Qualify Desktop update downloads with an isolated local server
使用隔离的本地服务器验证 Desktop 更新下载
Desktop update interaction depends on feed parsing, network failures, download integrity, and platform preparation. Simulated updater events cannot establish that these operations work together. Requiring cloud cache configuration and hardware signing for each feedback cycle makes local product validation depend on release infrastructure.
English
Problem
Desktop update interaction depends on feed parsing, network failures, download integrity, and platform preparation. Simulated updater events cannot establish that these operations work together. Requiring cloud cache configuration and hardware signing for each feedback cycle makes local product validation depend on release infrastructure.
Decision
The Windows local qualification command runs the built production coordinator inside Electron with the real ElectronHttpExecutor and NsisUpdater. An isolated application adapter supplies a test version and private update configuration. A loopback server supplies Nightly YAML, mandatory policy, and inert binary bytes. The installer, external browser, and clipboard calls are replaced with observations. Publisher verification is absent only from this private fixture configuration, never from production configuration.
Each invocation atomically allocates a temporary directory and an operating-system-assigned loopback port. Request barriers establish concurrency without sleeps. Server close and child exit are awaited before temporary files are removed. Independent processes can run concurrently without sharing updater cache or user data. The command requires built Desktop modules and stays separate from source-plane unit tests.
This testing decision supplements the release policy; it does not supersede signing, publication integrity, installed-artifact qualification, or the pending mandatory update proposal. Those records remain active.
Alternatives considered
Use only simulated updater events. They cover state transitions but cannot establish actual YAML parsing, Electron transport, downloaded bytes, or checksum rejection. Local qualification uses real dependency implementations; native-dialog and renderer component tests remain separate.
Require signed cloud-hosted packages for every local iteration. That couples ordinary interaction debugging to signing hardware and CDN configuration. Release qualification still requires those systems, but local download tests do not.
Relax production signing or execute a dummy installer. Neither is needed to observe download readiness and installation authorization. The fixture never loads production update configuration and records installation without launching downloaded bytes.
Consequences
The executable scenarios cover no update, downgrade rejection, HTTP 404/408 and invalid YAML, stalled-feed and stalled-download deadlines, same-address feed replacement, one in-flight check or download, explicit download authorization, SHA-512 mismatch, interrupted transfers, explicit retry, prepared-version retention, restart refusal, shutdown failure, and disposal during a pending check. Native-dialog regressions cover checking feedback, check failure, download refusal, and separate installation action; account-row tests own progress and persistent retry presentation.
The ordinary polling regression uses the real coordinator with a fake clock, instance-local random samples, and deferred network completion. It verifies completion-based jitter, capped exponential backoff, successful reset, manual failure visibility while joining an automatic check, retained download failures and prepared packages, and no timer rearming after disposal. Main-entry tests connect the same deadline to focus, resume, explicit IPC, and application shutdown. Removing failure backoff makes the timing assertion fail. These deterministic tests establish request timing, not fleet capacity or startup-burst distribution.
The actual mandatory-update window, sandboxed preload, and renderer button handlers run against the policy server and updater. A stalled policy request reaches its real deadline without clearing the block. Owner-local DOM expectations pin Chinese text and actions. Screenshot failures are recorded separately; missing screenshots cannot establish visual acceptance.
The workspace browser scenario checks built sidebar composition in both locales against the production presentation function. A page-local carrier supplies update state and records actions; the Host, client plugins, styles, and connection remain real. This separates slot placement and click-guard evidence from Electron IPC, task authorization, and installation evidence. Screenshots and result files use an invocation-specific ignored directory and do not replace release qualification.
The Host qualification runner loads the built Host and standard agent preset in a private profile. Real task registries, queued messages, pending tool questions/approvals, and Node jobs establish task-detection and admission-lock evidence without replacing Host composition. Scripted model output and held human answerers control the waiting points; cancellation and process exit establish completion. This evidence excludes the development launcher's dependency projection, Electron installation confirmation, and failed task shutdown. Concurrent invocations own separate home, project, and session directories.
The Electron workspace runner couples the compiled main entry and actual preload to a separate real Host. A private profile plugin controls real queued tasks and deliberately holds teardown; local HTTP delivery replaces release infrastructure, and the updater's installer call is intercepted. It verifies confirmation-time task changes, refused installation after failed teardown, replacement Host readiness, and fresh confirmation under ordinary and mandatory policy. Renderer actions wait for finite layout animations and reject obscured targets; window-listener cleanup retains web contents independently of destroyed windows. The runner does not prove bundled-main deployment or operating-system composition of overlapping windows.
The runner launches Electron without hiding its GUI and asserts main-window visibility before input or screenshots. Hidden or occluded renderers can suspend frame callbacks even when DOM queries complete. Each DOM, click-layout, and screenshot operation temporarily disables background throttling on its own WebContents, then restores and checks the original setting. Real animations and click-obscuration assertions remain enabled; production window configuration is unchanged. A main-process deadline and invocation-local trace bound each operation; failure diagnostics record window visibility and animation state with their own deadline. Diagnostic capture failure cannot replace the original test failure. Re-enabling Windows process hiding fails the visibility assertion, and concurrent unprotected windows reproduce a suspended frame wait.
The production DesktopUpdateHttpExecutor retains the library transport and adds an inactivity deadline to actual Electron requests. Electron 44 emits writable close before response headers, so that event cannot indicate HTTP completion. Response completion, abort, and errors release the timer; received bytes refresh it. Both stalled headers and stalled payloads fail and recover through an explicit retry in the local fixture.
The signed-download runner complements inert-byte tests with existing signed and unsigned executables, a public certificate, and real Windows Authenticode verification. Its private feed uses a synthetic version, and installation is intercepted. Correct-hash files with a wrong publisher or no signature are rejected and removed from cache; corrupt transfer bytes fail checksum verification first. Automatic checks retain the failure without another request, while explicit retry with a signed payload reaches readiness and requires separate restart approval. Inputs remain unchanged. Each process owns its port and cache; the deadline stops its complete Electron/PowerShell process tree before runtime cleanup. Missing publisher configuration makes the rejection assertion fail.
Optional old-installer input enables real differential reconstruction from private cache and original blockmaps. The server supports single and multipart byte ranges, with explicit missing-blockmap and rejected-range failures. Hash and signature verification cover reconstructed and full-fallback results; request records require actual byte reuse and reject full fallback as differential success. Disabling differential download fails the blockmap-request assertion. Historical input files remain read-only, and independent concurrent runs retain separate caches and ports.
The local download fixture injects ENOSPC into its executable write stream after writing partial bytes to disk. The actual updater must clear the partial executable, refuse installation, retain explicit retry, and download verified bytes after the original writer is restored. Only the private download path is intercepted, and stream closure is awaited. This qualifies handling of a filesystem write failure without exhausting a shared host volume.
This evidence does not establish actual installer execution, post-restart health, production CDN Range behavior, macOS updater preparation, or actual volume exhaustion. The test servers publish nothing and never read cloud credentials; production delivery and signed installed-version upgrades remain separate qualification.
中文
问题
Desktop 更新交互依赖清单解析、网络失败、下载完整性和平台准备。模拟 updater 事件无法证明这些操作能够协同工作。每轮反馈都要求配置云端缓存并完成硬件签名，会让本地产品验证依赖发布基础设施。
决策
Windows 本地验证命令 在 Electron 中运行构建后的生产协调器，并使用真实的 ElectronHttpExecutor 和 NsisUpdater。隔离的应用适配器提供测试版本和私有更新配置。回环服务器提供 Nightly YAML、强更策略和不可执行的二进制字节。安装器、外部浏览器和剪贴板调用被替换为观测记录。发布者验证仅在此私有 fixture（测试前置数据）配置中缺省，生产配置不受影响。
每次调用都原子分配临时目录和由操作系统分配的回环端口。请求屏障在不使用休眠的情况下建立并发条件。移除临时文件之前，测试会等待服务器关闭和子进程退出。独立进程可以并发运行，互不共享 updater 缓存或用户数据。此命令依赖构建后的 Desktop 模块，与源码侧单元测试分开运行。
此测试决策补充发布策略，并不取代签名、发布完整性、已安装产物验收或待实施的强制更新提案。这些记录保持有效。
考虑过的替代方案
仅使用模拟 updater 事件。 它们覆盖状态转换，但无法证明真实的 YAML 解析、Electron 传输、下载字节或校验和拒绝。本地验证使用真实依赖实现；原生弹窗和渲染器组件测试仍独立执行。
每轮本地迭代都要求云端托管的签名包。 这会让常规交互调试依赖签名硬件和 CDN 配置。发布验收仍需要这些系统，但本地下载测试不需要。
放宽生产签名要求或执行伪安装器。 观测下载就绪和安装授权不需要这两种做法。fixture 从不加载生产更新配置，只记录安装，不启动下载的字节。
后果
可执行场景覆盖无更新、拒绝降级、HTTP 404/408 和无效 YAML、清单与下载停滞截止时间、同地址清单替换、单一在途检查或下载、显式下载授权、SHA-512 不匹配、传输中断、显式重试、已准备版本保留、拒绝重启、关闭失败，以及检查未完成时的 dispose（资源释放）。原生弹窗回归覆盖检查反馈、检查失败、拒绝下载和独立安装操作；账户行测试负责进度和持久重试展示。
常规轮询回归使用真实协调器、模拟时钟、实例私有随机样本和可控的网络完成时机。它验证从完成时刻起算的抖动、有上限的指数退避、成功重置、手动复用自动检查时的失败可见性、下载失败和已准备安装包保留，以及 dispose 后不重新启动计时器。主入口测试将同一期限连接到窗口聚焦、系统恢复、显式 IPC 和应用退出。移除失败退避会使计时断言失败。这些确定性测试证明请求时序，不证明客户端总体容量或启动峰值分布。
真实的强更窗口、沙箱预加载和页面按钮处理会对接策略服务器与 updater。停滞的策略请求达到真实截止时间后不会清除阻塞。所属测试目录内的 DOM 预期输出约束中文文案和操作。截图失败单独记录；缺少截图不能证明视觉验收通过。
工作区浏览器场景 使用生产展示函数检查两种语言下构建后的侧栏组合。页面私有的载体提供更新状态并记录操作；Host、客户端插件、样式和连接保持真实。slot 位置与点击保护的证据因此与 Electron IPC、任务授权和安装证据分开。截图和结果文件使用每次调用独有的忽略目录，不替代发布验收。
Host 验证运行器 在私有 profile 中加载构建后的 Host 和 standard agent 预设。真实任务注册表、排队消息、等待答复的工具提问／审批和 Node job 提供任务探测与准入锁定证据，不替换 Host 组合。脚本化模型输出和暂停的人工答复器控制等待点；取消和进程退出确立完成状态。这些证据不包含开发启动器的依赖投影、Electron 安装确认和任务停止失败。并发调用各自拥有独立的 home、项目和会话目录。
Electron 工作区运行器 将编译后的主入口和真实预加载连接到独立的真实 Host。私有 profile 插件控制真实排队任务并刻意阻塞停止过程；本地 HTTP 分发替代发布基础设施，updater 的安装器调用被拦截。它验证确认期间任务变化、收尾失败后拒绝安装、替代 Host 就绪，以及普通和强更策略下的重新确认。页面操作等待有限布局动画完成，并拒绝被遮挡的目标；窗口监听清理独立保留 web contents，不依赖已销毁窗口。运行器不证明打包主入口部署或重叠窗口的操作系统合成效果。
运行器启动 Electron 时不隐藏 GUI，并在输入或截图之前断言主窗口可见。隐藏或被遮挡的渲染器即使能够完成 DOM 查询，也可能暂停帧回调。每次 DOM、点击布局和截图操作仅临时关闭自身 WebContents 的后台节流，随后恢复并检查原设置。真实动画和点击遮挡断言保持启用；生产窗口配置不变。每项操作都有主进程截止时间及本次运行独有的跟踪记录；失败诊断在独立截止时间内记录窗口可见性和动画状态。诊断捕获失败不能替换原始测试失败。重新启用 Windows 进程隐藏会使可见性断言失败，未保护的并发窗口可复现帧等待停滞。
生产使用的 DesktopUpdateHttpExecutor 保留依赖库传输，并为真实 Electron 请求添加无活动截止时间。Electron 44 会在响应头到达前触发可写流的 close，因此该事件不能代表 HTTP 完成。响应完成、中止和错误释放计时器；收到字节会刷新计时器。响应头停滞与负载停滞均在本地 fixture 中失败，并通过显式重试恢复。
签名下载运行器 使用已有的签名及未签名可执行文件、公开证书和真实 Windows Authenticode 验签，补充不可执行字节测试。私有清单使用合成版本号，安装调用被拦截。哈希正确但发布者错误或未签名的文件被拒绝，并从缓存移除；传输损坏的字节先被校验和验证拒绝。自动检查保留失败，不发送新请求；显式重试签名文件后进入就绪，并要求单独批准重启。输入文件保持不变。每个进程独占端口和缓存；截止时间触发时先停止完整的 Electron/PowerShell 进程树，再清理运行时目录。缺少发布者配置会使拒绝断言失败。
可选的旧安装器输入启用真实差分重建，使用私有缓存和原始 blockmap。服务器支持单段与多段字节范围，并显式提供缺少 blockmap 和拒绝 Range 的故障。哈希与签名验证覆盖重建及全量回退结果；请求记录要求实际复用字节，不允许全量回退冒充差分成功。禁用差分下载会使 blockmap 请求断言失败。历史输入文件保持只读，独立并发运行各自拥有缓存和端口。
本地下载 fixture 在向磁盘写入部分字节后，为其可执行文件写入流注入 ENOSPC。真实 updater 必须清理不完整的可执行文件、拒绝安装、保留显式重试，并在恢复原始写入器后下载通过校验的字节。只拦截私有下载路径，且等待流关闭。这验证文件系统写入失败的处理，不耗尽共享主机卷。
这些证据不能证明真实安装器执行、重启后健康状态、生产 CDN Range 行为、macOS updater 准备或真实卷耗尽。测试服务器不发布任何内容，也不读取云端凭据；生产分发和签名已安装版本升级仍需独立验收。