0.1.0-rc.26

locale:zh-CN

发布说明

• 0.1.0-rc.26 基于 RC25，合入发布前全部开放 PR，改善机器人回复的连贯性、任务核验收尾与桌面界面。正式版仍为 0.0.18。

修复与优化

• 等待、流式内容与正式答复沿用同一条回复，头像、公开段落和详情入口自然衔接；发送回执不再被误当成答复完成。
• 每条输入等待自己的答复；调度让出、断线恢复、停止和重试保留原交互位置与历史，重开时不凭空显示任务正在执行，也不自动重发未知结果的操作。
• 侧栏与对话头像共享当前回复状态，回复时活动、待机时闭眼休息；进行中的状态使用轻量扫光，并尊重减少动态效果设置。
• 内部核验结论和核验进度进入对应任务详情，主 Bot 在对话中说明用户关心的结果；旧记录与证据保留，内部 passed 等核验回执不再占据主对话。
• 对已经返回结果、长时间尚未完成验收的任务，沿用既有停滞事件提醒主 Bot 补完核验；真实待批和待答继续保留，不绕过验收门。
• 目标面板按统一生命周期表达技术阻塞与人工处理状态，避免持续显示“执行中”。
• 收紧单行状态和后台任务的间距，统一 macOS 整窗毛玻璃底色，移除侧栏分隔线；深浅主题使用中性底面及独立悬浮的主面板、详情面板及设置内容区，设置导航也不再绘制分隔线。外观设置可切换毛玻璃/实色，减少透明度时保留浮层布局。
• 启动动画改为独立全窗口布局，移除误继承聊天区内边距造成的双色白边；加载失败仍显示明确说明。
• 修复已激活日志的完整备份换位置后无法恢复的问题：用完整内容证明校验旧日志，保留原有标记、事件链与证据；大文件按有界块异步核对，正常读取不重复扫描。
• 修正本地打包命令，确保调用桌面安装包脚本。同步中英发布说明、使用文档和网站更新日志。

数据与验收边界

• 保留现有任务、对话、证据与其他未提交改动；主 Bot 默认读取公开用户交互，子链过程仍按需调查。
• RC25 旧备份若尚无内容证明，需要用仍可信的只读原件认证；原件不可用时保留数据并报告阻塞，不能仅凭时间和大小自动放行。RC26 会在原数据目录准备时补充证明，后续完整备份应一并携带。
• 本次只发布，不安装、不替换或重启当前应用。新版本在用户安排升级后生效。
• 发布前对最终组合源码执行全量回归、产品逐图自审，并使用候选安装包的实际模块在隔离数据副本验证恢复；自审不等于独立审查。
• 受控认知不代表所有云模型的质量和时延；Windows/Linux 原生交互安装、完整读屏和长期日用不在本轮验收范围内。

locale:en-US

Release notes

• 0.1.0-rc.26, based on RC25, includes every outstanding pull request for this release and improves continuous Bot replies, verification follow-up and desktop presentation. Stable remains 0.0.18.

Fixes and improvements

• Keep waiting, streamed content and the final answer in one reply, preserving the avatar, public paragraphs and details entry. A saved-input receipt does not count as an answer.
• Track each input's answer separately. Scheduling yields, connection recovery, stopping and retries preserve interaction placement and history. Reopening does not invent active execution or replay operations with unknown outcomes.
• Link sidebar and conversation avatars to the current reply: animate during replies and rest with closed eyes when idle. Active status shimmer respects reduced motion.
• Keep internal verification verdicts and progress in their task details while the main Bot explains the result to the user. Preserve historical records and evidence without showing internal passed receipts in the main conversation.
• Use the existing stall-event path to prompt verification of overdue, unaccepted results. Preserve real approval and question waits and all acceptance gates.
• Show lifecycle-based follow-up and human-action states in goal headers instead of leaving blocked work marked as executing.
• Tighten single-line status and background-work spacing. Unify the macOS glass backdrop and remove the sidebar divider. Light and dark themes use neutral surfaces with independent floating main/detail panels and Settings content, with divider-free Settings navigation. Appearance settings can switch glass/solid materials, and reduced transparency preserves the floating layout.
• Give startup its own full-window layout, removing the two-tone frame inherited from chat padding while retaining explicit loading failures.
• Restore byte-identical backups of activated journals after relocation using full-content proofs, while preserving activation markers, event chains and evidence. Hash large files asynchronously in bounded chunks; normal reads do not rescan them.
• Correct the local packaging command to invoke the desktop package script. Update bilingual release notes, usage documentation and the website changelog.

Data and validation boundaries

• Preserve existing tasks, conversations, evidence and other uncommitted work. Main-Bot context still defaults to public user interactions, with child execution details loaded on demand.
• A pre-seal RC25 backup requires certification against a genuine read-only source. If that source is unavailable, recovery remains blocked and the data is retained; matching size and timestamps alone are insufficient. RC26 prepares the proof in the original data directory so later complete backups can carry it.
• This operation publishes only; it does not install, replace or restart the running application. Users choose when to upgrade.
• Validate the final combined source with full regression, screenshot self-review and recovery through actual candidate-package modules over isolated data copies. Self-review is not independent review.
• Controlled cognition does not establish every cloud model's quality or latency. Native Windows/Linux installation flows, full screen-reader coverage and long-duration use remain outside this acceptance scope.

0.1.0-rc.25

locale:zh-CN

发布说明

• 0.1.0-rc.25 基于 RC24，合入发布前全部待合并 PR，修复日志与对话恢复、统一任务生命周期和界面圆角。正式版仍为 0.0.18。

修复与优化

• 主 Bot 默认读取用户输入与正式回复，子任务的执行过程按需打开。任务摘要仍来自宿主事实，不能代替权限、工具结果或证据。
• TUI 模型上下文保留未带新消息类型字段的旧版公开交互，继续排除私有执行回合，避免升级后漏读旧用户约束。
• 对话、任务卡与回复详情分别加载。任务状态读取失败或新消息刷新失败时，已有对话继续可读；首次加载失败明确提示，并可重试。
• 协调日志使用增量事件与分段存储，避免反复写入全部累计用量和进展；用量回执保持精确去重，大载荷保留原文与校验引用。
• 大日志在后台分块恢复，派生索引缺失时可重建。恢复校验来源、持久记录与租约；未知工具结果不自动重做，损坏或缺失的已提交记录不会被当成空任务继续执行。
• 修正完成回复的详情定位、异步问题的自由输入关联，以及窗口变窄时回复详情入口越界的问题。
• 任务列表、会话和取消入口共用生命周期决策；重启遗留的运行标记不再冒充活执行，已完成计划的预算门和待回答问题继续可见。修正兼容事实转发遗漏中断的问题。
• 核验读取独立报告摘要，避免等待验收时完整会话投影与核验事实递归，阻塞 Desktop/TUI 的正式回复。
• 统一控件、消息和浮层的连续圆角，保留开关与旋钮的圆形；移除机器人列表底部多余分隔线。
• 同步中英发布说明、升级文档和网站更新日志。

数据保留与升级

• 升级保留任务身份、对话、证据与原协调日志；新格式在取得项目独占租约并完成恢复校验后启用，不清空历史或重开任务。
• 发布不会安装或重启当前应用。 安装升级需要退出旧进程，不是热更新。正在执行任务时先保持旧应用运行，等待模型调用、工具、核验与受管后台工作自然结束后再安排升级。
• 升级前备份完整数据根，并同步升级使用同一数据的 Desktop、CLI/TUI 与 Gateway。不要让旧版宿主接管已经启用新日志格式的数据。回退到不支持新格式的版本必须恢复完整升级前备份；保留的旧日志不是升级后的最新执行状态。

验收范围与限制

• 发布候选执行全量测试、源码桌面交互与逐图审查，并使用候选安装包中的实际模块在隔离副本上验证恢复和数据保留。实现者自审不等于独立审查。
• 受控认知不代表所有云模型的质量或时延。完整读屏、Windows/Linux 交互安装与长期日用不在本轮验收范围内。
• 单条历史记录支持到 64 MiB；过程详情按选定回合加载。更细过程分页和超大正文分片仍有后续工作。

locale:en-US

Release notes

• 0.1.0-rc.25, based on RC24, integrates every outstanding pull request for the release, fixes journal and conversation recovery, and unifies task lifecycle decisions and interface corners. Stable remains 0.0.18.

Fixes and improvements

• The main Bot reads public user inputs and final replies by default; child execution details load on demand. Task summaries remain host facts and do not replace permissions, tool results or evidence.
• Preserve legacy public interactions without the newer message-kind field in TUI model context while excluding private execution turns, so upgrades retain earlier user constraints.
• Load conversations, task cards and reply details independently. Existing messages remain readable after task-status or refresh failures; initial failures show an explicit retry action.
• Use incremental, segmented coordination events instead of repeatedly writing accumulated usage and progress. Usage receipts retain exact deduplication; large payloads retain their content and verified references.
• Recover large journals in bounded background reads and rebuild missing derived indexes. Recovery checks source identity, durable records and ownership. Unknown tool outcomes are not automatically repeated, and missing committed records cannot become an empty executable task.
• Correct completed-reply detail anchors, free-text answers to asynchronous questions and reply-detail overflow when narrowing a window.
• Share one lifecycle decision across task lists, sessions and cancellation. Stale running flags after restart no longer prove live execution; budget and unanswered-question gates remain visible on completed plans. Preserve interruption facts through the compatibility adapter.
• Read verification report content independently of lifecycle projection, preventing recursive acceptance reads from blocking final replies in Desktop and TUI.
• Use consistent continuous corners for controls, messages and overlays while retaining circular switches and knobs; remove the extra Bot-list footer divider.
• Update bilingual release notes, upgrade documentation and the website changelog.

Data preservation and upgrading

• Preserve task identities, conversations, evidence and the original coordination journal. Activate the new format only after acquiring exclusive project ownership and completing recovery checks; do not reset history or recreate tasks.
• Publishing does not install or restart the current app. Installation requires the old process to exit. Keep it running until active model requests, tools, verification and managed background work have finished naturally before scheduling an upgrade.
• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway that share it together. Older hosts must not take ownership after activation. Downgrading to an incompatible version requires the complete pre-upgrade backup; the retained legacy journal is not the latest executable state.

Validation scope and limitations

• The release candidate runs full regression, source desktop interactions and screenshot review, plus recovery and preservation checks using actual candidate-package modules on isolated copies. Implementer self-review is not independent review.
• Controlled cognition does not establish quality or latency for every cloud model. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this acceptance scope.
• Individual history records support up to 64 MiB. Process details load one selected turn at a time; finer pagination and oversized-body chunking remain follow-up work.

0.1.0-rc.24

locale:zh-CN

发布状态

• 0.1.0-rc.24 整合 Bot 权限继承、任务关系与图标、会话排版和核验停滞治理。正式版仍为 0.0.18。

修复与优化

• Bot 的访问模式由本地统一管理，子任务继承同一份权限策略。完全访问模式下连续读取图片不再逐张重复提问；切换模式、取消和权限回执保持可追溯。

• 相互独立的任务按同级展示；子 Agent 使用可区分的图标，重复的结果返回记录合并呈现，保留真实任务来源。

• 调整聊天正文、标题和代码的文字层级。已结束的授权收成可展开的紧凑记录，连续记录有间距；待处理授权仍直接显示操作按钮。

• 修正回复详情中过程空态的多余留白。修复问题到期时被并发查询提前清除计时器，导致已打开的明细与计数未及时刷新的问题。

• 核验只在取得新的事实时更新进度。回复入口可展示已检查事实数量；连续 10 分钟没有事实进展的核验会收口并把原因交回主 Bot，避免长时间假运行。

• 核验报告格式修复最多补一次不带工具的请求，避免反复采集证据。更正任务交接后继续执行剩余工作，保留已有证据。

• 同步用户文档、累积更新日志和网站的中英发布记录。

升级与验收范围

• 产品回归使用 200 个机器人、1 万条消息、10 万条收件记录的隔离历史数据；发布前逐图自审并验证产品验收包。
• 验收覆盖源码桌面交互、迁移恢复和发布产物。受控模型场景不代表所有真实模型的表现；实现者自审不称为独立复核。
• 升级前备份完整数据根，并同步升级 Desktop、CLI/TUI 与 Gateway。完整读屏、Windows/Linux 交互安装和长期日用不在本轮验收范围内。

locale:en-US

Release status

• 0.1.0-rc.24 integrates Bot access inheritance, task relationships and icons, conversation typography, and stalled verification settlement. Stable remains 0.0.18.

Fixes and improvements

• Bot access modes use one locally governed policy inherited by child tasks. Full access no longer asks separately for every image. Mode changes, cancellation and permission receipts remain traceable.

• Independent tasks appear as peers. Child Agents have distinguishable icons, and repeated result-return events are consolidated while retaining their actual task sources.

• Refine body, heading and code typography. Completed authorization records collapse into compact expandable receipts with spacing between records; pending permissions keep their action buttons visible.

• Remove excess whitespace from the reply-process empty state. Fix a race where a concurrent query cleared an expired-question timer before it refreshed the open attention list and count.

• Verification progress advances only when new facts arrive. Reply activity can show the number of facts checked. A verification with no factual progress for 10 minutes settles and returns its reason to the main Bot instead of appearing to run indefinitely.

• Malformed verification reports receive at most one tool-free format repair request. Corrected task handoffs execute the remaining work while preserving existing evidence.

• Synchronize user documentation, the cumulative changelog and bilingual website release records.

Upgrade and validation scope

• Isolated product regression uses history containing 200 bots, 10,000 messages and 100,000 inbox records, followed by screenshot self-review and product-packet verification.
• Acceptance covers source desktop interactions, migration recovery and release artifacts. Controlled model scenarios do not establish behavior for every real model; implementer self-review is not independent review.
• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this acceptance scope.

0.1.0-rc.23

locale:zh-CN

发布状态

• npm 已发布并验证：@peer-agent/protocol、runtime-core、runtime-sdk 和 cli 的 rc 标签均为 0.1.0-rc.23；隔离安装、运行库导入、CLI 版本与二进制哈希核对通过。
• 0.1.0-rc.23 修复待处理事项计数、重复回复回执与详情入口。正式版仍为 0.0.18。

修复与优化

• “需要你”只统计当前有效的问题、授权与明确确认。已回答、已被后续决策替代或到期的旧问题自动退出，后台任务与历史错误不再增加数字。
• 点击全局数字或机器人角标查看具体事项，可直接回答、授权、确认或定位原文；真正的子任务提问也纳入同一份明细。
• 每条用户消息只显示一个轻量“已回复”，修复关联多个任务时出现多枚“已回答”。任务来源仍保留在回复详情。
• “回复详情”位于气泡左下，对齐正文文字；完成态在悬停或键盘聚焦时显示，运行时同一入口显示实际活动，预留高度避免跳动。
• 修复从待处理事项定位后，后续通知无法定位新消息的问题。保留 RC22 用户模型选择及已有性能修复。

升级与验收范围

• 问题收尾只关闭失效操作，不改写历史正文，不批准权限，也不伪造任务完成或取消。需要完整报告的人工复核仍打开原报告。
• 产品回归覆盖 200 个机器人、1 万消息的源码桌面交互，并由实现者逐图自审；发布包、签名与更新清单另行核验。
• 升级前备份完整数据根，并同步升级 Desktop、CLI/TUI 与 Gateway。完整读屏、Windows/Linux 交互安装及长期日用不在本轮验收范围内。

locale:en-US

Release status

• npm published and verified: The rc tags for @peer-agent/protocol, runtime-core, runtime-sdk and cli all resolve to 0.1.0-rc.23. Isolated installation, runtime imports, CLI version and binary hashes passed verification.
• 0.1.0-rc.23 fixes attention counts, duplicate reply receipts and the reply details entry. Stable remains 0.0.18.

Fixes and improvements

• Attention counts include only current questions, permissions and explicit confirmations. Answered, superseded or expired questions retire automatically; background work and historical errors no longer increase the count.
• Open the global count or a bot badge to inspect and answer, authorize, confirm or locate each item. Actionable child-task questions use the same list.
• Each user message shows one lightweight replied receipt, fixing repeated answered buttons when a reply references multiple tasks. Task sources remain in reply details.
• Reply details sits below the left edge of the bubble, aligned with its text. Completed entries appear on hover or keyboard focus; active replies show factual activity through the same entry. Reserved height avoids layout shifts.
• Notifications can locate a new message after navigation from attention details. Retain RC22 user model selections and previous performance fixes.

Upgrade and validation scope

• Retiring a question disables its old actions without rewriting history, granting permissions or fabricating task completion or cancellation. Manual reviews that require a full report still open that report.
• Product regression covers source desktop interactions with 200 bots and 10,000 messages, followed by implementer screenshot self-review. Published packages, signatures and update manifests receive separate checks.
• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this acceptance scope.

0.1.0-rc.22

locale:zh-CN

发布状态

• 0.1.0-rc.22 修复机器人把复核失败错误归因于“同一个模型、同一条渠道”。正式版仍为 0.0.18。

修复与优化

• 执行与复核可以使用相同模型和渠道。独立复核使用单独的审查上下文，固定的用户选择继续生效；不会为了取得“独立性”要求更换配置。
• 验证详情补充近期复核尝试的真实错误、未满足标准和缺失证据，区分历史记录与当前版本。机器人应根据这些记录说明问题，不能把“同源”标记当作失败原因。
• 移除委托工具内部无条件偏向其他来源的提示，保留用户配置的候选排序和故障切换范围。
• 同步中英文模型配置说明、README、发布说明和站点更新日志。保留 RC21 的大历史执行性能修复。

升级与验收范围

• 历史失败记录和验收结果保留；本次更新不会把旧任务自动改成通过。缺失的证据仍需补齐并通过正常复核。
• 同模型配置保留、验证失败诊断、重启读取及复核成功后的完成和签收有定向测试。隔离真实模型诊断与标准产品回归分别验证，不把合成任务记录当作真实项目交付。
• 升级前备份完整数据根，并同步升级 Desktop、CLI/TUI 与 Gateway。产品截图由实施者自审；完整读屏、Windows/Linux 交互安装及长期日用不在本轮验收范围内。

发布后验证

• 已下载核对全部 14 个公开文件、更新清单及 macOS 签名/公证；npm 实际安装通过，正式版 latest 仍为 0.0.18。
• 公开 RC22 包使用同模型、同渠道及 max 推理强度，真实任务完成独立上下文复核与宿主策略签收。单独的真实诊断回复正确说明缺失证据与历史轮次上限，没有要求换模型。
• 隔离生成历史包含 1100 个计划和约 185 MB 协调日志。完整任务运行 307.906 秒、采集 1243 组交互；输入/滚动 P99 约 24/34 ms，最大约 203/417 ms。此数值含自动化传输，不代表所有机器无峰值。
• 标准产品包 180 张图完成实施者自审并通过 Verify，保留 7 项既有轻微限制。使用隔离数据根运行公开 macOS 包，未替换日常安装；Windows/Linux 交互安装、完整读屏和长期日用仍未覆盖。

locale:en-US

Release status

• 0.1.0-rc.22 fixes incorrect explanations that attributed verification failure to using the same model and channel. Stable remains 0.0.18.

Fixes and improvements

• Execution and verification may use the same model and channel. Independent verification uses a separate review context while preserving fixed user selections; another model is not required to establish independence.
• Verification details now include recent recorded errors, failed criteria and missing evidence, distinguishing historical attempts from the current source version. The bot should explain these findings rather than treating shared model provenance as a failure reason.
• Remove the delegation tool's unconditional preference hint for another source while retaining user-configured candidate ordering and recovery boundaries.
• Synchronize bilingual model guidance, README, release notes and the documentation changelog. Retain RC21's large-history execution performance fixes.

Upgrade and validation scope

• Preserve historical failures and acceptance results. Updating does not automatically pass older tasks; missing evidence still requires normal verification.
• Focused tests cover fixed same-model selections, failure diagnostics, restart reads and completion and acceptance after successful review. Isolated real-model diagnostics and standard product regression are checked separately; generated task records do not constitute real project delivery.
• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. Product screenshots receive implementer self-review. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this acceptance scope.

Post-release validation

• Downloaded and verified all 14 public assets, update manifests and macOS signing/notarization. An isolated npm installation passed; stable latest remains 0.0.18.
• The public RC22 app completed independent-context verification and host policy acceptance with the same worker/verifier model, channel and max reasoning effort. A separate real-model diagnostic explained missing evidence and the historical loop limit without requesting another model.
• Generated isolated history contained 1,100 plans and approximately 185 MB of coordination records. The complete task ran for 307.906 seconds with 1,243 interaction pairs. Input/scroll P99 was approximately 24/34 ms, with maximums of 203/417 ms. These automation-inclusive measurements do not guarantee zero spikes on every machine.
• All 180 required product images received implementer self-review and passed Verify, retaining seven existing minor limitations. The public Mac app ran with isolated data without replacing the daily installation. Interactive Windows/Linux installation, full screen-reader coverage and long-term use remain outside this claim.

0.1.0-rc.21

locale:zh-CN

发布状态

• 0.1.0-rc.21 继续修复大历史项目执行时的全局间歇停顿。正式版仍为 0.0.18。
• RC20 公开构建的完整真实任务功能通过，但主进程单次停顿超过验收门槛；本版保留该失败记录并继续修复，不将 RC20 记为完整性能通过。

修复与优化

• 计划发现按既有持久变更日志增量失效，避免恢复、核验和调度反复逐一检查全部历史计划文件。日志异常时重新完整校验，每个已知任务仍读取当前计划和执行守卫。
• 临时任务的工具结果现在跟随指定数据根保存，避免测试或多环境运行的新材料写入日常用户缓存。旧记录和产物引用不迁移、不重写。
• 同步中英文更新说明与站点日志；保留真实大历史任务、独立核验、正式回复、取消及大记录提交的验收范围。

升级与验收范围

• 数据格式与权限链保持。升级前备份完整数据根，Desktop、CLI/TUI 与 Gateway 同步升级；正式版通道不变。
• 私有历史仅用于本地无凭据离线复现；真实模型任务使用生成历史，不上传个人记录。先前测试生成的少量工具缓存已按明确身份核对迁回对应隔离目录，并保留文件校验记录。
• 性能验收不代表所有机器零峰值；产品截图由实施者自审，完整读屏、Windows/Linux 交互安装和长期日用不在本轮断言内。

发布后实测（2026-10-20）

• 未修改公开签名 macOS 应用及包内助手：1100 计划、200 历史工作、185MB 日志、9.6MB 主对话及46MB子历史，真实 DeepSeek max 完整任务356秒、1436组交互通过。18项原测试断言、核验签收、正式回复、大提交、取消无重放及长回执均通过。
• 输入 P99/峰值42/211ms，滚动53/436ms，主进程峰值397ms，均低于原门槛；没有页面长任务或错误。自动化数据包含传输开销，不是硬件延迟或零峰值保证。
• 实际私有历史本地无凭据/无Provider离线125秒、845组通过：输入峰值96ms、滚动84ms、主进程92ms。没有替换日常安装。
• 14产物哈希、Mac签名/公证/Gatekeeper、更新清单、4个npm固定版本和隔离安装、站点检查均通过。
• 产品180图核对为实施者自审；保留7项既有轻微限制。未覆盖完整读屏、Windows/Linux交互安装或长期日用。
• 完整ADR与验收证据。

locale:en-US

Release status

• 0.1.0-rc.21 continues to address intermittent application-wide stalls during large-history task execution. Stable remains 0.0.18.
• The public RC20 build passed the complete real task workflow but exceeded the main-process latency gate. That failure is retained and is not described as a complete performance pass.

Fixes and improvements

• Incrementally invalidate plan discovery using the existing durable change journal instead of repeatedly inspecting every historical plan during recovery, verification and scheduling. Invalid observations trigger a full recheck; known tasks still read current plans and execution guards.
• Materialize tool results under the selected data root so temporary tasks and separate environments do not write new artifacts into the daily user cache. Existing records and artifact references are not migrated or rewritten.
• Synchronize bilingual release notes and the documentation changelog, retaining full-scale real-task, independent verification, delivery, cancellation and large-record acceptance scenarios.

Upgrade and validation scope

• Retain data formats and permission boundaries. Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. The stable channel is unchanged.
• Personal history is used only for local credential-free offline reproduction. Real-model tasks use generated history without uploading personal records. A small set of earlier generated test artifacts was identity-checked and moved back to its isolated directories with file-integrity records retained.
• Performance checks do not imply zero spikes on every machine. Product screenshots receive implementer self-review. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this claim.

Post-release validation (2026-10-20)

• The unmodified public signed macOS app and bundled helper passed a real DeepSeek max workflow lasting 356 seconds with 1436 input/scroll pairs, 1100 plans, 200 historical works, a 185 MB journal, 9.6 MB parent history and 46 MB child history. The original 18 tests, host verification/acceptance and delivery, large commits, cancellation without replay, and long authoritative receipts passed.
• Input P99/max: 42/211 ms. Scroll P99/max: 53/436 ms. Main-process maximum: 397 ms. All passed the unchanged gates, with no renderer long tasks or page errors. Automation figures include transport and are not hardware-only latency or a zero-spike guarantee.
• The local private-history copy ran without providers or credentials for 125 seconds and 845 pairs: input max 96 ms, scroll 84 ms, main 92 ms. The daily installation was not replaced.
• All 14 asset hashes, Mac signatures/notarization/Gatekeeper, updater manifests, four fixed npm versions, an actual isolated CLI installation, and public Pages checks passed.
• All 180 product screenshots were checked by the implementer, with seven existing minor limits retained. Full screen-reader coverage, interactive Windows/Linux installation and long-term use are outside this claim.
• ADR and complete evidence.

0.1.0-rc.20

验收补充（2026-10-20）：RC20 公开签名包的完整任务功能通过，但一次主循环峰值约 569ms，超过既定 500ms 门槛，最终性能验收尚未通过。后续修复将采用新 RC，保留本版 tag 和产物。
Validation update: the public signed RC20 build passed the complete functional workflow, but a 569ms main-loop peak exceeds the unchanged 500ms limit. Final performance acceptance is pending a follow-up RC; this tag and its assets remain unchanged.

locale:zh-CN

发布状态

• 0.1.0-rc.20 修复大历史项目后台恢复检查造成的全局间歇停顿。正式版仍为 0.0.18。

修复与优化

• 恢复检查按工作读取，一批事件共享当前任务观察，一次选取只做一次恢复准备，避免历史工作与事件相互放大。

• 输入归属改为批量读取；计划调度在没有异步等待和写入的同步阶段复用观察，状态变化后重新读取。

• 续接发现使用带文件版本校验的最小检查点投影；实际续跑和合并仍读取完整原文，保留取消、未知执行和租约限制。

• 常规聊天页不再反复传输内部执行轮次；打开回复详情时按消息身份读取完整过程，保留历史错误、派发和排队标记。

• 任务观察先过滤所属项目再加载完整计划；输入队列复用已校验的历史图片，文件或严格读取要求变化时重新校验。

• 补足大历史验收的历史工作和送达回执分布，检查输入、滚动的 P99 与单次峰值。

• 停止在每条协调日志提交后同步重写未被读取的完整快照缓存；保留日志及目录持久提交和全部恢复守卫。

• 按需回复详情继续携带同一回合的真实停止状态，准备中未派发步骤不会被误显示为仍在运行。

• 同一大历史版本的列表、卡片、回复状态与任务事实共享严格校验的只读解析输入，避免消息更新后重复全文读取；保留副本隔离、深冻结、容量上限与损坏/竞争失效检查。

• 修正 CSV 字段行中的 header 被误判为界面页眉、导致纯命令任务等待截图的情况；保留真实界面及混合任务的视觉验收。

• 工具记账不再复制其他历史工作的正文；任务活动与调度按所需身份和委托计划读取，减少派发、复核及交付时的主进程阻塞，保留原持久化与授权检查。

• 恢复调度与送达观察复用经过文件版本检查的完整不可变状态，避免重复复制历史账本；保留默认可变副本接口和取消后的重新核对。

升级与验收范围

• 沿用 RC19 数据格式和权限链；升级前备份完整数据根，Desktop、CLI/TUI 与 Gateway 同步升级。正式版通道不变。
• 实际历史仅用于无凭据的本地隔离复现；真实模型任务使用合成历史，不发送用户私有历史。
• 性能结果不代表零延迟峰值；产品截图为实施者自审，完整读屏、Windows/Linux 交互安装和长期日用不在本轮断言内。

后续修复

RC21已通过同规模公开签名包的完整性能与功能验收：升级至v0.1.0-rc.21。本页RC20未过性能门槛的记录继续保留。

locale:en-US

Release status

• 0.1.0-rc.20 fixes intermittent application-wide stalls caused by background recovery scans with large project histories. Stable remains 0.0.18.

Fixes and improvements

• Read recovery work individually, share current task observations within an event batch, and prepare recovery once per selection to avoid historical work/event amplification.

• Read pending input ownership in one batch. Reuse plan observations within synchronous scheduling phases with no writes or awaits; re-read after changes.

• Use bounded, file-version-checked checkpoint facts for continuation discovery. Actual continuation and merging retain full original checkpoints, cancellation, unknown-outcome and lease guards.

• Keep internal execution rounds out of refreshing chat pages. Load complete reply processes by message identity on demand, preserving historical diagnostics and dispatch/queue badges.

• Filter project ownership before hydrating watch plans. Reuse validated historical image inputs until their file version or strict-read requirements change.

• Extend large-history acceptance with historical works and delivered receipts, and check P99 and worst-case input/scroll latency.

• Stop synchronously rewriting an unused full-state snapshot cache after each coordination receipt; retain durable journal/directory commits and recovery guards.

• Keep exact-turn interrupted Activity in on-demand reply details so undispatched preparing steps retain their stopped status.

• Share strictly validated, immutable history input across list, card, reply and task projections after a history change; retain isolated outputs, bounded memory, and corruption/race invalidation.

• Stop treating explicit CSV field headers as UI pixels in command-only tasks; retain visual acceptance for real and mixed UI deliverables.

• Avoid copying unrelated historical work bodies for tool accounting. Read the identities and delegated plans needed for activity and scheduling, retaining durable commits and authorization checks.

• Share complete, file-version-checked immutable state for recovery and delivery observations, retaining mutable default reads and current-state checks after cancellation.

Upgrade and validation scope

• Retain RC19 data formats and permission boundaries. Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. The stable channel is unchanged.
• Personal history is used only in local credential-free isolated reproduction. Real-model tasks use synthetic history without uploading personal records.
• Performance results do not imply zero latency spikes. Screenshot review is implementer self-review. Full screen-reader coverage, interactive Windows/Linux installation and long-term use remain outside this claim.

Follow-up fix

RC21 passed the same-scale complete public-build workflow and performance gates. The RC20 failure record above is retained.

0.1.0-rc.19

locale:zh-CN

发布状态

• 0.1.0-rc.19 修复长回复被错误判为失败、有效复核总结被拦截，以及大历史项目执行时输入和滚动卡顿的几条关键路径。正式版仍为 0.0.18。

修复与优化

• 兼容流收到协议结束标记后立即交出回复和工具调用；取消、超时不再等待底层连接关闭，调用方取消信号持续覆盖整个响应体。

• 新用户输入可抢占没有活跃工具派发的后台模型请求，保留原工作检查点与事件归属；已派发或未知操作仍按原结果路径收尾。

• 主 Bot 从本地完整执行回执判断 post_reply 和等待结果，界面工具预览仍保持有界。长回复不会因为预览裁剪而丢失成功标记、正式消息和宿主来源版本。

• 已有真实工具结果的复核总结不再因出现“工具返回”、exitCode 或“我检查了”而误报伪工具协议。真正的伪调用标签、未执行行动旁白仍拦截；证据和正式签收门禁保留。

• 切换模型或连接后，对已完成且没有未知副作用的主回复，使用当前模型依据已有结果继续认知；不跨模型回放原生调用，不重放旧工具，不清空用量。未决执行继续明确阻塞。

• 历史错误折叠后仍可按原始消息身份打开详情，保留该次过程和原始错误；修复虚拟列表短暂卸载后引用反馈丢失。

• 计划发现只解析相关委派计划；一次只读查询共享事实快照；没有待恢复工作的机器人不再反复扫描整份任务名册。

• 主对话摘要与恢复卡读取使用版本校验的有界投影，追加消息无需重复读取整份历史；并发写入导致版本变化时有限重试，不返回过期事实。

• 合并流式通知、保留未变消息引用，减小虚拟列表窗口，并移除执行状态的持续背景重绘，降低运行时对输入和滚动的影响。

升级与验收范围

• 升级前备份完整数据根；Desktop、CLI/TUI 与 Gateway 同步升级。延续 RC18 协调日志兼容性约束，回退须恢复完整升级前备份。
• 大历史验收使用与现场规模相当的合成历史；用户真实历史只读用于本地定位，不发送到云模型。真实任务与受控 UI 回归分别记录。
• 截图为实施者自审，不称独立产品审查；任务复核使用独立上下文，可能仍为同源模型。Windows/Linux 交互安装、完整读屏、安装器交接及长期日用不在本轮断言内。

locale:en-US

Release status

• 0.1.0-rc.19 repairs failed delivery of long replies, false rejection of valid verifier summaries, and key input/scroll bottlenecks during execution with large project histories. Stable remains 0.0.18.

Fixes and improvements

• Settle compatible streams at protocol completion and release cancellation or timeout without waiting for transport cleanup. Caller cancellation remains active throughout the response body.

• Admit fresh user input ahead of background model requests with no active tool dispatch, preserving the original work checkpoint and event ownership. Dispatched or unknown operations retain their existing result handling.

• Use complete local execution receipts for post_reply and structured waits while keeping UI previews bounded. Preview clipping no longer removes success, canonical replies or host source revisions.

• Permit factual summaries after actual tool results without rejecting words such as exitCode or “I verified”. Literal pseudo calls and unexecuted action preambles remain blocked; Evidence and host acceptance remain required.

• After changing model or connection, safely continue reasoning over completed results with the current model. Never replay another model's native calls, repeat completed tools or reset usage. Unknown execution outcomes remain blocked.

• Resolve folded historical error details by canonical message identity, retaining the original process and error. Preserve reference feedback across temporary virtual-row unmounts.

• Parse only delegated plan candidates, share facts within one read query, and avoid full recovery-roster scans for bots with no eligible work.

• Use bounded, version-checked history and recovery projections. Metadata-only append receipts avoid rereading full history; concurrent version changes receive bounded retries without stale fallback.

• Coalesce streaming notifications, retain unchanged message identities, reduce the virtual window and remove continuous status-background repainting to improve interaction during execution.

Upgrade and validation scope

• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. RC18 coordination-journal compatibility restrictions still apply; rollback requires the complete pre-upgrade backup.
• Large-history acceptance uses synthetic history matching the observed scale. Personal history is read locally for diagnosis and is not sent to cloud models. Real-model tasks and controlled UI regression are recorded separately.
• Screenshots are implementer self-review, not independent product review. Task verification uses a separate context and may share the same model lineage. Interactive Windows/Linux installation, full screen-reader coverage, installer handoff and long-term use remain outside this claim.

0.1.0-rc.18

locale:zh-CN

发布状态

• 0.1.0-rc.18 修复 DeepSeek 工具调用后的继续执行、历史任务重复续接，以及回复已交付后仍反复唤醒的问题。正式版仍为 0.0.18，本版用于继续评测。

修复

• DeepSeek 思考模式保留服务实际返回的推理字段，工具结果返回后可以继续执行；旧模型历史不再伪装成新的工具调用，不补造思考内容或重放已完成操作。
• 恢复任务时保留原生调用历史，同时使用当前系统上下文，让最新用户输入、任务结果与交付规则进入本次续跑。
• 同一对话中、同一子任务的重复后台续接合并为一个负责人；新用户输入优先处理，取消、未知执行结果及权限边界继续保留。
• 已正式交付且事实未变化的结果进入持久化静默等待；实际任务状态、复核结果或用户问题变化后才重新处理。交付依据仍由本地宿主核对。
• 修复长协调日志在正常追加读取前重复扫描全文的问题，保留损坏尾部恢复与严格校验。

升级与验证范围

• 升级前备份完整数据根，Desktop、CLI/TUI 与 Gateway 同步升级。本版新增协调日志事件，旧版可能拒绝恢复新日志；回退须恢复完整升级前备份。
• 源码真实任务覆盖 DeepSeek 派工、读写与命令执行、独立上下文复核、正式回复、取消、重启恢复，以及重复历史续接收束。GPT 主 Bot 与 DeepSeek 子任务的首轮复核出现服务输出格式错误；同一任务显式续接后完成，未重做文件修改。
• 产品截图为实施者自审，非独立审查。小型隔离任务不代表任意项目、网络或长期日用；Windows/Linux 交互安装、完整读屏与安装器交接另行覆盖。
• 同步双语发布说明、README 版本入口和文档站更新日志。

locale:en-US

Release status

• 0.1.0-rc.18 fixes DeepSeek continuation after tool calls, duplicate historical continuations and repeated wakes after a result has already been delivered. Stable remains 0.0.18; this prerelease is for continued evaluation.

Fixes

• Preserve authentic reasoning fields returned by DeepSeek in thinking mode so execution can continue after tool results. Historical calls from older models become labelled history instead of new executable calls; reasoning is never fabricated and completed operations are not replayed.
• Restore native call history with the current assembled system context so new user input, task results and delivery rules reach the resumed turn.
• Coalesce duplicate background continuations for the same child task and parent conversation under one owner. Prioritize new user input while preserving cancellation, unknown outcomes and permission boundaries.
• Persist quiet waiting after verified delivery when the underlying facts are unchanged. Resume processing when task state, verification observations or user questions change, with delivery evidence still checked locally.
• Avoid rescanning a complete coordination journal before normal appended reads while retaining damaged-tail recovery and strict validation.

Upgrade and validation scope

• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. This release adds coordination journal events that older versions may refuse to restore. Rollback requires the complete pre-upgrade backup.
• Source real-model acceptance covers DeepSeek delegation, file and command tools, verification in an independent context, formal delivery, cancellation, restart recovery and duplicate historical continuation settlement. A GPT main Bot with DeepSeek workers initially encountered malformed verifier output; explicit continuation completed the same task without repeating file edits.
• Screenshot review is implementer self-review, not independent review. Small isolated tasks do not cover arbitrary projects, networks or long-term use. Interactive Windows/Linux installation, full screen-reader coverage and installer handoff remain separate.
• Synchronize bilingual release notes, README version links and the documentation-site changelog.

0.1.0-rc.17

locale:zh-CN

发布状态

• 0.1.0-rc.17 修复执行任务时的三处性能热点，以及子任务思考强度与机器人配置不一致的问题。正式版仍为 0.0.18，本版用于继续评测。

执行时的流畅度

• 读取任务进展时复用协调日志状态，只处理新追加的记录；文件替换、损坏和版本检查继续生效，长期历史不再被每次全量重放。
• 历史消息与机器人列表的头像保持静态，仅当前活动身份按状态显示动效，减少空闲绘制负担；减少动态效果设置继续有效。
• 大会话结束时，全文读取和序列化在后台 worker 中准备，主进程只完成短时间的校验与原子提交。回复完成、取消和执行槽位释放均等待实际保存；保存失败会明确报告，不自动重复工具操作。

子任务思考强度

• 创建子会话时同时传递该任务冻结的模型与思考强度，避免继承父会话旧强度。
• 已有任务按其自身冻结配置显示，不用当前机器人配置覆盖历史任务；执行模型原生支持的最高强度保留原值。

升级与验证范围

• 升级前备份完整数据根，Desktop、CLI/TUI 与 Gateway 同步升级；回退须恢复完整升级前备份。
• 源码验收覆盖真实模型派工、子任务读写及命令执行、复核交付、取消后继续对话。大日志和大对话分别测量；冷启动全量验证与旧同步存储兼容接口仍有成本。
• 产品截图由实施者自审，非独立审查。Windows/Linux 交互安装、完整读屏、任意项目的性能及长期日用不在本次断言内。
• 同步双语发布说明、README 版本入口和文档站更新日志。

locale:en-US

Release status

• 0.1.0-rc.17 fixes three performance hotspots during task execution and inconsistent reasoning effort in delegated conversations. Stable remains 0.0.18; this prerelease is for continued evaluation.

Responsiveness during execution

• Reuse validated coordination state and read appended records instead of replaying the entire history for each progress read. Replacement, corruption and revision checks remain enforced.
• Keep historical-message and Bot-list avatars static. Animate only eligible current activity, reducing idle rendering work while respecting reduced-motion settings.
• Prepare large terminal conversation rewrites in a background worker. The main process performs a short version check and atomic commit. Completion, cancellation and execution-slot release wait for durable saving; a save failure is reported without automatically replaying tools.

Delegated reasoning effort

• Pass both the frozen model and reasoning effort when creating a child conversation, avoiding inheritance of stale parent effort.
• Display existing tasks using their own frozen configuration, without replacing historical settings with the current Bot policy. Preserve native maximum effort where the execution model supports it.

Upgrade and validation scope

• Back up the complete data root and upgrade Desktop, CLI/TUI and Gateway together. Rollback requires the complete pre-upgrade backup.
• Source acceptance includes actual model delegation, child file edits and commands, verification and delivery, cancellation and subsequent conversation. Large histories are measured separately; cold full validation and legacy synchronous storage APIs still have a cost.
• Product screenshots receive implementer self-review, not independent review. Interactive Windows/Linux installation, full screen-reader coverage, arbitrary-project performance and long-term daily use remain outside this claim.
• Synchronize bilingual release notes, README version links and the documentation-site changelog.