Merge remote-tracking branch 'origin/master' into dshw/pr-2250

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.md
-2026-07-16-durable-per-step-time-context.md: 3305d3644fa3baf7e1522311b98b4eb29d08f631
-2026-07-16-durable-per-step-time-context.zh.md: dd7e63710ae99d1a04bc0e87d49976e28af1dae5
+2026-07-16-durable-per-step-time-context.md: e8fd04dd52f3c42de64cf64dd16bafa236dd396a
+2026-07-16-durable-per-step-time-context.zh.md: d2611848b732d0f07cba4508a4e24aa566545a3b

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.md

@@ -8,62 +8,67 @@ English | [中文](2026-07-16-durable-per-step-time-context.zh.md)
 
 A request-only clock can tell the model the current time, but replacing that value in the system prompt removes the evidence behind earlier time-sensitive reasoning. Multi-step turns need requests to retain the readings used by preceding steps. The request must remain reconstructable after restart, and automatic compaction must account for the same timing context the model receives.
 
-A process-local refresh cache makes displayed time depend on state that cannot survive resume or be reconstructed from the durable session. Durable interval scheduling can reduce append frequency without introducing that hidden state.
+A process-local refresh cache makes displayed time depend on state that cannot survive resume. Browser-originated natural language also needs a request-owned zone: a server process zone cannot infer the user's locality, while a mutable Session or connection default lets travel or concurrent tabs reinterpret another prompt.
 
 ## Decision
 
-`@deepseek-ai/dsh-time-context` is an opt-in function plugin in `packages/context/time-context/`. The `context/` group holds bounded request-context enrichments that define neither a tool nor a service, and shipped examples do not mount this plugin because its time-zone disclosure and token cost are deployment policy. It registers a prepended `agent/pre-step` listener and, when a reading is due and the downstream decision enters, returns one additional `UserMessage`. The message carries source `{ kind: 'plugin', plugin: 'time-context' }`; a suppressed, rejected, or failed attempt appends nothing.
+`@deepseek-ai/dsh-time-context` is an opt-in function plugin in `packages/context/time-context/`. Default compositions leave its disclosure and token cost disabled; the Schedule Web overlay mounts it so the model can interpret otherwise-unqualified dates and times in the browser zone attached to the current request.
 
-The listener samples before `step/start`, then settles its reading only in the final enter decision. AgentLoop records it after `step/start` and before request derivation. A downstream rejection or failure therefore prevents the reading from entering durable history.
+The plugin prepends an `agent/pre-step` listener and delegates first. When the downstream decision enters and a reading is due, it combines that decision's final messages with durable user messages already in the open turn, derives browser-zone provenance from exact `user-rpc` sources, and appends one reading to the decision. Rejection, listener failure, or an already-aborted signal records nothing. Steering claimed after the current batch keeps ordinary next-step ownership and receives a fresh reading when that step enters.
 
-The optional `timeZone` config resolves the Node process's IANA zone once at plugin load when omitted; an explicit value is validated by `Intl.DateTimeFormat`. The timestamp includes the numeric UTC offset and resolved IANA zone.
+Each Web prompt samples the browser's IANA zone. The Host validates and canonicalizes it before binding it to the exact durable user-message source. One unique zone in the open turn resolves the request; multiple zones produce a sorted `mixed` result; no zone is `unavailable`. A resolved request tells the model to interpret unqualified dates and times in that zone. Mixed or unavailable provenance tells it to ask the user to clarify.
 
-The optional `refreshIntervalMs` config is manually validated at plugin load as a non-negative safe integer. Omission or `0` injects on every eligible preparation attempt. A positive value scans the raw session events for the most recent `user/message` with this plugin's source and injects when none exists, wall time moved backward, or the event is at least the configured age. The raw event timestamp governs even after compaction shadows the message, so scheduling persists across turns and process resume without a timer or process-local cache.
+This message-bound provenance is not copied to `SessionHeader`, a connection default, or Schedule state. Time-context owns model guidance only. A tool accepting local calendar fields must still make its own explicit boundary; Schedule therefore requires `time_zone` rather than importing this plugin's reading ([decision](../simplification/2026-08-09-explicit-schedule-time-zone.md)).
+
+The resolved browser zone also formats the reading's timestamp. Mixed or unavailable requests use the configured `timeZone` fallback, or the Node process zone resolved once at plugin load when config is omitted, while retaining the clarify policy. Every fallback is validated through `Intl.DateTimeFormat`.
+
+Each reading uses the exact snapshot source `{ kind: 'plugin', plugin: 'time-context', form: 'snapshot', sections: [{ name: 'time-context', text: <same text> }] }`. The invariant companion checks the snapshot shape, re-derives current-turn browser provenance from the original user-rpc messages, and validates the rendered timestamp zone and elapsed baseline.
+
+The optional `refreshIntervalMs` config is a non-negative safe integer. Omission or `0` injects on every eligible entered step. A positive value scans raw Session events for the latest plugin reading and injects when none exists, wall time moved backward, or the event is old enough. The event timestamp governs after compaction and resume without a process-local cache. The Schedule Web overlay omits the interval so every request step gets current browser guidance.
 
 ### Text and elapsed baselines
 
-An injected first-step reading is:
+A resolved first-step reading is:
 
 ```text
-Time sampled while preparing turn <turn>, step 1: <timestamp>
+Time sampled while preparing turn <turn>, step 1: <timestamp-in-browser-zone>
+Browser time zone for this request: <iana-zone>. Interpret otherwise-unqualified dates and times in this zone.
 Elapsed since the preceding model-visible message: <duration-or-unavailable>.
 ```
 
-The baseline is the latest preceding user, assistant, tool-result, or steering message. This includes the accepted prompt that opened an ordinary message turn. If no model-visible message exists, the duration is `unavailable`.
+Mixed and unavailable variants replace the second line with an instruction to ask for clarification. The baseline is the latest durable preceding user, assistant, or tool-result message. The prompt proposed for this step has not been appended yet; a new Session can therefore report `unavailable`.
 
-An injected later-step reading is:
+A later-step reading changes the first line's step number and ends with:
 
 ```text
-Time sampled while preparing turn <turn>, step <step>: <timestamp>
 Elapsed since the preceding step context: <duration-or-unavailable>.
 ```
 
-Their baseline is the durable event timestamp of the preceding time-context message in the same turn. If interval suppression leaves no earlier same-turn reading, the duration is `unavailable`. Duration formatting uses compact whole-second units and clamps backward wall-clock movement to zero. The explicit turn and step make every retained reading attributable to its historical preparation attempt after later turns append more context.
+That baseline is the preceding time-context event in the open turn. Missing baselines report `unavailable`; duration formatting uses compact whole-second units and clamps backward wall-clock movement to zero.
 
-### Durability and request reconstruction
+### Durability and reconstruction
 
-Each reading remains a normal surface node until compaction shadows it; positive interval scheduling never removes existing readings. A later request therefore sees the cumulative unshadowed readings that affected earlier preparation and steps, rather than a system-prompt value rewritten in place.
+An entered step appends its returned messages followed by the time reading after `step/start`, before request derivation. A later preparation failure can leave the reading in history because it records entry, not successful transmission. Each reading remains a normal surface node until compaction shadows it. A positive interval can let a later request reuse existing history without adding a fresh reading.
 
-The plugin contributes nothing to system-prompt assembly. `request/header` contains no time-context text; request reconstruction obtains the complete durable surface prefix at each `step/start`. Readings and requests need not map one-to-one because interval suppression can enter a request without appending a reading, while rejection or failure appends neither. The plugin depends on the agent registry for its lifecycle listener and does not require the system-prompt service at runtime.
-
-## Testing
-
-Unit and real-loop tests pin formatting, both elapsed baselines, interval omission and zero, threshold boundaries, cross-turn and per-session scheduling, backward-clock behavior, invalid config, resumed raw-event lookup after compaction, aborted-signal behavior, later-listener cancellation and failure, listener disposal, source and surface metadata, cumulative multi-step visibility, and absence from request headers. A keyless subprocess e2e boots the real Loader with the Headless composition, drives two ordered one-shot turns, and verifies the persisted plugin-attributed messages externally.
+The plugin contributes nothing to system-prompt assembly or `request/header`. Request reconstruction obtains the complete durable surface prefix at each `step/start`, so historical requests recover the exact time and browser policy the model saw.
 
 ## Alternatives considered
 
-- **Keep the dynamic system-prompt section and process-local refresh cache** — rejected because replacement erases earlier readings, cache state is not replayable, and a frozen request envelope would make the value stale for an entire loop instance.
-- **Replace the preceding context surface node** — rejected because replacement preserves the old node's position or shadows intervening conversation; neither represents when the new reading became visible.
-- **Inject from a background timer** — rejected because idle time has no pending request to consume the value, and timer-driven injection would create durable turns solely to report time passing.
-- **Expose time only through a tool** — rejected because ordinary temporal reasoning would require an avoidable tool round trip and would not guarantee a reading before every step.
-- **Use `agent/session-prefix`** — rejected because one loop-instance prefix cannot represent distinct step timestamps and does not accumulate historically attributable readings.
-- **Mutate assembled requests or register independent prompt variables** — rejected because request-local insertion bypasses the durable surface and separate providers can sample different instants. One attributed context message records the timestamp and elapsed baseline atomically.
-- **Default to UTC or add a time-zone detection dependency** — rejected because an explicitly mounted plugin follows its process environment unless the operator selects an IANA zone, while no server-side library can infer a remote user's zone.
-- **Mount the plugin in shipped compositions or place it in `core/`** — rejected because disclosure, time zone, freshness, and history cost are deployment choices for an optional context leaf, not product-spine policy.
+- **Replace a dynamic system-prompt value** — rejected because replacement erases prior readings and changes reconstructed historical requests.
+- **Persist a Session default zone** — rejected because the browser fact belongs to one prompt; travel and concurrent tabs must not mutate shared meaning or spread zone state through Session, fork, and persistence contracts.
+- **Copy the browser zone into a second context authority** — rejected because the original user-rpc source already owns it and the invariant can re-derive policy directly.
+- **Let Schedule consume the reading implicitly** — rejected because prose context is not a stable typed default and would couple an absolute-time parser to AgentLoop history. The model instead passes an explicit offset or zone.
+- **Use only the process zone** — rejected because deployment locality cannot infer a remote user's zone. It remains a display fallback when request provenance is absent or mixed.
+- **Expose time only through a tool** — rejected because ordinary temporal reasoning would require an avoidable round trip and would not ensure a reading before each step.
+- **Mount time-context by default** — rejected because disclosure, freshness, and history cost remain composition policy.
+
+## Verification
+
+Unit and real-loop tests pin timestamp formatting, unique/mixed/missing browser derivation, fallback display, both elapsed baselines, interval boundaries, cross-turn and resumed scheduling, backward-clock behavior, steering ownership, cancellation, exact snapshot validation, and request reconstruction. Host/client tests pin browser sampling plus validation and canonicalization at prompt entry. The keyless assembled Schedule Web scenario sends a real browser prompt, observes the same zone in the model request, and verifies that the model supplies it explicitly to `schedule_create`.
 
 ## Consequences
 
-- Omission or `0` records every eligible preparation attempt; a positive interval reduces append frequency and history growth while preserving durable scheduling across resume.
-- Timing context remains append-only until compaction shadows older surface nodes, including a preparation reading left by a later cancellation or failure.
-- The first-step duration normally measures from the prompt that opened the turn, while later-step durations measure model and tool processing since the preceding step context.
-- An omitted `timeZone` still reflects the deployment process rather than a remote user, and elapsed time still uses durable harness append boundaries rather than client-origin timestamps. Supporting client-origin time requires a separate durable input contract.
+- Browser-zone meaning is request-local and durable without changing Session, fork, JSONL, or SQLite schemas.
+- The model receives the requested browser-local assumption on each Schedule Web request step; mixed or missing provenance asks instead of guessing.
+- Tools remain explicit: context helps the model choose fields but does not become a hidden package-seam default.
+- Timing context remains append-only until compaction; a positive interval reduces history growth but can omit fresh browser guidance on later requests.

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.zh.md

@@ -8,62 +8,67 @@ Status: implemented
 
 仅存在于请求中的时钟可以告诉模型当前时间，但在系统提示词中替换这个值会移除先前对时间敏感的推理所依据的证据。在包含多个步骤的轮次中，请求需要保留先前步骤使用的读数。系统必须能在重启后重建请求，自动压缩（compaction）也必须将模型实际收到的同一份时间上下文纳入考量。
 
-进程本地刷新缓存会使显示时间依赖于一种既无法在恢复后保留、也无法从持久会话重建的状态。持久的间隔调度可以减少追加频率，而不引入这种隐藏状态。
+进程本地刷新缓存会使显示时间依赖于无法在恢复后保留的状态。来自浏览器的自然语言也需要归属于请求的时区：服务端进程时区无法推断用户所在地，而可变的会话或连接默认值会让旅行或并发标签页重新解释另一条提示词。
 
 ## 决策
 
-`@deepseek-ai/dsh-time-context` 是位于 `packages/context/time-context/`、需要显式启用的函数插件。`context/` 分组容纳有界的请求上下文增强，这些增强既不定义工具也不定义服务；已交付示例不挂载此插件，因为时区披露与 token 成本属于部署策略。它注册一个前置的 `agent/pre-step` 监听器；当应生成读数且下游决策为进入时，返回一条额外的 `UserMessage`。该消息携带来源 `{ kind: 'plugin', plugin: 'time-context' }`；被抑制、被拒绝或失败的尝试不会追加任何内容。
+`@deepseek-ai/dsh-time-context` 是位于 `packages/context/time-context/`、需要显式启用的函数插件。默认组合不启用其披露内容与 token 成本；Schedule Web overlay 会挂载它，使模型能够按附加到当前请求的浏览器时区解释未明确限定时区的日期和时间。
 
-监听器在 `step/start` 之前采样，并仅在最终决定进入时确定该读数。AgentLoop 会在 `step/start` 之后、请求派生之前记录它。因此，下游拒绝或失败会阻止读数进入持久历史。
+该插件会前置一个 `agent/pre-step` 监听器，并先行委托下游。当下游决策进入步骤且需要生成读数时，插件会把该决策的最终消息与开放轮次中已有的持久用户消息合并，从确切的 user-rpc 来源派生浏览器时区来源信息，并向该决策追加一条读数。决策被拒绝、监听器失败或信号已经中止时，不会记录任何内容。在当前批次之后被认领的 steering（中途引导）仍归属于普通的下一步骤，并在该步骤进入时获得新读数。
 
-省略可选配置 `timeZone` 时，插件在加载时解析一次 Node 进程的 IANA 时区；显式值由 `Intl.DateTimeFormat` 校验。时间戳包含数字 UTC 偏移和解析后的 IANA 时区。
+每条 Web 提示词都会采样浏览器的 IANA 时区。Host 校验并规范化该值，再将其绑定到确切的持久用户消息来源。开放轮次中唯一一个时区可解析请求；多个时区会产生排序后的 `mixed` 结果；没有时区则为 `unavailable`。解析成功的请求会告诉模型，把未限定时区的日期和时间解释为该时区。来源信息混杂或不可用时，模型会收到要求用户澄清的指令。
 
-插件在加载时手动校验可选配置 `refreshIntervalMs`，其值必须为非负安全整数。省略或设为 `0` 时，每次符合条件的准备尝试都会注入。设为正数时，插件扫描原始会话事件，查找来源属于本插件的最新 `user/message`；不存在此类事件、系统挂钟向后移动，或该事件已达到配置时长时，插件执行注入。即使压缩已隐藏消息，调度仍以原始事件时间戳为准，因此该机制无需计时器或进程本地缓存，也能跨轮次和进程恢复持续生效。
+这种与消息绑定的来源信息不会复制到 `SessionHeader`、连接默认值或 Schedule 状态。Time-context 只负责模型指导。接受本地日历字段的工具仍必须自行定义显式边界；因此 Schedule 要求 `time_zone`，而不是导入该插件的读数（[决策](../simplification/2026-08-09-explicit-schedule-time-zone.md)）。
+
+解析后的浏览器时区也用于格式化读数中的时间戳。请求来源信息混杂或不可用时，使用配置的 `timeZone` 回退值；如果省略该配置，则使用插件加载时解析一次的 Node 进程时区，同时仍保留要求澄清的策略。每个回退值都经 `Intl.DateTimeFormat` 校验。
+
+每个读数都使用确切的快照来源 `{ kind: 'plugin', plugin: 'time-context', form: 'snapshot', sections: [{ name: 'time-context', text: <same text> }] }`。不变式配套模块会校验快照形状，从原始 user-rpc 消息重新派生当前轮次的浏览器来源信息，并校验渲染的时间戳时区与经过时长基线。
+
+可选配置 `refreshIntervalMs` 必须是非负安全整数。省略或设为 `0` 时，每个符合条件且已进入的步骤都会注入。设为正数时，插件会扫描原始会话事件，查找最新的插件读数；不存在读数、挂钟时间倒退或事件已达到相应时长时执行注入。事件时间戳在压缩和恢复后仍是判断依据，无需进程本地缓存。Schedule Web overlay 会省略该间隔，使每个请求步骤都获得当前浏览器时区指导。
 
 ### 文本与时长基线
 
-第一个步骤的注入读数为：
+已解析的第一步读数为：
 
 ```text
-Time sampled while preparing turn <turn>, step 1: <timestamp>
+Time sampled while preparing turn <turn>, step 1: <timestamp-in-browser-zone>
+Browser time zone for this request: <iana-zone>. Interpret otherwise-unqualified dates and times in this zone.
 Elapsed since the preceding model-visible message: <duration-or-unavailable>.
 ```
 
-基线是前一条用户消息、助手消息、工具结果或 steering（中途引导）消息。对于普通消息轮次，这包括开启轮次的已接受提示词。如果不存在模型可见消息，时长为 `unavailable`。
+混杂和不可用的变体会把第二行替换为要求澄清的指令。基线是最新一条在其之前持久化的用户、助手或工具结果消息。为该步骤拟议的提示词尚未追加，因此新会话可能报告 `unavailable`。
 
-后续步骤的注入读数为：
+后续步骤读数会改变第一行的步骤号，并以下行结束：
 
 ```text
-Time sampled while preparing turn <turn>, step <step>: <timestamp>
 Elapsed since the preceding step context: <duration-or-unavailable>.
 ```
 
-其基线是同一轮次中上一条时间上下文消息的持久事件时间戳。如果间隔抑制导致同一轮次中没有更早的读数，时长为 `unavailable`。时长采用紧凑的整秒单位，并在系统挂钟向后移动时钳制为零。显式的轮次号和步骤号使每个保留的读数在后续轮次追加更多上下文后，仍可归属于对应的历史准备尝试。
+其基线是开放轮次中的前一个 time-context 事件。缺少基线时报告 `unavailable`；时长采用紧凑的整秒单位，并在挂钟时间倒退时限制为零。
 
-### 持久性与请求重建
+### 持久性与重建
 
-每个读数都作为普通表层节点保留，直至压缩将其隐藏；正数间隔调度绝不会移除已有读数。因此，后续请求会看到影响先前准备过程和步骤且尚未被隐藏的累计读数，而不是一个被原地改写的系统提示词值。
+已进入的步骤会在 `step/start` 之后、请求派生之前，先追加其返回消息，再追加时间读数。后续准备失败时，读数可能留在历史中，因为它记录的是步骤进入，而不是成功传输。每个读数都作为普通表层节点保留，直至压缩将其遮蔽。正数间隔可以让后续请求复用现有历史，而不添加新读数。
 
-插件不向系统提示词组装贡献任何内容。`request/header` 不包含时间上下文文本；请求重建从每个 `step/start` 取得完整的持久表层前缀。读数与请求无需一一对应，因为间隔抑制可以让请求进入步骤而不追加读数，拒绝或失败则两者都不追加。插件通过 agent 注册表使用生命周期监听器，运行时不需要系统提示词服务。
+插件不向系统提示词组装或 `request/header` 贡献任何内容。请求重建会在每个 `step/start` 取得完整的持久表层前缀，因此历史请求可以还原模型看到的确切时间与浏览器策略。
 
-## 测试
+## 已考虑的替代方案
 
-单元测试和真实 agent loop（智能体循环）测试固定格式化、两种时长基线、间隔省略和零值、阈值边界、跨轮次和各会话独立调度、挂钟后退行为、无效配置、压缩后基于恢复会话的原始事件查找、已中止信号行为、后续监听器取消和失败、监听器 dispose（资源释放）、来源与表层元数据、多步骤累计可见性，以及请求头中不存在时间上下文。无密钥子进程 e2e 测试使用 Headless 组合启动真实 loader，依次驱动两个单次任务轮次，并从外部校验持久化且来源归属于插件的消息。
+- **替换动态系统提示词值**：不予采纳，因为替换会抹去先前读数，并改变重建后的历史请求。
+- **持久化会话默认时区**：不予采纳，因为浏览器事实只属于一条提示词；旅行与并发标签页不得修改共享含义，也不得把时区状态扩散到会话、fork 与持久化约定中。
+- **把浏览器时区复制到第二个上下文权威**：不予采纳，因为原始 user-rpc 来源已经拥有该值，不变式可以直接重新派生策略。
+- **让 Schedule 隐式消费读数**：不予采纳，因为自然语言上下文不是稳定的类型化默认值，而且这会把绝对时间解析器耦合到 AgentLoop 历史。模型会改为传入显式偏移量或时区。
+- **只使用进程时区**：不予采纳，因为部署所在地无法推断远程用户的时区。请求来源信息缺失或混杂时，它仍可作为显示回退值。
+- **只通过工具提供时间**：不予采纳，因为普通时间推理会产生本可避免的往返，也无法确保每个步骤之前都有读数。
+- **默认挂载 time-context**：不予采纳，因为披露内容、新鲜度与历史成本仍属于组合策略。
 
-## 考虑过的替代方案
+## 验证
 
-- **保留动态系统提示词区段和进程本地刷新缓存**——不予采纳，因为替换会抹去先前读数，缓存状态无法回放，而且冻结的请求内容集合会使该值在整个 agent loop 实例期间保持陈旧。
-- **替换前一条上下文表层节点**——不予采纳，因为替换会保留旧节点的位置或隐藏中间的会话内容；两者都不能表达新读数何时开始可见。
-- **通过后台计时器注入**——不予采纳，因为空闲期间没有待处理请求消费该值，而且计时器驱动的注入会仅为报告时间流逝而创建持久轮次。
-- **只通过工具提供时间**——不予采纳，因为普通时间推理会产生本可避免的工具往返，也不能保证每个步骤之前都有读数。
-- **使用 `agent/session-prefix`**——不予采纳，因为一个 loop 实例前缀无法表示不同的步骤时间戳，也不会累计具有历史归属的读数。
-- **修改已组装的请求或注册独立提示词变量**——不予采纳，因为请求内插入会绕过持久表层，不同提供方也可能在不同时间采样。一条带来源归属的上下文消息会原子地记录时间戳和时长基线。
-- **默认使用 UTC 或增加时区检测依赖**——不予采纳，因为显式挂载的插件默认遵循其进程环境，除非操作方选择 IANA 时区，而任何服务端库都无法推断远程用户的时区。
-- **在已交付组合中挂载插件，或把它放进 `core/`**——不予采纳，因为披露内容、时区、新鲜度和历史成本是可选上下文叶节点的部署选择，不是产品主干策略。
+单元测试和真实 agent loop（智能体循环）测试固定时间戳格式化、唯一／混杂／缺失浏览器时区的派生、回退显示、两种经过时长基线、间隔边界、跨轮次与恢复后的调度、挂钟倒退行为、steering 归属、取消、精确快照校验和请求重建。Host/client 测试固定浏览器采样，以及提示词进入时的校验与规范化。无密钥的组装 Schedule Web 场景发送一条真实浏览器提示词，在模型请求中观察到同一时区，并验证模型把该时区显式传给 `schedule_create`。
 
 ## 后果
 
-- 省略 `refreshIntervalMs` 或设为 `0` 时，每次符合条件的准备尝试都会留下记录；正数间隔会减少追加频率和历史增长，同时使持久调度在恢复后继续生效。
-- 时间上下文仅追加并保留到压缩隐藏旧表层节点为止，其中也包括后续取消或失败所留下的准备读数。
-- 第一个步骤的时长通常从开启轮次的提示词起算，后续步骤的时长则反映自上一条步骤上下文以来的模型与工具处理时间。
-- 省略 `timeZone` 时仍采用部署进程而非远程用户的时区，时长仍采用 harness 的持久追加边界而非客户端来源时间戳。若要支持客户端来源的时间，需要另行建立持久输入约定。
+- 浏览器时区含义归属于请求并可持久重建，无需更改会话、fork、JSONL 或 SQLite schema。
+- 模型在每个 Schedule Web 请求步骤中都会收到所请求的浏览器本地假设；来源信息混杂或缺失时会询问，而不是猜测。
+- 工具仍保持显式边界：上下文帮助模型选择字段，但不会成为包 seam 上隐藏的默认值。
+- 时间上下文仅追加并保留到压缩为止；正数间隔会减少历史增长，但也可能使后续请求缺少新的浏览器时区指导。

.agents/notes/implemented/feature/2026-07-31-telemetry-anonymous-user-id.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-31-telemetry-anonymous-user-id.md
-2026-07-31-telemetry-anonymous-user-id.md: 4d5d14cb63ebbf9d9f71c729da998ad9804e56bf
-2026-07-31-telemetry-anonymous-user-id.zh.md: 69c48076ceb97a4bbc3cc35a2aaa3e632b1b1b06
+2026-07-31-telemetry-anonymous-user-id.md: 6231065126d0884abe3ba6f45a40c42365fe520b
+2026-07-31-telemetry-anonymous-user-id.zh.md: 4f6bab8df5d479d2df42194581a4229c6de221e6

.agents/notes/implemented/feature/2026-07-31-telemetry-anonymous-user-id.md

@@ -10,7 +10,7 @@ Session telemetry is mounted by default ([default-mount Note](2026-07-31-web-tel
 
 ## Decision
 
-`getOrCreateAnonymousUserId()` returns the bare UUID line in `$DSH_HOME/.userid` (resolved by `resolveDshHome`, `$DSH_HOME` > `~/.dsh`), minting and persisting a random UUID v4 on first use; the backend constructor carries it as the Resource's `user.id` (the OTel semconv user attribute), once per export batch. The original implementation lived inside `session-telemetry-otel` because no second real consumer existed. `/feedback` later became that consumer, so [the shared-id decision](../architecture/2026-08-07-shared-feedback-telemetry-user-id.md) moves ownership to `@deepseek-ai/dsh-user-id` without changing the storage, anonymity, concurrency, or loss semantics recorded here.
+`getOrCreateAnonymousUserId()` returns the bare UUID line in `$DSH_HOME/.userid` (resolved by `resolveDshHome`, `$DSH_HOME` > `~/.dsh`), minting and persisting a random UUID v4 on first use; the backend constructor carries it as the Resource's `user.id` (the OTel semconv user attribute), once per export batch. The original implementation lived inside `session-telemetry-otel` because no second real consumer existed. `/feedback` later became that consumer, so [the shared-id decision](../architecture/2026-08-07-shared-feedback-telemetry-user-id.md) moves ownership to `@deepseek-ai/dsh-user-id` without changing the storage, anonymity, concurrency, or loss semantics recorded here. [Direct DeepSeek request identity](2026-08-11-deepseek-request-user-id-header.md) is a third consumer of the same id.
 
 | Ruling | Value | Rationale |
 |---|---|---|
@@ -22,8 +22,8 @@ Session telemetry is mounted by default ([default-mount Note](2026-07-31-web-tel
 | Write failure | Best-effort: return the in-memory id | Telemetry is never blocked by a read-only home |
 | Report position | Resource attribute, not per-record attributes | Once per batch suffices for Resource-dimension aggregation; per-record injection would touch the seam contract and grow the wire |
 | semconv dependency | `@opentelemetry/semantic-conventions` is not imported | One string constant does not justify a dependency |
-| Home | `@deepseek-ai/dsh-user-id`, shared by the OTel backend and `/feedback` | The second real consumer now exists; direct feedback must not depend on an exporter backend |
-| Separate switch | None | Either consumer can create the identity; `DSH_TELEMETRY_DISABLED` stops telemetry reporting but does not disable feedback acknowledgement |
+| Home | `@deepseek-ai/dsh-user-id`, shared by the OTel backend, `/feedback`, and direct DeepSeek requests | Consumers share one storage contract without depending on an exporter backend |
+| Separate switch | None | Any consumer can create the identity; `DSH_TELEMETRY_DISABLED` stops telemetry reporting but does not disable feedback acknowledgement or the DeepSeek request header |
 
 ## Alternatives considered
 
@@ -38,6 +38,6 @@ Session telemetry is mounted by default ([default-mount Note](2026-07-31-web-tel
 ## Consequences
 
 - One `$DSH_HOME` is one stable user in the OTel feed; separate homes are separate users by construction, with no cross-home linking mechanism.
-- The OTel feed and `/feedback` share `.userid`.
+- The OTel feed, `/feedback`, and direct DeepSeek requests share `.userid`.
 - Deleting `.userid` resets the identity (effective next launch); on an unwritable home each process holds its own in-memory id until the home becomes writable.
 - The [default-mount Note](2026-07-31-web-telemetry-default-mount.md)'s identity follow-up is closed for the anonymous-user-id part by this decision; hostname/surface dimensions, the redaction rule, and the usage-metrics track remain open.

.agents/notes/implemented/feature/2026-07-31-telemetry-anonymous-user-id.zh.md

@@ -10,7 +10,7 @@ session telemetry 已默认挂载（[默认挂载 Note](2026-07-31-web-telemetry
 
 ## Decision
 
-`getOrCreateAnonymousUserId()` 返回 `$DSH_HOME/.userid`（`resolveDshHome` 解析，`$DSH_HOME` > `~/.dsh`）中的裸 UUID 行，首用生成随机 UUID v4 并落盘；后端构造时把它作为 Resource 的 `user.id`（OTel semconv 标准用户属性）随每批导出携带一次。原始实现位于 `session-telemetry-otel`，因为当时不存在第二个真实消费方。`/feedback` 后来成为该消费方，因此[共享 id 决策](../architecture/2026-08-07-shared-feedback-telemetry-user-id.md)将所有权移交给 `@deepseek-ai/dsh-user-id`，但不改变本 Note 记录的存储、匿名、并发与丢失语义。
+`getOrCreateAnonymousUserId()` 返回 `$DSH_HOME/.userid`（`resolveDshHome` 解析，`$DSH_HOME` > `~/.dsh`）中的裸 UUID 行，首用生成随机 UUID v4 并落盘；后端构造时把它作为 Resource 的 `user.id`（OTel semconv 标准用户属性）随每批导出携带一次。原始实现位于 `session-telemetry-otel`，因为当时不存在第二个真实消费方。`/feedback` 后来成为该消费方，因此[共享 id 决策](../architecture/2026-08-07-shared-feedback-telemetry-user-id.md)将所有权移交给 `@deepseek-ai/dsh-user-id`，但不改变本 Note 记录的存储、匿名、并发与丢失语义。[直连 DeepSeek 请求身份](2026-08-11-deepseek-request-user-id-header.md)是同一 id 的第三个消费方。
 
 | 裁定 | 取值 | 理由 |
 |---|---|---|
@@ -22,8 +22,8 @@ session telemetry 已默认挂载（[默认挂载 Note](2026-07-31-web-telemetry
 | 写失败 | best-effort 返回内存 id | telemetry 永不因 home 只读被阻塞 |
 | 上报位置 | Resource 属性，非逐条 attributes | 每批一次即够接收端按 Resource 维度聚合；逐条注入要动 seam 约定且涨 wire 体积 |
 | semconv 依赖 | 不引 `@opentelemetry/semantic-conventions` 包 | 一个字符串常量不值一个依赖 |
-| 落点 | `@deepseek-ai/dsh-user-id`，由 OTel 后端与 `/feedback` 共享 | 第二个真实消费方已经出现；直接反馈不能依赖导出后端 |
-| 单独开关 | 无 | 任一消费方都可创建该身份；`DSH_TELEMETRY_DISABLED` 会停止遥测上报，但不会禁用反馈确认 |
+| 落点 | `@deepseek-ai/dsh-user-id`，由 OTel 后端、`/feedback` 与直连 DeepSeek 请求共享 | 消费方共用同一存储契约，且不依赖导出后端 |
+| 单独开关 | 无 | 任一消费方都可创建该身份；`DSH_TELEMETRY_DISABLED` 会停止遥测上报，但不会禁用反馈确认或 DeepSeek 请求头 |
 
 ## Alternatives considered
 
@@ -38,6 +38,6 @@ session telemetry 已默认挂载（[默认挂载 Note](2026-07-31-web-telemetry
 ## Consequences
 
 - 一个 `$DSH_HOME` 在 OTel 回流中是一个稳定用户；不同 home 在构造上就是不同用户，无跨 home 关联机制。
-- OTel 回流与 `/feedback` 共享 `.userid`。
+- OTel 回流、`/feedback` 与直连 DeepSeek 请求共享 `.userid`。
 - 删除 `.userid` 即重置身份（下次启动生效）；home 不可写时每进程各自持有一个内存 id 直至恢复可写。
 - [默认挂载 Note](2026-07-31-web-telemetry-default-mount.md) 的身份 follow-up 中「匿名用户 id」项由本决定关闭；hostname/surface 维度与脱敏规则、usage-metrics track 仍是待办。

.agents/notes/implemented/feature/2026-07-31-web-workspace-file-links.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-07-31-web-workspace-file-links.md
-2026-07-31-web-workspace-file-links.md: 977a24adf2f87db8a34027f36feccee44d231f9d
-2026-07-31-web-workspace-file-links.zh.md: c350fad4c74c3ed6121aa5cb1243b43140196c68
+2026-07-31-web-workspace-file-links.md: 611f012e0201fa9002ea473ef9a107841bf83bcc
+2026-07-31-web-workspace-file-links.zh.md: 5f450493e960604db8c11d679c8f279f7f5b1d1a

.agents/notes/implemented/feature/2026-07-31-web-workspace-file-links.md

@@ -14,11 +14,11 @@ Two distinct defects sat behind that. The transcript never said what a turn had
 
 ## Decision
 
-**A finished turn ends with the files it produced.** The row is its own plugin, `@deepseek-ai/dsh-client-ui-deliverables`, registered into the `conversation.chat.turnTail` hole the chat view renders between a closing message's body and its IconActions — ui-conversation owns the hole and the owner currency (nodes, closing seq, `openFile`), the plugin owns every policy. `producedForClosing` reads the paths off the mutation tools' own follow-along `locations` — a diff card, or a generic card whose `kind` is `edit` (the shape `str_replace_editor`'s insert presents) — so a turn's output is listed whether or not the closing message named it, and a new mutation tool joins by declaring what it does rather than by being added to a list. Reads, deletes, and failed calls contribute nothing; a path appears once per turn in first-seen order; accumulation resets on the turn boundary, so a turn that mutates and then ends without content text cannot spill into the next turn's row. One cordis.yml line composes the surface in or out; the unregistered hole renders nothing.
+**A finished turn ends with the files it produced.** The row is its own plugin, `@deepseek-ai/dsh-client-ui-deliverables`, registered into the `conversation.chat.turnTail` hole the chat view renders between a closing message's body and its IconActions — ui-conversation owns the hole and the owner currency (nodes, closing seq, `openFile`), the plugin owns every policy. `producedForClosing` reads the paths off the mutation tools' own follow-along `locations` — a diff card, or a generic card whose `kind` is `edit` (the shape `str_replace_editor`'s insert presents) — so a turn's output is listed whether or not the closing message named it, and a new mutation tool joins by declaring what it does rather than by being added to a list. Reads, deletes, and failed calls contribute nothing; a path appears once per turn in first-seen order; accumulation resets on the turn boundary, so a turn that mutates and then ends without content text cannot spill into the next turn's row. The single-line lane measures its chips and localized remainder, then shows the largest prefix that fits (up to six) plus `+ N files`. One cordis.yml line composes the surface in or out; the unregistered hole renders nothing.
 
 **The path link reads as a link.** Underlined at rest, not only on hover. This is the smaller half of the diff and the larger half of the fix.
 
-**Opening stays the Host's job, and prefers the default browser.** `host.openPath` hands the path to the operating system, which yields a `file://` document in a real browser: full page capabilities, and no reachability into `/api`, because a `file://` document is not same-origin with it. Measured on the reported artifact: `localStorage` works, the theme toggle flips, the tabs switch, and `fetch` to the API fails. For documents a browser renders — `.html`, `.htm`, `.xhtml`, `.svg` — the opener resolves the default *browser* rather than the type's default application when the platform can name one, because a developer who binds `.html` to an editor would otherwise click a produced page and get source code. macOS reads the LaunchServices `https` handler and desktop Linux reads `$BROWSER`; either falls back to the default application when no browser can be named. Windows uses its registered association, and WSL first translates the path before using that same Windows handoff.
+**Opening stays the Host's job, and prefers the default browser.** `host.openPath` hands the path to the operating system, which yields a `file://` document in a real browser: full page capabilities, and no reachability into `/api`, because a `file://` document is not same-origin with it. Measured on the reported artifact: `localStorage` works, the theme toggle flips, the tabs switch, and `fetch` to the API fails. For documents a browser renders — `.html`, `.htm`, `.xhtml`, `.svg` — the opener resolves the default *browser* rather than the type's default application when the platform can name one, because a developer who binds `.html` to an editor would otherwise click a produced page and get source code. macOS reads the LaunchServices `https` handler and desktop Linux reads `$BROWSER`; either falls back to the default application when no browser can be named. Windows uses its registered association, and WSL first translates the path before using that same Windows handoff. When files are hidden, **Show in folder** passes `.` through the same owner `openFile`; it appears only for a loopback page whose current `host.describe.canOpenPath` permits native opening. Other deployments omit it, with `nativeOpen: false` available when desktop detection would be a false positive.
 
 **Serving workspace files over HTTP is out of scope, and so are non-local clients.** Serving files from the harness itself — same-origin with `/api`, behind `CSP: sandbox`, or from a second listener whose own port gives served documents their own origin — was rejected with the product scope: previews for a browser that is not on the Host machine are not supported, so the Host opener answers the supported case completely and the HTTP machinery would answer only the unsupported one.
 
@@ -28,8 +28,9 @@ Two distinct defects sat behind that. The transcript never said what a turn had
 - **Same-origin HTTP serving without isolation** — measurably unsafe, and recorded so nobody retries it: a document served beside `/api` drove `settings.describe` to a `200` with full data and `session.list` to 35 KB of every session's transcript, from a page that need not be agent-authored at all (a read row makes every file in a cloned repository openable).
 - **`Content-Security-Policy: sandbox` over that same-origin serving** — closes the hole by taking the document's origin away, which measurably breaks the pages this feature exists to show: the reported artifact throws `SecurityError` on load, and because an uncaught exception aborts the rest of its `<script>`, every listener declared after that line — theme toggle, mobile menu, model tabs — never binds. Two of the four artifacts in the reporting user's workspace were dead pages under it, and they still rendered perfectly, so the breakage was invisible.
 - **Linkifying paths in the assistant's closing message** — the shape a user asks for ("put the link at the end"), but it makes rendering depend on the model spelling a path recognizably. The tool calls already carry `locations` as structured fact, so the produced-files row consumes that instead.
+- **Horizontal chip scrolling** — keeps every file in the DOM but makes the hidden tail undiscoverable, adds a nested horizontal gesture to the transcript, and provides no exact account of what is out of view. One measured line with a stable remainder preserves the answer's vertical rhythm and keeps the omission explicit.
 - **An embedded WebView in the desktop shell** — the strongest isolation available, since the preview then runs in a container the product owns rather than in the user's browser. It belongs to the desktop shell's own design, not to this surface, and is recorded here as the direction a future preview capability should take.
 
 ## Consequences
 
-Every existing file affordance changed at once: write, edit, read, and the generic single-file card all reach `openFile`, so the link fix and the browser preference apply to all of them without a per-row change. The keyless web lane (`apps/web/tests/produced-files.e2e.ts`) cold-seeds a recorded write turn and pins the row in the assembled application; it deliberately does not click, because the click hands a path to the Host's opener and would launch a real application on the machine running the suite. A produced file opens as a `file://` document, which cannot `fetch` its own siblings (a multi-file artifact that loads `./data.json` breaks, while `<script src>`, `<img>`, and CSS `@import` are unaffected) — the one capability HTTP serving had that this does not. A client reached over the network sees nothing when it clicks: `host.openPath` runs on the Host and is loopback-pinned by the `/api` trust fence. That is the scope decision showing through, not a defect, and it is why the row keeps the full path in its `title` for a reader who can only copy it. Markdown opens in whatever the platform hands `.md`, usually an editor rather than a renderer; rendering it inside the product is a separate, deferred surface.
+Every existing file affordance changed at once: write, edit, read, and the generic single-file card all reach `openFile`, so the link fix and browser preference apply without per-row changes. The assembled Web test covers overflow geometry and a one-click Host handoff without launching a native application. A produced `file://` document cannot `fetch` its own siblings (while `<script src>`, `<img>`, and CSS `@import` work), the one capability HTTP serving had that this does not. Remote clients keep the chips but omit the folder action; the full path remains in each chip's `title`. Markdown still opens in the platform's `.md` application; in-product rendering is separate work.

.agents/notes/implemented/feature/2026-07-31-web-workspace-file-links.zh.md

@@ -14,11 +14,11 @@ Status: implemented
 
 ## 决定
 
-**完成的一轮以它产出的文件收尾。** 该行是独立插件 `@deepseek-ai/dsh-client-ui-deliverables`，注册进 chat 视图在收尾消息正文与其 IconActions 之间渲染的 `conversation.chat.turnTail` 空位——ui-conversation 拥有空位与 owner 通货（节点、收尾 seq、`openFile`），插件拥有全部策略。`producedForClosing` 从改写工具自身的跟随文件 `locations` 中读出路径——diff 卡片，或 `kind` 为 `edit` 的 generic 卡片（即 `str_replace_editor` 的 insert 所呈现的形状）——因此无论收尾消息是否点名，这一轮的产出都会被列出；新的改写工具靠声明自己做了什么加入，而不是靠被加进某张名单。read、删除与失败的调用不贡献任何条目；同一路径在一轮内按首见顺序只出现一次；累积在 turn 边界重置，因此一轮若先改写文件、随后没有正文内容就结束，不会溢进下一轮的行里。cordis.yml 中的一行即可把该交互面组合进来或去掉；未注册的空位什么也不渲染。
+**完成的一轮以它产出的文件收尾。** 该行是独立插件 `@deepseek-ai/dsh-client-ui-deliverables`，注册进 chat 视图在收尾消息正文与其 IconActions 之间渲染的 `conversation.chat.turnTail` 空位——ui-conversation 拥有空位与 owner 通货（节点、收尾 seq、`openFile`），插件拥有全部策略。`producedForClosing` 从改写工具自身的跟随文件 `locations` 中读出路径——diff 卡片，或 `kind` 为 `edit` 的 generic 卡片（即 `str_replace_editor` 的 insert 所呈现的形状）——因此无论收尾消息是否点名，这一轮的产出都会被列出；新的改写工具靠声明自己做了什么加入，而不是靠被加进某张名单。read、删除与失败的调用不贡献任何条目；同一路径在一轮内按首见顺序只出现一次；累积在 turn 边界重置，因此一轮若先改写文件、随后没有正文内容就结束，不会溢进下一轮的行里。单行 lane 会测量 chip 和本地化剩余计数，再显示能放下的最大前缀（至多六个）及 `+ N 个文件`。cordis.yml 中的一行即可把该交互面组合进来或去掉；未注册的空位什么也不渲染。
 
 **路径链接读得出是链接。** 静止状态下就带下划线，而不只在悬停时。这是本次改动中更小的那一半，却是修复中更大的那一半。
 
-**打开仍然是 Host 的职责，并且优先选用默认浏览器。** `host.openPath` 把路径交给操作系统，得到的是真实浏览器里的一份 `file://` 文档：页面能力完整，且够不到 `/api`——因为 `file://` 文档与它并不同源。在所报告的那份产物上实测：`localStorage` 可用、主题切换生效、tabs 可切换，而对 API 的 `fetch` 失败。对浏览器能渲染的文档——`.html`、`.htm`、`.xhtml`、`.svg`——平台能够确定默认浏览器时，打开器解析的是默认**浏览器**而非该类型的默认应用，因为把 `.html` 绑给编辑器的开发者，否则点开一个产出的页面得到的会是源码。macOS 读取 LaunchServices 的 `https` 处理程序，桌面 Linux 读取 `$BROWSER`；无法确定浏览器时，两者都会回退到默认应用。Windows 使用其注册的文件关联，WSL 则先转换路径，再使用同一 Windows 交接。
+**打开仍然是 Host 的职责，并且优先选用默认浏览器。** `host.openPath` 把路径交给操作系统，得到的是真实浏览器里的一份 `file://` 文档：页面能力完整，且够不到 `/api`——因为 `file://` 文档与它并不同源。在所报告的那份产物上实测：`localStorage` 可用、主题切换生效、tabs 可切换，而对 API 的 `fetch` 失败。对浏览器能渲染的文档——`.html`、`.htm`、`.xhtml`、`.svg`——平台能够确定默认浏览器时，打开器解析的是默认**浏览器**而非该类型的默认应用，因为把 `.html` 绑给编辑器的开发者，否则点开一个产出的页面得到的会是源码。macOS 读取 LaunchServices 的 `https` 处理程序，桌面 Linux 读取 `$BROWSER`；无法确定浏览器时，两者都会回退到默认应用。Windows 使用其注册的文件关联，WSL 则先转换路径，再使用同一 Windows 交接。存在隐藏文件时，**在文件夹中显示**会把 `.` 经由同一 owner `openFile` 传递；它只在 loopback 页面的当前 `host.describe.canOpenPath` 允许原生打开时出现。其他部署会省略它；桌面探测误报时可配置 `nativeOpen: false`。
 
 **以 HTTP 提供工作区文件不在范围内，非本机客户端亦然。** 由 harness 自己提供文件——与 `/api` 同源、置于 `CSP: sandbox` 之后、或交给一个以自身端口给所服务文档独立源的第二监听器——随产品范围一并否决：不为“浏览器不在 Host 机器上”的场景提供预览，因此 Host 打开器完整回答受支持的场景，而那套 HTTP 机制只会回答不受支持的那个。
 
@@ -28,8 +28,9 @@ Status: implemented
 - **同源 HTTP 提供且不加隔离**——经实测不安全，记录在此以免有人重试：与 `/api` 并排提供的文档把 `settings.describe` 打到 `200` 并拿到完整数据，把 `session.list` 打到 35 KB 的全部会话转录，而这个页面根本不必由 agent 撰写（一条 read 行就让 clone 下来的仓库里任何文件变得可打开）。
 - **在那套同源提供之上加 `Content-Security-Policy: sandbox`**——它以剥夺文档的源来堵住这个洞，而这经实测会破坏本功能存在的意义所在的那类页面：所报告的产物在加载时抛 `SecurityError`，又因为未捕获异常会中止其 `<script>` 的其余部分，该行之后声明的所有监听器——主题切换、移动端菜单、模型 tabs——统统不会绑定。报告者工作区里四份产物有两份在它之下是死页面，而且它们渲染得完美无缺，所以这种破坏是看不见的。
 - **把路径在助手的收尾消息里链接化**——这是用户开口要的形状（“在结尾附上链接”），但它让渲染取决于模型是否把路径拼写得可识别。工具调用已经把 `locations` 作为结构化事实携带，产出文件行消费的正是它。
+- **让文件 chip 横向滚动**——这样会把每个文件都留在 DOM 中，却使隐藏的尾部难以发现，在 transcript 内增加一层横向手势，也无法精确说明视口外还有什么。经过测量的一行和稳定的剩余计数既保留回答的纵向节奏，也明确呈现省略量。
 - **桌面端外壳中的内嵌 WebView**——可得到的最强隔离，因为那时预览跑在产品自己拥有的容器里，而不是用户的浏览器里。它属于桌面端外壳自身的设计，而非本交互面，记录在此作为未来预览能力应走的方向。
 
 ## 影响
 
-现有的每一处文件交互都同时改变了：write、edit、read 与通用单文件卡片都汇到 `openFile`，因此链接可见性修复与浏览器优先策略无需逐行改动即适用于全部。无密钥 web 通道（`apps/web/tests/produced-files.e2e.ts`）冷播种一段录制的 write 轮次，在组装后的应用中钉住该行；它刻意不点击，因为点击会把路径交给 Host 打开器，从而在跑测试的机器上启动一个真实应用。产出的文件以 `file://` 文档打开，它无法 `fetch` 自己的同级文件（一个加载 `./data.json` 的多文件产物会坏，而 `<script src>`、`<img>` 与 CSS `@import` 不受影响）——这是 HTTP 提供曾有、而此处没有的那一项能力。经网络访问的客户端点击后看不到任何东西：`host.openPath` 在 Host 上运行，且被 `/api` 信任 fence 钉在回环。那是范围决定的显现，不是缺陷，也正因如此该行把完整路径保留在 `title` 中，供只能复制它的读者使用。markdown 会由平台交给 `.md` 的默认处理程序打开，通常是编辑器而非渲染器；在产品内渲染它是另一个被推迟的交互面。
+现有的每一处文件交互都同时改变了：write、edit、read 与通用单文件卡片都汇到 `openFile`，因此链接修复与浏览器优先策略无需逐行改动。组装层 Web 测试覆盖溢出几何和单次点击的 Host 交接，且不会启动原生应用。产出的 `file://` 文档无法 `fetch` 同级文件（但 `<script src>`、`<img>` 和 CSS `@import` 可用），这是 HTTP 提供曾有、而此处没有的能力。远程客户端保留 chip，但省略文件夹操作；每个 chip 的 `title` 仍保留完整路径。Markdown 仍由平台的 `.md` 应用打开；产品内渲染属于另一项工作。

.agents/notes/implemented/feature/2026-08-05-durable-web-schedule.i18n.yaml

@@ -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/2026-08-05-durable-web-schedule.md
+2026-08-05-durable-web-schedule.md: 689a9c985eb8c732740aa127a1fcf4c5107e5ae5
+2026-08-05-durable-web-schedule.zh.md: 070bf866ca38693db03609c93dc349ac2c110160

.agents/notes/implemented/feature/2026-08-05-durable-web-schedule.md

@@ -0,0 +1,86 @@
+# Agent Note: Durable Session-local reminders
+
+Status: implemented
+
+English | [中文](2026-08-05-durable-web-schedule.zh.md)
+
+## Problem
+
+A reminder created inside a conversation must remain attributable to that exact Session and survive a process restart. A process-local timer or inbox item cannot provide that durability, while a global scheduler or private database introduces a second identity, persistence, and lifecycle system.
+
+Busy Agents, long waits, wall-clock changes, cold Sessions, forks, persistence failures, absolute calendar input, and teardown make a simple timeout insufficient. The design must distinguish a durable record from its disposable live wait, keep a fork from inheriting its parent's active reminders, and avoid spreading Schedule-specific presentation or time-zone state across unrelated components.
+
+## Decision
+
+The [`examples/web-schedule`](../../../../examples/web-schedule/README.md) overlay explicitly loads `@deepseek-ai/dsh-time-context` and `@deepseek-ai/dsh-tool-schedule`; the default Web tree remains unchanged. Schedule observes only root Agents published after the plugin loads and installs its three tools plus one disposable owner in that Agent scope. Cold history reads, already-published roots, child Agents, and other hosts do not activate it.
+
+The user-visible boundary is `session-local`: the original Session runs an on-time reminder only while live, does no external notification while cold, and processes an overdue reminder after it becomes live again. Due work waits until the Agent is fully idle, then enters the ordinary next-turn queue through `followup()`; it never steers the current turn and has no independent Web receipt ([conversational delivery](../simplification/2026-08-09-conversational-schedule-delivery.md)).
+
+| Scenario | Durable fact | Live behavior | User-visible result |
+| --- | --- | --- | --- |
+| Create and manage | `schedule/change` create/delete in the original Session | Agent-scoped tools checkpoint before reads and after mutations | Stable id, UTC target, state, and `session-local` disclosure |
+| Due while busy | Active create remains in the fold | Owner waits for idle maintenance, queues one follow-up, then appends dispatch | A later ordinary conversation turn |
+| Several Every records are overdue | Each active record retains its earliest unaccepted anchor-aligned target | One decision selects each record's latest occurrence and advances it past now | One ordinary follow-up containing one occurrence per record |
+| Process stopped or Session cold | Active create remains persisted | No timer or background scan; resume rebuilds the owner | Future target waits; overdue target is attempted |
+| Fork | Parent events remain in the inherited prefix | Child fold starts at `seedLength` | Parent work does not become active in the child |
+
+### Session-log authority and tools
+
+The version-1 `schedule/change` stream is the only durable Schedule authority. A create record owns a Session-local, non-reused branded id, the trimmed prompt, its rule discriminator, and UTC target. Delete and one-shot dispatch are terminal transitions. Every dispatch stores its id and decision time so the fold advances that record directly past missed occurrences. The strict decoder and pure fold reject unknown versions, extra fields, reused ids, mismatched dispatch shapes, and transitions against inactive records. A normal Session folds its complete stream; a fork folds only events at or after `SessionHeader.seedLength`.
+
+The current rule union accepts a non-empty prompt and exactly one selector. `after_seconds` is a positive safe-integer delay whose record is `{ id, kind: 'after', prompt, afterSeconds, scheduledAt }`. `at` is either strict RFC 3339 with `Z` or a numeric offset, or structured `{ date, time, time_zone }` with an explicit zone; its record is `{ id, kind: 'at', prompt, scheduledAt }`. `every_seconds` is a safe integer of at least 300 whose `{ id, kind: 'every', prompt, everySeconds, scheduledAt }` record stays aligned to its creation-plus-interval sequence. One-shot dispatch stores only the id; Every dispatch stores `id + acceptedAt`. Tool values derive `scheduled` or `overdue` and include `deliveryMode: 'session-local'`.
+
+An Agent-scoped FIFO serializes management transactions and the live owner's due transaction from preflight through post-append barriers. Every tool read first awaits `ctx.sessions.flush(session)`. Create rejects input-shape failures before the FIFO when possible, preflights, allocates an id, appends, and checkpoints again. Delete validates its id before the FIFO, preflights before deciding whether it is active, and checkpoints again only after append. List and not-found delete never answer from an unconfirmed live suffix. Failed barriers return `persistence_uncertain` rather than guessing whether an eager write committed.
+
+Every successful management preflight asks the live owner to recompute. A later list can therefore confirm a retained create after a previous post-append rejection and arm it without a private persistence-retry timer.
+
+### Explicit absolute-time boundary
+
+Natural-language interpretation and Schedule parsing are deliberately separate ([time-zone simplification](../simplification/2026-08-09-explicit-schedule-time-zone.md)). Each browser prompt carries its Host-validated IANA zone only on that durable user message. Time-context tells the model to assume that zone for otherwise-unqualified dates and times. Schedule neither imports that plugin nor stores a Session zone: the model must turn its interpretation into an offset-bearing RFC 3339 value or a local object with explicit `time_zone`.
+
+Schedule validates exact calendar shapes, offsets, zone names, and a strictly future four-digit-year instant. A local time inside a daylight-saving gap is rejected; an overlap chooses its first, earlier instant. A successful create stores only canonical UTC `scheduledAt`, not the original offset, local fields, or zone.
+
+### Bounded fixed-rate semantics
+
+Every is a fixed-duration interval, not a calendar rule. The first target is creation time plus the interval. At a due decision, integer division selects the latest sequence point at or before the sampled wall clock and the first sequence point after it. The selected occurrence is presented once and the record advances directly to the future target, so a cold Session never accumulates a replay backlog and delayed model work never shifts the sequence.
+
+All distinct overdue Every records participate in one batch, each with one latest occurrence and one shared `acceptedAt`. There is no cross-record cooldown, gate, quota, or retained batch timestamp. A five-minute minimum bounds wake and model-request frequency. If the next sequence point would exceed the four-digit-year storage range, dispatch terminates that record.
+
+Calendar and Cron expressions are deliberately absent ([bounded recurrence simplification](../simplification/2026-08-09-bounded-fixed-rate-schedule.md)); supporting them would add a time-zone-sensitive calendar language, evaluator dependency, validation surface, and tzdata replay policy unrelated to fixed-rate reminders.
+
+### Live delivery lifecycle
+
+The Agent-scoped owner derives its earliest target from the durable fold. Long targets use bounded timer segments, and every wake reads the wall clock again, so a rollback cannot fire early and a forward jump becomes overdue. Due one-shots have priority and are admitted one at a time; otherwise every overdue Every record enters one batch in target and creation order. If a turn or maintenance task owns the Agent, `runMaintenance()` rejects the claim; the records stay active and one `whenIdle()` wait triggers another attempt. A rejected preflight or contained framing/enqueue failure also leaves them active without starting a private retry timer.
+
+The accepted path clears pending persistence and claims the true idle phase. It refolds the exact Session suffix, samples the decision clock, constructs fixed reminder framing with JSON-escaped values, synchronously queues one `followup()`, and appends dispatch before releasing maintenance. A one-shot appends an id-only terminal dispatch. A fixed-rate batch appends one `id + acceptedAt` transition per participating record. Waking input remains parked until release, so the message cannot be claimed before dispatch enters the log; afterward the owner checkpoints dispatch.
+
+Dispatch records queue admission, not model completion or user receipt. Framing or synchronous enqueue failure appends no dispatch. An append failure faults that owner because the message may already be queued. Agent or plugin disposal cancels timers, stops new work, unwinds tool registrations, and awaits in-flight work without deleting durable records. A crash after follow-up admission but before durable dispatch can repeat the reminder after recovery; the design makes no exactly-once promise.
+
+## Alternatives considered
+
+**Use `ctx.tasks`.** Tasks own process-local work, outcomes, and notifications rather than Session-log state and conversation follow-ups.
+
+**Store reminders in a private database or global scheduler.** This could run cold Sessions but requires a second identity map, startup scan, ownership lease, crash protocol, and notification policy.
+
+**Persist a Session time zone and infer local `at`.** This spreads one interpretive default through Session core, Host create/fork, persistence formats, clients, and mismatch recovery. Request-local model guidance plus an explicit tool boundary deletes that coupling.
+
+**Keep an independent durable Web receipt.** Dispatch is an internal queue fact, not the user's reminder. Rendering the ordinary assistant answer avoids a second delivery meaning and removes Schedule code from Host and client layers.
+
+**Add a general recurring-rule engine.** Fixed-duration intervals need only anchor arithmetic. A shared recurrence abstraction, global admission gate, and calendar evaluator would enlarge replay and runtime state without serving the retained product behavior.
+
+**Claim dispatch before `followup()` or add exactly-once fencing.** Claim-first can silently lose a reminder when enqueue fails. Cross-process exactly-once needs a lease, outbox, acknowledgement, and downstream idempotency boundary outside this Session-local scope.
+
+**Adopt existing roots or register global tools.** Late adoption makes plugin load order activate unseen timers and exposes tools outside the supported root composition.
+
+## Verification
+
+Package tests pin strict replay, one-shot and Every transitions, creation-anchor arithmetic, latest-only catch-up, multi-record batching, fork suffixes, id reuse, offset and local-calendar profiles, IANA validation, daylight-saving gaps and overlaps, time bounds, timer segmentation, wall-clock movement, overdue admission, fixed framing, enqueue and append failures, barrier recovery, registration rollback, and quiescent disposal at per-file 100% coverage. A property test compares Every calculation and replay across varied intervals and skipped spans. A production JSONL restart test proves one overdue reminder dispatches through the real Agent lifecycle and does not redispatch after another restart. Host/client tests pin browser-zone sampling and prompt-bound validation. Keyless assembled Web scenarios cover browser-local At and an overdue two-record Every batch through ordinary assistant follow-ups with no receipt UI.
+
+## Consequences
+
+- Reminder state survives restart through ordinary Session persistence without a new database or public service.
+- Cold Sessions do no work and send no external notification; reopening one may deliver overdue work.
+- Absolute input is deterministic without persistent Session-zone state or a dependency from Schedule to time-context.
+- Users see normal conversation output; dispatch never overstates model success or acknowledgement.
+- Each live root adds only fold-derived timers, an optional idle wait, and one in-flight operation.
+- Fixed-rate recurrence is bounded by a five-minute minimum, latest-only catch-up, and one batched occurrence per overdue record; calendar recurrence remains outside this product boundary.

.agents/notes/implemented/feature/2026-08-05-durable-web-schedule.zh.md

@@ -0,0 +1,86 @@
+# Agent Note: 持久、仅限 Session 内的提醒
+
+Status: implemented
+
+[English](2026-08-05-durable-web-schedule.md) | 中文
+
+## 问题
+
+在对话中创建的提醒必须始终归属于确切的那个 Session，并且跨进程重启存活。进程本地 timer 或 inbox 项无法提供这种持久性，而全局 scheduler 或私有数据库又会引入第二套身份、持久化和生命周期系统。
+
+繁忙的 Agent（智能体）、长等待、墙钟变化、cold Session、fork、持久化失败、绝对日历输入和资源释放，使简单 timeout 无法满足要求。设计必须区分持久记录与可丢弃的 live wait，阻止 fork 继承父 Session 的活动提醒，并避免把 Schedule 专属的呈现或时区状态扩散到无关组件。
+
+## 决策
+
+[`examples/web-schedule`](../../../../examples/web-schedule/README.md) overlay 显式加载 `@deepseek-ai/dsh-time-context` 与 `@deepseek-ai/dsh-tool-schedule`；默认 Web 配置树保持不变。Schedule 只观察插件加载后发布的根 Agent，并在该 Agent scope 中安装三个工具和一个可丢弃 owner。cold history 读取、已发布的根、child Agent 与其他 host 都不会激活它。
+
+用户可见边界是 `session-local`：原 Session 只有在 live 时才会准时运行提醒，cold 期间不发送任何外部通知；该 Session 再次 live 后才会处理 overdue 提醒。到期工作会等待 Agent 完全 idle，再通过 `followup()` 进入普通的下一轮队列；它绝不会中途引导当前轮次，也没有独立 Web 回执（[对话式交付](../simplification/2026-08-09-conversational-schedule-delivery.md)）。
+
+| 场景 | 持久事实 | live 行为 | 用户可见结果 |
+| --- | --- | --- | --- |
+| 创建与管理 | 原 Session 中的 `schedule/change` create／delete | Agent-scoped 工具在读取前、变更后执行 checkpoint | 稳定 id、UTC 目标、状态与 `session-local` 说明 |
+| 到期时繁忙 | 活动 create 仍在 fold 中 | owner 等待 idle maintenance，排入一个 follow-up，再追加 dispatch | 后续一个普通对话轮次 |
+| 多条 Every 记录逾期 | 每条活动记录都保留最早一个尚未接受且与锚点对齐的目标 | 一次决策选择每条记录的最新发生时点，并将其推进到当前时刻之后 | 一个普通 follow-up，其中每条记录各有一个发生时点 |
+| 进程停止或 Session cold | 活动 create 仍在 persistence 中 | 不存在 timer 或后台扫描；resume 重建 owner | 未来目标继续等待；overdue 目标会被尝试 |
+| fork | 父 event 留在继承前缀 | child fold 从 `seedLength` 开始 | 父工作不会在 child 中变为活动状态 |
+
+### Session 日志权威与工具
+
+版本 1 `schedule/change` stream 是唯一持久的 Schedule 权威。create 记录拥有一个 Session 内不复用的品牌 id、trim 后的提示词、规则判别字段和 UTC 目标。delete 与一次性 dispatch 是终结转换。Every dispatch 会存储 id 与决策时点，使 fold 将该记录直接推进到错过的发生时点之后。严格 decoder 与纯 fold 会拒绝未知版本、额外字段、重复使用的 id、形状不匹配的 dispatch，以及针对非活动记录的转换。普通 Session 折叠完整 stream；fork 只折叠 `SessionHeader.seedLength` 位置及其后的 event。
+
+当前规则 union 接受非空提示词和恰好一个 selector。`after_seconds` 是正的安全整数 delay，其记录为 `{ id, kind: 'after', prompt, afterSeconds, scheduledAt }`。`at` 可以是带 `Z` 或数值偏移量且严格符合 RFC 3339 的值，也可以是带显式时区的结构化 `{ date, time, time_zone }`；其记录为 `{ id, kind: 'at', prompt, scheduledAt }`。`every_seconds` 是不小于 300 的安全整数，其 `{ id, kind: 'every', prompt, everySeconds, scheduledAt }` 记录始终与从创建时刻加一个间隔开始的序列对齐。一次性 dispatch 只存储 id；Every dispatch 存储 `id + acceptedAt`。工具值派生 `scheduled` 或 `overdue`，并包含 `deliveryMode: 'session-local'`。
+
+一个 Agent-scoped FIFO 会将管理事务与 live owner 的到期事务从 preflight 到 post-append barrier 全程串行化。每项工具读取都会先等待 `ctx.sessions.flush(session)`。create 会尽可能在进入 FIFO 前拒绝输入形状错误，随后执行 preflight、分配 id、追加记录并再次 checkpoint。delete 会在进入 FIFO 前验证 id，在判断其是否活动前执行 preflight，并且只在追加后再次 checkpoint。list 与 not-found delete 绝不会根据未经确认的 live 后缀作答。barrier 失败会返回 `persistence_uncertain`，而不是猜测 eager write 是否已经提交。
+
+每次成功的管理 preflight 也会要求 live owner 重新计算。因此，如果先前的 post-append 被拒绝，后续 list 可以确认保留的 create 并将其 arm，而无需私有的 persistence 重试 timer。
+
+### 显式绝对时间边界
+
+自然语言解释与 Schedule 解析被有意分开（[时区简化](../simplification/2026-08-09-explicit-schedule-time-zone.md)）。每条浏览器提示词只在其对应的持久 user message 上携带由 Host 校验过的 IANA 时区。Time-context 会告诉模型，把未明确限定时区的日期和时间解释为该时区。Schedule 既不导入该插件，也不存储 Session 时区：模型必须把其解释结果转换为带偏移量的 RFC 3339 值，或带显式 `time_zone` 的本地对象。
+
+Schedule 会校验精确的日历形状、偏移量、时区名称，以及一个严格位于未来、年份为四位数的时点。落在夏令时缺口内的本地时间会被拒绝；遇到重叠时会选择第一次出现的较早时点。创建成功后只存储规范化后的 UTC `scheduledAt`，不会存储原始偏移量、本地字段或时区。
+
+### 有界固定速率语义
+
+Every 是固定时长间隔，而不是日历规则。第一个目标是创建时刻加上一个间隔。作出到期决策时，整数除法会选出不晚于所采样墙钟的最新序列点，以及其后的第一个序列点。选中的发生时点只呈现一次，记录会直接推进到未来目标，因此 cold Session 绝不会积累回放任务，延迟执行的模型工作也绝不会使该序列漂移。
+
+所有不同的逾期 Every 记录都会参与同一个批次，每条记录各自提供一个最新发生时点，并共享同一个 `acceptedAt`。系统不存在跨记录的冷却、门控、配额或保留的批次时间戳。至少 5 分钟的限制约束了唤醒与模型请求频率。如果下一个序列点会超出四位年份存储范围，dispatch 会终结该记录。
+
+日历表达式与 Cron 表达式被有意排除（[有界周期性简化](../simplification/2026-08-09-bounded-fixed-rate-schedule.md)）；支持这些表达式需要增加时区敏感的日历语言、求值器依赖、校验范围和 tzdata 回放策略，而这些都与固定速率提醒无关。
+
+### Live 交付生命周期
+
+Agent-scoped owner 从持久 fold 派生最早目标。超长目标使用有界 timer 分段，每次 wake 都会重新读取墙钟，因此回拨不会提前触发，前跳则会形成 overdue。已到期的一次性提醒优先，每次准入一条；否则，所有逾期 Every 记录会按目标时间和创建顺序进入同一个批次。如果 Agent 已被某个轮次或另一项 maintenance task 占用，`runMaintenance()` 会拒绝此次认领；这些记录保持活动，并由一次 `whenIdle()` wait 触发另一次尝试。被拒绝的 preflight 或被收容的 framing／入队失败同样会使其保持活动，但不会启动私有重试 timer。
+
+获得准入的路径会刷新所有 pending persistence 并认领真正的 idle phase。它会重新折叠确切的 Session 后缀、采样 decision clock、用经过 JSON 转义的值构造固定提醒 framing、同步排入一个 `followup()`，并在释放 maintenance 前追加 dispatch。一次性提醒会追加只含 id 的终结 dispatch。固定速率批次会为每条参与记录追加一个 `id + acceptedAt` 转换。触发唤醒的 input 会保持 parked，直到 maintenance 释放，因此在 dispatch 进入日志前，消息不会被认领；随后 owner 会为 dispatch 执行 checkpoint。
+
+dispatch 记录的是队列准入，而不是模型完成或用户收到提醒。framing 构造或同步入队失败不会追加 dispatch。append 失败会使该 owner fault，因为消息可能已经入队。Agent 或插件 dispose 会取消 timer、停止新工作、撤销工具注册，并等待进行中的工作，且不会删除持久记录。follow-up 获得准入后、持久 dispatch 前发生崩溃，可能使提醒在恢复后重复；本设计不作 exactly-once 承诺。
+
+## 已考虑的替代方案
+
+**使用 `ctx.tasks`。** Task 拥有进程本地工作、结果和通知，而不是 Session 日志状态和对话 follow-up。
+
+**把提醒存入私有数据库或全局 scheduler。** 这样可以运行 cold Session，却需要第二套身份映射、启动扫描、ownership lease、崩溃协议和通知策略。
+
+**持久化 Session 时区并推断本地 `at`。** 这会让一个解释默认值扩散到 Session core、Host create／fork、持久化格式、client 和不匹配恢复中。请求本地的模型指导与显式工具边界消除了这种耦合。
+
+**保留独立的持久 Web 回执。** dispatch 是内部队列事实，而不是用户的提醒。渲染普通 assistant 回答既避免了第二种交付含义，也从 Host 与 client 层移除了 Schedule 代码。
+
+**增加通用周期规则引擎。** 固定时长间隔只需要锚点运算。共享的周期抽象、全局准入门控和日历求值器会扩大回放与运行时状态，却不能服务于保留的产品行为。
+
+**在 `followup()` 前认领 dispatch，或增加 exactly-once fencing。** claim-first 会在入队失败时静默丢失提醒。跨进程 exactly-once 需要 lease、outbox、acknowledgement 与下游幂等边界，超出了此 Session-local 范围。
+
+**接管既有根或注册全局工具。** 晚接管会让插件加载顺序激活不可见的 timer，并把工具暴露到受支持的根组合之外。
+
+## 验证
+
+包测试以逐文件 100% coverage 固定严格回放、一次性与 Every 状态转换、创建锚点运算、只追赶最新一次、多记录批处理、fork 后缀、id 复用、偏移量与本地日历 profile、IANA 校验、夏令时缺口与重叠、时间边界、timer 分段、墙钟变化、overdue 准入、固定 framing、入队与 append 失败、barrier 恢复、注册 rollback 和完全停稳的 dispose。属性测试会在不同间隔与跳过跨度下比较 Every 计算与回放。production JSONL restart 测试证明一条 overdue 提醒会经过真实 Agent 生命周期 dispatch，并且再次 restart 后不会重复 dispatch。Host／client 测试固定浏览器时区采样与绑定到提示词的校验。无密钥组装 Web 场景覆盖浏览器本地 At，以及通过普通 assistant follow-up 交付的逾期双记录 Every 批次，两者都没有回执 UI。
+
+## 后果
+
+- 提醒状态通过普通 Session persistence 跨重启存活，无需新数据库或公开 service。
+- cold Session 不工作、不发送外部通知；重新打开后可能交付 overdue 工作。
+- 无需持久 Session 时区状态或从 Schedule 到 time-context 的依赖，绝对时间输入仍然具有确定性。
+- 用户看到普通对话输出；dispatch 绝不会夸大模型成功或 acknowledgement。
+- 每个 live 根只增加从 fold 派生的 timer、可选 idle wait 与一个 in-flight operation。
+- 固定速率周期性受到至少 5 分钟、只追赶最新一次，以及每条逾期记录只在一个批次中贡献一个发生时点的约束；日历周期性仍在此产品边界之外。

.agents/notes/implemented/feature/2026-08-06-manager-owned-subagent-settlement-delivery.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-06-manager-owned-subagent-settlement-delivery.md
-2026-08-06-manager-owned-subagent-settlement-delivery.md: e06fe2b4ca7dd9a975f70524979de09d13cdcf79
-2026-08-06-manager-owned-subagent-settlement-delivery.zh.md: 77e8b5e4bae9b05e08bf2c4a4bf997b688cd1157
+2026-08-06-manager-owned-subagent-settlement-delivery.md: 11669d4603e658d1478557f81c2d2b4d5495bca7
+2026-08-06-manager-owned-subagent-settlement-delivery.zh.md: f3684777ccb3800226803e998438147ebc1dbf16

.agents/notes/implemented/feature/2026-08-06-manager-owned-subagent-settlement-delivery.md

@@ -16,7 +16,7 @@ The signal already existed. `subagent/end` has carried `stopReason` and `lastAss
 
 The continuation manager delivers the account itself, from inside the disposal transaction that ends the Activation.
 
-When a resident Activation settles, `notifySettlement()` resolves the child's durable direct parent and sends it one user-role message: the epoch's outcome as a sentence the parent can act on, then the child's final assistant content, or a statement that it produced none. Delivery is unconditional for every child whose id a caller actually received. It does not consult whether the child reported, and it keeps no bookkeeping that could make the promise conditional — that unconditionality is what lets `tool-subagent` tell the model "you are told when it finishes, so never poll or wait on it" and have that be true. A materialization rolled back before its first accepted message stays silent, because the caller was told that child was not established.
+When a resident Activation settles, `notifySettlement()` resolves the child's durable direct parent and sends it one user-role message: the epoch's outcome as a sentence the parent can act on, then the child's final assistant content, or a statement that it produced none. Delivery is unconditional for every child whose id a caller actually received. It does not consult whether the child reported, and it keeps no bookkeeping that could make the promise conditional — that unconditionality is what lets `tool-subagent` promise a runtime notice containing the outcome and any final assistant message. A materialization rolled back before its first accepted message stays silent, because the caller was told that child was not established.
 
 ### Provenance
 
@@ -44,7 +44,7 @@ Delivery never blocks or fails teardown. A rejected send is logged and dropped,
 
 ### The epoch's own log is the whole account
 
-`epochStopReason()` reads the epoch's outcome from its own log, because teardown succeeding says nothing about whether the model errored, hit its ceiling, or was stopped. Reading turns alone got that wrong twice, in the same shape both times: a turn stopped before its first step leaves a `turn/end` indistinguishable from the balanced no-op turns a rejection or an emptied claim produces, so the filter that skipped those also skipped real endings and answered with the previous turn's clean completion. The durability checkpoint (`dsh-session-checkpoint-policy`, in every shipped profile) and prompt assembly both run at that boundary and both propagate, and `Inbox.claim()` has already taken the messages by then — so the parent was told a child finished while the delivery it was waiting on had been swallowed. Under a promise that says "you are told when it finishes, so never poll", that is the one failure a parent cannot detect and will not retry.
+`epochStopReason()` reads the epoch's outcome from its own log, because teardown succeeding says nothing about whether the model errored, hit its ceiling, or was stopped. Reading turns alone got that wrong twice, in the same shape both times: a turn stopped before its first step leaves a `turn/end` indistinguishable from the balanced no-op turns a rejection or an emptied claim produces, so the filter that skipped those also skipped real endings and answered with the previous turn's clean completion. The durability checkpoint (`dsh-session-checkpoint-policy`, in every shipped profile) and prompt assembly both run at that boundary and both propagate, and `Inbox.claim()` has already taken the messages by then — so the parent was told a child finished while the delivery it was waiting on had been swallowed. Under the advertised automatic settlement notice, that is the one failure a parent cannot detect and will not retry.
 
 The missing fact was never the turn's; it was the inbox's. `Inbox` logs every mutation with `removedCount` and marks a cancellation `outcome: 'canceled'`, which separates a turn claiming its input from work being dropped unrun. `foldConsumedWork()` in `dsh-agent` folds both vocabularies into one answer: the latest turn that accounts for consumed work — stepped, or claimed-then-failed, stopped, or rejected — and whether accepted work was cancelled after it with no turn opening over it. A `blocked` end over claimed input is an account too: the pre-step rejection that produced it — a hook deny, a policy plugin — discarded the messages the turn claimed, so the notice says the child declined rather than finished. Only a `blocked` turn that claimed nothing stays invisible.
 
@@ -60,7 +60,7 @@ Three assembled ACP scenarios cover the notice: a child that never reports, a ch
 
 `subagent-continuable` is the one that pins a failure. Its child's last turn dies on the forced durability checkpoint without entering a step, so that transcript is where the stop-reason rule above is visible end to end: the notice says the child *failed*, carries the earlier `SECOND_OK` as its last content rather than as a result, and the parent's own acknowledgement turn reaches the ACP client.
 
-A keyless headless Loader snapshot covers the user-visible path end to end. Its replay parent starts one continuable child with `run_in_background: true`, never calls `list_agents`, `send_message`, or Task tools, consumes the manager-authored `subagent-settled` notice, and produces its final answer. The child never calls `report`, so the transcript cannot pass through the cooperative report path. A test-only Loader fence holds the parent's post-spawn request until the real manager notice enters its inbox, removing platform scheduling from the transcript without synthesizing the notice.
+A keyless headless Loader snapshot covers the user-visible path end to end. Its replay parent omits `run_in_background` to exercise the continuable background default, never calls `list_agents`, `send_message`, or Task tools, consumes the manager-authored `subagent-settled` notice, and produces its final answer. The child never calls `report`, so the transcript cannot pass through the cooperative report path. A test-only Loader fence holds the parent's post-spawn request until the real manager notice enters its inbox, removing platform scheduling from the transcript without synthesizing the notice.
 
 `subagent-report` needed one more concession. With the shipped waking report default, that scenario has two independent parent wakes — the report and the settlement — and whether the second extends the first's turn or opens its own is a genuine coin flip that measured 50/50 across runs. No authored transcript can hold both orders. Its overlay therefore pins `reportDelivery: quiet`, leaving settlement as the only wake, and a snapshot-only pre-step fence holds the child until the parent's spawn turn ends so that wake opens one deterministic turn claiming both messages. The waking report default keeps its coverage in the report package's own tests.
 

.agents/notes/implemented/feature/2026-08-06-manager-owned-subagent-settlement-delivery.zh.md

@@ -16,7 +16,7 @@ Status: implemented
 
 继续执行管理器自己投递这份记账，就在结束 Activation 的那笔 dispose 事务内部完成。
 
-当驻留 Activation 结算时，`notifySettlement()` 解析该 child 持久化的直接父级，并向它发送一条用户角色消息：先是父级可据以行动的一句结果说明，然后是 child 的最终 assistant 内容，或一句说明它没有产出内容。对每个调用方真正拿到过 id 的 child，投递都是无条件的。它不查询 child 是否上报过，也不保留任何可能让这项承诺变成有条件的记账——正是这种无条件性，才让 `tool-subagent` 能够告诉模型「它结束时你会被告知，因此绝不要轮询或等待它」并且这句话为真。在第一条消息被接受之前就回滚的物化保持静默，因为调用方已被告知该 child 未建立。
+当驻留 Activation 结算时，`notifySettlement()` 解析该 child 持久化的直接父级，并向它发送一条用户角色消息：先是父级可据以行动的一句结果说明，然后是 child 的最终 assistant 内容，或一句说明它没有产出内容。对每个调用方真正拿到过 id 的 child，投递都是无条件的。它不查询 child 是否上报过，也不保留任何可能让这项承诺变成有条件的记账——正是这种无条件性，才让 `tool-subagent` 能够承诺一条包含结局与可能存在的最终 assistant 消息的运行时通知。在第一条消息被接受之前就回滚的物化保持静默，因为调用方已被告知该 child 未建立。
 
 ### 来源信息
 
@@ -44,7 +44,7 @@ Status: implemented
 
 ### epoch 自己的日志就是全部交代
 
-`epochStopReason()` 从 epoch 自己的日志读取结局，因为拆卸成功与否，对「模型是否报错、是否撞到上限、是否被停下」什么也没说明。只读轮次这件事已经错了两次，而两次的形状相同：在第一个 step 之前被停下的轮次，其 `turn/end` 与「拒绝」或「被清空的认领」产生的平衡空转轮次长得一模一样，于是那道用来跳过后者的过滤，也把真实的结局一起跳过了，转而用上一个轮次的干净收尾作答。持久化检查点（`dsh-session-checkpoint-policy`，存在于每个随附 profile 中）与提示词组装都运行在这个边界上、且都会向外传播，而此时 `Inbox.claim()` 已经把消息取走了——于是父级被告知 child 已完成，而它正在等待的那条投递已被吞掉。在「你会在它完成时被告知，所以永远不要轮询」这一承诺之下，这恰恰是父级无法察觉、也不会重试的那一种失败。
+`epochStopReason()` 从 epoch 自己的日志读取结局，因为拆卸成功与否，对「模型是否报错、是否撞到上限、是否被停下」什么也没说明。只读轮次这件事已经错了两次，而两次的形状相同：在第一个 step 之前被停下的轮次，其 `turn/end` 与「拒绝」或「被清空的认领」产生的平衡空转轮次长得一模一样，于是那道用来跳过后者的过滤，也把真实的结局一起跳过了，转而用上一个轮次的干净收尾作答。持久化检查点（`dsh-session-checkpoint-policy`，存在于每个随附 profile 中）与提示词组装都运行在这个边界上、且都会向外传播，而此时 `Inbox.claim()` 已经把消息取走了——于是父级被告知 child 已完成，而它正在等待的那条投递已被吞掉。在已公布的自动结算通知约定下，这恰恰是父级无法察觉、也不会重试的那一种失败。
 
 缺失的事实从来不属于轮次，而属于 inbox。`Inbox` 会把每次改动连同 `removedCount` 一起记入日志，并给取消标记 `outcome: 'canceled'`，这就把「某个轮次认领了它的输入」与「工作被丢弃且从未运行」区分开来。`dsh-agent` 中的 `foldConsumedWork()` 把两套词汇折叠成一个答案：能为已消费工作作出交代的最新轮次——进入过 step 的，或认领后失败、被停下或被拒绝的——以及此后是否有已接受的工作被取消、而没有任何轮次为它开启过。认领过输入、以 `blocked` 结束的轮次同样是一份交代：产生它的 pre-step 拒绝——hook deny、策略插件——把该轮次认领的消息一并丢弃了，因此通知会说 child 拒绝了任务，而不是完成了任务。只有没认领任何输入的 `blocked` 轮次保持不可见。
 
@@ -60,7 +60,7 @@ Status: implemented
 
 `subagent-continuable` 是其中固定失败结局的那个。它的 child 最后一个轮次在被强制的持久化检查点上死亡，且未进入任何 step，因此该 transcript 正是上面那条终止原因规则的端到端可见之处：通知说该 child **失败**，把此前的 `SECOND_OK` 作为它最后产出的内容而非结果携带，而父级自己的确认轮次会到达 ACP 客户端。
 
-另有一个无密钥的 headless Loader 快照端到端覆盖用户可见路径。其重放父级通过 `run_in_background: true` 启动一个可继续 child，从不调用 `list_agents`、`send_message` 或 Task 工具，消费管理器写入的 `subagent-settled` 通知，并给出最终答案。child 从不调用 `report`，因此该 transcript 不可能经由协作式上报路径通过。一个仅用于测试的 Loader 栅栏会把父级启动后的请求保持到真实管理器通知进入其 inbox 为止，从 transcript 中排除平台调度差异，但不会伪造该通知。
+另有一个无密钥的 headless Loader 快照端到端覆盖用户可见路径。其重放父级省略 `run_in_background` 以覆盖可继续后台默认路径，从不调用 `list_agents`、`send_message` 或 Task 工具，消费管理器写入的 `subagent-settled` 通知，并给出最终答案。child 从不调用 `report`，因此该 transcript 不可能经由协作式上报路径通过。一个仅用于测试的 Loader 栅栏会把父级启动后的请求保持到真实管理器通知进入其 inbox 为止，从 transcript 中排除平台调度差异，但不会伪造该通知。
 
 `subagent-report` 还需要多做一步让步。在随附的唤醒上报默认值下，该场景有两个互相独立的父级唤醒——上报与结算——而第二个究竟是延长第一个的轮次还是另开一个轮次，是一枚真正的硬币，多次运行实测约为五五开。任何手写 transcript 都无法同时容纳两种顺序。因此它的 overlay 固定 `reportDelivery: quiet`，使结算成为唯一唤醒；另一个仅用于快照的 pre-step 栅栏会把 child 保持到父级启动轮次结束，使这次唤醒开启一个确定轮次并同时认领两条消息。唤醒上报默认值的覆盖则保留在 report 包自身的测试中。
 

.agents/notes/implemented/feature/2026-08-10-web-plugin-configuration.i18n.yaml

@@ -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/2026-08-10-web-plugin-configuration.md
+2026-08-10-web-plugin-configuration.md: 7375f496c7af1a695243444fe56aca7262d3dedd
+2026-08-10-web-plugin-configuration.zh.md: 59d65db39bcc2306983f2a26dcf252164d7a6f37

.agents/notes/implemented/feature/2026-08-10-web-plugin-configuration.md

@@ -0,0 +1,50 @@
+# Agent Note: Plugin configuration in the web settings page
+
+Status: implemented
+
+English | [中文](2026-08-10-web-plugin-configuration.zh.md)
+
+## Problem
+
+Everything a plugin can be configured with lived in `cordis.yml`. A user who wanted a longer shell timeout, a different search endpoint, or fewer parallel tool calls had to find the composition file, know its shape, and restart — while the Models page had shown for months that a settings namespace can be edited from the browser and take effect immediately.
+
+The seam that made the Models page possible was already general: any plugin may register a namespace, and `settings.describe` serves its schema, its layers, and its revision. What was missing was on the two ends. No plugin outside the LLM adapters and the permission service had registered one, and there was no surface for a namespace that is not a model provider.
+
+## Decision
+
+Three host-plane plugins register their own settings namespace, and one browser-side section renders whatever the deployment exposes.
+
+**Layering, unchanged.** A section resolves as schema defaults → the plugin's composition entry → the user layer. Each plugin passes its `cordis.yml` entry as the `base` and reads its config through a source thunk, so a stored change reaches the next use and a detaching settings provider leaves the composition entry running. Constraints the schema cannot express — positive and finite, the timer bound on `graceMs`, the parallel cap being a positive integer — become the section validator, so a bad value is refused at the write instead of at the next command.
+
+**The shell namespace names the capability, not an implementation.** `BASH_SETTINGS_NAMESPACE` is exported by `@deepseek-ai/dsh-bash` because a host composes exactly one provider of `ctx.bash`: the win32 layer swaps the POSIX rows for the pwsh ones, and mounting both fails loud on a duplicate service registration. Both families therefore register the same namespace with their own schema and entry without ever colliding, and a `settings.yaml` carried between platforms keeps resolving on both — schemastery objects preserve keys the active schema does not declare.
+
+**A section is a subset when the plugin config is bigger than what a user owns.** `agent-loop` exposes only `maxParallelToolCalls`; its `agents` array is consumed once when the service starts, so a stored change there could only look like it had an effect.
+
+**The provider projects, rather than captures.** `web-search-deepseek` hands its provider a thunk instead of an options value, so an endpoint or model change reaches the next search without re-registering the provider — which would make the web seam's provider selection observable to the user as a flicker.
+
+**Exposure stays a Host allowlist.** The three namespaces join `WEB_SETTINGS_NAMESPACES`; registration alone still never crosses the transport, and a namespace absent from that list answers `settings-not-exposed` exactly as an unregistered one does.
+
+**The section knows no namespace.** `dsh-client-ui-plugin-config` declares a `settings.plugin.item` slot and renders the cards registered into it, so a plugin that ships a browser half owns its card and its controls. Each card binds its namespace through the client settings scope, which gained the two things a form needs: the raw `user` layer, whose key PRESENCE is what marks a field overridden, and `unset`, which clears one field back to the composition layer. A card renders nothing while its namespace is unavailable, so a deployment that does not compose the owning plugin shows no trace of it.
+
+**A card stages its edits and writes them on save.** Controls hold no draft of their own: the card's form owns the staged text, every control renders it, and only **Save** turns it into document mutations. A settings write is durable and revision-fenced, so a control that committed as it settled spent a revision on a value the user had not decided to store and could not preview; the reset stages the composed default the same way. Because the Host's validators own the constraints no schema can express, the form reads the section back after writing and reports a save that did not land instead of predicting the outcome, keeping those drafts for the user to correct. The credential control is staged with the rest even though it writes through the credentials domain, so one save covers everything the card shows.
+
+## Alternatives considered
+
+- **A registration-time exposure declaration replacing the allowlist.** The honest shape — the namespace's owner declares its own exposure, and a plugin distributed outside this repository can surface its configuration without a change in `packages/host/apiproxy`. Deferred because it changes the seam contract, every existing registration site, and the anti-enumeration semantics at once, and because a plugin exposing an arbitrary schema needs a fail-closed redaction path first: a secret reachable only through a union or transform is currently returned verbatim.
+- **A generic schema-driven form renderer.** Declined again for the reason recorded in the [web-config-plane note](../architecture/2026-07-30-web-config-plane.md): field truth without a presentation vocabulary produced an unusable card. Three plugins of hand-written controls cost about the same and read better, and the slot keeps the fourth plugin from having to negotiate with this package.
+- **Editing preset-mounted plugins from this page.** Out of scope, and not merely unbuilt: a preset's rows carry their configuration inline in `agent.cordis.yml` and cannot register a settings namespace at all, because a second session mounting the same preset would fail on a duplicate registration. A user layer shared across presets would also overwrite the fields a preset uses to define its agent's identity — its persona text, its delegation wiring — which are per-preset by design.
+- **One namespace per executor package instead of the capability-named `bash`.** Declined because the composed executor differs by platform while the settings document does not: a user who set a timeout on macOS would silently lose it on Windows.
+- **Writing the search key into the settings section.** Declined because the literal would then have to ride a `describe` response to be rendered. The card reports only whether a key is configured and writes through the credentials domain, addressed by the reference the section names.
+- **Committing each control as it settles, with no save.** Built first, and replaced: blur is not a decision. It spent a namespace revision per control, gave the user nothing to preview or undo before the write, and left an invalid draft silently discarded — a value the Host's validator refuses simply snapped back with no reason given. One save per card makes the write a gesture the user performs.
+- **Letting the provider read its options per property.** The thunk was read at each use site so read sites could stay unchanged, which quietly broke the contract the constructor states: `search()` awaits credential resolution and then reads the endpoint, model, and budget, so a settings write landing inside that await sent the key resolved from the old section to the endpoint named by the new one. Each operation now snapshots once at its entry and threads that snapshot into credential resolution.
+- **Validating the fields in the browser to keep the save honest.** Declined: the constraints live in the owning plugin's section validator, and restating them here would make two homes for one rule that could disagree per release. The card checks only what its own control can decide — that a numeric draft is a number — and lets the Host answer for the rest, which is why the save reads the section back.
+
+## Consequences
+
+A user edits the shell's command timeout and output cap, the agent loop's parallel tool-call cap, and the search provider's key, endpoint, and per-request budget from the settings page, with each field marking whether they set it and offering a reset.
+
+Two costs are real. Adding a fourth plugin still requires an entry in the apiproxy allowlist, so the page's reach is a Host decision rather than a plugin's. And the plugins the web deployment moved into the agent plane — the file tools, the skills, compaction, the todo tool — appear nowhere here, which is most of what a user might expect to find; their configuration remains the preset editor's.
+
+The bash and pwsh executors now expose `config` as a getter over a source thunk rather than a readonly field. Every read site was already per-call, so nothing else changed, but a subclass that captured `this.config` at construction would silently pin the composition entry.
+
+`verify-cordis-config` gained one check, paid for by this branch: merging master's rename of the client manifest field (`dshClient` → `dsh.client`) left this package declaring the old name, and the whole section vanished from the browser with no error anywhere — the row composed, the empty node half activated, and the browser roster scan simply never matched it. Nothing could catch that, because the composition file cannot tell a surface plugin from a Host plugin: the difference lives in the manifest. The gate now requires a `packages/client` package's `./client` export and its `dsh.client` declaration to agree in both directions. The check is scoped to that group because a Host package's `./client` export is the typed wire face its browser consumers import, not a plugin the roster serves.

.agents/notes/implemented/feature/2026-08-10-web-plugin-configuration.zh.md

@@ -0,0 +1,50 @@
+# Agent Note: Plugin configuration in the web settings page
+
+Status: implemented
+
+[English](2026-08-10-web-plugin-configuration.md) | 中文
+
+## 问题
+
+插件的一切可配置项都只存在于 `cordis.yml`。想要更长的 shell 超时、不同的搜索端点或更少的并行工具调用，用户必须找到组装文件、了解它的形状，然后重启——而 Models 页几个月来一直在证明：settings 命名空间可以在浏览器里编辑并立即生效。
+
+支撑 Models 页的那条 seam 本就是通用的：任何插件都可以注册命名空间，`settings.describe` 会提供它的 schema、分层与 revision。缺的是两端。除 LLM 适配器与权限服务外，没有插件注册过命名空间；而对于非模型提供方的命名空间，也没有任何表层。
+
+## 决策
+
+三个宿主平面插件各自注册 settings 命名空间，一个浏览器侧分区渲染该部署所暴露的一切。
+
+**分层不变。** 一个分节按 schema 默认值 → 插件的组装条目 → 用户层解析。每个插件把自己的 `cordis.yml` 条目作为 `base` 传入，并通过 source thunk 读取配置，因此存储的变更会作用于下一次使用，而脱离的 settings 提供方会让组装条目继续运行。schema 无法表达的约束——正有限、`graceMs` 的定时器上界、并行上限必须是正整数——成为分节的校验器，因此错误的值在写入时被拒绝，而不是到下一条命令时才失败。
+
+**shell 命名空间命名的是能力，而非某个实现。** `BASH_SETTINGS_NAMESPACE` 由 `@deepseek-ai/dsh-bash` 导出，因为一个宿主只组装一个 `ctx.bash` 提供方：win32 层会把 POSIX 行换成 pwsh 行，而同时挂载两者会因服务重复注册在加载期失败。因此两个家族都能用自己的 schema 与条目注册同一个命名空间而永不相撞；在平台间携带的 `settings.yaml` 也能在两边继续解析——schemastery 对象会保留当前 schema 未声明的键。
+
+**当插件配置大于用户所拥有的部分时，分节就是一个子集。** `agent-loop` 只暴露 `maxParallelToolCalls`；它的 `agents` 数组在服务启动时被消费一次，所以存储在那里的变更只会看起来生效。
+
+**提供方按次投影，而不是固化。** `web-search-deepseek` 交给提供方的是一个 thunk 而非 options 值，因此端点或模型的变更无需重新注册提供方即可作用于下一次搜索——重新注册会让 web seam 的提供方选择以闪断的形式被用户看到。
+
+**暴露仍是 Host 的白名单。** 这三个命名空间加入 `WEB_SETTINGS_NAMESPACES`；仅有注册依然不会跨越传输边界，而不在该名单中的命名空间会与未注册的命名空间得到完全相同的 `settings-not-exposed`。
+
+**该分区不认识任何命名空间。** `dsh-client-ui-plugin-config` 声明 `settings.plugin.item` slot 并渲染注册进来的卡片，因此带浏览器半侧的插件拥有自己的卡片与控件。每张卡片通过客户端 settings scope 绑定其命名空间，而该 scope 补上了表单所需的两样东西：原始 `user` 层——键的**存在**才标记字段被覆盖——以及把单个字段清回组装层的 `unset`。命名空间不可用时卡片什么都不渲染，因此未组装该插件的部署不会显示它的任何痕迹。
+
+**卡片暂存修改，保存时才写入。** 控件不持有自己的草稿：暂存文本归卡片的表单所有，所有控件渲染的都是它，只有**保存**才把它变成文档变更。settings 写入是持久且带 revision 栅栏的，因此「失焦即提交」的控件会为用户尚未决定存储、也无从预览的值花掉一个 revision；重置同样只是暂存组装默认值。schema 表达不了的约束归 Host 的校验器所有，所以表单在写入后回读分节、报告没有落盘的保存，而不是自行预测结果，并保留这些草稿供用户修改。密钥控件虽然经由 credentials 领域写入，也和其余字段一起暂存，因此一次保存覆盖卡片上的全部内容。
+
+## 备选方案
+
+- **用注册期的暴露声明取代白名单。** 这才是诚实的形状——命名空间的拥有方声明自己的暴露，在本仓库之外分发的插件也无需改动 `packages/host/apiproxy` 就能呈现自己的配置。之所以暂缓，是因为它会同时改变 seam 契约、全部现有注册点与防枚举语义；而且插件要暴露任意 schema，还得先有 fail-closed 的脱敏路径：目前只能经由 union 或 transform 抵达的 secret 会被原样返回。
+- **通用 schema 驱动的表单渲染器。** 再次否决，理由与 [web-config-plane 笔记](../architecture/2026-07-30-web-config-plane.md)所记一致：没有呈现词汇的字段真值产出的是无法使用的卡片。三个插件的手写控件成本相当而可读性更好，且该 slot 让第四个插件无需与本包协商。
+- **在本页编辑 preset 挂载的插件。** 超出范围，而且不只是「尚未实现」：preset 的行把配置内联在 `agent.cordis.yml` 中，且根本无法注册 settings 命名空间——同一 preset 挂载第二个会话时会因重复注册而失败。跨 preset 共享的用户层还会覆盖 preset 用来定义其 agent 身份的字段——人设文本、委派接线——而这些字段按设计就是各 preset 各自的。
+- **按执行器包各取一个命名空间，而非按能力命名的 `bash`。** 否决，因为被组装的执行器随平台不同，而设置文档不随平台不同：在 macOS 上设过超时的用户，到 Windows 上会悄无声息地失去它。
+- **把搜索密钥写进 settings 分节。** 否决，因为那样字面值就必须搭乘 `describe` 响应才能被渲染。卡片只报告是否已配置密钥，并按分节所命名的引用经由 credentials 领域写入。
+- **每个控件失焦即提交，不设保存。** 最初就是这么做的，后被替换：失焦不是决定。它每个控件花掉一个命名空间 revision，写入前不给用户任何预览或撤销的余地，还会把无效草稿悄悄丢弃——被 Host 校验器拒绝的值只是弹回原样，不给任何理由。每张卡片一个保存，才让写入成为用户执行的动作。
+- **让提供方按属性逐次读取 options。** 最初为了不改动读取点而在每个使用处读 thunk，这悄悄违背了构造函数自己声明的契约：`search()` 先 await 凭据解析，之后才读端点、模型与预算，因此落在那段 await 里的设置写入会把按旧分节解析出的密钥发往新分节命名的端点。现在每次操作在入口只快照一次，并把该快照传进凭据解析。
+- **在浏览器端校验字段，好让保存诚实。** 否决：这些约束住在拥有方插件的分节校验器里，在这里重述一遍就会让同一条规则有两个家，且可能随版本各说各话。卡片只判断自己的控件能判断的事——数字草稿是不是数字——其余交给 Host 回答，这正是保存要回读分节的原因。
+
+## 影响
+
+用户可以在设置页编辑 shell 的命令超时与输出上限、agent 循环的并行工具调用上限，以及搜索提供方的密钥、端点与单次请求预算，每个字段都标注是否由自己设定，并提供重置。
+
+有两项真实代价。加入第四个插件仍需要在 apiproxy 白名单里添一条，因此本页的覆盖面是 Host 的决定而非插件的决定。而 web 部署移入 agent 平面的那些插件——文件工具、技能、压缩、todo 工具——在这里一个都不出现，而它们恰恰是用户最可能期待找到的；它们的配置仍归 preset 编辑器。
+
+bash 与 pwsh 执行器现在把 `config` 暴露为 source thunk 之上的 getter，而不再是 readonly 字段。所有读取点本就是按次读取，因此别无变化；但若某个子类在构造期捕获 `this.config`，就会悄然把组装条目钉死。
+
+`verify-cordis-config` 新增一项检查，代价由本分支付过：合并 master 对客户端清单字段的重命名（`dshClient` → `dsh.client`）后，本包仍声明旧名，于是整个分区从浏览器上消失，且任何地方都不报错——行照常组装、空的 node 半侧照常激活，只是浏览器 roster 扫描永远匹配不到它。这一点无从被既有门禁发现，因为组装文件区分不了 surface 插件与 Host 插件：差别在清单里。现在门禁要求 `packages/client` 包的 `./client` 导出与 `dsh.client` 声明双向一致。之所以只限这一组：Host 包的 `./client` 导出是给浏览器消费方 import 的类型化 wire face，不是 roster 要服务的插件。

.agents/notes/implemented/feature/2026-08-11-background-first-continuable-delegation.i18n.yaml

@@ -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/2026-08-11-background-first-continuable-delegation.md
+2026-08-11-background-first-continuable-delegation.md: e90b0a97cd5eeb273c8084d231ced003a661b978
+2026-08-11-background-first-continuable-delegation.zh.md: b8f14e0c40c341fc91b2fd0dc8c3f00fc2e14b66

.agents/notes/implemented/feature/2026-08-11-background-first-continuable-delegation.md

@@ -0,0 +1,45 @@
+# Agent Note: Continuable delegation is background-first
+
+Status: implemented
+
+English | [中文](2026-08-11-background-first-continuable-delegation.zh.md)
+
+## Problem
+
+A continuable child already has a durable id, independent turns, follow-up messaging, and a manager-owned settlement notice. Treating an omitted `run_in_background` as foreground makes that lifecycle depend on the model restating `true` on every call. It also obscures the useful scheduling test: the parent should wait only when its next action requires the child's result.
+
+The child-scoped `report` prompt requires a self-contained final report, while [manager-owned settlement delivery](2026-08-06-manager-owned-subagent-settlement-delivery.md) independently sends the run outcome and closing message. A completed child can therefore wake its parent with a final report and again with settlement. Background-first scheduling must preserve both deliveries: the child-authored handoff remains mandatory guidance, while the manager-authored notice covers every terminal path regardless of model compliance.
+
+## Decision
+
+`tool-subagent` resolves an omitted `run_in_background` from the selected lifecycle policy. `backgroundMode: continuable` resolves omission to background and returns the durable child id immediately; explicit `false` selects foreground and waits for the result. `backgroundMode: one-shot` keeps its foreground default because background output still requires Task collection. `enableRunInBackground: false` continues to omit the parameter, reject forced `true`, and run in the foreground. No second default-selection config is added.
+
+The model-facing text divides responsibility by location:
+
+- the tool description states the call behavior, durable id, runtime settlement notice, follow-up through `send_message`, and the explicit foreground override;
+- the `run_in_background` parameter states the lifecycle-specific default and when to override it;
+- a `tool:<toolName>` system-prompt section tells the model to start independent delegations together, continue useful work while they run, and choose foreground only when the next action depends on the result. The section renders only when that tool remains visible in the assembly scope, so a child tool restriction removes the schema and its guidance together.
+
+The [continuable child report obligation](2026-08-06-continuable-child-report-obligation.md) remains unchanged: the child prompt requires one self-contained final report and earlier reports for findings that change the parent's next action. Manager-owned settlement remains unconditional and does not inspect whether a report arrived. The two messages may repeat final content, but they retain distinct authors and purposes: `report` is the child's explicit handoff, while settlement records how the run ended and preserves terminal output when the child cannot cooperate. `reportDelivery` remains deployment scheduling policy with `wakeup` as its default.
+
+The keyless headless `subagent-settlement` scenario omits `run_in_background`, receives the immediate child id, and reaches the final parent answer through the manager-authored settlement notice even though its fixture deliberately does not call `report`. Package tests separately pin explicit `false` as foreground, the parent scheduling text, and the child's mandatory-report prompt.
+
+## Alternatives considered
+
+**Replace the field with `run_in_foreground`.** Reversing the boolean makes the common case read positively, but creates a second vocabulary for the same scheduling choice and forces every existing caller and provider-facing transcript to change. Keeping `run_in_background` preserves one field and makes foreground the explicit exception.
+
+**Add a configurable background default.** A separate default can disagree with `backgroundMode`, the schema wording, and the installed prompt. The lifecycle policy already distinguishes a continuable activation from a one-shot Task, which is the distinction that determines whether background completion is delivered automatically.
+
+**Change only the prompt.** Prompt preference without runtime resolution still turns an omitted argument into foreground. The model must be able to rely on the advertised default rather than reproduce it perfectly on every tool call.
+
+**Suppress settlement after a final report arrives.** Conditional settlement reintroduces per-Activation bookkeeping and loses the unconditional runtime guarantee when a child reports progress and then fails. Settlement remains unconditional even when the resulting message overlaps a final report.
+
+**Use `report` only for progress before settlement.** This removes duplicate final content but also removes the explicit child-authored handoff from the child prompt. The final-report obligation remains, and runtime settlement remains its independent fallback and terminal record.
+
+## Consequences
+
+- An ordinary continuable call is non-blocking without spelling `run_in_background: true`; serialized delegation is an explicit `false` choice.
+- Independent subagent calls in one assistant message overlap under the tool loop's concurrency-safe dispatch, while dependent foreground calls can still be issued one at a time.
+- Parent guidance, tool schema, runtime resolution, and settlement delivery state the same default.
+- A compliant child reports one self-contained final result and may report important findings earlier. Every Activation also produces an unconditional settlement notice, so a completed run may deliver overlapping final content twice.
+- One-shot background Tasks and disabled-background tool instances retain their existing behavior.

.agents/notes/implemented/feature/2026-08-11-background-first-continuable-delegation.zh.md

@@ -0,0 +1,45 @@
+# Agent Note: 可继续委派采用后台优先
+
+Status: implemented
+
+[English](2026-08-11-background-first-continuable-delegation.md) | 中文
+
+## 问题
+
+可继续 child 已经具备持久化 id、独立轮次、后续消息以及由管理器负责的结算通知。如果把省略的 `run_in_background` 视为前台，模型就必须在每次调用时重复写出 `true`，才能得到这套生命周期。这样也会掩盖真正有用的调度判断：只有当 parent 的下一步动作需要 child 结果时，parent 才应等待。
+
+child 作用域的 `report` 提示词要求发送自包含的最终报告，而[由管理器负责的结算投递](2026-08-06-manager-owned-subagent-settlement-delivery.md)会独立发送本次运行的结束结果与收尾消息。已完成的 child 因而可能先用最终报告唤醒 parent，再用结算通知唤醒一次。后台优先调度会保留两次投递：由 child 编写的交接仍是强制提示词指引，由管理器生成的通知则不依赖模型是否遵循指令，覆盖每种终止路径。
+
+## 决策
+
+`tool-subagent` 根据选定的生命周期策略解析省略的 `run_in_background`。`backgroundMode: continuable` 会把省略解析为后台并立即返回持久化 child id；显式传入 `false` 会选择前台并等待结果。`backgroundMode: one-shot` 保留前台默认行为，因为它的后台输出仍需通过 Task 收集。`enableRunInBackground: false` 仍会省略该参数、拒绝强制传入的 `true` 并在前台运行。系统不增加第二个默认选择配置。
+
+面向模型的文本按位置划分职责：
+
+- 工具描述说明调用行为、持久化 id、运行时结算通知、通过 `send_message` 继续对话，以及显式前台覆盖；
+- `run_in_background` 参数说明具体生命周期的默认值以及何时覆盖；
+- `tool:<toolName>` 系统提示词 section 会告诉模型同时启动相互独立的委派、在它们运行时继续有用工作，并且仅当下一步动作依赖结果时选择前台。只有当该工具在组装作用域中仍可见时才会渲染这个 section，因此子级工具限制会同时移除 schema 与对应指引。
+
+[可继续 child 上报义务](2026-08-06-continuable-child-report-obligation.md)保持不变：child 提示词要求发送一份自包含的最终报告，并在发现会改变 parent 下一步动作的信息时提前报告。由管理器负责的结算仍然无条件执行，不检查报告是否已经到达。这两条消息可能重复最终内容，但作者和用途不同：`report` 是 child 的显式交接，结算则记录本次运行如何结束，并在 child 无法配合时保留终止输出。`reportDelivery` 仍是部署调度策略，默认值仍为 `wakeup`。
+
+无密钥 headless `subagent-settlement` 场景省略 `run_in_background`，收到立即返回的 child id；尽管 fixture（测试前置数据）有意不调用 `report`，它仍通过管理器生成的结算通知到达 parent 最终答案。包测试另行固定了显式 `false` 的前台语义、parent 调度文本以及 child 的强制报告提示词。
+
+## 考虑过的替代方案
+
+**把字段替换为 `run_in_foreground`。** 反转布尔值会让常见情形以肯定形式表达，却会为同一项调度选择创造第二套词汇，并迫使所有现有调用方与面向提供方的 transcript（文本记录）一起改变。保留 `run_in_background` 可以维持单一字段，并把前台作为显式例外。
+
+**增加可配置的后台默认值。** 独立默认值可能与 `backgroundMode`、schema 措辞和已安装提示词不一致。生命周期策略已经区分可继续 Activation 与一次性 Task，而这个区别正好决定了后台完成是否会自动投递。
+
+**只修改提示词。** 如果运行时解析不变，提示词偏好仍会让省略参数的调用进入前台。模型必须能够依赖公布的默认值，而不是在每次工具调用中完美复述它。
+
+**最终报告到达后抑制结算通知。** 条件结算会重新引入每次 Activation 的记账，并且当 child 先报告进度、随后失败时丢掉无条件运行时保证。即使生成的消息与最终报告重叠，结算仍然无条件执行。
+
+**只用 `report` 发送结算前的进度。** 这样可以消除重复的最终内容，但也会从 child 提示词中移除由 child 编写的显式交接。最终报告义务保持不变，运行时结算则继续作为它的独立后备和终止记录。
+
+## 后果
+
+- 普通可继续调用无需写出 `run_in_background: true` 即为非阻塞；串行委派需要显式选择 `false`。
+- 同一条 assistant 消息中的独立 subagent 调用会在工具循环的并发安全分发下重叠执行；有依赖的前台调用仍可逐个发出。
+- parent 指引、工具 schema、运行时解析和结算投递陈述同一个默认值。
+- 遵循指令的 child 会发送一份自包含的最终结果，也可以更早报告重要发现。每次 Activation 还会产生无条件结算通知，因此已完成的运行可能两次投递相互重叠的最终内容。
+- 一次性后台 Task 与禁用后台的工具实例保留现有行为。

.agents/notes/implemented/feature/2026-08-11-background-task-completion-wakes-an-idle-owner.i18n.yaml

@@ -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/2026-08-11-background-task-completion-wakes-an-idle-owner.md
+2026-08-11-background-task-completion-wakes-an-idle-owner.md: cbebf1fbe82d47db4ba6e39b318d2e8ed8d89e17
+2026-08-11-background-task-completion-wakes-an-idle-owner.zh.md: b5f3cdd594b04370dcfe99d83775f34ab8b8115c

.agents/notes/implemented/feature/2026-08-11-background-task-completion-wakes-an-idle-owner.md

@@ -0,0 +1,76 @@
+# Agent Note: Background task completion wakes an idle owner
+
+Status: implemented
+
+English | [中文](2026-08-11-background-task-completion-wakes-an-idle-owner.zh.md)
+
+## Problem
+
+`tool-tasks` promised the model "You are notified in-session when a task finishes — do not busy-poll or sleep on one." The promise held only while the model was still working. Completion delivered through `agent.inject()`, which appends to the next-step inbox without reserving a driver, so a task settling after its turn closed left the notice parked until something unrelated woke the agent. The common shape is exactly the one that breaks: the model starts a long command, tells the user it started it, ends its turn, and the command finishes into an inbox nobody will claim. The prompt told the model not to poll, and then nothing arrived.
+
+The gap was recorded as a limitation rather than reasoned about, so the fallback was `task_output(wait: true)` — the blocking wait the same prompt discourages.
+
+This supersedes one fact of the [background-task runtime decision](../architecture/2026-06-20-generic-long-running-tool-runtime.md) — that completion never wakes an idle owner — and adds teardown as a `reported` setter. That note keeps every other task-runtime decision and is updated in place rather than replaced.
+
+The delivery machinery was never the obstacle. `Agent.send(message, target, wakeup)` has covered the `target` × `wakeup` matrix since the [unified send decision](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md), and `wakeDriver()` already handles idle, maintenance, and cancelled-converging phases. The missing piece was the policy choice of which lane a completion takes, plus the bound that choice needs.
+
+## Decision
+
+An unreported completion picks its lane from what the owner is doing. A busy owner is injected, unchanged. An idle owner is woken with `followup()`.
+
+This adopts the delivery rule the [continuation manager](2026-08-06-manager-owned-subagent-settlement-delivery.md) already ships for subagent settlement, where "steering rather than injecting is deliberate … This is a correctness rule, not a deployment preference." The two paths do not overlap: `tool-subagent` registers a Task only for a one-shot background child and returns `continuable` before reaching that code, so a child is delivered by exactly one of the two mechanisms.
+
+### The busy owner keeps injection
+
+For a driver that is genuinely running, `steer()` and `inject()` are the same delivery: `wakeDriver()` returns early without latching for a running, unaborted phase. They differ only for an owner whose turn is cancelled but has not yet converged, where steering redirects to the next turn and replays the wake at convergence.
+
+Injection is correct there. A cancelled turn is a user pressing stop, and reopening one on their behalf launders an interrupt into a model request they did not ask for. The turn loop already covers the ordinary case: it cannot close while the next-step inbox holds anything, so a notice arriving before that check extends the current turn, and several tasks settling together cost one step rather than one turn each.
+
+### Waking is bounded, and the bound is not time
+
+`maxConsecutiveWakes` (default 3) caps the turns one owner may open this way; beyond it a notice degrades to injection and waits for the next turn. Claiming any user-authored message restores the budget — claiming, not arrival, because that is the point human input actually enters a step. Notices this plugin queued never refill it.
+
+The bound exists because this chain is self-exciting in a way subagent settlement is not. Settlement is bounded by how many children the model spawned; a woken turn can start the background task whose completion wakes it again, with nobody watching. `dsh run` needs no separate policy: its one user message is claimed in the first turn and never repeats, so the budget is spent monotonically and the process terminates.
+
+`completionDelivery: quiet` restores the old lane for idle owners. It exists for deterministic transcripts, and mirrors the `reportDelivery` switch on `tool-subagent-report` in name, values, and default.
+
+### Teardown claims the report
+
+`cancelForTeardown` now marks the record `reported`, exactly as `kill()` does after cancelling. The asymmetry was invisible while the notice was a harmless inject; a waking reporter turns it into one model request per teardown layer, on agents the host is destroying.
+
+`reported` was already the right bit — "a kill, read, or wait has reported or committed to report the terminal state" — and teardown is a kill without a caller. Using it keeps every observer of the settlement intact: `onTaskDone` still fires, so runtime invariants and the force-fail path stay covered, and only notice reporters go quiet.
+
+### Completion is announced last
+
+`settle()` released waiters, marked the record settled, and published the visible-set change *after* running completion listeners. A reporter that opens a turn does so synchronously, so that order let a woken turn's `turn/start` land before the settlement it was reacting to was committed, and before any `onTasksChanged` observer had seen it. Announcing completion last makes the reporter the final observer of a settlement every other observer has already seen.
+
+## Alternatives considered
+
+**A producer-declared wake bit on `TaskStart`,** matching Codex's `trigger_turn` and Kimi's `admission` enum. It is the better long-run shape — a `tail -f` stream and a two-hour build want different answers — but no current producer distinguishes them, and the repository requires a current owner and need for public surface. The natural trigger to add it is the first producer that wants one task to wake and another not to.
+
+**A general unsolicited-input queue** with priority lanes, as Claude Code uses to merge background tasks, cron, MCP push, and hooks into one drain. DSH's inbox already is that queue — durable `agent/inbox/spliced` splices over `next-turn`/`next-step` — so this would add a layer above an existing one to decide a single bit.
+
+**Refusing to reopen a turn that already produced a visible answer,** Codex's `MailboxDeliveryPhase` latch. That latch is the default this decision deliberately inverts: waking after the model has spoken is the entire point, and the wake budget is the bound instead.
+
+**A wall-clock window** on top of the counter. For an interactive agent the slow case is the wanted one — an hour-long build finishing and the agent resuming is the feature — and `dsh run` is already bounded by the counter it cannot refill. Worth revisiting only if an unattended long-lived deployment appears.
+
+**Suppressing `onTaskDone` entirely during owner drain,** symmetric with the service-wide `listenersClosed`. It reads cleaner and removes a signal that is not only for notices: the force-fail record and the runtime invariant both observe teardown settlements. The `reported` bit denies exactly the reporters and nothing else.
+
+## Consequences
+
+- Default behavior changes: an idle owner now spends a model request per completion, capped at `maxConsecutiveWakes` per owner between user messages. Deployments that want the old behavior set `completionDelivery: quiet`.
+- The `tool-tasks` prompt section needs no edit; "You are notified in-session when a task finishes" became true rather than aspirational.
+- `TaskSnapshot.reported` gains teardown as a fourth setter, documented at the Service Definition and in [the subsystem reference](../../../../docs/subsystems/tasks.md).
+- `settle()` announces completion after committing the record and publishing the visible-set change. Any listener relying on running before waiters were released or before `onTasksChanged` now runs after both.
+- The `tool-bash` real-composition test dropped its second user message: settlement alone carries the notice into a turn that collects the output. It asserts the durable outcome rather than a turn boundary, because whether the command outlives its turn is a race; the lane choice is pinned in `tool-tasks` unit tests instead.
+- Unit coverage pins idle wake, busy injection, quiet delivery, budget exhaustion, budget restore on user input, non-restore on plugin notices, and teardown silence.
+
+### Accepted risks
+
+A spent budget is restored only by user input. An unattended agent that exhausts it collects its remaining notices whenever something else opens a turn, and nothing re-arms it in the meantime.
+
+A notice pending on an idle owner under `quiet` still dies with that owner's disposal, unchanged from before: the disposal cancel clears the unclaimed inbox and the log keeps the insert/cancel pair as the record. The [settlement delivery note](2026-08-06-manager-owned-subagent-settlement-delivery.md) owns the offline-mailbox discussion this would need.
+
+Whether a completion extends the running turn or opens a new one is a genuine race for short-lived tasks, so no authored transcript can hold both orders. Assembled coverage asserts the outcome; the lane choice is pinned in unit tests.
+
+One microtask window survives: a settlement landing after the turn loop's last inbox check but before the driver commits its idle phase still reads `status === 'running'`, so it injects and nothing wakes. Steering would not close it either — `wakeDriver()` latches only for maintenance and post-cancel phases, not for a driver between its final check and its own retirement. Closing it needs an `agent-loop` boundary that publishes retirement before the last claim, which is a core-agent decision rather than a delivery-policy one.

.agents/notes/implemented/feature/2026-08-11-background-task-completion-wakes-an-idle-owner.zh.md

@@ -0,0 +1,76 @@
+# Agent Note: Background task completion wakes an idle owner
+
+Status: implemented
+
+[English](2026-08-11-background-task-completion-wakes-an-idle-owner.md) | 中文
+
+## 问题
+
+`tool-tasks` 对模型承诺「任务完成时你会在会话内收到通知——不要忙轮询，也不要 sleep 等待」。这个承诺只在模型仍在工作时成立。完成经由 `agent.inject()` 交付，它只向 next-step inbox 追加而不预留 driver，因此在轮次结束之后才结算的任务会把通知搁在那里，直到某件无关的事情唤醒 agent。最常见的形态恰恰就是会失效的那一种：模型启动一条长命令，告诉用户已经启动，结束轮次，而命令完成后进入了一个无人领取的 inbox。提示词让模型不要轮询，然后什么也没到。
+
+这个缺口被记为一条限制，而不是被推敲过，于是退路成了 `task_output(wait: true)`——同一段提示词并不鼓励的阻塞等待。
+
+本决策取代[后台任务运行时决策](../architecture/2026-06-20-generic-long-running-tool-runtime.md)中的一条事实——完成永不唤醒空闲所有者——并把 teardown 加为 `reported` 的置位方。那份 note 仍拥有其余全部任务运行时决策，因此就地更新而非替换。
+
+交付机制从来不是障碍。自[统一 send 决策](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md)起，`Agent.send(message, target, wakeup)` 就覆盖了 `target` × `wakeup` 矩阵，`wakeDriver()` 也已经处理 idle、maintenance 和已取消未收敛三种相位。缺的是「一次完成走哪条通道」这一策略选择，以及该选择所需的界。
+
+## 决策
+
+尚未报告的完成按所有者当时在做什么来选择通道。繁忙的所有者走注入，保持原样。空闲的所有者用 `followup()` 唤醒。
+
+这采纳了[延续管理器](2026-08-06-manager-owned-subagent-settlement-delivery.md)已经为 subagent 结算所采用的交付规则，那里写着「用 steer 而非 inject 是刻意的……这是一条正确性规则，不是部署偏好」。两条路径不重叠：`tool-subagent` 只为一次性后台子 agent 注册 Task，而 continuable 分支在抵达那段代码之前就已返回，因此一个子 agent 恰好由两种机制中的一种交付。
+
+### 繁忙的所有者保留注入
+
+对真正在运行的 driver 而言，`steer()` 与 `inject()` 是同一次交付：对于运行中且未中止的相位，`wakeDriver()` 会提前返回且不设置 latch。二者只在一种所有者上有区别——轮次已取消但尚未收敛，此时 steer 会重定向到下一轮并在收敛时重放唤醒。
+
+在那里注入才是对的。轮次被取消意味着用户按了停止，替他们重新开一轮等于把一次中断洗成了他们没有要求的模型请求。普通情形已由轮次循环覆盖：只要 next-step inbox 还有内容，轮次就无法结束，因此在该检查之前抵达的通知会延长当前轮次，同时结算的多个任务只花掉一步而不是各占一轮。
+
+### 唤醒有界，且该界不是时间
+
+`maxConsecutiveWakes`（默认 3）限制一个所有者由此开启的轮数；超出后通知降级为注入，等待下一轮。领取任何用户撰写的消息都会恢复预算——是领取而非抵达，因为那才是人类输入真正进入某一步的时刻。本插件自己排队的通知永远不会补充它。
+
+设界是因为这条链会自激，而 subagent 结算不会。结算受限于模型派生了多少子 agent；被唤醒的一轮却可能启动某个后台任务，而它的完成又会唤醒同一个所有者，且无人旁观。`dsh run` 不需要单独策略：它唯一的用户消息在第一轮就被领取且不会重复，因此预算单调消耗，进程必然终止。
+
+`completionDelivery: quiet` 为空闲所有者恢复旧通道。它的存在是为了确定性 transcript，并在名称、取值与默认值上都对齐 `tool-subagent-report` 的 `reportDelivery` 开关。
+
+### 销毁自行认领报告
+
+`cancelForTeardown` 现在会把记录标记为 `reported`，与 `kill()` 在取消之后所做的完全一致。当通知只是一次无害的注入时，这处不对称看不出来；而会唤醒的报告方会把它变成每个 teardown 层级一次模型请求，作用在宿主正要销毁的 agent 上。
+
+`reported` 本来就是正确的那个 bit——「kill、read 或 wait 已报告或承诺报告终止状态」——而 teardown 是一次没有调用方的 kill。用它可以让该结算的每一个观察者都保持完整：`onTaskDone` 仍会触发，因此运行时不变量与强制失败路径依旧被覆盖，只有通知报告方会安静下来。
+
+### 完成是最后才宣布的
+
+`settle()` 此前释放等待方、标记记录已结算并发布可见集变更的时机，都排在运行完成监听器**之后**。开启轮次的报告方是同步执行的，因此那个顺序会让被唤醒轮次的 `turn/start` 抢在它所响应的那次结算被提交之前落地，也抢在任何 `onTasksChanged` 观察者看到它之前。把完成放到最后宣布，使报告方成为该结算的最后一个观察者，而其他观察者都已先看到它。
+
+## 被否决的替代方案
+
+**在 `TaskStart` 上加生产方声明的唤醒位**，对应 Codex 的 `trigger_turn` 与 Kimi 的 `admission` 枚举。从长期看这是更好的形状——`tail -f` 流与两小时构建想要不同答案——但当前没有任何生产方需要区分它们，而仓库要求公共面必须有当下的所有者与需求。加它的自然触发点，是第一个「要让某个任务唤醒而另一个不唤醒」的生产方出现时。
+
+**一个通用的非请求输入队列**并带优先级通道，正如 Claude Code 用来把后台任务、cron、MCP 推送与 hook 合并进同一次排空。DSH 的 inbox 本身就是那个队列——`next-turn`/`next-step` 之上的持久 `agent/inbox/spliced` splice——因此这等于在既有层之上再加一层，只为决定一个 bit。
+
+**拒绝重开一个已经产出可见答复的轮次**，即 Codex 的 `MailboxDeliveryPhase` 闩锁。那条闩锁正是本决策刻意反转的默认值：在模型已经说完话之后唤醒它就是本特性的全部意义，界由唤醒预算来承担。
+
+**在计数之上再加墙钟窗口**。对交互式 agent 而言，慢的那种情形恰恰是想要的——一小时的构建结束、agent 接着干下去，这就是特性本身——而 `dsh run` 已被它无法补充的计数封顶。只有当出现无人值守的长生命周期部署时才值得重新考虑。
+
+**在 owner 排空期间整体压制 `onTaskDone`**，与服务级的 `listenersClosed` 对称。它读起来更干净，但会移走一个不只服务于通知的信号：强制失败记录与运行时不变量都会观察 teardown 结算。`reported` 位恰好只否决报告方，别的什么也不否决。
+
+## 影响
+
+- 默认行为改变：空闲所有者现在每次完成会花掉一次模型请求，按所有者、在两次用户消息之间由 `maxConsecutiveWakes` 封顶。想要旧行为的部署设置 `completionDelivery: quiet`。
+- `tool-tasks` 的提示词段落无需改动；「任务完成时你会在会话内收到通知」从愿景变成了事实。
+- `TaskSnapshot.reported` 新增 teardown 作为第四个置位方，记录在 Service Definition 与[子系统参考](../../../../docs/subsystems/tasks.md)中。
+- `settle()` 在提交记录并发布可见集变更之后才宣布完成。任何依赖「在释放等待方之前或在 `onTasksChanged` 之前运行」的监听器现在都排在两者之后。
+- `tool-bash` 的 real-composition 测试去掉了第二条用户消息：仅靠结算就能把通知带入一个收集输出的轮次。它断言持久结果而非轮次边界，因为命令是否活得比它的轮次久是一场竞态；通道选择改由 `tool-tasks` 单元测试钉住。
+- 单元覆盖钉住：空闲唤醒、繁忙注入、quiet 交付、预算耗尽、用户输入恢复预算、插件通知不恢复预算，以及 teardown 静默。
+
+### 已接受的风险
+
+已花掉的预算只由用户输入恢复。耗尽预算的无人值守 agent 要等到其他原因开启轮次时才收走剩余通知，在此期间没有任何机制为它重新充能。
+
+在 `quiet` 下待领于空闲所有者的通知仍会随该所有者释放而消亡，与此前一致：释放时的取消会清空未领取的 inbox，日志保留插入/取消这一对作为记录。[结算交付 note](2026-08-06-manager-owned-subagent-settlement-delivery.md) 承载这需要的离线信箱讨论。
+
+对短命任务而言，完成究竟是延长运行中的轮次还是开启新轮次是一场真实竞态，因此没有哪份编写的 transcript 能同时容纳两种顺序。组装态覆盖断言结果；通道选择由单元测试钉住。
+
+还残留一个微任务窗口：结算若落在轮次循环最后一次检查 inbox 之后、driver 提交 idle 相位之前，读到的仍是 `status === 'running'`，于是走注入且无人唤醒。改用 steer 也堵不上——`wakeDriver()` 只为 maintenance 与取消后的相位设置 latch，不为「最后一次检查与自身退休之间」的 driver 设置。要堵上它需要 `agent-loop` 在最后一次领取之前就发布退休状态，那属于核心 agent 的决策，而非交付策略。

.agents/notes/implemented/feature/2026-08-11-deepseek-request-user-id-header.i18n.yaml

@@ -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/2026-08-11-deepseek-request-user-id-header.md
+2026-08-11-deepseek-request-user-id-header.md: 246021d3ea69a6341def90f81051da78a267f051
+2026-08-11-deepseek-request-user-id-header.zh.md: 6861d2b6ed1294c8e29efb7c883f56f8aa1d31ca

.agents/notes/implemented/feature/2026-08-11-deepseek-request-user-id-header.md

@@ -0,0 +1,44 @@
+# Agent Note: DeepSeek request user and session identity headers
+
+Status: implemented
+
+English | [中文](2026-08-11-deepseek-request-user-id-header.zh.md)
+
+## Problem
+
+Direct DeepSeek requests already carried `x-deepseek-harness-session-id` when the caller supplied `GenerateOptions.sessionId`, which lets provider-side support and diagnostics correlate turns within one conversation. They lacked a stable identity across sessions even though the harness already persists an anonymous user id for telemetry and feedback. A separate id would break correlation, while putting it in the provider-neutral attribution helper would send a stable per-user identifier through every HTTP adapter.
+
+The user id is transport metadata, not model input. It must not enter the request body, prompt, token accounting, KV-cache identity, or session log. The destination is the adapter's resolved `baseURL`, which can be DeepSeek itself or a configured gateway, so the privacy boundary must be explicit.
+
+## Decision
+
+`dsh-llm-deepseek` sends `x-deepseek-harness-user-id` on every provider request sent after successful credential resolution. The value comes from `@deepseek-ai/dsh-user-id` and therefore matches the OpenTelemetry Resource `user.id` and `/feedback` acknowledgement for the same `$DSH_HOME`. The adapter continues to send `x-deepseek-harness-session-id` only when `GenerateOptions.sessionId` is present; the agent loop supplies the current durable `Session.id` for ordinary agent, title-generation, and compaction requests.
+
+The plugin resolves the user id lazily after credentials succeed and memoizes it for that plugin instance. A missing credential therefore does not create `.userid`, while the first authorized provider request can create it even when `DSH_TELEMETRY_DISABLED` is set. The direct adapter constructor accepts a `resolveUserId` dependency so wire behavior remains deterministic in unit tests.
+
+Both headers are model-hidden HTTP metadata sent to the resolved `baseURL`. They are absent from the JSON request body and do not become model-visible inputs or session events. A configured gateway receives them. Telemetry sharing controls only telemetry export and does not disable provider request identity.
+
+## Verification
+
+- The mock provider asserts that an authorized request carries the same user id returned by `getOrCreateAnonymousUserId()` and omits the session header when no session id is supplied.
+- The session-identity wire test asserts both headers and preserves the exact supplied session id.
+- A direct-adapter test asserts that user-id resolution happens once per stream, while the keyless configuration test proves a credential failure does not create `.userid`.
+- The real Loader composition test asserts that the assembled plugin uses the shared user-id package rather than a test-only value.
+- No keyless snapshot changes because the headers are not model-visible or user-visible transcript content.
+
+## Alternatives considered
+
+| Rejected | Reason |
+|---|---|
+| Add the id to generic `attributionHeaders()` | That helper is provider-neutral and static; a per-user value there would reach unrelated providers and violate its app-identity privacy contract |
+| Configure a fixed custom header in `cordis.yml` | Deployment configuration cannot derive the current session id and would expose a stable identity as mutable config instead of using its owning runtime contract |
+| Mint a DeepSeek-specific user id | Provider requests could not correlate with telemetry and feedback for the same harness home |
+| Disable the header with telemetry sharing | Provider request identity and telemetry export have different recipients and purposes; one switch would hide the actual privacy boundary |
+| Put the id in OpenAI-compatible `user` or `metadata` request fields | Body fields can affect provider schema, logging, caching, tokenization, or model-visible reconstruction; HTTP metadata preserves the intended boundary |
+
+## Consequences
+
+- DeepSeek support can correlate requests across sessions by one anonymous harness-home id and within a conversation by the durable session id.
+- The first authorized DeepSeek request may create `$DSH_HOME/.userid` independently of telemetry export.
+- Custom DeepSeek gateways receive the stable user id and any available session id, so operators must treat the configured `baseURL` as an identity recipient.
+- The request body, prompt, token count, KV-cache identity, and session log remain unchanged.

.agents/notes/implemented/feature/2026-08-11-deepseek-request-user-id-header.zh.md

@@ -0,0 +1,44 @@
+# Agent Note: DeepSeek 请求用户与会话身份头部
+
+Status: implemented
+
+[English](2026-08-11-deepseek-request-user-id-header.md) | 中文
+
+## 问题
+
+当调用方提供 `GenerateOptions.sessionId` 时，直连 DeepSeek 请求已携带 `x-deepseek-harness-session-id`，让提供方侧支持与诊断可以关联同一对话中的多个轮次。但请求缺少跨会话的稳定身份，而 harness 已为遥测与反馈持久化匿名用户 id。另行生成 id 会破坏关联；把它放进提供方无关的归属辅助函数，则会让每个 HTTP 适配器都发送稳定的逐用户标识。
+
+用户 id 是传输元数据，不是模型输入。它不得进入请求体、提示词、token 计量、KV cache 身份或会话日志。发送目标是适配器解析后的 `baseURL`，既可能是 DeepSeek 自身，也可能是配置的网关，因此必须明确隐私边界。
+
+## 决策
+
+`dsh-llm-deepseek` 在凭据解析成功后发出的每个提供方请求上发送 `x-deepseek-harness-user-id`。该值来自 `@deepseek-ai/dsh-user-id`，因此与同一 `$DSH_HOME` 的 OpenTelemetry Resource `user.id` 及 `/feedback` 确认一致。适配器继续仅在存在 `GenerateOptions.sessionId` 时发送 `x-deepseek-harness-session-id`；普通 agent、标题生成与压缩请求由 agent loop 提供当前持久化 `Session.id`。
+
+插件在凭据解析成功后惰性获取用户 id，并在该插件实例内缓存。缺少凭据不会创建 `.userid`；即使设置了 `DSH_TELEMETRY_DISABLED`，首个已授权的提供方请求仍可能创建它。直连适配器构造函数接收 `resolveUserId` 依赖，使线路行为可在单元测试中保持确定性。
+
+两个头部都是发送到解析后 `baseURL` 的模型不可见 HTTP 元数据。它们不在 JSON 请求体中，也不会成为模型可见输入或会话事件。配置的网关会收到它们。遥测共享只控制遥测导出，不会禁用提供方请求身份。
+
+## 验证
+
+- mock 提供方断言已授权请求携带 `getOrCreateAnonymousUserId()` 返回的同一用户 id，并在未提供会话 id 时省略会话头部。
+- 会话身份线路测试断言两个头部都存在，并原样保留传入的会话 id。
+- 直连适配器测试断言每条 stream 仅解析一次用户 id，keyless 配置测试则证明凭据失败不会创建 `.userid`。
+- 真实 Loader 组合测试断言组装后的插件使用共享 user-id 包，而非测试专用值。
+- 无需修改 keyless snapshot，因为这些头部不是模型可见或用户可见的 transcript 内容。
+
+## 考虑过的替代方案
+
+| 已否决 | 原因 |
+|---|---|
+| 把 id 加进通用 `attributionHeaders()` | 该辅助函数是提供方无关且静态的；加入逐用户值会把它发送给无关提供方，并违反其应用身份隐私契约 |
+| 在 `cordis.yml` 中配置固定自定义头部 | 部署配置无法推导当前会话 id，且会把稳定身份暴露为可变配置，而不是使用其所属运行时契约 |
+| 生成 DeepSeek 专用用户 id | 提供方请求将无法与同一 harness home 的遥测和反馈关联 |
+| 随遥测共享关闭该头部 | 提供方请求身份与遥测导出的接收方和目的不同；共用开关会掩盖真实隐私边界 |
+| 把 id 放进 OpenAI 兼容的 `user` 或 `metadata` 请求字段 | body 字段可能影响提供方 schema、日志、缓存、token 化或模型可见重建；HTTP 元数据可保留预期边界 |
+
+## 后果
+
+- DeepSeek 支持可以通过一个匿名 harness-home id 跨会话关联请求，并通过持久化 session id 关联同一对话。
+- 首个已授权 DeepSeek 请求可独立于遥测导出创建 `$DSH_HOME/.userid`。
+- 自定义 DeepSeek 网关会收到稳定用户 id 与可用的会话 id，因此运维方必须将配置的 `baseURL` 视为身份接收方。
+- 请求体、提示词、token 数、KV cache 身份和会话日志保持不变。

.agents/notes/implemented/feature/2026-08-11-minimal-profiles-bare-two-tool-runtime.i18n.yaml

@@ -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/2026-08-11-minimal-profiles-bare-two-tool-runtime.md
+2026-08-11-minimal-profiles-bare-two-tool-runtime.md: cf48c71214420443411bf91253b5e3a1c72a2c6c
+2026-08-11-minimal-profiles-bare-two-tool-runtime.zh.md: 5cfe819e1cd18f9c9a3c8d63ab9bf8e34b888a24

.agents/notes/implemented/feature/2026-08-11-minimal-profiles-bare-two-tool-runtime.md

@@ -0,0 +1,39 @@
+# Agent Note: Minimal profiles use the bare two-tool runtime
+
+Status: implemented
+
+English | [中文](2026-08-11-minimal-profiles-bare-two-tool-runtime.zh.md)
+
+## Problem
+
+The Web `minimal` preset and standalone JSON-RPC minimal composition exposed persistent `bash` and `str_replace_editor`, but their supporting services did not match the intended training runtime. Both mounted context compaction, while the Web preset inherited the host's sandboxed filesystem and the JSON-RPC composition mounted `fs-sandbox` plus filesystem policy. A long session could therefore replace history, and the editor advertised and enforced a filesystem policy that the bare local reference runtime does not have.
+
+The two launch paths also have different configuration owners. Web mounts a per-agent preset over a running host, while the Python SDK initializes a complete stdio JSON-RPC child process. Treating them as one interchangeable Cordis leaf would hide those lifecycle differences, and the SDK example had no environment path for selecting its model or system prompt.
+
+## Decision
+
+Both shipped minimal profiles expose exactly persistent `bash` and `str_replace_editor`, mount no context-compaction provider, and run the editor against `@deepseek-ai/dsh-fs-local`. The Web preset isolates `ctx.fs` inside the agent entry and mounts `fs-local` beside the editor, so other Web agents retain the host filesystem provider. Its persona remains the fixed complete prompt owned by the earlier [minimal-preset composition decision](../bug-fix/2026-08-10-minimal-preset-owns-rl-composition.md).
+
+The standalone [`minimal.cordis.yml`](../../../../examples/jsonrpc-agent/minimal.cordis.yml) remains a complete JSON-RPC process composition. It mounts `dsh-jsonrpc`, the local PTY and subprocess services required by persistent Bash, `fs-local`, the two tool consumers, and uncompressed JSONL persistence. It does not mount `token-meter`, `compact-basic`, `fs-sandbox`, or `fs-policy`. Persistent Bash still consumes the deployment's danger-full-access sandbox policy; the editor is not confined by that policy.
+
+`DSH_SYSTEM_PROMPT` selects the standalone persona. `DSH_MODEL` names the DeepSeek provider catalog entry, and `DSH_CONTEXT_WINDOW` supplies that entry's capacity. Because the SDK client owns the JSON-RPC `initialize` request, [`minimal.py`](../../../../examples/jsonrpc-agent/minimal.py) also uses `DSH_MODEL` as its default `model` argument; an explicit `--model` remains authoritative. Endpoint and credential variables stay owned by the DeepSeek adapter's existing environment-resolution path.
+
+## Verification
+
+The Web replay boots the complete Web host, creates the agent through the preset service, and asserts that the scoped filesystem is bare, no scoped compaction service exists, and the assembled request contains exactly the fixed prompt and two tools. It then executes persistent Bash and the editor against the real scoped services.
+
+The SDK replay boots the real JSON-RPC agent process through the SDK client, injects an environment-selected prompt, asserts the assembled prompt and exact two-tool catalog, and executes both tools. Python SDK bundled-runtime coverage initializes the standalone configuration through each available packaged carrier with environment-selected model, model capacity, and prompt values. Cordis validation checks that both configurations resolve their declared plugins and configuration fields.
+
+## Alternatives considered
+
+**Keep `compact-basic` mounted with a high threshold.** Rejected because even an inert-for-short-tests provider permits history replacement in longer sessions and leaves the minimal composition dependent on model-capacity metadata and the token meter.
+
+**Keep `fs-sandbox` in danger-full-access mode.** Rejected because the sandboxed provider still makes confinement and escalation part of the editor capability. The target runtime requires the bare local provider, whose lack of `sandboxMode` is composition truth.
+
+**Use one Cordis leaf for Web and Python SDK startup.** Rejected because a Web preset contributes agent-scoped services to an existing multi-session host, while the Python SDK must launch a complete process containing the JSON-RPC server and its process-wide dependencies.
+
+**Read `DSH_MODEL` only inside Cordis.** Rejected because Cordis configures the provider catalog but does not own the SDK client's JSON-RPC `initialize` request. The launcher must pass the same model to the client request for the environment value to select the routed model.
+
+## Consequences
+
+Minimal sessions never summarize or replace earlier history; callers must keep turns within the selected model's context capacity. The editor can address any absolute path visible to the runtime process, independently of the persistent shell's sandbox policy. The two launch paths share their model-facing tool and no-compaction guarantees while retaining different prompt and model configuration appropriate to their owners. The Python SDK path continues to communicate only through the bundled stdio JSON-RPC runtime.

.agents/notes/implemented/feature/2026-08-11-minimal-profiles-bare-two-tool-runtime.zh.md

@@ -0,0 +1,39 @@
+# Agent Note: minimal profile 使用裸双工具运行时
+
+Status: implemented
+
+[English](2026-08-11-minimal-profiles-bare-two-tool-runtime.md) | 中文
+
+## 问题
+
+Web `minimal` preset 与独立 JSON-RPC minimal 组合对外提供持久 `bash` 和 `str_replace_editor`，但支撑服务与目标训练运行时不一致。两者都挂载上下文压缩，而 Web preset 继承宿主的沙箱文件系统，JSON-RPC 组合则挂载 `fs-sandbox` 和文件系统策略。因此，长会话可能替换历史记录，编辑器也会宣告并实施裸本地参考运行时并不具备的文件系统策略。
+
+两条启动路径的配置所有者也不同。Web 在已运行的宿主上挂载逐 agent preset，Python SDK 则初始化一个完整的 stdio JSON-RPC 子进程。将二者视为可互换的同一个 Cordis leaf 会掩盖生命周期差异，而且 SDK 示例没有通过环境选择模型或系统提示词的入口。
+
+## 决策
+
+两种随附 minimal profile 都只对外提供持久 `bash` 与 `str_replace_editor`，不挂载上下文压缩提供方，并让编辑器使用 `@deepseek-ai/dsh-fs-local`。Web preset 在 agent entry 内隔离 `ctx.fs`，将 `fs-local` 与编辑器一起挂载，因此其他 Web agent 仍使用宿主文件系统提供方。其 persona 继续采用较早的 [minimal preset 组合决策](../bug-fix/2026-08-10-minimal-preset-owns-rl-composition.md)所拥有的固定 complete 提示词。
+
+独立的 [`minimal.cordis.yml`](../../../../examples/jsonrpc-agent/minimal.cordis.yml) 仍是完整的 JSON-RPC 进程组合。它挂载 `dsh-jsonrpc`、持久 Bash 所需的本地 PTY 和子进程服务、`fs-local`、两个工具消费方，以及未压缩的 JSONL 持久化。它不挂载 `token-meter`、`compact-basic`、`fs-sandbox` 或 `fs-policy`。持久 Bash 仍消费部署的 danger-full-access 沙箱策略；编辑器不受该策略限制。
+
+`DSH_SYSTEM_PROMPT` 选择独立组合的 persona。`DSH_MODEL` 命名 DeepSeek 提供方目录项，`DSH_CONTEXT_WINDOW` 提供该目录项的容量。由于 SDK 客户端拥有 JSON-RPC `initialize` 请求，[`minimal.py`](../../../../examples/jsonrpc-agent/minimal.py)也使用 `DSH_MODEL` 作为 `model` 参数的默认值；显式 `--model` 仍具有最高优先级。端点与凭据变量继续由 DeepSeek 适配器现有的环境解析路径持有。
+
+## 验证
+
+Web 回放会启动完整 Web 宿主，通过 preset 服务创建 agent，并断言作用域文件系统为裸后端、不存在作用域压缩服务，而且组装请求只包含固定提示词与两个工具。随后，它通过真实作用域服务执行持久 Bash 和编辑器。
+
+SDK 回放通过 SDK 客户端启动真实 JSON-RPC agent 进程，注入由环境选择的提示词，断言组装提示词与精确双工具目录，并执行两个工具。Python SDK 内置运行时覆盖会通过每种可用的打包载体，使用环境选择的模型、模型容量和提示词值初始化独立配置。Cordis 校验会检查两份配置能否解析声明的插件和配置字段。
+
+## 考虑过的替代方案
+
+**以较高阈值保留 `compact-basic`。** 不予采用，因为即便提供方在短测试中未触发，较长会话仍允许替换历史记录，而且 minimal 组合仍会依赖模型容量元数据与 token meter。
+
+**在 danger-full-access 模式下保留 `fs-sandbox`。** 不予采用，因为沙箱提供方仍会使限权与提权成为编辑器能力的一部分。目标运行时要求裸本地提供方，而其不具备 `sandboxMode` 正是组合事实。
+
+**为 Web 与 Python SDK 启动使用同一个 Cordis leaf。** 不予采用，因为 Web preset 向现有多会话宿主贡献 agent 作用域服务，而 Python SDK 必须启动包含 JSON-RPC 服务器及其进程级依赖的完整进程。
+
+**只在 Cordis 内读取 `DSH_MODEL`。** 不予采用，因为 Cordis 配置提供方目录，但不拥有 SDK 客户端的 JSON-RPC `initialize` 请求。launcher 必须向客户端请求传递同一个模型，环境值才能选择路由模型。
+
+## 后果
+
+Minimal 会话不会摘要或替换较早历史；调用方必须让会话轮次保持在所选模型的上下文容量内。编辑器可以访问运行时进程可见的任何绝对路径，且不受持久 shell 沙箱策略影响。两条启动路径共享面向模型的工具与无压缩保证，同时保留适合各自所有者的不同提示词和模型配置。Python SDK 路径继续仅通过内置 stdio JSON-RPC 运行时通信。