30.1
Released 2026-10-31中文:长回复完成后的滚动锚定真正生效——回复较长、屏幕只容得下结尾时,回复完成会自动回到正文开头供你从头阅读(此前该功能实际从未生效,视口始终停在消息末尾的时间戳处)。
修复
- 长回复完成自动回到正文开头:此前长回复完成后视口一直停在消息末尾,滚动锚定实际从未生效;现在长回复完成会平滑滚到正文开头,完整可见的短回复仍保持原位不打扰。
English: Scroll anchoring after long replies now actually works — when a long reply finishes, the view returns to the top of the reply so you can read it from the start (previously the feature never took effect and the viewport always stayed at the message footer).
Fixed
- Return to the top of long replies on completion: the viewport used to stay at the message footer after a long reply finished; it now smoothly scrolls to the start of the reply. Short replies that already fit on screen still stay put.
ChengYoung 1.30.1
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.30.1_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.30.0
Released 2026-10-30中文:提示文件链路、快捷入口、保存并开始新会话——提示文件弹窗现在能一眼看出哪些提示文件在生效、哪些没有;输入框和命令面板都加了「项目提示」入口,不必再进设置;改完提示文件可一键保存并开新会话,改动立即生效。
⚠ 行为变化:没有活跃会话时(新对话页),提示文件编辑器使用最近一次会话的目录(此前用启动时的目录)。
新增
- 提示文件链路:弹窗内新增只读区块,按读取顺序列出全局与项目各层的提示文件,并标出哪一份在生效;
- 快捷入口:输入框目录旁新增「项目提示」图标,命令面板新增「打开项目提示」;设置里的原入口保留;
- 保存并开始新会话:保存后直接在当前目录开新会话,改动立即生效(新会话不会带入当前对话记录)。
修复
- 「旧文件的新内容不再被读取」这句提示此前不显示:同目录同时存在
.cyhints与.goosehints时,该提示从未出现,现在会正常显示。
变更
- 提示文件帮助文档(三语)已同步更新,补充链路查看、清空提示文件后的回落行为与常见问题。
移除
- 本次发布没有移除既有功能。
安全
- 本次发布没有安全相关变更。
English: Hint file chain, quick access, and save & start a new session — the hint dialog now shows at a glance which hint files are in effect and which are not; the input box and the command palette both open the editor, so you no longer have to go through Settings; saving can start a fresh session in one step so your edits apply immediately.
⚠ Behavior change: with no active session (the new-chat screen), the hint editor uses your most recent session's directory instead of the window's start directory.
Added
- Hint file chain: a read-only section in the dialog lists global and project hint files in read order and marks which one is in effect;
- Quick access: a "Project hints" icon next to the directory switcher in the input box, plus an "Open project hints" command in the palette; the Settings entry stays;
- Save & start a new session: save and immediately start a new session in the same directory so the change applies at once (the new session does not carry the current conversation).
Fixed
- The "legacy file's new content is no longer read" notice now appears: when both
.cyhintsand.goosehintsexist in one directory, that notice was never shown before.
Changed
- The trilingual hint-file help is updated with the chain view, the fallback behaviour after clearing
.cyhints, and the troubleshooting section.
Removed
- No existing functionality was removed in this release.
Security
- No security-related changes in this release.
ChengYoung 1.30.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.30.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.29.0
Released 2026-10-29中文:火山 Coding Plan 完整支持、提示文件不再重复注入、发送消息自动回底——内置「火山方舟 Coding Plan」提供商的上下文窗口按真实容量(最高 1M)显示与计算;自定义提供商可声明默认上下文窗口;修复同一目录同时存在
.cyhints与.goosehints时提示内容被重复注入的问题;发送消息后自动回到最新位置并跟随回复。
⚠ 行为变化:① 同一目录两份提示文件并存时,此后以
.cyhints为准;旧名文件里的新增内容不再被本应用读取(编辑器会提示,并可用「重新导入」拉回);② 向上滚动阅读时发送消息,视图会自动回到最新位置并跟随回复。
新增
- 内置「火山方舟 Coding Plan」提供商:一键接入,上下文窗口按真实容量(最高 1M)显示与计算,混合大小写的模型 ID 也能正确命中;压缩提示按真实窗口触发(此前被标为 128k,约 90k 就提前压缩);
- 自定义提供商「默认上下文窗口」:表单新增该字段(留空 = 不声明),配置链完整保存与回显;
- 提示文件来源可见:编辑器会显示当前实际读取的文件;旧名文件仍生效(新文件为空)或被取代时会给出说明,被取代时可「重新导入」把旧内容覆盖写入
.cyhints。
修复
- 提示文件重复注入:同一目录同时存在
.cyhints与.goosehints时,此前两份内容会被重复注入对话;现在以.cyhints为准(为空(含仅空白)或不可读时回落到.goosehints)。升级不会创建或修改任何文件;只有在应用内编辑并保存提示文件时才会写入.cyhints(并带入原.goosehints的内容,原文件保留、其它客户端不受影响)。若上游客户端仍在更新.goosehints,其新增内容将不再被本应用读取——编辑器会提示并可「重新导入」(覆盖.cyhints)拉回; - 发送消息自动回到底部:向上滚动阅读历史时发送新消息,视图会自动回到最新位置并跟随回复(此前视口停在原位);长回复完成后的滚动锚定与阅读不被打扰的语义全面收敛。
变更
- 提示文件的读取优先级改为「品牌名优先、旧名回落或影子化」,适用于项目根、全局提示文件与工具访问到的子目录(见顶部行为变化①)。
移除
- 本次发布没有移除既有功能。
安全
- 本次发布没有安全相关变更。
English: Full Volcengine Coding Plan support, no more duplicate hint injection, and auto-scroll back on send — the built-in Volcengine Ark Coding Plan provider now shows and computes its context window at the real capacity (up to 1M); custom providers can declare a default context window; same-directory
.cyhintsand.goosehintsare no longer injected twice; sending a message returns the view to the latest position and follows the reply.
⚠ Behavior changes: ① when both hint files exist in one directory,
.cyhintsnow takes precedence — new content added to the legacy file is no longer read by this app (the editor shows a notice with a "Re-import" action); ② sending while scrolled up returns to the latest position and follows the reply.
Added
- Built-in Volcengine Ark Coding Plan provider: one-click setup, context window shown and computed at real capacity (up to 1M), mixed-case model IDs match correctly; compaction now triggers at the real window (previously pinned at 128k, compacting around 90k);
- Custom provider "default context window": new form field (blank = not declared), preserved and re-displayed end to end;
- Visible hint-file source: the editor shows which file is actually read; when the legacy file is still in effect (new file empty) or has been superseded, a notice explains it, with "Re-import" to overwrite
.cyhintswith the legacy content.
Fixed
- Duplicate hint injection: when both
.cyhintsand.goosehintsexisted in one directory, both were injected;.cyhintsnow takes precedence (falling back to.goosehintswhen it is empty or unreadable). Upgrading creates and modifies no files;.cyhintsis only written when you edit and save hints inside the app (carrying over the legacy content, keeping the original file intact for other clients). If another client keeps updating.goosehints, its new content is no longer read — the editor offers "Re-import" (overwriting.cyhints); - Auto-scroll back after sending: sending while scrolled up now returns to the latest position and follows the reply (previously the viewport stayed put); long-reply anchoring and reading-without-interruption are fully converged.
Changed
- Hint files are now read with branded-name precedence plus legacy fallback/shadowing, across the project root, global hint file, and tool-touched subdirectories (see behavior change ①).
Removed
- No existing functionality was removed in this release.
Security
- No security-related changes in this release.
ChengYoung 1.28.0
Released 2026-10-28中文:提供商密钥更可靠、更透明,回复滚动锚定更聪明——修复了在多个提供商之间切换后已保存密钥偶发消失的问题(密钥写入现在有写入校验与并发保护);「身份验证」页现在会列出全部已保存密钥(含自定义提供商);切换到缺少密钥的模型会提前预警;密钥相关的错误信息全面本地化。同时改进长回复的滚动锚定,并为快速启动器与全局快捷键补齐多处细节。
⚠ 行为变化:①把全局快捷键修改为被其它应用占用的组合时,会自动回滚到原有快捷键并在设置中标注(此前会落在「旧键失灵、新键无效」的状态);②其它应用中新配置的 API 密钥不再自动同步到本应用(首次迁移已完成)——两边使用请分别配置。
新增
- 「身份验证」页列出全部提供商密钥(含自定义提供商):此前自定义提供商(如火山方舟 Coding Plan)的 API 密钥已保存但在页面中不可见,无法判断「密钥到底配没配」;现在全部密钥如实列出,每条均可删除;
- 切换模型时的密钥预警:切换到「需要 API 密钥但尚未配置」的模型时会立即提示(并指引到设置页),不再出现「提示切换成功、发送才失败」的误导;
- 密钥缺失错误可读且可操作:模型因缺少 API 密钥无法回复时,错误信息已本地化并附「配置提供商」入口(此前为英文技术信息且无指引);
- 空回复可见提示:模型返回空内容(偶发的模型侧异常)时,会话中会显示「模型未返回内容」与重试指引,不再是空白一片;
- 全局快捷键失败自动回滚:把快捷键改为被其它应用占用的组合时,会自动恢复原有快捷键并在设置中标注「已回滚」——旧键不会失灵。
修复
- 切换提供商后密钥不再静默消失:修复了在多个提供商(尤其是自定义提供商)之间切换后,已保存的 API 密钥偶发丢失的问题——根因是密钥存储的并发写入缺少互斥与写后校验;现在写入有并发保护、写后回读校验,异常时会明确报错而不是稍后静默失效;
- 密钥删除后不再「复活」:删除某条密钥后重启应用不会再被自动迁移机制搬回来;
- 长回复的滚动锚定:回复较长、屏幕只容得下开头时,回复完成会正确锚定到正文开头(此前被误判为「已在视口内」而停在原位,需要手动滚动);
- 快速启动器:剪贴板粘贴失败时给出明确提示(此前静默无反馈)。
- 取消置顶时恢复隐藏模型可见性:取消置顶一个「已被本地隐藏」的模型时,会同步恢复其可见性,不再出现「取消置顶后模型凭空消失」的孤儿状态。
变更
- 快速启动器顶部拖动提示条的视觉高度由 12px 调整为 8px。
移除
- 本次发布没有移除既有功能。
安全
- 提供商密钥的写入增加回读校验与回滚(防止密钥静默丢失的存储层加固);密钥值不会进入日志或诊断输出(仅记录键名与数量)。
English: Provider keys are more reliable and visible, and scroll anchoring is smarter — fixed saved API keys occasionally disappearing after switching between providers (secret writes now have concurrency protection, read-back verification, and rollback); the Authentication page now lists all saved keys including custom providers; switching to a model without a configured key warns up front; key-related errors are localized with a direct path to fix them. Also improves scroll anchoring for long replies and polishes the quick launcher and global shortcuts.
⚠ Behavior changes: ① setting a global shortcut to a combination taken by another app now rolls back to your previous shortcut (instead of leaving all shortcuts dead); ② API keys configured in other applications are no longer auto-migrated (the one-time migration is complete) — configure them separately.
Added
- Authentication page lists all provider keys (including custom providers): custom provider keys (e.g. Volcengine Ark Coding Plan) were previously saved but invisible on this page, making it impossible to tell whether a key was configured; all keys are now listed and deletable;
- Missing-key warning on model switch: switching to a model whose provider has no API key now warns immediately (with a path to fix it) instead of showing "switched" and failing on send;
- Localized missing-key errors: when a reply fails due to a missing API key, the error is localized with a direct entry to configure the provider (previously an English technical message);
- Empty-reply placeholder: when the model returns an empty reply, the conversation shows a "no content returned" hint with retry guidance instead of a blank area;
- Global shortcut rollback: if a rebound shortcut fails to register (taken by another app), the previous shortcut is restored automatically and flagged in Settings.
Fixed
- Keys no longer disappear after switching providers: fixed saved API keys occasionally vanishing after switching between providers (especially custom ones) — the secret store lacked write locking and read-back verification; writes are now locked, verified by read-back, and rolled back with an explicit error instead of silently failing later;
- Deleted keys no longer come back: deleting a key no longer gets restored by the migration on the next start;
- Scroll anchoring for long replies: when a reply is taller than the viewport, completion now anchors to the top of the reply body (previously misjudged as "already visible" and left in place);
- Quick launcher: a clear toast when clipboard paste fails (previously silent).
- Unpin clears hidden state: unpinning a model that was locally hidden now restores its visibility, instead of leaving an invisible orphan.
Changed
- The quick launcher's drag hint strip visual height changed from 12px to 8px.
Removed
- Nothing was removed in this release.
Security
- Provider secret writes now perform read-back verification with rollback (storage hardening against silent key loss); secret values never appear in logs or diagnostics (key names and counts only).
ChengYoung 1.28.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.28.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.
ChengYoung 1.27.0
Released 2026-10-27中文:本地模型更可控、界面更可达——模型清单可隐藏/恢复不常用模型;置顶支持键盘操作;本地听写可选转写语言;对话字体大小四档可调;被其它应用占用的全局热键会在设置中明确提示。同时修复了一批界面文字颜色不生效的问题与快速启动器的多处交互缺陷。
⚠ 行为变化:快速启动器(600×80 小窗)输入框的右键菜单改为系统原生菜单——此前的自绘菜单在小窗口中被截断不可用,原菜单的「粘贴」功能由输入行右侧的粘贴按钮承担;另外,界面次要文字(时间戳、计数、说明行等)整体会比此前略浅(改为次要灰色——设计如此,非故障)。
新增
- 隐藏不常用的模型(可恢复):切换模型清单中,不常用的模型可通过行内按钮隐藏(不再出现在列表中);底部「已隐藏的模型」折叠条可随时一键恢复。已置顶的模型不受隐藏影响(置顶优先,置顶行不显示隐藏按钮);
- 置顶支持键盘:模型菜单中按
Ctrl+Enter(macOS 为⌘+Enter)可对当前高亮的模型切换置顶,与鼠标星标等效; - 本地听写转写语言:设置 → 聊天 → 听写(Local)新增「转写语言」选择,默认跟随界面语言(首次打开听写设置时自动初始化,可在设置中更改);
- 对话字体大小:设置 → 聊天新增四档(13/14/15/16px)——标题与列表等比缩放、层级不变,次要信息保持 12px;修改立即生效并跨窗口同步,默认 14px 与既往版本一致;
- 全局热键状态可见:全局快捷键被其它应用占用而无法注册时,设置 → 键盘快捷键会在对应键位行内明确提示,不再静默失效(仅提示,不会自动改键——请更换键位或关闭冲突应用)。
修复
- 次要文字颜色此前不生效,现已按层级显示:时间戳、计数、说明行等次要文字此前引用了未定义的颜色变量、实际渲染与正文混同;现已对齐语义色,并在浅色/深色/暖色三套主题下重新校准(界面次要文字整体会比此前略浅——设计如此,非故障);
- 本地模型偶发命令尾多余
$:本地小模型偶尔会把命令示例的$提示符带到命令末尾导致执行失败——已在命令解析层增加剥离兜底,并在系统提示词中约束命令格式(两份提示词保持一致并有防漂移检查); - 本地模型大输出命令的约束:系统提示词新增输出量约束(聚合/限量代替全量罗列),避免
tree之类命令的大输出撑爆上下文触发自动压缩; - 快速启动器多处交互缺陷:拖动把手失效、右键菜单被窗口截断(见顶部行为变化)、缺少粘贴按钮、高缩放屏下窗口落点偏右下、从主窗最小化状态唤出后「闪现即消失」——均已修复;
- 压缩条编辑体验:「token 上限」编辑态的控件顺序与百分比编辑态统一(保存按钮恒在最后);上下文数字与上限徽章新增悬停说明,区分「模型上下文窗口」与「自动压缩上限」两个概念。
变更
- 本地模型设置中的「上下文大小」新增说明:超过模型训练上下文的取值会降低生成质量与速度,且修改仅对新会话生效(当前会话保持加载时的窗口大小)。
移除
- 本次发布没有移除既有功能。
安全
- 无(本版本不含安全修复)。
English: More control over local models and a more accessible UI — hide/restore models in the switcher, pin via keyboard, choose the dictation transcription language (defaults to your interface language), four chat font sizes, and a clear in-app hint when a global shortcut is taken by another app (hint only — it won’t rebind keys for you). Also fixes secondary text (timestamps, counters, captions) that previously rendered in the primary text colour, and several quick-launcher interaction defects.
⚠ Behavior change: the quick launcher's right-click menu now uses the native system menu (the custom menu was truncated in the small window); paste moved to a dedicated button beside the input.
Added
- Hide models locally (restorable): hide rarely used models from the switcher; restore anytime from the "hidden models" section. Pinned models are exempt (pin wins over hide — pinned rows show no hide button);
- Keyboard pinning:
Ctrl+Enter(⌘+Enteron macOS) toggles the pin on the highlighted model in the menu; - Dictation transcription language: choose the transcription language for local dictation, defaulting to your interface language (initialized on first open; changeable in Settings);
- Chat font size: four tiers (13/14/15/16px) — headings and lists scale proportionally, secondary details stay 12px; applies immediately and syncs across windows, default 14px unchanged;
- Global shortcut status: when a global shortcut cannot be registered (taken by another app), Settings → Keyboard shortcuts says so on the exact row (hint only — it won’t rebind keys for you).
Fixed
- Secondary text (timestamps, counters, captions) now uses the intended muted colour instead of falling back to the primary text colour, recalibrated across light/dark/aura themes (it is intentionally lighter than before);
- Stray trailing
$in local-model commands: parse-layer stripping plus prompt-level format constraints (both prompt copies kept in sync); - Output-volume constraint for local models: system prompt now asks for aggregation/limits instead of full dumps that could overflow the context window;
- Quick launcher: fixed the broken drag handle, the truncated right-click menu (see behavior change), a missing paste button, a wrong window position on high-DPI displays, and the launcher vanishing right after being summoned from a minimized main window;
- Context bar editing: the token-cap editor now matches the percentage editor, with Save always last; hover hints distinguishing the model context window from the auto-compact cap.
Changed
- The local model settings' "context size" now explains that values above the model's trained context reduce quality/speed, and that changes apply to new sessions only.
Removed
- Nothing was removed in this release.
Security
- None (no security fixes in this release).
ChengYoung 1.27.0
Artifacts
- latest.json - Tauri updater manifest (minisign signature enables auto-update)
- SHA256SUMS.txt - SHA-256 checksums for the assets above
Verify the download
certutil -hashfile ChengYoung_1.27.0_x64-setup.exe SHA256 (compare with SHA256SUMS.txt)
Licensing
Apache License 2.0. See LICENSE and NOTICE bundled alongside the app executable inside the installed application.