fix(repository-cache): isolate Git package dependencies

This commit is contained in:
Tianyi Cui
2026-08-08 19:59:48 +08:00
parent 033fa4b0b1
commit a7832cbfbf
9 changed files with 98 additions and 37 deletions

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-08-08-host-owned-git-repository-plugin-preparation.md
2026-08-08-host-owned-git-repository-plugin-preparation.md: b459184f802981f91f611b0b781651cede0c9efd
2026-08-08-host-owned-git-repository-plugin-preparation.zh.md: 5e24efa59ce44ffa8214f9c5ef9ff722ac0ca73a
2026-08-08-host-owned-git-repository-plugin-preparation.md: 3f92c0b2d782735bfb548711bb13a7a8d03a3824
2026-08-08-host-owned-git-repository-plugin-preparation.zh.md: ef34c99b5d9f31ea13c5fddfee77c30e640d25c9

View File

@@ -10,13 +10,15 @@ The repository Plugin authoring contract depended on `scripts.prepare: "dsh-plug
The lifecycle choice also failed for a selectable `.dsh-plugin` inside a pnpm workspace. pnpm prepares a Git-hosted package by running the repository's preferred package manager before packing the selected subdirectory. A nested `pnpm install` joins the containing workspace and need not execute the unlisted `.dsh-plugin` package's `prepare` script. The install could therefore succeed and publish a cache generation containing only the source package metadata; real DSH startup failed later because `dsh-plugin.mjs` did not exist.
The same workspace discovery could suppress package-owned dependencies after the move to `prepack`. When the source repository carried a root pnpm lockfile but did not list the selected `.dsh-plugin` as a workspace importer, pnpm reported a successful workspace install without installing dependencies declared only by that package. Its TypeScript build then failed because Cordis and the MCP SDK were absent.
The checked-in headless fixture did not catch either defect because it mounted an already prepared wrapper. It proved runtime composition, not GitHub acquisition or package preparation.
## Decision
The authoring format requires a non-empty `scripts.prepack` that invokes `dsh-plugin-prepare` and needs no DSH dependency for that helper. The package may declare its own build and runtime dependencies and run compilation before the helper. pnpm's Git-hosted package preparation invokes `prepack` explicitly after its dependency-install step and before packlist selects the `.dsh-plugin` subtree, so the helper can validate built entries and still copy sibling repository assets such as `../skills` into the package.
`@deepseek-ai/dsh-repository-plugin` materializes short-lived POSIX and Windows command wrappers that invoke its own built `dsh-plugin-prepare` entry. `RepositoryCache` accepts caller-owned executable directories, resolves them absolutely, and prepends them to the credential-scrubbed lifecycle `PATH` passed to bundled pnpm. The command directory exists only for the installation transaction and is removed on success or failure. The repository remains trusted package-manager input: DSH supplies one command, but other lifecycle scripts and dependencies still execute under the existing trust contract.
`@deepseek-ai/dsh-repository-plugin` materializes short-lived POSIX and Windows command wrappers that invoke its own built `dsh-plugin-prepare` entry. `RepositoryCache` accepts caller-owned executable directories, resolves them absolutely, and prepends them to the credential-scrubbed lifecycle `PATH` passed to bundled pnpm. It also prepends a transaction-owned `pnpm` wrapper: the outer install still runs the pinned pnpm entry directly, while pnpm's hard-coded Git-package `pnpm install` reinvokes that same entry with `--ignore-workspace`. The selected package therefore owns dependency resolution even beneath another pnpm lockfile. Both command directories are removed after the child settles. The repository remains trusted package-manager input: DSH supplies the two host commands, but all package lifecycle scripts and dependencies still execute under the existing trust contract.
The Node 24 consumer lane passes an exact source derived from the pull request head repository and SHA. Because that repository is private, the workflow writes a job-scoped Git configuration that uses the read-only job token for GitHub HTTPS and rewrites pnpm's SSH fallback to that authenticated transport. Its built-entry acceptance launches the real `apps/cli/lib/bin.js run` command with a one-run patch selecting a `private: true` GitHub fixture. That fixture installs pinned npm dependencies, type-checks and bundles a TypeScript Cordis entry and MCP server in `prepack`, invokes the host helper, and proves the skill, MCP call, and code entry through real model requests and immutable-cache artifacts. The test fails if CI omits the exact source instead of silently skipping.
@@ -30,14 +32,17 @@ The Node 24 consumer lane passes an exact source derived from the pull request h
**Clone GitHub repositories in DSH and bypass pnpm's Git fetcher.** Rejected because it would duplicate ref resolution, subdirectory selection, dependency installation, packlist behavior, and cache integrity already owned by the pinned package manager.
**Add the in-repository CI fixture to this repository's pnpm workspace.** Rejected because that would repair only the proof fixture and leave an arbitrary selected package vulnerable to its containing repository's workspace membership and lockfile.
## Consequences
- A repository author can commit the fixed `.dsh-plugin/package.json` and source assets to GitHub without publishing either the Plugin or its preparation helper to npm.
- Private GitHub sources use the host's standard Git authentication. CI proves that path with a temporary read-only configuration rather than persistent runner credentials.
- `prepack`, not `prepare`, is part of the pre-release authoring format. It may contain package-owned build steps but must invoke the host helper; missing or empty lifecycle metadata fails installed-package validation instead of producing an ambiguous partial format.
- A selected package in a pnpm repository installs from its own manifest rather than an enclosing workspace. It must declare its dependencies and cannot rely on workspace-only hoisting; ordinary registry and relative `file:` dependencies remain package-owned inputs.
- Exact source strings still identify immutable cache generations; a changed ref or source configuration selects another generation.
- The host supplies only the preparation executable. Package dependencies, compilation, and the trusted `dsh.entry` contribution remain owned by the repository package and the [trusted-code decision](../architecture/2026-08-08-trusted-repository-package-code.md).
## Testing
`packages/ui/app-boot/tests/repository-cache.spec.ts` runs a local Git subpath through bundled pnpm with an injected command directory and proves that visible environment survives while credential-shaped variables are scrubbed. `packages/cordis/repository-plugin/tests/repository-plugin.spec.ts` pins helper-bearing `prepack` metadata and temporary command cleanup. `examples/headless-agent/tests/keyless-smoke.e2e.ts` keeps the checked-in prepared fixture on that source contract. `apps/cli/tests/github-repository-plugin.built.e2e.ts` is the product acceptance: fresh DSH home, exact authenticated private GitHub source, actual built `dsh run`, package-owned TypeScript build, real MCP execution, code-entry transformation, mock LLM request observation, and prepared cache inspection.
`packages/ui/app-boot/tests/repository-cache.spec.ts` runs a package excluded from its source repository's root pnpm lockfile through a local Git subpath and requires a relative `file:` build dependency during `prepack`; it also proves that visible environment survives while credential-shaped variables are scrubbed. `packages/cordis/repository-plugin/tests/repository-plugin.spec.ts` pins helper-bearing `prepack` metadata and temporary command cleanup. `examples/headless-agent/tests/keyless-smoke.e2e.ts` keeps the checked-in prepared fixture on that source contract. `apps/cli/tests/github-repository-plugin.built.e2e.ts` is the product acceptance: fresh DSH home, exact authenticated private GitHub source, actual built `dsh run`, package-owned TypeScript build, real MCP execution, code-entry transformation, mock LLM request observation, and prepared cache inspection.

View File

@@ -10,13 +10,15 @@ repository 插件的创作契约依赖 `scripts.prepare: "dsh-plugin-prepare"`
这种生命周期选择也无法支持 pnpm 工作区内可选的 `.dsh-plugin`。pnpm 会先运行 Git 托管仓库首选的包管理器,再打包选定的子目录,从而准备 Git 托管包。嵌套执行的 `pnpm install` 会加入外层工作区,而不一定执行未列入其中的 `.dsh-plugin` 包的 `prepare` 脚本。因此,安装可能成功并发布一个仅包含源包元数据的缓存 generation随后真实 DSH 启动因 `dsh-plugin.mjs` 不存在而失败。
迁移到 `prepack` 后,同一项 workspace 发现行为还可能抑制包自有依赖。如果源仓库带有根 pnpm lockfile却未把所选 `.dsh-plugin` 列为 workspace importerpnpm 会报告 workspace 安装成功,但不会安装仅由该包声明的依赖。随后其 TypeScript 构建会因缺少 Cordis 和 MCP SDK 而失败。
签入仓库的 headless fixture测试前置数据没有捕获任一缺陷因为它挂载的是已准备好的包装层。它证明的是运行时组合而不是 GitHub 获取或包准备。
## 决策
创作格式要求 `scripts.prepack` 非空且调用 `dsh-plugin-prepare`,使用该辅助程序无需 DSH 依赖。包可以声明自己的构建依赖与运行时依赖并在调用辅助程序前完成编译。pnpm 针对 Git 托管包的准备流程会在依赖安装步骤之后、打包清单选择 `.dsh-plugin` 子树之前显式调用 `prepack`,因此辅助程序可以校验构建入口,并继续把 `../skills` 等同仓库的相邻资源复制进包内。
`@deepseek-ai/dsh-repository-plugin` 会生成临时的 POSIX 和 Windows 命令包装脚本,用于调用其自有的已构建 `dsh-plugin-prepare` 入口。`RepositoryCache` 接受由调用方持有的可执行文件目录,将它们解析为绝对路径,再前置到传给随附 pnpm、已清除凭据的包生命周期 `PATH`该命令目录仅存在于安装事务期间,无论成功还是失败都会被移除。仓库仍是受信任的包管理器输入DSH 提供这一条命令,其他生命周期脚本和依赖仍按既有信任契约执行。
`@deepseek-ai/dsh-repository-plugin` 会生成临时的 POSIX 和 Windows 命令包装脚本,用于调用其自有的已构建 `dsh-plugin-prepare` 入口。`RepositoryCache` 接受由调用方持有的可执行文件目录,将它们解析为绝对路径,再前置到传给随附 pnpm、已清除凭据的包生命周期 `PATH`它还会前置一个由安装事务持有的 `pnpm` 包装命令:外层安装仍直接运行锁定的 pnpm 入口,而 pnpm 为 Git 包硬编码的 `pnpm install` 会通过 `--ignore-workspace` 重新调用同一入口。因此,即使位于另一个 pnpm lockfile 之下,所选包仍自行拥有依赖解析。两个命令目录都会在子进程结算后移除。仓库仍是受信任的包管理器输入DSH 提供这两条宿主命令,但所有包生命周期脚本和依赖仍按既有信任契约执行。
Node 24 消费方 CI 任务会传入从 PRPull Requesthead 仓库和 SHA 派生的精确源。由于该仓库为私有仓库,工作流会写入一份作业作用域的 Git 配置,使用该作业的只读 token 对 GitHub HTTPS 连接进行认证,并将 pnpm 的 SSH 回退路径重写为这一已认证的传输方式。其构建入口验收会启动真实的 `apps/cli/lib/bin.js run` 命令,并通过一个仅作用于当次运行的 patch 选择 `private: true` 的 GitHub fixture。该 fixture 安装固定版本的 NPM 依赖,在 `prepack` 中对 TypeScript Cordis 入口和 MCP server 进行类型检查与打包,调用宿主辅助程序,并通过真实模型请求和不可变缓存产物验证 skill技能、MCP 调用和代码入口。如果 CI 遗漏精确源,测试会失败,而不是静默跳过。
@@ -30,14 +32,17 @@ Node 24 消费方 CI 任务会传入从 PRPull Requesthead 仓库和 SHA
**在 DSH 中克隆 GitHub 仓库,并绕过 pnpm 的 Git 获取器。** 拒绝,因为这会重复实现已由锁定版本的包管理器负责的 ref 解析、子目录选择、依赖安装、打包清单行为和缓存完整性。
**把仓库内的 CI fixture 加入本仓库的 pnpm workspace。** 拒绝,因为这只能修复证明用的 fixture任意所选包仍会受其所在仓库的 workspace membership 与 lockfile 影响。
## 后果
- 仓库作者可以把修复后的 `.dsh-plugin/package.json` 和源资源提交到 GitHub而无需把插件或其准备辅助程序发布到 NPM。
- 私有 GitHub 源使用宿主的标准 Git 认证。CI 使用临时的只读配置而非运行器上的持久凭据来验证该路径。
- 预发布创作格式使用 `prepack` 而不是 `prepare`。其中可以包含包自有构建步骤,但必须调用宿主辅助程序;生命周期元数据缺失或为空会在已安装包校验时失败,而不会留下状态不明的半成品格式。
- pnpm 仓库中的所选包按自身 manifest 安装,而不是按外层 workspace 安装。它必须声明自己的依赖,不能依赖仅由 workspace 提升而可见的包;常规 registry 依赖与相对 `file:` 依赖仍是包自有输入。
- 精确源字符串仍标识不可变缓存 generation改变 ref 或源配置会选择另一个 generation。
- 宿主只提供准备阶段可执行文件。包依赖、编译和受信任的 `dsh.entry` 贡献仍由 repository 包和[受信任代码决策](../architecture/2026-08-08-trusted-repository-package-code.md)负责。
## 测试
`packages/ui/app-boot/tests/repository-cache.spec.ts`用注入的命令目录通过随附 pnpm 运行本地 Git 子路径,并证明可见环境变量得以保留,而名称符合凭据模式的变量会被清除。`packages/cordis/repository-plugin/tests/repository-plugin.spec.ts` 锁定包含辅助命令的 `prepack` 元数据和临时命令清理行为。`examples/headless-agent/tests/keyless-smoke.e2e.ts` 使签入仓库的已准备 fixture 继续符合该源格式契约。`apps/cli/tests/github-repository-plugin.built.e2e.ts` 是产品验收测试:全新的 DSH 主目录、精确且经过认证的私有 GitHub 源、实际构建产物的 `dsh run`、包自有 TypeScript 构建、真实 MCP 执行、代码入口转换、mock LLM大语言模型请求观测以及对已准备缓存的检查。
`packages/ui/app-boot/tests/repository-cache.spec.ts`让一个被源仓库根 pnpm lockfile 排除的包通过本地 Git 子路径运行,并要求 `prepack` 使用相对 `file:` 构建依赖;该测试还证明可见环境变量得以保留,而名称符合凭据模式的变量会被清除。`packages/cordis/repository-plugin/tests/repository-plugin.spec.ts` 锁定包含辅助命令的 `prepack` 元数据和临时命令清理行为。`examples/headless-agent/tests/keyless-smoke.e2e.ts` 使签入仓库的已准备 fixture 继续符合该源格式契约。`apps/cli/tests/github-repository-plugin.built.e2e.ts` 是产品验收测试:全新的 DSH 主目录、精确且经过认证的私有 GitHub 源、实际构建产物的 `dsh run`、包自有 TypeScript 构建、真实 MCP 执行、代码入口转换、mock LLM大语言模型请求观测以及对已准备缓存的检查。