docs(skill): document the user-explicit invocation path

Bilingual README updates for the four touched packages (ui-skill's claim
flow and deterministic-injection model experience, the apiproxy skills
domain, the shared renderSkillContent seam export, the catalog stitch
sentence), the implemented Agent Note triplet recording the decision and
its peer-product evidence, and the regenerated catalogs/graphs.

packages/client/ui-skill/README.i18n.yaml

@@ -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 packages/client/ui-skill/README.md
-README.md: f70bd2780f255cd8e0c64acb3da3863e10c4fa9d
-README.zh.md: 6eb6cbd3ae196a540e161a3a23f9df2136824f2e
+README.md: c888622bc92b038413c7d0ebf63abb61b483f6f5
+README.zh.md: 3bbbc90186726356c53f375bb664d678c4926988

packages/client/ui-skill/README.md

@@ -2,7 +2,9 @@
 
 English | [中文](README.zh.md)
 
-Skill reference source, browser half: registers the `/`-trigger `skill` source into `ctx.slash`. Ordinary-session candidates come from the `skill.list` RPC addressed by the per-call `ClientSessionContext` projection's `{sessionId}`, with the host resolving `cwd` from the session header. The host returns the intersection of model-invocable and user-invocable skills because this browser path inserts a model reference rather than loading the body directly. Catalog-addressed continuable children resolve no skill candidates locally because the existing skill RPC requires an attached session; viewing their persisted history must not activate them. Catalogs cache per ordinary session with a single-flight fetch; the scope-birth `warm` hook prewarms the session's entry and `connection/reset` clears everything. Results filter by `startsWith(query)`; picking a candidate lands the literal `/name ` text through the slash pipeline (decision 21 plain-text reference), and the source `codec` owns the reference's two projections: `clipboardText` → `/name`, `serialize` → the model form `<skill>name</skill>` invoked at submit time. The RPC rides the plugin's root-context connection captured at registration — the source never reads services off a per-call argument. The source implements no `matchSpace`/`matchEnter` hooks — skill references never enter command adjudication and ride ordinary prompts into the default sink.
+Skill invocation source, browser half: registers the `/`-trigger `skill` source into `ctx.slash`. Ordinary-session candidates come from the `skill.list` RPC addressed by the per-call `ClientSessionContext` projection's `{sessionId}`, with the host resolving `cwd` from the session header. The host serves every user-invocable skill; a `modelInvocable: false` entry (a `disable-model-invocation` skill, whose only entry point is this path) wears the user-only marker as a description prefix in the active language. Catalog-addressed continuable children resolve no skill candidates locally because the existing skill RPC requires an attached session; viewing their persisted history must not activate them. Catalogs cache per ordinary session with a single-flight fetch; the scope-birth `warm` hook prewarms the session's entry and `connection/reset` clears everything. Results filter by `startsWith(query)`.
+
+A menu pick or an entered `/name [args]` line claims the composer into an args-tolerant `skill.invoke` transaction (`matchEnter` strong-waits the catalog; an unknown name answers undefined and stays a plain prompt). Submit trims the args, keeps blank args off the wire, and folds an RPC refusal into the composer's error outcome; the host renders the skill body and injects it as a user message before starting the turn, so invocation is deterministic for every user-invocable skill. The RPC rides the plugin's root-context connection captured at registration — the source never reads services off a per-call argument. Draft chip visuals still derive from the `lexicon` scan; the legacy `<skill>name</skill>` reference codec is gone (decision 21 removal cut) and `matchSpace` stays unimplemented — menu and enter own the skill flows.
 
 A failed `skill.list` throws from `candidates`, which the slash shell logs and folds into a silent menu-group drop — the menu shows only pending/ready states.
 
@@ -14,23 +16,22 @@ The browser plugin also registers a keyed `skill` toolview in `conversation.chat
 
 ## Model Experience
 
-### Skill reference text in the user prompt
+### User-explicit skill invocation
 
 #### What the model sees
 
-A picked candidate lands the literal `/name ` in the draft (decision 21: plain text, no `<skill>` tag); the text reaches the model verbatim inside the ordinary user message (`session.prompt`), with no dedicated content block, prompt section, or host-side expansion. The association with the actual skill is model-side and non-deterministic: the session prefix already carries the skill catalog (rendered by `dsh-tool-skill`), and the reference's name matching a catalog entry is what invites the model to load it.
+A claimed invocation never ships the `/name` literal. The host (`skill.invoke`) renders the canonical `<skill_content>` block — the same `renderSkillContent` output the `skill` tool returns — appends the user's trailing text after a blank line, and injects the whole as one user-role message carrying the `skill-invocation` source, immediately starting a turn. Loading is deterministic: the model receives the full body without being asked to call the `skill` tool, and the catalog (rendered by `dsh-tool-skill`) tells it not to re-load an inline-injected skill.
 
 #### Token effect
 
-Conditional and tiny: only a pick (or hand-typing the same text) adds the reference's characters to that one user message. Menu browsing and the candidate fetch add zero model tokens.
+One invocation adds the rendered skill body plus the trailing text to that turn's user message — the same cost as the model loading the skill through the tool, paid unconditionally instead of at the model's discretion. Menu browsing and the candidate fetch add zero model tokens.
 
 #### KV Cache effect
 
-Append-only: the reference is part of a new user message appended after the reusable history prefix. This package never edits earlier request tokens.
+Append-only: the injected message lands after the reusable history prefix. This package never edits earlier request tokens.
 
 ## Known Limitations and Deferred Work
 
 - **Result-only history pages use the generic row** — keyed dispatch needs the paired call in the runtime window; pagination that leaves the call outside has no tool identity. This client presentation feature does not extend the history wire contract to recover it.
-- **Non-deterministic skill loading** — the reference is a collaboration cue, not a guarantee; the model may ignore it. The rework path when hit rate proves insufficient (a host-side `context/skill-reference` guidance package, or full-text injection) sits in the design ledger; the wire text shape would not change.
-- **First keystroke may race the prewarm** — the scope-birth warm launches the catalog fetch, but a menu opened before it settles shows no skill candidates for that keystroke. Accepted by design: skill references do not participate in enter adjudication, so nothing correctness-bearing waits on the catalog.
+- **Enter waits on the catalog once** — `matchEnter` strong-waits the session's first catalog fetch before answering, so an enter racing a cold cache resolves against the settled catalog rather than silently missing. A menu opened before the prewarm settles still shows no skill candidates for that keystroke.
 - **Text is the truth** — the reference is plain draft text; a hand-typed identical token is the same reference. Chip visuals derive from the lexicon scan; no occurrence identity or position tracking (componentized chips are a ledger item).

packages/client/ui-skill/README.zh.md

@@ -2,7 +2,9 @@
 
 [English](README.md) | 中文
 
-skill（技能）引用 source 的浏览器端：把 `/` 触发的 `skill` source 注册进 `ctx.slash`。普通会话的候选来自 `skill.list` RPC，以每次调用的 `ClientSessionContext` 投影中的 `{sessionId}` 寻址，host 从会话 header 解析 `cwd`。宿主返回模型可调用与用户可调用 skill 的交集，因为该浏览器路径插入的是模型引用，而不是直接加载正文。由目录寻址的可继续 subagent 在客户端解析为没有 skill 候选，因为现有 skill RPC 要求会话已挂载；查看其持久化历史不得激活它。目录按普通会话缓存，拉取走 single-flight；scope 创建时的 `warm` 钩子预热该会话的缓存项，`connection/reset` 清空全部缓存。结果按 `startsWith(query)` 过滤；pick 一个候选会把字面文本 `/name ` 经 slash 流水线落进草稿（决策 21 的纯文本引用），source 的 `codec` 拥有该引用的两种投影：`clipboardText` → `/name`，`serialize` → 提交时生成的模型形式 `<skill>name</skill>`。RPC 使用插件注册时捕获的根上下文连接——source 绝不从每次调用的参数上读取服务。source 不实现 `matchSpace`／`matchEnter` 钩子——skill 引用永不进入命令裁决，随普通提示词落入 default sink。
+skill（技能）调用 source 的浏览器端：把 `/` 触发的 `skill` source 注册进 `ctx.slash`。普通会话的候选来自 `skill.list` RPC，以每次调用的 `ClientSessionContext` 投影中的 `{sessionId}` 寻址，host 从会话 header 解析 `cwd`。宿主提供每一个用户可调用的 skill；`modelInvocable: false` 的条目（即 `disable-model-invocation` skill，此路径是其唯一入口）会以当前语言把仅限用户标记作为描述前缀带上。由目录寻址的可继续 subagent 在客户端解析为没有 skill 候选，因为现有 skill RPC 要求会话已挂载；查看其持久化历史不得激活它。目录按普通会话缓存，拉取走 single-flight；scope 创建时的 `warm` 钩子预热该会话的缓存项，`connection/reset` 清空全部缓存。结果按 `startsWith(query)` 过滤。
+
+菜单 pick 或回车提交的一行 `/name [args]` 会把 composer 认领进一个容忍参数的 `skill.invoke` 事务（`matchEnter` 强等目录；未知名称应答 undefined，保持为普通提示词）。提交时会修剪参数、让空白参数不上协议，并把 RPC 拒绝折叠进 composer 的错误结局；宿主在开启轮次之前渲染 skill 正文并将其作为用户消息注入，因此对每一个用户可调用的 skill，调用都是确定性的。RPC 使用插件注册时捕获的根上下文连接——source 绝不从每次调用的参数上读取服务。草稿 chip 视觉仍由 `lexicon` 扫描派生；旧的 `<skill>name</skill>` 引用 codec 已经移除（决策 21 的移除裁定），`matchSpace` 保持不实现——skill 流程归菜单与回车所有。
 
 `skill.list` 失败时 `candidates` 抛出异常，slash 壳层记录日志并折叠为静默的菜单组丢弃——菜单只显示 pending／ready 状态。
 
@@ -14,23 +16,22 @@ skill（技能）引用 source 的浏览器端：把 `/` 触发的 `skill` sourc
 
 ## 模型体验
 
-### 用户提示词中的 skill 引用文本
+### 用户显式 skill 调用
 
 #### 模型看到的内容
 
-被 pick 的候选会把字面文本 `/name ` 落进草稿（决策 21：纯文本，无 `<skill>` 标签）；该文本原样进入普通用户消息（`session.prompt`）到达模型，没有专用内容块、提示词 section 或 host 侧展开。与实际 skill 的关联在模型侧建立且具有非确定性：会话前缀已携带 skill 目录（由 `dsh-tool-skill` 渲染），引用名称与目录条目匹配，正是这一点引导模型去加载它。
+被认领的调用绝不会把字面文本 `/name` 发出去。宿主（`skill.invoke`）渲染规范的 `<skill_content>` 块——与 `skill` 工具返回的 `renderSkillContent` 输出相同——在一个空行之后追加用户的尾随文本，并把整体作为一条携带 `skill-invocation` 来源的 user 角色消息注入，随即开启一个轮次。加载是确定性的：模型无需被要求调用 `skill` 工具就能收到完整正文，目录（由 `dsh-tool-skill` 渲染）也会告诉它不要重新加载已内联注入的 skill。
 
 #### Token 影响
 
-有条件且极小：只有 pick（或手动键入相同文本）会把引用的字符加进那一条用户消息。浏览菜单和拉取候选不会增加任何模型 token。
+一次调用会把渲染后的 skill 正文连同尾随文本加进该轮次的用户消息——成本与模型经由工具加载该 skill 相同，只是无条件支付，而非由模型自行裁量。浏览菜单和拉取候选不会增加任何模型 token。
 
 #### KV Cache 影响
 
-仅追加：引用是追加在可复用历史前缀之后的新用户消息的一部分。该包绝不改写较早的请求 token。
+仅追加：注入的消息落在可复用历史前缀之后。该包绝不改写较早的请求 token。
 
 ## 已知限制与暂缓事项
 
 - **仅含结果的 history 页使用通用行**：键控分派要求配对调用位于 runtime 窗口内；分页将调用留在窗口外时，结果没有工具身份。这项客户端呈现功能不会为了恢复该身份而扩展 history 协议契约。
-- **skill 加载具有非确定性**：引用是协作线索，不是保证；模型可能忽略它。针对命中率不足情况的返工路径（host 侧 `context/skill-reference` 引导包，或全文注入）记录在设计台账中；协议中的文本形态不会改变。
-- **首次击键可能与预热竞速**：scope 创建时的预热会启动目录拉取，但目录落定之前打开的菜单，在那次击键下不会显示 skill 候选。这是设计上接受的取舍：skill 引用不参与回车裁决，因此没有任何攸关正确性的环节等待目录。
+- **回车对目录只等待一次**：`matchEnter` 在应答之前强等该会话的首次目录拉取，因此与冷缓存竞速的回车会对照已落定的目录解析，而不是静默错过。预热落定之前打开的菜单，在那次击键下仍不会显示 skill 候选。
 - **文本是唯一依据**：引用是普通的草稿文本；手动键入的相同 token 就是同一个引用。chip 视觉由 lexicon 扫描派生；没有 occurrence 身份或位置跟踪（组件化 chip 是台账事项）。

packages/host/apiproxy/README.i18n.yaml

@@ -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 packages/host/apiproxy/README.md
-README.md: 7ac7bdc6db2e2abbc60d1a8813e229c21ed39fe7
-README.zh.md: d6ece5caed752cf0cc59cc97017549ec2b1e66cb
+README.md: 8d7a24b0b8b897d94ed29d5dc9ed6e9efb250fc6
+README.zh.md: c988b7540ba719d02e50d6da9595353c93766835

packages/host/apiproxy/README.md

@@ -46,7 +46,7 @@ Directory picking delegates to the composed `ctx.directoryPicker` backend ([the
 
 `host.openPath` opens a filesystem path with the operating system's default application (`open` on macOS, `Invoke-Item` on Windows, and `xdg-open` on desktop Linux). For `.html`, `.htm`, `.xhtml`, and `.svg`, macOS and desktop Linux prefer a named default browser and fall back to that application handoff when none can be named. WSL translates every Linux path through `wslpath -w` and hands the resulting Windows/UNC path to Windows `Invoke-Item`, including browser-renderable documents, instead of assuming a Linux desktop association. The browser carrier applies the same loopback, same-origin restriction as `host.pickDirectory`.
 
-The `command.*` and `skill.*` domains expose the host command registry and skill catalog to clients. Every method addresses one session's agent by `sessionId` (a served session always has an Agent; `command.*` resumes cold sessions through the same path as `session.*`, while `skill.list` resolves the project root from the session header without touching the Agent registry). `skill.list` serves the browser's user-selected model-reference path, so it returns only skills that are both model-invocable and user-invocable; this domain has no direct skill-loading RPC. `command.execute` runs a slash-command line host-side with pure admission semantics: the response reports whether the line resolved to a handler plus the minted lifecycle `commandId` when it did (correlating the acknowledgment with the flow node), while the outcome rides the durably logged `command/run`/`command/done` lifecycle pair broadcast on the mux stream. Command handlers may legitimately outlast the 30-second transport health deadline, so `command.execute` carries only caller/connection cancellation; that signal cancels the running handler. `host/commands-changed` is the catalog invalidation frame: clients refetch `command.list` instead of diffing.
+The `command.*` and `skill.*` domains expose the host command registry and skill catalog to clients. Every method addresses one session's agent by `sessionId` (a served session always has an Agent; `command.*` resumes cold sessions through the same path as `session.*`, while `skill.list` resolves the project root from the session header without touching the Agent registry). `skill.list` serves the composer's invocation path: it returns every user-invocable skill with its `modelInvocable` flag, so menus can mark user-only (`disable-model-invocation`) entries whose only entry point this is. `skill.invoke` is the user-explicit loading RPC: it enforces user-invocation policy at this boundary (`skill-not-found` / `skill-not-invocable`), renders the canonical `<skill_content>` body via the shared `renderSkillContent`, appends the optional trailing `text`, injects the whole as a user-role message carrying the `skill-invocation` source, and starts a turn through the same route-served refusal gate as `session.prompt`. `command.execute` runs a slash-command line host-side with pure admission semantics: the response reports whether the line resolved to a handler plus the minted lifecycle `commandId` when it did (correlating the acknowledgment with the flow node), while the outcome rides the durably logged `command/run`/`command/done` lifecycle pair broadcast on the mux stream. Command handlers may legitimately outlast the 30-second transport health deadline, so `command.execute` carries only caller/connection cancellation; that signal cancels the running handler. `host/commands-changed` is the catalog invalidation frame: clients refetch `command.list` instead of diffing.
 
 The `settings.*`, `credentials.*`, and `llm.*` domains are the configuration-page wire. The settings domain serves the namespaces addressed by registered configurable providers (`ctx.llm.listConfigurableProviders()`) plus a small explicit allowlist — the Web preference `permission` and the product-owned `ui-onboarding`; adding a Settings registration alone never makes it remotely readable or writable. Any other namespace answers `settings-not-exposed` — the same answer an unregistered namespace gets, so no caller can enumerate the registry by probing. `settings.describe` returns each exposed namespace's serialized schemastery schema, redacted layered values (resolved/`base`/`user` — a field's presence in `user` marks it user-overridden), the `secrets` slot list, the section's `revision`, and the boolean `hasDocument` capability flag. The browser receives no Host path: pathless `settings.openDocument` asks the provider to materialize its document and then hands the Host-resolved result to the native opener, so no browser payload can select an arbitrary filesystem target. `settings.update`/`settings.replace` write the user layer; `settings.mutate` applies path ops (`set`/`unset`) against the section as stored, which is the removal path for a client holding the redacted view — rebuilding a section from it and replacing wholesale would delete the secrets the wire never returned. Any write may carry `expectedRevision`; a stale one answers `settings-conflict` with both revisions rather than overwriting the writer that landed first, and every other seam refusal folds into `settings-rejected`. Secret-role values never ride any response in any layer; a secret crosses the wire in exactly one direction — inside an `update`/`mutate` payload or `credentials.set`. `credentials.describe` returns value-free views (`configured`/`source`/`writable`), and `credentials.set`/`credentials.unset` map a shadowed-reference refusal onto `credential-rejected`. `llm.providers` merges the configurable-provider directory with live routes (dormant entries carry `active: false`; undeclared live routes append with no settings address) and `llm.models` is the session-independent catalog. `llm.discoverModels` interrogates a provider endpoint the page is still drafting: `settingsNs` selects the adapter family that knows how to read the listing, and the endpoint, protocol, and key come from the form rather than from storage. It writes nothing — the reply is candidates, and only a later `settings.mutate` decides what a route serves — so its `apiKey` is the third payload on which a secret may ride, alongside `settings.update`/`mutate` and `credentials.set`. The host never stores or returns it; like the other two it does ride the client's outgoing envelope, which `subscribeEnvelopes()` observers can see, and redacting that tap is a configuration-plane-wide change rather than this method's to make alone. Every refusal (an unserved namespace, a protocol with no readable listing, an unreachable endpoint, a rejected credential) folds into `model-discovery-failed`, whose message is the adapter's own text and whose details name the endpoint asked but never the credential offered. Three invalidation frames keep every surface converged without polling: `host/settings-changed {ns}` (`settings/document-updated` passthrough, so a raw change whose resolved value is unchanged still reaches clients), `host/credentials-changed {ref}` (reference names only, never values), and `host/models-changed` — fired by `llm/adapters-updated` and by a change to a configurable-provider namespace, whose settings carry that provider's catalog and endpoint; a `permission` or `ui-onboarding` change emits only its settings invalidation. The browser carrier restricts the whole configuration plane, reads and native actions included (`settings.describe`/`openDocument`/`update`/`replace`/`mutate`, `credentials.describe`/`set`/`unset`), to loopback same-origin requests — the `host.pickDirectory` privileged set. A composition without a settings or credential provider answers those domains with an actionable `internal` error naming the missing plugin.
 

packages/host/apiproxy/README.zh.md

@@ -46,7 +46,7 @@ Workspace 列表与 Session 列表是相互独立的重连基线。`workspace.cr
 
 `host.openPath` 会用操作系统的默认应用打开一个文件系统路径（macOS 为 `open`，Windows 为 `Invoke-Item`，桌面 Linux 为 `xdg-open`）。对于 `.html`、`.htm`、`.xhtml` 与 `.svg`，macOS 和桌面 Linux 会优先使用能够确定的默认浏览器；无法确定时回退到上述应用交接。WSL 会通过 `wslpath -w` 转换每个 Linux 路径，并将所得 Windows/UNC 路径交给 Windows `Invoke-Item`，浏览器可渲染的文档也不例外，而非假定存在 Linux 桌面文件关联。浏览器载体对其施加与 `host.pickDirectory` 相同的回环、同源限制。
 
-`command.*` 与 `skill.*` 领域向客户端暴露宿主命令注册表和技能目录。每个方法都通过 `sessionId` 寻址一个会话的 Agent（被服务的会话必有 Agent；`command.*` 经由与 `session.*` 相同的路径恢复冷会话，而 `skill.list` 从会话头解析项目根目录，不触碰 Agent 注册表）。`skill.list` 服务于浏览器中由用户选择的模型引用路径，因此仅返回模型和用户均可调用的 skill；该领域没有直接加载 skill 的 RPC。`command.execute` 在宿主侧运行一条斜杠命令行，语义为纯准入：响应报告该行是否解析到处理器，并在解析到时回带铸造的生命周期 `commandId`（将本次确认与流节点关联）；结局经由持久落账并在 mux 流广播的 `command/run`/`command/done` 生命周期事件对承载。命令处理器运行超过 30 秒的传输健康时限仍属正常，因此 `command.execute` 仅携带调用方／连接取消信号；该信号可取消正在运行的处理器。`host/commands-changed` 是目录失效帧：客户端重新拉取 `command.list` 而不是做差分。
+`command.*` 与 `skill.*` 领域向客户端暴露宿主命令注册表和技能目录。每个方法都通过 `sessionId` 寻址一个会话的 Agent（被服务的会话必有 Agent；`command.*` 经由与 `session.*` 相同的路径恢复冷会话，而 `skill.list` 从会话头解析项目根目录，不触碰 Agent 注册表）。`skill.list` 服务于 composer 的调用路径：它返回每一个用户可调用的 skill 及其 `modelInvocable` 标志，让菜单能够标出仅限用户（`disable-model-invocation`）的条目——此处是这类条目唯一的入口。`skill.invoke` 是用户显式加载 RPC：它在此边界强制执行用户调用策略（`skill-not-found`/`skill-not-invocable`），经共享的 `renderSkillContent` 渲染规范的 `<skill_content>` 正文，追加可选的尾随 `text`，把整体作为一条携带 `skill-invocation` 来源的 user 角色消息注入，并经由与 `session.prompt` 相同的「路由是否有适配器在服务」拒绝闸门开启一个轮次。`command.execute` 在宿主侧运行一条斜杠命令行，语义为纯准入：响应报告该行是否解析到处理器，并在解析到时回带铸造的生命周期 `commandId`（将本次确认与流节点关联）；结局经由持久落账并在 mux 流广播的 `command/run`/`command/done` 生命周期事件对承载。命令处理器运行超过 30 秒的传输健康时限仍属正常，因此 `command.execute` 仅携带调用方／连接取消信号；该信号可取消正在运行的处理器。`host/commands-changed` 是目录失效帧：客户端重新拉取 `command.list` 而不是做差分。
 
 `settings.*`、`credentials.*` 与 `llm.*` 领域是配置页协议。settings 领域服务于已注册可配置提供方所指向的 namespace（`ctx.llm.listConfigurableProviders()`），并额外服务于一份小型、显式的 allowlist——Web 偏好 `permission` 与产品持有的 `ui-onboarding`；仅新增一项 Settings 注册，绝不会使其可被远程读取或写入。其他任何 namespace 都只会得到 `settings-not-exposed`——未注册的 namespace 得到的是同一个答复，因此没有调用方能靠逐个探测把注册表枚举出来。`settings.describe` 为每个已暴露 namespace 提供其序列化 schemastery schema、脱敏后的分层值（resolved/`base`/`user`——字段出现在 `user` 中即标记其被用户覆盖）、`secrets` 槽位列表、该分节的 `revision`，以及布尔型 `hasDocument` 能力标志。浏览器不会收到 Host 路径：无路径参数的 `settings.openDocument` 会请求提供方准备文档，再把由 Host 解析出的结果交给原生打开器，因此任何浏览器载荷都无法选择任意文件系统目标。`settings.update`/`settings.replace` 写入用户层；`settings.mutate` 则在已存分节上施加路径 op（`set`/`unset`），这是持有脱敏视图的客户端的删除路径——据此重建分节再整体替换，会删掉协议从未回传过的那些机密。任何写入都可携带 `expectedRevision`；陈旧的期望值会以 `settings-conflict` 连同两个 revision 作答，而不是覆盖先落地的那个写方，其余每种 seam 拒绝则折叠为 `settings-rejected`。secret 角色的值绝不在任何一层搭乘任何响应；secret 只沿一个方向跨越协议——在 `update`/`mutate` 载荷或 `credentials.set` 之内。`credentials.describe` 返回不含值的视图（`configured`/`source`/`writable`），`credentials.set`/`credentials.unset` 则把被遮蔽引用的拒绝映射为 `credential-rejected`。`llm.providers` 把可配置提供方目录与存活路由合并（休眠条目携带 `active: false`；未声明的存活路由追加在后，不带 settings 地址），`llm.models` 则是与会话无关的目录。`llm.discoverModels` 询问页面尚在起草的提供方端点：`settingsNs` 选出懂得读取该列表的适配器家族，端点、协议与密钥则来自表单而非存储。它什么都不写——回复是候选，只有随后的 `settings.mutate` 才决定路由服务什么——因此其 `apiKey` 是 secret 可以搭乘的第三个、也是最后一个载荷（另两个是 `settings.update`/`mutate` 与 `credentials.set`），且绝不被存储或回显。host 从不存储或回传它；与另两者一样，它确实会搭乘客户端的出站信封，`subscribeEnvelopes()` 的观察者能看到——为该 tap 做脱敏是整个配置面的改动，而非本方法一家的事。每一种拒绝（无人服务的 namespace、没有可读列表的协议、不可达端点、被拒凭据）都折叠为 `model-discovery-failed`，其消息是适配器自己的文本，details 点名被询问的端点，绝不点名所提供的凭据。三个失效帧让每个面无需轮询即保持收敛：`host/settings-changed {ns}`（`settings/document-updated` 透传，因此解析值未变的原始变更同样能到达客户端）、`host/credentials-changed {ref}`（只带引用名，绝不带值），以及 `host/models-changed`——它由 `llm/adapters-updated` 和可配置提供方 namespace 的变更触发，因为该提供方的设置正承载着它的目录与端点；`permission` 或 `ui-onboarding` 变更只会发出自身的 settings 失效通知。浏览器载体把整个配置面（含读取与原生操作：`settings.describe`/`openDocument`/`update`/`replace`/`mutate` 与 `credentials.describe`/`set`/`unset`）限制为仅接受来自回环地址的同源请求——即 `host.pickDirectory` 所在的特权集合。未装 settings 或凭据 provider 的组合会以指名缺失插件、包含解决建议的 `internal` 错误应答这些领域。
 

packages/skill/skill/README.i18n.yaml

@@ -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 packages/skill/skill/README.md
-README.md: f538ae668ccff291be86348627d5547150f460df
-README.zh.md: d61a242d01df1e22270c1cb049b922536654bbd6
+README.md: 0c1b2249d8c46ad9ce8097ceeda2bd988c92eb21
+README.zh.md: 8fed350d00433206aecdb32819adc81c82745869

packages/skill/skill/README.md

@@ -37,6 +37,10 @@ This package owns the `ctx.skills` interface. It does not know whether skills co
 | `{ modelInvocable: false, userInvocable: true }` | excluded | included |
 | `{ modelInvocable: false, userInvocable: false }` | excluded | excluded |
 
+### Shared model-facing rendering
+
+`renderSkillContent(skill)` renders one loaded skill as the canonical `<skill_content>` block (escaped `name` attribute, resource hints, verbatim body). It is the single truth for both loading paths: `dsh-tool-skill` returns it as the `skill` tool result, and the host's user-explicit `skill.invoke` injects it as a user message, so the model sees one shape regardless of who initiated the load. `escapeText` is exported beside it for consumers embedding prose in the same markup frame. The package also declares the `skill-invocation` `MessageSource` kind ({ name, args? }) that user-explicit injection stamps on its messages — transcript consumers present the invocation from this metadata instead of re-parsing the body.
+
 `isModelInvocable(skill)` and `isUserInvocable(skill)` read the matching positive field directly. `ctx.skills.get()` remains the trusted, policy-neutral loading primitive, so every user- or model-facing consumer must enforce the predicate that matches its surface before exposing or loading a skill.
 
 ## Provider Contract

packages/skill/skill/README.zh.md

@@ -37,6 +37,10 @@
 | `{ modelInvocable: false, userInvocable: true }` | 排除 | 包含 |
 | `{ modelInvocable: false, userInvocable: false }` | 排除 | 排除 |
 
+### 共享的面向模型渲染
+
+`renderSkillContent(skill)` 把一个已加载 skill 渲染为规范的 `<skill_content>` 块（转义后的 `name` 属性、资源提示、原样正文）。它是两条加载路径的唯一真源：`dsh-tool-skill` 将其作为 `skill` 工具结果返回，宿主的用户显式 `skill.invoke` 将其作为用户消息注入，因此无论加载由谁发起，模型看到的都是同一种形态。`escapeText` 随之一并导出，供要在同一标记框架中嵌入文案的消费方使用。该包还声明 `skill-invocation` 这个 `MessageSource` kind（{ name, args? }），用户显式注入会把它打在自己的消息上——transcript（文本记录）消费方依据这份元数据呈现该次调用，而不是重新解析正文。
+
 `isModelInvocable(skill)` 和 `isUserInvocable(skill)` 分别直接读取对应的正向字段。`ctx.skills.get()` 仍是受信且与策略无关的加载原语，因此每个面向用户或模型的消费方都必须先执行与自身接口匹配的判定，再暴露或加载 skill。
 
 ## 提供方契约

packages/skill/tool-skill/README.i18n.yaml

@@ -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 packages/skill/tool-skill/README.md
-README.md: 8e0bff5d1c4853092d412b8f7f9528d4b00d9626
-README.zh.md: c6b815bef59eb1f14be0892078694f129366d004
+README.md: 5c6e592c670f324eb660dbe1fec168fd77e5b368
+README.zh.md: 202a621b1d4047c7d763de3b98c1a69c8c1ee1f7

packages/skill/tool-skill/README.md

@@ -36,7 +36,7 @@ Tool execution does not add a synthetic context message. Its freshly loaded resu
 
 #### What the model sees
 
-If model-invocable skills exist and this exact `skill` tool is visible, the agent receives the catalog template below as a durable user-role message before the first request, with one data-dependent entry per sorted skill. Later membership, description, or visibility changes append a complete replacement using the same `<available_skills>` envelope; deleting every skill appends an empty envelope with an explicit instruction not to use older names.
+If model-invocable skills exist and this exact `skill` tool is visible, the agent receives the catalog template below as a durable user-role message before the first request, with one data-dependent entry per sorted skill. Later membership, description, or visibility changes append a complete replacement using the same `<available_skills>` envelope; deleting every skill appends an empty envelope with an explicit instruction not to use older names. The template's closing sentence is the seam rule against double-loading: the host's user-explicit `skill.invoke` injects the same `renderSkillContent` output (shared from `@deepseek-ai/dsh-skill`) inline, and the catalog tells the model to follow that block instead of re-loading the skill through the tool; the replacement-catalog template carries the same sentence.
 
 ##### Skill catalog template
 
@@ -49,6 +49,7 @@ A skill is a reusable set of task-specific instructions. The following skills ar
 </available_skills>
 
 If the user names a skill, or the task clearly matches a skill's description, call the `skill` tool with the exact skill name before taking task actions. Load all applicable skills, then follow their full instructions. This catalog contains summaries only; do not infer or follow a skill's instructions until it has been loaded.
+A user may also invoke a skill directly; its <skill_content> block then appears in this conversation. Follow it, and do not call the `skill` tool again for that skill.
 </system-reminder>
 ```
 

packages/skill/tool-skill/README.zh.md

@@ -36,7 +36,7 @@
 
 #### 模型看到的内容
 
-如果存在模型可调用 skill，且可见的正是这个 `skill` 工具，agent 会在第一个请求之前收到下方目录模板，其中包含每个已排序 skill 的一条随数据而定的条目。该目录是一条持久的用户角色消息。后续成员关系、描述或可见性的变化会使用同一个 `<available_skills>` 信封追加完整替换；删除所有 skill 时，会追加一个空信封，并明确指示不得使用旧名称。
+如果存在模型可调用 skill，且可见的正是这个 `skill` 工具，agent 会在第一个请求之前收到下方目录模板，其中包含每个已排序 skill 的一条随数据而定的条目。该目录是一条持久的用户角色消息。后续成员关系、描述或可见性的变化会使用同一个 `<available_skills>` 信封追加完整替换；删除所有 skill 时，会追加一个空信封，并明确指示不得使用旧名称。模板的结尾一句是防止双重加载的 seam 规则：宿主的用户显式 `skill.invoke` 会把同一份 `renderSkillContent` 输出（共享自 `@deepseek-ai/dsh-skill`）内联注入，目录则告诉模型遵循该块，而不是再经工具重新加载该 skill；替换目录模板携带同一句话。
 
 ##### Skill 目录模板
 
@@ -49,6 +49,7 @@ A skill is a reusable set of task-specific instructions. The following skills ar
 </available_skills>
 
 If the user names a skill, or the task clearly matches a skill's description, call the `skill` tool with the exact skill name before taking task actions. Load all applicable skills, then follow their full instructions. This catalog contains summaries only; do not infer or follow a skill's instructions until it has been loaded.
+A user may also invoke a skill directly; its <skill_content> block then appears in this conversation. Follow it, and do not call the `skill` tool again for that skill.
 </system-reminder>
 ```