fix(skill): retain candidates across watcher failures

This commit is contained in:
Tianyi Cui
2026-07-29 21:08:10 +08:00
parent 48cf50fc55
commit 52d68a5383
23 changed files with 175 additions and 68 deletions
@@ -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/feature/2026-07-27-skill-catalog-hot-refresh.md
2026-07-27-skill-catalog-hot-refresh.md: e7cff2cb53a044ed0c4789cef3550652903f3586
2026-07-27-skill-catalog-hot-refresh.zh.md: 86519c93880f94b1b9d3bdc011aa09b04ca10686
2026-07-27-skill-catalog-hot-refresh.md: 7c81b287cecde60c42f7e1e3ec171244ffc18aa6
2026-07-27-skill-catalog-hot-refresh.zh.md: ca570a3c6a0402e6764824e3101e15054769b85d
@@ -12,11 +12,11 @@ Filesystem updates are also non-atomic from the observer's perspective. An edito
## Decision
The skill capability separates catalog membership from instruction-body loading. `ctx.skills.snapshot()` returns summaries plus a completeness bit. `ctx.skills.registerProvider(factory)` gives the synchronous factory one registration-scoped `{ signal, invalidate }` control: `invalidate()` dirties only that exact active registration and discards completed catalog caches, while the signal aborts when registration fails or is disposed. A provider or runtime generation change during discovery retries before returning. Incomplete observations are not cached. A late invalidation after disposal or replacement is a no-op because the capability has been revoked.
The skill capability separates catalog membership from instruction-body loading. `ctx.skills.snapshot()` returns summaries plus a completeness bit. `ctx.skills.registerProvider(factory)` gives the synchronous factory one registration-scoped `{ signal, invalidate }` control: `invalidate()` dirties only that exact active registration and discards completed catalog caches, while the signal aborts when registration fails or is disposed. Provider arrays are complete-discovery shorthand; an explicit incomplete observation can retain readable candidates for direct loads without becoming cacheable or authoritative for model-facing consumers. A provider or runtime generation change during discovery retries before returning. A late invalidation after disposal or replacement is a no-op because the capability has been revoked.
`@deepseek-ai/dsh-skill-local` directly depends on Chokidar and observes catalog-relevant host paths. Existing roots watch direct skill bundle directories, flat Markdown entries, and direct `SKILL.md` entry files. Additions, removals, and directory changes invalidate membership; file changes support frontmatter `name` and `description` refresh. Resource files below a bundle are ignored. Events in one microtask batch coalesce to one invalidation. Project watchers use a bounded least-recently-observed set.
A missing root is followed from its nearest existing ancestor one absent segment at a time with `fs.watchFile`, then handed to Chokidar once the real root exists. Before scanning, each discovery re-probes the retained root/ancestor mode. That independent probe re-establishes ancestor observation after deletion even when child removals invalidate and publish an authoritative empty catalog before, or without, a root `unlinkDir` event. Chokidar configuration exposes native-versus-polling mode, write stability, polling interval, symlink following, and project watcher capacity. First-party `write` and `edit` tool observations synchronously invalidate a relevant provider, so the next model step sees its own mutation without waiting for host delivery. Watch startup/runtime failures make discovery incomplete and retry; teardown closes watchers and ignores late callbacks.
A missing root is followed from its nearest existing ancestor one absent segment at a time with `fs.watchFile`, then handed to Chokidar once the real root exists. Before scanning, each discovery re-probes the retained root/ancestor mode. That independent probe re-establishes ancestor observation after deletion even when child removals invalidate and publish an authoritative empty catalog before, or without, a root `unlinkDir` event. Chokidar configuration exposes native-versus-polling mode, write stability, polling interval, symlink following, and project watcher capacity. First-party `write` and `edit` tool observations synchronously invalidate a relevant provider, so the next model step sees its own mutation without waiting for host delivery. Watch startup/runtime failures are logged and retried; discovery still returns readable candidates for direct loads but reports an incomplete observation. Teardown closes watchers and ignores late callbacks.
`@deepseek-ai/dsh-tool-skill` injects the first non-empty complete catalog as a durable sourced `user/message` on the first complete `agent/step` that observes one. At every `agent/step` it applies exact `skill` tool visibility, hashes the exact rendered text between the `<available_skills>` tags, and scans the read-only session events backwards without copying them for the newest recognizable visible catalog from this plugin. A changed digest appends a durable, complete replacement through `agent.inject()`, including an explicit empty catalog when all skills disappear. If no catalog remains visible but a recognizable one exists in historical events, compaction hid it and the next complete observation re-establishes the current catalog, including an empty tombstone. A current empty catalog with no historical publication emits nothing, while an incomplete snapshot preserves the last-good model view. The backward scan normally stops at the newest visible catalog; when compaction hides every catalog it pays an O(session-events) scan to recover that fact.
@@ -26,7 +26,7 @@ Instruction bodies keep progressive disclosure. Every `skill(name)` call asks th
## Verification
Registry tests pin registration-scoped invalidation, revocation, signal abort, contained observer failures, incomplete snapshots, generation retries, and stale-name rejection. Local-provider tests cover bundle and flat-file creation, removal, rename, root creation/deletion/recreation including an unobserved root `unlinkDir`, description changes, body-only edits, first-party observation, symlinks, polling options, watcher failures, event coalescing, bounded projects, teardown, and transient reads. Tool tests pin full replacement messages, empty tombstones, digest stability for body-only edits, incomplete-state retention, visibility, and resume metadata. TUI tests pin last-complete retention, authoritative empty removal, latest-wins refresh, teardown, and the already-open slash-draft race; a real Loader/PTY smoke adds a local skill after startup and observes its completion without restarting. A keyless assembled agent-spine snapshot creates a project skill through model-facing filesystem tools, observes its replacement catalog on the next request, and loads its current body with the real `skill` tool.
Registry tests pin registration-scoped invalidation, revocation, signal abort, contained observer failures, incomplete candidates, generation retries, and stale-name rejection. Local-provider tests cover bundle and flat-file creation, removal, rename, root creation/deletion/recreation including an unobserved root `unlinkDir`, description changes, body-only edits, first-party observation, symlinks, polling options, persistent watcher failures with loadable candidates, event coalescing, bounded projects, teardown, and transient reads. Tool tests pin full replacement messages, empty tombstones, digest stability for body-only edits, incomplete-state retention, visibility, and resume metadata. TUI tests pin last-complete retention, authoritative empty removal, latest-wins refresh, teardown, and the already-open slash-draft race; a real Loader/PTY smoke adds a local skill after startup and observes its completion without restarting. A keyless assembled agent-spine snapshot creates a project skill through model-facing filesystem tools, observes its replacement catalog on the next request, and loads its current body with the real `skill` tool.
## Alternatives considered
@@ -12,11 +12,11 @@ skill(技能)摘要是模型的路由输入,但本地 skill 可在会话
## 决策
skill 服务将目录成员关系与指令正文加载分离。`ctx.skills.snapshot()` 返回摘要及一个完整性位。`ctx.skills.registerProvider(factory)` 会向同步工厂提供一项注册作用域内的 `{ signal, invalidate }` 控制能力:`invalidate()` 只会将该精确活动注册标记为脏,并丢弃已完成目录缓存;注册失败或释放时,信号会中止。在发现期间,如果提供方或运行时 generation 发生变化,系统会先重试再返回。不完整的观察结果不会缓存。资源释放或替换后的延迟失效操作不会执行任何操作,因为该能力已被撤销。
skill 服务将目录成员关系与指令正文加载分离。`ctx.skills.snapshot()` 返回摘要及一个完整性位。`ctx.skills.registerProvider(factory)` 会向同步工厂提供一项注册作用域内的 `{ signal, invalidate }` 控制能力:`invalidate()` 只会将该精确活动注册标记为脏,并丢弃已完成目录缓存;注册失败或释放时,信号会中止。提供方返回的数组是完整发现的简写形式;显式的不完整观测可以保留可读候选项供直接加载,但不能缓存,也不能作为面向模型消费方的权威结果。在发现期间,如果提供方或运行时 generation 发生变化,系统会先重试再返回。资源释放或替换后的延迟失效操作不会执行任何操作,因为该能力已被撤销。
`@deepseek-ai/dsh-skill-local` 直接依赖 Chokidar,并观察与目录相关的宿主路径。已有根目录会监视其直属 skill bundle 目录、平铺的 Markdown 条目和直属 `SKILL.md` 条目文件。新增、移除和目录变更会使成员关系失效;文件变更还支持刷新 frontmatter 中的 `name` 和 `description`。bundle 内更深层的资源文件会被忽略。同一微任务批次中的事件会合并为一次失效。项目 watcher 使用有界集合,并按最久未观察顺序淘汰。
系统从缺失根目录最近的现有祖先开始,使用 `fs.watchFile` 每次跟进一层缺失路径片段;真实根目录出现后,再交给 Chokidar。每次发现操作都会在扫描前重新探测所保留的根目录/祖先模式。即使子项移除在根目录 `unlinkDir` 事件之前就触发失效并发布权威空目录,或者该事件根本没有到达,这项独立探测也会在删除后重新建立祖先观察。Chokidar 配置公开原生事件或轮询模式、写入稳定性、轮询间隔、符号链接跟随选项和项目 watcher 容量。第一方 `write` 和 `edit` 工具观察会同步使相关提供方失效,因此下一个模型步骤无需等待宿主事件投递,就能看到自身改动。watcher 启动或运行失败会使发现结果不完整并触发重试;资源销毁会关闭 watcher,并忽略延迟回调。
系统从缺失根目录最近的现有祖先开始,使用 `fs.watchFile` 每次跟进一层缺失路径片段;真实根目录出现后,再交给 Chokidar。每次发现操作都会在扫描前重新探测所保留的根目录/祖先模式。即使子项移除在根目录 `unlinkDir` 事件之前就触发失效并发布权威空目录,或者该事件根本没有到达,这项独立探测也会在删除后重新建立祖先观察。Chokidar 配置公开原生事件或轮询模式、写入稳定性、轮询间隔、符号链接跟随选项和项目 watcher 容量。第一方 `write` 和 `edit` 工具观察会同步使相关提供方失效,因此下一个模型步骤无需等待宿主事件投递,就能看到自身改动。watcher 启动或运行失败会被记录并触发重试;发现过程仍会返回可读候选项供直接加载,但会报告不完整观测。资源销毁会关闭 watcher,并忽略延迟回调。
`@deepseek-ai/dsh-tool-skill` 在 `agent/step` 首次观察到非空完整目录时,将该目录注入为一条持久且带来源的 `user/message`。每次 `agent/step`,它都会应用 `skill` 工具的精确可见性,对 `<available_skills>` 标签之间精确渲染的文本计算哈希,并从后向前扫描只读会话事件且不复制,以查找该插件发布的最新一条可识别且仍可见的目录。digest 变化时,插件通过 `agent.inject()` 追加一份持久的完整替换目录;所有 skill 消失时,也会追加显式空目录。如果没有目录仍然可见,但历史事件中存在可识别目录,则说明压缩(compaction)已将其遮蔽,下一次完整观察会重新建立当前目录,包括空 tombstone。如果当前目录为空且历史上从未发布目录,则不发送任何内容;不完整快照则保留最后一次完整的模型视图。反向扫描通常在最新且仍可见的目录处停止;当压缩遮蔽所有目录时,它会以一次 O(session-events) 扫描的成本确认这一事实。
@@ -26,7 +26,7 @@ TUI 将同一失效通知作为界面状态而非会话历史来消费。`skills
## 验证
注册表测试固定了注册作用域内的失效、能力撤销、信号中止、监听器失败隔离、不完整快照、generation 重试和陈旧名称拒绝。local-provider 测试覆盖 bundle 与平铺文件的创建、移除和重命名,以及根目录创建/删除/重建(包括未观测到根目录 `unlinkDir` 事件的情形)、描述变更、仅正文编辑、第一方观察、符号链接、轮询选项、watcher 失败、事件合并、项目 watcher 容量上限、资源销毁和暂时读取。工具测试固定了完整替换消息、空 tombstone、仅修改正文时 digest 稳定、不完整状态保留、可见性和恢复元数据。TUI 测试固定了上一份完整结果保留、权威空结果清除、刷新时以最新结果为准、资源销毁和已打开斜杠草稿的竞态;一项使用真实 Loader/PTY 的 smoke 测试会在启动后添加本地 skill,并观察其补全项出现,而无需重启。一个无密钥、装配完成的 agent-spine 快照测试通过面向模型的文件系统工具创建项目 skill,观察下一次请求中的替换目录,并使用真实 `skill` 工具加载当前正文。
注册表测试固定了注册作用域内的失效、能力撤销、信号中止、监听器失败隔离、不完整候选项、generation 重试和陈旧名称拒绝。local-provider 测试覆盖 bundle 与平铺文件的创建、移除和重命名,以及根目录创建/删除/重建(包括未观测到根目录 `unlinkDir` 事件的情形)、描述变更、仅正文编辑、第一方观察、符号链接、轮询选项、候选项仍可加载的持续 watcher 失败、事件合并、项目 watcher 容量上限、资源销毁和暂时读取。工具测试固定了完整替换消息、空 tombstone、仅修改正文时 digest 稳定、不完整状态保留、可见性和恢复元数据。TUI 测试固定了上一份完整结果保留、权威空结果清除、刷新时以最新结果为准、资源销毁和已打开斜杠草稿的竞态;一项使用真实 Loader/PTY 的 smoke 测试会在启动后添加本地 skill,并观察其补全项出现,而无需重启。一个无密钥、装配完成的 agent-spine 快照测试通过面向模型的文件系统工具创建项目 skill,观察下一次请求中的替换目录,并使用真实 `skill` 工具加载当前正文。
## 考虑过的替代方案