feat(web): split context injection into producer-declared forms

Injected context reached the transcript as one anonymous shape whose
expanded body was the whole message serialized as inline JSON, so the
escaping collapsed the only readable part — the model-facing prose —
into a single line.

`MessageSource` gains an optional producer-declared `form`, a small
semantic vocabulary of information shapes independent of `kind`: kind
says who produced the context, form says what shape it is, so several
producers may share one presentation. Two values ship.

`instructions` (workspace-context) lists the reconciled files above
their text and keeps the `<system-reminder>` framing verbatim, because
the framing is part of what the model read. `catalog` moves dsh-tool-skill
off the shared plugin kind onto a `skill-catalog` source carrying the
published name/description entries, and the body lists those instead of
re-parsing `<available_skills>` out of the prose. Catalog identity moves
with it: the republish digest now covers the durable entries, deleting
the text-slicing that recovered them from a logged message.

Everything else renders the opaque body — the model-facing text with its
real line breaks, then the remaining provenance as fields. That is the
documented default, not a leftover: a resumed, forked, or foreign log
must render whether or not its producer is mounted here, which is why
the classification lives in the durable source rather than a client-side
table keyed by producer.
This commit is contained in:
creatixchu
2026-08-05 12:15:03 +08:00
parent e5d4ae72fc
commit 340c9ba576
51 changed files with 779 additions and 174 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 packages/client/runtime/README.md
README.md: 17b3e73b7c2efa4aa71ff24c9e5a45eaf81771de
README.zh.md: 5fc5981ea3a1f896dddb008271c323e759505dad
README.md: 0d9935571bf84a3ce299a2dd37a81894412ab536
README.zh.md: 6b6c129ee24ee6a82a21e50924b170c4b9725445

View File

@@ -28,7 +28,7 @@ SlotsService gives the renderer separate bare observables for `useSessions` and
## The human transcript
`ConversationSnapshot.nodes` is the human transcript, not the model surface. `TranscriptAdapter` projects the raw window in log order — every append-origin surface event (`isAppendSurfaceEvent`) at its own log position, plus one `CompactionSummaryNode` marker per landed compaction checkpoint — and never consults surface order. `ConversationSnapshot.turnEnds` maps each completed turn in that window to its `turn/end` seq, retaining turn completion independently from the transcript so presentation can require a real boundary before enabling an action. A landed compaction therefore keeps the conversation it shadowed on the model side: the marker reports where the model stopped seeing that history instead of erasing it. Model-only replacement copies stay out: a pruned `tool/result` and a regenerated `assistant/message` rewrite one node for the model and mark no boundary. A checkpoint is a `user/message` carrying the compaction seam's plugin source that **replaced** a surface range; an appending plugin-sourced `user/message` is injected context, not a compaction. Each context node also carries a `provenance` view: `contextProvenance()` reads the durable source alone to decide whether the row is an `inject` or a cross-session `recall`, and to name its producer from the instruction paths, referenced session titles, or plugin id that source already records. The client holds no table of plugin ids, so a renamed or newly mounted producer stays identifiable without a client release and a resumed or foreign log projects exactly like a live one; a source with no readable kind degrades to an unnamed injection. The adapter's plugin literal is pinned to the seam's own declaration by a type-only import of the cordis-free [`dsh-compact/checkpoint`](../../compact/compact/README.md) leaf, so renaming it there fails `tsc` here; a **value** import of the package would fail the client purity gate, and the package **root** is unreachable even as a type (it reaches `dsh-session`'s root, whose `Context` merge collides the host `sessions` with this program's). `tests/compact-checkpoint-pin.spec.ts` covers the same drift behaviorally.
`ConversationSnapshot.nodes` is the human transcript, not the model surface. `TranscriptAdapter` projects the raw window in log order — every append-origin surface event (`isAppendSurfaceEvent`) at its own log position, plus one `CompactionSummaryNode` marker per landed compaction checkpoint — and never consults surface order. `ConversationSnapshot.turnEnds` maps each completed turn in that window to its `turn/end` seq, retaining turn completion independently from the transcript so presentation can require a real boundary before enabling an action. A landed compaction therefore keeps the conversation it shadowed on the model side: the marker reports where the model stopped seeing that history instead of erasing it. Model-only replacement copies stay out: a pruned `tool/result` and a regenerated `assistant/message` rewrite one node for the model and mark no boundary. A checkpoint is a `user/message` carrying the compaction seam's plugin source that **replaced** a surface range; an appending plugin-sourced `user/message` is injected context, not a compaction. Each context node also carries a `provenance` view: `contextProvenance()` reads the durable source alone to decide whether the row is an `inject` or a cross-session `recall`, and to name its producer from the instruction paths, referenced session titles, or plugin id that source already records. The client holds no table of plugin ids, so a renamed or newly mounted producer stays identifiable without a client release and a resumed or foreign log projects exactly like a live one; a source with no readable kind degrades to an unnamed injection. Beside it, `contextForm()` reads the producer-declared `ContextForm` — the second, independent axis: `kind` says who produced the context, `form` says what shape of information it is, so several producers may share one form. A form this UI version does not present projects as null and renders opaque. The adapter's plugin literal is pinned to the seam's own declaration by a type-only import of the cordis-free [`dsh-compact/checkpoint`](../../compact/compact/README.md) leaf, so renaming it there fails `tsc` here; a **value** import of the package would fail the client purity gate, and the package **root** is unreachable even as a type (it reaches `dsh-session`'s root, whose `Context` merge collides the host `sessions` with this program's). `tests/compact-checkpoint-pin.spec.ts` covers the same drift behaviorally.
Because the projection is log-ordered, the node array is seq-monotonic by construction: log-only `command/run` / `command/done` nodes splice in by seq, `Session` merges interrupted frozen nodes by their fractional seqs, and a window whose checkpoint cites a shadowed range outside it renders the marker with nothing logged. The marker's summary text comes from the checkpoint's `compact/summary` provenance; a window cut that left the provenance outside makes the row non-expandable rather than empty, and a later page that supplies it resolves the text. Performance contract: one append materializes at most one node and copies the projection only when it adds that node; an event that changes no node keeps the previous array reference (a chunk storm costs nothing), and unchanged nodes keep their object identity.

View File

@@ -28,7 +28,7 @@ SlotsService 分别为 renderer 提供 `useSessions` 与 `useWorkspaces` 的裸
## 面向人的 transcript(文本记录)
`ConversationSnapshot.nodes` 是面向人的 transcript,不是模型 surface。`TranscriptAdapter` 按日志顺序投影原始窗口——每个 append 来源的 surface 事件(`isAppendSurfaceEvent`)落在它自己的日志位置上,外加每次落地的压缩(compaction)检查点贡献一个 `CompactionSummaryNode` 标记——且从不查询 surface 顺序。`ConversationSnapshot.turnEnds` 把该窗口中的每个已完成轮次映射到其 `turn/end` seq;它独立于 transcript 保留轮次完成状态,使呈现层能够在启用操作前要求存在真实边界。于是一次落地的压缩会保留它在模型侧遮蔽掉的对话:标记报告模型从哪里开始看不见那段历史,而不是把它抹掉。仅模型可见的 replacement 副本不进入记录:被裁剪的 `tool/result` 和重新生成的 `assistant/message` 只为模型重写一个节点,不标记任何边界。检查点是携带压缩 seam 插件来源、且**替换**了一段 surface 范围的 `user/message`;一条 append 的插件来源 `user/message` 是注入上下文,不是压缩。每个上下文节点还携带一份 `provenance` 视图:`contextProvenance()` 只读取持久来源,据此判定该行是 `inject`(注入)还是跨会话的 `recall`(召回),并用该来源已经记录的指令文件路径、被引用会话标题或插件 id 命名其生产者。客户端不保存任何插件 id 表,因此重命名或新挂载的生产者无需客户端发版即可保持可辨识,恢复的会话日志与外部日志的投影结果和实时会话完全一致;没有可读 kind 的来源则降级为无名注入。适配器的插件字面量通过对无 cordis 的 [`dsh-compact/checkpoint`](../../compact/compact/README.md) 叶子做仅类型导入,钉在压缩 seam 自己的声明上:在那里改名会让此处 `tsc` 失败;而对该包(package)做**值**导入会被客户端纯度门禁拒绝,包的**根**即便作为类型也无法到达(它会到达 `dsh-session` 的根,其 `Context` 合并会让 host 的 `sessions` 与本程序的冲突)。`tests/compact-checkpoint-pin.spec.ts` 从行为侧覆盖同一漂移。
`ConversationSnapshot.nodes` 是面向人的 transcript,不是模型 surface。`TranscriptAdapter` 按日志顺序投影原始窗口——每个 append 来源的 surface 事件(`isAppendSurfaceEvent`)落在它自己的日志位置上,外加每次落地的压缩(compaction)检查点贡献一个 `CompactionSummaryNode` 标记——且从不查询 surface 顺序。`ConversationSnapshot.turnEnds` 把该窗口中的每个已完成轮次映射到其 `turn/end` seq;它独立于 transcript 保留轮次完成状态,使呈现层能够在启用操作前要求存在真实边界。于是一次落地的压缩会保留它在模型侧遮蔽掉的对话:标记报告模型从哪里开始看不见那段历史,而不是把它抹掉。仅模型可见的 replacement 副本不进入记录:被裁剪的 `tool/result` 和重新生成的 `assistant/message` 只为模型重写一个节点,不标记任何边界。检查点是携带压缩 seam 插件来源、且**替换**了一段 surface 范围的 `user/message`;一条 append 的插件来源 `user/message` 是注入上下文,不是压缩。每个上下文节点还携带一份 `provenance` 视图:`contextProvenance()` 只读取持久来源,据此判定该行是 `inject`(注入)还是跨会话的 `recall`(召回),并用该来源已经记录的指令文件路径、被引用会话标题或插件 id 命名其生产者。客户端不保存任何插件 id 表,因此重命名或新挂载的生产者无需客户端发版即可保持可辨识,恢复的会话日志与外部日志的投影结果和实时会话完全一致;没有可读 kind 的来源则降级为无名注入。与之并列的 `contextForm()` 读取生产方声明的 `ContextForm`,这是相互独立的第二根轴:`kind` 说明上下文由谁产生,`form` 说明它是何种形态的信息,因此多个生产方可以共用一种形态。本 UI 版本不呈现的形态投影为 null,按 opaque 渲染。适配器的插件字面量通过对无 cordis 的 [`dsh-compact/checkpoint`](../../compact/compact/README.md) 叶子做仅类型导入,钉在压缩 seam 自己的声明上:在那里改名会让此处 `tsc` 失败;而对该包(package)做**值**导入会被客户端纯度门禁拒绝,包的**根**即便作为类型也无法到达(它会到达 `dsh-session` 的根,其 `Context` 合并会让 host 的 `sessions` 与本程序的冲突)。`tests/compact-checkpoint-pin.spec.ts` 从行为侧覆盖同一漂移。
由于投影按日志顺序,节点数组天然按 seq 单调:仅日志的 `command/run` / `command/done` 节点按 seq 插入,`Session` 按分数 seq 归并被打断的冻结节点,而检查点所引范围落在窗口之外的窗口会渲染出标记且不打印任何日志。标记的摘要文本来自检查点的 `compact/summary` 溯源;窗口切分把溯源留在窗口外时该行不可展开而非空白,后续补上溯源的分页会解析出文本。性能契约:一次追加最多物化一个节点,并且仅在加入该节点时复制投影;不改变任何节点的事件保持上一次的数组引用(分片风暴零成本),未变化的节点保持其对象标识。

View File

@@ -53,7 +53,9 @@ export type {
export type {
ConversationContext, ConversationContextOriginKind,
} from './sessions/conversation-context.ts'
export type { ContextProvenanceView, ContextRole } from './sessions/context-provenance.ts'
export type {
ContextProvenanceView, ContextRole, KnownContextForm,
} from './sessions/context-provenance.ts'
export type {
ConversationPromptSnapshot, RequestInspectionSnapshot, RequestPromptChange, RequestView,
} from './sessions/request-inspection.ts'

View File

@@ -11,7 +11,7 @@ import type {
PartialAssistant, RunningToolCall,
} from '../sessions/conversation.ts'
import { toAssistantBlocks } from '../sessions/conversation.ts'
import { contextProvenance } from '../sessions/context-provenance.ts'
import { contextForm, contextProvenance } from '../sessions/context-provenance.ts'
import type {
ConversationContext, ConversationContextOriginKind,
} from '../sessions/conversation-context.ts'
@@ -128,6 +128,7 @@ function materializeNode(
kind: 'context', seq: event.seq, time: event.time,
content: event.data.content, source: event.data.source,
provenance: contextProvenance(event.data.source),
form: contextForm(event.data.source),
}
}
return {

View File

@@ -89,3 +89,28 @@ export function contextProvenance(source: unknown): ContextProvenanceView {
return { role: 'inject', label: kind }
}
}
/**
* Context forms this UI version renders with a dedicated presentation. The
* durable vocabulary (`ContextForm` in `dsh-llm`) may already be wider — an
* unrecognized or absent value degrades to the opaque presentation rather than
* dropping the row, so a log written by a newer or foreign producer still
* renders.
*/
const KNOWN_FORMS = ['instructions', 'catalog'] as const
/** One durable context form this UI version knows how to present. */
export type KnownContextForm = typeof KNOWN_FORMS[number]
/**
* Read the producer-declared form off one durable message source.
* @param source - the logged `user/message` source, exactly as recorded.
* @returns the form when this UI version presents it, otherwise null (opaque).
*/
export function contextForm(source: unknown): KnownContextForm | null {
const record = asRecord(source)
const form = record === null ? null : readString(record, 'form')
return form !== null && (KNOWN_FORMS as readonly string[]).includes(form)
? form as KnownContextForm
: null
}

View File

@@ -12,7 +12,7 @@ import type {
InboxItemId, RpcError, SessionId, SubagentAddress, ToolCallView, ToolResultView,
} from '@deepseek-ai/dsh-client-connection/client'
import type { PendingInteraction } from './pending.ts'
import type { ContextProvenanceView } from './context-provenance.ts'
import type { ContextProvenanceView, KnownContextForm } from './context-provenance.ts'
export type { TodoItem }
/** Request configuration recorded for one provider call. */
@@ -126,6 +126,8 @@ export interface ContextMessageNode {
source: unknown
/** Role and producer name projected from `source` ({@link contextProvenance}). */
provenance: ContextProvenanceView
/** Producer-declared information form ({@link contextForm}); null presents as opaque. */
form: KnownContextForm | null
}
/** Durable notice that a closed failed step is waiting for a model-request retry. */

View File

@@ -22,7 +22,7 @@ import type { COMPACT_CHECKPOINT_SOURCE } from '@deepseek-ai/dsh-compact/checkpo
import type { ToolCallView, ToolEventView, ToolResultView } from '@deepseek-ai/dsh-client-connection/client'
import type { CommandNode, CompactionSummaryNode, ConversationNode } from './conversation.ts'
import { toAssistantBlocks } from './conversation.ts'
import { contextProvenance } from './context-provenance.ts'
import { contextForm, contextProvenance } from './context-provenance.ts'
/**
* The compaction seam's checkpoint plugin, pinned to the seam's own declaration
@@ -62,6 +62,7 @@ function materializeNode(
kind: 'context', seq: event.seq, time: event.time,
content: event.data.content, source: event.data.source,
provenance: contextProvenance(event.data.source),
form: contextForm(event.data.source),
}
}
return {