Merge remote-tracking branch 'origin/master' into xtr/message-tool-call-id

2026-08-29 04:26:38 +00:00 · 2026-08-25 20:24:54 +08:00
parent 804db36546 e2a10b141e
commit f436c888aa
611 changed files with 17576 additions and 5986 deletions
@@ -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/architecture/2026-06-14-session-persistence.md
-2026-06-14-session-persistence.md: cef11271f26c304ad484d7851801bc69d0c1dfda
-2026-06-14-session-persistence.zh.md: b6c2467888d0d348aa492c265542565563b75fab
+2026-06-14-session-persistence.md: 62228bd2f5b25b13880a563818d08f3a2d52d956
+2026-06-14-session-persistence.zh.md: ebf004333c383336cd025aa8a4aabc9d1e07f0e5
@@ -29,7 +29,7 @@ Key durable, contested choices:

 Each key choice above records its rejected alternative where the choice is stated: a **chunk-filtered canonical log** (Codex's `policy.rs` shape) — breaks the contiguous-seq contract; **truncating a crashed turn** — silently destroys a long autonomous run's real work; an **in-log `session/meta` event as line 0** — metadata is not replayable state; **finite fractional `createdAt` values** — have no producer and diverge from integer Unix-millisecond storage and query columns; **adopting a non-pristine unversioned SQLite file** — can overwrite unrelated objects or identity; **hard-injecting `sessionPersistence` into the loop** — would pend non-persistent demos forever.

-Format versioning: the header carries a `version`; cold reads accept the current version or a complete static adjacent-version decoder path and reject future versions or missing steps. The format decoder owns historical header and event conversion, while the Coordinator owns operation-specific recovery after decoding ([Session log versioning](2026-08-10-session-log-version-mechanism.md)). The pre-release session format stays pinned at `SESSION_FORMAT_VERSION = 0` and carries no broad compatibility promise. Append-only + flush is robust to partial trailing writes (tolerated during cold preparation) but not to fsync-less power loss mid-line; a DB/WAL backend is the stronger option there.
+Format versioning: the header carries a `version`; cold reads reject any non-current version. The pre-release session format stays pinned at `SESSION_FORMAT_VERSION = 0` and carries no broad compatibility promise, while the coordinator may own an explicit narrow import upgrade when persisted user data requires it ([pre-identity message recovery](../bug-fix/2026-07-28-load-pre-identity-session-messages.md)). Append-only + flush is robust to partial trailing writes (tolerated during cold preparation) but not to fsync-less power loss mid-line; a DB/WAL backend is the stronger option there.

 ## Consequences

@@ -29,7 +29,7 @@ Status: implemented

 上述每个关键选择都在陈述处记录了被否决的替代方案：**过滤分片的规范日志**（Codex 的 `policy.rs` 形式）破坏连续 seq 约定；**截断崩溃的轮次**会静默销毁长时间自主运行中的真实工作；**日志内 `session/meta` 事件作为第 0 行**——元数据不是可回放状态；**有限的非整数 `createdAt` 值**没有生产方，且与整数 Unix 毫秒存储及查询列不一致；**接受非全新的未版本化 SQLite 文件**可能覆盖无关对象或应用标识；**将 `sessionPersistence` 硬注入循环**会让非持久化的演示永远挂起。

-格式版本控制：header 携带一个 `version`；冷读取接受当前版本或完整的静态相邻版本 decoder 路径，并拒绝未来版本或缺失步骤。Format decoder 负责历史 header 和 event 转换，Coordinator 只在解码后负责各操作自己的 recovery（[Session log 版本机制](2026-08-10-session-log-version-mechanism.zh.md)）。预发布阶段的会话格式仍固定为 `SESSION_FORMAT_VERSION = 0`，不承诺广泛兼容。仅追加 + 刷写对尾部的不完整写入具有健壮性（冷准备时可容忍），但无法抵御未使用 fsync 时在行写入中途断电；数据库/WAL 后端是该场景下更强的选项。
+格式版本控制：header 携带一个 `version`；冷读取拒绝任何非当前版本。预发布阶段的会话格式仍固定为 `SESSION_FORMAT_VERSION = 0`，不承诺广泛兼容；当持久化用户数据确有需要时，协调器可以负责显式且范围受限的导入升级（[消息标识机制引入前的消息恢复](../bug-fix/2026-07-28-load-pre-identity-session-messages.zh.md)）。仅追加 + 刷写对尾部的不完整写入具有健壮性（冷准备时可容忍），但无法抵御未使用 fsync 时在行写入中途断电；数据库/WAL 后端是该场景下更强的选项。

 ## 后果

@@ -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/architecture/2026-06-24-web-capability-seam.md
-2026-06-24-web-capability-seam.md: 7e7b09f19864bd2ad8ad9d69579c1d5c79600cde
-2026-06-24-web-capability-seam.zh.md: dbb41ee42d2c7503955ead2df32abe80b3a4f641
+2026-06-24-web-capability-seam.md: 8c6c088ea5d7345f9955892b2d6054cfae518bbd
+2026-06-24-web-capability-seam.zh.md: 4921c4a4647d180dbb00e6493a19d38584f99bdc
@@ -199,7 +199,7 @@ Full page retrieval remains the job of `web_fetch(url)`. Search snippets are dis

 ## Fetch request and result schema

-The `web_fetch` implementation is an anonymous public HTTP(S) fetch provider, `http`. It fetches bytes from a concrete URL, applies the basic transport hygiene below (http/https-only, credential rejection, byte/time caps, cross-origin redirect blocking), decodes textual content, and returns only the minimal model-useful result: final URL, status code, body, and truncation. It carries no browser cookies, editor credentials, git credentials, internal auth tokens, or implicit access to private services. (Full SSRF / private-network blocking is deferred — see [Deferred work](#deferred-work).)
+The `web_fetch` implementation is an anonymous public HTTP(S) fetch provider, `http`. It fetches bytes from a concrete URL, resolves and pins public destinations, applies the transport hygiene below, decodes textual content, and returns only the minimal model-useful result: final URL, status code, body, and truncation. It carries no browser cookies, editor credentials, git credentials, internal auth tokens, or implicit access to private services.

 The seam request stays smaller than OpenCode's model-facing tool:

@@ -235,12 +235,14 @@ The provider owns safe resource retrieval: URL validation, HTTP transport, redir
 The fetch provider's resource controls:

 - Only `http:` and `https:` URLs are accepted; credentials in URLs are rejected.
+- A literal address or the complete result of one hostname lookup must contain only globally reachable unicast IPv4 or IPv6 destinations. IPv6 resolution also discovers the active DNS64 prefix and rejects NAT64 addresses that translate to non-public IPv4. Loopback, private, link-local, carrier-grade NAT, multicast, reserved, transition, translation, and private IPv4-mapped IPv6 addresses are rejected.
+- The request retains that validated address set in an Undici lookup callback instead of resolving the hostname again. The original hostname remains the HTTP Host and TLS SNI value, while DNS rebinding cannot replace the connection destination after validation.
 - Maximum URL length, response byte cap, decoded body character cap, timeout, and redirect hop cap are enforced.
 - Abort signals propagate through network fetches and expensive decoding.
- Only same-origin redirects are followed automatically; a cross-origin redirect fails with `WEB_REDIRECT_BLOCKED`, requiring a fresh tool call and therefore a fresh provider/permission decision. (Claude Code's WebFetch uses this same model — it does not auto-follow a cross-host redirect; it returns the redirect target to the model for a fresh call.)
+- Only same-origin redirects are followed automatically; each followed hop performs a fresh public-address lookup and pins its own connection. A cross-origin redirect fails with `WEB_REDIRECT_BLOCKED`, requiring a fresh tool call and fresh public-address validation. (Claude Code's WebFetch uses this same model — it does not auto-follow a cross-host redirect; it returns the redirect target to the model for a fresh call.)
 - Requests carry an explicit product user agent rather than silently impersonating a browser.

-SSRF / private-network protection (blocking private, loopback, link-local, multicast, and otherwise non-public destinations, with DNS-resolve-then-validate to defeat rebinding and per-hop re-validation on redirects) is **deferred** — see [Deferred work](#deferred-work). Until it lands, `web_fetch` is an SSRF primitive and must not be enabled in a deployment that can reach sensitive internal network targets.
+The provider rejects an entire DNS answer set when any address is not public instead of silently filtering the unsafe members. This fail-closed rule prevents connection-family selection or fallback from reaching an address that did not satisfy the public-network policy.

 ## Tool consumer behavior

@@ -252,7 +254,7 @@ Tool registration is a minimal stable sync: on plugin startup the `dsh-tool-web`

 Provider availability changes affect execution results and diagnostics, not whether the model-facing schema exists. If a product wants no web tools at all, it disables `dsh-tool-web` or the individual web tool in config; if it wants web tools but the backend is misconfigured, the model sees a structured tool error at execution time.

-The prompt guidance explains the semantic split — `web_search` for discovery and current information, `web_fetch` when the model needs the content of a specific URL — and the prompt and tool result tell the model to cite relevant URLs with markdown links.
+The prompt guidance explains the semantic split — `web_search` for discovery and current information, `web_fetch` when the model needs the content of a specific URL — and the prompt and tool result tell the model to cite relevant URLs with markdown links. Every successful result labels provider-controlled text as external untrusted data. Fetch conversion removes active and hidden HTML content; unsafe conversion returns a fixed omission marker rather than raw HTML.

 The model-facing output is text-first because tool results are `ContentBlock[]`, but the seam outcome stays structured so UI presentation and future adapters do not have to scrape rendered text.

@@ -308,6 +310,18 @@ Rejected for the first version. Those providers often return extracted or summar

 Rejected for the seam. `prompt` turns fetch into LLM summarization and couples public-web retrieval to a model provider. The harness seam should fetch and decode deterministically; `dsh-tool-web` can later offer summaries as a presentation mode without making `ctx.web` depend on `ctx.llm`.

+### Validate DNS and then call an ordinary fetch
+
+Rejected because an ordinary fetch resolves the hostname again when it opens the connection. An attacker can return a public address during validation and a private address during the second lookup. Passing the validated answer set through the connection's lookup callback closes that rebinding interval while preserving hostname-based HTTP and TLS behavior.
+
+### Block private-looking hostname strings without pinning resolved addresses
+
+Rejected because hostname syntax does not establish the connection destination: an arbitrary public-looking name can resolve to loopback, a private range, or a cloud metadata address. Address classification belongs after resolution, and every address available to connection fallback must pass it.
+
+### Require per-call approval before public fetches
+
+Rejected for the shipped presets. Public-address validation blocks SSRF destinations, while per-call confirmation would interrupt ordinary browsing without controlling public data egress reliably: a model can reach the same public network through mounted shell tools. Deployments that require a dedicated confirmation step can add a `tools/pre-execute` policy or disable `web_fetch`.
+
 ## Consequences

 **The search schema is deliberately thin.** Exa and Perplexity both expose useful provider-specific controls; a control is added only once it can be defined provider-neutrally and enforced honestly by both tool registration and provider execution.
@@ -318,19 +332,16 @@ Rejected for the seam. `prompt` turns fetch into LLM summarization and couples p

 **Provider state can change after startup.** A tool can be visible in the request assembled at step start and lose its provider before execution. The execution path resolves again and fails with a structured error.

-**Fetch is a network boundary, not just a read-only tool.** `web_fetch` can reach sensitive network targets or exfiltrate data through URLs. Only the basic transport hygiene ships (http/https-only, credential rejection, byte/time caps, cross-origin redirect blocking); SSRF / private-network blocking is deferred (see [Deferred work](#deferred-work)), so until it lands `web_fetch` must not be enabled where it can reach internal targets.
+**Fetch is a network boundary, not just a read-only tool.** Public-address validation and connection pinning prevent `web_fetch` from reaching non-public destinations, but a model can still disclose data through a public URL and fetched text remains untrusted model input. The shipped `cordis`, `code`, and `standard` presets expose `web_fetch` in every sandbox and approval mode without per-call confirmation.

 **Large web content can damage context quality.** Providers enforce byte/character caps and report `truncated`; `tool-web` formats bounded model output with clear continuation or follow-up guidance.

 ## Deferred work

- SSRF / private-network protection for `web_fetch`: block private, loopback, link-local, multicast, and otherwise non-public destinations so `web_fetch` is not an SSRF primitive. Doing it correctly is more than a URL-string check — it needs DNS-resolve-then-connect-to-the-validated-IP (to defeat DNS rebinding / TOCTOU), per-hop re-validation across redirects, and IPv6 edge handling (private ranges, IPv4-mapped addresses). Neither reference implementation surveyed does IP-level blocking (OpenCode does a prefix check then fetches; Claude Code relies on a centralized hostname blocklist plus a "private URLs will fail" prompt), so there is no implementation to copy and this is the harness's only SSRF defense — it warrants its own focused design/spike. Until it lands, `web_fetch` must only be enabled in deployments that cannot reach sensitive internal targets.
 - A `pdf` `WebFetchBody` kind: the `http` provider decodes text-extractable PDFs (best-effort, capped, `truncated`) into a `{ kind: 'pdf'; content; pageCount? }` arm, and `tool-web` renders it. This is fetch, not `web_extract` — PDF retrieval is a concrete HTTP 200 plus deterministic local decoding, not provider-side extraction of a non-HTTP resource. Adding it is a coordinated change across `dsh-web` (declare the arm), the provider (decode + narrow "binary rejection" to "reject binary except text-extractable PDF"; scanned/image PDFs needing OCR stay out of scope), and `tool-web` (render). The closed `WebFetchBody` union makes the consumer side fail to compile until the new arm is handled.
 - Provider-backed extraction as a separate `web_extract` capability, rather than widening `web_fetch` silently.
- Permission policy integration: the permission system now exists ([sandbox and approval](../feature/2026-07-06-sandbox.md), [web permission presets](../feature/2026-07-23-web-permission-and-approval.md)) but bundles only sandbox mode and approval policy; web permission policy remains unintegrated.
 - Provider-neutral search controls beyond `query` and `maxResults`, once Exa and Perplexity can both honor them honestly.

 ## Open questions

 - Should product app packages probe web configuration at startup (treating `WEB_PROVIDER_CONFIGURED_MISSING`, `WEB_PROVIDER_CONFIGURED_UNAVAILABLE`, and `WEB_PROVIDER_AMBIGUOUS` as fatal when web is explicitly configured), or leave misconfiguration to surface at the first execution?
- Where should permission policy for public web access live in the shipped permission system ([sandbox and approval](../feature/2026-07-06-sandbox.md), [web permission presets](../feature/2026-07-23-web-permission-and-approval.md)): a dedicated web permission plugin on `tools/execute`, provider config, or both?
@@ -199,7 +199,7 @@ Exa 搜索将提供方扁平 `results[]` 的每一项映射为 `WebSearchSource`

 ## Fetch 请求与结果 schema

-`web_fetch` 的实现是一个匿名公开 HTTP(S) fetch 提供方 `http`。它从具体 URL 获取字节，应用下述基本传输卫生措施（仅 http/https、拒绝 URL 中的凭证、字节/时间上限、跨源重定向阻断），解码文本内容，并仅返回最小的模型可用结果：最终 URL、状态码、正文和截断标志。它不携带浏览器 cookie、编辑器凭证、git 凭证、内部认证令牌，也不隐式访问私有服务。（完整的 SSRF/私有网络阻断推迟——见[推迟工作](#deferred-work)。）
+`web_fetch` 的实现是一个匿名公开 HTTP(S) fetch 提供方 `http`。它从具体 URL 获取字节，解析并固定公开目的地址，应用下述传输卫生措施，解码文本内容，并仅返回最小的模型可用结果：最终 URL、状态码、正文和截断标志。它不携带浏览器 cookie、编辑器凭证、git 凭证、内部认证令牌，也不隐式访问私有服务。

 seam 请求比 OpenCode 的面向模型工具更小：

@@ -235,12 +235,14 @@ export type WebFetchBody =
 fetch 提供方的资源控制：

 - 仅接受 `http:` 和 `https:` URL；拒绝 URL 中的凭证。
+- 字面 IP 地址或 hostname 一次解析得到的完整结果只能包含全球可达的单播 IPv4 或 IPv6 目的地址。IPv6 解析还会发现当前 DNS64 前缀，并拒绝转换到非公开 IPv4 的 NAT64 地址。loopback、私有、link-local、运营商级 NAT、多播、保留、过渡、转换和映射到私有 IPv4 的 IPv6 地址都会被拒绝。
+- 请求通过 Undici lookup 回调保留这一组已验证地址，不会再次解析 hostname。原 hostname 仍作为 HTTP Host 与 TLS SNI 值，而 DNS rebinding 无法在验证后替换连接目的地址。
 - 强制执行最大 URL 长度、响应字节上限、解码正文字符上限、超时和重定向跳数上限。
 - Abort 信号传播到网络获取和高开销解码。
- 仅自动跟随同源重定向；跨源重定向以 `WEB_REDIRECT_BLOCKED` 失败，要求一次新的工具调用，从而触发新的提供方/权限决策。（Claude Code 的 WebFetch 使用同样的模型——它不自动跟随跨主机重定向，而是将重定向目标返回给模型以发起新调用。）
+- 仅自动跟随同源重定向；每个跟随的跳转都会重新解析公开地址，并把自己的连接固定到解析结果。跨源重定向以 `WEB_REDIRECT_BLOCKED` 失败，要求一次新的工具调用和新的公开地址校验。（Claude Code 的 WebFetch 使用同样的模型——它不自动跟随跨主机重定向，而是将重定向目标返回给模型以发起新调用。）
 - 请求携带显式的产品 User-Agent，而非静默伪装浏览器。

-SSRF/私有网络防护（阻断私有、回环、链路本地、多播及其他非公开目的地，通过先 DNS 解析再验证 IP 来防御 rebinding，并在重定向的每一跳重新验证）**推迟**——见[推迟工作](#deferred-work)。在其落地之前，`web_fetch` 是一个 SSRF 原语，不得在能触达敏感内部网络目标的部署中启用。
+只要 DNS 完整解析结果中存在任一非公开地址，提供方就会拒绝整个结果，而不是静默过滤不安全成员。该 fail-closed 规则可防止连接的地址族选择或回退触及未满足公开网络策略的地址。

 ## 工具消费方行为

@@ -252,7 +254,7 @@ SSRF/私有网络防护（阻断私有、回环、链路本地、多播及其他

 提供方可用性变化影响执行结果和诊断信息，而非面向模型的 schema 是否存在。如果产品完全不需要 web 工具，在配置中禁用 `dsh-tool-web` 或单个 web 工具即可；如果需要 web 工具但后端配置有误，模型在执行时看到结构化的工具错误。

-提示词引导解释了语义分工——`web_search` 用于发现和获取当前信息，`web_fetch` 用于模型需要特定 URL 内容的场景——提示词和工具结果告诉模型用 Markdown 链接引用相关 URL。
+提示词引导解释了语义分工——`web_search` 用于发现和获取当前信息，`web_fetch` 用于模型需要特定 URL 内容的场景——提示词和工具结果告诉模型用 Markdown 链接引用相关 URL。每个成功结果都会把提供方控制的文本标记为外部不可信数据。抓取转换会移除主动内容与隐藏 HTML 内容；无法安全转换时返回固定省略标记，而非原始 HTML。

 面向模型的输出以文本为先，因为工具结果是 `ContentBlock[]`，但 seam 的产出保持结构化，以便 UI 展示和未来的适配器无需解析渲染后的文本。

@@ -308,6 +310,18 @@ SSRF/私有网络防护（阻断私有、回环、链路本地、多播及其他

 在 seam 层面否决。`prompt` 将 fetch 变成 LLM 摘要，并将公开 web 获取耦合到模型提供方。harness seam 应当确定性地获取和解码；`dsh-tool-web` 日后可以将摘要作为展示模式提供，而无需让 `ctx.web` 依赖 `ctx.llm`。

+### 验证 DNS 后调用普通 fetch
+
+否决，因为普通 fetch 在打开连接时会再次解析 hostname。攻击者可以在验证时返回公开地址，在第二次解析时返回私有地址。把已验证解析结果通过连接的 lookup 回调传入，可以在保留基于 hostname 的 HTTP 与 TLS 行为的同时关闭这一 rebinding 时间窗口。
+
+### 只阻断看起来像私网的 hostname 字符串，不固定解析地址
+
+否决，因为 hostname 语法无法确定连接目的地址：任意看似公开的名称都可能解析到 loopback、私有网段或云 metadata 地址。地址分类必须在解析后执行，连接回退可使用的每个地址都必须通过校验。
+
+### 在公开抓取前要求逐次审批
+
+已交付的 preset 不采用这一方案。公开地址校验会阻断 SSRF 目的地址，而逐次确认会打断普通浏览，却不能可靠控制公开数据出站：模型可以通过已挂载的 shell 工具访问同一公开网络。要求专门确认步骤的部署可以添加 `tools/pre-execute` 策略或禁用 `web_fetch`。
+
 ## 后果

 **搜索 schema 刻意精简。** Exa 和 Perplexity 都暴露了有用的提供方特有控制；只有当某个控制能以提供方无关的方式定义、且工具注册和提供方执行都能诚实遵守时，才会添加。
@@ -318,7 +332,7 @@ SSRF/私有网络防护（阻断私有、回环、链路本地、多播及其他

 **提供方状态可能在启动后变化。** 一个工具可能在步骤开始时组装的请求中可见，但在执行前失去其提供方。执行路径重新解析并以结构化错误失败。

-**Fetch 是网络边界，不仅仅是只读工具。** `web_fetch` 能触达敏感网络目标或通过 URL 外泄数据。仅交付基本传输卫生措施（仅 http/https、拒绝凭证、字节/时间上限、跨源重定向阻断）；SSRF/私有网络阻断推迟（见[推迟工作](#deferred-work)），因此在其落地之前，`web_fetch` 不得在能触达内部目标的环境中启用。
+**Fetch 是网络边界，不仅仅是只读工具。** 公开地址校验与连接固定可防止 `web_fetch` 触达非公开目的地址，但模型仍可通过公开 URL 泄露数据，抓取文本也仍是不受信任的模型输入。已交付的 `cordis`、`code` 与 `standard` preset 会在所有 sandbox 和审批模式下暴露 `web_fetch`，无需逐次确认。

 **大量 web 内容可能损害上下文质量。** 提供方强制执行字节/字符上限并报告 `truncated`；`tool-web` 格式化有界的模型输出，附带清晰的继续或后续引导。

@@ -326,13 +340,10 @@ SSRF/私有网络防护（阻断私有、回环、链路本地、多播及其他

 ## 推迟工作

- `web_fetch` 的 SSRF/私有网络防护：阻断私有、回环、链路本地、多播及其他非公开目的地，使 `web_fetch` 不再是 SSRF 原语。正确实现不仅仅是 URL 字符串检查——需要先 DNS 解析再连接到已验证的 IP（防御 DNS rebinding/TOCTOU）、跨重定向的每跳重新验证，以及 IPv6 边缘处理（私有范围、IPv4 映射地址）。所调研的参考实现均未做 IP 级阻断（OpenCode 做前缀检查后直接 fetch；Claude Code 依赖集中式主机名黑名单加「私有 URL 会失败」的提示词），因此没有可复制的实现，且这是 harness 唯一的 SSRF 防线——值得一次专门的设计/spike。在其落地之前，`web_fetch` 只能在无法触达敏感内部目标的部署中启用。
 - `pdf` `WebFetchBody` 类别：`http` 提供方将可文本提取的 PDF 解码（尽力而为、有上限、`truncated`）为 `{ kind: 'pdf'; content; pageCount? }` 分支，`tool-web` 渲染它。这是 fetch 而非 `web_extract`——PDF 获取是具体的 HTTP 200 加确定性的本地解码，不是提供方侧对非 HTTP 资源的提取。添加它是跨 `dsh-web`（声明分支）、提供方（解码 + 将「二进制拒绝」收窄为「拒绝二进制，但可文本提取的 PDF 除外」；需要 OCR 的扫描/图片 PDF 不在范围内）和 `tool-web`（渲染）的协调变更。封闭的 `WebFetchBody` 联合类型使消费方在新分支被处理之前编译失败。
 - 提供方支撑的提取作为独立的 `web_extract` 能力，而非静默扩展 `web_fetch`。
- 权限策略集成：权限系统现已存在（[沙箱与审批](../feature/2026-07-06-sandbox.zh.md)、[web 权限预设](../feature/2026-07-23-web-permission-and-approval.zh.md)），但只捆绑了沙箱模式与审批策略；web 权限策略仍未集成。
 - `query` 和 `maxResults` 之外的提供方无关搜索控制，待 Exa 和 Perplexity 都能诚实遵守时再添加。

 ## 开放问题

 - 产品应用包是否应在启动时探测 web 配置（当 web 被显式配置时将 `WEB_PROVIDER_CONFIGURED_MISSING`、`WEB_PROVIDER_CONFIGURED_UNAVAILABLE` 和 `WEB_PROVIDER_AMBIGUOUS` 视为致命错误），还是将配置错误留到首次执行时浮出？
- 在已交付的权限系统（[沙箱与审批](../feature/2026-07-06-sandbox.zh.md)、[web 权限预设](../feature/2026-07-23-web-permission-and-approval.zh.md)）中，公开 web 访问的权限策略应放在哪里：`tools/execute` 上的专用 web 权限插件、提供方配置，还是两者兼有？
@@ -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/architecture/2026-07-05-reconstructable-requests.md
-2026-07-05-reconstructable-requests.md: 3f49ba71a6b98a84b05530c900e902b0cf9f6449
-2026-07-05-reconstructable-requests.zh.md: 7b8a9df65b60f975bc3ae60b2c1b0c3a8cc22e95
+2026-07-05-reconstructable-requests.md: 3786de02d06c0b6c094297ae89ac3f84053e408d
+2026-07-05-reconstructable-requests.zh.md: 851045aca7dababd0da859f3b04b721c65382fc3
@@ -22,9 +22,9 @@ Prefix-cache stability is corollary #1, not the headline: an append-only log pro

 **Messages.** `Session.deriveMessages()` is cached: each surface entry is projected exactly once, when first seen, through the public per-event function `deriveEventMessage(event)`; a surface rewrite (a compaction `replace` — `SurfaceManager.replaceGeneration`) rebuilds. Callers get a fresh array per call over shared, deep-frozen messages: mutating logged history through a projection is unrepresentable (it throws), replacing the old clone-per-call isolation. External reconstructors fold the same public function over a log prefix, so no two paths can disagree.

-`EpochHeader` records the request's non-history state: call config, rendered system prompt, and tool schemas, with empty values canonicalized to absence. `request/header` always writes a full snapshot: the first loop instance uses reason `initial`, later instances use `resume`, and an in-instance change uses `change`. `foldRequestHeader` selects the latest snapshot. Legacy `request/header-delta` events and the removed `fallback` reason are rejected when appended or loaded.
+`EpochHeader` records the request's non-history state: call config, rendered system prompt, and tool schemas, with empty values canonicalized to absence. Adapter-supplied effort and token defaults retain their `adapterDefaults` provenance; a Web model selection restored from the log omits an adapter-owned effort so the next resolution cannot reclassify the same effective config as an explicit selection and a false change. `request/header` always writes a full snapshot: the first loop instance uses reason `initial`, later instances use `resume`, an in-instance change uses `change`, and an unchanged envelope beginning an explicitly declared message series or following a surface replacement uses `series`. A `change` snapshot carries `startsSeries: true` when the changed request also starts a series, preserving the two independent facts without a duplicate header. Ordinary append-only later Turns, further same-series Steps, and retries inherit the latest snapshot. `foldRequestHeader` selects the latest snapshot. Legacy `request/header-delta` events and the removed `fallback` reason are rejected when appended or loaded.

-Each proposed step first claims its inbox batch and runs `agent/pre-step`. Rejection opens no step; enter opens `step/start` and records the final message batch as `user/message` events. The step then assembles the system prompt and tools, while `agent/request` may replace only the frozen call-config seed. The loop records the owed full header snapshot, builds `GenerateOptions` from derived messages and that header, and deep-freezes it while leaving `AbortSignal` live. The first call config starts from explicit `AgentOptions`, preserving fork overrides and resume reconfiguration; later calls start from the folded header.
+Each proposed step first claims its inbox batch and runs `agent/pre-step`. Rejection opens no step; enter opens `step/start`, records the final message batch as `user/message` events, and may use `startsRequestSeries: true` to declare a distinct series. The step then assembles the system prompt and tools, while `agent/request` may replace only the frozen call-config seed. The loop records the owed initial, resume, change, or series full snapshot, builds `GenerateOptions` from derived messages and that header, and deep-freezes it while leaving `AbortSignal` live. The first call config starts from explicit `AgentOptions`, preserving fork overrides and resume reconfiguration; later calls start from the folded header.

 **The open step is the reconstruction boundary.** Its entered `user/message` batch and any newly written `request/header` precede request dispatch. Injection after the atomic claim joins a later request, while a listener that must affect this request returns messages through `agent/pre-step`. Header reconstruction selects the step's `request/header`, or carries the prior snapshot when no new header is written.

@@ -42,6 +42,7 @@ Like MiniCode, the conversation advances append-only and resets only when model-
 - **Detect-and-report** (compare consecutive requests, warn on divergence): catches violations after the fact; a violating request is still constructible and ships. Rejected for interface-level unrepresentability.
 - **Event-driven assembly** (re-render only on change signals): a missed-signal bug class — a tool registered mid-session emits `tools/change`, not `system-prompt/change`, and a third-party provider may emit nothing. Per-step render + value compare is robust with zero signal discipline.
 - **A custom header-delta codec** (system line edits, name-keyed tool edits, whole config/prefix replacements): reduced repeated bytes but duplicated the representation and its diff/apply/fallback machinery. Full snapshots retain one replay representation.
+- **A lightweight series marker referencing the previous header**: reduced repeated prompt and tool bytes, but a window beginning at that marker could not render or reconstruct the request without fetching its predecessor. A self-contained full snapshot preserves one representation for persistence, partial history, and snapshot pinning.
 - **Narrative changed-field lists on header snapshots**: derivable by comparing consecutive snapshots. The `reason` remains because an instance boundary is not derivable from the snapshot values.

 ## Consequences
@@ -52,5 +53,5 @@ Like MiniCode, the conversation advances append-only and resets only when model-
 - `agent/pre-step` is the current-request message channel; direct inbox mutation is the eventual later-request channel.
 - Tool-result trimming needs no new mechanism: a logged single-entry surface replace (`start === end`) carrying a trimmed `tool/result` under the same `callId` — compaction-family, replay-correct, cache-bust batched by the same pressure logic.
 - Unreadable referenced attachment objects still fail model requests; [automatic attachment quarantine](../../proposed/bug-fix/2026-08-20-attachment-read-quarantine.md) records the proposed recovery without weakening byte-exact reconstruction.
- Session logs grow one `request/header` snapshot per loop instance plus snapshots on real changes. This is larger than a delta codec but small beside chunk-heavy logs and retains one replay representation. `SESSION_FORMAT_VERSION` stays `0`; legacy delta events are rejected rather than migrated.
- Snapshot expected outputs changed once (every transcript gains its header events); the fs-writing fixtures are stored in the normalized authored form with cwd-relative tool arguments, because replay only round-trips cwd-independent argument paths.
+- Session logs grow one `request/header` snapshot per loop instance, real change, and later model-message series. Repeating the full system prompt and tool catalog is larger than a delta codec but small beside chunk-heavy logs and retains one self-contained replay representation. `SESSION_FORMAT_VERSION` stays `0`; legacy delta events are rejected rather than migrated.
+- Snapshot fixtures include each repeated series header. Keyless refresh owns those deterministic log changes, while the snapshot harness pins prompt and tool sidecars only for the initial and actual change revisions and reuses the current revision for `series` snapshots. Filesystem-writing fixtures remain in normalized authored form with cwd-relative tool arguments because replay only round-trips cwd-independent argument paths.
@@ -22,9 +22,9 @@ Status: implemented

 **消息。** `Session.deriveMessages()` 带缓存：每个 surface 条目在首次出现时通过公开的逐事件函数 `deriveEventMessage(event)` 精确投影一次；surface 重写（压缩的 `replace`，即 `SurfaceManager.replaceGeneration`）触发重建。调用方每次获得一个新数组，底层是共享的深度冻结消息：通过投影变异已记录的历史是不可表达的（会抛异常），取代了旧的逐次调用克隆隔离。外部重建器对日志前缀折叠同一个公开函数，因此不可能有两条路径产生分歧。

-`EpochHeader` 记录请求的非历史状态：调用配置、渲染后的系统提示词和工具 schema，空值规范化为缺失。`request/header` 始终写入完整快照：首个循环实例使用 reason `initial`，后续实例使用 `resume`，实例内变更使用 `change`。`foldRequestHeader` 选择最新快照。旧的 `request/header-delta` 事件和已移除的 `fallback` reason 在追加或加载时都会被拒绝。
+`EpochHeader` 记录请求的非历史状态：调用配置、渲染后的系统提示词和工具 schema，空值规范化为缺失。适配器提供的推理强度与 token 默认值会保留其 `adapterDefaults` 来源信息；Web 从日志恢复模型选择时会省略适配器持有的推理强度，因此下一次解析不会把相同的有效配置重新归类为显式选择并产生虚假变更。`request/header` 始终写入完整快照：首个循环实例使用 reason `initial`，后续实例使用 `resume`，实例内变更使用 `change`，内容未变的封装显式开启消息序列或跟随表层替换时使用 `series`。如果发生变化的请求同时开启序列，`change` 快照会携带 `startsSeries: true`，无需重复 header 即可保留这两个独立事实。普通的仅追加后续 Turn、同一序列内后续的 Step 与重试沿用最新快照。`foldRequestHeader` 选择最新快照。旧的 `request/header-delta` 事件和已移除的 `fallback` reason 在追加或加载时都会被拒绝。

-每个拟议步骤先领取其 inbox 批次，再运行 `agent/pre-step`。reject 不打开步骤；enter 打开 `step/start`，并把最终消息批次记录为 `user/message` 事件。随后步骤组装系统提示词与工具，`agent/request` 只能替换冻结的调用配置种子。循环记录所需的完整 header 快照，从派生消息与该 header 构建 `GenerateOptions`，对其深度冻结但保持 `AbortSignal` 活跃。首次调用配置从显式的 `AgentOptions` 出发，保留 fork 覆盖和恢复重配置；后续调用从折叠后的 header 出发。
+每个拟议步骤先领取其 inbox 批次，再运行 `agent/pre-step`。reject 不打开步骤；enter 打开 `step/start`，把最终消息批次记录为 `user/message` 事件，并可使用 `startsRequestSeries: true` 声明独立序列。随后步骤组装系统提示词与工具，`agent/request` 只能替换冻结的调用配置种子。循环记录所需的 initial、resume、change 或 series 完整快照，从派生消息与该 header 构建 `GenerateOptions`，对其深度冻结但保持 `AbortSignal` 活跃。首次调用配置从显式的 `AgentOptions` 出发，保留 fork 覆盖和恢复重配置；后续调用从折叠后的 header 出发。

 **已打开步骤是重建边界。** 进入步骤的 `user/message` 批次与任何新写入的 `request/header` 都位于请求分派之前。原子领取后发生的注入加入后续请求；必须影响本次请求的监听器则通过 `agent/pre-step` 返回消息。header 重建选择该步骤的 `request/header`，或在无新 header 写入时沿用前一个快照。

@@ -42,6 +42,7 @@ Status: implemented
 - **检测并报告**（比较连续请求，发散时告警）：事后捕获违规；违规请求仍可构造并发出。因违规必须在接口层面不可表达而否决。
 - **事件驱动组装**（仅在变更信号时重新渲染）：存在漏信号的 bug 类别——会话中途注册的工具发出 `tools/change` 而非 `system-prompt/change`，第三方提供方可能什么都不发。逐步骤渲染加值比较在零信号纪律下即可稳健工作。
 - **自定义 header-delta 编解码器**（系统行编辑、按名称键控的工具编辑、完整配置/前缀替换）：减少了重复字节，却复制了表示及其 diff/apply/fallback 机制。完整快照只保留一种回放表示。
+- **引用前一个 header 的轻量 series 标记**：减少重复的提示词与工具字节，但从该标记开始的窗口若不再读取前序，就无法渲染或重建请求。自包含的完整快照让持久化、局部历史和快照固定共用一种表示。
 - **Header 快照上的叙事性变更字段列表**：可以通过比较连续快照推导。`reason` 仍保留，因为实例边界无法从快照值推导。

 ## 后果
@@ -52,5 +53,5 @@ Status: implemented
 - `agent/pre-step` 是当前请求的消息通道；直接修改 inbox 则是最终进入后续请求的通道。
 - 工具结果裁剪无需新机制：一个已记录的单条目 surface replace（`start === end`），携带同一 `callId` 下裁剪后的 `tool/result`——属压缩家族，回放正确，缓存失效由相同的压力逻辑批量处理。
 - 无法读取的被引用附件对象仍会让模型请求失败；[附件自动隔离](../../proposed/bug-fix/2026-08-20-attachment-read-quarantine.zh.md)记录了不削弱字节精确重建的拟议恢复方案。
- 会话日志每个循环实例增长一个 `request/header` 快照，并在真正变更时增加快照。它比 delta 编解码器更大，但相对分片密集型日志仍然很小，并只保留一种回放表示。`SESSION_FORMAT_VERSION` 保持 `0`；旧的 delta 事件被拒绝而非迁移。
- 快照预期输出变更一次（每个 transcript（文本记录）增加其 header 事件）；写入文件系统的 fixture（测试前置数据）以规范化的撰写形式存储，工具参数使用 cwd 相对路径，因为回放只对 cwd 无关的参数路径做往返。
+- 会话日志会为每个循环实例、真实变更和后续模型消息序列增加一个 `request/header` 快照。重复完整系统提示词与工具目录比 delta 编解码器更大，但相对分片密集型日志仍然很小，并保留一种自包含的回放表示。`SESSION_FORMAT_VERSION` 保持 `0`；旧的 delta 事件被拒绝而非迁移。
+- 快照 fixture 包含每个重复的 series header。无密钥 refresh 负责这些确定性日志变化；快照 harness 只为 initial 与真实 change 修订固定提示词和工具 sidecar，并让 `series` 快照复用当前修订。写入文件系统的 fixture 继续以规范化的撰写形式存储，工具参数使用 cwd 相对路径，因为回放只对 cwd 无关的参数路径做往返。
@@ -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/architecture/2026-07-29-projected-token-usage-and-request-context.md
-2026-07-29-projected-token-usage-and-request-context.md: f2179885512bcb216ecb191ce98b535db571807a
-2026-07-29-projected-token-usage-and-request-context.zh.md: e4435b6245d1e20b51fc2cc1d73151ced8d94731
+2026-07-29-projected-token-usage-and-request-context.md: 063f2300f378f6f7763bce87b11add5da3093230
+2026-07-29-projected-token-usage-and-request-context.zh.md: 37b8741d09e9ec56f6b9f273e05460b2deb4f6f9
@@ -14,7 +14,9 @@ Context occupancy needs a numerator and a denominator that no existing surface c

 Both values are ordinary durable session-projection state. `@deepseek-ai/dsh-token-meter` registers two units when `ctx.sessionProjections` is present.

-`tokenUsage` folds the complete durable log into uncached input, output, cache-read, and cache-write buckets. An `assistant/chunk` usage sample survives a later failed request; an `assistant/message` usage value for the same `(turn, step)` replaces the earlier sample instead of double-counting it. Reasoning stays an output subdivision. Compaction and surface replacement do not erase earlier billing.
+`tokenUsage` folds the complete durable log into uncached input, output, cache-read, and cache-write buckets. An `assistant/chunk` usage sample survives a later failed request; an `assistant/message` usage value replaces the earlier sample from the same model attempt instead of double-counting it. A matching `llm/retry-started` boundary ends that replacement scope, so a retry with the same `(turn, step)` contributes a new attempt. Reasoning stays an output subdivision. Compaction and surface replacement do not erase earlier billing.
+
+Token-meter also owns the shared pure attempt/Turn fold over durable events. It applies the same retry boundary while adding the stricter completeness and exact-total checks required by an exact per-Turn disclosure. A presentation consumer may select a complete Turn window and invoke that fold, but does not own or duplicate the accounting semantics.

 `contextPressure` carries optional `pressureTokens` — the newest provider-reported prompt size, summing uncached input plus cache reads and writes, excluding output — and optional `contextWindow` from the newest `request/context` record. Neither field is synthesized before its source exists.

@@ -56,4 +58,4 @@ Token totals stay stable across pagination, compaction, replay, restart, and rec

 Occupancy is approximate in the ways documented above. It is available immediately after restore or reconnect, since both fields are durable, at the cost of describing the last recorded request rather than an exact current boundary.

-Each session log gains one small `request/context` record per route or advertised-capacity change. The token-meter projection is the canonical owner of durable session-projection usage semantics; the TUI retains its live per-step map because it does not mount the generic projection seam, and the standalone browser fixture mirrors the unit. ApiProxy carries no token-specific code, owns no per-session metrics cache, and performs no measurement. The browser keeps two generic projection values and no connection-local telemetry, and streaming text deltas still do not force the stats line to recompute.
+Each session log gains one small `request/context` record per route or advertised-capacity change. Token-meter is the canonical owner of durable usage semantics, including retry-attempt separation in the cumulative projection and the reusable exact attempt/Turn fold; Web Chat only selects a complete loaded Turn and renders the fold result. The TUI retains its live per-step map because it does not mount the generic projection seam, and the standalone browser fixture mirrors the unit. ApiProxy carries no token-specific code, owns no per-session metrics cache, and performs no measurement. The browser keeps two generic projection values and no connection-local telemetry, and streaming text deltas still do not force the stats line to recompute.
@@ -14,7 +14,9 @@ Web 统计行原先从当前已加载的会话节点推导 token 总量。该窗

 这两个值都是普通的持久会话投影状态。当 `ctx.sessionProjections` 存在时，`@deepseek-ai/dsh-token-meter` 会注册两个单元。

-`tokenUsage` 将完整持久日志归并为未缓存输入、输出、缓存读取和缓存写入四类计数项。即使后续请求失败，`assistant/chunk` 用量样本仍会保留；同一 `(turn, step)` 的 `assistant/message` 用量值会替换先前样本，不会重复计数。推理（reasoning）仍是输出的细分项。压缩和表层替换不会抹除先前的计费用量。
+`tokenUsage` 将完整持久日志归并为未缓存输入、输出、缓存读取和缓存写入四类计数项。即使后续请求失败，`assistant/chunk` 用量样本仍会保留；`assistant/message` 用量值会替换同一次模型 attempt 的先前样本，不会重复计数。匹配的 `llm/retry-started` 边界会结束该替换作用域，因此复用同一 `(turn, step)` 的重试会贡献一次新的 attempt。推理（reasoning）仍是输出的细分项。压缩和表层替换不会抹除先前的计费用量。
+
+token-meter 还拥有在持久事件上运行的共享纯 attempt／Turn fold。它采用相同的重试边界，并增加精确单轮次 disclosure 所需的更严格完整性与精确总量检查。展示消费方可以选择完整 Turn 窗口并调用该 fold，但不拥有或复制记账语义。

 `contextPressure` 携带可选的 `pressureTokens`（提供方报告的最新提示词规模，为未缓存输入加缓存读取与写入之和，不含输出），以及来自最新一条 `request/context` 记录的可选 `contextWindow`。在各自来源出现前，两个字段都不会被合成。

@@ -56,4 +58,4 @@ token 总量在分页、压缩、回放、重启和重连期间保持稳定，

 占用率在上文记录的意义上是近似值。由于两个字段都是持久的，它在恢复或重连后立即可用；代价是它描述的是最后一条已记录的请求，而不是精确的当前边界。

-每个会话日志会为每次路由或已公布容量变化增加一条小型 `request/context` 记录。token-meter 投影是持久会话投影用量语义的正典所有方；TUI 未挂载通用投影 seam，因此保留自己的实时逐步骤 map，而独立浏览器 fixture（测试前置数据）会镜像该单元。ApiProxy 不携带任何 token 专用代码，不拥有逐会话指标缓存，也不执行测量。浏览器只保留两个通用投影值，不保留连接本地的遥测数据；流式文本增量仍不会迫使统计行重新计算。
+每个会话日志会为每次路由或已公布容量变化增加一条小型 `request/context` 记录。token-meter 是持久用量语义的正典所有方，包括累计投影中的重试 attempt 分离，以及可复用的精确 attempt／Turn fold；Web Chat 只选择已完整加载的 Turn 并渲染 fold 结果。TUI 未挂载通用投影 seam，因此保留自己的实时逐步骤 map，而独立浏览器 fixture（测试前置数据）会镜像该单元。ApiProxy 不携带任何 token 专用代码，不拥有逐会话指标缓存，也不执行测量。浏览器只保留两个通用投影值，不保留连接本地的遥测数据；流式文本增量仍不会迫使统计行重新计算。
@@ -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/architecture/2026-08-09-client-conversation-node-assembly.md
-2026-08-09-client-conversation-node-assembly.md: e6c0e790a361265870a04ee63301b9f11940c648
-2026-08-09-client-conversation-node-assembly.zh.md: 702ddba0019e125d3976727f841db775276b3b77
+2026-08-09-client-conversation-node-assembly.md: ea2505d4a72f483a9df6fcd78d7e5c9a96b02f5c
+2026-08-09-client-conversation-node-assembly.zh.md: b87f127d753cadf2805ed5cd948fc58ad01830aa
@@ -260,6 +260,7 @@ Page size, the number of history loads, and RAF coalescing affect only when evid
 | Next-turn Inbox / `inbox-next-turn` | Splice Event seq | Each `agent/inbox/spliced` targeting next-turn | None | Apply the current splice to the pending/claimed instantaneous state from `reader.previous(ownKind)` |
 | Next-step Inbox / `inbox-next-step` | Splice Event seq | Each `agent/inbox/spliced` targeting next-step | None | Build the same per-instruction instantaneous state; Message reads its claimed set |
 | Message / `input-message` | Message ID | Append-surface `user/message` | None | Use source for a context message, or read the nearest next-step Inbox to distinguish user from steering |
+| Request Prompt / `request-prompt` | Header Event seq | Each `request/header` | None | Read the preceding Request Prompt through Reader, retain the full prompt state, and classify system/tool changes |
 | Assistant / `assistant-step` | `turn:step` | `step/start` | `assistant/chunk`, final `assistant/message`, and same-step Retry | Aggregate blocks, usage, first-token time, final evidence, and retry-hidden state, then publish same-key Step data |
 | Tool / `tool-call` | Root call ID | Root `tool/call` | Root result and Code Dispatch start/result | Aggregate the root, children, and parent Map; Dispatch Events route exactly through `rootCallId` |
 | Command / `command` | Command ID | `command/run` | `command/done` and compact lifecycle/checkpoint Events carrying a source command ID | Aggregate command outcome and manual-compaction evidence |
@@ -276,6 +277,7 @@ Page size, the number of history loads, and RAF coalescing affect only when evid
 |---|---|---|---|
 | Inbox | `none` | No Node | Recompute instantaneous states along the Reader chain when prepend supplies earlier splices |
 | Message | Immediate by default | `user`, `steering`, or `context` | Window-gap repair can reclassify the same message key |
+| Request Prompt | Immediate by default | One `system-prompt` for every header carrying a non-empty system field | A step's first header anchors before its request messages; a later same-step series anchors after its surface rewrite; prepend of the preceding header can correct a partial-window anchor |
 | Assistant | RAF for chunks, immediate for final, none for pure usage/finish | Same-key `assistant-step` with running/settled/interrupted status | Matches support fallback without `step/start`; Location close produces interruption presentation |
 | Tool | Immediate by default | One recursive `tool-call` root containing all `subCalls` | A result-only history window supports fallback; running→settled retains its key |
 | Command | Immediate by default | Ordinary `command` or integrated `manual-compaction` | Checkpoint arrival may change the anchor without changing the Context key |
@@ -288,6 +290,8 @@ Page size, the number of history loads, and RAF coalescing affect only when evid

 Inbox demonstrates that every Event can be a start-only instantaneous-state Context; not every business requires a start/update pair. Reader links each state to the prior same-kind Context instead of inventing a lifecycle ID for the entire Inbox.

+Request Prompt demonstrates shared pure interpretation without shared target State: Chat and Trajectory call `inspectRequestPrompt()` from their own Definitions. The function canonicalizes the full header and classifies model-visible system/tool differences; each target then chooses its own output. Chat materializes every header carrying a non-empty system field, including `series` snapshots that repeat an unchanged header for an explicitly declared series or a post-replacement request, while Trajectory retains the complete request fact and its change classification. Ordinary append-only later Turns do not write another unchanged header. The first header in a Step follows the provider envelope rather than the header Event position: step one uses the owning Turn start and later steps use their Step start, placing the system field before the request's user-role messages; a later header in the same Step stays at its own Event after the surface rewrite that began the new series. When the preceding header is outside a partial window, a non-`initial` header stays at its own Event until prepend supplies that predecessor. Every header is a full snapshot, so a first loaded `resume`, `change`, or `series` header can render its system field without fabricating a comparison to unloaded history.
+
 Retry, Assistant, and Turn Tail demonstrate independent claims on one Event. Each Definition updates only its own State and produces its own atomic Chat Node.

 Assistant, Turn Tail, and Deliverables demonstrate layered Location data composition. Assistant writes `assistant-step` data for each Step; Turn Tail derives `turn-tail` data from those Step values; Deliverables independently maintains `deliverables` data for the same Turn. Consumers read only declaration-merged keys, do not scan another business's Nodes, and cannot obtain the provider's Context State.
@@ -260,6 +260,7 @@ Chat `order` 的结构性变化仍可能重排当前可见 key；纯 data 更新
 | Next-turn Inbox / `inbox-next-turn` | splice Event seq | 每条目标为 next-turn 的 `agent/inbox/spliced` | 无 | 从 `reader.previous(ownKind)` 的 pending/claimed 瞬间态应用当前 splice |
 | Next-step Inbox / `inbox-next-step` | splice Event seq | 每条目标为 next-step 的 `agent/inbox/spliced` | 无 | 同样形成逐指令瞬间态，claimed 集合供 Message 读取 |
 | Message / `input-message` | message ID | append-surface `user/message` | 无 | 根据 source 生成 context message，或读取最近 next-step Inbox 判断 user/steering |
+| Request Prompt / `request-prompt` | header Event seq | 每条 `request/header` | 无 | 通过 Reader 读取前一条 Request Prompt，保留完整 prompt 状态，并判定 system/tool 变化 |
 | Assistant / `assistant-step` | `turn:step` | `step/start` | `assistant/chunk`、final `assistant/message`、同 step Retry | 聚合 blocks、usage、首 token 时间、final 和 retry 隐藏状态，并发布同 key Step data |
 | Tool / `tool-call` | root call ID | root `tool/call` | root result、Code Dispatch start/result | 聚合 root、children 和 parent Map；Dispatch Event 用 `rootCallId` 精确路由 |
 | Command / `command` | command ID | `command/run` | `command/done`、带 source command ID 的 compact lifecycle/checkpoint | 聚合 command outcome 和手动压缩证据 |
@@ -276,6 +277,7 @@ Chat `order` 的结构性变化仍可能重排当前可见 key；纯 data 更新
 |---|---|---|---|
 | Inbox | `none` | 不生成 Node | prepend 补前序 splice 时沿 Reader 链重算瞬间态 |
 | Message | 默认 immediate | `user`、`steering` 或 `context` | window gap 修复可让同一 message key 重新分类 |
+| Request Prompt | 默认 immediate | 每条带非空 system 字段的 header 都生成一个 `system-prompt` | Step 首条 header 锚定在请求消息之前；同 step 后续序列锚定在表层改写之后；prepend 补入前序 header 后可纠正部分窗口的锚点 |
 | Assistant | chunk 为 RAF，final immediate，纯 usage/finish 为 none | 同 key `assistant-step`，状态为 running/settled/interrupted | 缺 `step/start` 可先用 Matches fallback；Location close 生成中断表现 |
 | Tool | 默认 immediate | 一个递归 `tool-call` root，包含全部 `subCalls` | result-only 历史窗口可 fallback；running→settled 保持 key |
 | Command | 默认 immediate | 普通 `command` 或集成 `manual-compaction` | checkpoint 到达可改变 anchor，但不改变 Context key |
@@ -288,6 +290,8 @@ Chat `order` 的结构性变化仍可能重排当前可见 key；纯 data 更新

 Inbox 展示了“每条 Event 都是一个 start-only 瞬间态 Context”，不是所有业务都需要 start/update 配对。它通过 Reader 与前一个同 kind Context 形成连续 fold，而非给整个 Inbox 人工制造生命周期 ID。

+Request Prompt 展示了如何在不共享 target State 的前提下共用纯解释逻辑：Chat 与 Trajectory 各自在自己的 Definition 中调用 `inspectRequestPrompt()`。该函数规范化完整 header，并判定面向模型的 system/tool 差异；随后每个 target 自行选择产物。Chat 会物化每条带非空 system 字段的 header，包括为显式声明的序列或表层替换后的请求重复未变 header 的 `series` 快照；Trajectory 则保留完整请求事实及其变化分类。普通的仅追加后续 Turn 不会再次写入未变 header。一个 Step 中的首条 header 遵循提供方信封，而不是 header Event 位置：step one 使用所属 Turn start，后续 step 使用各自的 Step start，把 system 字段放到该请求的 user-role 消息之前；同一 Step 的后续 header 保留在开启新序列的表层改写之后。部分窗口未包含前序 header 时，非 `initial` header 会保留在自身 Event，直到 prepend 补入该前序 header。每条 header 都是完整快照，因此已加载窗口中的首条 `resume`、`change` 或 `series` header 无需凭空构造与未加载历史的比较，也能渲染其 system 字段。
+
 Retry、Assistant 和 Turn Tail 展示了同一 Event 被多个 Definition 独立认领。每个 Definition 只更新自己的 State，最终分别生成原子 Chat Node。

 Assistant、Turn Tail 和 Deliverables 展示了 Location data 的分层组合。Assistant 负责写好每个 Step 的 `assistant-step` data；Turn Tail 从这些 Step values 计算 `turn-tail` data；Deliverables 独立维护同一 Turn 的 `deliverables` data。消费者只读取声明合并后的 key，不扫描其他业务 Node，也不取得提供方的 Context State。
@@ -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/architecture/2026-08-09-headless-direct-core-entry-point.md
-2026-08-09-headless-direct-core-entry-point.md: 8ed979794afa008588d1b849f0074e8696e6e43f
-2026-08-09-headless-direct-core-entry-point.zh.md: 512d4b88c921431fe26afd9f62c34a1939ac5bdd
+2026-08-09-headless-direct-core-entry-point.md: 9c17b8d418924c38174b4d958fd54b057b118019
+2026-08-09-headless-direct-core-entry-point.zh.md: d95978a832d52b26b1139cabb4b23ade93ce0da3
@@ -6,7 +6,7 @@ English | [中文](2026-08-09-headless-direct-core-entry-point.zh.md)

 ## Problem

-The `headless` product contract is one local task with final assistant text on stdout, a success-sensitive exit code, empty stderr on success, and no listening port. A composition containing Workspace Host services, ApiProxy, HTTP, the Web runtime, or browser plugins contradicts that contract and makes local completion depend on an unrelated transport tree.
+The `headless` product contract is one local task with final assistant text on stdout, a success-sensitive exit code, no listening port, and the stderr reasoning projection owned by [headless reasoning progress](../feature/2026-08-21-headless-reasoning-progress.md). A composition containing Workspace Host services, ApiProxy, HTTP, the Web runtime, or browser plugins contradicts that contract and makes local completion depend on an unrelated transport tree.

 The direct entry point still needs the same deployment model state as Web-created Agents. A separate provider/model default would give one deployment two answers, while deriving completion before the Agent and Session persistence are quiescent permits stdout and the exit code to observe incomplete state.

@@ -14,17 +14,17 @@ The direct entry point still needs the same deployment model state as Web-create

 The shipped `headless` profile contains `dsh-base` and `dsh-headless`. The base supplies the disabled module-HMR default; the headless bundle supplies its persona and tool mode, mounts the Code Mode worker explicitly, and inserts `headless-runner` without overriding that policy. Its tree contains no `@deepseek-ai/dsh-host-*` package, ApiProxy, HTTP server, Web runtime, or browser client. Code Mode and Session persistence are one-shot Agent capabilities independent of Web presentation.

-`headless-runner` is a direct core entry point. After Loader settlement, it reads `ctx.agentDefaultModel.currentSelection()`, creates a fresh persisted Agent through `ctx.agents.create`, installs that `ModelSelection` in the Agent scope, waits for startup quiescence, anchors the Session sequence, submits one ordinary user message, and waits for quiescence again. It awaits `ctx.sessions.flush`, folds its durable event interval for the last non-empty assistant text and final `turn/end` reason, writes the text plus one newline to stdout, and requests bounded launcher shutdown with exit 0 exactly when the reason is `completed`. A terminal `error` reason writes its durable code and message to stderr; unexpected driver failures also use stderr and exit 1.
+`headless-runner` is a direct core entry point. After Loader settlement, it reads `ctx.agentDefaultModel.currentSelection()`, creates a fresh persisted Agent through `ctx.agents.create`, installs that `ModelSelection` in the Agent scope, waits for startup quiescence, anchors the Session sequence, submits one ordinary user message, and waits for quiescence again. It awaits `ctx.sessions.flush`, folds its durable event interval for the last non-empty assistant text and final `turn/end` reason, writes the text plus one newline to stdout, and requests bounded launcher shutdown with exit 0 exactly when the reason is `completed`. [Headless reasoning progress](../feature/2026-08-21-headless-reasoning-progress.md) owns the live stderr projection; a terminal `error` reason writes its durable code and message there, and unexpected driver failures also use stderr and exit 1.

 `@deepseek-ai/dsh-agent-default-model` owns the transport-independent default used for an Agent without a session-local selection. `AgentDefaultModelConfig` provides `ctx.agentDefaultModel` and registers the `agent-default-model` Settings section. Composition config supplies `{provider, model}`; user settings may also supply `reasoningEffort`. `currentSelection()` returns the live complete selection and `saveSelection()` writes it as a complete section, so a selection without an effort clears any stored effort. `dsh-base` supplies the composition entry. Direct and ApiProxy entry points consume this service; ApiProxy alone owns session-local precedence, model validation, and persistence of accepted Web selections.

 `loadProfile` recognizes the exact installation-owned headless tuple (`dsh-base`, `dsh-web-app`, `dsh-headless`) and normalizes it to the shipped headless template while preserving every other manifest field. Extra, missing, or reordered bundle lists are user-owned and remain untouched.

-This note owns the headless transport and completion contracts. [Apps own their command lines](2026-08-06-app-owned-command-line.md) owns the current `dsh --profile headless` grammar; the former [`dsh run` decision](../../archived/feature/2026-08-08-dsh-run-headless-command.md) records the superseded launcher-owned grammar, [GUI layering and RPC protocol](2026-07-19-gui-layering-and-rpc-protocol.md) owns browser gateway boundaries, [web config-tree boot and transport layering](2026-07-24-web-config-tree-boot-and-transport-layering.md) owns the Web tree, and [the default model follows the picker](../feature/2026-08-07-default-model-follows-the-picker.md) owns persistence of the shared Agent default.
+This note owns the headless transport and completion contracts; [headless reasoning progress](../feature/2026-08-21-headless-reasoning-progress.md) owns successful stderr output. [Apps own their command lines](2026-08-06-app-owned-command-line.md) owns the current `dsh --profile headless` grammar; the former [`dsh run` decision](../../archived/feature/2026-08-08-dsh-run-headless-command.md) records the superseded launcher-owned grammar, [GUI layering and RPC protocol](2026-07-19-gui-layering-and-rpc-protocol.md) owns browser gateway boundaries, [web config-tree boot and transport layering](2026-07-24-web-config-tree-boot-and-transport-layering.md) owns the Web tree, and [the default model follows the picker](../feature/2026-08-07-default-model-follows-the-picker.md) owns persistence of the shared Agent default.

 ## Verification

-Package tests use the real Session store and Agent registry around a scripted Agent factory to pin idle-to-idle aggregation, late asynchronous completion, terminal model diagnostics, other non-completed exits, direct failures, Loader-time disposal, and flush-before-exit ordering. The keyless assembled snapshots drive `dsh --profile headless` through a replayed tool round trip, record a `user/message` with `source.kind: 'user'`, and expose a terminal model failure on stderr. Built-bin acceptance reaches a mock provider through the published entry and requires final text on stdout, exit 0, and empty stderr. Config-dump acceptance excludes every Host, Web, and Client package from the shipped headless tree; PTY shutdown coverage requires no observation line and bounded disposal.
+Package tests use the real Session store and Agent registry around a scripted Agent factory to pin idle-to-idle aggregation, late asynchronous completion, terminal model diagnostics, other non-completed exits, direct failures, Loader-time disposal, and flush-before-exit ordering. The keyless assembled snapshots drive `dsh --profile headless` through a replayed tool round trip, record a `user/message` with `source.kind: 'user'`, and expose both reasoning progress and a terminal model failure on stderr. Built-bin acceptance reaches a mock DeepSeek endpoint through the published entry and requires streamed reasoning on stderr, final text on stdout, and exit 0. Config-dump acceptance excludes every Host, Web, and Client package from the shipped headless tree; PTY shutdown coverage requires no observation line and bounded disposal.

 ## Alternatives considered

@@ -39,6 +39,6 @@ Package tests use the real Session store and Agent registry around a scripted Ag

 ## Consequences

-`dsh --profile headless` provides a local Agent task rather than browser observation, Host APIs, or HTTP. Users who need those capabilities choose `dsh web`. Successful stderr is empty, completion follows durable flush, and the persisted Session remains available to later tooling. Its initial user message records `source.kind: 'user'` and therefore carries no ApiProxy `rpcId`.
+`dsh --profile headless` provides a local Agent task rather than browser observation, Host APIs, or HTTP. Users who need those capabilities choose `dsh web`. Text-only successful runs leave stderr empty, reasoned runs stream the provider-reported content there, completion follows durable flush, and the persisted Session remains available to later tooling. Its initial user message records `source.kind: 'user'` and therefore carries no ApiProxy `rpcId`.

 ApiProxy carrier coverage stays in the ApiProxy package. Custom one-shot profiles may include Host or Web bundles explicitly, while the shipped profile and the recognized installation-owned tuple are Web-free.
@@ -6,7 +6,7 @@ Status: implemented

 ## 问题

-`headless` 的产品约定是一个本地任务：最终 assistant 文本写入 stdout，退出状态反映成功与否，成功时 stderr 为空，并且不打开监听端口。包含 Workspace Host 服务、ApiProxy、HTTP、Web 运行时或浏览器插件的组合违背这一约定，也使本地完成状态依赖无关的传输树。
+`headless` 的产品约定是一个本地任务：最终 assistant 文本写入 stdout，退出状态反映成功与否，不打开监听端口，并由 [headless 推理进度](../feature/2026-08-21-headless-reasoning-progress.zh.md)负责 stderr 推理投影。包含 Workspace Host 服务、ApiProxy、HTTP、Web 运行时或浏览器插件的组合违背这一约定，也使本地完成状态依赖无关的传输树。

 直接入口仍需要与 Web 所创建 Agent 相同的部署模型状态。独立的提供方／模型默认值会让同一部署产生两种答案，而在 Agent 与会话持久化完全停稳之前推导完成状态，会让 stdout 与退出状态观察到不完整状态。

@@ -14,17 +14,17 @@ Status: implemented

 随附的 `headless` profile 包含 `dsh-base` 与 `dsh-headless`。base 提供默认禁用模块 HMR（热模块替换）的策略；headless 组合包提供自身的 persona 与工具模式、显式挂载 Code Mode worker，并在不覆盖该策略的情况下插入 `headless-runner`。其插件树不包含任何 `@deepseek-ai/dsh-host-*` 包、ApiProxy、HTTP server、Web 运行时或浏览器客户端。Code Mode 与会话持久化均为独立于 Web 呈现的一次性 Agent 能力。

-`headless-runner` 是直接使用核心服务的入口。Loader 完全加载后，它读取 `ctx.agentDefaultModel.currentSelection()`，通过 `ctx.agents.create` 创建一个新的持久化 Agent，在 Agent 作用域中安装该 `ModelSelection`，等待启动工作完全停稳，锚定会话事件序号，提交一条普通用户消息，再次等待完全停稳。随后，它等待 `ctx.sessions.flush`，折叠自身持有的持久事件区间，以取得最后一条非空 assistant 文本和最终 `turn/end` 结束原因，将文本连同一个换行写入 stdout，并且仅在结束原因为 `completed` 时请求启动器以退出状态 0 有界关闭。结束原因为 `error` 时，其持久化错误码与消息写入 stderr；驱动器的意外失败也写入 stderr 并以 1 退出。
+`headless-runner` 是直接使用核心服务的入口。Loader 完全加载后，它读取 `ctx.agentDefaultModel.currentSelection()`，通过 `ctx.agents.create` 创建一个新的持久化 Agent，在 Agent 作用域中安装该 `ModelSelection`，等待启动工作完全停稳，锚定会话事件序号，提交一条普通用户消息，再次等待完全停稳。随后，它等待 `ctx.sessions.flush`，折叠自身持有的持久事件区间，以取得最后一条非空 assistant 文本和最终 `turn/end` 结束原因，将文本连同一个换行写入 stdout，并且仅在结束原因为 `completed` 时请求启动器以退出状态 0 有界关闭。[Headless 推理进度](../feature/2026-08-21-headless-reasoning-progress.zh.md)负责实时 stderr 投影；结束原因为 `error` 时，其持久化错误码与消息写入 stderr，驱动器的意外失败也写入 stderr 并以 1 退出。

 `@deepseek-ai/dsh-agent-default-model` 拥有与传输无关的默认值，供没有会话级选择的 Agent 使用。`AgentDefaultModelConfig` 提供 `ctx.agentDefaultModel` 并注册 `agent-default-model` Settings 分节。组合配置提供 `{provider, model}`，用户设置还可以提供 `reasoningEffort`。`currentSelection()` 返回当前的完整选择，`saveSelection()` 则写入完整分节，因此不含强度的选择会清除已存强度。`dsh-base` 提供组合条目。直接入口与 ApiProxy 入口均消费该服务；只有 ApiProxy 负责会话级优先级、模型校验与已接受 Web 选择的持久化。

 `loadProfile` 识别安装过程拥有的精确 headless 元组（`dsh-base`、`dsh-web-app`、`dsh-headless`），将其规范化为随附的 headless 模板，并保留 manifest（元数据清单）的其他所有字段。带额外项、缺少项或顺序不同的组合包列表归用户所有，保持不变。

-本 Agent Note 负责 headless 的传输与完成约定。[应用持有自己的命令行](2026-08-06-app-owned-command-line.zh.md)负责当前的 `dsh --profile headless` 语法；原 [`dsh run` 决策](../../archived/feature/2026-08-08-dsh-run-headless-command.md)记录已被取代的启动器持有语法，[GUI 分层与 RPC 协议](2026-07-19-gui-layering-and-rpc-protocol.zh.md)负责浏览器网关边界，[Web 配置树启动与传输分层](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)负责 Web 插件树，[默认模型跟随选择器](../feature/2026-08-07-default-model-follows-the-picker.zh.md)负责共享 Agent 默认值的持久化。
+本 Agent Note 负责 headless 的传输与完成约定；[headless 推理进度](../feature/2026-08-21-headless-reasoning-progress.zh.md)负责成功运行时的 stderr 输出。[应用持有自己的命令行](2026-08-06-app-owned-command-line.zh.md)负责当前的 `dsh --profile headless` 语法；原 [`dsh run` 决策](../../archived/feature/2026-08-08-dsh-run-headless-command.md)记录已被取代的启动器持有语法，[GUI 分层与 RPC 协议](2026-07-19-gui-layering-and-rpc-protocol.zh.md)负责浏览器网关边界，[Web 配置树启动与传输分层](2026-07-24-web-config-tree-boot-and-transport-layering.zh.md)负责 Web 插件树，[默认模型跟随选择器](../feature/2026-08-07-default-model-follows-the-picker.zh.md)负责共享 Agent 默认值的持久化。

 ## 验证

-包测试围绕脚本化 Agent 工厂使用真实的会话存储与 Agent 注册表，固定空闲态到空闲态的聚合、延迟异步完成、终止态模型诊断、其他未完成退出、直接失败、Loader 加载期间的 dispose（资源释放），以及退出前 flush 的顺序。组装后的无密钥快照通过回放的工具往返驱动 `dsh --profile headless`，记录一条带 `source.kind: 'user'` 的 `user/message`，并在 stderr 暴露终止态模型失败。构建后二进制验收通过已发布入口访问 mock 提供方，并要求最终文本出现在 stdout、退出状态为 0 且 stderr 为空。配置转储验收排除随附 headless 树中的所有 Host、Web 与 Client 包；PTY 关闭覆盖要求不出现观察行，并在有界时间内完成 dispose。
+包测试围绕脚本化 Agent 工厂使用真实的会话存储与 Agent 注册表，固定空闲态到空闲态的聚合、延迟异步完成、终止态模型诊断、其他未完成退出、直接失败、Loader 加载期间的 dispose（资源释放），以及退出前 flush 的顺序。组装后的无密钥快照通过回放的工具往返驱动 `dsh --profile headless`，记录一条带 `source.kind: 'user'` 的 `user/message`，并在 stderr 暴露推理进度与终止态模型失败。构建后二进制验收通过已发布入口访问 mock DeepSeek 端点，并要求推理流出现在 stderr、最终文本出现在 stdout 且退出状态为 0。配置转储验收排除随附 headless 树中的所有 Host、Web 与 Client 包；PTY 关闭覆盖要求不出现观察行，并在有界时间内完成 dispose。

 ## 考虑过的替代方案

@@ -39,6 +39,6 @@ Status: implemented

 ## 后果

-`dsh --profile headless` 提供本地 Agent 任务，而不是浏览器观察、Host API 或 HTTP。需要这些能力的用户选择 `dsh web`。成功时 stderr 为空，完成结果在持久化 flush 后推导，持久化会话仍可供后续工具使用。初始用户消息记录 `source.kind: 'user'`，因此不携带 ApiProxy `rpcId`。
+`dsh --profile headless` 提供本地 Agent 任务，而不是浏览器观察、Host API 或 HTTP。需要这些能力的用户选择 `dsh web`。没有推理内容的成功运行会保持 stderr 为空，有推理内容的运行则在那里流式输出提供方报告的内容；完成结果在持久化 flush 后推导，持久化会话仍可供后续工具使用。初始用户消息记录 `source.kind: 'user'`，因此不携带 ApiProxy `rpcId`。

 ApiProxy 载体覆盖保留在 ApiProxy 包中。自定义一次性 profile 可以显式包含 Host 或 Web 组合包；随附 profile 与可识别的安装过程所属元组均不含 Web。
@@ -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/architecture/2026-08-10-product-subagent-providers-in-shared-host.md
-2026-08-10-product-subagent-providers-in-shared-host.md: 196e28c1263c4b6d71eaeb59b9ba8457b36f3ff4
-2026-08-10-product-subagent-providers-in-shared-host.zh.md: 1451890e1b250a2095e3366c59d6ce0873b55fe9
+2026-08-10-product-subagent-providers-in-shared-host.md: eca1d5b6b9b0e46b39c6c2ef382bfc261d014e44
+2026-08-10-product-subagent-providers-in-shared-host.zh.md: 374a3ef72e298bb621fd71d306a4fca16c8000b4
@@ -14,9 +14,9 @@ The placement decision must preserve two independent facts. Loading a provider m

 Product providers remain process-scoped host-plane registrations. The [production-install exclusion decision](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.md) supersedes only this note's former base-bundle installation choice: production `dsh-base` neither depends on nor mounts them. A Profile that opts in installs the selected provider Bundle; its patch mounts the default instance, and the Profile may mount additional named instances on the host plane. The [named-instance decision](../feature/2026-08-18-product-subagent-named-instances.md) owns each row's registry identity: both products accept multiple unique `providerName` values while preserving `codex` and `claude-code` as their defaults. Loading either plugin only registers a dormant backend; the corresponding Codex or Claude process starts on the first actual delegation call. Agent Presets independently contribute ordinary `dsh-tool-subagent` rows whose `provider` and `toolName` values expose exactly the configured instances needed by one agent without changing the Host registry.

-Each provider package owns its directly installable Bundle patch and private product runtime. This note continues to own process-wide Host placement whenever either provider is installed. The provider-contract note continues to own each product protocol, result mapping, cancellation, process-tree lifecycle, and evidence tiers. The [Agent Preset architecture](2026-08-03-per-session-agent-presets.md) continues to own the Host/Agent split, preset authoring, and the rule that edits affect only newly composed sessions.
+Each provider package owns its directly installable Bundle patch and private product runtime. This note continues to own process-wide Host placement whenever either provider is installed. The provider-contract note continues to own each product protocol, result mapping, cancellation, process-tree lifecycle, and evidence tiers. The [named-instance decision](../feature/2026-08-18-product-subagent-named-instances.md) owns the optional Claude Code model and other per-instance configuration. The [Agent Preset architecture](2026-08-03-per-session-agent-presets.md) continues to own the Host/Agent split, preset authoring, and the rule that edits affect only newly composed sessions.

-Each Bundle delegates executable selection to its package-owned product runtime: the Codex package runs its declared wrapper, while the Claude Code package lets its pinned Agent SDK select the private native executable. Neither provider consults or falls back to a host product command. Profile loading creates no product state, probes no version or authentication, and may supply each mounted Provider instance's deployment configuration, including the product-specific `permissionMode` values owned by the [non-interactive permissions decision](../feature/2026-08-15-product-subagent-noninteractive-permissions.md), without moving those choices into an Agent Preset or model-facing tool. Missing platform payloads and product failures remain local to the attempted delegation.
+Each Bundle delegates executable selection to its package-owned product runtime: the Codex package runs its declared wrapper, while the Claude Code package lets its pinned Agent SDK select the private native executable. Neither provider consults or falls back to a host product command. Profile loading creates no product state, probes no version or authentication, and may supply each mounted Provider instance's deployment configuration, including an optional opaque model where supported and the product-specific `permissionMode` values owned by the [non-interactive permissions decision](../feature/2026-08-15-product-subagent-noninteractive-permissions.md), without moving those choices into an Agent Preset or model-facing tool. Missing platform payloads and product failures remain local to the attempted delegation.

 ## Verification

@@ -14,9 +14,9 @@ Status: implemented

 产品提供方仍是进程级的 host plane（宿主平面）注册。[生产安装排除决策](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md)只取代本说明原先由 base bundle 安装提供方的选择：生产 `dsh-base` 既不依赖也不挂载它们。选择产品集成的 Profile 会安装目标提供方 Bundle；其 patch 挂载默认实例，而 Profile 可以在 host plane 挂载更多命名实例。[命名实例决策](../feature/2026-08-18-product-subagent-named-instances.zh.md)负责每个配置项的注册身份：两个产品都接受多个唯一的 `providerName`，同时保留 `codex` 与 `claude-code` 作为默认值。加载任一插件只会注册一个休眠后端；对应的 Codex 或 Claude 进程直到第一次实际委派调用时才启动。Agent Preset 通过普通 `dsh-tool-subagent` 配置项的 `provider` 与 `toolName` 准确公开单个 agent 所需的已配置实例，而无需更改 Host 注册表。

-每个提供方包都拥有可直接安装的 Bundle patch 与私有产品运行时。本说明继续负责每个已安装提供方的进程级 Host 放置。提供方约定说明继续负责每个产品的协议、结果映射、取消、进程树生命周期与证据层级。[Agent Preset 架构](2026-08-03-per-session-agent-presets.zh.md)继续负责宿主与 agent 的划分、preset 创作，以及改动只影响新组装会话的规则。
+每个提供方包都拥有可直接安装的 Bundle patch 与私有产品运行时。本说明继续负责每个已安装提供方的进程级 Host 放置。提供方约定说明继续负责每个产品的协议、结果映射、取消、进程树生命周期与证据层级。[命名实例决策](../feature/2026-08-18-product-subagent-named-instances.zh.md)负责可选 Claude Code 模型及其他逐实例配置。[Agent Preset 架构](2026-08-03-per-session-agent-presets.zh.md)继续负责宿主与 agent 的划分、preset 创作，以及改动只影响新组装会话的规则。

-每个 Bundle 都把可执行文件选择交给包自有的产品运行时：Codex 包运行自身声明的 wrapper，Claude Code 包则让锁定的 Agent SDK 选择私有原生可执行文件。两个提供方都不会查询或回退宿主产品命令。加载 Profile 不会创建产品状态、探测版本或测试身份验证；它可以提供每个已挂载 Provider 实例的部署配置，包括由[非交互权限决策](../feature/2026-08-15-product-subagent-noninteractive-permissions.zh.md)负责的产品专属 `permissionMode` 值，但不会把这些选择移入 Agent Preset 或面向模型的工具。平台载荷缺失和产品故障仍局限于发生问题的那次委派。
+每个 Bundle 都把可执行文件选择交给包自有的产品运行时：Codex 包运行自身声明的 wrapper，Claude Code 包则让锁定的 Agent SDK 选择私有原生可执行文件。两个提供方都不会查询或回退宿主产品命令。加载 Profile 不会创建产品状态、探测版本或测试身份验证；它可以提供每个已挂载 Provider 实例的部署配置，包括产品支持时可选的不透明模型，以及由[非交互权限决策](../feature/2026-08-15-product-subagent-noninteractive-permissions.zh.md)负责的产品专属 `permissionMode` 值，但不会把这些选择移入 Agent Preset 或面向模型的工具。平台载荷缺失和产品故障仍局限于发生问题的那次委派。

 ## 验证

@@ -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/architecture/2026-08-10-session-log-version-mechanism.md
-2026-08-10-session-log-version-mechanism.md: dfbe5c1926cf683a34ec6694f188f57c44b9ca10
-2026-08-10-session-log-version-mechanism.zh.md: 00d58757d3ea4bf1689a0847613557613d40ebf6
+2026-08-10-session-log-version-mechanism.md: 81108ceaf23405c8f2def9aaef88505d635808a3
+2026-08-10-session-log-version-mechanism.zh.md: cbb127420e2695853fdc2ad0bb98a7a0bf230b5b
@@ -14,21 +14,13 @@ Session logs must be upgradable after release, and the runtime that ships first

 **The writer decides bumps, not the reader.** A bump is required exactly when an old runtime could no longer handle a new log with full semantic correctness. "Parses without error" is not the bar: silently skipping content that shapes reconstruction is a wrong read. Only structural changes qualify — header shape, event envelope, core event semantics, the surface mechanism (`SurfaceEventType` set, `SurfaceOp` variants). When unsure, bump: a near-identity upgrader is almost free, a missed bump silently corrupts old readers.

-**Read rules by direction.** Equal version: decode normally. Newer than the reader: refuse, name the direction ("written by a newer harness — upgrade"), and point at the raw log artifact so the user can still see the text (`SessionFormatUnsupportedError`, distinct from `SessionPersistenceCorruptionError` because nothing is damaged). Older than the reader: require a complete chain of static n→n+1 `SessionFormatMigration` classes; a missing migration refuses the read and names the gap. The registry is part of the build rather than Cordis composition, so one build has the same durable read capability under every plugin set.
-
-**Format migration is the decoder, not a Coordinator repair branch.** Backends expose parsed durable data as `unknown` through a repeatable `StoredSessionSource`: one raw header, one exact revision, and `readEvents()` factories that create independently consumable `AsyncIterable` streams bound to that revision. Each migration class carries static adjacent `from`/`to` versions. One fresh instance handles one decode attempt: `header()` runs once, `event()` maps each input record to exactly one lossless-JSON output with the same seq, and optional `finish()` validates accumulated state after EOF. Instance fields may retain header and earlier-event facts without sharing state across sessions, concurrent reads, or revision retries. Header-only reads stop after `header()` and never call `finish()`, so that method validates EOF state rather than releasing resources. Any version conversion reads the complete event stream and applies the requested suffix only after all migrations; an equal-version read retains backend suffix seek. The decoder validates each output header version and each migration's seq preservation, then applies current `SessionHeader` and `SessionEvent` validation only after the complete chain.
-
-**A future format bump adds one format-owned migration.** The change adds `format-migrations/vN-to-vN+1.ts`, exports its class from the static `SESSION_FORMAT_MIGRATIONS` array, and increments `SESSION_FORMAT_VERSION`. The migration owns every old header and event variant it accepts, its instance state, and explicit failure for malformed input. It cannot add, remove, reorder, or renumber events: durable references use seq as event identity. A format change that alters facts consumed by a projection increments that projection's `stateVersion`; unchanged projections retain their cache rows. Backends and the Coordinator do not gain version-specific branches. Historical variants that never changed the version remain isolated in the format-v0 compatibility decoder and are not a template for later version migrations. This decoder maps the historical `compact/start`, `compact/summary`, `compact/end`, and `compact/prune` names to canonical `compaction/*` events while preserving the rest of each record.
-
-**Recovery and writeback consume current-format data.** `inspect()` and `readFrom()` decode only in memory. Cold `prepare()`/`load()` first decode the whole source, add the current recovery closers, and replace the exact old revision with that complete balanced current-format stream. Live HMR adoption uses the same replacement primitive after seed verification but does not synthesize closers for a turn still owned by the live Session. A successful replacement or revision conflict discards the prepared object and reopens the stored source before continuing.
-
-**Replacement is an internal backend compare-and-swap.** `replaceStored(expectedRevision, meta, events)` accepts a streaming current-format log and checks storage identity plus the source revision at the commit boundary. JSONL writes and fsyncs a sibling temporary artifact, rechecks the source revision immediately before the atomic replace, atomically replaces the path (using the Windows write-through replacement primitive there), and syncs the parent directory on POSIX; like every other coordinator freshness check, the recheck adds no cross-process writer exclusion — JSONL assumes one live writer per session. SQLite stages the event iterator, then rechecks and replaces the header and event rows in one transaction. A failed commit leaves one complete old or new log; retaining a permanent pre-upgrade copy is a separate recovery policy, not part of the format migration API.
+**Read rules by direction.** Equal version: read normally. Newer than the reader: refuse, name the direction ("written by a newer harness — upgrade"), and point at the raw log artifact so the user can still see the text (`SessionFormatUnsupportedError`, distinct from `SessionPersistenceCorruptionError` because nothing is damaged). Older than the reader: convert in memory through the chain of n→n+1 upgraders for viewing; persist the converted log only when the session is actually continued (atomic temp-file replace, original kept as backup). A step whose upgrader cannot be written is left empty, which cuts off every version at or below it — those degrade to raw-text viewing.

 **A per-event `ignorable` marker covers vocabulary growth, so ordinary event additions never bump the version.** The event vocabulary is decided by which plugins are mounted, which a single version integer cannot describe. A reader meeting an unrecognized event type refuses to interpret the log unless the event carries `ignorable: true` in its envelope. The default is *required*: forgetting the marker over-refuses a resumable session (an inconvenience), while a default of ignorable would make the same mistake silently resume a gutted one (a safety failure). The architecture makes this sound: model-visible content flows only through the three `surfaceOp`-marked surface event types plus the `request/header`/`request/context` folds, so the dangerous unknowns are exactly the non-surface events that change how the rest of the log is read (`session/end-seed` is the existing example).

 ## Consequences

-Format v0 carries direction-aware refusal with the raw-log path; the unknown-event guard against a generated known-vocabulary list (`KNOWN_SESSION_EVENT_TYPES`, emitted by `gen-persistence-catalog` from every `SessionEventMap` merge and kept fresh by `verify-persistence-catalog`); the `ignorable` envelope field accepted by seed validation, both backends (a dedicated SQLite column, `SCHEMA_VERSION` 15), and the BFF wire schema; and the static streaming migration decoder with an empty adjacent-version registry. `SESSION_FORMAT_VERSION` remains 0 until a real v0→v1 step lands. The decoder and backend replacement APIs therefore have direct tests without manufacturing a format bump. Writers do not yet set `ignorable` because no producer needs it. Until a registration surface exists, an out-of-repo plugin's events refuse resume under first-party readers; the refusal is loud rather than silent. The unknown-type guard is read-side only: `appendCore` keeps rejecting retired legacy shapes but does not vocabulary-check new types, because an append-time refusal would stall a live session's durability mid-flight, which costs more than a loud refusal at the log's next load. The JSONL backend additionally refuses a foreign version from the raw header line before validating today's header fields or decoding any event record, so a structurally different future format still reports the upgrade direction instead of "corrupt"; SQLite gates whole-file structure through its own `SCHEMA_VERSION` pragma first.
+What shipped in v0 (release 0812): direction-aware refusal with the raw-log path; the unknown-event guard against a generated known-vocabulary list (`KNOWN_SESSION_EVENT_TYPES`, emitted by `gen-persistence-catalog` from every `SessionEventMap` merge and kept fresh by `verify-persistence-catalog`); the `ignorable` envelope field accepted by seed validation, both backends (a dedicated SQLite column, `SCHEMA_VERSION` 15), and the BFF wire schema. The upgrader chain itself is deferred until the first real v0→v1 step exists to test it against; writers do not yet set `ignorable` (no producer needs it), so `Session.append` gains that surface with its first user. Until a registration surface exists, an out-of-repo plugin's events refuse resume under first-party readers — the pre-release stance accepts that, and the refusal is loud rather than silent. The unknown-type guard is read-side only: `appendCore` keeps rejecting retired legacy shapes but does not vocabulary-check new types, because an append-time refusal would stall a live session's durability mid-flight, which costs more than a loud refusal at the log's next load. The JSONL backend additionally refuses a foreign version from the raw header line before validating this format version's header shape or decoding any event row, so a structurally different future format still reports the upgrade direction instead of "corrupt"; SQLite gates whole-file structure through its own `SCHEMA_VERSION` pragma first.

 ## Alternatives considered

@@ -36,6 +28,3 @@ Format v0 carries direction-aware refusal with the raw-log path; the unknown-eve
 - **Default-ignorable unknown events** — inverts the failure mode of a forgotten marker from visible over-refusal into silent corruption.
 - **Auto-migrating on view** — rewriting the artifact on open turns a read into a destructive write: a converter bug corrupts logs at browse time, and a same-directory older runtime loses access because a newer one merely looked.
 - **Per-plugin runtime registration of known event types** — would make the known set composition-dependent, so a leaner same-version composition would refuse logs a fuller one wrote. The generated repo-wide list keeps same-version reads uniform; out-of-repo plugin events are outside it by construction, and a registration surface for them is deferred until such a consumer exists.
- **Materializing migrations as header and event arrays** — makes the framework proportional to complete log size in memory even when each transformation is record-local. Repeatable revision-bound readers plus one-at-a-time event transforms preserve retry semantics without imposing that allocation.
- **Version-specific conversion in `PersistenceCoordinator`** — mixes format decoding with operation-specific crash recovery and duplicates behavior across inspect, suffix read, cold continuation, and live adoption. The shared decoder produces only current-format data; each consumer retains its own recovery intent.
- **A mandatory permanent backup for every upgrade** — is not needed for atomicity and cannot promise the same physical representation across JSONL and SQLite. Backends may add recovery copies as a separate product policy without changing migrations.
@@ -14,21 +14,13 @@ Session log 在发布后必须能升级格式，而最先发布的运行时决

 **升不升版本由写入方决定，与读取方能力无关。**当且仅当老运行时无法在语义上完全正确地处理新日志时才必须升版本。"解析不报错"不是标准：静默跳过影响重建的内容就是读错。只有结构性变更够得上这条线：header 形状、事件信封、核心事件语义、surface 机制（`SurfaceEventType` 集合、`SurfaceOp` 变体）。拿不准就升：近似恒等的升级器几乎没有成本，漏升一次会让老读取器静默读坏。

-**读取规则按方向区分。**版本相等：正常解码。比读取器新：拒绝，说明方向（"由更新的 harness 写入，请升级"），并给出原始日志文件的路径，用户仍能看到文本（`SessionFormatUnsupportedError`，与 `SessionPersistenceCorruptionError` 区分，因为数据没有损坏）。比读取器旧：要求静态 n→n+1 `SessionFormatMigration` 类组成完整链路，缺失任何 migration 都会拒绝并指出断点。注册表属于 build 而不是 Cordis composition，因此同一个 build 在任何插件组合下都具有相同的持久化读取能力。
-
-**格式迁移就是 decoder，不是 Coordinator 的修复分支。**后端通过可重复读取的 `StoredSessionSource` 把解析后的持久化数据作为 `unknown` 暴露：一个原始 header、一个精确 revision，以及每次产生独立 `AsyncIterable` 且绑定该 revision 的 `readEvents()` factory。每个 migration class 用静态且相邻的 `from`／`to` 标识版本。每次 decode 都创建一个新实例：`header()` 调用一次；`event()` 把每条输入记录映射为一条 seq 相同、可无损表示为 JSON 的输出；可选的 `finish()` 在 EOF 后验证累计状态。实例字段可以保留 header 与之前事件的事实，而不会在 Session、并发读取或 revision retry 之间共享状态。只读 header 时在 `header()` 后结束，绝不调用 `finish()`，因此该方法用于验证 EOF 状态而不是释放资源。只要发生版本转换，就读取完整事件流，并在所有 migration 完成后才应用请求的 suffix；版本相等时仍保留 backend suffix seek。Decoder 验证每一步输出的 header version 和每个 migration 是否保持 seq，完整链路结束后才执行当前 `SessionHeader` 和 `SessionEvent` 校验。
-
-**以后每次 format bump 只增加一个格式 migration。**改动新增 `format-migrations/vN-to-vN+1.ts`，把它的 class 导出到静态 `SESSION_FORMAT_MIGRATIONS` 数组，并递增 `SESSION_FORMAT_VERSION`。Migration 自己负责它接受的所有旧 header 和 event 变体、实例状态，以及对畸形输入的明确失败。它不能增加、删除、重排事件或重编号：持久引用以 seq 作为事件身份。如果格式变化影响了某个 projection 消费的事实，就递增该 projection 的 `stateVersion`；未受影响的 projection 保留 cache 记录。Backend 和 Coordinator 不增加版本特判。没有改变版本号的历史变体继续隔离在 format-v0 compatibility decoder 中，不作为后续版本 migration 的模板。该 decoder 将历史 `compact/start`、`compact/summary`、`compact/end`、`compact/prune` 名称映射为规范的 `compaction/*` 事件，并保留每条记录的其余内容。
-
-**Recovery 和写回只消费当前格式数据。**`inspect()` 和 `readFrom()` 只在内存中解码。Cold `prepare()`/`load()` 先解码完整 source，补充当前 recovery closers，再用完整、平衡的当前格式 stream 替换精确的旧 revision。Live HMR adoption 在 seed 校验后使用同一个 replacement primitive，但不会为仍由 live Session 掌握的 turn 合成 closer。替换成功或 revision 冲突后都会丢弃 prepared object，重新打开持久化 source 后再继续。
-
-**Replacement 是 backend 内部的 compare-and-swap。**`replaceStored(expectedRevision, meta, events)` 接受流式当前格式日志，并在提交边界检查存储身份和 source revision。JSONL 写入并 fsync 同目录临时 artifact，在原子替换路径前立即复核 source revision，然后原子替换（Windows 使用 write-through replacement primitive），并在 POSIX 上同步父目录；与协调器的其他新鲜性检查一样，复核不提供跨进程写者排他——JSONL 假定每个 session 同时只有一个 live writer。SQLite 先暂存 event iterator，再在一个事务中复核并替换 header 与 event rows。提交失败后只会留下完整旧日志或完整新日志；永久保留升级前副本是独立的恢复策略，不属于 format migration API。
+**读取规则按方向区分。**版本相等：正常读。比读取器新：拒绝，说明方向（"由更新的 harness 写入，请升级"），并给出原始日志文件的路径，用户仍能看到文本（`SessionFormatUnsupportedError`，与 `SessionPersistenceCorruptionError` 区分，因为数据没有损坏）。比读取器旧：查看时经 n→n+1 升级器链在内存中逐级转换；只有会话真正被继续时才把转换落盘（临时文件原子替换，原文件留备份）。写不出升级器的那一步留空，这会切断该步及更早所有版本的升级路径，它们降级为只能看原文。

 **逐事件的 `ignorable` 标记吸收词汇表增长，普通的新增事件永远不用升版本。**事件词汇表由挂载了哪些插件决定，单个版本整数描述不了它。读取器遇到不认识的事件类型时拒绝解读日志，除非该事件的信封带 `ignorable: true`。默认为必需：忘写标记的后果是把一个本可恢复的会话拒绝过头（体验问题），而默认可忽略会让同样的疏忽静默恢复出残缺会话（安全事故）。架构保证了这条规则成立：模型可见内容只经三种带 `surfaceOp` 标记的 surface 事件加 `request/header`、`request/context` 折叠进入重建，危险的未知事件恰好是那些不进 surface 但改变日志其余部分解读方式的事件（`session/end-seed` 是现存例子）。

 ## 影响

-Format v0 包含：分方向的拒绝并带原始日志路径；基于生成的已知词汇清单（`KNOWN_SESSION_EVENT_TYPES`，由 `gen-persistence-catalog` 从所有 `SessionEventMap` 声明合并生成，`verify-persistence-catalog` 保证新鲜）的未知事件守卫；`ignorable` 信封字段被种子校验、两个后端（SQLite 专用列，`SCHEMA_VERSION` 15）和 BFF 线上 schema 接受；以及使用空相邻版本注册表的静态流式 migration decoder。`SESSION_FORMAT_VERSION` 保持 0，直到真实 v0→v1 步骤合入。Decoder 和 backend replacement API 因此可以直接测试，不需要制造一次 format bump。写入侧目前不写 `ignorable`，因为还没有生产者需要它。在注册表面出现之前，仓库外插件的事件在第一方读取器下无法恢复会话；拒绝是显式的而非静默的。未知类型守卫只在读取侧生效：`appendCore` 继续拒绝已淘汰的 legacy 形状，但不对新类型做词汇检查，因为写入时拒绝会让活跃会话的持久化中途停摆，代价大于下次加载时的显式拒绝。JSONL 后端还会在校验当前 header 字段、解码任何 event record 之前，直接从原始 header 行拒绝外来版本，因此结构完全不同的未来格式仍会报告升级方向而不是"损坏"；SQLite 则先由自己的 `SCHEMA_VERSION` pragma 把关整个文件的结构。
+v0（0812 发布）交付的内容：分方向的拒绝并带原始日志路径；基于生成的已知词汇清单（`KNOWN_SESSION_EVENT_TYPES`，由 `gen-persistence-catalog` 从所有 `SessionEventMap` 声明合并生成，`verify-persistence-catalog` 保证新鲜）的未知事件守卫；`ignorable` 信封字段被种子校验、两个后端（SQLite 专用列，`SCHEMA_VERSION` 升到 15）和 BFF 线上 schema 接受。升级器链本身推迟到第一个真实的 v0→v1 变更出现、有真实对象可测时再建；写入侧目前不写 `ignorable`（还没有生产者需要它），`Session.append` 的这一表面随第一个使用者一起落地。在注册表面出现之前，仓库外插件的事件在第一方读取器下无法恢复会话，预发布立场接受这一点，而且拒绝是显式的而非静默的。未知类型守卫只在读取侧生效：`appendCore` 继续拒绝已淘汰的 legacy 形状，但不对新类型做词汇检查，因为写入时拒绝会让活跃会话的持久化中途停摆，代价大于下次加载时的显式拒绝。JSONL 后端还会在校验本格式版本的 header 形状、解码任何事件行之前，直接从原始 header 行拒绝外来版本，因此结构完全不同的未来格式仍会报告升级方向而不是"损坏"；SQLite 则先由自己的 `SCHEMA_VERSION` pragma 把关整个文件的结构。

 ## 曾考虑的替代方案

@@ -36,6 +28,3 @@ Format v0 包含：分方向的拒绝并带原始日志路径；基于生成的
 - **未知事件默认可忽略**：把忘写标记的后果从可见的过度拒绝反转成静默损坏。
 - **查看时自动迁移落盘**：打开即改写把读操作变成破坏性写操作，转换器的 bug 会在浏览时损坏日志，同目录的旧版本运行时也会因为新版本只是看了一眼就失去访问能力。
 - **插件运行时注册已知事件类型**：会让已知集依赖插件组合，同版本的精简组合会拒绝完整组合写出的日志。生成的全仓库清单保证同版本读取行为一致；仓库外插件的事件按构造就在清单之外，为它们提供注册表面推迟到真有这样的消费者时再做。
- **把 migration 物化为 header 和 event 数组**：即使每步转换只依赖单条 record，也会让框架内存占用与完整日志大小成正比。可重复、绑定 revision 的 reader 加逐事件转换保留重试语义，又不强制这笔分配。
- **在 `PersistenceCoordinator` 内写版本转换**：会把格式解码和各操作不同的 crash recovery 混在一起，并在 inspect、suffix read、cold continuation 和 live adoption 间复制行为。共享 decoder 只产出当前格式数据，各 consumer 保留自己的 recovery intent。
- **每次升级都强制永久备份**：原子性不依赖永久副本，而且 JSONL 与 SQLite 无法承诺相同的物理表示。Backend 可以把恢复副本作为独立产品策略加入，不需要修改 migration。
@@ -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/2026-07-23-web-permission-and-approval.md
-2026-07-23-web-permission-and-approval.md: 9fba57e37e0d26a6cb40a81e9330ecfaa9e881b9
-2026-07-23-web-permission-and-approval.zh.md: 0a030d60dcf94e83adc41a21aee850d839d1af01
+2026-07-23-web-permission-and-approval.md: f533adc43e7aafe24d09f7998ba8f7336dea3c79
+2026-07-23-web-permission-and-approval.zh.md: 85e70bac153b4687cb46c1953784f184b8311637
@@ -30,4 +30,4 @@ Client-side, `Session` gained `permissions` and `setPermission`, and approval an

 ## Consequences

-Web sessions start confined (`workspace-write` + `ask` by default) and a sandbox-denial escalation reaches the browser as an answerable card; the deployment can widen or narrow the default through `BootHostOptions.sandbox` without touching the assembly. Question answering uses the same registry pattern (ui-user-questions over the question pending table), and Session navigation identifies approval, plan-review, and ordinary question waits before the user opens them. The permission select reads once per mount; live refresh from another client's switch is deferred. Coverage: proxy registry and permission RPC unit suites, session-object and fixture unit suites, the keyless web smoke for fixture-mode approval and preset switching, and real-composition plan-review and question snapshots that pin the pending sidebar status through resolution.
+Web sessions start confined (`workspace-write` + `ask` by default), and a sandbox-denial escalation reaches the browser through the approval channel. The deployment can widen or narrow the default through `BootHostOptions.sandbox` without touching the assembly. Question answering uses the same registry pattern (ui-user-questions over the question pending table), and Session navigation identifies approval, plan-review, and ordinary question waits before the user opens them. The permission select reads once per mount; live refresh from another client's switch is deferred. Coverage includes proxy registry and permission RPC suites, session-object and fixture suites, and the keyless Web smoke for approval answering and preset switching.
@@ -30,4 +30,4 @@ Web 承载层组合与 acp-agent 相同的沙箱化产品路径：`dsh-sandbox-l

 ## 后果

-Web 会话从受限状态启动（默认 `workspace-write` + `ask`），一次沙箱拒绝的升级会以可应答的卡片形式抵达浏览器；部署方可以通过 `BootHostOptions.sandbox` 放宽或收紧默认值，无需触动装配。问题应答使用同一注册表模式（ui-user-questions 基于问题 pending 表），Session 导航会在用户打开会话前识别审批、计划审阅与普通问题等待。权限选择在每次挂载时读取一次；来自另一个 client 切换的实时刷新暂缓实现。覆盖率：proxy 注册表与权限 RPC 的单元测试套件、会话对象与 fixture 的单元测试套件、针对 fixture 模式审批应答与预设切换的无密钥 Web 冒烟测试，以及真实组合的 plan-review 与问题快照；这些快照会固定 pending 侧边栏状态直至解决。
+Web 会话从受限状态启动（默认 `workspace-write` + `ask`），sandbox 拒绝升级会通过审批通道抵达浏览器。部署方可以通过 `BootHostOptions.sandbox` 放宽或收紧默认值，无需触动装配。问题应答使用同一注册表模式（ui-user-questions 基于问题 pending 表），Session 导航会在用户打开会话前识别审批、计划审阅与普通问题等待。权限选择在每次挂载时读取一次；来自另一个 client 切换的实时刷新暂缓实现。覆盖包括 proxy 注册表与权限 RPC 单元测试套件、会话对象与 fixture 单元测试套件，以及针对审批应答与 preset 切换的无密钥 Web 冒烟测试。
@@ -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/2026-07-27-typescript-sdk-and-sdk-subagent-backend.md
-2026-07-27-typescript-sdk-and-sdk-subagent-backend.md: 8bfd6654e449bb762b153a1b0b950a504cf25a4b
-2026-07-27-typescript-sdk-and-sdk-subagent-backend.zh.md: f59126f70fb0b1a88ff874840a49993d4dc2f988
+2026-07-27-typescript-sdk-and-sdk-subagent-backend.md: e1e0d044089d3c7aacdbe98e8556669f28d552c5
+2026-07-27-typescript-sdk-and-sdk-subagent-backend.zh.md: 1d1b5bf5b8ae4d5403eee4d8e98c2ad9b8d0c861
@@ -12,20 +12,20 @@ The stdio JSON-RPC serving surface (`@deepseek-ai/dsh-sdk-jsonrpc-server`, the [

 Three packages, layered exactly like the existing Python stack, plus one Service Provider registration:

- **`@deepseek-ai/dsh-sdk-protocol`** (`packages/sdk/protocol/`) — the wire made shared and nominal. `JsonRpcLineTransport` moves here verbatim from `dsh-sdk-jsonrpc-server` (which now imports it), and `types.ts` names every payload the server speaks: `InitializeParams/Result`, `SessionPromptParams/Result`, the four notification payloads, and the `HarnessSdkRequestMap`/`HarnessSdkNotificationMap` indexes. The package root explicitly exports that complete interface and provides no source-module deep imports. The server's `notify()` call sites are typed against these named payloads, so server drift breaks compilation, not clients. One behavioral change: an error response now rejects with `JsonRpcResponseError` carrying the wire `code`/`data` (the Python client already preserved these; the old transport threw a bare `Error` with only the message).
- **`@deepseek-ai/dsh-sdk-client`** (`packages/sdk/client/`) — the TypeScript twin of `python/sdk`: `HarnessClient` (spawn, frame, fan out notifications, typed error surfaces, close-to-quiescence via the shared dispose ladder) under `DeepSeekHarness`/`HarnessSession` (lazy start, memoized `initialize`, `run()` pairing one `session/prompt` with its `session.finished`). Its package-root consumer interface explicitly exports both client layers, caller-facing types, and the protocol-owned `JsonRpcResponseError`; source modules, normalization helpers, and the notification producer stay internal. `RunResult.events` contains only the root session's typed events, while `notifications` retains session ids across the root and descendants discovered from `subagent.started`; session-tree scoping is client-side, mirroring `client.py`. The launch interface resolves the same-version `@deepseek-ai/dsh` dependency and selects a named profile, with optional `dshBin`, ordered patches, an explicit Harness home, process cwd, environment, and timeouts; arbitrary command/argv launch remains an internal fake-runtime adapter. A clean checkout without `lib/bin.js` uses that package's source entry through an absolute `tsx/esm` loader and an internal patch that omits build-generated Typert contribution loading, which the SDK protocol does not consume. `env` replaces rather than merges and is read when `start()` spawns, so callers own credential policy and can finish preparing it before first use. `RunResult` carries the structured `reason` (Python exposes only `status`); teardown walks a private stdin-EOF → SIGTERM → SIGKILL ladder to actual exit (the client runs outside any harness context, so it cannot ride `ctx.subprocess`).
- **`@deepseek-ai/dsh-subagent-dsh-sdk`** (`packages/subagent/subagent-dsh-sdk/`) — the second out-of-process `SubagentProvider`, structured as `subagent-acp`'s sibling: same all-false capabilities and `inheritsParentContext: false`, same publish-after-handshake ownership transaction, same result-never-rejects flattening through an `onError` sink, same parent-namespace run id. The child answer is read from streamed `session.event`s — the last complete `assistant/message`, else accumulated `text-delta` chunks, so partial answers survive cancellation. Stop reasons map from the child's structured `TurnEndReason` (`completed`/`max-tokens`/`aborted` pass through; everything else, including a settled-without-turn child, is `error`). Its `dshBin`/profile/patch/home config selects an isolated SDK application, `provider`/`model` feeds the child's `initialize`, and `env` supplies explicit child-only values such as its API key.
+- **`@deepseek-ai/dsh-sdk-protocol`** (`packages/sdk/protocol/`) — the wire made shared and nominal. `JsonRpcLineTransport` lives here, and `types.ts` names every payload the server speaks: `InitializeParams/Result`, `SessionPromptParams/Result`, the four notification payloads, and the `HarnessSdkRequestMap`/`HarnessSdkNotificationMap` indexes. `InitializeParams` carries provider, model, optional adapter-owned reasoning effort, and optional output cap. The package root explicitly exports that complete interface and provides no source-module deep imports. The server's `notify()` call sites are typed against these named payloads, so server drift breaks compilation, not clients. Error responses reject with `JsonRpcResponseError` carrying the wire `code`/`data`, matching the Python client.
+- **`@deepseek-ai/dsh-sdk-client`** (`packages/sdk/client/`) — the TypeScript twin of `python/sdk`: `HarnessClient` (spawn, frame, fan out notifications, typed error surfaces, close-to-quiescence via the shared dispose ladder) under `DeepSeekHarness`/`HarnessSession` (lazy start, memoized `initialize`, `run()` pairing one `session/prompt` with its owned activity). Its package-root consumer interface explicitly exports both client layers, caller-facing types, and the protocol-owned `JsonRpcResponseError`; source modules, normalization helpers, and the notification producer stay internal. `RunResult.events` contains only the root session's typed events, while `notifications` retains session ids across the root and descendants discovered from `subagent.started`; session-tree scoping is client-side, mirroring `client.py`. The launch interface resolves the same-version `@deepseek-ai/dsh` dependency and selects a named profile, with optional `dshBin`, ordered patches, an explicit Harness home, process cwd, environment, and timeouts; arbitrary command/argv launch remains an internal fake-runtime adapter. `initialize` carries provider, model, optional reasoning effort, and optional output cap. A clean checkout without `lib/bin.js` uses that package's source entry through an absolute `tsx/esm` loader and an internal patch that omits build-generated Typert contribution loading, which the SDK protocol does not consume. `env` replaces rather than merges and is read when `start()` spawns, so callers own credential policy and can finish preparing it before first use. Teardown walks a private stdin-EOF → SIGTERM → SIGKILL ladder to actual exit because the client runs outside any harness context.
+- **`@deepseek-ai/dsh-subagent-dsh-sdk`** (`packages/subagent/subagent-dsh-sdk/`) — the second out-of-process `SubagentProvider`, structured as `subagent-acp`'s sibling but advertising `agentOptions: true`: each run merges provider/model/reasoning/maxTokens over instance defaults and sends only those fields through the child `initialize`. Other start capabilities remain false, and `inheritsParentContext: false`. The provider retains the same publish-after-handshake ownership transaction, result-never-rejects flattening through an `onError` sink, and parent-namespace run id. The child answer is read from streamed `session.event`s — the last complete `assistant/message`, else accumulated `text-delta` chunks, so partial answers survive cancellation. Stop reasons map from the child's structured `TurnEndReason` (`completed`/`max-tokens`/`aborted` pass through; everything else, including a settled-without-turn child, is `error`). Its `dshBin`/profile/patch/home config selects an isolated SDK application, while `env` supplies explicit child-only values such as its API key.
 - **The subagent seam grows `out-of-process.ts`**: the provider-side vocabulary both out-of-process backends share — `NO_START_CAPABILITIES`, timing-bound validation, child cwd resolution (config override, else the delegating parent session's workspace), the never-reject `settleRunResult`, and the `subprocessRunHandle` publication. Process mechanics (spawn, env scrub, tree-scoped teardown) live in the `dsh-subprocess` seam; `subagent-acp` spawns through `ctx.subprocess`, while this backend spawns through the SDK client (the subprocess README's documented exception for SDK-managed transports) and applies the seam's `scrubbedParentEnv()` itself.

-`dsh-sdk-jsonrpc-server` keeps serving unchanged (the wire is byte-identical). TypeScript and Python clients both consume the shared protocol through `dsh --profile sdk`; the Python wheel packages that CLI and its closed dependency tree.
+`dsh-sdk-jsonrpc-server` validates the exact provider/model/effort route during `initialize`, stores only explicitly supplied effort and token values, and creates every SDK root Agent from that fixed process-wide route. Because JSON-RPC requests can dispatch concurrently, it rejects `session/prompt` until one initialization has completed successfully, preventing pending or invalid routes from falling back to constructor defaults. TypeScript and Python clients both expose the same initialization fields through `dsh --profile sdk`; the Python wheel packages that CLI and its closed dependency tree.

 ## Testing

 Four tiers, per [testing policy](../../../../docs/testing.md):

 - **Keyless unit** — `sdk-client` drives a scripted fake runtime (`tests/fake-runtime.ts`, env-scripted, protocol-only — the Python `test_client.py` pattern) over real stdio; `subagent-dsh-sdk` drives the same fake through the real provider. 100% per-file coverage on all three packages.
- **Keyless Loader composition** — `subagent-dsh-sdk/tests/loader-composition.e2e.ts` boots its package-owned test composition (`packages/subagent/subagent-dsh-sdk/tests/fixtures/loader/`), where the child is a real second `dsh --profile sdk` runtime with its own isolated home and ordered patch; asserts the parent tool result and the child's own persisted transcript both carry the parent session's cwd.
- **Keyless snapshot** — `snapshots/sdk/sdk.snapshot.ts` drives the real `dsh --profile sdk` runtime through the real `dsh-sdk-client`, replaying recorded fixtures through an ordered `llm-replay` patch. Each scenario pins the normalized notification stream, SDK turn result, and persisted parent and child logs. This also closes the protocol-tier gap the single-exe note's Python-side snapshot left on the vitest side.
+- **Keyless Loader composition** — `subagent-dsh-sdk/tests/loader-composition.e2e.ts` boots its package-owned test composition (`packages/subagent/subagent-dsh-sdk/tests/fixtures/loader/`), where the child is a real second `dsh --profile sdk` runtime with its own isolated home and ordered patch; the tool result and the child's persisted request header both prove provider, model, reasoning effort, maxTokens, and parent-session cwd.
+- **Keyless snapshot** — `snapshots/sdk/sdk.snapshot.ts` drives the real `dsh --profile sdk` runtime through the real `dsh-sdk-client`, replaying recorded fixtures through ordered `llm-replay` patches. The DSH SDK scenario uses deterministic parent and child adapters to pin a model-selected route through the delegation tool, a second SDK runtime, and the child's persisted request header; every scenario pins the normalized notification stream, SDK turn result, and applicable parent and child logs.
 - **With-key e2e** — the snapshot suite's `DSH_SNAPSHOT=record` mode is the live-API path (it produced the committed fixtures); the composition e2e needs no key by design.

 ## Alternatives considered
@@ -12,20 +12,20 @@ stdio JSON-RPC 对外服务接口（`@deepseek-ai/dsh-sdk-jsonrpc-server`，见[

 三个包，分层与既有 Python 栈完全一致，外加一个 Service Provider 注册：

- **`@deepseek-ai/dsh-sdk-protocol`**（`packages/sdk/protocol/`）—— 把线协议做成共享且具名。`JsonRpcLineTransport` 从 `dsh-sdk-jsonrpc-server` 原样移入（后者现在导入它），`types.ts` 为服务器所说的每个载荷命名：`InitializeParams/Result`、`SessionPromptParams/Result`、四个通知载荷，以及 `HarnessSdkRequestMap`/`HarnessSdkNotificationMap` 索引。该包根显式导出这一完整接口，且不提供指向源模块的深层导入。服务器的 `notify()` 调用点以这些具名载荷标注类型，服务器漂移会先破坏编译而不是破坏客户端。一处行为变化：错误响应现在以携带线上 `code`/`data` 的 `JsonRpcResponseError` 拒绝（Python 客户端本就保留这些；旧传输只抛携带消息的裸 `Error`）。
- **`@deepseek-ai/dsh-sdk-client`**（`packages/sdk/client/`）—— `python/sdk` 的 TypeScript 孪生：`HarnessClient`（spawn、分帧、通知扇出、有类型的错误表面、经共享 dispose（资源释放）阶梯关闭至完全停稳）之上是 `DeepSeekHarness`/`HarnessSession`（惰性启动、记忆化 `initialize`、`run()` 把一个 `session/prompt` 与其 `session.finished` 配对）。其包根消费方接口显式导出两层客户端、面向调用方的类型，以及协议包所拥有的 `JsonRpcResponseError`；源模块、规范化辅助函数和通知投递端都保留为内部实现。`RunResult.events` 只包含根会话的类型化事件，而 `notifications` 则保留根会话及从 `subagent.started` 发现的后代各自的会话 id；基于 `subagent.started` 血缘边的会话树范围限定在客户端完成，镜像 `client.py`。启动接口解析同版本 `@deepseek-ai/dsh` 依赖并选择具名 profile，可选配置包括 `dshBin`、有序 patch、显式 Harness home、进程 cwd、环境和超时；任意 command/argv 启动只作为内部 fake-runtime 适配器。干净 checkout 中若不存在 `lib/bin.js`，client 会通过绝对 `tsx/esm` loader 使用该包的源码入口，并应用一个省略构建期生成 Typert 贡献加载的内部 patch；SDK 协议不消费这些贡献。`env` 整体替换而非合并，并在 `start()` spawn 时读取，因此凭据策略归调用方，且调用方可在首次使用前完成环境准备。`RunResult` 携带结构化 `reason`（Python 只暴露 `status`）；拆除走私有的 stdin-EOF → SIGTERM → SIGKILL 阶梯直到真正退出（client 运行在任何 harness 上下文之外，无法搭乘 `ctx.subprocess`）。
- **`@deepseek-ai/dsh-subagent-dsh-sdk`**（`packages/subagent/subagent-dsh-sdk/`）—— 第二个进程外 `SubagentProvider`，采用与 `subagent-acp` 对等的结构：同样的全 false 能力与 `inheritsParentContext: false`，同样的握手后发布所有权事务，同样通过 `onError` sink 将结果归一为绝不拒绝，同样的父命名空间 run id。子答案从流式 `session.event` 读取——最后一条完整 `assistant/message`，否则累积的 `text-delta` 块，部分答案在取消时得以保留。停止原因由子进程的结构化 `TurnEndReason` 映射（`completed`/`max-tokens`/`aborted` 直通；其余一切、包括未运行任何轮次便已结束的子进程，都是 `error`）。其 `dshBin`／profile／patch／home 配置选择隔离的 SDK 应用，`provider`／`model` 写入子进程 `initialize`，`env` 则提供子进程专用的显式值，例如其 API key。
+- **`@deepseek-ai/dsh-sdk-protocol`**（`packages/sdk/protocol/`）—— 把协议格式做成共享且具名。`JsonRpcLineTransport` 位于此处，`types.ts` 为服务器所说的每个载荷命名：`InitializeParams/Result`、`SessionPromptParams/Result`、四个通知载荷，以及 `HarnessSdkRequestMap`/`HarnessSdkNotificationMap` 索引。`InitializeParams` 携带提供方、模型、可选且由适配器持有的推理强度，以及可选输出上限。该包根显式导出完整接口，且不提供指向源模块的深层导入。服务器的 `notify()` 调用点以这些具名载荷标注类型，服务器漂移会先破坏编译而不是破坏客户端。错误响应以携带协议 `code`/`data` 的 `JsonRpcResponseError` 拒绝，与 Python 客户端一致。
+- **`@deepseek-ai/dsh-sdk-client`**（`packages/sdk/client/`）—— `python/sdk` 的 TypeScript 孪生：`HarnessClient`（spawn、分帧、通知扇出、有类型的错误表面、经共享 dispose（资源释放）阶梯关闭至完全停稳）之上是 `DeepSeekHarness`/`HarnessSession`（惰性启动、记忆化 `initialize`、`run()` 持有一次完整活动区间）。其包根消费方接口显式导出两层客户端、面向调用方的类型，以及协议包所拥有的 `JsonRpcResponseError`；源模块、规范化辅助函数和通知投递端都保留为内部实现。`RunResult.events` 只包含根会话的类型化事件，而 `notifications` 则保留根会话及从 `subagent.started` 发现的后代各自的会话 id；基于 `subagent.started` 血缘边的会话树范围限定在客户端完成，镜像 `client.py`。启动接口解析同版本 `@deepseek-ai/dsh` 依赖并选择具名 profile，可选配置包括 `dshBin`、有序 patch、显式 Harness home、进程 cwd、环境和超时；任意 command/argv 启动只作为内部 fake-runtime 适配器。`initialize` 携带提供方、模型、可选推理强度与可选输出上限。干净 checkout 中若不存在 `lib/bin.js`，client 会通过绝对 `tsx/esm` loader 使用该包的源码入口，并应用一个省略构建期生成 Typert 贡献加载的内部 patch；SDK 协议不消费这些贡献。`env` 整体替换而非合并，并在 `start()` spawn 时读取，因此凭据策略归调用方，且调用方可在首次使用前完成环境准备。拆除走私有的 stdin-EOF → SIGTERM → SIGKILL 阶梯直到真正退出，因为 client 运行在任何 harness 上下文之外。
+- **`@deepseek-ai/dsh-subagent-dsh-sdk`**（`packages/subagent/subagent-dsh-sdk/`）—— 第二个进程外 `SubagentProvider`，采用与 `subagent-acp` 对等的结构，但声明 `agentOptions: true`：每次运行都会把提供方／模型／推理强度／maxTokens 合并到实例默认值之上，并且只把这些字段送入子进程 `initialize`。其他启动能力保持 false，`inheritsParentContext: false`。提供方保留握手后发布所有权事务、通过 `onError` sink 将结果归一为绝不拒绝，以及父命名空间 run id。子答案从流式 `session.event` 读取——最后一条完整 `assistant/message`，否则累积的 `text-delta` 块，部分答案在取消时得以保留。停止原因由子进程的结构化 `TurnEndReason` 映射（`completed`/`max-tokens`/`aborted` 直通；其余一切、包括未运行任何轮次便已结束的子进程，都是 `error`）。其 `dshBin`／profile／patch／home 配置选择隔离的 SDK 应用，`env` 则提供子进程专用的显式值，例如其 API key。
 - **subagent seam 新增 `out-of-process.ts`**：两个进程外后端共享的 provider 侧词汇——`NO_START_CAPABILITIES`、时限校验、子进程 cwd 解析（配置覆盖、否则发起委托的父会话工作区）、绝不拒绝的 `settleRunResult`、以及 `subprocessRunHandle` 发布。进程机制（spawn、环境清理、进程树清理）属于 `dsh-subprocess` seam；`subagent-acp` 经 `ctx.subprocess` spawn 子进程，本后端则经 SDK 客户端 spawn 子进程（subprocess README 记载的 SDK 托管传输例外）并自行应用该 seam 的 `scrubbedParentEnv()`。

-`dsh-sdk-jsonrpc-server` 的服务不变（协议字节完全一致）。TypeScript 与 Python 客户端都通过 `dsh --profile sdk` 消费共享协议；Python wheel 会打包该 CLI 及其封闭依赖树。
+`dsh-sdk-jsonrpc-server` 会在 `initialize` 期间校验确切的提供方／模型／推理强度路由，只保存显式提供的推理强度与 token 值，并使用这条固定的进程级路由创建每个 SDK 根 Agent。由于 JSON-RPC 请求可能并发分派，它会在一次初始化成功完成前拒绝 `session/prompt`，避免待定或非法路由回退到构造期默认值。TypeScript 与 Python 客户端都通过 `dsh --profile sdk` 公开同一组初始化字段；Python wheel 会打包该 CLI 及其封闭依赖树。

 ## 测试

 四层，依[测试政策](../../../../docs/testing.zh.md)：

 - **免密钥单元**——`sdk-client` 通过真实 stdio 驱动脚本化伪运行时（`tests/fake-runtime.ts`，环境变量脚本化、纯协议——即 Python `test_client.py` 的模式）；`subagent-dsh-sdk` 经真实提供方驱动同一伪运行时。三个包全部 100% 逐文件覆盖。
- **免密钥 Loader 组合**——`subagent-dsh-sdk/tests/loader-composition.e2e.ts` 启动包自有测试组合（`packages/subagent/subagent-dsh-sdk/tests/fixtures/loader/`），其中子进程是真实的第二个 `dsh --profile sdk` 运行时，拥有独立 home 与有序 patch；断言父工具结果与子进程自己持久化的 transcript（文本记录）都携带父会话 cwd。
- **免密钥快照**——`snapshots/sdk/sdk.snapshot.ts` 通过真实 `dsh-sdk-client` 驱动真实 `dsh --profile sdk` 运行时，并通过有序 `llm-replay` patch 回放已录制 fixture（测试前置数据）。每个场景都钉住规范化通知流、SDK 轮次结果，以及持久化的父日志与子日志。这也补上了单文件可执行 Note 的 Python 侧快照在 vitest 侧留下的协议层缺口。
+- **免密钥 Loader 组合**——`subagent-dsh-sdk/tests/loader-composition.e2e.ts` 启动包自有测试组合（`packages/subagent/subagent-dsh-sdk/tests/fixtures/loader/`），其中子进程是真实的第二个 `dsh --profile sdk` 运行时，拥有独立 home 与有序 patch；工具结果与子进程持久化请求 header 都会证明提供方、模型、推理强度、maxTokens 与父会话 cwd。
+- **免密钥快照**——`snapshots/sdk/sdk.snapshot.ts` 通过真实 `dsh-sdk-client` 驱动真实 `dsh --profile sdk` 运行时，并通过有序 `llm-replay` patch 回放已录制 fixture（测试前置数据）。DSH SDK 场景使用确定性的父级和子级适配器，把模型选择的路由固定在委派工具、第二个 SDK 运行时及子级持久化请求 header 中；每个场景都钉住规范化通知流、SDK 轮次结果和适用的父子日志。
 - **带密钥 e2e**——快照套件的 `DSH_SNAPSHOT=record` 模式即真实 API 路径（已提交 fixture 由它产出）；组合 e2e 设计上无需密钥。

 ## 考虑过的替代方案
@@ -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/2026-07-28-sdk-max-output-tokens.md
-2026-07-28-sdk-max-output-tokens.md: 72a9e87484ca87e0a750f52d7e46db7aee436d21
-2026-07-28-sdk-max-output-tokens.zh.md: 820008ec7293cfee20c0a9c26037f746c9e081c2
+2026-07-28-sdk-max-output-tokens.md: 1d2915b7f7169b0784c648aad5900a85fac4c977
+2026-07-28-sdk-max-output-tokens.zh.md: ba59f745bc921d3cc0d5c01f83808dd495220bc7
@@ -14,7 +14,7 @@ The high-level SDKs expose one optional process-wide output cap: Python names it

 Each SDK-created root Agent receives the cap through `AgentOptions.maxTokens`. Agent Loop places that value in the initial `LlmCallConfig`; final call preparation preserves the explicit value or materializes an exact-model adapter default, logs the effective cap in the request header, and reconstructs every dispatched conversation request from that durable header. Omitting the SDK option therefore allows the selected adapter or provider route default to apply.

-In-process subagents inherit the parent's provider, model, and output cap. An explicit `SubagentStartRequest.agentOptions.maxTokens`, including one configured by `dsh-tool-subagent`, overrides the inherited value for that child and its descendants. Out-of-process providers own the configuration of their separate runtime; `subagent-dsh-sdk` therefore exposes its own optional `maxTokens` and forwards it through that child runtime's SDK handshake.
+In-process subagents inherit the parent's provider, model, and output cap. An explicit `SubagentStartRequest.agentOptions.maxTokens`, including one configured by `dsh-tool-subagent`, overrides the inherited value for that child and its descendants. `subagent-dsh-sdk` owns a separate runtime per run: request `maxTokens` overrides its optional instance default, and the resolved cap crosses that child runtime's SDK handshake.

 Compaction, session-title generation, web search, and other auxiliary calls keep their independently owned output limits. `maxTokensAsSuccess` remains outcome mapping only: it does not set or alter the cap.

@@ -30,4 +30,4 @@ Compaction, session-title generation, web search, and other auxiliary calls keep

 SDK callers can bound model output without editing Cordis composition, and direct Agent creation uses the same validated `AgentOptions` contract. The cap is visible in durable request headers and reaches provider adapters as `GenerateOptions.maxTokens`; DeepSeek serialization maps it to `max_tokens`.

-One SDK runtime has one default cap. A caller needing different caps runs separate runtime instances or explicitly overrides an in-process child through its agent options. Reaching the cap still produces the existing `max-tokens` stop reason, whose `ok` or `error` mapping remains deployment policy.
+One SDK runtime has one default cap. A caller needing different caps runs separate runtime instances or uses a subagent provider that advertises `agentOptions`; DSH SDK naturally creates one such runtime per child run. Reaching the cap still produces the existing `max-tokens` stop reason, whose `ok` or `error` mapping remains deployment policy.
@@ -14,7 +14,7 @@ Python 与 TypeScript SDK 可以选择提供方和模型，却无法限制对话

 每个由 SDK 创建的根 Agent 都通过 `AgentOptions.maxTokens` 获得该上限。agent loop（智能体循环）将它放入初始 `LlmCallConfig`；最终调用准备会保留显式值，或填入确切模型的适配器默认值，再将生效上限记录到请求 header，并从该持久化 header 重建每次分派的对话请求。因此，省略 SDK 选项时会应用所选适配器或提供方路由的默认值。

-进程内 subagent 继承父级的提供方、模型和输出上限。显式的 `SubagentStartRequest.agentOptions.maxTokens`（包括通过 `dsh-tool-subagent` 配置的值）会覆盖该子级及其后代的继承值。进程外提供方自行持有其独立运行时的配置；因此 `subagent-dsh-sdk` 公开独立的可选 `maxTokens`，并通过该子运行时自己的 SDK 握手传入。
+进程内 subagent 继承父级的提供方、模型和输出上限。显式的 `SubagentStartRequest.agentOptions.maxTokens`（包括通过 `dsh-tool-subagent` 配置的值）会覆盖该子级及其后代的继承值。`subagent-dsh-sdk` 为每次运行持有独立运行时：请求 `maxTokens` 会覆盖可选的实例默认值，解析后的上限再经过该子运行时自己的 SDK 握手。

 压缩、会话标题生成、网页搜索和其他辅助调用继续使用各自持有的独立输出上限。`maxTokensAsSuccess` 仍然只负责结果映射，不会设置或改变上限。

@@ -30,4 +30,4 @@ Python 与 TypeScript SDK 可以选择提供方和模型，却无法限制对话

 SDK 调用方无需修改 Cordis 组合即可限制模型输出，直接创建 Agent 也使用同一套经过校验的 `AgentOptions` 约定。该上限在持久化请求 header 中可见，并以 `GenerateOptions.maxTokens` 到达提供方适配器；DeepSeek 序列化会将其映射为 `max_tokens`。

-一个 SDK 运行时只有一个默认上限。需要不同上限的调用方应运行独立的运行时实例，或通过 agent options 显式覆盖某个进程内子级。达到上限时仍产生现有的 `max-tokens` 停止原因；将其映射为 `ok` 还是 `error` 仍由部署策略决定。
+一个 SDK 运行时只有一个默认上限。需要不同上限的调用方应运行独立的运行时实例，或使用声明 `agentOptions` 的 subagent 提供方；DSH SDK 会自然地为每次子级运行创建一个这样的运行时。达到上限时仍产生现有的 `max-tokens` 停止原因；将其映射为 `ok` 还是 `error` 仍由部署策略决定。
@@ -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/2026-07-29-persistent-bash-str-replace-editor.md
-2026-07-29-persistent-bash-str-replace-editor.md: e8e37b7e534773429a9c6fe0f63bb8d5460de364
-2026-07-29-persistent-bash-str-replace-editor.zh.md: 71034ba615e09e09ec03212b6d5535959df73f4a
+2026-07-29-persistent-bash-str-replace-editor.md: 1cab6e1b37a642dcbb07c4006a8c7851a8cf592c
+2026-07-29-persistent-bash-str-replace-editor.zh.md: 0edaa0c4d594a94d2860c552504d12f3cc7fdc63
@@ -12,7 +12,7 @@ Some deployments need a one-call Bash schema whose shell state survives across m

 `@deepseek-ai/dsh-tool-bash-persistent` consumes `ctx.terminals` and registers one `bash(command)` tool. It lazily creates one interactive shell per exact Agent and serializes that owner's calls. Cwd, exported variables, activated environments, functions, and background jobs persist. Random private markers delimit command output. Retained scrollback is paged backward to recover the command's original prefix; a dropped prefix is reported explicitly. A nonzero wrapped command appends `[exit code: N]`; a shell that dies before reporting that status instead appends `[shell exited: code N]`, `[shell killed by signal: SIG]`, or `[shell exited]` when the backend supplies neither. `maxOutputChars` bounds retained command output, while fixed diagnostics can extend the returned string. Timeout or cancellation closes the shell before another call can reuse uncertain state, and model-visible timeout/exit results disclose that reset. Cancellation always resets and discards the result, even when a complete status marker is already observable, so state changes the model never saw cannot survive. The configurable description defaults to persistence facts only, so network and package-mirror claims remain deployment-owned.

-`@deepseek-ai/dsh-tool-str-replace-editor` independently consumes `ctx.fs` and registers `str_replace_editor` with `view`, `create`, `str_replace`, and `insert`. It provides numbered text views, filtered two-level directory listings, unique literal replacement, canonical insertion boundaries, and bounded output. Paths are absolute; file views preserve content tabs so copied text remains valid literal replacement input; mutations preserve tabs outside the requested edit; and the public schema and failures use only `old_str`. The plugin can compose with persistent Bash, one-shot Bash, sandboxed Bash, or no shell.
+`@deepseek-ai/dsh-tool-str-replace-editor` independently consumes `ctx.fs` and registers `str_replace_editor` with `view`, `create`, `str_replace`, and `insert`. It provides numbered text views, filtered two-level directory listings, unique literal replacement, canonical insertion boundaries, and bounded output. Paths are absolute; file views preserve content tabs so copied text remains valid literal replacement input; mutations preserve tabs outside the requested edit; and the public schema and failures use only `old_str`. Command-specific fields accept `null` placeholders: execution treats them as omitted when the selected command does not use them, preserves required-field checks, treats `view_range: null` as a full view, and rejects `str_replace.new_str: null` so only omission requests deletion. The plugin can compose with persistent Bash, one-shot Bash, sandboxed Bash, or no shell.

 `dsh-system-prompt` accepts `includeHarnessIdentity: false`, while `dsh-agent-spine-demo` forwards that setting and accepts `toolBash: false`. A deployment can therefore own an exact persona and replace the spine's native Bash without duplicate prompt or tool registrations. Existing defaults remain unchanged.

@@ -30,6 +30,8 @@ The shipped [`minimal` agent preset](../../../../packages/preset/agent-presets/p

 **Modify native read/write/edit.** Rejected because it would distort their general-purpose contracts instead of adding an independently composable editor.

+**Reject every present `null` command field.** Rejected because model-generated calls may serialize placeholders for optional fields that the selected command does not use. The selected command still rejects `null` for required fields and for the deletion-sensitive `str_replace.new_str` field.
+
 ## Consequences

-Profiles can reproduce an external agent by configuring persona and descriptions while the underlying packages remain general. Persistent Bash requires an owning Agent and real PTY backend. Shell exit, timeout, or cancellation loses state. The editor delegates security and mutation policy to the mounted filesystem stack. A minimal Web agent retains Web permissions but must close its persistent shell before changing modes. Runtime-wheel consumers still need no Node installation; Linux wheels contain one executable, while macOS wheels also contain its private native helper.
+Profiles can reproduce an external agent by configuring persona and descriptions while the underlying packages remain general. Persistent Bash requires an owning Agent and real PTY backend. Shell exit, timeout, or cancellation loses state. The editor delegates security and mutation policy to the mounted filesystem stack. Nullable branches increase the command-specific fields' schema cost so unused placeholders do not force retries; execution keeps the selected command's required and deletion semantics explicit. A minimal Web agent retains Web permissions but must close its persistent shell before changing modes. Runtime-wheel consumers still need no Node installation; Linux wheels contain one executable, while macOS wheels also contain its private native helper.
@@ -12,7 +12,7 @@ Status: implemented

 `@deepseek-ai/dsh-tool-bash-persistent` 消费 `ctx.terminals` 并注册一个 `bash(command)` 工具。它为每个精确 Agent 惰性创建一个交互式 shell，并串行化该所有者的调用。Cwd、导出的变量、已激活环境、函数和后台任务会保留。随机私有标记划分命令输出；保留的 scrollback 会向前分页，以恢复命令真正的输出前缀，若前缀已被丢弃则明确告知。经封装的命令以非零状态结束时，会追加 `[exit code: N]`；若 shell 在报告该状态前终止，则改为追加 `[shell exited: code N]`、`[shell killed by signal: SIG]`，或在后端既未提供退出码也未提供信号时追加 `[shell exited]`。`maxOutputChars` 限制保留的命令输出，而固定诊断可能使返回字符串更长。超时或取消会先关闭 shell，避免下一次调用复用状态不确定的会话，模型可见的超时／退出结果也会说明该重置。取消始终会重置 shell 并丢弃结果，即使已经能观察到完整状态标记也是如此，从而不会让模型未曾看到的状态变更得以保留。可配置描述默认只声明持久性事实，因此网络和软件包镜像等声明仍归部署所有。

-`@deepseek-ai/dsh-tool-str-replace-editor` 独立消费 `ctx.fs`，注册包含 `view`、`create`、`str_replace` 与 `insert` 的 `str_replace_editor`。它提供带行号文本查看、过滤后的两层目录列表、唯一字面量替换、规范插入边界和有界输出。路径必须为绝对路径；文件查看会保留内容中的制表符，因此复制的文本仍可作为有效的字面量替换输入；变更会保留请求编辑范围之外的制表符；公开 schema 与错误则只使用 `old_str`。它可以与持久 Bash、一次性 Bash、沙箱 Bash 或无 shell 组合。
+`@deepseek-ai/dsh-tool-str-replace-editor` 独立消费 `ctx.fs`，注册包含 `view`、`create`、`str_replace` 与 `insert` 的 `str_replace_editor`。它提供带行号文本查看、过滤后的两层目录列表、唯一字面量替换、规范插入边界和有界输出。路径必须为绝对路径；文件查看会保留内容中的制表符，因此复制的文本仍可作为有效的字面量替换输入；变更会保留请求编辑范围之外的制表符；公开 schema 与错误则只使用 `old_str`。命令专属字段接受 `null` 占位参数：当前命令不使用该字段时，执行会将其视为未提供；必填检查保持不变；`view_range: null` 表示查看完整文件；`str_replace.new_str: null` 会被拒绝，只有省略该字段才表示删除。它可以与持久 Bash、一次性 Bash、沙箱 Bash 或无 shell 组合。

 `dsh-system-prompt` 接受 `includeHarnessIdentity: false`；`dsh-agent-spine-demo` 会转发该设置，并接受 `toolBash: false`。因此部署可以拥有精确 persona，并替换 spine 的原生 Bash，而不会重复注册提示词或工具。既有默认值不变。

@@ -30,6 +30,8 @@ Status: implemented

 **修改原生 read/write/edit。** 被拒绝，因为这会扭曲其通用约定，而不是增加一个可独立组合的编辑器。

+**拒绝每个已提供的 `null` 命令字段。** 被拒绝，因为模型生成的调用可能为当前命令不使用的可选字段序列化占位参数。当前命令仍会拒绝必填字段以及对删除操作有影响的 `str_replace.new_str` 字段为 `null`。
+
 ## 后果

-Profile 可以通过配置 persona 和描述复现外部 Agent，而底层包保持通用。持久 Bash 需要拥有它的 Agent 与真实 PTY 后端；shell 退出、超时或取消会丢失状态。编辑器把安全与变更策略委托给挂载的文件系统栈。minimal Web agent 保留 Web 权限，但必须先关闭持久 shell 才能更改权限模式。运行时 wheel 包的消费方仍无需安装 Node；Linux wheel 包包含一个可执行文件，macOS wheel 包还包含其私有原生 helper。
+Profile 可以通过配置 persona 和描述复现外部 Agent，而底层包保持通用。持久 Bash 需要拥有它的 Agent 与真实 PTY 后端；shell 退出、超时或取消会丢失状态。编辑器把安全与变更策略委托给挂载的文件系统栈。可为 `null` 的分支增加了命令专属字段的 schema 成本，使未使用的占位参数不会迫使模型重试；执行仍明确保留当前命令的必填与删除语义。minimal Web agent 保留 Web 权限，但必须先关闭持久 shell 才能更改权限模式。运行时 wheel 包的消费方仍无需安装 Node；Linux wheel 包包含一个可执行文件，macOS wheel 包还包含其私有原生 helper。
@@ -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/2026-07-31-even-out-shipped-tool-rosters.md
-2026-07-31-even-out-shipped-tool-rosters.md: 97a9fdaedb97de77c195c319f14f972aac850726
-2026-07-31-even-out-shipped-tool-rosters.zh.md: 130573f0ddf0b94b4dcb017f58f1e0935e53e844
+2026-07-31-even-out-shipped-tool-rosters.md: 8d1af039c99fe6616745340fd1ef78b62e15b0ca
+2026-07-31-even-out-shipped-tool-rosters.zh.md: b79486516dd3f7b54e27a3e7870ce42cd84a2ebd
@@ -20,14 +20,10 @@ Two rows stay surface-specific. `tmux-context` is TUI-only because a browser sur

 ### What stays unmounted, and why

-Three capabilities stay out on the evidence their own packages record, and are listed here so "we forgot" and "we decided against" stay distinguishable.
+Two capabilities stay out on the evidence their own packages record, and are listed here so "we forgot" and "we decided against" stay distinguishable.

 **`dsh-tool-cordis`** lets the model write JavaScript and mount it as a temporary plugin. Its README states the limit: "The sandbox is containment for honest code, not a security boundary — host-realm helpers on the sandbox global are reachable, so mount code can reach Node" ([Known limitations](../../../../packages/extensions/tool-cordis/README.md)). The `node:vm` realm lives inside the harness process while `dsh-sandbox-local` confines only the argv it spawns, so on the Web surface both the sandbox and the approval seam are bypassed rather than enforced.

-**`dsh-web-fetch-http`** stays unmounted and `dsh-tool-web` keeps `fetch: false`. SSRF protection is deferred in the implementation ([`policy.ts`](../../../../packages/web/web-fetch-http/src/policy.ts) validates protocol, credentials, and length only) and the package says so: "this provider is an SSRF primitive and **must not be enabled** in a deployment that can reach sensitive internal network targets" ([README](../../../../packages/web/web-fetch-http/README.md)). The model chooses the target, which includes the harness's own gateway on loopback, private ranges, and cloud metadata endpoints.
-
-Withholding it narrows the surface without removing the reach: `bash` is mounted, so `curl` gets the same page, as a live run confirmed. What the absence buys is the removal of an argument-shaped request primitive that needs no shell — and with it the accidental path where a summarization request quietly reaches loopback. A deployment that must contain outbound traffic needs a network-level control.
-
 **The LSP trio** stays out for an operational reason rather than a security one: `command` resolves from `PATH` at plugin load, so a missing language server fails the whole boot rather than one tool. It becomes mountable once absence degrades to a skipped registration.

 ### MCP is a dependency, not a row
@@ -20,14 +20,10 @@ Status: implemented

 ### 什么保持不挂，以及为什么

-有三项能力基于其自身包所记录的证据保持在外,列在这里是为了让「我们忘了」和「我们决定不要」保持可区分。
+有两项能力基于其自身包所记录的证据保持在外,列在这里是为了让「我们忘了」和「我们决定不要」保持可区分。

 **`dsh-tool-cordis`** 让模型写一段 JavaScript 并挂成临时插件。它的 README 写明了这个界限:「The sandbox is containment for honest code, not a security boundary — host-realm helpers on the sandbox global are reachable, so mount code can reach Node」([Known limitations](../../../../packages/extensions/tool-cordis/README.zh.md))。`node:vm` 的 realm 就在 harness 进程内,而 `dsh-sandbox-local` 只约束它 spawn 出去的 argv,因此在 Web surface 上,沙箱与批准接缝是被绕过而非被执行。

-**`dsh-web-fetch-http`** 保持不挂,`dsh-tool-web` 保持 `fetch: false`。SSRF 防护在实现中是 deferred 状态([`policy.ts`](../../../../packages/web/web-fetch-http/src/policy.ts) 只校验协议、凭据与长度),包里也直说了:「this provider is an SSRF primitive and **must not be enabled** in a deployment that can reach sensitive internal network targets」([README](../../../../packages/web/web-fetch-http/README.zh.md))。目标由模型选择,其中包括 harness 自己跑在环回地址上的网关、内网段和云元数据端点。
-
-不挂载它收窄的是接触面而非可达性：`bash` 是挂着的,`curl` 照样能拿到同一个页面——一次真实运行确认了这点。这个缺席买到的是去掉一个无需 shell、以参数成形的请求原语,以及随之而来的那条意外路径:一次「帮我总结这个页面」悄悄打到环回地址。真要收住出站流量的部署需要的是网络层管控。
-
 **LSP 三件套**留在外面是运维原因而非安全原因:`command` 在插件加载时从 `PATH` 解析,因此缺少语言服务器会让整次启动失败,而不只是失去一个工具。等到「缺失」退化为「跳过注册」之后,它就可以挂了。

 ### MCP 是依赖,不是配置行
@@ -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/2026-07-31-web-default-search.md
-2026-07-31-web-default-search.md: 1efa6fc939a883221e3158705236bc001313303d
-2026-07-31-web-default-search.zh.md: a782c4b58b6504c87932b379d294448b141f1305
+2026-07-31-web-default-search.md: efec6e1e94089d3bbd79296ff0eb2cb55ce005b3
+2026-07-31-web-default-search.zh.md: 825f0ee3f899c108039f6a0db59f0dfe2822cb72
@@ -10,13 +10,13 @@ The harness had a complete Web capability family—provider registry, DeepSeek/E

 ## Decision

-`apps/cli/config/base.cordis.yml` explicitly mounts `dsh-web` with `searchProvider: deepseek-official`, `dsh-web-search-deepseek`, and `dsh-tool-web` with `fetch: false` and `searchTimeoutMs: 60000`. It does not mount `dsh-web-fetch-http` or select a fetch provider. The shared base makes only `web_search` a default for TUI, browser, and headless sessions. The explicit search provider id keeps selection independent of registration order and leaves personal or `--config` overlays able to replace or disable the rows. The one-minute shipped budget covers an auxiliary DeepSeek Messages request plus server-side retrieval while leaving `dsh-tool-web`'s provider-neutral 30-second default unchanged for custom compositions.
+`apps/cli/config/base.cordis.yml` explicitly mounts `dsh-web` with `searchProvider: deepseek-official` and `fetchProvider: http`, `dsh-web-search-deepseek`, `dsh-web-fetch-http`, and `dsh-tool-web` with `fetch: false` and `searchTimeoutMs: 60000`. The shared base therefore keeps only `web_search` visible unless a product preset enables fetch; the shipped Web `cordis`, `code`, and `standard` presets do so. Explicit provider ids keep selection independent of registration order and leave personal or `--config` overlays able to replace or disable the rows. The one-minute shipped budget covers an auxiliary DeepSeek Messages request plus server-side retrieval while leaving `dsh-tool-web`'s provider-neutral 30-second default unchanged for custom compositions. The [Web capability seam decision](../architecture/2026-06-24-web-capability-seam.md) owns the public-fetch security policy and Web preset default.

 DeepSeek search uses the same `DEEPSEEK_API_KEY` credential reference as the official conversation adapter. The provider resolves that reference inside every search through the optional `ctx.credentials` service; only a composition without the seam falls back to the launching process environment, and a non-empty literal `apiKey` remains the programmatic last resort. A stored or rotated Web Models key therefore reaches the next search without restarting or retaining the value on the provider. Because `WebSearchProvider.available()` is synchronous, it treats an installed resolver as locally usable and missing dynamic credentials fail the operation with the provider-specific `WEB_PROVIDER_CREDENTIAL_MISSING` code while the stable tool schema stays registered.

 Search keeps its endpoint distinct from chat completions: `DEEPSEEK_SEARCH_BASE_URL` overrides the Anthropic-compatible base, while `DEEPSEEK_BASE_URL` continues to configure conversation requests. Each `web_search` performs an auxiliary DeepSeek Messages call with the native search server tool. Immediately before dispatch, the provider appends a log-only `web/deepseek-search-llm-request` event to the initiating Agent session with the resolved endpoint, API version, and exact secret-free JSON body. Credential preflight remains provider-local and races caller cancellation; neither concern expands the generic Web or credentials seams.

-The default mount does not create a Web-specific permission policy. `web_search` executes outside the shell/filesystem sandbox and approval presets, following `dsh-tool-web`'s existing contract. It does not mount `web_fetch` or a local fetch provider, so the default does not grant model-selected arbitrary URL retrieval. The shipped `workspace-write` default governs file mutations only; a restricted-network product stance requires a `tools/pre-execute` policy or capability-specific network confinement rather than implying that filesystem access mode governs Web calls.
+The default mount does not create a Web-specific permission policy. `web_search` and enabled `web_fetch` calls execute outside the shell/filesystem sandbox and approval presets, following `dsh-tool-web`'s existing contract. The HTTP provider restricts fetches to validated public destinations, but it does not constrain public data egress. The shipped `workspace-write` default governs file mutations only; a restricted-network product stance requires a `tools/pre-execute` policy or capability-specific network confinement rather than implying that filesystem access mode governs Web calls.

 ## Alternatives considered

@@ -30,8 +30,8 @@ The default mount does not create a Web-specific permission policy. `web_search`

 **Raise `dsh-tool-web`'s provider-neutral timeout.** Rejected because custom providers and deployments own different latency expectations; the shipped DeepSeek composition owns this deployment budget.

-**Enable search and fetch together.** Rejected because default `web_fetch` would allow model-selected anonymous outbound HTTP(S) retrieval to arbitrary URLs. Search covers discovery; deployments that accept broader retrieval can opt into `dsh-web-fetch-http` and set `dsh-tool-web`'s `fetch` option to `true` in their overlay.
+**Enable fetch on every shared-base surface.** Rejected because the shared base serves products with different network postures. It mounts the public-only provider but keeps the tool opt-in; the shipped Web presets deliberately enable it, while another product can leave it hidden or add stricter network policy.

 ## Consequences

-Native model requests on every shipped surface carry only the `web_search` schema and search-only prompt guidance; Web/headless Code Mode exposes the same search capability beneath `run_code`. The prompt tells the model to use returned snippets and never advertises the disabled `web_fetch` tool. Search adds a complete auxiliary model call and may use the native server tool multiple times; its exact secret-free request remains reconstructable from the initiating session log. The default offers search-result snippets and source metadata but no arbitrary page retrieval; deployments that need full-page fetch must opt in. The Web snapshot lane boots the shipped tree, drives a replayed `web_search` call through the real DeepSeek provider against a local Messages fixture, asserts the durable auxiliary request and structured result, and pins the settled browser presentation. The TUI/Web composition smokes pin the shared `web_search` roster and absence of `web_fetch`; the built composition dump pins the one-minute shipped search budget; provider tests pin missing, stored, and rotated credential behavior plus literal and ambient compatibility.
+Native model requests on every shared-base surface carry the `web_search` schema and search guidance; Web/headless Code Mode exposes the same search capability beneath `run_code`. Search adds a complete auxiliary model call and may use the native server tool multiple times; its exact secret-free request remains reconstructable from the initiating session log. The shipped Web `cordis`, `code`, and `standard` presets additionally expose `web_fetch` with public-address enforcement and no per-call approval. The Web snapshot lane boots the shipped tree, drives a replayed `web_search` call through the real DeepSeek provider against a local Messages fixture, asserts the durable auxiliary request and structured result, and pins the settled browser presentation. Composition smokes pin the shared search roster and per-preset fetch choices; the built composition dump pins the one-minute shipped search budget; provider tests pin missing, stored, and rotated credential behavior plus literal and ambient compatibility.
@@ -10,13 +10,13 @@ Status: implemented

 ## 决策

-`apps/cli/config/base.cordis.yml` 明确挂载 `dsh-web`，配置 `searchProvider: deepseek-official`，同时挂载 `dsh-web-search-deepseek`，并以 `fetch: false` 和 `searchTimeoutMs: 60000` 挂载 `dsh-tool-web`。它不挂载 `dsh-web-fetch-http`，也不选择抓取提供方。共享 base 只将 `web_search` 设为 TUI、浏览器与无头会话的默认工具。显式搜索提供方 id 使选择不受注册顺序影响，同时个人覆盖层或 `--config` 覆盖层仍可替换或禁用这些配置项。已交付的一分钟预算用于覆盖一次辅助 DeepSeek Messages 请求及服务端检索，同时保持 `dsh-tool-web` 提供方无关的 30 秒默认值不变，以供自定义组合使用。
+`apps/cli/config/base.cordis.yml` 明确挂载 `dsh-web`，配置 `searchProvider: deepseek-official` 与 `fetchProvider: http`，同时挂载 `dsh-web-search-deepseek`、`dsh-web-fetch-http`，并以 `fetch: false` 和 `searchTimeoutMs: 60000` 挂载 `dsh-tool-web`。因此，共享 base 只会暴露 `web_search`，除非产品 preset 启用抓取；已交付的 Web `cordis`、`code` 与 `standard` preset 会启用抓取。显式提供方 id 使选择不受注册顺序影响，同时个人覆盖层或 `--config` 覆盖层仍可替换或禁用这些配置项。已交付的一分钟预算用于覆盖一次辅助 DeepSeek Messages 请求及服务端检索，同时保持 `dsh-tool-web` 提供方无关的 30 秒默认值不变，以供自定义组合使用。[Web 能力 seam 决策](../architecture/2026-06-24-web-capability-seam.zh.md)负责公开抓取安全策略与 Web preset 默认值。

 DeepSeek 搜索使用与官方会话适配器相同的 `DEEPSEEK_API_KEY` 凭据引用。提供方在每次搜索内部通过可选的 `ctx.credentials` 服务解析该引用；只有未挂载该 seam 的组合才会回退到启动进程的环境变量，非空的 `apiKey` 字面值仍作为程序化配置的最后兜底。因此，由 Web 的 Models 页存储或轮换的密钥无需重启即可用于下一次搜索，提供方也无需保留该值。由于 `WebSearchProvider.available()` 是同步方法，它会将已安装解析器视为本地可用；若动态凭据缺失，操作会以提供方专属错误码 `WEB_PROVIDER_CREDENTIAL_MISSING` 失败，而稳定的工具 schema 仍保持注册。

 搜索端点与 chat completions 保持独立：`DEEPSEEK_SEARCH_BASE_URL` 覆盖 Anthropic 兼容基址，`DEEPSEEK_BASE_URL` 则继续配置会话请求。每次 `web_search` 都会发起一次辅助 DeepSeek Messages 调用，并携带原生搜索服务器工具。发出请求前一刻，提供方会向发起请求的 agent（智能体）会话追加仅用于日志的 LLM（大语言模型）请求事件 `web/deepseek-search-llm-request`，其中包含已解析端点、API 版本，以及不含密钥的精确 JSON 请求体。凭据预检仍留在提供方内部，并与调用方取消存在竞态；这两项关注点都不会扩展通用 Web seam 或凭据 seam。

-默认挂载不会创建 Web 专用权限策略。`web_search` 在 bash／文件系统沙箱及审批预设之外执行，并遵循 `dsh-tool-web` 的现有约定。组合不挂载 `web_fetch` 或本地抓取提供方，因此默认配置不会允许模型自行选择任意 URL 进行抓取。已交付的 `workspace-write` 默认值只管辖文件修改；若产品采取受限网络策略，就需要添加 `tools/pre-execute` 策略或按能力限制网络访问，而不能暗示文件系统访问模式会管辖 Web 调用。
+默认挂载不会创建 Web 专用权限策略。`web_search` 与已启用的 `web_fetch` 调用会在 bash／文件系统沙箱及审批 preset 之外执行，并遵循 `dsh-tool-web` 的现有约定。HTTP 提供方把抓取限制到已验证的公开目的地址，但不限制公开数据出站。已交付的 `workspace-write` 默认值只管辖文件修改；若产品采取受限网络策略，就需要添加 `tools/pre-execute` 策略或按能力限制网络访问，而不能暗示文件系统访问模式会管辖 Web 调用。

 ## 考虑过的替代方案

@@ -30,8 +30,8 @@ DeepSeek 搜索使用与官方会话适配器相同的 `DEEPSEEK_API_KEY` 凭据

 **提高 `dsh-tool-web` 的提供方无关超时。** 不予采纳：自定义提供方和部署有各自不同的延迟预期；这一部署预算应归已交付的 DeepSeek 组合所有。

-**同时启用搜索和抓取。** 不予采纳：默认启用 `web_fetch` 会允许模型自行选择任意 URL，执行匿名出站 HTTP(S) 抓取。搜索负责发现信息；接受更广泛抓取范围的部署可以在覆盖层中选择启用 `dsh-web-fetch-http`，并将 `dsh-tool-web` 的 `fetch` 选项设为 `true`。
+**在每个共享 base surface 上启用抓取。** 不予采纳：共享 base 服务于网络策略不同的产品。它会挂载仅限公网的提供方，但保持工具按需启用；已交付的 Web preset 会有意启用该工具，其他产品则可以继续隐藏它或添加更严格的网络策略。

 ## 后果

-每个已交付界面的原生模型请求都只会携带 `web_search` schema，以及仅用于搜索的提示词指引；Web／无头 Code Mode 通过 `run_code` 公开相同的搜索能力。该提示词要求模型使用返回的 snippet，且绝不会向模型提及已禁用的 `web_fetch` 工具。搜索会增加一次完整的辅助模型调用，并可能多次使用原生服务器工具；发起会话的日志仍可精确重建其不含密钥的请求。默认配置会提供搜索结果 snippet 与来源元数据，但不支持任意页面抓取；需要抓取完整页面的部署必须自行选择启用抓取。Web 快照通道会启动已交付配置树，使用本地 Messages fixture（测试前置数据），经由真实 DeepSeek 提供方驱动一次回放的 `web_search` 调用，断言持久化的辅助请求与结构化结果，并固定最终浏览器呈现。TUI/Web 组合冒烟测试固定了共享的 `web_search` 清单及不提供 `web_fetch` 这一事实；构建后组合配置的转储固定了已交付的一分钟搜索预算；提供方测试固定缺失、已存储及已轮换凭据的行为，以及字面值与环境变量的兼容性。
+每个共享 base surface 的原生模型请求都会携带 `web_search` schema 与搜索指引；Web／无头 Code Mode 通过 `run_code` 公开相同的搜索能力。搜索会增加一次完整的辅助模型调用，并可能多次使用原生服务器工具；发起会话的日志仍可精确重建其不含密钥的请求。已交付的 Web `cordis`、`code` 与 `standard` preset 还会暴露 `web_fetch`，实施公开地址强制校验且无需逐次审批。Web 快照通道会启动已交付配置树，使用本地 Messages fixture（测试前置数据），经由真实 DeepSeek 提供方驱动一次回放的 `web_search` 调用，断言持久化的辅助请求与结构化结果，并固定最终浏览器呈现。组合冒烟测试会固定共享搜索清单与各 preset 的抓取选择；构建后组合配置的转储固定已交付的一分钟搜索预算；提供方测试固定缺失、已存储及已轮换凭据的行为，以及字面值与环境变量的兼容性。
@@ -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/2026-08-04-claude-code-and-codex-subagent-backends.md
-2026-08-04-claude-code-and-codex-subagent-backends.md: 9b47fcf49d47d2c3561245fa1e16ff8c5da0a35c
-2026-08-04-claude-code-and-codex-subagent-backends.zh.md: 8ba4d259872558ced89083bb53f9228f46c3d45c
+2026-08-04-claude-code-and-codex-subagent-backends.md: 8af99c9171e073e5901390fed29047b350fe4924
+2026-08-04-claude-code-and-codex-subagent-backends.zh.md: 5abc0b16077bcf0be7835b2cf687a1709b53fc81
@@ -12,7 +12,7 @@ The product integrations must not become second owners for task text, cwd, cance

 ## Decision

-The harness publishes two sibling one-shot provider packages whose default registry names are `codex` and `claude-code`. This note owns their product protocols, result mapping, and process lifecycle; the [named-instance decision](2026-08-18-product-subagent-named-instances.md) owns Profile-selected provider identity and static tool binding, the [production-install exclusion decision](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.md) owns their independent optional Bundles and host-plane placement, the [product one-shot background decision](2026-08-12-product-subagent-one-shot-background-tasks.md) owns the model-visible scheduling choice, the [non-interactive permissions decision](2026-08-15-product-subagent-noninteractive-permissions.md) owns each product Provider's Profile-selected mode and safe permission decisions, and the [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) owns version-pinned product categories, lifecycle stages, and process outcomes exposed through the same diagnostic. Both packages accept multiple named instances. Loading either provider starts no product process, and each tool accepts only a standalone text task; product and instance selection remain deployment configuration.
+The harness publishes two sibling one-shot provider packages whose default registry names are `codex` and `claude-code`. This note owns their product protocols, result mapping, and process lifecycle; the [named-instance decision](2026-08-18-product-subagent-named-instances.md) owns Profile-selected provider identity, optional instance model where supported, and static tool binding; the [production-install exclusion decision](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.md) owns their independent optional Bundles and host-plane placement; the [product one-shot background decision](2026-08-12-product-subagent-one-shot-background-tasks.md) owns the model-visible scheduling choice; the [non-interactive permissions decision](2026-08-15-product-subagent-noninteractive-permissions.md) owns each product Provider's Profile-selected mode and safe permission decisions; and the [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns coarse product action categories while the [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) continues to describe Codex's current detailed categories. Both packages accept multiple named instances. Loading either provider starts no product process, and each tool accepts only a standalone text task; product and instance selection remain deployment configuration.

 Both providers report `inheritsParentContext: false`, advertise no optional start capabilities, and pass the parent Session cwd without copying the parent conversation. Their documented tools use `backgroundMode: 'one-shot'` and `maxDepth: 'provider-managed'`: the consumer keeps foreground collection as the default and may place the same run in the generic Job runtime, while recursion policy stays with the out-of-process product. Every call creates a fresh product process and a non-resumable product conversation. `ctx.subagents` owns named-request resolution and paired lifecycle events; `dsh-tool-subagent` owns model-visible scheduling and foreground-versus-Job adaptation; `ctx.jobs` and `dsh-tool-jobs` own Job ids, state, output, controls, notices, and parent-owner cancellation; each product provider owns native result mapping, while `dsh-subprocess` owns credential scrubbing, process-tree termination, and whole-tree exit observation.

@@ -48,11 +48,11 @@ Codex 0.147.0 speaks the Responses protocol, while DeepSeek's public OpenAI-comp

 ## Claude Code provider

-`@deepseek-ai/dsh-subagent-claude-code` registers a Profile-selected provider name that defaults to `claude-code` and invokes `@anthropic-ai/claude-agent-sdk@0.3.220`. The provider omits `pathToClaudeCodeExecutable`, so the SDK selects Claude Code 2.1.220 from the matching OS, CPU, and Linux-libc platform package in its own optional dependency closure. The provider does not resolve or fall back to a host `claude`; an omitted, unsupported, missing, or damaged platform payload fails the first delegation at the SDK startup boundary. The provider uses the official `query()` entrypoint and passes the SDK's native `claude` or `claude.exe` command, arguments, cwd, environment, and forwarded signal from `spawnClaudeCodeProcess` to `dsh-subprocess`; its private `SpawnedProcess` adapter exposes only the stream, event, kill, and exit facts the SDK requires.
+`@deepseek-ai/dsh-subagent-claude-code` registers a Profile-selected provider name that defaults to `claude-code` and invokes `@anthropic-ai/claude-agent-sdk@0.3.241`. The provider omits `pathToClaudeCodeExecutable`, so the SDK selects Claude Code 2.1.241 from the matching OS, CPU, and Linux-libc platform package in its own optional dependency closure. The provider does not resolve or fall back to a host `claude`; an omitted, unsupported, missing, or damaged platform payload fails the first delegation at the SDK startup boundary. The provider uses the official `query()` entrypoint and passes the SDK's native `claude` or `claude.exe` command, arguments, cwd, environment, and forwarded signal from `spawnClaudeCodeProcess` to `dsh-subprocess`; its private `SpawnedProcess` adapter exposes only the stream, event, kill, and exit facts the SDK requires.

-The public configuration contains a non-empty `providerName`, an explicit `env` overlay, a positive finite `disposeGraceMs` no greater than the repository's shared `MAX_TIMER_DELAY_MS`, and a five-value native `permissionMode` that defaults to `dontAsk`. Each named instance retains those resolved values for its own runs. Each run creates its own `AbortController`, sets `persistSession: false`, disables `AskUserQuestion`, and passes the resolved mode to the SDK; only `bypassPermissions` receives the SDK's explicit dangerous confirmation. The provider deliberately omits `settingSources`, so the SDK reads the host's normal user, project, and local Claude settings relative to the parent Session cwd. It neither copies nor filters those settings and does not create or modify login state. Remaining permission prompts are denied, MCP elicitation is declined, and blocking dialogs fail closed instead of waiting for a user interface the provider does not own.
+The public configuration contains a non-empty `providerName`, an optional non-empty `model`, an explicit `env` overlay, a positive finite `disposeGraceMs` no greater than the repository's shared `MAX_TIMER_DELAY_MS`, and a five-value native `permissionMode` that defaults to `dontAsk`. Each named instance retains those resolved values for its own runs. An explicit model is passed unchanged through `Options.model`; omission leaves that field absent so native settings choose the model. Each run creates its own `AbortController`, sets `persistSession: false`, disables `AskUserQuestion`, and passes the resolved mode to the SDK; only `bypassPermissions` receives the SDK's explicit dangerous confirmation. The provider deliberately omits `settingSources`, so the SDK reads the host's normal user, project, and local Claude settings relative to the parent Session cwd. It neither copies nor filters those settings and does not create or modify login state. Remaining permission prompts are denied, MCP elicitation is declined, and blocking dialogs fail closed instead of waiting for a user interface the provider does not own.

-The provider publishes only after both the SDK `Query` and a live managed CLI handle exist. It consumes the complete SDK stream and completes only when a `result` message has `subtype: "success"`, `is_error: false`, and a nonblank `result`, and the iterator then ends normally. The [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) owns every non-success category, stage, process outcome, and its ordering with a contributing permission decision. Local cancellation wins and becomes `aborted` without either diagnostic fact.
+The provider publishes only after both the SDK `Query` and a live managed CLI handle exist. It consumes the complete SDK stream and completes only when a `result` message has `subtype: "success"`, `is_error: false`, and a nonblank `result`, and the iterator then ends normally. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns every non-success action category, stage, process outcome, and its ordering with a contributing permission decision. Local cancellation wins and becomes `aborted` without either diagnostic fact.

 Startup rollback and published disposal close the SDK query, abort the per-run controller, invoke shared process-tree termination, and wait for whole-tree exit. `Query.close()` expresses graceful protocol intent but does not replace the subprocess owner's exit proof. An unpublished failure exposes only fixed `query-start` facts; a published process failure can expose its independent exit code and signal; an independent cleanup rejection exposes `teardown`. Original SDK, Host, and cleanup errors remain on internal cause chains and logs rather than entering the diagnostic.

@@ -66,7 +66,7 @@ The Codex evidence pins `@openai/codex@0.147.0`, `codex-cli 0.147.0`, and all si

 The Codex credentialed e2e registers the production provider, starts the same real app-server, and requests one random nonce through the test-private bridge described above. It fixes the external endpoint and model, stores no credential or request payload, requires exactly one completed upstream response, compares the trimmed product answer byte-for-byte with the nonce, and waits for every managed handle to exit.

-The Claude Code evidence pins Agent SDK 0.3.220, Claude Code 2.1.220, and all eight SDK platform packages. Its real-product spec lets the SDK select the installed payload, asserts that the shared subprocess argv begins with that package's native CLI, and observes the exact `x-api-key`, original task, byte-exact final answer, native permission modes, suite-owned denied and bypassed writes, and whole-tree exit. Package tests prove that production never resolves host `PATH`, omits the executable override, and forwards the SDK-selected Windows `claude.exe` without a batch shim. This evidence proves the pinned official SDK/CLI integration rather than compatibility with independently installed Claude versions; the [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) owns failure and process-outcome evidence. Loader coverage resolves both products through their optional Bundle patches while starting neither product.
+The Claude Code evidence pins Agent SDK 0.3.241, Claude Code 2.1.241, and all eight SDK platform packages. Its real-product spec lets the SDK select the installed payload, asserts that the shared subprocess argv begins with that package's native CLI, and observes omitted-model inheritance, two explicit instance models, the exact `x-api-key`, original task, byte-exact final answer, native permission modes, suite-owned denied and bypassed writes, and whole-tree exit. Package tests prove that production never resolves host `PATH`, omits the executable override, and forwards the SDK-selected Windows `claude.exe` without a batch shim. This evidence proves the pinned official SDK/CLI integration rather than compatibility with independently installed Claude versions; the [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns failure and process-outcome evidence. Loader coverage resolves both products through their optional Bundle patches while starting neither product.

 The Claude Code credentialed e2e maps the key and fixed official endpoint only in the provider's in-memory environment, uses the documented `deepseek-v4-pro[1m]` and `deepseek-v4-flash` model variables, and traverses the production provider, official SDK, and real CLI. It compares the trimmed result with a random nonce and proves whole-tree exit without calling the Messages API directly from the test.

@@ -82,7 +82,7 @@ The project owner's distribution authorization is scoped to the official `@anthr

 **Product doubles as required evidence.** Doubles cover exhaustive private protocol branches but do not prove package exports, official distributions, authentication, or real process behavior. Required evidence drives each official product against a loopback model fixture.

-**Plugin-managed login, product home, models, settings, sandbox rules, or fine-grained permission policy.** Those choices would create another authority beside each product's native configuration and enlarge a one-shot provider into account management. Each product exposes only one native non-interactive mode choice in addition to environment and teardown configuration; neither Provider mirrors product rules or adds a human interaction channel.
+**Plugin-managed login, product home, model discovery or fallback, settings, sandbox rules, or fine-grained permission policy.** Those choices would create another authority beside each product's native configuration and enlarge a one-shot provider into account management. A Provider may pass one opaque Profile-selected model override where the official product supports it, but it does not discover, validate, alias, or fall back between models. Neither Provider mirrors product rules or adds a human interaction channel.

 **Continuation, progress, product-native background state, and shared parent context.** The provider payload remains one final answer for one self-contained task. The generic Job layer may add its id, status, notice, collection, and cancellation results, but product sessions, resume, follow-up, intermediate messages, parent transcript transfer, structured output, and provider-specific background state need separate user contracts and are not prebuilt.

@@ -90,6 +90,6 @@ The project owner's distribution authorization is scoped to the official `@anthr

 Users delegate through Profile-configured one-shot tools backed by the official product integrations. Explicit Profile installation and host-plane provider placement are owned by the [production-install exclusion decision](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.md); named instance identity and tool binding are owned by the [named-instance decision](2026-08-18-product-subagent-named-instances.md); per-Preset tool exposure and foreground-default optional Job scheduling are owned by the [product one-shot background decision](2026-08-12-product-subagent-one-shot-background-tasks.md). This note's provider lifecycle keeps native settings and behavior while shared services retain the sole ownership of job settlement and process-tree quiescence.

-Every delegation pays for a fresh product process and independent model context. Successful product payload remains final assistant text; a failed product run may separately expose the shared safe diagnostic containing provider-owned permission facts or version-pinned structured failure facts. Background scheduling additionally exposes generic Job ids, status, completion notices, and collection or cancellation results. Both products use Bundle-pinned platform CLIs plus native account and workspace settings and the selected Provider permission mode. Credentialed e2e runs also spend external API quota and depend on the official DeepSeek endpoint; deterministic protocol, failure, cancellation, and approval coverage remains in the keyless tier. The providers do not resume sessions, stream progress, accept new human interaction, roll back tool or file side effects, or impose a wall-clock timeout.
+Every delegation pays for a fresh product process and independent model context. Successful product payload remains final assistant text; a failed product run may separately expose the shared safe diagnostic containing provider-owned permission facts and safe product failure categories. Background scheduling additionally exposes generic Job ids, status, completion notices, and collection or cancellation results. Both products use Bundle-pinned platform CLIs plus native account and workspace settings and the selected Provider permission mode; a supported optional instance model overrides only that run's native model selection. Credentialed e2e runs also spend external API quota and depend on the official DeepSeek endpoint; deterministic protocol, failure, cancellation, and approval coverage remains in the keyless tier. The providers do not resume sessions, stream progress, accept new human interaction, roll back tool or file side effects, or impose a wall-clock timeout.

 Compatibility is pinned by package-level unit coverage, keyless real-product loopback tests, credentialed DeepSeek nonce tests, public Loader composition, built-package and NodeNext consumer checks, generated documentation and notices, and the repository CI matrix. A supported product or DeepSeek endpoint/model baseline change must refresh those facts; production performs no separate runtime version probe.
@@ -12,7 +12,7 @@ Status: implemented

 ## 决策

-harness 交付两个同级的一次性提供方包，其默认注册名称分别为 `codex` 与 `claude-code`。本说明负责它们的产品协议、结果映射和进程生命周期；[命名实例决策](2026-08-18-product-subagent-named-instances.zh.md)负责 Profile 选择的提供方身份与静态工具绑定，[生产安装排除决策](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md)负责各自独立的可选 Bundle 与 host plane（宿主平面）放置，[产品一次性后台任务决策](2026-08-12-product-subagent-one-shot-background-tasks.zh.md)负责模型可见的调度选择，[非交互权限决策](2026-08-15-product-subagent-noninteractive-permissions.zh.md)负责各产品提供方的 Profile 模式选择与安全权限决定，[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)则负责通过同一诊断公开锁定产品版本的类别、生命周期阶段与进程结果。两个包都接受多个命名实例。加载任一提供方都不会启动产品进程，而且每个工具只接受独立文本任务；产品与实例选择仍属于部署配置。
+harness 交付两个同级的一次性提供方包，其默认注册名称分别为 `codex` 与 `claude-code`。本说明负责它们的产品协议、结果映射和进程生命周期；[命名实例决策](2026-08-18-product-subagent-named-instances.zh.md)负责 Profile 选择的提供方身份、支持时的可选实例模型与静态工具绑定；[生产安装排除决策](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md)负责各自独立的可选 Bundle 与 host plane（宿主平面）放置；[产品一次性后台任务决策](2026-08-12-product-subagent-one-shot-background-tasks.zh.md)负责模型可见的调度选择；[非交互权限决策](2026-08-15-product-subagent-noninteractive-permissions.zh.md)负责各产品提供方的 Profile 模式选择与安全权限决定；[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责粗粒度产品行动类别，而[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)继续描述 Codex 当前的详细类别。两个包都接受多个命名实例。加载任一提供方都不会启动产品进程，而且每个工具只接受独立文本任务；产品与实例选择仍属于部署配置。

 这两个提供方都报告 `inheritsParentContext: false`，不声明任何可选的启动能力，并传递父会话 cwd，但不会复制父级对话。文档所示的工具使用 `backgroundMode: 'one-shot'` 与 `maxDepth: 'provider-managed'`：消费方默认在前台收集结果，也可把同一次运行放入通用 Job 运行时，而递归策略仍由进程外产品负责。每次调用都会创建一个全新的产品进程和一次不可续接的产品对话。`ctx.subagents` 负责具名请求解析与成对生命周期事件；`dsh-tool-subagent` 负责模型可见的调度以及前台与 Job 适配；`ctx.jobs` 和 `dsh-tool-jobs` 负责 Job id、状态、输出、控制、通知与父级 owner 取消；各产品提供方负责原生结果映射，`dsh-subprocess` 则负责凭证清洗、进程树终止以及整棵进程树的退出观测。

@@ -48,11 +48,11 @@ Codex 0.147.0 使用 Responses 协议，而 DeepSeek 的公开 OpenAI 兼容端

 ## Claude Code 提供方

-`@deepseek-ai/dsh-subagent-claude-code` 注册由 Profile 选择、默认值为 `claude-code` 的提供方名称，并调用 `@anthropic-ai/claude-agent-sdk@0.3.220`。提供方会省略 `pathToClaudeCodeExecutable`，因此 SDK 会从自己的 optional dependency 闭包中，按操作系统、CPU 与 Linux libc 选择携带 Claude Code 2.1.220 的匹配平台包。提供方既不会解析也不会回退宿主 `claude`；省略 optional dependency、不受支持的平台，以及缺失或损坏的平台载荷，都会在第一次委派的 SDK 启动边界失败。提供方使用官方 `query()` 入口点，并把 SDK 的 `spawnClaudeCodeProcess` 给出的原生 `claude` 或 `claude.exe` 命令、参数、cwd、环境和转发的信号交给 `dsh-subprocess`；其私有 `SpawnedProcess` 适配器只公开 SDK 所需的流、事件、终止和退出事实。
+`@deepseek-ai/dsh-subagent-claude-code` 注册由 Profile 选择、默认值为 `claude-code` 的提供方名称，并调用 `@anthropic-ai/claude-agent-sdk@0.3.241`。提供方会省略 `pathToClaudeCodeExecutable`，因此 SDK 会从自己的 optional dependency 闭包中，按操作系统、CPU 与 Linux libc 选择携带 Claude Code 2.1.241 的匹配平台包。提供方既不会解析也不会回退宿主 `claude`；省略 optional dependency、不受支持的平台，以及缺失或损坏的平台载荷，都会在第一次委派的 SDK 启动边界失败。提供方使用官方 `query()` 入口点，并把 SDK 的 `spawnClaudeCodeProcess` 给出的原生 `claude` 或 `claude.exe` 命令、参数、cwd、环境和转发的信号交给 `dsh-subprocess`；其私有 `SpawnedProcess` 适配器只公开 SDK 所需的流、事件、终止和退出事实。

-公开配置包含非空的 `providerName`、显式的 `env` 覆盖项、须为正有限值且不得大于仓库共享 `MAX_TIMER_DELAY_MS` 的 `disposeGraceMs`，以及默认使用 `dontAsk` 的五值原生 `permissionMode`。每个命名实例会为自己的运行保留这些已解析值。每次运行都会创建自己的 `AbortController`，设置 `persistSession: false`、禁用 `AskUserQuestion`，并把已解析模式传给 SDK；只有 `bypassPermissions` 会取得 SDK 的显式危险确认。提供方故意省略 `settingSources`，因此 SDK 会相对于父会话 cwd 读取宿主机常规的用户、项目和本地 Claude 设置。它既不复制也不过滤这些设置，也不会创建或修改登录状态。其余权限提示会被拒绝，MCP elicitation 会被拒绝，阻塞对话会快速失败，而不会等待本提供方不负责的用户界面。
+公开配置包含非空的 `providerName`、可选的非空 `model`、显式的 `env` 覆盖项、须为正有限值且不得大于仓库共享 `MAX_TIMER_DELAY_MS` 的 `disposeGraceMs`，以及默认使用 `dontAsk` 的五值原生 `permissionMode`。每个命名实例会为自己的运行保留这些已解析值。显式模型会原样传入 `Options.model`；省略时不设置该字段，由原生设置选择模型。每次运行都会创建自己的 `AbortController`，设置 `persistSession: false`、禁用 `AskUserQuestion`，并把已解析模式传给 SDK；只有 `bypassPermissions` 会取得 SDK 的显式危险确认。提供方故意省略 `settingSources`，因此 SDK 会相对于父会话 cwd 读取宿主机常规的用户、项目和本地 Claude 设置。它既不复制也不过滤这些设置，也不会创建或修改登录状态。其余权限提示会被拒绝，MCP elicitation 会被拒绝，阻塞对话会快速失败，而不会等待本提供方不负责的用户界面。

-只有在 SDK `Query` 与受管的活动 CLI 句柄都已存在后，提供方才会发布运行。它会消费完整的 SDK 流；只有 `result` 消息具有 `subtype: "success"`、`is_error: false` 和非空白 `result`，且迭代器随后正常结束时，运行才会完成。[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)负责所有非成功类别、阶段、进程结果，以及它们与参与失败的权限决定之间的顺序。本地取消会胜出并成为 `aborted`，且不附带这两类诊断事实。
+只有在 SDK `Query` 与受管的活动 CLI 句柄都已存在后，提供方才会发布运行。它会消费完整的 SDK 流；只有 `result` 消息具有 `subtype: "success"`、`is_error: false` 和非空白 `result`，且迭代器随后正常结束时，运行才会完成。[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责所有非成功行动类别、阶段、进程结果，以及它们与参与失败的权限决定之间的顺序。本地取消会胜出并成为 `aborted`，且不附带这两类诊断事实。

 启动回滚和已发布运行的资源释放都会关闭 SDK query、中止该次运行的控制器、调用共享的进程树终止机制，并等待整棵进程树退出。`Query.close()` 表达优雅的协议关闭意图，但不能取代子进程责任方的退出证明。未发布失败只公开固定的 `query-start` 事实；已发布进程失败可以分别公开退出码与信号；独立清理拒绝则公开 `teardown`。原始 SDK、Host 与清理错误只保留在内部 cause 链和日志中，不进入诊断。

@@ -66,7 +66,7 @@ Codex 证据会锁定 `@openai/codex@0.147.0`、`codex-cli 0.147.0` 与六个平

 带密钥 Codex e2e 会注册生产提供方，启动同样的真实 app-server，并通过上述测试专用桥接层请求一个随机数。该测试固定外部端点与模型，不存储任何凭据或请求载荷，要求上游恰好完成一次响应，将去除首尾空白后的产品答案与该随机数逐字节比较，并等待所有受管句柄退出。

-Claude Code 证据会锁定 Agent SDK 0.3.220、Claude Code 2.1.220 与八个 SDK 平台包。真实产品测试会让 SDK 选择已安装载荷，断言共享子进程 argv 以该包的原生 CLI 开头，并观测确切的 `x-api-key`、原始任务、逐字节完全一致的最终回答、原生权限模式、测试拥有范围内的拒绝写入与 bypass 写入，以及整棵进程树退出。包测试还会证明生产运行从不解析宿主 `PATH`、省略可执行文件覆盖，并直接转发 SDK 所选的 Windows `claude.exe` 而不经过 batch shim。这项证据证明锁定的官方 SDK/CLI 集成，而不证明与独立安装的 Claude 版本兼容；[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)负责失败与进程结果证据。Loader 覆盖会通过各自的可选 Bundle patch 解析两个产品，且不会启动任一产品。
+Claude Code 证据会锁定 Agent SDK 0.3.241、Claude Code 2.1.241 与八个 SDK 平台包。真实产品测试会让 SDK 选择已安装载荷，断言共享子进程 argv 以该包的原生 CLI 开头，并观测省略模型继承、两个显式实例模型、确切的 `x-api-key`、原始任务、逐字节完全一致的最终回答、原生权限模式、测试拥有范围内的拒绝写入与 bypass 写入，以及整棵进程树退出。包测试还会证明生产运行从不解析宿主 `PATH`、省略可执行文件覆盖，并直接转发 SDK 所选的 Windows `claude.exe` 而不经过 batch shim。这项证据证明锁定的官方 SDK/CLI 集成，而不证明与独立安装的 Claude 版本兼容；[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责失败与进程结果证据。Loader 覆盖会通过各自的可选 Bundle patch 解析两个产品，且不会启动任一产品。

 带密钥 Claude Code e2e 仅在提供方的内存环境中映射密钥与固定的官方端点，把模型变量设为文档所示的 `deepseek-v4-pro[1m]` 与 `deepseek-v4-flash`，并实际经过生产提供方、官方 SDK 与真实 CLI。它将去除首尾空白后的结果与一个随机数比较，并证明整棵进程树退出，且测试不会直接调用 Messages API。

@@ -82,7 +82,7 @@ Claude Code 证据会锁定 Agent SDK 0.3.220、Claude Code 2.1.220 与八个 SD

 **以产品替身作为强制证据。** 替身可以穷尽覆盖私有协议分支，但无法证明包导出、官方发行版、身份验证或真实进程行为。强制证据会驱动每个官方产品连接回环模型 fixture。

-**由插件管理登录、产品主目录、模型、设置、沙箱规则或细粒度权限策略。** 这些选择会在每个产品的原生配置之外建立另一套权威来源，并将一次性提供方扩张为账户管理功能。两个产品除环境和清理配置外都只公开一个原生非交互模式选择；任一提供方都不会镜像产品规则或增加人工交互通道。
+**由插件管理登录、产品主目录、模型发现或 fallback、设置、沙箱规则或细粒度权限策略。** 这些选择会在每个产品的原生配置之外建立另一套权威来源，并将一次性提供方扩张为账户管理功能。官方产品支持时，提供方可以传入一个不透明的 Profile 模型覆盖，但不会发现、校验、解释别名或在模型间 fallback。任一提供方都不会镜像产品规则或增加人工交互通道。

 **续接、进度、产品原生后台状态和共享父级上下文。** 提供方载荷仍是一项自包含任务的一个最终回答。通用 Job 层可以额外提供 id、状态、通知、收集与取消结果，但产品会话、恢复、后续交互、中间消息、父级 transcript（文本记录）传递、结构化输出和提供方专属后台状态都需要独立的用户约定，当前实现不会预先构建这些功能。

@@ -90,6 +90,6 @@ Claude Code 证据会锁定 Agent SDK 0.3.220、Claude Code 2.1.220 与八个 SD

 用户通过由 Profile 配置、并由官方产品集成支持的一次性工具进行委派。显式 Profile 安装与 host plane 提供方放置由[生产安装排除决策](../simplification/2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md)负责；命名实例身份与工具绑定由[命名实例决策](2026-08-18-product-subagent-named-instances.zh.md)负责；按 Preset 暴露工具以及默认前台且可选通用 Job 的调度方式由[产品一次性后台任务决策](2026-08-12-product-subagent-one-shot-background-tasks.zh.md)负责。本说明规定的提供方生命周期会保留原生设置与行为，而共享服务继续独占作业结算与进程树完全停稳的责任。

-每次委派都要承担新建产品进程和独立模型上下文的开销。成功的产品载荷仍只有最终 assistant 文本；失败的产品运行可以另行公开共享安全诊断，其中包含由提供方拥有的权限事实，或锁定版本产品提供的结构化失败事实。后台调度还会额外公开通用 Job id、状态、完成通知以及收集或取消结果。两个产品都使用 Bundle 锁定的平台 CLI，并保留原生账户与工作区设置以及所选提供方权限模式。带密钥 e2e 运行还会消耗外部 API 配额，并依赖 DeepSeek 官方端点；对协议、失败、取消与审批的确定性覆盖仍由无密钥层级承担。提供方不会恢复会话、以流式方式传送进度、接受新的人工交互、回滚工具或文件副作用，也不会施加按实际经过时间触发的超时。
+每次委派都要承担新建产品进程和独立模型上下文的开销。成功的产品载荷仍只有最终 assistant 文本；失败的产品运行可以另行公开共享安全诊断，其中包含由提供方拥有的权限事实与安全产品失败类别。后台调度还会额外公开通用 Job id、状态、完成通知以及收集或取消结果。两个产品都使用 Bundle 锁定的平台 CLI，并保留原生账户与工作区设置以及所选提供方权限模式；受支持的可选实例模型只覆盖该次运行的原生模型选择。带密钥 e2e 运行还会消耗外部 API 配额，并依赖 DeepSeek 官方端点；对协议、失败、取消与审批的确定性覆盖仍由无密钥层级承担。提供方不会恢复会话、以流式方式传送进度、接受新的人工交互、回滚工具或文件副作用，也不会施加按实际经过时间触发的超时。

 兼容性由包级单元测试覆盖率、无密钥真实产品回环测试、带密钥 DeepSeek 随机数测试、公开 Loader 组合、已构建包与 NodeNext 消费方检查、生成的文档与声明以及仓库 CI 矩阵共同锁定。更改受支持的产品基线或 DeepSeek 端点／模型基线时必须刷新这些事实；生产环境不会另行执行运行时版本探测。
@@ -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/2026-08-11-workspace-sidebar-order-and-folding.md
-2026-08-11-workspace-sidebar-order-and-folding.md: ad079cfc71d6efff7679ce3b8512167bce95e6c8
-2026-08-11-workspace-sidebar-order-and-folding.zh.md: 1c3c4345f59a7e4ed1d801bdb7ffe0937e4733b4
+2026-08-11-workspace-sidebar-order-and-folding.md: 4582631da79c946a2d3a713bc0d3a085ebed4cd0
+2026-08-11-workspace-sidebar-order-and-folding.zh.md: 6dfd443599b6c6696bbe12c19ebcde411d86a517
@@ -20,15 +20,15 @@ The client installs a Workspace drag optimistically. Request and frame generatio

 ### Session folding and view order

-Each Workspace persists one browser-local open state: closed means zero Session rows and open means up to five. When more Sessions exist, **Show more** reveals the remainder only for the current mount; closing the whole Workspace clears this transient expansion, so reopening returns to five. The current Session's group opens automatically only when the user has not already stored an explicit state for that Workspace. Creating a Session from a Workspace row opens the target group before starting the Session, keeping the new row visible when state propagation completes. After a ready Workspace baseline changes, the browser removes expansion, order, and observed-timestamp records for ids absent from that baseline while retaining the Ungrouped and flat-list accounts.
+Each Workspace persists one browser-local open state: closed means zero Session rows and open means up to five non-blank rows plus the selected blank New Session as one provisional extra row. **Show more** reveals only the hidden remainder for the current mount; closing the whole Workspace clears this transient expansion, so reopening returns to the bounded folded projection. The current Session's group opens automatically only when the user has not already stored an explicit state for that Workspace. Creating a Session from a Workspace row opens the target group before starting the Session, keeping the new row visible when state propagation completes. After a ready Workspace baseline changes, the browser removes expansion, order, and observed-timestamp records for ids absent from that baseline while retaining the Ungrouped and flat-list accounts.

 The combined view menu offers **Manual** and **Last updated** in grouped and flat presentation, with one browser-local persisted order per account. A real Workspace initializes from `WorkspaceView.sessionIds`; Ungrouped and the cross-Workspace flat list initialize from recency and have no Host Session account. Entering Last updated performs one complete recency sort; a later user prompt or steer promotes that Session once, and dragging may edit the resulting order. Returning to Manual preserves the current order and only disables later activity promotion. Manual-mode drags for a real Workspace also write the Host Session account, while Ungrouped and flat-list drags and activity promotion remain browser-local. Flat rows omit an empty leading status slot because they have no parent hierarchy, while a visible status retains its slot.

-When New Session creation selects a blank Session, the browser promotes it once in both its grouped account and the flat-list account. This explicit creation promotion does not advance `updatedAt`; later drag ordering treats the blank like any other Session, and the first prompt does not undo a Manual-mode drag.
+When New Session creation selects a blank Session, the browser promotes it once in both its grouped account and the flat-list account. This explicit creation promotion does not advance `updatedAt`. While the Session remains blank, grouped folding does not charge it against the five non-blank rows; the first prompt returns it to the ordinary quota without undoing a Manual-mode drag.

 ### Drag and compact chrome

-Workspace hit testing uses the complete rendered group section, including visible Session rows. One insertion boundary is shared by the preceding group's lower half and the following group's upper half, and the indicator is an absolutely positioned line with a joined right-facing chevron that does not affect layout. A tree-body overlay draws the first boundary at the same negative offset outside the scrolling clip, so the leading chevron remains visible without moving the list. During a Workspace or Session drag, document-level `dragover` and `drop` handlers accept the native operation; if release occurs outside the Workspace list, `dragend` commits the last valid marker.
+Workspace hit testing uses the complete rendered group section, including visible Session rows. One insertion boundary is shared by the preceding group's lower half and the following group's upper half, and the indicator is an absolutely positioned line with a joined right-facing chevron that does not affect layout. A tree-body overlay draws the first boundary at the same negative offset outside the scrolling clip, so the leading chevron remains visible without moving the list. A collapsed Session drag resolves its insertion boundary from rendered rows, places the source before any hidden account members at that boundary, and rejects a result that would hide the source. During a Workspace or Session drag, document-level `dragover` and `drop` handlers accept the native operation; if release occurs outside the Workspace list, `dragend` commits the last valid marker.

 Search is a header action while collapsed and expands across the title and trailing actions. An outside click collapses a query that is empty after trimming but retains a non-empty query; while the rail search gesture is still in flight the outside-click listener stays unmounted ([rail-search self-dismissal](../bug-fix/2026-08-18-rail-search-outside-click-self-dismissal.md)). Compact Workspace and Session rows, a 24px bottom fade, and the absence of per-Workspace Session counts preserve vertical space without removing navigation affordances.

@@ -40,7 +40,7 @@ Search is a header action while collapsed and expands across the title and trail

 **Always show every Session in an open Workspace.** One large Workspace would continue to crowd out the rest, and remembering only the whole-group open state would not bound its height.

-**Persist the expanded-remainder state.** A Workspace reopened much later could unexpectedly occupy the full sidebar. Only the zero-or-five state represents a stable navigation preference; revealing the remainder is a local inspection.
+**Persist the expanded-remainder state.** A Workspace reopened much later could unexpectedly occupy the full sidebar. Only the bounded folded state represents a stable navigation preference; revealing the remainder is a local inspection.

 **Use numeric drop indices or header-only hit testing.** Indices drift when rows change during a drag, while header midpoints disagree with the visible boundary when a Workspace is expanded. Anchor ids and full-section geometry remain stable under both conditions.

@@ -50,10 +50,10 @@ Search is a header action while collapsed and expands across the title and trail

 - Workspace order is durable and shared through the Host, while grouping, open state, per-account Session view order, and query state remain browser-local presentation preferences. Ungrouped and the flat list support the same drag and promotion rules, but their orders are browser-local because neither has one Workspace account.
 - Last updated performs a complete recency sort on entry, then preserves manual adjustments until a user prompt or steer advances one Session and moves it to the front. Returning to Manual preserves every current position.
- A newly selected blank New Session row enters grouped and flat orders first once, then follows the same drag and activity rules as every other Session.
- Opening a Workspace never shows more than five Sessions without an explicit **Show more** gesture, and closing it resets only that transient gesture.
+- A newly selected blank New Session row enters grouped and flat orders first once. Grouped folding renders it in addition to five non-blank rows until its first prompt, then applies the ordinary quota.
+- Opening a Workspace never shows more than five non-blank Sessions without an explicit **Show more** gesture; the selected blank New Session may add one provisional row. Closing the Workspace resets only the transient gesture.
 - The Host Session account retains the manual-order meaning established by [Session List Browsing and Manual Workspace Order](2026-07-25-session-list-browsing-and-manual-order.md).

 ## Testing

-Domain and Host tests cover durable Workspace moves, no-op and invalid anchors, restart recovery, full-order RPC responses, order frames, and one Workspace snapshot per Host-stream baseline. Runtime tests cover optimistic order, frame/response precedence, overlapping rejection rollback to Host-confirmed order, reconnect baselines, and New Session target priority. UI tests cover five-row folding, transient expansion reset, pruning persisted state after Workspace removal, order-preserving mode switches, one-time recent-update and New Session promotion, Manual drag retention after the first prompt, browser-local Ungrouped and flat-list drag persistence, hierarchy-free flat-row leading spacing, selected view indicators, expanded-section Workspace hit testing, an unclipped first insertion boundary, outside-list Workspace and Session drops, search collapse rules, and compact CSS dimensions.
+Domain and Host tests cover durable Workspace moves, no-op and invalid anchors, restart recovery, full-order RPC responses, order frames, and one Workspace snapshot per Host-stream baseline. Runtime tests cover optimistic order, frame/response precedence, overlapping rejection rollback to Host-confirmed order, reconnect baselines, and New Session target priority. UI tests cover five-row folding, the blank-row quota and hidden count, collapsed drag anchors across hidden rows, transient expansion reset, pruning persisted state after Workspace removal, order-preserving mode switches, one-time recent-update and New Session promotion, Manual drag retention after the first prompt, browser-local Ungrouped and flat-list drag persistence, hierarchy-free flat-row leading spacing, selected view indicators, expanded-section Workspace hit testing, an unclipped first insertion boundary, outside-list Workspace and Session drops, search collapse rules, and compact CSS dimensions. A shipped-composition Web snapshot pins five established rows beside the provisional New Session.
@@ -20,15 +20,15 @@ Workspace 注册表持有持久 `workspaceIds` 顺序，并提供采用 DOM `ins

 ### Session 折叠与视图顺序

-每个 Workspace 持久化一项浏览器本地打开状态：关闭表示零条 Session 行，打开表示最多五条。存在更多 Session 时，**展开其余**只在当前挂载期间显示剩余项；关闭整个 Workspace 会清除此临时展开，因此重新打开时恢复为五条。只有在用户尚未为该 Workspace 存储明确状态时，当前 Session 所在分组才会自动打开。从 Workspace 行创建 Session 时会在启动 Session 前打开目标分组，使状态传播完成后新行保持可见。就绪的 Workspace 基线发生变化后，浏览器会移除基线中不存在 id 的展开状态、顺序和已观察时间戳记录，同时保留 Ungrouped 和单列表记账。
+每个 Workspace 持久化一项浏览器本地打开状态：关闭表示零条 Session 行，打开表示最多五条非空白行，再把当前选中的空白“新会话”作为一条临时额外行。**展开其余**只在当前挂载期间显示仍被隐藏的条目；关闭整个 Workspace 会清除此临时展开，因此重新打开时恢复为有界折叠投影。只有在用户尚未为该 Workspace 存储明确状态时，当前 Session 所在分组才会自动打开。从 Workspace 行创建 Session 时会在启动 Session 前打开目标分组，使状态传播完成后新行保持可见。就绪的 Workspace 基线发生变化后，浏览器会移除基线中不存在 id 的展开状态、顺序和已观察时间戳记录，同时保留 Ungrouped 和单列表记账。

 组合视图菜单在分组和单列表呈现中都提供**手动排序**和**最近更新**，每个记账各自持有一份浏览器本地持久顺序。真实 Workspace 从 `WorkspaceView.sessionIds` 初始化；Ungrouped 和跨 Workspace 的单列表从最近更新时间顺序初始化，且没有 Host Session 记账。进入最近更新时会执行一次完整的时间排序；后续 user prompt 或 steer 会将对应 Session 置顶一次，拖拽仍可编辑所得顺序。返回手动排序会保留当前顺序，只停用后续活动置顶。真实 Workspace 在手动模式下的拖拽还会写入 Host Session 记账，而 Ungrouped 和单列表的拖拽与活动置顶保留在浏览器本地。单列表没有父级层次，因此不显示空的左侧状态槽；存在可见状态时仍保留该槽。

-创建“新会话”并选中空白 Session 时，浏览器会在其分组记账和单列表记账中各置顶一次。这次明确的创建置顶不会推进 `updatedAt`；后续拖拽把空白 Session 当作普通 Session，首条提示词落地也不会撤销手动模式下的拖拽。
+创建“新会话”并选中空白 Session 时，浏览器会在其分组记账和单列表记账中各置顶一次。这次明确的创建置顶不会推进 `updatedAt`。Session 保持空白期间，分组折叠不会让它占用五条非空白行配额；首条提示词落地后，该行恢复普通配额，同时不会撤销手动模式下的拖拽。

 ### 拖拽与紧凑界面

-Workspace 命中测试使用完整渲染分组区段，包括可见 Session 行。前一分组的下半部与后一分组的上半部共享同一条插入边界，指示器是一条带有相连右向尖角且不影响布局的绝对定位横线。树主体覆盖层会在滚动裁切区外以相同的负偏移绘制第一条边界，因此左侧尖角保持可见，列表位置也不会改变。Workspace 或 Session 拖拽期间，文档级 `dragover` 与 `drop` 处理器会接受原生操作；若在 Workspace 列表外松手，`dragend` 会提交最后一个有效标记。
+Workspace 命中测试使用完整渲染分组区段，包括可见 Session 行。前一分组的下半部与后一分组的上半部共享同一条插入边界，指示器是一条带有相连右向尖角且不影响布局的绝对定位横线。树主体覆盖层会在滚动裁切区外以相同的负偏移绘制第一条边界，因此左侧尖角保持可见，列表位置也不会改变。折叠状态的 Session 拖拽按渲染行确定插入边界，把来源行放在该边界处所有隐藏记账成员之前，并拒绝会隐藏来源行的结果。Workspace 或 Session 拖拽期间，文档级 `dragover` 与 `drop` 处理器会接受原生操作；若在 Workspace 列表外松手，`dragend` 会提交最后一个有效标记。

 搜索在折叠时是区头操作，展开后占据标题与尾部操作的空间。查询经清除首尾空白后为空时，点击外部会收起搜索；非空查询则会保留；轨道搜索手势仍在进行期间，外部点击监听器保持未挂载（[轨道搜索自我收起](../bug-fix/2026-08-18-rail-search-outside-click-self-dismissal.zh.md)）。紧凑的 Workspace 与 Session 行、24px 底部渐隐以及取消每个 Workspace 的 Session 数量共同节省纵向空间，同时保留导航入口。

@@ -40,7 +40,7 @@ Workspace 命中测试使用完整渲染分组区段，包括可见 Session 行

 **打开 Workspace 时始终显示全部 Session。** 大型 Workspace 仍会挤占其他分组；只记忆整个分组的打开状态无法限制其高度。

-**持久化展开剩余状态。** 很久以后重新打开 Workspace 时，它可能意外占满侧边栏。只有零条或五条状态属于稳定导航偏好；显示剩余项只是一次本地查看。
+**持久化展开剩余状态。** 很久以后重新打开 Workspace 时，它可能意外占满侧边栏。只有有界折叠状态属于稳定导航偏好；显示剩余项只是一次本地查看。

 **使用数字下标或只按组头命中拖拽。** 拖拽期间行发生变化会使下标漂移；Workspace 展开时，组头中点与可见边界不一致。锚点 id 与完整区段几何在两种情况下都保持稳定。

@@ -50,10 +50,10 @@ Workspace 命中测试使用完整渲染分组区段，包括可见 Session 行

 - Workspace 顺序通过 Host 持久并共享；分组方式、打开状态、每个记账的 Session 视图顺序和查询状态仍是浏览器本地呈现偏好。Ungrouped 和单列表支持相同的拖拽与置顶规则，但因没有单一 Workspace 记账，其顺序只保存在浏览器本地。
 - 最近更新模式会在进入时执行完整时间排序，随后保持手动调整，直到 user prompt 或 steer 推进某条 Session 并将其置顶。返回手动排序会保留所有当前位置。
- 新选中的空白“新会话”行会在分组和单列表顺序中各置顶一次，之后遵循与其他 Session 相同的拖拽和活动规则。
- 未执行明确的**展开其余**手势时，打开 Workspace 最多显示五条 Session；关闭分组只重置这项临时手势。
+- 新选中的空白“新会话”行会在分组和单列表顺序中各置顶一次。分组折叠会在五条非空白行之外额外渲染该行，直至首条提示词落地后恢复普通配额。
+- 未执行明确的**展开其余**手势时，打开 Workspace 最多显示五条非空白 Session；当前选中的空白“新会话”可以增加一条临时行。关闭分组只重置这项临时手势。
 - Host Session 记账继续采用[会话列表浏览与 Workspace 手动排序](2026-07-25-session-list-browsing-and-manual-order.zh.md)确立的手动顺序含义。

 ## 测试

-领域与 Host 测试覆盖持久 Workspace 移动、无操作与无效锚点、重启恢复、完整顺序 RPC 响应、顺序帧以及每条 Host stream 基线只读取一份 Workspace 快照。运行时测试覆盖乐观顺序、帧／响应优先级、重叠拒绝后恢复 Host 已确认顺序、重连基线以及 New Session 目标优先级。UI 测试覆盖五行折叠、临时展开重置、Workspace 移除后清理持久状态、保持顺序的模式切换、一次性最近更新与“新会话”置顶、首条提示词落地后保留手动拖拽、浏览器本地 Ungrouped 与单列表拖拽持久化、无层级单列表行左侧间距、当前视图标记、展开区段的 Workspace 命中、未裁切的第一条插入边界、列表外 Workspace 与 Session 松手、搜索收起规则和紧凑 CSS 尺寸。
+领域与 Host 测试覆盖持久 Workspace 移动、无操作与无效锚点、重启恢复、完整顺序 RPC 响应、顺序帧以及每条 Host stream 基线只读取一份 Workspace 快照。运行时测试覆盖乐观顺序、帧／响应优先级、重叠拒绝后恢复 Host 已确认顺序、重连基线以及 New Session 目标优先级。UI 测试覆盖五行折叠、空白行配额与隐藏数量、跨隐藏行的折叠拖拽锚点、临时展开重置、Workspace 移除后清理持久状态、保持顺序的模式切换、一次性最近更新与“新会话”置顶、首条提示词落地后保留手动拖拽、浏览器本地 Ungrouped 与单列表拖拽持久化、无层级单列表行左侧间距、当前视图标记、展开区段的 Workspace 命中、未裁切的第一条插入边界、列表外 Workspace 与 Session 松手、搜索收起规则和紧凑 CSS 尺寸。真实组合 Web 快照固定五条既有行与临时“新会话”并列显示。
@@ -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/2026-08-15-product-subagent-noninteractive-permissions.md
-2026-08-15-product-subagent-noninteractive-permissions.md: 8788fba3492e08090dd038fc3e7377f6bd1e29cd
-2026-08-15-product-subagent-noninteractive-permissions.zh.md: cbf3c3cd14fcecd2c24e335a8b71cc3b5370e247
+2026-08-15-product-subagent-noninteractive-permissions.md: f9fe1e0af4f65b83a53127a64f6475861a13f0a1
+2026-08-15-product-subagent-noninteractive-permissions.zh.md: 2acb1bf095a23022fc761d254f0e515e1acf1588
@@ -26,7 +26,7 @@ Claude Code defaults to `dontAsk` and accepts only the native non-interactive mo
 | `plan` | Use planning mode, deny execution approval, and return the completed plan as the final answer. |
 | `bypassPermissions` | Set the SDK's explicit dangerous confirmation and bypass permission checks. |

-The Provider continues to omit `settingSources`: Claude Code remains the owner of user, project, and local settings, authentication, tools, and sandbox behavior outside the selected mode.
+The Provider continues to omit `settingSources`: an optional instance-level model is a separate direct SDK override, while Claude Code remains the owner of user, project, and local settings, authentication, tools, and sandbox behavior outside the selected mode.

 Every query disables `AskUserQuestion`. Non-bypass permission callbacks deny instead of returning the SDK's indefinitely blocking `null`; plan mode also places `ExitPlanMode` in `disallowedTools`, so native allow rules cannot switch the unattended query back to execution. MCP elicitation is declined; the supported refusal dialog is cancelled; undeclared dialog kinds use the SDK's no-dialog failure behavior. A native `permission_denied` message records the same operation-local fact. These paths do not create an approval session, queue, cache, or retry loop.

@@ -44,7 +44,7 @@ The Provider overrides only those thread fields. `CODEX_HOME`, project configura

 ### Failure diagnostic

-`SubagentResult` carries an optional `diagnostic` for provider-authored, non-assistant failure detail. A Provider removes tool inputs, file contents, environment values, credentials, and raw protocol payloads before producing it. The shared out-of-process result boundary limits the complete text to 4096 UTF-8 bytes and marks truncation without splitting a character. The [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) owns non-permission product categories, lifecycle stages, and process outcomes carried by the same field.
+`SubagentResult` carries an optional `diagnostic` for provider-authored, non-assistant failure detail. A Provider removes tool inputs, file contents, environment values, credentials, and raw protocol payloads before producing it. The shared out-of-process result boundary limits the complete text to 4096 UTF-8 bytes and marks truncation without splitting a character. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns Claude Code's non-permission action categories, while the [structured failure-facts decision](2026-08-18-product-subagent-failure-facts.md) continues to own Codex's current categories; both retain lifecycle stages and process outcomes in the same field.

 Each product's permission fact contains only the effective mode, request category, unattended decision, and a fixed safe reason. Claude Code derives those facts from SDK callbacks and `permission_denied` messages. Codex derives them from app-server requests, declined items, `sandboxError`, and two fixed permission signatures in a bounded stderr tail; raw stderr is still forwarded to the Host but never copied into the diagnostic. Both Providers place their structured failure line before the latest contributing permission fact. A successful result returns only the strict final answer; local cancellation remains `aborted` without permission detail; an unpublished startup failure still rejects `start()`. The Provider never adds either diagnostic fact to assistant output, structured output, or `subagent/end.lastAssistantMessage`.

@@ -63,7 +63,7 @@ The foreground consumer presents the stop-reason headline, then the optional dia

 ## Verification

-Package tests pin every allowed and rejected Config value, the exact SDK and app-server field mappings, dangerous confirmations, unattended terminal responses, diagnostic sanitization and UTF-8 bound, successful-result omission, concurrent-run isolation, foreground ordering, Job detail, stderr observer disposal, and process cleanup. The real Claude Agent SDK/CLI fixture proves its safe default, restricted denial, explicit bypass, and whole-tree quiescence. The real Codex app-server fixture proves that thread-level `never` overrides ambient `on-request`, automatic review starts, dangerous bypass writes only inside suite-owned temporary storage, fixed stderr signatures produce safe diagnostics, and the wrapper/native tree exits. Loader composition proves non-default modes can be published without starting either product, and the keyless ACP snapshot records each product's failure diagnostic through foreground and Job presentation while the model-facing product tool schemas contain no permission parameter.
+Package tests pin every allowed and rejected Config value, the exact SDK and app-server field mappings, dangerous confirmations, unattended terminal responses, diagnostic sanitization and UTF-8 bound, successful-result omission, concurrent-run isolation, foreground ordering, Job detail, stderr observer disposal, and process cleanup. The real Claude Agent SDK 0.3.241 and Claude Code 2.1.241 fixture proves its safe default, restricted denial, explicit bypass, and whole-tree quiescence. The real Codex app-server fixture proves that thread-level `never` overrides ambient `on-request`, automatic review starts, dangerous bypass writes only inside suite-owned temporary storage, fixed stderr signatures produce safe diagnostics, and the wrapper/native tree exits. Loader composition proves non-default modes can be published without starting either product, and the keyless ACP snapshot records each product's failure diagnostic through foreground and Job presentation while the model-facing product tool schemas contain no permission parameter.

 ## Alternatives considered

@@ -26,7 +26,7 @@ Claude Code 默认使用 `dontAsk`，而且只接受锁定版本 Agent SDK 支
 | `plan` | 使用规划模式，拒绝执行审批，并把完整计划作为最终答案返回。 |
 | `bypassPermissions` | 设置 SDK 的显式危险确认并跳过权限检查。 |

-提供方继续省略 `settingSources`：除所选模式以外，用户、项目和本地设置、身份验证、工具与沙箱行为仍由 Claude Code 拥有。
+提供方继续省略 `settingSources`：可选的实例级模型是独立的直接 SDK 覆盖；除所选模式以外，用户、项目和本地设置、身份验证、工具与沙箱行为仍由 Claude Code 拥有。

 每次 query 都禁用 `AskUserQuestion`。非 bypass 模式的权限回调会拒绝请求，而不会返回 SDK 中会无限阻塞的 `null`；plan 模式还会把 `ExitPlanMode` 放入 `disallowedTools`，因此原生 allow 规则无法把无人值守 query 切回执行模式。MCP elicitation 会被拒绝；已支持的拒绝对话会被取消；未声明的对话类型使用 SDK 的无对话失败行为。原生 `permission_denied` 消息会记录同一份当前运行事实。这些路径不会创建审批会话、队列、缓存或重试循环。

@@ -44,7 +44,7 @@ Codex 默认使用 `never`，并接受 Codex 0.147.0 公开的三种原生非交

 ### 失败诊断

-`SubagentResult` 携带可选的 `diagnostic`，用于提供方产生且不属于 assistant 内容的失败说明。提供方在生成它之前会排除工具输入、文件内容、环境值、凭证与原始协议载荷。共享的进程外结果边界会把完整文本限制在 4096 个 UTF-8 字节以内，并在不切断字符的前提下标记截断。[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)负责由同一字段承载的非权限产品类别、生命周期阶段与进程结果。
+`SubagentResult` 携带可选的 `diagnostic`，用于提供方产生且不属于 assistant 内容的失败说明。提供方在生成它之前会排除工具输入、文件内容、环境值、凭证与原始协议载荷。共享的进程外结果边界会把完整文本限制在 4096 个 UTF-8 字节以内，并在不切断字符的前提下标记截断。[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责 Claude Code 的非权限行动类别，[结构化失败事实决策](2026-08-18-product-subagent-failure-facts.zh.md)继续负责 Codex 的当前类别；二者都在同一字段中保留生命周期阶段与进程结果。

 每个产品的权限事实都只包含有效模式、请求类别、无人值守决定与固定的安全原因。Claude Code 从 SDK 回调和 `permission_denied` 消息取得这些事实。Codex 从 app-server 请求、被拒绝的 item、`sandboxError` 与每次运行有界 stderr 尾部中的两个固定权限签名取得事实；原始 stderr 仍会转发给 Host，但绝不会复制进诊断。两个提供方都会把结构化失败行放在最新参与失败的权限事实之前。成功结果只返回严格的最终答案；本地取消仍以 `aborted` 结算且不附带权限说明；未发布的启动失败仍会拒绝 `start()`。提供方绝不会把任一诊断事实写入 assistant 输出、结构化输出或 `subagent/end.lastAssistantMessage`。

@@ -63,7 +63,7 @@ Codex 默认使用 `never`，并接受 Codex 0.147.0 公开的三种原生非交

 ## Verification

-包测试固定所有允许与拒绝的 Config 值、准确的 SDK 与 app-server 字段映射、危险确认、无人值守终态、诊断脱敏与 UTF-8 上限、成功结果不携带诊断、并发运行隔离、前台顺序、Job detail、stderr observer 释放和进程清理。真实 Claude Agent SDK/CLI fixture 证明其安全默认、受限拒绝、显式 bypass 与整棵进程树完全停稳。真实 Codex app-server fixture 证明线程级 `never` 覆盖环境中的 `on-request`、自动评审可以启动、危险绕过只在测试拥有的临时存储中写入、固定 stderr 签名产生安全诊断，而且 wrapper／native 进程树会退出。Loader 组装证明非默认模式可以在不启动任一产品的情况下发布；无密钥 ACP snapshot 则记录每个产品的失败诊断如何经过前台与 Job 呈现，同时面向模型的产品工具 schema 不包含权限参数。
+包测试固定所有允许与拒绝的 Config 值、准确的 SDK 与 app-server 字段映射、危险确认、无人值守终态、诊断脱敏与 UTF-8 上限、成功结果不携带诊断、并发运行隔离、前台顺序、Job detail、stderr observer 释放和进程清理。真实 Claude Agent SDK 0.3.241 与 Claude Code 2.1.241 fixture 证明其安全默认、受限拒绝、显式 bypass 与整棵进程树完全停稳。真实 Codex app-server fixture 证明线程级 `never` 覆盖环境中的 `on-request`、自动评审可以启动、危险绕过只在测试拥有的临时存储中写入、固定 stderr 签名产生安全诊断，而且 wrapper／native 进程树会退出。Loader 组装证明非默认模式可以在不启动任一产品的情况下发布；无密钥 ACP snapshot 则记录每个产品的失败诊断如何经过前台与 Job 呈现，同时面向模型的产品工具 schema 不包含权限参数。

 ## Alternatives considered

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-17-web-system-prompt-opaque-body.md
+2026-08-17-web-system-prompt-opaque-body.md: 9b1992dd8eb1062a3b35666b332178985aa91653
+2026-08-17-web-system-prompt-opaque-body.zh.md: 6f2fbba9d3f286e8c3073197a144d347f732cbc1
@@ -0,0 +1,29 @@
+# Agent Note: System prompt expands into the opaque context body
+
+Status: implemented
+
+English | [中文](2026-08-17-web-system-prompt-opaque-body.zh.md)
+
+## Problem
+
+The Chat `System prompt` row shares `DisclosureRow` chrome with context injection and needs an expanded body for the request's system field. Rendering that field as Markdown would restyle it — headings, emphasis, lists — so a reader would see a rendered document the model never received. Context injection already solves the same job with a 141px code-block scrollport and `<pre>` text that keeps the bytes and line breaks the model read, so the row needs that presentation, not a second one.
+
+## Decision
+
+`SystemPromptRow` mounts the same expanded body as an opaque context injection. It reuses `ContextInjectionRow.module.css` for the 141px Figma 10:2482 scrollport and renders the durable `request/header` system string through `OpaqueBody` as one text block, so the disclosure shows model-facing text with its real line breaks and the same 20_000-character display bound. The row stays collapsed by default and still has no streaming path. It does not grow a producer label, form marker, or source-field list: the system field is one joined string on the header, not a sourced `user/message`.
+
+## Alternatives considered
+
+**Render settled Markdown in a card-styled body.** The chrome could match, but Markdown rewrites what the model read. A heading or bold span is a different document from the request bytes.
+
+**Split the joined system string into snapshot sections.** The durable header stores only the assembled text. Inventing section boundaries in the client would attribute prose the log does not name, and a resumed or foreign header could not reconstruct them.
+
+**Render through `ContextInjectionRow` itself.** That row is for sourced user-role messages: it titles a role, shows a producer, and chooses a form body. The system field is a different durable fact and has none of those fields.
+
+## Consequences
+
+The two disclosures now share one expanded-body chrome and one text presentation, so a later change to the 141px scrollport or the opaque bound applies to both. The cost is that a long system prompt scrolls inside 141px instead of 360px, and Markdown markup in the prompt stays visible as characters.
+
+## Testing
+
+`packages/client/ui-chat/tests/system-prompt-row.client.spec.tsx` expands and collapses the row and pins the opaque `[data-context-text]` bytes, including Markdown markers that must not become a heading. `apps/web/tests/replay-round-trip.e2e.ts` still opens the assembled disclosure and reads the persona line from that body.
@@ -0,0 +1,29 @@
+# Agent Note: System prompt expands into the opaque context body
+
+Status: implemented
+
+[English](2026-08-17-web-system-prompt-opaque-body.md) | 中文
+
+## Problem
+
+Chat 的 `系统提示词` 行和上下文注入共用 `DisclosureRow` 外壳，其展开内容区需要呈现请求的 system 字段。如果把该字段渲染成 Markdown——标题、强调、列表——读者看到的将是模型从未收到的排版文档。上下文注入已经用 141px 代码块滚动区和保留模型所见字节与换行的 `<pre>` 文本解决了同一件事，因此该行需要复用这一呈现，而不是再造一套。
+
+## Decision
+
+`SystemPromptRow` 展开后挂载与不透明上下文注入相同的内容区。它复用 `ContextInjectionRow.module.css` 的 Figma 10:2482 的 141px 滚动区，并把持久 `request/header` 的 system 字符串作为一块文本交给 `OpaqueBody`，因此展开后看到的是带真实换行的模型可见文本，以及相同的 20_000 字符显示上限。该行默认折叠，仍然没有流式路径。它不增加生产者标签、form 标记或 source 字段列表：system 字段是 header 上的一段拼接字符串，不是带 source 的 `user/message`。
+
+## Alternatives considered
+
+**在卡片式内容区里渲染结算后的 Markdown。** 外壳可以对齐，但 Markdown 会改写模型读到的内容。标题或加粗是另一份文档，不是请求里的字节。
+
+**把拼接后的 system 字符串拆成 snapshot 分段。** 持久 header 只保存组装后的文本。客户端臆造分段边界会把日志未命名的正文归到某个子系统，恢复或外来 header 也无法重建这些分段。
+
+**直接走 `ContextInjectionRow`。** 那一行面向带 source 的 user-role 消息：它标角色、显示生产者，并按 form 选内容区。system 字段是另一件持久事实，没有这些字段。
+
+## Consequences
+
+两处展开现在共用同一套内容区外壳和同一套文本展示，因此之后改 141px 滚动区或不透明显示上限会同时作用到两边。代价是较长的系统提示词在 141px 而不是 360px 内滚动，提示词里的 Markdown 标记会以字符形式可见。
+
+## Testing
+
+`packages/client/ui-chat/tests/system-prompt-row.client.spec.tsx` 会展开并折叠该行，并钉住不透明 `[data-context-text]` 字节，包括不得变成标题的 Markdown 标记。`apps/web/tests/replay-round-trip.e2e.ts` 仍会打开组装后的展开行，并从该内容区读出 persona 行。
@@ -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/2026-08-18-model-selected-subagent-routes.md
-2026-08-18-model-selected-subagent-routes.md: 1602e3ac90870edbd0206cd87fdf97ecc34cad41
-2026-08-18-model-selected-subagent-routes.zh.md: 0dad8b9d030e6de65cb3fa1e0e93ad7c28bcc5c1
+2026-08-18-model-selected-subagent-routes.md: bf4788b141370933197d9ec1a1ad3c8e76a6740c
+2026-08-18-model-selected-subagent-routes.zh.md: 1d83e2e91ffe87fff7f8e9d1988320cb2bb8f2f7
@@ -14,9 +14,9 @@ The model also needs a bounded way to discover live providers and model-owned ef

 `dsh-tool-subagent` exposes optional `provider`, `model`, and `reasoning_effort` fields only when its instance enables `enableModelSelection`, or its Agent-scoped `modelSelectionSettings` instance resolves an enabled Session decision, and the bound subagent provider advertises `SubagentCapabilities.agentOptions`. No route allowlist is required. Registered LLM provider routes are available for child selection; this tool does not add a second authorization policy over the deployment's LLM registry. Disabled instances omit and reject model-facing selection, while configured `Config.agentOptions` remain deployment-owned defaults. Either selection mode against a provider without the capability fails the plugin mount.

-Provider and model form one route and must be supplied together. An effort may be supplied alone when configured or parent values provide the effective route. Model arguments override `Config.agentOptions`, and configured fields override the parent Agent's latest logged request selection; creation options supply the fallback before its first request and retain the configured output-token limit. Reasoning-effort identifiers remain adapter-owned. An unchanged route inherits an omitted effort, while changing provider or model without naming an effort clears the lower layer's route-owned value so the selected model resolves its own default. `AgentOptions` carries the resulting effort into the child loop, whose request header logs the effective value. A continuable descriptor records it with the resolved provider and model so a child that has not logged its first request can cold-resume with the same selection.
+Provider and model form one route and must be supplied together. An effort may be supplied alone when configured, parent, or provider-owned route defaults provide the effective route. Static `provider.agentRouteDefaults`, when present, establish the provider/model baseline; `Config.agentOptions` and model arguments overlay it before route-aware effort clearing. Providers without static defaults use compatible fields from the parent Agent's latest logged request selection, with creation options supplying the fallback before its first request and retaining the configured output-token limit. Reasoning-effort identifiers remain adapter-owned. An unchanged route inherits an omitted effort only from the selected baseline; changing provider or model without naming an effort clears the lower layer's route-owned value so the selected model resolves its own default. `AgentOptions` carries the resulting effort into the child loop, whose request header logs the effective value. A continuable descriptor records it with the resolved provider and model so a child that has not logged its first request can cold-resume with the same selection.

-An explicit or configured provider, model, or effort resolves through `ctx.llm.resolveCallConfig()` before child creation. That lookup owns provider registration, exact-model metadata, reasoning-effort validation, and adapter defaults. The tool checks cancellation again after the asynchronous lookup and before creating a child or background job. Calls with no model-facing selection and no configured route fields preserve the existing provider path without requiring the optional LLM service.
+An explicit or configured provider, model, or effort resolves through `ctx.llm.resolveCallConfig()` after the provider baseline and request precedence are complete. Providers with static route defaults suppress parent-effort inheritance when the request omits effort, preserving the selected model's default. The LLM lookup owns provider registration, exact-model metadata, reasoning-effort validation, and adapter defaults. After the asynchronous lookup, the tool checks cancellation and confirms the same provider instance remains registered before creating a child or background job, so HMR cannot combine one provider's defaults with another provider's process. Calls with no model-facing selection and no configured route fields preserve the existing provider path without requiring the optional LLM service.

 An enabled definition registers `list_subagent_models`. With no arguments the tool lists registered providers; with `provider` it calls that adapter's advisory model catalog; with `provider` and `model` it resolves the exact model and returns its reasoning efforts and default. At most one instance in a tool scope enables selection because the discovery name is global. Shipped product compositions put `modelSelectionSettings: true` on the primary Agent-scoped `subagent` instance and register the Host-owned `subagent-model-selection` settings namespace with `enabled: false`. A new top-level Session samples that preference during composition and logs an enabled decision as `subagent/model-selection-enabled` before any model request. A child Session inherits the live parent's decision, and a resumed Session uses its existing marker instead of the current preference. Therefore a settings edit affects only subsequently composed top-level Sessions. The fixed discovery definition remains available without the optional LLM service, while discovery and selected-route calls fail until that service is present. An unlisted model remains selectable when the adapter accepts its id.

@@ -24,7 +24,7 @@ Shipped `subagent_fork` instances leave `enableModelSelection` disabled even tho

 The delegation definition is static across adapter registration and catalog changes, so live topology neither expands every parent request nor invalidates its cache prefix. The discovery result enters the transcript only when called. A custom inheritance-capable instance that enables selection warns that changing provider or model can prevent provider-side reuse of the inherited conversation prefix.

-`SubagentCapabilities.agentOptions` remains the transport truth. The service rejects a request carrying those options before calling a provider that advertises `false`. Both in-process providers advertise `true`; the current ACP, Codex, Claude Code, and DSH SDK transports advertise `false`. Tool configuration that supplies `agentOptions`, statically enables model selection, or makes it settings-controlled also fails when its bound provider lacks the capability.
+`SubagentCapabilities.agentOptions` remains the transport truth. The service rejects a request carrying those options before calling a provider that advertises `false`. Both in-process providers and the DSH SDK transport advertise `true`; DSH SDK publishes its provider/model defaults as detached immutable data for Consumer preflight, while `start()` independently applies the same Config defaults plus maxTokens for direct callers and child initialization. ACP, Codex, and Claude Code advertise `false`. Tool configuration that supplies `agentOptions`, statically enables model selection, or makes it settings-controlled also fails when its bound provider lacks the capability.

 ## Alternatives considered

@@ -51,10 +51,10 @@ The delegation definition is static across adapter registration and catalog chan
 - An enabled delegation tool can select any live child LLM route without deployment selector configuration; disabled instances omit and reject model-facing route fields.
 - The primary delegation-tool instance defaults selection off, exposes a Models-page opt-in for new Sessions, and registers `list_subagent_models` only in Sessions whose durable decision is enabled; its catalog rows do not restrict delegation.
 - Shipped fork tools inherit the parent's provider and model and omit model-facing route fields so the inherited conversation prefix remains eligible for KV Cache reuse.
- Omission retains configured defaults and compatible inheritance from the parent's latest logged request; a route change without an explicit effort uses the selected model's default.
+- Omission retains configured defaults plus static provider route defaults or compatible parent inheritance; a route change without an explicit effort uses the selected model's default.
 - Adapter catalog and topology changes leave the delegation definition and its prompt-cache prefix unchanged.
- Out-of-process subagent providers reject configured and model-selected Agent options until they implement and advertise the capability.
- Unit coverage owns the default-off Host preference, new-Session sampling, child inheritance, resumed decisions, opt-in schema and execution enforcement, merge precedence, route-aware effort inheritance, preflight cancellation, live discovery, diagnostics, definition stability, capability rejection, and optional-service behavior. A shipped headless snapshot pins inheritance from a logged parent selection; the shipped examples also own the assembled keyless model-visible schemas.
+- DSH SDK children accept configured and model-selected Agent routes; ACP, Codex, and Claude Code reject them until they implement and advertise the capability.
+- Unit coverage owns the default-off Host preference, new-Session sampling, child inheritance, resumed decisions, opt-in schema and execution enforcement, merge precedence, route-aware effort inheritance, preflight cancellation, live discovery, diagnostics, definition stability, capability rejection, and optional-service behavior. A shipped headless snapshot pins inheritance from a logged parent selection; the shipped examples own the assembled keyless model-visible schemas, and the SDK Loader and snapshot evidence pin the complete route through a separate child runtime.

 ## Related decisions

@@ -14,9 +14,9 @@ Status: implemented

 只有实例启用 `enableModelSelection`，或其 Agent 作用域的 `modelSelectionSettings` 实例解析出已启用的 Session 决定，且绑定的 subagent 提供方声明 `SubagentCapabilities.agentOptions` 时，`dsh-tool-subagent` 才公开可选的 `provider`、`model` 与 `reasoning_effort` 字段，不要求配置路由允许列表。已注册的 LLM 提供方路由都可供子级选择；本工具不会在部署的 LLM 注册表之上增加第二套授权策略。禁用的实例会省略并拒绝面向模型的选择，而配置的 `Config.agentOptions` 仍是部署方所有的默认值。如果提供方缺少该能力，任一种选择模式都会使插件挂载失败。

-提供方与模型共同组成一条路由，必须一起提供。如果配置值或父级值能够提供生效路由，则可以只提供推理强度。模型参数覆盖 `Config.agentOptions`，配置字段覆盖父 Agent 最新记录的请求选择；首个请求之前由创建选项提供回退，并保留其中配置的输出 token 上限。推理强度 ID 仍由 adapter 所有。路由不变时会继承省略的强度；更换提供方或模型但没有指定强度时，会清除下层路由自有的值，使所选模型解析自己的默认值。`AgentOptions` 把结果强度传入子级循环，其请求 header 会记录生效值。可继续描述符会把它与解析后的提供方和模型一同记录，使尚未写入首个请求的子级能以相同选择冷恢复。
+提供方与模型共同组成一条路由，必须一起提供。如果配置值、父级值或提供方持有的路由默认值能够提供生效路由，则可以只提供推理强度。静态的 `provider.agentRouteDefaults` 在存在时构成 provider／model 基线；`Config.agentOptions` 与模型参数会在路由相关强度清除之前覆盖它。没有静态默认值的提供方会使用父 Agent 最新记录请求中的兼容字段，首个请求之前由创建选项提供回退，并保留其中配置的输出 token 上限。推理强度 ID 仍由 adapter 所有。只有所选基线的路由不变时才会继承省略的强度；更换提供方或模型但没有指定强度时，会清除下层路由自有的值，使所选模型解析自己的默认值。`AgentOptions` 把结果强度传入子级循环，其请求 header 会记录生效值。可继续描述符会把它与解析后的提供方和模型一同记录，使尚未写入首个请求的子级能以相同选择冷恢复。

-显式或配置的提供方、模型或强度会在创建子级前通过 `ctx.llm.resolveCallConfig()` 解析。该查询负责提供方注册、精确模型元数据、推理强度校验和 adapter 默认值。异步查询完成后、创建子级或后台 job 之前，工具会再次检查取消状态。既没有面向模型的选择、也没有配置路由字段的调用会保留原有提供方路径，不要求可选 LLM 服务存在。
+显式或配置的提供方、模型或强度会在提供方基线与请求优先级完成后，通过 `ctx.llm.resolveCallConfig()` 解析。具有静态路由默认值的提供方会在请求省略强度时禁止继承父级强度，从而保留所选模型的默认值。LLM 查询负责提供方注册、精确模型元数据、推理强度校验和 adapter 默认值。异步查询完成后、创建子级或后台 job 之前，工具会再次检查取消状态，并确认同一个提供方实例仍处于注册状态，因此 HMR 不会把一个提供方的默认值与另一个提供方的进程组合。既没有面向模型的选择、也没有配置路由字段的调用会保留原有提供方路径，不要求可选 LLM 服务存在。

 启用的定义会注册 `list_subagent_models`。无参数调用列出已注册提供方；提供 `provider` 时调用该适配器的建议性模型目录；同时提供 `provider` 与 `model` 时解析精确模型，并返回其推理强度和默认值。因为发现工具使用全局名称，一个工具作用域最多由一个实例启用选择。随附产品组合在 Agent 作用域的主 `subagent` 实例上设置 `modelSelectionSettings: true`，并注册默认 `enabled: false` 的 Host 自有 `subagent-model-selection` settings namespace。新的顶层 Session 会在组合期间读取该偏好，并在任何模型请求之前把启用决定记录为 `subagent/model-selection-enabled`。子 Session 继承在线父级的决定；恢复的 Session 使用已有标记，而不是当前偏好。因此，设置修改只影响之后组合的顶层 Session。即使缺少可选 LLM 服务，固定发现定义仍保持可用；发现调用和所选路由调用会在该服务出现前失败。只要适配器接受某个未列出的模型 ID，仍可选择该模型。

@@ -24,7 +24,7 @@ Status: implemented

 委派定义不会随 adapter 注册和目录变化而改变，因此实时拓扑既不会扩大每个父级请求，也不会使缓存前缀失效。只有调用发现工具时，目录结果才进入 transcript。自定义的上下文继承实例如果启用选择，其描述会警告，更改提供方或模型可能阻止提供方复用继承的对话前缀。

-`SubagentCapabilities.agentOptions` 仍是传输事实。如果请求携带这些选项，而提供方声明为 `false`，服务会在调用提供方前拒绝。两个进程内提供方声明为 `true`；当前 ACP、Codex、Claude Code 与 DSH SDK 传输声明为 `false`。工具配置提供 `agentOptions`、静态启用模型选择或让它受 settings 控制时，如果绑定的提供方缺少该能力，也会失败。
+`SubagentCapabilities.agentOptions` 仍是传输事实。如果请求携带这些选项，而提供方声明为 `false`，服务会在调用提供方前拒绝。两个进程内提供方与 DSH SDK 传输声明为 `true`；DSH SDK 会把 provider／model 默认值作为分离且不可变的数据公开给 Consumer 预检，而 `start()` 会为直接调用方与子运行时初始化独立应用同一份 Config 默认值及 maxTokens。ACP、Codex 与 Claude Code 声明为 `false`。工具配置提供 `agentOptions`、静态启用模型选择或让它受 settings 控制时，如果绑定的提供方缺少该能力，也会失败。

 ## 考虑过的替代方案

@@ -51,10 +51,10 @@ Status: implemented
 - 启用的委派工具无需部署选择器配置，即可选择任意实时子级 LLM 路由；禁用的实例会省略并拒绝面向模型的路由字段。
 - 主委派工具实例默认关闭选择，为新 Session 提供 Models 页面 opt-in，并且只在持久决定已启用的 Session 中注册 `list_subagent_models`；其目录条目不会限制委派。
 - 随附 fork 工具会继承父级的提供方与模型，并省略面向模型的路由字段，使继承的对话前缀仍可供 KV Cache 复用。
- 省略选择时保留配置默认值，并从父级最新记录的请求中进行兼容继承；改变路由但不显式指定强度时，使用所选模型的默认值。
+- 省略选择时保留配置默认值，并使用静态提供方路由默认值或来自父级最新记录请求的兼容继承；改变路由但不显式指定强度时，使用所选模型的默认值。
 - adapter 目录和拓扑变化不会改变委派定义及其 prompt 缓存前缀。
- 进程外 subagent 提供方在实现并声明该能力前，会拒绝配置和模型选择的 Agent 选项。
- 单元测试覆盖默认关闭的 Host 偏好、新 Session 读取、子级继承、恢复决定、选择启用时的 schema 与执行强制、合并优先级、路由相关强度继承、预检取消、实时发现、诊断、定义稳定性、能力拒绝与可选服务行为。随附的 headless 快照固定从父级已记录选择继承的行为；随附示例还覆盖组装后无密钥、模型可见的 schema。
+- DSH SDK 子级接受配置和模型选择的 Agent 路由；ACP、Codex 与 Claude Code 在实现并声明该能力前仍会拒绝。
+- 单元测试覆盖默认关闭的 Host 偏好、新 Session 读取、子级继承、恢复决定、选择启用时的 schema 与执行强制、合并优先级、路由相关强度继承、预检取消、实时发现、诊断、定义稳定性、能力拒绝与可选服务行为。随附的 headless 快照固定从父级已记录选择继承的行为；随附示例覆盖组装后无密钥、模型可见的 schema，SDK Loader 与快照证据固定完整路由经过独立子运行时的链路。

 ## 相关决策

@@ -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/2026-08-18-product-subagent-failure-facts.md
-2026-08-18-product-subagent-failure-facts.md: 50d8e918f288a6b8a9b90474499b2ed20731643f
-2026-08-18-product-subagent-failure-facts.zh.md: abdbf8c0ebddb1fb30cef3c7e80fcf7040e45d2d
+2026-08-18-product-subagent-failure-facts.md: e47fc3246ce20e988cd693b2eb04f785d37bcb27
+2026-08-18-product-subagent-failure-facts.zh.md: 5efd15e0e22d38b0100a96fc020ba3671fd257ed
@@ -12,7 +12,7 @@ Copying SDK error text, app-server payloads, or stderr into the result would exp

 ## Decision

-Each product Provider owns the mapping from its pinned official error union, current operation, and managed process outcome to one fixed safe diagnostic line. `SubagentResult` remains unchanged: consumers receive the existing bounded `diagnostic` string and do not parse its product-private fields.
+Each product Provider owns the mapping from its pinned official structured failures, current operation, and managed process outcome to one fixed safe diagnostic line. `SubagentResult` remains unchanged: consumers receive the existing bounded `diagnostic` string and do not parse its product-private fields. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) supersedes this note's complete Claude Code subtype mirror; this note continues to own the current detailed Codex categories until that provider adopts the same simplification.

 ### Safe diagnostic

@@ -28,14 +28,7 @@ Successful results and local cancellation expose no failure fact. Raw product er

 ### Claude Code facts

-Agent SDK 0.3.220 defines four error subtypes: `error_during_execution`, `error_max_turns`, `error_max_budget_usd`, and `error_max_structured_output_retries`. The Claude Code Provider preserves each exact subtype as the category while keeping the shared stop reason `error`. An error-marked or blank success uses `invalid-success`, a missing result uses `missing-result`, a process exit before an SDK terminal result uses `process-exit`, and an unrecognized value or exception uses `unknown` without copying the value.
-
-| Stage | Owned operation | Observable failure |
-| --- | --- | --- |
-| `query-start` | SDK query construction, native platform-payload startup, and unpublished rollback | `start()` rejects with fixed safe facts and any process outcome observed before rollback |
-| `query-run` | Published SDK message iteration and strict terminal-result validation | The run resolves as `error` with the exact known subtype or a fixed result category |
-| `process` | Managed CLI exits before the SDK supplies a terminal result | The run resolves as `error` with `process-exit` and the available exit code and signal |
-| `teardown` | Query close and managed process-tree release | `dispose()` rejects independently with fixed safe facts after cleanup still reaches its final exit wait |
+The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) exclusively owns Claude Code categories, stages, process facts, permission ordering, and verification for Agent SDK 0.3.241 and Claude Code 2.1.241. This note carries no separate Claude category contract.

 ### Codex facts

@@ -56,7 +49,7 @@ Codex app-server 0.147.0 defines eleven string categories and five object varian

 | Fact or resource | Owner | Consumer behavior |
 | --- | --- | --- |
-| Product error category | Pinned official SDK or app-server version | The Provider maps only the declared structured union and uses `unknown` outside it |
+| Codex error category | Codex Provider over its pinned official app-server | The Provider preserves its current structured category and uses `unknown` outside the recognized set |
 | Current failure stage | Product Provider operation | Derived at the failure site; never persisted or used as a recovery state |
 | Exit code and signal | `dsh-subprocess` process handle | The Provider displays observed values without inferring missing ones |
 | Diagnostic bytes and delivery | `dsh-subagent`, foreground tool, and Job runtime | The same bounded text is presented separately from assistant output in both scheduling modes |
@@ -64,7 +57,7 @@ Codex app-server 0.147.0 defines eleven string categories and five object varian

 ## Verification

-Claude Code package tests pin all four SDK subtypes, invalid success, missing result, unknown values and exceptions, all four stages, independent exit code and signal fields, permission-fact ordering, sanitization, successful-result and cancellation omission, concurrent-run isolation, and cleanup completion. Codex package tests pin all sixteen error-info variants, HTTP status presence and absence, all six stages, unknown fallback, stop-reason preservation, permission ordering, sanitization, cancellation, concurrency, and cleanup aggregation. The real SDK/CLI fixture produces an actual Claude `error_max_turns`; the real app-server fixture produces an actual Codex `internalServerError`; both fixtures cover process/protocol failure and whole-tree quiescence. The keyless ACP snapshot records each product's exact diagnostic in foreground error output, a background completion notice, and `job_output`.
+Claude Code verification is owned by the [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md). Codex package tests pin all sixteen current error-info variants, HTTP status presence and absence, all six stages, unknown fallback, stop-reason preservation, permission ordering, sanitization, cancellation, concurrency, and cleanup aggregation. The real app-server fixture produces an actual Codex `internalServerError` and covers process/protocol failure and whole-tree quiescence. The keyless ACP snapshot records the Codex diagnostic in foreground error output, a background completion notice, and `job_output`.

 ## Alternatives considered

@@ -80,8 +73,8 @@ Claude Code package tests pin all four SDK subtypes, invalid success, missing re

 ## Consequences

-The parent can distinguish important Claude Code limits and Codex budget, usage, service, policy, request, connection, stream, rollback, sandbox, and active-turn failures without receiving raw product text. Foreground and background scheduling preserve the same fact because both consume one `SubagentResult`.
+The parent can distinguish the current Codex budget, usage, service, policy, request, connection, stream, rollback, sandbox, and active-turn categories without receiving raw product text. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns the corresponding Claude result. Foreground and background scheduling preserve the same fact because both consume one `SubagentResult`.

-The diagnostic is display text rather than a new public protocol. Callers may present it but must not branch on its punctuation or product-private category names. A pinned product-version upgrade must update the Provider mapping and evidence when its official error union changes.
+The diagnostic is display text rather than a new public protocol. Callers may present it but must not branch on its punctuation or product-private category names. A pinned product-version upgrade revalidates the Provider mapping and evidence without requiring every official error member to remain model-visible.

 This decision adds no product session persistence, retry policy, recovery state, stderr classifier, authentication or configuration taxonomy, progress stream, or human interaction path.
@@ -12,7 +12,7 @@ Status: implemented

 ## Decision

-每个产品提供方分别拥有从锁定版本官方错误联合、当前操作和受管进程结果到一行固定安全诊断的映射。`SubagentResult` 保持不变：消费方仍接收现有的有界 `diagnostic` 字符串，而且不解析其中由产品私有的字段。
+每个产品提供方分别拥有从锁定版本官方结构化失败、当前操作和受管进程结果到一行固定安全诊断的映射。`SubagentResult` 保持不变：消费方仍接收现有的有界 `diagnostic` 字符串，而且不解析其中由产品私有的字段。[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)已经取代本说明对 Claude Code 完整 subtype 的镜像；在 Codex 采用同一简化前，本说明继续负责其当前详细类别。

 ### 安全诊断

@@ -28,14 +28,7 @@ Product subagent failure (product: <product>; stage: <stage>; category: <categor

 ### Claude Code 事实

-Agent SDK 0.3.220 定义四种错误子类型：`error_during_execution`、`error_max_turns`、`error_max_budget_usd` 和 `error_max_structured_output_retries`。Claude Code 提供方会把每种准确子类型保留为类别，同时维持共享终止原因 `error`。标记为错误或内容空白的成功消息使用 `invalid-success`，缺失结果使用 `missing-result`，SDK 给出终态结果前发生的进程退出使用 `process-exit`，无法识别的值或异常使用 `unknown`，且不会复制原值。
-
-| 阶段 | 归属操作 | 可观察失败 |
-| --- | --- | --- |
-| `query-start` | SDK query 构造、原生平台载荷启动与未发布回滚 | `start()` 以固定安全事实和回滚前已观测到的进程结果拒绝 |
-| `query-run` | 已发布 SDK 消息迭代与严格终态结果校验 | 运行以 `error` 兑现，并携带准确已知子类型或固定结果类别 |
-| `process` | SDK 提供终态结果之前受管 CLI 已退出 | 运行以 `error` 兑现，并携带 `process-exit` 以及可用的退出码和信号 |
-| `teardown` | Query 关闭与受管进程树释放 | `dispose()` 独立拒绝并携带固定安全事实，同时清理仍会完成最终退出等待 |
+[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)独占负责 Agent SDK 0.3.241 与 Claude Code 2.1.241 的 Claude Code 类别、阶段、进程事实、权限顺序与验证。本说明不再承载独立的 Claude 类别约定。

 ### Codex 事实

@@ -56,7 +49,7 @@ Codex app-server 0.147.0 定义十一种字符串类别与五种对象 variant

 | 事实或资源 | Owner | 消费方行为 |
 | --- | --- | --- |
-| 产品错误类别 | 锁定版本的官方 SDK 或 app-server | 提供方只映射已声明的结构化联合，并对联合外值使用 `unknown` |
+| Codex 错误类别 | Codex 提供方及其锁定的官方 app-server | 提供方保留当前结构化类别，并在已识别集合之外使用 `unknown` |
 | 当前失败阶段 | 产品提供方操作 | 只在失败点派生；绝不持久化，也不作为恢复状态 |
 | 退出码与信号 | `dsh-subprocess` 进程句柄 | 提供方展示已观测值，不推测缺失值 |
 | 诊断字节与送达 | `dsh-subagent`、前台工具与 Job 运行时 | 两种调度模式都把同一份有界文本与 assistant 输出分开呈现 |
@@ -64,7 +57,7 @@ Codex app-server 0.147.0 定义十一种字符串类别与五种对象 variant

 ## Verification

-Claude Code 包测试固定四种 SDK 子类型、无效成功、缺失结果、未知值与异常、四个阶段、相互独立的退出码与信号字段、权限事实顺序、脱敏、成功结果与取消时省略诊断、并发运行隔离和清理完成。Codex 包测试固定全部十六种 error-info variant、HTTP status 存在与缺失、六个阶段、unknown 回退、终止原因保持不变、权限顺序、脱敏、取消、并发与清理聚合。真实 SDK/CLI fixture 会产生真实的 Claude `error_max_turns`，真实 app-server fixture 会产生真实的 Codex `internalServerError`；两个 fixture 都覆盖进程／协议失败与整棵进程树完全停稳。无密钥 ACP snapshot 会在前台错误输出、后台完成通知和 `job_output` 中记录两个产品各自的准确诊断。
+Claude Code 验证由[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责。Codex 包测试固定当前全部十六种 error-info variant、HTTP status 存在与缺失、六个阶段、unknown 回退、终止原因保持不变、权限顺序、脱敏、取消、并发与清理聚合。真实 app-server fixture 会产生实际 Codex `internalServerError`，并覆盖进程／协议失败与整棵进程树完全停稳。无密钥 ACP snapshot 会在前台错误输出、后台完成通知和 `job_output` 中记录 Codex 诊断。

 ## Alternatives considered

@@ -80,8 +73,8 @@ Claude Code 包测试固定四种 SDK 子类型、无效成功、缺失结果、

 ## Consequences

-父 agent 可以区分重要的 Claude Code 限制，以及 Codex 预算、用量、服务、策略、请求、连接、stream、回滚、sandbox 和 active-turn 失败，而不会收到原始产品文本。前台与后台调度会保留同一事实，因为二者都消费同一个 `SubagentResult`。
+父 agent 可以区分当前 Codex 的预算、用量、服务、策略、请求、连接、stream、回滚、sandbox 与 active-turn 类别，而不会收到原始产品文本。[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)负责对应的 Claude 结果。前台与后台调度会保留同一事实，因为二者都消费同一个 `SubagentResult`。

-诊断只是展示文本，不是新的公开协议。调用方可以呈现它，但不得根据其标点或产品私有类别名称进行分支。锁定产品版本升级并改变官方错误联合时，必须同步更新提供方映射与证据。
+诊断只是展示文本，不是新的公开协议。调用方可以呈现它，但不得根据其标点或产品私有类别名称进行分支。锁定产品版本升级时必须重新验证提供方映射与证据，但不要求每个官方错误成员都继续模型可见。

 本决策不增加产品会话持久化、重试策略、恢复状态、stderr 分类器、身份验证或配置分类体系、进度流或人工交互路径。
@@ -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/2026-08-18-product-subagent-named-instances.md
-2026-08-18-product-subagent-named-instances.md: 759d3941ff8404138954c409f0fd4949e357200e
-2026-08-18-product-subagent-named-instances.zh.md: 6faf0e70f639cbc6528e27b800b8e5f99f0d6c86
+2026-08-18-product-subagent-named-instances.md: fd4f95dc9de0cc03a95f8f1dcbd6dcded258dc5f
+2026-08-18-product-subagent-named-instances.zh.md: 1d8395a21824501c5f26bdbec9c9d7f25d3bf20f
@@ -6,17 +6,17 @@ English | [中文](2026-08-18-product-subagent-named-instances.zh.md)

 ## Problem

-A Profile can mount one Cordis plugin package in multiple rows, but the Codex and Claude Code product providers previously registered every row under one fixed product name. A second row therefore failed as a duplicate before its distinct permission mode, environment, or process-release settings could become usable. Deriving an implicit name from those settings would create a second identity rule, while choosing a provider during a tool call would let model input select deployment authority.
+A Profile can mount one Cordis plugin package in multiple rows, but the Codex and Claude Code product providers previously registered every row under one fixed product name. A second row therefore failed as a duplicate before its distinct model where supported, permission mode, environment, or process-release settings could become usable. Deriving an implicit name from those settings would create a second identity rule, while choosing a provider during a tool call would let model input select deployment authority.

 The existing subagent registry already owns unique provider names, reversible registration, lifecycle events, and holder-owned published runs. The existing `dsh-tool-subagent` configuration already binds one provider name to one model-visible tool name. Product providers need to expose the missing Profile-owned identity without adding another registry or selection protocol.

 ## Decision

-Each product provider Config owns a non-empty `providerName`; the defaults remain `codex` and `claude-code`. The resolved name is fixed when the plugin row loads and becomes the Provider object's `name`; registration, lookup, lifecycle events, run logs, and HMR removal therefore use the same value. Each mounted row retains its own `permissionMode`, `env`, `disposeGraceMs`, and run resources.
+Each product provider Config owns a non-empty `providerName`; the defaults remain `codex` and `claude-code`. The resolved name is fixed when the plugin row loads and becomes the Provider object's `name`; registration, lookup, lifecycle events, run logs, and HMR removal therefore use the same value. Each mounted row retains its own `permissionMode`, `env`, `disposeGraceMs`, and run resources. Claude Code rows also retain an optional non-empty `model`; an explicit value is fixed for every query from that instance, while omission leaves native settings authoritative.

 Profiles may mount multiple Codex or Claude Code rows when every row uses a distinct `providerName`. Each `dsh-tool-subagent` row continues to bind its existing `provider` field to that exact name and exposes an independently configured `toolName`. Tool calls carry no provider selector, alias, or permission input. A duplicate provider name fails through the existing `DUPLICATE_PROVIDER` path and leaves the first registration intact.

-Removing one provider row blocks new starts and removes only tools bound to that name. Runs already published by the removed instance remain owned by their holders and settle or dispose independently. Sibling instances remain registered and keep their own environment, native permission mode, cancellation controller, product process, and cleanup grace.
+Removing one provider row blocks new starts and removes only tools bound to that name. Runs already published by the removed instance remain owned by their holders and settle or dispose independently. Sibling instances remain registered and keep their own configured model where supported, environment, native permission mode, cancellation controller, product process, and cleanup grace.

 ### Ownership and lifecycle

@@ -25,11 +25,11 @@ Removing one provider row blocks new starts and removes only tools bound to that
 | Provider instance name | Product Provider Config | One immutable registry name per mounted row, with the existing default when omitted |
 | Name uniqueness and lifecycle events | `ctx.subagents` | Duplicate registration fails; disposal removes only the matching name |
 | Model-visible tool name and binding | `dsh-tool-subagent` Config | One static tool resolves one configured provider name |
-| Permission, environment, and process cleanup | One Provider instance | Concurrent runs and sibling instances do not share deployment configuration or run resources |
+| Model where supported, permission, environment, and process cleanup | One Provider instance | Concurrent runs and sibling instances do not share deployment configuration or run resources |

 ## Verification

-Both product packages pin their default and custom names, empty-name rejection, duplicate rollback, actual-name diagnostics, two concurrent instances with different permission modes, environments, and cleanup grace, cancellation isolation, and removal of one instance while its published run remains valid. The official product loopback tests run two named instances in one Host against separate model fixtures and prove independent unload and process-tree quiescence. Public Loader compositions mount two rows and two distinct tools for each product without starting either product, while keyless ACP snapshots pin the four-tool combined roster and the absence of a dynamic provider parameter.
+Both product packages pin their default and custom names, empty-name rejection, duplicate rollback, actual-name diagnostics, two concurrent instances with different permission modes, environments, and cleanup grace, cancellation isolation, and removal of one instance while its published run remains valid. Claude Code additionally pins empty-model rejection, omitted-model inheritance, two explicit models, and exact per-query isolation through the real SDK/CLI fixture. Public Loader compositions mount two rows and two distinct tools for each product without starting either product, while keyless ACP snapshots pin the four-tool combined roster and the absence of dynamic provider or model parameters.

 ## Alternatives considered

@@ -43,6 +43,6 @@ Both product packages pin their default and custom names, empty-name rejection,

 ## Consequences

-A Profile can expose several Codex and Claude Code tools backed by separate native permission modes and environments while existing configurations continue to resolve `codex` and `claude-code`. Provider and tool names remain independent configuration facts, so changing one requires updating the binding that refers to it.
+A Profile can expose several Codex and Claude Code tools backed by separate native permission modes and environments, and Claude Code instances may additionally fix separate models, while existing configurations continue to resolve `codex` and `claude-code`. Provider and tool names remain independent configuration facts, so changing one requires updating the binding that refers to it.

-The design adds no runtime renaming, model-visible selector, generated tool name, persistent instance directory, shared process pool, or compatibility alias. Correct multi-instance configurations require unique provider names and unique tool names; duplicate tool-name waiting remains a separate limitation.
+The design adds no runtime renaming, model-visible provider or model selector, generated tool name, persistent instance directory, shared process pool, model discovery, fallback, or compatibility alias. Correct multi-instance configurations require unique provider names and unique tool names; duplicate tool-name waiting remains a separate limitation.
@@ -6,17 +6,17 @@ Status: implemented

 ## 问题

-Profile 可以用多个配置项挂载同一个 Cordis 插件包，但 Codex 与 Claude Code 产品提供方此前会把每个配置项都注册到一个固定产品名称下。因此，第二个配置项会在其独立权限模式、环境或进程释放设置可用前因名称重复而失败。根据这些设置隐式派生名称会建立第二套身份规则，而在工具调用期间选择提供方会让模型输入决定部署权限。
+Profile 可以用多个配置项挂载同一个 Cordis 插件包，但 Codex 与 Claude Code 产品提供方此前会把每个配置项都注册到一个固定产品名称下。因此，第二个配置项会在其支持时使用的独立模型、权限模式、环境或进程释放设置可用前因名称重复而失败。根据这些设置隐式派生名称会建立第二套身份规则，而在工具调用期间选择提供方会让模型输入决定部署权限。

 现有 subagent 注册表已经拥有提供方名称唯一性、可逆注册、生命周期事件和由持有方拥有的已发布运行。现有 `dsh-tool-subagent` 配置也已经把一个提供方名称绑定到一个模型可见工具名称。产品提供方只需公开缺失的 Profile 所有身份，无需增加另一套注册表或选择协议。

 ## 决策

-每个产品提供方 Config 都拥有非空的 `providerName`；默认值仍分别为 `codex` 与 `claude-code`。插件配置项加载时会固定解析后的名称，并把它作为 Provider 对象的 `name`；注册、查找、生命周期事件、运行日志和 HMR（热模块替换）移除因此使用同一个值。每个已挂载配置项保留自己的 `permissionMode`、`env`、`disposeGraceMs` 和运行资源。
+每个产品提供方 Config 都拥有非空的 `providerName`；默认值仍分别为 `codex` 与 `claude-code`。插件配置项加载时会固定解析后的名称，并把它作为 Provider 对象的 `name`；注册、查找、生命周期事件、运行日志和 HMR（热模块替换）移除因此使用同一个值。每个已挂载配置项保留自己的 `permissionMode`、`env`、`disposeGraceMs` 和运行资源。Claude Code 配置项还会保留可选的非空 `model`；显式值为该实例的每次 query 固定，省略时仍以原生设置为权威。

 当每个配置项使用不同的 `providerName` 时，Profile 可以挂载多个 Codex 或 Claude Code 配置项。每个 `dsh-tool-subagent` 配置项继续用已有的 `provider` 字段绑定这个准确名称，并公开独立配置的 `toolName`。工具调用不携带提供方选择器、别名或权限输入。重复提供方名称沿用现有 `DUPLICATE_PROVIDER` 路径失败，而且不会替换第一个注册项。

-移除一个提供方配置项会阻止新的启动，并且只移除绑定到该名称的工具。该实例已经发布的运行仍由其持有方拥有，并会独立结算或 dispose（资源释放）。兄弟实例继续保持注册，并保留各自的环境、原生权限模式、取消控制器、产品进程和清理宽限期。
+移除一个提供方配置项会阻止新的启动，并且只移除绑定到该名称的工具。该实例已经发布的运行仍由其持有方拥有，并会独立结算或 dispose（资源释放）。兄弟实例继续保持注册，并保留各自支持时配置的模型、环境、原生权限模式、取消控制器、产品进程和清理宽限期。

 ### 所有权与生命周期

@@ -25,11 +25,11 @@ Profile 可以用多个配置项挂载同一个 Cordis 插件包，但 Codex 与
 | 提供方实例名称 | 产品提供方 Config | 每个已挂载配置项拥有一个不可变注册名称；省略时使用现有默认值 |
 | 名称唯一性与生命周期事件 | `ctx.subagents` | 重复注册失败；资源释放只移除匹配名称 |
 | 模型可见工具名称与绑定 | `dsh-tool-subagent` Config | 一个静态工具解析一个已配置的提供方名称 |
-| 权限、环境与进程清理 | 一个提供方实例 | 并发运行与兄弟实例不共享部署配置或运行资源 |
+| 支持时的模型、权限、环境与进程清理 | 一个提供方实例 | 并发运行与兄弟实例不共享部署配置或运行资源 |

 ## 验证

-两个产品包测试都会固定默认与自定义名称、空名称拒绝、重复注册回滚、实际名称诊断、使用不同权限模式、环境与清理宽限期的两个并发实例、取消隔离，以及移除一个实例后其已发布运行仍然有效。官方产品回环测试会在同一个 Host 中针对独立模型 fixture（测试前置数据）运行两个命名实例，并证明独立卸载与进程树完全停稳。公共 Loader 组合会为每个产品挂载两个配置项与两个不同工具，而且不启动任一产品；无密钥 ACP 快照固定最终四工具组合，并证明没有动态提供方参数。
+两个产品包测试都会固定默认与自定义名称、空名称拒绝、重复注册回滚、实际名称诊断、使用不同权限模式、环境与清理宽限期的两个并发实例、取消隔离，以及移除一个实例后其已发布运行仍然有效。Claude Code 还会固定空模型拒绝、省略模型继承、两个显式模型，以及通过真实 SDK／CLI fixture 证明逐 query 精确隔离。公共 Loader 组合会为每个产品挂载两个配置项与两个不同工具，而且不启动任一产品；无密钥 ACP 快照固定最终四工具组合，并证明没有动态提供方或模型参数。

 ## 考虑过的替代方案

@@ -43,6 +43,6 @@ Profile 可以用多个配置项挂载同一个 Cordis 插件包，但 Codex 与

 ## 结果

-Profile 可以公开多个由不同原生权限模式与环境支持的 Codex 与 Claude Code 工具，而现有配置仍会解析为 `codex` 与 `claude-code`。提供方名称与工具名称继续是彼此独立的配置事实，因此修改其中一项时必须同时更新引用它的绑定。
+Profile 可以公开多个由不同原生权限模式与环境支持的 Codex 与 Claude Code 工具，Claude Code 实例还可分别固定模型，而现有配置仍会解析为 `codex` 与 `claude-code`。提供方名称与工具名称继续是彼此独立的配置事实，因此修改其中一项时必须同时更新引用它的绑定。

-本设计不增加运行时改名、模型可见选择器、自动生成的工具名称、持久实例目录、共享进程池或兼容别名。正确的多实例配置要求提供方名称与工具名称都保持唯一；重复工具名称的等待问题仍是独立限制。
+本设计不增加运行时改名、模型可见的提供方或模型选择器、自动生成的工具名称、持久实例目录、共享进程池、模型发现、fallback 或兼容别名。正确的多实例配置要求提供方名称与工具名称都保持唯一；重复工具名称的等待问题仍是独立限制。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-21-headless-reasoning-progress.md
+2026-08-21-headless-reasoning-progress.md: 6c4a3574b63ef316fd456f24b473406054ed30f3
+2026-08-21-headless-reasoning-progress.zh.md: e19fffc33fb5a55432ba2b6cb50550a0cfbd00b8
@@ -0,0 +1,37 @@
+# Agent Note: headless streams provider reasoning to stderr
+
+Status: implemented
+
+English | [中文](2026-08-21-headless-reasoning-progress.zh.md)
+
+## Problem
+
+The one-shot headless runner waits for complete Agent quiescence before printing the final assistant text. Reasoning-capable providers already expose their reasoning as durable `assistant/chunk` events, but a long reasoned response leaves the terminal silent until the run completes. The final answer must remain the only stdout payload so command substitution and other consumers keep a stable result channel.
+
+The earlier [direct core entry-point decision](../architecture/2026-08-09-headless-direct-core-entry-point.md) required empty stderr on every successful run. That clause prevents live reasoning progress and is superseded by this note; its transport, durability, and completion decisions remain unchanged.
+
+## Decision
+
+`headless-runner` observes the exact Session it creates after startup quiescence and before submitting the task. Once the owned interval opens with `turn/start`, each non-empty `assistant/chunk.reasoning-delta` is written immediately to stderr. A contiguous reasoning phase starts with `dsh: reasoning:` on its own line; deltas retain provider order without token-boundary decoration. Reasoning block boundaries and usage metadata keep that phase open; a later non-reasoning block or output delta, stream finish, new turn, or listener disposal terminates it with one newline when the provider supplied none.
+
+This output is a transient projection of the existing durable Session event stream. The runner still derives final text and exit status from the flushed log rather than from progress-presentation state. The LLM adapter, agent loop, Session event types, persistence format, and SDK projections do not change.
+
+Reasoning progress is not TTY-gated and has no separate flag. A redirected stderr stream and a supervisor receive the same provider-reported content as an attached terminal. A successful run without reasoning still writes nothing to stderr; terminal model and driver errors keep their existing `dsh:` diagnostics after any open reasoning phase is terminated.
+
+## Verification
+
+The package test holds the Agent active after a reasoning delta and observes stderr before idle, then pins newline ownership for provider-terminated and unterminated phases plus terminal errors. The owner-local product expectation drives the shipped headless profile through a reasoning-plus-tool round and pins both stderr and the persisted Session. Recorded-session replay reconstructs expected stderr from scalar and packed chunk rows, closes sections on packed text and tool-call output, and uses the raw run log before fixture path tokenization in record modes. Built-bin acceptance sends `reasoning_content` through the native DeepSeek SSE adapter and requires reasoning on stderr while stdout remains the final answer.
+
+## Alternatives considered
+
+**Dump reasoning after quiescence.** Folding reasoning from the persisted log would preserve content but leave the terminal silent during the long-running interval that motivates the feature.
+
+**Wrap the LLM stream.** Tapping `ctx.llm.stream()` would place a presentation concern in the request path and duplicate the authoritative chunks that the agent loop already appends to the Session.
+
+**Print a spinner or periodic heartbeat.** A timer reports process liveness rather than provider progress, adds an interval policy, and still hides reasoning that the provider already supplies. Time before the first reasoning delta remains silent and can be addressed separately if providers buffer their first token.
+
+**Enable output only on a TTY or explicit flag.** Headless runs under CI and supervisors need the same progress signal, while implicit TTY-dependent behavior makes redirected runs differ from interactive runs. Callers that do not want reasoning logs redirect stderr.
+
+## Consequences
+
+Reasoning-capable successful runs now write provider-reported content to stderr, so log collectors may retain substantially more and potentially sensitive model output. Stdout remains one final assistant result, text-only success keeps stderr empty, errors remain line-separated, and no new configuration or durable format is introduced. Silence before the provider emits its first non-empty reasoning delta remains an explicit limitation.
@@ -0,0 +1,37 @@
+# Agent Note: headless 将提供方推理流式写入 stderr
+
+Status: implemented
+
+[English](2026-08-21-headless-reasoning-progress.md) | 中文
+
+## 问题
+
+一次性 headless runner 会等待 Agent（智能体）完全停稳，再打印最终 assistant 文本。具备推理能力的提供方已经把推理作为持久化的 `assistant/chunk` 事件暴露，但耗时较长的推理响应会让终端在运行完成前始终保持静默。最终答案必须继续作为 stdout 中唯一的载荷，使命令替换和其他消费方保持稳定的结果通道。
+
+此前的[直接使用核心服务入口决策](../architecture/2026-08-09-headless-direct-core-entry-point.zh.md)要求每次成功运行都保持 stderr 为空。该条款会阻止实时推理进度，因此由本 Agent Note 取代；其中关于传输、持久性与完成状态的其他决策保持不变。
+
+## 决策
+
+`headless-runner` 在启动工作完全停稳后、提交任务前，观察其创建的精确 Session。自身持有的区间以 `turn/start` 打开后，每个非空的 `assistant/chunk.reasoning-delta` 都会立即写入 stderr。一段连续推理以独占一行的 `dsh: reasoning:` 开始；各分片保持提供方顺序，不添加 token 边界装饰。推理块边界与用量元数据会保持该段打开；之后出现非推理块或输出分片、流结束、新轮次或 listener dispose（资源释放）时，如果提供方没有输出末尾换行，runner 会用一个换行终止该段。
+
+该输出是既有持久化会话事件流的瞬时投影。runner 仍从 flush 后的日志而不是进度呈现状态推导最终文本与退出状态。LLM（大语言模型）适配器、agent loop（智能体循环）、Session 事件类型、持久化格式与 SDK 投影均不改变。
+
+推理进度不按 TTY 启用，也没有单独 flag。重定向的 stderr 流与监督进程会收到和已连接终端相同的提供方报告内容。没有推理内容的成功运行仍不会写入 stderr；终止态模型错误与驱动器错误继续在任何已打开推理段终止后输出既有的 `dsh:` 诊断。
+
+## 验证
+
+包测试在推理分片后保持 Agent 活跃，并在 idle 前观察 stderr；测试同时固定由提供方终止和未终止的推理段换行归属，以及终止态错误。产品自有期望通过包含推理与工具调用的轮次驱动随附 headless profile，并固定 stderr 与持久化 Session。录制会话回放从标量及压缩分片记录重建预期 stderr，在压缩文本或工具调用输出处关闭推理段，并在录制模式下于 fixture 路径标记化之前使用原始运行日志。构建后二进制验收通过原生 DeepSeek SSE（Server-Sent Events）适配器发送 `reasoning_content`，要求推理出现在 stderr，同时 stdout 仍只包含最终答案。
+
+## 考虑过的替代方案
+
+**完全停稳后再输出推理。** 从持久化日志折叠推理能够保留内容，但在导致本功能产生的长时间运行区间内，终端仍会保持静默。
+
+**包装 LLM 流。** 截取 `ctx.llm.stream()` 会把呈现职责放入请求路径，并重复处理 agent loop 已经追加到 Session 的权威分片。
+
+**打印 spinner 或周期性心跳。** 定时器报告的是进程存活状态，而不是提供方进度；它还会新增间隔策略，并继续隐藏提供方已经给出的推理。首个推理分片前的时间仍保持静默；如果提供方会缓冲首个 token，可以另行处理。
+
+**仅在 TTY 或显式 flag 下启用输出。** CI 与监督进程中的 headless 运行需要相同的进度信号，而隐式依赖 TTY 会让重定向运行与交互式运行产生差异。不需要推理日志的调用方可以重定向 stderr。
+
+## 后果
+
+具备推理能力的成功运行会把提供方报告的内容写入 stderr，因此日志收集器可能保留明显更多且可能敏感的模型输出。stdout 仍只包含一个最终 assistant 结果，没有推理内容的成功运行保持 stderr 为空，错误继续与推理内容分行，并且本决策不引入新配置或持久化格式。提供方发出首个非空推理分片前保持静默，这是明确的限制。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-21-out-of-process-subagent-minimal-diagnostics.md
+2026-08-21-out-of-process-subagent-minimal-diagnostics.md: 533ace5a13df75fb594e0cecc65a743df27b6baf
+2026-08-21-out-of-process-subagent-minimal-diagnostics.zh.md: fe8adf764b240d77cfcde95999ee6689bf11b4a4
@@ -0,0 +1,73 @@
+# Agent Note: Out-of-process subagents expose minimal actionable diagnostics
+
+Status: implemented
+
+English | [中文](2026-08-21-out-of-process-subagent-minimal-diagnostics.zh.md)
+
+## Problem
+
+An ACP child can stop because it reached a remote limit, denied a required permission, lost its protocol transport, or exited as a process. The shared result historically reduced these outcomes to a stop reason such as `error`, while startup and cleanup rejection messages could expose the original exception. A parent could not choose between narrowing the task, adjusting permission policy, or repairing the child deployment without Host logs.
+
+Copying exceptions, stderr, task content, tool input, paths, environment values, credentials, or protocol payloads into `SubagentResult.diagnostic` would make untrusted child text model-visible. Reusing a complete product-specific error union would also duplicate independently versioned authorities in the provider-neutral [subagent seam](2026-06-21-subagent-capability-seam.md).
+
+## Decision
+
+Each out-of-process provider owns a small mapping from facts it already receives at its protocol and process lifecycle points to fixed safe display text. The ACP provider implements that rule from its closed stop reasons, current operation, closed tool kind, configured permission policy, selected permission outcome, and the managed subprocess exit code or signal. Consumers continue to use the existing optional `SubagentResult.diagnostic`; they do not parse its punctuation or provider-private category names.
+
+### Safe failure text
+
+Generic error diagnostics have this fixed field order:
+
+```text
+Subagent failure (provider: <provider>; stage: <stage>; category: <category>; stop reason: <reason>; exit code: <code>; signal: <signal>)
+```
+
+Unavailable optional fields are omitted. The complete result is limited to 4096 UTF-8 bytes by the shared settlement boundary. Successful results and local cancellation carry no failure diagnostic. Partial assistant output remains in `SubagentResult.output` and is presented separately.
+
+When an ACP permission request contributes to a non-completed result, a fixed line records `policy`, the closed ACP tool `request` kind, and `decision`. Tool titles, raw input, locations, option names, and metadata are excluded. For `max-tokens`, `refusal`, or remote `aborted`, the public stop reason already carries the terminal fact, so the permission line is the complete diagnostic; generic error paths append it after the failure line. A diagnostic-bearing remote `aborted` result keeps its public stop reason; the one-shot Job adapter treats it as failed, while diagnostic-free local cancellation remains killed.
+
+### ACP facts
+
+| Stage | Owned operation | Safe categories and facts |
+| --- | --- | --- |
+| `initialize` | Parent workspace resolution and ACP initialize | `configuration`, `transport`, or `process-exit` |
+| `new-session` | ACP `session/new` and returned session-id validation | `protocol`, `transport`, or `process-exit` |
+| `prompt` | ACP prompt request, remote stop reason, and permission callback | `remote-limit`, `transport`, `unknown`, or a permission-only diagnostic |
+| `process` | Child-process spawn failure, or a managed child exits before a prompt terminal response | `process-start`, or `process-exit` plus independently observed exit code and signal |
+| `teardown` | EOF quiescence and managed process-tree termination | Fixed teardown facts; the original cleanup failure remains internal |
+
+`max_turn_requests` remains the shared `error` stop reason and adds `remote-limit`. An unknown stop reason remains `error` and becomes the fixed `unknown` category without copying the value. `max_tokens`, `refusal`, and `cancelled` keep their existing shared stop reasons; they add a diagnostic only when a permission decision must be explained.
+
+### Ownership and lifecycle
+
+| Fact or resource | Owner | Consumer behavior |
+| --- | --- | --- |
+| ACP stop reason and tool kind | ACP server and SDK | The provider maps only closed values and uses fixed unknown fallbacks |
+| Current failure stage and latest permission decision | One ACP run | Derived at the failure point and discarded with the run; concurrent runs share no diagnostic state |
+| Exit code and signal | `dsh-subprocess` handle | Displayed only after the managed outcome is observed; stderr is never parsed |
+| Diagnostic bytes and presentation | `dsh-subagent`, foreground tool, and Job runtime | The same bounded text stays separate from assistant output in foreground and one-shot background modes |
+| Raw failure | Child runtime, Error cause chain, and Host logger | Available for Host diagnosis only, never copied into the parent model result |
+
+Startup publishes no run until initialize and new-session succeed. A startup failure rolls the private child back to quiescence before rejecting with safe facts. A published run settles its result without rejection, and `dispose()` independently reports a safe teardown failure while still using the backend's existing whole-tree cleanup ladder.
+
+## Verification
+
+ACP package tests drive a real stdio protocol child and pin every stop-reason mapping, remote-limit and unknown fallbacks, permission allow/deny facts, configuration, initialize, new-session, prompt, process, and teardown stages, startup rollback, successful-result and local-cancellation omission, partial output, concurrent-run isolation, Host-only raw errors, process quiescence, and the shared multibyte diagnostic limit. A Loader composition proves the real configured provider reaches the model-visible foreground result. The keyless ACP snapshot pins the same diagnostic and permission fact in foreground error output and one-shot background `job_output` detail.
+
+## Alternatives considered
+
+**Return raw exceptions, stderr, or protocol payloads.** These values can contain task content, tool input, paths, environment values, credentials, and upstream prose. Fixed allowlisted facts preserve the actionable distinction without expanding the model-visible trust boundary.
+
+**Add a shared structured error enum.** ACP and other process-backed providers own different lifecycle points and closed termination vocabularies. A shared enum would invent false equivalence and force unrelated consumers to track provider releases.
+
+**Parse exception messages or stderr into categories.** Free-form text is neither stable nor safe. Only closed protocol values, typed errors, current call sites, and managed process outcomes qualify as diagnostic inputs.
+
+**Change existing stop reasons.** The stop reason remains the provider-neutral terminal result. The optional diagnostic explains why a non-completed result needs a different next action without adding new public result states.
+
+**Add retries, recovery state, or interactive approval.** Diagnostics report a failure; they do not own remediation. Retry policy, session recovery, and human interaction require separate user contracts and lifecycle owners.
+
+## Consequences
+
+The parent can distinguish an ACP remote limit, permission involvement, protocol or transport failure, deployment/process failure, and teardown failure without receiving child-controlled text. Startup and cleanup errors use the same safe facts as published results, while Host observation retains the original cause.
+
+The diagnostic remains display text rather than a public protocol. Consumers may present it but must not branch on its format. This decision adds no retry policy, recovery controller, shared provider-error enum, stderr classifier, authentication taxonomy, session persistence, progress stream, or new ACP capability.
@@ -0,0 +1,73 @@
+# Agent Note: 进程外 subagent 公开最小可行动诊断
+
+Status: implemented
+
+[English](2026-08-21-out-of-process-subagent-minimal-diagnostics.md) | 中文
+
+## Problem
+
+ACP 子进程可能因为达到远端限制、拒绝必需权限、失去协议传输或进程退出而停止。共享结果以往只把这些结果压成 `error` 等结束原因，而启动和清理拒绝的消息还可能暴露原始异常。父 agent 若不读取 Host 日志，就无法决定应缩小任务、调整权限策略还是修复子运行时部署。
+
+若把异常、stderr、任务内容、工具输入、路径、环境值、凭证或协议 payload 复制进 `SubagentResult.diagnostic`，不受信任的子进程文本就会变成模型可见内容。若复用完整的产品专属错误联合，又会在提供方无关的 [subagent seam](2026-06-21-subagent-capability-seam.zh.md) 中复制彼此独立版本化的权威。
+
+## Decision
+
+每个进程外提供方分别拥有一份小型映射，把其协议与进程生命周期位置已经收到的事实转换成固定安全展示文本。ACP 提供方使用闭集结束原因、当前操作、闭集工具种类、已配置权限策略、选中的权限结果，以及受管子进程退出码或信号来实现该规则。消费方继续使用现有可选 `SubagentResult.diagnostic`，且不解析其标点或提供方私有 category 名称。
+
+### 安全失败文本
+
+通用 error 诊断采用以下固定字段顺序：
+
+```text
+Subagent failure (provider: <provider>; stage: <stage>; category: <category>; stop reason: <reason>; exit code: <code>; signal: <signal>)
+```
+
+不可用的可选字段会被省略。共享结算边界会把完整结果限制在 4096 个 UTF-8 字节以内。成功结果和本地取消不携带失败诊断。部分 assistant 输出继续保留在 `SubagentResult.output` 中，并与诊断分开呈现。
+
+当 ACP 权限请求参与非完成结果时，一个固定行会记录 `policy`、ACP 闭集工具 `request` 种类和 `decision`。工具标题、raw input、位置、选项名称与 metadata 均被排除。对于 `max-tokens`、`refusal` 或远端 `aborted`，公共结束原因已经携带终态事实，因此权限行就是完整诊断；通用 error 路径则把它附在失败行之后。带诊断的远端 `aborted` 结果仍保持公共结束原因；一次性 Job adapter 会把它判为 failed，而不带诊断的本地取消仍是 killed。
+
+### ACP 事实
+
+| Stage | 归属操作 | 安全 category 与事实 |
+| --- | --- | --- |
+| `initialize` | 父工作区解析与 ACP initialize | `configuration`、`transport` 或 `process-exit` |
+| `new-session` | ACP `session/new` 与返回 session id 校验 | `protocol`、`transport` 或 `process-exit` |
+| `prompt` | ACP prompt 请求、远端结束原因与权限回调 | `remote-limit`、`transport`、`unknown` 或仅权限诊断 |
+| `process` | 子进程 spawn 失败，或受管子进程先于 prompt 终态响应退出 | `process-start`，或 `process-exit` 以及分别观测到的退出码与信号 |
+| `teardown` | EOF 停稳与受管进程树终止 | 固定 teardown 事实；原始清理失败仍留在内部 |
+
+`max_turn_requests` 继续映射到共享 `error`，并附加 `remote-limit`。未知结束原因继续映射到 `error`，category 固定为 `unknown`，不会复制原值。`max_tokens`、`refusal` 与 `cancelled` 保持既有共享结束原因；只有需要解释权限决定时才会附加诊断。
+
+### 所有权与生命周期
+
+| 事实或资源 | Owner | 消费方行为 |
+| --- | --- | --- |
+| ACP 结束原因与工具种类 | ACP server 与 SDK | 提供方只映射闭集值，并对闭集外值使用固定 unknown 回退 |
+| 当前失败 stage 与最新权限决定 | 单次 ACP 运行 | 只在失败点派生，并随运行丢弃；并发运行不共享诊断状态 |
+| 退出码与信号 | `dsh-subprocess` 句柄 | 仅在观测到受管结果后展示；绝不解析 stderr |
+| 诊断字节与呈现 | `dsh-subagent`、前台工具与 Job 运行时 | 前台和一次性后台模式都把同一份有界文本与 assistant 输出分开 |
+| 原始失败 | 子运行时、Error cause 链与 Host logger | 只供 Host 排障，绝不复制进父模型结果 |
+
+启动只有在 initialize 与 new-session 成功后才发布运行。启动失败会先把私有子进程回滚到完全停稳，再以安全事实拒绝。已发布运行的结果不会拒绝，而 `dispose()` 会独立报告安全 teardown 失败，并继续使用后端既有的整棵进程树清理阶梯。
+
+## Verification
+
+ACP 包测试通过真实 stdio 协议子进程固定全部结束原因映射、远端限制与 unknown 回退、权限 allow/deny 事实、configuration、initialize、new-session、prompt、process 与 teardown stage、启动回滚、成功结果与本地取消省略、部分输出、并发运行隔离、仅 Host 可见的原始错误、进程完全停稳，以及共享多字节诊断限制。Loader 组合证明真实配置的提供方会到达模型可见前台结果。无密钥 ACP snapshot 会在前台错误输出与一次性后台 `job_output` detail 中固定同一份诊断与权限事实。
+
+## Alternatives considered
+
+**返回原始异常、stderr 或协议 payload。** 这些值可能包含任务内容、工具输入、路径、环境值、凭证和上游文本。固定白名单事实能够保留可行动差异，而不扩大模型可见信任边界。
+
+**增加共享结构化错误 enum。** ACP 与其他进程外提供方拥有不同生命周期位置和闭集终止词汇。共享 enum 会制造虚假的统一，并迫使无关消费方跟随提供方版本。
+
+**解析异常消息或 stderr 来分类。** 自由文本既不稳定也不安全。只有闭集协议值、typed 错误、当前调用位置与受管进程结果可以成为诊断输入。
+
+**修改既有结束原因。** 结束原因继续表示提供方无关的终态结果。可选诊断说明非完成结果为何要求不同的下一步，而不增加新的公共结果状态。
+
+**增加重试、恢复状态或交互审批。** 诊断只负责报告失败，不拥有修复动作。重试策略、会话恢复与人工交互需要独立用户约定和生命周期责任方。
+
+## Consequences
+
+父 agent 可以区分 ACP 远端限制、权限参与、协议或传输失败、部署/进程失败与 teardown 失败，同时不会接收子进程控制的文本。启动和清理错误与已发布结果使用同一套安全事实，而 Host 观测仍保留原始 cause。
+
+诊断仍是展示文本，不是公共协议。消费方可以呈现它，但不得按格式分支。本决策不增加重试策略、恢复控制器、共享提供方错误 enum、stderr 分类器、认证分类、会话持久化、进度流或新的 ACP 能力。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-24-route-priced-image-request-pressure.md
+2026-08-24-route-priced-image-request-pressure.md: 45a29211730474369607ed5fb933f380d640bf27
+2026-08-24-route-priced-image-request-pressure.zh.md: cf005a3ee343edf5774d554a4ec78cb876703774
@@ -0,0 +1,39 @@
+# Agent Note: Route-priced image request pressure
+
+Status: implemented
+
+English | [中文](2026-08-24-route-priced-image-request-pressure.zh.md)
+
+## Problem
+
+The token meter priced an `ImageBlock` as the structural JSON of its durable reference — roughly forty tokens — while a DeepSeek request image costs up to 384 visual tokens, so an image-dense session could carry hundreds of thousands of unbilled estimated tokens. Provider usage anchors only completed requests: the first multimodal request, images added after the anchor, and offload-set changes all fed automatic compaction a pressure figure that was wrong by orders of magnitude, triggering it far too late (context overflow) or, after a route change, too early. The [version-one simplification](../simplification/2026-07-29-simplify-web-image-input-v1.md) had deliberately rejected a provider-neutral tile formula and deferred visual pricing until a provider-aware estimator had a concrete consumer.
+
+## Decision
+
+Compaction pressure is now priced by the routed model's own request projection. `LlmAdapter.imageRequestPricing(provider, model)` is an optional synchronous hook returning an `LlmImageRequestPricing` for one exact route, resolved through `ctx.llm.imageRequestPricing()`; the base adapter declares none and unknown providers degrade to `undefined`, never throw. Each ordered image occurrence resolves to an `LlmImageRequestPrice`: the provider's visual tokens for a retained image plus the model-visible text the wire actually carries (request-preview handle, offload placeholder, or text-only substitution), with the text left to the caller's own estimator so no provider fixes a text tokenization.
+
+The DeepSeek adapter implements the hook from its connection snapshot (`request-pricing.ts`): uncatalogued and text-only models price every occurrence as its `textOnlyImageText` substitution; image-capable models reproduce the serializer's first-stage oldest-first offload through the shared `offloadedImagePrefixCount()`, build handle and placeholder text through the same execution-world access resolution the serializer uses, and price retained images at their `requestImageDimensions` projection with `deepSeekImageTokens()` — a verbatim port of the provider's published v4 vision calculator (14px patches, 3:1 downsampling, 384-token cap, minimum-pixel scale-up, 8:1 width clamp), priced at the worst-case pad-to-4 alignment. The pure geometry moved from `attachment-local` to `dsh-attachment` so provider and pricing share it.
+
+The token meter's surface fold stores route-neutral facts per node — the fixed-heuristic price, the image-free price, and the durable image occurrences — and `measure()` prices the surface under the effective envelope's route on every call. The anchor holds its raw materials (surface snapshot, provider-output price, usage) instead of a precomputed baseline, so a matching header reprices both the anchor and the current surface under one route and the signed delta compares like with like; the usage-versus-estimated choice happens per measurement against the route-priced anchor. Public `TokenSurfaceNode` carries both `tokens` (route-priced; read by trigger, retention, range selection, and the summary-shrink comparison) and `heuristicTokens` (fixed; the shadow-price protocol's unit, so `compaction/summary` and `compaction/prune` stay consistent with the O(1) projection fold's own appends). The `contextPressure` and `contextBreakdown` projections deliberately stay on the fixed heuristic.
+
+The test-support replay adapter declares a flat per-model `imageRequestTokens` so keyless assembled scenarios exercise the seam; the `image-compaction` ACP snapshot proves six inline images push the second turn's pre-step measurement over an automatic threshold that the text-only heuristic stays under, and that the triggered compaction shadows the image message at its heuristic price.
+
+## Alternatives considered
+
+**Price images inside the provider-neutral estimator.** Rejected by the [version-one note](../simplification/2026-07-29-simplify-web-image-input-v1.md) and still wrong: visual pricing varies by provider, model, detail mode, and preprocessing, and a hard-coded figure would look authoritative on routes it does not describe. The hook keeps every constant in the adapter that owns the route.
+
+**Correct pressure only from provider usage.** Usage cannot price the first multimodal request, an image added after the anchor, or a changed offload set — exactly the cases that made compaction fire too late. Usage stays the anchor for completed requests; the route projection prices the increment.
+
+**Reproduce the full serialization pipeline, including prepared-version bytes and the base64 fallback budgets.** The second-stage offload depends on encoded request bytes that only exist after asynchronous image preparation. The pricing reproduces the deterministic first stage from durable byte lengths; a fallback request can only offload more and cost less, so the estimate stays conservative without I/O in a synchronous hook.
+
+**Route-price the shadow-price protocol too.** Logged `shadowedTokenCount` feeds the O(1) projection fold, whose appends are priced by the fixed heuristic; pricing replacements by route would make the persisted running total drift. Keeping the protocol on `heuristicTokens` preserves the fold's by-construction agreement.
+
+**Fold route pricing into the meter's replay state.** A fold keyed to one route would have to replay on every route change and could not answer a `requestHeader` override for a different model. Storing route-neutral node facts and pricing at `measure()` keeps replay single-pass and measurement O(surface), which the contract already promises.
+
+## Consequences
+
+Automatic compaction now triggers on the pressure the routed model's next request will actually carry: image-dense DeepSeek sessions compact before overflow instead of after it, text-only routes charge substitution text instead of phantom visual tokens, and offloaded images cost their placeholder. The worst-case alignment pad overprices an image by at most three tokens, and the unreproduced base64-fallback budgets can only overprice — both errors are conservative; an execution-world access path that changes between pricing and the request shifts a descriptor's text price by its own length, and provider usage remains the authoritative anchor once a request completes. The published v4 calculator constants live in `llm-deepseek` alone; if the provider revises its vision projection, that one module and its pinned vectors are the change site. Measurement cost gains one pricing resolution and one image-occurrence walk per call, still O(surface).
+
+## Testing
+
+Formula vectors in `image-tokens.spec.ts` pin the published calculator's outputs, including the aspect-clamp, scale-up floor, one-column solver, odd-grid trim, and second-pass convergence cases, cross-checked against the reference implementation over a dimension grid and 50,000-point fuzz during development. `request-pricing.spec.ts` covers text-only substitution, the low-detail preset, and count- and byte-driven offload boundaries. Token-meter specs cover the first multimodal estimate, post-anchor image deltas over usage, text-only repricing under a header override, pricer-less neutrality, occurrence-count mismatch, and nested tool-result images. Compaction specs prove trigger, retention, range selection, and the summary-shrink comparison read the route price while the logged shadow price stays heuristic, including a summary that only route-priced shrink accepts. Access-resolution threading is covered at the pricing function and the adapter override. The keyless `image-compaction` ACP snapshot exercises the assembled application end to end.
@@ -0,0 +1,39 @@
+# Agent Note: 按路由定价的图片请求压力
+
+Status: implemented
+
+[English](2026-08-24-route-priced-image-request-pressure.md) | 中文
+
+## Problem
+
+token 计量服务把 `ImageBlock` 按其持久引用的 JSON 结构计价，约四十个 token，而一张 DeepSeek 请求图片最多消耗 384 个视觉 token，因此图片密集的会话可能携带数十万个未计入估算的 token。provider usage 只锚定已完成的请求：首次多模态请求、锚点之后新增的图片、offload 集合的变化，都会让自动 compaction 拿到数量级错误的压力值，触发得过晚（上下文溢出）或在路由切换后过早。[版本一简化](../simplification/2026-07-29-simplify-web-image-input-v1.zh.md)曾有意否决 provider-neutral 的 tile 公式，把视觉定价推迟到 provider-aware 估算器出现具体消费方之时。
+
+## Decision
+
+compaction 压力现在按路由模型自身的请求投影定价。`LlmAdapter.imageRequestPricing(provider, model)` 是可选的同步钩子，为一条确切路由返回 `LlmImageRequestPricing`，经 `ctx.llm.imageRequestPricing()` 解析；基类不声明定价，未注册的 provider 降级为 `undefined` 而绝不抛出。每个按序的图片出现处解析为一个 `LlmImageRequestPrice`：保留图片的提供方视觉 token，加上线上实际携带的模型可见文本（请求预览句柄、offload 占位文本或纯文本替换），文本交由调用方自己的估算器计价，避免任何提供方固定一种文本 token 化。
+
+DeepSeek 适配器基于连接快照实现该钩子（`request-pricing.ts`）：未编目和纯文本模型把每个出现处按其 `textOnlyImageText` 替换计价；支持图片的模型通过共享的 `offloadedImagePrefixCount()` 复现序列化器第一阶段的最旧优先 offload，经序列化器同一套执行环境访问解析构建句柄与占位文本，并按 `requestImageDimensions` 投影尺寸用 `deepSeekImageTokens()` 为保留图片计价，后者是提供方公布的 v4 视觉计算器的逐句移植（14px patch、3:1 降采样、384 token 上限、最小像素放大、8:1 宽度钳制），按最坏的 pad-to-4 对齐计价。纯几何函数从 `attachment-local` 上移到 `dsh-attachment`，供提供方与定价共享。
+
+token 计量服务的表层 fold 为每个节点存储与路由无关的事实：固定启发式价格、去图价格与持久图片出现处；`measure()` 在每次调用时按生效 envelope 的路由为表层定价。锚点保存原始材料（表层快照、提供方输出价格、usage）而非预先计算的基线，因此匹配的标头会把锚点与当前表层放在同一路由下重新定价，带符号 delta 的比较口径一致；usage 与估算的选择在每次计量时针对路由定价锚点做出。公开的 `TokenSurfaceNode` 同时携带 `tokens`（路由定价；触发、保留、选段与摘要收缩比较读取它）和 `heuristicTokens`（固定值；影子价协议的计量单位，使 `compaction/summary` 与 `compaction/prune` 与 O(1) 投影 fold 自身的追加保持一致）。`contextPressure` 与 `contextBreakdown` 投影有意保持固定启发式规则。
+
+test-support 的回放适配器按模型声明固定的 `imageRequestTokens`，让 keyless 装配场景走通这条 seam；`image-compaction` ACP 快照证明六张内联图片把第二轮 pre-step 计量推过自动阈值，而纯文本启发式保持在阈值之下，且被触发的 compaction 按启发式价格遮蔽了图片消息。
+
+## Alternatives considered
+
+**在 provider-neutral 估算器里为图片定价。** 已被[版本一 note](../simplification/2026-07-29-simplify-web-image-input-v1.zh.md)否决且依然错误：视觉定价随提供方、模型、细节档位与预处理而不同，写死的数字在它不描述的路由上会显得权威却错误。钩子把每个常量留在拥有该路由的适配器里。
+
+**只用 provider usage 校正压力。** usage 无法为首次多模态请求、锚点后新增图片或变化的 offload 集合定价，而这些正是让 compaction 触发过晚的情形。usage 仍是已完成请求的锚点；增量由路由投影定价。
+
+**复现完整序列化管线，包括请求版本字节与 base64 回退预算。** 第二阶段 offload 依赖异步图片准备之后才存在的编码字节。定价复现由持久字节长度决定的确定性第一阶段；回退请求只会 offload 更多、花费更少，因此估算在同步无 I/O 的钩子里保持保守。
+
+**让影子价协议也按路由定价。** 记录的 `shadowedTokenCount` 供 O(1) 投影 fold 消费，而该 fold 的追加按固定启发式计价；替换若按路由定价会让持久化的累计值漂移。协议保持在 `heuristicTokens` 上，维持 fold 的构造性一致。
+
+**把路由定价并入计量服务的回放状态。** 绑定单一路由的 fold 在路由每次变化时都得重放，也无法回答指向另一模型的 `requestHeader` 覆盖。存储与路由无关的节点事实并在 `measure()` 时定价，保持单遍回放与契约已承诺的 O(surface) 计量。
+
+## Consequences
+
+自动 compaction 现在按路由模型下一次请求实际携带的压力触发：图片密集的 DeepSeek 会话在溢出之前而非之后压缩，纯文本路由收取替换文本而非幻影视觉 token，被 offload 的图片按占位文本计费。最坏对齐 pad 对单图最多多计三个 token，未复现的 base64 回退预算只会多计——两种误差都偏保守；执行环境访问路径若在定价与请求之间变化，只会按其自身长度改变描述文本的价格，请求完成后 provider usage 仍是权威锚点。公布的 v4 计算器常量只存在于 `llm-deepseek`；提供方若修订其视觉投影，改动点就是这一个模块与其钉死的向量。每次计量多一次定价解析与一次图片出现处遍历，仍为 O(surface)。
+
+## Testing
+
+`image-tokens.spec.ts` 的公式向量钉死公布计算器的输出，覆盖宽高比钳制、放大下限、单列求解、奇数网格裁剪与第二遍收敛的用例，开发期间与参考实现在尺寸网格及五万点模糊测试上对拍。`request-pricing.spec.ts` 覆盖纯文本替换、低细节预设以及数量与字节驱动的 offload 边界。token-meter 测试覆盖首次多模态估算、usage 之上的锚后图片 delta、标头覆盖下的纯文本重定价、无定价器时的中性行为、出现处数量不匹配与嵌套工具结果图片。compaction 测试证明触发、保留、选段与摘要收缩比较读取路由价格而记录的影子价保持启发式，包括一个只有路由定价收缩才接受的摘要。访问解析的传递在定价函数与适配器覆写两处都有覆盖。keyless 的 `image-compaction` ACP 快照端到端验证装配后的应用。
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-24-web-per-turn-token-usage.md
+2026-08-24-web-per-turn-token-usage.md: 91aab0f2c261e2141ee964c828e7209ba2b3f72f
+2026-08-24-web-per-turn-token-usage.zh.md: f9c424fa0f84b802e98280bea9eaf4038bf31f19
@@ -0,0 +1,31 @@
+# Agent Note: Exact Web per-Turn token usage
+
+Status: implemented
+
+English | [中文](2026-08-24-web-per-turn-token-usage.zh.md)
+
+## Problem
+
+Web Chat exposes cumulative session token usage near the composer, but that value cannot explain the cost of one completed Turn. A paged history window may begin inside a Turn, retries may consume several model calls, streaming and final events may repeat one attempt's usage, and optional cache fields do not prove an exact total. Displaying a partial subtotal as Turn usage would make recorded provider facts look more complete than they are.
+
+## Decision
+
+The shared `TokenUsage` value carries optional `totalTokens` for one model call. Adapters publish it only from an exact provider total or authoritative aggregate prompt and output counters. DeepSeek checks its prompt-plus-completion aggregate against any wire total, and pi-ai preserves its provided total.
+
+Token-meter owns a browser-safe pure Turn-local fold over durable session events, shared with its retry-aware cumulative usage projection. `step/start` and `llm/retry-started` open actual attempts; a final assistant message replaces the same attempt's streaming sample; terminal failures, retries, and step boundaries close attempts without double counting. Every started attempt must close with safe non-negative integer usage and an exact total. Optional cache, reasoning, and route aggregates appear only when every contributing attempt reports them, and reasoning remains a subset of output.
+
+Web Chat selects a Turn only when its loaded match window includes `turn/start`, passes that complete durable-event window to the token-meter fold, and renders the result. A complete, exact result appears through a local-state `DisclosureRow` above the existing actions; incomplete or contradictory evidence produces no row. Chat owns no token-accounting state machine.
+
+## Alternatives considered
+
+**Subtract neighboring cumulative session values.** Rejected because pagination, compaction, retry coverage, and projection completeness can make adjacent values incomparable; subtraction would infer data that no call reported.
+
+**Publish historical per-Turn values through a new client session projection.** Rejected because the loaded per-Turn view already has the durable attempt events it needs, while a history-growing projection would add transport, persistence, and versioning costs. Reusing token-meter's pure fold keeps one accounting owner without adding another wire value.
+
+**Show known buckets without an exact total.** Rejected because a lower-bound subtotal presented in a completed Turn footer is indistinguishable from a complete bill.
+
+## Consequences
+
+New provider records can expose exact per-Turn accounting without a new transport or persisted UI state. Older sessions and adapters without enough evidence simply omit the disclosure. Model routes disappear as a group when any billed attempt lacks attribution, while trustworthy token totals remain visible.
+
+Focused adapter, token-meter fold/projection, component, pagination, and assembled Web replay tests pin total preservation, retry-attempt separation, fail-closed validation, optional-field omission, interaction, and full-window publication. The cumulative projection and exact Turn fold now share token-meter ownership; the projection remains a whole-log bucket view, while the fold alone makes the stricter exactness and completeness claim required by the disclosure.
@@ -0,0 +1,31 @@
+# Agent Note: Web 单轮次精确 token 用量
+
+Status: implemented
+
+[English](2026-08-24-web-per-turn-token-usage.md) | 中文
+
+## Problem
+
+Web Chat 在编辑框附近显示会话累计 token 用量，但该值无法解释一个已完成轮次的消耗。分页历史窗口可能从轮次中间开始，重试可能消耗多次模型调用，流式事件与最终事件可能重复携带同一次 attempt 的用量，而可选 cache 字段也不能证明精确总量。将局部小计显示成轮次用量，会让已记录的提供方事实显得比实际更完整。
+
+## Decision
+
+共享 `TokenUsage` 值为一次模型调用携带可选的 `totalTokens`。适配器只从提供方精确总量，或权威的提示词与输出聚合计数发布该字段。DeepSeek 会将提示词加输出的聚合值与协议提供的总量核对，pi-ai 则保留其提供的总量。
+
+token-meter 拥有一份可安全用于浏览器的纯轮次局部 fold，并与其具备重试感知能力的累计用量投影共享记账所有权。`step/start` 与 `llm/retry-started` 打开真实 attempt；最终 assistant 消息替换同一 attempt 的流式样本；终止失败、重试与步骤边界关闭 attempt，且不会重复计数。每个已开始的 attempt 都必须以安全的非负整数用量和精确总量关闭。只有每个参与聚合的 attempt 都报告时，才会显示可选的 cache、推理与路由聚合值；推理仍是输出的子集。
+
+Web Chat 只选择已加载匹配窗口包含 `turn/start` 的 Turn，将该完整的持久事件窗口交给 token-meter fold，再渲染结果。完整且精确的结果通过现有 actions 上方、仅保留本地状态的 `DisclosureRow` 显示；证据不完整或矛盾时不显示该行。Chat 不拥有 token 记账状态机。
+
+## Alternatives considered
+
+**对相邻的会话累计值做减法。** 不采用，因为分页、压缩、重试覆盖范围与投影完整性可能让相邻值无法比较；减法会推断任何调用都未报告的数据。
+
+**通过新的客户端会话投影发布历史单轮次值。** 不采用，因为已加载的单轮次视图已经拥有所需的持久 attempt 事件，而随历史增长的投影会增加传输、持久化与版本成本。复用 token-meter 的纯 fold，可以在不新增 wire 值的前提下保持唯一记账所有方。
+
+**缺少精确总量时仍显示已知 bucket。** 不采用，因为在已完成轮次 footer 中展示的下界小计与完整账单无法区分。
+
+## Consequences
+
+新的提供方记录无需新增传输接口或持久化 UI 状态，即可显示精确的单轮次记账。证据不足的旧会话与适配器只会省略 disclosure。任一计费 attempt 缺少归属时，模型路由会整体消失，可信 token 总量仍可显示。
+
+定向的适配器、token-meter fold／投影、组件、分页与组装 Web 回放测试固定了总量保留、重试 attempt 分离、fail-closed 校验、可选字段省略、交互与完整窗口发布。累计投影与精确 Turn fold 现在同归 token-meter 所有；投影仍是完整日志的 bucket 视图，只有 fold 会作出 disclosure 所需的更严格精确性与完整性声明。
@@ -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/process/2026-08-08-native-windows-pull-request-ci.md
-2026-08-08-native-windows-pull-request-ci.md: a1a783df2a700eec5436abbb144849629bbecebc
-2026-08-08-native-windows-pull-request-ci.zh.md: 42db5aa974c1d1dca6c6151d6c058f39571dcc59
+2026-08-08-native-windows-pull-request-ci.md: 24f1a5fa0c72255f678b837c910eb779445dee27
+2026-08-08-native-windows-pull-request-ci.zh.md: 805974192577b835b565bd42688c574473ac63bf
@@ -18,9 +18,9 @@ Every pull request also starts an ordinary independent `windows-native` job name

 The native job is deliberately absent from `all-checks-passed.needs` and does not use `continue-on-error`: the aggregate neither waits for it nor changes conclusion because of it, while the job retains its own unmasked result. Workspace build, production-site, and 100%-per-file coverage failures make the native job fail. Static, documentation, package, built-artifact, lint, and snapshot inventories run in the same job as observational gates: their failures remain visible without changing the native aggregate result because Linux owns their blocking verdict.

-The 16-core lane admits four concurrent outer gates. Workspace build and production-site validation start immediately. Instrumented and exempt-heavy coverage both wait for the complete build: the instrumented corpus includes packer assertions over built `lib/` output, while the exempt gate's temporary Oxlint contract probes must not race source compilation. Every observational gate waits for both coverage gates to settle, regardless of outcome, before entering an available slot; its own `needs` edges still require their predecessors to pass. This also keeps later static gates that create temporary contract files from racing either coverage scan. [In-job partitioned coverage](2026-08-18-in-job-partitioned-coverage.md) uses eight single-worker shards, while the exempt-heavy gate receives two workers from `DSH_COVERAGE_MAX_WORKERS=6`, for about ten active coverage execution units after build. `publint` is capped at eight workers when the observational inventory starts. Every Vitest project uses forked workers because Node 24's CJS lexer fatal reproduced in shared worker threads on Windows and POSIX. Both coverage gates set Vitest's default per-test and polling budgets to 30 seconds because unrelated process, Git, SQLite, watcher, grammar, and static-gate fixtures can exceed 15 seconds only under the complete lane's concurrent Windows instrumentation. The SQLite busy-journal pacing fixture injects two busy results followed by success under the normal busy budget and observes each inter-attempt delay, keeping schema-setup scheduling outside its timing assertion. The script-only translation-pairing merge suite runs in the exempt-heavy gate because it imports only `scripts/` sources and child processes; V8 instrumentation contributes no threshold coverage there but magnifies Git-process latency. Lefthook concurrency fixtures retain their outcomes with 30-second case budgets and a 10-second process-ready probe, while the installer allows five seconds for a preempted lock owner to publish its record after exclusive creation. Directory-picker composition gives its debounced config write an explicit 15-second poll budget; workspace-context composition fixtures use a test-owned signal without an unrelated one-second deadline. These lane-scoped budgets preserve asserted outcomes, while the 120-minute job deadline still bounds a stuck run. The LSP sources and the ACL-sandbox sources remain in the Windows denominator: stub-based failure-path suites carry every in-process ACL-sandbox file to 100%, and only the runner entry stays excluded — it executes exclusively as a spawned child outside the instrumented run, its behavior pinned end-to-end by the runner suite. Narrow annotated V8 ignores cover only unreachable branches (peer-platform arms and lifecycle-unreachable guards), with their behavior tests retained on the owning platform.
+The 16-core lane admits four concurrent outer gates. Workspace build and production-site validation start immediately. Instrumented and exempt-heavy coverage both wait for the complete build: the instrumented corpus includes packer assertions over built `lib/` output, while the exempt gate's temporary Oxlint contract probes must not race source compilation. Every observational gate waits for both coverage gates to settle, regardless of outcome, before entering an available slot; its own `needs` edges still require their predecessors to pass. This also keeps later static gates that create temporary contract files from racing either coverage scan. [In-job partitioned coverage](2026-08-18-in-job-partitioned-coverage.md) uses four single-worker shards, while the exempt-heavy gate receives two workers from `DSH_COVERAGE_MAX_WORKERS=6`, for about six active coverage execution units after build. `publint` is capped at eight workers when the observational inventory starts. Every Vitest project uses forked workers because Node 24's CJS lexer fatal reproduced in shared worker threads on Windows and POSIX. Both coverage gates set Vitest's default per-test and polling budgets to 30 seconds because unrelated process, Git, SQLite, watcher, grammar, and static-gate fixtures can exceed 15 seconds only under the complete lane's concurrent Windows instrumentation. The SQLite busy-journal pacing fixture injects two busy results followed by success under the normal busy budget and observes each inter-attempt delay, keeping schema-setup scheduling outside its timing assertion. The script-only translation-pairing merge suite runs in the exempt-heavy gate because it imports only `scripts/` sources and child processes; V8 instrumentation contributes no threshold coverage there but magnifies Git-process latency. Lefthook concurrency fixtures retain their outcomes with 30-second case budgets and a 10-second process-ready probe, while the installer allows five seconds for a preempted lock owner to publish its record after exclusive creation. Directory-picker composition gives its debounced config write an explicit 15-second poll budget; workspace-context composition fixtures use a test-owned signal without an unrelated one-second deadline. These lane-scoped budgets preserve asserted outcomes, while the 120-minute job deadline still bounds a stuck run. The LSP sources and the ACL-sandbox sources remain in the Windows denominator: stub-based failure-path suites carry every in-process ACL-sandbox file to 100%, and only the runner entry stays excluded — it executes exclusively as a spawned child outside the instrumented run, its behavior pinned end-to-end by the runner suite. Narrow annotated V8 ignores cover only unreachable branches (peer-platform arms and lifecycle-unreachable guards), with their behavior tests retained on the owning platform.

-The 16-core allocation is the measured capacity point for this inventory. Six-worker coverage trials produced complete passes in 6 minutes 27 seconds and 7 minutes 50 seconds, while exact-head trials with four, three, and two concurrent workers inside one instrumented Vitest process exposed unreliable fixtures and worker exits. Separate single-worker child processes retain process isolation. Historical sixteen-shard samples reduced instrumented coverage to 112.66–122.01 seconds. Under the current post-build graph, sixteen instrumented shards plus two exempt workers would schedule eighteen coverage execution units on a 16-core runner before any production-site tail or system overhead; eight shards plus two exempt workers schedule ten. Eight deliberately trades some latency for that headroom. A 32-core comparison reduced aggregate gate time by only 1.47 seconds and still triggered the CJS-lexer fatal inside a fork worker, so additional cores did not provide a reliable wall-clock improvement.
+The 16-core allocation is the measured capacity point for this inventory. Six-worker coverage trials produced complete passes in 6 minutes 27 seconds and 7 minutes 50 seconds, while exact-head trials with four, three, and two concurrent workers inside one instrumented Vitest process exposed unreliable fixtures and worker exits. Separate single-worker child processes retain process isolation. Historical sixteen-shard samples reduced instrumented coverage to 112.66–122.01 seconds. Under the current post-build graph, sixteen instrumented shards plus two exempt workers would schedule eighteen coverage execution units on a 16-core runner before any production-site tail or system overhead; four shards plus two exempt workers schedule six. Four deliberately trades some single-job latency for lower process-creation pressure under high self-hosted concurrency. A 32-core comparison reduced aggregate gate time by only 1.47 seconds and still triggered the CJS-lexer fatal inside a fork worker, so additional cores did not provide a reliable wall-clock improvement.

 The first native run exposed two failures hidden by the compatibility lane. Documentation projection tests derived an image basename by splitting only on `/`; they now use Node's platform basename. Chokidar consumers received `%TEMP%` through the `C:\\Users\\RUNNER~1` 8.3 alias while libuv returned the long directory name, tripping its Windows event-path assertion. Shared settings and credentials watchers, plus Cordis module and exact-config HMR, now canonicalize the existing native watch base or deepest existing ancestor before opening the watcher and preserve a missing suffix, while file access and diagnostics retain the configured path. Module HMR attaches listeners and awaits the main watcher's ready event before plugin startup settles, so an immediate post-boot edit cannot race the initial scan. HMR acceptance derives expected identities through the same asynchronous native realpath operation, avoiding a synchronous Windows spelling that can retain the 8.3 alias.

@@ -18,9 +18,9 @@ Status: implemented

 原生作业被刻意排除在 `all-checks-passed.needs` 之外，且不使用 `continue-on-error`：聚合流程既不等待它，也不会因它改变结论；该作业则保留自身未被掩盖的结果。工作区构建、生产网站和逐文件 100% 覆盖率检查失败会使原生作业失败。静态检查、文档、包、构建产物、lint 与快照清单在同一作业内作为观测性门禁运行；其失败保持可见，但不会改变原生聚合结果，因为这些检查的阻断性判定由 Linux 负责。

-16 核通道最多同时运行 4 道外层门禁。工作区构建与生产网站验证会立即启动。插桩覆盖率与豁免重型覆盖率都等待完整构建：插桩语料包含针对已构建 `lib/` 输出的打包器断言，豁免门禁的临时 Oxlint 约定探针则不得与源码编译竞态。每道观测性门禁只等待两道覆盖率门禁以任意结果结算后再进入可用槽位；各门禁自身的 `needs` 边仍要求前置门禁通过。这也使随后创建临时约定文件的静态门禁不会与任一覆盖率扫描竞态。[job 内分区覆盖率](2026-08-18-in-job-partitioned-coverage.zh.md)使用 8 个单 worker 分片，豁免重型门禁则从 `DSH_COVERAGE_MAX_WORKERS=6` 获得 2 个 worker，因此构建完成后约有 10 个活动覆盖率执行单元。观测性清单启动时，`publint` 最多使用 8 个 worker。每个 Vitest 项目都使用 fork worker，因为 Node 24 的 CJS lexer 致命故障可在 Windows 与 POSIX 的共享 worker 中复现。两项覆盖率门禁都将 Vitest 默认的单测试和轮询时间预算设为 30 秒，因为在完整通道并发的 Windows 插桩下，多个互不相关的进程、Git、SQLite、watcher、语法和静态门禁 fixture（测试前置数据）可能超过 15 秒。SQLite busy-journal 节奏 fixture 会在普通 busy 预算内先注入两次 busy 结果，再返回成功，并观察每次尝试之间的延迟，使 schema 设置的调度时间不进入该断言。translation-pairing 合并套件只导入 `scripts/` 源码和子进程，因此放入豁免重型套件门禁；V8 插桩不会为它贡献任何阈值覆盖率，却会放大 Git 进程延迟。Lefthook 并发 fixture 保留原有结果，采用 30 秒单用例预算与 10 秒进程就绪探测；安装器则允许被抢占的 lock 持有者在独占创建后用 5 秒发布记录。directory-picker 组合为防抖配置写入提供显式的 15 秒轮询预算；workspace-context 组合 fixture 使用测试自有、没有无关 1 秒截止时间的信号。这些只属于该通道的预算保留了原有断言结果，120 分钟的 job 截止时间仍会约束卡死的运行。LSP 源码与 ACL 沙箱源码仍计入 Windows 分母：基于 stub 的失败路径套件把每个进程内 ACL 沙箱文件都带到 100%，只有 runner 入口保持排除——它只作为 spawn 出的子进程在插桩运行之外执行，其行为由 runner 套件端到端钉住。窄范围且带注释的 V8 ignore 只覆盖不可达分支（另一平台专属分支、生命周期内不可达的防御守卫），其行为测试仍保留在所属平台。
+16 核通道最多同时运行 4 道外层门禁。工作区构建与生产网站验证会立即启动。插桩覆盖率与豁免重型覆盖率都等待完整构建：插桩语料包含针对已构建 `lib/` 输出的打包器断言，豁免门禁的临时 Oxlint 约定探针则不得与源码编译竞态。每道观测性门禁只等待两道覆盖率门禁以任意结果结算后再进入可用槽位；各门禁自身的 `needs` 边仍要求前置门禁通过。这也使随后创建临时约定文件的静态门禁不会与任一覆盖率扫描竞态。[job 内分区覆盖率](2026-08-18-in-job-partitioned-coverage.zh.md)使用 4 个单 worker 分片，豁免重型门禁则从 `DSH_COVERAGE_MAX_WORKERS=6` 获得 2 个 worker，因此构建完成后约有 6 个活动覆盖率执行单元。观测性清单启动时，`publint` 最多使用 8 个 worker。每个 Vitest 项目都使用 fork worker，因为 Node 24 的 CJS lexer 致命故障可在 Windows 与 POSIX 的共享 worker 中复现。两项覆盖率门禁都将 Vitest 默认的单测试和轮询时间预算设为 30 秒，因为在完整通道并发的 Windows 插桩下，多个互不相关的进程、Git、SQLite、watcher、语法和静态门禁 fixture（测试前置数据）可能超过 15 秒。SQLite busy-journal 节奏 fixture 会在普通 busy 预算内先注入两次 busy 结果，再返回成功，并观察每次尝试之间的延迟，使 schema 设置的调度时间不进入该断言。translation-pairing 合并套件只导入 `scripts/` 源码和子进程，因此放入豁免重型套件门禁；V8 插桩不会为它贡献任何阈值覆盖率，却会放大 Git 进程延迟。Lefthook 并发 fixture 保留原有结果，采用 30 秒单用例预算与 10 秒进程就绪探测；安装器则允许被抢占的 lock 持有者在独占创建后用 5 秒发布记录。directory-picker 组合为防抖配置写入提供显式的 15 秒轮询预算；workspace-context 组合 fixture 使用测试自有、没有无关 1 秒截止时间的信号。这些只属于该通道的预算保留了原有断言结果，120 分钟的 job 截止时间仍会约束卡死的运行。LSP 源码与 ACL 沙箱源码仍计入 Windows 分母：基于 stub 的失败路径套件把每个进程内 ACL 沙箱文件都带到 100%，只有 runner 入口保持排除——它只作为 spawn 出的子进程在插桩运行之外执行，其行为由 runner 套件端到端钉住。窄范围且带注释的 V8 ignore 只覆盖不可达分支（另一平台专属分支、生命周期内不可达的防御守卫），其行为测试仍保留在所属平台。

-16 核配置是这项清单经实测选定的容量规格。使用 6 个 coverage worker 的试验分别以 6 分 27 秒和 7 分 50 秒跑出完整通过结果，而在单个插桩 Vitest 进程内使用 4 个、3 个和 2 个并发 worker 的分支头精确试验暴露出不稳定的 fixture 与 worker 退出。相互独立的单 worker 子进程保留进程隔离。历史上的 16 分片样本把插桩覆盖率缩短到 112.66–122.01 秒。在当前的构建后拓扑中，16 个插桩分片加 2 个豁免 worker 会在 16 核运行器上调度 18 个覆盖率执行单元，且尚未计入生产网站的尾部工作或系统开销；8 个分片加 2 个豁免 worker 则调度 10 个。8 个分片刻意用部分延迟换取这份余量。32 核对比仅将聚合门禁时间缩短 1.47 秒，且仍在 fork worker 内触发 CJS lexer 致命故障，因此增加核心数没有带来可靠的墙钟时间改善。
+16 核配置是这项清单经实测选定的容量规格。使用 6 个 coverage worker 的试验分别以 6 分 27 秒和 7 分 50 秒跑出完整通过结果，而在单个插桩 Vitest 进程内使用 4 个、3 个和 2 个并发 worker 的分支头精确试验暴露出不稳定的 fixture 与 worker 退出。相互独立的单 worker 子进程保留进程隔离。历史上的 16 分片样本把插桩覆盖率缩短到 112.66–122.01 秒。在当前的构建后拓扑中，16 个插桩分片加 2 个豁免 worker 会在 16 核运行器上调度 18 个覆盖率执行单元，且尚未计入生产网站的尾部工作或系统开销；4 个分片加 2 个豁免 worker 则调度 6 个。4 个分片刻意用部分单 job 延迟换取自托管高并发下更低的进程创建压力。32 核对比仅将聚合门禁时间缩短 1.47 秒，且仍在 fork worker 内触发 CJS lexer 致命故障，因此增加核心数没有带来可靠的墙钟时间改善。

 首次原生运行暴露出两项被兼容性通道掩盖的故障。文档投影测试此前只按 `/` 拆分来派生图片 basename；现在改为使用 Node 根据平台计算的 basename。Chokidar 消费方收到的 `%TEMP%` 以 `C:\\Users\\RUNNER~1` 这个 8.3 别名表示，而 libuv 返回的是长目录名，导致其 Windows 事件路径断言失败。共享的设置 watcher 与凭据 watcher，以及 Cordis 的模块 HMR（热模块替换）与精确配置 HMR，现在都会在打开 watcher 前规范化现有的原生监听基准路径或层级最深的现有祖先路径，并保留尚不存在的后缀；文件访问和诊断仍使用配置路径。模块 HMR 会挂接监听器并等待主 watcher 的 ready 事件，之后插件启动才会完成，因此启动后立即发生的编辑无法与初始扫描形成竞态。HMR 验收通过相同的异步原生 realpath 操作派生预期身份，避免同步 Windows 路径写法仍保留 8.3 别名。

@@ -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/process/2026-08-18-in-job-partitioned-coverage.md
-2026-08-18-in-job-partitioned-coverage.md: 2cf06f6dd3eb9cdfd563aaa01ecd4ea9b8fb376f
-2026-08-18-in-job-partitioned-coverage.zh.md: 266feed66a28833488973db66fbd1f3476bf9ba7
+2026-08-18-in-job-partitioned-coverage.md: 9b8e6308ec746ee9a31465e57f58457130178b85
+2026-08-18-in-job-partitioned-coverage.zh.md: cf4d50996d9ecd4cb14ee762df11ade4aa63dc31
@@ -12,7 +12,7 @@ The optimization must retain every test and the merged per-file 100% thresholds.

 ## Decision

-The ordinary `pnpm run test:coverage` command remains one Vitest invocation. Linux coverage CI fixes `DSH_COVERAGE_PARTITIONS=4`, while native Windows fixes it at 8; no elapsed-time trigger changes either count while a run is in progress. The [coverage-exempt heavy suite](2026-07-31-coverage-exempt-heavy-suites.md) remains a separate uninstrumented gate beside the instrumented work.
+The ordinary `pnpm run test:coverage` command remains one Vitest invocation. Linux coverage CI fixes `DSH_COVERAGE_PARTITIONS=4`; native Windows now also fixes it at 4 to reduce process-creation pressure under high self-hosted concurrency. No elapsed-time trigger changes either count while a run is in progress. The [coverage-exempt heavy suite](2026-07-31-coverage-exempt-heavy-suites.md) remains a separate uninstrumented gate beside the instrumented work.

 When partitioning is enabled, `scripts/run-gates.ts` selects `pnpm run test:coverage:partitioned` for the instrumented gate. `scripts/coverage-partitions.ts` starts the configured Vitest children concurrently, each with one worker and one `--shard=<index>/<count>` option. Partition mode suppresses thresholds and coverage reporters in each child, gives every child a separate report directory, and writes one blob report per process.

@@ -30,7 +30,7 @@ A normal failed test still emits a blob through `--coverage.reportOnFailure`, al

 `scripts/coverage-partitions.spec.ts` pins argument construction, package-script separator removal, one-worker partitions, the single merged threshold command, failed-test merging, failure diagnostics before complete-blob validation, waiting for sibling partitions after a spawn failure, and link-safe cleanup. `scripts/run-gates.spec.ts` pins opt-in selection, invalid-count rejection, both native Windows coverage gates' complete-build dependency, the complete Windows inventory with its blocking split, and unbuffered streamed output. React fake-timer cases that can move between partitions advance timers inside `act()`; geometry-dependent portal tests stub their element rectangles so a different shard schedule cannot turn deferred updates or jsdom coordinates into coverage-only failures.

-Completed native Windows comparisons measured two partitions near 405 seconds and sixteen partitions at 112.66–122.01 seconds under the earlier gate ordering; those values compare partition latency, not the current peak. The current post-build phase runs eight instrumented partition processes beside two exempt workers, for ten coverage execution units. Sixteen partitions would raise that phase to eighteen before any still-running production-site work or system overhead. Eight keeps separate-process isolation while leaving capacity headroom. Two Linux samples measured the conservative two-partition configuration at 276.68 and 282.27 seconds; that configuration was stable but halved the ordinary path's four instrumented workers. Four partitions restore that fan-out, for six total coverage execution units on the 16-core hosted runner and at most 36 across the failover VM's six runner instances. These values come from completed runs or fixed capacity bounds; an unfinished run crossing an arbitrary elapsed-time mark is not evidence for increasing concurrency.
+Completed native Windows comparisons measured two partitions near 405 seconds and sixteen partitions at 112.66–122.01 seconds under the earlier gate ordering; those values compare partition latency, not the current peak. The current post-build phase runs four instrumented partition processes beside two exempt workers, for six coverage execution units. Sixteen partitions would raise that phase to eighteen before any still-running production-site work or system overhead. Four partitions keep separate-process isolation and match Linux, at the cost of a longer single-job coverage wall time; the trade-off is accepted to reduce vitest worker startup failures under high self-hosted concurrency. Two Linux samples measured the conservative two-partition configuration at 276.68 and 282.27 seconds; that configuration was stable but halved the ordinary path's four instrumented workers. Four partitions restore that fan-out, for six total coverage execution units on the 16-core hosted runner and at most 36 across the failover VM's six runner instances. These values come from completed runs or fixed capacity bounds; an unfinished run crossing an arbitrary elapsed-time mark is not evidence for increasing concurrency.

 ## Alternatives considered

@@ -38,7 +38,7 @@ Completed native Windows comparisons measured two partitions near 405 seconds an

 **Raise the Vitest worker count inside one instrumented process.** Rejected because completed Windows trials at higher fan-out exposed worker exits, fixture instability, and Node 24 CJS lexer failures. Separate single-worker processes preserve isolation while still executing the selected partitions concurrently.

-**Use one partition count on every host.** Rejected because Linux's four-process run and Windows's eight-process run have different startup costs and resource ceilings. Each fixed configuration requires its own completed end-to-end evidence.
+**Use one partition count on every host.** Previously rejected because Linux's four-process run and Windows's eight-process run had different startup costs and resource ceilings. This change aligns both to four partitions after Windows high-concurrency runs exposed worker startup failures at eight.

 **Apply thresholds independently in each partition.** Rejected because every partition intentionally sees only part of the suite and would report false uncovered files. Threshold ownership belongs to the merged report.

@@ -12,7 +12,7 @@ Status: implemented

 ## 决策

-普通的 `pnpm run test:coverage` 命令仍只启动一次 Vitest。Linux 覆盖率 CI 将 `DSH_COVERAGE_PARTITIONS` 固定为 4，原生 Windows 则固定为 8；运行期间不会由任何耗时触发器改变这两个数量。[覆盖率豁免重型套件](2026-07-31-coverage-exempt-heavy-suites.zh.md)仍作为独立的无插桩门禁与插桩工作并排运行。
+普通的 `pnpm run test:coverage` 命令仍只启动一次 Vitest。Linux 覆盖率 CI 将 `DSH_COVERAGE_PARTITIONS` 固定为 4；原生 Windows 现在也固定为 4，以降低自托管高并发下的进程创建压力。运行期间不会由任何耗时触发器改变这两个数量。[覆盖率豁免重型套件](2026-07-31-coverage-exempt-heavy-suites.zh.md)仍作为独立的无插桩门禁与插桩工作并排运行。

 启用分区后，`scripts/run-gates.ts` 为插桩门禁选择 `pnpm run test:coverage:partitioned`。`scripts/coverage-partitions.ts` 按配置数量并发启动 Vitest 子进程，每个进程只用 1 个 worker，并各自接收一个 `--shard=<index>/<count>` 选项。分区模式会在各子进程中关闭阈值与覆盖率报告器，为每个子进程分配独立报告目录，并让每个进程写出 1 份 blob 报告。

@@ -30,7 +30,7 @@ Status: implemented

 `scripts/coverage-partitions.spec.ts` 固定了参数构造、包脚本分隔符移除、单 worker 分区、唯一一次合并阈值命令、失败测试合并、完整 blob 校验前的失败诊断、spawn 失败后等待兄弟分区，以及链接安全清理。`scripts/run-gates.spec.ts` 固定了显式启用、非法数量拒绝、两道原生 Windows 覆盖率门禁对完整构建的依赖、完整 Windows 清单及其阻断性划分，以及不缓冲的流式输出。可能在分区间移动的 React fake-timer 用例会在 `act()` 内推进计时器；依赖几何位置的 portal 测试会固定元素矩形，使不同分片调度不会把延迟更新或 jsdom 坐标变成只在覆盖率运行中出现的失败。

-已完成的原生 Windows 对比中，双分区耗时约 405 秒，16 分区耗时 112.66–122.01 秒；这些数据来自先前的门禁顺序，只用于比较分区延迟，不代表当前峰值。当前的构建后阶段会让 8 个插桩分区进程与 2 个豁免 worker 并行，共形成 10 个覆盖率执行单元。若改为 16 个分区，则在尚未结束的生产网站工作或系统开销计入之前，该阶段就会达到 18 个执行单元。8 个分区既保留独立进程隔离，也为其他工作留出容量余量。两个 Linux 样本中，保守的双分区配置耗时 276.68 秒和 282.27 秒；该配置运行稳定，却把普通路径原有的 4 个插桩 worker 减半。4 个分区恢复这份并发，使 16 核托管 runner 上的覆盖率执行单元总数为 6，故障切换虚拟机的 6 个 runner 实例最多合计 36 个执行单元。这些数值来自完整运行或固定容量上限；运行尚未结束时跨过任意耗时刻度，不构成增加并发的证据。
+已完成的原生 Windows 对比中，双分区耗时约 405 秒，16 分区耗时 112.66–122.01 秒；这些数据来自先前的门禁顺序，只用于比较分区延迟，不代表当前峰值。当前的构建后阶段会让 4 个插桩分区进程与 2 个豁免 worker 并行，共形成 6 个覆盖率执行单元。若改为 16 个分区，则在尚未结束的生产网站工作或系统开销计入之前，该阶段就会达到 18 个执行单元。4 个分区保留独立进程隔离并与 Linux 对齐，代价是单 job 覆盖率墙钟更长；这是为了降低自托管高并发下 vitest worker 启动失败而接受的取舍。两个 Linux 样本中，保守的双分区配置耗时 276.68 秒和 282.27 秒；该配置运行稳定，却把普通路径原有的 4 个插桩 worker 减半。4 个分区恢复这份并发，使 16 核托管 runner 上的覆盖率执行单元总数为 6，故障切换虚拟机的 6 个 runner 实例最多合计 36 个执行单元。这些数值来自完整运行或固定容量上限；运行尚未结束时跨过任意耗时刻度，不构成增加并发的证据。

 ## 曾考虑的替代方案

@@ -38,7 +38,7 @@ Status: implemented

 **提高单个插桩进程内的 Vitest worker 数。** 不予采用，因为已完成的 Windows 高扇出试验暴露了 worker 退出、fixture（测试前置数据）不稳定和 Node 24 CJS lexer 故障。相互独立的单 worker 进程既保留隔离，也能让所选分区并发执行。

-**在每种宿主上使用相同的分区数量。** 不予采用，因为 Linux 的 4 进程运行与 Windows 的 8 进程运行具有不同的启动成本与资源上限。每种固定配置都必须取得自己的端到端完整证据。
+**在每种宿主上使用相同的分区数量。** 先前不予采用，因为 Linux 的 4 进程运行与 Windows 的 8 进程运行具有不同的启动成本与资源上限。本次变更在 Windows 高并发运行暴露 8 分片 worker 启动失败后，将两者统一为 4 分区。

 **在每个分区内独立应用阈值。** 不予采用，因为每个分区有意只看到套件的一部分，会误报未覆盖文件。阈值归合并报告所有。

@@ -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/2026-07-12-simplify-session-log-representation.md
-2026-07-12-simplify-session-log-representation.md: 3efb531d3c0822d7444d1270eac4da2617c12447
-2026-07-12-simplify-session-log-representation.zh.md: 8c05ac6512a8aa8e7fefc56ba410bc81dc3be277
+2026-07-12-simplify-session-log-representation.md: a0c86b66af78b4c94991d38656f03609297e2314
+2026-07-12-simplify-session-log-representation.zh.md: ceaf90236a47c141e3a4bb2cc78d415b3b9ac2ba
@@ -18,7 +18,7 @@ The implementation retains append and replacement `sourceEventSeqs`, the `tool/c

 `SurfaceManager.nodes` is a `readonly number[]` of event sequences; the public `SurfaceNode` shape, node links, and seq-to-node map are removed. The internal replace-generation signal remains. The complete `foldSurface()` read used by session-query returns the same number-array representation plus replacement metadata without making the incremental manager retain history. Tool-pairing balance and compaction use event sequences and surface positions; the compact-owned per-cut balance cache does not depend on node links.

-Request headers use canonical full snapshots only. Initial and resume anchors remain full snapshots even when unchanged; an in-instance change appends another full `request/header` with reason `change`. The delta event, codec types, diff/apply helpers, and codec-only `fallback` reason are removed. Request reconstruction selects the latest snapshot.
+Request headers use canonical full snapshots only. Initial and resume anchors remain full snapshots even when unchanged; an in-instance change appends another full `request/header` with reason `change`; and an unchanged envelope beginning an explicitly declared message series or following a surface replacement appends a full snapshot with reason `series`. Ordinary append-only later Turns, further Steps, and retries in that model-message series inherit the latest snapshot. The delta event, codec types, diff/apply helpers, and codec-only `fallback` reason are removed. Request reconstruction selects the latest snapshot.

 `SESSION_FORMAT_VERSION` remains pinned at `0`, so seed, append, and persistence-load validation explicitly reject old v0 `request/header-delta` events and full snapshots carrying the removed `fallback` reason. There is no compatibility fold or migration. JSONL and SQLite tests pin this fail-loud boundary, and the ACP snapshot harness represents legitimate mid-session changes as full pinned headers and full readable prompts.

@@ -18,7 +18,7 @@ Status: implemented

 `SurfaceManager.nodes` 是由事件序号组成的 `readonly number[]`；公共 `SurfaceNode` 形状、node 链接和 seq-to-node map 均已移除。内部替换 generation 信号保留。session-query 使用的完整 `foldSurface()` 读取会返回相同的数字数组表示和替换元数据，而无需让增量 manager 保留历史。工具配对 balance 和压缩（compaction）使用事件序号与 surface 位置；由 compact 拥有的每个切点的 balance cache 不依赖 node 链接。

-请求头只使用规范的完整快照。初始与恢复锚点即使没有变化也仍是完整快照；实例内变化会追加另一个完整 `request/header`，reason 为 `change`。delta 事件、codec 类型、diff/apply 辅助函数，以及仅供 codec 使用的 `fallback` reason 均已移除。请求重建选择最新快照。
+请求头只使用规范的完整快照。初始与恢复锚点即使没有变化也仍是完整快照；实例内变化会追加另一个完整 `request/header`，reason 为 `change`；未变的信封显式开启消息序列或跟随 surface 替换时，会追加 reason 为 `series` 的完整快照。普通的仅追加后续 Turn、同一模型消息序列内的后续 Step 与重试沿用最新快照。delta 事件、codec 类型、diff/apply 辅助函数，以及仅供 codec 使用的 `fallback` reason 均已移除。请求重建选择最新快照。

 `SESSION_FORMAT_VERSION` 仍固定为 `0`，因此 seed、追加和持久化加载验证会显式拒绝旧 v0 `request/header-delta` 事件，以及携带已删除 `fallback` reason 的完整快照。不存在兼容性 fold 或迁移。JSONL 与 SQLite 测试固定了这一失败即报错的边界；ACP（Agent Client Protocol）快照 harness 则把合法的会话中途变更表示为固定的完整请求头和完整可读提示词。

@@ -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/2026-07-29-simplify-web-image-input-v1.md
-2026-07-29-simplify-web-image-input-v1.md: e7847dc2ae18f48146cb2b686dcda64beebf2350
-2026-07-29-simplify-web-image-input-v1.zh.md: 47f6bb5face5b896b5184f7198eef7d99ee25afe
+2026-07-29-simplify-web-image-input-v1.md: f13abfda3be890f80a8ff852acbc68118b91941c
+2026-07-29-simplify-web-image-input-v1.zh.md: c92974efe5f84da5b91c09ea4acf820e7853ca7d
@@ -32,6 +32,6 @@ The attachment seam exposes its limits plus storage-free `validateImage`, `saveI

 ## Consequences

-The feature retains the two batch limits and one storage-free validation method required by multi-image prompts, while removing unrelated public fields, lifecycle operations, policy snapshots, and route-assembly branches. Provider/model selection remains composition or profile configuration. Pre-request token pressure may undercount visual input until a provider-aware estimator is designed, while reported usage remains exact.
+The feature retains the two batch limits and one storage-free validation method required by multi-image prompts, while removing unrelated public fields, lifecycle operations, policy snapshots, and route-assembly branches. Provider/model selection remains composition or profile configuration. Pre-request token pressure keeps the structural heuristic only on routes without declared image pricing; the [route-priced estimator](../feature/2026-08-24-route-priced-image-request-pressure.md) supplies the provider-aware figure, and reported usage remains exact.

 Reintroducing any removed surface requires a concrete consumer and its failure, lifecycle, replay, and testing contract rather than compatibility with this pre-release shape.
@@ -32,6 +32,6 @@ Status: implemented

 ## 后果

-该功能保留了多图片提示词所需的两个批次上限和一个不触碰存储的校验方法，同时移除了无关的公开字段、生命周期操作、策略快照和路由组装分支。提供方／模型选择仍属于组合或 profile 配置。在设计出提供方感知型估算器之前，请求前的 token 压力计算可能少计视觉输入，而上报的用量仍保持精确。
+该功能保留了多图片提示词所需的两个批次上限和一个不触碰存储的校验方法，同时移除了无关的公开字段、生命周期操作、策略快照和路由组装分支。提供方／模型选择仍属于组合或 profile 配置。请求前的 token 压力只在未声明图片定价的路由上保留结构启发式；[按路由定价的估算器](../feature/2026-08-24-route-priced-image-request-pressure.zh.md)提供提供方感知的数值，上报的用量仍保持精确。

 重新引入任何已移除表面时，都必须有具体消费方，并为其定义失败、生命周期、回放和测试契约，而不是为了兼容这一预发布形态。
@@ -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/2026-08-12-production-dsh-excludes-product-subagent-providers.md
-2026-08-12-production-dsh-excludes-product-subagent-providers.md: faa14a5848a421389d8c427f844c4dedc118a0fd
-2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md: 0b99cf8aebba4aa44fb42e635c94cc64b4d3b458
+2026-08-12-production-dsh-excludes-product-subagent-providers.md: aaa6ddc70a0320c85d0d784441dfc82d74831a90
+2026-08-12-production-dsh-excludes-product-subagent-providers.zh.md: 37b861ad879995ac7dbd91903cf92332a83f9d92
@@ -16,7 +16,7 @@ The two Bundles remain independent. The Codex Bundle owns the pinned official wr

 ## Verification

-Package tests pin both Bundle manifests, published patches, exact self-provider rows, and product runtime dependencies. Claude coverage pins Agent SDK 0.3.220, Claude Code 2.1.220, all eight platform packages, SDK-selected execution, and missing-payload failure without host fallback. Codex coverage pins wrapper 0.147.0, all six platform aliases, package-declared execution, native descendant quiescence, and the same missing-payload behavior. Workspace validation derives each published patch from its Bundle declaration rather than a package catalog. Package/base assertions plus actual pnpm production evidence prove the default and selected-product dependency boundaries, while real Bundle-patch and Agent-Preset composition covers none, either product, both, the tool-grant intersection, later-Session adoption, and zero startup processes.
+Package tests pin both Bundle manifests, published patches, exact self-provider rows, and product runtime dependencies. Claude coverage pins Agent SDK 0.3.241, Claude Code 2.1.241, all eight platform packages, SDK-selected execution, and missing-payload failure without host fallback. Codex coverage pins wrapper 0.147.0, all six platform aliases, package-declared execution, native descendant quiescence, and the same missing-payload behavior. Workspace validation derives each published patch from its Bundle declaration rather than a package catalog. Package/base assertions plus actual pnpm production evidence prove the default and selected-product dependency boundaries, while real Bundle-patch and Agent-Preset composition covers none, either product, both, the tool-grant intersection, later-Session adoption, and zero startup processes.

 ## Alternatives considered

@@ -16,7 +16,7 @@ Status: implemented

 ## 验证

-包测试会固定两个 Bundle 的 manifest、发布 patch、准确的自身提供方行与产品运行时依赖。Claude 覆盖会固定 Agent SDK 0.3.220、Claude Code 2.1.220、八个平台包、SDK 所选执行路径，以及载荷缺失时不回退宿主命令的失败。Codex 覆盖会固定 wrapper 0.147.0、六个平台 alias、包声明的执行路径、原生后代进程停稳，以及同样的载荷缺失行为。工作区验证会从 Bundle 声明派生每份发布 patch，而非维护包目录。包与 base 断言加上实际 pnpm 生产证据会证明默认与所选产品的依赖边界；真实 Bundle patch 与 Agent Preset 组装则覆盖未安装、任一单包、双包、工具授权交集、后续 Session 采纳以及零启动进程。
+包测试会固定两个 Bundle 的 manifest、发布 patch、准确的自身提供方行与产品运行时依赖。Claude 覆盖会固定 Agent SDK 0.3.241、Claude Code 2.1.241、八个平台包、SDK 所选执行路径，以及载荷缺失时不回退宿主命令的失败。Codex 覆盖会固定 wrapper 0.147.0、六个平台 alias、包声明的执行路径、原生后代进程停稳，以及同样的载荷缺失行为。工作区验证会从 Bundle 声明派生每份发布 patch，而非维护包目录。包与 base 断言加上实际 pnpm 生产证据会证明默认与所选产品的依赖边界；真实 Bundle patch 与 Agent Preset 组装则覆盖未安装、任一单包、双包、工具授权交集、后续 Session 采纳以及零启动进程。

 ## 考虑过的替代方案

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# 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/2026-08-21-product-subagent-minimal-diagnostics.md
+2026-08-21-product-subagent-minimal-diagnostics.md: a4d51c325738e6e3e6c74561deb877c9bca3c016
+2026-08-21-product-subagent-minimal-diagnostics.zh.md: 409b548e0e6169479cec892d349dc8c969b24637
@@ -0,0 +1,59 @@
+# Agent Note: Product subagents expose minimal actionable diagnostics
+
+Status: implemented
+
+English | [中文](2026-08-21-product-subagent-minimal-diagnostics.zh.md)
+
+## Problem
+
+The Claude Code and Codex providers receive structured failures from independently versioned product runtimes. Mirroring every member of those upstream error unions into model-visible diagnostics makes each runtime upgrade expand the Provider contract even when the parent agent would take the same next action for several categories.
+
+Raw SDK errors, app-server payloads, stderr, commands, paths, task content, environment values, and credentials cannot replace the structured mapping because they cross the product process boundary without a stable safety guarantee. The parent still needs enough information to distinguish a product limit, access restriction, service or transport problem, invalid result, and managed process failure.
+
+## Decision
+
+Each product Provider derives a small action category from the current operation and safe structured product facts. The existing `SubagentResult.diagnostic` string remains the only public representation: consumers display it but do not parse it, and the shared subagent result boundary continues to enforce the complete 4096-byte UTF-8 limit.
+
+Claude Code maps Agent SDK results into five categories:
+
+| Category | Safe input |
+| --- | --- |
+| `limit` | Turn, budget, or structured-output retry limits |
+| `product-error` | A general SDK execution failure |
+| `invalid-result` | An error-marked or blank success, or no terminal result |
+| `process` | The managed CLI exits before a terminal result |
+| `unknown` | Startup, teardown, unrecognized SDK values, or failures without a more specific safe fact |
+
+The diagnostic also retains the derived `query-start`, `query-run`, `process`, or `teardown` stage and independently observed exit code and signal. A contributing permission decision follows the failure line. Successful completion and local cancellation expose no failure diagnostic, and original SDK text remains only on the internal cause chain and in Host observation.
+
+Codex currently retains the version-pinned categories documented by the [structured failure-facts decision](../feature/2026-08-18-product-subagent-failure-facts.md). The two providers may adopt the minimal mapping independently because neither shares a public error enum or asks a consumer to branch on category text.
+
+### Ownership and lifecycle
+
+| Fact or operation | Owner | Result |
+| --- | --- | --- |
+| Product error interpretation | Official product runtime | The Provider consumes only structured facts exposed by its pinned integration |
+| Action category and stage | One product Provider run | Derived at the failure site and discarded after result settlement |
+| Exit code and signal | `dsh-subprocess` handle | Displayed independently when observed, without inferring missing values |
+| Permission decision | Product Provider permission callbacks or protocol | Appended only when it contributed to the failed run |
+| Diagnostic delivery and byte limit | `dsh-subagent` and its foreground or Job consumers | One bounded optional string remains separate from assistant output |
+
+## Verification
+
+Claude Code package tests cover every coarse category, all four stages, unknown structured values, permission ordering, raw-text exclusion, success and cancellation omission, concurrent-run isolation, and independent exit code and signal fields. The real Agent SDK 0.3.241 and Claude Code 2.1.241 fixture produces an actual max-turns limit, process failure, permission denial, strict final answer, cancellation, and whole-tree quiescence. Loader and keyless product compositions continue to expose static tools without a diagnostic parser or model-visible category input.
+
+## Alternatives considered
+
+**Preserve every upstream enum member.** Several members lead to the same parent action, while the display-only consumer gains no behavior from their exact names. Keeping them would make routine product upgrades redefine a larger Harness-facing promise.
+
+**Return raw product text or classify general stderr.** Free-form text can contain sensitive task and environment data and has no stable versioned meaning. Only structured product facts, Provider stage, permission decisions, and managed process outcomes qualify as input.
+
+**Add a shared structured error type.** The products version their failures independently, and current consumers only present a bounded string. A shared enum would move product release churn into the provider-neutral seam without a consumer that needs it.
+
+**Retry or fall back based on the category.** The Provider owns one one-shot run and has no retry authority, fallback model, persistent product session, or recovery state. The category informs the parent; it does not start another product action.
+
+## Consequences
+
+Claude Code upgrades no longer require a model-visible promise for every SDK error subtype. Parents still distinguish a limit, general product failure, invalid result, managed process exit, and unknown failure while retaining stage, permission, and process facts.
+
+The diagnostic remains safe display text rather than a recovery protocol. This change adds no raw error forwarding, fallback model, automatic retry, product session persistence, public structured result field, or dynamic provider and model selection.
@@ -0,0 +1,59 @@
+# Agent Note: 产品 subagent 公开最小可行动诊断
+
+Status: implemented
+
+[English](2026-08-21-product-subagent-minimal-diagnostics.md) | 中文
+
+## Problem
+
+Claude Code 与 Codex 提供方会从彼此独立版本化的产品运行时接收结构化失败。如果把这些上游错误联合的每个成员都镜像进模型可见诊断，那么即使多个类别会让父 agent 采取同一种下一步动作，每次运行时升级仍会扩大提供方约定。
+
+原始 SDK 错误、app-server payload、stderr、命令、路径、任务内容、环境值与凭证都不能取代结构化映射，因为这些内容跨越产品进程边界时没有稳定的安全保证。父 agent 仍需要足够信息来区分产品限制、访问限制、服务或传输问题、无效结果与受管进程失败。
+
+## Decision
+
+每个产品提供方都会根据当前操作与安全的结构化产品事实派生少量行动类别。现有 `SubagentResult.diagnostic` 字符串仍是唯一公共表示：消费方只展示而不解析，共享 subagent 结果边界继续对完整文本执行 4096 字节 UTF-8 上限。
+
+Claude Code 会把 Agent SDK 结果映射为五种类别：
+
+| 类别 | 安全输入 |
+| --- | --- |
+| `limit` | 轮次、预算或结构化输出重试限制 |
+| `product-error` | 一般 SDK 执行失败 |
+| `invalid-result` | 标记为错误或内容空白的成功结果，或缺失终态结果 |
+| `process` | 受管 CLI 在终态结果前退出 |
+| `unknown` | 启动、清理、无法识别的 SDK 值，或没有更具体安全事实的失败 |
+
+诊断还会保留派生的 `query-start`、`query-run`、`process` 或 `teardown` 阶段，以及分别观测到的退出码与信号。参与失败的权限决定会跟在失败行之后。成功完成与本地取消不公开失败诊断，原始 SDK 文本只留在内部 cause 链和 Host 观测中。
+
+Codex 当前仍保留[结构化失败事实决策](../feature/2026-08-18-product-subagent-failure-facts.zh.md)记录的锁定版本类别。两个提供方可以分别采用最小映射，因为它们既不共享公共错误枚举，也不要求消费方根据类别文本分支。
+
+### 所有权与生命周期
+
+| 事实或操作 | 责任方 | 结果 |
+| --- | --- | --- |
+| 产品错误解释 | 官方产品运行时 | 提供方只消费其锁定集成公开的结构化事实 |
+| 行动类别与阶段 | 单次产品提供方运行 | 在失败位置派生，并在结果结算后丢弃 |
+| 退出码与信号 | `dsh-subprocess` 句柄 | 观测到时分别展示，不推断缺失值 |
+| 权限决定 | 产品提供方权限回调或协议 | 只在参与失败运行时追加 |
+| 诊断传递与字节上限 | `dsh-subagent` 及其前台或 Job 消费方 | 一个有界可选字符串继续与 assistant 输出分离 |
+
+## Verification
+
+Claude Code 包测试覆盖全部粗粒度类别、四个阶段、未知结构化值、权限顺序、原始文本排除、成功与取消时省略、并发运行隔离，以及彼此独立的退出码与信号字段。真实 Agent SDK 0.3.241 与 Claude Code 2.1.241 fixture 会产生实际 max-turns 限制、进程失败、权限拒绝、严格最终答案、取消与整棵进程树完全停稳。Loader 与无密钥产品组合继续公开静态工具，不增加诊断解析器或模型可见类别输入。
+
+## Alternatives considered
+
+**保留每个上游枚举成员。** 多个成员会触发同一种父级动作，而只负责展示的消费方无法从准确名称获得额外行为。继续保留会让常规产品升级重新定义更大的 Harness 可见承诺。
+
+**返回原始产品文本或分类一般 stderr。** 自由文本可能包含敏感任务与环境数据，也没有稳定的版本化含义。只有结构化产品事实、提供方阶段、权限决定与受管进程结果可以成为输入。
+
+**增加共享结构化错误类型。** 两个产品独立版本化失败，而当前消费方只呈现有界字符串。共享枚举会在没有消费需求的情况下，把产品发布变动移入提供方无关的 seam。
+
+**根据类别重试或 fallback。** 提供方只拥有一次 one-shot 运行，没有重试权限、fallback 模型、持久产品会话或恢复状态。类别只用于通知父级，不会启动另一项产品动作。
+
+## Consequences
+
+Claude Code 升级不再要求为每个 SDK 错误 subtype 提供模型可见承诺。父级仍可区分限制、一般产品失败、无效结果、受管进程退出与未知失败，并保留阶段、权限与进程事实。
+
+诊断仍是安全展示文本，而不是恢复协议。本改动不增加原始错误转发、fallback 模型、自动重试、产品会话持久化、公共结构化结果字段或动态提供方与模型选择。
@@ -397,63 +397,154 @@ jobs:
        if: always()
        run: wineserver -k 2>/dev/null || true

-  # Every pull request also gets a real Windows-kernel signal. This job keeps
-  # its own unmasked conclusion but is deliberately absent from
-  # all-checks-passed.needs, so it never delays or changes that required
-  # verdict. Under normal operation it runs on the hosted larger runner; under
-  # Windows failover (DSH_CI_FAILOVER_WINDOWS=selfhosted) it retargets onto the
-  # in-house self-hosted Windows pool. Dependabot PRs are excluded from the
-  # self-hosted pool and stay queued for the hosted runner — see the failover
-  # runbook. This Windows switch is independent of the Linux
-  # DSH_CI_FAILOVER_LINUX variable that retargets the three required Linux jobs
-  # and the all-checks-passed verdict above.
-  windows-native:
+  # Every pull request also gets real Windows-kernel signals. The former
+  # monolithic windows-native job is split into smaller jobs so one slow
+  # coverage gate does not hold up build/static results, while the total
+  # per-job process count stays lower. Observational checks are non-blocking.
+  # Dependabot PRs are excluded from the self-hosted pool and stay queued for
+  # the hosted runner.
+  windows-build:
    if: github.event_name == 'pull_request'
    runs-on: >-
      ${{ vars.DSH_CI_FAILOVER_WINDOWS == 'selfhosted'
          && github.event.pull_request.user.login != 'dependabot[bot]'
          && fromJSON('["self-hosted", "dsh-win-ci", "windows"]')
          || 'dsh-windows-2025-16core' }}
-    name: windows node 24 / native complete
-    timeout-minutes: 120
-    env:
-      DSH_COVERAGE_MAX_WORKERS: '6'
-      DSH_COVERAGE_PARTITIONS: '8'
-      # Instrumented process and polling fixtures can exceed Vitest's defaults
-      # under the complete lane's concurrent gate load.
-      DSH_COVERAGE_TEST_TIMEOUT_MS: '30000'
-      DSH_GATE_CONCURRENCY: '4'
-      DSH_PUBLINT_CONCURRENCY: '8'
+    name: windows node 24 / build
+    timeout-minutes: 60
    steps:
      - uses: actions/checkout@v6
        with:
          persist-credentials: false
-
      - name: Enable Developer Mode (symlink support)
        shell: pwsh
        run: >-
          reg add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\AppModelUnlock"
          /t REG_DWORD /f /v "AllowDevelopmentWithoutDevLicense" /d "1"
-
      - uses: pnpm/action-setup@v4
        with:
          dest: ${{ runner.temp }}/setup-pnpm-js
-
      - uses: actions/setup-node@v6
        with:
          node-version: ${{ env.PRIMARY_NODE_VERSION }}
-
-      # Extracting the many-file pnpm store cache is slower than a clean
-      # install on hosted Windows runners, and saving it adds latency after
-      # the gates. The self-hosted VM's persistent store makes caching
-      # redundant.
      - name: Install (immutable)
        shell: pwsh
        run: pnpm install --frozen-lockfile
-
-      - name: Run complete native Windows gate inventory
+      - name: Run blocking Windows builds
        shell: pwsh
-        run: pnpm run check:ci:windows-complete
+        run: pnpm run check:ci:windows-blocking
+
+  windows-coverage:
+    if: github.event_name == 'pull_request'
+    runs-on: >-
+      ${{ vars.DSH_CI_FAILOVER_WINDOWS == 'selfhosted'
+          && github.event.pull_request.user.login != 'dependabot[bot]'
+          && fromJSON('["self-hosted", "dsh-win-ci", "windows"]')
+          || 'dsh-windows-2025-16core' }}
+    name: windows node 24 / coverage
+    timeout-minutes: 120
+    env:
+      DSH_COVERAGE_MAX_WORKERS: '6'
+      DSH_COVERAGE_PARTITIONS: '4'
+      DSH_COVERAGE_TEST_TIMEOUT_MS: '30000'
+      DSH_GATE_CONCURRENCY: '3'
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+      - name: Enable Developer Mode (symlink support)
+        shell: pwsh
+        run: >-
+          reg add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\AppModelUnlock"
+          /t REG_DWORD /f /v "AllowDevelopmentWithoutDevLicense" /d "1"
+      - uses: pnpm/action-setup@v4
+        with:
+          dest: ${{ runner.temp }}/setup-pnpm-js
+      - uses: actions/setup-node@v6
+        with:
+          node-version: ${{ env.PRIMARY_NODE_VERSION }}
+      - name: Install (immutable)
+        shell: pwsh
+        run: pnpm install --frozen-lockfile
+      - name: Build before coverage
+        shell: pwsh
+        run: pnpm run build
+      - name: Run Windows coverage
+        shell: pwsh
+        run: pnpm run check:ci:coverage
+
+  windows-native-tests:
+    if: github.event_name == 'pull_request'
+    runs-on: >-
+      ${{ vars.DSH_CI_FAILOVER_WINDOWS == 'selfhosted'
+          && github.event.pull_request.user.login != 'dependabot[bot]'
+          && fromJSON('["self-hosted", "dsh-win-ci", "windows"]')
+          || 'dsh-windows-2025-16core' }}
+    name: windows node 24 / native tests
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+      - name: Enable Developer Mode (symlink support)
+        shell: pwsh
+        run: >-
+          reg add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\AppModelUnlock"
+          /t REG_DWORD /f /v "AllowDevelopmentWithoutDevLicense" /d "1"
+      - uses: pnpm/action-setup@v4
+        with:
+          dest: ${{ runner.temp }}/setup-pnpm-js
+      - uses: actions/setup-node@v6
+        with:
+          node-version: ${{ env.PRIMARY_NODE_VERSION }}
+      - name: Install (immutable)
+        shell: pwsh
+        run: pnpm install --frozen-lockfile
+      - name: Run Windows-specific native tests
+        shell: pwsh
+        run: >-
+          pnpm exec vitest run
+          --no-file-parallelism
+          --testTimeout 30000
+          packages/shell/tool-pwsh/tests/loader.spec.ts
+          packages/workflow/workflow-worker-thread/tests/workflow-worker-thread.spec.ts
+          packages/workflow/tool-ralph/tests/integration.spec.ts
+          packages/subprocess/subprocess-local/tests/process-exit.spec.ts
+          packages/session/session-persistence-sqlite/tests/differential.spec.ts
+
+  windows-observational:
+    if: github.event_name == 'pull_request'
+    continue-on-error: true
+    runs-on: >-
+      ${{ vars.DSH_CI_FAILOVER_WINDOWS == 'selfhosted'
+          && github.event.pull_request.user.login != 'dependabot[bot]'
+          && fromJSON('["self-hosted", "dsh-win-ci", "windows"]')
+          || 'dsh-windows-2025-16core' }}
+    name: windows node 24 / observational
+    timeout-minutes: 60
+    env:
+      DSH_PUBLINT_CONCURRENCY: '8'
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+      - name: Enable Developer Mode (symlink support)
+        shell: pwsh
+        run: >-
+          reg add "HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\AppModelUnlock"
+          /t REG_DWORD /f /v "AllowDevelopmentWithoutDevLicense" /d "1"
+      - uses: pnpm/action-setup@v4
+        with:
+          dest: ${{ runner.temp }}/setup-pnpm-js
+      - uses: actions/setup-node@v6
+        with:
+          node-version: ${{ env.PRIMARY_NODE_VERSION }}
+      - name: Install (immutable)
+        shell: pwsh
+        run: pnpm install --frozen-lockfile
+      - name: Run Windows observational gates
+        shell: pwsh
+        run: pnpm run check:ci:windows-observational

  # Single stable required check for branch protection: require "all checks
  # passed" instead of enumerating matrix legs whose names change as lanes and
@@ -479,7 +570,7 @@ jobs:
          && github.event.pull_request.user.login != 'dependabot[bot]'
          && fromJSON('["self-hosted", "linux", "x64", "vm-backup"]')
          || 'ubuntu-latest' }}
-    needs: [node-24, node-24-coverage, node-24-consumers, node-compat, python-sdk, python-runtime, windows]
+    needs: [node-24, node-24-coverage, node-24-consumers, node-compat, python-sdk, python-runtime, windows, windows-build, windows-native-tests]
    if: always() && github.event_name == 'pull_request'
    steps:
      - name: Fail if any needed job did not succeed
@@ -67,6 +67,7 @@ External packages that a workspace package resolves at runtime. The tier covers
 | [`eventsource-parser`](https://github.com/rexxars/eventsource-parser) | MIT |
 | [`fflate`](https://github.com/101arrowz/fflate) | MIT |
 | [`immer`](https://github.com/immerjs/immer) | MIT |
+| [`ipaddr.js`](https://github.com/whitequark/ipaddr.js) | MIT |
 | [`js-yaml`](https://github.com/nodeca/js-yaml) | MIT |
 | [`katex`](https://github.com/KaTeX/KaTeX) | MIT |
 | [`koffi`](https://github.com/Koromix/koffi) | MIT |
@@ -97,6 +98,7 @@ External packages that a workspace package resolves at runtime. The tier covers
 | [`tsx`](https://github.com/privatenumber/tsx) | MIT |
 | [`turndown`](https://github.com/mixmark-io/turndown) | MIT |
 | [`typescript`](https://github.com/microsoft/TypeScript) | Apache-2.0 |
+| [`undici`](https://github.com/nodejs/undici) | MIT |
 | [`use-sync-external-store`](https://github.com/facebook/react) | MIT |
 | [`ws`](https://github.com/websockets/ws) | MIT |
 | [`yaml`](https://github.com/eemeli/yaml) | ISC |
@@ -158,6 +158,8 @@ flowchart LR
  cfg --> plugin_dsh_base_web
  plugin_dsh_base_web_search_deepseek["web-search-deepseek<br/>@deepseek-ai/dsh-web-search-deepseek"]
  cfg --> plugin_dsh_base_web_search_deepseek
+  plugin_dsh_base_web_fetch_http["web-fetch-http<br/>@deepseek-ai/dsh-web-fetch-http"]
+  cfg --> plugin_dsh_base_web_fetch_http
  plugin_dsh_base_tool_web["tool-web<br/>@deepseek-ai/dsh-tool-web"]
  cfg --> plugin_dsh_base_tool_web
  plugin_dsh_base_tools["tools<br/>@deepseek-ai/dsh-tools"]
@@ -249,6 +251,7 @@ flowchart LR
 | `repeat-tool-reminder` | `@deepseek-ai/dsh-repeat-tool-reminder` |
 | `web` | `@deepseek-ai/dsh-web` |
 | `web-search-deepseek` | `@deepseek-ai/dsh-web-search-deepseek` |
+| `web-fetch-http` | `@deepseek-ai/dsh-web-fetch-http` |
 | `tool-web` | `@deepseek-ai/dsh-tool-web` |
 | `tools` | `@deepseek-ai/dsh-tools` |
 | `system-prompt` | `@deepseek-ai/dsh-system-prompt` |
@@ -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 apps/cli/reference/README.md
-README.md: eb33816b3ac859e8173e62635f74abbe13fb8462
-README.zh.md: ed84927aa4f211926fd32750385268dc5153b3a9
+README.md: fe8d6ef0bb296f0807de4a3ec2756016bbb510c2
+README.zh.md: e8c353f33bc9760fd6da74af33a85111cf9012aa
@@ -30,7 +30,7 @@ The shipped apps own these command lines:
 | `sdk-minimal` | no options; stdio carries the same JSON-RPC protocol |
 | `acp` | no options; stdio carries Agent Client Protocol |

-A one-shot task (`dsh --profile headless "run the tests"`) creates one fresh persisted Agent through the core registry, submits the task, waits for quiescence, and flushes the Session before deriving the last non-empty assistant text and final `turn/end` reason from its durable interval. It prints the text on stdout and exits 0 for `completed`, else 1. An invocation with no task is a usage error from that app. The shipped headless profile mounts no ApiProxy, Host, HTTP server, Web runtime, or browser client; a successful run writes nothing to stderr and opens no listening port.
+A one-shot task (`dsh --profile headless "run the tests"`) creates one fresh persisted Agent through the core registry, submits the task, waits for quiescence, and flushes the Session before deriving the last non-empty assistant text and final `turn/end` reason from its durable interval. It streams non-empty provider reasoning deltas to stderr under a `dsh: reasoning:` heading, prints only the final text on stdout, and exits 0 for `completed`, else 1; a successful response with no reasoning leaves stderr empty. An invocation with no task is a usage error from that app. The shipped headless profile mounts no ApiProxy, Host, HTTP server, Web runtime, or browser client, and opens no listening port.

 Inspect the composed tree without booting it:

@@ -89,7 +89,7 @@ New sessions in base-backed profiles default to the `workspace-write` permission

 ## Shared deployment behavior

-The base bundle mounts the native DeepSeek adapter, settings and credential providers, stable `web_search`, and disabled session telemetry. Provider credentials resolve from the inherited environment, `$DSH_HOME/.credentials.yaml`, the invoking directory's `.env`, then `$DSH_HOME/.env`; the managed document is never materialized into `process.env`, while both `.env` files are ordinary launch environment layers. Search uses `DEEPSEEK_API_KEY` and accepts `DEEPSEEK_SEARCH_BASE_URL`; `web_fetch` is disabled unless a patch layer inserts a provider and enables it.
+The base bundle mounts the native DeepSeek adapter, settings and credential providers, stable `web_search`, the public-only HTTP fetch provider, and disabled session telemetry. Provider credentials resolve from the inherited environment, `$DSH_HOME/.credentials.yaml`, the invoking directory's `.env`, then `$DSH_HOME/.env`; the managed document is never materialized into `process.env`, while both `.env` files are ordinary launch environment layers. Search uses `DEEPSEEK_API_KEY` and accepts `DEEPSEEK_SEARCH_BASE_URL`. The Web app's `cordis`, `code`, and `standard` agent presets expose `web_fetch` in every sandbox and approval mode without per-call confirmation; the provider still rejects non-public destinations before connecting.

 Session telemetry stays local by default. `DSH_TELEMETRY_MODE=FULL` streams every projected session event as OTLP/HTTP logs, while `DSH_TELEMETRY_MODE=FEEDBACK_ONLY` uploads a session-log suffix only when feedback is recorded. `DSH_TELEMETRY_OTLP_URL` selects another collector, and any non-empty `DSH_TELEMETRY_DISABLED` remains an authoritative hard opt-out. The shipped base has no telemetry redaction rule, so explicitly enabled exports can contain message text, tool arguments and results, and workspace paths; the [default-off Agent Note](../../../.agents/notes/implemented/feature/2026-08-10-telemetry-default-off.md) owns that deployment decision.

@@ -30,7 +30,7 @@
 | `sdk-minimal` | 无选项；stdio 携带相同的 JSON-RPC 协议 |
 | `acp` | 无选项；stdio 携带 Agent Client Protocol |

-一次性任务（`dsh --profile headless "run the tests"`）通过核心注册表创建一个全新的持久化 Agent（智能体），提交任务、等待完全停稳并对会话执行 flush，再从其持久化事件区间中推导最后一个非空 assistant 文本与最终 `turn/end` 原因。它在 stdout 打印文本，并在原因为 `completed` 时以 0 退出，否则以 1 退出。没有任务的调用是该应用的用法错误。随附 headless profile 不挂载 ApiProxy、Host、HTTP 服务器、Web 运行时或浏览器客户端；成功运行不会向 stderr 写入任何内容，也不会打开监听端口。
+一次性任务（`dsh --profile headless "run the tests"`）通过核心注册表创建一个全新的持久化 Agent（智能体），提交任务、等待完全停稳并对会话执行 flush，再从其持久化事件区间中推导最后一个非空 assistant 文本与最终 `turn/end` 原因。它在 `dsh: reasoning:` 标题下将非空的提供方推理分片流式写入 stderr，只在 stdout 打印最终文本，并在原因为 `completed` 时以 0 退出，否则以 1 退出；没有推理内容的成功响应会保持 stderr 为空。没有任务的调用是该应用的用法错误。随附 headless profile 不挂载 ApiProxy、Host、HTTP 服务器、Web 运行时或浏览器客户端，也不会打开监听端口。

 可在不启动的情况下检查组合出的配置树：

@@ -89,7 +89,7 @@ dsh web --help

 ## 共享部署行为

-基础组合包挂载原生 DeepSeek 适配器、settings 与凭据提供方、稳定的 `web_search` 和已禁用的会话遥测。提供方凭据依次从继承环境、`$DSH_HOME/.credentials.yaml`、调用目录的 `.env` 和 `$DSH_HOME/.env` 解析；受管文档从不物化进 `process.env`，而两个 `.env` 文件都是普通启动环境层。搜索使用 `DEEPSEEK_API_KEY` 并接受 `DEEPSEEK_SEARCH_BASE_URL`；只有 patch 层插入提供方并启用 `web_fetch` 后，该工具才可用。
+基础组合包挂载原生 DeepSeek 适配器、settings 与凭据提供方、稳定的 `web_search`、仅限公网的 HTTP fetch 提供方，以及已禁用的会话遥测。提供方凭据依次从继承环境、`$DSH_HOME/.credentials.yaml`、调用目录的 `.env` 和 `$DSH_HOME/.env` 解析；受管文档从不物化进 `process.env`，而两个 `.env` 文件都是普通启动环境层。搜索使用 `DEEPSEEK_API_KEY` 并接受 `DEEPSEEK_SEARCH_BASE_URL`。Web app 的 `cordis`、`code` 与 `standard` agent preset 会在所有 sandbox 和审批模式下暴露 `web_fetch`，无需逐次确认；提供方仍会在连接前拒绝非公开目的地址。

 会话遥测默认留在本地。`DSH_TELEMETRY_MODE=FULL` 将每条已投影会话事件作为 OTLP/HTTP 日志流式发送，`DSH_TELEMETRY_MODE=FEEDBACK_ONLY` 则仅在记录反馈时上传会话日志后缀。`DSH_TELEMETRY_OTLP_URL` 选择其他 collector。任何非空的 `DSH_TELEMETRY_DISABLED` 都是具有最终效力的遥测强制关闭开关。随附基础配置没有遥测脱敏规则，因此显式启用的导出可能包含消息文本、工具参数和结果，以及 workspace 路径；相关部署决策见[默认关闭 Agent Note](../../../.agents/notes/implemented/feature/2026-08-10-telemetry-default-off.zh.md)。

@@ -560,8 +560,9 @@ describe.skipIf(!existsSync(dshBin))('dsh BUILT bin (node lib/bin.js, no tsx)',
  it('runs the headless profile through its app-owned task positional', async () => {
    const apiKey = 'built-dsh-headless-key'
    const server = await startMockLlmServer({
-      sequence: ['success'],
+      sequence: ['reasoning_success'],
      apiKey,
+      reasoningText: 'Inspecting the published entry.',
      successText: 'published headless profile reached the mock',
    })
    const home = mkdtempSync(join(tmpdir(), 'dsh-built-headless-'))
@@ -574,7 +575,7 @@ describe.skipIf(!existsSync(dshBin))('dsh BUILT bin (node lib/bin.js, no tsx)',
      })
      expect(result.code, result.stderr).toBe(0)
      expect(result.stdout).toBe('published headless profile reached the mock')
-      expect(result.stderr).toBe('')
+      expect(result.stderr).toBe('dsh: reasoning:\nInspecting the published entry.')
      expect(server.requests.length).toBeGreaterThan(0)
      expect(server.requests.every(request => request.path === '/chat/completions')).toBe(true)
      expect(JSON.stringify(server.requests.map(request => request.body))).toContain('answer from the published entry')
--- a/Show More
+++ b/Show More