@deepseek-ai/dsh-llm 0.1.2-rc.1 → 0.1.5-alpha.1

package/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/llm/llm/README.md
-README.md: 6baa020a9503f32a71231d30e78ec7a3d8862f2b
-README.zh.md: f985736da44a76b063f16915a53cb19979fe1962
+README.md: c2854cc29f98aa860b1e4064d0c12ced3114f61c
+README.zh.md: c4a5979e7c8003e897f53971a62abbdd97eca801

package/README.md

@@ -9,7 +9,7 @@ English | [中文](README.zh.md)
 
 ## Summary
 
-`@deepseek-ai/dsh-llm` is the provider-neutral model-call service at the center of the harness's LLM capability. Every composition that streams a request to a model provider goes through it, and it owns the shared vocabulary — messages, content blocks, and raw stream chunks — that the agent loop, session log, and every plugin speak. With it you can register provider adapters, stream one model call, list and discover models, resolve exact-model metadata and call defaults, and capture each provider's retry policy; every request is logged so it stays reconstructable from the session log. It executes no retries and owns no provider wire logic: adapters translate their provider's format, and the optional `dsh-llm-retry` package re-runs failed requests at durable step boundaries. Requests are deep-frozen before dispatch, so middleware and adapters can read them but never rewrite them.
+Use `@deepseek-ai/dsh-llm` to stream model calls through configured provider adapters, discover models, and resolve model capabilities and call defaults. Every dispatched request remains reconstructable from the session log. Requests are deep-frozen before dispatch, so extensions and adapters can read them but cannot rewrite them. Each stream is one provider attempt: provider-specific translation stays with its adapter, while the optional `@deepseek-ai/dsh-llm-retry` package re-runs failed requests. Streams always end with a terminal result, so callers can handle success, failure, and cancellation consistently.
 
 ## Table of Contents
 
@@ -42,7 +42,7 @@ Mount the service and at least one adapter, then select the provider by name in
     apiKeyEnv: DEEPSEEK_API_KEY
 ```
 
-A stream returns token-level chunks and always ends with one terminal `finish` chunk; `BlockAssembler` turns the chunks into content blocks and messages, and the loop logs each chunk for replay:
+A stream returns token-level chunks and always ends with one terminal `finish` chunk. `BlockAssembler` turns the chunks into content blocks and messages; `AssistantStreamAccumulator` preserves their exact timestamps and token boundaries in a compact representation that the loop embeds in one durable attempt settlement:
 
 ```text
 for await (const chunk of ctx.llm.stream({
@@ -61,8 +61,9 @@ After a successful mount, `ctx.llm.listProviders()` reports the registered route
 - **Stream one model call** — `ctx.llm.stream(options)` yields raw chunks (token-level deltas) for any registered provider and model; consumers assemble them with `BlockAssembler`.
 - **Register provider adapters** — an adapter owns one or more provider routes, and its registration captures that route's retry policy; registering the same route twice fails with `DUPLICATE_ADAPTER`.
 - **Expose and activate providers through configuration** — adapters declare configurable-provider routes plus a settings namespace, so configuration surfaces can activate dormant providers and edit connection facts without a restart.
-- **Discover and resolve models** — list the models an adapter advertises, interrogate an endpoint for the models it serves, and resolve one exact model's context window, output default, reasoning efforts, and input modalities.
+- **Discover and resolve models** — list the models an adapter advertises, interrogate an endpoint for the models it serves, and resolve one exact model's context window, output default, reasoning efforts, input modalities, and system prompt update mode: `LlmResolvedModelInfo.systemPromptUpdate` is `'in-history'` when the model reads the latest `system` message at any position as the effective system prompt and absent when only a leading system message is read; `normalizeModelInfo` rejects any other value with `INVALID_MODEL_INFO`.
 - **Validate call config** — an explicit or configured reasoning effort is checked against the exact model before any provider I/O, and an adapter-configured output cap is materialized when the request omits one.
+- **Read an embedded Assistant stream without expanding it** — `assistantStreamFirstTokenTime` (first token), `assistantStreamHasVisibleContent` (any visible content), and `assistantStreamHasVisibleText` (any visible text) answer their questions from the compact records with early exit; `lastAssistantStreamChunk` scans backward to the last raw chunk of one type, `assistantStreamChunks` and `joinAssistantStreamText` scan the whole stream, and `assembleAssistantStream` feeds a `BlockAssembler` one joined delta per run with the same blocks, usage, and replay state as the per-member expansion. `runFirstTokenTime` and `runFirstVisibleTime` do the early-exit scan for one packed run, and `isTokenDelta`, `isVisibleChunk`, and `chunkHasVisibleText` define the token and visibility rules for a single chunk. `expandAssistantStream` remains the validating path for records read at a durable boundary; it is not memoized, because a retained expansion costs roughly ten times the compact stream for as long as the event lives.
 
 ### Failures and recovery
 
@@ -90,16 +91,19 @@ The service is built on one separation: **the logical contract is provider-neutr
 | [`src/types.ts`](src/types.ts) | The `StreamChunk` protocol, content-block map, finish reasons, and shared vocabulary |
 | [`src/message.ts`](src/message.ts) | Immutable message constructors shared by delivery, history, and requests |
 | [`src/assembler.ts`](src/assembler.ts) | `BlockAssembler`: incremental chunk-to-block assembly |
+| [`src/assistant-stream.ts`](src/assistant-stream.ts) | Compact timed Assistant stream accumulation, strict validation, exact expansion, and record-level readers |
 | [`src/call-config.ts`](src/call-config.ts) | Call-config validation, adapter-default materialization, and request freezing |
 | [`src/retry-policy.ts`](src/retry-policy.ts) | Provider-owned retry policy resolution (normal and always modes) |
 | [`src/error.ts`](src/error.ts) | `HarnessError`/`LlmError` taxonomy and provider-neutral failure codes |
-| [`src/content.ts`](src/content.ts) | Shared image-content helpers, including request-image offloading |
+| [`src/content.ts`](src/content.ts) | Shared file and image projection helpers, including request-image offloading |
 | [`src/api-key.ts`](src/api-key.ts) | Credential format check shared by every adapter |
 | [`src/adapter-failure.ts`](src/adapter-failure.ts) | Failure normalization into terminal finish chunks |
 
 ### Main flow
 
-A request is validated against its exact model's capability — context window, output default, reasoning efforts, and input modalities — and any adapter-configured defaults are materialized, then the whole request is deep-frozen. `prepareCall()` binds those facts, detached context, and retry policy to the exact adapter generation that performs terminal dispatch, so HMR or dynamic settings cannot combine one generation's image capability with another generation's endpoint. An image-capable adapter projects durable references into route-specific request versions; `resolveImageAttachmentAccess()` separately maps an attachment provider's optional host object into the current tool execution world without changing the request image or its `variantId`. A text-only route receives deterministic per-image placeholders, including nested tool-result images, without rewriting append-only session history. `offloadRequestImagesWithPolicy()` removes oldest images deterministically by raw or base64 size and count or byte quanta; the pure `offloadedImagePrefixCount()` exposes that decision so route-owned request pricing can reproduce it without building the projection. Adapters that charge visual tokens declare per-route `imageRequestPricing`, which `ctx.llm.imageRequestPricing(provider, model)` resolves synchronously for the token meter. Dispatch goes through the `llm/stream` waterfall, then chunks return as token-level deltas and every adapter outcome reaches the consumer as one terminal `finish` chunk.
+A request is validated against its exact model's capability — context window, output default, reasoning efforts, input modalities, and `systemPromptUpdate` mode — and any adapter-configured defaults are materialized, then the whole request is deep-frozen. `prepareCall()` binds those facts, detached context, and retry policy to the exact adapter generation that performs terminal dispatch, so HMR or dynamic settings cannot combine one generation's image capability with another generation's endpoint. An image-capable adapter projects durable references into route-specific request versions; `resolveImageAttachmentAccess()` separately maps an attachment provider's optional host object into the current tool execution world without changing the request image or its `variantId`. A text-only route receives deterministic per-image placeholders, including nested tool-result images, without rewriting append-only session history. Durable `FileBlock` references never reach any adapter: request assembly replaces each one, nested tool-result occurrences included, with deterministic handle text naming the file and its saved read-only path, resolved through the mounted attachment and filesystem providers. `ctx.llm.fileRequestText(ref)` exposes that exact synchronous projection to request measurement. `offloadRequestImagesWithPolicy()` removes oldest images deterministically by raw or base64 size and count or byte quanta; the pure `offloadedImagePrefixCount()` exposes that decision so route-owned request pricing can reproduce it without building the projection. Adapters that charge visual tokens declare per-route `imageRequestPricing`, which `ctx.llm.imageRequestPricing(provider, model)` resolves synchronously for the token meter. Dispatch goes through the `llm/stream` waterfall, then chunks return as token-level deltas and every adapter outcome reaches the consumer as one terminal `finish` chunk.
+
+File detection reads current content, including nested tool results, on every request without caching message identities or freeze state. The [file-scan decision](../../../.agents/notes/implemented/simplification/2026-09-07-file-content-scan.md) records the measured traversal cost.
 
 ### Invariants
 
@@ -107,6 +111,7 @@ A request is validated against its exact model's capability — context window,
 - **Replay state travels only within one adapter** — assistant replay state rides along only when the same adapter instance owns the historical and target routes; otherwise it is dropped before dispatch.
 - **Prepared calls are one-shot** — a prepared call can be dispatched exactly once, and its call-config fields must match the prepared config.
 - **Image projection follows the captured route** — durable `ImageBlock` references become route-specific request versions only for image-capable models; text-only models receive stable placeholders.
+- **File projection is unconditional** — no provider receives file bytes; every route gets one deterministic handle line per `FileBlock`, and the model reads the saved copy with its file tools on demand.
 - **Protocol ordering** — `usage` precedes `finish`, tool arguments stay raw JSON strings, and nothing follows the terminal `finish`.
 - **Registry mutations are atomic** — route and directory registration validates the whole candidate set before anything moves, so a refused change leaves the previous state serving.
 
@@ -119,7 +124,7 @@ A request is validated against its exact model's capability — context window,
 
 Read these pages when the package-level contract is not enough. They move from the shared types to the concrete adapters, the retry executor, and the measurement service.
 
-- [LLM streaming subsystem](../../../docs/subsystems/llm-streaming.md) — the message and block types, the assembled model request, the `StreamChunk` protocol, and the adapter contract.
+- [LLM streaming subsystem](../../../docs/subsystems/llm-streaming.md) — the message and block types, compact Assistant stream records, the `StreamChunk` protocol, and the adapter contract.
 - [llm-deepseek adapter](../llm-deepseek/README.md) — the direct DeepSeek chat-completions implementation.
 - [llm-pi-ai adapter](../llm-pi-ai/README.md) — the pi-ai-backed multi-provider implementation.
 - [llm-retry](../llm-retry/README.md) — the retry executor that re-runs failed model requests.

package/README.zh.md

@@ -9,7 +9,7 @@ kind: "package-reference"
 
 ## 概述
 
-`@deepseek-ai/dsh-llm` 是位于 harness LLM 能力核心的提供方无关模型调用服务。任何向模型提供方发起流式请求的组合都会经过它，它拥有 agent loop（智能体循环）、会话日志和所有插件共同使用的共享词汇——消息、内容块与原始流式分片。借助它，你可以注册提供方适配器、流式发起一次模型调用、列出与发现模型、解析精确模型元数据与调用默认值，并捕获每个提供方的重试策略；每个请求都会被记录，因此始终可以从会话日志重建。它不执行重试，也不拥有任何提供方协议逻辑：适配器翻译各自提供方的格式，可选包 `dsh-llm-retry` 在持久步骤边界上重跑失败的请求。请求在分发前会被深度冻结，因此 middleware 与适配器只能读取，绝不能改写。
+使用 `@deepseek-ai/dsh-llm` 可通过已配置的提供方适配器流式调用模型、发现模型，并解析模型能力与调用默认值。每个已分发请求都可以从会话日志重建。请求在分发前会被深度冻结，因此扩展与适配器可以读取但不能改写。每个流只尝试调用提供方一次：提供方特定的转换由对应适配器完成，可选包 `@deepseek-ai/dsh-llm-retry` 负责重跑失败的请求。流始终以终止结果结束，因此调用方可以一致地处理成功、失败与取消。
 
 ## 目录
 
@@ -42,7 +42,7 @@ kind: "package-reference"
     apiKeyEnv: DEEPSEEK_API_KEY
 ```
 
-流会返回 token 级分片，并始终以一个终止 `finish` 分片结束；`BlockAssembler` 把分片组装为内容块与消息，loop 记录每个分片以供回放：
+流会返回 token 级分片，并始终以一个终止 `finish` 分片结束。`BlockAssembler` 把分片组装为内容块与消息；`AssistantStreamAccumulator` 在紧凑表示中保留其精确时间戳与 token 边界，loop 再把它嵌入一个持久 attempt settlement：
 
 ```text
 for await (const chunk of ctx.llm.stream({
@@ -61,8 +61,9 @@ for await (const chunk of ctx.llm.stream({
 - **流式发起一次模型调用**——`ctx.llm.stream(options)` 为任何已注册提供方与模型产出原始分片（token 级增量）；消费方用 `BlockAssembler` 组装。
 - **注册提供方适配器**——一个适配器拥有一个或多个提供方路由，其注册会捕获该路由的重试策略；重复注册同一路由会以 `DUPLICATE_ADAPTER` 失败。
 - **通过配置暴露并激活提供方**——适配器声明可配置提供方路由与 settings namespace，配置界面因此可以激活休眠提供方并编辑连接事实，无需重启。
-- **发现与解析模型**——列出适配器公布的模型、询问端点它提供哪些模型，并解析某个精确模型的上下文窗口、输出默认值、推理（reasoning）强度与输入模态。
+- **发现与解析模型**——列出适配器公布的模型、询问端点它提供哪些模型，并解析某个精确模型的上下文窗口、输出默认值、推理（reasoning）强度、输入模态与系统提示词更新模式：当模型把任意位置最新的 `system` 消息读作有效系统提示词时，`LlmResolvedModelInfo.systemPromptUpdate` 为 `'in-history'`；只读取开头 system 消息时该字段缺失；`normalizeModelInfo` 以 `INVALID_MODEL_INFO` 拒绝任何其他值。
 - **校验调用配置**——显式或配置的推理强度会在任何提供方 I/O 之前对照精确模型校验；请求省略输出上限时，会填入适配器配置的输出上限。
+- **不展开即读取内嵌 Assistant 流**——`assistantStreamFirstTokenTime`（首 token）、`assistantStreamHasVisibleContent`（任一可见内容）与 `assistantStreamHasVisibleText`（任一可见文本）以提前退出从紧凑记录回答各自的问题；`lastAssistantStreamChunk` 反向扫描到某一类型的最后一个原始 chunk，`assistantStreamChunks` 与 `joinAssistantStreamText` 扫描整个流，`assembleAssistantStream` 向 `BlockAssembler` 每个 run 喂一段拼接 delta，blocks／usage／replayState 与逐成员展开相同。`runFirstTokenTime` 与 `runFirstVisibleTime` 对单个打包 run 做提前退出扫描，`isTokenDelta`、`isVisibleChunk` 与 `chunkHasVisibleText` 定义单个 chunk 的 token 与可见性规则。`expandAssistantStream` 仍是持久边界读取记录的校验路径；它不被记忆化，因为保留的展开在事件生命周期内约花费紧凑流的十倍内存。
 
 ### 失败与恢复
 
@@ -90,16 +91,19 @@ for await (const chunk of ctx.llm.stream({
 | [`src/types.ts`](src/types.ts) | `StreamChunk` 协议、内容块映射、结束原因与共享词汇 |
 | [`src/message.ts`](src/message.ts) | 投递、历史与请求共享的不可变消息构造函数 |
 | [`src/assembler.ts`](src/assembler.ts) | `BlockAssembler`：分片到块的增量组装 |
+| [`src/assistant-stream.ts`](src/assistant-stream.ts) | 紧凑带时间 Assistant stream 的累积、严格校验、精确展开与记录级读取器 |
 | [`src/call-config.ts`](src/call-config.ts) | 调用配置校验、适配器默认值填入与请求冻结 |
 | [`src/retry-policy.ts`](src/retry-policy.ts) | 提供方自有重试策略解析（normal 与 always 模式） |
 | [`src/error.ts`](src/error.ts) | `HarnessError`/`LlmError` 分类体系与提供方无关失败 code |
-| [`src/content.ts`](src/content.ts) | 共享图片内容辅助函数，包括请求图片卸载 |
+| [`src/content.ts`](src/content.ts) | 共享文件与图片投影辅助函数，包括请求图片卸载 |
 | [`src/api-key.ts`](src/api-key.ts) | 每个适配器共享的凭据格式校验 |
 | [`src/adapter-failure.ts`](src/adapter-failure.ts) | 把失败归一化为终止 finish 分片 |
 
 ### 主流程
 
-请求会对照其精确模型的能力——上下文窗口、输出默认值、推理强度与输入模态——校验，填入任何适配器配置的默认值，然后整个请求被深度冻结。`prepareCall()` 把这些事实、分离的上下文与重试策略绑定到执行最终分发的精确适配器代次，因此 HMR 或动态设置无法把一个代次的图片能力与另一代次的端点混用。支持图片的适配器把持久引用投影为路由专用请求版本；`resolveImageAttachmentAccess()` 会单独把附件提供方的可选宿主对象映射进当前工具执行世界，而不改变请求图片或其 `variantId`。纯文本路由接收确定性的逐图片占位符，包括嵌套工具结果图片，而不会改写仅追加会话历史。`offloadRequestImagesWithPolicy()` 按原始字节或 base64 大小以及图片数或字节步长，确定性地从最旧图片开始移除；纯函数 `offloadedImagePrefixCount()` 公开同一决策，使路由所属的请求定价无需构建投影即可复现它。对视觉 token 收费的适配器声明按路由的 `imageRequestPricing`，`ctx.llm.imageRequestPricing(provider, model)` 为 token meter 同步解析它。分发经过 `llm/stream` waterfall，随后分片以 token 级增量返回，每个适配器结果都以唯一一个终止 `finish` 分片到达消费方。
+请求会对照其精确模型的能力——上下文窗口、输出默认值、推理强度、输入模态与 `systemPromptUpdate` 模式——校验，填入任何适配器配置的默认值，然后整个请求被深度冻结。`prepareCall()` 把这些事实、分离的上下文与重试策略绑定到执行最终分发的精确适配器代次，因此 HMR 或动态设置无法把一个代次的图片能力与另一代次的端点混用。支持图片的适配器把持久引用投影为路由专用请求版本；`resolveImageAttachmentAccess()` 会单独把附件提供方的可选宿主对象映射进当前工具执行世界，而不改变请求图片或其 `variantId`。纯文本路由接收确定性的逐图片占位符，包括嵌套工具结果图片，而不会改写仅追加会话历史。持久 `FileBlock` 引用永远不会到达任何适配器：请求组装把每个引用（包括嵌套工具结果中的出现）替换为确定性 handle 文本，指出文件与其只读保存路径，路径经由挂载的附件与文件系统提供方解析。`ctx.llm.fileRequestText(ref)` 向请求计量公开相同的同步投影。`offloadRequestImagesWithPolicy()` 按原始字节或 base64 大小以及图片数或字节步长，确定性地从最旧图片开始移除；纯函数 `offloadedImagePrefixCount()` 公开同一决策，使路由所属的请求定价无需构建投影即可复现它。对视觉 token 收费的适配器声明按路由的 `imageRequestPricing`，`ctx.llm.imageRequestPricing(provider, model)` 为 token meter 同步解析它。分发经过 `llm/stream` waterfall，随后分片以 token 级增量返回，每个适配器结果都以唯一一个终止 `finish` 分片到达消费方。
+
+文件检测在每次请求时读取当前内容，包括嵌套工具结果，不缓存消息身份或冻结状态。[文件扫描决策](../../../.agents/notes/implemented/simplification/2026-09-07-file-content-scan.zh.md)记录了实测遍历成本。
 
 ### 不变式
 
@@ -107,6 +111,7 @@ for await (const chunk of ctx.llm.stream({
 - **回放状态只在同一适配器内流动**——仅当同一适配器实例同时拥有历史路由与目标路由时，assistant 回放状态才会随行；否则在分发前被丢弃。
 - **已准备调用是一次性的**——已准备调用只能分发一次，且其调用配置字段必须与准备好的配置一致。
 - **图片投影遵循捕获的路由**——只有支持图片的模型会把持久 `ImageBlock` 引用转换为路由专用请求版本；纯文本模型接收稳定占位符。
+- **文件投影无条件进行**——没有任何提供方会收到文件字节；每条路由对每个 `FileBlock` 都得到一行确定性 handle 文本，模型在需要时用文件工具读取保存的副本。
 - **协议顺序**——`usage` 先于 `finish`，工具参数保持原始 JSON 字符串，终止 `finish` 之后不再有任何内容。
 - **注册表变更具有原子性**——路由与目录注册会在任何变动前整体校验候选集合，因此被拒绝的变更会让此前状态继续服务。
 
@@ -119,7 +124,7 @@ for await (const chunk of ctx.llm.stream({
 
 当包级约定不够用时阅读以下页面。它们从共享类型逐步进入具体适配器、重试执行器与计量服务。
 
-- [LLM 流式子系统](../../../docs/subsystems/llm-streaming.zh.md)——消息与块类型、组装后的模型请求、`StreamChunk` 协议与适配器约定。
+- [LLM 流式子系统](../../../docs/subsystems/llm-streaming.zh.md)——消息与块类型、紧凑 Assistant stream record、`StreamChunk` 协议与适配器约定。
 - [llm-deepseek 适配器](../llm-deepseek/README.zh.md)——DeepSeek chat-completions 直连实现。
 - [llm-pi-ai 适配器](../llm-pi-ai/README.zh.md)——基于 pi-ai 的多提供方实现。
 - [llm-retry](../llm-retry/README.zh.md)——重跑失败模型请求的重试执行器。