runner-owned temporary storage for PR CI
PR CI 使用 runner 管理的临时存储
The Linux failover pool runs multiple runner instances on one VM. PR coverage and snapshot processes use the operating-system temporary directory for transformed modules and fixtures. Files outside the runner's temporary directory escape its job cleanup, including when cancellation prevents process-level disposal. Exhausting that shared directory makes unrelated PRs fail before tests execute.
English
Problem
The Linux failover pool runs multiple runner instances on one VM. PR coverage and snapshot processes use the operating-system temporary directory for transformed modules and fixtures. Files outside the runner's temporary directory escape its job cleanup, including when cancellation prevents process-level disposal. Exhausting that shared directory makes unrelated PRs fail before tests execute.
Decision
The static, coverage, and consumer jobs in PR CI export TMPDIR=runner.temp through GITHUB_ENV in their first step before any setup or test process starts. Node, Vite, tsx, and temporary test consumers inherit the runner-owned location. Each runner owns its directory and GitHub Actions clears its removable contents at job start and completion; fixtures still allocate unique children and retain their own cleanup.
npm keeps its configured persistent cache, normally $HOME/.npm on POSIX, without a per-job override in the main CI or release workflows. The pnpm store remains shared at $HOME/.local/share/pnpm/store. Both retain cross-runner reuse under the package managers’ concurrent-access support; shared-cache capacity and filesystem failures remain operational responsibilities. The consumer job places Playwright browser downloads and installation locks beside RUNNER_TEMP; hosted cache restore uses that same location.
The release rehearsal decision applies the same lifetime rule to release consumers. The failover runbook continues to own runner selection and shared-host capacity. This change does not retarget jobs, reduce concurrency, retry tests, weaken assertions, or modify master-only CI.
Recorded ACP completion order
The ACP diagnostic scenario holds its scripted background response until job_output owns the completion wait. Without that synchronization, a fast child can publish a legitimate job notice between the recorded parent steps. A scenario-local wrapper releases the child after the jobs service registers the completion waiter; the mock watches an exclusive marker in the private test workspace and closes the watcher after release. The fixture restores the wrapped method on disposal. The recorded Session bytes and production job-notice behavior stay unchanged.
Workspace-grant fixture placement
The headless session-sandbox-root fixture declares workspace.parent: outside-temp, not a home-filesystem dependency. Its allocator uses a sibling of the canonical platform temp root where the parent is writable and avoids system temporary grants, otherwise home, and rejects a cwd already covered by automatic temporary write grants. On the failover runner this keeps the test on the data volume without making its write succeed through a temporary-directory exemption. The filesystem-sandbox containment tests use the same allocator for their workspace and denied sibling; they register cleanup immediately after successful acquisition. Atomic workspace allocation, recorded Session bytes, and the independent expected file remain unchanged.
Live verification and browser fixture inputs
The installed-wheel live SDK test externally replaces the created file with a fresh host-only challenge before asking the model to verify it; the verification prompt does not reveal that value. Both turns must contain model-requested tool calls, and the verifier compares the returned value and actual file bytes.
The reference-composer fixture maps the known home-abbreviated workspace display to its existing cwd token and waits for the current exact suggestion set before selecting; neither host paths nor stale suggestions determine its result. The shared browser timezone, Inspector subscription synchronization, and PowerShell completion behavior follow the existing platform-test decision.
The advanced Python snapshot pauses only its matching workflow child’s first pre-step until the parent’s durable workflow membership event is observed. The fixture supports either event-arrival order and cancels pending waits on abort or disposal. This pins the scenario’s cross-session ordering without sorting notifications or changing production scheduling.
The queue snapshot moves the pointer away from the Stop/Send control and waits for its Send tooltip to close before capture. Workspace-management tests select the sole non-blank Session by its actions affordance, not row position, and select that Session before asserting that archiving it removes the empty Ungrouped bucket. Hover behavior, queue contents, durable archive identity, and reload assertions remain unchanged. The concurrent spill isolation test keeps each root paired with its run result rather than assuming filesystem allocation completion order matches input order.
Alternatives considered
Delete shared temporary files from a PR job. Another runner may still own those files. Repository jobs must not reclaim a shared directory by pathname or age.
Retry tests or enlarge timeouts. Neither recovers storage or gives residual files a cleanup owner.
Switch every job to hosted runners. This avoids the affected VM but leaves the failover path defective and changes the operator's independent pool selection.
Consequences
Output honoring TMPDIR follows the job lifetime instead of accumulating in unmanaged host storage. Package-manager and browser caches remain persistent. This does not reclaim existing shared temporary files, guarantee filesystem capacity, or clean files the runner account cannot remove. Operators still own historical residue, disk provisioning, and jobs outside this PR workflow.
Linux bwrap and Landlock workspace-write profiles grant literal /tmp and the workspace, not an inherited TMPDIR outside it; confined fixtures must place temporary writes in those granted paths. The snapshot spill helper separates fixed-length logical locators from atomically allocated live storage. A fixture-only adapter delegates saves to the real local spill provider and resolves only locators saved by that run to their live files. Recorded preview lengths, omission counts, and retrieval assertions remain unchanged; no files are allocated at the logical /tmp/dsh-acp-snap-* prefix. This change does not widen product sandbox grants.
The parsed-workflow cases in ci-workflow.spec.ts require the assignment on all three workers and reject step-level overrides. They fail against the unmodified workflow. Independent-process smoke checks and repeated PR runs validate the actual tooling; the YAML assertions alone do not prove host capacity.
中文
问题
Linux 故障切换池在同一台虚拟机上运行多个 runner 实例。PR 覆盖率和快照进程使用操作系统临时目录存放转换后的模块与测试夹具。runner 临时目录之外的文件不受作业清理管理，取消执行阻止进程自行清理时也不例外。该共享目录耗尽会让无关 PR 在测试执行前就失败。
决策
PR CI 的静态检查、覆盖率和消费者作业在任何准备或测试进程启动前，在首个步骤通过 GITHUB_ENV 导出 TMPDIR=runner.temp。Node、Vite、tsx 和临时测试消费者继承 runner 管理的位置。每个 runner 管理自己的目录，GitHub Actions 在作业开始和完成时清除其中可删除的内容；测试夹具仍分配唯一子目录，并保留自身清理逻辑。
npm 保留配置的持久化缓存，在 POSIX 上通常为 $HOME/.npm；主 CI 和发布工作流不设置每作业覆盖。pnpm store 保持共享于 $HOME/.local/share/pnpm/store。两者依靠包管理器的并发访问支持保留跨 runner 复用；共享缓存容量及文件系统故障仍由运维负责。消费者作业将 Playwright 浏览器下载和安装锁放在 RUNNER_TEMP 旁；托管缓存恢复使用同一位置。
发布演练决策 对发布消费者采用相同的生命周期规则。故障切换运行手册 继续负责 runner 选择和共享主机容量。本变更不调整作业目标、不降低并发、不重试测试、不削弱断言，也不修改仅在 master 上执行的 CI。
ACP 完成顺序的录制
ACP 诊断场景 暂停脚本化的后台响应，直到 job_output 开始等待完成。没有这种同步，快速子进程可能在录制的父步骤之间发布合法的作业通知。场景本地 wrapper 在 jobs 服务注册完成等待器后释放子进程；mock 在测试私有 workspace 中监听独占创建的标记，并在释放后关闭 watcher。夹具在销毁时恢复被包装的方法。录制的 Session 字节和生产作业通知行为保持不变。
Workspace 授权夹具的位置
Headless 的 session-sandbox-root 夹具声明 workspace.parent: outside-temp，而不是依赖 home 所在文件系统。分配器在父目录可写且避开系统临时授权时选择规范化平台临时根目录的同级目录，否则使用 home，并拒绝已被自动临时写授权覆盖的 cwd。在故障切换 runner 上，这让测试留在数据卷中，同时不会让写入借助临时目录豁免而成功。文件系统沙箱的包含关系测试使用同一分配器创建 workspace 及被拒绝的同级目录，并在成功获取目录后立即注册清理。原子 workspace 分配、录制的 Session 字节以及独立预期文件保持不变。
在线验证与浏览器夹具输入
已安装 wheel 的在线 SDK 测试在要求模型验证前，由外部将创建的文件替换为新的、仅主机知道的挑战值；验证提示不暴露该值。两个 turn 必须包含模型请求的工具调用，验证器同时比较返回值及真实文件字节。
Reference-composer 夹具将已知的 home 缩写 workspace 显示映射到既有 cwd token，并在选择前等待当前精确建议集；主机路径或过时建议都不决定测试结果。共享浏览器时区、Inspector 订阅同步及 PowerShell 完成行为遵循既有平台测试决策。
高级 Python 快照仅暂停其匹配的 workflow 子进程首次 pre-step，直到观察到父 Session 的持久化 workflow 成员事件。夹具支持事件先到或等待先建立两种顺序，并在取消或销毁时结束未完成等待。这固定了场景的跨 Session 顺序，而不排序通知或改变生产调度。
Queue 快照在捕获前将指针移离 Stop/Send 控件，并等待其 Send tooltip 关闭。Workspace-management 测试通过操作按钮定位唯一非空 Session，而不依赖行位置；在断言归档会移除空的 Ungrouped 分组前，先选中该 Session。Hover 行为、队列内容、持久化归档身份及重载断言保持不变。并发 spill 隔离测试将每个根目录与其运行结果关联，而不假设文件系统分配完成顺序与输入顺序一致。
考虑过的替代方案
由 PR 作业删除共享临时文件。 其他 runner 可能仍在使用这些文件。仓库作业不得按路径或文件年龄回收共享目录。
重试测试或增大超时。 两者都不能恢复存储空间，也不能为残留文件指定清理责任方。
将所有作业切换到托管 runner。 这能避开受影响的虚拟机，但故障切换路径的缺陷仍在，也会改变运维人员独立选择的执行池。
影响
遵循 TMPDIR 的输出随作业生命周期清理，而不累积于无人管理的主机存储。包管理器和浏览器缓存保持持久化。本方案不回收既有共享临时文件、不保证文件系统容量，也不清理 runner 账号无权删除的文件。历史残留、磁盘配置以及本 PR 工作流以外的作业仍由运维人员负责。
Linux bwrap 和 Landlock 的 workspace-write profile 允许写入字面路径 /tmp 和 workspace，而不允许写入其外部继承的 TMPDIR；受限测试夹具必须将临时写入放在这些已授权路径中。快照 spill helper 将固定长度的逻辑定位符与原子分配的实际存储分开。仅用于夹具的适配器将保存操作委托给真实的本地 spill provider，并仅将本次运行已保存的定位符解析到实际文件。录制的预览长度、省略计数及检索断言保持不变；逻辑 /tmp/dsh-acp-snap-* 前缀下不分配文件。本变更不扩大产品沙箱授权。
ci-workflow.spec.ts 的 YAML 解析用例要求三个 worker 都包含该赋值，并拒绝步骤级别的覆盖。它们在未修改的工作流上失败。独立进程 smoke 检查和重复 PR 运行验证实际工具链；仅有 YAML 断言不能证明主机容量充足。