Merge remote-tracking branch 'origin/master' into claude/web-llm-pi-ai-config-385e24

# Conflicts:
#	packages/llm/llm-pi-ai/tests/adapter.spec.ts
#	packages/llm/llm/README.i18n.yaml

.agents/notes/archived/feature/2026-07-29-addressable-queue-operations.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
+2026-07-29-addressable-queue-operations.md: 258e4e4e50a4c6386a2e8c28402fe32cb8870e1a
+2026-07-29-addressable-queue-operations.zh.md: b4370d87892df2e0477c0897b45b4c7a1030978e

.agents/notes/archived/feature/2026-07-29-addressable-queue-operations.md

@@ -0,0 +1,44 @@
+# Agent Note: Address pending queue occurrences for edit and removal
+
+Status: implemented
+Archived: 2026-07-31
+
+English | [中文](2026-07-29-addressable-queue-operations.zh.md)
+
+## Problem
+
+The Web queue rendered pending messages but could not edit or delete one row. `MessageId` was insufficient as an address because callers may enqueue the same immutable message more than once. The browser also inferred queue retirement from turn and status events, so a row operation racing with driver claim had no authoritative outcome.
+
+## Decision
+
+**Each accepted FIFO occurrence has its own identity.** AgentLoop mints an opaque `InboxItemId` and publishes an `InboxItem` containing that id, the identified `UserMessage`, and its acceptance-time `queued | steering` placement. Reusing one `MessageId` creates distinct inbox identities. Injection bypasses the FIFOs and receives no inbox identity.
+
+**Mutation ends at driver claim.** `Agent.updateInbox(id, action)` synchronously searches the pending queued FIFO. Edit replaces frozen content while preserving `InboxItemId`, `MessageId`, source, wake policy, and position. Remove emits the occurrence’s terminal discard. Steering and driver-claimed occurrences return `not-found`, so queue operations never rewrite active-turn input or durable history.
+
+**The live ledger is authoritative.** `agent/inbox/enqueue`, `update`, `dequeue`, and `discard` maintain a Host mirror of queued occurrences. A synchronously re-entrant update or terminal event may reach the mirror before its outer enqueue listener; the mirror retains that unseen outcome for the current dispatch and folds it into the enqueue, so listener registration order cannot publish stale content or a ghost row. The wire sends complete `session/queue` snapshots rather than incremental guesses. Reconnect sends the current baseline, and every queued mutation or terminal event replaces it. The client applies no optimistic edit and never retires a row from durable turn events or status changes.
+
+**Queue addresses require a live Agent.** `session.updateQueue` queries only the mounted Agent registry and never resumes a cold session: an `InboxItemId` is process-local and cannot name work after restart or disposal. A missing Agent and a driver-claimed occurrence both return `queue-item-not-found`.
+
+**Web actions address Queue only.** The Host excludes pending steering from `session/queue`; steering retains its existing durable transcript path after consumption. QueueDock hides while empty, renders one pending occurrence directly, and defaults two or more occurrences to a collapsed `"<n> 条排队消息"` header that expands or collapses the complete list. The header exposes `aria-expanded` and `aria-controls`; the expanded list scrolls within a 180px height bound. An active edit or mutation keeps its rows visible, and emptying the queue restores the collapsed default for the next queue. Visible rows expose edit and delete, but no send-now control. The UI derives queue row and mutation types from the runtime `SessionFace` contract rather than importing the connection plugin, so plugin cooperation continues through services and snapshots. Edit is available only when all content blocks are text; the editor cannot silently drop non-text blocks. An editing row exposes only save and cancel, with Enter and Escape as their keyboard equivalents. Delete removes the exact occurrence.
+
+## Alternatives considered
+
+**Address rows by `MessageId`.** Rejected because one immutable message may be sent repeatedly; editing or deleting by message identity would affect an ambiguous occurrence.
+
+**Apply optimistic browser mutations.** Rejected because driver claim and another client can win before the Host action. Waiting for the authoritative snapshot makes the ownership boundary visible and lets `queue-item-not-found` report a real race.
+
+**Include pending steering in the queue mutation protocol.** Rejected because QueueDock has no steering interaction, and editing or deleting active-turn input would widen this feature beyond its current consumer. A dedicated steering interaction owns that delivery contract.
+
+**Expose a protocol-only promotion operation.** Rejected because no product interaction reorders Queue. A public operation without a current consumer would add ordering semantics and tests for speculative use.
+
+**Resume a cold Agent for a queue operation.** Rejected because durable session identity does not preserve the process-local inbox capability. Resuming can only produce `not-found` after creating unrelated live state.
+
+## Verification
+
+AgentLoop contract tests hold prompt admission while editing and removing exact queued occurrences, reject mutations of steering occurrences, and verify the resulting independent turn and terminal lifecycle events. Host schema and proxy tests cover queued-only authoritative snapshots, synchronous re-entrant mutation order, reconnect, cold-Agent rejection, typed not-found errors, and the RPC transport. Client runtime and QueueDock tests cover non-optimistic projection, single-row presentation, default multi-row collapse, interaction-forced visibility, reset after emptying, expansion, text-only editing, save and cancel affordances, removal, retirement races, and disabled mixed-content editing. Keyless browser scenarios capture the default collapsed header before expanding the queue and driving its exposed edit and delete actions through the built Web composition and real HTTP/SSE wire.
+
+## Consequences
+
+Queued work gains precise row operations without becoming durable session history. Occurrence identity is a live process-local capability and disappears at claim, cancellation, disposal, or restart; reconnect recovers only queued items still held by the live Agent. Editing excludes mixed content until an editor can preserve every block, while pending steering remains outside this operation surface.
+
+The protocol now carries full queue snapshots on each change. Queues are expected to remain short, so deterministic recovery and multi-client convergence are preferred over an incremental mutation protocol.

.agents/notes/archived/feature/2026-07-29-addressable-queue-operations.zh.md

@@ -0,0 +1,44 @@
+# Agent Note: 为待处理队列项提供编辑与移除操作
+
+Status: implemented
+Archived: 2026-07-31
+
+[English](2026-07-29-addressable-queue-operations.md) | 中文
+
+## 问题
+
+Web 队列能够渲染待处理消息，但无法编辑或删除其中某一行。`MessageId` 不足以充当寻址标识，因为调用方可以多次将同一条不可变消息加入队列。浏览器还会根据轮次和状态事件推断队列项已退役，因此当行操作与驱动器认领发生竞态时，系统无法给出权威结果。
+
+## 决策
+
+**每次获准进入 FIFO 的项都有独立标识。** AgentLoop 会铸造不透明的 `InboxItemId`，并发布一个 `InboxItem`，其中包含该 id、已有标识的 `UserMessage`，以及接受时确定的 `queued | steering` 放置方式。复用同一个 `MessageId` 会创建不同的 inbox 标识。注入绕过 FIFO，因此不会获得 inbox 标识。
+
+**变更边界止于驱动器认领。** `Agent.updateInbox(id, action)` 会同步搜索待处理的 queued FIFO。编辑会替换已冻结的内容，同时保留 `InboxItemId`、`MessageId`、来源、唤醒策略和位置。移除会发出该次入队项的终态 discard。steering（中途引导）项和已被驱动器认领的项会返回 `not-found`，因此队列操作绝不会改写活动轮次输入或持久历史。
+
+**实时账本是权威状态。** `agent/inbox/enqueue`、`update`、`dequeue` 和 `discard` 共同维护 queued 入队项的 Host 镜像。同步可重入的 update 或终态事件可能先于外层 enqueue 监听器到达镜像；镜像会在当前分发期间保留这一尚不可见的结果，并在处理 enqueue 时把它合并进去，因此监听器注册顺序不会导致系统发布陈旧内容或不存在的行。协议发送完整的 `session/queue` 快照，而非增量猜测。重连会发送当前基线，每次 queued 变更或终态事件都会整体替换它。客户端不会进行乐观编辑，也绝不根据持久轮次事件或状态变化退役队列行。
+
+**Queue 寻址要求 Agent 存活。** `session.updateQueue` 只查询已挂载的 Agent 注册表，绝不恢复冷会话：`InboxItemId` 属于进程本地标识，无法在重启或资源释放后继续指向工作。Agent 缺失和单次入队项已被驱动器认领这两种情况都返回 `queue-item-not-found`。
+
+**Web 操作只面向 Queue。** Host 从 `session/queue` 中排除待处理 steering；steering 消费后仍沿用既有的持久 transcript（文本记录）路径。QueueDock 在队列为空时隐藏，只有一个待处理项时直接渲染该行，存在两个或更多待处理项时则默认收起为可展开或收起完整列表的 `"<n> 条排队消息"` 表头。表头暴露 `aria-expanded` 和 `aria-controls`；展开后的列表以 180px 为高度上限，并可滚动。存在进行中的编辑或变更时，列表行会保持可见；队列清空后，下一次出现队列时会恢复默认收起状态。可见行暴露编辑和删除操作，不提供立即发送控件。UI 从运行时 `SessionFace` 契约派生队列行与变更类型，而不是导入连接插件，因此插件仍通过服务和快照协作。仅当所有内容块都是文本时才提供编辑功能；编辑器不能静默丢弃非文本块。编辑中的行只展示保存和取消操作，对应的键盘操作分别是 Enter 和 Escape。删除会移除对应的精确入队项。
+
+## 考虑过的替代方案
+
+**通过 `MessageId` 寻址行。** 不予采纳，因为同一条不可变消息可以重复发送；按消息标识编辑或删除会无法确定应操作哪一次入队。
+
+**在浏览器中进行乐观变更。** 不予采纳，因为驱动器认领或另一个客户端可能先于 Host 操作完成。等待权威快照可以显式呈现所有权边界，并让 `queue-item-not-found` 报告真实竞态。
+
+**将待处理 steering 纳入队列变更协议。** 不予采纳，因为 QueueDock 没有 steering 交互，而编辑或删除活动轮次输入会把此功能扩展到当前消费方之外。应由专用 steering 交互负责该投递契约。
+
+**暴露仅协议层的前移操作。** 不予采纳，因为当前没有产品交互会重新排序 Queue。公开一个没有当前消费方的操作，会为了推测性用途引入排序语义和测试。
+
+**为队列操作恢复冷 Agent。** 不予采纳，因为持久会话标识不会保留进程本地的 inbox 寻址凭据。恢复只能在创建无关的实时状态后得到 `not-found`。
+
+## 验证
+
+AgentLoop 契约测试会在编辑和移除精确 queued 入队项时阻塞提示词接纳，拒绝对 steering 入队项的变更，并验证所得独立轮次及终态生命周期事件。Host schema 与代理测试覆盖仅含 queued 项的权威快照、同步可重入变更顺序、重连、拒绝冷 Agent、类型化 not-found 错误和 RPC 传输。客户端运行时和 QueueDock 测试覆盖非乐观投影、单行呈现、多行默认收起、交互期间强制保持可见、清空后重置、展开、仅文本编辑、保存与取消入口、移除、退役竞态，以及禁用混合内容编辑。无密钥浏览器场景会先捕获默认收起的表头，再展开队列，并通过构建后的 Web 组合和真实 HTTP／SSE 协议操作其公开的编辑和删除。
+
+## 后果
+
+queued 工作获得精确的行操作，但不会因此成为持久会话历史。单次入队标识是进程本地的实时寻址凭据，会在认领、取消、dispose 或重启时消失；重连只能恢复仍由活跃 Agent 持有的 queued 项。编辑会排除混合内容，直至编辑器能够保留每个块；待处理 steering 则不属于此操作接口。
+
+现在，协议会在每次变更时携带完整队列快照。队列预期保持较短，因此系统优先选择确定性恢复和多客户端收敛，而非增量变更协议。

.agents/notes/archived/manifest.json

@@ -199,6 +199,9 @@
     "feature/2026-07-28-dsh-meta-source-workspace.i18n.yaml": "sha256:f7c6b5db53c32c7475f4f6bb3ae189ed8d67f2163196ab315d6f81fde3aa37d9",
     "feature/2026-07-28-dsh-meta-source-workspace.md": "sha256:ee8b2f6055b27957fa27258f26a07183b7102933af68c64d0df418e32d3d8754",
     "feature/2026-07-28-dsh-meta-source-workspace.zh.md": "sha256:0b10db368e04c24be03569a56ec4b69fed66b70c39a3c4f168b5f1c912efcd96",
+    "feature/2026-07-29-addressable-queue-operations.i18n.yaml": "sha256:0a067b38dd3c02ac41148ed03a7a6fa4c3dee88d5a4b9bad325ff76ebc03c443",
+    "feature/2026-07-29-addressable-queue-operations.md": "sha256:f4d38d6cc49ad23cda2a287ba7ea5ce0e5fdb0edd41f33e68633f439654ca9fa",
+    "feature/2026-07-29-addressable-queue-operations.zh.md": "sha256:9aedd7e241cf46ec7a78ae999a7fb105d73da4baf57c9ecf7ae745ee9b98182b",
     "feature/2026-07-29-tui-hidden-mode-assistant-fold.i18n.yaml": "sha256:0865835802348b730542adbe6b7db613750f3786993c6a14dbb2f47686c13c70",
     "feature/2026-07-29-tui-hidden-mode-assistant-fold.md": "sha256:a5fefebd802e2d9c3c79c7852c1c34c7bbef3f2ac2150d224608b9ec44e966ad",
     "feature/2026-07-29-tui-hidden-mode-assistant-fold.zh.md": "sha256:21bccd1e07ec8dc73b618f428461848bb90b6235afe0b842afb0afab2d5cc575",

.agents/notes/implemented/architecture/2026-06-11-content-block-vocabulary.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/architecture/2026-06-11-content-block-vocabulary.md
-2026-06-11-content-block-vocabulary.md: d926c28e7e197aff28c7b1c09d085febf866832b
-2026-06-11-content-block-vocabulary.zh.md: c547cd87acf61107d1b5ff2960878da7ea8cfc53
+2026-06-11-content-block-vocabulary.md: 5228724bb9101307db9929aaf7831b477c2a6022
+2026-06-11-content-block-vocabulary.zh.md: ffcbbc13dfe9176941f4838b0078d3850403a16c

.agents/notes/implemented/architecture/2026-06-11-content-block-vocabulary.md

@@ -12,7 +12,7 @@ The harness needs one internal language for messages that the loop, session log,
 
 Own the vocabulary: messages are arrays of typed content blocks (`text`, `reasoning`, `tool-call`, `tool-result`), with the union derived from the merge-extensible `ContentBlockMap` so plugins add block types via declaration merging. The same merge-extensible-map pattern types every "stringly" field (`MessageSource`, `FinishReason`, `TurnTrigger`, `TurnEndReason`). Streaming is a raw chunk protocol; `BlockAssembler` is the single shared assembly implementation. Adapters translate to provider wire formats — mapping cost lives in adapters, where it belongs.
 
-In-session context injection (`context/message`) and mid-turn steering (`steering/message`) originally rendered as tagged user-role envelopes (the system-reminder pattern) rather than a new role, so adapters carry zero burden. Both now project as plain user content with no wrapper; see [the injected-content-envelope Agent Note](../simplification/2026-07-20-unwrap-injected-content-envelopes.md). Live-adapter validation confirms this rendering for current DeepSeek behavior; a future provider-specific mismatch belongs in that adapter rather than a new canonical role.
+In-session context injection (`context/message`) and mid-turn steering originally rendered as tagged user-role envelopes (the system-reminder pattern) rather than a new role, so adapters carry zero burden. Both now project as plain user content with no wrapper; see [the injected-content-envelope Agent Note](../simplification/2026-07-20-unwrap-injected-content-envelopes.md). Live-adapter validation confirms this rendering for current DeepSeek behavior; a future provider-specific mismatch belongs in that adapter rather than a new canonical role.
 
 ## Alternatives considered
 

.agents/notes/implemented/architecture/2026-06-11-content-block-vocabulary.zh.md

@@ -12,7 +12,7 @@ harness 需要一套统一的内部消息语言，供 agent loop（智能体循
 
 自主拥有词汇：消息是类型化内容块的数组（`text`、`reasoning`、`tool-call`、`tool-result`），其联合类型派生自可合并扩展的 `ContentBlockMap`，插件通过声明合并添加新的块类型。同一可合并扩展映射模式为所有「字符串化」字段提供类型（`MessageSource`、`FinishReason`、`TurnTrigger`、`TurnEndReason`）。流式输出采用原始分片协议；`BlockAssembler` 是唯一的共享组装实现。适配器负责转换为提供方的协议格式（wire format）——映射成本留在适配器中，正是它该在的地方。
 
-会话内上下文注入（`context/message`）和轮次中途 steering（中途引导）（`steering/message`）最初渲染为带标签的 user-role 信封（system-reminder 模式），而非引入新角色，因此适配器无需承担额外负担。如今两者都投影为无包装的普通用户内容；见[注入内容信封 Agent Note](../simplification/2026-07-20-unwrap-injected-content-envelopes.md)。实际适配器验证已确认此渲染方式符合当前 DeepSeek 的行为；如果未来某提供方出现不兼容，应在该适配器内处理，而非引入新的规范角色。
+会话内上下文注入（`context/message`）和轮次中途 steering 最初渲染为带标签的 user-role 信封（system-reminder 模式），而非引入新角色，因此适配器无需承担额外负担。如今两者都投影为无包装的普通用户内容；见[注入内容信封 Agent Note](../simplification/2026-07-20-unwrap-injected-content-envelopes.md)。实际适配器验证已确认此渲染方式符合当前 DeepSeek 的行为；如果未来某提供方出现不兼容，应在该适配器内处理，而非引入新的规范角色。
 
 ## 曾考虑的替代方案
 

.agents/notes/implemented/architecture/2026-06-11-event-sourced-sessions.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/architecture/2026-06-11-event-sourced-sessions.md
-2026-06-11-event-sourced-sessions.md: 15ba7b23d5eae48e7dee2328b5924493d54aeeb0
-2026-06-11-event-sourced-sessions.zh.md: 011d139f112ca86e0894878c5ffb0e9ea255a664
+2026-06-11-event-sourced-sessions.md: 01f9628c1cfc000aca8654caf5edeff09411fdcc
+2026-06-11-event-sourced-sessions.zh.md: ec5c3e766dfa97c5612827023c5cc66bf01a8e6c

.agents/notes/implemented/architecture/2026-06-11-event-sourced-sessions.md

@@ -14,7 +14,7 @@ A `Session` is an append-only log of typed `SessionEvent`s — the single source
 
 Appends are synchronous (the hot path never blocks on I/O); `session/event` is a sync notification; persistence plugins buffer write-behind and drain at the awaited `session/flush` checkpoint fired at every turn end.
 
-Ordering contract: the loop appends to the session *before* emitting the corresponding Cordis event, and the `agent/step-result` waterfall runs before the `assistant/message` append so the log records the message tool dispatch actually used. Regression tests pin that ordering.
+Ordering contract: the loop claims inbox messages before `agent/pre-step`, opens `step/start` only after an enter decision, then appends the returned `user/message` batch before request derivation. Provider output is assembled and appended as `assistant/message` before tool dispatch, so the durable log records the exact message the tools follow. Regression tests pin that ordering.
 
 ## Alternatives considered
 

.agents/notes/implemented/architecture/2026-06-11-event-sourced-sessions.zh.md

@@ -14,7 +14,7 @@ MVP 要求严格的基于事件的追踪，以及完全可回放的会话（严
 
 追加操作是同步的（热路径从不阻塞于 I/O）；`session/event` 是同步通知；持久化插件在后台缓冲写入，并在每个轮次结束时触发的 `session/flush` 检查点处等待排空。
 
-顺序契约：agent loop（智能体循环）*先*追加到会话，再发出对应的 Cordis 事件；`agent/step-result` waterfall（瀑布式事件）在 `assistant/message` 追加之前运行，因此日志记录的是工具调度实际使用的消息。回归测试固定了这一顺序。
+顺序契约：agent loop（智能体循环）先领取 inbox 消息，再运行 `agent/pre-step`；只有 enter 决策才打开 `step/start`，随后在请求派生前追加返回的 `user/message` 批次。提供方输出组装并以 `assistant/message` 追加后才分派工具，因此持久日志记录工具实际遵循的确切消息。回归测试固定了这一顺序。
 
 ## 曾考虑的替代方案
 

.agents/notes/implemented/architecture/2026-06-11-microkernel-event-taxonomy.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/architecture/2026-06-11-microkernel-event-taxonomy.md
-2026-06-11-microkernel-event-taxonomy.md: 8bf05b7deba5f054d4ec8ecf104c3b8798e42d4e
-2026-06-11-microkernel-event-taxonomy.zh.md: c0985bc8685f5ed9ca6eba3c3e47dd0c7e713dad
+2026-06-11-microkernel-event-taxonomy.md: 202595fed125966a5d77920536e7f4ee88f875fe
+2026-06-11-microkernel-event-taxonomy.zh.md: 899c96d86cb7e37d90df349ce5f3f932e0a72f95

.agents/notes/implemented/architecture/2026-06-11-microkernel-event-taxonomy.md

@@ -12,10 +12,10 @@ The product principle is "everything is a plugin": hooks, /goal, /loop, dynamic
 
 Pure Cordis event taxonomy. The loop's extension seams are typed events with deliberate dispatch modes:
 
-- **waterfall** (around-middleware) where plugins transform, veto, recover, or wrap: `agent/prompt-submit`, `agent/request`, `agent/request-error`, `agent/step-result`, `agent/turn-continuation`, `tools/pre-execute`, `tools/execute`, `tools/post-execute`, `llm/stream`, `system-prompt/assemble`.
-- **serial** (awaited in listener order; a bail value stops later listeners) for ordered checkpoints: every `agent/pre-step` and `agent/post-step` listener runs when all abstain, while the first stop returned from `agent/turn-stop` makes the terminal decision final.
+- **waterfall** (around-middleware) where plugins transform, veto, recover, or wrap: `agent/pre-step`, `agent/request`, `agent/request-error`, `tools/pre-execute`, `tools/execute`, `tools/post-execute`, `llm/stream`, `system-prompt/assemble`.
+- **serial** (awaited in listener order) for ordered checkpoints such as `agent/turn-stopping`.
 - **parallel** (awaited fan-out) where every listener must get an independent chance: the `session/flush` durability checkpoint.
-- **emit** (synchronous fire-and-forget) for notifications: turn/step boundaries, stream chunks, lifecycle, errors, and the contained immutable `tools/result` observation.
+- **emit** (synchronous fire-and-forget) for notifications: inbox transitions, lifecycle, errors, and the contained immutable `tools/result` observation. Durable session events own turn and step boundaries.
 
 The event vocabulary lives in interface packages (dsh-agent declares the agent/* events); `@deepseek-ai/dsh-agent-loop` is the only concrete loop plugin and is itself swappable — nothing outside it may depend on it.
 

.agents/notes/implemented/architecture/2026-06-11-microkernel-event-taxonomy.zh.md

@@ -12,10 +12,10 @@ Status: implemented
 
 纯 Cordis 事件分类体系。agent loop（智能体循环）的扩展 seam 是带类型的事件，具有明确的分发模式：
 
-- **waterfall（瀑布式事件）**（around-middleware）：插件可变换、否决、恢复或包装：`agent/prompt-submit`、`agent/request`、`agent/request-error`、`agent/step-result`、`agent/turn-continuation`、`tools/pre-execute`、`tools/execute`、`tools/post-execute`、`llm/stream`、`system-prompt/assemble`。
-- **serial**（按监听器顺序依次 await；bail 值会阻止后续监听器执行）：用于有序检查点。当所有监听器均未返回 bail 值时，`agent/pre-step` 和 `agent/post-step` 的每个监听器都会运行，而 `agent/turn-stop` 返回的第一个 stop 值即为最终的终止决策。
+- **waterfall（瀑布式事件）**（around-middleware）：插件可变换、否决、恢复或包装：`agent/pre-step`、`agent/request`、`agent/request-error`、`tools/pre-execute`、`tools/execute`、`tools/post-execute`、`llm/stream`、`system-prompt/assemble`。
+- **serial**（按监听器顺序依次 await）：用于 `agent/turn-stopping` 等有序检查点。
 - **parallel**（await 扇出）：每个监听器都必须获得独立执行的机会：`session/flush` 持久性检查点。
-- **emit**（同步 fire-and-forget）：用于通知：轮次/步骤边界、流分片、生命周期、错误，以及受错误隔离的 `tools/result` 观测；该观测接收不可变的最终结果。
+- **emit**（同步 fire-and-forget）：用于 inbox 转换、生命周期、错误，以及包含不可变 `tools/result` 观测的事件。轮次与步骤边界由持久会话事件拥有。
 
 事件词汇定义在接口包中（dsh-agent 声明 agent/* 事件）；`@deepseek-ai/dsh-agent-loop` 是唯一的具体循环插件，且自身可替换——外部不得依赖它。
 

.agents/notes/implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.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/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md
-2026-06-18-agent-lifecycle-and-ownership-seams.md: f190b4ba2b7f22d29f473c8a2725401ff371488e
-2026-06-18-agent-lifecycle-and-ownership-seams.zh.md: 4c55323450b0e4b2aa5a4354c8b26163260c539a
+2026-06-18-agent-lifecycle-and-ownership-seams.md: 93247a6da7446a5a67db33423d2b766ce4cf3308
+2026-06-18-agent-lifecycle-and-ownership-seams.zh.md: 7c0f27e1e6ddeec91a8031ef7b1d9fd965a4463a

.agents/notes/implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md

@@ -47,4 +47,4 @@ The bash owner-token comparison relies on the shared `Agent.id`/`SessionId` bein
 
 ## Consequences
 
-This touched public interfaces (`Agent`, `AgentFactory`, the bash seam) deliberately, not as a local ACP patch. The simple synchronous `Agent.send()` ergonomics were preserved; the async lifecycle path is additive, for owners that need it.
+This touched public interfaces (`Agent`, `AgentFactory`, the bash seam) deliberately, not as a local ACP patch. Synchronous agent delivery remains simple; the async lifecycle path is additive for owners that need it.

.agents/notes/implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.zh.md

@@ -47,4 +47,4 @@ bash 所有者 token 比较依赖共享的 `Agent.id`/`SessionId` 在存活 agen
 
 ## 后果
 
-本变更有意触及公开接口（`Agent`、`AgentFactory`、bash seam），而非作为 ACP 的局部补丁。同步 `Agent.send()` 的简洁易用性得以保留；异步生命周期路径是增量添加的，供需要它的所有者使用。
+本变更有意触及公开接口（`Agent`、`AgentFactory`、bash seam），而非作为 ACP 的局部补丁。同步 agent 交付仍然简单；异步生命周期路径是增量添加的，供需要它的所有者使用。

.agents/notes/implemented/architecture/2026-06-18-session-surface.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/architecture/2026-06-18-session-surface.md
-2026-06-18-session-surface.md: 80034881d0112076759a68737b5931c8ff659d15
-2026-06-18-session-surface.zh.md: b7fd67eb0749b0d111f2941059ed2d300875c6e0
+2026-06-18-session-surface.md: eeac53534c70099e4102aff9ef226702ea939654
+2026-06-18-session-surface.zh.md: c58d3da049cd6c18e564e596354f5d1831c4756f

.agents/notes/implemented/architecture/2026-06-18-session-surface.md

@@ -27,7 +27,7 @@ export type SurfaceOp =
   | { op: 'replace'; start: number; end: number }  // shadow [start, end] inclusive
 ```
 
-1. **Append** — add the new event seq to the tail. Used by `user/message`, `assistant/message`, `tool/result`, `context/message`, `steering/message`. The loop passes `surfaceOp: 'append'` on all such appends and records `sourceEventSeqs` where applicable: every successful `assistant/message` records its complete `assistant/chunk` source set, including `[]`, while `tool/result` records its `tool/call` source.
+1. **Append** — add the new event seq to the tail. Used by `user/message`, `assistant/message`, `tool/result`, `context/message`. The loop passes `surfaceOp: 'append'` on all such appends and records `sourceEventSeqs` where applicable: every successful `assistant/message` records its complete `assistant/chunk` source set, including `[]`, while `tool/result` records its `tool/call` source.
 
 2. **Replace** — remove entries from `start` through `end` (both inclusive) and insert the new event seq in their place. Both `start` and `end` must be present in the current surface; `start === end` replaces one entry. The event's `sourceEventSeqs` must contain every shadowed surface seq. The shadowed events remain in the log but are no longer on the surface.
 

.agents/notes/implemented/architecture/2026-06-18-session-surface.zh.md

@@ -27,7 +27,7 @@ export type SurfaceOp =
   | { op: 'replace'; start: number; end: number }  // shadow [start, end] inclusive
 ```
 
-1. **Append**：在尾部追加新事件的 seq。`user/message`、`assistant/message`、`tool/result`、`context/message`、`steering/message` 使用此操作。agent loop（智能体循环）在所有此类追加上传入 `surfaceOp: 'append'`，并在适用时记录 `sourceEventSeqs`：每个成功的 `assistant/message` 都记录完整的 `assistant/chunk` 来源集合（包括 `[]`），而 `tool/result` 记录其 `tool/call` 来源。
+1. **Append**：在尾部追加新事件的 seq。`user/message`、`assistant/message`、`tool/result`、`context/message` 使用此操作。agent loop（智能体循环）在所有此类追加上传入 `surfaceOp: 'append'`，并在适用时记录 `sourceEventSeqs`：每个成功的 `assistant/message` 都记录完整的 `assistant/chunk` 来源集合（包括 `[]`），而 `tool/result` 记录其 `tool/call` 来源。
 
 2. **Replace**：移除从 `start` 到 `end`（两端包含）的条目，并在其位置插入新事件的 seq。`start` 和 `end` 都必须存在于当前 surface；`start === end` 表示替换单个条目。该事件的 `sourceEventSeqs` 必须包含所有被遮蔽的 surface seq。被遮蔽的事件仍留在日志中，但不再出现在 surface 上。
 

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.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/architecture/2026-06-21-bounded-llm-request-recovery.md
-2026-06-21-bounded-llm-request-recovery.md: 24725dcf300cf69e9cc72580d0c8afe937d4e2b9
-2026-06-21-bounded-llm-request-recovery.zh.md: 92132704304c4a91908e48df1fdc7cf9c59efe11
+2026-06-21-bounded-llm-request-recovery.md: 587f2d26eee922e91c8797018964f983890eb8ec
+2026-06-21-bounded-llm-request-recovery.zh.md: be52ece63ce794cb13cdf657b73bae6e9f42cf6d

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md

@@ -4,11 +4,11 @@ Status: implemented
 
 English | [中文](2026-06-21-bounded-llm-request-recovery.zh.md)
 
-The [per-provider request retry policy](../feature/2026-07-24-provider-retry-policies.md) extends this foundation with exact-provider configuration and an explicit unbounded mode. This note continues to own structured failure facts, the closed-step recovery boundary, normal mode's transient defaults, visible single attempts, and durable retry status.
+The [per-provider request retry policy](../feature/2026-07-24-provider-retry-policies.md) extends this foundation with exact-provider configuration and an explicit unbounded mode. This note continues to own structured failure facts, the closed-step recovery boundary, normal mode's transient defaults, visible single attempts, and durable retry status. [Terminal LLM stream failures](2026-07-29-terminal-llm-stream-failures.md) supersedes its thrown-error identity and stream-sidecar mechanism.
 
 ## Problem
 
-`dsh-llm` can report provider failures either by throwing during adapter dispatch or iteration or by ending with `finish { kind: 'error' | 'aborted' }`. The final adapter boundary tags thrown failures so `dsh-agent-loop` can distinguish them from middleware and result-processing defects, and the loop normalizes both delivery forms into `agent/request-error` after closing the failed step. An unhandled failure is terminal; a handling listener repairs policy-owned state, returns `{ kind: 'retry' }`, and stops waterfall delegation. The [retry-action decision](../simplification/2026-07-27-request-error-retry-action.md) owns this return contract.
+Provider adapters can fail by throwing during dispatch or iteration or by ending with `finish { kind: 'error' | 'aborted' }`. The final adapter boundary normalizes thrown values to that terminal finish protocol before `dsh-agent-loop` receives them; middleware and result-processing defects remain thrown. The loop offers a terminal model-request failure to `agent/request-error`. An unhandled failure is terminal; a handling listener repairs policy-owned state, returns `{ kind: 'retry' }`, and stops waterfall delegation. The [retry-action decision](../simplification/2026-07-27-request-error-retry-action.md) owns this return contract.
 
 That boundary is already safe for another request attempt. Raw `assistant/chunk` events carry the failed `turn` and `step`, message derivation ignores them unless a successful `assistant/message` cites them, tool calls are dispatched only after a successful terminal finish and assembly, and a retry opens a new numbered turn from the durable log. The harness therefore does not need a second response lifecycle or tentative-output protocol to keep two attempts separate.
 
@@ -40,9 +40,9 @@ interface LlmFailure {
 
 `code` remains the provider-neutral machine-routing taxonomy established by `HarnessError`; the new fields are observations from the provider boundary. `ProviderRequestId` is owned and constructed by `dsh-llm`, then serializes as its provider-issued string. The payload deliberately has no `retryable`, `failover`, `partialOutput`, provider, model, phase, or route id fields. Retryability belongs to policy, provider/model are already in the durable request header, and partial output is derived from the failed step's `assistant/chunk` events.
 
-`LlmError` carries `failure: LlmFailure` and preserves `failure.code === error.code`. `FinishReasonMap.error` and `FinishReasonMap.aborted` carry the same payload instead of parallel failure shapes. An adapter-thrown `Error` keeps its exact object identity: the final-adapter scope associates the normalized facts with that object in call-local sidecar state and rethrows it unchanged; a non-`Error` throw is wrapped as today. `llmFailureOf(stream, error)` retrieves those facts alongside the existing provenance check, while an in-band finish without an error object becomes a new `LlmError`. This preserves listeners that key on error type or identity while giving all final-adapter failures, including unknown SDK exceptions, an `UNKNOWN` terminal payload.
+`LlmError` carries `failure: LlmFailure` and preserves `failure.code === error.code`. `FinishReasonMap.error` and `FinishReasonMap.aborted` carry the same payload instead of parallel failure shapes. The final adapter boundary detaches those facts from adapter-thrown values and emits the appropriate terminal finish; unknown SDK exceptions receive an `UNKNOWN` payload. Exact thrown-object identity does not cross the LLM stream seam.
 
-The agent loop keeps `RequestError` as that exact error object and passes `LlmFailure` as a separate argument to `agent/request-error`; it does not mutate possibly frozen third-party errors. It also uses the payload when converting an in-band finish and when recording an unrecovered `turn/end.reason`.
+The agent loop passes the terminal finish's `LlmFailure` to `agent/request-error` and uses the same payload when recording an unrecovered `turn/end.reason`.
 
 Adapters extract structured facts before falling back to message inspection. They validate HTTP status, parse `Retry-After` seconds or dates into a positive finite millisecond delay, brand the provider request id when exposed, and distinguish their own timeout from the caller's abort. Provider-specific codes and messages may refine a mapping, but no recovery listener parses them.
 
@@ -106,8 +106,8 @@ If recovery is exhausted, the final failure is stored once on `turn/end.reason`
 
 ## Verification
 
-- `LlmFailure` is the single serializable payload for thrown, error-finish, and aborted-finish final-adapter failures; normalization preserves stable code, status, retry delay, branded provider request id, error cause, and caller-abort versus adapter-timeout classification where available.
-- An adapter-thrown `Error` reaches `agent/request-error` as the exact same object while its sidecar `LlmFailure` reaches the adjacent argument; tests retain the existing identity assertion for extensible and frozen third-party errors.
+- `LlmFailure` is the single serializable payload for adapter throws, error finishes, and aborted finishes; normalization preserves stable code, status, retry delay, branded provider request id, and caller-abort versus adapter-timeout classification where available.
+- Adapter throws become terminal failure chunks before reaching consumers; middleware and consumer exceptions remain thrown outside model-request recovery.
 - DeepSeek and pi-ai adapter tests cover representative 400, 401/403, 429, 5xx, connection, malformed/truncated stream, timeout, abort, retry-after seconds/date, request-id, and unknown-SDK-error paths without recovery policy parsing message text.
 - Pi-ai pins the SDK option to zero retries and performs one observed wire attempt for a retryable provider response; separate tests make removing either boundary fail.
 - `agent/request-error` carries current failure facts, immutable prior-retried failure facts, and the serving registration's immutable retry policy; a success clears the history, and alternating transient/context-overflow integration tests prove the two policies consume only their own finite budgets.

.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.zh.md

@@ -4,11 +4,11 @@ Status: implemented
 
 [English](2026-06-21-bounded-llm-request-recovery.md) | 中文
 
-[按提供方配置的请求重试策略](../feature/2026-07-24-provider-retry-policies.md)在此基础上增加了确切提供方配置与显式无界 mode。本说明继续负责结构化失败事实、已关闭步骤的恢复边界、normal mode 的暂时性默认值、可见的单次尝试和持久重试状态。
+[按提供方配置的请求重试策略](../feature/2026-07-24-provider-retry-policies.md)在此基础上增加了确切提供方配置与显式无界 mode。本说明继续负责结构化失败事实、已关闭步骤的恢复边界、normal mode 的暂时性默认值、可见的单次尝试和持久重试状态。[LLM 流的终止失败](2026-07-29-terminal-llm-stream-failures.md)取代了其中关于抛出错误身份和 stream sidecar 的机制。
 
 ## 问题
 
-`dsh-llm` 可能在适配器分发或迭代时抛出异常，也可能以 `finish { kind: 'error' | 'aborted' }` 结束，以这两种形式报告提供方失败。最终适配器边界会标记抛出的失败，使 `dsh-agent-loop` 能将其与中间件和结果处理缺陷区分开。循环关闭失败步骤后，会把两种交付形式统一规范化为 `agent/request-error`。未被处理的失败是终态；处理失败的监听器修复策略自有状态，返回 `{ kind: 'retry' }`，并停止 waterfall（瀑布式事件）委托。[重试动作决策](../simplification/2026-07-27-request-error-retry-action.md)规定这一返回契约。
+提供方适配器可能在分发或迭代时抛出异常，也可能以 `finish { kind: 'error' | 'aborted' }` 结束。最终适配器边界会在 `dsh-agent-loop` 接收前把抛出值规范化为该终止 finish 协议；middleware 与结果处理缺陷仍会抛出。loop 会将终止模型请求失败交给 `agent/request-error`。未被处理的失败是终态；处理失败的监听器修复策略自有状态，返回 `{ kind: 'retry' }`，并停止 waterfall 委托。[重试动作决策](../simplification/2026-07-27-request-error-retry-action.md)规定这一返回契约。
 
 该边界已能安全地再次发起请求。原始 `assistant/chunk` 事件携带失败的 `turn` 和 `step`；除非某条成功的 `assistant/message` 引用这些事件，否则消息派生会忽略它们。只有终止性 finish 成功且组装完成后，系统才会分发工具调用；重试则会从持久日志开启新的编号轮次。因此，harness 无需引入第二套响应生命周期或暂定输出协议，即可分隔两次尝试。
 
@@ -40,9 +40,9 @@ interface LlmFailure {
 
 `code` 仍是 `HarnessError` 建立的提供方无关机器路由分类体系；新字段是在提供方边界观测到的事实。`ProviderRequestId` 由 `dsh-llm` 拥有并构造，序列化后为提供方发放的字符串。该载荷有意不包含 `retryable`、`failover`、`partialOutput`、提供方、模型、阶段或路由 id 字段。是否可重试属于策略，提供方／模型已位于持久请求头中，部分输出则从失败步骤的 `assistant/chunk` 事件派生。
 
-`LlmError` 携带 `failure: LlmFailure`，并保持 `failure.code === error.code`。`FinishReasonMap.error` 和 `FinishReasonMap.aborted` 携带同一载荷，而不是并行的失败形状。适配器抛出的 `Error` 保留其精确的对象标识：最终适配器 scope 在调用局部的伴随状态中把规范化事实与该对象关联，然后原样重新抛出；非 `Error` 抛出值则依旧被包装。`llmFailureOf(stream, error)` 会在现有来源检查旁取回这些事实，而没有错误对象的带内 finish 则会成为新的 `LlmError`。这既保留了按错误类型或标识分流的监听器，又使所有最终适配器失败（包括未知 SDK 异常）都获得 `UNKNOWN` 终止载荷。
+`LlmError` 携带 `failure: LlmFailure`，并保持 `failure.code === error.code`。`FinishReasonMap.error` 和 `FinishReasonMap.aborted` 携带同一载荷，而不是并行的失败形状。最终适配器边界会从适配器抛出值中分离这些事实，并发出相应的终止 finish；未知 SDK 异常会获得 `UNKNOWN` 载荷。精确的抛出对象身份不会跨越 LLM stream seam。
 
-agent loop（智能体循环）会保留 `RequestError` 作为该精确的错误对象，并将 `LlmFailure` 作为独立参数传给 `agent/request-error`；它不会改动可能已冻结的第三方错误。在转换带内 finish 以及记录未恢复的 `turn/end.reason` 时，循环也会使用该载荷。
+agent loop（智能体循环）会将终止 finish 的 `LlmFailure` 传给 `agent/request-error`，并在记录未恢复的 `turn/end.reason` 时使用同一载荷。
 
 适配器会先提取结构化事实，再回退到消息检查。它们会验证 HTTP 状态，将 `Retry-After` 的秒数或日期解析为正的有限毫秒延迟，在提供方公开请求 id 时将其品牌化，并区分自身超时与调用方中止。提供方专用 code 和消息可以细化映射，但恢复监听器不会解析它们。
 
@@ -106,8 +106,8 @@ agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次
 
 ## 验证
 
-- `LlmFailure` 是最终适配器抛出失败、错误 finish 和中止 finish 使用的唯一可序列化载荷；在可用时，规范化保留稳定 code、状态、重试延迟、品牌化的提供方请求 id、错误原因，以及调用方中止与适配器超时之间的分类。
-- 适配器抛出的 `Error` 会以完全相同的对象抵达 `agent/request-error`，其伴随的 `LlmFailure` 则抵达相邻参数；测试保留针对可扩展及冻结第三方错误的现有对象标识断言。
+- `LlmFailure` 是适配器抛出、错误 finish 和中止 finish 使用的唯一可序列化载荷；在可用时，规范化保留稳定 code、状态、重试延迟、品牌化的提供方请求 id，以及调用方中止与适配器超时之间的分类。
+- 适配器抛出值会在抵达消费方前成为终止失败 chunk；middleware 与消费方异常仍在模型请求恢复之外抛出。
 - DeepSeek 和 pi-ai 适配器测试覆盖具有代表性的 400、401/403、429、5xx、连接、格式错误／截断流、超时、中止、Retry-After 秒数／日期、请求 id 和未知 SDK 错误路径，恢复策略无需解析消息文本。
 - pi-ai 将 SDK 选项固定为零次重试，并针对可重试的提供方响应执行一次可观测的线路请求尝试；独立测试确保移除任一边界都会失败。
 - `agent/request-error` 携带当前失败事实、不可变的先前已重试失败事实，以及实际服务注册所对应的不可变重试策略；成功会清除历史，暂时性失败／上下文溢出交替发生的集成测试证明两种策略只消耗各自的有限预算。

.agents/notes/implemented/architecture/2026-06-30-event-domain-semantics.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/architecture/2026-06-30-event-domain-semantics.md
-2026-06-30-event-domain-semantics.md: 75c1cac11d1bfc9aa7fba9c523eab8c0475027e8
-2026-06-30-event-domain-semantics.zh.md: dec9b1589b79071e06e9eaea24048606aa46c6bf
+2026-06-30-event-domain-semantics.md: 14102b105e7bcaa6a00ac9c406933f772fc45a6f
+2026-06-30-event-domain-semantics.zh.md: 91e490b57bcdc1ed95f0b8b40cc9e16ef4d36f29

.agents/notes/implemented/architecture/2026-06-30-event-domain-semantics.md

@@ -21,7 +21,7 @@ This vocabulary is the foundation for interception decisions, the durable `hook/
 **Three domains, one job each, with a single boundary rule.**
 
 - **`session/*` — the durable, replayable FACT log.** Owns `SessionEventMap`; every entry is JSON-only (no live objects). One `session/event` emit per append, plus the `session/flush` parallel durability checkpoint. It is also the live transcript feed: a consumer that wants to render or react to what happened subscribes here, so live rendering and replay projections share one path.
-- **`agent/*` — the LIVE runtime surface.** Always carries the live `Agent`. Two shapes: INTERCEPTION waterfalls (`agent/request`, `agent/step-result`, `agent/turn-continuation`) that mutate or veto, and TRANSIENT emits (`agent/status`, `agent/error`, `agent/created`/`agent/disposed`, `agent/queued`) that notify with the `Agent` in hand. Turn and step BOUNDARIES are NOT here — they are durable session events read off `session/event`, and so are the token stream (`assistant/chunk`) and mid-turn steering (`steering/message`).
+- **`agent/*` — the LIVE runtime surface.** Always carries the live `Agent`. Interception waterfalls (`agent/pre-step`, `agent/request`, `agent/request-error`) transform, reject, or recover; awaited `agent/turn-stopping` observes the stop boundary; transient emits report lifecycle, status, inbox insertion/claim/discard, and errors. Turn and step BOUNDARIES are NOT here — they are durable session events read off `session/event`, as are the token stream (`assistant/chunk`) and mid-turn steering (a `user/message`).
 - **`tools/*` — the tool registry + execution seam.**
 
 **The boundary rule:** a durable, replayable fact is a `SessionEvent`; a live interception or a transient/live-object signal is an `agent`/`tools` Cordis event. A turn or step boundary is a durable fact, so it lives in the session log and is read off the `session/event` feed — it is NOT mirrored as an `agent/*` emit.
@@ -33,7 +33,7 @@ This vocabulary is the foundation for interception decisions, the durable `hook/
 - The loop no longer emits any boundary mirror; `closeStep` appends `step/end` only and `closeTurn` appends `turn/end` only. `Session.append` owns post-commit observer containment, so a throwing boundary observer cannot change the turn outcome or starve later consumers; an acceptance or internal validation failure still escapes before the boundary enters the log.
 - Tests that observed boundaries via the removed emits now observe the durable `turn/start`/`turn/end`/`step/start`/`step/end` session events — the behavior they pin (boundary ordering, step counting) is unchanged; only the feed they read moved to the canonical one. The tests that exercised a *throwing turn-boundary emit listener* were deleted, because that code path no longer exists (there is no emit to throw from). Per [AGENTS.md "tests document behavior, not golden truth"](../../../../AGENTS.md), the behavior and its test moved (or died) together.
 - The loop marks the step open (`stepOpen = true`) only after `append('step/start')` returns. Internal dispatch validation runs before the log push and may reject without opening a step; post-commit `session/event` observer failures are contained inside `Session.append`. The marker therefore represents exactly the committed boundary that owes a later `step/end`.
-- The full realization of this is [the simplification Agent Note "Stop mirroring durable boundaries as agent events"](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md): all four boundary mirrors are removed and every consumer reads boundaries off `session/event`. `agent/steering` (not a boundary mirror) stayed outside that Agent Note's scope and was removed by its own follow-up, [Remove the `agent/steering` mirror emit](../../archived/simplification/2026-07-04-remove-agent-steering-mirror.md) — it mirrored the durable `steering/message`.
+- The full realization of this is [the simplification Agent Note "Stop mirroring durable boundaries as agent events"](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md): all four boundary mirrors are removed and every consumer reads boundaries off `session/event`. `agent/steering` (not a boundary mirror) stayed outside that Agent Note's scope and was removed by its own follow-up, [Remove the `agent/steering` mirror emit](../../archived/simplification/2026-07-04-remove-agent-steering-mirror.md) — it mirrored the durable mid-turn steering `user/message`.
 - The cordis events catalog (`docs/cordis-catalog/events.md`) is regenerated to drop the mirror events.
 
 <!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->

.agents/notes/implemented/architecture/2026-06-30-event-domain-semantics.zh.md

@@ -21,7 +21,7 @@ harness 通过 Cordis 事件分类体系扩展 agent loop（智能体循环）
 **三个域，各司其职，以一条边界规则统一。**
 
 - **`session/*`——持久的、可回放的事实日志。** 拥有 `SessionEventMap`；每条记录仅含 JSON（无活对象）。每次追加触发一次 `session/event` emit，加上 `session/flush` 并行持久性检查点。它同时也是实时 transcript（文本记录）源：想渲染或响应已发生事件的消费方在此订阅，因此实时渲染与回放投影共享同一路径。
-- **`agent/*`——运行时实时表面。** 始终携带活的 `Agent`。两种形态：拦截 waterfall（瀑布式事件）（`agent/request`、`agent/step-result`、`agent/turn-continuation`）可变更或否决；瞬态 emit（`agent/status`、`agent/error`、`agent/created`/`agent/disposed`、`agent/queued`）在持有 `Agent` 的情况下通知。轮次和步骤边界不在此处——它们是持久的会话事件，从 `session/event` 读取；token 流（`assistant/chunk`）和轮次中途的 steering（中途引导）（`steering/message`）同理。
+- **`agent/*`——运行时实时表面。** 始终携带活的 `Agent`。拦截 waterfall（瀑布式事件）（`agent/pre-step`、`agent/request`、`agent/request-error`）负责变换、拒绝或恢复；awaited `agent/turn-stopping` 观察停止边界；瞬态 emit 报告生命周期、状态、inbox 插入／领取／丢弃与错误。轮次和步骤边界不在此处——它们是持久的会话事件，从 `session/event` 读取；token 流（`assistant/chunk`）和中途 steering（以 `user/message` 呈现）同理。
 - **`tools/*`——工具注册表与执行 seam。**
 
 **边界规则：** 持久的、可回放的事实是 `SessionEvent`；实时拦截或瞬态/活对象信号是 `agent`/`tools` Cordis 事件。轮次或步骤边界是持久事实，因此存在于会话日志中并从 `session/event` 源读取——不会被镜像为 `agent/*` emit。
@@ -33,7 +33,7 @@ harness 通过 Cordis 事件分类体系扩展 agent loop（智能体循环）
 - 循环不再 emit 任何边界镜像；`closeStep` 仅追加 `step/end`，`closeTurn` 仅追加 `turn/end`。`Session.append` 负责 post-commit observer 隔离，因此抛出异常的边界 observer 无法改变轮次结果或饿死后续消费方；事件接纳失败或内部校验失败仍会在边界进入日志之前向外抛出。
 - 之前通过已移除 emit 观察边界的测试，现在观察持久的 `turn/start`/`turn/end`/`step/start`/`step/end` 会话事件——它们所锁定的行为（边界顺序、步骤计数）不变；只是读取的源移到了规范源。那些测试*抛出异常的轮次边界 emit 监听器*的用例被删除，因为该代码路径不再存在（没有 emit 可供抛出）。按照 [AGENTS.md「测试记录行为，而非黄金真相」](../../../../AGENTS.md)，行为与其测试一同迁移（或一同消亡）。
 - 循环仅在 `append('step/start')` 返回后才标记步骤已打开（`stepOpen = true`）。内部分发校验在日志推入之前运行，可能在不打开步骤的情况下拒绝；post-commit `session/event` observer 的失败被隔离在 `Session.append` 内部。因此该标记精确表示已提交的、欠一个后续 `step/end` 的边界。
-- 完整实现见[简化 Agent Note「停止将持久边界镜像为 agent 事件」](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md)：全部四个边界镜像被移除，所有消费方从 `session/event` 读取边界。`agent/steering`（不是边界镜像）不在该 Agent Note 范围内，由其后续 Agent Note [移除 `agent/steering` 镜像 emit](../../archived/simplification/2026-07-04-remove-agent-steering-mirror.md) 单独移除——它镜像的是持久的 `steering/message`。
+- 完整实现见[简化 Agent Note「停止将持久边界镜像为 agent 事件」](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md)：全部四个边界镜像被移除，所有消费方从 `session/event` 读取边界。`agent/steering`（不是边界镜像）不在该 Agent Note 范围内，由其后续 Agent Note [移除 `agent/steering` 镜像 emit](../../archived/simplification/2026-07-04-remove-agent-steering-mirror.md) 单独移除——它镜像的是持久的中途 steering `user/message`。
 - Cordis 事件目录（`docs/cordis-catalog/events.md`）重新生成以移除镜像事件。
 
 <!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->

.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.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/architecture/2026-07-05-reconstructable-requests.md
-2026-07-05-reconstructable-requests.md: 153d37a2faf2265134d5ff9e88f0bbfa275328e0
-2026-07-05-reconstructable-requests.zh.md: 25ce6bfdb56db80c92e8293206484a2de505d662
+2026-07-05-reconstructable-requests.md: 2f559a3052b9fb84f788975a64799e4f020b0d3e
+2026-07-05-reconstructable-requests.zh.md: 26abdc024a166856e51ebf09f086c7868fc8236d

.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md

@@ -22,11 +22,11 @@ Prefix-cache stability is corollary #1, not the headline: an append-only log pro
 
 **Messages.** `Session.deriveMessages()` is cached: each surface entry is projected exactly once, when first seen, through the public per-event function `deriveEventMessage(event)`; a surface rewrite (a compaction `replace` — `SurfaceManager.replaceGeneration`) rebuilds. Callers get a fresh array per call over shared, deep-frozen messages: mutating logged history through a projection is unrepresentable (it throws), replacing the old clone-per-call isolation. External reconstructors fold the same public function over a log prefix, so no two paths can disagree.
 
-`EpochHeader` records the request's non-history state: call config, rendered system prompt, tool schemas, and session prefix, with empty values canonicalized to absence. `request/header` always writes a full snapshot: the first loop instance uses reason `initial`, later instances use `resume`, and an in-instance change uses `change`. `foldRequestHeader` selects the latest snapshot. Legacy `request/header-delta` events and the removed `fallback` reason are rejected when appended or loaded.
+`EpochHeader` records the request's non-history state: call config, rendered system prompt, and tool schemas, with empty values canonicalized to absence. `request/header` always writes a full snapshot: the first loop instance uses reason `initial`, later instances use `resume`, and an in-instance change uses `change`. `foldRequestHeader` selects the latest snapshot. Legacy `request/header-delta` events and the removed `fallback` reason are rejected when appended or loaded.
 
-Each step rebuilds prompt assembly. On the instance's first step, `agent/session-prefix` extends a frozen empty seed with request-only opener messages; the result is frozen and cached for that loop instance before the generic `agent/pre-step` checkpoint and boundary snapshot. The first call config starts from explicit `AgentOptions`, preserving fork overrides and resume reconfiguration; later calls start from the folded header. `agent/request` may replace only that frozen config seed, while model-visible content enters through logged channels. The loop records the owed header event—the prefix's only durable home—builds `GenerateOptions` from prefix, snapshot, and header, and deep-freezes it while leaving `AbortSignal` live. Per-instance state is only the cached prefix and whether its anchoring snapshot has been written.
+Each proposed step first claims its inbox batch and runs `agent/pre-step`. Rejection opens no step; enter opens `step/start` and records the final message batch as `user/message` events. The step then assembles the system prompt and tools, while `agent/request` may replace only the frozen call-config seed. The loop records the owed full header snapshot, builds `GenerateOptions` from derived messages and that header, and deep-freezes it while leaving `AbortSignal` live. The first call config starts from explicit `AgentOptions`, preserving fork overrides and resume reconfiguration; later calls start from the folded header.
 
-**`step/start` is the reconstruction boundary.** A step derives messages from events before that sequence. Injection after the snapshot joins the next request, and reentrant appends are rejected during event publication. `agent/pre-step(agent, turn, step, signal)` remains the generic seam for content needed by the current request. Header reconstruction selects the step's `request/header`, or carries the prior snapshot when no new header is written.
+**The open step is the reconstruction boundary.** Its entered `user/message` batch and any newly written `request/header` precede request dispatch. Injection after the atomic claim joins a later request, while a listener that must affect this request returns messages through `agent/pre-step`. Header reconstruction selects the step's `request/header`, or carries the prior snapshot when no new header is written.
 
 **Enforcement.** The `dsh-agent-loop/invariant` companion registers with `ctx.invariants` and, when selected, independently rebuilds each loop request through a fresh `Session`, so the live cache cannot vouch for itself, then compares messages and folded header fields at `llm/stream`. The loop records the exact frozen request through `markAgentLoopRequest()` in `dsh-llm`; the process-local identity lets the companion and other request observers recognize conversation work, while direct one-shots remain excluded regardless of their frozen shape or session id. Correctness depends on sequence-bounded reconstruction rather than listener order. A with-key e2e requires positive cache-read tokens after the first request; per-step usage is the production signal, and a header change or compaction appears as a cache-read drop on the next step.
 
@@ -47,9 +47,9 @@ Like MiniCode, the conversation advances append-only and resets only when model-
 ## Consequences
 
 - A request that is not explained by the log cannot be constructed by accident — not by the loop, not by a listener; mutating a built request throws; every header change is a durable, diffable log event.
-- Choosing between the advisory channels is a change-frequency decision, and the design makes the stable one structural: an `agent/session-prefix` contribution is composed once per loop instance and reused verbatim, so it extends the cacheable prefix at zero marginal cost and CANNOT bust the provider cache mid-session; content that changes mid-session flows through the append-only history channels — `agent.inject()` and tool/prompt-submit `additionalContexts` — each a durable `context/message` paid once and prefix-cached thereafter, at the price of accumulating in history and the log. Route session-frozen openers to the prefix and change notices to the history channels; a per-step request-only tail slot was deliberately dropped (no consumer, and a durable append covers every current update pattern).
+- Model-visible context uses logged message channels. `agent.inject()` and tool `additionalContexts` enter the inbox for a later claim, while `agent/pre-step` returns context that must settle with the current claimed batch. Each entered value is a durable sourced `user/message`, paid once and prefix-cached thereafter at the price of accumulating in history until compaction.
 - What still costs full price at the provider is inherent and logged: compaction (its `compact/*` events and replacement entry), a real prompt, tool, or config change (`request/header` with reason `change`), or a process boundary with drift (a differing `resume` snapshot). The provider's own reasoning-content exclusion is managed server-side.
-- The `step/start`-listener behavior change (above) is the one observable semantics change for plugins; `agent/pre-step` is the current-request seam.
+- `agent/pre-step` is the current-request message seam; direct inbox mutation is the eventual later-request seam.
 - Tool-result trimming (planned) needs no new mechanism: a logged single-entry surface replace (`start === end`) carrying a trimmed `tool/result` under the same `callId` — compaction-family, replay-correct, cache-bust batched by the same pressure logic.
 - Session logs grow one `request/header` snapshot per loop instance plus snapshots on real changes. This is larger than a delta codec but small beside chunk-heavy logs and retains one replay representation. `SESSION_FORMAT_VERSION` stays `0`; legacy delta events are rejected rather than migrated.
 - Snapshot expected outputs changed once (every transcript gains its header events); the fs-writing fixtures are stored in the normalized authored form with cwd-relative tool arguments, because replay only round-trips cwd-independent argument paths.

.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.zh.md

@@ -22,11 +22,11 @@ Status: implemented
 
 **消息。** `Session.deriveMessages()` 带缓存：每个 surface 条目在首次出现时通过公开的逐事件函数 `deriveEventMessage(event)` 精确投影一次；surface 重写（压缩的 `replace`，即 `SurfaceManager.replaceGeneration`）触发重建。调用方每次获得一个新数组，底层是共享的深度冻结消息：通过投影变异已记录的历史是不可表达的（会抛异常），取代了旧的逐次调用克隆隔离。外部重建器对日志前缀折叠同一个公开函数，因此不可能有两条路径产生分歧。
 
-`EpochHeader` 记录请求的非历史状态：调用配置、渲染后的系统提示词、工具 schema 和会话前缀，空值规范化为缺失。`request/header` 始终写入完整快照：首个循环实例使用 reason `initial`，后续实例使用 `resume`，实例内变更使用 `change`。`foldRequestHeader` 选择最新快照。旧的 `request/header-delta` 事件和已移除的 `fallback` reason 在追加或加载时都会被拒绝。
+`EpochHeader` 记录请求的非历史状态：调用配置、渲染后的系统提示词和工具 schema，空值规范化为缺失。`request/header` 始终写入完整快照：首个循环实例使用 reason `initial`，后续实例使用 `resume`，实例内变更使用 `change`。`foldRequestHeader` 选择最新快照。旧的 `request/header-delta` 事件和已移除的 `fallback` reason 在追加或加载时都会被拒绝。
 
-每个步骤重建提示词组装。在实例的首个步骤中，`agent/session-prefix` 以一个冻结的空种子为基础，用仅限请求的开场消息进行扩展；结果在通用 `agent/pre-step` 检查点与边界快照之前被冻结并缓存于该循环实例。首次调用配置从显式的 `AgentOptions` 出发，保留 fork 覆盖和恢复重配置；后续调用从折叠后的 header 出发。`agent/request` 只能替换那个冻结的配置种子，模型可见内容通过已记录的通道进入。循环记录应写入的 header 事件（前缀唯一的持久归宿），从前缀、快照和 header 构建 `GenerateOptions`，对其深度冻结但保持 `AbortSignal` 活跃。每实例状态仅有缓存的前缀和锚定快照是否已写入。
+每个拟议步骤先领取其 inbox 批次，再运行 `agent/pre-step`。reject 不打开步骤；enter 打开 `step/start`，并把最终消息批次记录为 `user/message` 事件。随后步骤组装系统提示词与工具，`agent/request` 只能替换冻结的调用配置种子。循环记录所需的完整 header 快照，从派生消息与该 header 构建 `GenerateOptions`，对其深度冻结但保持 `AbortSignal` 活跃。首次调用配置从显式的 `AgentOptions` 出发，保留 fork 覆盖和恢复重配置；后续调用从折叠后的 header 出发。
 
-**`step/start` 是重建边界。** 一个步骤从该序列之前的事件推导消息。快照之后的注入加入下一次请求，事件发布期间的重入追加被拒绝。`agent/pre-step(agent, turn, step, signal)` 仍是当前请求所需内容的通用 seam。header 重建选择该步骤的 `request/header`，或在无新 header 写入时沿用前一个快照。
+**已打开步骤是重建边界。** 进入步骤的 `user/message` 批次与任何新写入的 `request/header` 都位于请求分派之前。原子领取后发生的注入加入后续请求；必须影响本次请求的监听器则通过 `agent/pre-step` 返回消息。header 重建选择该步骤的 `request/header`，或在无新 header 写入时沿用前一个快照。
 
 **强制执行。** `dsh-agent-loop/invariant` 配套插件向 `ctx.invariants` 注册，并在被选用时通过一个全新的 `Session` 独立重建每个循环请求，使活跃缓存无法为自身背书，然后在 `llm/stream` 处比较消息和折叠后的 header 字段。循环通过 `dsh-llm` 的 `markAgentLoopRequest()` 记录精确的冻结请求；这一进程内标识让配套插件和其他请求观察者识别对话工作，而直接的一次性调用无论其冻结形状或会话 id 如何都保持排除。正确性依赖于序列有界的重建，而非监听器顺序。带密钥的 e2e 要求首次请求之后有正值的 cache-read token；逐步骤用量是生产信号，header 变更或压缩表现为下一步骤的 cache-read 下降。
 
@@ -47,10 +47,10 @@ Status: implemented
 ## 后果
 
 - 一个日志无法解释的请求不可能被意外构造——无论是循环还是监听器；变异已构建的请求会抛异常；每个 header 变更都是持久的、可 diff 的日志事件。
-- 在建议性通道之间的选择取决于内容的变更频率，而本设计将稳定通道固化在结构中：`agent/session-prefix` 的贡献在每个循环实例中只组合一次并逐字复用，因此以零边际成本扩展可缓存前缀，且不可能在会话中途使提供方缓存失效；会话中途变化的内容通过仅追加的历史通道流入——`agent.inject()` 以及工具/prompt-submit 的 `additionalContexts`——每条都是持久的 `context/message`，付出一次代价后即被前缀缓存，代价是在历史和日志中累积。将在会话期间固定不变的开场内容路由到前缀，将变更通知路由到历史通道；逐步骤的仅限请求尾部槽位被有意放弃（无消费方，且持久追加覆盖了当前所有更新模式）。
-- 在提供方处仍需全价计算的内容是固有的且已记录的：压缩（其 `compact/*` 事件和替换条目）、真正的提示词、工具或配置变更（reason 为 `change` 的 `request/header`），或带漂移的进程边界（不同的 `resume` 快照）。提供方自身对思考内容的排除由服务端管理。
-- `step/start` 监听器行为变更（见上文）是对插件唯一可观察的语义变更；`agent/pre-step` 是当前请求的 seam。
-- 工具结果裁剪（计划中）无需新机制：一个已记录的单条目 surface replace（`start === end`），携带同一 `callId` 下裁剪后的 `tool/result`——属压缩家族，回放正确，缓存失效由相同的压力逻辑批量处理。
+- 模型可见上下文使用已记录消息通道。`agent.inject()` 与工具 `additionalContexts` 进入 inbox，等待后续领取；必须与当前已领取批次一起结算的上下文由 `agent/pre-step` 返回。每个进入步骤的值都是带来源的持久 `user/message`，只付出一次代价并在后续成为可缓存前缀，代价是会在历史中累积直至压缩。
+- 在提供方处仍需全价计算的内容是固有的且已记录的：压缩（其 `compact/*` 事件和替换条目）、真正的提示词、工具或配置变更（reason 为 `change` 的 `request/header`），或带漂移的进程边界（不同的 `resume` 快照）。提供方自身的 reasoning-content 排除由服务端管理。
+- `agent/pre-step` 是当前请求的消息 seam；直接修改 inbox 则是最终进入后续请求的 seam。
+- 工具结果裁剪（计划中）无需新机制：一个已记录的单条目 surface replace（`start === end`），携带同一 `callId` 下裁剪后的 `tool/result`——属压缩家族，回放正确，缓存击穿由相同的压力逻辑批量处理。
 - 会话日志每个循环实例增长一个 `request/header` 快照，并在真正变更时增加快照。它比 delta 编解码器更大，但相对分片密集型日志仍然很小，并只保留一种回放表示。`SESSION_FORMAT_VERSION` 保持 `0`；旧的 delta 事件被拒绝而非迁移。
 - 快照预期输出变更一次（每个 transcript（文本记录）增加其 header 事件）；写入文件系统的 fixture（测试前置数据）以规范化的撰写形式存储，工具参数使用 cwd 相对路径，因为回放只对 cwd 无关的参数路径做往返。
 - FIXME(call-config-shape)：重新审视 `LlmCallConfig` 的确切字段集——哪些字段对缓存而言真正属于 epoch 级别（`model` 毫无疑问；采样标量出于谨慎放在那里），以及当适配器需要时，提供方特定的额外项（推理（reasoning）选项、额外 body 参数）应归属何处。

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.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/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md
-2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md: 51d488db28c57426c75c9ed1cfc90892261c0224
-2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md: 1aa827bb0263ed9d104d566a7b0b0c8006ad7586
+2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md: 6fbd5e2c9d57da3f25c72c652ca50eb45b84323c
+2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md: c46b3eeeb446165192ef44de938c9a5a657a02f8

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md

@@ -12,17 +12,17 @@ Successful calls are not the only pressure signal. A provider can reject a reque
 
 ## Decision
 
-### Successful pressure moves to a durable post-step checkpoint
+### Successful pressure runs at the next pre-step boundary
 
-`agent/pre-step` is narrowed to `(agent, turn, step, signal)`. It remains a generic serial checkpoint before `step/start`, but it carries no compaction-only prompt or prefix fields.
+`agent/pre-step` receives the exclusive claimed message batch plus `{ turn, step, signal }` and returns the final reject/enter decision. It carries no compaction-only prompt or prefix fields.
 
-The loop fires awaited serial `agent/post-step(agent, turn, step, signal)` after assistant output, every dispatched or synthetic tool result, post-tool context, and steering are durable, but before `step/end`. This placement gives pressure policy the complete successful-call state without splitting an assistant tool call from its result. A propagated listener failure is an ordinary turn failure; it never enters model-request recovery. Compact-basic contains its expected operational failures as described below.
+Compact-basic wraps `agent/pre-step` before each proposed request. At a continuation boundary the preceding assistant output, every dispatched or synthetic tool result, post-tool context, and steering are already durable, so pressure policy sees the complete successful-call state without splitting an assistant tool call from its result. At the initial boundary a headerless session has no completed routed request and produces no pressure work. Compact-basic contains operational failures, warns, and delegates without rejecting the proposed step.
 
 `dsh-compact-basic` reads the exact latest routed model from the durable request header only to establish that a completed route exists, then asks the singleton `ctx.tokenMeter` to measure the canonical logged envelope and current surface. It does not fall back to `AgentOptions.model` for automatic pressure. A headerless session has no completed routed request to assess and produces no work; any durable non-empty model name uses the same estimator. Operational measurement or summarization failures warn and continue from the latest durable surface: full history before any replacement, or the pruned surface if pruning already landed.
 
 ### Request recovery is limited to the final model boundary
 
-`RequestError` and the `agent/request-error` waterfall represent failures after the final adapter has been selected. Each returned stream handle owns a private failure set that preserves the original thrown error identity across dispatch, iterator construction, and iteration without leaking nested-call provenance into an outer call. Terminal in-band `error` or `aborted` finishes enter the same path. Prompt assembly, request middleware, request logging, result processing, tools, step listeners, and cleanup remain ordinary failures.
+`agent/request-error` represents terminal failures from the final adapter boundary. Adapter selection, dispatch, iterator construction, and iteration throws become terminal `error` or `aborted` finishes before the agent loop consumes them; adapter-emitted terminal finishes enter the same path. Prompt assembly, request middleware, request logging, result processing, tools, step listeners, and cleanup remain ordinary failures. [Terminal LLM stream failures](2026-07-29-terminal-llm-stream-failures.md) owns this normalization boundary.
 
 The failed step closes before recovery runs. A handling listener repairs durable state, returns `{ kind: 'retry' }`, and stops waterfall delegation. The loop then closes the failed turn and opens one retry turn from the durable log without an intervening idle notification. Retry policy and attempt counts remain plugin-owned; compact-basic clears its per-agent overflow count when the chain reaches terminal `agent/settled`. Both DeepSeek adapters normalize recognized provider context-limit failures to `CONTEXT_WINDOW_EXCEEDED`. The [retry-action decision](../simplification/2026-07-27-request-error-retry-action.md) owns the return boundary.
 
@@ -42,11 +42,11 @@ The default summarizer resolves explicit configuration, then the latest logged r
 
 ## Testing
 
-Unit tests cover final-adapter failure provenance and identity, closed-turn retry numbering and reset, cancellation and disposal, step-boundary ordering, routed-envelope pressure, pressure-gated pruning, pruning-only relief, pruned-input summarization, balanced overflow reduction, durable prune progress before later failure, generation proof, caps, delegation, and auxiliary-call routing. Real-loop tests cover thrown and in-band overflow through pruning or summary compaction to a reconstructed retry request.
+Unit tests cover the final-adapter normalization boundary, closed-turn retry numbering and reset, cancellation and disposal, step-boundary ordering, routed-envelope pressure, pressure-gated pruning, pruning-only relief, pruned-input summarization, balanced overflow reduction, durable prune progress before later failure, generation proof, caps, delegation, and auxiliary-call routing. Real-loop tests cover thrown and in-band overflow through pruning or summary compaction to a reconstructed retry request.
 
 ## Alternatives considered
 
-- **Keep provisional pre-step pressure and add more arguments** — rejected because later routing and request mutation remain outside any earlier snapshot, while generic lifecycle becomes coupled to one plugin.
+- **Add compaction-only fields to pre-step** — rejected because the canonical durable session and token meter already own the measurement input; the generic lifecycle need not carry a second envelope.
 - **Retry the same numbered step** — rejected because recovery appends durable events after the failed boundary. A new step preserves balanced nesting and reconstructability.
 - **Retry whenever `compactIfNeeded` returns a result** — rejected because a custom backend can report success without changing model-visible state. `replaceGeneration` is the authoritative proof.
 - **Let compact-basic parse provider wording** — rejected because classification belongs at adapters and must cover both thrown and in-band delivery.
@@ -54,8 +54,8 @@ Unit tests cover final-adapter failure provenance and identity, closed-turn retr
 
 ## Consequences
 
-Post-step pressure describes the completed routed request, including durable tool results and request-only prefix fields. Optional model-free pruning removes predictable tool-output bulk before summary selection and can independently create retry-worthy progress. Canonical overflow supplies the backstop when no successful usage anchor exists. Recovery is bounded, cancellation-owned, and monotonic: it retries only after a visible surface generation change.
+The next pre-step pressure check describes the preceding completed routed request, including durable tool results and newly claimed input. Optional model-free pruning removes predictable tool-output bulk before summary selection and can independently create retry-worthy progress. Canonical overflow supplies the backstop when no successful usage anchor exists. Recovery is bounded, cancellation-owned, and monotonic: it retries only after a visible surface generation change.
 
-The cost is one additional serial checkpoint on successful steps and adapter-maintained overflow classification. Provider wording and heuristic character density remain maintenance risks. Surface compaction still cannot repair an envelope that alone exceeds the window, split an indivisible non-tool node, or repair a tool unit whose non-prunable remainder remains oversized. The optional pruner can repair an otherwise indivisible tool pair when removable text-bearing tool-result content is the bulk.
+The cost is pressure work in the shared pre-step waterfall and adapter-maintained overflow classification. Provider wording and heuristic character density remain maintenance risks. Surface compaction still cannot repair an envelope that alone exceeds the window, split an indivisible non-tool node, or repair a tool unit whose non-prunable remainder remains oversized. The optional pruner can repair an otherwise indivisible tool pair when removable text-bearing tool-result content is the bulk.
 
-This Agent Note supersedes only the pre-step automatic-trigger portion of the [compaction capability-seam Agent Note](../feature/2026-06-18-compaction-capability-seam.md). The service split, standalone token meter, balanced range contract, log-recorded lock, summary replacement, and sole `summarize()` subclass hook remain unchanged.
+The [claimed pre-step lifecycle](2026-07-31-claimed-pre-step-inbox-lifecycle.md) supersedes this note's former post-step trigger. The service split, standalone token meter, balanced range contract, log-recorded lock, summary replacement, and sole `summarize()` subclass hook remain unchanged.

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md

@@ -12,17 +12,17 @@ Status: implemented
 
 ## 决策
 
-### 成功压力移动到持久 post-step 检查点
+### 成功压力在下一个 pre-step 边界运行
 
-`agent/pre-step` 收窄为 `(agent, turn, step, signal)`。它仍是 `step/start` 之前的通用串行检查点，但不再携带压缩专用的提示词或前缀字段。
+`agent/pre-step` 接收独占的已领取消息批次与 `{ turn, step, signal }`，并返回最终 reject/enter 决策。它不携带压缩专用的提示词或前缀字段。
 
-循环在 assistant 输出、所有已分发或合成的工具结果、工具后上下文与 steering 都持久化之后、`step/end` 之前，触发等待式串行 `agent/post-step(agent, turn, step, signal)`。该位置让压力策略看到完整的成功调用状态，同时不会拆开 assistant 工具调用与其结果。向外传播的监听器失败属于普通 turn 失败，绝不会进入模型请求恢复；compact-basic 会按下文所述在内部处理其预期的操作性失败。
+Compact-basic 会在每个拟议请求之前包装 `agent/pre-step`。在续步边界，前一条 assistant 输出、所有已分发或合成的工具结果、工具后上下文与 steering 都已经持久化，因此压力策略能看到完整的成功调用状态，同时不会拆开 assistant 工具调用与其结果。初始边界上的无 header 会话尚无已完成路由请求，因此不执行压力工作。Compact-basic 会在内部处理操作性失败、发出警告并继续委托，不会 reject 拟议步骤。
 
 `dsh-compact-basic` 从持久请求头读取精确的最新实际路由模型，只用它确认已经存在已完成的路由，随后让单例 `ctx.tokenMeter` 计量规范日志信封与当前表层。自动压力不会回退到 `AgentOptions.model`。没有请求头的会话尚无已完成路由请求可供判断，因此不执行工作；任意持久记录的非空模型名都使用同一个估算器。操作性的计量或摘要失败会发出警告，并从最新持久表层继续：任何替换发生前使用完整历史；若剪枝已经落盘，则使用已剪枝表层。
 
 ### 请求恢复只覆盖最终模型边界
 
-`RequestError` 与 `agent/request-error` waterfall（瀑布式事件）表示最终适配器已经选定之后的失败。每个返回的流句柄都绑定一个私有失败集合；该集合在分发、异步迭代器构造与迭代过程中保留原始抛出错误的身份，同时防止把嵌套调用的错误来源误归到外层调用。终止性的带内 `error` 或 `aborted` finish 进入同一路径。提示词装配、请求中间件、请求日志、结果处理、工具、step 监听器与清理仍属于普通失败。
+`agent/request-error` 表示来自最终适配器边界的终止失败。适配器选择、分发、iterator 构造与迭代抛出会在 agent loop 消费前成为终止 `error` 或 `aborted` finish；适配器直接发出的终止 finish 进入同一路径。提示词装配、请求 middleware、请求日志、结果处理、工具、step 监听器与清理仍属于普通失败。[LLM 流的终止失败](2026-07-29-terminal-llm-stream-failures.md)规定这一规范化边界。
 
 恢复运行前，失败 step 已经关闭。负责处理的监听器修复持久状态、返回 `{ kind: 'retry' }`，并停止 waterfall 委托。循环随后关闭失败 turn，并从持久日志开启一个重试 turn，中间不发布空闲通知。重试策略与尝试计数由插件自己拥有；compact-basic 在链路到达终态 `agent/settled` 时清除对应 agent 的溢出计数。两个 DeepSeek 适配器都把识别出的提供方上下文限制错误规范化为 `CONTEXT_WINDOW_EXCEEDED`。[重试动作决策](../simplification/2026-07-27-request-error-retry-action.md)规定这一返回边界。
 
@@ -42,11 +42,11 @@ Status: implemented
 
 ## 测试
 
-单元测试覆盖最终适配器失败的来源与身份、已关闭 turn 的重试编号与重置、取消与销毁、step 边界顺序、已路由信封压力、压力门控剪枝、剪枝独立解除压力、从已剪枝输入生成摘要、平衡溢出缩减、后续失败前已落盘的剪枝进展、generation 证明、上限、委托与辅助调用路由。真实循环测试覆盖抛出式和带内溢出经剪枝或摘要压缩后重建重试请求的过程。
+单元测试覆盖最终适配器规范化边界、已关闭 turn 的重试编号与重置、取消与销毁、step 边界顺序、已路由信封压力、压力门控剪枝、剪枝独立解除压力、从已剪枝输入生成摘要、平衡溢出缩减、后续失败前已落盘的剪枝进展、generation 证明、上限、委托与辅助调用路由。真实循环测试覆盖抛出式和带内溢出，并验证剪枝或摘要压缩后的重试请求从替换表层重建。
 
 ## 考虑过的替代方案
 
-- **保留临时 pre-step 压力并增加更多参数**——不予采纳，因为后续路由与请求变换仍在更早快照之外，同时通用生命周期会耦合到单个插件。
+- **向 pre-step 增加压缩专用字段**——不予采纳，因为规范持久会话与 token meter 已拥有计量输入；通用生命周期不需要携带第二份信封。
 - **重试相同编号的 step**——不予采纳，因为恢复会在失败边界之后追加持久事件。新 step 保持边界配对与可重建性。
 - **只要 `compactIfNeeded` 返回结果就重试**——不予采纳，因为自定义后端可能报告成功却没有改变模型可见状态。`replaceGeneration` 才是权威证明。
 - **让 compact-basic 解析提供方措辞**——不予采纳，因为分类属于适配器，而且必须同时覆盖抛出式与带内交付。
@@ -54,8 +54,8 @@ Status: implemented
 
 ## 后果
 
-Post-step 压力描述已完成的路由请求，包括持久工具结果与仅存在于请求中的前缀字段。可选的无模型剪枝会在选择摘要前移除可预测的工具输出体积，也能独立产生足以重试的进展。当成功 usage 锚点不存在时，规范化溢出提供兜底路径。恢复有明确上限、以取消为准，并保持单调：只有模型可见的表层 generation 变化后才重试。
+下一个 pre-step 的压力检查描述前一个已完成的路由请求，包括持久工具结果与新领取输入。可选的无模型剪枝会在选择摘要前移除可预测的工具输出体积，也能独立产生足以重试的进展。当成功 usage 锚点不存在时，规范化溢出提供兜底路径。恢复有明确上限、以取消为准，并保持单调：只有模型可见的表层 generation 变化后才重试。
 
-代价是成功 step 增加一个串行检查点，并需要适配器持续维护溢出分类。提供方措辞与启发式字符密度仍是维护风险。表层压缩依然无法修复仅信封本身就超出窗口的情况，也不能拆分不可分割的非工具节点，或修复非可剪枝剩余部分仍然过大的工具单元。若可移除的文本工具结果是主要体积，可选剪枝器仍可修复原本不可分割的工具配对。
+代价是在共享 pre-step waterfall 中执行压力工作，并需要适配器持续维护溢出分类。提供方措辞与启发式字符密度仍是维护风险。表层压缩依然无法修复仅信封本身就超出窗口的情况，也不能拆分不可分割的非工具节点，或修复非可剪枝剩余部分仍然过大的工具单元。若可移除的文本工具结果是主要体积，可选剪枝器仍可修复原本不可分割的工具配对。
 
-本 Agent Note 只取代[压缩能力 seam Agent Note](../feature/2026-06-18-compaction-capability-seam.md) 中的 pre-step 自动触发部分。服务拆分、独立 token meter、平衡范围契约、日志记录锁、摘要替换与唯一 `summarize()` 子类 hook 均保持不变。
+[已领取 pre-step 生命周期](2026-07-31-claimed-pre-step-inbox-lifecycle.md)取代了本记录原先的 post-step 触发方式。服务拆分、独立 token meter、平衡范围契约、日志记录锁、摘要替换与唯一 `summarize()` 子类 hook 均保持不变。

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.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/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md
-2026-07-10-single-file-executable-sdk-runtime-distribution.md: f749d6a72b4c32a189a9f848595076457819d9b9
-2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md: c7a915fc6542d212fa2c1db99ca38710c6862932
+2026-07-10-single-file-executable-sdk-runtime-distribution.md: f3da981c478ef08672a82f23ab9cd42e0f38ebab
+2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md: 5a8e15aab79875bdb08e6347199924c4215a2705

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md

@@ -27,7 +27,7 @@ Terminology reminder: pkg's `/snapshot` VFS has nothing to do with this repo's t
 
 The deterministic protocol implementation (`server.ts` / `transport.ts`) lands as two packages on the existing `acp/acp` + `examples/acp-demo` pattern — the serving surface is itself a plugin:
 
-- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md) (`@deepseek-ai/dsh-jsonrpc`): the pure protocol plugin; on apply it mounts `HarnessSdkServer` plus a line-delimited JSON-RPC transport on the process stdio, with disposal through `ctx.effect()`. Whether to serve is decided by `cordis.yml`; a yml that does not mount it is a legitimate process that does not serve. Protocol-level exit belongs to the plugin (after answering the `shutdown` request it disposes its own fiber, then `exit(0)`; an HMR-style unload only stops the service without exiting the process).
+- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md) (`@deepseek-ai/dsh-jsonrpc`): the pure protocol plugin; on apply it mounts `HarnessSdkServer` plus a line-delimited JSON-RPC transport on the process stdio, with disposal through `ctx.effect()`. Whether to serve is decided by `cordis.yml`; a yml that does not mount it is a legitimate process that does not serve. Protocol-level exit belongs to the plugin (after answering and flushing the `shutdown` response it disposes the root runtime so persistence drains, then `exit(0)`; an HMR-style unload only stops the service without exiting the process).
 - [`packages/examples/jsonrpc-demo`](../../../../packages/examples/jsonrpc-demo/README.md) (`@deepseek-ai/dsh-jsonrpc-demo`): a thin app bin — `installFailLoud` + `loadEnv` + config discovery + `boot()` from [`dsh-app-boot`](../../../../packages/ui/app-boot/src/index.ts), done once boot completes; the server is brought up by the `dsh-jsonrpc` entry in the yml. Its only dependency is app-boot. Process-level exit belongs to the bin (stdin EOF/SIGTERM → dispose then 0, SIGINT → 130).
 
 Config discovery has two channels and fails loudly when both are missing: the `DSH_CORDIS_CONFIG` environment variable first (the SDK client convention), then an argv positional argument; no default path and no built-in fallback whatsoever — "the plugins actually booted are decided by an external cordis.yml" is a hard semantic.

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md

@@ -27,7 +27,7 @@ exe 使用 [@yao-pkg/pkg](https://github.com/yao-pkg/pkg)（vercel/pkg 归档后
 
 确定性协议实现（`server.ts` / `transport.ts`）按 `acp/acp` + `examples/acp-demo` 的既有模式落为两包——对外服务接口本身也是插件：
 
-- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md)（`@deepseek-ai/dsh-jsonrpc`）：纯协议插件；执行 `apply` 时，在进程 stdio 上挂载 `HarnessSdkServer` 与按行传输的 JSON-RPC 层，资源释放走 `ctx.effect()`。是否提供服务由 `cordis.yml` 决定；未挂载该插件的配置会启动一个不提供此服务的合法进程。协议级退出归插件所有（应答 `shutdown` 请求后 dispose（资源释放）自身 fiber，再调用 `exit(0)`；HMR（热模块替换）式卸载只停止服务，不退出进程）。
+- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md)（`@deepseek-ai/dsh-jsonrpc`）：纯协议插件；执行 `apply` 时，在进程 stdio 上挂载 `HarnessSdkServer` 与按行传输的 JSON-RPC 层，资源释放走 `ctx.effect()`。是否提供服务由 `cordis.yml` 决定；未挂载该插件的配置会启动一个不提供此服务的合法进程。协议级退出归插件所有（应答并刷新 `shutdown` 响应后 dispose 根运行时以排空持久化，再调用 `exit(0)`；HMR 式卸载只停止服务，不退出进程）。
 - [`packages/examples/jsonrpc-demo`](../../../../packages/examples/jsonrpc-demo/README.md)（`@deepseek-ai/dsh-jsonrpc-demo`）：轻量应用入口——`installFailLoud` + `loadEnv` + 配置发现 + [`dsh-app-boot`](../../../../packages/ui/app-boot/src/index.ts) 的 `boot()`；`boot()` 完成后入口即完成，服务器由 `cordis.yml` 中的 `dsh-jsonrpc` 条目启动。它只依赖 `app-boot`。进程级退出归 `bin` 所有（stdin EOF/SIGTERM → dispose 后返回 0，SIGINT → 130）。
 
 配置发现有两个通道，均缺失时立即报错：优先使用 `DSH_CORDIS_CONFIG` 环境变量（SDK 客户端约定），其次使用 argv 位置参数；没有默认路径或内置回退——「实际启动的插件由外部 `cordis.yml` 决定」是硬语义。

.agents/notes/implemented/architecture/2026-07-12-agent-scope-runtime-design.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/architecture/2026-07-12-agent-scope-runtime-design.md
-2026-07-12-agent-scope-runtime-design.md: d6b865977a76061784c88dbad089fa5963be8c7e
-2026-07-12-agent-scope-runtime-design.zh.md: 5b6b9b1ca582d83a267a30f8f76e38ea87d52e9b
+2026-07-12-agent-scope-runtime-design.md: b93c291957fab98ab9d0d04eb816bb01a15a0b51
+2026-07-12-agent-scope-runtime-design.zh.md: 8b1332c23ce3b6c0332fcde40b4f4ac8dd20aa45

.agents/notes/implemented/architecture/2026-07-12-agent-scope-runtime-design.md

@@ -150,7 +150,7 @@ sequenceDiagram
 Every teardown request joins one memoized path. The order is:
 
 1. Deactivate creation or driving and let synchronous publication finish.
-2. Stop and drain the driver, including idle injection flushes.
+2. Stop and drain the driver, discarding any injection that remains pending.
 3. Detach the agent.
 4. Detach the session.
 5. Dispose the agent scope.
@@ -238,7 +238,7 @@ For a native call, the observer deletes the stage and commits its value only whe
 
 For a Code Mode SDK call, the inner successful result records `{ parentToken, value }` rather than committing. The observer waits for the `run_code` execution whose token matches `parentToken` and commits only if that outer final result also succeeds. Program failure, runtime abort, or outer post-policy denial discards the pending value.
 
-Once a value is pending or committed, a scoped monotonic guard denies later tool calls. After commit, the ordinary serial `agent/turn-stop` listener returns a stop decision after continuation and steering have already folded. A schema-validation failure remains an ordinary `INVALID_ARGS` tool error and leaves the child able to retry within the same turn.
+Once a value is pending or committed, a scoped monotonic guard denies later tool calls. The successful structured-output execution calls `exec.concludeTurn()`, so its own immutable result carries `concludesTurn: true` and the loop ends the tool loop at that step. A schema-validation failure remains an ordinary `INVALID_ARGS` tool error and leaves the child able to retry within the same turn.
 
 Pure Code Mode's registry contribution omits `structured_output` from native wire schemas and exposes it through the generated SDK. The assembly waterfall may deliberately change that presentation; execution still validates against the child-scoped definition, and the listener owns the consistency of any alternate model-visible route it creates.
 
@@ -250,9 +250,9 @@ Prompt assembly is intentionally cooperative, but three execution facts need one
 |---|---|---|
 | Tool pre-policy | Deny monotonically | A later listener must not re-allow an already denied call |
 | Tool result | Observe the immutable committed outcome | Structured output must commit only the result that actually escaped the pipeline |
-| Turn continuation | Stop after ordinary continuation folding | A committed terminal output must end the turn |
+| Turn continuation | Conclude through the committed tool result | A committed terminal output must end the turn |
 
-`ToolGuard` is the monotonic policy registry. Committed tool observation is the contained `tools/result` point described above. Terminal structured output listens on the ordinary serial `agent/turn-stop` fold after normal continuation and steering decisions; no public `strictSerial()` dispatcher is needed for the typed listener contract.
+`ToolGuard` is the monotonic policy registry. Committed tool observation is the contained `tools/result` point described above. Terminal structured output marks its own execution with `concludesTurn`, so terminality is data on the authoritative result rather than a separate hook decision.
 
 ### Skill and approval services trust typed callers
 

.agents/notes/implemented/architecture/2026-07-12-agent-scope-runtime-design.zh.md

@@ -150,7 +150,7 @@ sequenceDiagram
 每个拆除请求加入一条记忆化路径。顺序为：
 
 1. 停用创建或驱动，让同步发布完成。
-2. 停止并排空 driver，包括空闲注入刷新。
+2. 停止并排空 driver，丢弃仍处于待处理状态的注入。
 3. 分离 agent。
 4. 分离会话。
 5. Dispose agent 作用域。
@@ -238,7 +238,7 @@ Scope 直接解决了真正的隔离问题。结构化输出贡献注册在子
 
 对于 Code Mode SDK 调用，内层成功结果记录 `{ parentToken, value }` 而非提交。观察者等待 token 匹配 `parentToken` 的 `run_code` 执行，仅在该外层最终结果也成功时才提交。程序失败、运行时中止或外层 post-policy 拒绝会丢弃待定值。
 
-一旦值处于待定或已提交状态，作用域单调守卫拒绝后续工具调用。提交后，普通串行的 `agent/turn-stop` 监听器在 continuation 和 steering（中途引导）已折叠之后返回停止决策。Schema 验证失败仍然是普通的 `INVALID_ARGS` 工具错误，子级可以在同一轮次内重试。
+一旦值处于待定或已提交状态，作用域单调守卫拒绝后续工具调用。成功的结构化输出执行会调用 `exec.concludeTurn()`，因此其自身不可变结果携带 `concludesTurn: true`，循环在该步骤结束工具循环。Schema 验证失败仍然是普通的 `INVALID_ARGS` 工具错误，子级可以在同一轮次内重试。
 
 纯 Code Mode 的注册表贡献从原生 wire schema 中省略 `structured_output`，并通过生成的 SDK 暴露它。Assembly waterfall 可以有意改变该展示；执行仍然针对子作用域定义进行验证，监听器拥有其创建的任何替代模型可见路由的一致性。
 
@@ -250,9 +250,9 @@ Scope 直接解决了真正的隔离问题。结构化输出贡献注册在子
 |---|---|---|
 | 工具 pre-policy | 单调拒绝 | 后续监听器不得重新允许已被拒绝的调用 |
 | 工具结果 | 观察不可变的已提交结果 | 结构化输出必须仅提交实际逃出流水线的结果 |
-| 轮次 continuation | 在普通 continuation 折叠之后停止 | 已提交的终端输出必须结束轮次 |
+| 轮次 continuation | 通过已提交工具结果终止 | 已提交的终端输出必须结束轮次 |
 
-`ToolGuard` 是单调策略注册表。已提交的工具观察是上述被隔离的 `tools/result` 点。终端结构化输出监听普通串行的 `agent/turn-stop` 折叠，在正常 continuation 和 steering 决策之后；类型化的监听器契约不需要公开的 `strictSerial()` dispatcher。
+`ToolGuard` 是单调策略注册表。已提交的工具观察是上述被隔离的 `tools/result` 点。终端结构化输出在自身执行上标记 `concludesTurn`，因此终止性成为权威结果上的数据，而不是独立 hook 决策。
 
 ### Skill 和 approval 服务信任类型化调用方
 

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.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/architecture/2026-07-14-provider-routed-llm-adapters.md
-2026-07-14-provider-routed-llm-adapters.md: 1bd9197667f6e49c5025c98b4a77500f78595c2b
-2026-07-14-provider-routed-llm-adapters.zh.md: c91858a54ea20340b008099ae816e07a47ac6381
+2026-07-14-provider-routed-llm-adapters.md: 27277280e423553f79d5a34f512b673413f495ff
+2026-07-14-provider-routed-llm-adapters.zh.md: aeb09a500d5750ef2793bc9a7fc09834055a56a4

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.md

@@ -40,7 +40,7 @@ pi-ai's common stream options do not expose stop sequences. `dsh-llm-pi-ai` reje
 
 Assistant messages carry provider-neutral provenance containing the request's `provider` and `model`, plus an optional JSON-serializable adapter replay state. A successful `assistant/message` session event records this provenance and `deriveMessages()` returns it with the assistant message. User, system, context, and tool-result messages carry no assistant provenance. The provider/model fields are authoritative loop data; an adapter owns only its opaque replay-state payload.
 
-A terminal successful `finish` chunk may carry replay state, and `BlockAssembler` retains it alongside usage and finish reason. The loop attaches it to the assistant provenance only when the post-`agent/step-result` content is structurally equal to the assembled provider output. A listener that rewrites content keeps the provider/model provenance but loses the now-stale replay state. Error and aborted responses do not produce a normal assistant message and therefore do not enter future model history.
+A terminal successful `finish` chunk may carry replay state, and `BlockAssembler` retains it alongside usage and finish reason. The loop attaches that state to the assembled assistant provenance without exposing a response-rewrite hook. Error and aborted responses do not produce a normal assistant message and therefore do not enter future model history.
 
 The pi-ai replay state is a versioned, minimal projection of its successful `AssistantMessage`: source API/provider/model, response id/model, stop reason, and index-aligned text, thinking, and tool-call signatures. It does not duplicate text or tool arguments already carried by Harness content blocks, and it omits diagnostics, timestamps, usage, and errors. On a later request, `LlmService` gives replay state to the target adapter only when the historical provider and target provider are currently owned by the same adapter instance. That adapter combines the logged Harness content with replay state when it can restore the historical response, and owns any required cross-model or cross-provider conversion. An adapter receiving replay state with an unknown version or mismatched block shape fails explicitly; a different adapter receives only provider-neutral content and provenance.
 

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.zh.md

@@ -40,7 +40,7 @@ pi-ai 的通用流选项不支持停止序列。若 harness `stop` 选项已定
 
 助手消息携带提供方无关的来源信息，其中包含请求的 `provider` 和 `model`，以及可选的 JSON 可序列化适配器回放状态。成功的 `assistant/message` 会话事件记录这些来源信息，`deriveMessages()` 返回助手消息时也会包含这些信息。用户、system、context 与工具结果消息不携带助手来源信息。provider/model 字段是 agent loop 的权威数据；适配器仅拥有其不透明回放状态 payload。
 
-成功的终止 `finish` 分片可以携带回放状态，`BlockAssembler` 会将其与 token 用量和结束原因一起保留。只有当 `agent/step-result` 处理后的内容与提供方组装输出在结构上相等时，agent loop 才会把回放状态附加到助手来源信息。监听器重写内容后，provider/model 来源信息仍会保留，但已经陈旧的回放状态会被移除。错误或中止响应不会生成正常助手消息，因此不会进入后续模型历史。
+成功的终止 `finish` 分片可以携带回放状态，`BlockAssembler` 会将其与 token 用量和结束原因一起保留。agent loop 会把该状态附加到组装后的助手来源信息，不再暴露响应改写 hook。错误或中止响应不会生成正常助手消息，因此不会进入后续模型历史。
 
 pi-ai 回放状态是其成功 `AssistantMessage` 的带版本最小投影，包含源 API/provider/model、响应 ID/model、停止原因，以及按索引对齐的文本、thinking 和工具调用签名。它不会重复 harness 内容块中已有的文本或工具参数，也不包含诊断信息、时间戳、用量或错误。后续请求中，只有历史提供方和目标提供方当前归同一个适配器实例所有时，`LlmService` 才会把回放状态交给目标适配器。适配器在能够恢复历史响应时，将 harness 记录的内容与回放状态组合，并负责所需的跨模型或跨提供方转换。适配器收到未知版本或块形状不匹配的回放状态时会显式失败；其他适配器只能收到提供方无关的内容与来源信息。
 

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.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/architecture/2026-07-15-replay-token-meter-service.md
-2026-07-15-replay-token-meter-service.md: 3496364663c1f73b8161461d1a229b19d9730c6d
-2026-07-15-replay-token-meter-service.zh.md: 666422e4eefbed5268dfd7c8ddec7d1e79b8d4e3
+2026-07-15-replay-token-meter-service.md: c0f4b467ad0013dd4ac0a0301281b011ea8c261c
+2026-07-15-replay-token-meter-service.zh.md: c1d81dc0e76ee687ced5c23d26197627551e1ced

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.md

@@ -36,7 +36,7 @@ Automatic compaction uses one unified measurement for each threshold-and-retenti
 
 Compact policy has service-wide defaults: threshold ratio `0.8`, retained-tail ratio `0.16`, `summarizationProvider: ''`, `summarizationModel: ''`, `maxTokens: 8192`, `compactionRetries: 1`, `maxOverflowRetries: 1`, and `auto: true`. Top-level fields apply to every routed target; exact provider/model entries in `modelPolicies` partially override them. Pressure scales ratios against capacity resolved from the owning adapter, and `retainTokens` may replace `retainRatio`; retention must remain below the resulting threshold. The summarization provider and model must both be set or both be empty; an empty pair resolves the latest logged request target, then the `AgentOptions` pair.
 
-Automatic pressure runs at `agent/post-step` and measures the canonical durable envelope produced under the provider/model actually selected by `agent/request`. A headerless session has no completed routed request to assess and produces no work; any routed target can use the singleton estimator. Canonical overflow recovery uses the same measurement for forced range selection and retries only after a proven surface replacement.
+Automatic pressure runs at `agent/pre-step` before request derivation and measures the canonical durable envelope produced under the provider/model actually selected by the preceding `agent/request`. A headerless session has no completed routed request to assess and produces no work; any routed target can use the singleton estimator. Canonical overflow recovery uses the same measurement for forced range selection and retries only after a proven surface replacement.
 
 ## Testing
 

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.zh.md

@@ -36,7 +36,7 @@ Usage 会对互不重叠的输入、缓存读取、缓存写入与输出 bucket
 
 压缩策略采用服务级默认值：阈值比例 `0.8`、保留尾部比例 `0.16`、`summarizationProvider: ''`、`summarizationModel: ''`、`maxTokens: 8192`、`compactionRetries: 1`、`maxOverflowRetries: 1` 与 `auto: true`。顶层字段适用于每个路由目标；`modelPolicies` 中的精确提供方/模型项可以部分覆盖这些字段。压力检查根据所属适配器解析的容量缩放比例，`retainTokens` 可以替代 `retainRatio`；保留值必须小于最终阈值。摘要提供方与模型必须同时设置或同时为空；空组合先解析最近记录的请求目标，再使用 `AgentOptions` 中的组合。
 
-自动压力检查运行在 `agent/post-step`，并计量 `agent/request` 实际所选提供方/模型产生的规范持久信封。没有请求头的会话尚无已完成的路由请求可供判断，因此不执行工作；任意路由目标都可使用这个单例估算器。规范化溢出恢复使用同一计量结果强制选择范围，并且只有在表层替换得到证明后才重试。
+自动压力检查在请求派生前运行于 `agent/pre-step`，并计量前一个 `agent/request` 实际所选提供方/模型产生的规范持久信封。没有请求头的会话尚无已完成的路由请求可供判断，因此不执行工作；任意路由目标都可使用这个单例估算器。规范化溢出恢复使用同一计量结果强制选择范围，并且只有在表层替换得到证明后才重试。
 
 ## 测试
 

.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.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/architecture/2026-07-16-explicit-turn-cancellation.md
-2026-07-16-explicit-turn-cancellation.md: 15085a1da2cf183bace9957a4bedb3ea466aa472
-2026-07-16-explicit-turn-cancellation.zh.md: 8a7890a27e0d0cb1d3a9afd3935e8b5e785e2d6f
+2026-07-16-explicit-turn-cancellation.md: cce649976c9f4f596d5306b9fe8c3fd49a0e1adc
+2026-07-16-explicit-turn-cancellation.zh.md: 6f8b83fdb42af03c97dc2e8a9345a01acc6019fc

.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md

@@ -16,11 +16,11 @@ Agent owns the runtime-only `AgentCancelCause` union `{ kind: 'user' } | { kind:
 
 An interrupted live turn ends with the coarse durable `{ kind: 'aborted' }` outcome. The terminal event records what happened to the turn, while the runtime signal identifies who requested cancellation; it does not duplicate `user` or `parent` into replay. Session seed/load rejects legacy aborted records with a reason or any other extra field, so replay cannot reintroduce caller-owned cancellation detail. The process-local `agent/cancel-requested` notification is not durable; a future audit requirement uses a separate durable control-request event so a request and its eventual outcome remain distinct. Durable events contain no stack, signal, error object, free-form cancellation text, or backend-private detail.
 
-AgentLoop privately owns one `TurnCancellation` per prospective turn. It installs the holder before notifying `agent/status = running`, retains its single `AbortController` through prompt processing, prompt assembly, every step, model and tool execution, continuation, and `agent/turn-stop`, then clears the exact holder immediately before publishing `turn/end`. Terminal event observers and the following durability flush therefore cannot cancel already-completed turn work even though driver status may remain `running` until the flush settles. Every participating method, event, and request value receives that same explicit signal; the next turn receives a fresh signal.
+AgentLoop privately owns one `TurnCancellation` per prospective turn. It installs the holder before notifying `agent/status = running`, retains its single `AbortController` through inbox claim, `agent/pre-step`, prompt assembly, every step, model and tool execution, and `agent/turn-stopping`, then clears the exact holder immediately before publishing `turn/end`. Terminal event observers and the following durability flush therefore cannot cancel already-completed turn work even though driver status may remain `running` until the flush settles. Every participating method, event, and request value receives that same explicit signal; the next turn receives a fresh signal.
 
 The driver keeps only a cause-less pre-run marker for queued work cancelled before a turn is claimed. An effective `cancel()` emits the observe-only `agent/cancel-requested` notification with its resolved typed cause before clearing queued and steering work or aborting the holder; notification failures cannot veto the stop, and an idle call emits nothing. Work synchronously queued by a notification observer is included in that clear, while work queued by a later signal abort observer belongs to the next turn. If a `running` listener synchronously cancels old work and sends a replacement, the driver discards the aborted holder and creates a fresh one for the replacement. Repeated cancellation is first-wins for the active holder, while later calls may still clear newly queued pending work.
 
-The explicit event signatures keep their positional form and place `signal` immediately before a waterfall's final `next`. Prompt submission, request configuration, step-result processing, continuation, and terminal stop join the pre-existing explicit signal seams for pre-step, session prefix, model generation, tool execution, approval, and subagent or workflow requests. Hook bridges must also supply `RunHookOptions.signal`, so a turn cancellation reaches the bash executor's process-group kill and join boundary. `SystemPrompt.assemble()` carries `signal?: AbortSignal` in `AssembleContext` because that object is an explicit request value that can also represent signal-less assembly outside a turn. Listeners may cooperate with the signal but must not retain it to control another turn.
+The explicit event signatures keep their positional form and place `signal` inside `PreStepContext` or immediately before a waterfall's final `next`. Pre-step entry, request configuration, request-error recovery, model generation, tool execution, approval, turn stopping, and subagent or workflow requests all receive the current signal. Hook bridges must also supply `RunHookOptions.signal`, so a turn cancellation reaches the bash executor's process-group kill and join boundary. `SystemPrompt.assemble()` carries `signal?: AbortSignal` in `AssembleContext` because that object is an explicit request value that can also represent signal-less assembly outside a turn. Listeners may cooperate with the signal but must not retain it to control another turn.
 
 `ctx.agents` continues to carry only the initiating Agent. Ambient Agent presence does not imply liveness, a current turn, or cancellation authority. The cause reader is private to the loop and states the machine-private slot invariant (only `cancel()` aborts a turn controller, always with a canonical frozen cause) instead of re-validating the reason structurally; no public helper reads a cause off an arbitrary signal. Concurrent Agents isolate both their initiator identities and their turn signals; a child driver shadows the parent initiator while its parent request signal still travels through the subagent seam.
 
@@ -30,7 +30,7 @@ Cancellation remains cooperative. The loop checks interruption before and after
 
 ## Verification
 
-Contract tests verify the typed caller union, frozen detachment, default and first-wins behavior, the coarse Session JSON round trip and legacy-record rejection, ACP `user`, in-process subagent `parent`, and disposal precedence. Loop tests make cooperative listeners wait on the signal at prompt submission, system-prompt assembly, session prefix, pre-step, request, model stream, step result, tool execution, continuation, and terminal stop; they assert one signal within a turn, a fresh signal across turns, and no cancellation authority during terminal publication or a blocked durability flush. A real hook bridge test cancels and reaps a blocked prompt hook before idle.
+Contract tests verify the typed caller union, frozen detachment, default and first-wins behavior, the coarse Session JSON round trip and legacy-record rejection, ACP `user`, in-process subagent `parent`, and disposal precedence. Loop tests make cooperative listeners wait on the signal at pre-step, system-prompt assembly, request, model stream, request-error recovery, tool execution, and turn stopping; they assert one signal within a turn, a fresh signal across turns, and no cancellation authority during terminal publication or a blocked durability flush. A real hook bridge test cancels and reaps a blocked prompt hook before idle.
 
 Initiator-scope tests assert that every hook still observes the exact Agent and no ambient turn signal, concurrent Agents retain independent identities and signals, and a nested child driver shadows only identity. Race tests cover idle cancellation, pre-run cancellation, replacement submission from a `running` listener, repeated cancellation, and cancel-versus-dispose quiescence.
 

.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.zh.md

@@ -16,11 +16,11 @@ Agent 拥有仅用于运行时的 `AgentCancelCause` 联合类型 `{ kind: 'user
 
 正在运行的轮次被中断后，以粗粒度的持久化结果 `{ kind: 'aborted' }` 结束。终态事件记录轮次发生了什么，运行时 signal 标识谁请求了取消；回放不会重复保存 `user` 或 `parent`。会话 seed/load 会拒绝携带取消原因或任何其他额外字段的旧式中止记录，因此回放无法重新引入由调用方持有的取消细节。仅限进程内的 `agent/cancel-requested` 通知不会持久化；未来若有审计需求，应使用独立的持久化控制请求事件，让请求与最终结果保持为两项事实。持久化事件不包含调用栈、signal、错误对象、自由文本取消原因或后端私有细节。
 
-AgentLoop 为每个待启动轮次私有地持有一个 `TurnCancellation`。它在通知 `agent/status = running` 前安装该持有者，使其中唯一的 `AbortController` 持续覆盖提示词处理、提示词组装、每个步骤、模型与工具执行、继续决策和 `agent/turn-stop`；随后在发布 `turn/end` 前立即清除所安装的那个持有者。因此，即使驱动器状态可能在持久化刷新结算前保持 `running`，终态事件观察者及其后的持久化刷新也无法取消已完成的轮次工作。所有参与的方法、事件和请求值都会收到同一个显式 signal；下一个轮次会收到全新的 signal。
+AgentLoop 为每个待启动轮次私有地持有一个 `TurnCancellation`。它在通知 `agent/status = running` 前安装该持有者，使其中唯一的 `AbortController` 持续覆盖 inbox 领取、`agent/pre-step`、提示词组装、每个步骤、模型与工具执行以及 `agent/turn-stopping`；随后在发布 `turn/end` 前立即清除所安装的那个持有者。因此，即使驱动器状态可能在持久化刷新结算前保持 `running`，终态事件观察者及其后的持久化刷新也无法取消已完成的轮次工作。所有参与的方法、事件和请求值都会收到同一个显式 signal；下一个轮次会收到全新的 signal。
 
 对于轮次被认领前已取消的排队工作，驱动器只保留一个不携带取消原因的运行前标记。实际生效的 `cancel()` 会先发出仅供观察的 `agent/cancel-requested` 通知并携带最终确定的类型化取消原因，然后才清除排队工作和 steering（中途引导）工作或中止持有者；通知失败不能阻止此次停止，空闲状态下调用则不发出任何通知。通知观察者同步加入队列的工作也会被这次清除，而稍后由 signal 中止观察者加入队列的工作属于下一个轮次。若 `running` 监听器同步取消旧工作并发送替代提示词，驱动器会丢弃已中止的持有者，并为替代提示词创建全新的持有者。同一活跃持有者上的重复取消遵循首次请求优先，后续调用仍可清除新入队的待处理工作。
 
-显式事件签名保留位置参数形式，并把 `signal` 放在 waterfall（瀑布式事件）的最后一个参数 `next` 之前。提示词提交、请求配置、步骤结果处理、继续决策和终止停止加入已有的步骤前处理、会话前缀、模型生成、工具执行、审批以及 subagent 或工作流请求的显式 signal seam。钩子桥接器也必须提供 `RunHookOptions.signal`，使轮次取消能够到达 Bash 执行器终止进程组并等待其退出的边界。`SystemPrompt.assemble()` 在 `AssembleContext` 中携带 `signal?: AbortSignal`，因为该对象是显式请求值，也可表示轮次之外不携带 signal 的组装。监听器可以配合该 signal 取消，但不得保留它来控制其他轮次。
+显式事件签名保留位置参数形式，并把 `signal` 放入 `PreStepContext`，或放在 waterfall（瀑布式事件）的最后一个参数 `next` 之前。pre-step 进入决策、请求配置、请求错误恢复、模型生成、工具执行、审批、轮次停止以及 subagent 或工作流请求都会收到当前 signal。钩子桥接器也必须提供 `RunHookOptions.signal`，使轮次取消能够到达 Bash 执行器终止进程组并等待其退出的边界。`SystemPrompt.assemble()` 在 `AssembleContext` 中携带 `signal?: AbortSignal`，因为该对象是显式请求值，也可表示轮次之外不携带 signal 的组装。监听器可以配合该 signal 取消，但不得保留它来控制其他轮次。
 
 `ctx.agents` 仍只携带发起 Agent。环境中的 Agent 并不代表存活、当前轮次或取消权限。cause 读取器是 loop 私有的，它直接陈述机器私有的 slot 不变量（只有 `cancel()` 会中止轮次控制器，且总是携带规范的冻结 cause），而不是对 reason 做结构化再校验；不存在从任意 signal 读取 cause 的公开辅助函数。并发 Agent 会同时隔离各自的发起方身份和轮次 signal；子驱动会遮蔽父发起方，而父请求 signal 仍通过 subagent seam 传递。
 
@@ -30,7 +30,7 @@ Agent dispose（资源释放）会在活跃持有者上请求仅用于运行时
 
 ## 验证
 
-契约测试验证类型化调用方联合类型、冻结且与调用方分离、默认行为与首次请求优先行为、粗粒度的会话 JSON 往返与旧式记录拒绝、ACP `user`、进程内 subagent `parent` 以及 dispose 优先级。AgentLoop 测试让协作式监听器在提示词提交、系统提示词组装、会话前缀、步骤前处理、请求、模型流、步骤结果、工具执行、继续决策和终止停止处等待 signal；并断言同一轮次使用一个 signal，不同轮次使用全新的 signal，终态发布期间和持久化刷新受阻期间不存在取消权限。真实钩子桥接器测试会在报告空闲状态前取消并回收受阻的提示词钩子。
+契约测试验证类型化调用方联合类型、冻结且与调用方分离、默认行为与首次请求优先行为、粗粒度的会话 JSON 往返与旧式记录拒绝、ACP `user`、进程内 subagent `parent` 以及 dispose 优先级。AgentLoop 测试让协作式监听器在 pre-step、系统提示词组装、请求、模型流、请求错误恢复、工具执行和轮次停止处等待 signal；并断言同一轮次使用一个 signal，不同轮次使用全新的 signal，终态发布期间和持久化刷新受阻期间不存在取消权限。真实钩子桥接器测试会在报告空闲状态前取消并回收受阻的提示词钩子。
 
 发起方作用域测试断言所有钩子仍观察到同一个 Agent 且没有环境中的轮次 signal，并发 Agent 保持独立的身份与 signal，嵌套子驱动只遮蔽身份。竞态测试覆盖空闲状态取消、运行前取消、从 `running` 监听器提交替代提示词、重复取消以及取消与 dispose 竞争下的完全停稳。
 

.agents/notes/implemented/architecture/2026-07-22-unified-send-and-coalesced-user-messages.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/architecture/2026-07-22-unified-send-and-coalesced-user-messages.md
-2026-07-22-unified-send-and-coalesced-user-messages.md: 4d0cbeff0c8a07362caa1ec18493267a9f0d2823
-2026-07-22-unified-send-and-coalesced-user-messages.zh.md: 4b593bd578840dc51309310dcb8478fb0dd1e1f5
+2026-07-22-unified-send-and-coalesced-user-messages.md: f32d6ca65d5236e1fabdd177cdf54e36929c853f
+2026-07-22-unified-send-and-coalesced-user-messages.zh.md: 7647742166bb760139506f93af672f323b3b7b97

.agents/notes/implemented/architecture/2026-07-22-unified-send-and-coalesced-user-messages.md

@@ -12,23 +12,23 @@ Separately, `context/message` and `user/message` had converged: the surface proj
 
 ## Decision
 
-**One primitive, three preset aliases.** The `Agent` interface's `send(message, { target, wakeup })` covers the (`target` × `wakeup`) matrix. Its complete `UserMessage` owns identity, role, model-facing `content`, and producer `source`; the complete `SendOptions` owns only routing policy. `followup` (`next-turn`/wakeup), `steer` (`next-step`/wakeup), and `inject` (`next-step`/no-wakeup) each accept that one message and fix the policy. `wakeup` means "make the model run": wake a parked driver for a `next-turn` item, or force a continuation for a running `next-step` item. `next-turn`/no-wakeup (queue without waking) is representable with no alias and no current caller.
+**One primitive, three preset aliases.** The `Agent` interface's `send(message, target, wakeup)` covers the (`target` × `wakeup`) matrix. Its complete `UserMessage` owns identity, role, model-facing `content`, and producer `source`; the remaining arguments own only routing policy. `followup` (`next-turn`/wakeup), `steer` (`next-step`/wakeup), and `inject` (`next-step`/no-wakeup) each accept that one message and fix the policy. `wakeup` reserves a driver when the agent is idle; an already active driver receives no second reservation and can claim the input only if it reaches a later pre-step boundary. `next-turn`/no-wakeup (queue without waking) is representable with no alias and no current caller.
 
-**inject keeps its mechanism.** The `next-step`/no-wakeup path is exactly the old `inject`: durable model-facing context appended at the current log position, deferred while prompt admission or a turn owns the next safe boundary, and appended directly outside that window. It bypasses the FIFOs entirely, while its required `UserMessage.source` preserves the caller's explicit provenance.
+**inject is a non-waking next-step delivery.** It always appends the complete message to the next-step inbox and records that insertion in a durable `agent/inbox/spliced` event. The driver claims it at a later pre-step and records it as model-visible `user/message` only when the final decision returns it in the entering batch; idle injection remains pending until another delivery wakes the driver. Its required `UserMessage.source` preserves the caller's explicit provenance.
 
-**context/message is gone.** Injected context is now a `user/message`; context producers supply the appropriate non-user `source` explicitly, and typed source variants carry any domain-specific durable provenance. The surface, derivation, and `SurfaceEventType` drop `context/message`; consumers that need "is this a human prompt?" read `source.kind === 'user'` instead of the event type.
+**context/message is gone.** Injected context uses one `UserMessage` value in the inbox and becomes a `user/message` event if admitted; context producers supply the appropriate non-user `source` explicitly, and typed source variants carry any domain-specific durable provenance. The surface, derivation, and `SurfaceEventType` drop `context/message`; consumers that need "is this a human prompt?" read `source.kind === 'user'` instead of the event type.
 
-**Goal replay disambiguates by round, not type.** A goal state change is a round-zero goal-sourced `user/message` whose source carries the complete change; a positive round is an admitted continuation prompt. `decodeGoalEvent` takes a `user/message` and fails loud when goal-state content and its typed source disagree.
+**Goal continuation attribution uses positive rounds.** Goal lifecycle state commits through the domain-owned `goal/change` event defined by the later [goal-owned durable event decision](2026-07-31-goal-owned-durable-events.md). A positive round advances only from an admitted continuation `user/message`; goal persistence does not use injection or inbox state.
 
 **`send` does not return identity.** Callers already own the complete message and its opaque `MessageId`; creation and freezing are owned by the [identified immutable message decision](2026-07-28-identified-immutable-message-values.md), not by routing.
 
-**Inbox lifecycle events carry occurrence identities.** `agent/inbox/enqueue` (an item entered a FIFO), `agent/inbox/update` (a pending queued item was edited), `agent/inbox/dequeue` (the driver claimed one), and `agent/inbox/discard` (pending items were dropped) carry an `InboxItem`: an occurrence-local `InboxItemId`, the accepted `UserMessage`, and the resolved `queued | steering` placement captured at acceptance. The occurrence identity lets observers and reconnect mirrors distinguish repeated sends of the same `MessageId` without reconstructing routing from later status or session history. Injection never touches a FIFO and emits none of these. Every FIFO entry publishes one enqueue and exactly one terminal dequeue or discard; updates are non-terminal. The `dsh-agent` invariant companion asserts this FIFO conservation.
+**Inbox mutations have one durable projection and three minimal live notifications.** Every append, prepend, edit, remove, cancellation, and claim records normalized `agent/inbox/spliced` coordinates. Insertions emit `agent/inbox/inserted { message }`; ordinary removals carry durable `outcome: 'canceled'` and emit `agent/inbox/discarded { message }`; the loop's atomic `claim()` records pure deletion splices and then emits `agent/inbox/claimed { message, turn }`. `MessageId` is the sole occurrence identity and remains unique across both pending lists. The live payloads deliberately omit placement, outcome, and batch envelopes because the durable splice owns those facts.
 
-**Admission accepts next-step input without becoming a turn.** The loop opens a private next-step acceptance window before `agent/prompt-submit`, keeps it open through the turn, and closes it before `turn/end`. Steering and injection received during admission therefore remain together in the outbox and join an allowed turn. If admission blocks or fails, a context-only caller batch takes idle injection's immediate append, while steering and context staged beside it remain available to retry; neither path writes the rejected prompt. When a later prompt is admitted, retained outbox input enters its turn before that prompt, while input accepted during the current admission remains after the prompt. Closing the window before `turn/end` preserves the rule that reentrant late steering becomes an independent queued turn. `Agent.acceptsNextStep` exposes whether a `next-step` send would currently join this window; `status` remains the broader activity signal rather than a routing predicate.
+**Pre-step claims next-step input without making it a separate turn.** Steering and injection always enter the same next-step inbox; steering wakes the driver, while injection does not. At a turn boundary the driver atomically claims pending next-step input before one queued prompt, and between steps it claims only next-step input. Claiming records pure deletion splices and emits `agent/inbox/claimed { message, turn }` once per message. `agent/pre-step` then rejects the proposed step or returns its complete entering batch. Rejection and listener failure leave the claimed batch removed; input arriving after the claim waits for a later boundary.
 
-**One accepted message keeps one representation.** Durable user-role input and additional model-facing context both use the identified, frozen `UserMessage` directly. The loop stores that value beside private routing state rather than copying its identity, content, or source into another public shape. A queued message that becomes steering keeps the same message value in the outbox, while injected and tool-produced context each carry their own identified message. The [identified immutable message decision](2026-07-28-identified-immutable-message-values.md) supersedes this note's former `UserMessageData`/`AgentMessage` hierarchy and extends the representation to assistant and tool-result messages.
+**One accepted message keeps one representation.** Durable user-role input and additional model-facing context both use the identified, frozen `UserMessage` directly. The loop stores that value beside private routing state rather than copying its identity, content, or source into another public shape. Steering, injection, and tool-produced context each keep their identified messages in the next-step inbox. The [identified immutable message decision](2026-07-28-identified-immutable-message-values.md) supersedes this note's former `UserMessageData`/`AgentMessage` hierarchy and extends the representation to assistant and tool-result messages.
 
-**Idle wakeup follows acceptance.** Before publishing enqueue, a waking queued send installs quiescence ownership and schedules driver admission for a microtask that runs after the id returns. Every send in one synchronous caller stack therefore resolves placement against the same pre-admission state, while reentrant cancellation or teardown cannot retire before the scheduled admission settles. Two idle `steer()` calls remain two FIFO turns instead of the first opening an admission window that captures the second.
+**Idle wakeup follows insertion.** A waking send inserts its input, then enters the running driver before returning. The first pre-step may claim that input immediately; later synchronous sends therefore join the running loop and wait for a later boundary. Cancellation belongs to the running turn signal from wakeup onward; no distinct pre-run phase intervenes.
 
 **cancel gains keepInbox.** `cancel(cause, { keepInbox? })`; callers choose the cause explicitly, and `keepInbox: true` aborts the active turn while preserving queued and steering items (no discard event, and un-started work is not dropped).
 
@@ -36,14 +36,14 @@ Separately, `context/message` and `user/message` had converged: the surface proj
 
 - **A dedicated `MessageSource` kind `context`** for injected content. Rejected because `plugin` already means "not a human," so a fourth kind would add a parallel axis the authority checks would have to learn. Plugin-produced injected context supplies its plugin source explicitly.
 - **A typed discriminant field on `UserMessage`** (e.g. `origin: 'prompt' | 'context'`) to replace the event-type split. Rejected in favor of `source`, which every consumer already carries and which the goal system already keyed on; a second discriminant would duplicate that fact.
-- **Keeping `agent/queued` alongside the inbox events.** Rejected as a mirror: `agent/inbox/enqueue` is the same enqueue-time signal with the resolved placement, and the dequeue/discard events complete the FIFO lifecycle the single event could not describe.
-- **Derive inbox placement from agent status or the session log.** Rejected because `running` includes admission and settlement, while reconnect baselines need the original acceptance result even when the earlier turn boundary is absent. The producer already owns the exact routing decision.
+- **Keeping `agent/queued` alongside the inbox events.** Rejected as a mirror: `agent/inbox/inserted` is the live insertion signal, while claimed/discarded notifications describe exits and the durable splice retains placement.
+- **Derive inbox placement from agent status.** Rejected because `running` includes pre-step processing and settlement. The producer already supplies the exact target to the durable splice.
 
 ## Consequences
 
-The delivery surface is now one primitive plus three self-documenting presets, and the (`target` × `wakeup`) matrix makes previously-unreachable combinations explicit. One durable message type serves prompts, injected context, and goal rounds, so the surface projection and every "human prompt?" check simplify to a `source` test. The `Agent` contract remains an interface, so alternate implementations and object-literal test fakes implement the same minimal structural surface. The goal fold's channel split moved from event type to `source.round`, and every consumer that filtered `context/message` now filters `user/message` by source. An idle injection appends `user/message` between turns without opening a turn or running the model.
+The delivery surface is now one primitive plus three self-documenting presets, and the (`target` × `wakeup`) matrix makes previously-unreachable combinations explicit. One identified message value serves prompts, injected context, and goal rounds, so every "human prompt?" check simplifies to a `source` test. The `Agent` contract remains an interface, so alternate implementations and object-literal test fakes implement the same minimal structural surface. Positive goal rounds fold from admitted `user/message` events, while goal lifecycle state remains outside the delivery surface. An idle injection remains pending without opening a turn or running the model, then becomes `user/message` when a later waking delivery's pre-step returns it in the entering batch.
 
-`wakeup` is the "should the model run" signal, so the inbox distinguishes waking queued work from anything available to dequeue: a lone `next-turn`/no-wakeup item stays parked at idle and rides along the next waking send, and `whenIdle`/`cancel` settle quiescence off the waking signal. Every FIFO exit publishes exactly one lifecycle event, while domain-specific durable facts travel in typed message sources rather than a parallel metadata channel. The direct pending-item representation keeps public lifecycle events correlated without maintaining a second steering wrapper or allowing its durable data to diverge. The later [addressable queue operations](../feature/2026-07-29-addressable-queue-operations.md) decision adds live mutations over that occurrence identity without changing the one-message-per-turn or durable-message contracts.
+`wakeup` is the "should the model run" signal, so the inbox distinguishes waking queued work from anything available to claim: a lone `next-turn`/no-wakeup item stays parked at idle and rides along the next waking send, and `whenIdle`/`cancel` settle quiescence off the waking signal. Every insertion and exit publishes its matching live notification, while domain-specific durable facts travel in typed message sources rather than a parallel metadata channel. The direct pending-message representation keeps durable splices and live events correlated without maintaining a second steering wrapper or allowing its data to diverge. The later [claimed pre-step inbox lifecycle](2026-07-31-claimed-pre-step-inbox-lifecycle.md) decision keeps live queue mutations addressed by `MessageId` and separates single-message lifecycle notifications from the durable whole-queue splice projection.
 
 ## Related
 

.agents/notes/implemented/architecture/2026-07-22-unified-send-and-coalesced-user-messages.zh.md

@@ -12,23 +12,23 @@ agent 的对外驱动接口逐渐长出三个近乎平行的动词——`send`
 
 ## 决策
 
-**一个原语，三个预设别名。** `Agent` 接口的 `send(message, { target, wakeup })` 覆盖 (`target` × `wakeup`) 矩阵。完整的 `UserMessage` 持有标识、角色、模型可见 `content` 与生产方 `source`；完整的 `SendOptions` 只持有路由策略。`followup`（`next-turn`/wakeup）、`steer`（`next-step`/wakeup）和 `inject`（`next-step`/no-wakeup）都接收这一条消息并固定策略。`wakeup` 意为“让模型运行”：为一个 `next-turn` 队列项唤醒处于停泊状态的驱动器，或为一个运行中的 `next-step` 队列项强制继续执行。`next-turn`/no-wakeup（入队但不唤醒）可以表达，只是没有别名，也没有当前调用方。
+**一个原语，三个预设别名。** `Agent` 接口的 `send(message, target, wakeup)` 覆盖 (`target` × `wakeup`) 矩阵。完整的 `UserMessage` 持有标识、角色、模型可见 `content` 与生产方 `source`；其余参数只持有路由策略。`followup`（`next-turn`/wakeup）、`steer`（`next-step`/wakeup）和 `inject`（`next-step`/no-wakeup）都接收这一条消息并固定策略。`wakeup` 会在 agent 空闲时保留一个驱动器；已经活跃的驱动器不会获得第二次保留，只有在抵达后续 pre-step 边界时才能领取该输入。`next-turn`/no-wakeup（入队但不唤醒）可以表达，只是没有别名，也没有当前调用方。
 
-**inject 保留其机制。** `next-step`/no-wakeup 路径正是旧的 `inject`：持久的面向模型上下文会追加到当前日志位置；当提示词准入流程或某个轮次占用下一个安全边界时，它会延迟处理，而在该窗口之外则直接追加。它完全绕过 FIFO 队列，而必填的 `UserMessage.source` 会保留调用方显式提供的来源信息。
+**inject 是不会唤醒的 next-step 投递。** 它始终把完整消息追加到 next-step inbox，并在持久 `agent/inbox/spliced` 事件中记录该插入。驱动器会在后续 pre-step 领取它，并且只有最终决策把它放入进入步骤的批次时，才会将其记录为模型可见的 `user/message`；空闲注入会保持待处理，直到其他投递唤醒驱动器。必填的 `UserMessage.source` 会保留调用方显式提供的来源信息。
 
-**context/message 已移除。** 注入的上下文现在是一条 `user/message`；上下文生产方显式提供合适的非 `user` 类别 `source`，类型化 source 变体携带所有特定于领域的持久来源信息。对外接口、派生逻辑和 `SurfaceEventType` 都不再包含 `context/message`；需要判断“这是不是一条人类提示词？”的消费方改为读取 `source.kind === 'user'`，而不是事件类型。
+**context/message 已移除。** 注入的上下文在 inbox 中使用同一个 `UserMessage` 值，并在获准时成为 `user/message` 事件；上下文生产方显式提供合适的非 `user` 类别 `source`，类型化 source 变体携带所有特定于领域的持久来源信息。对外接口、派生逻辑和 `SurfaceEventType` 都不再包含 `context/message`；需要判断“这是不是一条人类提示词？”的消费方改为读取 `source.kind === 'user'`，而不是事件类型。
 
-**goal 回放靠轮次而非类型来区分。** 一次 goal 状态变更是一条第 0 轮、来源为 goal 的 `user/message`，其 source 携带完整变更；正数轮次则是一条已准入的继续执行提示词。`decodeGoalEvent` 接收一条 `user/message`，并在 goal 状态内容与其类型化 source 不一致时立即报错。
+**Goal 继续执行归属使用正数 Round。** Goal 生命周期状态通过后续 [Goal 自有持久事件决策](2026-07-31-goal-owned-durable-events.md)定义的领域自有 `goal/change` 事件提交。正数 Round 只从已准入的继续执行 `user/message` 推进；goal 持久化不使用注入或 inbox 状态。
 
 **`send` 不返回标识。** 调用方已经持有完整消息及其不透明的 `MessageId`；消息的创建与冻结由[带标识的不可变消息值决策](2026-07-28-identified-immutable-message-values.md)负责，而不是由路由负责。
 
-**Inbox 生命周期事件携带单次入队标识。** `agent/inbox/enqueue`（一个队列项进入某个 FIFO）、`agent/inbox/update`（待处理的 queued 项被编辑）、`agent/inbox/dequeue`（驱动器认领一个项）和 `agent/inbox/discard`（待处理项被丢弃）都会携带一个 `InboxItem`：仅属于本次入队的 `InboxItemId`、已接受的 `UserMessage`，以及生产方在接受消息时捕获的已解析 `queued | steering` 放置方式。单次入队标识让观察方和重连镜像能够区分同一 `MessageId` 的多次发送，无需根据后续状态或会话历史重建路由。注入从不触及 FIFO，也不发出这些事件中的任何一个。每次 FIFO 入队都会发布一个 enqueue，并且恰好发布一个终态 dequeue 或 discard；update 不是终态。`dsh-agent` 的不变量配套断言这种 FIFO 守恒。
+**Inbox 变更只有一份持久投影和三种最小实时通知。** 每次 append、prepend、编辑、删除、取消与领取都会记录规范化的 `agent/inbox/spliced` 坐标。插入会发出 `agent/inbox/inserted { message }`；普通删除携带持久 `outcome: 'canceled'`，并发出 `agent/inbox/discarded { message }`；循环的原子 `claim()` 会记录纯删除 splice，随后发出 `agent/inbox/claimed { message, turn }`。`MessageId` 是唯一的单次出现标识，并在两个待处理列表间保持唯一。实时载荷刻意不携带 placement、outcome 或批次封套，因为这些事实由持久 splice 持有。
 
-**准入接受 next-step 输入，但不会因此成为一个轮次。** 循环会在 `agent/prompt-submit` 前打开一个私有的 next-step 接受窗口，使其贯穿整个轮次，并在 `turn/end` 前关闭。因此，在准入期间收到的 steering 和注入会一起留在 outbox 中并加入已准入的轮次。如果准入被阻止或失败，仅含调用方上下文的批次会采用空闲注入的立即追加行为，而 steering 及与其一同暂存的上下文仍可重试；两种路径都不会写入被拒绝的提示词。后续提示词获准时，保留在 outbox 中的输入会先于该提示词进入其轮次，而当前准入期间接受的输入则留在提示词之后。在 `turn/end` 前关闭窗口，可以保留这样的规则：可重入的晚到 steering 会成为一个独立的排队轮次。`Agent.acceptsNextStep` 会公开一次 `next-step` 发送当前是否会加入该窗口；`status` 仍是更宽泛的活动信号，而非路由判据。
+**pre-step 会领取 next-step 输入，但不会为它单独创建轮次。** steering 和注入始终进入同一个 next-step inbox；steering 会唤醒驱动器，注入则不会。在轮次边界，驱动器会原子领取待处理的 next-step 输入，再领取一条排队提示词；在步骤之间则只领取 next-step 输入。领取会记录纯删除 splice，并针对每条消息发出一次 `agent/inbox/claimed { message, turn }`。随后 `agent/pre-step` 会拒绝拟议步骤，或返回进入步骤的完整批次。reject 与监听器失败都会让已领取批次保持已删除；领取后才到达的输入会等待后续边界。
 
-**一条已接受消息只保留一种表示。** 持久的用户角色输入和附加的模型可见上下文都直接使用带标识且冻结的 `UserMessage`。循环把该值与私有路由状态存放在一起，不会将其标识、内容或来源复制到另一种公开形状中。一条成为 steering 的排队消息会在 outbox 中保留同一个消息值，而注入和工具产生的上下文则各自携带带标识的消息。[带标识的不可变消息值决策](2026-07-28-identified-immutable-message-values.md)取代了本记录此前的 `UserMessageData`/`AgentMessage` 层级，并将这一表示扩展到 assistant 消息和工具结果消息。
+**一条已接受消息只保留一种表示。** 持久的用户角色输入和附加的模型可见上下文都直接使用带标识且冻结的 `UserMessage`。循环把该值与私有路由状态存放在一起，不会将其标识、内容或来源复制到另一种公开形状中。steering、注入和工具产生的上下文都会在 next-step inbox 中保留各自带标识的消息。[带标识的不可变消息值决策](2026-07-28-identified-immutable-message-values.md)取代了本记录此前的 `UserMessageData`/`AgentMessage` 层级，并将这一表示扩展到 assistant 消息和工具结果消息。
 
-**空闲唤醒在接受之后发生。** 在发布 enqueue 前，一次会唤醒驱动器的排队发送会先取得完全停稳所有权，并把驱动器准入调度到一个会在该次发送返回 id 后运行的微任务中。因此，同一同步调用栈中的每次发送都会基于同一份准入前状态解析放置方式，而可重入的取消或拆除在已调度的准入结算前无法完成退役。空闲时的两次 `steer()` 调用会保留为两个 FIFO 轮次，而不会因第一次调用打开准入窗口而把第二次吸纳进去。
+**空闲唤醒在插入之后发生。** 会唤醒的发送会先插入输入，再于返回前进入 running 驱动器。首次 pre-step 可能立即领取该输入；因此，后续同步发送会加入正在运行的循环，并等待更晚的边界。自唤醒开始，取消就归属于 running 轮次信号，中间不会插入独立的预运行 phase。
 
 **cancel 新增 keepInbox。** `cancel(cause, { keepInbox? })`；调用方显式选择 cause，且 `keepInbox: true` 会中止活跃轮次，同时保留排队项和 steering 项（不发出 discard 事件，尚未启动的工作也不会被丢弃）。
 
@@ -36,14 +36,14 @@ agent 的对外驱动接口逐渐长出三个近乎平行的动词——`send`
 
 - **为注入内容设立专门的 `MessageSource` 类别 `context`。** 不予采纳，因为 `plugin` 已经表示“不是人类”，因此第四种类别会增加一条平行的轴，让授权检查不得不去学习它。由插件产生的注入上下文会显式提供其 plugin 来源。
 - **在 `UserMessage` 上设一个类型化的判别字段**（例如 `origin: 'prompt' | 'context'`）来取代事件类型的区分。不予采纳，转而采用 `source`——每个消费方都已经携带它，goal 系统也已经以它为键；第二个判别字段会重复这一事实。
-- **在 inbox 事件之外保留 `agent/queued`。** 作为镜像而被否决：`agent/inbox/enqueue` 是同一个入队时刻的信号，只是带有已解析的放置方式，而 dequeue/discard 事件补全了单个事件无法描述的 FIFO 生命周期。
-- **根据 agent 状态或会话日志推导 inbox 放置方式。** 不予采纳，因为 `running` 同时涵盖准入与结算，而重连基线即使缺少此前的轮次边界，也需要最初的接受结果。生产方已经拥有精确的路由决策。
+- **在 inbox 事件之外保留 `agent/queued`。** 作为镜像而被否决：`agent/inbox/inserted` 已经是实时插入信号，claimed/discarded 通知描述退出，而持久 splice 保留 placement。
+- **根据 agent 状态推导 inbox 放置方式。** 不予采纳，因为 `running` 同时涵盖 pre-step 处理与结算。生产方已经把精确目标写入持久 splice。
 
 ## 后果
 
-投递接口现在是一个原语加三个自解释的预设，(`target` × `wakeup`) 矩阵把此前无法表达的组合显式化。一种持久消息类型同时服务提示词、注入的上下文和 goal 轮次，因此对外接口的投影和每一处“是否人类提示词？”检查都简化为一次 `source` 判断。`Agent` 契约仍是接口，因此其他实现和对象字面量形式的测试替身只需实现同一个最小结构接口。goal 折叠的通道区分从事件类型改到了 `source.round`；此前过滤 `context/message` 的每个消费方现在改为按来源过滤 `user/message`。空闲状态下的注入会在两个轮次之间追加 `user/message`，既不打开轮次，也不运行模型。
+投递接口现在是一个原语加三个自解释的预设，(`target` × `wakeup`) 矩阵把此前无法表达的组合显式化。同一个带标识消息值同时服务提示词、注入的上下文和 Goal Round，因此每一处“是否人类提示词？”检查都简化为一次 `source` 判断。`Agent` 契约仍是接口，因此其他实现和对象字面量形式的测试替身只需实现同一个最小结构接口。正数 Goal Round 从已准入的 `user/message` 事件折叠，而 goal 生命周期状态位于投递接口之外。空闲注入会保持待处理，不打开轮次也不运行模型；后续会唤醒的投递在 pre-step 将其放入进入步骤的批次时，它才成为 `user/message`。
 
-`wakeup` 是“模型是否应当运行”的信号，因此 inbox 会区分能唤醒的排队工作与任何可 dequeue 的项：一个孤立的 `next-turn`/no-wakeup 队列项会在空闲状态保持停泊，并随下一次会唤醒驱动器的 send 一同出队；`whenIdle`/`cancel` 则依据唤醒信号判断何时达到完全停稳。每一次 FIFO 退出都恰好发布一个生命周期事件，特定于领域的持久事实则通过类型化消息 source 传递，而非通过平行的元数据通道。直接使用待处理项的表示方式，使公开生命周期事件保持可关联，既无需维护第二个 steering 包装层，也避免其持久数据发生分歧。后续的[可寻址队列操作](../feature/2026-07-29-addressable-queue-operations.md)决策在该单次入队标识上增加了实时变更，但不改变单消息单轮次或持久消息契约。
+`wakeup` 是“模型是否应当运行”的信号，因此 inbox 会区分能唤醒的排队工作与任何可领取的项：一个孤立的 `next-turn`/no-wakeup 队列项会停泊在空闲状态，并随下一次唤醒 send 一同带出，而 `whenIdle`/`cancel` 依据唤醒信号来结算完全停稳。每次插入与退出都会发布对应的实时通知，特定于领域的持久事实则通过类型化消息 source 传递，而非通过平行的元数据通道。直接使用待处理消息的表示方式，使持久 splice 与实时事件保持可关联，既无需维护第二个 steering 包装层，也避免数据发生分歧。后续的[已领取 pre-step inbox 生命周期](2026-07-31-claimed-pre-step-inbox-lifecycle.md)决策保留通过 `MessageId` 寻址的实时队列变更，并把单消息生命周期通知与持久的整体队列 splice 投影分离。
 
 ## 相关
 

.agents/notes/implemented/architecture/2026-07-24-separate-context-injection-from-turn-execution.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/architecture/2026-07-24-separate-context-injection-from-turn-execution.md
-2026-07-24-separate-context-injection-from-turn-execution.md: bf3ae2ecbd2205a4c49e8004ffc694f89a2460a3
-2026-07-24-separate-context-injection-from-turn-execution.zh.md: fb40af6fbfae5a6ca842ad28fd1a4e8c12ff3256
+2026-07-24-separate-context-injection-from-turn-execution.md: bb28d96cf1494d94ad7a7d4a4714e536c7072d2f
+2026-07-24-separate-context-injection-from-turn-execution.zh.md: fa9b9275cae67f4d5a34dac8a0236558596ab28f

.agents/notes/implemented/architecture/2026-07-24-separate-context-injection-from-turn-execution.md

@@ -18,29 +18,29 @@ Idle `inject()` exposed a second mismatch. Injection did not request model execu
 
 `inject()` is the only caller-facing operation for supplementary model-facing input, and a turn means one execution of the model loop.
 
-`SendOptions` contains only `target` and `wakeup`. A caller that owns context delivers an identified, frozen `UserMessage` through `inject()` and submits the direct message independently with `send()` or `steer()`.
+A caller that owns context delivers an identified, frozen `UserMessage` through `inject()` and submits the direct message independently with `followup()` or `steer()`.
 
-Prompt and tool extension points still return `additionalContexts`. These values are outputs of the extension point, not attachments captured from a caller's inbox item. Prompt admission runs before `run()` opens a turn. An allowed prompt and its returned additional contexts enter the new turn as separate messages; a blocked prompt writes neither and opens no turn. Tool-produced additional contexts enter the outbox after the corresponding tool results.
+An entering pre-step returns the complete `PreStepDecision.messages` batch for the request being finalized. Tool extension points still return `additionalContexts`, which enter the next-step inbox only after the corresponding tool results. These values are extension-point outputs, not attachments captured from a caller's inbox item.
 
-Every additional context is an independent `user/message` whose `source` records provenance. There is no `context/message`, prompt-prefix placement, stable request delimiter, or prompt envelope. Transcript and UI consumers distinguish direct user messages from injected context by `source`.
+Every additional context is an independent `UserMessage` whose `source` records provenance. Inbox insertion is durable immediately; admission later records the same value as `user/message`. There is no `context/message`, prompt-prefix placement, stable request delimiter, or prompt envelope. Transcript and UI consumers distinguish direct user messages from injected context by `source`.
 
 ## Injection lifecycle
 
-During prompt admission or an open turn, `inject()` stages context in the loop outbox. The private next-step acceptance window opens before `agent/prompt-submit` and closes before `turn/end`, so steering and context accepted for one boundary reach the same following request while a `turn/end` listener's late steering becomes a queued prompt. The loop drains the outbox at a safe step boundary, preserving tool protocol adjacency: context accepted during an assistant tool-call batch appears only after that batch's complete ordered results.
+`inject()` always inserts context into the non-waking `next-step` inbox and commits that queue mutation as `agent/inbox/spliced`. A running driver claims it at the nearest later pre-step boundary. An idle driver leaves it pending until `followup()` or `steer()` supplies waking work; cancellation or disposal may discard it first without erasing the durable queue history.
 
-Outside that window, `inject()` appends its `user/message` immediately. It does not increment turn numbering, emit `turn/start` or `turn/end`, change agent status, or run the model; persistence observes the append through `session/event`.
+The loop claims the current next-step batch before running `agent/pre-step`, so an injection that arrives after that claim may miss the request already being finalized. The next boundary claims it instead. An enter decision appends its returned messages inside the owning turn before the request consumes them. Context produced during an assistant tool-call batch therefore appears after that batch's complete ordered results.
 
-If prompt admission blocks or fails, a caller-staged context-only batch appends immediately without a turn. Steering and context staged beside it remain in the outbox for a later admitted prompt; cancellation or disposal may discard them. Hook-produced `additionalContexts` never materialize because they belong to the rejected admission decision.
+If pre-step rejects or throws, its claimed injected context, steering, and queued prompt stay removed and no returned batch is appended. Messages inserted after that atomic claim are unaffected and remain pending.
 
-The session invariant permits `user/message` between turns while continuing to require turn enclosure for core execution events, steering, assistant output, and tools. Merge-extensible event relations belong to their declaring plugin rather than a core default. Persistence, recovery, resume, fork, and compaction treat valid between-turn events as committed session history rather than an interrupted or discardable turn tail.
+The loop appends injected `user/message` events only from entered batches inside a turn. Core execution events, steering, assistant output, and tools remain turn-enclosed; merge-extensible event relations belong to their declaring plugin rather than a core default.
 
 ## Extension and caller semantics
 
-`PromptDecision.content` continues to replace only the direct prompt. `PromptDecision.additionalContexts` and tool-result `additionalContexts` retain FIFO order and individual provenance, but no longer select placement. A waterfall listener that delegates with `next()` must preserve downstream prompt content and additional contexts unless it intentionally returns replacements.
+The enter branch's `PreStepDecision.messages` is the complete batch for the proposed step. A waterfall listener that delegates with `next()` preserves downstream messages unless it intentionally replaces them; additions follow natural waterfall return order. Tool-result `additionalContexts` retain FIFO order and individual provenance.
 
-Caller-driven injection and hook-produced additional context deliberately have different admission ownership. A hook's additional contexts materialize only after that hook allows the prompt or tool result. Outside a next-step acceptance window, a caller that invokes `inject(context)` and then `send(prompt)` commits context independently; callers requiring all-or-nothing behavior use a domain-specific admission wrapper.
+Caller-driven injection and current-step context deliberately use different timing. `inject()` joins the next pre-step available and cannot promise that a request already being finalized will consume it. A listener that must affect that exact request returns the context in `PreStepDecision.messages`; downstream rejection or failure then prevents it from materializing.
 
-Cross-session references use that domain composition: TUI prepares the snapshot, then either adds it to the prompt's admission decision outside an acceptance window or injects it beside steering during one. The target log contains two simple messages, so later source mutation cannot change replay and transcript consumers do not need a prompt envelope. This supersedes the attachment mechanism in the [cross-session reference decision](../feature/2026-07-21-cross-session-references.md) while retaining its snapshot and trust-boundary rules.
+Cross-session references use that domain composition: TUI prepares the snapshot, returns it from the idle direct message's pre-step beside that message, or injects it before waking steering during a running turn. The target log contains two simple messages, so later source mutation cannot change replay and transcript consumers do not need a prompt envelope. This supersedes the attachment mechanism in the [cross-session reference decision](../feature/2026-07-21-cross-session-references.md) while retaining its snapshot and trust-boundary rules.
 
 This decision preserves the caller-owned framing decision from [unwrapped injected content](../simplification/2026-07-20-unwrap-injected-content-envelopes.md) and the one-item turn rule from [one send, one turn](../simplification/2026-07-17-one-send-one-turn.md). The later [standalone log-only event decision](../simplification/2026-07-28-remove-synthetic-log-only-turns.md) applies the same execution-only meaning to plugin-owned records.
 
@@ -48,27 +48,27 @@ This decision preserves the caller-owned framing decision from [unwrapped inject
 
 **Keep `SendOptions.contexts` as an atomic attachment.** This preserves all-or-nothing delivery when prompt admission blocks, but it keeps context inside inbox lifecycle state and requires every queue transition and observation event to carry it. The generic agent API should not encode a domain transaction that most callers can express as context injection followed by message delivery.
 
-**Keep a distinct `context/message` session event.** A separate event makes the out-of-turn exception narrower, but user-role model input would again have two event types with identical projection. `user/message.source` already carries the distinction needed by policy, transcript, and replay consumers.
+**Keep a distinct `context/message` session event.** User-role model input would again have two event types with identical projection. `user/message.source` already carries the distinction needed by policy, transcript, and replay consumers.
 
-**Keep one-shot turns for idle injection.** This retains universal turn enclosure and a convenient flush boundary, but it makes turn counts and turn observers report work that never ran the model. Durability is an independent session concern and can be awaited without fabricating execution.
+**Keep one-shot turns for idle injection.** Durable inbox insertion already records idle context without opening a turn. A synthetic turn would make turn counts and observers report work that never ran the model; non-waking context remains pending until real waking work supplies a request.
 
 **Keep `prompt-prefix` as an optional placement.** Prefix baking can make the context and request appear in one provider message, but it introduces a second representation of the direct prompt and spreads placement handling across admission, steering, logging, replay, and UI code. Producers that require textual framing may include it in their own context content.
 
-**Let hooks call `inject()` directly instead of returning additional contexts.** Direct injection would erase the extension point's admission ownership: a listener could append context before a downstream listener blocks the operation. Returning `additionalContexts` keeps the waterfall result authoritative while sharing the same post-admission outbox path.
+**Let prompt hooks call `inject()` instead of returning messages.** An injection may miss the request whose prompt is already being finalized and would escape a downstream block of that decision. Returning the complete message batch keeps current-request context under the waterfall's authority.
 
 ## Verification
 
-- `SendOptions` and steering inbox records contain no attached contexts; `agent/inbox/enqueue` reports only the message plus its resolved queued-or-steering placement.
+- Delivery inputs and steering inbox records contain no attached contexts; `agent/inbox/inserted` reports only the inserted message, while the durable splice retains its target list.
 - `UserMessage` is the shared identified, frozen shape across prompt interception, tool execution, hook bridges, guards, and context producers.
 - Prompt-prefix placement, prompt envelopes, and `context/message` are absent from public types, durable events, projection, and UI replay.
-- Idle `inject()` appends one sourced `user/message` without a turn or model call.
-- Admission-time and active-turn injection drain at safe boundaries after complete tool-result batches and before the request that consumes them.
-- Blocked prompt admission opens no turn and appends neither the prompt nor hook-produced additional contexts; caller context alone falls back to an idle append, while a steering boundary remains available to retry.
-- Unit, persistence/resume, invariant, host/client queue, and TUI coverage pin event order, admission ownership, and reconnect classification.
+- Idle `inject()` immediately appends one durable inbox insertion but no model-visible `user/message`; a later waking delivery may start pre-step processing.
+- Active-turn injection is claimed at the nearest later pre-step boundary, after complete tool-result batches and before the request that consumes it.
+- Rejected or failed pre-step drops its claimed batch; input inserted after the claim remains pending.
+- Unit, persistence/resume, invariant, and TUI coverage pin event order, claim ownership, and durable replay.
 
 ## Consequences
 
-- One surface event is valid outside turns, so persistence scanning, crash repair, forking, compaction, and session queries distinguish execution enclosure from session history.
+- Idle injection is not model-visible until a later pre-step enters it and may be discarded by cancellation or disposal, while its durable inbox lifecycle remains recorded.
 - Consecutive user-role messages replace one baked prompt message; provider adapters preserve that ordering.
-- Outside an acceptance window, `inject()` followed by a blocked `send()` leaves context without its intended direct prompt unless the caller supplies domain-specific admission ownership.
+- Exact-current-request context must be returned from `agent/pre-step`; ordinary injection provides only nearest-later-boundary delivery.
 - The public delivery contract and inbox records remain small: no context attachment, context-placement metadata, prompt envelope, or duplicate durable event type.

.agents/notes/implemented/architecture/2026-07-24-separate-context-injection-from-turn-execution.zh.md

@@ -18,29 +18,29 @@ agent API 曾用三种相互重叠的方式表示面向模型的补充输入：
 
 `inject()` 是调用方交付补充模型输入的唯一操作，而轮次表示一次模型循环执行。
 
-`SendOptions` 只包含 `target` 和 `wakeup`。拥有上下文的调用方通过 `inject()` 交付带标识且冻结的 `UserMessage`，再独立使用 `send()` 或 `steer()` 提交直接消息。
+拥有上下文的调用方通过 `inject()` 交付带标识且冻结的 `UserMessage`，再独立使用 `followup()` 或 `steer()` 提交直接消息。
 
-提示词和工具扩展点仍可返回 `additionalContexts`。这些值是扩展点的输出，而不是从调用方收件箱条目捕获的附件。提示词准入在 `run()` 打开轮次之前执行。获准的提示词及其返回的额外上下文会作为独立消息进入新轮次；提示词被阻止时，两者都不写入，也不打开轮次。工具产生的额外上下文则在对应工具结果之后进入 outbox。
+返回 enter 的 pre-step 会为正在最终确定的请求返回完整的 `PreStepDecision.messages` 批次。工具扩展点仍可返回 `additionalContexts`，这些上下文只会在对应工具结果之后进入 next-step inbox。这些值是扩展点的输出，而不是从调用方 inbox 条目捕获的附件。
 
-每项额外上下文都是独立的 `user/message`，并由 `source` 记录来源。不再有 `context/message`、prompt-prefix 放置方式、稳定请求分隔符或提示词封套。transcript 与 UI 消费方通过 `source` 区分直接用户消息和注入上下文。
+每项额外上下文都是独立的 `UserMessage`，并由 `source` 记录来源。inbox 插入会立即持久化；后续准入会将同一个值记录为 `user/message`。不再有 `context/message`、prompt-prefix 放置方式、稳定请求分隔符或提示词封套。transcript 与 UI 消费方通过 `source` 区分直接用户消息和注入上下文。
 
 ## 注入生命周期
 
-提示词准入期间或轮次处于打开状态时，`inject()` 会将上下文暂存在 loop outbox 中。私有的 next-step 接受窗口在 `agent/prompt-submit` 前打开，并在 `turn/end` 前关闭，因此同一边界接受的 steering 和上下文会进入同一个后续请求，而 `turn/end` 监听器提交的晚到 steering 则成为排队提示词。agent loop 会在安全的步骤边界排空 outbox，同时保持工具协议要求的相邻关系：在助手工具调用批次期间接受的上下文，只能出现在该批次所有有序结果之后。
+`inject()` 始终把上下文插入不会唤醒的 `next-step` inbox，并以 `agent/inbox/spliced` 提交该队列变更。运行中的驱动器会在最近的后续 pre-step 边界领取它。idle 驱动器会让它保持待处理，直至 `followup()` 或 `steer()` 提供可唤醒工作；在此之前，取消或 dispose（资源释放）可能将其丢弃，但不会抹除持久队列历史。
 
-在该窗口之外，`inject()` 会立即追加对应的 `user/message`。它不会增加轮次编号、发出 `turn/start` 或 `turn/end`、改变 agent 状态，也不会运行模型；持久化通过 `session/event` 观察这次追加。
+循环会先领取当前 next-step 批次，再运行 `agent/pre-step`，因此领取后到达的注入可能赶不上正在最终确定的请求，而由下一次边界领取。enter decision 返回的消息会在所属轮次内、消费它们的请求之前追加。在助手工具调用批次期间产生的上下文因此只会出现在该批次全部有序结果之后。
 
-如果提示词准入被阻止或失败，调用方暂存的仅含上下文的批次会立即追加，且不产生轮次。steering 及与其一同暂存的上下文会留在 outbox 中，供后续获准提示词使用；取消或 dispose（资源释放）可能丢弃它们。钩子产生的 `additionalContexts` 属于被拒绝的准入决策，因此永远不会落入日志。
+如果 pre-step reject 或抛错，其已领取的注入上下文、steering 与排队提示词都会保持已删除，也不会追加返回批次。原子领取后插入的消息不受影响，继续保持待处理。
 
-会话不变量允许 `user/message` 位于两个轮次之间，同时继续要求核心执行事件、steering、助手输出和工具事件均受轮次边界约束。可通过声明合并来扩展的事件关系由声明它们的插件拥有，而不是采用核心默认规则。持久化、崩溃恢复、会话恢复、fork 和压缩（compaction）会把合法的轮次间事件当作已提交会话历史，而不是中断轮次或可丢弃的日志尾部。
+loop 只会在轮次内从进入步骤的批次追加注入的 `user/message`。核心执行事件、steering、助手输出和工具事件仍受轮次边界约束；可合并扩展事件的关系由声明它们的插件拥有，而不是采用核心默认规则。
 
 ## 扩展点与调用方语义
 
-`PromptDecision.content` 仍只替换直接提示词。`PromptDecision.additionalContexts` 和工具结果的 `additionalContexts` 保留 FIFO 顺序及各自来源，但不再选择放置方式。waterfall（瀑布式事件）监听器调用 `next()` 委托时，必须保留下游返回的提示词内容和额外上下文，除非它有意返回替代值。
+enter 分支的 `PreStepDecision.messages` 是拟议步骤的完整批次。waterfall（瀑布式事件）监听器调用 `next()` 委托时，会保留下游消息，除非有意替换；新增消息遵循 waterfall 的自然返回顺序。工具结果的 `additionalContexts` 保留 FIFO 顺序及各自来源。
 
-调用方主动注入与钩子产生的额外上下文具有不同的准入归属。钩子的额外上下文只会在该钩子允许提示词或工具结果后落入日志。在 next-step 接受窗口之外，调用方执行 `inject(context)` 后再执行 `send(prompt)` 时，会独立提交上下文；需要全有或全无语义的调用方应使用领域专用的准入包装层。
+调用方主动注入与当前步骤上下文刻意采用不同的时序。`inject()` 会加入下一个可用 pre-step，无法保证正在最终确定的请求会消费它。必须影响该请求的监听器在 `PreStepDecision.messages` 中返回上下文；下游 reject 或失败时，该上下文不会落入日志。
 
-跨会话引用采用这种领域组合方式：TUI 先准备快照，然后在接受窗口之外将其加入提示词准入决策，或在窗口期间将其注入到 steering 旁。目标日志包含两条简单消息，因此来源会话后续变化不会改变回放，transcript 消费方也不需要提示词封套。本决策取代[跨会话引用决策](../feature/2026-07-21-cross-session-references.md)中的附件机制，但保留其快照与信任边界规则。
+跨会话引用采用这种领域组合方式：TUI 先准备快照，然后在 idle 直接消息的 pre-step 中把快照与该消息一同返回，或在 running 轮次中先注入快照再唤醒 steering。目标日志包含两条简单消息，因此来源会话后续变化不会改变回放，transcript 消费方也不需要提示词封套。本决策取代[跨会话引用决策](../feature/2026-07-21-cross-session-references.md)中的附件机制，但保留其快照与信任边界规则。
 
 本决策保留[移除注入内容封套](../simplification/2026-07-20-unwrap-injected-content-envelopes.md)确立的由调用方决定内容框架的原则，以及[一次 send、一个轮次](../simplification/2026-07-17-one-send-one-turn.md)确立的单条目轮次规则。后续的[独立纯日志事件决策](../simplification/2026-07-28-remove-synthetic-log-only-turns.md)将同样的「轮次仅表示执行」语义应用于插件所属记录。
 
@@ -48,27 +48,27 @@ agent API 曾用三种相互重叠的方式表示面向模型的补充输入：
 
 **保留 `SendOptions.contexts` 作为原子附件。** 提示词准入阻止消息时，这种方式能保留全有或全无交付，但也会让上下文继续成为收件箱生命周期状态的一部分，并迫使每次队列转换和观察事件携带它。大多数调用方都可以通过先注入上下文、再交付消息来表达需求，通用 agent API 不应内置领域事务。
 
-**保留独立的 `context/message` 会话事件。** 独立事件可以缩小轮次外事件的例外范围，但面向模型的 user-role 输入会再次拥有两个投影完全相同的事件类型。`user/message.source` 已能为策略、transcript 和回放消费方提供所需区分。
+**保留独立的 `context/message` 会话事件。** 面向模型的 user-role 输入会再次拥有两个投影完全相同的事件类型。`user/message.source` 已能为策略、transcript 和回放消费方提供所需区分。
 
-**为空闲注入保留一次性轮次。** 这种方式能保留通用轮次封闭和方便的刷新边界，却会让轮次计数与轮次观察方报告从未运行模型的工作。持久性是独立的会话关注点，无需伪造执行即可等待。
+**为空闲注入保留一次性轮次。** 持久 inbox 插入已经能在不打开轮次的情况下记录空闲上下文。合成轮次会让轮次计数与观察方报告从未运行模型的工作；不会唤醒的上下文会保持待处理，直至真实的可唤醒工作提供请求。
 
 **保留 `prompt-prefix` 可选放置方式。** 前缀烘焙可以让上下文和请求位于同一条提供方消息中，但它会引入直接提示词的第二种表示，并把放置处理扩散到准入、steering、日志、回放和 UI 代码。需要文本框架的生产方可以直接把它写入自身上下文内容。
 
-**让钩子直接调用 `inject()`，而不是返回额外上下文。** 直接注入会破坏扩展点的准入归属：下游监听器阻止操作之前，上游监听器就可能已经追加上下文。返回 `additionalContexts` 能维持 waterfall 结果的最终权威性，同时复用准入后的 outbox 路径。
+**让提示词钩子调用 `inject()`，而不是返回消息。** 注入可能赶不上提示词正在最终确定的请求，也会逃逸下游对该 decision 的阻止。返回完整消息批次能让当前请求上下文继续受 waterfall 约束。
 
 ## 验证
 
-- `SendOptions` 与 steering 收件箱记录不包含附加上下文；`agent/inbox/enqueue` 只报告消息及其已解析的 queued 或 steering 放置方式。
+- 投递输入与 steering inbox 记录不包含附加上下文；`agent/inbox/inserted` 只报告插入消息，目标列表由持久 splice 保留。
 - `UserMessage` 是提示词拦截、工具执行、hook bridge、guard 和上下文生产方共享的带标识且冻结的形状。
 - 公共类型、持久事件、投影和 UI 回放中均不存在 prompt-prefix 放置方式、提示词封套与 `context/message`。
-- 空闲 `inject()` 在不产生轮次或模型调用的情况下，追加一条带来源的 `user/message`。
-- 准入期间和活跃轮次中的注入会在完整工具结果批次之后的安全边界排空，并在消费它们的请求之前进入日志。
-- 被阻止的提示词准入不会打开轮次，也不会追加提示词或钩子产生的额外上下文；仅有调用方上下文时会回退为空闲追加，而带 steering 的边界仍可重试。
-- 单元测试、持久化与恢复测试、不变量测试、宿主／客户端队列测试和 TUI 覆盖会固定事件顺序、准入归属和重连分类。
+- idle 状态下的 `inject()` 会立即追加一条持久 inbox 插入记录，但不会追加模型可见的 `user/message`；后续可唤醒投递可能开始 pre-step 处理。
+- 活跃轮次中的注入会在最近的后续 pre-step 边界领取，并位于完整工具结果批次之后、消费它的请求之前。
+- pre-step reject 或失败会丢弃其已领取批次；领取后插入的 inbox 工作继续保持待处理。
+- 单元测试、持久化与 resume 测试、不变量测试和 TUI 覆盖会固定事件顺序、领取归属和持久回放。
 
 ## 后果
 
-- 一个表层事件可以合法位于轮次之外，因此持久化扫描、崩溃恢复、fork、压缩和会话查询需要区分执行封闭与会话历史。
+- idle 注入要到后续 pre-step 让它进入步骤后才会对模型可见，并可能被取消或 dispose 丢弃，而其持久 inbox 生命周期仍会保留记录。
 - 两条连续的 user-role 消息会取代一条烘焙后的提示词消息；提供方适配器会保留这一顺序。
-- 在接受窗口之外，`inject()` 后跟一个被阻止的 `send()` 会留下缺少预期直接提示词的上下文，除非调用方提供领域专用的准入归属。
+- 必须影响当前请求的上下文要从 `agent/pre-step` 返回；普通注入只保证由最近的后续边界交付。
 - 公共投递契约和收件箱记录保持精简：没有上下文附件、上下文放置元数据、提示词封套或重复的持久事件类型。

.agents/notes/implemented/architecture/2026-07-25-web-client-session-scope-and-provide-channel.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/architecture/2026-07-25-web-client-session-scope-and-provide-channel.md
-2026-07-25-web-client-session-scope-and-provide-channel.md: 353cf35c9d6f5fa93a97fb0be60303ad6cef4d14
-2026-07-25-web-client-session-scope-and-provide-channel.zh.md: b6d1a20073e1a43414d43b830ccc8ac2b1573bb2
+2026-07-25-web-client-session-scope-and-provide-channel.md: aeefbe22a397e3d7ffb9f6427a3c70c8c8e8b940
+2026-07-25-web-client-session-scope-and-provide-channel.zh.md: 056d50d45cef891e0e635d8bb4f2e73064ccdb87

.agents/notes/implemented/architecture/2026-07-25-web-client-session-scope-and-provide-channel.md

@@ -4,7 +4,7 @@ Status: implemented
 
 English | [中文](2026-07-25-web-client-session-scope-and-provide-channel.zh.md)
 
-> Scope: the client Agent scope (actx) and targeted events, the client/host materialization parity model, the blank-session bit and reuse (`connectWorkspace`), the per-session provisioning channel (`sessions.provide`), the read-only queue mirror (`session/queued`), and the host wire smalls that carry these capabilities (the summary `blank` column, the `host/session-added` frame field, and the `host/commands-changed` frame). The input state machine and the slash pipeline live in the [input machine note](2026-07-25-web-input-machine-and-slash-pipeline.md); the command business surfaces live in the [command surfaces note](2026-07-25-web-command-surfaces-and-assembly.md).
+> Scope: the client Agent scope (actx) and targeted events, the client/host materialization parity model, the blank-session bit and reuse (`connectWorkspace`), the per-session provisioning channel (`sessions.provide`), and the host wire smalls that carry these capabilities (the summary `blank` column, the `host/session-added` frame field, and the `host/commands-changed` frame). The input state machine and the slash pipeline live in the [input machine note](2026-07-25-web-input-machine-and-slash-pipeline.md); the command business surfaces live in the [command surfaces note](2026-07-25-web-command-surfaces-and-assembly.md).
 
 ## Problem
 
@@ -101,7 +101,6 @@ Slot scope is the closed set `root | session-maybe | session`:
 
 ### The read-only queue mirror
 
-- The MuxFrame `session/queued`: the Session holds a read-only inbox mirror (previews truncated; steering retired by source match). The host stamps the agent-loop's acceptance-time steering classification on live and replayed frames, so a reconnect baseline does not depend on replaying an earlier `turn/start`. Queue frames never enter history — pure stream state, cleared on reconnect and refilled from the new baseline; the never-instantiated window is buffered and replayed through the manager pendingBuffers.
 - Queue semantics: running does not lock input; ordinary messages queue through `session.prompt {mode:'queue'}`, and commands never queue.
 
 ### Host wire smalls

.agents/notes/implemented/architecture/2026-07-25-web-client-session-scope-and-provide-channel.zh.md

@@ -4,7 +4,7 @@ Status: implemented
 
 [English](2026-07-25-web-client-session-scope-and-provide-channel.md) | 中文
 
-> 范围：client Agent scope（actx）与定向事件、client/host 实体化对等模型、空会话 blank 位与复用（`connectWorkspace`）、逐会话供数通道（`sessions.provide`）、队列只读镜像（`session/queued`），以及承载这些能力的 host wire 小件（summary `blank` 列、`host/session-added` 帧字段、`host/commands-changed` 帧）。输入状态机与 slash 管线见[输入状态机 note](2026-07-25-web-input-machine-and-slash-pipeline.md)；命令业务面见[命令业务面 note](2026-07-25-web-command-surfaces-and-assembly.md)。
+> 范围：client Agent scope（actx）与定向事件、client/host 实体化对等模型、空会话 blank 位与复用（`connectWorkspace`）、per-session 供数通道（`sessions.provide`），以及承载这些能力的 host wire 小件（summary `blank` 列、`host/session-added` 帧字段、`host/commands-changed` 帧）。输入状态机与 slash 管线见[输入状态机 note](2026-07-25-web-input-machine-and-slash-pipeline.md)；命令业务面见[命令业务面 note](2026-07-25-web-command-surfaces-and-assembly.md)。
 
 ## 问题
 
@@ -101,7 +101,6 @@ slot scope 是闭集 `root | session-maybe | session`：
 
 ### 队列只读镜像
 
-- MuxFrame `session/queued`：Session 持只读 inbox 镜像（预览截断、steering（中途引导）按 source 匹配退休）。宿主会在实时和回放帧中标记 agent loop（智能体循环）接受消息时的 steering 分类，因此重连基线不依赖回放更早的 `turn/start`。queue 帧不进 history，纯流状态——重连清空、新基线重灌；未实例化窗口经 manager pendingBuffers 缓冲重放。
 - 队列语义：running 不锁输入；普通消息经 `session.prompt {mode:'queue'}` 排队，命令永不排队。
 
 ### host wire 小件

.agents/notes/implemented/architecture/2026-07-28-identified-immutable-message-values.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/architecture/2026-07-28-identified-immutable-message-values.md
-2026-07-28-identified-immutable-message-values.md: cdb0f1aadc4796b5aa0642a3994d3e3e4ab67bd9
-2026-07-28-identified-immutable-message-values.zh.md: b179feb5d0648706293048e1131df2a954b0d511
+2026-07-28-identified-immutable-message-values.md: 472de8b133ba323c3e1ff5e53c8dacb3d66525c5
+2026-07-28-identified-immutable-message-values.zh.md: 6082fd5fb65f4acd0759a6c4b49be8110a559a73

.agents/notes/implemented/architecture/2026-07-28-identified-immutable-message-values.md

@@ -12,15 +12,15 @@ This made identity a routing side effect rather than a message invariant. Produc
 
 ## Decision
 
-`@deepseek-ai/dsh-llm` owns one `Message` value with required `id`, `role`, `content`, and `source`. `MessageId` is opaque and shared by user, assistant, and tool-result messages. A message receives its id at creation, before routing, prompt admission, durable append, or request projection. The same id survives every representation boundary.
+`@deepseek-ai/dsh-llm` owns one `Message` value with required `id`, `role`, `content`, and `source`. `MessageId` is opaque and shared by user, assistant, and tool-result messages. A message receives its id at creation, before inbox routing, claim, pre-step rewriting, durable append, or request projection. The same id survives every representation boundary.
 
 `createMessage(input)` is the canonical role-generic creation boundary. It mints a `MessageId`, detaches the supplied role, content, and source, and deep-freezes the complete value before returning it. `createUserMessage({ content, source })` fixes the user role for prompt and context producers. `createAssistantMessage({ content, source })` fixes both the assistant role and the model source kind, so model-output producers supply only content and model provenance. All creation helpers exclude an input id so callers cannot accidentally present creation as import. `freezeMessage(message)` is the separate import or transformation boundary: it detaches and deep-freezes a message whose identity already exists, without minting a replacement.
 
 The helpers live in `dsh-llm` beside the base message vocabulary because their complete contracts depend only on that vocabulary. `createToolResultMessage()` belongs with the other creation helpers: it couples a tool call id to the exact user-role tool-result block and source without depending on session state or events. `dsh-session` consumes complete messages rather than owning their construction.
 
-The `Agent` interface accepts a complete `UserMessage`. `send`, `followup`, `steer`, and `inject` never allocate or return identity; they freeze an imported value whose id the caller already holds. Prompt admission receives that message directly. A content rewrite creates a frozen replacement with the same id, while an additional context is a separately created `UserMessage` with its own id.
+The `Agent` interface accepts a complete `UserMessage` through `followup`, `steer`, and `inject`. These operations never allocate or return identity; they freeze an imported value whose id the caller already holds. Inbox claims and `agent/pre-step` receive that message directly. A content rewrite creates a frozen replacement with the same id, while an additional context is a separately created `UserMessage` with its own id.
 
-Durable message-producing events store complete messages. `user/message` stores its `UserMessage` directly; `assistant/message`, `tool/result`, and `steering/message` wrap their role-specialized message beside event-local position, usage, failure, or presentation facts. Session derivation returns those frozen values instead of reconstructing anonymous messages. Assistant assembly creates a model-sourced message when a response completes, and tool execution creates a tool-sourced message when a result is committed.
+Durable message-producing events store complete messages. `user/message` stores its `UserMessage` directly; `assistant/message` and `tool/result` wrap their role-specialized message beside event-local position, usage, failure, or presentation facts. Session derivation returns those frozen values instead of reconstructing anonymous messages. Assistant assembly creates a model-sourced message when a response completes, and tool execution creates a tool-sourced message when a result is committed.
 
 Any operation that changes only the representation of an existing semantic message preserves its id and returns another frozen value. An operation that creates a new semantic message mints a new id. Compaction content rewrites therefore preserve the rewritten tool-result identity, while a summary checkpoint is a new message.
 
@@ -28,7 +28,7 @@ Any operation that changes only the representation of an existing semantic messa
 
 **Keep ids optional on the base message.** This would minimize fixture migration and allow provider or persistence shapes to remain anonymous. It would also preserve the original ambiguity: every consumer would need to branch on whether identity exists, and no type would prove that admission, logging, or projection retained it.
 
-**Let `Agent.send()` allocate the id.** This keeps identity scoped to inbox correlation but makes the agent call the earliest point at which a producer can name its own message. Prompt construction, UI attachments, and synchronous enqueue/discard coordination then need content matching or an out-of-band token before `send()` returns.
+**Let agent delivery allocate the id.** This keeps identity scoped to inbox correlation but makes the agent call the earliest point at which a producer can name its own message. Prompt construction, UI attachments, and synchronous enqueue/discard coordination then need content matching or an out-of-band token before delivery returns.
 
 **Let each durable event allocate a new id.** This gives persisted messages identities but deliberately breaks correlation with the live input and makes replayed requests appear to contain different messages. Identity belongs to the semantic value, not to each envelope that carries it.
 
@@ -38,7 +38,7 @@ Any operation that changes only the representation of an existing semantic messa
 
 Every message producer must choose creation or import explicitly, and tests construct complete values rather than partial content/source records. UUID generation moves outward to the first semantic creation point, so deterministic fixtures that provide an existing id use `freezeMessage()` instead of `createMessage()`.
 
-Live inbox events, durable events, derived history, and model requests can correlate one message without content equality or envelope-specific ids. Prompt admission and UI attachment cleanup can compare `MessageId` before a turn exists. Deep freezing prevents a producer, hook, or observer from changing the value after identity is established.
+Live inbox events, durable events, derived history, and model requests can correlate one message without content equality or envelope-specific ids. Pending-input policy and UI attachment cleanup can compare `MessageId` before a turn exists, while claims retain that identity inside the open turn. Deep freezing prevents a producer, hook, or observer from changing the value after identity is established.
 
 The shared representation removes the old `UserMessageData`/`AgentMessage` split and folds provider provenance into typed message sources. Event envelopes still own facts that are not message semantics, such as turn and step position, token usage, internal tool failure identity, and presentation metadata.
 
@@ -46,5 +46,5 @@ The message and helper unit tests pin immediate identity, detachment, deep immut
 
 ## Related
 
-- [Unify agent delivery on send(target × wakeup) and coalesce injected context into user/message](2026-07-22-unified-send-and-coalesced-user-messages.md) — this note supersedes its input-representation and agent-assigned-id details while retaining its routing decision.
+- [Unified agent delivery routing and coalesced injected context](2026-07-22-unified-send-and-coalesced-user-messages.md) — this note supersedes its input-representation and agent-assigned-id details while retaining its routing decision.
 - [Reconstructable requests](2026-07-05-reconstructable-requests.md) — the session log remains the authority for every model-visible input.

.agents/notes/implemented/architecture/2026-07-28-identified-immutable-message-values.zh.md

@@ -12,15 +12,15 @@ harness 曾存在多种形似消息的表示，各自采用不同的标识规则
 
 ## 决策
 
-`@deepseek-ai/dsh-llm` 持有唯一一种 `Message` 值，其 `id`、`role`、`content` 和 `source` 均为必填。`MessageId` 是不透明标识，由用户消息、assistant 消息和工具结果消息共享。消息在创建时就会获得 id，早于路由、提示词准入、持久追加或请求投影。同一个 id 会跨越每个表示边界。
+`@deepseek-ai/dsh-llm` 持有唯一一种 `Message` 值，其 `id`、`role`、`content` 和 `source` 均为必填。`MessageId` 是不透明标识，由用户消息、assistant 消息和工具结果消息共享。消息在创建时就会获得 id，早于 inbox 路由、领取、pre-step 改写、持久追加或请求投影。同一个 id 会跨越每个表示边界。
 
 `createMessage(input)` 是角色通用的规范创建边界。它会生成 `MessageId`，将传入的角色、内容和来源与调用方对象解除引用关系，并在返回完整值前将其深度冻结。`createUserMessage({ content, source })` 为提示词和上下文生产方固定 user 角色。`createAssistantMessage({ content, source })` 同时固定 assistant 角色与模型来源类别，因此模型输出生产方只需提供内容和模型溯源信息。所有创建辅助函数的输入都不包含 id，因此调用方不会意外地把新消息的创建伪装成已有消息的导入。`freezeMessage(message)` 是独立的导入或转换边界：它会将已有标识的消息与调用方对象解除引用关系并深度冻结，不会生成替代标识。
 
 这些辅助函数位于基础消息词汇旁的 `dsh-llm` 中，因为它们的完整契约只依赖该词汇。`createToolResultMessage()` 与其他创建辅助函数同属此处：它使用同一个工具调用 id，将工具来源与确切的 user-role 工具结果块耦合起来，不依赖会话状态或事件。`dsh-session` 只消费完整消息，不负责构造它们。
 
-`Agent` 接口接收完整的 `UserMessage`。`send`、`followup`、`steer` 和 `inject` 绝不会分配或返回标识；它们会冻结导入的值，而调用方已经持有该值的 id。提示词准入会直接接收该消息。改写内容时会创建具有相同 id 的冻结替代值，而每个附加上下文都是单独创建的 `UserMessage`，拥有自己的 id。
+`Agent` 接口通过 `followup`、`steer` 和 `inject` 接收完整的 `UserMessage`。这些操作绝不会分配或返回标识；它们会冻结导入的值，而调用方已经持有该值的 id。inbox 领取和 `agent/pre-step` 会直接接收该消息。改写内容时会创建具有相同 id 的冻结替代值，而每个附加上下文都是单独创建的 `UserMessage`，拥有自己的 id。
 
-产生持久消息的事件会存储完整消息。`user/message` 直接存储其 `UserMessage`；`assistant/message`、`tool/result` 和 `steering/message` 则将各自角色专用的消息与事件本地的位置、用量、失败或呈现事实包装在一起。会话派生会返回这些冻结值，而不是重建匿名消息。assistant 组装会在响应完成时创建模型来源的消息，工具执行会在提交结果时创建工具来源的消息。
+产生持久消息的事件会存储完整消息。`user/message` 直接存储其 `UserMessage`；`assistant/message` 和 `tool/result` 则将各自角色专用的消息与事件本地的位置、用量、失败或呈现事实包装在一起。会话派生会返回这些冻结值，而不是重建匿名消息。assistant 组装会在响应完成时创建模型来源的消息，工具执行会在提交结果时创建工具来源的消息。
 
 仅改变已有语义消息表示的操作会保留其 id，并返回另一个冻结值。创建新语义消息的操作则会生成新 id。因此，压缩（compaction）中的内容改写会保留被改写工具结果的标识，而摘要检查点是一条新消息。
 
@@ -28,7 +28,7 @@ harness 曾存在多种形似消息的表示，各自采用不同的标识规则
 
 **让基础消息的 id 保持可选。** 这能减少 fixture（测试前置数据）迁移，并允许提供方或持久化形状继续保持匿名，但也会保留原有歧义：每个消费方都必须根据标识是否存在执行分支，且没有任何类型能证明准入、记录或投影保留了标识。
 
-**让 `Agent.send()` 分配 id。** 这会将标识限定在 inbox 关联范围内，却也会让 agent 调用成为生产方可以标识自身消息的最早时机。这样一来，在 `send()` 返回前，提示词构造、UI 附件和同步入队／丢弃协调都需要进行内容匹配，或使用带外 token。
+**让 agent 交付分配 id。** 这会将标识限定在 inbox 关联范围内，却也会让 agent 调用成为生产方可以标识自身消息的最早时机。这样一来，在交付返回前，提示词构造、UI 附件和同步入队／丢弃协调都需要进行内容匹配，或使用带外 token。
 
 **让每个持久事件分配新 id。** 这能为持久消息提供标识，却会有意切断它与实时输入的关联，并让回放请求表现得像包含了不同消息。标识属于语义值，而不是承载它的每个封装。
 
@@ -38,7 +38,7 @@ harness 曾存在多种形似消息的表示，各自采用不同的标识规则
 
 每个消息生产方都必须显式选择创建或导入，测试也会构造完整值，而不是不完整的内容／来源记录。UUID 的生成会前移至最初的语义创建点，因此提供已有 id 的确定性 fixture 会使用 `freezeMessage()`，而不是 `createMessage()`。
 
-实时 inbox 事件、持久事件、派生历史和模型请求可以关联同一条消息，无需比较内容或使用封装专用 id。提示词准入和 UI 附件清理可以在轮次存在之前比较 `MessageId`。深度冻结可以防止生产方、钩子或观察方在标识建立后更改消息值。
+实时 inbox 事件、持久事件、派生历史和模型请求可以关联同一条消息，无需比较内容或使用封装专用 id。待处理输入策略和 UI 附件清理可以在轮次存在之前比较 `MessageId`，领取后则会在已打开的轮次内保留该标识。深度冻结可以防止生产方、钩子或观察方在标识建立后更改消息值。
 
 共享表示移除了旧的 `UserMessageData`/`AgentMessage` 划分，并将提供方溯源信息纳入带类型的消息来源。事件封装仍持有不属于消息语义的事实，例如轮次与步骤位置、token 用量、内部工具失败标识和呈现元数据。
 
@@ -46,5 +46,5 @@ harness 曾存在多种形似消息的表示，各自采用不同的标识规则
 
 ## 相关
 
-- [统一通过 send(target × wakeup) 交付 agent 消息，并将注入上下文合并到 user/message](2026-07-22-unified-send-and-coalesced-user-messages.md)——本记录取代其中的输入表示和由 agent 分配 id 的细节，同时保留其路由决策。
+- [统一 agent 交付路由，并合并注入上下文](2026-07-22-unified-send-and-coalesced-user-messages.md)——本记录取代其中的输入表示和由 agent 分配 id 的细节，同时保留其路由决策。
 - [可重建的请求](2026-07-05-reconstructable-requests.md)——会话日志仍是每项模型可见输入的权威来源。

.agents/notes/implemented/architecture/2026-07-29-terminal-llm-stream-failures.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/architecture/2026-07-29-terminal-llm-stream-failures.md
+2026-07-29-terminal-llm-stream-failures.md: 1e26973360f07c212016c6a44103448a3510a75b
+2026-07-29-terminal-llm-stream-failures.zh.md: d3eeb0534f6cb8d4d1ad167cca089314eb02b513

.agents/notes/implemented/architecture/2026-07-29-terminal-llm-stream-failures.md

@@ -0,0 +1,37 @@
+# Agent Note: Terminal LLM stream failures
+
+Status: implemented
+
+English | [中文](2026-07-29-terminal-llm-stream-failures.zh.md)
+
+This note supersedes only the thrown-error identity and call-local sidecar mechanism in [bounded LLM request recovery](2026-06-21-bounded-llm-request-recovery.md) and [after-call context-overflow recovery](2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md). Those notes continue to own structured failure facts, retry policy, durable attempts, and compaction recovery.
+
+## Problem
+
+An adapter failure had two public representations: an exception from selection, dispatch, iterator construction, or iteration, and an in-band `finish { kind: 'error' | 'aborted' }`. `LlmService` tagged thrown objects in a stream-keyed sidecar so the agent loop could distinguish them from middleware and consumer failures. The consumer still needed a catch around iteration, signal checks, chunk logging, and assembly; correctness therefore depended on proving which statement threw and consulting metadata attached to the exact returned iterable.
+
+Retry policy had the same indirect ownership. It was discovered through the stream sidecar after dispatch even though `prepareCall()` had already captured the serving registration. A wrapper-owned route and an adapter-owned route consequently shared one opaque lookup API despite having different authority.
+
+## Decision
+
+`LlmService` is the normalization boundary for one adapter attempt. It catches only final-adapter selection, synchronous dispatch, iterator construction, and `next()` failures, converts the thrown value to immutable `LlmFailure`, and emits one terminal `finish`. Caller cancellation or an `ABORTED` failure selects the aborted reason; every other adapter failure selects error. An adapter may also emit either terminal reason directly.
+
+The adapter-owned catch ends before each yielded chunk. Errors from `llm/stream` middleware, nested calls, adapter cleanup, chunk consumers, logging, signal checks, and assembly remain thrown as defects or lifecycle failures; they never enter model-request recovery. A transport failure after partial deltas may leave blocks open, so the stream invariant permits open blocks only for terminal error or aborted finishes. No assistant message or tool call is assembled from that incomplete output.
+
+`PreparedLlmCall` exposes the immutable retry policy captured with its config and registration. One-shot reuse and config mismatch remain synchronous `INVALID_PREPARED_CALL` misuse errors. A route served entirely by `llm/stream` middleware has no prepared registration and therefore no serving policy.
+
+The agent loop consumes one failure representation. It iterates and logs chunks without a classification catch, inspects the terminal finish, and passes its failure facts plus the prepared policy to `agent/request-error`. The public `isLlmAdapterFailure`, `llmFailureOf`, and `llmRetryPolicyOf` sidecar APIs are absent.
+
+## Alternatives considered
+
+**Keep call-local error tagging.** This preserves thrown object identity, but makes every consumer catch a region containing its own fallible work and couples classification to the identity of an iterable wrapper. The original error object has no durable role in recovery; normalized facts are the useful boundary value.
+
+**Require every adapter to emit failure chunks and forbid throws.** Library iterators, transports, and JavaScript dispatch can still throw. Requiring every adapter to reproduce the same catch boundary duplicates ownership and does not protect a direct `LlmService` consumer from an incomplete implementation.
+
+**Catch every iteration error in the agent loop.** The loop cannot reliably distinguish provider failure from middleware, session append, cancellation, or assembly failure without restoring the same sidecar provenance mechanism. Classification belongs where the adapter call is made.
+
+**Return a `Result` before streaming.** A pre-stream result cannot represent a transport failure after partial output without adding a second response lifecycle. The existing terminal chunk already represents both early and late attempt outcomes.
+
+## Consequences
+
+All `LlmService.stream()` consumers receive adapter operational failures through one typed terminal protocol, while programming and lifecycle failures retain ordinary exception semantics. Recovery gives up exact thrown-object identity and exposes only detached provider-neutral facts. The stream service owns slightly more adapter plumbing, but consumers delete provenance catches and stream-keyed metadata. Prepared calls carry their policy explicitly, and middleware-only routing remains visibly policy-free.

.agents/notes/implemented/architecture/2026-07-29-terminal-llm-stream-failures.zh.md

@@ -0,0 +1,37 @@
+# Agent Note: LLM 流的终止失败
+
+Status: implemented
+
+[English](2026-07-29-terminal-llm-stream-failures.md) | 中文
+
+本说明仅取代[有界 LLM 请求恢复](2026-06-21-bounded-llm-request-recovery.md)与[调用后上下文溢出恢复](2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md)中关于抛出错误身份和调用局部 sidecar 的机制。上述说明继续规定结构化失败事实、重试策略、持久尝试与压缩恢复。
+
+## Problem
+
+适配器失败曾有两种公共表示：选择、分发、iterator 构造或迭代抛出的异常，以及带内的 `finish { kind: 'error' | 'aborted' }`。`LlmService` 会在以 stream 为 key 的 sidecar 中标记抛出对象，使 agent loop 能将其与 middleware 和消费方失败区分开。消费方仍需用 catch 包围迭代、signal 检查、chunk 日志记录和组装；正确性因此取决于证明是哪条语句抛错，并查询附着于精确返回 iterable 的元数据。
+
+重试策略也采用同样的间接归属。尽管 `prepareCall()` 已捕获服务注册，策略仍要在分发后通过 stream sidecar 查找。因此，由 wrapper 提供服务的路由与由适配器提供服务的路由共用一个不透明查询 API，尽管两者的权威不同。
+
+## Decision
+
+`LlmService` 是一次适配器尝试的规范化边界。它只捕获最终适配器选择、同步分发、iterator 构造与 `next()` 失败，将抛出值转换为不可变 `LlmFailure`，并发出一个终止 `finish`。调用方取消或 `ABORTED` 失败选择 aborted reason；其他适配器失败选择 error。适配器也可以直接发出这两种终止 reason。
+
+适配器所属的 catch 会在每个 chunk 被 yield 前结束。来自 `llm/stream` middleware、嵌套调用、适配器清理、chunk 消费方、日志记录、signal 检查与组装的错误仍作为缺陷或生命周期失败抛出；它们绝不进入模型请求恢复。部分 delta 之后的传输失败可能留下未关闭块，因此流 invariant 只允许终止 error 或 aborted finish 带有未关闭块。不会从这些不完整输出组装 assistant 消息或工具调用。
+
+`PreparedLlmCall` 公开随其配置和注册捕获的不可变重试策略。一次性句柄复用与配置不匹配仍是同步的 `INVALID_PREPARED_CALL` 误用错误。完全由 `llm/stream` middleware 提供服务的路由没有准备完成的注册，因此也没有服务策略。
+
+agent loop 只消费一种失败表示。它不再使用分类 catch，而是直接迭代并记录 chunk、检查终止 finish，再把其中的失败事实与准备完成的策略传给 `agent/request-error`。公共的 `isLlmAdapterFailure`、`llmFailureOf` 和 `llmRetryPolicyOf` sidecar API 不再存在。
+
+## Alternatives considered
+
+**保留调用局部错误标记。** 这会保留抛出对象身份，但要求每个消费方捕获一段包含自身易失败工作的区域，并让分类依赖 iterable wrapper 的身份。原始错误对象在持久恢复中没有作用；规范化事实才是有用的边界值。
+
+**要求所有适配器发出失败 chunk，并禁止抛出。** 库 iterator、transport 与 JavaScript 分发仍可能抛错。要求每个适配器复制同一 catch 边界会重复归属，也无法保护 `LlmService` 的直接消费方免受不完整实现影响。
+
+**在 agent loop 中捕获所有迭代错误。** 如果不恢复同一套 sidecar 溯源机制，loop 无法可靠区分提供方失败与 middleware、session append、取消或组装失败。分类属于发起适配器调用的边界。
+
+**在流式输出前返回 `Result`。** 流前结果无法表示部分输出之后的传输失败，除非增加第二套响应生命周期。现有终止 chunk 已能表示早期和后期尝试结果。
+
+## Consequences
+
+所有 `LlmService.stream()` 消费方都通过一种带类型的终止协议接收适配器运行失败，而编程与生命周期失败保留普通异常语义。恢复放弃精确抛出对象身份，只暴露与原对象分离的提供方无关事实。流服务承担略多的适配器管道工作，但消费方删除了溯源 catch 与以 stream 为 key 的元数据。准备完成的调用显式携带策略，而仅由 middleware 路由的调用仍明确没有策略。

.agents/notes/implemented/architecture/2026-07-30-followup-enqueue-and-owned-runs.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/architecture/2026-07-30-followup-enqueue-and-owned-runs.md
+2026-07-30-followup-enqueue-and-owned-runs.md: 54f4af75eeb29b06504fa2629b0665f3e5f4c4ee
+2026-07-30-followup-enqueue-and-owned-runs.zh.md: 62c0d998a8bf94df8cabfccf23ea8b2a4da92c4d

.agents/notes/implemented/architecture/2026-07-30-followup-enqueue-and-owned-runs.md

@@ -0,0 +1,42 @@
+# Agent Note: Follow-up enqueue and owned run boundaries
+
+Status: implemented
+
+English | [中文](2026-07-30-followup-enqueue-and-owned-runs.zh.md)
+
+## Problem
+
+`Agent.followup()` identifies and queues a user message, but one follow-up does not own the activity that follows it. Steering, injected context, tool continuations, recovery, and later queued messages can all contribute before the agent next becomes idle. A `MessageId` can therefore prove inbox admission, but it cannot identify which assistant message or `turn/end` is the result of that input.
+
+The [one-send-one-turn decision](../simplification/2026-07-17-one-send-one-turn.md) already rejects a per-send completion handle at the core seam. Protocol and SDK layers that pair one prompt request with a turn result manufacture that missing relationship downstream. The pairing becomes ambiguous as soon as activity admits more input, and it exposes turn mechanics as if they were a prompt-level outcome.
+
+## Decision
+
+Keep `Agent.followup(message): void` as an enqueue-only operation. `Agent.whenIdle()` and `agent/status` remain whole-agent lifecycle observations; neither settles an individual message. Inbox durability records the identified message and its admission or cancellation, without assigning later output to it.
+
+The low-level SDK protocol answers `session/prompt` as soon as enqueue succeeds with `{ messageId }`. It streams durable facts through `session.event`, publishes whole-agent transitions through `session.status`, and has no `session.finished`. A low-level client may observe that receipt and later idleness, but receives no prompt result.
+
+High-level automation APIs return a `RunResult` only when they explicitly own an activity interval. The TypeScript and Python SDK `run()` methods collect from the submitted message's durable inbox receipt through the next whole-agent `idle`; their `finalResponse` is the last committed assistant message in that interval, not a response causally attributed to the submitted prompt. The one-shot CLI owns the analogous idle-to-idle interval. An isolated child-agent run may report a result because its caller owns the complete child lifecycle and any steering belongs to that run.
+
+ACP must return a protocol `stopReason`. Its bridge serializes one in-flight prompt per ACP session, waits for whole-agent idle, and otherwise reports the generic `end_turn`. Token-limit endings are not attributed to the prompt: they settle as `end_turn`. A model error on the prompt's correlated turn does reject the prompt immediately (the error is attributed by its owning turn), and a turnless slot (admission discarded the prompt) settles as `cancelled` at idle alongside explicit ACP cancellation or disposal.
+
+Goal continuation retains `MessageId` only to recognize its durable queued and admitted goal message. It advances from durable goal state at whole-agent idle, without mapping the message to a turn result.
+
+## Alternatives considered
+
+**Map `MessageId` to the turn that admits it.** A turn may consume steering and injected context and may continue through multiple model/tool steps. The mapping identifies admission, not causal ownership of the resulting output or stop reason.
+
+**Return a per-follow-up completion handle.** A handle would imply a result boundary that the shared agent lifecycle does not have. It would either omit work that influenced the activity or silently absorb unrelated later input.
+
+**Use the last `turn/end` observed before idle.** This is a useful run-level observation for an explicitly owned interval, but naming it as the submitted message's outcome recreates the false causal claim.
+
+## Verification
+
+- Agent and inbox tests pin enqueue-only follow-up, durable admission or cancellation, and whole-agent idle observation.
+- SDK protocol, TypeScript SDK, and Python SDK tests pin the `{ messageId }` receipt, `session.status`, the absence of `session.finished`, and receipt-to-idle `RunResult` collection without prompt-level `status` or `reason`.
+- ACP, one-shot CLI, goal continuation, and subagent tests pin the distinct activity ownership each integration possesses.
+- Consumer tests pin that no production integration derives a follow-up result by correlating `MessageId` with `turn/end`.
+
+## Consequences
+
+An owned activity interval can include steering, injected context, or other work submitted before idleness, so its final response and events are deliberately broader than the initiating message. Prompt-level model error and token-limit classifications disappear from SDK and ACP results; callers that need those facts must inspect the durable event stream without claiming causal attribution. Concurrent automation on one session requires an explicit serialization or ownership policy rather than an implicit per-prompt result.

.agents/notes/implemented/architecture/2026-07-30-followup-enqueue-and-owned-runs.zh.md

@@ -0,0 +1,42 @@
+# Agent Note: follow-up 入队与自有运行边界
+
+Status: implemented
+
+[English](2026-07-30-followup-enqueue-and-owned-runs.md) | 中文
+
+## 问题
+
+`Agent.followup()` 会标识一条用户消息并将其排入队列，但单次 follow-up 并不拥有随后发生的活动。在 agent（智能体）下一次进入 idle 前，steering（中途引导）、注入的上下文、工具续行、恢复和后续排队消息都可能参与活动。因此，`MessageId` 可以证明 inbox 已准入，但不能标识哪一条 assistant 消息或哪一个 `turn/end` 是该输入的结果。
+
+[one-send-one-turn 决策](../simplification/2026-07-17-one-send-one-turn.md) 已经在核心 seam 中排除了按 send 返回完成句柄的设计。凡是把一项提示词请求与一个轮次结果配对的协议层和 SDK 层，都会在下游人为构造这一缺失的关系。一旦活动准入更多输入，该配对就会产生歧义，还会把轮次机制暴露为提示词级结果。
+
+## 决策
+
+保留 `Agent.followup(message): void`，使其仅执行入队。`Agent.whenIdle()` 和 `agent/status` 仍用于观察整个 agent 的生命周期；二者都不结算单条消息。Inbox 持久性会记录已标识消息及其准入或取消，但不会把后续输出归属于该消息。
+
+底层 SDK 协议在入队成功后立即以 `{ messageId }` 响应 `session/prompt`。它通过 `session.event` 传输持久事实，通过 `session.status` 发布整个 agent 的状态转换，且不包含 `session.finished`。底层客户端可以观察该回执和之后的 idle，但不会收到提示词结果。
+
+只有明确拥有一个活动区间时，高层自动化 API 才返回 `RunResult`。TypeScript 和 Python SDK 的 `run()` 方法从已提交消息的持久 inbox 回执开始收集，直至整个 agent 下一次进入 `idle`；其 `finalResponse` 是该区间内最后一条已提交的 assistant 消息，而不是按因果关系归属于已提交提示词的响应。单次 CLI（命令行界面）拥有相应的 idle 到 idle 区间。隔离的子 agent 运行可以报告结果，因为调用方拥有完整的子级生命周期，任何 steering 都属于该运行。
+
+ACP（Agent Client Protocol）必须返回协议规定的 `stopReason`。其桥接层串行处理每个 ACP 会话中唯一一个正在处理的提示词，等待整个 agent 进入 idle，其他情况均报告通用的 `end_turn`。token 上限的轮次结束不归因于提示词：它们以 `end_turn` 结算。与该提示词关联的轮次上的模型错误会立即以该错误 reject 提示词（错误按其所属轮次归因），而 turnless 槽位（准入已丢弃提示词）会在 idle 时以 `cancelled` 结算，与显式 ACP 取消或资源释放并列。
+
+Goal 续行只保留 `MessageId`，用于识别持久排队和已准入的 goal 消息。它在整个 agent 进入 idle 时根据持久 goal 状态推进，不把消息映射到轮次结果。
+
+## 考虑过的替代方案
+
+**将 `MessageId` 映射到准入它的轮次。** 一个轮次可能使用 steering 和注入的上下文，还可能经过多个模型／工具步骤继续执行。该映射只能标识准入，不能确立结果输出或停止原因的因果归属。
+
+**返回按 follow-up 区分的完成句柄。** 这样的句柄暗示共享 agent 生命周期中存在并不实际成立的结果边界。它要么遗漏影响活动的工作，要么在不作说明的情况下吸收后续无关输入。
+
+**使用进入 idle 前观察到的最后一个 `turn/end`。** 对于明确拥有的区间，这是一项有用的运行级观测；但如果将其命名为已提交消息的结果，就会再次作出错误的因果声明。
+
+## 验证
+
+- Agent 与 inbox 测试固定 follow-up 仅入队、持久准入或取消以及整个 agent 的 idle 观测。
+- SDK 协议、TypeScript SDK 和 Python SDK 测试固定 `{ messageId }` 回执、`session.status`、不存在 `session.finished`，以及不含提示词级 `status` 或 `reason` 的回执到 idle `RunResult` 收集。
+- ACP、单次 CLI、goal 续行和 subagent 测试固定各集成实际拥有的不同活动边界。
+- 消费方测试固定生产集成都不会通过关联 `MessageId` 与 `turn/end` 来推导 follow-up 结果。
+
+## 后果
+
+自有活动区间可以包含进入 idle 前提交的 steering、注入上下文或其他工作，因此其最终响应和事件有意比初始消息涵盖更广。SDK 和 ACP 结果不再包含提示词级模型错误和 token 上限分类；需要这些事实的调用方必须检查持久事件流，但不能声称这些事实具有因果归属。在同一会话上并发执行自动化操作时，必须采用显式串行或所有权策略，不能依赖隐式的按提示词结果。

.agents/notes/implemented/architecture/2026-07-31-claimed-pre-step-inbox-lifecycle.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/architecture/2026-07-31-claimed-pre-step-inbox-lifecycle.md
+2026-07-31-claimed-pre-step-inbox-lifecycle.md: 5184c00335b3084a8eb6d58fa33a0af8d3a16ed6
+2026-07-31-claimed-pre-step-inbox-lifecycle.zh.md: 4966ccc8a3a7535022c04f9445ca75102ccca71e

.agents/notes/implemented/architecture/2026-07-31-claimed-pre-step-inbox-lifecycle.md

@@ -0,0 +1,41 @@
+# Agent Note: Claim inbox input before one pre-step decision
+
+Status: implemented
+
+English | [中文](2026-07-31-claimed-pre-step-inbox-lifecycle.zh.md)
+
+## Problem
+
+The loop previously split one step boundary across prompt preparation, prompt admission, and a serial step hook. Claimed input could be retained or discarded by an admission result, and live queue events carried shapes that duplicated durable inbox state. Plugins had to choose whether to mutate the inbox, rewrite a submitted batch, or append directly to session history, while observers could not rely on one exact ordering.
+
+Occurrence-local inbox wrappers also duplicated the identity already carried by every `UserMessage`. They made insertion, editing, claiming, cancellation, reconnect projection, and step entry one combined protocol even though the append-only session already owned the durable queue projection.
+
+## Decision
+
+Before every proposed step, `Inbox.claim(target)` atomically removes the complete batch: all `next-step` messages and, at a turn boundary, one `next-turn` message. At the initial boundary the loop first commits `turn/start`, so the claim and its single `agent/pre-step` decision have durable turn ownership. Claiming records normalized `agent/inbox/spliced` pure deletions with no outcome. The loop then emits `agent/inbox/claimed { message, turn }` once per claimed message and awaits the waterfall with that exclusive batch and `{ turn, step, signal }`.
+
+`PreStepDecision` is `{ kind: 'reject' } | { kind: 'enter'; messages: UserMessage[] }`. Reject opens no step, leaves the claimed batch removed, and closes the turn as blocked without any step events. Empty entry, cancellation, and failure before `step/start` likewise close a balanced no-step turn. Enter supplies the complete batch appended as `user/message` events after `step/start`. A listener wrapping `next()` preserves downstream changes unless it intentionally replaces them, so all message rewrites settle once in the final return value. There is no `agent/prompt-prepare`, `agent/prompt-submit`, or `agent/step` seam.
+
+The durable inbox remains two `UserMessage[]` lists addressed by `MessageId`. `append`, `prepend`, and `splice` take a target, while `replace(messageId, newMessage)` and `remove(messageId)` locate the pending message across both lists before committing a normalized splice. Replacement may change identity and emits the old message as discarded followed by the new message as inserted. Every insertion emits `agent/inbox/inserted { message }`; an ordinary removal records `outcome: 'canceled'` and emits `agent/inbox/discarded { message }`. Claiming is the loop's internal step-boundary operation on the inbox and records pure deletions without notifications or an outcome, so the loop can publish claimed events itself. These live events add no placement, outcome, or batch fields.
+
+The two event surfaces have separate consumers. Observers following one message use `agent/inbox/inserted`, `claimed`, and `discarded`. Whole-queue consumers, including the Web queue projection and reconnect baseline, use the durable `agent/inbox/spliced` stream; UI edits and removals route through `Inbox.splice()` or another Inbox mutation method so the same projection records every change.
+
+Plugins that need current-step atomic rewriting return messages from `agent/pre-step`. Plugins that only need later context may mutate `agent.inbox` directly. Workspace context uses both paths: asynchronous filesystem projections stage one replaceable `next-step` item, while the next entering pre-step folds that item or a newly composed baseline into its final batch and removes the pending copy. Rejection keeps the item queued.
+
+The archived [addressable queue occurrence decision](../../archived/feature/2026-07-29-addressable-queue-operations.md) describes the superseded occurrence-wrapper design. `MessageId` now owns addressability, while the retained Host queue mirror derives its snapshots from the durable splice projection.
+
+## Alternatives considered
+
+**Keep separate prepare and admit hooks.** This lets preparation mutate the inbox before claiming and admission rewrite afterward, but it creates two ordering surfaces for one boundary and makes cancellation ownership ambiguous.
+
+**Let rejection requeue the claimed batch.** This preserves retry-like behavior but turns a veto into hidden queue mutation, duplicates later work unless every race is fenced, and prevents claim from being an atomic ownership transfer.
+
+**Put placement and outcome on every live event.** Durable splices already own those facts. Repeating them on live notifications creates a second contract that can drift and is unnecessary for consumers holding the exact message identity.
+
+## Verification
+
+Agent-loop coverage pins turn-start-before-claim-before-pre-step ordering, exact live event payloads, balanced no-step rejection, final-batch rewriting, input inserted after a claim, listener failure, and cancellation. Inbox and consumer tests pin pure claim deletions, canceled ordinary removals, workspace-context staging, replacement, and same-step entry, plan/goal/hook behavior, UI cleanup, compaction, checkpointing, and resumed durable projection. Generated event and type catalogs expose only the new seam and payloads.
+
+## Consequences
+
+The loop has one awaited decision before each step and one ownership transfer for its input. Claimed messages never return to the inbox implicitly; later insertions remain independent. Live events are symmetrical with other inbox notifications without mirroring durable metadata, and plugins can choose exact-current-step rewriting or ordinary later inbox delivery explicitly.

.agents/notes/implemented/architecture/2026-07-31-claimed-pre-step-inbox-lifecycle.zh.md

@@ -0,0 +1,41 @@
+# Agent Note：在单一 pre-step 决策前领取 inbox 输入
+
+Status: implemented
+
+[English](2026-07-31-claimed-pre-step-inbox-lifecycle.md) | 中文
+
+## 问题
+
+循环此前把一个步骤边界拆成提示词准备、提示词准入与串行 step hook。准入结果可以保留或丢弃已领取输入，实时队列事件还携带了与持久 inbox 状态重复的形状。插件不得不在修改 inbox、改写已提交批次与直接追加会话历史之间选择，而观察方无法依赖一套明确顺序。
+
+单次出现专属的 inbox wrapper 也重复了每个 `UserMessage` 已有的标识。它把插入、编辑、领取、取消、重连投影与步骤进入合并成一套协议，但仅追加会话本就拥有持久队列投影。
+
+## 决策
+
+每个拟议步骤之前，`Inbox.claim(target)` 会原子移除完整批次：全部 `next-step` 消息，以及轮次边界上的一条 `next-turn` 消息。在首次边界，循环会先提交 `turn/start`，使领取及其唯一一次 `agent/pre-step` 决策拥有持久轮次归属。领取会记录规范化、不带 outcome 的纯删除 `agent/inbox/spliced`。随后，循环针对每条已领取消息发出一次 `agent/inbox/claimed { message, turn }`，并用该独占批次与 `{ turn, step, signal }` 等待 waterfall（瀑布式事件）。
+
+`PreStepDecision` 为 `{ kind: 'reject' } | { kind: 'enter'; messages: UserMessage[] }`。reject 不会打开步骤，会让已领取批次保持已删除，并将轮次关闭为 blocked，且不产生任何步骤事件。空的 enter、取消以及 `step/start` 前的失败同样会关闭一个边界平衡的无步骤轮次。enter 提供在 `step/start` 后以 `user/message` 追加的完整批次。包装 `next()` 的监听器会保留下游变更，除非有意替换，因此全部消息改写只在最终返回值中一次性结算。系统不再存在 `agent/prompt-prepare`、`agent/prompt-submit` 或 `agent/step` seam。
+
+持久 inbox 仍是两份通过 `MessageId` 寻址的 `UserMessage[]` 列表。`append`、`prepend` 与 `splice` 接受 target；`replace(messageId, newMessage)` 与 `remove(messageId)` 则在提交规范化 splice 前，通过 `MessageId` 跨两份列表定位待处理消息。替换可以改变标识，并先将旧消息作为 discarded 发布，再将新消息作为 inserted 发布。每次插入发出 `agent/inbox/inserted { message }`；普通删除记录 `outcome: 'canceled'` 并发出 `agent/inbox/discarded { message }`。领取是循环在 inbox 上的内部步骤边界操作，记录不带通知或 outcome 的纯删除，因此循环可以自行发布 claimed 事件。这些实时事件不增加 placement、outcome 或批次字段。
+
+两类事件表面服务不同消费方。跟踪单条消息的观察方使用 `agent/inbox/inserted`、`claimed` 与 `discarded`。包括 Web 队列投影和重连基线在内的整体队列消费方使用持久 `agent/inbox/spliced` 流；UI 编辑与移除经 `Inbox.splice()` 或其他 Inbox 变更方法进入，从而让同一投影记录所有变化。
+
+必须对当前步骤进行原子改写的插件从 `agent/pre-step` 返回消息。只需要稍后上下文的插件可以直接修改 `agent.inbox`。Workspace context 同时使用两条路径：异步文件系统投影会暂存一条可替换的 `next-step` 消息，而下一次进入步骤的 pre-step 会把该消息或新组合的基线折入最终批次，并移除仍待处理的副本。reject 会让该条目继续排队。
+
+已归档的[可寻址队列项决策](../../archived/feature/2026-07-29-addressable-queue-operations.md)描述了已被取代的单次出现 wrapper 设计。现在由 `MessageId` 负责寻址，而保留的 Host 队列镜像根据持久 splice 投影派生快照。
+
+## 曾考虑的替代方案
+
+**保留分离的 prepare 与 admit hook。** 这样准备阶段可以在领取前修改 inbox，准入阶段可以在领取后改写，但同一边界会出现两个顺序表面，取消归属也会变得模糊。
+
+**reject 时把已领取批次重新入队。** 这看似保留重试行为，却会让否决隐式修改队列；若不为每个竞态加围栏，还会复制后续工作，并使 claim 无法成为原子所有权转移。
+
+**在每个实时事件上携带 placement 与 outcome。** 持久 splice 已经拥有这些事实。实时通知重复它们会建立可能漂移的第二份契约，而持有确切消息标识的消费方并不需要这些字段。
+
+## 验证
+
+Agent-loop 覆盖固定先 `turn/start`、再领取、后 pre-step 的顺序、实时事件的确切载荷、边界平衡的无步骤 reject、最终批次改写、领取后插入的输入、监听器失败与取消。Inbox 和消费方测试固定纯领取删除、普通删除的 canceled 结果、workspace-context 的暂存、替换与同一步骤进入、plan/goal/hook 行为、UI 清理、压缩、检查点以及恢复后的持久投影。生成的事件与类型目录只公开新的 seam 与载荷。
+
+## 后果
+
+循环在每个步骤前只有一个需等待的决策，对输入也只有一次所有权转移。已领取消息不会隐式返回 inbox；后续插入保持独立。实时事件与其他 inbox 通知保持对称，但不镜像持久元数据；插件可以显式选择精确的当前步骤改写，或普通的后续 inbox 投递。

.agents/notes/implemented/feature/2026-07-29-addressable-queue-operations.i18n.yaml → .agents/notes/implemented/architecture/2026-07-31-goal-owned-durable-events.i18n.yaml

@@ -1,6 +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/2026-07-29-addressable-queue-operations.md
-2026-07-29-addressable-queue-operations.md: 04bcaa26be2f4f70c4003f6cc2fce4798499d205
-2026-07-29-addressable-queue-operations.zh.md: a16bf8667294de8d08da54ddb0d317bf9938e4eb
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-07-31-goal-owned-durable-events.md
+2026-07-31-goal-owned-durable-events.md: ac0358469958319f7629adb5e96845b7e0013297
+2026-07-31-goal-owned-durable-events.zh.md: 9b45cdd7990bfc1fdfbb63650e40b41e5ba2918d

.agents/notes/implemented/architecture/2026-07-31-goal-owned-durable-events.md

@@ -0,0 +1,33 @@
+# Agent Note: Goal-owned durable events
+
+Status: implemented
+
+English | [中文](2026-07-31-goal-owned-durable-events.zh.md)
+
+## Problem
+
+Goal state and inbox state have different lifecycles. A goal mutation must survive restart and fork whether or not any related model context is admitted, while an inbox message may be edited, claimed, rejected, or discarded as part of step scheduling. Encoding a goal mutation inside a round-zero inbox message made queue placement the domain commit point and required replay to reconcile insertion, admission, message identity, source metadata, and rendered content.
+
+The goal domain needs durable state, but it does not need ownership of pending model input. Continuation scheduling still needs the inbox; goal persistence does not.
+
+## Decision
+
+`@deepseek-ai/dsh-goal` owns a durable `goal/change` session event. Each event carries the complete post-mutation goal snapshot or a revisioned clear tombstone. `GoalService` appends that event synchronously, then emits `goal/changed`; strict replay and the `goal` session projection fold only `goal/change` for lifecycle state.
+
+`GoalMessageSource` identifies only positive admitted continuation rounds. A matching `user/message` advances `roundsStarted`; ordinary user messages and inbox splice events do not change goal state. The goal package never inserts, claims, removes, or inspects inbox messages. `@deepseek-ai/dsh-goal-session` remains responsible for queuing and tracking its own continuation prompts through the public inbox lifecycle.
+
+Activation remains process-local. The service associates the synchronously appended event sequence with the requested activation while its cache observes the event; replayed or externally appended changes default to disarmed. The session log remains the only durable authority.
+
+The domain does not automatically project each mutation into model input. Goal tools return current state, and continuation prompts include the objective and round state when work is actually scheduled. Any future always-visible goal context is a separate context plugin that owns its inbox message rather than a persistence side effect.
+
+## Alternatives considered
+
+- **Keep round-zero goal messages as the durable record.** Rejected because it couples domain commits to queue mutation and requires the goal fold to understand claim and admission reconciliation even though queue outcomes cannot roll back domain state.
+- **Derive goal state only from model-visible messages.** Rejected because a mutation may be valid and durable without opening a step, and cancellation or policy rejection must not erase it.
+- **Store goals in a separate database.** Rejected because the ordered session log already supplies persistence, replay, and fork inheritance without a second atomicity boundary.
+
+## Consequences
+
+Goal state is independent of inbox placement and admission. Replay has one mutation path, projections advance directly on `goal/change`, and continuation messages carry only round attribution. The model does not receive a mutation-only `<goal_state>` message; model-visible state appears through goal tools and scheduled continuation prompts. Direct session writers remain trusted and can append malformed changes, which the strict fold and invariant companion reject.
+
+Focused goal, goal-session, command, TUI, and client-fixture tests pin durable replay, positive-round accounting, inbox independence, projection updates, and restored-session behavior. The keyless process test inspects the persisted `goal/change` event and verifies that creation alone starts no continuation round.

.agents/notes/implemented/architecture/2026-07-31-goal-owned-durable-events.zh.md

@@ -0,0 +1,33 @@
+# Agent Note: Goal 自有的持久事件
+
+Status: implemented
+
+[English](2026-07-31-goal-owned-durable-events.md) | 中文
+
+## 问题
+
+Goal 状态与 inbox 状态具有不同的生命周期。无论相关模型上下文是否获准进入步骤，goal 变更都必须在重启与 fork 后保留；inbox 消息则可能在步骤调度期间被编辑、领取、拒绝或丢弃。把 goal 变更编码到 Round 为 0 的 inbox 消息中，会让队列放置成为领域提交点，并迫使回放对账插入、准入、消息标识、来源元数据与渲染内容。
+
+Goal 领域需要持久状态，但不需要拥有待处理的模型输入。继续执行调度仍然需要 inbox；goal 持久化不需要。
+
+## 决策
+
+`@deepseek-ai/dsh-goal` 拥有持久的 `goal/change` 会话事件。每个事件携带变更后的完整 goal 快照，或带修订号的清除墓碑。`GoalService` 同步追加该事件，再发出 `goal/changed`；严格回放与 `goal` 会话投影只折叠 `goal/change` 来获得生命周期状态。
+
+`GoalMessageSource` 只标识已准入且为正数的继续执行 Round。匹配的 `user/message` 会推进 `roundsStarted`；普通用户消息与 inbox splice 事件不会改变 goal 状态。Goal 包不会插入、领取、移除或检查 inbox 消息。`@deepseek-ai/dsh-goal-session` 仍通过公开 inbox 生命周期负责排队和跟踪自己的继续执行提示词。
+
+激活态仍只存在于进程中。服务在缓存观察事件时，将同步追加的事件序号与目标激活态关联；回放或外部追加的变更默认处于 `disarmed`。会话日志仍是唯一的持久权威。
+
+该领域不会自动把每次变更投影为模型输入。Goal 工具返回当前状态；真正调度工作时，继续执行提示词包含目标描述与 Round 状态。未来如果需要始终可见的 goal 上下文，应由独立上下文插件拥有其 inbox 消息，而不是把它作为持久化副作用。
+
+## 考虑过的替代方案
+
+- **继续以 Round 为 0 的 goal 消息作为持久记录。** 不予采纳，因为这会把领域提交与队列变更绑定，并要求 goal 折叠理解领取和准入对账，尽管队列结果不能回滚领域状态。
+- **只从模型可见消息派生 goal 状态。** 不予采纳，因为变更可以在不打开步骤的情况下有效且持久，取消或策略拒绝也不能擦除它。
+- **把 goal 存入独立数据库。** 不予采纳，因为有序会话日志已经提供持久化、回放与 fork 继承，无需引入第二个原子性边界。
+
+## 后果
+
+Goal 状态不依赖 inbox 放置与准入。回放只有一条变更路径，投影直接由 `goal/change` 推进，继续执行消息只携带 Round 归属。模型不会收到仅用于变更的 `<goal_state>` 消息；模型可见状态来自 goal 工具与已调度的继续执行提示词。直接写入会话的插件仍受信任，并且可以追加畸形变更；严格折叠与 invariant 配套模块会拒绝这些变更。
+
+聚焦的 goal、goal-session、command、TUI 与 client fixture 测试固定持久回放、正数 Round 计数、inbox 独立性、投影更新和恢复会话行为。无密钥进程测试检查持久的 `goal/change` 事件，并验证仅创建 goal 不会启动继续执行 Round。

.agents/notes/implemented/bug-fix/2026-07-21-semantic-session-checkpoints.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/bug-fix/2026-07-21-semantic-session-checkpoints.md
-2026-07-21-semantic-session-checkpoints.md: 927a4c5d6d2aad5dea460ea29f97c1686e9d5398
-2026-07-21-semantic-session-checkpoints.zh.md: c896ce55b23f08832cb5c80bf7aae264e366fee9
+2026-07-21-semantic-session-checkpoints.md: 697d878dccbfab1ded9f73e134d594ba6f0665e1
+2026-07-21-semantic-session-checkpoints.zh.md: c6184b0b00db16f9e318f4cbd148b1f0525f03a5

.agents/notes/implemented/bug-fix/2026-07-21-semantic-session-checkpoints.md

@@ -10,9 +10,9 @@ Persistence buffered every synchronous `session/event` until the loop's final tu
 
 ## Decision
 
-`dsh-session-checkpoint-policy` owns semantic durability barriers as a zero-config plugin beside a persistence backend. At `agent/step`, it flushes pending prompt input or the preceding response/result batch before the next request is derived. It wraps `llm/stream` lazily and flushes the live session after `request/header` is logged but before the adapter stream is constructed. It wraps top-level `tools/execute` after ordered pre-execute policy and flushes the recorded `tool/call` before the tool body; nested dispatches reuse the outer model-visible call. The loop's final `turn/end` checkpoint remains the closing boundary and settles before another queued turn or idle observation.
+`dsh-session-checkpoint-policy` owns semantic durability barriers as a zero-config plugin beside a persistence backend. At `agent/pre-step`, it flushes pending prompt input or the preceding response/result batch before the next request is derived. It wraps `llm/stream` lazily and flushes the live session after `request/header` is logged but before the adapter stream is constructed. It wraps top-level `tools/execute` after ordered pre-execute policy and flushes the recorded `tool/call` before the tool body; nested dispatches reuse the outer model-visible call. The loop's final `turn/end` checkpoint remains the closing boundary and settles before another queued turn or idle observation.
 
-Persistence and checkpoint scheduling remain separate Cordis plugins. A backend makes requested `session/flush` boundaries durable but does not choose them; loading it without this policy is valid and retains the loop's coarser checkpoints. First-party persisted apps and runtimes explicitly mount both, while a specialized deployment may intentionally omit or replace the policy. Registration order governs whether events appended by other `agent/step` listeners precede this checkpoint; prompt input and the preceding loop-owned assistant message and ordered results are already in the log.
+Persistence and checkpoint scheduling remain separate Cordis plugins. A backend makes requested `session/flush` boundaries durable but does not choose them; loading it without this policy is valid and retains the loop's coarser checkpoints. First-party persisted apps and runtimes explicitly mount both, while a specialized deployment may intentionally omit or replace the policy. Registration order governs whether events appended by other `agent/pre-step` listeners precede this checkpoint; prompt input and the preceding loop-owned assistant message and ordered results are already in the log.
 
 Checkpoint failure and cancellation are fail-closed at effect boundaries. A rejected request checkpoint prevents adapter dispatch; a rejected tool checkpoint becomes an error result without invoking the tool body. If cancellation lands while the tool checkpoint is pending, the policy rechecks the signal and returns the canonical `ABORTED_BEFORE_DISPATCH` result. A rejected between-step checkpoint closes the turn before another model request. A rejected final turn checkpoint is reported live and does not prevent later queued work. Persistence serialization continues to belong to the coordinator, so concurrent tool checkpoints cannot duplicate event sequences.
 

.agents/notes/implemented/bug-fix/2026-07-21-semantic-session-checkpoints.zh.md

@@ -10,9 +10,9 @@ Status: implemented
 
 ## 决策
 
-`dsh-session-checkpoint-policy` 以零配置插件的形式与持久化后端共同加载，并负责语义持久性屏障。在 `agent/step` 时，该插件会在推导下一个请求前刷新待持久化的提示词输入或前一批响应/结果。该插件惰性包装 `llm/stream`，在记录 `request/header` 之后、构造适配器流之前，刷新当前会话。该插件还在有序的执行前策略之后包装顶层 `tools/execute`，在进入工具主体前刷新已记录的 `tool/call`；嵌套分发则复用外层模型可见调用。循环的最终 `turn/end` 检查点仍是轮次的收尾边界，并会在处理另一个已排队轮次或观察到空闲状态之前完成。
+`dsh-session-checkpoint-policy` 以零配置插件的形式与持久化后端共同加载，并负责语义持久性屏障。在 `agent/pre-step` 时，该插件会在推导下一个请求前刷新待持久化的提示词输入或前一批响应/结果。该插件惰性包装 `llm/stream`，在记录 `request/header` 之后、构造适配器流之前，刷新活动会话。该插件还在有序的执行前策略之后包装顶层 `tools/execute`，在进入工具主体前刷新已记录的 `tool/call`；嵌套分发则复用外层模型可见调用。循环的最终 `turn/end` 检查点仍是轮次的收尾边界，并会在处理另一个已排队轮次或观察到空闲状态之前完成。
 
-持久化与检查点调度仍是相互独立的 Cordis 插件。后端使请求的 `session/flush` 边界持久化，但不选择边界；只加载后端而不加载本策略仍是有效组合，并保留循环提供的较粗检查点。第一方持久化应用与运行时会显式加载两者，专用部署则可以有意省略或替换本策略。注册顺序决定其他 `agent/step` 监听器追加的事件是否先于本检查点；提示词输入以及前一批由循环自身记录的助手消息与有序结果都已在日志中。
+持久化与检查点调度仍是相互独立的 Cordis 插件。后端使请求的 `session/flush` 边界持久化，但不选择边界；只加载后端而不加载本策略仍是有效组合，并保留循环提供的较粗检查点。第一方持久化应用与运行时会显式加载两者，专用部署则可以有意省略或替换本策略。注册顺序决定其他 `agent/pre-step` 监听器追加的事件是否先于本检查点；提示词输入以及前一批由循环自身记录的助手消息与有序结果都已在日志中。
 
 检查点失败与取消在副作用边界上采取失败关闭策略。请求检查点被拒绝时，系统不会分发给适配器；工具检查点被拒绝时，系统会返回错误结果，不调用工具主体。如果在工具检查点等待期间收到取消，策略会重新检查信号，并返回标准的 `ABORTED_BEFORE_DISPATCH` 结果。步骤间检查点被拒绝时，系统会在发起下一个模型请求前结束该轮次。轮次的最终检查点被拒绝时，系统会实时报告该失败，但不会阻止后续排队工作。持久化写入的串行化仍由协调器负责，因此并发的工具检查点不会产生重复的事件序列。
 

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.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/bug-fix/2026-07-29-human-transcript-append-origin.md
-2026-07-29-human-transcript-append-origin.md: 4d8d66ea625d02f098da906ae58af1a2e5e41819
-2026-07-29-human-transcript-append-origin.zh.md: d4c1c4a83d868af5e27d7bda2a22827cf9d6c5b1
+2026-07-29-human-transcript-append-origin.md: 9804d3d8d67a4ad00c3d395c2081fd47e03c254a
+2026-07-29-human-transcript-append-origin.zh.md: 909f0756dd29c9deddb840a188eb19a773f92d29

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.md

@@ -6,7 +6,7 @@ English | [中文](2026-07-29-human-transcript-append-origin.zh.md)
 
 ## Problem
 
-The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message`, `assistant/message`, and `steering/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only provenance and the replacement that cites it.
+The terminal and the host history gateway both treated the model-visible surface as the human transcript. A successful compaction replaces a surface range with one checkpoint node, so the moment that replacement landed the terminal dropped every message it shadowed — conversation the user had already read — and re-ran that destructive rebuild on any later replacement. The same confusion reached pagination: `maxMessages` counted every `user/message` and `assistant/message` in the window, so a model-only replacement copy consumed a page slot the human never filled, and the cut could land between a compaction's log-only provenance and the replacement that cites it.
 
 Nothing was lost from the log. `Session.events` still held every original message and full tool result; the surface only decides what the model is sent next. The defect was entirely in the projection.
 

.agents/notes/implemented/bug-fix/2026-07-29-human-transcript-append-origin.zh.md

@@ -6,7 +6,7 @@ Status: implemented
 
 ## Problem
 
-终端与宿主历史网关都把模型可见的 surface 当作人类可读记录（transcript）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message`、`assistant/message` 和 `steering/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志溯源信息与引用它的替换之间。
+终端与宿主历史网关都把模型可见的 surface 当作人类可读记录（transcript）。一次成功的压缩（compaction）会用一个检查点节点替换一段 surface 范围，因此该替换一落地，终端就丢弃了它所遮蔽的每条消息——那些是用户已经读过的对话——并在此后任何替换到来时重新执行这次破坏性重建。同样的混淆也波及分页：`maxMessages` 统计窗口内的每个 `user/message` 和 `assistant/message`，于是仅供模型使用的替换副本占用了一个人类从未填充的页面额度，而切分点还可能落在压缩的仅日志溯源信息与引用它的替换之间。
 
 日志本身没有丢失任何内容。`Session.events` 仍保存着每条原始消息和完整的工具结果；surface 只决定接下来发送给模型的内容。缺陷完全在投影层。
 

.agents/notes/implemented/bug-fix/2026-08-04-load-pre-react-loop-sessions.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/bug-fix/2026-08-04-load-pre-react-loop-sessions.md
+2026-08-04-load-pre-react-loop-sessions.md: b7dcad1ff7fe0aa8239ac03f52b50dfa55417515
+2026-08-04-load-pre-react-loop-sessions.zh.md: 14c6dcc7b4da016ebef424a66fae0b96fb469cd4

.agents/notes/implemented/bug-fix/2026-08-04-load-pre-react-loop-sessions.md

@@ -0,0 +1,40 @@
+# Agent Note: Load sessions from the pre-react-loop format
+
+Status: implemented
+
+English | [中文](2026-08-04-load-pre-react-loop-sessions.zh.md)
+
+## Problem
+
+The react-loop simplification changed durable events while retaining `SESSION_FORMAT_VERSION` 0. Stored sessions from the change's base contain `steering/message` and `turn/start.trigger`; their terminal reasons also use coarse `aborted`, separate `disposed`, and two older error payloads. Current surface and turn invariants cannot replay those records directly.
+
+The new durable inbox is not part of this compatibility problem. The base emitted process-local inbox notifications but no `agent/inbox/*` session events, so replaying old history as pending work would resurrect already claimed or discarded prompts.
+
+## Decision
+
+`PersistenceCoordinator` recognizes the exact pre-react-loop shapes after backend decoding and projects them into the current read view. It removes the obsolete `turn/start.trigger`, converts `steering/message` to the same identified `user/message`, maps old failure facts into the current structured error, folds `disposed` into an aborted turn with the `disposed` cause, and represents coarse aborted records with the persistence-only `{ kind: 'legacy' }` cause because their caller is unavailable.
+
+The coordinator applies the projection to `load`, `inspect`, adoption, HMR prefix comparison, and `readFrom`. A seek-capable `readFrom` normally reads only its suffix; when that suffix contains a legacy event needing an earlier replacement identity, the coordinator loads and normalizes the complete prefix before returning the requested seq range.
+
+The importer does not synthesize inbox splices. A resumed pre-react-loop agent begins with empty pending lists, matching the base runtime's inability to persist pending inbox work. The stored artifact remains append-only and later events use the current format.
+
+## Alternatives considered
+
+**Treat the same-version records as unsupported.** This follows the pre-release default but strands sessions produced by the PR base even though the removed steering content and terminal facts have complete mappings.
+
+**Replay old inbox notifications into durable splices.** Those notifications were not session events and do not provide a trustworthy pending-state snapshot. Inferring insertions without every claim and discard would re-run consumed work.
+
+**Assign coarse aborted records to an existing caller.** Mapping them to `user`, `parent`, or `hook` would manufacture provenance. A dedicated `legacy` cause keeps the stop classification without making a false audit claim.
+
+**Rewrite stored JSONL and SQLite records.** A rewrite would violate the append-only contract and require backend-specific atomic migration machinery for a read compatibility boundary.
+
+## Consequences
+
+Sessions written in the refactor's base format resume through the current AgentLoop with their steering content, turn boundaries, error facts, and stop classification intact. The shared coordinator contract covers in-memory, JSONL, and SQLite `load`/`inspect`/`readFrom`, including the SQLite suffix fallback; an assembled JSONL Agent resume verifies that the historical transcript is visible while both new inbox lists start empty.
+
+This exception supports the base format, not intermediate formats produced during development of the refactor. In particular, it defines no migration for earlier experimental `agent/inbox/spliced` payloads. Exact-shape recognition keeps malformed current-looking records on their rejection path instead of guessing them into validity.
+
+## Related
+
+- [Load sessions persisted before message identity](2026-07-28-load-pre-identity-session-messages.md) — owns deterministic identities and the general read-only import boundary for another same-version format change.
+- [Session persistence as an abstract service](../architecture/2026-06-14-session-persistence.md) — owns append-only backend storage and resume.

.agents/notes/implemented/bug-fix/2026-08-04-load-pre-react-loop-sessions.zh.md

@@ -0,0 +1,40 @@
+# Agent Note: 加载 react-loop 重构前格式的会话
+
+Status: implemented
+
+[English](2026-08-04-load-pre-react-loop-sessions.md) | 中文
+
+## 问题
+
+react-loop 简化在保持 `SESSION_FORMAT_VERSION` 为 0 的同时更改了持久事件。该变更基线所存储的会话包含 steering（中途引导）事件 `steering/message` 和 `turn/start.trigger`；其终止原因还使用粗粒度 `aborted`、独立的 `disposed` 和两种旧版错误载荷。当前 surface 和轮次不变量无法直接回放这些记录。
+
+新的持久 inbox 不属于此兼容性问题。该基线会发出进程本地 inbox 通知，但不会产生 `agent/inbox/*` 会话事件，因此将旧历史回放为待处理工作会让已经领取或丢弃的提示词再次执行。
+
+## 决策
+
+`PersistenceCoordinator` 会在后端解码后识别 react-loop 重构前的确切形状，并将其投影为当前读取视图。它移除已废弃的 `turn/start.trigger`，把 `steering/message` 转换为同一条带标识的 `user/message`，将旧版失败事实映射为当前结构化错误，把 `disposed` 折叠为带 `disposed` 原因的已中止轮次，并用仅供持久化导入使用的 `{ kind: 'legacy' }` 原因表示粗粒度中止记录，因为无法获得其调用方。
+
+协调器会把该投影应用于 `load`、`inspect`、接管、HMR 前缀比较和 `readFrom`。可寻址的 `readFrom` 通常只读取后缀；如果后缀包含需要更早替换标识的旧版事件，协调器会先加载并规范化完整前缀，再返回所请求的 seq 范围。
+
+导入器不会合成 inbox splice。恢复后的 react-loop 重构前 agent 从空的待处理列表开始，这与基线运行时无法持久化待处理 inbox 工作的行为一致。已存储产物仍然仅追加，后续事件使用当前格式。
+
+## 考虑过的替代方案
+
+**将同版本记录视为不受支持。** 这符合预发布阶段的默认立场，但会使 PR 基线产生的会话无法恢复，尽管已移除的 steering 内容和终止事实都有完整映射。
+
+**将旧 inbox 通知回放为持久 splice。** 这些通知不是会话事件，也无法提供可信的待处理状态快照。如果无法获知每一次领取和丢弃，就推断插入操作，会让已消费的工作再次执行。
+
+**将粗粒度中止记录归因于现有调用方。** 将其映射到 `user`、`parent` 或 `hook` 会虚构来源。专用的 `legacy` 原因既能保留停止分类，也不会产生虚假的审计事实。
+
+**重写已存储的 JSONL 和 SQLite 记录。** 重写会违反仅追加契约，并要求为读取兼容边界建立后端专用的原子迁移机制。
+
+## 后果
+
+以重构基线格式写入的会话可以通过当前 AgentLoop 恢复，并完整保留 steering 内容、轮次边界、错误事实和停止分类。共享协调器契约覆盖内存、JSONL 和 SQLite 的 `load`／`inspect`／`readFrom`，包括 SQLite 后缀回退；组装后的 JSONL agent 恢复用例会验证历史 transcript（文本记录）可见，同时两个新 inbox 列表都从空状态开始。
+
+此例外支持基线格式，不支持重构开发期间产生的中间格式。具体而言，它没有为更早的实验性 `agent/inbox/spliced` 载荷定义迁移。通过确切形状识别，当前格式外观相似但结构错误的记录仍会走拒绝路径，不会被猜测性地转换为有效记录。
+
+## 相关资料
+
+- [加载消息标识机制引入前持久化的会话](2026-07-28-load-pre-identity-session-messages.md)：负责另一项同版本格式变更的确定性标识和通用只读导入边界。
+- [以抽象服务实现会话持久化](../architecture/2026-06-14-session-persistence.md)：负责仅追加后端存储和恢复。

.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.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/2026-06-18-compaction-capability-seam.md
-2026-06-18-compaction-capability-seam.md: ef37313bc6fb984689793fa5a3e7ac4d9238ea88
-2026-06-18-compaction-capability-seam.zh.md: 6d094062ad88bdb587128351fc7217de353ebbd8
+2026-06-18-compaction-capability-seam.md: 27dbde9f2349681cf47c4d25b16399b26ed9e1ca
+2026-06-18-compaction-capability-seam.zh.md: 1fe9ece2861bd6d75633a866a4a11eaadbf7ef26

.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.md

@@ -10,7 +10,7 @@ A long-running agent conversation grows without bound. As the event log accumula
 
 The [session surface](../architecture/2026-06-18-session-surface.md) was built as the foundation for exactly this — an ordered projection over the event log with a `surfaceOp: { op: 'replace', start, end }` operation purpose-built to shadow a range of entries and insert a replacement, with `sourceEventSeqs` recording provenance so the decision replays deterministically. What remained was the plugin that *decides what to compact and produces the summary*.
 
-Two forces shape the design. First, compaction policy and reusable token measurement vary independently: measurement belongs to the LLM-family [`ctx.tokenMeter` service](../architecture/2026-07-15-replay-token-meter-service.md), while summarization can be a model call, a template, or a remote service. Second, `SurfaceEventType` is closed to five event types (`user/message`, `assistant/message`, `tool/result`, `context/message`, `steering/message`); only those may carry `surfaceOp`. A bespoke `compaction/*` event therefore **cannot** itself appear on the surface — the compiler and Session's always-on append/seed boundary reject `surfaceOp` on it.
+Two forces shape the design. First, compaction policy and reusable token measurement vary independently: measurement belongs to the LLM-family [`ctx.tokenMeter` service](../architecture/2026-07-15-replay-token-meter-service.md), while summarization can be a model call, a template, or a remote service. Second, `SurfaceEventType` is closed to the message-producing event types (`user/message`, `assistant/message`, `tool/result`); only those may carry `surfaceOp`. A bespoke `compaction/*` event therefore **cannot** itself appear on the surface — the compiler and Session's always-on append/seed boundary reject `surfaceOp` on it.
 
 ## Decision
 
@@ -19,7 +19,7 @@ Two forces shape the design. First, compaction policy and reusable token measure
 Per the [capability-seams Agent Note](../architecture/2026-06-13-capability-seams.md), compaction ships as separate packages so the contract, the algorithm, and (later) the consumer surface evolve independently:
 
 1. **Interface** — `@deepseek-ai/dsh-compact`: an abstract `CompactService` owning the `ctx.compact` key, the `CompactionResult` vocabulary, the `compact/*` session events, the manual failure taxonomy, and the canonical checkpoint message source. It declares `compactIfNeeded()`, `compactNow()`, and `compactRegion()` as **abstract** — the contract states *what* compaction does, not *how*.
-2. **Implementation** — `@deepseek-ai/dsh-compact-basic`: a concrete `BasicCompactService` that consumes `ctx.tokenMeter` and owns the tail→head retention walk, summarization via `ctx.llm.stream()`, the surface replacement, the lock, post-step pressure, and canonical context-overflow recovery. `summarize()` is its sole subclass hook; pricing and replay stay with the meter.
+2. **Implementation** — `@deepseek-ai/dsh-compact-basic`: a concrete `BasicCompactService` that consumes `ctx.tokenMeter` and owns the tail→head retention walk, summarization via `ctx.llm.stream()`, the surface replacement, the lock, pre-step pressure, and canonical context-overflow recovery. `summarize()` is its sole subclass hook; pricing and replay stay with the meter.
 3. **Model-free companion** — `@deepseek-ai/dsh-compact-tool-result-prune`: a concrete optional service that rewrites oversized current `tool/result` nodes before the backend selects a summary range. It is not a second compaction implementation and does not implement `CompactService`.
 4. **Human consumer** — `@deepseek-ai/dsh-command-compact` registers argument-free `/compact` through `ctx.commands` and calls the backend-independent `compactNow()` operation. It is direct human control, not a model-facing tool.
 
@@ -33,18 +33,18 @@ This is not a coupling smell — it is the contract's domain. The "only cordis"
 
 An earlier draft put the full algorithm (the retention walk, token-summing, text extraction) as concrete methods on the interface. That recouples the contract to one strategy: a backend that wants a different retention policy or event sequence would have to fight inherited concrete code. Making all three operations abstract puts every *how* decision in the backend and keeps the interface a statement of *what*. Token measurement is not a compaction hook at all; the singleton service lets multiple consumers share one per-session replay fold.
 
-`compactIfNeeded(agent, trigger, signal)` takes an explicit `'pressure' | 'context-overflow'` trigger and cancellation. It reads only the latest durable routed request; no header means no work, while any routed provider/model target uses the singleton estimator. `compactNow(agent, signal)` reserves idle turn admission and performs one useful balanced reduction even below pressure, returning `null` without writes when none exists. `compactRegion(start, end, agent, signal?)` uses `agent.session` as its single session identity and keeps an optional signal for explicit callers. The default summarizer resolves its target from explicit config, the latest logged routed target, then agent options, and records the provider/model pair after any `llm/stream` routing. It replays the routed request's prefix and appends the compaction directive as a trailing user message so the provider's warm KV cache is reused — see the [summary prefix-cache Agent Note](../bug-fix/2026-07-21-compaction-summary-prefix-cache-reuse.md). The call sets the provider-neutral `GenerateOptions.purpose` to `compaction`; adapters may map that purpose to model-hidden transport metadata, and the DeepSeek adapter sends `x-deepseek-harness-compact: 1`.
+`compactIfNeeded(agent, trigger, signal)` takes an explicit `'pressure' | 'context-overflow'` trigger and cancellation. It reads only the latest durable routed request; no header means no work, while any routed provider/model target uses the singleton estimator. `compactNow(agent, signal)` requires an idle agent and performs one useful balanced reduction even below pressure, returning `null` without writes when none exists. `compactRegion(start, end, agent, signal?)` uses `agent.session` as its single session identity and keeps an optional signal for explicit callers. The default summarizer resolves its target from explicit config, the latest logged routed target, then agent options, and records the provider/model pair after any `llm/stream` routing. It replays the routed request's prefix and appends the compaction directive as a trailing user message so the provider's warm KV cache is reused — see the [summary prefix-cache Agent Note](../bug-fix/2026-07-21-compaction-summary-prefix-cache-reuse.md). The call sets the provider-neutral `GenerateOptions.purpose` to `compaction`; adapters may map that purpose to model-hidden transport metadata, and the DeepSeek adapter sends `x-deepseek-harness-compact: 1`.
 
 ### Automatic pressure runs after successful durable step work
 
-Successful-call pressure cannot run at pre-step because final `agent/request` routing, provider output, tool results, buffered context, and steering do not exist there. Serial `agent/post-step(agent, turn, step, signal)` fires after those facts are durable and before `step/end`. `dsh-compact-basic` measures the canonical logged request through `ctx.tokenMeter`, so the next request sees any replacement without a speculative envelope override. Once pressure qualifies, optional `ctx.toolResultPrune` rewriting runs before summary selection; compact-basic remeasures the durable surface and skips summarization if pruning restores safe pressure.
+Successful-call pressure runs at the next `agent/pre-step`, after the preceding response, tool results, buffered context, and steering are durable and before the next request is derived. `dsh-compact-basic` measures the canonical logged request through `ctx.tokenMeter`, so the next request sees any replacement without a speculative envelope override. Once pressure qualifies, optional `ctx.toolResultPrune` rewriting runs before summary selection; compact-basic remeasures the durable surface and skips summarization if pruning restores safe pressure.
 
 Canonical provider context overflow takes a separate path. The failed step closes and `agent/request-error` receives the original request error. Compact-basic owns its per-agent overflow count, prunes before forcing one useful balanced reduction, and returns `{ kind: 'retry' }` only if `session.surface.replaceGeneration` increases, including pruning-only progress when no summary range exists. The loop then closes the failed turn, opens a new numbered retry turn, and reconstructs its request from the durable log. No replacement, a recovery failure before any replacement, cancellation, an exhausted cap, or an unrelated error preserves the original provider failure. If pruning already advanced the generation before later summary work fails, recovery retries from that durable pruned surface unless cancellation or disposal wins. The complete lifecycle decision is in the [after-call recovery Agent Note](../architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md).
 
 ```
-assistant/message → tool/result/context/steering
-await serial agent/post-step          ⟵ pressure compaction inside the successful step
-step/end
+assistant/message → tool/result/context/steering → step/end
+claim the next batch → await waterfall agent/pre-step  ⟵ pressure compaction before the next request
+enter → next step/start
 
 provider overflow → step/end
 await waterfall agent/request-error  ⟵ forced compaction between attempts
@@ -53,7 +53,7 @@ retry → next numbered step/start      ⟵ derives from the replacement surface
 
 ### Retention is turn-agnostic; tool-pairing balance is the only structural guard
 
-Auto-compaction checks after **every successful** step, not once per turn. This is load-bearing for runaway-turn survival: a tool-heavy ReAct turn appends an `assistant/message` + a `tool/result` per step, so the surface grows within a turn. The post-step check can compact early closed tool pairs before continuation opens the next step, and provider-confirmed overflow remains the backstop when a request crosses the limit first.
+Auto-compaction checks after **every successful** step, not once per turn. This is load-bearing for runaway-turn survival: a tool-heavy ReAct turn appends an `assistant/message` + a `tool/result` per step, so the surface grows within a turn. The next pre-step check can compact early closed tool pairs before continuation opens another step, and provider-confirmed overflow remains the backstop when a request crosses the limit first.
 
 `compactIfNeeded` retains the smallest tail of whole surface units whose estimated size reaches the resolved retained-token budget and compacts older nodes. A unit is a complete closed step or one no-step message. If the token cutoff lands inside a step, retention expands until the cut is tool-pairing balanced. Balance is checked on surface order, not log sequence, because replacement summaries have new sequence numbers at old surface positions. `dsh-compact` exports the before/after edge helpers; their per-session cache folds only appended surface-tail nodes while `replaceGeneration` is unchanged, does no event reads for log-only growth, and rebuilds current membership and balances after replacement. `compactRegion` rejects boundaries that split a tool call from its result. The in-flight turn receives no special retention.
 
@@ -96,7 +96,7 @@ The `compact/start … compact/end` bracket is justified by two roles:
 1. **Crash-detectable orphan + provenance** (primary). Summarization is a slow model call persisted *after* `compact/start`. A crash mid-summarization leaves a `compact/start` with no matching `compact/end` — a detectable orphan. Releasing the lock last (rather than first) converts the crash window from *silent corruption* into that detectable orphan.
 2. **Prevents concurrent compaction.** Every automatic, manual, and explicit-range entry point refuses a live unmatched `compact/start`. The bracket is the single lock; no process-local mutex duplicates it.
 
-The lock excludes another compaction, not unrelated facts. Its markers are time points rather than an exclusive container, so idle injected context may appear between a standalone manual start and end. Automatic work requires whole-surface stability inside its turn. Manual work revalidates only the selected positional span, letting append-only context outside it remain visible after replacement.
+The lock excludes another compaction, not unrelated facts. Its markers are time points rather than an exclusive container, so durable inbox splices may appear between a standalone manual start and end. Automatic work requires whole-surface stability inside its turn. Manual work revalidates only the selected positional span, letting append-only context outside it remain visible after replacement.
 
 The lifecycle boundary makes crash state unambiguous:
 
@@ -111,7 +111,7 @@ The lifecycle boundary makes crash state unambiguous:
 ## Alternatives considered
 
 - **The full algorithm as concrete interface methods** — rejected because it recouples the contract to one retention strategy. All three operations are abstract; reusable measurement is a separate LLM-family service and `summarize()` is basic's sole hook.
-- **Compaction on `agent/request` or provisional `agent/pre-step` inputs** — rejected because neither proves the final durable request and both couple generic lifecycle to compaction-specific envelope data. Post-step replay plus canonical overflow recovery covers both successful and rejected calls.
+- **Compaction on `agent/request` or a compaction-specific loop callback** — rejected because the former observes a provisional request and the latter couples generic lifecycle to compaction policy. Pre-step replay of the prior durable request plus canonical overflow recovery covers successful and rejected calls.
 - **A `compact` boolean or untyped request metadata map** — rejected because multiple auxiliary call kinds would become mutually exclusive flags, while an open bag would discard compiler-checked vocabulary. One typed `purpose` discriminant extends with additional call kinds without adding another `GenerateOptions` field.
 - **A separate `compact/error` event** — rejected: `compact/end` keeps an `error?` field, mirroring `tool/result`'s self-contained error — one event tells success from failure without correlating a sibling.
 - **Teaching core turn-repair about `compact/*`** — rejected: the general end-seed boundary already distinguishes prior-lifecycle history, and patching core for every future `xxx/start … xxx/end` pair is exactly the coupling the capability-seam architecture exists to avoid.
@@ -119,7 +119,7 @@ The lifecycle boundary makes crash state unambiguous:
 ## Consequences
 
 - **Packages**: `packages/compact/compact` supplies the interface, `compact-basic` supplies the backend, `compact-tool-result-prune` supplies optional deterministic rewriting, and `command-compact` supplies human `/compact`. `packages/llm/token-meter` owns replay-aware measurement independently.
-- **Automatic seams**: `agent/post-step` (`@mode serial`) handles successful-call pressure and `agent/request-error` (`@mode waterfall`) handles final request failures after the failed step closes. Generic `agent/pre-step` remains a four-argument checkpoint with no compaction-only prompt/prefix payload.
+- **Automatic seams**: `agent/pre-step` (`@mode waterfall`) handles pressure before request derivation and `agent/request-error` (`@mode waterfall`) handles final request failures after the failed step closes. Pre-step receives the claimed batch and `PreStepContext`, with no compaction-only prompt/prefix payload.
 - **`SessionEventMap`** gains `compact/start` / `compact/summary` / `compact/end` by declaration merging (merge-extensible); `SurfaceEventType` is **not** touched. These are session events, not cordis `Events`, so the event-taxonomy gate needs no entry.
 - **`dsh-compact`** owns `COMPACT_CHECKPOINT_SOURCE`, `isCompactCheckpointSource(source)`, `toolPairingBalancedBefore(session, seq)`, and `toolPairingBalancedAfter(session, seq)`. The marker identifies replacement summaries across backend implementations. The cached surface-edge checks prevent `compactRegion` and `compactIfNeeded` from splitting a tool-call/result pair, validate current membership by seq, answer both edges from one per-cut balance sequence, and reject stale or missing seqs and orphan results.
 - **`dsh-session`** validates positional replacement, complete provenance, and content-only single-node `tool/result` rewrites through its one surface manager. Its invariant companion treats fresh appended tool results as executions that require an open step and pending call, while the compaction companion owns numeric-turn versus standalone-null bracket relations.
@@ -128,7 +128,7 @@ The lifecycle boundary makes crash state unambiguous:
 ## Testing
 
 - **Unit:** Real Loader and invariant plugins cover whole-unit retention, pruning configuration and replay, rich-block ordering, metadata preservation, convergence, both `compact/end` outcomes, open-tail refusal, pruning-only and summarized overflow recovery, generation proof, caps, and original-error preservation.
-- **Loop:** Tests pin post-step after durable tool results and before `step/end`, actual `agent/request` routing, closed failed steps, fresh retry numbering, and complete thrown/in-band overflow → compaction → reconstructed retry composition.
-- **Manual:** Admission, marker ordering, injection retention, live/stale orphan classification, cancellation, close/flush failures, command mapping, and the queued TUI journey are pinned without a model key.
+- **Loop:** Tests pin pre-step after the preceding `step/end` and before the next `step/start`, actual `agent/request` routing, closed failed steps, fresh retry numbering, and complete thrown/in-band overflow → compaction → reconstructed retry composition.
+- **Manual:** Maintenance serialization, marker ordering, injection retention, live/stale orphan classification, cancellation, close/flush failures, command mapping, and the queued TUI journey are pinned without a model key.
 - **With-key e2e:** A real model and bash session with lowered limits triggers compaction, records a complete `compact/start…end` pair, shrinks the surface, and finishes the task.
 - **Snapshot gap:** Runaway-turn compaction cannot yet replay because the summarization call records no `assistant/chunk` events or `sessionId`; interleaved summarization-call replay remains follow-up work.

.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.zh.md

@@ -10,7 +10,7 @@ Status: implemented
 
 [会话接口面](../architecture/2026-06-18-session-surface.md)正是为此而构建的基础设施：一份建立在事件日志之上的有序投影，带有专门设计的 `surfaceOp: { op: 'replace', start, end }` 操作，用于遮蔽一段条目并插入替换内容，`sourceEventSeqs` 记录溯源信息以便决策可确定性地回放。剩下的是那个*决定压缩什么、并产出摘要*的插件。
 
-两股力量塑造了设计。第一，压缩策略与可复用的 token 测量独立变化：测量归 LLM 系列的 [`ctx.tokenMeter` 服务](../architecture/2026-07-15-replay-token-meter-service.md)所有，摘要生成则可以使用模型调用、模板或远程服务。第二，`SurfaceEventType` 封闭为五种事件类型（`user/message`、`assistant/message`、`tool/result`、`context/message`、`steering/message`）；只有这些类型可以携带 `surfaceOp`。因此一个专用的 `compaction/*` 事件**不能**出现在 surface 上，编译器与 Session 始终启用的 append/seed 边界都会拒绝在其上附加 `surfaceOp`。
+两股力量塑造了设计。第一，压缩策略与可复用的 token 测量独立变化：测量归 LLM 系列的 [`ctx.tokenMeter` 服务](../architecture/2026-07-15-replay-token-meter-service.md)所有，摘要生成则可以使用模型调用、模板或远程服务。第二，`SurfaceEventType` 封闭为产生消息的事件类型（`user/message`、`assistant/message`、`tool/result`）；只有这些类型可以携带 `surfaceOp`。因此一个专用的 `compaction/*` 事件**不能**出现在 surface 上，编译器与 Session 始终启用的 append/seed 边界都会拒绝在其上附加 `surfaceOp`。
 
 ## 决策
 
@@ -19,7 +19,7 @@ Status: implemented
 遵循[能力 seam Agent Note](../architecture/2026-06-13-capability-seams.md)，压缩以独立包发布，使契约、算法和（后续的）消费方 surface 各自独立演进：
 
 1. **接口** — `@deepseek-ai/dsh-compact`：抽象 `CompactService`，拥有 `ctx.compact` 键、`CompactionResult` 词汇、`compact/*` 会话事件、手动失败分类体系以及规范的检查点消息来源。它将 `compactIfNeeded()`、`compactNow()` 和 `compactRegion()` 声明为**抽象方法**——契约说明压缩*做什么*，而非*怎么做*。
-2. **实现** — `@deepseek-ai/dsh-compact-basic`：具体的 `BasicCompactService`，消费 `ctx.tokenMeter`，并拥有尾→头保留遍历、通过 `ctx.llm.stream()` 生成摘要、surface 替换、锁、步骤后压力处理和规范的上下文溢出恢复。`summarize()` 是其唯一的子类钩子；计价与回放仍归 meter 所有。
+2. **实现** — `@deepseek-ai/dsh-compact-basic`：具体的 `BasicCompactService`，消费 `ctx.tokenMeter`，并拥有尾→头保留遍历、通过 `ctx.llm.stream()` 生成摘要、surface 替换、锁、步骤前压力处理和规范的上下文溢出恢复。`summarize()` 是其唯一的子类钩子；计价与回放仍归 meter 所有。
 3. **无模型配套服务** — `@deepseek-ai/dsh-compact-tool-result-prune`：一个具体的可选服务，在后端选择摘要范围之前，重写当前过大的 `tool/result` 节点。它不是第二种压缩实现，也不实现 `CompactService`。
 4. **面向用户的消费方** — `@deepseek-ai/dsh-command-compact` 通过 `ctx.commands` 注册无参数 `/compact`，并调用后端无关的 `compactNow()` 操作。它是供用户直接控制的命令，不是面向模型的工具。
 
@@ -33,18 +33,18 @@ Status: implemented
 
 早期草案将完整算法（保留遍历、token 求和、文本提取）作为接口上的具体方法。这会将契约重新耦合到一种策略：想要不同保留策略或事件排序的后端必须与继承来的具体代码对抗。将三个操作都设为抽象，把所有*怎么做*的决策放在后端，并让接口保持为*做什么*的声明。token 测量根本不是压缩钩子；单例服务使多个消费方能够共享逐会话的回放折叠。
 
-`compactIfNeeded(agent, trigger, signal)` 接受显式的 `'pressure' | 'context-overflow'` 触发原因与取消信号。它只读取最新的持久化已路由请求；没有 header 就不执行工作，任何已路由的提供方/模型目标都使用单例估算器。`compactNow(agent, signal)` 会预留空闲轮次接纳，即使未达到压力也进行一次有效的平衡缩减；不存在这种范围时返回 `null`，且不写入任何内容。`compactRegion(start, end, agent, signal?)` 将 `agent.session` 作为唯一会话身份，并为显式调用方保留可选 signal。默认摘要器依次从显式配置、最新记录的已路由目标和 agent 选项解析目标，并在任何 `llm/stream` 路由后记录提供方/模型对。它回放已路由请求的前缀，并将压缩指令追加为尾部 user 消息，从而复用提供方的热 KV Cache；见[摘要前缀缓存 Agent Note](../bug-fix/2026-07-21-compaction-summary-prefix-cache-reuse.md)。该调用将提供方无关的 `GenerateOptions.purpose` 设为 `compaction`；适配器可以将此用途映射为对模型隐藏的传输元数据，DeepSeek 适配器会发送 `x-deepseek-harness-compact: 1`。
+`compactIfNeeded(agent, trigger, signal)` 接受显式的 `'pressure' | 'context-overflow'` 触发原因与取消信号。它只读取最新的持久化已路由请求；没有 header 就不执行工作，任何已路由的提供方/模型目标都使用单例估算器。`compactNow(agent, signal)` 要求 agent 处于 idle，即使未达到压力也进行一次有效的平衡缩减；不存在这种范围时返回 `null`，且不写入任何内容。`compactRegion(start, end, agent, signal?)` 将 `agent.session` 作为唯一会话身份，并为显式调用方保留可选 signal。默认摘要器依次从显式配置、最新记录的已路由目标和 agent 选项解析目标，并在任何 `llm/stream` 路由后记录提供方/模型对。它回放已路由请求的前缀，并将压缩指令追加为尾部 user 消息，从而复用提供方的热 KV cache；见[摘要前缀缓存 Agent Note](../bug-fix/2026-07-21-compaction-summary-prefix-cache-reuse.md)。该调用将提供方无关的 `GenerateOptions.purpose` 设为 `compaction`；适配器可以将此用途映射为对模型隐藏的传输元数据，DeepSeek 适配器会发送 `x-deepseek-harness-compact: 1`。
 
 ### 成功的持久步骤工作完成后运行自动压力检查
 
-成功调用的压力检查不能在步骤前运行，因为最终的 `agent/request` 路由、提供方输出、工具结果、缓冲上下文与 steering 当时尚不存在。串行的 `agent/post-step(agent, turn, step, signal)` 会在这些事实持久化后、`step/end` 之前触发。`dsh-compact-basic` 通过 `ctx.tokenMeter` 测量规范的已记录请求，因此下一个请求无需推测性覆盖信封即可看到任何替换。压力达到条件后，可选的 `ctx.toolResultPrune` 重写在摘要范围选择前运行；compact-basic 重新测量持久 surface，如果修剪恢复到安全压力便跳过摘要生成。
+成功调用的压力检查在下一个 `agent/pre-step` 运行；此时前一响应、工具结果、缓冲上下文与 steering 已经持久化，而下一个请求尚未派生。`dsh-compact-basic` 通过 `ctx.tokenMeter` 测量规范的已记录请求，因此下一个请求无需推测性覆盖信封即可看到任何替换。压力达到条件后，可选的 `ctx.toolResultPrune` 重写在摘要范围选择前运行；compact-basic 重新测量持久 surface，如果修剪恢复到安全压力便跳过摘要生成。
 
 规范的提供方上下文溢出走另一条路径。失败步骤先关闭，`agent/request-error` 接收原始请求错误。compact-basic 自行持有按 agent 计的溢出次数，在强制执行一次有效且平衡的缩减前先修剪，且仅当 `session.surface.replaceGeneration` 增加时才返回 `{ kind: 'retry' }`；这包括没有摘要范围时仅修剪取得的进展。随后循环关闭失败轮次，开启新的编号重试轮次，并从持久日志重建请求。没有替换、任何替换前的恢复失败、取消、耗尽的上限或无关错误都会保留原始提供方失败。如果修剪已经推进 generation，而后续摘要工作失败，恢复会从该持久的已修剪 surface 重试，除非取消或资源释放胜出。完整生命周期决策见[调用后恢复 Agent Note](../architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md)。
 
 ```
-assistant/message → tool/result/context/steering
-await serial agent/post-step          ⟵ pressure compaction inside the successful step
-step/end
+assistant/message → tool/result/context/steering → step/end
+claim the next batch → await waterfall agent/pre-step  ⟵ pressure compaction before the next request
+enter → next step/start
 
 provider overflow → step/end
 await waterfall agent/request-error  ⟵ forced compaction between attempts
@@ -53,7 +53,7 @@ retry → next numbered step/start      ⟵ derives from the replacement surface
 
 ### 保留是轮次无关的；工具配对平衡是唯一的结构守卫
 
-自动压缩在**每个成功的**步骤之后检查，而非每轮一次。这对失控轮次存活至关重要：工具密集型的 ReAct 轮次每步追加一个 `assistant/message` + 一个 `tool/result`，因此 surface 会在一轮之内增长。步骤后检查可以在后续步骤开启前压缩早期已关闭的工具对；如果请求率先越过限制，由提供方确认的溢出仍是兜底机制。
+自动压缩在**每个成功的**步骤之后检查，而非每轮一次。这对失控轮次存活至关重要：工具密集型的 ReAct 轮次每步追加一个 `assistant/message` + 一个 `tool/result`，因此 surface 会在一轮之内增长。下一个 pre-step 检查可以在继续执行打开另一步骤之前压缩早期已关闭的工具对；如果请求率先越过限制，由提供方确认的溢出仍是兜底机制。
 
 `compactIfNeeded` 保留估算大小达到解析后保留 token 预算的最小完整 surface 单元尾部，压缩更早的节点。一个单元是一个完整的已关闭步骤或一条无步骤消息。如果 token 截断点落在步骤内部，保留范围会扩展直到切割点满足工具配对平衡。平衡按 surface 顺序检查，而非日志序号，因为替换摘要在旧的 surface 位置拥有新的序号。`dsh-compact` 导出前后边缘辅助函数；只要 `replaceGeneration` 不变，其逐会话缓存就只折叠新增的 surface 尾部节点，面对仅日志增长时不读取事件，并在替换后重建当前成员关系与平衡。`compactRegion` 拒绝将工具调用与其结果拆分的边界。进行中的轮次不享受特殊保留。
 
@@ -96,7 +96,7 @@ compact/end      → log-only. Releases the lock (carries `error` on a recoverab
 1. **可检测的崩溃孤儿 + 来源追溯**（首要）。摘要生成是一次慢速模型调用，持久化在 `compact/start` *之后*。摘要生成中途崩溃会留下一个没有匹配 `compact/end` 的 `compact/start`——一个可检测的孤儿。最后释放锁（而非最先）将崩溃窗口从*静默损坏*转变为可检测的孤儿。
 2. **防止并发压缩。** 每个自动、手动和显式范围入口点都会拒绝活动的未匹配 `compact/start`。该标记对就是唯一的锁；没有进程本地 mutex 重复承担同一职责。
 
-该锁只排除另一项压缩，不排除无关事实。其标记是时间点，而不是排他的容器，因此空闲注入的上下文可以出现在独立手动 start 与 end 之间。自动工作要求其轮次内的整个 surface 保持稳定。手动工作只重新验证所选位置 span，使其外部的仅追加上下文在替换后保持可见。
+该锁只排除另一项压缩，不排除无关事实。其标记是时间点，而不是排他的容器，因此持久 inbox splice 可以出现在独立手动 start 与 end 之间。自动工作要求其轮次内的整个 surface 保持稳定。手动工作只重新验证所选位置 span，使其外部的仅追加上下文在替换后保持可见。
 
 生命周期边界使崩溃状态含义明确：
 
@@ -111,7 +111,7 @@ compact/end      → log-only. Releases the lock (carries `error` on a recoverab
 ## 曾考虑的替代方案
 
 - **完整算法作为接口的具体方法**——否决，因为它将契约重新耦合到一种保留策略。三个操作都是抽象的；可复用测量属于单独的 LLM 系列服务，`summarize()` 是 basic 唯一的钩子。
-- **在 `agent/request` 或临时 `agent/pre-step` 输入上执行压缩**——否决，因为两者都无法证明最终的持久请求，而且都会将通用生命周期耦合到压缩专属的信封数据。步骤后回放与规范溢出恢复同时覆盖成功和被拒绝的调用。
+- **在 `agent/request` 或压缩专属的 loop 回调上执行压缩**——否决，因为前者观察的是临时请求，后者会将通用生命周期耦合到压缩策略。对先前持久请求进行 pre-step 回放，再加上规范溢出恢复，即可覆盖成功和被拒绝的调用。
 - **`compact` 布尔值或无类型的请求元数据 map**——否决，因为多个辅助调用种类会变成互斥标志，而开放 map 会丢弃由编译器检查的词汇。一个类型化的 `purpose` 判别字段可以扩展其他调用种类，而无需再为 `GenerateOptions` 添加字段。
 - **单独的 `compact/error` 事件**——否决：`compact/end` 保留 `error?` 字段，与 `tool/result` 的自包含错误一致——一个事件即可区分成功与失败，无需关联兄弟事件。
 - **教导核心轮次修复识别 `compact/*`**——否决：通用 end-seed 边界已经能够区分先前生命周期的历史；为每个未来的 `xxx/start … xxx/end` 插件对修补核心模块，恰好是能力 seam 架构存在的意义所要避免的耦合。
@@ -119,7 +119,7 @@ compact/end      → log-only. Releases the lock (carries `error` on a recoverab
 ## 后果
 
 - **包**：`packages/compact/compact` 提供接口，`compact-basic` 提供后端，`compact-tool-result-prune` 提供可选的确定性重写，`command-compact` 提供面向用户的 `/compact`。`packages/llm/token-meter` 独立拥有回放感知的测量。
-- **自动 seam**：`agent/post-step`（`@mode serial`）处理成功调用的压力，`agent/request-error`（`@mode waterfall`）处理失败步骤关闭后的最终请求失败。通用 `agent/pre-step` 保持为四参数检查点，不携带压缩专属的提示词/前缀 payload。
+- **自动 seam**：`agent/pre-step`（`@mode waterfall`）在请求派生前处理压力，`agent/request-error`（`@mode waterfall`）处理失败步骤关闭后的最终请求失败。pre-step 接收已领取批次与 `PreStepContext`，不携带压缩专属的提示词/前缀 payload。
 - **`SessionEventMap`** 通过可合并扩展的声明合并获得 `compact/start` / `compact/summary` / `compact/end`；`SurfaceEventType` **未被**触及。这些是会话事件，不是 cordis `Events`，因此事件分类门禁无需新增条目。
 - **`dsh-compact`** 拥有 `COMPACT_CHECKPOINT_SOURCE`、`isCompactCheckpointSource(source)`、`toolPairingBalancedBefore(session, seq)` 与 `toolPairingBalancedAfter(session, seq)`。该标记用于跨后端实现识别替换摘要。带缓存的 surface 边缘检查会防止 `compactRegion` 和 `compactIfNeeded` 拆分工具调用/结果对，按 seq 校验当前成员关系，从每个切割点的一条平衡序列回答两侧边缘，并拒绝陈旧或缺失的 seq 与孤立结果。
 - **`dsh-session`** 通过唯一的 surface 管理器校验位置替换、完整溯源信息和仅内容的单节点 `tool/result` 重写。其不变式配套插件将新追加的工具结果视为执行，要求存在已打开的步骤与待处理调用，而压缩配套组件拥有数字轮次归属与独立 `null` 归属标记对之间的关系。
@@ -128,7 +128,7 @@ compact/end      → log-only. Releases the lock (carries `error` on a recoverab
 ## 测试
 
 - **单元测试：** 使用真实 Loader 和 invariant 插件覆盖完整单元保留、修剪配置与回放、富块顺序、元数据保留、收敛、`compact/end` 的两种结果、开放尾部拒绝、仅修剪与带摘要的溢出恢复、generation 证明、上限和原始错误保留。
-- **循环测试：** 测试固定步骤后处理发生在持久工具结果之后、`step/end` 之前，使用实际 `agent/request` 路由，关闭失败步骤，分配新的重试编号，并覆盖完整的抛出/带内溢出 → 压缩 → 重建重试组合。
-- **手动测试：** 无需模型密钥即可固定接纳、标记顺序、注入保留、活动／陈旧未匹配标记分类、取消、闭合／flush 失败、命令映射以及排队 TUI 流程。
+- **循环测试：** 测试固定 pre-step 发生在前一个 `step/end` 之后、下一个 `step/start` 之前，使用实际 `agent/request` 路由，关闭失败步骤，分配新的重试编号，并覆盖完整的抛出/带内溢出 → 压缩 → 重建重试组合。
+- **手动测试：** 无需模型密钥即可固定 maintenance 串行化、标记顺序、注入保留、活动／陈旧未匹配标记分类、取消、闭合／flush 失败、命令映射以及排队 TUI 流程。
 - **带密钥 e2e：** 真实模型和 bash 会话在降低的限制下触发压缩，记录完整的 `compact/start…end` 对，缩小 surface，并完成任务。
 - **快照缺口：** 失控轮次压缩尚无法回放，因为摘要调用未记录 `assistant/chunk` 事件或 `sessionId`；交错摘要调用的回放仍是后续工作。

.agents/notes/implemented/feature/2026-06-24-workspace-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/2026-06-24-workspace-context.md
-2026-06-24-workspace-context.md: 8baced0143abb38ff34d16a072761ec016a53d6e
-2026-06-24-workspace-context.zh.md: 3697d5bb00b2b166f8cd6c19e515aed1d0a183c7
+2026-06-24-workspace-context.md: df7ae58f8a42a8c1aac8113a429ef0f50b2fb795
+2026-06-24-workspace-context.zh.md: aecc57a5c63b4c6282260863f4183944c14b5cf8

.agents/notes/implemented/feature/2026-06-24-workspace-context.md

@@ -14,7 +14,7 @@ The lifecycle has two distinct classes of content. The initial applicable chain
 
 ## Decision
 
-The implementation lives in `packages/context/workspace-context` as `@deepseek-ai/dsh-workspace-context`. It is a request-context extension, not a core service or a filesystem backend. The shared demo spine and Host Runtime mount it from an explicit `{ maxBytes } | false` deployment choice; `dsh web` enables a 65,536-byte budget while the Host Runtime's headless consumer disables it. The plugin consumes `agent/step`, `tools/post-execute`, and the optional `ctx.fs` capability.
+The implementation lives in `packages/context/workspace-context` as `@deepseek-ai/dsh-workspace-context`. It is a request-context extension, not a core service or a filesystem backend. The shared demo spine and Host Runtime mount it from an explicit `{ maxBytes } | false` deployment choice; `dsh web` enables a 65,536-byte budget while the Host Runtime's headless consumer disables it. The plugin consumes `agent/pre-step`, `tools/post-execute`, and the optional `ctx.fs` capability.
 
 The plugin does not statically inject `fs`. Providerless product trees therefore boot normally and the plugin no-ops until a filesystem provider exists. All production reads go through that provider. Candidate probes resolve each path and stat the result, so a final-component symlink is followed to its target: a link to a regular file loads, while a missing path or a non-file target is a confirmed absence. Following repository-owned links across the trust boundary is a deliberate reversal of the original no-follow probe; the [instruction-symlink follow note](2026-07-21-follow-instruction-symlinks.md) owns that decision and its residual risk. The step signal and dynamic tool execution signal propagate through resolution, metadata probes, and streaming reads, so cancellation does not wait for an unrelated filesystem scan. A resolve or stat exception is classified as unavailable: it skips only that candidate and is never interpreted as the deletion of an already-loaded scope.
 
@@ -28,9 +28,9 @@ The user-global file is fixed at `$DSH_HOME/AGENTS.md`, is not affected by eithe
 
 ### Baseline Injection
 
-At the first `agent/step` of an agent-loop instance, the plugin injects one sourced user-role message before the request is derived. It loads the user-global file first, then finds the project root by walking upward from `agent.session.header.cwd` to a configured root marker (default `.git`), then loads one candidate from each directory from the root to the cwd. A `.git` file and a `.git` directory are both valid markers, covering linked worktrees and submodules. Without a marker, the cwd itself is the root.
+At the first `agent/pre-step` of an agent-loop instance, the plugin composes one sourced user-role baseline. When the downstream decision enters a nonempty first-step batch, the plugin folds the baseline into that final batch right after the claimed prompt, so it becomes durable with the direct prompt and reaches the first request. Rejection or an empty first-step decision leaves the baseline in the next-step inbox for a later wakeup. The plugin loads the user-global file first, then finds the project root by walking upward from `agent.session.header.cwd` to a configured root marker (default `.git`), then loads the configured candidates from each directory from the root to the cwd. A `.git` file and a `.git` directory are both valid markers, covering linked worktrees and submodules. Without a marker, the cwd itself is the root.
 
-The injection becomes a durable `user/message` with a typed `workspace-instructions` source. Its `baseline: true` marker distinguishes the complete startup or resume baseline from later deltas, and its change list persists the included scopes and content digests. In the product spine workspace instructions are registered before the skills catalog, so their `agent/step` listener injects first. The loop drains both messages before deriving the first request.
+The baseline becomes a durable `user/message` with a typed `workspace-instructions` source. Its `baseline: true` marker distinguishes the complete startup or resume baseline from later deltas, and its change list persists the included scopes and content digests. If a previously queued workspace baseline is still pending, the plugin removes that exact message and prepends its replacement instead of accumulating duplicates.
 
 A resumed agent creates a new loop instance and injects a baseline composed from current files before its first request. This permits current baseline content on resume without mutating an earlier history event. A resume and a hot plugin remount both face a log that may already hold a baseline; they are told apart by `agent/session-start`, which a startup or resume emits before the first step while a remount attaches to an already-live session and never sees it. A remount retains the existing baseline only when its typed event remains in the current visible surface, and still rebuilds scope and provider-version tracking from current files. If compaction has shadowed that event, the remount injects a current baseline. A resume always re-composes.
 
@@ -68,7 +68,7 @@ There is intentionally no watcher. Detection occurs at the next successful struc
 
 **Use a global `ctx.systemPrompt.section()`.** Rejected because one Cordis context can host sessions with different cwd values, while repository-owned text is lower-authority context rather than top-authority provider system content.
 
-**Inject the baseline on every `agent/step`.** Rejected because repeated history injection wastes tokens and complicates duplicate state. A per-mount session guard gives one visible baseline event while it remains on the surface; dynamic append-only messages handle changes and compaction re-arming.
+**Always leave prepared workspace context in the inbox.** Rejected because context prepared during pre-step would then survive the current request and start a second model step by itself. The inbox remains the staging and rejection fallback, while an entering pre-step owns atomic delivery with its final batch.
 
 **Load both `AGENTS.md` and `CLAUDE.md` in one directory.** Rejected because repositories in transition commonly duplicate guidance across both files. Ordered candidates make precedence explicit and configurable.
 

.agents/notes/implemented/feature/2026-06-24-workspace-context.zh.md

@@ -14,7 +14,7 @@ Status: implemented
 
 ## 决策
 
-该实现在 `packages/context/workspace-context` 中，包名为 `@deepseek-ai/dsh-workspace-context`。它是请求上下文扩展，不是核心服务或文件系统后端。共享 demo 主干与 Host Runtime 根据显式的 `{ maxBytes } | false` 部署选择挂载它；`dsh web` 启用 65,536 字节预算，Host Runtime 的 headless 消费方则禁用它。该插件使用 `agent/step`、`tools/post-execute` 和可选的 `ctx.fs` 功能。
+该实现在 `packages/context/workspace-context` 中，包（package）名为 `@deepseek-ai/dsh-workspace-context`。它是请求上下文扩展，不是核心服务或文件系统后端。共享 demo 主干与 Host Runtime 根据显式的 `{ maxBytes } | false` 部署选择挂载它；`dsh web` 启用 65,536 字节预算，Host Runtime 的 headless 消费方则禁用它。该插件使用 `agent/pre-step`、`tools/post-execute` 和可选的 `ctx.fs` 功能。
 
 插件不会静态注入 `fs`。因此，不带提供方的产品树仍能正常启动；在文件系统提供方出现之前，插件保持无操作。所有生产读取都通过该提供方完成。候选项探测会解析每个路径并对结果执行 stat，因此会跟随最终路径组件的符号链接至其目标：指向普通文件的链接会被加载，缺失路径或非文件目标则确认为不存在。允许仓库拥有的链接跨越信任边界，是对最初不跟随探测方式的刻意反转；[跟随指令符号链接记录](2026-07-21-follow-instruction-symlinks.md)负责说明该决策及其残余风险。步骤信号与动态工具执行信号会贯穿解析、元数据探测和流式读取，因此取消不会等待无关的文件系统扫描。解析或 stat 异常归类为不可用：它只跳过该候选项，绝不被解释为已经加载的作用域被删除。
 
@@ -28,9 +28,9 @@ Status: implemented
 
 ### 基线注入
 
-在 agent loop（智能体循环）实例的第一个 `agent/step`，插件会在派生请求前注入一条带来源的 user 角色消息。它先加载用户全局文件，再从 `agent.session.header.cwd` 向上遍历至配置的根标记（默认为 `.git`）以确定项目根目录，随后从根目录至 cwd 的每级目录各加载一个候选项。`.git` 文件与 `.git` 目录都是有效标记，因而能覆盖链接 worktree 和 submodule。找不到标记时，cwd 本身就是根目录。
+在 agent loop（智能体循环）实例的第一个 `agent/pre-step`，插件会组合一条带来源的 user 角色基线。当下游决策让非空的第一步批次进入时，插件会将基线折入最终批次、紧随已领取的直接提示词之后，使其与直接提示词一同成为持久记录并抵达第一次请求。reject 或空的第一步决策会将基线留在 next-step inbox，等待后续唤醒。插件先加载用户全局文件，再从 `agent.session.header.cwd` 向上遍历至配置的根标记（默认为 `.git`）以确定项目根目录，随后从根目录至 cwd 的每级目录加载已配置候选项。`.git` 文件与 `.git` 目录都是有效标记，因而能覆盖链接 worktree 和 submodule。找不到标记时，cwd 本身就是根目录。
 
-该注入成为一条持久 `user/message`，并携带带类型的 `workspace-instructions` 来源。其 `baseline: true` 标记将完整的启动或恢复基线与后续增量区分开来，变更列表则持久保存已纳入的作用域和内容 digest。在产品主干中，工作区指令的注册先于 skill（技能）目录，所以其 `agent/step` 监听器先注入。循环会在派生第一次请求前 drain 这两条消息。
+该基线会成为一条持久 `user/message`，并携带带类型的 `workspace-instructions` 来源。其 `baseline: true` 标记将完整的启动或恢复基线与后续增量区分开来，变更列表则持久保存已纳入的作用域和内容 digest。若先前排队的 workspace 基线仍在等待，插件会删除该确切消息并 prepend 替代值，而不会累积副本。
 
 恢复 agent 会创建新的循环实例，并在其第一次请求前注入由当前文件组合的基线。这样，恢复时可以使用当前基线内容，而无需修改先前的历史事件。恢复与插件热重挂都会面对日志中可能已存在基线的情况；二者通过 `agent/session-start` 区分：启动或恢复会在第一步前发出该事件，而热重挂附着到一个已存活的会话、永远不会看到它。只有当基线的类型化事件仍在当前可见表层中时，热重挂才保留既有基线，同时仍会根据当前文件重建 scope 与提供方版本跟踪。如果压缩（compaction）已遮蔽该事件，热重挂会注入当前基线。恢复则始终重新组合。
 
@@ -68,7 +68,7 @@ shell 命令不会触发发现。本地 bash 调用会启动全新的 shell，
 
 **使用全局 `ctx.systemPrompt.section()`。** 不予采纳，因为同一个 Cordis 上下文可以承载 cwd 不同的多个会话，而仓库自身拥有的文本属于低权威用户上下文，不是最高权威的提供方系统内容。
 
-**在每次 `agent/step` 时注入基线。** 不予采纳，因为重复注入历史会浪费 token，并使重复状态复杂化。逐挂载会话防护会在基线事件仍留在表面期间提供一条可见基线事件；动态仅追加消息负责处理变更和压缩后的重新启用。
+**始终把准备好的 workspace 上下文留在 inbox 中。** 不予采纳，因为在 pre-step 中准备的上下文会因此在当前请求结束后继续留存，并自行启动第二个模型步骤。inbox 仍作为暂存区以及 reject 时的后备，而进入步骤的 pre-step 负责随最终批次原子投递。
 
 **在一个目录中同时加载 `AGENTS.md` 和 `CLAUDE.md`。** 不予采纳，因为正在迁移的仓库通常会在两个文件中重复指引。按顺序排列的候选项让优先级显式且可配置。
 

.agents/notes/implemented/feature/2026-06-30-hook-bridges.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/2026-06-30-hook-bridges.md
-2026-06-30-hook-bridges.md: 99c6b1941a10e198ec3028f5fe505dabfff9abbe
-2026-06-30-hook-bridges.zh.md: 72c275128dbf2bf673beeb6b17daf657d728e418
+2026-06-30-hook-bridges.md: a31116818c57e18d04e656f371efd1101a63b6af
+2026-06-30-hook-bridges.zh.md: 8827408b67e2dd19f867a8658309f00231ba06c5

.agents/notes/implemented/feature/2026-06-30-hook-bridges.md

@@ -6,7 +6,7 @@ English | [中文](2026-06-30-hook-bridges.zh.md)
 
 ## Problem
 
-The harness's extension surface is its typed interception seams ([the interception-seams Agent Note](2026-06-30-interception-seams.md)): a "native hook" is just an ordinary cordis plugin subscribing to `agent/session-start`, `agent/prompt-submit`, `tools/pre-execute`, `tools/post-execute`, `agent/turn-stopping`, `subagent/start`, or `subagent/end`. But users arrive with **existing** Claude Code (CC) and Codex hook configs — a `hooks.json` (or a settings file's `hooks` key) full of shell-command hooks — and want those to run unmodified. This Agent Note introduces the two **bridge plugins** that translate that external shell-hook protocol onto the typed seams, built on the shared wire-protocol library ([the hook-protocol-lib Agent Note](2026-06-30-hook-protocol-lib.md)).
+The harness's extension surface is its typed interception seams ([the interception-seams Agent Note](2026-06-30-interception-seams.md)): a "native hook" is just an ordinary cordis plugin subscribing to `agent/session-start`, `agent/pre-step`, `tools/pre-execute`, `tools/post-execute`, `agent/turn-stopping`, `subagent/start`, or `subagent/end`. But users arrive with **existing** Claude Code (CC) and Codex hook configs — a `hooks.json` (or a settings file's `hooks` key) full of shell-command hooks — and want those to run unmodified. This Agent Note introduces the two **bridge plugins** that translate that external shell-hook protocol onto the typed seams, built on the shared wire-protocol library ([the hook-protocol-lib Agent Note](2026-06-30-hook-protocol-lib.md)).
 
 The framing that shapes the whole design: **a bridge is a compatibility adapter, not a power tool.** Anything a bridge does (block a tool, inject context, force continuation, observe a subagent) a native cordis plugin does more powerfully — typed returns, full `ctx`, no serialization boundary. The bridge's reason to exist is to run the explicitly supported subset of external CC/Codex command hooks. That keeps each bridge thin: parse the config, pick a matcher mode, build the per-event payload, call `runHook` + `mergeHookOutputs` from the shared lib, and map the neutral outcome onto a seam Decision. The package READMEs own the exact current unsupported-event and partial-field inventory against the official protocols.
 
@@ -24,7 +24,7 @@ Each bridge maps the neutral `MergedHookOutcome` from the shared lib onto the se
 | Seam | CC | Codex |
 |---|---|---|
 | `agent/session-start` (emit) | additionalContext → `agent.inject()` | plain-stdout output → additionalContext → `agent.inject()` |
-| `agent/prompt-submit` | `deny`→`block`; context-only→delegate+fold | `block`→`block`; context-only→delegate+fold |
+| `agent/pre-step` | `deny`→`reject`; context-only→delegate+fold into `enter` | `block`→`reject`; context-only→delegate+fold into `enter` |
 | `tools/pre-execute` | `deny`→`deny`; `ask`→`ask` | `block`→`deny` (no allow/ask) |
 | `tools/post-execute` | `deny`→`block`+feedback; context-only→delegate+fold | same |
 | `agent/turn-stopping` | blocking Stop → next-step steering | same |
@@ -37,11 +37,11 @@ The CC bridge's `ask` result is a real permission path, not a terminal bridge de
 
 Every bridge `inject()` and additional-context input explicitly passes `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }`. Unit coverage pins the resulting `user/message.source` as the plugin rather than the user.
 
-`UserPromptSubmit` runs during admission, before any turn opens. It therefore writes no turn-scoped `hook/invoked` / `hook/result` pair: a block leaves no transcript, while allowed additional context is durably represented by its sourced `user/message`. The Codex payload still receives the candidate next `turn_id`; rejection does not consume that number.
+`UserPromptSubmit` runs at pre-step after `turn/start`, so every invocation writes its turn-scoped `hook/invoked` / `hook/result` pair. Rejection leaves the claimed input removed, closes the turn as blocked with no step, and retains the hook pair as its durable decision evidence. The Codex payload receives that open turn's `turn_id`.
 
 ### Adding context is not a veto — delegate, then prepend
 
-A hook that only attaches `additionalContext` (no block/deny) is NOT a decision the bridge should return on its own: returning `allow`/`accept` from a waterfall listener WITHOUT calling `next()` short-circuits every later `agent/prompt-submit` / `tools/post-execute` listener, so a policy/sandbox plugin registered after the bridge would never see the prompt. Each bridge therefore delegates via `next()` before adding its context to the downstream decision. Both seams carry ordered `additionalContexts` arrays, so the bridge prepends its separately sourced entry while preserving every downstream source, envelope, and metadata field; a downstream prompt block still drops all context because the prompt never reaches the model, while post-tool block semantics may explicitly retain contexts. Code Mode ferries the same array through the outer `run_code` result. Only a real `deny`/`block` from the hook itself short-circuits. Tests assert a later listener can still block a prompt a context-only hook allowed and that retained prompt and post-tool contexts remain separate.
+A hook that only attaches `additionalContext` (no block/deny) is NOT a decision the bridge should return on its own: returning `enter` from a waterfall listener WITHOUT calling `next()` short-circuits every later `agent/pre-step` / `tools/post-execute` listener, so a policy/sandbox plugin registered after the bridge would never see the prompt. Each bridge therefore delegates via `next()` before adding its context to a downstream enter decision. The bridge preserves every downstream message, while a downstream pre-step rejection drops the whole claimed batch because no step opens. Post-tool decisions retain their independent ordered `additionalContexts` semantics, including Code Mode deferral through the outer `run_code` result. Only a real `deny`/`block` from the hook itself short-circuits. Tests assert a later listener can still reject a prompt after a context-only hook and that retained prompt and post-tool contexts remain separate.
 
 ### CLAUDE_PROJECT_DIR defaults to the session workspace
 

.agents/notes/implemented/feature/2026-06-30-hook-bridges.zh.md

@@ -6,7 +6,7 @@ Status: implemented
 
 ## 问题
 
-harness 的扩展面是其类型化的拦截 seam（见[拦截 seam Agent Note](2026-06-30-interception-seams.md)）：所谓「原生钩子」不过是一个普通的 Cordis 插件，订阅 `agent/session-start`、`agent/prompt-submit`、`tools/pre-execute`、`tools/post-execute`、`agent/turn-stopping`、`subagent/start` 或 `subagent/end`。但用户带着**既有的** Claude Code（CC）和 Codex 钩子配置到来，一个 `hooks.json`（或 settings 文件中的 `hooks` 键）里满是 shell 命令钩子，并希望它们原样运行。本 Agent Note 引入两个**桥接插件**，将外部 shell 钩子协议翻译到类型化 seam 上，构建于共享的协议格式（wire format）库之上（见 [hook-protocol-lib Agent Note](2026-06-30-hook-protocol-lib.md)）。
+harness 的扩展面是其类型化的拦截 seam（见[拦截 seam Agent Note](2026-06-30-interception-seams.md)）：所谓「原生钩子」不过是一个普通的 Cordis 插件，订阅 `agent/session-start`、`agent/pre-step`、`tools/pre-execute`、`tools/post-execute`、`agent/turn-stopping`、`subagent/start` 或 `subagent/end`。但用户带着**既有的** Claude Code（CC）和 Codex 钩子配置到来，一个 `hooks.json`（或 settings 文件中的 `hooks` 键）里满是 shell 命令钩子，并希望它们原样运行。本 Agent Note 引入两个**桥接插件**，将外部 shell 钩子协议翻译到类型化 seam 上，构建于共享的协议格式（wire format）库之上（见 [hook-protocol-lib Agent Note](2026-06-30-hook-protocol-lib.md)）。
 
 贯穿整个设计的定位：**桥接是兼容性适配器，不是高级工具。** 桥接能做的事（阻止工具、注入上下文、强制继续、观察 subagent），原生 Cordis 插件都能做得更强——类型化返回值、完整 `ctx`、无序列化边界。桥接存在的理由是运行外部 CC/Codex 命令钩子中被明确支持的子集。这使每个桥接保持精简：解析配置、选择匹配模式、构建每事件的 payload、调用共享库的 `runHook` + `mergeHookOutputs`，再将中性结果映射为 seam Decision。各包的 README 维护着当前不支持的事件和部分字段的完整清单，以官方协议为参照。
 
@@ -24,7 +24,7 @@ harness 的扩展面是其类型化的拦截 seam（见[拦截 seam Agent Note](
 | Seam | CC | Codex |
 |---|---|---|
 | `agent/session-start`（emit） | additionalContext → `agent.inject()` | 纯 stdout 输出 → additionalContext → `agent.inject()` |
-| `agent/prompt-submit` | `deny`→`block`；仅上下文→delegate+fold | `block`→`block`；仅上下文→delegate+fold |
+| `agent/pre-step` | `deny`→`reject`；仅上下文→委托并折叠到 `enter` | `block`→`reject`；仅上下文→委托并折叠到 `enter` |
 | `tools/pre-execute` | `deny`→`deny`；`ask`→`ask` | `block`→`deny`（无 allow/ask） |
 | `tools/post-execute` | `deny`→`block`+feedback；仅上下文→delegate+fold | 同上 |
 | `agent/turn-stopping` | 阻塞的 Stop → 下一步 steering（中途引导） | 同上 |
@@ -37,11 +37,11 @@ CC 桥接的 `ask` 结果是一条真正的权限路径，而非终态桥接决
 
 每个桥接的 `inject()` 和 additional-context 输入都显式传入 `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }`。单元测试固定验证结果中的 `user/message.source` 为插件而非用户。
 
-`UserPromptSubmit` 在准入阶段运行，早于任何轮次开启。因此它不写入任何轮次范围的 `hook/invoked` / `hook/result` 对：阻止不会留下 transcript（文本记录），而被允许的额外上下文由其带来源的 `user/message` 持久呈现。Codex payload 仍会收到候选的下一个 `turn_id`；拒绝不会消耗该编号。
+`UserPromptSubmit` 在 `turn/start` 之后的 pre-step 运行，因此每次调用都会写入轮次范围的 `hook/invoked` / `hook/result` 对。reject 会让已领取输入保持删除，将轮次关闭为 blocked 且不包含步骤，并保留该 hook 对作为持久决策证据。Codex payload 会收到这个已打开轮次的 `turn_id`。
 
 ### 添加上下文不是否决——先 delegate，再 prepend
 
-仅附加 `additionalContext`（没有 block/deny）的钩子并不是桥接可以独自返回的决策：在 waterfall 监听器中不调用 `next()` 就返回 `allow`/`accept`，会短路其后的每个 `agent/prompt-submit` / `tools/post-execute` 监听器，使注册在桥接之后的策略/沙箱插件看不到该提示词。因此，每个桥接都会先通过 `next()` 委托，再将自身上下文加入下游决策。两个 seam 都携带有序的 `additionalContexts` 数组，因此桥接会在保留所有下游来源、信封和元数据字段的同时，前置加入其独立来源的条目；下游提示词阻止仍会丢弃所有上下文，因为提示词从未到达模型，而工具后阻止语义可以显式保留上下文。Code Mode 会通过外层 `run_code` 结果转送同一数组。只有钩子本身真正返回 `deny`/`block` 才会短路。测试断言：上下文钩子允许后，较晚的监听器仍能阻止提示词，且保留的提示词和工具后上下文仍彼此分离。
+仅附加 `additionalContext`（没有 block/deny）的钩子并不是桥接可以独自返回的决策：在 waterfall 监听器中不调用 `next()` 就返回 `enter`，会短路其后的每个 `agent/pre-step` / `tools/post-execute` 监听器，使注册在桥接之后的策略/沙箱插件看不到该提示词。因此，每个桥接都会先通过 `next()` 委托，再将自身上下文加入下游 enter 决策。桥接会保留所有下游消息；下游 pre-step reject 会丢弃整个已领取批次，因为步骤从未打开。工具后决策仍保留独立的有序 `additionalContexts` 语义，包括 Code Mode 通过外层 `run_code` 结果延迟上下文。只有钩子本身真正返回 `deny`/`block` 才会短路。测试断言：仅上下文钩子之后，较晚的监听器仍能 reject 提示词，且保留的提示词和工具后上下文仍彼此分离。
 
 ### CLAUDE_PROJECT_DIR 默认为会话工作区
 

.agents/notes/implemented/feature/2026-06-30-hook-protocol-lib.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/2026-06-30-hook-protocol-lib.md
-2026-06-30-hook-protocol-lib.md: ce25f40e96ffd5c319d9845e36eab5130cec5857
-2026-06-30-hook-protocol-lib.zh.md: 42843c1ef17f1586bab12ca2ab071aa9998f2631
+2026-06-30-hook-protocol-lib.md: 2edff0d501cd7695873c68054eeb3a1e9942eced
+2026-06-30-hook-protocol-lib.zh.md: 42e526fdf8073465d5f6faf99a6c92fbe09505f5

.agents/notes/implemented/feature/2026-06-30-hook-protocol-lib.md

@@ -21,7 +21,7 @@ A new `packages/hooks/` group with `hook-protocol` as a pure library. It owns fo
 - **Merge** — `mergeHookOutputs(outputs)`, folding multiple matched hooks into one most-restrictive `MergedHookOutcome`: permission precedence **deny > ask > allow**, halt sticky on the first `continue:false`, block reasons joined `\n\n`, context/system-messages accumulated in order.
 - **`hook/*` session events** — `hook/invoked` / `hook/result`, declaration-merged into `SessionEventMap` (log-only, like `compact/*` — NOT `SurfaceEventType`s), with `appendHookInvoked`/`appendHookResult` helpers so the invoked/result pairing and owner-defined execution relation stay consistent across bridges. `appendHookResult` also owns the durable record's semantics — the decision string (the hook's parsed decision, else `'stop'` on `continue:false`, else `'pass'`) and the 500-character `stderrSummary` truncation derive from the `HookOutput` here, not per-bridge.
 
-**Per-dialect (the bridge plugins):** building each event's stdin payload (CC's base+per-event field sets vs Codex's snake_case with `turn_id`/`model` extras), the dialect's env + `${CLAUDE_PLUGIN_ROOT}` substitution (CC) vs none (Codex), and mapping the neutral `HookOutput`/`MergedHookOutcome` onto the harness's seam-specific typed Decisions (`PreToolDecision`, `PromptDecision`, `ContinuationDecision`, `PostToolDecision`).
+**Per-dialect (the bridge plugins):** building each event's stdin payload (CC's base+per-event field sets vs Codex's snake_case with `turn_id`/`model` extras), the dialect's env + `${CLAUDE_PLUGIN_ROOT}` substitution (CC) vs none (Codex), and mapping the neutral `HookOutput`/`MergedHookOutcome` onto the harness's seam-specific typed Decisions (`PreToolDecision`, `PreStepDecision`, `ContinuationDecision`, `PostToolDecision`).
 
 ## Alternatives considered
 

.agents/notes/implemented/feature/2026-06-30-hook-protocol-lib.zh.md

@@ -21,7 +21,7 @@ Status: implemented
 - **合并** — `mergeHookOutputs(outputs)`，将多个匹配钩子的输出折叠为一个最严格的 `MergedHookOutcome`：权限优先级 **deny > ask > allow**，halt 在首个 `continue:false` 时粘滞，阻止原因以 `\n\n` 拼接，context/system-messages 按序累积。
 - **`hook/*` 会话事件** — `hook/invoked` / `hook/result`，通过声明合并进入 `SessionEventMap`（仅日志，如 `compact/*`——不是 `SurfaceEventType`），配有 `appendHookInvoked`/`appendHookResult` 辅助函数，确保 invoked/result 配对与由所有者定义的执行关系在各桥接插件间保持一致。`appendHookResult` 还负责定义持久化记录的语义：decision 字符串（钩子解析出的 decision，否则 `continue:false` 时为 `'stop'`，否则为 `'pass'`）和 500 字符的 `stderrSummary` 截断均从本库的 `HookOutput` 派生，而非各桥接插件各自实现。
 
-**方言专属（桥接插件）：** 构建每个事件的 stdin payload（CC 的 base+per-event 字段集 vs Codex 的 snake_case 加 `turn_id`/`model` 额外字段）、方言的 env 与 `${CLAUDE_PLUGIN_ROOT}` 替换（CC）vs 两者皆无（Codex），以及将方言无关的 `HookOutput`/`MergedHookOutcome` 映射为 harness seam 专属的类型化 Decision（`PreToolDecision`、`PromptDecision`、`ContinuationDecision`、`PostToolDecision`）。
+**方言专属（桥接插件）：** 构建每个事件的 stdin payload（CC 的 base+per-event 字段集 vs Codex 的 snake_case 加 `turn_id`/`model` 额外字段）、方言的 env 与 `${CLAUDE_PLUGIN_ROOT}` 替换（CC）vs 无替换（Codex），以及将方言无关的 `HookOutput`/`MergedHookOutcome` 映射为 harness seam 专属的类型化 Decision（`PreToolDecision`、`PreStepDecision`、`ContinuationDecision`、`PostToolDecision`）。
 
 ## 曾考虑的替代方案
 

.agents/notes/implemented/feature/2026-06-30-interception-seams.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/2026-06-30-interception-seams.md
-2026-06-30-interception-seams.md: 4658983e1f098ecd199eecec4408e7c2f134cbf7
-2026-06-30-interception-seams.zh.md: 9c7a8509915d583d8f58d9d6d50509fdfcfca42f
+2026-06-30-interception-seams.md: 629a1aed509bd9bce9a2da89ce84b17a1db8e6b6
+2026-06-30-interception-seams.zh.md: d6958c9d1e7a8af8fa06d859d1905719a19cd43d

.agents/notes/implemented/feature/2026-06-30-interception-seams.md

@@ -16,7 +16,7 @@ The canonical surface separates transformable policy, around-dispatch control, a
 
 **Agent events** (`dsh-agent`):
 - `agent/session-start(agent, source)` — emit, once before turn 1, carrying a `SessionStartSource` (`startup` for a fresh/forked create, `resume` for a reloaded persisted session; `clear`/`compact` reserved). A pure notification — it CANNOT block startup (a deliberate gap: a bridge logs/injects, it does not gate startup). A listener seeds context via `agent.inject()`.
-- `agent/prompt-submit(agent, content, source, signal, next) → PromptDecision` — waterfall, fired for one claimed queued message before the loop opens a turn or appends `user/message`. The explicit admission signal is placed before the final `next`; `allow` optionally rewrites the prompt `content` or attaches separately sourced `additionalContexts[]`, while `block` discards the candidate without creating session history.
+- `agent/pre-step(agent, messages, context, next) → PreStepDecision` — waterfall, fired before every proposed step after the loop has atomically removed its exclusive inbox batch. `PreStepContext` carries that request's `turn`, `step`, and cancellation `signal`; `messages` is empty for a tool continuation with no intervening input. `enter` returns the complete message batch, including any current-request context a listener contributes; `reject` opens no step and leaves the claimed messages removed.
 
 **`agent/turn-stopping`** is an awaited notification at the natural stop boundary. A listener that needs another step calls `agent.steer()` with explicitly sourced model-facing content; the loop then re-reads the outbox and either continues or closes the turn.
 
@@ -35,7 +35,7 @@ Core dispatch and the tool body sit inside normalization boundaries, so tool, li
 
 ### Three load-bearing loop decisions
 
-1. **Run prompt policy before opening the turn.** A blocked prompt creates no turn or durable event. On allow, the loop stages the rewritten prompt followed by every returned `additionalContexts` entry, opens the turn, and drains that outbox before the first step. Each claimed ordinary-send item is the sole direct prompt in its turn under the [one-send-one-turn simplification](../simplification/2026-07-17-one-send-one-turn.md).
+1. **Run pre-step policy at every proposed step.** The loop opens the turn before the initial claim and decision, so rejection closes a durable blocked turn with no step or model-visible message. A tool continuation with no newly claimed input still submits an empty batch, allowing per-request context producers to add logged messages to that exact request. On enter, the loop opens the step and appends the returned batch as `user/message` events before request derivation. Each claimed follow-up remains the sole direct prompt in its turn under the [one-send-one-turn simplification](../simplification/2026-07-17-one-send-one-turn.md).
 
 2. **Post-tool `additionalContexts` and asynchronous injections enter the active-batch FIFO and append when that batch settles.** `content`/`feedback` shape the result `execute()` returns, but each context is a separate sourced `user/message`, and a single step or composite tool can produce many. Appending context immediately would interleave `result(c1) → context → result(c2)` or place nested context before its outer result, breaking tool-call/result adjacency. `ToolRunContext.deferContext()` therefore collects nested-dispatch context through failures, `execute()` surfaces the ordered array on `ToolExecutionResult`, and the loop accepts it into the same FIFO as `agent.inject()` calls made during execution. The FIFO appends after every recorded result when the batch settles, including before an interrupted turn closes. An accepted outer call preserves deferred contexts before decision contexts; an outer block discards deferred contexts and exposes only contexts explicitly supplied by the blocking decision.
 
@@ -56,4 +56,4 @@ The seam package does **not** declare `hook/*` session events (the durable hook-
 
 ## Consequences
 
-The canonical interception surface is uniformly typed without giving every extension the same power: hooks return decisions, execution wrappers wrap, terminal guards only deny, and final observers only observe. The loop owns session-start, pre-turn prompt admission, post-tool context buffering, and stopping; `dsh-tools` owns identity sealing and the five-phase execution pipeline. Their contracts are documented in [architecture.md](../../../../docs/architecture.md), package READMEs, [core interception decisions](../../../../docs/core-data-structures/core.md#interception-decisions), and [tool structures](../../../../docs/core-data-structures/tools.md). The ACP bridge settles an admission rejection as `cancelled` after the agent becomes idle with no owned turn, while hook-driven snapshots verify the observable bridge behavior end to end.
+The canonical interception surface is uniformly typed without giving every extension the same power: hooks return decisions, execution wrappers wrap, terminal guards only deny, and final observers only observe. The loop owns session-start, pre-step claim settlement, post-tool context buffering, and stopping; `dsh-tools` owns identity sealing and the five-phase execution pipeline. Their contracts are documented in [architecture.md](../../../../docs/architecture.md), package READMEs, [core interception decisions](../../../../docs/core-data-structures/core.md#interception-decisions), and [tool structures](../../../../docs/core-data-structures/tools.md). The ACP bridge settles an initial pre-step rejection from its blocked no-step turn as `end_turn`, while hook-driven snapshots verify the observable bridge behavior end to end.

.agents/notes/implemented/feature/2026-06-30-interception-seams.zh.md

@@ -15,8 +15,8 @@ harness 需要一套钩子子系统：用户像 Claude Code（CC）和 Codex 那
 规范表面将可变换策略、环绕调度控制与仅观测通知分离。策略 waterfall（瀑布式事件）返回小型的、seam 专属的**类型化 Decision 联合类型**；包装层返回规范化结果；通知接收不可变快照，无法影响结果。覆盖的钩子点包括 `session-start`、`prompt-submit`、`pre-tool`、`post-tool`、通过 continuation 实现的 `stop`，同时将非钩子的执行策略留作独立可组合。
 
 **Agent 事件**（`dsh-agent`）：
-- `agent/session-start(agent, source)` ——emit，在第 1 轮次之前触发一次，携带 `SessionStartSource`（`startup` 表示全新／fork 创建，`resume` 表示重新加载的持久化会话；`clear`/`compact` 保留）。纯通知，不能阻塞启动（这是有意的空白：桥接可以记录／注入，但不管控启动）。监听器通过 `agent.inject()` 注入上下文。
-- `agent/prompt-submit(agent, content, source, signal, next) → PromptDecision` ——waterfall，针对一条取得所有权的排队消息触发，早于循环开启轮次或追加 `user/message`。显式准入 signal 位于最后的 `next` 之前；`allow` 可以重写提示词 `content` 或附加来源各自独立的 `additionalContexts[]`，而 `block` 会丢弃该候选消息，不产生会话历史。
+- `agent/session-start(agent, source)` ——emit，在第 1 轮次之前触发一次，携带 `SessionStartSource`（`startup` 表示全新/fork 创建，`resume` 表示重新加载的持久化会话；`clear`/`compact` 保留）。纯通知，不能阻塞启动（这是有意的空白：桥接可以记录/注入，但不管控启动）。监听器通过 `agent.inject()` 注入上下文。
+- `agent/pre-step(agent, messages, context, next) → PreStepDecision` ——waterfall，在每个拟议步骤之前、循环原子移除其独占 inbox 批次后触发。`PreStepContext` 携带该请求的 `turn`、`step` 与取消 `signal`；没有中途输入的工具续步会收到空批次。`enter` 返回完整消息批次，其中包括监听器为当前请求贡献的上下文；`reject` 不打开步骤，并让已领取消息保持已删除。
 
 **`agent/turn-stopping`** 是自然停止边界上的一次 awaited 通知。需要再执行一步的监听器调用 `agent.steer()`，传入来源显式的 steering（中途引导）内容供模型使用；循环随后重新读取 outbox，继续执行或关闭轮次。
 
@@ -35,7 +35,7 @@ harness 需要一套钩子子系统：用户像 Claude Code（CC）和 Codex 那
 
 ### 三个承重的循环决策
 
-1. **在开启轮次之前运行提示词策略。** 被阻止的提示词不会创建轮次，也不产生持久事件。允许时，循环先暂存重写后的提示词，再暂存每个返回的 `additionalContexts` 条目，然后开启轮次并在第一个步骤之前排空该 outbox。依照[一次 send 对应一个轮次的简化](../simplification/2026-07-17-one-send-one-turn.md)，每个取得所有权的 ordinary-send 条目都是其轮次中唯一的直接提示词。
+1. **在每个拟议步骤运行 pre-step 策略。** 循环会在首次领取和决策之前打开轮次，因此 reject 会关闭一个持久、blocked 且不含步骤或模型可见消息的轮次。即使工具续步没有新取得所有权的输入，也会提交空批次，使逐请求上下文生产方可以把带日志的消息加入这一次请求。enter 时，循环先开启步骤，再把返回批次作为 `user/message` 追加，然后派生请求。依照[一次 send 对应一个轮次的简化](../simplification/2026-07-17-one-send-one-turn.md)，每个已领取 follow-up 仍是其轮次中唯一的直接提示词。
 
 2. **工具执行后的 `additionalContexts` 与异步注入进入活跃批次 FIFO，并在该批次结算时追加。** `content`/`feedback` 塑造 `execute()` 返回的结果，但每项上下文都是一条独立的带来源 `user/message`，而单个步骤或组合工具可以产生许多上下文。立即追加上下文会产生 `result(c1) → context → result(c2)` 的交错，或把嵌套上下文放在外层结果之前，破坏工具调用／工具结果邻接性。因此 `ToolRunContext.deferContext()` 会在失败路径上也收集嵌套调度上下文，`execute()` 在 `ToolExecutionResult` 上暴露有序数组，循环再把它接纳到与执行期间 `agent.inject()` 调用相同的 FIFO 中。FIFO 在批次结算时，在所有已记录结果之后追加，其中也包括被中断轮次关闭之前。被接受的外层调用将 deferred contexts 保留在 decision contexts 之前；被外层阻止时则丢弃 deferred contexts，只暴露阻止 decision 显式提供的上下文。
 
@@ -56,4 +56,4 @@ seam 包**不**声明 `hook/*` 会话事件（持久的钩子调用日志）；
 
 ## 后果
 
-规范拦截表面采用统一的类型体系，同时不给每个扩展相同的权力：钩子返回 decision，执行包装层封装执行过程，终结 guard 只能拒绝，最终观测者只能观测。循环负责 session-start、轮次前的提示词准入、工具执行后上下文缓冲和 stopping；`dsh-tools` 负责身份封存与五阶段执行流水线。它们的契约记录在 [architecture.md](../../../../docs/architecture.md)、各包 README、[核心拦截 decision](../../../../docs/core-data-structures/core.md#interception-decisions) 与[工具结构](../../../../docs/core-data-structures/tools.md)中。ACP（Agent Client Protocol）桥接在 agent 空闲且不再拥有轮次后，将准入拒绝结算为 `cancelled`，而钩子驱动的快照端到端验证可观测的桥接行为。
+规范拦截表面具有统一的类型化，同时不给每个扩展相同的权力：钩子返回 decision，执行包装层做包装，终结 guard 只能拒绝，最终观测者只能观测。循环负责 session-start、pre-step 领取结算、工具执行后上下文缓冲和 stopping；`dsh-tools` 负责身份封存与五阶段执行流水线。它们的契约记录在 [architecture.md](../../../../docs/architecture.md)、各包 README、[核心拦截 decision](../../../../docs/core-data-structures/core.md#interception-decisions) 与[工具结构](../../../../docs/core-data-structures/tools.md)中。ACP 桥接会把 blocked 无步骤轮次中的首次 pre-step reject 结算为 `end_turn`，而钩子驱动的快照端到端验证可观测的桥接行为。

.agents/notes/implemented/feature/2026-07-05-skill-system.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/2026-07-05-skill-system.md
-2026-07-05-skill-system.md: 36a2ae220db24c3183253c18957769ab5d5c068c
-2026-07-05-skill-system.zh.md: 3a673d44d74ab38732dd5f5bda4d607b174db251
+2026-07-05-skill-system.md: dd2fb1d22949f55ea7cb2c9f280e7cfbfcbbb226
+2026-07-05-skill-system.zh.md: 96656a8e1dfc2ae1ce7301ba29e7739349b6aab6

.agents/notes/implemented/feature/2026-07-05-skill-system.md

@@ -22,7 +22,7 @@ Each skill is either `<name>/SKILL.md` or `<name>.md` with YAML frontmatter. `na
 
 Local skill filesystem I/O goes through `ctx.fs` when a filesystem service is loaded: project-root lookup probes `.git` with `resolve` and `stat`, root discovery uses `listDir`, and skill reads use `readText`. The Node filesystem remains a fallback for minimal contexts that mount `dsh-skill-local` without the fs seam. Missing roots, unreadable or malformed skill files, and transient provider `list()` failures degrade to warn-and-skip so one bad source does not make every agent request fail; malformed candidates still fail fast because they are provider contract violations.
 
-`dsh-tool-skill` injects one durable user-role `<system-reminder>` catalog as a sourced `user/message` at the session's first `agent/step`, and only when that agent's tool view resolves this plugin's exact `skill` registration. The catalog contains sorted skill name and description only; it excludes bodies, paths, sources, providers, and routing hints. Descriptions are whitespace-normalized, XML-escaped, and capped by `catalogDescriptionMaxLength`, whose default is `500` and minimum is `3`. Full skill bodies are never included in the catalog. (The catalog originally rode the request-only [session-prefix seam](../../archived/feature/2026-07-07-session-prefix.md), archived; the [unified sourced-message decision](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md) moved it into durable history.)
+`dsh-tool-skill` injects one durable user-role `<system-reminder>` catalog as a sourced `user/message` at the session's first `agent/pre-step`, and only when that agent's tool view resolves this plugin's exact `skill` registration. The catalog contains sorted skill name and description only; it excludes bodies, paths, sources, providers, and routing hints. Descriptions are whitespace-normalized, XML-escaped, and capped by `catalogDescriptionMaxLength`, whose default is `500` and minimum is `3`. Full skill bodies are never included in the catalog. (The catalog originally rode the request-only [session-prefix seam](../../archived/feature/2026-07-07-session-prefix.md), archived; the [unified sourced-message decision](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md) moved it into durable history.)
 
 The registry's `list()` returns every winning summary, while model and user consumers apply the invocation predicates owned by the [independent invocation-policy decision](2026-07-28-skill-invocation-policy.md). The `skill({ name })` tool loads one model-invocable skill for the current agent cwd and returns a tool result containing `<skill_content name="...">`, `<skill_resources>`, and `<skill_instructions>`. `resourceBase` supplies a directory, URL, or opaque provider-managed base for explicitly referenced scripts, references, and assets; resources load only as needed, without directory enumeration. An unresolved name reports that the skill is unknown or no longer available; invalid names and skills with `invocation.modelInvocable: false` retain distinct tool errors. The tool result is the model-visible disclosure path.
 

.agents/notes/implemented/feature/2026-07-05-skill-system.zh.md

@@ -22,7 +22,7 @@ DeepSeek Harness 使用同一原语，使项目特定的评审、插件编写和
 
 本地 skill 的文件系统 I/O 在加载了文件系统服务时通过 `ctx.fs` 进行：项目根目录查找使用 `resolve` 和 `stat` 探测 `.git`，根目录发现使用 `listDir`，skill 读取使用 `readText`。Node 文件系统作为后备，供在不挂载 fs seam 的最小上下文中加载 `dsh-skill-local` 时使用。缺失的根目录、不可读或格式错误的 skill 文件、以及提供方 `list()` 的瞬态失败均降级为警告并跳过，使一个坏源不会导致所有 agent 请求失败；格式错误的候选项仍然快速失败，因为它们违反了提供方契约。
 
-`dsh-tool-skill` 在会话的第一个 `agent/step` 注入一个持久化的 user-role `<system-reminder>` 目录，作为带来源的 `user/message`，且仅当该 agent 的工具视图解析到本插件精确的 `skill` 注册时才注入。该目录仅包含排序后的 skill 名称与描述；不包含正文、路径、来源、提供方和路由提示。描述经过空白规范化、XML 转义，并受 `catalogDescriptionMaxLength` 上限约束，其默认值为 `500`，最小值为 `3`。完整的 skill 正文从不包含在目录中。（目录最初通过仅请求的[会话前缀 seam](../../archived/feature/2026-07-07-session-prefix.md)（已归档）传递；[统一带来源消息的决策](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md)将其移入持久化历史。）
+`dsh-tool-skill` 在会话的第一个 `agent/pre-step` 注入一个持久化的 user-role `<system-reminder>` 目录，作为带来源的 `user/message`，且仅当该 agent 的工具视图解析到本插件精确的 `skill` 注册时才注入。该目录仅包含排序后的 skill 名称与描述；不包含正文、路径、来源、提供方和路由提示。描述经过空白规范化、XML 转义，并受 `catalogDescriptionMaxLength` 上限约束，其默认值为 `500`，最小值为 `3`。完整的 skill 正文从不包含在目录中。（目录最初通过仅请求的[会话前缀 seam](../../archived/feature/2026-07-07-session-prefix.md)（已归档）传递；[统一带来源消息的决策](../architecture/2026-07-22-unified-send-and-coalesced-user-messages.md)将其移入持久化历史。）
 
 注册表的 `list()` 返回全部胜出摘要，而模型与用户消费方应用[独立调用策略决策](2026-07-28-skill-invocation-policy.md)定义的调用判定。`skill({ name })` 工具为当前 agent cwd 加载一个模型可调用的 skill，返回包含 `<skill_content name="...">`、`<skill_resources>` 和 `<skill_instructions>` 的工具结果。`resourceBase` 提供一个目录、URL 或不透明的提供方管理的基路径，用于显式引用的脚本、参考资料和资产；资源仅按需加载，不进行目录枚举。无法解析的名称报告该 skill 未知或不再可用；无效名称和 `invocation.modelInvocable` 为 `false` 的 skill 保留不同的工具错误。工具结果是面向模型的可见披露路径。
 

.agents/notes/implemented/feature/2026-07-06-sandbox.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/2026-07-06-sandbox.md
-2026-07-06-sandbox.md: 6934691e9b3f37e50bb02dfec0240a03ec9b5f30
-2026-07-06-sandbox.zh.md: cc5b7e5be0f697d77732efbf55667c258b69a95f
+2026-07-06-sandbox.md: aed5ac1ceb02130ce97a8c83c0f77869fdc32146
+2026-07-06-sandbox.zh.md: db95b1a5b7a7cae1e0fcdd8deba9dcb6ad020a67

.agents/notes/implemented/feature/2026-07-06-sandbox.md

@@ -40,7 +40,7 @@ The swap is invisible to every consumer of `ctx.bash`: the bash tools, hook comm
 
 Misconfiguration fails loud: `mode` outside the closed vocabulary is rejected at plugin load, and a host with no usable backend throws the structured `SANDBOX_UNAVAILABLE` at `confine()` rather than degrading to unconfined execution. If the selected runner rejects with attributable `ENOENT` or `EACCES`, the consumer reports the same infrastructure error from the spawn channel before any command starts; other spawn errors retain local command-start semantics while still running nothing. `runnerCommand` on `dsh-sandbox-local` is the operator's explicit assertion of a bwrap-compatible runner (chain and probes skipped); it doubles as the deterministic fake-runner seam for keyless tests.
 
-Denied file effects return a `[sandbox: file access denied under <mode> mode]` marker and instructions not to work around the denial. A confining executor adds paired `sandbox_permissions` and `justification` fields for one approved retry that must be strictly wider than the session's effective mode. A grant widens only that retry; rejection executes nothing, returns `the user rejected escalating this command to "<mode>"`, and permits no re-ask. The owner-derived runtime context states the current file policy without replacing those enforcement boundaries. When `dsh-permission` is composed with a UI adapter, one preset selects both knob values; unmatched values fold to `custom`. The [ACP automation composition](../../../../examples/acp-agent/README.md) does not mount that UI service and selects its deployment mode explicitly.
+Denied file effects return a `[sandbox: file access denied under <mode> mode]` marker and instructions not to work around the denial. A confining executor adds paired `sandbox_permissions` and `justification` fields for one approved retry that must be strictly wider than the session's effective mode. A grant widens only that retry; rejection executes nothing, returns `the user rejected escalating this command to "<mode>"`, and permits no re-ask. The owner-derived pending policy context states the current file policy without replacing those enforcement boundaries. When `dsh-permission` is composed with a UI adapter, one preset selects both knob values; unmatched values fold to `custom`. The [ACP automation composition](../../../../examples/acp-agent/README.md) does not mount that UI service and selects its deployment mode explicitly.
 
 ### Design detail
 
@@ -72,7 +72,7 @@ Backend profiles share the mode contract but differ in necessary host grants. La
 
 `dsh-bash-sandbox` extends `LocalBashExecutor`, hands `ctx.sandbox` the exact `['bash', '-c', command]` argv, and directly spawns the provider result. This leaves shell semantics and `BASH_ENV` on the inner Bash after the shipped native runner establishes confinement. A provider error propagates unchanged. A pre-process rejection is runner-owned only when the caller-owned workdir is independently usable and Node reports `ENOENT` or `EACCES` with positive provenance for provider argv[0]; a bare `syscall: 'spawn'` without an exact error path, other codes, invalid workdirs, resource failures, unrelated syscalls, and unstructured rejections retain local command-start semantics. Foreground execution converts runner-owned rejections to `SANDBOX_UNAVAILABLE` with the original detail; an asynchronous background rejection stamps `runnerFailed: true`, `denied: false`. A `SubprocessService` that synchronously throws the same provenanced shape makes background start throw `SANDBOX_UNAVAILABLE`, while other synchronous errors propagate unchanged. After a process starts, foreground and background use one runner-failure classifier that requires the rule's exit-code gate and a remaining fatal line after informational exclusions. A match outranks denial: foreground execution throws `SANDBOX_UNAVAILABLE` with that fatal line as detail; a settled `BashProcess` stamps `sandbox.runnerFailed`, and the bash producer renders it through generic `task_output`.
 
-The model sees the current effective file policy in the owner-derived `sandbox:policy` runtime context, while the static tool description explains the denial marker (`[sandbox: file access denied under <mode> mode]`), encourages attempting commands that may be denied, and forbids retrying around a denial; when the escalation fields are advertised, a denied result additionally carries the escalation hint itself, so the sanctioned same-turn retry is prompted at the decision point rather than depending on the model recalling the description (§ Escalation). [The current-policy decision](2026-07-30-current-sandbox-policy-context.md) owns the context's rationale and boundaries.
+The model sees the current effective file policy in the owner-derived `sandbox:policy` context, while the static tool description explains the denial marker (`[sandbox: file access denied under <mode> mode]`), encourages attempting commands that may be denied, and forbids retrying around a denial; when the escalation fields are advertised, a denied result additionally carries the escalation hint itself, so the sanctioned same-turn retry is prompted at the decision point rather than depending on the model recalling the description (§ Escalation). [The current-policy decision](2026-07-30-current-sandbox-policy-context.md) owns the context's rationale and boundaries.
 
 #### Escalation: one approved wider retry after a denial
 
@@ -105,7 +105,7 @@ interface SessionEventMap {
 
 Each owner exports the same three-piece kit: the event declaration, a pure fold (`effectiveSandboxMode(events)` / `effectiveApprovalPolicy(events)` — a `findLast`, typed to the domain's closed union), and THE write path (`setSandboxMode(session, mode)` / `setApprovalPolicy(session, policy)` — a switch IS its event; nothing mutates state out of band). No shared owner service, no generic facts map, no registry: a third knob copies the ~40-line pattern into its own package. Execution follows the fold on both sides — the bash tool's per-call stamp reads it as the middle rung of the § Escalation precedence chain, and the approval seam's `'never'` gate is [the approval Agent Note](2026-07-06-approval-seam.md)'s side of the same pattern.
 
-Sandbox and approval policy are rendered as ordered contributions to one runtime-context snapshot before every request. The loop records the complete snapshot as a sourced `user/message`; both `'ask'` and `'never'` are explicit, so neither owner needs switch narration or last-told state.
+Before each proposed step, sandbox and approval policy are rendered as ordered contributions to one desired policy-context message. The listener reconciles that message against session history, the claimed batch, and its pending `next-step` inbox entry. Once claimed and entered, the loop records the complete sourced `user/message`; both `'ask'` and `'never'` are explicit, so neither owner needs switch narration or last-told state.
 
 **The optional UI surface** is `PermissionService`: a deployment-defined preset table whose entries bundle one sandbox mode with one approval policy. The shipped `workspace-write` and `danger-full-access` presets write through to both domain setters; a knob combination outside the table is reported as `custom`. UI adapters may expose that table as a selector. The automation-only ACP transport advertises no configuration selector and mounts no permission-preset service.
 
@@ -149,9 +149,9 @@ Each phase gets its full design when picked up, validated against the code at th
 - **Per-session dynamic tool schemas** — rejected: schemas are registry-global by design (one assembly vocabulary, the pinned-header snapshot contract), and re-registering per session would buy only what the execution-time strict-wider check already guarantees, at the cost of a per-session schema surface and header churn on every switch.
 - **Hard-match the retry to a prior denial** — rejected: command-string identity is fragile (quoting, `workdir`, env prefixes, a pipeline retried as its failing stage) — false-rejects honest retries or is trivially satisfied; the real boundary is the human seeing command + justification. Revisit only if `allow_always` grant storage ever needs machine-checkable scopes.
 - **A generic `env/state` facts map with an owner service** — rejected: approval and sandbox compose independently, so neither's state may drag in a third package; single-key folds are one `findLast` each, dissolving the owner service; no invariant spans the knobs, so atomic multi-key patches bought nothing.
-- **Narrate each switch through `agent.inject()` plus a bus event** — rejected: independent notices expose owner ordering and intermediate combinations, while one assembly pass can materialize the complete current state atomically at the request boundary.
+- **Narrate each switch through `agent.inject()` plus a bus event** — rejected: independent notices expose owner ordering and intermediate combinations, while one pre-step composition can enqueue the complete current state atomically.
 - **State sandbox mode in the stable system prompt** — shipped first, then removed on live evidence: with `Bash commands run under the "read-only" file sandbox.` in every request, the model refused to ATTEMPT denied-then-escalatable work (five of twelve turns in the first manual session ended with zero tool calls), turning the sandbox into a soft lockout. The denial marker names the mode at the moment it matters and the escalation fields carry the recovery. The absence decision is superseded by [the current-policy decision](2026-07-30-current-sandbox-policy-context.md); this measurement and causal observation remain the evidence that any replacement must counter-test.
-- **Track "last told" with its own bookkeeping events** — rejected: the latest sourced runtime-context `user/message` records the exact full snapshot the model saw. Materializing that snapshot from current owner contributions replaces a second bookkeeping stream — events are needed only where they ARE the store.
+- **Track "last told" with its own bookkeeping events** — rejected: session history records the exact policy context the model saw, while the claimed batch and pending inbox entry show what is entering or queued. Recomputing the desired message replaces a second bookkeeping stream — events are needed only where they ARE the store.
 - **Independent sandbox and approval selectors** — rejected: one deployment-defined permission preset keeps the two policy knobs coherent for UI clients that expose runtime switching.
 
 ## Consequences
@@ -160,12 +160,12 @@ What shipped pins — the tiers in Testing hold each:
 
 - A denied command retried with `sandbox_permissions` + `justification` prompts the user through the composed answerer chain; a grant runs THAT call under the wider mode (result facts say so) while every other call keeps its own effective mode; every non-grant outcome produces its distinct error text and executes nothing.
 - The escalation fields exist exactly when the mounted executor confines; a request that is not strictly wider than the call's effective mode fails closed with its own text and prompts no one; a deployment with no ApprovalService fails escalating calls closed and leaves plain calls untouched.
-- One sourced runtime-context message states the complete current sandbox and approval policies atomically; the whole exchange — context snapshots, headers, knob events, approval notices, approvals, and results — reconstructs from the session log alone, with no policy bookkeeping events beyond the two knob events.
-- One preset selection records only changed knob values, while a no-op selection records nothing; the next request snapshots both current values atomically, and a committed sandbox switch is honored by the next call's stamp.
-- A resumed session's overrides enter its first new runtime-context snapshot with no catch-up state; a composition default changed while the process was down likewise appears in that snapshot.
+- One sourced policy-context message states the complete current sandbox and approval policies atomically; the whole exchange — context messages, headers, knob events, approval notices, approvals, and results — reconstructs from the session log alone, with no policy bookkeeping events beyond the two knob events.
+- One preset selection records only changed knob values, while a no-op selection records nothing; the next pre-step upserts both current values atomically, and a committed sandbox switch is honored by the next call's stamp.
+- A resumed session's overrides enter its first new policy-context message with no catch-up state; a composition default changed while the process was down likewise appears in that message.
 - Two concurrent sessions never see each other's state or notices.
 - Two concurrent project sessions in one Cordis context resolve independent workspace roots; bash and fs writes succeed inside the calling session's cwd and fail against its neighbor's cwd.
-- Policy ownership stays in plugins through `systemPrompt.context`, `SessionEventMap` merging, and capability-owned resolution; the generic loop change materializes every owner's ordered context as one sourced message.
+- Policy ownership stays in plugins through `SessionEventMap` merging, inbox mutation from `agent/pre-step`, and capability-owned resolution; the generic loop only claims and records the final entered batch.
 
 Costs and accepted limits:
 
@@ -191,8 +191,8 @@ Costs and accepted limits:
 - **Does the sandbox restrict network or process visibility?** No — `SandboxMode` claims FILE effects only; the bwrap profile deliberately does not unshare pid, and no backend claims network. Whether network restriction becomes its own knob is left open in § The seam.
 - **Which tools actually run confined?** OS subprocesses through `ctx.bash` — the bash tools, and hook commands transitively — plus the filesystem tools (`read`/`write`/`edit`) through the sandboxed `ctx.fs` provider (the [cross-family fs sandbox RFC](2026-07-14-cross-family-fs-sandbox.md)): bash confines via the OS runner, fs via an in-process path fence, both keying off the same `ctx.sandboxPolicy` mode. web/todo stay in-process and unfenced (web's only effect is network, outside the file-effect mode vocabulary).
 - **Does a granted escalation persist?** No. The grant is consumed by the exact foreground or background call that asked; every neighboring call keeps its own effective mode. A later background denial surfaces through `task_output` and may ground a new exact-command retry.
-- **When does a runtime mode switch take effect?** Once its session event commits, the next runtime-context snapshot and the next capability resolution fold the new mode. The sourced context message records what the model was told, and any later denial names the same policy at the point of use.
-- **What survives a restart — and what if the operator changed the config default while the process was down?** Overrides replay from the session log (`effective = fold ?? config`), so a resumed session keeps its modes with zero catch-up machinery; a default that drifted offline enters the next full runtime-context snapshot.
+- **When does a runtime mode switch take effect?** Once its session event commits, the next pre-step policy-context reconciliation and the next capability resolution fold the new mode. The sourced context message records what the model was told, and any later denial names the same policy at the point of use.
+- **What survives a restart — and what if the operator changed the config default while the process was down?** Overrides replay from the session log (`effective = fold ?? config`), so a resumed session keeps its modes with zero catch-up machinery; a default that drifted offline enters the next full policy-context message.
 - **What does `enforcement: 'partial'` on a result mean?** The selected backend enforces the subset its kernel ABI governs — e.g. Landlock before ABI v3 does not govern path truncate — and says so structurally instead of refusing the host; the probe's report line distinguishes the cases. The bwrap and Seatbelt profiles govern every promised file effect by construction, so they always report `full`.
 
 ## Prior art

.agents/notes/implemented/feature/2026-07-06-sandbox.zh.md

@@ -40,7 +40,7 @@ harness 是一个 SDK，因此约束必须是开发者可组合的能力：是
 
 配置错误会显式导致失败：`mode` 不在封闭词汇中时在插件加载时被拒绝；主机上没有可用后端时在 `confine()` 阶段抛出结构化的 `SANDBOX_UNAVAILABLE`，而非降级为无约束执行。如果所选 runner 以可归因的 `ENOENT` 或 `EACCES` 拒绝，消费方会在任何命令开始前通过 spawn 通道报告同一基础设施错误；其他 spawn 错误仍保留本地命令启动语义，同时也不会运行任何内容。`dsh-sandbox-local` 上的 `runnerCommand` 是运维人员对一个 bwrap 兼容 runner 的显式断言（跳过链和探测）；它同时充当 keyless 测试的确定性 fake-runner seam。
 
-被拒绝的文件操作返回 `[sandbox: file access denied under <mode> mode]` 标记，并附带不要绕过拒绝的指令。约束执行器添加配对的 `sandbox_permissions` 和 `justification` 字段，用于一次经批准的重试，该重试必须严格宽于会话的有效模式。授权仅放宽该次重试；拒绝则不执行任何内容，返回 `the user rejected escalating this command to "<mode>"`，且不允许再次请求。由归属方派生的运行时上下文会说明当前文件策略，但不会取代这些强制执行边界。当 `dsh-permission` 与某个 UI 适配器一起组合时，一个 preset 会同时选定两个配置项的值；不匹配的组合折叠为 `custom`。[ACP 自动化组合](../../../../examples/acp-agent/README.md)不挂载该 UI 服务，而是显式选定其部署模式。
+被拒绝的文件操作返回 `[sandbox: file access denied under <mode> mode]` 标记，并附带不要绕过拒绝的指令。约束执行器添加配对的 `sandbox_permissions` 和 `justification` 字段，用于一次经批准的重试，该重试必须严格宽于会话的有效模式。授权仅放宽该次重试；拒绝则不执行任何内容，返回 `the user rejected escalating this command to "<mode>"`，且不允许再次请求。由归属方派生的待处理策略上下文会说明当前文件策略，但不会取代这些强制执行边界。当 `dsh-permission` 与某个 UI 适配器一起组合时，一个 preset 同时选定两个旋钮值；不匹配的组合折叠为 `custom`。[ACP 自动化组合](../../../../examples/acp-agent/README.md)不挂载该 UI 服务，而是显式选定其部署模式。
 
 ### 设计细节
 
@@ -72,7 +72,7 @@ Landlock launcher 源码和包工作区位于 `native/landlock-run`，与 harnes
 
 `dsh-bash-sandbox` 扩展 `LocalBashExecutor`，把精确的 `['bash', '-c', command]` argv 交给 `ctx.sandbox`，并直接 spawn 提供方返回的 argv。这样，随附的原生 runner 建立约束后，shell 语义与 `BASH_ENV` 仍由内层 Bash 处理。提供方错误原样传播。进程启动前，只有当调用方拥有的 workdir 经独立验证可用，并且 Node 报告 `ENOENT` 或 `EACCES`，且带有明确指向提供方 argv[0] 的来源信息时，拒绝才会归因于 runner；没有精确错误路径的裸 `syscall: 'spawn'`、其他错误码、无效 workdir、资源失败、无关 syscall 与无结构拒绝保留本地命令启动语义。前台执行会将可归因于 runner 的拒绝转为 `SANDBOX_UNAVAILABLE` 并附上原始详细信息；异步后台拒绝则盖章 `runnerFailed: true`、`denied: false`。如果 `SubprocessService` 同步抛出同样带有来源信息的 `ENOENT`／`EACCES` 形态，后台启动会抛出 `SANDBOX_UNAVAILABLE`；其他同步错误原样传播。进程启动后，前台与后台共用一个 runner 失败分类器：先排除信息性行，再要求规则的退出码门控与余下的一行致命诊断同时匹配。匹配结果优先于拒绝：前台执行抛出 `SANDBOX_UNAVAILABLE`，并以该致命行作为详细信息；结算后的 `BashProcess` 会盖章 `sandbox.runnerFailed`，bash 生产者再通过通用 `task_output` 渲染它。
 
-模型会在归属方派生的 `sandbox:policy` 运行时上下文中看到当前有效的文件策略；静态工具描述则解释拒绝标记（`[sandbox: file access denied under <mode> mode]`），鼓励尝试可能被拒绝的命令，并禁止绕过拒绝重试。当升级字段被公布时，被拒绝的结果还会携带升级提示本身，使按规定进行的同轮次重试在决策点得到提示，而非依赖模型回忆描述（§ 升级机制）。[当前策略决策](2026-07-30-current-sandbox-policy-context.md)负责该上下文的理由与边界。
+模型会在归属方派生的 `sandbox:policy` 上下文中看到当前有效的文件策略；静态工具描述则解释拒绝标记（`[sandbox: file access denied under <mode> mode]`），鼓励尝试可能被拒绝的命令，并禁止绕过拒绝重试。当升级字段被公布时，被拒绝的结果还会携带升级提示本身，使被认可的同轮次重试在决策点获得提示，而非依赖模型回忆描述（§ 升级机制）。[当前策略决策](2026-07-30-current-sandbox-policy-context.md)负责该上下文的理由与边界。
 
 #### 升级机制：拒绝后一次经批准的更宽重试
 
@@ -105,7 +105,7 @@ interface SessionEventMap {
 
 每个拥有者导出相同的三件套：事件声明、纯 fold（`effectiveSandboxMode(events)` / `effectiveApprovalPolicy(events)`——一个 `findLast`，类型化到领域的封闭联合），以及唯一的写入路径（`setSandboxMode(session, mode)` / `setApprovalPolicy(session, policy)`——切换即其事件；没有任何东西在带外修改状态）。无需共享的归属服务、通用 facts map 或注册表：第三个配置项只需将约 40 行模式复制到自己的包中。执行在两侧都遵循 fold——bash 工具的按调用盖章将其作为 § 升级机制优先级链的中间层读取，approval seam 的 `'never'` 门控是[批准 Agent Note](2026-07-06-approval-seam.md) 同一模式的另一侧。
 
-沙箱策略与批准策略会在每次请求前渲染为同一份运行时上下文快照中的有序贡献。循环会将完整快照记录为一条带来源的 `user/message`；`'ask'` 与 `'never'` 都会明确写入，因此两个归属方都无需切换叙述或「上次告知」状态。
+每次拟议步骤之前，沙箱策略与批准策略都会渲染为一条目标策略上下文消息中的有序贡献。监听器将该消息与会话历史、已领取批次及其待处理的 `next-step` inbox 条目协调。消息一旦被领取并进入步骤，循环就会记录完整且带来源的 `user/message`；`'ask'` 与 `'never'` 都会明确写入，因此两个归属方都无需切换叙述或「上次告知」状态。
 
 **可选的 UI 界面**是 `PermissionService`：一张部署定义的 preset 表，每个条目捆绑一个沙箱模式与一个批准策略。随附的 `workspace-write` 和 `danger-full-access` preset 写入两个领域 setter；preset 表之外的旋钮组合报告为 `custom`。UI 适配器可以把该表暴露为选择器。仅面向自动化的 ACP 传输层不公布任何配置选择器，也不挂载权限 preset 服务。
 
@@ -149,9 +149,9 @@ fs/web/todo 在进程内执行，因此它们的沙箱语义是各自 seam 层
 - **按会话动态工具 schema**：否决。schema 设计上是注册表全局的（一套 assembly 词汇、固定 header 快照契约），按会话重新注册只能买到执行时严格放宽检查已保证的东西，代价是按会话的 schema 表面和每次切换的 header 变动。
 - **将重试硬匹配到先前的拒绝**：否决。命令字符串同一性脆弱（引号、`workdir`、env 前缀、作为失败阶段重试的管道）——要么误拒诚实的重试，要么被轻易满足；真正的边界是人看到命令 + 理由。仅在 `allow_always` 授权存储需要机器可检查的范围时才重新考虑。
 - **通用 `env/state` facts map 加拥有者服务**：否决。approval 和沙箱独立组合，因此任何一方的状态都不应拖入第三个包；单键 fold 各自是一个 `findLast`，拥有者服务自然消解；没有跨旋钮的不变式，因此原子多键补丁无收益。
-- **通过 `agent.inject()` 加总线事件逐次叙述切换**：否决。独立通知会暴露归属方顺序和中间组合，而一次组装过程可以在请求边界以原子方式具体化完整的当前状态。
+- **通过 `agent.inject()` 加总线事件逐次叙述切换**：否决。独立通知会暴露归属方顺序和中间组合，而一次 pre-step 组合可以原子排队完整的当前状态。
 - **在稳定系统提示词中声明沙箱模式**：先行交付，随后根据线上证据移除：每次请求都带有 `Bash commands run under the "read-only" file sandbox.` 时，模型会拒绝尝试本可在被拒后升级的工作（首次人工会话的十二个轮次中有五个以零工具调用结束），使沙箱变成软锁死。拒绝标记会在相关时刻指出模式，升级字段则承载恢复路径。[当前策略决策](2026-07-30-current-sandbox-policy-context.md)取代了省略策略的决策；这项测量和因果观察仍是任何替代方案必须进行反证测试的依据。
-- **用专门的簿记事件追踪「上次告知」**：否决。最新一条带来源的运行时上下文 `user/message` 会记录模型看到的确切完整快照。根据当前归属方贡献具体化该快照，取代了第二条簿记流——事件仅在它们本身即为存储时才需要。
+- **用专门的簿记事件追踪「上次告知」**：否决。会话历史记录模型看到的确切策略上下文，已领取批次与待处理 inbox 条目则表明正在进入或已经排队的内容。重新计算目标消息取代了第二条簿记流——事件仅在它们本身即为存储时才需要。
 - **相互独立的沙箱与批准选择器**：否决。一个部署定义的权限 preset 让两个策略旋钮对暴露运行时切换的 UI 客户端保持一致。
 
 ## 后果
@@ -160,12 +160,12 @@ fs/web/todo 在进程内执行，因此它们的沙箱语义是各自 seam 层
 
 - 被拒绝的命令以 `sandbox_permissions` + `justification` 重试时，通过组合的应答器链提示用户；授权使该次调用在更宽模式下运行（结果事实如此报告），而其他所有调用保持各自的有效模式；每种非授权结果产生各自不同的错误文本且不执行任何内容。
 - 升级字段恰好在已挂载的执行器约束时存在；不严格宽于调用有效模式的请求以自身文本失败关闭且不提示任何人；没有 ApprovalService 的部署对升级调用失败关闭，对普通调用不影响。
-- 一条带来源的运行时上下文消息会以原子方式声明完整的当前沙箱策略与批准策略；整个交互——上下文快照、header、旋钮事件、批准通知、批准与结果——仅从会话日志即可重建，除两个旋钮事件外没有策略簿记事件。
-- 一次 preset 选择只记录发生变化的旋钮值，而无操作的选择不记录任何内容；下一个请求会把两个当前值共同纳入一份原子快照，已提交的沙箱切换由下一次调用的盖章兑现。
-- 恢复会话的覆盖项会进入其首个新运行时上下文快照，无需追赶状态；进程停止期间变更的组合默认值也会出现在该快照中。
+- 一条带来源的策略上下文消息会以原子方式声明完整的当前沙箱策略与批准策略；整个交互——上下文消息、header、旋钮事件、批准通知、批准与结果——仅从会话日志即可重建，除两个旋钮事件外没有策略簿记事件。
+- 一次 preset 选择只记录发生变化的旋钮值，而无操作的选择不记录任何内容；下一次 pre-step 会原子 upsert 两个当前值，已提交的沙箱切换由下一次调用的盖章兑现。
+- 恢复会话的覆盖项会进入其首条新策略上下文消息，无需追赶状态；进程停止期间变更的组合默认值也会出现在该消息中。
 - 两个并发会话永远看不到彼此的状态或通知。
 - 同一个 Cordis 上下文中的两个并发项目会话解析各自独立的工作区根目录；bash 和 fs 写入在调用方会话的 cwd 内成功，对其相邻会话的 cwd 则失败。
-- 策略归属仍通过 `systemPrompt.context`、`SessionEventMap` 合并和由能力归属方拥有的解析留在插件中；通用循环变更会将每个归属方的有序上下文具体化为一条带来源的消息。
+- 策略归属仍通过 `SessionEventMap` 合并、从 `agent/pre-step` 变更 inbox，以及由能力归属方拥有的解析留在插件中；通用循环只领取并记录最终进入步骤的批次。
 
 代价与已接受的限制：
 
@@ -191,8 +191,8 @@ fs/web/todo 在进程内执行，因此它们的沙箱语义是各自 seam 层
 - **沙箱限制网络或进程可见性吗？** 不——`SandboxMode` 仅声称文件操作；bwrap profile 刻意不 unshare pid，没有后端声称网络。网络限制是否成为自己的旋钮留在 § seam 中开放。
 - **哪些工具实际在约束下运行？** 通过 `ctx.bash` 的 OS 子进程——bash 工具及传递性的钩子命令——再加上通过沙箱化 `ctx.fs` 提供方运行的文件系统工具（`read`/`write`/`edit`，见[跨工具族 fs 沙箱 Agent Note](2026-07-14-cross-family-fs-sandbox.md)）：bash 通过 OS runner 约束，fs 通过进程内路径围栏约束，二者都以同一个 `ctx.sandboxPolicy` 模式为键。web/todo 仍在进程内且不受限制（web 的唯一效果是网络，不在文件效果模式词汇内）。
 - **授权的升级会持久化吗？** 不会。授权由发起请求的确切前台或后台调用消费；每个相邻调用保留自己的有效模式。后续的后台拒绝通过 `task_output` 呈现，并且可以作为一次新的精确命令重试的依据。
-- **运行时模式切换何时生效？** 一旦其会话事件提交，下一个运行时上下文快照与下一次能力解析都会折叠新模式。带来源的上下文消息会记录模型收到的内容，之后的任何拒绝都会在使用点命名同一策略。
-- **重启后什么存活——如果运维人员在进程停止期间改了配置默认值呢？** 覆盖从会话日志回放（`effective = fold ?? config`），因此恢复的会话以零追赶机制保持其模式；离线漂移的默认值会进入下一份完整运行时上下文快照。
+- **运行时模式切换何时生效？** 一旦其会话事件提交，下一次 pre-step 策略上下文协调与下一次能力解析都会折叠新模式。带来源的上下文消息会记录模型收到的内容，之后的任何拒绝都会在使用点命名同一策略。
+- **重启后什么存活——如果运维人员在进程停止期间改了配置默认值呢？** 覆盖从会话日志回放（`effective = fold ?? config`），因此恢复的会话以零追赶机制保持其模式；离线漂移的默认值会进入下一条完整策略上下文消息。
 - **结果上的 `enforcement: 'partial'` 是什么意思？** 所选后端强制其内核 ABI 管控的子集——例如 ABI v3 之前的 Landlock 不管控路径 truncate——并以结构化方式如此声明而非拒绝主机；探测的报告行区分各种情况。bwrap 和 Seatbelt profile 构造上管控所有承诺的文件操作，因此始终报告 `full`。
 
 ## 先例