Skip to content

CoreMind 1.4.0

Latest

Choose a tag to compare

@github-actions github-actions released this 10 Nov 05:53
ca46c49

CoreMind v1.4.0

Version: 1.4.0
Tag: v1.4.0
Commit: ca46c49

简体中文

本说明对应 1.4.0 发布线,包含未独立发布的 1.3.0 修复;1.3.0 不单独发布。源码、工程 CI 和文档同步不等于正式交付,最终资格及公开安装状态必须由同版本、最终提交与准确制品证据确认。

关键亮点

  • 在唯一 Node Runtime 内提供原子任务提交、可恢复阶段、持久等待与稳定键调用,沿用权限、预算、取消和未知效果恢复边界。
  • TypeScript 与同步/异步 Python 经同一 Worker 使用任务与会话树能力,支持真实进程重启、分支选择和按节点 fork。
  • 纳入框架状态保护、并行终止、步骤输出、Python 注解、评测、Web/CLI 收尾、重试分类、Windows 会话路径和 CI 失败传播修复。
  • 同步 1.4.0 版本、双语使用/迁移/发布材料与公开制品身份。将底层运行库升级至 1.5.0 并移除 #359 的历史风险接受项,独立开发依赖风险保留原到期日。

本版本完整变更日志

本节记录 1.4.0 发布线;源码和发布准备不等于公开交付,实际状态以同版本 Release/Registry 与最终提交、准确制品的资格证据为准。1.4.0 包含未独立发布的 1.3.0 修复及本轮持久任务、会话树和跨语言能力;不补发 1.3.0。1.2.0 用户请按迁移指南升级,最终固定正文见发布说明。

安全与依赖

  • Linux 严格沙箱启动: 由 Bubblewrap 直接启动受信 PID 1,避免 Ubuntu 26.04 上重复创建用户命名空间导致的启动拒绝;保留文件隔离、零能力、断网与原 seccomp 过滤器,并禁止嵌套用户命名空间。Bubblewrap 需支持 --disable-userns 和 --as-pid-1,缺少前置条件时拒绝执行。 (#380)

  • 框架状态保护: 统一授权按 canonical 目标保护框架管理的持久状态,普通业务工具不能通过 ask、assisted、full 或显式 allow 改写这些状态。现存多硬链接文件的普通工具写入也会被拒绝;合法业务文件和宿主维护接口保留原职责。这是路径保护,不是任意宿主代码的 OS 沙箱。 (#324)

  • 开发构建依赖: 将 Vitest 测试链的 Vite 固定为 8.2.4,仅将 VitePress 1.6.4 子树的 Vite 覆盖为 6.4.5,修复编辑器启动相关漏洞;不引入预发行文档框架,也不新增风险豁免。变更限于开发/文档构建依赖,不改变 Runtime 权限或沙箱边界。 (#350)

  • 测试依赖安全修复: 将 vitest 与 @vitest/coverage-v8 同步升级至 4.1.12,使测试链纳入 @vitest/mocker 的手工 mock 代码生成转义与 Windows redirect 路径规范化补丁(GHSA-qxg6-89mg-rqv8、GHSA-xp26-3454-mjgp)。变更仅限开发/测试依赖,不改变 Runtime 权限或沙箱边界,也不新增风险豁免。 (#355)

  • 生产与构建依赖加固。 fast-uri 更新至 3.1.9,固定 Linux 后端开发源的 Zod 更新至 4.6.6、node-forge 更新至 1.4.2,文档链的 Vite 更新至 6.4.5。 (#359)

  • 底层运行库 1.5.0 迁移。 将八个依赖成员锁定到同一版本,纳入新增工具参数校验包,采用官方传递版本 Undici 8.12.3 与 minimatch 11.0.3,并同步 CoreMind 网络工具的 Undici 版本。两项传递依赖超出 #359 登记的九条公告影响范围,相应风险接受项移除。独立开发依赖风险仍按原范围审计,到期日为 2026-11-14。详见安全说明。 (#379)

能力提升

  • 任务原子提交: 新增 internal TaskCommitSession,将 checkpoint、typed document 与文本 entry 作为同一持久事实提交;revision、内容绑定 commitId、Schema 身份和单 writer 约束支持安全重开与崩溃恢复。旧 Session/Run/Workspace 数据不迁移,原子性不涵盖外部业务副作用。 (#362)
  • 持久任务阶段与迁移: 公共 TaskSession 支持 start/query/resume/cancel、阶段检查点与显式状态迁移;模型和工具复用既有 Runtime/Run/ChildRun、权限、预算与取消。重启从已提交阶段继续,无法确认的 Run 保留人工核对,终态任务不重做。 (#363)
  • 持久等待与幂等唤醒: 原子保存下一阶段、检查点和绝对期限;重开后 nextWake 重新计算等待,持久 sleepId/wakeId 约束重复或过期通知。宿主负责进程启动、alarm 与显式恢复,不增加后台调度服务或扩大原超时。 (#364)
  • 稳定键调用与结果复用: runCalls、executeTool 和调用查询将结果绑定父任务、阶段、版本及稳定键,并从原 Run/任务事实恢复。参数、能力或调用顺序分歧会拒绝并审计;未知外部效果仍由原恢复门拦截,所有键共享原权限、预算、取消和 ChildRun 上限。 (#365)
  • 公共会话树与分支: ChatSession 提供树查询、命名分支、持久选择及按消息节点 fork;继续对话读取所选历史,原写锁与活动 Run 限制保留。fork 产生新会话和来源关联,不复制 Run/控制/授权身份。 (#366)
  • 跨语言任务与会话: Protocol v2 增加独立能力与绑定身份的 DTO,TypeScript 与同步/异步 Python 经唯一 Node Worker 调用宿主注册的任务及会话导航。默认 Worker 不开放未注册任务,RPC 不传源码、模块路径或凭据;工具权限与项目代码信任分别校验,实际信任开关进入任务权限身份。原顶层错误分类与任务子码、v1 行为及所有超时保留。 (#367)

缺陷修复

Runtime 与 Workflow
  • 阿里云输出参数兼容: 内置北京 qwen-plus 使用 max_tokens,命名 Agent 切换模型时重新选择上限字段,避免继承不兼容参数;其他自定义端点保留原行为。字段修复不代表真实认证已通过。 (#381)

  • 会话路径恢复: 同一物理工作目录的 Windows 长短名或目录链接可重开并 fork 既有 v4 会话;核对目录与内部/公开文件身份后沿用原存储和 metadata.cwd,避免误入 v3 迁移,不搬移或改写历史。不同物理目录与复制文件仍拒绝。 (#367)

  • 重试状态码边界: 请求 ID 或长数字中的 503、524 等子串不再把未知 Provider 错误误分类为瞬态;Loop 保留原错误及脱敏审计,暂停等待人工处理。独立状态码、其他瞬态原因、配额排除及原重试预算保持不变。 (#360)

  • 并行终止与准入: 普通 parallel 子步骤首次失败或暂停时,取消同组其他步骤并等待收尾,保留首个错误;Run 结束后不再准入新模型调用。已开始操作保留收据,取消不等于回滚;全成功仍按声明顺序聚合。 (#325)

  • 步骤输出边界: Workflow prompt/call 保留同一 Session 历史作为模型上下文,但步骤输出、saveAs、后续插值与 transcript 只包含本步骤新增的 assistant 文本,不重复历史回答。 (#327)

  • v2 启动错误身份: 已接受 Run 的 Runtime 初始化失败按共享错误分类持久化稳定错误码、对应终态与安全消息,供 query/events 查询,不再统一改为 agent_failed。未知异常以 unclassified_error 保守暂停并要求人工审计,不公开凭据或原始堆栈,也不表示可直接恢复。 (#329)

  • 租约持久化开销: 对未改写的内置 FileRunStore,租约关键 Fact 后不再重复执行全局刷盘,降低慢磁盘下 MCP 关键路径的冗余同步开销;仍等待 critical acknowledgement 并拦截失败,释放前安全屏障保持。自定义 Store、子类以及改写 commit 或 barrier 的实例沿用原屏障行为。 (#352)

Python 工具桥
  • 延迟参数注解: v1 @client.tool 自动 Schema 可解析 future/字符串形式的 int、str、bool 和既有支持的容器注解;无法解析的前向引用明确要求显式 parameters。保留 Node 侧先转换再校验的既有规则,转换后仍未通过 Schema 的参数在 callable 执行前拒绝;显式 Schema 兼容,复杂 typing 支持范围不扩大。 (#326)
评测
  • Diff grader: preserveExisting: false 只关闭既有脏内容保护,allowedPaths、forbiddenPaths、requiredPaths、maxChangedFiles 仍按基线与终态的状态或内容差异执行。状态和内容均未变化的既有脏文件不计,原 untracked 文件删除计入变化;原 stage/unstage 计数及默认 true 保护不变。 (#328)
Web 工具
  • 响应正文收尾: web-fetch/web-search 遇非 2xx 或正文读取失败时,等待正文取消/释放的清理尝试结束后才结算本地网络活动;清理异常保留原 HTTP/读取错误。成功、取消、权限与网络策略不变,web-fetch 原 2 MiB 正文上限保留。 (#330)
CLI
  • readline 显式退出: /exit、!exit 与 TUI 一致,先请求取消活动 Run 并等待收尾;普通管道 EOF 仍让已开始的回答自然完成。空闲退出、审批及 MCP 登录等待的既有行为保持,取消不回滚已开始工具副作用。 (#331)
  • doctor 凭据策略: 有配置时复用所选 Provider 的既有校验,合法匿名/Header-only 自定义端点不要求无关 API key;显式凭据缺失仍失败,内置鉴权要求不放宽。无配置通用诊断保留;检查不加载项目脚本、不发 Provider 请求,通过不代表凭据有效、真实连通或认证。 (#332)
工程验收
  • DeepSeek 人民币认证预算: 固定 deepseek-flash、关闭思考并限制单次输出 1,024 token;全次认证按人民币高峰价格在外发前预留完整请求上界,父子调用、失败与取消共用账本,仅完整可信终态释放差额。新增明确批准的人民币上限,原美元估算门、零重试、超时和真实鉴权验收保留;不代表已完成真实认证。 (#385)
  • 阿里云认证费用边界: 真实认证新增默认拒绝的确认项,仅在操作人核实北京 qwen-plus 有效免费额度且“用完即停”已生效后放行,并在证据中记录固定费用保护声明;不把零费率视为零费用,不代替账号核实,保留原预算、零重试、候选身份和全部超时。 (#381)
  • CI 失败传播: 拆分 Windows 工程门的多命令步骤,防止主 Python 等前序失败被后续成功覆盖;保留原命令、顺序、验收和超时,并补充真实 PowerShell 退出码反例。 (#376)
  • Coding Kernel 解释器身份: 优先使用 setup-python 提供的 pythonLocation 解析 Python 绝对路径,避免裸命令依赖 PATH 搜索;未设置时保留原行为,原 30 秒上限、退出码与错误分类保持。 (#357)
  • 取消竞态边界: 验收夹具按真实取消准入边界及未完成 Call/Turn 身份核对关闭事件,修正场景标签导致的正常收尾误判,并检查真正迟到的已准入活动。增加确定进入审批再取消的回归、严格反例及双平台固定种子门;保留完整 1000 种子、原超时与全部不变量,生产取消与权限规则不变。 (#348)

文档与发布工程

  • 安全策略版本状态。 双语 SECURITY 支持表链接最新正式 Release,记录已发布版本与既有加固、私密披露方式和安全边界,并区分早期底层依赖的风险接受与当前开发依赖审计范围。 (#333, #359, #379)
  • 统一 1.4.0 交付材料: 纳入原 1.3.0 发布准备、安装导航与 Worker 生成物同步,统一根与 8 个 npm 包、Python、内部精确依赖、锁文件、模块与 TTY 模板到 1.4.0。同步双语 README、设计/指南/SOP/Skill、示例、1.2→1.4 迁移、路线图和文档站,固定发布正文包含完整本版本变更。1.3.0 不独立发布,历史文档与证据保留。 (#334, #346, #358, #368)
  • 能力操作材料: 任务提交、阶段、等待、稳定键调用、会话树和跨语言入口均配套双语指南、模块合同、SOP/Skill、离线可运行示例及真实进程恢复回归;工程检查不代替最终 Provider 或正式发布资格。 (#362, #363, #364, #365, #366, #367)

兼容与发布边界

保留 Config v2/v3、Protocol v1/v2 和唯一 Node Runtime。新任务/会话 DTO 与 capability 需要同步升级 SDK 和 bundled Worker,并通过新 v2 Schema 指纹检查。默认 Worker 不接收任意可执行任务;宿主代码信任不等于沙箱或工具授权。ToolPolicy 的可选 protectedPaths 和授权身份的可选 signal?: AbortSignal 保持既有调用兼容。

旧 Session/Run/Workspace 不统一迁移;新的任务事实和会话选择记录由新 SDK 管理,旧 SDK 不接管。回滚前必须停止所有 writer,恢复升级前的程序、SDK、Worker 与完整状态备份;未知外部效果不能无条件重放,也不承诺外部 exactly-once、自动副作用回滚或分布式锁/调度。详见1.2→1.4 迁移。

正式发布仍须绑定最终 main、当前版本、Runtime/Worker 和准确制品的严格 Provider、双平台工程/稳定性/覆盖率/真实 PTY,以及公开下载与回装证据。旧认证、历史例外、单项工程 CI 和文档同步均不能代替本轮资格;实际状态以同版本 Release/Registry 及最终资格证据为准。

升级与回滚

要求 Node.js ≥ 22.19、Python ≥ 3.10;Git 工具要求 Git ≥ 2.36。平台范围仍为 Windows/Linux。按1.2→1.4 迁移指南停止旧进程并备份配置、Session、RunState、Checkpoint 和任务事实,统一升级 SDK 与 Worker,再核对权限、结果和错误边界。回滚须先停止新 writer,并恢复旧程序及升级前完整备份;本地恢复不撤销外部操作。

验证范围与已知限制

各修复和能力的定向回归、真实离线进程与合并 CI 不替代 1.4.0 最终发布资格。须在冻结 main 后完成同版本、最终提交、Runtime 摘要和 bundled Worker 绑定的严格 Provider、双平台稳定性、覆盖率、真实 PTY 和公开逐制品回装;没有这些证据时不能标记已认证或已发布。真实 Provider 外发、费用、模型、凭据来源和最长运行时间须获得绑定最终提交及 Runtime 包 SHA256 的明确批准,旧授权不代用。

此前本地 initialize 响应超时尚未获得独立根因解释;本轮会话路径修复不作该历史观察已消除的声明。框架不承诺外部 exactly-once、自动回滚副作用、后台分布式调度或任意宿主代码的 OS 沙箱;真实外部 OAuth/MCP 互通及全局成对性能收益未认证。底层依赖迁移的发布证据包括无仓库 overrides 的 npm 消费者安装图和最终提交的生产与完整审计。

版本、Tag、清单与 8 个 npm 包、wheel、独立源码 ZIP 共 10 件附件身份必须一致,断点恢复只复用原不可变制品。真实依赖、版权、许可证和 NOTICE 义务保持,见目标 Tag 的第三方声明及许可证;目标 Tag 须在正式发布时存在。

English

These notes describe the 1.4.0 release line and include the unpublished 1.3.0 fixes; version 1.3.0 is not released separately. Source, engineering CI and synchronized documentation do not establish formal delivery. Final qualification and public installability require evidence for the same version, final commit and exact artifacts.

Key highlights

  • Provide atomic task commits, recoverable phases, durable waits and keyed calls within one Node Runtime, preserving permissions, budgets, cancellation and unknown-effect recovery controls.
  • Use tasks and session trees from TypeScript and sync/async Python through one Worker, with real-process restart, branch selection and message-node forks.
  • Include framework-state protection, parallel termination, step outputs, Python annotations, evaluation, Web/CLI settlement, retry classification, Windows session-path recovery and CI failure propagation fixes.
  • Synchronize 1.4.0 metadata, bilingual usage/migration/release materials, and artifact identity. Upgrade the underlying runtime to 1.5.0 and remove the historical #359 risk acceptances while retaining the separate development dependency risk expiry.

Complete changelog for this version

This section records the 1.4.0 release line. Source and preparation do not establish public delivery; actual availability depends on same-version Release/registry records and qualification for the final commit and exact artifacts. Version 1.4.0 includes the unreleased 1.3.0 fixes plus durable tasks, session trees and cross-language capabilities; 1.3.0 will not be released separately. Users on 1.2.0 should follow the migration guide. See the fixed release notes.

Security and Dependencies

  • Strict Linux sandbox startup: Bubblewrap directly launches a trusted PID 1, avoiding denied nested user-namespace setup on Ubuntu 26.04. File isolation, zero capabilities, network isolation and the existing seccomp filter remain, with further user namespaces disabled. Bubblewrap must support --disable-userns and --as-pid-1; missing prerequisites fail closed. (#380)

  • Framework-state protection: unified authorization protects framework-managed persistent state by canonical target. Ordinary business tools cannot bypass it through ask, assisted, full, or an explicit allow. Ordinary tool writes to existing multiply hard-linked files are also rejected; legitimate business files and host maintenance interfaces retain their roles. This is path protection, not an OS sandbox for arbitrary host code. (#324)

  • Development/build dependencies: pin Vite in the Vitest test chain to 8.2.4 and override Vite to 6.4.5 only within the VitePress 1.6.4 subtree to fix editor-launch vulnerabilities. No prerelease documentation framework or new risk exception is introduced. The change is limited to development/documentation build dependencies and does not change Runtime permissions or sandbox boundaries. (#350)

  • Test-dependency security fixes: upgrade vitest and @vitest/coverage-v8 together to 4.1.12, bringing the @vitest/mocker manual-mock code-generation escaping and Windows redirect-path normalization patches into the test chain (GHSA-qxg6-89mg-rqv8, GHSA-xp26-3454-mjgp). The change is limited to development/test dependencies, does not change Runtime permissions or sandbox boundaries, and adds no risk exception. (#355)

  • Production and build dependency hardening. Update fast-uri to 3.1.9, Zod to 4.6.6 and node-forge to 1.4.2 in the pinned Linux backend build source, and Vite to 6.4.5 in the documentation chain. (#359)

  • Underlying runtime 1.5.0 migration. Lock all eight family members to one version, including the new tool argument validation package. Use the official transitive versions Undici 8.12.3 and minimatch 11.0.3, and align CoreMind network tools' Undici version. These transitive versions fall outside all nine advisory ranges recorded in #359, so the corresponding risk acceptances are removed. Separate development dependency risks retain their audit scope and original 2026-11-14 expiry. See the security policy. (#379)

Improvements

  • Atomic task commits: internal TaskCommitSession commits checkpoints, typed documents and text entries as one durable fact. Revisions, content-bound commit IDs, schema identity and a single writer support safe reopen and crash recovery. Legacy Session/Run/Workspace data stays unchanged; atomicity does not extend to external business effects. (#362)
  • Durable task phases and migration: public TaskSession adds start/query/resume/cancel, phase checkpoints and explicit state migration. Models and tools reuse the existing Runtime/Run/ChildRun, permissions, budgets and cancellation. Reopening continues from a committed phase; unresolved Runs require human reconciliation and terminal tasks do not run again. (#363)
  • Durable waits and idempotent wakeups: atomically persist the next phase, checkpoint and absolute deadline. Reopening recomputes waiting through nextWake; persistent sleepId/wakeId bindings constrain duplicate or stale notifications. Hosts own process startup, alarms and explicit resumption; no background scheduler or longer timeout is introduced. (#364)
  • Keyed calls and result reuse: runCalls, executeTool and call queries bind results to the parent task, phase, version and stable key and recover them from original Run/task facts. Argument, capability or call-order divergence is rejected and audited. Unknown effects retain the original recovery gate; all keys share existing permissions, budgets, cancellation and ChildRun limits. (#365)
  • Public session trees and branches: ChatSession adds tree inspection, named branches, durable selection and message-node forks. Continued chat reads the selected history under existing writer ownership and active-Run restrictions. Forks create a new session and source link without copying Run/control/authorization identities. (#366)
  • Cross-language tasks and sessions: Protocol v2 adds negotiated capabilities and identity-bound DTOs. TypeScript and sync/async Python use one Node Worker for host-registered tasks and session navigation. The default Worker exposes no unregistered tasks; RPC transports no source code, module paths or credentials. Tool permissions and project-code trust remain separate, and actual trust controls enter task authority identity. Existing top-level error classification, task subcodes, v1 behavior and timeouts remain. (#367)

Bug Fixes

Runtime and Workflow
  • Alibaba output-parameter compatibility: built-in Beijing qwen-plus uses max_tokens. Named Agents reselect the limit field when switching models, preventing incompatible parameter inheritance; other custom endpoints retain their behavior. The fix does not establish live certification. (#381)

  • Session path recovery: reopen and fork existing v4 sessions through Windows long/short paths or directory links to the same physical workspace. Directory and internal/public-file identity checks retain original storage and metadata.cwd without entering v3 migration or moving or rewriting history. Different physical directories and copied files remain rejected. (#367)

  • Retry status-code boundaries: digits such as 503 or 524 inside request IDs or longer numbers no longer misclassify unknown Provider errors as transient. Loop preserves the original error and redacted audit and pauses for human review. Standalone status codes, other transient reasons, quota exclusions, and existing retry budgets retain their behavior. (#360)

  • Parallel termination and admission: the first failure or pause in an ordinary parallel substep cancels its siblings and waits for settlement while preserving the first error. No new model calls are admitted after the Run ends. Started operations retain receipts; cancellation is not rollback. All-success results still aggregate in declaration order. (#325)

  • Step-output boundaries: Workflow prompt/call retains the same Session history as model context, while step output, saveAs, subsequent interpolation, and transcript contain only assistant text newly produced by that step, without repeating earlier replies. (#327)

  • V2 startup-error identity: Runtime initialization failures after Run acceptance persist the shared classification's stable code, corresponding outcome status, and safe message for query/events, rather than replacing every failure with agent_failed. Unknown exceptions pause as unclassified_error and require human audit, without exposing credentials or raw stacks or establishing that direct resume is safe. (#329)

  • Lease durability overhead: for an unmodified built-in FileRunStore, avoid redundant global flushes after critical lease Facts, reducing redundant synchronous I/O on the MCP critical path with slow disks. Still await the critical acknowledgement and block on failure; retain the safety barrier before lease release. Custom Stores, subclasses, and instances with overridden commit or barrier retain their existing barrier behavior. (#352)

Python tool bridge
  • Deferred parameter annotations: v1 @client.tool automatic schemas resolve future/string annotations for int, str, bool, and existing supported containers. Unresolvable forward references explicitly require parameters. Existing Node conversion still precedes validation; arguments that remain invalid are rejected before callable execution. Explicit schemas remain compatible without expanding complex typing support. (#326)
Evaluation
  • Diff grader: preserveExisting: false only disables protection of existing dirty content. allowedPaths, forbiddenPaths, requiredPaths, and maxChangedFiles still use baseline-to-final status or content differences. Dirty files with unchanged status and content do not count; deleting a file untracked at baseline does. Existing stage/unstage counting and default true protection remain unchanged. (#328)
Web tools
  • Response-body settlement: on non-2xx responses or body-read failures, web-fetch/web-search await body cancellation/release cleanup attempts before settling local network activity. Cleanup failures preserve the original HTTP/read error. Success, cancellation, permissions, network policy, and the existing web-fetch 2 MiB body limit remain unchanged. (#330)
CLI
  • Explicit readline exit: /exit and !exit match the TUI by requesting active-Run cancellation before waiting for cleanup. Ordinary piped EOF still lets an already-started response finish naturally. Existing idle-exit, approval-wait, and MCP-login-wait behavior remains; cancellation does not roll back started tool effects. (#331)
  • Doctor credential policy: configured diagnostics reuse the selected Provider's existing checks. Legitimate anonymous/Header-only custom endpoints require no unrelated API key; explicit missing credentials still fail, and built-in authentication requirements are not relaxed. General diagnostics without a config remain. Checks neither load project scripts nor send Provider requests; passing does not prove credential validity, connectivity, or certification. (#332)
Engineering checks
  • DeepSeek CNY certification budget: fix deepseek-flash, disable thinking and cap each output at 1,024 tokens. Reserve the full request bound at peak CNY prices before dispatch in one ledger shared by parent, child, failed and cancelled calls; release unused reserves only after a complete, trustworthy terminal response. Require an explicitly approved CNY cap while retaining the USD estimate gate, zero retries, timeouts and live authentication checks. This does not establish live certification. (#385)
  • Alibaba certification billing boundary: live certification requires a confirmation that defaults to rejection. The operator must verify valid Beijing qwen-plus free quota and an effective stop-on-exhaustion control; evidence records a fixed billing-control declaration. Zero rates do not establish zero charges, and the declaration does not replace account verification. Existing budgets, zero retries, candidate identity and all timeouts remain. (#381)
  • CI failure propagation: split multi-command Windows engineering steps so later success cannot mask earlier failures, including Python SDK tests. Preserve commands, ordering, acceptance and timeouts, with a real PowerShell exit-code counterexample. (#376)
  • Coding Kernel interpreter identity: prefer the absolute Python path from setup-python's pythonLocation over a bare command resolved through PATH. Preserve the fallback when it is unset, the original 30-second limit, exit codes and error classification. (#357)
  • Cancellation-race boundaries: the acceptance fixture validates cleanup events against the actual cancellation-admission boundary and pending Call/Turn identities, correcting false failures based on scenario labels while checking genuinely late admitted activity. Add a gated approval-cancellation regression, strict negative cases, and a fixed-seed gate on both platforms. Retain all 1,000 seeds, original timeouts, and every invariant; production cancellation and permission rules remain unchanged. (#348)

Documentation and Release Engineering

  • Security-policy version status. Bilingual SECURITY tables link to the latest official Release and document shipped versions and hardening, private disclosure, and security boundaries. They distinguish historical underlying dependency risk acceptance from the current development dependency audit scope. (#333, #359, #379)
  • Synchronized 1.4.0 delivery materials: include the earlier 1.3.0 release preparation, installation/navigation updates and Worker generation, and synchronize the root, eight npm packages, Python, exact internal dependencies, lockfile, modules and TTY templates to 1.4.0. Align bilingual READMEs, design/guides/SOP/Skills, examples, the 1.2→1.4 migration, roadmap and documentation site. Fixed release notes include the complete changelog. Version 1.3.0 is not released separately; historical documents and evidence remain. (#334, #346, #358, #368)
  • Operational materials: task commits, phases, waits, keyed calls, session trees and cross-language entry points include bilingual guidance, module contracts, SOP/Skills, runnable offline examples and real-process recovery regressions. Engineering checks do not replace final Provider or publication qualification. (#362, #363, #364, #365, #366, #367)

Compatibility and Publication Boundaries

Retain Config v2/v3, Protocol v1/v2 and one Node Runtime. New task/session DTOs and capabilities require the SDK and bundled Worker to upgrade together and match the new v2 schema fingerprint. The default Worker accepts no arbitrary executable task; host-code trust is neither a sandbox nor tool authorization. Existing calls remain compatible with optional ToolPolicy protectedPaths and optional authorization identity signal?: AbortSignal.

Legacy Session/Run/Workspace records are not migrated wholesale. New task facts and session-selection entries belong to the new SDK; older SDKs must not take them over. Before rollback, stop every writer and restore pre-upgrade programs, SDKs, Worker and complete state backups. Unknown external effects cannot be replayed unconditionally. External exactly-once, automatic effect rollback and distributed locking/scheduling are not promised. See the 1.2→1.4 migration guide.

Formal publication still requires strict Provider, both-platform engineering/stability/coverage/real PTY and public download/reinstall evidence bound to final main, this version, Runtime/Worker and exact artifacts. Older certifications, historical exceptions, individual engineering runs and documentation synchronization do not qualify this release. Actual availability is determined by same-version Release/registry records and final qualification evidence.

Upgrade and rollback

Requirements remain Node.js ≥ 22.19 and Python ≥ 3.10, with Git ≥ 2.36 for Git tools; supported platforms remain Windows/Linux. Follow the 1.2→1.4 migration guide: stop old processes, back up configuration, Sessions, RunState, Checkpoints and task facts, upgrade the SDK and Worker together, and verify permission, result and error boundaries. Rollback requires stopping new writers and restoring old programs plus complete pre-upgrade state backups. Local restoration does not undo external operations.

Verification and known limitations

Targeted regressions, real offline processes and merge CI do not replace final 1.4.0 release qualification. After freezing main, obtain strict Provider, both-platform stability, coverage, real PTY and exact-artifact public reinstall evidence bound to the same version, final commit, Runtime digest and bundled Worker. Without this evidence, the release is not certified or published. Real Provider external transfer, cost, model, credential source and maximum runtime require explicit approval bound to the final commit and Runtime package SHA256; older authorization cannot be reused.

The earlier local initialize-response timeout still lacks an independent root-cause explanation; this session-path fix does not claim to eliminate that historical observation. The framework does not promise external exactly-once, automatic effect rollback, a background distributed scheduler or an OS sandbox for arbitrary host code. Real external OAuth/MCP interoperability and global paired performance gains are not certified. Release evidence for the underlying runtime migration includes an npm consumer graph without repository overrides and production and full audits for the final commit.

Version, tag, manifest and all ten assets—eight npm packages, one wheel and one standalone source ZIP—must identify the same release. Publication recovery reuses the original immutable artifacts. Actual dependency, copyright, license and NOTICE obligations remain; see the target tag's third-party notices and license. The target tag must exist at formal publication.