Changelog
版本历史与发布说明
历史摘要:较早条目描述对应版本的行为,而非当前安装。完整规范更新日志: CHANGELOG.md.
v4.12.3
Latest 2026年11月10日-
FIX
MiniMax 安装不再写入全局 hook 状态。
~/.minimax/hook-settings/state.json位于项目之外,因此只有项目自身的条目会变化,且仅当项目拥有(或曾经拥有)MiniMax hooks 时才会变化。 -
FIX
hook 包装器中的 MiniMax 载荷契约现在与实测结果一致。 mcode 0.7.6 将 Claude 风格的
hookSpecificOutput.permissionDecision: "deny"和顶层{"decision": "block"}映射为同一个内部拒绝,且不带原因的 block 会被丢弃。 - FIX 内部 hook 失败不再静默。 退出状态 2 会阻止调用并从 stderr 获取原因,任何其他非零状态都会作为 hook 诊断上报,而无法评估的 PreToolUse 事件现在会失败关闭,而不是放行调用。
-
FIX
MiniMax 层级 effort 已记录在案,而不是被静默丢弃。 实测的
mcode命令行暴露了--model,但没有 effort 标志,因此由层级选择模型。 -
FIX
两个 hook 写入器在同一个
~/.minimax/hooks.json上的共存现在已有测试覆盖。 宿主 Jev hook 与独立的 Human Outcome Reporting 会保留对方的条目,两种顺序均如此。 - FIX 版本一致性门禁现在能以任意名词形式读取俄语目录三元组。 词形变化(“32 команды”、“82 навыка”)会与磁盘内容比较,而不是被报告为模式漂移;过时的前导提及也不再能掩盖同一行后面出现的计数漂移。
- FIX 派生计数的页面还会额外检查硬编码数字。 类别词旁边的硬编码目录大小现在会使门禁失败,并且精确的跨根表面计数会再次被断言。
-
FIX
精确 trim 上的无 Redis 防护不再能通过引号绕过。 近似形式(
MAXLEN ~、MAXLEN "~"、MAXLEN '~')会被拒绝,注释文本会被忽略,精确形式会被固定。 - FIX 4.12.0 章节列出了该版本中交付的全部十六个拉取请求,标题与发布正文逐字一致。
- DOCS 版本 4.12.2 已打标签但从未发布。发布工作流从已签名标签消息内的标记中读取版本递增级别,而发布标签受保护,禁止更新和删除,因此该标签无法携带该标记。4.12.2 的更改原样包含在 4.12.3 中,变更日志仅列出已发布的版本。v4.12.2 标签已于 2026 年 11 月 11 日移除:它从未发布,因此被删除,其更改包含在 4.12.3 中。
-
DOCS
版本 4.12.3 已于 2026 年 11 月 10 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布文件;源归档 SHA-256:
a511017d6748f29b06f8cb1986a2854cb8e94ce2595066533d0f3e592ba11b51。
v4.12.1
2026年11月10日- FIX 变更日志仅列出已发布的版本。 4.11.0 部分(从未打标签、从未发布)已并入 4.12.0,而 4.10.0 部分则标注了其实际发布日期。
-
FIX
版本一致性校验不再在已对齐的工作区上失败。 它会读取两种发布表述,并按照站点构建的方式检查 datarim.club 的各个界面:
pages/about.php和pages/home.php必须从目录中推导出计数,而首屏文案则按智能体/命令/技能三元组进行校验。 -
FIX
仓库↔站点漂移检测器不再将准确的引用报告为漂移。 嵌套技能(
skills/fleet/l1-basic/SKILL.md发布为fleet-l1-basic)、插件命令和已实现的标志均已解析;此前有七处准确的引用被报告为差异。 -
FIX
fleet 审计日志裁剪会精确保留所请求的条目数。 它此前使用近似的
XTRIM,该命令只删除完整的宏节点,因此流长度仍超过--max-len。 -
DOCS
版本 4.12.1 已于 2026 年 11 月 10 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布资产;源归档 SHA-256:
303e6dc19ee48adde16a7afeb5e34f23fa1fe4ca9d05db42262ff9e4cd87930a。
v4.12.0
2026年11月9日-
NEW
一等公民的 MiniMax Code(
minimax/mcode)支持。 项目安装(./install.sh --client minimax)、更新(./update.sh)、Jev(jev --agent=minimax、jevminimax、.minimax/hooks.json、~/.minimax/hooks.json),以及不依赖 Datarim 的独立 Human Outcome Reporting(--agents minimax安装到~/.minimax/skills/human-outcome-reporting/和~/.minimax/hooks.json),现在都与 Claude Code、Codex CLI、Cursor 和 Google Antigravity 一样支持 MiniMax Code。MiniMax 没有 markdown 命令加载器,因此/dr-*命令以技能形式出现在.minimax/skills/下。 -
FIX
Jev 安全底线在 MiniMax 上确实能拦截破坏性命令。 MiniMax Code 运行时以其自身的原生格式解析 hook 输出:拒绝会以顶层
{"decision": "block", "reason": "..."}载荷加上退出状态2的形式传递。该运行时接受 Claude 风格的hookSpecificOutput.permissionDecision,但从不强制执行,因此 MiniMax 上的安全底线仅依赖原生载荷;提示性上下文通过hookSpecificOutput.additionalContext传递给模型。 -
DOCS
版本 4.12.0 已于 2026 年 11 月 9 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布资产;源归档 SHA-256:
32c367bfb021fd05cd7670364e77a86cf479c541acc8df54ad08956022fca675。
v4.10.0
2026年11月3日- NEW 有界的 CLI 输入清单工作流。 工具可选择、验证、哈希、比较、合并、检查冲突,并格式化结构化或可读的清单差异。
- NEW 摘要工具会解析声明的 schema、转发声明的退出状态,并聚合摘要汇总。 其输出仅报告所提供的数据,并不证明生产者已运行或已接受其结果。
- NEW QA 辅助工具会报告初始化任务是否存在,并为追加 I/O 失败保留诊断信息。
-
DOCS
版本 4.10.0 已于 2026 年 11 月 3 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布资产;源归档 SHA-256:
d60f0877cb6ea11cdd705cb1436181986c855ceb4b124d85fefe7a6650e684c3。
v4.9.0
2026年10月29日-
NEW
原生支持 Google Antigravity(
antigravity/agy)。 项目安装(./install.sh --client antigravity)、更新(./update.sh)、Jev(jev --agent=antigravity、jevagy、jevantigravity、.agents/hooks.json、~/.gemini/config/hooks.json),以及不依赖 Datarim 的独立 Human Outcome Reporting(--agents antigravity,claude,codex,cursor安装至~/.gemini/config/skills/human-outcome-reporting/和~/.gemini/config/hooks.json),现已在 Claude Code、Codex CLI 和 Cursor 之外同时支持 Google Antigravity。 -
NEW
/dr-qa与/dr-compliance中的穷尽式单遍审计与单轮批量修复。 两个阶段都不会在首次失败后短路(工作区不干净或 HEAD 漂移时的 Step 0 溯源门禁除外):两者均在一遍中评估 100% 的适用层级、门禁、预期项与领域检查清单,并将所有差异汇总为编号的Consolidated Remediation Batch([QA-BATCH-1..N]/[COMP-BATCH-1..M]),使/dr-do与/dr-auto能在一次返工迭代中解决并本地复验全部发现项。 -
DOCS
版本 4.9.0 已于 2026 年 10 月 29 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布资产;源归档 SHA-256:
90ffb678b47335033b0f0cce5f98982acfa0838772b432d1149522d2f45e31fb。
v4.8.1
2026年10月26日-
FIX
团队中立模式保留项目根目录指令。安装和更新不会改动根目录的
AGENTS.md及其他指令文件。对旧版受管根块的更新会停止,直到手动清理完毕。 - FIX 发布门禁会检查由所有者持有的私有清单修订版本。打标签前会检查源码与软件包;签名后,构建会在发布前检查全部 11 个生成资源,包括签名包。日志中不包含私有值。
-
DOCS
版本 4.8.1 已于 2026 年 10 月 26 日发布。安装前请使用发布指南验证源归档、校验和、签名及证明。发布资产;源归档 SHA-256:
d45939231a729fda6fd371d9bbddbc8aedeabb5525a98bf03fb3362a56eade65。
v4.8.0
2026年10月26日-
NEW
共享仓库获得
--team-neutral。 Datarim 命令保持显式;只有datarim/AGENTS.md、datarim/config.yaml和datarim/local/对 Git 可见。全新安装不会改动项目根目录的指令;升级可能会移除其旧的管理块。 -
NEW
QA 与合规报告可与队友共享。
delivery.reports使用{task_id}选择目录。发布时会与工作副本进行字节级一致性校验;未配置时,沿用原有行为并给出提示。 - FIX 收尾遵循项目自身的规则。 QA 检查所报告的失败与常规成功路径;合规在约定变更后运行已发现的校验器并检查活动规则存储;归档记录项目义务与未知结果。
-
DOCS
版本 4.8.0 已于 2026 年 10 月 26 日发布。源归档、校验和、签名、SBOM 及证明可从发布页面获取。源归档 SHA-256:
918551ff7b94789dcb9126149dda59ba89816e5a4380558414fa177cc5ba047f。更新前请遵循发布验证指南。共享仓库请参阅团队中立指南。
v4.7.0
2026年10月25日-
NEW
自主演进现在会在通过合规检查后决定提案。
/dr-auto和/dr-orchestrate会记录每次决策、理由和门禁结果。B 类框架变更需通过 PRD 和独立审查。 -
NEW
/dr-finalize使用资源所有权台账关闭任务。 创建的工作树、分支和临时路径会在创建时记录;清理时会检查台账和资源状态。 -
NEW
datarim/local/中的本地扩展在更新后仍会保留。 安装时会将其与上游组合,报告基线漂移,并在替换直接编辑过的运行时文件之前停止。 - FIX 九项 QA、验收和合规检查得到加强。 PASS 需要实时证据和 QA 章程覆盖,而每条保留的 FAIL 行都需要处置和重放证据。
-
DOCS
版本 4.7.0 已于 2026 年 10 月 25 日发布。源归档、校验和、签名、SBOM 及证明可从发布页面获取。源归档 SHA-256:
a8d8eae504f953283441d09dcc32ec29db7d4dd97427786785e62885fbfb361e。更新前请遵循发布验证指南。
v4.6.1
2026年10月22日- NEW 在测试层级的运行结果用于生产影响预测之前,相关设置会在两个层级之间以只读方式进行比较。 密钥值会被排除;不安全或不可用的比较会使该预测失效。
-
NEW
当验收契约被取代时,
/dr-qa与合规检查会逐条比较所有用例文本。 实质性的语义差异会记录在 Return-to-Source 中。 -
DOCS
版本 4.6.1 已于 2026 年 10 月 22 日发布。源归档、校验和、签名、SBOM 及证明可从发布页面获取。源归档 SHA-256:
4af94d25e2badda8b599c162e982b0bdea9047c2ff0ff3fbb6417d405d3fad9d。更新前请遵循发布验证指南。
v4.6.0
2026年10月20日-
FIX
L3/L4 任务在开始实现前会检查是否存在 PRD。
/dr-do和/dr-auto会提前运行现有门禁,并在缺少必需 PRD 时路由到/dr-prd,而不是等到合规检查时才发现。 -
NEW
对非
V-AC的linked_ac值提供早期提示。 当某个值既不是V-AC-N令牌也不是表示无链接的短横线时,会发出提示linked-ac-not-vac-token,此时仍可低成本地编辑预期。退出码保持不变。 - FIX 验收用例在每个计划步骤之后都必须保持可观测。 当后续步骤移除了被观测的机制时,用例应指明该移除之后的观测方式。
-
DOCS
版本 4.6.0 已于 2026 年 10 月 20 日发布。源归档、校验和、签名、SBOM 及证明可从发布页面获取。源归档 SHA-256:
addf0f39fc8cada1c7c00961184585f65d24d6b4f59f9f62b9e5c9edae9642a8。更新前请遵循发布验证指南。
v4.5.0
2026年10月19日- NEW 验证尝试与其声明的输入清单绑定。 被取代的尝试仍会保留记录,且选定的尝试在其输入发生变化后无法通过。
-
NEW
/dr-verify校验发现与解决记录,而/dr-qa额外提供 QA 章程模板与覆盖率检查器。 未解决的发现到达迭代上限时仍保持未解决状态。 -
NEW
/dr-design与/dr-qa覆盖适用的 UI 状态和开发者旅程。 评分标准得分仅作为评审辅助,不代表实测的用户结果。 -
FIX
发布仅接受已验证的签名标签,且其目标须为受保护
main的祖先。 发布任务在发布前会再次验证该标签。 - FIX 更正后的合规报告会在其预留路径上重新校验。 状态仅在验证通过后变更;未解决的结果仍保持阻断。
- DOCS 维护者指南现已定义源与归档的隐私证据要求。 文档并不声称某个特定归档已通过评审,也不声称运行时已被接受。
-
DOCS
版本 4.5.0 已于 2026 年 10 月 19 日发布。源归档、校验和、签名、SBOM 及证明可从发布页面获取。源归档 SHA-256:
78aa1f64ead4fd738aa2daf0dd335d045b5dc6bc2322b27070314211aa12fdf3。更新前请遵循发布验证指南。
v4.4.0
2026年10月18日- NEW 产品发现新增决策记录 v2 与策略师范围模式。 证据接受、业务成果、建议与行动权限保持相互独立。仅凭 GO 并不授权执行;v1 记录仍可读取,严格读取器会拒绝缺少反证的 v2 GO。
-
FIX
/dr-quick保持读写边界。 查询保持信息性 / UNCERTIFIED;当意图变为编辑或保存结果时,变更型 QCK 会在首次写入前启动,并带有全新路由、接受用例与预检回执。闭包检查使用阶段quick;变更型 QCK 阶段快照紧随 CTA。 - FIX 快速任务学习仍为条件性提案。 仅针对已修正并验证的具体错误或逻辑不匹配进行记录。A 类保持待批准,B 类保持搁置直至其所需的契约/PRD 更新完成,未解决的类别保持未解决;已应用需要应用证据。
-
DOCS
/dr-auto在单一终端摘要中携带反思状态。 待处理、搁置、已应用与已拒绝状态均被保留;缺失或过期的反思报告为未测量。这是源契约,不构成实时聊天投递。参见快速任务证据与学习指南。 -
DOCS
版本 4.4.0 已发布。 源代码归档、校验和、cosign 签名、SBOM 和证明可从发布页面获取。源代码归档 SHA-256:
7ee2f70a368daeb6f32672d93223f031cc2be1ca0453f79c8dfd8e6e105d9154。安装前请遵循发布验证指南。
v4.3.7
2026年10月6日- FIX 工具周围的回复语言保持一致。 普通叙述(包括首条和末条消息)使用回复语言;可复用文档保留独立的产物语言。请求更改某个文档的语言不会改变周围的说明。
-
NEW
Claude 在工具之后收到当前偏好。 自有的同步
PostToolUse和PostToolUseFailure回调使用共享解析器。工具输出、决策、权限和外部钩子均保持原样。 - DOCS 上下文投递限制已明确。 工具后输入上限为 1 MiB;启动和提示输入保持 64 KiB。无效或超限数据不会成为指令。权限拒绝、执行前校验失败和取消不会触发失败工具回调;上下文投递不保证生成语言。
- FIX 检查不依赖开发者的语言偏好。 两个断言英文标签的检查显式选择英文。安装程序完整性哈希已更新;个人设置和安全集成保持原样。
-
DOCS
独立报告已更新至 0.3.3。 安装指南、技能卡片和模板链接均跟随当前版本。编排、
/dr-auto、Jev 和/dr-quick仍是核心入口;原始徽标保持不变。
v4.3.6
2026年10月6日- FIX 历史截止时间检查衡量的是预期阶段。回归计时从历史验证开始,到终端响应写入后结束。更早的工作保留其自身的全局截止时间;四秒测试上限和拒绝条件保持不变。
- FIX 回归检测每个子进程的截止时间重置。跨多次 Git 调用的短延迟和调用计数器检查同一个共享时间预算。生产验证器行为保持不变。
-
FIX
HTTP 服务器如实关闭每个连接。内置服务器每个连接处理一个请求,并声明
Connection: close,以便客户端为下一个请求打开新连接。契约检查保留其现有范围。 -
FIX
UTF-8 响应保留其精确长度。HTTP 正文长度按字节计算,因此多字节文本不会收到过短的
Content-Length。请求解析器保留其区域设置。 -
DOCS
权限检查适用于 GNU 和 BSD。 版本回归使用适当的
stat语法,同时保留严格的文件和目录模式,并拒绝组可写记录。 -
FIX
私有项目语言设置保留在其自身项目中。最近的
datarim/config.local.yaml会选择该项目,而不是继承祖先框架;格式错误的设置仍会报错。 -
FIX
Claude 在新提示前刷新偏好设置。一个受管的同步
UserPromptSubmit会在每次提交提示前提供当前的回复语言和产物语言。它不会在每次工具续接时运行,也不保证模型遵守。原生钩子信任和权限边界保持不变。 - FIX 回调执行当前辅助程序源码。启动和提示回调不会执行预先存在的辅助程序字节码缓存,不会写入该缓存,并保留外部缓存。
-
DOCS
独立报告已更新至 0.3.2。安装指南和模板源链接随版本发布更新。文件检查不衡量行为、启动激活或每轮激活;需要实际观察。编排、
/dr-auto、Jev 和/dr-quick仍是突出的入口点。已批准的徽标保留其原始字节。
v4.3.3
2026年10月5日- FIX 策略师门禁记录校验可在 macOS 上运行。 文件身份从实际打开的描述符读取,而非 BSD 虚拟路径元数据。所有权、权限和文件替换检查仍然保留。
-
FIX
校验选定的 PRD 和计划修订版本。 预期、规范、追踪、阶段检查以及确定性的
/dr-verify底线读取任务描述中的prd和plan字段。支持显式版本化文档;缺少指针时保留旧版文件选择。 - FIX 过时副本不再掩盖错误。 当任务选择较新的修订版本时,干净的旧版文件无法在校验期间替代它。损坏的指针或任务身份不匹配会产生选择错误;校验既不会猜测最新修订版本,也不会回退到其他文件。
- FIX 拒绝格式错误的输入。 检查涵盖标识符、扁平描述符字段、重复键、引号和允许的路径。绝对路径、路径遍历、跨任务文档以及所选路径中的符号链接均被拒绝。选择检查不认证任意文档元数据,也不锁定并发编辑者。
- CHANGE 官方标识随框架包发布。 签名版本包含已批准的徽标及浅色/深色变体;原始文件保留其精确字节。网站保持相同的标识。
- DOCS 校验器契约说明了文档选择及其限制。 目录保持为 19 个智能体、29 个命令、80 个技能和 43 个模板资源。独立的 Human Outcome Reporting 保持 0.3.1;语言偏好和四个核心入口点仍然可用。
v4.3.2
2026年10月4日-
FIX
在首条消息前确定偏好。 独立报告会为受支持的 Codex 和 Claude Code 客户端注册同步的
SessionStart回调。它们分别提供独立的回复语言和产物语言;每次启动或恢复时都会读取设置,无需重新安装。Cursor 保留其原生的对话启动集成。 -
DOCS
安装与激活是两回事。 新的 Codex 钩子定义需要通过
/hooks进行原生信任审查。安装程序不授予信任、不更改权限,并保留外部钩子。文件检查无法确认钩子是否执行或首条消息的语言;请观察一个新会话。 - FIX 回复中的注释仍属于产物。 生成的注释和文档摘录中的正文与示例使用产物语言;周围的说明使用回复语言。不会添加未经请求的翻译示例。明确的双语或翻译请求、逐字输入、代码和机器值保留其例外。
-
DOCS
目录与安装随版本发布。 Human Outcome Reporting 0.3.1、卡片和模板链接使用 4.3.2 源。编排、
/dr-auto、/dr-quick以及独立的 Jev 安全底线仍然可用。
v4.3.1
2026年10月4日-
FIX
英文文档通过原生预期检查。共享解析器接受现有的
Operator expectations、Status history、Current status及相关字段标签,保留规范机器键、俄语文档,并严格拒绝无效输入。 - FIX 图谱检查每一条验收链接。单个愿望中的多个 V-AC 引用会被逐一检查;append-log 和 append-merge 区段中的愿望会被保留。
- DOCS 独立的语言偏好仍然可用。4.3.0 引入的回复与产物设置无需重新安装设置。安装、源和产物卡片均链接到修正后的版本。
v4.3.0
2026年10月4日-
FEAT
回复语言和文档语言可独立选择。
language.replies和language.artifacts均默认为en。一个辅助程序会读取个人、共享项目和私有项目设置;更改这些设置无需重新安装。 - FEAT 相同的偏好设置在没有框架时同样有效。 独立的 Human Outcome Reporting 会读取个人设置,Cursor 会在新对话中接收当前语言上下文。原生客户端规则、安全性和权限仍具有最高效力。
- CHANGE 固定报告标签使用语言目录。 目录覆盖缺失时会披露为英文回退。引文、代码、机器值和提供的文本不会被自动翻译。
-
DOCS
配置和限制一并记录在文档中。 操作指南、参考、教程和卡片说明了来源优先级、
Artifact language迁移以及如何检查新会话。文件检查与观察到的模型行为是两回事。 - FIX 检查端到端覆盖新增语言。 CI 清单包含每个新用例;完整的已签名法语/阿拉伯语证据链会拒绝缺失、多余和重复的条目。配置边界和本地扫描器设置已修正,且未削弱准入规则。
v4.2.3
2026年10月4日- FIX 文字和数字保留正常空格。Human Outcome Reporting 0.2.3 要求日期、计数和角色编号可读。咨询性检查会标记明显的拼接,同时保留代码和字面标识符。
- FIX 原生测试夹具相互隔离。发布后控制与无关的场景准备相分离;故意变更必须由预期场景检测到。设置错误和超时不能计为成功验证。
- FIX 公开示例中已移除私有来源信息。剩余的内部项目引用改用中性契约;移除未使用的私有项目检查器。
- FIX 发布说明链接在历史迁移后仍然有效。根历史发布链接到其当前带标签的变更日志,而不是与验证后计划退役的标签进行比较。普通发布保留自动生成的说明。
- CHANGE 单一公开主分支。当前源代码始于一个已签名的根提交。先前的公开标签和发布下载仅在独立后继验证通过后才可退役;旧的拉取请求引用、复刻和缓存可能保留较旧的对象。
- DOCS 历史遵循规范变更日志。重复条目被合并而不丢失发布信息;网站覆盖当前 CHANGELOG 中保留的每个唯一版本,并明确标注未发布的候选版本。
- DOCS 从已验证的干净克隆更新。迁移指南保留本地工作并避免合并无关历史。Orchestration、/dr-auto、/dr-quick 和 Jev 仍是主要入口点。
v4.2.2
Candidate only 2026年10月3日- CHANGE 已打标签的候选版本,而非稳定版本。根提交与标签均已创建,但验证发现一处原生夹具失败,且仍存在私有项目来源信息。未发布任何稳定版本资产,也未执行任何安装。相关修正将在 4.2.3 中发布。
v4.2.1
2026年10月3日- FIX 源隐私修正。归档测试示例中剩余的两处内部名称引用已替换为中性示例;旧版内部项目推广内容已从网站移除。这是一个独立的补丁版本;已签名的 4.2.0 版本不会被重写。
- DOCS 链接指向修正后的版本。安装、连接图和模板卡片均使用 4.2.1。4.2.0 中引入的四项核心能力和人工报告功能仍然可用;目录仍包含 19 个智能体、29 个命令、80 个技能和 43 个模板资源。
v4.2.0
2026年10月3日- FEAT 报告阐明人类可读的结果。 Human Outcome Reporting 将原始请求、交付结果、已验证条件和证据串联起来,并让缺失的检查与局限性清晰可见。它既可在 Datarim 内部使用,也可独立用于 Claude Code、Codex 和 Cursor。
-
FEAT
/dr-explain再次解释。 请求对现有结果或术语进行更简单的解释,不执行工作,也不改变任务状态。 - CHANGE 完整优先于简短。 Human Summary 保留其熟悉的四个标题,应用共享的报告契约,但不设硬性字数上限。命令、智能体、模板和连接图均使用相同规则。
- FIX 发布辅助工具要求提供证据。 检查绑定到确切的源代码修订版本;未知状态绝不视为成功。创建并验证签名标签,仅推送其确切的 ref,推送或发布后安装失败仍视为失败。
- DOCS 安装与目录均为最新。 独立报告与项目框架有各自的说明。模板卡片链接到发布源;目录包含 19 个智能体、29 个命令和 80 个技能。
v4.1.0
2026年9月30日-
CHANGE
Claude Code 原生读取
AGENTS.md。安装器从不创建CLAUDE.md;--claude-import和--no-claude-import已弃用,为空操作。更新时仅移除记录为安装器所有的旧版链接。
v4.0.5
2026年9月25日-
FIX
dr-orchestrateworks on macOS. The open-file identity check readstat -L /dev/fd/N, which on macOS reports the device of/dev/fdrather than the file’s own, so every context-window setup and every audit append was refused (“requires both explicit trust flags”, “append path changed during open”). The descriptor is now read with fstat. -
FIX
The audit ledger no longer needs util-linux
flock, which macOS does not ship; without it the same lock is taken through perl. -
FIX
A failing context-window test no longer leaves its tmux session behind. On macOS 49 of the plugin’s 92 failing tests now pass; the other 43 depend on GNU
realpath -mand onflockin the resolver and learned-rules store. Linux is unchanged.
v4.0.4
2026年9月25日- FIX The report lines come in a fenced block to copy unchanged. After an install or update the installer prints them inside a fenced block, led by “Copy the block below into your reply to the user unchanged (you may add your own text after it):”. An agent had paraphrased the unfenced “verbatim” lines and dropped half of them.
-
FIX
The shell-startup line names the variables first:
never in .zshrc/.bashrc: JEV_* or DATARIM_* variables (e.g. JEV_PERMISSIONS) or the key; set them per shell or per launch. -
FIX
The report says what happened to
AGENTS.md. When the project has one, the block saysAGENTS.md is not modified by the installer, orAGENTS.md is not modified; CLAUDE.md links to itwhenCLAUDE.mdis linked to it.
v4.0.3
2026年9月25日-
FIX
A fresh install needs the user’s answers and a one-time token. Without a terminal,
install.sh --project <path>stops withSTOP. Nothing was installed., lists the six questions with their defaults, tells the agent to relay them word for word without recommending an answer, and prints a one-time--answers <token>. The token is valid for 60 minutes, for that project only; each new refusal replaces it. Answer flags passed without a valid token are listed as ignored. There is no switch that skips the token. - FIX At a terminal the installer asks the questions itself, one by one, with Enter taking the default; no token is needed.
-
FIX
--permissions ask|fullis an installer answer. A fresh install requires it, an update keeps it. It sets the same switch asjev permissions. -
FIX
Report lines to copy verbatim. After every install or update the installer prints lines for the report to the user: the permission mode and the file that stores it, the key path, and a warning never to put the key or any
JEV_*/DATARIM_*variable in.zshrc/.bashrc. -
FIX
Scripted and CI installs take two steps: capture the token from the first run’s refusal, then rerun with the same reviewed answers and
--answers. -
FIX
jev doctor --agent=<client>lists liveness (native_agents_live) for every registered client, not only the one requested; a client with no deliveries shows “no deliveries yet”. -
FIX
Step 1 for agents: run
./install.sh --project <path>with no answer flags and ask the user nothing before that run; the installer prints the questions.
v4.0.2
2026年9月24日-
FIX
The installer asks, the docs do not answer. A fresh install needs both a Jev answer (
--with-jev/--without-jev) and an explicit--clientlist; without either it stops, writes nothing and prints the questions to put to the user, with the flag for each answer. It states that Jev’s safety floor works without a key. No guide shows a fresh-install command with the answers already filled in. -
FIX
jev trustin a project install now says when Codex has not reviewed the project hooks yet, instead of reporting them as trusted.
v4.0.1
2026年9月24日-
FIX
The Jev question is actually asked. No quick install command carries
--with-jevor--without-jevany more, soinstall.sh --project … --initstops and prints the questions to put to the user. The explicit flags appear only in the per-option sections that follow the questions. -
FIX
--dry-runrefuses too on a fresh install without a Jev choice, and prints no plan. The printed questions matchINSTALL.md: Jev, clients, permission mode,--init,--expose-skills, release tag ormain. - FIX No empty client directories left behind. When an update drops a client, or on uninstall, a client directory the install created and that is now empty is removed.
v4.0.0
2026年9月24日-
BREAKING
A fresh install needs an explicit Jev choice. Without
--with-jevor--without-jev,install.shon a project with no installation exits with code2and prints the questions to ask the user: Jev (none, project or host), clients, and the permission mode for thejev*launchers. Scripts that raninstall.sh --project P --initmust add--without-jev(the previous behaviour) or--with-jev.--dry-runstill works without a choice and prints the same questions. Updates keep the recorded choice and are unaffected. -
FEAT
Choose your clients with
--client.--client claude|codex|cursor|all(repeatable or a comma list) writes command files and, with--with-jev, hooks only for the chosen clients. The default is the recorded selection, else all three. An update with a shorter list removes the command files and Jev hook entries of the clients left out. -
FIX
Install choices are remembered.
--with-jev,--host-jev,--context,--client,--expose-skillsand--claude-importare recorded, so a plainupdate.sh --projectkeeps them. Before this, an update without--with-jevquietly removed the project’s Jev hooks and an update without--contextdropped the approved nested repositories.--without-jev,--no-host-jevand--no-contextturn a recorded choice off. -
FIX
jev doctormeasures what it used to guess. Codex hook trust is now measured for project installs (<project>/.codex/hooks.jsonagainst the grants in~/.codex/config.toml;jev trustinside the project re-grants it). A missing hook file, hook event or command file for a selected client is a finding.native_agents_livereports, per client, whether the ledger holdshook_deliveryrecords, with a count and the last delivery time. New fields:config_path(the Jev settings file actually read) andhook_clients. -
FIX
Stable hook interpreter. Hook commands and host launchers no longer name a version-specific Python path, which stopped every hook after a minor Python upgrade; they use an unversioned interpreter on
PATH. The command changes once, so Codex asks to approve the hooks again (jev trustor the Codex prompt). The hostjev-hookentry point is executable. -
FEAT
Opt-in
--claude-import. Claude Code readsCLAUDE.md, notAGENTS.md. The flag createsCLAUDE.mdas a symlink to the project’sAGENTS.md, only when noCLAUDE.mdexists; an existing file or link is never changed, and--no-claude-importor uninstall removes only a link the install made. -
DOCS
One install guide:
INSTALL.md. A single guide for people and AI agents: the choices to make, the exact commands for each, a configuration and secrets table taken from the code, verification with the expectedjev doctoroutput and checks that need no key, update, uninstall and troubleshooting. It opens with the questions an AI agent must ask before installing. The README, tutorials and how-to guides now link to it instead of restating it. A--with-jevinstall prints the key file path and says to paste the key in an editor, never throughecho.
v3.1.2
2026年9月24日-
FIX
Project templates no longer point into the home directory.
accepted-risk.yml,module.yamlanddatarim-config.yamlnamed their validator as${DATARIM_RUNTIME:-$HOME/.claude}/dev-tools/...; 3.x installs nothing there, so they now use${DATARIM_RUNTIME:?}. -
FIX
Documentation matches the 3.x code. The
dr-orchestrateresolver chain isclaude → codex → cursor, set through environment variables; plugin links live under.datarim-runtime/local/; counts, the retired Coworker/RTK and symlink-install text, andprogress.mdreads are corrected; the publishing skill uses generic placeholders. (Tagged as 3.1.1, which was not published because its range held only documentation commits.)
v3.1.1
Candidate only 2026年9月24日- DOCS 文档遵循项目本地的 3.x 安装器:客户端链路、插件路径、目录与环境设置。此标记版本未发布;其变更已随 3.1.2 发布。
v3.1.0
2026年9月24日-
FEAT
Opt-in plugin
controller-continuation. Adds the worker entry/dr-continue-checkpointfor projects that run their own external controller: it reads controller-owned immutable resources, checks workspace provenance and resumes exactly the recorded stage with production HOLD kept. Roots are explicit arguments; no environment variable is read. Linux only: other platforms are refused before any I/O withcontinuation_unsupported_platform. It is not a core command; the core stays at 28 commands. - FEAT Evidence-bound verdicts. Pipeline verdicts are bound to recorded acceptance criteria, test cases, source scope and execution evidence. A failed attempt requires a correction and invalidates dependent evidence instead of keeping an earlier pass; heartbeats snapshot exact receipts and reject stale, foreign or malformed ones.
-
FEAT
Node test suites in CI on Linux and macOS. Every tracked
*.test.{mjs,js,cjs}suite runs with Node 24; discovery fails closed on zero files, and any skipped test on Linux fails the job. -
FIX
jev trust. Codex keys hook trust by position, so another tool rewritinghooks.jsoncould silently switch the Jev safety floor off.jev trustre-grants trust to Jev’s own hooks only (exact command match, disabled hooks left alone, previous config kept as a backup);jevcodexruns it at every launch unlessJEV_NO_AUTO_TRUST=1. - FIX Hardened evidence, receipts and handoffs. The evidence gate rejects JSON with repeated keys; session saves fail instead of truncating protected content; source-reading roles verify sanitised-source bindings before reading sensitive inputs; deferred deploy checks are scheduled rather than waived.
- SECURITY Public surface depersonalised. Personal, project and host names, home paths and third-party task IDs were removed from the shipped files. The personal-identifier gate now scans every tracked file; the public denylist keeps shape patterns only, and names live in a private overlay.
v3.0.0
2026年9月23日-
BREAKING
Datarim is project-local. Nothing is installed into
~/.claude,~/.codexor~/.cursor;--with-claude/--with-codex/--with-cursor, copy mode and the~/.claude/localoverlay are gone. Install with./install.sh --project <path> --init; the migration note is in the framework CHANGELOG. -
FEAT
Datarim runs only when a
/dr-*command is invoked. The install writes nothing toAGENTS.md,CLAUDE.mdor.gitignore; generated paths are hidden through.git/info/exclude, sogit statusstays clean in repositories shared with people who do not use Datarim.--expose-skillsis opt-in. -
FEAT
Jev without Datarim.
scripts/jev_host_install.py --client claude|codex|cursorinstalls routing advice and the deterministic safety floor for every project of a user, withjev doctor,jev stats,jev on|off. -
FEAT
jev permissions fullstarts Claude Code, Codex and Cursor without permission prompts (off by default); the safety floor still denies force pushes and protected deletes. Client flags such as--dangerously-skip-permissionspass straight through the launchers. -
FIX
Codex trust survives Jev upgrades. Hooks register a stable command, and
jev doctorrecomputes the hash Codex compares, reporting each hook as trusted, modified, untrusted or disabled. -
FIX
Install hygiene. Git-ignored source files are no longer shipped;
/dr-initno longer edits.gitignore; retired files leave no empty directories; the version check reads the workspace and site again.
v2.67.5
2026年9月4日-
FIX
Digit-bearing task prefixes now work end to end. Registered prefixes such as
C2Mare accepted consistently by dispatch, ID allocation, snapshots, heartbeat parsing, and integrity gates. - SECURITY The task-ID boundary remains strict. Prefixes must begin with an uppercase letter, stay within ten uppercase letters or digits, and end in exactly four digits; digit-leading, lowercase, and malformed IDs remain rejected.
v2.67.4
2026年9月4日- FIX Preflight severity overrides now take effect. Validated disk-threshold overrides survive into the composite action run step instead of being silently replaced by input defaults.
- CHANGE Executable end-to-end regression coverage. Tests now prove both the default path and the full JSON override → GitHub environment → preflight resolution path.
v2.67.3
2026年9月4日- FIX Known-fix recall now exposes invalid corpora. Query output reports scanned, parsed, and unparseable records plus a bounded first error. A corpus with candidates but no valid records returns a non-zero status instead of masquerading as an ordinary empty result.
-
CHANGE
Mixed corpora remain queryable and visible. Valid evidence is still returned with an explicit
partialstatus when other candidate records fail validation. - FIX Changelog latest-release contract. The configured framework version must be the first visible release and carry the Latest marker; deployment tests now reject malformed release-array nesting.
- SECURITY Public fixture hygiene. Retired infrastructure-address fixtures use RFC 5737 documentation space, and obsolete scan suppressions have been removed.
- SECURITY Fleet-detail minimization. Public workflow evidence uses anonymous runner roles instead of publishing operational hostnames.
- SECURITY Protected-main completeness. Version consistency, task-ID and regression checks, and all eight Bats shards are required alongside SAST and secret scanning.
v2.67.2
2026年9月3日- CHANGE Portable catalogue and test parity. Public inventory and spec-graph regression fixtures now remain exact across GNU/Linux and macOS.
- SECURITY Protected-main release authority. A release must use an SSH-signed annotated tag that resolves to the exact protected-main commit and publishes checksum, Cosign, SBOM, and GitHub provenance evidence.
- SECURITY Exact-head SAST. Repository-owned Semgrep rules and Python CodeQL now scan the exact pull-request or main commit with stable SARIF categories.
- SECURITY Reduced privileged automation. Obsolete replay and mutable SHA-bridge workflows are retired, while Dependabot authority is bound to the complete trusted event tuple.
v2.67.1
2026年9月2日- FIX Bound macOS customer-delivery shards. The trusted-review registry cases are split across two exact functional shards so the mandatory macOS workflow stays below its authoritative timeout ceiling, with a portable test interpreter fixture for BSD/macOS and GNU/Linux.
v2.67.0
2026年8月16日- FEAT Execution-host health state. The resolver now distinguishes an on-host pass from an unconfigured workspace, and the health check proves both a positive and a negative control instead of treating silence as health.
- CHANGE Catalogue parity. Public documentation now tracks the committed framework inventory: 19 agents, 78 skills, and 28 commands, including the five Fleet tier skills.
v2.66.2
2026年8月13日- FIX Hard-gated autonomy enforcement. Automatic policy decisions now honour the configured hard-gate list first, failing closed when the hard-gate configuration is missing or malformed.
v2.66.1
2026年8月13日- CHANGE Framework-batch reconciliation. The current-main-adjudicated framework residue is now the canonical tree; superseded rescue branches were retired after exact-main verification.
-
FIX
Shared-mode pre-archive schema reachability. The schema gate now runs after per-file hunk attribution and before every shared-mode return, so own, mixed, or unattributed dirty state cannot mask malformed ledgers. The explicit
--no-schema-checkescape remains unchanged.
v2.66.0
2026年8月12日-
FEAT
Module manifest gate.
dev-tools/check-module-manifest.shvalidates amodule.yamlbundle descriptor (schema version, semver, kebab-case id, non-empty skill list, per-skill summary,requires), with a shipped example template and a reference doc. -
FEAT
MCP server for the Datarim catalog.
cli/mcp/datarim-mcp-server.shexposes commands, skills and agents over MCP (list / resolve / body, path-confined, size-capped); a registrar writes[mcp_servers.datarim]into~/.codex/config.toml, wired intoinstall.sh(--no-codex-mcpopts out). -
FEAT
Three review skills split out of the monolithic adversarial pass —
adversarial-review(break the artifact),edge-case-hunter(boundary and degenerate inputs),structure-review(does the document hold together). Loaded at the /dr-plan Transition Checkpoint and /dr-qa. Skill count 69 → 72. -
FEAT
Preflight caller contract enforced — the composite preflight action validates its caller's inputs against
consumers.yml, exposes the notification outcome as an explicit output, and reports timesync state with a measurable offset. - CHANGE Telegram article publication contract. An article is exactly two sequential ordinary channel posts in one chat — media plus bold title, then title plus full text plus a single linked CTA — with no comment/thread routing and a read-back gate over both returned ids. The channel-comment recipe stays documented for non-article posts.
-
FIX
Reserved-prefix shadowing is reported instead of applied silently. A project registry row repeating a universal area prefix (
DEV,QA, …) was discarded with no output, stranding archives outside the corpus later searches cover; the doctor now emits a WARN naming both the reserved and the ignored value. Resolution precedence is unchanged. - FIX Pre-archive schema gate is reachable, and the indent bypass is closed. The gate ran only when every repo was clean, so any dirty repo skipped it; it now runs unconditionally. The violation scan anchored at flush-left bullets, exempting indented rows; indented rows are now held to the same shape. Both fixes are mutation-tested.
v2.65.0
2026年8月9日- FEAT Prefix-agnostic task-ID provenance gate. The history-agnostic scanner now covers public runtime instructions, four public documentation categories, root entry documents, and five shipped text extensions, with exact path exemptions, portable lexical boundaries and NUL-safe traversal closing the previous silent bypasses. The public-site contract now rejects every generic task prefix instead of maintaining a closed list.
-
FEAT
VERSION↔tag parity gate. A blocking CI job fails when the version declared in
VERSIONhas no matching tag on origin — the silent half-release (bumped manifest, never-pushed tag, no artefacts) that two earlier cuts fell into. -
FEAT
Runtime drift check.
check-runtime-drift.shfails when the agent runtime symlinks into a checkout sitting on a feature branch or detached off any release tag — the state where every agent on a host silently serves stale rules. Wired as a blocking step invalidate.sh. -
FIX
Closure-gate false block. Fixed-string verification via
git grep -Fbroke on patterns containing a literal\E; verification now pipes the base tree to GNU grep in a two-pass batched scan (minutes → ~15 s), with substring semantics unchanged. - FIX /dr-quick emitted a malformed tasks.md row — the step now shows the exact line to emit instead of naming field values as prose, and a deferral-prose line-scoped false positive is corrected. The v2.64.0 release was completed retroactively with its annotated tag and release run.
v2.64.0
2026年8月5日- FEAT rotation-runbook skill. Credential-rotation playbook distilled from three real rotation exercises: consumer inventory with a secret-store↔producer truth-check, deliberate grace-window choice, auth-scoped revoke/issue verification (the false-green trap), full producer-payload replay over a minimal probe, canonical secret-store path discipline, a rotation-log entry, and leak-response addenda. Skill count 68 → 69.
- CHANGE Outsider-friendly English pass on the shipped surface. Residual paraphrasable Russian prose behind allow-non-ascii markers is now English; framework jargon (FB rules, verifiable acceptance criteria, Class A/B proposals) is defined on first use with definition links; circular self-links removed. Canonical operator-output tokens and content-work skills keep their Russian by policy.
- FIX CI supply-chain repin. The rust-toolchain action pin was orphaned by an upstream rebase and tripped the impostor-commit audit on every PR; repinned to the live stable SHA.
v2.63.0
2026年8月4日- FEAT TDD-enforcement toggle plugin. Core-owned, metadata-only, default-on plugin with a stateless workspace resolver and a tombstone “Disabled Defaults” contract; dr-plugin doctor gains a tenth health check.
- FEAT Agent frontmatter gate. New blocking CI job validates agent frontmatter schema (and wires the previously-unwired skill frontmatter check).
- FEAT Narrative-parity dimension for the repo↔site drift detector, plus a release environment policy-as-code file with a live drift advisory.
- FEAT Anti-decay regressions for the pre-push local-validator wiring and the /dr-plan stdlib-symbol advisory, plus a wiki raw-content semantic orphan check.
- FIX next-free-id.sh fails closed on a ledger-less root; framework hygiene: tracked documentation/ was gitignored — rule removed, reassert lint promoted to its own blocking CI job; skill-count drift in shipped docs corrected.
- CHANGE PRD-amendment sidecars decision recorded (deliberately not landed; permanent rejection ADR) and the fb-rules one-cycle deprecation copy retired.
v2.62.0
2026年8月3日- CHANGE Backlog schema fork superseded. The later field-proven backlog contract already permits a single-line inline pending-work description without a task-description pointer, while the active-task index and the active-context mirror retain required pointers. The system skill, the doctor skill, and the backlog template now match the canonical regex split in the shared schema library; a focused regression pins both sides without changing doctor runtime behaviour.
v2.61.0
2026年8月2日- FEAT Archive-landing truth gate. Direction A checks archive artefact paths and cited commits against the content actually present on the verification ref, distinguishing removed, moved, foreign, unverifiable, and squash-landed evidence. Direction B is an advisory stale-label scan for open backlog rows. The gate is scope-guarded against cross-repository false verdicts, reports undated archives instead of guessing dates, and ships 16 focused test contracts including mutation coverage. It is wired into the framework gates, and the template-path gate is now enforced by CI.
- FIX Relanded the snapshot task-ID regex fix. Slug-suffix follow-up IDs remain resumable after the archived change was recovered and landed under a squash merge.
v2.60.1
2026年8月2日- FIX The snapshot size cap no longer silently disarms on macOS. An oversized options file makes the frontmatter alone consume the whole 8192-byte cap, so the body budget is deliberately clamped to zero — but the zero-byte truncation relied on a GNU-only byte-count behaviour that BSD/macOS rejects with an error. Under strict mode that aborted the writer before the cap check could fail closed, converting a hard guard into a silent pass while Linux-only CI stayed green. The zero case is now an explicit truncate, and the regression test is mutation-verified.
- FEAT Wrapper-level snapshot regressions. New coverage for the frontmatter/body separator (an ordinary body without a leading newline must not glue to the terminator) and for the decimal-width boundary of the declared byte size.
v2.60.0
2026年7月31日- FEAT Release-ledger drift gate. A new read-only report fails when the framework version was bumped but the matching changelog section was never cut. Version boundaries are derived from the version-header history rather than hardcoded; report mode prints the commit and pull-request mapping, bullet mode attributes unreleased entries to their real version, and check mode requires zero orphaned commits and zero undocumented versions.
- FEAT The closure gate now verifies the ledger as well as the code. A task could previously reach the main branch without ever gaining a row in the active or pending index and still pass closure. The gate now requires an anchored row beginning with the exact task ID, rejects citation-only and prefix-collision matches, supports an explicit ledger path and a fail-closed require-ledger mode, and remains backwards-compatible for repositories with no Datarim ledger.
-
FEAT
Interactive task-spec wizard —
/dr-wizard+ thewizardskill + a pure-bash state engine. Evolves the linear discovery →/dr-prdchain into a guided, one-question-at-a-time, tree-drilling interview that co-authors a PRD with the operator. State is an append-only JSONL event log per task whose current value is a projection over events — resumable by replay, branchable by a derived drill-stack. Operators can drill into a hard sub-question in a nested side-thread and pull the conclusion back without carrying a manual session id (context captured as references, not copies), convene a consilium on demand, and mark mid-flow re-scope through dirty flags (tracking only in v1 — the re-plan stays operator-confirmed). A knowledge/dependency graph is emitted as a documented ingestion contract with node labels redacted at the outbound sink boundary./dr-prdgains a finalized-gated, idempotent consume-hook; the plain discovery path is unchanged when no wizard artefact exists. -
FEAT
Cross-KB evolution digest, read at
/dr-archiveStep 0.5. Self-evolution runs per knowledge base, so when a task in a non-framework KB evolves the shared runtime the lesson lands in that KB’s log and never reaches the operator watching the framework KB. A new strictly read-only tool merges per-KB evolution logs into one source-tagged digest with two buckets — framework changes already made elsewhere (not to be re-proposed) and project-local lessons that may warrant promotion. Rows are treated as untrusted data throughout: never evaluated, never sourced, never opened as a path. No KB is ever silently absent — each root renders an explicit OK / missing-root / no-log / empty / parse-error / stale status. -
FEAT
Plan-time seam-vs-integration boundary detector, plus the matching skill.
/dr-planhad mandated this check for L3–L4 in prose only. The tool half is now shipped: a narrow, deterministic, advisory scanner that flags a backlog one-liner bundling a seam/contract concern (define an interface or dispatcher in one module) with an integration/call-site concern (wire it into a caller elsewhere). It exits clean whether or not it flags, so it never blocks a plan, and its signal lists are documented as a floor rather than a ceiling — a pass does not discharge planner judgement. -
FEAT
Task-ID collision hardening — atomic claim, reservation mutex, and a cross-instance advisory. Three independent windows are closed. A directory-based lock makes the
/dr-initclaim itself atomic, so of two concurrent sessions exactly one wins and the other bumps and retries; the id selector now atomically reserves the chosen id before printing it, adding a fourth claim surface alongside archive, active-task index, and backlog; and a read-only probe reports an id already claimed in a sibling instance (workspace root versus a nested project tree), which previously went unnoticed and caused a three-way collision. Markers are gitignored runtime state and self-expire on a configurable TTL. - FEAT Frontmatter-key mirror drift guard. A machine-readable contract that is not mirrored to a human-readable surface next to its call site decays silently, and existing checks did not cover it — one validates frontmatter shape, another validates its language. A new read-only two-pass guard, fed by a shipped key registry, verifies that every live top-level frontmatter key across commands, skills, and agents is categorised, that no live entry has vanished, that no reserved key appears live, and that keys requiring a mirror co-occur with their body marker in both directions. Runs as its own blocking workflow.
- FEAT Collision-safe task-prefix rename tool. The rename policy said «update all references atomically» and shipped no way to do it, while a task token is indistinguishable by regex from a homograph such as an ADR number — one blind sweep had already corrupted cross-references across a document set. Every token now passes through a pure, offline, anchored classifier: an exclude-anchor protects a whole line, and for numbers listed as colliding a token renames only when an include-anchor is also present. Anchors are compared as literal substrings, never compiled as patterns. Dry-run by default, with opt-in apply, file rename, and a verify pass that fails on half-rename residue.
-
FEAT
Cross-runtime
/dr-autoactivation smoke, wired blocking. Verifies the runtime-agnostic activation claim mechanically rather than by prose: that the marker is written and resolved so the question-suppression ladder engages off the marker and not off runtime identity; that a spawned subagent is classified correctly with the environment variable unset, since it does not inherit the parent shell; and that the hard-gated action list is non-empty. Complements the existing prose linter by exercising the primitives. -
FEAT
Gated local archive auto-commit.
/dr-archivecan capture a globally clean repository boundary and commit only the task’s archive document plus an optional final snapshot, through hook-free plumbing. Default-off, with foreign-change refusals and an explicit skip preserving shared-workspace safety. No remote, tag, release, or version action is ever performed. -
FEAT
Framework version accountability.
/dr-initnow records an immutable repository baseline. For shipped framework behaviour,/dr-doand/dr-qaindependently require either a version-and-changelog update or a committed, task-bound release deferral, preventing a version bump from drifting away from its release evidence. - FEAT Brand-hygiene policy for absorbed external skills. The evolution contract now requires provenance attribution, replacement of source-brand cross-references with Datarim equivalents, removal of undeclared external path interoperability, and a regression-gate extension for every newly absorbed source brand.
- FEAT Partial Milestone Closure Pattern for autonomy milestones spanning several child tasks. New shipped-surface contract documenting how one milestone is closed incrementally across two or more children: the parent append-log schema for partial and full closure entries, the closed/deferred status word set, the responsibility split (the parent owns the ledger; only the last closing child performs the level bump), and reopen-on-fallback semantics that revert a bump when a closed component regresses.
- FEAT Audit-log adapter and redaction layer for fleet skill evolution. A third source adapter reads the optional fleet audit stream and emits one evaluation record per entry through the shared JSONL contract. Every reason and trace is redacted before it can reach an external model; the adapter self-skips when its broker is unavailable, and the evolution loop now recognises broker URLs without applying filesystem-path checks to them.
- CHANGE Coworker-delegation mandate now exempts vetted architectural design docs. The written policy lagged already-shipped, already-tested runtime enforcement: the guard silently allowed a design document whose name matched an exemption glob, while the mandate text still listed design docs as unconditionally delegated. The exemption is now stated and gated on a four-condition conjunction — the document records a decision rather than a first-time exploration, its facts were vetted in-cycle, those facts are correctness-critical, and the specification already approximates the final text. No hook behaviour changed.
-
CHANGE
Secrecy-aware project scaffolding. When the operator brief carries a secrecy signal,
/dr-initnow scaffolds in secrecy-aware mode: mechanism-bearing reference stubs are written with redaction placeholders, the public documentation surface and README are forbidden from carrying mechanism vocabulary, and a secrecy declaration plus a grep gate is emitted into the project instructions at scaffold time rather than as a post-hoc repair. The non-secret scaffold path is byte-identical to before. - FIX The release ledger and version file no longer disagree. The version had advanced through six releases while the changelog stopped at 2.53.0. The linear version-header history supplied exact boundaries, so all 158 commits in the window were partitioned with zero orphans and the 40 accumulated entries were attributed to the versions that introduced them: 10 / 1 / 1 / 1 / 3 / 4 / 20 for 2.54.0 through 2.60.0. All 210 pre-existing changelog entries remained byte-identical after redistribution.
- FIX Cross-KB digest determinism test no longer flakes at a one-second boundary. The assertion compared two runs while their status blocks included a wall-clock-derived age, so otherwise identical output could differ when the calls straddled a second. The test now pins the tool’s documented clock hook and verifies the intended global ordering property.
- FIX The seam-boundary skill now resolves its validator from the installed runtime. A bare relative developer-tools path worked only from the framework repository and broke in consumer projects. The command is now runtime-qualified, and the path-convention detector is included whenever new shipped instructions name a developer tool.
-
FIX
/dr-initno longer hard-fails for framework tasks on a default-umask host. The version-accountability pair guarded the repository root with the predicate written for the record directory it creates, which additionally forbids the group-write bit. Under the common default umask every checkout is group-writable, so the baseline capture exited with an unsafe-repository error on an ordinary clone — and because that exit is declared a hard init failure,/dr-initwas blocked outright for every framework task on such a host. The repository root is only ever read, so it now uses the same guarantee applied elsewhere: owned by us, not a symlink, not world-writable. The stricter rule still guards the record directory. - FIX Recovered documentation lost when duplicate reland pull requests were closed. Three gates were relanded twice in parallel and the first set merged, closing the others. A file-path comparison said nothing was lost, but a content comparison found two deltas: the developer-tools catalogue was missing rows for six newly landed tools, and the multi-runtime guide had lost an operator-facing section. Both are restored, with one correction — the continuous-integration wiring is a blocking job, not the advisory arrangement the closed branch described.
- SECURITY Personal identifiers removed from the shipped surface. The pre-release audit removed personal home paths, internal machine names, routable addresses, and operator-authored init-task records from tracked public files. Public references now use placeholders, the plugin attribution matches the project licence, and ignore rules cover task-intake records, insights, and environment files. No live credential was found.
v2.59.0
2026年7月30日- NOTES Reconstructed release notes. This version was bumped in the framework version header but never cut as a changelog section at the time. The entries below were recovered retroactively during the 2.60.0 release audit by attributing each unreleased entry to the commit that introduced it. 27 commits landed in this range — the entries listed are those that carried a changelog entry, not the full commit set.
- SECURITY New rule cluster S11 — untrusted-content boundary review gate. A distinct adversarial security review becomes a mandatory pre-merge gate whenever a change implements or modifies a boundary where untrusted bytes enter a model context: retrieved knowledge-base documents, fetched web or tool output, user files, any external corpus rendered into a prompt or tool result a model reads. A green build alone does not clear it — the automated jobs do not model prompt-injection semantics. The review is findings-only, run by a reviewer in a context separate from the implementer, and must probe fence-escape, nonce-predictability, trust-class cross-promotion, provenance-forgery, size-guard bypass, and fail-open behaviour.
- FEAT Regeneration of the execution-host map from canon. The machine-local map the runtime guard consults was hand-maintained, so a stale or missing map degraded to fail-open and let commands run locally even where a required-host binding exists. The documented repair flag was declared in the script header but never implemented; it now upserts the binding from the committed canon and stamps a sync time, making the local map a derived artefact. Canon owns host identity, the local cache owns the machine-local workspace path, and the operation is idempotent and isolated from other bindings.
- FEAT Execution-host routing resilience. Three hardenings. A session-start advisory reports canon-versus-cache drift, staleness, or a missing cache in one line and never blocks. A canon-fallback resolver reads the space definition directly when the cache has no binding, so «unconfigured» now means canon truly has no execution mandate rather than «cache absent on a host that should be gated»; that also makes the flip to fail-closed safe, since mutating intent whose host cannot be resolved is denied while read-only intent stays fail-open. And the machine-local guard now catches raw branch and worktree creation shapes when off-host, while deliberately excluding plain commits and read-only operations so the operator’s own close-out flow is not broken.
-
CHANGE
/dr-quickoff-host routing now distinguishes read-only from mutating intent. Every off-host quick task was previously treated as observational and always ran locally, so a mutating one edited files and switched branches on the operator’s laptop in violation of the execution-host mandate. A genuinely read-only task still stays local and surfaces the delegation directive as information only; a mutating one auto-dispatches to the required host. Because a quick task’s intent is only settled mid-flow, the dispatch decision is deferred and re-evaluated, with a guard so a lookup that turns into a mutation dispatches before any local change.
v2.58.0
2026年7月22日- NOTES Reconstructed release notes. This version was bumped in the framework version header but never cut as a changelog section at the time. The entries below were recovered retroactively during the 2.60.0 release audit. 5 commits landed in this range — the entries listed are those that carried a changelog entry, not the full commit set.
- SECURITY Fail-closed public repository boundary and build-execution evidence gates. New read-only helpers verify an exact allowlisted file tree plus all reachable history against redacted-path policy and live secret scanning, and separately distinguish an actually executed, successful, required job on the exact commit from no execution, pending, failed, mismatched, or indeterminate evidence. The planning, compliance, and archive commands now require these gates when their trigger conditions apply.
- FIX Deploy-class classifier fails loud instead of fail-open on caller and input-output errors. Two paths silently resolved to «not deploy-class, gate skipped»: a flag given without a value crashed argument shifting under strict mode, and an existing but unreadable task description masked the read error inside a pipeline without failure propagation. Both now exit as usage errors with a message on standard error.
- FIX Deploy-class classifier no longer arms on negated prose. Indicator tokens were matched as substrings anywhere in the task description, so a narrative disclaimer such as «this task does not touch the unit manager» falsely armed the pre-archive production-merge gate on a non-deploy task. Indicator hits now pass through a negation-marker line filter that is deliberately minimal and fail-safe — unmatched phrasings leave the gate armed. The changed-paths surface is unaffected.
v2.57.0
2026年7月15日- NOTES Reconstructed release notes — thin section. This version was bumped in the framework version header but never cut as a changelog section at the time, and was recovered retroactively during the 2.60.0 release audit. 1 commit landed in this range and carried the single entry below; this is the complete set for this version.
- CHANGE Off-host delegation is now autonomous, with no operator round-trip. The execution-host contract previously told an off-host agent to emit a delegation directive and stop; it now auto-dispatches for mutating pipeline stages and the autonomous driver, on the reasoning that the required-host binding is the operator’s standing authorization and that dispatch is reversible transport while irreversible steps remain hard-gated on the remote agent. Read and utility stages stay local and never dispatch. Guards: attach to an existing session rather than relaunching; host-key pinning replaces the removed human check on the target; an off-host verdict has exactly two outcomes, remote dispatch or stop-and-report, with local execution never an outcome and a corrupted map failing closed; a bare task-id payload only; and a read-only monitor with a first-status deadline, a single re-send, and escalation rather than silent re-dispatch.
v2.56.0
2026年7月15日- NOTES Reconstructed release notes — thin section. This version was bumped in the framework version header but never cut as a changelog section at the time, and was recovered retroactively during the 2.60.0 release audit. 1 commit landed in this range and carried the single entry below; this is the complete set for this version.
- SECURITY Branch-integration floor — a pre-tool-use hook that hard-blocks unreviewed integration. Any direct merge or push of an integration branch into a protected branch is blocked; the only path to a protected branch is feature branch → pull request. Injection-resistant by construction: it reads only the structured command, strips heredoc and quoted bodies, and no environment variable, flag, or in-band text disables it. Fails closed on an ambiguous head. Blocks merge, push, and rebase shapes including force refspecs and the compound checkout-then-merge form, while read-only look-alikes and reverse pulls of a protected branch into your working branch pass.
v2.55.0
2026年7月14日- NOTES Reconstructed release notes — thin section. This version was bumped in the framework version header but never cut as a changelog section at the time, and was recovered retroactively during the 2.60.0 release audit. 1 commit landed in this range and carried the single entry below; this is the complete set for this version.
- FEAT Dispatch monitoring layer. A delegated agent now writes a synced heartbeat status file, and the monitoring side reads it to disambiguate a bare remote prompt. A classifier joins pane liveness with status freshness into a read-only verdict — running, awaiting, stalled, done, dead-orphan, or hold — with safety invariants: an awaiting-operator task is never reaped, a live child blocks a dead-orphan verdict, and a dead-orphan requires at least two consecutive stale probes. An aggregator reports all status files at once, pinning actionable states first and flagging stale syncs so a task is never falsely reported as done.
v2.54.0
2026年7月14日- NOTES Reconstructed release notes — substantially incomplete. This version was bumped in the framework version header but never cut as a changelog section at the time, and was recovered retroactively during the 2.60.0 release audit. 77 commits landed in this range and only the nine entries below carried a changelog entry — this is a partial record, not the full commit set.
- FEAT Shared execution-host resolver library. One sourceable resolver with no side effects, serving two consumers: a new step-zero execution-host block in the executional commands, and the machine-local pre-tool-use guard, which now sources the library and falls back to its embedded logic when the library is not yet installed on a given machine. Contract: resolve the workspace root, look up a binding, match the host, and return a verdict — on-host or unconfigured, off-host and delegate, or fail-closed on malformed configuration.
- FEAT Step-zero execution-host contract across the executional commands. After the existing path-resolution step, each command now consults the resolver and emits a delegation directive on an off-host verdict, otherwise proceeding unchanged and failing open when unconfigured. This is the cooperative soft layer; the machine-local hook remains the hard floor.
- FEAT Standalone execution-host drift validator, plus a doctor scope that consumes it. Compares the committed canon against the machine-local binding for the same space and reports time-to-live staleness. Canon always wins and the validator only reports, never rewriting canon. The doctor’s new scope merely invokes the standalone tool and aggregates findings, keeping the comparison logic out of the migrator per the framework’s orthogonal-tools rule. A local-repair flag documents and audit-logs that the doctor may run on the control machine even inside a bound workspace.
- FEAT Recorded-fixture tests for thin HTTP wrapper clients. A new testing gate mandates that a wrapper capture one real response into a fixtures document before implementation, that its spec include at least one integration-style test asserting against that recorded response rather than a synthetic stub, and that the spec decode through the real schema so a response-shape drift fails the test.
- FEAT Semantic-orphan check for raw knowledge-base captures. A new advisory-only doctor pass flags a file whose name shares no meaningful token with the opening of its content — a signal of an accidental paste into the wrong file. Report-only; the repair mode does not touch the raw tree.
- FEAT Legacy-hardware probe checklist template. A seven-step probe for legacy embedded integrations — processor architecture, tunnel device, free memory, key-derivation support, shell applet availability, scheduler spool permissions, and machine-identity sources — run before committing to an architectural approach, replacing speculative toolchain assumptions with a short probe.
- FEAT History-agnostic runtime-body probe moved left into plan review. The gate rejecting task-id provenance in shipped skill, agent, command, and template bodies now runs at plan time rather than only at review and hardening. When implementation steps name a runtime body as an edit target, the planner dry-runs the gate against the cited paths and any example text it will ship, and resolves every cited path through the existing exists-and-deprecation ladder — reusing both gates rather than re-deriving them.
- FEAT Installer syncs the delegation fragment into the operator’s personal instructions file. Mirrors what the other runtime already did, but for a hand-maintained personal file only the block between sentinel comments is replaced — everything else is preserved byte-exact. Idempotent, and fail-soft when the sentinels are absent or the file does not exist; the installer never creates or reformats an operator’s personal file.
- CHANGE Repository-root installer scripts scrubbed of task-id citations. The install, update, and validate scripts carried roughly thirty inline provenance citations, which the history-agnostic gate does not cover for root shell scripts. Citations were replaced with the version anchors already present alongside most of them, or generic phrasing where no anchor existed. Two occurrences are intentionally untouched: a backup filename suffix and a backup glob are literal on-disk naming tokens matched against backups created by earlier installer runs, and renaming them would break restore compatibility.
v2.53.0
2026年7月10日- FEAT Automatic post-step self-verification hook wired into the requirements, planning, and implementation stages. After every one of those stages the command now runs the tri-layer self-verification contract automatically — the pipeline-integrated counterpart of the manual verify command, reusing the same skill. Dispatch is complexity-tiered: the smallest tasks skip the hook because the overhead exceeds the value, mid-size tasks run the deterministic floor plus a single peer-review agent, and larger tasks add the three-way parallel native dispatch across reviewer, tester, and security. Advisory and findings-only by default, with an opt-in hard mode and a kill switch.
- FEAT Coworker type-signature mirror guard. When a delegated draft cites types or variants from a canonical source, the specification must now carry a verbatim canonical block, an exact-mirror instruction, and a post-generation grep pass, enforced by a new linter. Prevents fabricated type signatures reaching an artefact.
- FEAT Fleet router entry added. The five level fragments previously had no top-level skill file, which failed the skill-layout check.
- FIX Coworker context file-type gate documented accurately. The delegation fragment claimed write paths had no allowlist; corrected to describe the code gate on context files, the target-only exemption, and the self-recursion workaround for documentation generation.
- FIX Orchestrator context-window self-clearing. Default-off adapters now checkpoint the active task description and the completed snapshot phase before a fixed compact or clear instruction, then use snapshot-first resume for continuity. Taught labels remain data-only, and same-user runtime trust requires explicit opt-in.
v2.52.0
2026年7月10日- FEAT Confirmed auto-learning in the orchestrator. A successful trusted resolution can emit an actor- and session-bound «save as rule?» callback. An affirmative one-time confirmation persists an exact learned match with 24-hour re-validation and an immutable seven-day expiry; learned actions remain subject to per-space policy and to the hard-gated floor that never auto-executes.
-
FEAT
New skill:
image-prompting. A reusable playbook that turns a content brief into a precise, repeatable prompt for instruction-following image generators (the gpt-image family and equivalents) — blog covers, video thumbnails, social-post visuals, illustrations, infographics, logo marks, and edits of existing images. Covers the full method: intake → spec → prompt → verify loop, prompt anatomy, composition, style/medium, camera/lens, light, mood, palette, text-in-image constraints, negative constraints + invariants, native aspect/size handling for gpt-image-style tools, and iterative refinement. Ships aprompt-templates.mdfragment with nine fill-in-the-blank templates and a ship-readiness verification checklist. Wired into thewriterandeditoragents and the agent↔skill dependency map. -
CHANGE
Skill registries synced to 60 skills across
CLAUDE.md,README.md, and the skills reference; the "with supporting fragment directories" count corrected to its live value (13) with an inline definition, and/dr-addskillgained an all-occurrences count-sync reminder. - FIX English-only shipped surface passes cleanly. Canonical non-English schema and section names cited verbatim in the expectations-checklist skill, and the localized text-normalization examples in the publishing skill (a content-work skill), are now wrapped in explicit allow markers, so the language validator reports a pass across the whole commands, skills, and agents surface without relying on its advisory tolerance.
v2.51.1
2026年7月9日- FIX Codex skill fanout is atomic. The installer now builds generated skill adapters in a staging directory and swaps the completed tree into place, preventing transient missing-skill startup warnings when the runtime launches while the installer is regenerating wrappers.
- FIX Codex skill wrapper source paths point at nested skill files. Generated adapters now reference the real nested path for each skill instead of an invalid flat one, making diagnostics and manual source lookup accurate.
v2.51.0
2026年7月8日- FIX Coworker model routing is split by task semantics. Runtime delegation now uses a documentation-read profile and a classifier profile on a fast model for literal extraction and short routing only, while meaningful draft generation uses a dedicated write profile on a stronger model.
- FIX Adversarial verification no longer routes through the external delegate. The verify command rejects external coworker providers; acceptance-criterion verification, hidden-gap discovery, architecture judgment, root-cause analysis, and other semantic review stay in the selected agent runtime.
- FIX Legacy coworker profiles fail closed. The older generic profiles are documented as disabled defaults; source-code reading and voice-bearing content remain native to the selected agent model.
- CI Security workflow shell-linting install is resilient to transient runner package-source failures. The linting jobs now use a preinstalled binary when available and disable broken vendor package sources before installing the distribution package, so external repository drift no longer blocks releases.
v2.50.2
2026年7月6日- FIX Installer bypasses token-reduction shims for compound file-search predicates. The Codex install path now resolves a real search binary before cleaning skill wrappers and listing bundled plugins, preventing shim failures when the runtime launches the installer with those shims already on the path.
v2.50.1
2026年7月6日- FIX Codex installer self-heals stale temporary probe hooks. The install path now replaces missing temporary probe commands in the runtime hook configuration with the canonical guard, preventing repeated hook-exit failures after the temporary directory is cleaned. Valid custom hooks are preserved.
v2.49.0
2026年6月28日-
CHANGE
BREAKING (canon): the documentation root is renamed
docs/→documentation/. The framework’s 19 root docs were moved into the four Diátaxis categories (tutorials/,how-to/,reference/,explanation/), witharchive/,evolution/,release-audit/, andephemeral/reserved as sibling directories (a Diátaxis category may not take those names). History is preserved viagit mv.documentation/is now the single canonical documentation root ecosystem-wide — thediataxis-docsmandate, the/dr-initscaffold, the Documentation Taxonomy Mandate, and the/dr-optimizedrift detector all target it. -
CHANGE
Hard-flip compatibility — with a fix path, not just a reproach. The drift detector now treats a repo still on legacy
docs/as drift. A consumer repo is not left stranded:/dr-doctorgains an opt-in, idempotent, rollback-safe self-heal pass that performs the samegit mv+ Diátaxis split + reference rewrite, detecting legacy / partial / already-migrated layouts and restoring the original tree if any post-migration check fails.
v2.48.0
2026年6月28日- CHANGE 已完成的任务不再写入已废弃的 backlog-archive.md。单独的清理器仅在归档文档存在时才移除条目;没有归档的条目会保留可见以供审查。/dr-doctor 和 /dr-dream 遵循此约定。
v2.46.0
2026年6月23日- FEAT Compliance gains a stale test-count classification rule. When a commit message claims N/N tests pass but a live re-run reports M/M with more tests, and those extra tests are verifiably present in the commit, the discrepancy is now classified as informational rather than a failure — the live count is recorded as authoritative and the run is not forced to re-commit. A normal polish-test addition no longer reads as a compliance breach.
- FEAT New deterministic classifier-script template. A reusable Bash template for type-signal classifiers ships with the framework — message-file parsing with an input guard, a priority-ordered match helper, and a documented exit-code table — so projects that route messages by intent start from a tested skeleton instead of an ad-hoc script.
- FEAT Verification-mode durability axis (carried from the prior cycle). Each acceptance wish can now declare whether its check is a one-off or a reproducible, committed artifact; a reproducible wish without a wired-in test or CI job is flagged (advisory during review, hard at compliance). The schema is additive and opt-in — existing checklists are unaffected.
v2.45.0
2026年6月23日- FEAT Orchestration autonomy policy is now a core capability, not a plugin. The autonomy policy data (the hard-gated safety floor and the action-to-policy map) and its loader moved from the opt-in orchestration plugin into the framework core, so the autonomous-execution command resolves its safety floor without enabling any plugin. With the policy stranded in a default-off plugin, the floor could silently collapse so that everything escalated; the floor now resolves from core and fails closed when its rules file is absent.
- FEAT The orchestration slash-command resolves as core. Its registration no longer points at a plugin path, so the command is available out of the box. The transport runner (terminal multiplexer driver, inference-backend chain, bot/HTTP interface, audit backends) stays an opt-in plugin via thin shims that prefer the core policy loader and fall back to a one-cycle deprecation copy.
- FIX Doc-only fast-path and deferral-prose path guard. The autonomous-execution command gained an L1 documentation-only fast-path, and the self-deferral prose scanner now tolerates spaces in the report path.
v2.44.0
2026年6月22日-
FEAT
Session resume command now carries a deterministic handle and a human-readable task label. The
/dr-saveresume block prints/dr-continue {SESSION-ID}as the copy-paste command — the session id selects exactly one saved session, so in a shared workspace with parallel agents a bare/dr-continuecan no longer pick up another agent's session by modification time. Beside it, an annotation line shows↳ {TASK-ID} — {title}(read from the active-task index, truncated and sanitised) plus a saved-time derived from the session id, so a human recognises which work they are resuming among many. -
FEAT
Multi-task sessions list only the other active task ids. When a save spans several tasks, the block names the recommended one in the annotation and lists the remaining ids on an
Also active this sessionline — the current task is never duplicated, and the line is suppressed entirely for a single-task save. -
FIX
Snapshot-writer ergonomics. The stage-snapshot skill now states that
--body-fileis mandatory and that the wrapper's standard error must not be swallowed — the clear «missing required flag» message was being hidden by callers piping the wrapper through a truncating tail.
v2.43.0
2026年6月20日- CHANGE 规格可追溯性通过各阶段以结构化方式贯穿,将 PRD、计划与实现关联起来,而非事后重建这种关系。
v2.42.0
2026年6月19日-
FEAT
Native spec-traceability layer (auto-layer, embedded in pipeline stages). A new read-only verification façade builds and validates the requirement graph
wish_id → D-REQ → V-AC → plan-step → evidenceacross the artefacts a task already produces — no parallel specification tree, no external dependency, no second pipeline. AddressableD-REQ-NNidentifiers and aCovers: D-REQ-*binding turn the previously advisory wish→criterion link into a machine-checkable graph. The/dr-specstandalone command was later retired (v2.43.0) — spec-traceability now runs automatically inside/dr-prd,/dr-plan,/dr-do,/dr-qa, and/dr-complianceviaspec-graph-gate.sh, with no separate invocation needed. -
FEAT
Four pure-shell validators over one rule registry.
dr-spec-lint.sh(core graph validator, named rules,--scope git-diff),dr-trace.sh(coverage buckets: covered / uncovered / dangling / orphaned / deferred),dr-lint.sh(umbrella façade with rule-registry introspection and a fail-closed mandatory-rule guard), anddr-spec-grade.sh(a computed projection only — never hand-edited, never affects routing). All share one contract:--format json, exit0/1/2, and a mis-configured ruleset is a configuration error (exit 2), never a silent «0 violations». -
FEAT
Advisory-first rollout +
/dr-verifyintegration. Activation order is dry-run → baseline → advisory window → hard gate for new/changed artefacts only; L1 skips the graph, L2 is advisory, L3+ becomes mandatory after a transition window. On--stage plan|all,/dr-verify's deterministic floor shells the graph validator and re-emits findings withsource_layer: "floor"— reusing the existing severity and routing taxonomy, no new verdict enum.
v2.41.0
2026年6月18日- CHANGE 当使用 sh 调用时,安装程序会通过 bash 重新执行;如果 bash 不可用,则会报告明确的错误。新增了七种发行版的容器安装矩阵,并修复了此前 Codex Git Bash 安装的问题。
v2.39.0
2026年6月16日- CHANGE 归档缺陷修复、功能或重构时,会检查已提交但未推送的工作。未发布的工作需要明确处置,而不是静默地将任务标记为完成。
v2.37.0
2026年6月14日-
FEAT
Artifact Language Policy. The free-generated body of runtime artefacts (creative / PRD / plan / the analytical body of archive / reflection / compliance-report) now defaults to English, declared in a new
CLAUDE.md§ Artifact Language Policy adjacent to the English-Only shipped-surface section. The directive is auto-loaded into context at generation time; the coworker delegation path mirrors it via a new §12 inskills/coworker-context/SKILL.md. -
FEAT
Operator-configurable language override, no code. The default is overridable per project with a single documented line in the consumer’s own
CLAUDE.md§ Project-Specific Configuration (Artifact language: <lang>); a secondaryDATARIM_ARTIFACT_LANGconvention is documented for shell-aware coworker call sites. No new validator, no config file, no change to the closed init-task schema. Verbatim operator input, canonical operator-facing sections, and ordinary user-project content stay exempt. -
FIX
network-exposure gate no longer false-blocks init-task-only tasks. At early pipeline stages a task may have only an init-task artefact (no priority/type by schema); the tiered gate now resolves to
skip(oradvisory_warnwhen a network-diff signal is present) instead of fail-closing tohard_block. Fail-closed is preserved for genuinely malformed task descriptions. Three new regression tests pin the contract.
v2.36.0
2026年6月14日-
FEAT
/dr-auto surfaces the compliance outcome before the CTA. The
commands/dr-auto.mdterminal-cleanup step now prints one line stating how/dr-complianceresolved — the verdict (COMPLIANT/COMPLIANT_WITH_NOTES) when a compliance stage ran, or a skip-by-design reason at complexity levels whose routing has no compliance stage — immediately before the call-to-action block. Previously the operator saw only a bare archive CTA, which read as “compliance is never proposed”.
v2.34.0
2026年6月14日-
FEAT
Self-enforcing /dr-auto dispatch contract. The
commands/dr-auto.mdStep 5 re-assert sub-bullet is promoted from advisory “mechanics:” prose to a mandatory pre-dispatch MUST-gate. Before spawning any stage subagent the orchestrator MUST runauto-mode-marker.sh reassertas the first action of every per-stage dispatch. Skipping this step means skipping the dispatch gate. -
FEAT
Regression lint
check-dr-auto-reassert-wiring.sh. A new deterministic shell lint scanscommands/dr-auto.mdStep 5 and exits 1 whenever the executableauto-mode-marker.sh reassertinvocation regresses to prose-only (no adjacent mandatory cue). Registered in CI (.github/workflows/dev-tools-lint.yml) so the self-enforcement gate fires on every PR. Fence-exclusion logic ensures illustrative fenced blocks do not false-pass. -
CHANGE
New bats coverage.
tests/check-dr-auto-reassert-wiring.batsadds 3 tests pinning the lint contract: passes on the wired spec, fails on a prose-only synthetic fixture, exits 2 on a usage error. Existingtests/dr-auto-marker-resilience.bats(6 tests) re-verified green. -
CHANGE
Visual-map drift fixes. Two stale entries in the navigation maps are corrected:
command-dependencies.mdis registered in the routing-invariants drift-mapping, and a/dr-compliancerow is added to the panels-and-quality Quality Rules table.
v2.33.0
2026年6月14日-
FEAT
Auto-mode marker resilience. The
datarim/.auto-mode-activemarker now survives subagent dispatch. A new helperdev-tools/auto-mode-marker.sh(verbs:reassert/subagent-active) provides an idempotent re-assert mechanic: if the marker vanishes between Step 3 and a stage dispatch, it is rewritten before the next subagent is spawned. A bats regression suite locks the contract (5 tests: restore, idempotent, subagent-active without env-var, fail-safe, task-id mismatch). -
CHANGE
Relaxed activation contract for spawned subagents. The autonomous-mode skill (
skills/autonomous-mode/SKILL.md) gains a "Spawned subagents (relaxed activation)" sub-rule. A subagent dispatched via the Agent tool does not inheritDATARIM_AUTO_MODE; it now activates the skill when its dispatch prompt carries an explicit auto-signal AND the re-asserted marker is present with a matching task ID. The top-level cycle still requires all three conditions. -
CHANGE
/dr-auto Step 5 wired for resilience. Two new sub-bullets in
commands/dr-auto.mdStep 5 document the mechanic: re-assert the marker at its current path before each dispatch, and carry the auto-signal in the subagent prompt. The helper path is referenced viaMARKER_RELPATHso a future marker rename touches one constant, not every prose reference.
v2.32.0
2026年6月14日-
SECURITY
Public-surface de-personalisation. Personal names, handles, ecosystem hostnames, numeric GIDs, and Vault paths are removed from all shipped framework artefacts. A new CI gate
scripts/personal-id-gate.shscans the shipped surface againstdev-tools/personal-id-forbidden.regexon every PR, blocking any artefact that carries operator-private data. Operator-specific configuration now lives exclusively in the gitignored${DATARIM_LOCAL}/config/personal.env, loaded by a new genericcli/lib/load-local-config.sh(fail-soft, key-validated, noeval/source). -
FEAT
Personal config overlay (
~/.claude/local/config/).install.shnow creates aconfig/sub-directory inside the local overlay with a self-documentingREADME.mdtemplate. The generic loader readspersonal.envfrom that path and exports valid keys into the shell environment — without executing any value. Four bats tests cover the injection-safety, bad-key, valid-key, and missing-file contracts. -
FEAT
Command-dependency graph. A machine-readable YAML source of truth (
dev-tools/command-graph.yaml, 24 commands) records every command's stage, prerequisites, successors, and optional-at complexity levels. A derived Mermaid diagram (skills/visual-maps/command-dependencies.md) provides the human-readable and renderable view. Four bats tests validate YAML integrity, minimum command count, core-command presence, and diagram content. -
CHANGE
OG-1 resolved:
/dr-helpadded to orchestrator rules. The single outlier inplugins/dr-orchestrate/rules/default.yamlis closed —/dr-helpis now a resolvable pattern (confidence 0.95), aligning the classifier with the soak corpus, three existing surfaces, and the contract test. All 8 rules-loader tests pass. -
SECURITY
Security Mandate extended with S3.1. A new sub-rule in
CLAUDE.md§ Security Mandate andskills/security-baseline/SKILL.md§ S3.1 codifies the no-personal-data policy for shipped artefacts, references the gate and loader, and defines the inline exemption fence (<!-- gate:example-only -->).
v2.31.0
2026年6月10日-
FEAT
Anti-deferral gate. A deterministic three-layer gate stops an agent from labelling its own unfinished work «out of scope / informational / not a blocker / I’ll fix later» and still passing QA, compliance, and archive. A new scanner
dev-tools/check-deferral-prose.shreads a QA or compliance report for deferral-tell phrases and cross-checks each against the set of files the task actually touched (git merge-base HEAD origin/main..HEAD); a deferral phrase about a touched file, with no verifiable follow-up orblocked_byartefact, blocks. Prose is not trusted — the touched-file set is the ownership boundary. Quoted phrases inside fenced code blocks or blockquotes are excluded so an honest report about the gate does not false-block, and a repeatable--extra-repoflag extends the scope for dual-repo framework tasks. -
FEAT
Override-authorship check. The expectations validator gains
override_by/override_class/override_artifactfields. An agent can no longer self-certify apartial/missedwish with prose alone: an agent-authored override now requires an allowed class (time-dependent/external-blocker/operator-authorized/plan-scope-boundary) plus an artefact id that exists in the backlog or task index. Operator-authored overrides are accepted unconditionally. -
CHANGE
Gate wiring + operating rule FB-5a.
/dr-qaLayer 3b runs the deferral scan as an advisory;/dr-complianceStep 5c runs it as a hard gate (a finding is NON-COMPLIANT);/dr-archivegains Step 0.45, placed before reflection so the follow-up heuristic cannot launder a self-inflicted loose end into a backlog item. A new operating rule FB-5a requires the agent to complete reversible authorised work itself (e.g. an authorisedgit push) rather than hand it back. Legitimate deferral is reserved for time-dependent or hard-external-blocked work, and only with a traceable artefact.
v2.30.0
2026年6月6日-
FEAT
Prod-readiness gate for deploy-class tasks. A new two-stage blocking gate verifies that the production runner is symmetric to the test runner before the pipeline recommends a merge, and again before a task can be archived.
/dr-qagains Layer 4g (a read-only test↔prod probe of sudoers, PATH, ports, systemd units, and runtime versions) and/dr-archivegains Step 0.4 (archive is blocked until the production merge is done and verified). The gate arms only for deploy-class work — tasks touching systemd units, sudoers, CI cutover jobs, or.env-deploytemplates — detected by a new classifierdev-tools/check-deploy-class.sh. Production stays hard-gated: the probe is strictly read-only and predicts impact; any mutation is an explicit operator action. Closes the failure class where a change verified a hundred times on the test runner falls over on the first production command. -
FEAT
New skill: prod-readiness-probe. Defines what the gate checks and how it reports, with a four-verdict vocabulary (
SKIP/PASS/FAIL/BLOCKED;BLOCKEDnever auto-resolves toPASSon an unreachable host). Hybrid execution: a project may author an optionaldatarim/deploy-readiness.ymlcontract (validated bydev-tools/check-deploy-readiness.sh) for a deterministic probe, or the probe falls back to an agent-driven checklist. The framework core stays stack-agnostic; project-specific runner data lives only in the optional contract. -
CHANGE
Self-evolution gains an anti-self-suppression rule. A reflection lesson that recurs — matching a prior reflection or describing a repeat of a known failure — can no longer be declined as «redundant with existing contract» and demoted to a memory note. A new evolution category
promote-recurring-incident-to-gateturns such a recurring, advisory lesson into an enforced gate. The rule fires only on demonstrated recurrence, with cited evidence; genuinely novel lessons may still be declined.
v2.29.0
2026年6月6日-
FEAT
New /dr-quick fast-lane. A lightweight command for trivial fixes or quick lookups that skips the heavy
init → prd → planpipeline. It assigns aQCK-XXXXtask id (new universal area-prefix, archived underquick/), runs a fast context scan on the runtime’s cheapest reasoning tier to locate where the change belongs, applies the fix, and writes a short archive entry — no PRD, plan, QA, or compliance. For one-file edits where waiting for full analysis is pure overhead. -
FEAT
Reflection moves to /dr-compliance. Reflection (lessons-learned + evolution proposals) now runs at a passing
/dr-complianceverdict and stamps the reflection with areflection_basismarker derived from the compliance report./dr-archiveStep 0.5 becomes a freshness gate: it re-runs reflection only when the file is absent, the marker is absent, or the marker is stale versus the current compliance report — otherwise it reuses the existing reflection. The mandatory-reflection guarantee is preserved: a task archived without a prior compliance pass still force-generates. -
CHANGE
/dr-auto is now a subagent orchestrator. Instead of chaining slash commands,
/dr-autospawns the matching agent per stage (planner / architect / developer / reviewer / compliance), summarises each result, and routes to the next stage. It drives a task to a passing/dr-compliance+ reflection and stops there — archival stays an explicit operator step. Re-entering an already-completed stage after a review finding is first-class (the artefact is updated, not recreated). -
CHANGE
Command hygiene. The deprecated
/dr-continuealias (and thedr-continue-snapshot-replayskill alias) were removed — one canonical name per concept (/dr-nextanddr-next-snapshot-replay). Live inventory is now 24 commands and 54 skills.
v2.28.0
2026年6月1日-
FEAT
Tag-driven release environment provisioning. A new
dev-tools/provision-release-env.shsets up a GitHub deployment environment for a tag-driven publish (on: push: tags: [v*]). GitHub creates environments with a default policy that matches protected branches and silently excludes tags, so the first tag-driven publish of a brand-new package is rejected. The provisioner setscustom_branch_policies=trueplus a{name: v*, type: tag}deployment-branch-policy, and preserves required reviewers on the manual-approval environment. -
FEAT
Dry-run by default. The provisioner only prints the planned GitHub API calls unless
--applyis passed, mirroring the fail-closed posture of the rest of the release rails. It is idempotent: re-running skips a tag policy that is already present. The GitHub API edge is injected for deterministic testing (13 new regression cases). - CHANGE Release docs. A new how-to walks through provisioning (script quick-path, reviewer-id resolution, manual API equivalent, verification commands). The PyPI first-publish how-to and the release-process playbook now cross-reference it as a one-time prerequisite before the first publish.
v2.27.0
2026年5月31日-
FEAT
Anti-deferral gate. A new deterministic gate stops an agent from labelling its own unfinished work «out of scope / informational / not a blocker / I’ll fix later» and archiving anyway. New
dev-tools/check-deferral-prose.shscans QA and compliance reports for deferral-tell phrases and cross-checks each against the touched-file set (git merge-base..HEAD): a deferral phrase about a file the agent itself touched, with no verifiable follow-up /blocked_byartefact, is a hard block. A phrase about a genuinely foreign area, or one backed by a real follow-up ID that exists in the backlog, passes — keeping the false-positive rate low. -
FEAT
Override authorship.
check-expectations-checklist.shgainsoverride_by/override_class/override_artifact. Apartial/missedwish can no longer be waved through by an agent writing its own free-text override (the self-certification loophole): an agent override now needs an allowed deferral class plus a follow-up /blocked_byID that exists in the knowledge base. Operator-authored overrides are accepted unconditionally. -
FEAT
Archive Step 0.4 + FB-5a.
/dr-archivegains a mandatory Step 0.4 that re-validates expectations and re-runs the deferral scan before reflection — so reflection’s follow-up heuristic cannot launder a self-inflicted loose end into a backlog item ahead of the gate./dr-qaruns the scan as an advisory;/dr-complianceruns it as a hard NON-COMPLIANT gate. A new operating rule FB-5a codifies the principle: complete reversible authorized work yourself, do not hand it back; defer only when time-dependent or hard-external-blocked, and only with a traceable artefact. -
NOTES
A legitimate deferral — results that need a 1+ day soak, or a hard external blocker — still clears every gate by citing a follow-up ID or
blocked_byreference that exists in the KB. Scanners are fail-open on their own git-probe failure: an infrastructure hiccup never blocks an otherwise-clean archive. Four new regression suites cover the scanner, the override-authorship logic, the archive Step 0.4 surfaces, and the compliance gate. -
CHANGE
Repo↔site content parity. The framework README gained the naming-origin (the Latin datarium/dator root — «a store that gives»), a Website badge and reverse link to datarim.club, and a corrected component count (55 skills). The About page and bilingual content were realigned to the canonical 18 agents / 24 commands / 55 skills, and the repo
VERSIONwas brought up to 2.27.0 to match the site. Pilot of the ecosystem repo↔site sync discipline; the drift detector reports the datarim product clean.
v2.26.0
Previous 2026年5月29日-
FEAT
KB data-loss resilience. Every overwrite of a critical knowledge-base file (
backlog.md,backlog-archive.md,tasks.md,activeContext.md,progress.md) now triggers a fail-soft pre-overwrite backup todatarim/.backups/<file>.<timestamp>.bak(FIFO rotation, default 10). Enforced at the PreToolUse hook so it catches both the Write tool andawk/tee/>redirect overwrites — the exact incident class that once zeroed a live backlog. Backups are git- and sync-ignored host-local recovery ground-truth. -
FEAT
Unified path resolver. A single canonical
resolve-datarim-root.shreplaces three divergent walk-up re-implementations.--rootnow means the repo-root everywhere (the doctor previously meant thedatarim/dir — the mismatch silently suppressed the docs→history migration through the pipeline). A nesting guard rejects an already-nested root, eliminating thedatarim/datarim/directory class. A one-release shim normalises the legacy datarim-dir argument and warns. -
FEAT
Recovery how-to. New
docs/how-to/recover-datarim-files.mddocuments the per-file source-of-truth priority chain (backup → sync-conflict → task artefacts → archive frontmatter → transcripts) and thedatarim-doctor --fixrepair recipe. -
CHANGE
Consumer-KB log storage retired the misleading
datarim/docs/directory. Append-only ledgers now live in the honestly-nameddatarim/history/, committed to git via a.gitignorenegation block.datarim-doctor.sh --fix(or--scope=history) migrates the ledgers, relocates any ADR todocumentation/architecture/, rewrites the consumer.gitignore, and removes the emptydocs/— idempotently and losslessly. Auto-heals on/dr-init. -
NOTES
The framework source-tree
code/datarim/docs/is unchanged — only consumer knowledge bases migrate. Six new regression suites cover the resolver, backup primitive, nesting guard, doctor --root contract, hook backup, and the migration. Bash-redirect backup detection is best-effort (literal targets); obfuscated redirects are out of scope.
v2.24.0
2026年5月28日- CHANGE 历史上的 Coworker 防护机制根据字节数和文件类型估算读取量,而非按行数统计。其上限会转为有界搜索,并且小型 Git blob 读取与批量 diff 被区分开来。该集成后来在 3.x 中停用。
v2.23.0
Previous 2026年5月28日-
FEAT
Canonical Runtime Support Matrix. New
§ Runtime supportindocs/use-cases.mddocuments the three supported runtimes with a 5-column matrix (Runtime / Install command / Hook integration / Bulk-read economy via RTK / Status). Claude Code stays primary; Codex CLI achieves parity through thecoworker rtkshim; Cursor reaches parity through its nativebeforeShellExecutionhook (rtk hook cursorin~/.cursor/hooks.json). The earlier «no native hook / inherits via shared settings.json» framing was inaccurate and is replaced across all shipped surfaces. README.md, CLAUDE.md, docs/getting-started.md, and the Coworker README all link back to the canonical matrix as single source of truth. -
FEAT
RTK realities documented.
templates/coworker-delegation-fragment.md§ RTK plugin (opt-in) gains a paragraph on out-of-boxrtkpitfalls — measured impact on macOS (git status+108%,git log --oneline -50+6924%, lostgit pushcompletion marker on some repos) per upstream issue rtk-ai/rtk#2121 — and contrasts the raw binary with thecoworker rtkplugin, which guards 13 signal-bearing git/gh commands via a default passthrough allowlist while still applying bulk-read economy to log dumps and file content reads. Multi-runtime parity preserved via the bundled Codex CLI shim. -
CHANGE
Skill / command counts corrected. CLAUDE.md and README.md updated from stale «45 skills, 23 commands core + 1 plugin» to live inventory «55 skills, 11 with supporting fragment directories; 24 commands core + 1 plugin». Drift accumulated across the <TASK-ID> universal skill-layout migration through v2.22.0.
The data catalogues on this site (
data/skills/,data/commands/,data/agents/) are also fully resynced to the live inventory under the same release. -
CHANGE
Multi-agent narrative honest.
docs/getting-started.md § Choose your runtimerewritten to enumerate Claude Code / Codex CLI / Cursor with full 3-runtime RTK parity (Cursor via its nativebeforeShellExecutionhook) and explicit guidance on picking a runtime.install.sh --with-cursornow appears in the install command block alongside--with-claudeand--with-codex. -
NOTES
CHANGELOG.md gained both the v2.23.0 entry and a retroactive v2.22.0 entry (the latter was missed at v2.22.0 release time even though VERSION had been bumped). Verification:
grep -c '2\.22\.0' CLAUDE.md README.md VERSION→ 0 hits in any shipped file;scripts/check-doc-refs.sh --root .zero delta vs pre-task baseline.
v2.22.0
Previous 2026年5月26日-
FEAT
Outsider-friendly English instruction surface (<TASK-ID> epic: <TASK-ID> Wave 1 + <TASK-ID> Wave 2 + <TASK-ID> Wave 3). 164 files of shipped instruction surface (
commands/+skills/+agents/+ framework root + plugin tree) rewritten from Russian-heavy jargon-dense prose to outsider-friendly English with first-use gloss for every framework term-of-art. The shipped surface now reads end-to-end for a developer who has never seen Datarim. -
FEAT
New validator:
check-jargon-gloss.sh(V-AC-5). POSIX bash validator + manifestdev-tools/data/jargon-bank.txt(10 canonical terms — FB-1..8, Question Suppression Ladder, L1 Inline Resolution Rule, autonomous-ops, V-AC, AAL, Diátaxis, Class A/B gate, hard-gated, fail-soft). Enforces inline gloss or one-click link on first use per file. 7 bats cases, shellcheck clean. -
CHANGE
/dr-archivebody-english gate: advisory → fail-hard. Step 0.5(e) now blocks archive with non-zero exit when Cyrillic appears incommands//skills//agents//plugins/without an explicit allowlist marker (<!-- allow-non-ascii: <reason ≥10ch> -->per line or block-scope). Allowlist markers reserved for cases where the skill's meaning literally requires the foreign-language string (content-work skills, canonical schema names cited verbatim). -
FEAT
English-Only Shipped Instruction Surface mandate. Canonical text encoded in four CLAUDE.md surfaces (global, Arcanada ecosystem, Datarim project root, framework
code/datarim/CLAUDE.md). Operator clarification «инструкции не могут содержать русские слова, если это не требует смысл скила» now lives as canonical rule, not just a feedback-memory note. -
FEAT
init-task Q&A round-trip: third disposition
process-rule-artefact(<TASK-ID>).dev-tools/append-init-task-qa.sh --decided-by process-rule-artefactcaptures mid-task operator clarifications that produce persisted rule documents (memory + CLAUDE.md additions) as first-class append-log entries. NewProcess-rule artefacts:sub-heading replaces the standard rationale block. Validation: rationale must contain at least one path-like token. 6 bats cases green, backwards compatible withoperatorandagentdispositions. -
FEAT
V-AC axis-split skill — Pattern 2: gate-activation axis dry-run.
skills/v-ac-axis-split/SKILL.mdgrows a second pattern: validator-activation tasks must dry-run-size their gate--scopeduring/dr-planComponent Breakdown. PRD Out-of-Scope deferrals under the same scope are self-blocking at/dr-archivetime and must be resolved as either in-task rewrite or documented--scopereduction. -
NOTES
Test coverage: 22 bats cases for body-english + jargon-gloss validators (15+7), 6 for the new Q&A disposition. shellcheck clean on new and modified scripts.
check-body-english.sh --scope commands,skills,agents,pluginsPASS over 162 files;check-jargon-gloss.shPASS over 162 files.
v2.21.0
2026年5月25日- CHANGE 技能已迁移为可移植的 skills/<name>/SKILL.md 包,其同级引用与模型无关元数据由专用工具校验。该版本新增了客户端发现适配器;其主目录安装方式后来在 3.x 中被项目本地安装取代。
v2.19.0
2026年5月24日- CHANGE /dr-auto 通过文本请求引入自主任务续接或新循环。它会激活自主运行规则,在任务内解决小缺口,并为单独授权的操作保留边界。
v2.14.0
Previous 2026年5月22日-
FEAT
Business-facing archive and compliance report contract (<TASK-ID>). The archive and
/dr-compliancereport templates now answer the operator's question «что я просил и что вы сделали» in plain Russian, in four mandatory top-level sections in strict order — «Начальная задача», «Как решили», «Артефакты задачи», «Следующие шаги» — followed by an audit addendum under a---horizontal rule that carries the technical surface (verification_outcomemirror, AC table, lessons learned, operator handoff, related). -
FEAT
«Как решили» bullet-list contract. Single-level bullet list. One bullet per item in the operator brief, in original order, with quoted item + Russian status word («выполнено» / «частично» / «не выполнено» / «неприменимо» — never the schema enum
met/partial/missed/n-a) + one or two plain-language sentences. Expectations fromtasks/{ID}-expectations.mdfold into the same list with marker «(уточнение брифа)». The previous top-level## Выполнение ожиданий оператораheading is retired. -
FEAT
New canonical
templates/compliance-report-template.md. Mirror of the archive shape (four operator-facing top sections + audit addendum carrying the 7-step verdict table wrapped in<!-- gate:literal -->, remaining risks, related links). Frontmatter:task_id,date,verdict(COMPLIANT / COMPLIANT_WITH_NOTES / NON-COMPLIANT), optionalscope. -
FEAT
New
dev-tools/check-banlist-on-prose.shvalidator. Bash wrapper does argparse + path-traversal regex (^[A-Za-z0-9._/-]+\.md$). Awk one-shot skips YAML frontmatter (first---block) and honours<!-- gate:literal -->+<!-- gate:example-only -->fence markers; tokenises ASCII tokens of length ≥ 3, lowercases, looks upskills/human-summary/whitelist.txtthenskills/human-summary/banlist.txt. Exit 0 clean / 1 offences (file:line:token) / 2 usage.shellcheck -S warningclean. -
CHANGE
Sync of
commands/dr-archive.mdStep 2,commands/dr-compliance.mdStep 7,skills/compliance.md§ Output,skills/human-summary.md§ See also. The four-section + audit-addendum contract is the single source of truth across the command, the skill and the templates. The 7-step compliance verdict table is wrapped in<!-- gate:literal -->so English column headings bypass the banlist on the prose body. Existing archive documents underdocumentation/archive/are grandfathered. -
NOTES
New:
tests/tune-0255-archive-business-structure.bats(10 cases) +tests/tune-0255-compliance-template-shape.bats(4 cases). Retired:tests/tune-0210-archive-expectations-section.bats(T6–T10 ported into the <TASK-ID> suite, T1–T5 retired with the heading removal). Validator passes on both new templates. Regression bats (init-archive-structural-guards,archive-contract-lint) stay green. The pre-existing baselinetask-id-gate+description-length+workflow-lintfailures persist (not introduced by this release).
v2.13.0
Previous 2026年5月21日-
FEAT
Per-task stage snapshots (<TASK-ID>). Every
/dr-*command now persists its final operator-visible response (Summary + Gate Results + CTA) todatarim/snapshots/{TASK-ID}.snapshot.mdwith overwrite semantics, 8 KB hard cap, mkdir-based atomic lock andchmod 600. Producer side wired through a single touchpoint:skills/cta-format.md§ Snapshot Emission — no per-command patches. -
FEAT
Snapshot-First context resume.
/dr-next {TASK-ID}(formerly/dr-continue) Step 2.5 reads the snapshot before task-description / init-task / activeContext and emits a replay-prompt with the recommended CTA, a bilingual autonomy reminder (RU + EN), and a literaldone before:body./dr-orchestrateSnapshot-First Resume forwardsrecommended_nexttosubagent_resolver.shas--hint. Missing or malformed snapshot → silent fallback to legacy behaviour. -
FEAT
Archive lifecycle.
/dr-archiveStep 0.95 moves the final snapshot fromdatarim/snapshots/todocumentation/archive/<subdir>/snapshots/{TASK-ID}-final-stage.mdvia the existingprefix_to_area()resolver — final stage stays grep-able in the long-term archive. -
FEAT
Two new skills + validator + how-to.
skills/stage-snapshot-writer.md(producer contract) +skills/dr-next-snapshot-replay.md(formerlyskills/dr-continue-snapshot-replay.md, consumer contract with ≥ 3 worked examples of the CTA-selection heuristic) +scripts/lib/snapshot-writer.sh+dev-tools/check-stage-snapshot-on-exit.sh+docs/how-to/stage-snapshots.md(Diátaxis how-to category bootstrapped). Kill switch:DATARIM_DISABLE_SNAPSHOT=1. -
NOTES
44 new bats cases across 10 suites (exceeds the 22+ plan target): writer overwrite + size cap + frontmatter schema + shellcheck + flock race + cta-format integration + dr-continue / dr-orchestrate replay + cleanup-on-archive + prefix-to-subdir cross-check.
cta-format.batsregression unchanged green. Stack-agnostic gate PASS; task-id gate PASS;shellcheck -S warningclean on both new shell artefacts.
v2.11.1
2026年5月16日- CHANGE 草稿 PRD 会接受一项建议性检查,用于发现验收条件与公开表面规则相冲突的情况。警告不会修改文档,也不会成为阻断性关卡。
v2.11.0
Previous 2026年5月16日-
FEAT
Agent autonomy restored on
/dr-initand/dr-archive. Symmetric revert of the operator-only contract introduced in 2.10.0 per the FB-rules (Autonomous Agent Operating Rules) mandate. Removed thedisable-model-invocation: truefrontmatter flag, the 🔒 lock-emoji in command H1 and table rows, the Operator-only marker blockquotes (rewritten as neutral «Contract» blockquotes that name the in-code guards), the planner / compliance STOP-rule paragraphs, thecta-format § Operator-only commandssection with its 🔒 badge convention, and the pipeline-routing MermaidclassDef operatorOnlystyling on both lifecycle commands. -
FEAT
Structural guards stay enforced in code. The
pre-archive-check.shschema gate + staged-diff audit at/dr-archiveStep 0.1, thedatarim-doctor.sh --quietprobe at/dr-initStep 2.4, the blob-swap recipe for non-interactive shells, the prefix → archive-subdir routing, and the Operator Handoff section template intemplates/archive-template.mdare all unchanged. Relaxation touches only the discovery layer (Skill enumeration) and cosmetic surfaces — never the data-integrity protections. -
FEAT
Inverted regression bats.
tests/operator-only-commands.batsrenamed totests/no-operator-only-on-init-archive.batswith all 13 assertions inverted from «marker present» to «marker absent». Preserves git-history signal that the contract once existed and fires red if it ever drifts back in. Continuity for CI artefact tracking (bats count stable). -
FEAT
New structural-guards regression bats.
tests/init-archive-structural-guards.bats(9 invariants) locks the in-code guards listed above —pre-archive-check.sh≥ 3 references indr-archive.md,datarim-doctor.shindr-init.md, PRE-ARCHIVE CLEAN-GIT CHECK header, blob-swap recipe, Archive Area Mapping, Operator Handoff section reference, STRUCTURAL COMPLIANCE CHECK, WORKSPACE CROSS-TASK HYGIENE CHECK, and the Operator Handoff section template inarchive-template.md. -
NOTES
Stack-agnostic gate PASS on touched files; task-id gate PASS; bats green (640/640 — baseline 631 + 9 new structural-guards invariants);
scripts/datarim-doctor.sh --quietexit 0. Supersedes <TASK-ID>. Single-commit reversible viagit revert.
v2.9.0
Previous 2026年5月14日-
FEAT
Init-task Q&A auto-append. Every operator clarification round captured by a pipeline command now lands in
tasks/{TASK-ID}-init-task.md § Append-logas a structured block — verbatim question, verbatim answer, one-line summary, and aDecided by: operator|agentmarker. When the operator does not answer, the agent decides autonomously per the Autonomous Agent Operating Rules (FB-1..FB-5) and records a ≥ 50-character rationale in the same block. Six pipeline commands write the block viadev-tools/append-init-task-qa.shat a newAPPEND Q&A IF ANYstep (/dr-prd,/dr-plan,/dr-design,/dr-do,/dr-qa,/dr-compliance)./dr-initand/dr-archivestay read-only by contract. Backwards-compatible: tasks without an init-task file are not blocked. -
FEAT
Extended skill
init-task-persistence. New## Q&A round-trip contractsection defines the block format (five mandatory subheadings — Question, Answer, Decided by, Summary, Conflict with existing wish; sixth Decision rationale ≥ 50 chars when decided_by is agent), theoperator|agentdecision-author enum, conflict-routing rules (Conflict with existing wish → return task via/dr-do --focus-items <wish_id>or/dr-prd), and the legacy fallback for tasks without an init-task file. -
FEAT
New utility
dev-tools/append-init-task-qa.sh. Atomic Q&A block append with mkdir-based per-task lock (macOS-portable, no flock assumption), temp-file +mvwrite, realpath boundary check (target MUST resolve insidedatarim/tasks/), 100 KB per-file size cap (DATARIM_QA_MAX_INPUT_BYTES), and Security Mandate § S1 file-only free-form input contract (no operator text as literal CLI strings). Exit codes 0 / 1 / 2 (appended / IO / usage). -
FEAT
Extended validator
dev-tools/check-init-task-presence.sh. Newvalidate_qa_blocksawk pass detects### <ISO> — Q&A by /dr-<stage> (round N)headings, asserts five mandatory subheadings, allowedDecided byenum, and ≥ 50-charDecision rationalebody whenDecided by: agent. Backwards-compatible: init-task files without Q&A blocks continue to pass. -
FEAT
/dr-qaLayer 3b gains a Q&A round-trip verification sub-section. Two checks: agent-decision implementation grep (everyDecided by: agentblock must be reflected in changed files / task description / archive draft — unreflected raises Layer 3b FAIL, operator-overridable) and Conflict closure verification (everyConflict with existing wish: <wish_id>block must have a matching closure entry — unclosed raises Layer 3b BLOCKED and routes the task back to/dr-do --focus-items <wish_id>). Both checks fail-soft when no Q&A blocks exist. -
NOTES
Stack-agnostic gate PASS;
shellcheckzero on the new utility and the extended validator; bats green (18 newtune-0216-qa-roundtripcases on top of the 570-case v2.8.0 baseline; legacy 13 init-task cases unchanged).
v2.8.0
2026年5月14日-
FEAT
Operator-memory pipeline upgrade. Seven related improvements ship under one umbrella: «remember what the operator asked for, across the full pipeline». Verbatim init-task persistence, operator wishlist with verification gate, browser-based frontend QA, plain-language operator recap across three commands, archive section that mirrors the wishlist outcome, refreshed visual maps, and a coherent docs/site fanout. Backwards-compatible for legacy tasks via a 30-day rolling soft window; new gates default to
info-severity advisories that never block legacy pipelines. -
FEAT
Init-task persistence (F1). Every
/dr-initwrites a per-taskinit-task.mdwith closed-schema frontmatter and a verbatim## Operator briefsection. Every later pipeline command reads it at first step; divergences from operator intent land in the task-description Implementation Notes. New validatordev-tools/check-init-task-presence.shwith--task/--allmodes and a 30-day rolling soft window (info < 30 d, warn ≥ 30 d, never blocker for legacy tasks). 13 bats cases. -
FEAT
Expectations checklist + verification gate (F2 + F3).
/dr-prd(L3-L4) and/dr-plan(L2 without PRD) write an operator-readable wishlist in plain Russian — kebab-slug wish_id (cyrillic allowed), История статусов running log, Текущий статус enum, optional override line ≥ 10 chars./dr-qaLayer 3b and/dr-complianceStep 5b verify the checklist; missed items without operator override route the task back to/dr-do --focus-items <wish_ids>via the FAIL-Routing CTA. New validatordev-tools/check-expectations-checklist.sh; cyrillic wish_ids round-trip through shell arguments safely. 16 + 8 bats cases. -
FEAT
Browser-based frontend QA (F4). When a task touches any frontend markup,
/dr-qaLayer 4f resolves an available browser tool (CLI → MCP → env-browser → none), acquires a per-task lock, opens the local dev surface, and writes screenshot + trace + summary intodatarim/qa/playwright-{TASK-ID}/run-<ISO-ts>/. Skipped silently for non-frontend tasks. Three headed modes: default headless, lenient--headed(no display ⇒ finding + fall through), strict--headed-strict(no display ⇒ exit 2). New detectordev-tools/detect-playwright-tooling.sh; mkdir-fallback lock; path-traversal guard. 15 bats cases. -
FEAT
Plain-language reports across all three operator-facing commands (F5, absorbs the standalone preview from v2.7.0).
/dr-qa,/dr-compliance, and/dr-archiveend with a four-sub-section recap («Что было сделано» / «Что получилось» / «Что не получилось / осталось открытым» / «Что дальше») between the technical block and the next-step block. Anglicism banlist (around 50 ASCII tokens) + universal-term whitelist (around 30 terms —JSON,OAuth,HTTP,CLI,RFC,CI/CD, …) + per-paragraph<!-- gate:literal -->escape hatch (≤ 2 fenced paragraphs per recap). Severity ladder: 1st offence ⇒ info, 3rd ⇒ warn, 5th ⇒ block. Archive documents written before this contract are grandfathered and never re-validated. 24 bats cases. -
FEAT
Archive expectations section (F6). Every archive document carries a new
## Выполнение ожиданий оператораsection between Final Acceptance Criteria and Known Outstanding State / Operator Handoff. Each wish from the expectations checklist is rendered as a single-level bullet with the plain-language status word (выполнено / частично / не выполнено / неприменимо — never the raw schema enum) plus one or two sentences of comment sourced from the most recent История статусов reason. No tables; banlist applies; missing expectations file ⇒ explicit «Чек-лист ожиданий не заводился» line preserves the canonical archive shape. 10 bats cases. -
FEAT
Visual maps refreshed (F7). The fragment-index visual maps gain three new artefact nodes (
init-task,expectations,playwright-run) in a new «Artifact Flow Across the Pipeline» Mermaid diagram, four new skill nodes (init-task-persistence,expectations-checklist,playwright-qa,human-summary) in the Agent ↔ Skill dependency graph, and Layer 3b + Layer 4f branches in the/dr-qastage flow. 18 bats cases. -
FEAT
Skill count: 41 → 45. Four new operator-facing skills ship:
init-task-persistence,expectations-checklist,playwright-qa,human-summary. All consumer surfaces (CLAUDE.md, README.md, docs/skills.md, public site) brought to the same count. -
NOTES
Stack-agnostic gate PASS; gate sweeps green;
shellcheckzero on new shell tools; security gates green. Bats green (13 + 16 + 8 + 15 + 24 + 10 + 18 = 104 new cases, on top of 466 existing — 570 total, 565 passing with 5 pre-existing failures unrelated to this release).
v2.7.0
2026年5月13日-
FEAT
Topic-overlap advisory in
/dr-initStep 2.5b. A non-blocking check compares the fresh task description against the pending backlog and surfaces any item that shares ≥2 keyword stems. RU + EN tokenisation, hand-curated stopwords, crude suffix stemmer — Python 3 stdlib only, no pip dependencies. Operator picks duplicate / refine-scope / orthogonal before committing to a new task ID. Exit 0 by contract — pipeline ordering unchanged. -
FEAT
dev-tools/check-topic-overlap.py— top-N significant stems against pending backlog titles;text(default) andjsonoutput formats;--include-statusflag for pilots that need to scanin_progressitems too. -
FEAT
dev-tools/data/stopwords-{en,ru}.txt— ≥200 entries per language, including Datarim domain noise. -
FEAT
commands/dr-init.md— Step 2.5b inserted after the workspace-hygiene check. Skips silently whenpython3is absent, the backlog has no pending items, or the detector is missing on an older install. -
FEAT
tests/dr-init-topic-overlap.bats+...-fp-budget.bats+...-latency.bats— PRD cases a/b/c, FP rate <10%, TP rate ≥4/5, ≤300 ms on a 500-item synthetic backlog. -
FEAT
Human-readable operator summary at the end of
/dr-complianceand/dr-archive. A new skill defines a 4-sub-section recap (what was done / what worked / what didn't work or is still open / what's next) that both operator-facing commands now emit between their technical block (verdict / archive write) and the CTA block. The recap follows the operator's most recent message language (Russian default for Arcanada consumers, English otherwise), bans tables and jargon, and is capped at 150–400 words. The technical output is unchanged. -
FEAT
skills/human-summary.md— contract: 4 fixed sub-headings, length budget 150–400 words, anti-patterns (tables, English loanwords in Russian text, bare task IDs, multi-level nested lists, acronyms without expansion, emoji, mixed-language summaries), RU and EN mini-examples. -
FEAT
commands/dr-compliance.mdStep 8 — HUMAN SUMMARY. Runs on every verdict; on NON-COMPLIANT the «what didn't work» sub-section carries the failure detail in plain language and «what's next» mirrors the FAIL-Routing CTA without command syntax. -
FEAT
commands/dr-archive.mdStep 8 — HUMAN SUMMARY. Sourced from the just-written archive document plus the reflection file. Chat-only — archive and reflection are not mutated. -
FEAT
tests/test-human-summary-contract.bats— 9 spec-regression tests guarding skill existence, four mandated sub-headings, RU+EN mini-examples, length budget declaration, and cross-references from both commands. - NOTES Stack-agnostic gate PASS; task-id gate PASS; bats green (topic-overlap 6/6 + human-summary 9/9).
v2.6.1
2026年5月12日- CHANGE /dr-doctor 能识别复合标识符和额外的旧版格式。只有当归档文档通过校验时,才会移除归档注释;位置错误的 Backlog 部分会被保留,以便手动更正。
v2.3.0
2026年5月11日-
FEAT
First non-core plugin:
dr-orchestratePhase 1, a lean tmux runner for advancing a Datarim pipeline one bounded step at a time. - SECURITY Fail-closed input floor with a character allowlist, terminal-escape rejection, two cooldown layers, and a temporary pane block after repeated violations.
- SECURITY Operator configuration stays outside git with mode-0600 enforcement. Audit events store a SHA-256 digest of matched pane text instead of the raw text.
- TEST Six Bats suites cover the runner, tmux management, parser, security controls, audit sink, and plugin registration. Key injection remains disabled until the operator opts in.
v2.2.0
2026年5月10日-
FEAT
<TASK-ID> —
skills/diataxis-docs.md(NEW). Documentation Taxonomy Mandate: 4 closed Diátaxis categories —tutorials/(learning),how-to/(problem-solving),reference/(lookup),explanation/(understanding). Closed mapping table covers architecture / testing / deployment / gotchas / api / cli / config / faq / glossary / troubleshooting / examples. 6 anti-patterns explicitly forbidden. Stack-agnostic (no SSG/CMS lock-in). -
FEAT
<TASK-ID> —
templates/docs-diataxis/{tutorials,how-to,reference,explanation}/README.md(NEW, 4 stub files). Per-category onboarding stubs ("when to write here" / "when NOT to write here" / naming convention) used by/dr-initproject scaffolding. -
FEAT
<TASK-ID> —
/dr-optimizeStep 6 docs-drift detector (commands/dr-optimize.md). Filesystem-presence + threshold ≥3 docs check (Bash, exemption-aware). On drift proposesINFRA-* — Diátaxis docs reorg для <repo>in backlog. Soft warning only; hard CI gate intentionally deferred to a separate backlog item, trigger: ≥3 live consumers post-mandate (same detector flips toexit 1). -
CHANGE
<TASK-ID> —
skills/project-init.mdStep 4. Scaffold default replaces flatdocs/{architecture,testing,deployment,gotchas}.mdwithdocs/{tutorials,how-to,reference,explanation}/4-category split. Legacy stubs auto-mapped per skill mapping table: testing/deployment/gotchas →how-to/, architecture →reference/. Backwards-compat smooth (idempotency rule preserves existing files; never overwrite). -
CHANGE
<TASK-ID> —
code/datarim/CLAUDE.md§ Documentation Taxonomy Mandate (NEW section, between Security Mandate and Defensive Invariants). Consumer workspaceCLAUDE.md§ Documentation Taxonomy Mandate added (after Operational Resilience). Both sections point toskills/diataxis-docs.mdas single source of truth. -
DOCS
<TASK-ID> — Public-surface sync:
data/skills/diataxis-docs.phpNEW (EN+RU),config.php2.1.0 → 2.2.0,pages/getting-started.phpDiátaxis project-scaffolding section,pages/home.php+content/{en,ru}.phpstats counts swept skills 39 → 40, frameworkdocs/skills.md+docs/getting-started.md+README.md+ project-level CLAUDE.md/README.md updated. -
CHANGE
<TASK-ID> reframed as first consumer of <TASK-ID> mandate.
datarim.clubremains pilot site for live deployment of the 4-category taxonomy. - NOTES v2.2.0 cumulative — Class B (operating-model / contract change). PRD-<TASK-ID> ratifies the mandate; comparative analysis vs Google dev docs / Write the Docs / Microsoft Concept-Task-Reference / "no mandate" / "discoverability lint only" alternatives recorded for first time. Hard CI gate deliberately deferred — empirical drift data first.
v2.1.0
2026年5月10日-
FEAT
<TASK-ID> —
/dr-verifytri-layer architecture (v2). Layer 1 deterministic floor (dev-tools/dr-verify-floor.sh, shell pipeline, zero LLM cost) + Layer 2 cross-model peer-review (DeepSeek default viacoworker, ~14× cheaper than Sonnet, clean external context — no self-agreement bias) + Layer 3 native runtime dispatch (Claude 3-agent canonical; Codex single-prompt demoted to[experimental]fallback). Findings carry explicitsource_layertag for tri-layer provenance and dedupe. -
FEAT
<TASK-ID> — New
--floor-only+--peer-provider={deepseek,groq,openrouter,...}flags on/dr-verify.--task-idpropagation MANDATORY at Layer 2 — without it downstream token-cost / prospective-rate tooling cannot filter logs by task. -
FEAT
<TASK-ID> — Canonical archive frontmatter template
templates/archive-template.md(NEW). Addsverification_outcomeblock schema (caught_by_verify,missed_by_verify,false_positive,n_a,dogfood_window) — single source of truth for prospective dogfood measurement./dr-archiveStep 2 instructs operator to fill the block. -
FEAT
<TASK-ID> — Token-cost tooling:
dev-tools/measure-invocation-token-cost.sh(per-task aggregation from~/.local/state/coworker/log/<YYYY-MM-DD>.jsonl, OpenTelemetry-style dotted keys, provider breakdown) +dev-tools/measure-prospective-rate.sh(archive frontmatter aggregator, computescaught_per_5_taskswithdecision_hint). Oldmeasure-verify-cost.shdeprecated side-by-side (broken parser shape against current coworker log format). -
FEAT
<TASK-ID> —
/dr-verifyprovider auto-resolution chain (zero-flag UX). New helperdev-tools/resolve-peer-provider.shimplements a 6-step chain: CLI flag → per-project./datarim/config.yaml→ per-user XDG~/.config/datarim/config.yaml→ coworker--profile codedefault → cross-Claude-family subagent fallback → same-model isolated last resort./dr-verify {TASK-ID}now runs end-to-end without an explicit--peer-providerflag. -
FEAT
<TASK-ID> — Cross-Claude-family fallback. New
agents/peer-reviewer.mdSonnet-tier subagent dispatched at chain step #5 in clean isolated context. Covered by Claude subscription, no per-user external API key required. New 3-tierpeer_review_modetaxonomy:cross_vendor/cross_claude_family/same_model_isolated. -
FEAT
<TASK-ID> —
templates/datarim-config.yaml(NEW). Per-project datarim-config schema (peer-review provider, cost cap, AAL targets, runtime preferences). Whitelistdeepseek | moonshot | openrouter | sonnet | haiku | opus | noneblocks malicious-PR typosquat injection at parse time (exit 1 on unknown). -
FEAT
<TASK-ID> — JSONL emission discipline (Layer 2 reviewer prompts).
skills/self-verification.md§ Layer 2 mandates suppression of PASS-as-finding entries: findings array carries only defects or incorrect-premise items. Confirmations compress into final-line summary{cleared_iter1: [...], total_new_findings: N}. -
FEAT
<TASK-ID> —
/dr-planStep 6.5 extension. PRD AC verification commands smoke-checked at plan time against the implemented CLI surface (or pre-implementation skeleton). AC ↔ V-AC semantic match check ensures Validation Checklist rows verify what the AC actually asserts, not just verbatim mirror the AC number. -
FEAT
<TASK-ID> — Per-mode rate aggregation in
dev-tools/measure-prospective-rate.sh. Emitscross_vendor_rate,cross_claude_family_rate,same_model_isolated_ratekeys for per-mode dogfood measurement. -
CHANGE
<TASK-ID> — Brand-hygiene cleanup. Active runtime cross-references to external
superpowers:*skill namespace replaced with local Datarim skill names inskills/systematic-debugging.md(3 refs) andskills/finishing-a-development-branch.md(2 refs). External worktree-manager path-interop strings (~/.config/superpowers/worktrees/) removed from cleanup-eligibility list — Datarim runtime owns only.worktrees/andworktrees/. Lineage from v2.0.0 absorption preserved unchanged in CHANGELOG / PRDs / docs (MIT attribution). -
FEAT
<TASK-ID> —
/dr-planPhase 4 architectural-superseding probe. Mandatory first sub-step before component breakdown: read archives referenced viaSpawned from/Source:and answer whether the architectural problem is already solved by a sibling task. A 30-second grep at planning time prevents dedicated-host plans for problems already absorbed elsewhere. -
FEAT
<TASK-ID> follow-up —
skills/evolution.md§ Pattern: Split-Architecture Metrics for Absorption Tasks. Aggregate token budgets fail when absorption adds on-demand files; replaced with idle hot-path + per-existing-file + on-demand-exempt buckets. -
FEAT
<TASK-ID> —
skills/ai-quality.md§ Pipeline-Position-Aware AC Formulation. When AC asserts HTTP status, trace request through full middleware/filter chain; if status is downstream of any validator, phrase as semantic gate, not literal status. -
FEAT
<TASK-ID> —
skills/testing.md§ Reporting Test Counts in Audit Output. QA/Compliance MUST derive per-spec test counts via mechanical extractor (framework-neutral).skills/compliance.mdStep 7 — stale-base merge-result gate: simulate 3-way merge viagit merge-treebefore flagging a regression from PR diff vs base. -
FEAT
<TASK-ID> —
agents/developer.mdresilience-pattern defaults. Circuit-breakererrorFilter: 4xx excluded except 408/429 (downstream pressure signals); breaker.close → self-heal observability event with explicit listener-binding enumeration. L3–L4 tasks: post-final-TDD design-conformance audit listing every event/lifecycle binding against the referenced ADR. -
DOCS
<TASK-ID> + <TASK-ID> — Public-surface ↔ runtime sync (5-way):
data/commands/dr-verify.php+data/agents/peer-reviewer.php(NEW EN+RU),code/datarim/docs/commands.mdrow,docs/agents.mdroster (peer-reviewer added), framework CLAUDE.md /dr-verify rewrite + agent table, README.md feature mention. Counts swept: agents 17→18, skills 27→39, commands 20→22, templates 18→23. -
NOTES
v2.1.0 cumulative — Class B-lite additive (no breaking changes). <TASK-ID> inherits scope from <TASK-ID> v1 → v2 PRD revision; <TASK-ID> extends without contract change. Findings-only mode preserved at all layers — no auto-fix added. Cross-Claude-family dispatch is first measured tier: empirical bias delta vs same-model self-critique remains under observation in the active dogfood window. Codex CLI degraded mode (
CODEX_RUNTIME=1): chain step #5 skipped, step #6 taken with stderr WARN propagated to audit-log.
v2.0.0
2026年5月9日-
FEAT
<TASK-ID> — Multi-runtime framework (Claude + Codex). New
install.shflag-based architecture:--with-claude,--with-codex,--project DIR,--yes,--dry-run,--force(no flags = print help). Shim creation: install.sh creates~/.claude/or~/.codex/directories even without runtime binary.--project DIRcopy mode rejects system paths (/etc,/usr,/bin,/sbin,/System) with exit 3. Lockfile~/.${runtime}/.install.lockblocks concurrent runs. -
FEAT
<TASK-ID> —
AGENTS.mdsymlink →CLAUDE.md. Codex CLI and other runtimes that readAGENTS.mdby convention work out of the box, no build step or fork required. -
FEAT
<TASK-ID> — 14 superpowers skills absorbed. 4 verbatim ports (
finishing-a-development-branch,receiving-code-review,systematic-debugging,verification-before-completion); 8 intent-layer rewrites (brainstorming,dispatching-parallel-agents,executing-plans,requesting-code-review,subagent-driven-development,using-git-worktrees,writing-plans,writing-skills) — Claude tool literals (Task()/TodoWrite()/<system-reminder>) replaced with functional prose for runtime-agnostic readability; 2 merges (test-driven-development→testing.md§ Discipline;using-superpowers→datarim-system.md§ Skill Discovery). -
FEAT
<TASK-ID> — Per-skill
runtime: [claude, codex]+current_aal/target_aalfrontmatter on all 38 top-level skills. Classification per AAL Mandate: human-driven idea/design tools 1/2; disciplined process gates 2/3; orchestration / execution-loop skills 2/4. New validatordev-tools/check-skill-frontmatter.sh. -
FEAT
<TASK-ID> — Token-budget regression gate
dev-tools/measure-skill-token-cost.sh. AC-4a idle hot-path:skills/datarim-system.md≤+16% chars vs v1.23.0 baseline (frozen as.datarim/baseline-v1.23.0.tokens). AC-4b per-existing-file: ≤+30% chars. AC-4c on-demand absorbed files exempt (loaded only when invoked). -
FEAT
<TASK-ID> —
CHANGELOG.mdKeep-a-Changelog format introduced; v2.0.0 entry covers full <TASK-ID> ship. -
DOCS
<TASK-ID> — Honest positioning: multi-runtime framework (Claude + Codex), not "vendor-neutral". Cursor / Goose / Aider — future milestones, not current scope. Codex disclaimer: experience may differ — no
Task/TodoWriteprimitives; intent-layer rewrites preserve runtime-agnostic readability. Public Surface Scan covers 12 files: README, CLAUDE.md, AGENTS.md, CHANGELOG, docs/*, Projects/Datarim/{README,CLAUDE}, datarim.club {config.php, content/{en,ru}.php, getting-started.php, changelog.php}, ~/.claude/CLAUDE.md global. -
CI
<TASK-ID> — 41 bats tests GREEN: 29/29
tests/install.bats(legacy invariants <TASK-ID>/0077/0091 preserved) + 12/12tests/install-tune-0114.bats(no-flags help, --with-claude/--with-codex/multi, --project happy + 3 system-path rejects + missing-arg, --dry-run x2, lockfile concurrency, legacy --copy WARN). -
NOTES
<TASK-ID> — No breaking changes for existing Claude installs. Refresh via
./install.sh --with-claude. Sub-tasks unblocked: <TASK-ID> (Adversarial Review skill split), <TASK-ID> (Diátaxis reorg), <TASK-ID> (/dr-status pull-mode), <TASK-ID> (Party Mode → Consilium-lite). Follow-ups spawned: <TASK-ID> (project-local evolution learning), <TASK-ID> (Module Manifest — separate task).
v1.24.0
2026年5月7日- FEAT <TASK-ID> — Secure-by-default Network Exposure Baseline. New skill `network-exposure-baseline.md` formalises a four-tier bind-target model (Tier 0 unix-socket / Tier 1 loopback / Tier 2 Tailscale CGNAT / Tier 3 public-with-justification). Tier 3 requires inline `x-exposure-justification` + `x-exposure-expires` (≤90 days). Public-by-default = breach-by-default; loopback/Tailscale binds become the dev/prod default.
- FEAT <TASK-ID> — Verifier `dev-tools/network-exposure-check.sh` parses docker-compose ports, redis.conf bind, postgresql.conf listen_addresses, systemd `.socket` ListenStream and Dockerfile EXPOSE; emits text or SARIF. Tiered gate `dev-tools/network-exposure-gate.sh` reads task-description frontmatter (priority/type) and returns `hard_block` / `advisory_warn` / `skip`; missing/malformed frontmatter resolves fail-closed.
- FEAT <TASK-ID> — Pipeline integration: `/dr-prd`, `/dr-plan`, `/dr-do`, `/dr-archive` auto-load the skill on network-surface diffs; pre-commit gate in `/dr-do` (Step 8.5) blocks Tier 3 violations or requires `--skip-exposure-gate` override (advisory cases only) with an ops notification event + § Decisions note. Workspace `CLAUDE.md` § Backend Stack Standards + § CI/CD Convention updated with Tier model + mandatory lint workflow.
- CI <TASK-ID> — Reusable workflow `templates/network-exposure-lint.yml` (`workflow_call` with six contract inputs). Drop-in for any service repo; `lint failure ⇒ deploy blocked`. V-AC9 perf budget 30 s; combined run (compose+redis+postgres+systemd) measured p95 ≈ 0.14 s — ~200× headroom.
- TEST <TASK-ID> — 59/59 cumulative bats GREEN: 16 P1 verifier (boundary IPv4/IPv6, mapped IPv6, TTL, sub-second perf), 19 tiered-gate, 11 pipeline-command integration, 13 CI workflow integration (structural assertions + 4 fixture-repo simulated runs + V-AC9 perf assertion). actionlint + semgrep p/default + p/secrets + gitleaks all clean.
- SECURITY <TASK-ID> — Phase 4 audit of Arcanada-one repos surfaced three production violations spawned as <TASK-ID> (a transcription service PROD compose: `3700:3700` short-form, P0), <TASK-ID> (an auth service dev compose: postgres/redis/app loopback bind, P1), <TASK-ID> (an ops-bot dev compose: postgres/redis loopback bind, P1). Verifier capability gaps (env-interpolation defaults, `${TAILSCALE_IP}:PORT` recognition) tracked as <TASK-ID> / <TASK-ID>.
- DOCS <TASK-ID> — `getting-started.php` adds Network Exposure Baseline section (Tier model badges + local check + reusable CI workflow). Workspace + framework CLAUDE.md updated; data/skills/network-exposure-baseline.php published for site n-way sync.
v1.23.0
2026年5月4日-
FEAT
<TASK-ID> — Plugin System Core: opt-in plugin architecture for Datarim. New CLI `/dr-plugin {list,enable,disable,sync,doctor}` with manifest-driven runtime symlinks. Local plugin sources (absolute path) installed as namespaced symlinks under `~/.claude/{skills,agents,commands,templates}/
/`. Override mechanism allows root-position installation via `overrides:` field in `plugin.yaml`. `enabled-plugins.md` is the single source of truth for the active set; `dr-plugin sync` reconciles runtime against manifest (orphan removal, broken-symlink recreate, disabled-orphan restore). - FEAT <TASK-ID> — `dr-plugin doctor` ships with 9 health checks: manifest-syntax, inventory-consistency, broken-symlinks, orphan-files, override-integrity, dependency-graph (cycle/dangling DFS), git-state (uncommitted manifest), snapshot-cleanup (>30d), and skill-registry (frontmatter `name:` ↔ basename validation — closes the Skill-tool resolution gap surfaced in Round 4 dogfooding). `--fix` flag auto-repairs checks 2-5 via `cmd_sync` and check 8 via snapshot purge.
- SECURITY <TASK-ID> — Pre-mutation snapshot/rollback protocol on every `enable` (V-8). FIFO cap `DR_PLUGIN_SNAPSHOT_MAX=50`; age-based purge via `DR_PLUGIN_SNAPSHOT_AGE_DAYS=30`. Conflict pre-scan rejects symlink loops and cross-plugin shadowing. Critical-core overrides (`evolution`, `datarim-system`, `pre-archive-check`) emit operator-warning to stderr but allow (V-SEC-4). Path traversal, embedded credentials in URLs, and CRLF in YAML rejected at validation gate.
- TEST <TASK-ID> — `tests/dr-plugin.bats` 77/77 GREEN across Phases A-D (T1-T135). Coverage: ID/source validation, YAML parsing, list/enable/disable/sync/doctor flows, snapshot/restore, fault injection rollback, override integrity, dependency cycles, skill-registry frontmatter validation. shellcheck -S warning clean.
- DOCS <TASK-ID> — `commands/dr-plugin.md` slash-command spec, `templates/plugin.yaml.template` schema (schema_version 1), `templates/enabled-plugins.md.template` runtime manifest schema, README + CLAUDE.md + getting-started.md mention plugin system as opt-in extension mechanism. `data/commands/dr-plugin.php` site page (EN+RU).