Merge pull request #2002 from deepseek-harness/feature/issue-1470-skill-invoke
feat(skill): deterministic user-explicit skill invocation at the pre-step gesture boundary
This commit is contained in:
Yichen Jiang
@@ -2449,7 +2449,8 @@ function createFixtureWorld(options: FixtureOptions): FixtureWorld {
|
||||
if (missing !== undefined) return missing
|
||||
return ok(request, {
|
||||
skills: [
|
||||
{ name: 'fixture-demo', description: 'fixture 技能样本', whenToUse: '仅供 UI 目录渲染验收' },
|
||||
{ name: 'fixture-demo', description: 'fixture 技能样本', whenToUse: '仅供 UI 目录渲染验收', modelInvocable: true },
|
||||
{ name: 'fixture-user-only', description: 'fixture 仅用户技能样本', modelInvocable: false },
|
||||
],
|
||||
})
|
||||
},
|
||||
|
||||
@@ -163,6 +163,7 @@ export class FakeApiClient implements IApiClient {
|
||||
onSkillList: (payload: unknown) => Promise<RpcResponse<{ skills: SkillEntry[] }>>
|
||||
= () => Promise.resolve(ok({ skills: [] }))
|
||||
|
||||
|
||||
readonly commands: IApiClient['commands'] = {
|
||||
list: (payload: unknown) => this.record('command.list', payload, this.onCommandList(payload)),
|
||||
execute: (payload: unknown) => this.record('command.execute', payload, this.onCommandExecute(payload)),
|
||||
|
||||
@@ -83,6 +83,9 @@ export function contextProvenance(source: unknown): ContextProvenanceView {
|
||||
return { role: 'inject', label: joined(collect(record, 'changes', 'path')) ?? kind }
|
||||
case 'plugin':
|
||||
return { role: 'inject', label: readString(record, 'plugin') ?? kind }
|
||||
// A user-explicit skill invocation names the skill it injected.
|
||||
case 'skill-invocation':
|
||||
return { role: 'inject', label: readString(record, 'name') ?? kind }
|
||||
// Documented default arm of the merge-extensible source map: an unknown
|
||||
// producer still identifies itself by its own durable kind.
|
||||
default:
|
||||
|
||||
@@ -57,10 +57,11 @@ function materializeNode(
|
||||
stepTimings: ReadonlyMap<string, AssistantStepMetadata>,
|
||||
): ConversationNode {
|
||||
switch (event.type) {
|
||||
case 'user/message':
|
||||
// Injected context (plugin/goal source) folds to a context node, not a
|
||||
// user message; only a direct human prompt is a user node. A compaction
|
||||
// checkpoint never reaches here (isCompactCheckpoint routes it away).
|
||||
case 'user/message': {
|
||||
// Injected context (plugin/goal/skill-invocation source) folds to a
|
||||
// context node, not a user message; only a direct human prompt is a
|
||||
// user node. A compaction checkpoint never reaches here
|
||||
// (isCompactCheckpoint routes it away).
|
||||
if (event.data.source.kind !== 'user') {
|
||||
return {
|
||||
kind: 'context', seq: event.seq, time: event.time,
|
||||
@@ -80,6 +81,7 @@ function materializeNode(
|
||||
kind: 'user', seq: event.seq, time: event.time,
|
||||
content: event.data.content, source: event.data.source,
|
||||
}
|
||||
}
|
||||
case 'assistant/message':
|
||||
return {
|
||||
kind: 'assistant', seq: event.seq, time: event.time,
|
||||
|
||||
@@ -198,6 +198,7 @@ export class FakeApiClient implements IApiClient {
|
||||
onSkillList: (payload: unknown) => Promise<RpcResponse<{ skills: SkillEntry[] }>>
|
||||
= () => Promise.resolve(ok({ skills: [] }))
|
||||
|
||||
|
||||
readonly commands: IApiClient['commands'] = {
|
||||
list: (payload: unknown) => this.record('command.list', payload, this.onCommandList(payload)),
|
||||
execute: (payload: unknown) => this.record('command.execute', payload, this.onCommandExecute(payload)),
|
||||
|
||||
@@ -164,6 +164,28 @@ describe('TranscriptAdapter', () => {
|
||||
expect(adapter.nodes().map(node => node.kind)).toEqual(['user', 'user', 'context'])
|
||||
})
|
||||
|
||||
it('materializes a skill-invocation injection as a named instructions context', () => {
|
||||
const adapter = new TranscriptAdapter()
|
||||
adapter.reset([
|
||||
at(0, { type: 'user/message', surfaceOp: 'append', data: createUserMessage({
|
||||
content: [{ type: 'text', text: '/hidden-demo check the fixture' }],
|
||||
source: { kind: 'user' },
|
||||
}) }),
|
||||
at(1, { type: 'user/message', surfaceOp: 'append', data: createUserMessage({
|
||||
content: [{ type: 'text', text: '<skill_content name="hidden-demo">body</skill_content>' }],
|
||||
source: { kind: 'skill-invocation', name: 'hidden-demo', form: 'instructions' } as never,
|
||||
}) }),
|
||||
])
|
||||
const nodes = adapter.nodes()
|
||||
// The gesture stays a user bubble; the injected body folds to a context
|
||||
// row named after the skill, presented as instructions.
|
||||
expect(nodes.map(node => node.kind)).toEqual(['user', 'context'])
|
||||
expect(nodes[1]).toMatchObject({
|
||||
provenance: { role: 'inject', label: 'hidden-demo' },
|
||||
form: 'instructions',
|
||||
})
|
||||
})
|
||||
|
||||
it('skips events core does not call surface-eligible, marker or not', () => {
|
||||
// The transcript is the append-origin surface, so log-only events (a chunk,
|
||||
// a turn boundary, a compact/* provenance record) and a future type core
|
||||
|
||||
@@ -137,29 +137,27 @@ function TurnErrorItem({ node, t }: {
|
||||
/**
|
||||
* Display projection of reference forms in a user bubble (free geometry — no
|
||||
* textarea alignment constraint here); everything else stays plain text. The
|
||||
* logged model text remains the single truth; this is presentation only. Two
|
||||
* shapes decorate: legacy `<skill>name</skill>` spans (pre-decision-21
|
||||
* history) and plain-text `/name` / `@name` word-boundary tokens (decision
|
||||
* 21: the sent text IS the reference — the bubble uses the same plainest
|
||||
* token scan as the composer, minus the lexicon: sent tokens were validated
|
||||
* at compose time, so shape alone decorates).
|
||||
* logged model text remains the single truth; this is presentation only.
|
||||
* Plain-text `/name` / `@name` word-boundary tokens decorate (decision 21:
|
||||
* the sent text IS the reference — the bubble uses the same plainest token
|
||||
* scan as the composer, minus the lexicon: sent tokens were validated at
|
||||
* compose time, so shape alone decorates).
|
||||
*/
|
||||
function projectUserText(text: string): ReactNode {
|
||||
const re = /<skill>([^<]+)<\/skill>|(^|\s)([/@][\w-]+)(?=\s|$)/g
|
||||
const re = /(^|\s)([/@][\w-]+)(?=\s|$)/g
|
||||
const parts: ReactNode[] = []
|
||||
let cursor = 0
|
||||
let m: RegExpExecArray | null
|
||||
while ((m = re.exec(text)) !== null) {
|
||||
const legacy = m[1] !== undefined
|
||||
const tokenStart = legacy ? m.index : m.index + (m[2]?.length ?? 0)
|
||||
const label = legacy ? `/${m[1]}` : m[3] ?? ''
|
||||
const tokenStart = m.index + (m[1]?.length ?? 0)
|
||||
const label = m[2] ?? ''
|
||||
if (tokenStart > cursor) parts.push(<MessageText key={cursor} text={text.slice(cursor, tokenStart)} />)
|
||||
parts.push(
|
||||
<span key={tokenStart} className={css.refChip} data-ref-chip={label.startsWith('@') ? 'subagent' : 'skill'}>
|
||||
{label}
|
||||
</span>,
|
||||
)
|
||||
cursor = legacy ? m.index + m[0].length : tokenStart + label.length
|
||||
cursor = tokenStart + label.length
|
||||
}
|
||||
if (parts.length === 0) return <MessageText text={text} />
|
||||
if (cursor < text.length) parts.push(<MessageText key={cursor} text={text.slice(cursor)} />)
|
||||
|
||||
@@ -864,6 +864,7 @@ describe('MessageItem arms', () => {
|
||||
view.rerender(<MessageItem t={t} node={node} retryActive />)
|
||||
expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 1s')
|
||||
})
|
||||
|
||||
})
|
||||
|
||||
describe('formatMessageClock', () => {
|
||||
|
||||
@@ -73,6 +73,7 @@ describe('producedForClosing derivation', () => {
|
||||
expect(producedForClosing(nodes, 999)).toEqual([])
|
||||
})
|
||||
|
||||
|
||||
it('counts a generic edit and never spills across the turn boundary', () => {
|
||||
const inserted = (seq: number, callId: string, path: string): ToolResultNode => ({
|
||||
...toolResult(seq, callId, 'str_replace_editor'),
|
||||
|
||||
@@ -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: bdd772662acda1f8cf1b7d8a7c5532f9b37123dd
|
||||
README.zh.md: 959ff0ede6d545150fb22710c8af75859966caa9
|
||||
|
||||
@@ -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 pick lands the literal `/name ` text and the prompt ships the same literal (decision 21) — this source implements no adjudication hooks and no reference codec (the legacy `<skill>name</skill>` form is gone with the removal cut). Determinism lives host-side: the pre-step gesture boundary (`dsh-tool-skill`) recognizes whitespace-bounded `/name` tokens naming user-invocable skills anywhere in a user message and injects the rendered `<skill_content>` for every front end, so a menu pick, a hand-typed token, and a TUI/ACP prompt all load the skill the same way. A name shared with a host command still resolves to the command: adjudication claims the line client-side before it ever becomes a prompt — deliberate precedence, matching peer products. The list RPC rides the plugin's root-context connection captured at registration — the source never reads services off a per-call argument; draft chip visuals derive from the `lexicon` scan.
|
||||
|
||||
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.
|
||||
The user's message reaches the model verbatim, `/name` literal included. The host's pre-step boundary (`dsh-tool-skill`) then appends the canonical `<skill_content>` block — the same `renderSkillContent` output the `skill` tool returns — as injected instructions context at the end of that step's injections, closest to the model's answer. Loading is deterministic: the model receives the full body without being asked to call the `skill` tool, and the catalog 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 to that turn as injected context — 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.
|
||||
- **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).
|
||||
- **Text is the truth** — the reference is plain draft text; a hand-typed identical token is the same reference, and the host gesture boundary judges the sent text, not the menu interaction. Chip visuals derive from the lexicon scan; no occurrence identity, position tracking, or structured reference payload on the prompt wire (both are ledger items).
|
||||
- **A menu opened before the prewarm settles** shows no skill candidates for that keystroke; the next keystroke re-polls the settled cache.
|
||||
|
||||
@@ -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。
|
||||
|
||||
pick 会落下字面文本 `/name `，提示词发出的就是同一段字面文本（决策 21）——本 source 不实现任何裁决钩子，也没有引用 codec（旧的 `<skill>name</skill>` 形式已随移除裁定消失）。确定性在宿主侧：pre-step 手势边界（`dsh-tool-skill`）识别用户消息中任意位置、以空白为界、指名用户可调用 skill 的 `/name` token，并为每一种前端注入渲染后的 `<skill_content>`，因此菜单 pick、手动键入的 token 与 TUI/ACP 提示词都以同一种方式加载 skill。与宿主命令同名的名称仍解析为命令：裁决在客户端把该行认领走，它根本不会成为提示词——这是有意的优先级，与同行产品一致。列表 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 视觉由 `lexicon` 扫描派生。
|
||||
|
||||
`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。
|
||||
用户消息原样到达模型，字面文本 `/name` 也包含在内。随后宿主的 pre-step 边界（`dsh-tool-skill`）把规范的 `<skill_content>` 块——与 `skill` 工具返回的 `renderSkillContent` 输出相同——作为注入的指令上下文追加在该步骤各项注入的末尾，最贴近模型的回答。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 正文作为注入上下文加进该轮次——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。
|
||||
- **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）。
|
||||
- **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，宿主手势边界评判的是发出的文本，而不是菜单交互。chip 视觉由 lexicon 扫描派生；没有 occurrence 身份、位置跟踪，也没有提示词协议上的结构化引用载荷（两者都是台账事项）。
|
||||
- **预热落定之前打开的菜单**：在那次击键下不显示 skill 候选；下一次击键会重新轮询已落定的缓存。
|
||||
|
||||
@@ -2,13 +2,16 @@
|
||||
* Skill reference plugin, browser half: registers the '/' skill source —
|
||||
* candidates from the skill.list RPC addressed by the per-call session
|
||||
* projection's sessionId (sessions are always agent-backed; the host
|
||||
* resolves cwd from the session header), pick inserts the literal `/name `
|
||||
* text (decision 21: the draft carries plain text, chip visuals are derived
|
||||
* by scanning against the source lexicon, and the prompt ships the same
|
||||
* literal — no `<skill>` tag). The RPC rides the plugin's root-context
|
||||
* connection captured at registration — the source never reads services off
|
||||
* a per-call argument. No adjudication hooks: skill references ride
|
||||
* ordinary prompts and never enter command adjudication.
|
||||
* resolves cwd from the session header). A pick lands the literal `/name `
|
||||
* text and the prompt ships the same literal (decision 21); determinism
|
||||
* lives host-side — the pre-step boundary (`dsh-tool-skill`) recognizes a
|
||||
* leading `/name` naming a user-invocable skill and injects the rendered
|
||||
* body for every front end, including `disable-model-invocation` skills the
|
||||
* model-side catalog never lists (issue #1470). 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>` reference codec is gone (decision
|
||||
* 21 removal cut).
|
||||
*
|
||||
* Catalog fetches are cached per session (the small twin of the ui-command
|
||||
* directory): the per-keystroke candidates re-poll filters a settled
|
||||
@@ -119,6 +122,10 @@ export function apply(ctx: ClientContext): void {
|
||||
for (const key of [...fetches.keys()]) invalidate(key)
|
||||
}
|
||||
|
||||
// The bound translate resolves against the registered dictionaries with the
|
||||
// locale service's own fallback ladder; candidate-time reads stay plain text.
|
||||
const t = ctx.locale.bind(NS)
|
||||
|
||||
const source: SlashSource = {
|
||||
trigger: '/',
|
||||
name: 'skill',
|
||||
@@ -129,7 +136,12 @@ export function apply(ctx: ClientContext): void {
|
||||
if (signal.aborted) return []
|
||||
return skills
|
||||
.filter(skill => skill.name.startsWith(query))
|
||||
.map(skill => ({ name: skill.name, description: skill.description }))
|
||||
.map(skill => ({
|
||||
name: skill.name,
|
||||
// The user-only marker rides the description (the menu's only
|
||||
// secondary text); `hint` is the claim-state ghost text, not a badge.
|
||||
description: skill.modelInvocable ? skill.description : `${t('menu.userOnly')} · ${skill.description}`,
|
||||
}))
|
||||
},
|
||||
warm(session) {
|
||||
// Fire-and-forget scope-birth prewarm; the shared fetch reports
|
||||
@@ -150,16 +162,14 @@ export function apply(ctx: ClientContext): void {
|
||||
}
|
||||
},
|
||||
onPick({ candidate }) {
|
||||
// Decision 21: plain-text reference — the literal lands in the draft
|
||||
// and ships to the model verbatim (trailing space closes the token).
|
||||
// Legacy path (decision 21), retained for the removal cut, no longer reached:
|
||||
// return { insert: { source: 'skill', ref: candidate.name, label: candidate.name, clipboardText: `/${candidate.name}` } }
|
||||
// Decision 21: the pick lands plain text and the prompt ships the same
|
||||
// literal. Determinism no longer rides the client — the host's
|
||||
// pre-step boundary (dsh-tool-skill) recognizes the leading /name and
|
||||
// injects the rendered body for every front end. A name shared with a
|
||||
// host command still resolves to the command: adjudication claims the
|
||||
// line client-side before it ever becomes a prompt.
|
||||
return { text: `/${candidate.name} ` }
|
||||
},
|
||||
codec: {
|
||||
clipboardText: ref => `/${ref}`,
|
||||
serialize: ref => Promise.resolve(`<skill>${ref}</skill>`),
|
||||
},
|
||||
}
|
||||
const slash = ctx.get('slash') as SlashServiceContract
|
||||
ctx.on('connection/reset', clearAll)
|
||||
|
||||
@@ -9,6 +9,7 @@ export const zh = {
|
||||
'row.failed': 'skill 加载失败',
|
||||
'row.stopped': 'skill 加载已中止',
|
||||
'row.instructions': '说明',
|
||||
'menu.userOnly': '仅用户',
|
||||
} satisfies Record<string, string>
|
||||
|
||||
/** The skill namespace key union. */
|
||||
@@ -20,4 +21,5 @@ export const en = {
|
||||
'row.failed': 'Skill load failed',
|
||||
'row.stopped': 'Skill load stopped',
|
||||
'row.instructions': 'Instructions',
|
||||
'menu.userOnly': 'user-only',
|
||||
} satisfies Record<SkillKey, string>
|
||||
|
||||
@@ -20,11 +20,15 @@ import type { ClientSessionContext, SlashSource } from '@deepseek-ai/dsh-client-
|
||||
import { apply, inject } from '../src/client/index.ts'
|
||||
import { SkillRow as SkillToolRow } from '../src/client/SkillRow.tsx'
|
||||
|
||||
type SkillRow = { name: string; description: string; whenToUse?: string }
|
||||
type SkillRow = { name: string; description: string; whenToUse?: string; modelInvocable?: boolean }
|
||||
type ListResult =
|
||||
| { ok: true; value: { skills: SkillRow[] } }
|
||||
| { ok: false; error: { code: string; message: string; details: object } }
|
||||
type ListFn = (payload: object, signal?: AbortSignal) => Promise<{ result: ListResult }>
|
||||
type InvokeResult =
|
||||
| { ok: true; value: { accepted: true } }
|
||||
| { ok: false; error: { code: string; message: string; details: object } }
|
||||
type InvokeFn = (payload: object) => Promise<{ result: InvokeResult }>
|
||||
|
||||
interface PresentationCapture {
|
||||
slots: SlotsService
|
||||
@@ -49,16 +53,19 @@ function providePresentation(ctx: Context): PresentationCapture {
|
||||
capture.dictionaries.push({ namespace, dictionaries })
|
||||
return () => { capture.localeDisposed = true }
|
||||
},
|
||||
// Minimal bound-translate fake: zh dictionary lookup, key passthrough on miss.
|
||||
bind: () => (key: string) => key === 'menu.userOnly' ? '仅用户' : key,
|
||||
})
|
||||
return capture
|
||||
}
|
||||
|
||||
/** Boot the plugin over fake slash/connection faces; returns the captured source and its ctx. */
|
||||
async function bench(list: ListFn, addressed?: SessionId) {
|
||||
async function bench(list: ListFn, addressed?: SessionId, invoke?: InvokeFn) {
|
||||
const ctx = new Context()
|
||||
let captured: SlashSource | undefined
|
||||
ctx.provide('slash', { registerSource: (src: SlashSource) => { captured = src; return () => {} } })
|
||||
ctx.provide('connection', { api: { skills: { list } } })
|
||||
const defaultInvoke: InvokeFn = () => Promise.resolve({ result: { ok: true as const, value: { accepted: true as const } } })
|
||||
ctx.provide('connection', { api: { skills: { list, invoke: invoke ?? defaultInvoke } } })
|
||||
ctx.provide('sessions', {
|
||||
subagentAddress: (id: SessionId) => id === addressed
|
||||
? { parentSessionId: sid('parent'), childSessionId: id, mode: 'continuable' as const }
|
||||
@@ -70,9 +77,9 @@ async function bench(list: ListFn, addressed?: SessionId) {
|
||||
}
|
||||
|
||||
const CATALOG: SkillRow[] = [
|
||||
{ name: 'commit-helper', description: 'commit flow' },
|
||||
{ name: 'code-review', description: 'review flow', whenToUse: 'reviews' },
|
||||
{ name: 'deploy', description: 'deploy flow' },
|
||||
{ name: 'commit-helper', description: 'commit flow', modelInvocable: true },
|
||||
{ name: 'code-review', description: 'review flow', whenToUse: 'reviews', modelInvocable: true },
|
||||
{ name: 'deploy', description: 'deploy flow', modelInvocable: true },
|
||||
]
|
||||
|
||||
const listOk = (skills: SkillRow[]): ListFn => () => Promise.resolve({ result: { ok: true as const, value: { skills } } })
|
||||
@@ -117,12 +124,14 @@ describe('apply', () => {
|
||||
'row.failed': 'skill 加载失败',
|
||||
'row.stopped': 'skill 加载已中止',
|
||||
'row.instructions': '说明',
|
||||
'menu.userOnly': '仅用户',
|
||||
},
|
||||
en: {
|
||||
'row.running': 'Loading skill',
|
||||
'row.failed': 'Skill load failed',
|
||||
'row.stopped': 'Skill load stopped',
|
||||
'row.instructions': 'Instructions',
|
||||
'menu.userOnly': 'user-only',
|
||||
},
|
||||
},
|
||||
}])
|
||||
@@ -313,8 +322,8 @@ describe('lexicon', () => {
|
||||
})
|
||||
})
|
||||
|
||||
describe('pick and codec', () => {
|
||||
it('onPick returns the literal /name text with a closing space (decision 21)', async () => {
|
||||
describe('pick lands plain text (decision 21)', () => {
|
||||
it('onPick returns the literal /name text with a closing space', async () => {
|
||||
const { source } = await bench(listOk(CATALOG))
|
||||
const outcome = source.onPick({
|
||||
candidate: { name: 'commit-helper', description: 'commit flow' },
|
||||
@@ -326,18 +335,27 @@ describe('pick and codec', () => {
|
||||
expect(outcome).toEqual({ text: '/commit-helper ' })
|
||||
})
|
||||
|
||||
it('codec projects clipboard `/name` and serializes the model form <skill>name</skill>', async () => {
|
||||
const { source } = await bench(listOk(CATALOG))
|
||||
expect(source.codec!.clipboardText('deploy')).toBe('/deploy')
|
||||
await expect(source.codec!.serialize('deploy', new AbortController().signal))
|
||||
.resolves.toBe('<skill>deploy</skill>')
|
||||
})
|
||||
})
|
||||
|
||||
describe('adjudication', () => {
|
||||
it('never participates: no matchSpace/matchEnter hooks on the skill source', async () => {
|
||||
it('keeps the legacy reference codec removed and stays out of adjudication', async () => {
|
||||
const { source } = await bench(listOk(CATALOG))
|
||||
// Determinism lives host-side (the pre-step gesture boundary), so the
|
||||
// source neither claims lines nor serializes reference markup.
|
||||
expect(source.codec).toBeUndefined()
|
||||
expect(typeof source.matchSpace).toBe('undefined')
|
||||
expect(typeof source.matchEnter).toBe('undefined')
|
||||
})
|
||||
})
|
||||
|
||||
describe('user-only marking', () => {
|
||||
it('prefixes the description of candidates the model cannot invoke', async () => {
|
||||
const rows: SkillRow[] = [
|
||||
{ name: 'shared-skill', description: 'both surfaces', modelInvocable: true },
|
||||
{ name: 'user-only-skill', description: 'user surface only', modelInvocable: false },
|
||||
]
|
||||
const { source } = await bench(listOk(rows))
|
||||
const candidates = await source.candidates(proj('s1'), req(''))
|
||||
expect(candidates).toEqual([
|
||||
{ name: 'shared-skill', description: 'both surfaces' },
|
||||
{ name: 'user-only-skill', description: '仅用户 · user surface only' },
|
||||
])
|
||||
})
|
||||
})
|
||||
|
||||
@@ -527,6 +527,7 @@ describe('dsh-agent-spine-demo bundle', () => {
|
||||
</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>",
|
||||
"type": "user/message",
|
||||
},
|
||||
|
||||
@@ -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: 5506cbef7b778a870e1e28c3f9fdf1713f89d65f
|
||||
README.zh.md: de31f653944097e9b47a966f56c418dc9fa9b1b9
|
||||
|
||||
@@ -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 menu: it returns every user-invocable skill with its `modelInvocable` flag, so menus can mark user-only (`disable-model-invocation`) entries whose only entry point the slash gesture is. Listing is the skill domain's only RPC — invocation itself is an ordinary `session.prompt` whose whitespace-bounded `/name` tokens `dsh-tool-skill` recognizes at the pre-step boundary and answers with injected `<skill_content>` context, so every front end (web, TUI, ACP, hand-typed text) shares one deterministic path with no dedicated invocation wire. `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.
|
||||
|
||||
|
||||
@@ -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 的菜单：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——斜杠手势是这类条目唯一的入口。列表是 skill 领域唯一的 RPC——调用本身就是一次普通的 `session.prompt`，`dsh-tool-skill` 会在 pre-step 边界识别其中以空白为界的 `/name` token，并以注入的 `<skill_content>` 上下文作答，因此每一种前端（web、TUI、ACP、手动键入的文本）共享同一条确定性路径，没有专设的调用协议。`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。
|
||||
|
||||
|
||||
@@ -18,6 +18,7 @@ import type { SessionPersistence } from '@deepseek-ai/dsh-session-persistence'
|
||||
import { SessionQueryError, type SessionSearchCursor } from '@deepseek-ai/dsh-session-query'
|
||||
import { SubagentError } from '@deepseek-ai/dsh-subagent'
|
||||
import type { SubagentListEntry as CatalogSubagentListEntry } from '@deepseek-ai/dsh-subagent'
|
||||
import { isUserInvocable } from '@deepseek-ai/dsh-skill'
|
||||
import type { Workspace, WorkspaceRecord } from '@deepseek-ai/dsh-workspace'
|
||||
import {
|
||||
workspaceDomainState, workspaceRecord, WorkspaceId as brandWorkspaceId,
|
||||
@@ -1246,6 +1247,34 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
|
||||
return llm === undefined || llm.listProviders().some(entry => entry.id === provider)
|
||||
}
|
||||
|
||||
/**
|
||||
* Resolve the addressed agent for a turn-starting method and refuse when no
|
||||
* adapter serves its current route: a route nothing serves cannot start a
|
||||
* turn, and letting it try spends the whole pre-step path to fail inside
|
||||
* the adapter with a message about registration. Refusing here names the
|
||||
* model the session is pointed at while the draft is still in the composer.
|
||||
* This is `session.prompt`'s enforcement boundary: a client that disables
|
||||
* its input is an affordance, and the method stays callable regardless.
|
||||
*/
|
||||
async function turnAgentFor<T>(
|
||||
request: RpcRequest<unknown>, sessionId: SessionId,
|
||||
): Promise<{ agent: Agent } | { refused: RpcResponse<T> }> {
|
||||
const found = await agentFor(sessionId)
|
||||
if ('error' in found) return { refused: err(request, found.error) }
|
||||
const agent = found.agent
|
||||
const target = targetFor(agent).current
|
||||
if (!routeServed(target.provider)) {
|
||||
return {
|
||||
refused: err(request, {
|
||||
code: 'model-unavailable',
|
||||
message: `no adapter serves provider "${target.provider}"; select a model for this session`,
|
||||
details: { provider: target.provider, model: target.model },
|
||||
}),
|
||||
}
|
||||
}
|
||||
return { agent }
|
||||
}
|
||||
|
||||
/** Missing-service report shared by the settings domain (skills-domain stance). */
|
||||
function settingsAbsent(): RpcError {
|
||||
return { code: 'internal', message: 'settings service is absent: this deployment does not mount a settings provider (e.g. @deepseek-ai/dsh-settings-local) in its composition', details: {} }
|
||||
@@ -1782,23 +1811,9 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
|
||||
|
||||
async prompt(request) {
|
||||
const { sessionId, mode, content } = request.payload
|
||||
const found = await agentFor(sessionId)
|
||||
if ('error' in found) return err(request, found.error)
|
||||
const agent = found.agent
|
||||
// A route no adapter serves cannot start a turn, and letting it try
|
||||
// spends the whole pre-step path to fail inside the adapter with a
|
||||
// message about registration. Refusing here names the model the
|
||||
// session is pointed at while the draft is still in the composer.
|
||||
// This is the enforcement boundary: a client that disables its input
|
||||
// is an affordance, and this method stays callable regardless.
|
||||
const target = targetFor(agent).current
|
||||
if (!routeServed(target.provider)) {
|
||||
return err(request, {
|
||||
code: 'model-unavailable',
|
||||
message: `no adapter serves provider "${target.provider}"; select a model for this session`,
|
||||
details: { provider: target.provider, model: target.model },
|
||||
})
|
||||
}
|
||||
const resolved = await turnAgentFor<{ accepted: true }>(request, sessionId)
|
||||
if ('refused' in resolved) return resolved.refused
|
||||
const agent = resolved.agent
|
||||
// The rpcId rides MessageSource into user/message (merge declaration in api/sessions.ts; provisional correlation).
|
||||
const source: MessageSource = { kind: 'user', rpcId: request.rpcId }
|
||||
try {
|
||||
@@ -2359,13 +2374,13 @@ export function createApiProxy(ctx: Context, defaults: ApiProxyDefaults): ApiPro
|
||||
return err(request, { code: 'internal', message: 'skill registry is absent: this deployment does not mount @deepseek-ai/dsh-skill in its composition (cordis.yml or explicit assembly)', details: {} })
|
||||
}
|
||||
try {
|
||||
const skills = (await skillRegistry.list({ cwd }))
|
||||
.filter(skill => skill.invocation.modelInvocable && skill.invocation.userInvocable)
|
||||
const skills = (await skillRegistry.list({ cwd })).filter(isUserInvocable)
|
||||
return ok(request, {
|
||||
skills: skills.map(skill => ({
|
||||
name: skill.name,
|
||||
description: skill.description,
|
||||
...skill.whenToUse === undefined ? {} : { whenToUse: skill.whenToUse },
|
||||
modelInvocable: skill.invocation.modelInvocable,
|
||||
})),
|
||||
})
|
||||
} catch (error: unknown) {
|
||||
|
||||
@@ -14,6 +14,7 @@ export const skillEntrySchema = z.object({
|
||||
name: z.string().min(1),
|
||||
description: z.string(),
|
||||
whenToUse: z.string().optional(),
|
||||
modelInvocable: z.boolean(),
|
||||
}) satisfies z.ZodType<Wire<SkillEntry>>
|
||||
|
||||
/** skill.list request payload. */
|
||||
|
||||
@@ -10,16 +10,24 @@ import type { RpcRequest, RpcResponse } from './rpc.ts'
|
||||
|
||||
/** Skill catalog row (wire projection of the host SkillSummary; provider/source vocabulary stays host-side). */
|
||||
export interface SkillEntry {
|
||||
/** Kebab-case identifier referenced as `<skill>name</skill>` in prompts. */
|
||||
/** Kebab-case identifier the user references as `/name` in the composer. */
|
||||
readonly name: string
|
||||
/** Short routing description. */
|
||||
readonly description: string
|
||||
/** Optional extra routing guidance. */
|
||||
readonly whenToUse?: string
|
||||
/** False marks a user-only skill (`disable-model-invocation`): invocable here, absent from the model catalog. */
|
||||
readonly modelInvocable: boolean
|
||||
}
|
||||
|
||||
/** Skill-domain unary methods (the map key skill.* of RpcMethodMap). */
|
||||
/**
|
||||
* Skill-domain unary methods (the map key skill.* of RpcMethodMap). Listing
|
||||
* is the domain's only RPC: invocation itself is a plain `session.prompt`
|
||||
* whose leading `/name` token the host recognizes at the pre-step boundary
|
||||
* (`dsh-tool-skill` injects the rendered body there), so every client shares
|
||||
* one deterministic path with no dedicated invocation wire.
|
||||
*/
|
||||
export interface SkillsApi {
|
||||
/** Lists skills usable by the browser's user-selected model-reference path. */
|
||||
/** Lists the user-invocable skill catalog for the session's project. */
|
||||
list(request: RpcRequest<{ sessionId: SessionId }>): Promise<RpcResponse<{ skills: readonly SkillEntry[] }>>
|
||||
}
|
||||
|
||||
@@ -228,7 +228,10 @@ describe('skill.list', () => {
|
||||
// touch (or resume through) the Agent registry.
|
||||
const session = ctx.sessions.create(undefined, { meta: { cwd: '/proj' } })
|
||||
const value = expectOk(await api.skills.list(request({ sessionId: session.id })))
|
||||
expect(value.skills).toEqual([{ name: 'commit-helper', description: 'Git commits', whenToUse: 'when committing' }])
|
||||
expect(value.skills).toEqual([
|
||||
{ name: 'commit-helper', description: 'Git commits', whenToUse: 'when committing', modelInvocable: true },
|
||||
{ name: 'user-only', description: 'User-only', modelInvocable: false },
|
||||
])
|
||||
expect(seenCwds).toEqual(['/proj'])
|
||||
expect(ctx.agents.get(session.id)).toBeUndefined()
|
||||
})
|
||||
|
||||
@@ -196,7 +196,7 @@ function fakeApi(overrides: Partial<{ muxFrames: MuxFrame[]; hostFrames: HostFra
|
||||
},
|
||||
skills: {
|
||||
async list(request) {
|
||||
return { rpcId: request.rpcId, result: { ok: true, value: { skills: [{ name: 'commit-helper', description: 'Git commits' }] } } }
|
||||
return { rpcId: request.rpcId, result: { ok: true, value: { skills: [{ name: 'commit-helper', description: 'Git commits', modelInvocable: true }] } } }
|
||||
},
|
||||
},
|
||||
goals: {
|
||||
@@ -381,7 +381,7 @@ describe('unary round trip (handler ⇄ client, no network)', () => {
|
||||
const miss = await c.commands.execute({ sessionId: 's' as never, line: '/nope' })
|
||||
expect(miss.result).toEqual({ ok: true, value: { matched: false } })
|
||||
const skills = await c.skills.list({ sessionId: 's' as never })
|
||||
expect(skills.result).toEqual({ ok: true, value: { skills: [{ name: 'commit-helper', description: 'Git commits' }] } })
|
||||
expect(skills.result).toEqual({ ok: true, value: { skills: [{ name: 'commit-helper', description: 'Git commits', modelInvocable: true }] } })
|
||||
})
|
||||
|
||||
it('lets command.execute finish after the 30-second default unary deadline', async () => {
|
||||
|
||||
@@ -395,12 +395,15 @@ describe('skills domain schemas', () => {
|
||||
expect(() => skillListRequestSchema.parse({})).toThrow()
|
||||
expect(skillListValueSchema.parse({ skills: [] }).skills).toEqual([])
|
||||
const value = skillListValueSchema.parse({ skills: [
|
||||
{ name: 'commit-helper', description: 'Git commits', whenToUse: 'when committing' },
|
||||
{ name: 'bare', description: 'No guidance' },
|
||||
{ name: 'commit-helper', description: 'Git commits', whenToUse: 'when committing', modelInvocable: true },
|
||||
{ name: 'bare', description: 'No guidance', modelInvocable: false },
|
||||
] })
|
||||
expect(value.skills[0]?.whenToUse).toBe('when committing')
|
||||
expect(value.skills[1]?.whenToUse).toBeUndefined()
|
||||
expect(() => skillEntrySchema.parse({ name: '', description: 'd' })).toThrow()
|
||||
expect(value.skills[1]?.modelInvocable).toBe(false)
|
||||
expect(() => skillEntrySchema.parse({ name: '', description: 'd', modelInvocable: true })).toThrow()
|
||||
// modelInvocable is required wire data: an entry without it fails.
|
||||
expect(() => skillEntrySchema.parse({ name: 'n', description: 'd' })).toThrow()
|
||||
})
|
||||
})
|
||||
|
||||
|
||||
@@ -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: 3dc2bcfa5775736717bdebcb92329d5655198234
|
||||
README.zh.md: d11f90d5a8356f06df63aa249a1f8b5851f36f5f
|
||||
|
||||
@@ -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 injects it at the user-explicit gesture boundary, 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, form: 'instructions' }) 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
|
||||
|
||||
@@ -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，并在用户显式的手势边界将其注入，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, form: 'instructions' }），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
|
||||
|
||||
@@ -26,6 +26,7 @@
|
||||
"license": "BSD-3-Clause",
|
||||
"peerDependencies": {
|
||||
"@deepseek-ai/dsh-invariants": "^0.0.1",
|
||||
"@deepseek-ai/dsh-llm": "^0.0.1",
|
||||
"cordis": "^4.0.0-rc.7"
|
||||
},
|
||||
"dependencies": {
|
||||
@@ -33,6 +34,7 @@
|
||||
},
|
||||
"devDependencies": {
|
||||
"@deepseek-ai/dsh-invariants": "workspace:^",
|
||||
"@deepseek-ai/dsh-llm": "workspace:^",
|
||||
"cordis": "^4.0.0-rc.7"
|
||||
}
|
||||
}
|
||||
|
||||
@@ -10,6 +10,7 @@
|
||||
*/
|
||||
|
||||
import { Context, Service } from 'cordis'
|
||||
import { assertNever } from '@deepseek-ai/dsh-llm'
|
||||
import z from 'schemastery'
|
||||
import type Schema from 'schemastery'
|
||||
|
||||
@@ -119,6 +120,97 @@ export function isUserInvocable(skill: Pick<SkillSummary, 'invocation'>): boolea
|
||||
return skill.invocation.userInvocable
|
||||
}
|
||||
|
||||
/**
|
||||
* Durable source for the context message a user-explicit skill invocation
|
||||
* injects: the user's own words ride a plain user message, and the rendered
|
||||
* skill body follows as injected `instructions`-form context carrying this
|
||||
* source, so transcript consumers present the injection from metadata
|
||||
* instead of re-parsing the model-facing text.
|
||||
*/
|
||||
export interface SkillInvocationSource {
|
||||
readonly kind: 'skill-invocation'
|
||||
/** Invoked skill name, validated user-invocable at the injecting boundary. */
|
||||
readonly name: string
|
||||
/** Injected skill bodies are instructions for the model to follow. */
|
||||
readonly form: 'instructions'
|
||||
}
|
||||
|
||||
declare module '@deepseek-ai/dsh-llm' {
|
||||
interface MessageSourceMap {
|
||||
/** A user-explicit skill invocation injected by the host. */
|
||||
'skill-invocation': SkillInvocationSource
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Render one loaded skill for the model. The output is shared verbatim by the
|
||||
* `skill` tool result and the user-explicit invocation injection, so the model
|
||||
* sees one canonical `<skill_content>` shape on both paths. The name rides an
|
||||
* escaped attribute; the body is embedded verbatim (skills are trusted local
|
||||
* content, and user-supplied invocation text stays outside this wrapper).
|
||||
* @param skill - name, provider, optional resource base, and body to render.
|
||||
* @returns the complete model-facing `<skill_content>` block.
|
||||
*/
|
||||
export function renderSkillContent(skill: Pick<SkillDefinition, 'name' | 'provider' | 'resourceBase' | 'content'>): string {
|
||||
const resourceHint = renderResourceHint(skill)
|
||||
return [
|
||||
`<skill_content name="${escapeAttr(skill.name)}">`,
|
||||
'<skill_resources>',
|
||||
...resourceHint,
|
||||
'</skill_resources>',
|
||||
'',
|
||||
'<skill_instructions>',
|
||||
skill.content,
|
||||
'</skill_instructions>',
|
||||
'</skill_content>',
|
||||
].join('\n')
|
||||
}
|
||||
|
||||
function renderResourceHint(skill: Pick<SkillDefinition, 'provider' | 'resourceBase'>): string[] {
|
||||
const base = skill.resourceBase
|
||||
if (base === undefined) {
|
||||
return [
|
||||
`Resources for this skill are managed by provider "${escapeText(skill.provider)}".`,
|
||||
'Load referenced resources only as needed.',
|
||||
]
|
||||
}
|
||||
switch (base.kind) {
|
||||
case 'directory':
|
||||
return [
|
||||
`Base directory for this skill: ${escapeText(base.path)}`,
|
||||
'Resolve relative paths mentioned by this skill against the base directory before using them. Load referenced resources only as needed.',
|
||||
]
|
||||
case 'url':
|
||||
return [
|
||||
`Base URL for this skill: ${escapeText(base.url)}`,
|
||||
'Resolve relative URLs mentioned by this skill against the base URL before using them. Load referenced resources only as needed.',
|
||||
]
|
||||
case 'opaque':
|
||||
return [
|
||||
`Resources for this skill: ${escapeText(base.description)}`,
|
||||
'Load referenced resources only as needed.',
|
||||
]
|
||||
/* v8 ignore start -- SkillResourceBase is a closed union; a future kind must fail compilation here. */
|
||||
default:
|
||||
return assertNever(base, 'SkillResourceBase.kind')
|
||||
/* v8 ignore stop */
|
||||
}
|
||||
}
|
||||
|
||||
function escapeAttr(value: string): string {
|
||||
return value.replaceAll('&', '&').replaceAll('"', '"').replaceAll('<', '<')
|
||||
}
|
||||
|
||||
/**
|
||||
* Escape model-facing prose embedded inside skill markup so provider-supplied
|
||||
* text cannot open or close framing tags.
|
||||
* @param value - raw prose to embed.
|
||||
* @returns the escaped text.
|
||||
*/
|
||||
export function escapeText(value: string): string {
|
||||
return value.replaceAll('&', '&').replaceAll('<', '<').replaceAll('>', '>')
|
||||
}
|
||||
|
||||
/** One catalog observation plus whether discovery completed within a stable catalog revision. */
|
||||
export interface SkillCatalogSnapshot {
|
||||
/** Sorted invocation-neutral summaries collected in this observation. */
|
||||
|
||||
@@ -3,6 +3,7 @@ import { Context } from 'cordis'
|
||||
import SkillService, {
|
||||
isModelInvocable,
|
||||
isUserInvocable,
|
||||
renderSkillContent,
|
||||
type SkillCandidate,
|
||||
type SkillDefinition,
|
||||
type SkillInvocationPolicy,
|
||||
@@ -1013,3 +1014,65 @@ describe('SkillService registry', () => {
|
||||
expect(await ctx.skills.get('same-skill')).toBeUndefined()
|
||||
})
|
||||
})
|
||||
|
||||
describe('renderSkillContent', () => {
|
||||
it('renders a directory-based skill with the shared wrapper', () => {
|
||||
const text = renderSkillContent({
|
||||
name: 'demo-skill',
|
||||
provider: 'memory',
|
||||
resourceBase: { kind: 'directory', path: '/tmp/demo' },
|
||||
content: 'Do the thing.',
|
||||
})
|
||||
expect(text).toBe([
|
||||
'<skill_content name="demo-skill">',
|
||||
'<skill_resources>',
|
||||
'Base directory for this skill: /tmp/demo',
|
||||
'Resolve relative paths mentioned by this skill against the base directory before using them. Load referenced resources only as needed.',
|
||||
'</skill_resources>',
|
||||
'',
|
||||
'<skill_instructions>',
|
||||
'Do the thing.',
|
||||
'</skill_instructions>',
|
||||
'</skill_content>',
|
||||
].join('\n'))
|
||||
})
|
||||
|
||||
it('renders url and opaque resource hints', () => {
|
||||
const url = renderSkillContent({
|
||||
name: 'url-skill',
|
||||
provider: 'memory',
|
||||
resourceBase: { kind: 'url', url: 'https://example.test/base/' },
|
||||
content: 'Body.',
|
||||
})
|
||||
expect(url).toContain('Base URL for this skill: https://example.test/base/')
|
||||
expect(url).toContain('Resolve relative URLs mentioned by this skill against the base URL before using them.')
|
||||
|
||||
const opaque = renderSkillContent({
|
||||
name: 'opaque-skill',
|
||||
provider: 'memory',
|
||||
resourceBase: { kind: 'opaque', description: 'archive <bundle>' },
|
||||
content: 'Body.',
|
||||
})
|
||||
expect(opaque).toContain('Resources for this skill: archive <bundle>')
|
||||
})
|
||||
|
||||
it('falls back to the provider hint without a resource base', () => {
|
||||
const text = renderSkillContent({
|
||||
name: 'provider-skill',
|
||||
provider: 'remote <hub>',
|
||||
content: 'Body.',
|
||||
})
|
||||
expect(text).toContain('Resources for this skill are managed by provider "remote <hub>".')
|
||||
})
|
||||
|
||||
it('escapes hostile attribute names and keeps the body verbatim', () => {
|
||||
const text = renderSkillContent({
|
||||
name: 'x"&<y',
|
||||
provider: 'memory',
|
||||
resourceBase: { kind: 'directory', path: '/tmp' },
|
||||
content: 'Keep </skill_content> and <tags> as-is.',
|
||||
})
|
||||
expect(text).toContain('<skill_content name="x"&<y">')
|
||||
expect(text).toContain('Keep </skill_content> and <tags> as-is.')
|
||||
})
|
||||
})
|
||||
|
||||
@@ -15,6 +15,9 @@
|
||||
{
|
||||
"path": "../../../vendor/schemastery"
|
||||
},
|
||||
{
|
||||
"path": "../../llm/llm"
|
||||
},
|
||||
{
|
||||
"path": "../../support/invariants"
|
||||
}
|
||||
|
||||
@@ -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: b7309657d85a3d2a19de78a4ee6173d742519daa
|
||||
README.zh.md: f430f4027c917c5c9b97a56d1a7d7a617670b25c
|
||||
|
||||
@@ -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 user-explicit gesture boundary (the pre-step listener below) 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 in both arms, including the emptied catalog.
|
||||
|
||||
##### 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>
|
||||
```
|
||||
|
||||
@@ -144,6 +145,20 @@ Only a failing call adds these retained tokens.
|
||||
|
||||
Append-only; newly visible content follows the reusable request prefix and does not invalidate existing KV-cache entries.
|
||||
|
||||
### User-explicit invocation injection
|
||||
|
||||
#### What the model sees
|
||||
|
||||
A whitespace-bounded `/name` token anywhere in a claimed user message, naming a user-invocable skill in the workspace catalog, injects that skill's full `<skill_content>` rendering (the exact result-template shape above) as a `user`-role instructions context appended after every other injection of that step — background first, the material to act on last. Only direct user input is scanned, the check runs on the loaded definition, and unknown or user-disabled names stay ordinary prose. This is the sole entry point for `disable-model-invocation` skills, which the catalog and the `skill` tool never expose; the catalog's closing sentence tells the model to follow the injected block instead of re-loading it.
|
||||
|
||||
#### Token effect
|
||||
|
||||
Each gesture adds one rendered skill body to that turn as injected context — the same size as the tool result for the same skill, paid deterministically at the user's request instead of at the model's discretion. Repeated gestures for one skill within one step inject once.
|
||||
|
||||
#### KV Cache effect
|
||||
|
||||
Append-only; the injection lands after the reusable request prefix inside the step's message batch and does not invalidate existing KV-cache entries.
|
||||
|
||||
## Known Limitations and Deferred Work
|
||||
|
||||
- **The catalog omits `whenToUse`, source, and provider metadata** — routing is based only on name and a capped description; `whenToUse` remains provider metadata and is not rendered by the loaded wrapper either.
|
||||
|
||||
@@ -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：用户显式的手势边界（下文的 pre-step 监听器）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；替换目录模板的两个臂——包括清空后的目录——都携带同一句话。
|
||||
|
||||
##### 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>
|
||||
```
|
||||
|
||||
@@ -144,6 +145,20 @@ Load referenced resources only as needed.
|
||||
|
||||
仅追加；新可见内容位于可重用请求前缀之后，不会使现有 KV Cache 条目失效。
|
||||
|
||||
### 用户显式调用注入
|
||||
|
||||
#### 模型看到的内容
|
||||
|
||||
已认领用户消息中任意位置、以空白为界、指名工作区目录中某个用户可调用 skill 的 `/name` token，会把该 skill 的完整 `<skill_content>` 渲染（与上文结果模板完全相同的形态）作为 `user` 角色的指令上下文注入，追加在该步骤所有其他注入之后——背景在前，模型要着手处理的材料在最后。只扫描直接的用户输入，检查在已加载定义上进行，未知名称和用户不可调用的名称保持为普通行文。这是 `disable-model-invocation` skill 唯一的入口，目录和 `skill` 工具永不暴露这类 skill；目录的结尾一句会告诉模型遵循注入块，而不是重新加载它。
|
||||
|
||||
#### Token 影响
|
||||
|
||||
每次手势会把一份渲染后的 skill 正文作为注入上下文加进该轮次——尺寸与同一 skill 的工具结果相同，按用户的请求确定性地支付，而非由模型自行裁量。同一步骤内对同一 skill 的重复手势只注入一次。
|
||||
|
||||
#### KV Cache 影响
|
||||
|
||||
仅追加；注入落在该步骤的消息批次中、可重用请求前缀之后，不会使现有 KV Cache 条目失效。
|
||||
|
||||
## 已知限制与暂缓事项
|
||||
|
||||
- **目录省略 `whenToUse`、来源和提供方元数据**：路由只基于名称和有长度上限的描述；`whenToUse` 仍是提供方元数据，加载后的包装层也不渲染它。
|
||||
|
||||
@@ -9,12 +9,15 @@ import type { Context } from 'cordis'
|
||||
import z from 'schemastery'
|
||||
import type { Agent, PreStepDecision } from '@deepseek-ai/dsh-agent'
|
||||
import { defineTool } from '@deepseek-ai/dsh-tools'
|
||||
import { assertNever, createUserMessage } from '@deepseek-ai/dsh-llm'
|
||||
import { createUserMessage } from '@deepseek-ai/dsh-llm'
|
||||
import type { UserMessage } from '@deepseek-ai/dsh-session'
|
||||
import {
|
||||
escapeText,
|
||||
isModelInvocable,
|
||||
isSkillName,
|
||||
type SkillDefinition,
|
||||
isUserInvocable,
|
||||
renderSkillContent,
|
||||
type SkillInvocationSource,
|
||||
type SkillSummary,
|
||||
} from '@deepseek-ai/dsh-skill'
|
||||
|
||||
@@ -160,6 +163,49 @@ export function apply(ctx: Context, config: Config = {}): void {
|
||||
throw new Error('dsh-tool-skill: registered skill tool is not visible in the global registry')
|
||||
}
|
||||
|
||||
// User-explicit skill invocation: a claimed user message whose first line
|
||||
// starts with `/<name>` naming a user-invocable skill is a deterministic
|
||||
// load gesture. The rendered body enters this step as injected
|
||||
// instructions context appended after every other injection — background
|
||||
// first (workspace rules, runtime policy, the catalog), the material the
|
||||
// model must act on last, closest to its answer. Registration order makes
|
||||
// that placement deterministic: this listener registers before the catalog
|
||||
// listener, so the waterfall hands it the catalog-bearing list to extend.
|
||||
// Only `source.kind === 'user'` messages are scanned — external text
|
||||
// cannot forge the gesture — and a token naming no user-invocable skill
|
||||
// stays ordinary prose (the command registry is a different closed
|
||||
// namespace, resolved client-side before a line ever becomes a prompt).
|
||||
// This is the only entry point for `disable-model-invocation` skills; the
|
||||
// catalog and the `skill` tool below never see them.
|
||||
ctx.on('agent/pre-step', async (
|
||||
{ agent, messages, signal },
|
||||
next,
|
||||
): Promise<PreStepDecision> => {
|
||||
const decision = await next()
|
||||
if (decision.kind === 'reject') return decision
|
||||
const names = invokedSkillNames(messages)
|
||||
if (names.length === 0) return decision
|
||||
signal.throwIfAborted()
|
||||
const lookup = { cwd: agent.session.header.cwd, signal }
|
||||
const injections: UserMessage[] = []
|
||||
for (const name of names) {
|
||||
const skill = await ctx.skills.get(name, lookup)
|
||||
signal.throwIfAborted()
|
||||
// Unknown names and user-disabled skills stay plain prose: the
|
||||
// gesture was never a claim this boundary recognizes. The check sits
|
||||
// on the loaded definition — the single lookup that produces what is
|
||||
// actually injected.
|
||||
if (skill === undefined || !isUserInvocable(skill)) continue
|
||||
const source: SkillInvocationSource = { kind: 'skill-invocation', name, form: 'instructions' }
|
||||
injections.push(createUserMessage({
|
||||
content: [{ type: 'text', text: renderSkillContent(skill) }],
|
||||
source,
|
||||
}))
|
||||
}
|
||||
if (injections.length === 0) return decision
|
||||
return { kind: 'enter', messages: [...decision.messages, ...injections] }
|
||||
})
|
||||
|
||||
// Register after the tool so reverse teardown removes guidance first. Exact definition
|
||||
// identity prevents a scoped shadow merely named `skill` from inheriting this catalog.
|
||||
ctx.on('agent/pre-step', async (
|
||||
@@ -203,52 +249,6 @@ export function apply(ctx: Context, config: Config = {}): void {
|
||||
})
|
||||
}
|
||||
|
||||
function renderSkillContent(skill: Pick<SkillDefinition, 'name' | 'provider' | 'resourceBase' | 'content'>): string {
|
||||
const resourceHint = renderResourceHint(skill)
|
||||
return [
|
||||
`<skill_content name="${escapeAttr(skill.name)}">`,
|
||||
'<skill_resources>',
|
||||
...resourceHint,
|
||||
'</skill_resources>',
|
||||
'',
|
||||
'<skill_instructions>',
|
||||
skill.content,
|
||||
'</skill_instructions>',
|
||||
'</skill_content>',
|
||||
].join('\n')
|
||||
}
|
||||
|
||||
function renderResourceHint(skill: Pick<SkillDefinition, 'provider' | 'resourceBase'>): string[] {
|
||||
const base = skill.resourceBase
|
||||
if (base === undefined) {
|
||||
return [
|
||||
`Resources for this skill are managed by provider "${escapeText(skill.provider)}".`,
|
||||
'Load referenced resources only as needed.',
|
||||
]
|
||||
}
|
||||
switch (base.kind) {
|
||||
case 'directory':
|
||||
return [
|
||||
`Base directory for this skill: ${escapeText(base.path)}`,
|
||||
'Resolve relative paths mentioned by this skill against the base directory before using them. Load referenced resources only as needed.',
|
||||
]
|
||||
case 'url':
|
||||
return [
|
||||
`Base URL for this skill: ${escapeText(base.url)}`,
|
||||
'Resolve relative URLs mentioned by this skill against the base URL before using them. Load referenced resources only as needed.',
|
||||
]
|
||||
case 'opaque':
|
||||
return [
|
||||
`Resources for this skill: ${escapeText(base.description)}`,
|
||||
'Load referenced resources only as needed.',
|
||||
]
|
||||
/* v8 ignore start -- SkillResourceBase is a closed union; a future kind must fail compilation here. */
|
||||
default:
|
||||
return assertNever(base, 'SkillResourceBase.kind')
|
||||
/* v8 ignore stop */
|
||||
}
|
||||
}
|
||||
|
||||
function renderCatalogMessage(entries: SkillCatalogSource['entries']): UserMessage {
|
||||
return createUserMessage({
|
||||
content: [{
|
||||
@@ -262,6 +262,7 @@ function renderCatalogMessage(entries: SkillCatalogSource['entries']): UserMessa
|
||||
'</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>',
|
||||
].join('\n'),
|
||||
}],
|
||||
@@ -277,9 +278,11 @@ function renderCatalogUpdate(entries: SkillCatalogSource['entries']): UserMessag
|
||||
const availability = entries.length === 0
|
||||
? [
|
||||
'No skills are currently available through the `skill` tool. Do not use names from earlier skill catalogs.',
|
||||
'A user may still invoke a skill directly; its <skill_content> block then appears in this conversation. Follow it, and do not call the `skill` tool for it.',
|
||||
]
|
||||
: [
|
||||
'Use only names in this replacement catalog. If the user names a listed skill, or the task clearly matches its description, call the `skill` tool with the exact name before acting.',
|
||||
'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.',
|
||||
]
|
||||
return createUserMessage({
|
||||
content: [{
|
||||
@@ -394,10 +397,33 @@ function assertPositiveInteger(name: string, value: number, minimum = 1): void {
|
||||
}
|
||||
}
|
||||
|
||||
function escapeAttr(value: string): string {
|
||||
return value.replaceAll('&', '&').replaceAll('"', '"').replaceAll('<', '<')
|
||||
}
|
||||
/**
|
||||
* A whitespace-bounded `/name` token (the public skill-name grammar) anywhere
|
||||
* in the text — the same word-boundary shape the transcript chip decoration
|
||||
* uses, so a gesture reads as one wherever it sits in the sentence. A second
|
||||
* `/` or any non-boundary character breaks the match, which keeps file paths
|
||||
* (`/usr/bin`) and fractions (`5/8`) out.
|
||||
*/
|
||||
const SKILL_GESTURE = /(^|\s)\/([a-z0-9]+(?:-[a-z0-9]+)*)(?=\s|$)/g
|
||||
|
||||
function escapeText(value: string): string {
|
||||
return value.replaceAll('&', '&').replaceAll('<', '<').replaceAll('>', '>')
|
||||
/**
|
||||
* `/name` gesture tokens from the claimed user messages, deduplicated in
|
||||
* first-seen order. Every text block of direct user input is scanned; no
|
||||
* other source can forge a gesture.
|
||||
* @param messages - the step's claimed batch.
|
||||
* @returns candidate skill names, unvalidated against the registry.
|
||||
*/
|
||||
function invokedSkillNames(messages: readonly UserMessage[]): string[] {
|
||||
const names: string[] = []
|
||||
for (const message of messages) {
|
||||
if ((message.source as { kind?: unknown }).kind !== 'user') continue
|
||||
for (const block of message.content) {
|
||||
if (block.type !== 'text') continue
|
||||
for (const match of block.text.matchAll(SKILL_GESTURE)) {
|
||||
const name = match[2]
|
||||
if (name !== undefined && !names.includes(name)) names.push(name)
|
||||
}
|
||||
}
|
||||
}
|
||||
return names
|
||||
}
|
||||
|
||||
@@ -287,6 +287,7 @@ describe('dsh-tool-skill', () => {
|
||||
'</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>',
|
||||
].join('\n'),
|
||||
}],
|
||||
@@ -914,3 +915,134 @@ describe('dsh-tool-skill', () => {
|
||||
expect(vanishedBlock.text).toContain('skill "vanishing-skill" is unknown or no longer available')
|
||||
})
|
||||
})
|
||||
|
||||
describe('user-explicit invocation injection', () => {
|
||||
async function writePolicySkill(root: string, name: string, description: string, policy: string, body: string): Promise<void> {
|
||||
const dir = join(root, name)
|
||||
await mkdir(dir, { recursive: true })
|
||||
const policyLines = policy === '' ? '' : `${policy}\n`
|
||||
await writeFile(join(dir, 'SKILL.md'), `---\nname: ${name}\ndescription: ${description}\n${policyLines}---\n\n${body}\n`)
|
||||
}
|
||||
|
||||
function gesture(text: string): UserMessage {
|
||||
return createUserMessage({ content: [{ type: 'text', text }], source: { kind: 'user' } })
|
||||
}
|
||||
|
||||
async function invokeHarness(): Promise<{ ctx: Context; agent: Agent }> {
|
||||
const home = await tempDir('invoke')
|
||||
const skillsRoot = join(home, '.agents', 'skills')
|
||||
await writePolicySkill(skillsRoot, 'hidden-demo', 'User-only demo', 'disable-model-invocation: true', 'Say the magic word: PINEAPPLE.')
|
||||
await writePolicySkill(skillsRoot, 'shared-skill', 'Ordinary skill', '', 'Shared instructions.')
|
||||
await writePolicySkill(skillsRoot, 'model-only-skill', 'Model only', 'user-invocable: false', 'Model-only instructions.')
|
||||
const ctx = await setup(home)
|
||||
return { ctx, agent: agentForCwd(home) }
|
||||
}
|
||||
|
||||
it('injects a user-invocable skill named by a leading /token, after every other injection', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const first = gesture('/hidden-demo what does this do')
|
||||
const second = gesture('plain follow-up prose')
|
||||
const decision = await proposeStep(ctx, agent, [first, second])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
const kinds = decision.messages.map(message => (message.source as { kind: string }).kind)
|
||||
// Background injections (the catalog here) sit between the claimed batch
|
||||
// and the invoked body: the material the model must act on comes last.
|
||||
expect(kinds.slice(0, 2)).toEqual(['user', 'user'])
|
||||
expect(kinds.at(-1)).toBe('skill-invocation')
|
||||
expect(kinds.indexOf('skill-catalog')).toBeLessThan(kinds.indexOf('skill-invocation'))
|
||||
const injection = decision.messages.at(-1)!
|
||||
expect(injection.source).toMatchObject({ kind: 'skill-invocation', name: 'hidden-demo', form: 'instructions' })
|
||||
const block = injection.content[0]
|
||||
if (block?.type !== 'text') throw new Error('expected text injection')
|
||||
expect(block.text).toContain('<skill_content name="hidden-demo">')
|
||||
expect(block.text).toContain('Say the magic word: PINEAPPLE.')
|
||||
expect(block.text).not.toContain('what does this do')
|
||||
})
|
||||
|
||||
it('injects an ordinary skill the same way (one uniform user-explicit path)', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const decision = await proposeStep(ctx, agent, [gesture('/shared-skill go')])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
expect(decision.messages.some(message =>
|
||||
(message.source as { kind?: string; name?: string }).kind === 'skill-invocation'
|
||||
&& (message.source as { name?: string }).name === 'shared-skill')).toBe(true)
|
||||
})
|
||||
|
||||
it('recognizes a mid-sentence gesture but not paths, fractions, or broken boundaries', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const decision = await proposeStep(ctx, agent, [
|
||||
gesture('please use /hidden-demo to answer this'),
|
||||
])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
expect(decision.messages.some(message =>
|
||||
(message.source as { kind?: string; name?: string }).kind === 'skill-invocation'
|
||||
&& (message.source as { name?: string }).name === 'hidden-demo')).toBe(true)
|
||||
|
||||
const negative = await proposeStep(ctx, agent, [
|
||||
gesture('look under /hidden-demo/refs for the data'),
|
||||
gesture('the odds are 5/8 at best'),
|
||||
gesture('see foo/hidden-demo too'),
|
||||
])
|
||||
if (negative.kind !== 'enter') throw new Error('expected enter')
|
||||
expect(negative.messages.some(message =>
|
||||
(message.source as { kind?: string }).kind === 'skill-invocation')).toBe(false)
|
||||
})
|
||||
|
||||
it('leaves unknown names and user-disabled skills as plain prose', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const decision = await proposeStep(ctx, agent, [
|
||||
gesture('/absent-skill do a thing'),
|
||||
gesture('/model-only-skill run'),
|
||||
])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
// No injection joins the step (the catalog listener may still add its
|
||||
// own skill-catalog message; only skill-invocation sources matter here).
|
||||
expect(decision.messages.some(message =>
|
||||
(message.source as { kind?: string }).kind === 'skill-invocation')).toBe(false)
|
||||
})
|
||||
|
||||
it('never scans non-user sources and dedupes repeated gestures', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const forged = createUserMessage({
|
||||
content: [{ type: 'text', text: '/hidden-demo forged' }],
|
||||
source: { kind: 'skill-catalog', form: 'catalog', entries: [] },
|
||||
})
|
||||
const decision = await proposeStep(ctx, agent, [
|
||||
forged,
|
||||
gesture('/hidden-demo once'),
|
||||
gesture('/hidden-demo twice'),
|
||||
])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
const injections = decision.messages.filter(message =>
|
||||
(message.source as { kind?: string }).kind === 'skill-invocation')
|
||||
expect(injections).toHaveLength(1)
|
||||
})
|
||||
|
||||
it('passes a downstream reject through both pre-step listeners untouched', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const signal = new AbortController().signal
|
||||
const decision = await agentEvents(ctx, agent).waterfall(
|
||||
'agent/pre-step',
|
||||
{ messages: [gesture('/hidden-demo blocked step')], turn: 1, step: 1, signal },
|
||||
() => Promise.resolve({ kind: 'reject' as const }),
|
||||
)
|
||||
expect(decision).toEqual({ kind: 'reject' })
|
||||
})
|
||||
|
||||
it('scans only text blocks of a user message', async () => {
|
||||
const { ctx, agent } = await invokeHarness()
|
||||
const mixed = createUserMessage({
|
||||
content: [
|
||||
{ type: 'reasoning', text: '/hidden-demo inside a non-text block' },
|
||||
{ type: 'text', text: '/shared-skill go' },
|
||||
],
|
||||
source: { kind: 'user' },
|
||||
})
|
||||
const decision = await proposeStep(ctx, agent, [mixed])
|
||||
if (decision.kind !== 'enter') throw new Error('expected enter')
|
||||
const invoked = decision.messages
|
||||
.filter(message => (message.source as { kind?: string }).kind === 'skill-invocation')
|
||||
.map(message => (message.source as { name: string }).name)
|
||||
expect(invoked).toEqual(['shared-skill'])
|
||||
})
|
||||
})
|
||||
|
||||
Reference in New Issue
Block a user