fix(web): address the second review round on context provenance

- An empty replacement catalog is a real catalog: `renderCatalogUpdate()`
  publishes zero entries when the last skill disappears, and falling back
  would hide that every earlier name was retired.
- The opaque fallback keeps a `form` declaration this version cannot present.
  It is the one place a newer or foreign log's declared shape would otherwise
  vanish from the UI entirely, since the row marker is also absent there.
- An instruction change with an unrecognized `action` disqualifies the record.
  The action decides the word the row shows, so an unknown one would be
  presented as loaded or updated.
- The catalog list bounds itself and reports the withheld count. Entry count
  is unbounded and the scrollport bounds height, not node count.
- A catalog message keeps content blocks this version does not know, instead
  of dropping model-visible content the extensible union may carry.
- `core.md` defines `ContextFormed`, the interface actually carrying the
  optional field, beside `ContextForm`.
- The superseded-in-part bullet states the affected clauses as one rule rather
  than enumerating them; two rounds of enumeration each missed some, which is
  the shape being fragile rather than the list being wrong.
- The note records the one migration case that does not self-heal: an
  old-format catalog as the only one, with an empty current view, leaves a
  stale catalog nothing replaces.
This commit is contained in:
creatixchu
2026-08-05 16:51:01 +08:00
parent aeb718688e
commit ccd27f3775
16 changed files with 160 additions and 59 deletions

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-05-context-form-vocabulary.md
2026-08-05-context-form-vocabulary.md: 20577b0ad0eaa9eb8dc8528cedd18d15439ad299
2026-08-05-context-form-vocabulary.zh.md: 154e8ae54fc6b27ce782533b4246d48190639128
2026-08-05-context-form-vocabulary.md: c111acfec4fa85fe9e25adeb615371d65098f20f
2026-08-05-context-form-vocabulary.zh.md: d4da218c4c6026eb875513c2acbfcd18ced8faa1

View File

@@ -27,7 +27,7 @@ The vocabulary is semantic, never visual. A value states that the content is a f
Entries record the published fact **unescaped**. The pseudo-XML escaping belongs to the `<available_skills>` frame, which exists for the model, so it is applied when rendering that frame and never stored; otherwise a consumer would have to know the frame's encoding to display a description containing `<`, and the same frame knowledge this decision removes would leak back in another shape. `escapeText` is deterministic and injective, so digesting the unescaped entries preserves republish semantics exactly, and the model-facing text stays byte-identical.
That move also relocates catalog **identity**: the republish digest now covers the durable entries rather than the rendered text, so the model-facing framing can no longer decide whether a republish is needed, and the text-slicing that recovered entries from a logged message is gone. A resumed session whose newest catalog predates this change republishes once, which the pre-release stance permits.
That move also relocates catalog **identity**: the republish digest now covers the durable entries rather than the rendered text, so the model-facing framing can no longer decide whether a republish is needed, and the text-slicing that recovered entries from a logged message is gone. A resumed session whose newest catalog predates this change republishes once, which the pre-release stance permits. One case does not self-heal: if that old-format catalog is the only one and the current view has no skills, the plugin sees no published catalog and emits no tombstone, so the model keeps a stale catalog nothing replaces. The pre-release stance ("backends reject old on-disk formats") permits it; it is recorded here rather than left to the optimistic path.
Both readers are **all-or-nothing**: one unreadable entry disqualifies the record rather than being dropped, because a body that replaces the model-facing text must not present a confident but incomplete account of what the model read. The row's form marker reports what actually rendered, not what was declared.

View File

@@ -27,7 +27,7 @@ Status: implemented
条目记录的是**未转义**的发布事实。伪 XML 转义属于 `<available_skills>` 这层为模型而设的框架,因此只在渲染该框架时施加、从不存储;否则消费方要正确展示含 `<` 的描述就得知道框架的编码方式,本决策刚移除的框架知识会换一种形式泄漏回来。`escapeText` 确定且单射,故对未转义条目取 digest 与此前完全等价,重新发布语义不变,面向模型的文本逐字节不变。
这次迁移同时挪动了目录的**身份**:重新发布用的 digest 现在覆盖持久条目而非渲染文本,于是面向模型的包装再也无法左右是否需要重新发布,那段从已记录消息里切出条目的文本切分逻辑也随之删除。若恢复的会话中最新目录早于本次改动,会重新发布一次——发布前阶段的姿态允许这样做。
这次迁移同时挪动了目录的**身份**:重新发布用的 digest 现在覆盖持久条目而非渲染文本,于是面向模型的包装再也无法左右是否需要重新发布,那段从已记录消息里切出条目的文本切分逻辑也随之删除。若恢复的会话中最新目录早于本次改动,会重新发布一次——发布前阶段的姿态允许这样做。有一种情形不会自愈:当那份旧格式目录是唯一的一份、且当前视图没有任何 skill 时,插件看不到已发布目录,也就不会发出 tombstone模型手里会留着一份无人替换的过期目录。发布前阶段的姿态「后端拒绝旧的磁盘格式」允许这一点此处如实记录而不是只写乐观路径。
两个读取器都是**全有或全无**:一条不可读的条目即判定整条记录不可用,而不是把它丢掉——会替换掉面向模型文本的内容区,不得给出自信但残缺的「模型读到了什么」。行上的形态标记报告的是实际渲染出的形态,而非声明的形态。

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/simplification/2026-07-31-web-ui-no-steer-entry-or-interjection-chrome.md
2026-07-31-web-ui-no-steer-entry-or-interjection-chrome.md: 9dc87cd69f5d274ba5e8d66c30805cada7f1fa1a
2026-07-31-web-ui-no-steer-entry-or-interjection-chrome.zh.md: e650b660efe0bf308a4b1a8c72e9c89a133ac769
2026-07-31-web-ui-no-steer-entry-or-interjection-chrome.md: dfd79a1b1b44778275b4326cf5f51f6e69ce2a7c
2026-07-31-web-ui-no-steer-entry-or-interjection-chrome.zh.md: 9b470dda8de45ca754a04900ffac0535ebfabd96

View File

@@ -28,7 +28,7 @@ Keep host and runtime steering intact. Remove only the Web UI entry and chrome:
## Consequences
- **Superseded in part.** Composer steering shipped afterwards, so the first and third Decision bullets and the first consequence below no longer describe master: `InputMachine` carries the `queue`/`steer` mode union again, Web users can steer, and the steering bubble regained its user IconActions before carrying a caption again. The [context-source and steer marks decision](../feature/2026-08-04-web-context-source-and-steer-marks.md) then supplied the product decision this note's reintroduction clause required and restored the `插话` caption on steering bubbles. Host steering ownership, `ConversationService.send`'s queue-only contract, and the rejection of hiding `steering/message` from the transcript still hold.
- **Superseded in part.** Every clause below about the steer ENTRY, the `queue`/`steer` mode union, the interjection caption, its locale strings, and the goldens that pinned their absence no longer describes master — in the Decision that is bullets 1 and 3 through 5, and in the consequences bullets 1 and 3. Composer steering shipped afterwards, and the [context-source and steer marks decision](../feature/2026-08-04-web-context-source-and-steer-marks.md) then supplied the product decision this note's reintroduction clause required. What still holds: host steering ownership, `ConversationService.send`'s queue-only contract, and the rejection of hiding `steering/message` from the transcript.
- Web users cannot steer from the composer or `ctx.conversation.send`; stop/cancel and Queue remain the only mid-turn controls.
- Host-wire and non-Web clients can still steer; the Web client shows those messages without labeling them as interjections.
- Reintroducing a dedicated steer UI would need a new product decision; do not revive the mode union or badge without one.

View File

@@ -28,7 +28,7 @@ Status: implemented
## 后果
- **部分被取代。** composer steering 在此之后落地,因此 Decision 的第一条与第三条、以及下面的第一条后果都已不再描述 master`InputMachine` 重新携带 `queue``steer` mode 联合类型Web 用户可以 steersteering 气泡也早在重新带上标注之前就恢复了用户 IconActions。随后[上下文来源与 steer 标识决策](../feature/2026-08-04-web-context-source-and-steer-marks.md)提供了本 note 重新引入条款所要求的产品决策,并恢复了 steering 气泡上的 `插话` 标注。host 侧 steering 的归属、`ConversationService.send` 仅排队的契约,以及不在 transcript 中隐藏 `steering/message` 的取舍,仍然成立
- **部分被取代。** 下文中所有关于 steer **入口**`queue``steer` mode 联合类型、插话标注、其 locale 字符串、以及钉住这些「不存在」的黄金基线的子句,均已不再描述 master——在 Decision 中是第 1 条与第 3 至 5 条,在后果中是第 1 条与第 3 条。composer steering 在此之后落地,随后[上下文来源与 steer 标识决策](../feature/2026-08-04-web-context-source-and-steer-marks.md)提供了本 note 重新引入条款所要求的产品决策。仍然成立的是:host 侧 steering 的归属、`ConversationService.send` 仅排队的契约,以及不在 transcript 中隐藏 `steering/message` 的取舍。
- Web 用户无法从 composer 或 `ctx.conversation.send` steer中途控制只剩停止取消与 Queue。
- Host 线缆与非 Web 客户端仍可 steerWeb 客户端展示这些消息时不再标成插话。
- 若要重新引入专用 steer UI需要新的产品决策没有决策就不要复活 mode 联合类型或徽章。

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write docs/core-data-structures/core.md
core.md: d91e84a7066b004567ed9f668ded52ee486a1c99
core.zh.md: f9cf2c14fe1021d7862d8bb14696044b7d8f7fbb
core.md: 7761930b7f133600e89befd1bda96cfc7d2f05f5
core.zh.md: 1a68b47229e3c961e772e7b0db33ded91b025e2e

View File

@@ -193,6 +193,13 @@ type ContextForm =
| 'catalog'
```
```ts type-equiv
/** Optional producer-declared {@link ContextForm}, mixed into the source shapes that carry one. */
interface ContextFormed {
readonly form?: ContextForm
}
```
## Streaming
Adapters emit a raw **chunk** protocol; the loop logs the chunks (replay fidelity) while feeding the same chunks through a `BlockAssembler` to rebuild blocks and messages. `StreamChunk` is a closed discriminated union over `type` — `block-start`, `text-delta`, `reasoning-delta`, `tool-call-delta`, `block-end`, `usage`, `finish`.

View File

@@ -199,6 +199,13 @@ type ContextForm =
| 'catalog'
```
```ts type-equiv
/** Optional producer-declared {@link ContextForm}, mixed into the source shapes that carry one. */
interface ContextFormed {
readonly form?: ContextForm
}
```
## 流式输出
适配器发出原始**分片**协议;循环记录分片(回放保真度),同时将同一批分片送入 `BlockAssembler` 以重建块和消息。`StreamChunk` 是基于 `type` 的封闭判别联合——`block-start`、`text-delta`、`reasoning-delta`、`tool-call-delta`、`block-end`、`usage`、`finish`。

View File

@@ -13,6 +13,9 @@ import css from './ContextBody.module.css'
/** Model-facing text stays bounded at the disclosure, not at the producer. */
const MAX_CHARS = 20_000
/** Rows a list body materializes before summarizing the remainder. */
const MAX_ENTRIES = 200
type Translate = ChatViewSlotProps['t']
/** One durable source narrowed to the readable-record shape; null for anything else. */
@@ -61,14 +64,22 @@ function fieldValue(value: unknown, t: Translate): string {
}
/**
* Provenance fields as a key/value list. `kind` is omitted because the row
* header already names the producer, and `form` because the presentation the
* reader is looking at IS that value.
* Provenance fields as a key/value list. `kind` is always omitted because the
* row header already names the producer. `form` is omitted only when a
* dedicated body rendered for it — then the presentation the reader is looking
* at IS that value. On the opaque fallback the declaration is kept, because
* that is the one place a form this version cannot present would otherwise
* disappear from the UI entirely.
*/
function SourceFields({ source, t }: { source: unknown; t: Translate }): ReactNode {
function SourceFields({ source, formRendered, t }: {
source: unknown
formRendered: boolean
t: Translate
}): ReactNode {
const record = asRecord(source)
if (record === null) return null
const rows = Object.entries(record).filter(([key]) => key !== 'kind' && key !== 'form')
const hidden = formRendered ? ['kind', 'form'] : ['kind']
const rows = Object.entries(record).filter(([key]) => !hidden.includes(key))
if (rows.length === 0) return null
return (
<dl className={css.fields} data-context-fields>
@@ -82,6 +93,28 @@ function SourceFields({ source, t }: { source: unknown; t: Translate }): ReactNo
)
}
/**
* Content blocks this UI version does not know, kept visible rather than
* dropped: the block union is merge-extensible, so a newer or foreign log may
* carry a shape this build has no presentation for.
* @param props - The unrecognized blocks and the locale seat.
* @returns One generic JSON block per unknown entry.
*/
function UnknownBlocks({ blocks, t }: { blocks: readonly unknown[]; t: Translate }): ReactNode {
return (
<>
{blocks.map((block, index) => (
<JsonBlock
key={index}
label={t('message.unknownBlock')}
payload={block}
truncatedLabel={total => t('json.truncated', { total })}
/>
))}
</>
)
}
/**
* The model-facing content of one context, shared by every form that shows it:
* the text with its real line breaks, then any block this UI version does not
@@ -97,14 +130,7 @@ function ModelFacingContent({ content, t }: {
return (
<>
{text !== '' && <pre className={css.text} data-context-text>{boundedText(text, t)}</pre>}
{rest.map((block, index) => (
<JsonBlock
key={index}
label={t('message.unknownBlock')}
payload={block}
truncatedLabel={total => t('json.truncated', { total })}
/>
))}
<UnknownBlocks blocks={rest} t={t} />
</>
)
}
@@ -124,14 +150,14 @@ export function OpaqueBody({ content, source, t }: {
return (
<>
<ModelFacingContent content={content} t={t} />
<SourceFields source={source} t={t} />
<SourceFields source={source} formRendered={false} t={t} />
</>
)
}
/** One reconciled instruction file, as the durable source records it. */
interface InstructionChange {
action: string
action: 'set' | 'replace' | 'remove'
path: string
digest?: string
}
@@ -157,14 +183,13 @@ function instructionChanges(source: unknown): InstructionChange[] | null {
const path = change['path']
if (typeof path !== 'string' || path === '') return null
const action = change['action']
// The action decides which word the row shows, so an unrecognized one is
// not a readable change — it would be presented as loaded or updated.
if (action !== 'set' && action !== 'replace' && action !== 'remove') return null
const digest = change['digest']
if (seen.has(path)) continue
seen.add(path)
changes.push({
action: typeof action === 'string' ? action : '',
path,
...typeof digest === 'string' ? { digest } : {},
})
changes.push({ action, path, ...typeof digest === 'string' ? { digest } : {} })
}
return changes.length === 0 ? null : changes
}
@@ -228,7 +253,9 @@ function catalogEntries(source: unknown): CatalogEntry[] | null {
if (typeof name !== 'string' || name === '' || typeof description !== 'string') return null
entries.push({ name, description })
}
return entries.length === 0 ? null : entries
// An empty list is a real catalog: a replacement with no entries retires
// every earlier name. Only an unreadable shape falls back.
return entries
}
/**
@@ -249,11 +276,15 @@ export function CatalogBody({ content, source, t }: {
const entries = catalogEntries(source)
if (entries === null) return <OpaqueBody content={content} source={source} t={t} />
const update = asRecord(source)?.['update'] === true
// Entry count is unbounded (a provider may publish any number of skills), and
// the scrollport bounds height, not node count — so the list bounds itself.
const shown = entries.slice(0, MAX_ENTRIES)
const { rest } = partitionContent(content)
return (
<>
{update && <p className={css.catalogNotice} data-context-catalog-update>{t('message.context.catalog.replaced')}</p>}
<ul className={css.entries} data-context-entries>
{entries.map((entry, index) => (
{shown.map((entry, index) => (
// Index key: a hand-edited or foreign log may repeat a name, and a
// duplicate React key would drop a row the model did see.
<li key={index} className={css.entry}>
@@ -262,6 +293,14 @@ export function CatalogBody({ content, source, t }: {
</li>
))}
</ul>
{shown.length < entries.length && (
<p className={css.catalogNotice} data-context-entries-truncated>
{t('message.context.catalog.more', { count: entries.length - shown.length })}
</p>
)}
{/* The block union is merge-extensible: a catalog message carrying an
unknown block still shows it rather than dropping model-visible content. */}
<UnknownBlocks blocks={rest} t={t} />
</>
)
}

View File

@@ -59,6 +59,7 @@ export const zh = {
'message.context.instructions.updated': '已更新',
'message.context.instructions.removed': '已移除',
'message.context.catalog.replaced': '替换目录',
'message.context.catalog.more': '…还有 {count} 条',
'message.steering': '插话',
'message.compaction': '上下文已压缩',
'message.compaction.expand': '点击查看压缩摘要',
@@ -178,6 +179,7 @@ export const en = {
'message.context.instructions.updated': 'updated',
'message.context.instructions.removed': 'removed',
'message.context.catalog.replaced': 'Replacement catalog',
'message.context.catalog.more': '… {count} more',
'message.steering': 'Interjection',
'message.compaction': 'Context compacted',
'message.compaction.expand': 'View compaction summary',

View File

@@ -411,13 +411,34 @@ describe('MessageItem arms', () => {
.toMatch(/… 已截断,共 \d+ 字符$/)
})
it('a catalog whose source carries no entries falls back to the opaque body', () => {
it('an empty replacement catalog stays a catalog: it retires every earlier name', () => {
// `renderCatalogUpdate` legitimately publishes zero entries when the last
// skill disappears; falling back would hide that the catalog was cleared.
const view = render(
<MessageItem t={t} node={{
kind: 'context',
seq: 3,
content: [{ type: 'text', text: 'catalog prose' }],
source: { kind: 'skill-catalog', form: 'catalog', entries: [] },
source: { kind: 'skill-catalog', form: 'catalog', update: true, entries: [] },
provenance: { role: 'inject', label: 'skill-catalog' },
form: 'catalog',
} as never}
/>,
)
fireEvent.click(view.getByRole('button', { name: /^上下文注入\s*skill-catalog$/ }))
expect(view.container.querySelector('[data-context-catalog-update]')?.textContent).toBe('替换目录')
expect(view.container.querySelectorAll('[data-context-entries] li')).toHaveLength(0)
expect(view.container.querySelector('[data-context-injection-body]')?.getAttribute('data-context-form'))
.toBe('catalog')
})
it('a catalog whose entries are unreadable falls back to the opaque body', () => {
const view = render(
<MessageItem t={t} node={{
kind: 'context',
seq: 3,
content: [{ type: 'text', text: 'catalog prose' }],
source: { kind: 'skill-catalog', form: 'catalog', entries: 'not-a-list' },
provenance: { role: 'inject', label: 'skill-catalog' },
form: 'catalog',
} as never}
@@ -428,51 +449,71 @@ describe('MessageItem arms', () => {
expect(view.container.querySelector('[data-context-text]')?.textContent).toBe('catalog prose')
})
it('a recalled session titles its row by role and names the sessions it read', () => {
it('bounds a large catalog and says how many rows it withheld', () => {
const entries = Array.from({ length: 205 }, (_, index) => ({ name: `s-${index}`, description: 'd' }))
const view = render(
<MessageItem t={t} node={{
kind: 'context', seq: 3, content: [{ type: 'text', text: 'catalog prose' }],
source: { kind: 'skill-catalog', form: 'catalog', entries },
provenance: { role: 'inject', label: 'skill-catalog' },
form: 'catalog',
} as never}
/>,
)
fireEvent.click(view.getByRole('button', { name: /^上下文注入\s*skill-catalog$/ }))
expect(view.container.querySelectorAll('[data-context-entries] li')).toHaveLength(200)
expect(view.container.querySelector('[data-context-entries-truncated]')?.textContent).toBe('…还有 5 条')
})
it('a catalog keeps a content block this version does not know', () => {
const view = render(
<MessageItem t={t} node={{
kind: 'context',
seq: 3,
content: [{ type: 'text', text: 'snapshot' }],
source: { kind: 'session-reference', version: 1, references: [{ label: '重构 loader' }] },
provenance: { role: 'recall', label: '重构 loader' },
form: null,
content: [{ type: 'text', text: 'prose' }, { type: 'future-block', payload: 1 }],
source: { kind: 'skill-catalog', form: 'catalog', entries: [{ name: 'a', description: 'b' }] },
provenance: { role: 'inject', label: 'skill-catalog' },
form: 'catalog',
} as never}
/>,
)
expect(view.getByRole('button', { name: /^跨会话召回\s*重构 loader$/ })).toBeTruthy()
expect(view.queryByText('上下文注入')).toBeNull()
fireEvent.click(view.getByRole('button', { name: /^上下文注入\s*skill-catalog$/ }))
expect(view.getByText(/未知内容块/)).toBeTruthy()
})
it('keeps the producer name visible while the context body is expanded', () => {
it('an instruction change with an unrecognized action falls back whole', () => {
// The action decides the word the row shows, so an unknown one cannot be
// presented as loaded or updated.
const view = render(
<MessageItem t={t} node={{
kind: 'context',
seq: 3,
content: [{ type: 'text', text: 'instructions' }],
source: { kind: 'workspace-instructions', changes: [{ path: 'AGENTS.md' }] },
provenance: { role: 'inject', label: 'AGENTS.md' },
form: null,
content: [{ type: 'text', text: 'instruction prose' }],
source: { kind: 'workspace-instructions', form: 'instructions', changes: [{ action: 'merge', path: 'A.md' }] },
provenance: { role: 'inject', label: 'workspace-instructions' },
form: 'instructions',
} as never}
/>,
)
const disclosure = view.getByRole('button', { name: /^上下文注入\s*AGENTS\.md$/ })
fireEvent.click(disclosure)
expect(disclosure.getAttribute('aria-expanded')).toBe('true')
expect(view.container.querySelector('[data-context-source]')?.textContent).toBe('AGENTS.md')
fireEvent.click(view.getByRole('button', { name: /^上下文注入\s*workspace-instructions$/ }))
expect(view.container.querySelector('[data-context-files]')).toBeNull()
expect(view.container.querySelector('[data-context-text]')?.textContent).toBe('instruction prose')
})
it('a context source that names no producer shows the role alone', () => {
it('the opaque fallback keeps a form declaration this version cannot present', () => {
// Otherwise a newer or foreign log's declared shape vanishes from the UI.
const view = render(
<MessageItem t={t} node={{
kind: 'context', seq: 3, content: [{ type: 'text', text: 'x' }], source: null,
provenance: { role: 'inject', label: null },
kind: 'context', seq: 3, content: [{ type: 'text', text: 'x' }],
source: { kind: 'plugin', plugin: 'later', form: 'a-later-form' },
provenance: { role: 'inject', label: 'later' },
form: null,
} as never}
/>,
)
expect(view.getByRole('button', { name: '上下文注入' })).toBeTruthy()
expect(view.container.querySelector('[data-context-source]')).toBeNull()
fireEvent.click(view.getByRole('button', { name: /^上下文注入\s*later$/ }))
const fields = [...view.container.querySelectorAll('[data-context-fields] dt')].map(node => node.textContent)
expect(fields).toEqual(['plugin', 'form'])
})
it('unknown nodes retain the generic JSON row', () => {

View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write packages/skill/tool-skill/README.md
README.md: deccda0ece1ffe2cbdb576a8af3801f28695d8d5
README.zh.md: eaf442d9e123a2d56bf09abd9faab47a001e01b9
README.md: c8c8b0d18665893104bc8ac65c1a90cae7816724
README.zh.md: 696e18955bf6f622251a0ca875f53e66419ec944

View File

@@ -10,7 +10,7 @@ Requires `ctx.agents`, `ctx.tools`, and `ctx.skills` (`inject: ['agents', 'tools
At every `agent/step`, the plugin calls `ctx.skills.snapshot()` for the calling session's cwd, forwards the step abort signal to discovery, applies exact `skill` tool visibility, and renders the ordered `name` and `description` entries. When no prior catalog exists and that view is non-empty, it injects an initial durable user-role `<system-reminder>` before the request. Catalog messages contain only those summaries; skill bodies, paths, sources, providers, and `whenToUse` hints remain outside the catalog.
Every catalog message carries the `skill-catalog` source: a `catalog`-form context whose `entries` record exactly the `name` and `description` pairs it published, plus `update` on a replacement. The digest covers those durable entries, not the rendered prose, so the surrounding `<system-reminder>` framing — written for the model — cannot decide whether a republish is needed, and a consumer presenting the list never re-parses the `<available_skills>` block. The plugin scans durable session events backwards without copying them and derives the comparison baseline from the newest visible `skill-catalog` message. When the digest changes, `agent.inject()` records a durable user-role message containing the complete replacement catalog; an empty replacement explicitly retires earlier names. If no catalog remains visible but a recognizable historical catalog exists, compaction hid it and the next complete observation re-establishes the current catalog. An incomplete provider snapshot emits nothing and preserves the last-good model view for retry on the next step. If no prior catalog exists and the current view is empty, no tombstone is necessary.
Every catalog message carries the `skill-catalog` source: a `catalog`-form context whose `entries` record exactly the `name` and `description` pairs it published, plus `update` on a replacement. The digest covers those durable entries, not the rendered prose, so the surrounding `<system-reminder>` framing — written for the model — cannot decide whether a republish is needed, and a consumer presenting the list never re-parses the `<available_skills>` block. The plugin scans durable session events backwards without copying them and derives the comparison baseline from the newest visible `skill-catalog` message it can read; an unreadable record is skipped like any foreign one. When the digest changes, `agent.inject()` records a durable user-role message containing the complete replacement catalog; an empty replacement explicitly retires earlier names. If no catalog remains visible but a recognizable historical catalog exists, compaction hid it and the next complete observation re-establishes the current catalog. An incomplete provider snapshot emits nothing and preserves the last-good model view for retry on the next step. If no prior catalog exists and the current view is empty, no tombstone is necessary.
The catalog is omitted when no model-invocable skills are initially available, and also when that agent's tool view restricts away the shipped `skill` tool or resolves a same-name scoped shadow instead. Visibility changes participate in the digest, keeping prompt guidance, model-visible schema, and executable dispatch aligned.

View File

@@ -10,7 +10,7 @@
每次 `agent/step`,该插件都会使用调用会话的 cwd 调用 `ctx.skills.snapshot()`,将步骤中止信号转发到发现流程,应用 `skill` 工具的精确可见性,并按顺序渲染 `name``description` 条目。如果先前不存在目录且该视图非空,插件会在请求之前注入初始的持久用户角色 `<system-reminder>`。目录消息只包含这些摘要skill 正文、路径、来源、提供方和 `whenToUse` 提示仍位于目录之外。
每条目录消息都携带 `skill-catalog` 来源:一份 `catalog` 形态的上下文,其 `entries` 精确记录本次发布的 `name``description` 对,替换目录另带 `update`。digest 覆盖的是这些持久条目而非渲染出的散文,因此为模型书写的 `<system-reminder>` 包装无法左右是否需要重新发布,消费方展示该列表时也不必再解析 `<available_skills>` 块。插件从后向前扫描持久会话事件且不复制,并以最新一条仍可见的 `skill-catalog` 消息作为比较基线。digest 变化时,`agent.inject()` 会记录一条包含完整替换目录的持久用户角色消息空替换会显式停用较早的名称。如果没有目录仍然可见但历史中存在可识别目录则说明压缩compaction已将其遮蔽下一次完整观察会重新建立当前目录。提供方快照不完整时插件不会发送任何内容并会保留最后一次完整的模型视图以便在下一步骤重试。若不存在先前目录且当前视图为空则不需要 tombstone。
每条目录消息都携带 `skill-catalog` 来源:一份 `catalog` 形态的上下文,其 `entries` 精确记录本次发布的 `name``description` 对,替换目录另带 `update`。digest 覆盖的是这些持久条目而非渲染出的散文,因此为模型书写的 `<system-reminder>` 包装无法左右是否需要重新发布,消费方展示该列表时也不必再解析 `<available_skills>` 块。插件从后向前扫描持久会话事件且不复制,并以最新一条仍可见且可读`skill-catalog` 消息作为比较基线;不可读的记录与外来记录一样被跳过。digest 变化时,`agent.inject()` 会记录一条包含完整替换目录的持久用户角色消息空替换会显式停用较早的名称。如果没有目录仍然可见但历史中存在可识别目录则说明压缩compaction已将其遮蔽下一次完整观察会重新建立当前目录。提供方快照不完整时插件不会发送任何内容并会保留最后一次完整的模型视图以便在下一步骤重试。若不存在先前目录且当前视图为空则不需要 tombstone。
如果最初没有模型可调用 skill则省略目录如果该 agent智能体的工具视图排除了随附的 `skill` 工具,或解析出同名的作用域内遮蔽项,也会省略目录。可见性变更参与 digest 计算,使提示词指引、模型可见 schema 和可执行分派保持对齐。

View File

@@ -31,6 +31,11 @@
"symbol": "ContextForm",
"source": "packages/llm/llm/src/message.ts"
},
{
"doc": "docs/core-data-structures/core.md",
"symbol": "ContextFormed",
"source": "packages/llm/llm/src/message.ts"
},
{
"doc": "docs/core-data-structures/core.md",
"symbol": "FinishReasonMap",