Varin 0.9.30

更新摘要

Varin 0.9.30 将 Agent 执行切换到原生 Rust 运行时，扩展多 Agent 协作与按需工具发现，并集中改进长会话、快速检索和聊天内 HTML 的使用体验。

原生运行时与上下文

• 由原生 Rust 运行时统一管理对话历史、输入队列、模型请求、工具执行和子任务，移除旧 Pi 运行时入口。完善取消、恢复和工具结果交付，修复内核故障后 Host 服务无法继续工作的问题。
• 改进首条消息的准备与持续流式输出，修复运行状态、工具回执和界面更新之间的衔接问题。
• 恢复后台上下文压缩与摘要读取能力。压缩预算跟随所选模型的上下文窗口，支持按比例配置预留空间；摘要保留近期原文，并可按需查阅历史、工具输出和任务记录。
• 新增可选的消息时间信息，默认关闭。开启后可使用设备时区或指定时区，时间随用户消息保存。

多 Agent 协作

• 实现、验证和检索角色可分别启用并配置模型。主 Agent 可自主完成任务，也可安排实现与独立验证并行推进。
• Agent 支持用途名、稳定的八位短码和改名；已有 Agent 可接收新任务继续工作，历史报告按原轮次保留。
• 增加双向代码交换：Agent 可提交选定文件改动，也可从其他 Agent 拉取改动；使用三方合并协调接收方已有修改。派发或续接任务时，可选择其他 Agent 的阶段成果作为文件基础。
• 本地只读子 Agent 共享派发来源对应的实际工作区，保留不同 worktree 的目录身份，减少整目录捕获与复制。已交付的结束任务自动收纳到历史；工作概览、Agent 列表和上下文状态表同步反映当前相关任务。

工具发现、输出与检索

• 常用工具直接提供，其他可用能力通过稳定索引和 tool_describe 按需获取说明，发现工具时保持声明列表稳定。支持直接调用已披露工具，并保留 tool_invoke 通用调用入口；不可用能力从模型目录中过滤。
• 通用模式将已启用的研究工具放入按需披露目录，科研模式直接提供这些工具。工具可用性与配置保持一致，修复说明显示可用、实际调用却失败的不一致。
• 长工具结果改用结构化预览，保留状态、错误、总数、游标和完整记录边界；完整输出继续保存，可按引用和位置续读。修复脚本长输出未按展示预算裁剪、异常中断后已打印内容丢失的问题。
• 快速检索在观察器或索引尚未准备完成时使用已有检索路径，减少准备过程对查询的阻塞。检索卡片持续显示阶段、耗时、候选数量和片段位置；多轮检索指令加强证据核对与结果整合。
• 修复文件列表分页与过滤、部分读取的版本标识、进程工作目录及显式路径处理等问题，减少重复核查和额外模型往返。

会话与交互体验

• 长历史采用自动预取与滚动位置保持，减少向上阅读时的手动加载。导航刻度和 Token 统计以完整 Agent 轮次组织，并独立于虚拟列表的当前渲染范围。
• HTML 内容融入消息正文，支持流式预览、源码查看和内容引用。修复会话身份绑定错误造成的预览、源码空白，以及创建过程展示不完整的问题。
• 恢复顶部上下文占用、右侧上下文面板和输入框权限控制；修复运行期间工作概览无法点击的问题，文件改动汇总在 Agent 轮次结束后展示。
• 全局“新建会话”默认创建无项目聊天，项目内入口保留项目范围。新会话在首次发送前保持为本地草稿。

模型与服务设置

• 重整模型服务与能力编辑，提供服务预设、设备账号登录和后台更新的服务／模型目录；支持区分展示名称与实际模型或部署 ID。
• 图像与决策协议移入对应模型设置，预设自动填入标准请求路径，输入框允许自定义。嵌入、重排和决策模型也可配置上下文窗口。
• 统一不同模型协议的工具 Schema 适配，修复无效正则导致的请求拒绝，并修正辅助模型设置中直接显示提供商内部 ID 的问题。

桌面端与终端

• 修复 Windows PowerShell 7（pwsh）发现，自动选择优先使用 pwsh，随后为 Windows PowerShell 和 CMD。设置显示实际 Shell 路径并支持刷新；macOS/Linux 保留用户 Shell 偏好，登录模式可单独选择。
• UTF-8 用于 Varin 自有传输和环境探测，普通终端保留用户与外部程序的编码设置。改善启动失败后的恢复展示，修正窗口关闭按钮的右上角覆盖范围，并修复 macOS 原生助手构建与架构校验。

升级提示： 本版使用原生 Rust 运行时，旧 Pi 运行时的内部会话历史不会自动导入。升级前请备份应用数据，并保留需要查阅的旧版历史。

---

Release highlights

Varin 0.9.30 moves Agent execution to its native Rust runtime, expands multi-agent collaboration and tool discovery, and improves long conversations, code retrieval and interactive HTML in chat.

Native runtime and context

• Use the native Rust runtime to manage conversation history, queued input, model requests, tools and child tasks, replacing the previous Pi runtime entry points. Improve cancellation, recovery and result delivery, including Host service recovery after kernel failure.
• Improve first-message preparation and continuous streaming. Repair coordination between execution state, tool receipts and UI updates.
• Restore background context compaction and summary lookups. Calculate budgets from the selected model's context window with a configurable reserve ratio; retain recent messages and let the summarizer look up history, tool output and task records when needed.
• Add optional message time information, disabled by default. Use the device timezone or a configured timezone and retain the timestamp with the user message.

Multi-agent collaboration

• Enable implementation, verification and retrieval roles independently and configure their models. The main Agent can work directly or run implementation and independent verification in parallel.
• Give Agents descriptive names and stable eight-character codes, rename them and continue work with an existing Agent. Preserve reports for their original execution rounds.
• Add code exchange in both directions: submit selected file changes or pull changes from another Agent, using three-way merging to reconcile existing edits. Start or continue a task from another Agent's saved stage.
• Let local read-only child Agents share the actual source workspace while preserving each worktree's directory identity, reducing full-directory capture and copying. Archive settled tasks automatically and keep the work overview, Agent list and context status table consistent with current relevant work.

Tool discovery, output and retrieval

• Declare common tools directly and expose other available capabilities through a stable index and on-demand tool_describe results. Keep tool declarations stable during discovery, support direct calls to disclosed tools and retain tool_invoke as a general entry point. Filter unavailable capabilities out of the model catalog.
• Make enabled research tools discoverable on demand in general mode and declare them directly in research mode. Keep availability consistent with configuration and repair cases where a tool was described as available but could not be invoked.
• Preview large results structurally, preserving status, errors, totals, cursors and complete record boundaries. Retain full output for subsequent reads by reference and position. Apply the display budget to script output and preserve printed content after interruption.
• Use available retrieval paths while watchers or indexes are still preparing. Show live stages, elapsed time, candidate counts and excerpt locations in retrieval cards. Improve the multi-step retrieval instructions for evidence checking and synthesis.
• Repair file-list pagination and filtering, version identifiers on partial reads, process working directories and explicit paths. Reduce redundant checks and extra model round trips.

Conversations and interactive content

• Prefetch earlier history automatically and preserve reading position. Organize navigation markers and token usage by complete Agent rounds, independently of the virtual list's rendered range.
• Integrate HTML into the message flow with streaming previews, source viewing and content references. Repair conversation identity binding that caused empty previews and source views, and improve display while content is being created.
• Restore the header context indicator, context sidebar and composer permission controls. Keep the work overview clickable during execution and show the file-change summary after the Agent round ends.
• Create projectless chats from the global New conversation action while keeping project-specific actions scoped to their project. Keep new conversations as local drafts until the first message is sent.

Models and service settings

• Reorganize model connections and capability editing, with service presets, device account sign-in and service/model catalogs that refresh in the background. Separate display names from actual model or deployment IDs.
• Configure image and decision protocols within each model. Presets fill in standard request paths that remain editable. Expose context-window settings for embedding, rerank and decision models as well.
• Adapt tool schemas consistently across model protocols, fix invalid regular expressions that caused request rejection, and display provider names instead of internal IDs in auxiliary model settings.

Desktop and terminals

• Repair PowerShell 7 (pwsh) discovery on Windows. Auto selection prefers pwsh, then Windows PowerShell and CMD. Show the resolved shell path and provide refresh in settings; retain user shell preferences on macOS/Linux and make login mode a separate option.
• Use UTF-8 for Varin-owned transport and environment probes while preserving encoding choices in ordinary terminals and external programs. Improve startup recovery, extend the close button to the window's top-right corner and fix macOS native helper builds and architecture verification.

Upgrade note: This release uses the native Rust runtime. Conversation history from the previous Pi runtime is not imported automatically. Back up application data before upgrading and retain any older history you need to consult.

安装包 / Downloads

平台 / Platform	架构 / Architecture	下载 / Download
Windows	x64	Installer (.exe)
Windows	ARM64	Installer (.exe)
Linux	x64	AppImage
Linux	ARM64	AppImage
macOS	Apple Silicon	DMG · ZIP

以上五个目标均已通过构建与启动验收。Varin-local-semantic-* 为可选本地语义组件，按平台和架构选择；基础安装包可独立安装。

All five targets passed the release build and startup checks. Varin-local-semantic-* archives are optional local semantic components; choose the matching platform and architecture. The base application installs independently.

完整提交记录 / Full changelog

Varin 0.9.25

更新摘要

本版改进工具并发、子任务启动和 Windows Computer Use，修复会话标题、记忆删除与工作概览刷新，完善原生 MCP 设置。

工具与子任务

• 调整只读工具的调度，减少长检索对同批其他工具的阻塞；请求取消贯穿后台准备和文件盘点。
• LSP 按工作区预热，隔离代码视图各自准备。修复 Worker 启动时目录准备依赖尚未登记的会话，导致子任务启动失败的问题。
• 文件监听按实际内容修订核对语言视图，避免延迟事件清除已经更新的诊断；语言扩展激活失败保留明确的失败状态。
• 修复首次语义索引构建失败被显示为空结果的问题。
• Windows 文件工具识别 MSYS 盘符路径和临时目录路径；批量编辑对照原文匹配并检测重叠，编辑失败只显示一次错误。
• Todo 使用明确的进度状态并简化输入；文档多页读取、隐藏文件检索及 Shell 输出标记处理得到修正。
• 修复已结束子任务的待输入状态及过期接续任务控件，减少已结束队友和重复消息对上下文的干扰。

Computer Use

• Windows 输入接入原生控制器，改进窗口激活、输入取消和焦点恢复，并修复鼠标输入参数类型错误。
• 区分取消电脑作业、人工接管与交还。取消通知 Agent 并清理作业展示；同一桌面的子任务操作由主任务协调。
• 未派发输入的动作拒绝按拒绝结果记录，取消后重新观察现场再开始新作业。
• 完善鼠标移动、点击和键盘输入反馈，以及工作概览中的电脑作业状态。
• 修复 computer.run 的异步脚本执行；观察文本支持截断和续读，截图改为按需请求。

会话、记忆与设置

• 修复配置了小模型后仍不生成会话标题的问题；侧栏新增会话右键菜单。
• 工作概览随已提交的会话状态刷新 Todo、子任务和其他工作项，重连后恢复未回答的问题卡片。
• 修复打包版记忆删除请求解析；记忆提示词使用稳定快照，新增内容先通过工具回执提供，在上下文压缩时更新记忆块。
• 精简工具结果、失败提示和代码检索说明，把模型用途集中展示在代码检索设置中。
• 原生 MCP 设置接管配置来源、服务器和工具管理，移除基础集成中的旧适配器管理入口。
• 调整输入框为凸起表面与柔和焦点光晕，移除 Agent 生成期间的流光效果；完善会话导航和 Agent/IDE 切换。
• 更新 Pi、Electron、PDF、Canvas 等依赖，并适配相关运行时契约。

---

Release highlights

Varin 0.9.25 improves concurrent tools, child-task startup, Windows Computer Use, conversation updates and native MCP settings.

Tools and child tasks

• Adjust read-tool scheduling so long retrieval calls do not unnecessarily hold up other tools in the batch. Propagate cancellation through preparation and file inventory.
• Prewarm workspace language services and prepare isolated code views separately. Fix Worker startup materialization looking up a session before its running-state binding exists.
• Compare watched disk content with the current language binding so delayed events do not clear fresh diagnostics. Preserve explicit language-extension activation failures.
• Report an initial semantic-index build failure instead of an empty result.
• Recognize MSYS drive and temporary paths in Windows file tools. Match batch edits against original content, reject overlaps and display edit failures once.
• Clarify Todo progress states and simplify plan input. Repair multi-page document input, hidden-file search and Shell output framing.
• Refresh finished-task attention and expired continuation controls; reduce repeated settled-teammate and message material.

Computer Use

• Use native Windows input control with improved activation, cancellation and focus restoration. Fix the mouse-input parameter type.
• Separate work cancellation, explicit user takeover and return of control. Notify the Agent when work is cancelled and clear its display; coordinate child-task access to each desktop.
• Record rejected input as rejected when nothing was dispatched; re-observe after cancellation before starting fresh work.
• Improve pointer, click and keyboard feedback and computer-work status in the work overview.
• Repair asynchronous computer.run scripts. Page observation text and request screenshots explicitly.

Conversations, memory and settings

• Generate conversation titles with the configured small model and add contextual menus to the sidebar.
• Refresh the work overview from committed conversation state and restore unanswered question cards after reconnect.
• Repair packaged memory-deletion requests. Keep a stable memory prompt snapshot, return new notes in tool receipts and refresh the memory block at compaction.
• Shorten tool results, failure guidance and retrieval wording; explain retrieval model use in its settings panel.
• Complete native MCP configuration-source, server and tool settings and retire foundational adapter management.
• Refine the elevated composer and diffuse focus halo, remove generation glow sweeps and improve conversation navigation and Agent/IDE transitions.
• Update Pi, Electron, PDF, Canvas and related runtime dependencies.

完整提交记录 / Full changelog

Varin 0.9.24

更新摘要

本版完善多 Agent 协作和用户提问，调整子线程、任务概览与用量展示，并减少索引的重复维护。

多 Agent 协作与提问

• 常规子 Agent 收敛为 Worker 和检索两类，保留用户自定义配置及科研能力。主 Agent 负责整体设计、
  关键实现、协调与最终整合；协作提示词按实际启用的角色和工具生成，不再在关闭角色后继续推荐它。
• 同一任务的主线、同级和后代可按范围读取彼此会话，通过消息协调接口与依赖。
  事件等待支持显式期限或不限时等待，普通主会话也可持久等待并在条件满足后接续。
• 新增 submit_code，向可写成员或主线提交选定文件或精确代码片段，使用原生三方整合保留双方的其他修改。
  待应用、已应用、冲突及失败分别呈现，重复提交和重启对账沿用原操作身份；修复其工具摘要缺失。
• 提问默认立即返回，Agent 可继续不依赖回答的工作；也可设置最长 600 秒的等待。
  未回答和超时都不会被当成选择或授权，稍后回复送回原会话或原子线程。
• 提问改为底部小卡片，默认停留 60 秒，显式等待时使用等待期限。关闭卡片保留草稿和问题；
  工作概览显示未回答项，可重新打开或明确结束问题。重复回复不重复启动执行。
• 修复首次打开会话时读取未绑定 Pi 会话，导致输入框显示 No Pi session is open 的问题。
• 修复设置部署目录边界时，私有源快照目录被误当成用户工作区，导致 Web 与容器启动中断的问题。

会话与设置

• 子线程在侧栏显示于主会话下方，提供树状连接；搜索可展开匹配的后代，不再把子会话散放到“最近”。
• 精简子任务卡片，按需收起已结束任务；点击任务打开会话面板，支持跳转到子会话和主会话。
  子任务侧栏提供搜索和状态筛选，压缩历史使用更明确的名称。
• 工作概览默认展开除“来源”外的各节，并按会话记住面板开关和用户的展开选择。
• 用量改为本会话按提供商和模型汇总的 Token 消耗，包括输入、输出、缓存读取与缓存写入。
  移除原提供商额度设置、查询轮询和托盘额度汇总。
• 调整 Agent Harness 设置顺序，将页面命名为“上下文管理”和“Computer Use”，
  文件编辑器及会话选项显示本地化标签，不再直接显示内部枚举值。
• 移除 Fleet 设置页及其插件聚合桥接、代码片段功能及资源库分类；任务和后台命令仍通过现有会话界面查看。
• 工作区历史只展示有实际记录的条目，共享存储大小单独汇总；清理显示真实回收字节和失败原因，
  隐藏后端未支持的历史删除与保留策略入口。

索引与性能

• 文件监听恢复采用增量核对，稳定目录跳过嵌入模型准备和索引检查点写入，
  进度区分索引构建与代码关系维护。
• 默认本地嵌入组件改用多语言 Bekko a8m（384 维），支持配置组件和自适应 CPU 调度。
  原生编码及分词移到独立 Worker，后台批次之间让出执行机会，长生成行减少重复分词探测。
• 同一起点的隔离 Worker 在原始内容身份一致且原生根仍有效时复用基线，
  内容变化时重新捕获，避免反复建立相同的工作起点。

工程与安全

• Bot PR 的 CI、Docker 验证和依赖报告共用后台队列，减少批量更新对主线与发布构建的资源争抢。
• 更新受影响的 Git、HTTP 压缩、代理地址及源码映射依赖，统一移动端 Capacitor 运行库。
  为尚无上游修复版的 sprintf-js 应用并验证精度处理补丁；合并和变基仍无需打开编辑器。

---

Release highlights

Varin 0.9.24 improves Agent collaboration, user questions, conversation navigation and index maintenance.

Collaboration and user input

• Consolidate ordinary subagents into Worker and retrieval profiles while retaining custom profiles and
  research capabilities. The main Agent owns overall design, important implementation and integration;
  generated guidance follows the enabled profiles and actual tools.
• Read related main, sibling and descendant conversations by range, coordinate through messages and wait
  on events with an optional deadline. Ordinary main conversations retain their waits for continuation.
• Submit selected file changes or exact snippets with submit_code. Native three-way integration keeps
  unrelated edits; application, conflict and failure receipts distinguish acceptance from completion.
  Replays and restart reconciliation keep the original operation identity. Fix its missing tool summary.
• Submit questions without waiting by default, or wait up to 600 seconds. Missing answers are never
  choices or approvals; later replies reach the original conversation or child thread.
• Replace blocking question dialogs with a bottom card. It stays for 60 seconds unless an explicit wait
  supplies the deadline. Hiding preserves drafts; work overview retains unanswered questions for reopening
  or cancellation. Duplicate replies do not start duplicate execution.
• Fix first-session initialization reporting No Pi session is open before native session binding.
• Keep private source-view storage separate from user workspace recovery so deployment directory
  boundaries do not reject Web/container startup.

Conversations and settings

• Nest child conversations beneath their parents in the sidebar, with tree connectors and descendant search.
  Simplify subtask cards, compact finished work and open conversation panels with parent/child navigation.
• Expand work-overview sections except Sources by default and remember the panel and section choices per conversation.
• Replace provider quota tracking with current-conversation Token totals by provider/model, including input,
  output, cache reads and cache writes; remove quota polling, settings and tray summaries.
• Reorder Harness settings, clarify Context Management and Computer Use, and localize editor/session option labels.
• Remove Fleet settings and its aggregation bridges, along with reusable text snippets and the Library category.
  Show shared recovery storage once, omit empty histories and report actual cleanup outcomes.

Indexing and performance

• Reconcile watcher recovery incrementally; unchanged inventories skip model preparation and checkpoints.
  Distinguish index construction from code-relation maintenance in progress.
• Default local embeddings to multilingual Bekko a8m (384 dimensions), with configurable components and adaptive
  CPU scheduling. Run native encoding/tokenization in a separate Worker and yield between background batches.
• Reduce repeated tokenization probes for generated long lines and reuse verified native work baselines when
  source identities match. Changed contents still require a new capture.

Engineering and security

• Queue Bot PR CI, Docker verification and dependency reports in one background slot to reduce runner contention.
• Update affected Git, HTTP compression, proxy-address and source-map dependencies, and align mobile Capacitor runtimes.
  Apply and verify the local sprintf-js precision repair pending an upstream fix; merge/rebase continuation remains noninteractive.

完整提交记录 / Full changelog

Varin 0.9.23

Release highlights

Varin 0.9.23 adds user-editable Agent instructions and scoped lightweight memory, separates
ordinary Agent memory from Bot memory, and improves tool execution and long-session performance.

Agent instructions and memory

• Edit the built-in system instructions in Settings, with global, project and conversation scopes.
  Density, style and autonomy presets insert editable text; the full preview includes project
  instructions, skills and memory. Inspect the last model request's system text and restore defaults.
• Add, edit, delete and move ordinary Agent notes between global, project and conversation scopes.
  Matching notes load directly into each subsequent model request. Selected chat text can be saved
  without a background model call, and concurrent settings edits report conflicts.
• Keep Bot instructions, long-term recall and automatic memory organization independent. Ordinary
  Agents no longer expose recall or run background memory organization.
• Remove Magic Prompts, prompt-template editors and command-template expansion, along with their
  obsolete configuration and tests. Git, review and plan actions, skills and native instruction files remain.
• Simplify default model instructions and tool descriptions while preserving parsed output contracts.
• Remove the four built-in configuration Skills and their startup generator; configuration guidance
  is provided by the settings tools on demand.
• Filter native Agent choices, dispatch/send parameters and generated team instructions by work focus
  and enabled state. Refresh them at run boundaries; disabled research roles cannot bypass the switch by inheriting a model.

Tool reliability and performance

• Let quick retrieval run independently of file reads and writes; cancel abandoned context preparation.
  Yield background association refreshes between files so plan updates and environment observations
  are not held behind one workspace-wide association pass.
• Decouple file-write completion from language diagnostics and index checkpoints, and preserve
  request-stage, queue and late-completion diagnostics for investigating remaining timeouts.
• Fix persisted webpage/PDF snapshot reads, storage record contracts, user-memory storage ownership,
  optional tool parameters, diagnostic line numbers and live-session MCP settings access.
• Keep research tools scoped to research work focus. Recover completed Responses reasoning text
  when needed and hide empty thinking disclosures.
• Reduce repeated document/thread catalog scans, IPC serialization and cross-process retrieval data.
  Isolate live chat updates from historical rows; reuse captured file hashes and compiled language queries.

Upgrade notes

• Ordinary Agent notes use a new independent store. Existing knowledge records are not automatically
  imported into the new memory list; Bot long-term memory remains on its existing storage path.
• System-instruction and note changes take effect at the next model request. Native instruction files
  keep their normal Pi loading behavior.
• Earlier builds generated varin-research-environment, varin-multi-agent-models,
  varin-retrieval-setup and varin-remote-experiments in the Pi skills directory.
  Delete these guides manually if present; this release stops generating them.

---

更新摘要

Varin 0.9.23 开放系统提示词编辑和分范围轻记忆，将普通 Agent 与 Bot 的记忆分开，
并改善工具执行、后台存储和长会话性能。

Agent 提示词与记忆

• 在设置中编辑官方内置系统提示词，支持全局、项目和会话范围。信息密度、语言风格、自主程度预设
  会插入可编辑文字；完整预览包含项目指令、技能和记忆，可查看上次模型请求的系统文本并恢复默认。
• 普通 Agent 记忆支持新增、编辑、删除和调整范围；每次后续模型请求直接加载适用的记忆。
  选中的聊天文字可直接保存，无需后台模型提取；并发编辑会明确报告冲突。
• Bot 保留独立指令、长期召回与自动整理。普通 Agent 不再暴露 recall，不再后台整理记忆。
• 移除魔法提示词、提示词模板编辑与命令模板展开，清理对应配置和过时测试。
  Git、审阅、计划操作、技能及原生指令文件继续可用。
• 精简默认模型指令和工具描述，保留实际需要解析的输出契约。
• 移除四份内置配置 Skill 及启动生成器；配置说明由设置工具按需提供。
• 子 Agent 派发参数和自动团队提示词按工作侧重与启用状态筛选，在运行边界刷新；
  关闭的科研角色也无法通过继承模型绕过开关。

工具可靠性与性能

• 快速检索与文件读写独立执行，取消已放弃的上下文准备；后台关联刷新在文件之间让出执行机会，
  避免计划更新和环境观察一直排在整工作区关联解析之后。
• 文件写入完成不再同步等待语言诊断与索引检查点；保留请求阶段、排队和迟到完成诊断，便于定位剩余超时。
• 修复网页/PDF 快照持久化后的读取、存储记录契约、用户记忆库归属、可选工具参数、诊断行号和实时会话 MCP 设置读取。
• 科研工具仅在科研工作侧重下提供；补取 Responses 完成事件中的思考文本，隐藏空的思考折叠项。
• 减少文档/线程目录重复扫描、IPC 重复序列化和跨进程检索数据搬运；隔离流式聊天更新与历史行，
  复用文件捕获哈希与已编译的语言查询。

升级说明

• 普通 Agent 轻记忆使用独立存储，旧知识库记录不会自动导入新记忆列表；Bot 长期记忆继续使用原存储。
• 提示词与记忆修改从下一次模型请求生效，原生指令文件保持 Pi 的加载行为。
• 旧版在 Pi 技能目录生成的 varin-research-environment、varin-multi-agent-models、
  varin-retrieval-setup、varin-remote-experiments 需手动删除；新版不再自动生成。

完整提交记录 / Full changelog

Varin 0.9.22

Release highlights

Varin 0.9.22 improves long-running sessions, tool execution, indexing and chat interaction, with expanded Bot, Agent, multilingual voice and provider controls. This release includes changes through 65dd3615.

Stability and performance

• Wait for runtime readiness on first launch, preventing premature initialization failures that required a manual retry.
• Prevent slow context and background-service requests from blocking subsequent tools, preserving cancellation and failure outcomes.
• Stabilize embedding configuration identity, reuse unchanged index content after restart, and resume incremental maintenance and interrupted builds.
• Move native index storage off the desktop main thread, coalesce file updates and reduce repeated history processing during streaming.
• Fix the white flash after the splash animation and improve chat following near the bottom.

Chat and task controls

• Refine typography and compact tool disclosures; show actions and usage once per turn and make conversation width adjustable.
• Stream file additions and deletions with their actual application status.
• Present quick retrieval in one expandable card, with live progress and source previews after completion.
• Manage queued messages individually, choose queueing or steering, and apply prepared context summaries immediately.
• Add chat context menus and memory extraction from selected text; align the composer with the chat column and simplify the work overview, directory hints and Git navigation.

Projects, retrieval, models and Agents

• Create named projects with multiple folders and index selected directories. Bot-private directories stay outside the index.
• Add, pause, resume, check and remove index directories, including their generated index resources.
• Preserve retrieval material obtained before deadlines; improve candidate preparation, source-window delivery and text coverage for unsupported languages.
• Configure chat, embedding, reranking and fast-decision capabilities independently, select custom API formats, and discover or import their models.
• Replace failed model-catalog spinners with explicit failure and retry states; improve projectless inference, embedding retries and reranker document formatting.
• Manage built-in, custom and plugin Agents in a list/detail editor, including enablement, names, models, temperature, prompts and tools.
• Remove automatic review triggers and merge gates while keeping review Agents available on demand.

Bots, desktop cooperation and voice

• Dedicated Bot mode supports naming, pinning, renaming, profile and memory management, sleep/wake, archive/restore and deletion.
• Bot deletion stops owned work and removes owned sessions and generated resources, with progress, errors and retry controls.
• Bind work and computer targets independently, with cross-environment file operations, authenticated service forwarding, desktop follow-ups and durable operation evidence.
• Connect Chromium CDP and LibreOffice UNO to the same visible Linux desktop. Linux VM and remote desktop features remain experimental.
• Fix unpacked Windows driver paths and UTF-8 output.
• Add Whisper large-v3 Turbo, Qwen3-ASR 0.6B and SenseVoice Small; default local dictation to multilingual Whisper Turbo with language selection and unified model management.

Runtime, security and maintenance

• Upgrade bundled Pi to 1.0.0, integrating native runtime references and image usage reporting.
• Include Wasmtime 48.0.4 and dependency security fixes, and repair Pi patch staging in Docker builds.
• Remove obsolete and duplicate tests, unused helpers and repeated builds while preserving meaningful regressions; fix native acceptance issues involving directory aliases, terminal prompts and timing.

---

更新摘要

Varin 0.9.22 改善长时间运行、工具执行、索引维护和聊天交互，完善 Bot、多 Agent、多语言语音和提供商配置。本次发布包含截至 65dd3615 的更新。

稳定性与性能

• 修复首次启动时服务尚未就绪便进入初始化失败页面、需要手动重试的问题。
• 防止慢速上下文和后台服务请求阻塞后续工具调用，保留取消与失败结果。
• 稳定嵌入配置的索引身份；重启后复用未变化的内容，继续增量维护及中断的构建，减少无故重建。
• 将原生索引存储移出桌面主线程，合并突发文件更新，减少流式聊天中的重复历史处理。
• 修复开屏动画结束时的白屏闪烁，并改善聊天跟随底部的行为。

聊天与任务操作

• 调整文字排版和工具折叠显示，辅助操作与用量统计按整轮呈现；会话宽度支持调节。
• 文件编辑实时展示绿色添加、红色删除及实际应用状态。
• 快速检索使用统一卡片：运行时滚动展示进度，完成后预览来源片段，可展开查看。
• 逐条管理排队消息，选择排队或补充当前任务；支持立即应用已准备的上下文摘要。
• 加入聊天右键菜单和选中文本提取记忆；输入框与聊天列对齐，简化工作概览、工作目录提示和 Git 导航。

项目、检索、模型与 Agent

• 项目可命名并包含多个文件夹；后台索引按选择的目录维护，Bot 私有目录不参与索引。
• 索引目录支持新增、暂停／恢复、检测更新和删除对应的索引资源。
• 快速检索保留截止前取得的材料，改进候选准备、来源片段交付，以及未支持语言的文本覆盖。
• 聊天、嵌入、重排、快速决策分别配置；自定义提供商可选择 API 格式，并为启用的能力拉取和导入模型。
• 修复模型目录失败后的持续转圈，提供明确的失败状态与重试入口；改善无项目推理配置、嵌入重试和重排材料格式。
• 用列表和详情编辑器管理内置、自定义及插件 Agent，配置启用状态、名称、模型、温度、提示词和工具。
• 移除自动审阅触发与合并门禁，保留按需派发的审阅 Agent。

Bot、电脑协作与语音

• 独立 Bot 模式支持命名、置顶、重命名、档案与记忆管理、休眠／唤醒、归档／恢复及删除。
• Bot 删除会停止所属工作并清理所属会话与生成资源，提供处理状态、错误和重试入口。
• 分别绑定工作环境与电脑目标，支持跨环境文件操作、认证服务转发、桌面事件跟随和持久操作证据。
• Chromium CDP 与 LibreOffice UNO 接入同一可见 Linux 桌面；Linux VM 和远端桌面仍为实验性能力。
• 修复 Windows 电脑驱动的解包路径及 UTF-8 输出。
• 新增 Whisper large-v3 Turbo、Qwen3-ASR 0.6B 和 SenseVoice Small；本地听写默认采用多语言 Whisper Turbo，支持语言选择并统一模型管理。

运行时、安全与维护

• 内置 Pi 升级至 1.0.0，接入原生运行时引用与图像用量统计。
• 纳入 Wasmtime 48.0.4 及相关依赖安全修复，修正 Docker 构建中的 Pi 补丁装配。
• 清理过时和重复测试、废弃辅助代码及重复构建流程，保留关键回归保护；修正原生验收中的目录别名、终端提示符与计时问题。

完整提交记录 / Full changelog

Varin 0.9.21

Release highlights

Varin 0.9.21 improves chat recovery and code indexing, and introduces Bot work, evidence-backed memory, and computer use in the workbench.

Chat, context, and retrieval

• Message submission recovers from stalled runtime or authentication requests. Context compaction detects stalled streaming responses separately from long non-streaming waits, can restart a stalled worker with the same frozen history, and shows retries in chat. Its timeout and retry settings are under Context & Knowledge.
• Code Retrieval offers automatic, LLM, fast decision model, reranker, and source-ranking judgment. Semantic File Index has a dedicated settings page for folders, storage, embedding model, request pacing, progress, and disk usage. Oversized embedding batches are split when a provider rejects them.
• explore can target selected folders. Cross-directory file access uses the target resource root. Desktop web requests use the system networking path for better compatibility with system proxies and fake-IP setups.

Bots, memory, and computer use

• Bot identities can own work and model/persona preferences. The workbench links Bot work to its associated desktop and registered output files.
• Active memory edits and background organization share source coverage. Recall can show the original source excerpt; corrections, forgetting, branch changes, and interrupted organization preserve that evidence relationship.
• Computer use adds a desktop view, agent and human control handoff, and versioned file references for completed work. Remote Linux desktops and managed Debian/libvirt guests have provisioning and upgrade paths. These Linux VM and remote desktop paths remain experimental until verified on real target hosts; macOS and Wayland computer control also remain limited.

Packaging and maintenance

• Cloud and VM bundles include the patched Pi runtime. Dependency security updates and the Windows ARM64 TriviumDB loader repair are included.
• The standard desktop build targets Windows x64/ARM64, Linux x64/ARM64, and macOS Apple Silicon. macOS Intel remains an optional manual target.

---

更新摘要

Varin 0.9.21 改善聊天恢复与代码索引，并在工作台加入 Bot 工作、可追溯来源的记忆和电脑操作。

聊天、上下文与检索

• 发送消息遇到运行时或认证请求卡住时可以恢复。上下文压缩分别检测流式输出停滞与非流式响应久等，可用同一份冻结历史重启 worker，并在聊天中显示重试；等待和重试设置位于“上下文与知识”。
• 代码检索可选自动、LLM、快速决策模型、重排模型或来源排序。独立的“语义文件索引”设置页可管理目录、存储、嵌入模型、请求节奏、进度和磁盘占用；嵌入批次过大被提供方拒绝时会拆分。
• explore 可以指定检索目录。跨目录文件访问按目标资源根判断；桌面网页请求走系统网络路径，改善系统代理和 fake-IP 环境中的可用性。

Bot、记忆与电脑操作

• Bot 身份可以关联工作、人格与模型偏好；工作台可从 Bot 工作回到关联桌面并查看登记的成果文件。
• 主动记忆编辑与后台整理共用来源覆盖。召回可追读原始片段；纠正、遗忘、分支变化和整理中断时保留对应的证据关系。
• 电脑操作加入桌面观看、Agent 与人工控制交接，以及带版本校验的成果文件引用。远端 Linux 桌面和托管 Debian/libvirt guest 已接入准备与升级路径；这些路径尚待真实目标主机验证，属于实验性能力。macOS 与 Wayland 的电脑控制仍有平台限制。

打包与维护

• 云端和 VM 包含修补后的 Pi 运行时；纳入依赖安全更新和 Windows ARM64 TriviumDB 加载修复。
• 默认桌面构建覆盖 Windows x64/ARM64、Linux x64/ARM64 和 macOS Apple Silicon；macOS Intel 仍为可选手动目标。

完整提交记录 / Full changelog

Varin 0.9.20

Release highlights

Varin 0.9.20 makes cross-project file work more natural and keeps chat responsive during long tasks and context compaction.

Sessions and files

• Relative paths use the session working directory. Each shell call starts there or in an explicit cwd; a cd in one call does not silently redirect the next. The separate mutable session operation directory has been removed.
• Sessions without a bound project workspace can still use shell and file capabilities from their launch directory. Cross-project reads and writes resolve against each target's resource root.
• Writes to editor drafts in another root check the draft version and report conflicts instead of overwriting newer edits. Resource views and semantic inventory are reconciled as files change.

Retrieval and task state

• Search defaults to the session directory and can explicitly cover multiple projects or external resources. grep reports zero matches and scan coverage honestly; explore identifies changed or missing source candidates instead of presenting stale content as current.
• Diagnostics use the target resource path. Todo and knowledge ownership can be recovered after a task settles or the Host restarts.

Chat and compaction

• /compact shows progress immediately, with a link to its worker transcript. Preparation runs in the background while chat and active tasks continue. A prepared summary is adopted when a later model request needs context space, and duplicate triggers reuse work already in progress.
• Ordinary messages no longer wait on a pre-send compaction check. Streamed thinking updates as it arrives instead of appearing late after the answer has begun.

Runtime and network reliability

• The runtime reconnects automatically and restores session state from snapshots, entries, and statistics. Stop requests distinguish requested, accepted, and unknown outcomes so late replies do not mislabel a newer run.
• Shell payload framing improves recovery from heredocs, quoting, and syntax errors. Accepted commands can be queried by execution identity; wait timing and remote disconnect cancellation have also been corrected.
• Web reading, web search, and research search share an outbound network policy. Proxy configuration, DNS, proxy authentication, and fake-IP failures receive clearer error types, with a read-only network diagnostic tool.

The standard desktop build covers Windows x64/ARM64, Linux x64/ARM64, and macOS Apple Silicon. macOS Intel remains a manual build option.

---

更新摘要

0.9.20 让会话更自然地处理跨项目文件，也让长任务中的发送、思考和上下文压缩更连贯。

会话与文件

• 相对路径以会话工作目录为基准；每条 Shell 命令也从该目录或显式指定的 cwd 开始。移除了需要单独切换的会话操作目录，上一条命令中的 cd 不会悄悄改变下一条命令的落点。
• 没有绑定项目工作区的会话仍能在启动目录使用 Shell 和相关文件能力。跨项目读写按目标文件所属的资源根处理，不再要求把所有文件放进同一个工作区。
• 跨根编辑器草稿的写入会核对版本；草稿已被更新时报告冲突，不会静默覆盖。相关资源视图与语义索引也会随文件变化核对。

检索与任务状态

• 检索默认聚焦会话工作目录，也能明确查询多个项目或外部资源。grep 的零命中和扫描覆盖如实报告；explore 遇到已变化或消失的候选文件会标出缺口，不把旧内容当作当前结果。
• 诊断工具统一使用目标资源的路径；待办和知识库的会话归属在任务结束或 Host 重启后仍可恢复。

聊天与上下文压缩

• /compact 发出后即可看到压缩进度，并能打开压缩子会话查看过程。压缩在后台准备；聊天和正在执行的任务可以继续，准备好的摘要会在后续模型请求确实需要空间时采用，重复触发会复用进行中的工作。
• 普通消息不再被发送前压缩检查卡住。流式思考内容会随生成及时刷新，避免正文已经输出后，思考区才补出先前的文字。

运行可靠性与联网

• 运行时断线后自动重连，并从会话快照、消息和统计补齐状态；停止请求会区分已请求、已接受和结果未知，避免把迟到的回执当作当前任务的结果。
• Shell 命令的正文与控制标记分离，改善 heredoc、引号和语法错误后的恢复。已受理命令可凭执行标识查询输出；等待预算和远程断线取消也得到修正。
• 网页读取、搜索和科研搜索共用出口策略；代理配置、DNS、代理认证与 fake-IP 等失败会给出更准确的原因，并提供只读网络诊断。

默认桌面包覆盖 Windows x64/ARM64、Linux x64/ARM64 和 macOS Apple Silicon。macOS Intel 保留为手动构建选项。

完整提交记录 / Full changelog

Varin 0.9.19

Varin 0.9.19 Update Summary

Features and Improvements

• UI/UX: Redesigned the agent work overview, floated the panel, added animations and clear actions. Smoothed timeline live-edge scrolling and extended the chat scrollbar to the window bottom.
• Installer: Sped up Windows payload extraction.

Bug Fixes

• Harness and Shell: Recovered shell respawn when the tracked cwd is unusable and hardened shell cwd recovery.
• Agent: Stopped output immediately on abort.
• UI: Refreshed work overview changes.

Chores and CI

• Refreshed cloud runtime lock version.
• Isolated native process tree test suite.
• Removed stale shell test import.

---

Varin 0.9.19 更新摘要

新特性与改进

• UI/UX：重新设计了 Agent 工作概览，将面板改为浮动，新增动画与清除功能；平滑了时间线实时边缘滚动，并将聊天滚动条延伸至窗口底部。
• 安装程序：加速了 Windows 载荷提取。

问题修复

• Harness 与 Shell：修复了当跟踪的 cwd 不可用时 shell 的重生问题，并加强了 shell cwd 的恢复机制。
• Agent：修复了中止时未立即停止输出的问题。
• UI：修复了工作概览的变更刷新问题。

维护与 CI

• 刷新了云运行时锁版本。
• 隔离了原生进程树测试套件。
• 移除了陈旧的 shell 测试导入。

查看完整变更记录

Varin 0.9.18

Update Summary (Varin 0.9.18)

Features

• PDF: Unified source-preserving document reading and visual interaction.
• Harness: Exposed explicit context compaction.

Bug Fixes

• Knowledge: Avoided catalog-wide payload scans for session context.
• Knowledge: Isolated persistence from application host.
• Knowledge: Removed trailing blank line.
• UI: Kept chat timeline pinned during streaming.
• UI: Constrained chat timeline flex height.
• UI: Hid return-to-latest control at timeline end.
• Electron: Aligned prepaint splash mark with cube top.
• CI: Stabilized production build and research UI mock.

Docs, Tests, and Chores

• Docs: Defined office and daily work continuity.
• Docs: Normalized office design document.
• Test: Awaited durable file change occurrence.
• Chore: Prepared release for Varin 0.9.18.

---

更新摘要 (Varin 0.9.18)

新特性

• PDF：统一了保留源文件的文档阅读与视觉交互体验。
• Harness：暴露了显式的上下文压缩功能。

问题修复

• 知识库：优化会话上下文，避免全目录级别的载荷扫描。
• 知识库：将持久化逻辑与应用宿主环境隔离。
• 知识库：移除了文件末尾的多余空行。
• UI：在流式传输期间保持聊天时间线固定。
• UI：约束了聊天时间线的 Flex 高度。
• UI：在时间线末尾隐藏了“返回最新”控件。
• Electron：对齐了预绘制启动画面标记与立方体顶部。
• CI：稳定了生产环境构建与研究型 UI 模拟。

文档、测试与维护

• 文档：定义了办公与日常工作连续性。
• 文档：规范化了办公设计文档。
• 测试：增加了等待持久文件更改发生的测试。
• 维护：准备了 Varin 0.9.18 版本的发布工作。

Full Changelog: v0.9.17...v0.9.18

Varin 0.9.17

Varin 0.9.17

Fixes

• Fix session Tok/s statistics: Calculate based on the actual output duration from message_start to message_end per model call. For multiple calls, use the average of call-level rates. Tool waiting times and inter-call intervals are no longer mixed into the denominator.
• Add default-off session "Next Step Selection": Can be individually enabled in model assignments and bound to a model. Generates zero or more editable candidates at once after a round completes.
• Clarify the knowledge suggestion model slot as models.knowledgeSuggestions, separated from the session Next Step Selection.
• Remove legacy automatic session review and duplicate suggestion setting sources.

---

Varin 0.9.17

修复

• 修复会话 Tok/s 统计：按每次模型调用从 message_start 到 message_end 的实际输出时长计算，多次调用取调用级速率平均值，不再把工具等待和调用间隔混入分母。
• 新增默认关闭的会话“下一步选择”：可在模型分工中单独启用并绑定模型，一轮完成后一次性生成零条或多条可编辑候选。
• 知识建议模型槽位明确为 models.knowledgeSuggestions，与会话下一步选择分开。
• 移除旧的自动会话回顾和重复的建议设置来源。

完整变更：v0.9.16...v0.9.17