Merge master into PR #4553 and resolve runtime helper import

This commit is contained in:
07akioni committed 2026-09-21 13:39:09 +08:00
commit e0b5ab5ebd
1448 files changed
+43359 -16856

No files matched your search

@@ -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/2026-09-16-movable-mandatory-update-window.md
2026-09-16-movable-mandatory-update-window.md: 4695074429b76b467e6ec90201632dbb5a007e68
2026-09-16-movable-mandatory-update-window.zh.md: 3b416c53dc02abce1276bb70a13bb45b4c1afa44
2026-09-16-movable-mandatory-update-window.md: da996e5b37f990ff773e3ec88de0baed3c4b54f3
2026-09-16-movable-mandatory-update-window.zh.md: 15fd4eb2520a39d8259321a747b33c833b56cbb1
@@ -1,6 +1,7 @@
# Agent Note: Movable mandatory-update window
Status: implemented
Archived: 2026-09-18
English | [中文](2026-09-16-movable-mandatory-update-window.zh.md)
@@ -1,6 +1,7 @@
# Agent Note: 可移动的强更窗口
Status: implemented
Archived: 2026-09-18
[English](2026-09-16-movable-mandatory-update-window.md) | 中文
@@ -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/archived/feature/2026-08-25-unified-task-list.md
2026-08-25-unified-task-list.md: f2f955e36a78eee8fa8ff8a804293bcc37a44acd
2026-08-25-unified-task-list.zh.md: 5a4c51a6cc3d0f331d2559debb73dc02e9cf12c5
@@ -0,0 +1,49 @@
# Agent Note: One session-header task list — jobs joined with activities in the UI projection only
Status: implemented
Archived: 2026-09-03
English | [中文](2026-08-25-unified-task-list.zh.md)
Superseded: the two-roster join described here was removed with the activity seam — the job rows are now the whole roster; see [jobs absorb the record](../architecture/2026-09-01-jobs-absorb-activity-record.md).
## Problem
The [activity observation seam](2026-08-24-activity-observation-seam.md) shipped its Web surface as a second session-header button beside the [background-job list](2026-08-08-web-background-job-display.md). A background bash run registers in both planes — `ctx.jobs` for control and `ctx.activities` for observation — so the same work rendered twice: once in "background jobs" with lifecycle and duration, once in "activities" with expandable live output. Two near-identical popovers with different names for overlapping row sets read as a product accident, and the flat activity rows (live and settled interleaved, single-line, no duration) had no visual hierarchy.
## Decision
The service planes stay orthogonal; the join happens once, per row, in the UI projection.
- **`ctx.jobs` and `ctx.activities` remain separate services.** Jobs are the control plane: model-visible ids, a consuming `readOutput` cursor the model drains, `stopping`, `kill`, and the `reported` completion-notice contract. Activities are the transient observation plane: non-consuming absolute offsets, invisible to the model, gone on restart. Merging the services would force one registry to carry both a consuming and a non-consuming cursor and would blur the "model-visible ⟺ logged" line for a plane designed to stay off the session log.
- **`dsh-client-ui-activity` renders the single merged list; `dsh-client-ui-jobs` is deleted** (pre-release, no compatibility shim), and the merged entry takes the job list's slot order. Each `jobsBySession` row is joined with the activity carrying its `correlation.jobId`: the job supplies identity, lifecycle (`stopping` exists only there), duration, and the model-visible `detail`; the activity supplies the expandable output panel. An activity without a job (a workflow run) keeps its own row; an activity whose job is not in the projection is not dropped. Without the activity registry, job rows still render — they just offer no panel, so the merged control degrades to exactly the old job list.
- **User-facing vocabulary is "task" (任务)** — `N tasks running` / `N 个任务进行中` — because every row is running work to the user, while "background job" excludes workflow runs and "activity" names the internal plane, not the user concept. The dictionary namespace and package names stay `activity`: internal names follow the naming ledger, not display copy.
- **Hierarchy**: live rows first (label over a kind-badge-and-status second line, ticking duration), then settled rows compacted to one de-emphasized line; each non-empty section carries its own heading (Running / Finished). Rows without an observable activity render as static rows with no expansion affordance.
Deliberately not joined: subagent delegation rows stay bare job rows (the subagent panel is their surface), and foreground commands stay in their tool cards — neither gains an activity producer.
## Alternatives considered
- **Merge the services** — rejected for the cursor-semantics and model-visibility reasons above.
- **Keep the job list and bolt output onto it** — leaves workflow runs (activities without jobs) homeless, forcing a second list to survive anyway.
## Refinements from browser acceptance
Driving the merged list in the real app produced four presentation decisions, all scoped to the expanded panel and the header band:
- **The header band order is preset → tasks → subagent catalog.** The root-session descendant count moved out of the breadcrumb (`header.lineage`) into the actions band (order 30) and dropped its `/` separator — a root session has no hierarchy to breadcrumb; child sessions keep their `root / child` switcher where the `/` is real. `SubagentCatalogAction` owns the moved seat; the lineage renderer now renders nothing on roots.
- **The panel's copy control copies the command** (`TerminalBlock.copyText`), not the output: a long command ellipsizes in the row, so the control is its only complete source, while output is selectable text below. Supplying `copyText` also keeps the control rendered before any output exists.
- **The panel scrolls instead of folding**: commands and output lines wrap in full (`--dsl-terminal-command-whitespace` / `--dsl-terminal-line-whitespace: pre-wrap`, so the output region never scrolls sideways) and the output region caps at a fixed height (`--dsl-terminal-output-max-height`) with `maxLines: Infinity` disabling the head/tail fold.
- **The panel draws no run-state dot** (`TerminalBlock.runStateDot: false`) — the row above it already carries the same state — and reclaims the dot gutter via `--dsl-terminal-gutter`. The panel's left spine was dropped with it.
All four land as opt-ins on `TerminalBlock` (props or `--dsl-terminal-*` variables), so tool cards keep their shipped fold/ellipsis/dot behavior unchanged.
## Testing
Component suites cover the join (job lifecycle wins, activity supplies the panel, bare-job rows, standalone activities), sections, durations, and the panel options; `pnpm vitest run packages/client/ui-activity packages/client/ui-primitives packages/client/ui-subagent` runs them. The keyless web e2e pair (`background-job-list`, `live-activity-stream`) replays the merged list end to end against ARIA goldens; `agent-preset-selection` pins the header band order.
## Consequences
- The two web e2e scenarios (`background-job-list`, `live-activity-stream`) now assert one "Tasks" list; the job-list scenario's row gained a live-output panel, which is the join working end to end.
- A human kill control on live rows is still blocked on the jobs `reported` contract question recorded in the job-display note.
- The `job` locale namespace is gone with its package; `activity` owns the merged copy including the duration vocabulary.
@@ -0,0 +1,49 @@
# Agent Note: 会话头部单一任务列表——jobs 与 activity 只在 UI 投影层合并
Status: implemented
Archived: 2026-09-03
[English](2026-08-25-unified-task-list.md) | 中文
Superseded: 本文描述的双名册 join 已随 activity seam 一并移除——job 行现在就是完整名册;见 [jobs 吸收 record](../architecture/2026-09-01-jobs-absorb-activity-record.zh.md)。
## 问题
[activity 观察面](2026-08-24-activity-observation-seam.zh.md)的 Web 呈现落地为[后台任务列表](2026-08-08-web-background-job-display.zh.md)旁边的第二个会话头部按钮。一次后台 bash 会同时登记在两个面里——`ctx.jobs` 负责控制、`ctx.activities` 负责观察——于是同一份工作渲染了两次:一次在"后台任务"里带生命周期与时长,一次在"活动"里带可展开的实时输出。两个行集合高度重叠、名字却不同的近似弹层,读起来像产品事故;而活动列表的扁平行(live 与 settled 混排、单行、无时长)也缺乏视觉层次。
## 决策
服务面保持正交;合并只在 UI 投影层逐行发生一次。
- **`ctx.jobs` 与 `ctx.activities` 仍是两个服务。** jobs 是控制面:模型可见的 id、模型消耗式的 `readOutput` 游标、`stopping`、`kill` 与 `reported` 完成通知契约。activity 是瞬态观察面:非消耗式绝对偏移、模型不可见、重启即无。合并服务会迫使一个注册表同时携带消耗式与非消耗式两种游标,并模糊这个刻意不进 session log 的面与"模型可见 ⟺ 已记录"不变量的边界。
- **`dsh-client-ui-activity` 渲染唯一的合并列表;`dsh-client-ui-jobs` 删除**(预发布阶段,不留兼容层),合并入口占用原 job 列表的 slot 顺位。每个 `jobsBySession` 行与携带其 `correlation.jobId` 的 activity 逐行 join:job 提供身份、生命周期(`stopping` 只存在于此)、时长与模型可见 `detail`;activity 提供可展开的输出面板。没有 job 的 activity(workflow 运行)保留自己的行;job 不在投影里的 activity 也不会被丢弃。没有 activity 注册表时 job 行照常渲染——只是没有面板,合并控件退化为与旧 job 列表完全等价。
- **用户可见词汇统一为"任务"(task)**——`N 个任务进行中` / `N tasks running`——因为对用户而言每一行都是在跑的工作,而"后台任务"排除了 workflow 运行、"活动"命名的是内部面而非用户概念。词典命名空间与包名保持 `activity`:内部名遵循命名台账,不跟随展示文案。
- **层次**:进行中的行在前(命令为主行、kind 徽章加状态为副行、时长每秒跳动),已结束的行折为降权单行;每个非空分组带自己的标题(进行中 / 已结束)。没有可观察 activity 的行渲染为无展开交互的静态行。
刻意不合并的:subagent 委托行保持裸 job 行(subagent 面板是它的呈现面),前台命令留在工具卡片里——两者都不接活动生产者。
## 曾考虑的替代方案
- **合并服务层**——因上述游标语义与模型可见性理由否决。
- **保留 job 列表、把输出挂上去**——workflow 运行(无 job 的 activity)无处安放,第二个列表终究得活下来。
## 浏览器验收带来的细化
在真实应用里驱动合并列表得出四个呈现决策,全部限定在展开面板与头部条带:
- **头部条带顺序为 preset → 任务 → subagent 目录。** 根会话的后代数量从面包屑(`header.lineage`)移入操作区(order 30)并去掉 `/` 分隔符——根会话没有可面包屑的层级;子会话保留 `根 / 子` 切换器,那里的 `/` 是真实层级。移位后的席位由 `SubagentCatalogAction` 拥有;谱系 renderer 在根会话上不再渲染任何东西。
- **面板的复制控件复制命令**(`TerminalBlock.copyText`)而非输出:长命令在行内被省略号截断,控件是它唯一的完整来源,而输出本身就是下方可选中的文本。提供 `copyText` 还让控件在任何输出出现之前就保持渲染。
- **面板以滚动取代折叠**:命令与输出行完整换行(`--dsl-terminal-command-whitespace` / `--dsl-terminal-line-whitespace: pre-wrap`,输出区因此永不横向滚动),输出区以固定高度封顶(`--dsl-terminal-output-max-height`),`maxLines: Infinity` 禁用头尾折叠。
- **面板不画运行状态点**(`TerminalBlock.runStateDot: false`)——上方的行已携带同一状态——并通过 `--dsl-terminal-gutter` 收回状态点落区。面板左侧竖线随之移除。
四项全部以 `TerminalBlock` 的可选项落地(prop 或 `--dsl-terminal-*` 变量),工具卡片保持既有的折叠/省略/状态点行为不变。
## 测试
组件套件覆盖 join(job 生命周期优先、activity 提供面板、裸 job 行、独立 activity)、分段、时长与面板选项;`pnpm vitest run packages/client/ui-activity packages/client/ui-primitives packages/client/ui-subagent` 运行它们。无密钥 web e2e 对(`background-job-list`、`live-activity-stream`)按 ARIA golden 端到端回放合并列表;`agent-preset-selection` 钉住头部条带顺序。
## 后果
- 两个 web e2e 场景(`background-job-list`、`live-activity-stream`)现在断言同一个"任务"列表;job 场景的行获得了实时输出面板,这正是 join 端到端生效的证据。
- live 行上的人工 kill 控件仍被 job-display note 记录的 jobs `reported` 契约问题阻塞。
- `job` 词典命名空间随包一起消失;`activity` 拥有合并后的全部文案,包括时长词汇。
+6
View File
@@ -751,6 +751,9 @@
"bug-fix/2026-09-09-content-sized-diagram-previews.i18n.yaml": "sha256:a4404fe8e35ca5b6371f9cdb58fea6724cf3e7737bb2db8a3a321d534454b17c",
"bug-fix/2026-09-09-content-sized-diagram-previews.md": "sha256:5239e98e5c057d6f7e10c3f10a0f67273778d99b0c4f10b31a1921c3c6474b80",
"bug-fix/2026-09-09-content-sized-diagram-previews.zh.md": "sha256:c1ee3790a9d5c3561e1d93d5e649afa904c7d19b9d77248c85ef16daccd953e2",
"bug-fix/2026-09-16-movable-mandatory-update-window.i18n.yaml": "sha256:e526ce14ae9557e24364e3686778b130090445ef2152d976b7cc93e76eac6401",
"bug-fix/2026-09-16-movable-mandatory-update-window.md": "sha256:67546cea0b3de2b3e32957a508dbde097d91dd919fe067c8dbecb2c65b50c366",
"bug-fix/2026-09-16-movable-mandatory-update-window.zh.md": "sha256:20a0aade6ee71c89e84d70e23c89bb51e8a38c674d103dd43022feabb8d27be3",
"bug-fix/2026-09-11-session-controller-fork-turn-cut.i18n.yaml": "sha256:4442529ced1c8e48d30f7b50c65ff03eb92dce9815beb438cedbf29f234bce47",
"bug-fix/2026-09-11-session-controller-fork-turn-cut.md": "sha256:a996f24a896c7fc5d81e015db3eae27d8a46c853287572e815623ce35eae1b12",
"bug-fix/2026-09-11-session-controller-fork-turn-cut.zh.md": "sha256:6486f0eacf5067ef9850dc32f1c27b2109a7d4e9e3c498366f5c38742bb53ac5",
@@ -1375,6 +1378,9 @@
"feature/2026-08-25-loaded-turn-chat-navigation.i18n.yaml": "sha256:86bd02d56cf95ac465d0f6e326ccf75432ffb9c3db47373ae6a795195cff9ea7",
"feature/2026-08-25-loaded-turn-chat-navigation.md": "sha256:0ee1cb6da76b93e467d747aa8ba07697ba648cb213ba52202c51ee1e85dff1a8",
"feature/2026-08-25-loaded-turn-chat-navigation.zh.md": "sha256:5d3e567027fc65241856bf9f9cf0d73a20d53214c49bfbc884ab2ddd19503918",
"feature/2026-08-25-unified-task-list.i18n.yaml": "sha256:fec0d4a3ecf59aafd440bc801c686b14d63152c937f51bebe0da61ff615f6f33",
"feature/2026-08-25-unified-task-list.md": "sha256:30bd4e338a3de2f59af98d512c7878928c61846e6700d6012fc45e51c0bf94df",
"feature/2026-08-25-unified-task-list.zh.md": "sha256:88fce0a52b5387ea98c1f68742c03b2f475c5465c680befb32c587066e52cfce",
"feature/2026-08-26-web-trigger-menu-presentation-polish.i18n.yaml": "sha256:69179f3056c6d5fc35119e83dd7a5af30213502dac3c1f2d4fd43eae96ee2ecf",
"feature/2026-08-26-web-trigger-menu-presentation-polish.md": "sha256:c54e3bbe528d6be711541c61e777e6e4ed4fbc1374df97d846b015d3693679f4",
"feature/2026-08-26-web-trigger-menu-presentation-polish.zh.md": "sha256:7db49018eebbba539838d734a80fbb330e2b73fea3b343900591a3605c6c03b5",
@@ -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/2026-06-20-generic-long-running-tool-runtime.md
2026-06-20-generic-long-running-tool-runtime.md: c386d3b338c94e0e5bbf8e99c2c7a30971ba92b7
2026-06-20-generic-long-running-tool-runtime.zh.md: 002514740a15c5961960d20976fc18f7fd8071ff
2026-06-20-generic-long-running-tool-runtime.md: 5960eb268656d7e316c763631451db1c6d1c6e49
2026-06-20-generic-long-running-tool-runtime.zh.md: ea63ae82962a735b3a96e1803708172a33f83e9d
@@ -27,7 +27,7 @@ The literal types live on the [tasks subsystem page](../../../../docs/subsystems
The process-local provider also owns bounded admission, whose rationale is recorded in the [bounded background job admission decision](../../archived/bug-fix/2026-08-11-bounded-background-job-admission.md). Its positive-safe-integer `maxConcurrentJobsPerOwner` config defaults to `10`; `start()` derives each exact `Agent` object's active count from `running` and `stopping` records, while every unowned task shares one service bucket. Capacity rejection occurs before `run()` and id allocation, and producer `done` settlement is the only event that releases a stopping task's place. The provider does not queue, preempt, or retain a second mutable count.
`outputLimitBytes` is producer-owned presentation policy, not a registry buffer. The registry validates and projects it unchanged into `JobSnapshot`; generic control APIs apply the cap to complete model-facing output after adding their own status or notice metadata. Omitting it preserves the existing controller behavior, so the runtime does not impose a hidden default on unrelated producer families.
`outputLimitBytes` is producer-owned presentation policy, not a registry buffer. The registry validates and projects it unchanged into `JobView`; generic control APIs apply the cap to complete model-facing output after adding their own status or notice metadata. Omitting it preserves the existing controller behavior, so the runtime does not impose a hidden default on unrelated producer families.
A model-facing producer exposes that committed id in its canonical success value, normally `{ kind: 'background', jobId }`; Native rendering may keep human-readable prose. A pre-aborted background call fails rather than returning a no-op because no task exists to satisfy the promised handle. Once registration publishes the id, cancellation belongs to the task's own controller and the job runtime: later cancellation of the producing tool call must not kill the published task. `job_kill`, owner disposal, and service teardown request cancellation; foreground execution remains coupled to the call's `exec.signal`.
@@ -62,7 +62,7 @@ For contract-compliant producers, `AgentHandle.dispose()` resolves only after ow
- `read(id, caller?)` for a consuming stream delta or an idempotent final result.
- `kill(id, caller?, reason?)` for cancellation.
- `wait(id, timeoutMs, caller?, signal?)` for bounded terminal waiting.
- `onJobDone(listener)` for effect-scoped observation with exact-owner delivery and listener containment.
- `events.subscribe(filter, listener)` for effect-scoped observation of registration, progress, output, settlement, and removal, filtered by owner or owner scope, with listener containment.
- `attachController(name)` for the task-controller availability fence.
`wait` returns the terminal snapshot when the task settles or the live snapshot when its timeout expires. Aborting a wait cancels only that wait. If settlement has already assigned terminal delivery to the waiter, the terminal snapshot still wins. Waiters unregister synchronously on abort so a same-tick settlement cannot suppress a completion notice on behalf of a reader that receives nothing.
@@ -81,7 +81,7 @@ Stream reads share one task-scoped consuming cursor because the owning model is
The system prompt tells the model to retain job ids, continue independent work instead of busy-polling or duplicating a running task, collect relevant tasks before its final answer, and kill work that no longer matters. Completion delivers a logged message to the exact owner's session. A busy owner is injected; an idle owner is woken, under the bounded policy the [idle-owner wake decision](../feature/2026-08-11-background-job-completion-wakes-an-idle-owner.md) owns.
The runtime marks a terminal task `reported` when a read or wait delivers it, when a live waiter has claimed delivery at settlement, or when the model explicitly kills it. Reported tasks do not inject redundant completion notices. Listener failures are logged independently, do not stop later listeners, and are not awaited by waiters or teardown. When a snapshot carries `outputLimitBytes`, `dsh-tool-jobs` preserves UTF-8 boundaries and reuses an existing producer truncation marker rather than duplicating it. Reads reserve status suffixes and retain the output tail; completion notices reserve the stable `background job <id>` prefix and `job_output` instruction before truncating variable kind, label, status, detail, or the truncation marker itself, so the minimum PTY cap still identifies the task to collect. The job controller resolves the caller-visible producer cap in a prepended pre-execute listener before policy can deny or short-circuit dispatch, then applies it through the task definitions' last-mile `finalizeContent` callback so normalized tool errors, outer pipeline failures, and single-text policy results cannot escape the bound; deliberately structured multi-block policy results retain policy ownership of their shape and size.
`dsh-tool-jobs` keeps the delivery ledger: a wait claims the task when it starts (and withdraws if the task outlives the wait) and a `job_kill` claims it when accepted, so a settlement the model already received through a tool result injects no redundant completion notice; teardown settlements and unowned tasks are never announced ([seam consolidation](2026-09-03-jobs-seam-consolidation.md)). Listener failures are logged independently, do not stop later listeners, and are not awaited by waiters or teardown. When a view carries `outputLimitBytes`, `dsh-tool-jobs` preserves UTF-8 boundaries and reuses an existing producer truncation marker rather than duplicating it. Reads reserve status suffixes and retain the output tail; completion notices reserve the stable `background job <id>` prefix and `job_output` instruction before truncating variable kind, label, status, detail, or the truncation marker itself, so the minimum PTY cap still identifies the task to collect. The job controller resolves the caller-visible producer cap in a prepended pre-execute listener before policy can deny or short-circuit dispatch, then applies it through the task definitions' last-mile `finalizeContent` callback so normalized tool errors, outer pipeline failures, and single-text policy results cannot escape the bound; deliberately structured multi-block policy results retain policy ownership of their shape and size.
## Producer opt-in
@@ -105,7 +105,7 @@ Separate bash and subagent output/stop tools duplicate ids, isolation, cleanup,
### An immediate abstract task-runtime backend
The current `JobStart.run()` contract passes in-process callbacks and exact `Agent` objects. A durable backend changes identity, restart, ownership, and observation semantics, so at introduction time the registry stayed one concrete service rather than freezing the wrong boundary. The [task-registry contract Agent Note](../../archived/architecture/2026-07-26-job-registry-seam.md) later separated the contract from the process-local implementation without changing these in-process semantics.
The current `JobSpec.run()` contract returns in-process hooks and pushes output through an in-process `JobHandle`, and owners resolve to live `Agent`s in-process. A durable backend changes identity, restart, ownership, and observation semantics, so at introduction time the registry stayed one concrete service rather than freezing the wrong boundary. The [task-registry contract Agent Note](../../archived/architecture/2026-07-26-job-registry-seam.md) later separated the contract from the process-local implementation without changing these in-process semantics.
### Consumer-owned authorization or cleanup events
@@ -27,7 +27,7 @@ Status: implemented
进程内 Service Provider 还拥有有界准入,其理由记录在[有界后台任务准入决策](../../archived/bug-fix/2026-08-11-bounded-background-job-admission.md)中。它的 `maxConcurrentJobsPerOwner` 配置必须是正的安全整数,默认值为 `10`;`start()` 从 `running` 与 `stopping` 记录派生每个确切 `Agent` 对象的活动数量,而全部无 owner 任务共享一个服务级桶。容量拒绝发生在 `run()` 与 id 分配之前,处于 stopping 的任务只有在生产方 `done` 结算时才释放名额。Service Provider 不排队或抢占任务,也不保留第二份可变计数。
`outputLimitBytes` 是生产方拥有的呈现策略,而非注册表缓冲区。注册表校验该值,并将其原样投影到 `JobSnapshot`;通用任务控制器添加自身的状态或通知元数据后,再将该上限应用于完整的面向模型输出。省略该值时保持现有控制器行为,因此运行时不会向无关的生产方类别施加隐式默认值。
`outputLimitBytes` 是生产方拥有的呈现策略,而非注册表缓冲区。注册表校验该值,并将其原样投影到 `JobView`;通用任务控制器添加自身的状态或通知元数据后,再将该上限应用于完整的面向模型输出。省略该值时保持现有控制器行为,因此运行时不会向无关的生产方类别施加隐式默认值。
面向模型的生产方会在规范成功值中暴露已提交的 id,通常为 `{ kind: 'background', jobId }`;Native 渲染仍可保留便于人类阅读的行文。预先被中止的后台调用会失败,而不是返回空操作,因为不存在可履行所承诺句柄的任务。一旦注册过程发布 id,取消就归任务自身的控制器与任务运行时所有:随后取消生产工具调用不得终止已发布的任务。`job_kill`、所有者资源释放和服务拆除会请求取消;前台执行仍与调用的 `exec.signal` 耦合。
@@ -62,7 +62,7 @@ job id 在运行时全局可见且可预测,因此注册表会授权每次访
- `read(id, caller?)`:消费式流增量或幂等的最终结果。
- `kill(id, caller?, reason?)`:取消。
- `wait(id, timeoutMs, caller?, signal?)`:有界的终止等待。
- `onJobDone(listener)`:effect 作用域内的观察,具有精确所有者投递和监听器隔离。
- `events.subscribe(filter, listener)`:effect 作用域内对注册、进度、输出、结算与移除的观察,按所有者或所有者 scope 过滤,具有监听器隔离。
- `attachController(name)`:任务控制器可用性防线。
`wait` 在任务完成时返回终止快照,在等待超时时返回当前快照。中止一次等待只取消该次等待。如果结算已经将终止投递分配给该等待方,终止快照仍然优先。等待方在中止时同步注销,因此同一 tick 内发生结算时,不会代表一个实际未收到任何内容的读取方压制完成通知。
@@ -81,7 +81,7 @@ job id 在运行时全局可见且可预测,因此注册表会授权每次访
系统提示词要求模型保留 job id、在后台工作运行时继续处理独立工作而非忙轮询或重复启动同一任务、在给出最终答案前收集相关任务,并终止不再重要的工作。完成时,系统会向确切所有者的会话交付一条已记录的消息:繁忙的所有者走注入,空闲的所有者会被唤醒,其有界策略由[空闲所有者唤醒决策](../feature/2026-08-11-background-job-completion-wakes-an-idle-owner.zh.md)负责。
当读取或等待交付终止任务、尚在等待的等待方在结算时认领了投递,或模型显式终止任务时,运行时将终止任务标为 `reported`。已报告的任务不会注入冗余的完成通知。监听器失败会独立记录,不会阻止后续监听器,也不会被等待方或资源销毁过程等待。当快照携带 `outputLimitBytes` 时,`dsh-tool-jobs` 会保持 UTF-8 边界,并复用生产方已有的截断标记,而不会重复添加。读取会为状态后缀预留空间并保留输出尾部;完成通知会先为稳定的 `background job <id>` 前缀与 `job_output` 指令预留空间,再截断可变的 kind、label、status、detail,乃至截断标记本身,因此 PTY 的最小上限仍能标识需要收集的任务。任务控制器在策略有机会拒绝或短路分发之前,于最先执行的 pre-execute 监听器中解析调用方可见的生产方上限;随后通过任务定义最后一道的 `finalizeContent` 回调应用该上限,使规范化的工具错误、外层流水线失败与单文本策略结果都无法绕过该边界;经特意结构化的多块策略结果仍由策略拥有其形状与大小。
`dsh-tool-jobs` 持有投递台账:等待在开始时认领任务(任务比等待活得更久则撤回),`job_kill` 被接受时也认领,因此模型已经通过工具结果收到的结算不会再注入冗余的完成通知;teardown 结算与无主任务从不播报([seam 收敛](2026-09-03-jobs-seam-consolidation.zh.md))。监听器失败会独立记录,不会阻止后续监听器,也不会被等待方或资源销毁过程等待。当视图携带 `outputLimitBytes` 时,`dsh-tool-jobs` 会保持 UTF-8 边界,并复用生产方已有的截断标记,而不会重复添加。读取会为状态后缀预留空间并保留输出尾部;完成通知会先为稳定的 `background job <id>` 前缀与 `job_output` 指令预留空间,再截断可变的 kind、label、status、detail,乃至截断标记本身,因此 PTY 的最小上限仍能标识需要收集的任务。任务控制器在策略有机会拒绝或短路分发之前,于最先执行的 pre-execute 监听器中解析调用方可见的生产方上限;随后通过任务定义最后一道的 `finalizeContent` 回调应用该上限,使规范化的工具错误、外层流水线失败与单文本策略结果都无法绕过该边界;经特意结构化的多块策略结果仍由策略拥有其形状与大小。
## 生产方显式启用
@@ -105,7 +105,7 @@ bash seam 暴露 `resolve`、`run` 和 `start`。`start(spec)` 返回一个 `She
### 立即抽象任务运行时后端
当前 `JobStart.run()` 约定传入进程内回调与确切的 `Agent` 对象。持久化后端会改变身份、重启、所有权与观察语义,因此在引入之时注册表保持为单一具体服务,而非固化错误的边界。[任务注册表约定 Agent Note](../../archived/architecture/2026-07-26-job-registry-seam.md)后来在不改变这些进程内语义的前提下,将约定与进程内实现分离。
当前 `JobSpec.run()` 约定返回进程内钩子并经进程内的 `JobHandle` 推送输出,所有者也在进程内解析为存活的 `Agent`。持久化后端会改变身份、重启、所有权与观察语义,因此在引入之时注册表保持为单一具体服务,而非固化错误的边界。[任务注册表约定 Agent Note](../../archived/architecture/2026-07-26-job-registry-seam.md)后来在不改变这些进程内语义的前提下,将约定与进程内实现分离。
### 由消费方负责授权或清理事件
@@ -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: 09631349226837d7f93112fe5e9c517f2fee099c
2026-07-16-explicit-turn-cancellation.zh.md: baf039d49578895d53ffae4c6b88d29b3d1563b6
2026-07-16-explicit-turn-cancellation.md: d59accf2dc3f894ffab226e5ddc9d2e746f4d93e
2026-07-16-explicit-turn-cancellation.zh.md: 8fb0728eb2bbe8fa01dbefebe7e02aa8c836617a
@@ -12,7 +12,7 @@ The [initiating Agent scope decision](2026-07-15-agent-initiator-scope.md) inten
## Decision
Agent owns the runtime-only `AgentCancelCause` union `{ kind: 'user' } | { kind: 'parent' }`; `agent.cancel()` defaults to `user`. TypeScript enforces that vocabulary at this typed same-process boundary, with no runtime validator, fallback, or special compatibility contract for untyped callers. An active `TurnCancellation` copies the typed discriminant into a fresh frozen signal reason; idle cancellation has no holder to mutate and does not arm later work.
Agent owns the runtime-only `AgentCancelCause` union `{ kind: 'user' } | { kind: 'parent' } | { kind: 'hook'; reason: string } | { kind: 'disposed' }`, and `agent.cancel()` requires one. TypeScript enforces that vocabulary at this typed same-process boundary, with no runtime validator, fallback, or special compatibility contract for untyped callers. The holder aborts with the caller's own object rather than a detached copy, so a transport reaching that signal can extend it: Node's fetch assigns a `stack` onto the reason it receives. The loop therefore copies the declared fields where the cause becomes durable data. Freezing the cause at `cancel()` is not an alternative — undici performs that assignment from strict-mode code, so a frozen reason makes `fetch` reject with `TypeError: Cannot add property stack` in place of the abort reason. Idle cancellation has no holder to mutate and does not arm later work.
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.
@@ -22,7 +22,7 @@ The driver keeps only a cause-less pre-run marker for queued work cancelled befo
The explicit event signatures pass a single payload object: agent-scoped events carry `agent` and `signal` in the payload with `next` last, and the remaining APIs keep `signal` immediately before a waterfall's final `next`. `PreStepContext` and `RequestFailureContext` are retired, with their fields folded into the `agent/pre-step` and `agent/request-error` payloads ([payload-object events](../../archived/architecture/2026-08-06-agent-event-payload-objects.md)). 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 provider-managed termination and join boundary; the [native-containment decision](2026-08-28-subprocess-native-containment.md) owns the supported local scope, Job, and fallback mechanics. `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.
`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 typed 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.
Agent disposal requests the runtime-only `{ kind: 'disposed' }` interruption on the active holder. If cancellation already won the controller reason, the reason cannot be rewritten, so terminal classification first checks lifecycle state: disposed wins, then a supported `user` or `parent` cause becomes the coarse aborted outcome, and unrelated exceptions retain the existing error path. ACP cancellation maps to `user`; in-process spawn and fork propagation map to `parent`. Remote ACP subagents retain their existing wire protocol.
@@ -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 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.
Contract tests verify the typed caller union, field copying where the cause becomes durable data, 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.
@@ -12,7 +12,7 @@ Status: implemented
## 决策
Agent 拥有仅用于运行时的 `AgentCancelCause` 联合类型 `{ kind: 'user' } | { kind: 'parent' }`;`agent.cancel()` 默认使用 `user`。TypeScript 在这个类型化的同进程边界中强制执行该词汇,不提供运行时校验器、后备行为,也不为无类型调用方提供特殊兼容性约定。活跃的 `TurnCancellation` 会把类型化判别字段复制为一个全新且已冻结的 signal 原因;空闲状态下没有可修改的持有者,也不会让后续工作预先进入取消状态。
Agent 拥有仅用于运行时的 `AgentCancelCause` 联合类型 `{ kind: 'user' } | { kind: 'parent' } | { kind: 'hook'; reason: string } | { kind: 'disposed' }`,`agent.cancel()` 必须显式传入其一。TypeScript 在这个类型化的同进程边界中强制执行该词汇,不提供运行时校验器、后备行为,也不为无类型调用方提供特殊兼容性约定。持有者中止时携带的是调用方自己的那个对象,而非一份分离的拷贝,因此能接触到该 signal 的传输层可以向其添加属性:Node 的 fetch 会给它收到的 reason 赋一个 `stack`。所以循环在 cause 成为持久数据的那一步拷贝已声明字段。在 `cancel()` 处冻结 cause 并非可选方案——undici 是在严格模式代码中执行该赋值的,被冻结的 reason 会让 `fetch` 以 `TypeError: Cannot add property stack` 取代中止原因而拒绝。空闲状态下没有可修改的持有者,也不会让后续工作预先进入取消状态。
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。
@@ -22,7 +22,7 @@ AgentLoop 为每个待启动轮次私有地持有一个 `TurnCancellation`。它
显式事件签名传递单个 payload 对象:agent 作用域事件在 payload 中携带 `agent` 和 `signal`,`next` 位于最后;其余 API 保持 `signal` 紧邻 waterfall(瀑布式事件)的最终 `next` 之前。`PreStepContext` 与 `RequestFailureContext` 已退役,其字段并入 `agent/pre-step` 与 `agent/request-error` 的 payload([payload-object 事件](../../archived/architecture/2026-08-06-agent-event-payload-objects.md))。进入 pre-step 时、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 执行器由提供方管理的终止与等待边界;[原生 containment 决策](2026-08-28-subprocess-native-containment.zh.md)负责受支持本地路径上的 scope、Job 与 fallback 机制。`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。
`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,且总是携带类型化的 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。
Agent dispose(资源释放)会在活跃持有者上请求仅用于运行时的 `{ kind: 'disposed' }` 中断。若取消已经先占用控制器的中断原因,该原因便无法改写,因此终态分类会先检查生命周期状态:资源释放结果优先,之后受支持的 `user` 或 `parent` 取消原因形成粗粒度的中止结果,其他异常保留现有错误路径。ACP(Agent Client Protocol)取消映射为 `user`;进程内 spawn 和 fork 的传播映射为 `parent`。远程 ACP subagent 保持现有协议。
@@ -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
约定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。
约定测试验证类型化调用方联合类型、在 cause 成为持久数据处拷贝字段、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。
@@ -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: 32be87f3a0a4f8da92100ec7351af788862532f7
2026-07-25-web-client-session-scope-and-provide-channel.zh.md: b2e7efae98efb69d5f34d6719a382221ffc7a00d
2026-07-25-web-client-session-scope-and-provide-channel.md: 76561bc0e330b3e3a26607569ae327217d0a0b67
2026-07-25-web-client-session-scope-and-provide-channel.zh.md: d5d92b2ca53ba78929dd8591a6ded7028e4fed24
@@ -26,7 +26,7 @@ The reference-owned lifetime and Provider targeting now follow [Client Session r
Host-side `session.create(workspaceId)` produces Session + Agent + cwd in one piece (an atomic bundle, never split); the client side is the mirror of that birth — the instant a session row enters the list mirror, the client mints its Agent scope (actx + provide + the full input surface mounted):
- Session identity is the host's true form from birth: the sessionId arrives via the `session.create` response / the `host/session-added` frame, and every client-side address (the scope tag, slot store keys, RPC addressing) uses that same id.
- The materialization moment = the instant the user picks a Workspace (cwd settled): the client calls `session.create({workspaceId})` on the spot and receives the complete entity.
- A manual Workspace pick or successful [first-use startup initialization](../feature/2026-09-20-default-workspace.md) establishes cwd before the client calls `session.create({workspaceId})` and receives the complete entity.
- "New Session with no workspace picked" is a **pure view state** (a navigation position) corresponding to no session/scope entity; until the pick, the composer is locked whole (no slash, no plain text).
- A "blank session" is an ordinary materialized session with no turn yet; to Agent-scope plugins on the host (goal/plan/skill/…) it remains an ordinary Session, so slash/plan are naturally live.
@@ -26,7 +26,7 @@ web client 只有一张全局会话面:slot 全部从根上下文渲染,插
host 侧 `session.create(workspaceId)` 一体产出 Session + Agent + cwd(作为不可拆分的原子整体);client 侧就是这次出生的镜像——会话行进入 list mirror 的瞬间,client 为它铸 Agent scope(actx + provide + 输入面全套挂上):
- 会话身份自出生即为 host 真身:sessionId 由 `session.create` 响应 / `host/session-added` 帧带来,client 侧一切寻址(scope tag、slot store 键、RPC 地址)用的都是同一个 id。
- 实体化时点 = 用户选定 Workspace(cwd 确定)的瞬间:client 当场调 `session.create({workspaceId})`,拿到完整实体。
- 手动选定 Workspace 或完成[首次使用的启动初始化](../feature/2026-09-20-default-workspace.zh.md) 后,cwd 已确定,client 再调 `session.create({workspaceId})`,拿到完整实体。
- 「New Session 且未选 workspace」是**纯视图态**(一个导航位置),不对应任何会话/scope 实体;选定之前 composer 整体锁死(无 slash、无纯文本)。
- 「空会话」是尚无轮次的普通实体化会话;对 host 上的 Agent-scope 插件(goal/plan/skill(技能)/…)而言,它仍是普通 Session,因此 slash/plan 自然可用。
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-02-typert-remote-method-calls.md
2026-08-02-typert-remote-method-calls.md: b599b3f9a353738bbb44a4da9b5a72be8309e2fb
2026-08-02-typert-remote-method-calls.zh.md: 64dc84f0ae07b10d50452ef7e36c30db102953cf
2026-08-02-typert-remote-method-calls.md: 2a12aaf78f638d8af786d2ca817a3fee07a2c614
2026-08-02-typert-remote-method-calls.zh.md: bf4a7b0179b4dc878a0a53c7076c02e6f1bf0273
@@ -31,8 +31,8 @@ The Remote consumer projection contains `.d.ts`, `.d.ts.map`, and `.js` files. T
| `@deepseek-ai/dsh-typert-protocol` | Declares only the minimal `ctx.typert` protocol | `TypertRemoteService`, decorators, binding fallback, descriptors, lookup/Context, and the Remote map; no dependency on the compiler, Zod, Connection, or Browser |
| Typert registry | `ctx.typert` | Separately stores reflection for the current environment, imported Remote contributions, lookup providers, and Context providers |
| Typert generator/loader | No new business service | Generates three kinds of `lib` artifacts from the Host/Client Programs and registers the current environment's artifacts with `ctx.typert` |
| API Gateway's Host face | `ctx.typertGateway` | Associates Host definitions with live Services, decodes parameters, resolves receivers, and invokes methods |
| Connection | `ctx.connection` | Exclusively owns the HTTP Server/future WebSocket, the shared `/api` route, RPC envelope, rpcId, serialization, trust, error transport, Typert interception, and owner-registered exact Fetch routes on the same channel |
| API Gateway's Host face | `ctx.typertGateway` | Associates Host definitions with live Services, decodes parameters, resolves receivers, invokes methods, and projects unary results into carrier-neutral fields and attachments |
| Connection | `ctx.connection` | Exclusively owns the HTTP Server/future WebSocket, the shared `/api` route, RPC envelope, rpcId, physical serialization, trust, error transport, Typert interception, and owner-registered exact Fetch routes on the same channel |
| API Gateway's Client face | `ctx.remote`, `ctx.remote.<namespace>` | Mounts Remote contributions, materializes each namespace as a traced `remote.<namespace>` child Service, and delegates canonical calls to `ctx.connection.rpc` |
| API Remotes | No new service | Owns Host Agent/Session lookup policy and serves as the only Client business facade, selecting and mounting `/remote` contributions while exposing the selected API declarations |
| Agent/Session owning packages | Existing domain services | Provide both static interface merges and runtime lookup/Context providers |
@@ -181,7 +181,7 @@ Consequently, `SessionId`, the Agent wire ID, the request, and the result all re
Remote methods themselves use declaration-map navigation. Typert anchors `InvocationModel.location` to the decorated Host method-name token and emits a source-map segment on the corresponding property of the namespace interface. For an adapter-backed endpoint, after the TypeScript editor resolves `ctx.remote.models.list` to its generated declaration, `typert.remote-client.d.ts.map` takes it to the Host Service's `remoteExportList` entry point. That entry point explicitly calls the existing, unrenamed `list()` method; the map does not misidentify the decorator, class, or full signature as the method definition.
Typert generates a wire Zod codec for the same symbol key. The Host Gateway uses parameter and identity codecs to validate input. Client Remote trusts its generated TypeScript arguments and successful Host results instead of executing invocation codecs. If a complex type cannot produce a strict codec, the LIB build fails instead of degrading to `unknown` or unchecked JSON.
Typert generates a wire Zod codec for the same symbol key. The Host Gateway uses parameter and identity codecs to validate input. Client Remote trusts its generated TypeScript arguments and JSON success results. Binary success results use their generated codec after multipart decoding, as specified by [binary Remote transfer](2026-09-17-workspace-file-binary-transfer.md). If a complex type cannot produce a strict codec, the LIB build fails instead of degrading to `unknown` or unchecked JSON.
Named business types referenced by Remote methods must be exported from public, type-only subpaths. If the only reachable entry also imports Host Services, Cordis `Context` merges, or Host-only implementations, the build fails and requires the business package to provide a safe type entry. Primitives, literals, and simple compositions explicitly supported by Typert need no additional names.
@@ -514,7 +514,7 @@ Canonical public types require business DTOs to have type-only entries, which ma
Type imports and runtime contributions have different effects. `import type {}` extends only the static Remote surface. If a real calling environment omits the value contribution, the Client Remote Service must fail with an explicit "Remote not mounted" error.
Generated Host and Client artifacts carry matching Zod factories, but Client Remote does not materialize invocation schemas. Canonical symbol keys, the same generated model, and Host wire validation keep the two sides aligned without comparing schema object identities across realms.
Generated Host and Client artifacts carry matching Zod factories; Client Remote materializes only binary result schemas. Canonical symbol keys, the same generated model, and Host wire validation keep the two sides aligned without comparing schema object identities across realms.
A consumer may import a Remote contract that is not currently mounted on the Host. The types mean "this protocol capability was selected by the consumer," not that a corresponding Service currently exists in the target process; an unavailable endpoint must fail explicitly at runtime.
@@ -31,8 +31,8 @@ Remote 消费端投影同时包含 `.d.ts`、`.d.ts.map` 和 `.js`。`.d.ts` 只
| `@deepseek-ai/dsh-typert-protocol` | 只声明 `ctx.typert` 的最小协议 | `TypertRemoteService`、decorator、binding 回退、descriptor、lookup/Context 和 Remote map;不依赖 compiler、Zod、Connection 或 Browser |
| Typert registry | `ctx.typert` | 分开保存当前环境 reflection、导入的 Remote contribution、lookup provider 和 Context provider |
| Typert generator/loader | 无新增业务服务 | 从 Host/Client Program 生成三类 `lib` 产物,并把当前环境产物注册到 `ctx.typert` |
| API Gateway 的 Host face | `ctx.typertGateway` | 关联 Host definition 与活 Service,解码参数、解析 receiver 并调用方法 |
| Connection | `ctx.connection` | 独占 HTTP Server/未来 WebSocket、共享 `/api` route、RPC envelope、rpcId、序列化、trust、错误传输、Typert 拦截,以及各 owner 在同一 channel 上注册的精确 Fetch route |
| API Gateway 的 Host face | `ctx.typertGateway` | 关联 Host definition 与活 Service,解码参数、解析 receiver、调用方法,并将一元结果投影为载体无关的字段与附件 |
| Connection | `ctx.connection` | 独占 HTTP Server/未来 WebSocket、共享 `/api` route、RPC envelope、rpcId、物理序列化、trust、错误传输、Typert 拦截,以及各 owner 在同一 channel 上注册的精确 Fetch route |
| API Gateway 的 Client face | `ctx.remote`、`ctx.remote.<namespace>` | mount Remote contribution,把每个 namespace 实体化为可追踪的 `remote.<namespace>` 子 Service,并把规范调用交给 `ctx.connection.rpc` |
| API Remotes | 无新增服务 | 负责 Host Agent/Session lookup 策略,并作为 Client 业务的唯一 facade,选择并挂载 `/remote` contribution,同时暴露所选 API 声明 |
| Agent/Session owning 包 | 既有领域服务 | 同时提供静态 interface merge 与运行时 lookup/Context provider |
@@ -181,7 +181,7 @@ import type { CreateGoalRequest, CreateGoalResult } from '@deepseek-ai/dsh-goal/
Remote 方法本身使用 declaration map 导航。Typert 把 `InvocationModel.location` 固定在 Host 被装饰方法的方法名 token,并在 namespace interface 的对应属性上写入 source-map segment。对于由适配器支撑的 endpoint,TypeScript editor 从 `ctx.remote.models.list` 取得生成 declaration 后,再沿 `typert.remote-client.d.ts.map` 跳到 Host Service 的 `remoteExportList` 远程出口。该出口继续显式调用不改名的存量 `list()`,map 不把 decorator、class 或整个签名误当成方法定义位置。
Typert 为同一 symbol key 生成 wire Zod codec。Host Gateway 用参数与 identity codec 校验输入;Client Remote 信任生成的 TypeScript 参数与成功的 Host 结果,不执行调用 codec。复杂类型无法生成严格 codec 时,LIB 构建失败,不降级为 `unknown` 或无校验 JSON。
Typert 为同一 symbol key 生成 wire Zod codec。Host Gateway 用参数与身份 codec 校验输入。Client Remote 信任生成的 TypeScript 参数与 JSON 成功结果。二进制成功结果在 multipart 解码后执行生成的结果 codec,具体由[二进制 Remote 传输](2026-09-17-workspace-file-binary-transfer.zh.md)规定。复杂类型若无法生成严格 codec,LIB 构建失败,不降级为 `unknown` 或未校验 JSON。
Remote 方法引用的命名业务类型必须从纯类型公共 subpath 导出。如果唯一可达入口会带入 Host Service、Cordis `Context` merge 或 Host-only 实现,构建失败并要求业务包提供安全的类型出口。原始值、字面量和 Typert 明确支持的简单组合不需要额外命名。
@@ -514,7 +514,7 @@ SRC 弱 descriptor 不验证普通 JSON 内部结构。Host Remote 签名变化
类型 import 与运行时 contribution 是两种不同效果。`import type {}` 只扩展静态 Remote surface;真实调用环境遗漏 value contribution 时,Client Remote Service 必须以明确的「Remote 未挂载」错误失败。
生成的 Host 与 Client 产物携带匹配的 Zod factory,但 Client Remote 不实例化调用 schema。规范 symbol key、同一生成模型和 Host wire 校验让两侧保持一致,而无需跨 realm 比较 schema 对象 identity。
生成的 Host 与 Client 产物携带匹配的 Zod factory;Client Remote 仅实例化二进制结果 schema。规范 symbol key、同一生成模型和 Host wire 校验让两侧保持一致,而无需跨 realm 比较 schema 对象 identity。
消费端可以导入 Host 当前未挂载的 Remote contract。类型表示「该协议能力已被消费端选择」,不保证目标进程当前存在对应 Service;运行时 endpoint 不可用必须明确失败。
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-10-remote-event-delivery.md
2026-08-10-remote-event-delivery.md: 1706cc525bd7707874353a6658ec2f52597044a9
2026-08-10-remote-event-delivery.zh.md: ac79642e90791892870c8920e02867faea180b07
2026-08-10-remote-event-delivery.md: 3eaa26fb92c40b3a20cab11d4522bcf68b8a207f
2026-08-10-remote-event-delivery.zh.md: d41589f1b800da98342210d2e429a2b67b6c61f2
@@ -195,3 +195,7 @@ The few required Client symbols are mirrored on the test side: `scaffold.ts` exp
- **Capabilities deliberately omitted.** Payload projection and redaction are unsupported, scopes other than Agent are unsupported, and ordinary notifications are not replayed. Recoverable state needs a query, cursor, or opening baseline; a waterfall is replayed only while its original Host invocation remains pending.
- **Some Client packages remain in the Host graph.** Twelve projects, including `connection`, `runtime`, and `ui-slots`, remain reachable through unsplit `directory-picker-browse`/`-native` and `api/gateway → client/connection`. They compile and no longer pull in API Remotes' Client face, so this change does not split them. Direct `dsh-client-runtime/client` imports in two chat e2e files rely on Runtime's current presence in that graph rather than a general guarantee.
- **The package intentionally publishes no invariant companion.** A prior revision asserted delivery form on the live event bus, coupling diagnostics to the allowlist and causing Rolldown to emit a third bundle chunk omitted by the mechanically derived publication list. The Host-face `TypertForwardableEventEntry` assertion already rejects those mismatches at compile time, and the package README records why no independent runtime relation remains.
## Related
- [Remote duplex stream](2026-09-19-remote-duplex-stream.md): every Remote stream carries a Client uplink on the same logical stream; the duplex `$events` listed under its Deferred section replaces the unary `$events/result` path once it lands.
@@ -197,3 +197,7 @@ Client 要求首项是带非空 `clientId` 与 `host.home` 的 `ready`;后续
- **放弃的能力**:不支持投影或脱敏载荷,不支持 Agent 以外的 Scope,也不为普通通知提供重放。需要可靠恢复的状态必须拥有查询、cursor 或 opening baseline;waterfall 只重放仍处于同一次 Host 调用生命周期内的 pending request。
- **仍有 client 包留在 host 图里**:12 个工程(`connection`、`runtime`、`ui-slots` 等)经未拆分的 `directory-picker-browse`/`-native` 与 `api/gateway → client/connection` 仍可达 host 图。它们都能编译且不再牵连 api/remotes 的 client face,因此没有阻塞本次改动;拆分那些包能减少几个,但经评估后不做。两个 chat e2e 直接引 `dsh-client-runtime/client` 依赖 `runtime` 本来就在图里——属偶然而非保证。
- **本包不发布 invariant companion**:早先的修订曾在活事件总线上断言投递形状(`thisArg === null`、`mode === 'emit'`),这让诊断逻辑与名单值耦合,并使 rolldown 把它提成第三个 bundle chunk——而机械推导的发布文件清单并不携带它。Host 面的 `TypertForwardableEventEntry` 断言已在编译期拒绝这些偏离,包 README 也记录了不再存在独立运行时关系的原因。
## 相关
- [Remote 双工流](2026-09-19-remote-duplex-stream.zh.md):所有 Remote 流在同一逻辑流上承载客户端上行;其「后续」列出的双工 `$events` 落地后取代 unary `$events/result` 路径。
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-25-electron-desktop-packaging-and-updates.md
2026-08-25-electron-desktop-packaging-and-updates.md: f58e8dc471faeb717dd79d32df4afc1b5cc3458f
2026-08-25-electron-desktop-packaging-and-updates.zh.md: 2d2d5c68f28974b606693d5b6727c864c92a0763
2026-08-25-electron-desktop-packaging-and-updates.md: a985e1c91a1ccdb9c4aab0bc519974f9b6fafd4a
2026-08-25-electron-desktop-packaging-and-updates.zh.md: 17ace865477d4fae3650b56d6725be07c7bd0e0f
@@ -75,6 +75,8 @@ The [immediate-window decision](2026-09-09-desktop-immediate-window-and-direct-s
`DSH_DESKTOP_AUTO_UPDATE_ENV` selects the test deployment by default or the production deployment for both the target-specific generic-provider URL and COS destination. Release automation supplies the test HTTPS origin through `DOWNLOAD_TEST_ORIGIN` and each deployment's bucket through `DOWNLOAD_TEST_COS_BUCKET` or `DOWNLOAD_PROD_COS_BUCKET`; keeping mutable test routing and COS storage identities out of source lets deployment infrastructure change without a code release, while the public production origin remains fixed. Packaging resolves only the public updater URL, disables electron-builder publishing, removes every COS credential field from its subprocess environment, and writes a completion record only after electron-builder and every signing or notarization hook succeeds. Target upload additionally requires the selected bucket, then requires the completion record, root dsh version, Desktop version, version-derived channel metadata, artifact names, sizes, and SHA-512 values to agree before it reads the selected credentials or sends data. It uploads immutable versioned updater payloads and any separate blockmaps before replacing the channel metadata emitted by electron-builder, and it never deletes historical objects. Stable versions use the `latest` metadata name; prereleases use the first semantic-version prerelease identifier. NSIS embeds its blockmap in the signed executable; the macOS ZIP carries a separate blockmap. Both let electron-updater download changed blocks when supported, while application replacement and the local pnpm package operation remain separate operations.
Test releases require `DOWNLOAD_TEST_RELEASE_ID` from the platform dotenv file. Their feeds and binaries share `dsh-desk/<release-id>/`; production retains its fixed directories. The ID is part of the packaged update URL, so upload rejects a different ID. Internal distribution can supply a new installer after rotation; installed clients retain their existing feed. Random directories reduce guessing but provide no authorization for link holders. The [release configuration](../../../../apps/desktop/README.md#upload-updates) defines generation and reuse.
## Security and release policy
Core dsh and the private Desktop Host come only from the signed application resource tree. Plugin installation forwards package specs to pnpm, including local and remote sources, but never accepts raw pnpm commands. pnpm owns dependency resolution and the profile’s `allowBuilds` policy; the Host loads activated bundles.
@@ -75,6 +75,8 @@ Electron 更新只使用一个 `electron-updater` 发布流和签名 `electron-b
`DSH_DESKTOP_AUTO_UPDATE_ENV` 默认为测试部署,也可以选择生产部署,并同时决定目标专用的 generic-provider URL 与 COS 目标。发布自动化通过 `DOWNLOAD_TEST_ORIGIN` 提供测试 HTTPS origin,并通过 `DOWNLOAD_TEST_COS_BUCKET` 或 `DOWNLOAD_PROD_COS_BUCKET` 提供各部署的 bucket;可变的测试路由与 COS 存储身份不写入源码,部署基础设施变更时无需发布新代码,而公开的生产 origin 仍固定。打包只解析公开更新 URL、禁止 electron-builder 发布、从子进程环境中删除每个 COS 凭据字段,并且只有在 electron-builder 以及每个签名或公证 hook 成功后才写入完成记录。目标上传还必须提供所选 bucket,随后会先要求完成记录、根 dsh 版本、Desktop 版本、根据版本得出的频道元数据、产物名称、大小与 SHA-512 全部一致,再读取所选凭据或发送数据。它先上传不可变且带版本的更新载荷与所有独立 blockmap,最后替换 electron-builder 生成的频道元数据,并且不会删除历史对象。稳定版本使用 `latest` 元数据名称,预发布版本则使用语义化版本的第一个预发布标识符。NSIS 把 blockmap 嵌入已签名的可执行文件,macOS ZIP 则使用独立 blockmap;两者都让 electron-updater 在平台支持时只下载变化的数据块,而应用替换与本地 pnpm 包操作仍是两个独立操作。
测试发布必须从平台 dotenv 文件提供 `DOWNLOAD_TEST_RELEASE_ID`。清单和安装包共用 `dsh-desk/<release-id>/`,生产环境保留固定目录。ID 是打包更新 URL 的一部分,因此上传会拒绝不同的 ID。轮换后可通过内部分发提供新安装包,已安装客户端保留原清单地址。随机目录降低被猜中的概率,但不对持有链接者进行鉴权。[发布配置](../../../../apps/desktop/README.zh.md#upload-updates)定义生成与复用方式。
## 安全与发布策略
核心 dsh 和私有 Desktop Host 只来自签名应用的资源树。插件安装把包规格交给 pnpm,包括本地和远程来源,但不接受原始 pnpm 命令。pnpm 负责依赖解析和 profile 的 `allowBuilds` 策略;Host 加载已启用的 bundle。
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-08-31-released-session-format-migrations.md
2026-08-31-released-session-format-migrations.md: 273b51ac7ce9c507c2322499788dc2084b857415
2026-08-31-released-session-format-migrations.zh.md: 69132ec1763ad12b51480e150c3968aa1dcacf61
2026-08-31-released-session-format-migrations.md: e21793e0ea3d0c5376cbb3949c5d6ef599b9afb1
2026-08-31-released-session-format-migrations.zh.md: 637d8fcf0206661df9a6f1b3b470ad8a55746720
@@ -131,7 +131,7 @@ Existing write handles retain the process-local claim and kernel-backed cross-pr
### Parent catalog prerequisites
V3→V4 completes parent child discovery from retained direct-child headers and own descriptors. Storage supplies compact descriptor evidence to the catalog assembly, which binds it into the V3→V4 stage factory; the Stage waits for the parent's final inherited cut before validating catalog payloads and deriving available own discovery facts. Nested seed markers discard inherited catalog candidates without interpreting their payloads. An existing catalog entry survives an unavailable descriptor, including a child that failed before its first step. Descriptor v1 denotes continuable mode; v2/v3 explicitly record mode. Available unambiguous discovery fields must agree with the parent; absent entries require one supported descriptor. Missing, unsupported, or multiple descriptors preserve the logs without backfilling that child, as recorded in [incomplete child evidence](../bug-fix/2026-09-19-v3-incomplete-child-catalog-evidence.md). JSONL rechecks related membership and source revisions when preparing, reusing, and publishing. An unsupported generation or unreadable header elsewhere in that scan refuses migration because it cannot prove unrelatedness. Source drift invalidates preparation; read open retries once automatically, while write open refuses publication. Diagnostics retain the offending child path and distinguish unsupported migration evidence from malformed child data. The storage decoder classifies physical JSON and compression failures before the collector adds child context; filesystem reads and cancellation keep their original errors. The V0–V3-only catalog avoids recursive child migration. Current V4 reads skip that scan but apply the same own-catalog field, uniqueness, and current delivery ownership checks as strict restoration. See the [format specification](../../../../packages/session/session-format-v3-to-v4/README.md).
Runtime child-failure handling in this paragraph is partially superseded by [session-local catalog preparation](../bug-fix/2026-09-19-session-local-subagent-migration.md): opening the parent still completes its catalog from direct-child reads, but unreadable children do not block healthy parent history. V3→V4 completes parent child discovery from retained direct-child headers and own descriptors. Storage supplies compact descriptor evidence to the catalog assembly, which binds it into the V3→V4 stage factory; the Stage waits for the parent's final inherited cut before validating catalog payloads and deriving available own discovery facts. Nested seed markers discard inherited catalog candidates without interpreting their payloads. An existing catalog entry survives an unavailable descriptor, including a child that failed before its first step. Descriptor v1 denotes continuable mode; v2/v3 explicitly record mode. Available unambiguous discovery fields must agree with the parent; new complete entries require one supported descriptor. Missing, unsupported, or multiple descriptors preserve the logs and retain unknown-mode membership without inventing discovery fields, as recorded in [incomplete child evidence](../bug-fix/2026-09-19-v3-incomplete-child-catalog-evidence.md). JSONL rechecks related membership and source revisions when preparing, reusing, and publishing. Unreadable or unsupported headers are excluded from discovery; failures reading a recognized child preserve its header identity as unknown-mode membership. Source drift invalidates preparation; read open retries once automatically, while write open refuses publication. Diagnostics retain the offending child path and distinguish unsupported migration evidence from malformed child data. The collector reports child decoding and descriptor failures as warnings with the child path; cancellation and source-consistency checks still stop preparation. The V0–V3-only catalog avoids recursive child migration. Current V4 reads skip that scan but apply the same own-catalog field, uniqueness, and current delivery ownership checks as strict restoration. See the [format specification](../../../../packages/session/session-format-v3-to-v4/README.md).
Raw parent/child migration tests retain 24 historical creation-time conflicts as rejection evidence. Detached controls align only the child creation time and prove event-preserving restoration, including a published child without a descriptor. Headless/ACP and SDK snapshot adapters validate live parent/child clocks before normalization and preserve their equality during refresh; committed predecessor generations remain unchanged.
@@ -131,7 +131,7 @@ POSIX publication 使用 hard-link creation 加目录 sync;Windows 使用 no-o
### 父目录前置事实
V3→V4 使用保留的直属子 Session header 和自身 descriptor 补齐父 Session 的子代理发现记录。存储将紧凑的 descriptor 证据提供给 catalog 组装处,由其绑定到 V3→V4 stage 工厂;Stage 等待父 Session 的最终继承截点确定后,才校验目录 payload 并导出可获得的自身发现事实。嵌套的 seed 标记丢弃继承目录候选,不解释其 payload。已有目录记录允许 descriptor 不可用,包括在首次 step 之前失败的子 Session。Descriptor v1 表示 continuable 模式;v2/v3 显式记录模式。可获得且明确的发现字段必须与父目录一致;缺失记录要求一个受支持的 descriptor。Descriptor 缺失、版本不受支持或存在多个时,保留日志但不补填对应子 Session,理由见[不完整子目录证据](../bug-fix/2026-09-19-v3-incomplete-child-catalog-evidence.zh.md)。JSONL 在准备、复用和发布时重新检查关联成员与来源修订。扫描中其他任何不支持代际或不可读 header 都会拒绝迁移,因为无法证明它与父 Session 无关。来源变化使准备缓存失效;只读打开自动重试一次,写打开则拒绝发布。诊断保留出错子日志的路径,并区分不支持的迁移证据与畸形子数据。存储解码器先将物理 JSON 与压缩错误归类,再由收集器补充子日志信息;文件系统读取与取消保留原始错误。仅支持 V0–V3 的目录避免递归迁移子 Session。当前 V4 读取跳过该扫描,但会执行与严格恢复相同的自身目录字段、唯一性及当前投递归属检查。见[格式规范](../../../../packages/session/session-format-v3-to-v4/README.zh.md)。
本段的运行时子会话错误处理部分由[逐会话目录准备](../bug-fix/2026-09-19-session-local-subagent-migration.zh.md)取代:打开父会话仍通过读取直属子会话补齐目录,但不可读子会话不再阻断健康父历史。V3→V4 使用保留的直属子 Session header 和自身 descriptor 补齐父 Session 的子代理发现记录。存储将紧凑的 descriptor 证据提供给 catalog 组装处,由其绑定到 V3→V4 stage 工厂;Stage 等待父 Session 的最终继承截点确定后,才校验目录 payload 并导出可获得的自身发现事实。嵌套的 seed 标记丢弃继承目录候选,不解释其 payload。已有目录记录允许 descriptor 不可用,包括在首次 step 之前失败的子 Session。Descriptor v1 表示 continuable 模式;v2/v3 显式记录模式。可获得且明确的发现字段必须与父目录一致;新增完整记录要求一个受支持的 descriptor。Descriptor 缺失、版本不受支持或存在多个时,保留日志与未知模式成员关系,不编造发现字段,理由见[不完整子目录证据](../bug-fix/2026-09-19-v3-incomplete-child-catalog-evidence.zh.md)。JSONL 在准备、复用和发布时重新检查关联成员与来源修订。不可读或不支持的 header 不参与发现;读取已识别子会话失败时,其 header 身份保留为未知模式成员关系。来源变化使准备缓存失效;只读打开自动重试一次,写打开则拒绝发布。诊断保留出错子日志的路径,并区分不支持的迁移证据与畸形子数据。收集器将子解码和 descriptor 失败报告为带子路径的警告;取消与来源一致性检查仍中止准备。仅支持 V0–V3 的目录避免递归迁移子 Session。当前 V4 读取跳过该扫描,但会执行与严格恢复相同的自身目录字段、唯一性及当前投递归属检查。见[格式规范](../../../../packages/session/session-format-v3-to-v4/README.zh.md)。
原始父子迁移测试保留 24 个历史创建时间冲突作为拒绝证据。独立的内存对照只对齐子创建时间,证明恢复保留事件,包括已发布但没有 descriptor 的子 Session。Headless/ACP 与 SDK 快照适配器在规范化之前校验实时父子时钟,并在刷新时保持两者相等;已提交的前代文件保持不变。
@@ -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/2026-09-01-jobs-absorb-activity-record.md
2026-09-01-jobs-absorb-activity-record.md: 508608f7eb694bf131bad9a25e856d32c7079559
2026-09-01-jobs-absorb-activity-record.zh.md: 2c157d951042f1cd6f8a3a66129773bc99cd5b02
@@ -0,0 +1,48 @@
# Agent Note: The job registry absorbs the observation record; the standalone activity seam is removed
Status: implemented
Superseded: the record declaration (`JobStart.record`, `RecordingJob`, `readRecord`, `pumpJobOutput`) and the split wire (`SessionJob.record` on the session control stream, `ctx.jobOutput`) described below were consolidated into one output ring and direct operations on `ctx.jobs` — see [the jobs seam consolidation](2026-09-03-jobs-seam-consolidation.md). The argument for absorbing the standalone activity seam into `ctx.jobs` still holds.
English | [中文](2026-09-01-jobs-absorb-activity-record.zh.md)
## Problem
The [activity observation seam](../feature/2026-08-24-activity-observation-seam.md) shipped live output streaming as a second registry beside `ctx.jobs`: producers registered the same work twice (a job for lifecycle, an activity for observation), kept the two terminal states consistent by hand (`observeBackgroundActivity` waited for `proc.done` before mapping the outcome so a pump failure could not freeze a wrong terminal state), correlated the rows through `ActivityCorrelation.jobId`, and wrapped every observation call in registry-absent degradation branches. The Web client maintained two rosters (session-control `jobs` frames and the activity control stream) and joined them per row. Measured before the merge: eight pairing call sites, and one producer population — background bash/pwsh (both registries), PTY sends and subagent delegations (jobs only), foreground workflow (activity only).
The one activity without a job was the foreground workflow mirror. Everything else the split enabled — independent observers, absolute-offset reads, bounded retention — is a property of the record, not of the registry split.
## Decision
`ctx.jobs` owns the observation record; `packages/activity/`, `packages/api/activity-controller`, and the correlation vocabulary are removed.
- **`JobStart.record?: true`** declares an observable output record. `run(job)` now receives the job's producer face — `RecordingJob { id, append(text, {channel?, gapBefore?}), updateDetail(detail) }` for a `record: true` start, the plain `RunningJob { id, updateDetail(detail) }` otherwise — so the id is issued before the starter runs (a throwing starter still registers nothing; its ordinal is skipped). Writes staged inside the starter surface at the registration commit. `updateDetail` works for every job and gives `job_list` a live progress line; `append` without a record declaration logs and drops.
- **There is no `end`.** Job settlement — the producer outcome, a kill, or teardown — is the record's only close: it trims retention to the settled cap and fires the final `onOutput` signal. The dual-settlement pairing (`ActivityHandle.end` first-wins against teardown force-ends) is deleted, not reimplemented.
- **`readRecord(id, from, caller)`** is the non-consuming multi-reader view (absolute UTF-8 offsets, `lossy` below the retained window), fenced like every other job read; it never marks the job `reported`. The model-facing `readOutput` cursor is untouched — the two projections serve different readers and stay separate by design.
- **`jobs-local`** absorbs the chunk ring (`retainBytes` 256 KiB live, `settledRetainBytes` 16 KiB settled), and `pumpJobOutput` replaces `pumpActivityOutput` beside the seam. Producers fold the pump's final drain into `hooks.done` so the record holds its last bytes before settlement closes it.
- **The wire splits by role.** `SessionJob.record` (present exactly for record jobs) marks a row observable on the session control stream; `job.follow({sessionId?, jobId, from?})` on the `job` namespace of `api-job-controller` streams anchor/output/status frames with the fenced read resolved from the request's session. `api-activity-controller` is deleted; its roster stream is redundant (the jobs frames are the roster) and its observe machinery lives in `api-job-controller` (`observe.ts`, client `ctx.jobOutput`), which resolves its Remote faces in its own apply.
- **`ui-activity` returns to its upstream name `ui-jobs`** and renders one roster: `jobsBySession` rows, expandable exactly when `record` is present. The two-roster join is deleted.
- **Foreground workflow loses its live panel deliberately.** `tool-workflow`'s activity mirror is removed; a foreground run surfaces through its recorded run/member lifecycle events only, and the per-line `workflow/phase` / `workflow/log` narration has no observer until workflow gains `run_in_background` and registers a record job. Jobs stay a pure background registry — no `foreground` mode bit, no model-invisible rows, no run-less rows.
## Alternatives considered
- **A `foreground: true` job mode** to keep the foreground workflow panel: it needs two coupled enforcement points (reported-at-birth and a `job_list` filter) whose divergence double-delivers or leaks rows to the model, for one edge feature replaceable by `run_in_background`.
- **Serving model reads from the record** (deleting `readOutput`): model reads are producer-formatted (truncation and spill notices, sandbox markers) and consuming; the record is raw, channel-labeled, and non-consuming. Unifying them moves producer-specific formatting into the registry or noise into the observer stream.
- **Keeping the split with shared implementation** (a common registry library): it removes the duplicated skeleton but keeps the real costs — double registration, terminal-state pairing, correlation, a second wire roster, and a second package family.
The [prior seam note](../feature/2026-08-24-activity-observation-seam.md)'s argument against durable session events and control-stream framing for live output still holds and carries over unchanged: the record is process-local observation state, never a session event, so "model-visible ⟺ logged" is untouched.
## Review corrections
Review of the merged design settled five points that the sections above leave implicit:
- **The pump's wait holds constant resources.** `pumpJobOutput` subscribes to the producer's `done` once and keeps one pending timer; settlement clears the timer and wakes the current wait. Racing the same pending promise every poll round retained a reaction per round until the job ended — a day-long job at the default cadence accumulated over a million closures.
- **Client disposal waits for carrier quiescence.** The `ctx.jobOutput` effect disposer is async and awaits every open `RemoteStream.dispose()`, because Cordis awaits async disposers and a fiber that reported unloaded while an old iterator was still closing could overlap the next plugin instance under HMR.
- **The roster flags a record, it does not count bytes.** `SessionJob.record: true` replaced `outputTotal`: `append` commits no roster change (only lifecycle commits fire `onJobsChanged`), so a mirrored count went stale the moment output arrived, and broadcasting the roster per append was rejected as traffic for a number no row needed. The live offsets ride the observation stream's `opened` anchor.
- **The record is a best-effort live preview.** Producers copy stdout and stderr per poll round, so writes inside one window land stdout first; pipe capture cannot recover the true interleaving (only a PTY or `2>&1` can), and the client concatenates chunks regardless of `channel`. `channel` stays on the wire because it is what makes that limit visible and what a distinct stderr rendering would need.
- **The declaration types the face.** `JobStart` is a union discriminated by `record`: `RecordingJobStart.run` receives `RecordingJob`, `PlainJobStart.run` receives `RunningJob` without `append`, and the registry builds the plain face without the method — output the registry has nowhere to keep is unrepresentable instead of warned about and dropped.
- **Web e2e keep-alive is a barrier, not a clock.** The background command holds on a file the test owns in the job's cwd and the scenario kills it explicitly, so no CI stall can settle the job first; each scenario is one test, never a chain of tests sharing a job id.
## Consequences
Every observable row is a killable job; a pure-observation surface with no lifecycle would need a new home (harness diagnostics belong to the inspector plane, not here). A future remote job provider must implement lifecycle and record together. The record's retention config lives on `jobs-local`; the observation wire's cadence config (`observeFlushMs`, `observeMaxFrameBytes`) lives on `api-job-controller`.
@@ -0,0 +1,48 @@
# Agent Note:job 注册表吸收观测 record;独立的 activity seam 被移除
Status: implemented
Superseded:下文描述的 record 声明(`JobStart.record`、`RecordingJob`、`readRecord`、`pumpJobOutput`)与拆分的线路(会话控制流上的 `SessionJob.record`、`ctx.jobOutput`)已收敛为一个输出环与 `ctx.jobs` 上的直接操作——见 [jobs seam 收敛](2026-09-03-jobs-seam-consolidation.zh.md)。把独立 activity seam 吸收进 `ctx.jobs` 的论证仍然成立。
[English](2026-09-01-jobs-absorb-activity-record.md) | 中文
## 问题
[activity 观测 seam](../feature/2026-08-24-activity-observation-seam.zh.md) 把实时输出流做成了 `ctx.jobs` 旁边的第二个注册表:生产者要把同一份工作登记两次(job 管生命周期,activity 管观测),手工保持两个终态一致(`observeBackgroundActivity` 要等 `proc.done` 才映射 outcome,防止 pump 失败把错误终态冻上去),通过 `ActivityCorrelation.jobId` 关联两行,并给每个观测调用包上注册表缺失的降级分支。Web 客户端维护两份名册(session-control 的 `jobs` 帧与 activity control 流)并逐行做 join。合并前实测:八个配对调用点;生产者人口为——后台 bash/pwsh(双注册表)、PTY 发送与 subagent 委托(仅 jobs)、前台 workflow(仅 activity)。
唯一没有 job 的 activity 是前台 workflow 镜像。拆分带来的其他一切——独立观察者、绝对偏移读取、有界保留——都是 record 的性质,不是注册表拆分的性质。
## 决策
`ctx.jobs` 拥有观测 record;`packages/activity/`、`packages/api/activity-controller` 与 correlation 词表被移除。
- **`JobStart.record?: true`** 声明可观测输出 record。`run(job)` 现在收到 job 的生产者面——`record: true` 的 start 收到 `RecordingJob { id, append(text, {channel?, gapBefore?}), updateDetail(detail) }`,其余收到普通的 `RunningJob { id, updateDetail(detail) }`——因此 id 在 starter 运行前签发(starter 抛出仍然不登记任何东西;序号被跳过)。starter 内暂存的写入在登记提交时可见。`updateDetail` 对所有 job 可用,让 `job_list` 也能看到实时进度行;未声明 record 的 `append` 记日志并丢弃。
- **没有 `end`。** job 结算——生产者 outcome、kill 或 teardown——是 record 唯一的关闭:裁剪到结算保留量并发出最后一个 `onOutput` 信号。双结算配对(`ActivityHandle.end` 与 teardown 强结的 first-wins 对账)被删除,而非重新实现。
- **`readRecord(id, from, caller)`** 是非消耗多读者视图(绝对 UTF-8 偏移,低于保留窗口为 `lossy`),围栏与其他 job 读取一致;永不置 `reported`。模型面的 `readOutput` 游标原样保留——两个投影服务不同读者,刻意保持分离。
- **`jobs-local`** 吸收块环(`retainBytes` 运行期 256 KiB,`settledRetainBytes` 结算后 16 KiB),`pumpJobOutput` 在 seam 旁替换 `pumpActivityOutput`。生产者把 pump 的末次排空折进 `hooks.done`,让 record 在结算关闭前握有最后的字节。
- **wire 按角色拆分。** `SessionJob.record`(恰在 record job 上存在)在会话控制流上标记行可观测;`api-job-controller` 的 `job` 命名空间上的 `job.follow({sessionId?, jobId, from?})` 流式发送 anchor/output/status 帧,围栏读取者由请求的 session 解析。`api-activity-controller` 删除;其名册流是冗余的(jobs 帧就是名册),observe 机制位于 `api-job-controller`(`observe.ts`,客户端 `ctx.jobOutput`),它在自己的 apply 里解析所依赖的 Remote 面。
- **`ui-activity` 改回上游名字 `ui-jobs`**,渲染单一名册:`jobsBySession` 行,恰在 `record` 存在时可展开。双名册 join 被删除。
- **前台 workflow 有意失去实时面板。** `tool-workflow` 的 activity 镜像被移除;前台 run 只通过已记录的 run/member 生命周期事件呈现,逐行的 `workflow/phase` / `workflow/log` 叙述在 workflow 获得 `run_in_background` 并登记 record job 之前没有观察者。jobs 保持纯后台注册表——没有 `foreground` 模式位、没有模型不可见行、没有无 run 行。
## 曾考虑的替代方案
- **`foreground: true` job 模式**以保留前台 workflow 面板:它需要两个耦合的执行点(出生即 reported 与 `job_list` 过滤),两者一旦分叉就会双投递或向模型泄漏行,而它换来的只是一个可被 `run_in_background` 替代的边缘特性。
- **用 record 服务模型读取**(删除 `readOutput`):模型读取是生产者格式化的(截断与 spill 提示、sandbox 标记)且消耗型;record 是原始、带 channel 标签、非消耗的。统一二者要么把生产者特有格式化搬进注册表,要么把噪声混入观察者流。
- **保留拆分但共享实现**(公共注册表库):它消除重复骨架,但保留真正的成本——双重登记、终态配对、correlation、第二条 wire 名册与第二个包族。
[先前 seam note](../feature/2026-08-24-activity-observation-seam.zh.md) 反对用持久会话事件或 control 流承载实时输出的论证仍然成立并原样沿用:record 是进程本地观测状态,从不是会话事件,「模型可见 ⟺ 已记录」不受影响。
## 评审修正
对已合并设计的评审确定了上文各节未言明的五点:
- **泵的等待资源恒定。** `pumpJobOutput` 对生产者的 `done` 只订阅一次,同一时刻只保留一个待触发的 timer;结算清掉 timer 并唤醒当前等待。此前每个轮询轮次都对同一个 pending promise 赛跑,每轮留下一个 reaction 直到任务结束——默认节奏下跑一天的任务会累积上百万个闭包。
- **客户端卸载等待载体静止。** `ctx.jobOutput` 的 effect disposer 是异步的,await 每条打开的 `RemoteStream.dispose()`:Cordis 会 await 异步 disposer,而一个在旧 iterator 仍在关闭时就宣告卸载完成的 fiber,在 HMR 下可能与下一个插件实例重叠。
- **名册标记 record,不计字节。** `SessionJob.record: true` 取代了 `outputTotal`:`append` 不提交名册变更(只有生命周期提交触发 `onJobsChanged`),镜像的计数在输出到达的瞬间就陈旧了;而为一个没有任何行需要的数字按 append 广播整份名册被否决。实时偏移随观测流的 `opened` 锚帧到达。
- **record 是尽力而为的实时预览。** 生产者按轮询轮次复制 stdout 与 stderr,同一窗口内的写入先 stdout 落地;pipe 捕获无法还原真实交错(只有 PTY 或 `2>&1` 可以),客户端拼接 chunk 时也不区分 `channel`。`channel` 留在 wire 上,因为它正是让这一限制可见、也是将来区分渲染 stderr 所需要的东西。
- **声明决定生产者面的类型。** `JobStart` 是按 `record` 判别的联合类型:`RecordingJobStart.run` 收到 `RecordingJob`,`PlainJobStart.run` 收到没有 `append` 的 `RunningJob`,注册表给普通 job 构造的面根本没有这个方法——注册表无处存放的输出成为不可表达,而不是警告后丢弃。
- **Web e2e 的保活是 barrier 而非时钟。** 后台命令守在测试拥有的、位于 job cwd 的文件上,由场景显式 kill,因此 CI 卡顿不可能先把任务结算掉;每个场景是一个测试,绝不是共享 job id 的测试链。
## 后果
每个可观测行都是可 kill 的 job;没有生命周期的纯观测面需要新的归属(harness 诊断属于 inspector 平面,不在这里)。未来的远端 job provider 必须同时实现生命周期与 record。record 的保留配置在 `jobs-local`;观测 wire 的节奏配置(`observeFlushMs`、`observeMaxFrameBytes`)在 `api-job-controller`。
@@ -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/2026-09-03-jobs-seam-consolidation.md
2026-09-03-jobs-seam-consolidation.md: 2742ff1a80ec4cc4d8e5948f63d306abb9753733
2026-09-03-jobs-seam-consolidation.zh.md: 1fa61919a27362ce42419b57e0960e9be86ea07e
@@ -0,0 +1,35 @@
# Agent Note: One output ring, one projection, one event stream for background jobs
Status: implemented
English | [中文](2026-09-03-jobs-seam-consolidation.zh.md)
## Problem
The job seam had grown by accretion. Producers declared their output twice — a consuming `readOutput` hook the model read and an optional `record: true` ring observers read — and the registry kept two vocabularies for one job (`JobSnapshot` for the model, `SessionJob` on the session control stream, `JobWireChunk` on the observation wire). Three listener families (`onJobDone`, `onJobsChanged`, `onOutput`) delivered overlapping facts with different owner filters, and the registry itself tracked `reported`, a bit that only the model-facing tool could interpret. Review of the merged design called the interface layer the weakest part: a reader could not tell which of `read`, `readRecord`, `JobStart`, `RecordingJob`, and `updateDetail` was the contract and which was an accident of history.
## Decision
- **One ring per job.** `JobSpec` replaces `JobStart`: a spec names optional pull `output` sources (`JobOutputSource`, the subprocess `readFrom` family) that the registry pumps at its own cadence (`pumpPollMs` on `dsh-jobs-local`), and every producer receives a `JobHandle` whose `append` pushes chunks into the same ring. The model's consuming read and every observer's absolute-offset read serve the same bytes; `readOutput` and the `record` discriminator are gone. Shell producers hand the registry their `observed` offset readers and fold sandbox facts into the terminal `detail`; the terminal producer adapts its consuming send reader through a byte-counting source; the subagent returns its answer as `JobOutcome.result`, which the first read after settlement carries once.
- **Two cursors, one storage.** `JobRegistry.read(id)` is the stateful model cursor; `JobRegistry.readAt(id, from)` is the stateless observer read. Model reads exclude `log` chunks (producer narration for observers) and render stdout, then one `[stderr]` section, exactly as the shell tools always did.
- **`updateProgress` replaces `updateDetail`.** `JobView.progress` is the live line and is cleared at settlement; `JobView.detail` is the terminal reason, with the model's `job_kill` reason merged in.
- **One projection.** `JobView` (client-safe leaf `@deepseek-ai/dsh-jobs/view`) is what the model tools, the browser roster, and the observation frames consume; `owner` is the session id, never the `Agent`.
- **Direct caller-scoped operations.** `JobRegistry.list`, `get`, `read`, `readAt`, `kill`, and `wait` accept the caller session on each call. Omitting the caller permits only unowned jobs. The output ring and browser streams do not require a separate caller-bound operation object.
- **One event stream.** `events.subscribe(filter, listener)` with `{ owner }`, `{ owners: 'scope' }`, or `{ owners: 'all' }` replaces the three listener families. `settled` names its cause (`producer`, `kill`, `teardown`); `output` carries only the id and the new total.
- **The notice ledger moves to `dsh-tool-jobs`.** The registry no longer tracks `reported`. The tool claims a job when a wait starts (and withdraws on timeout or abort) and when `job_kill` is accepted; the settlement event drops the entry, teardown settlements and unowned jobs are skipped, and the destination is the agent registered for the owner session at settlement.
- **The roster leaves the session control stream.** `dsh-api-job-controller` owns `job.list({ sessionId })` — whole-set frames after each coalesced burst of lifecycle commits, never per append — beside `job.follow`, and its client installs `ctx.jobs` (`watchRows`, `observe`, one snapshot). `dsh-api-session-controller` carries queues and projections only; `ui-jobs` watches its session's roster while mounted, and a row is expandable while live or once it left retained output behind.
`attachController` and the owner-relative `servesOwner` gate stay as they were; whether a producer may offer `run_in_background` without a mounted controller is deferred, since the gap predates this change.
## Alternatives considered
- **Keep three listeners with a shared owner filter.** It preserves the accreted names but not the property reviewers asked for — one place to learn what the registry announces — and `onJobDone`'s exact-owner delivery is what forced the registry to keep an `Agent` in every projection.
- **Keep `reported` in the registry.** Only the tool knows which deliveries reach the model (a wait's tool result, a kill's acknowledgement); the registry could only approximate it, and the approximation was the source of the double-notice and missed-notice bugs the review listed.
- **Keep the roster on the session control stream.** It saved one Remote stream but coupled the session controller to the job registry and mirrored a `record` flag the ring's byte count now answers directly (`output.total`).
- **Serve `log` chunks to the model too.** Producer narration is for observers; the model already receives the same facts through `progress` and the terminal `detail`.
## Consequences
The model-visible text is unchanged: status lines, tool descriptions and schemas, and notice wording are byte-identical, and the only new model-visible fact is the kill reason a `killed` status line now carries. A producer that needs to stream output picks pull sources or `append`; it never formats a consuming delta. The Web client renders every live job as expandable, so a job that never writes shows an empty running panel instead of a static row. `JobKindMap` merges are host-side, so a browser program types `JobView.kind` as the kinds it knows and treats the rest as opaque strings, which is all the roster does with it.
Review follow-ups after the master merge tightened four of these edges. A job's first event is always `registered`: the announcement precedes the pump, whose first drain runs synchronously and may announce output. A settled ring keeps every byte the model cursor has not consumed until the first terminal model read, so a job that finishes before its first `job_output` loses nothing the live cap retained. `dsh-tool-jobs` holds one claim per wait call with a `live` flag that the caller's abort clears synchronously from the abort event, so a settlement announced before the registry's rejection reaches the tool still delivers its notice, and a timed-out wait drops only its own claim while a concurrent wait keeps the settlement covered. The browser records an observation failure that precedes the anchor as an errored view, so the panel shows the notice instead of staying blank. The `@deepseek-ai/dsh-jobs/invariant` companion checks the announced event protocol per job (registered first, one settlement, removal last) and each announced projection against the registry's own read, which is the cross-observation the [companion rule](../simplification/2026-08-28-omit-unneeded-invariant-companions.md) requires; the earlier single-view field checks are gone. The spill file a pull source keeps is job metadata (`JobView.output.spillPaths`, reported by every pump read and withdrawn once a read stops naming it), not per-chunk metadata, and the model's read renders the shell tools' dropped-output notice for a producer-side gap as well as for ring eviction, naming those files even after the gap chunk itself was evicted — the `before this read` variant this change had introduced is gone. An observation generation that sees its job removed mid-stream (the owner's teardown) closes with the removed job's terminal projection instead of a failed read. A rejected background spawn produces no process output, so the shell providers now serve the `spawn failed: …` note as the whole observed stderr stream (the consuming read still folds it in once), and `ShellProcess.observed` with both readers is required: the model's first `job_output` after such a failure reads exactly what master's consuming read rendered.
@@ -0,0 +1,35 @@
# Agent Note:后台 job 的一个输出环、一个投影、一条事件流
Status: implemented
[English](2026-09-03-jobs-seam-consolidation.md) | 中文
## 问题
job seam 是逐步堆出来的。生产者要把输出声明两次——模型读取的消耗式 `readOutput` 钩子,以及观察者读取的可选 `record: true` 环——注册表也为同一个 job 保留了两套词汇(模型侧的 `JobSnapshot`、会话控制流上的 `SessionJob`、观测线路上的 `JobWireChunk`)。三个监听器家族(`onJobDone`、`onJobsChanged`、`onOutput`)用不同的拥有者过滤投递彼此重叠的事实,注册表自己还跟踪着只有模型侧工具才解释得了的 `reported` 位。合并后设计的评审把接口层称作最薄弱的部分:读者分不清 `read`、`readRecord`、`JobStart`、`RecordingJob` 与 `updateDetail` 里哪些是约定、哪些是历史遗留。
## 决策
- **每个 job 一个环。** `JobSpec` 取代 `JobStart`:spec 声明可选的拉取式 `output` 源(`JobOutputSource`,子进程的 `readFrom` 家族),由注册表按自己的节奏(`dsh-jobs-local` 的 `pumpPollMs`)泵送;每个生产者都收到一个 `JobHandle`,其 `append` 把块推进同一个环。模型的消耗式读取与每个观察者的绝对偏移读取服务同一批字节;`readOutput` 与 `record` 判别字段被移除。shell 生产者把它们的 `observed` 偏移读取器交给注册表,并把沙箱事实折进终态 `detail`;terminal 生产者通过一个计字节的源适配其消耗式发送读取器;subagent 把回答作为 `JobOutcome.result` 返回,结算后的第一次读取携带它一次。
- **两个游标,一份存储。** `JobRegistry.read(id)` 是有状态的模型游标;`JobRegistry.readAt(id, from)` 是无状态的观察者读取。模型读取排除 `log` 块(给观察者的生产者叙述),并渲染 stdout,再接一段 `[stderr]`,与 shell 工具一贯的做法完全一致。
- **`updateProgress` 取代 `updateDetail`。** `JobView.progress` 是实时行,结算时清除;`JobView.detail` 是终态原因,合并了模型 `job_kill` 给出的原因。
- **一个投影。** `JobView`(客户端安全叶子 `@deepseek-ai/dsh-jobs/view`)是模型工具、浏览器名册与观测帧共同消费的形状;`owner` 是会话 id,从不是 `Agent`。
- **直接传入调用方的操作。** `JobRegistry.list`、`get`、`read`、`readAt`、`kill` 与 `wait` 在每次调用时接收调用方会话。省略调用方时仅允许访问无主 job。输出环与浏览器流不需要独立的调用方操作对象。
- **一条事件流。** 带 `{ owner }`、`{ owners: 'scope' }` 或 `{ owners: 'all' }` 过滤的 `events.subscribe(filter, listener)` 取代三个监听器家族。`settled` 标出原因(`producer`、`kill`、`teardown`);`output` 只携带 id 与新的 total。
- **播报台账移到 `dsh-tool-jobs`。** 注册表不再跟踪 `reported`。工具在等待开始时认领 job(超时或中止则撤回),在 `job_kill` 被接受时也认领;结算事件删除条目,teardown 结算与无主 job 被跳过,投递目标是结算时登记在拥有者会话下的 agent。
- **名册离开会话控制流。** `dsh-api-job-controller` 在 `job.follow` 旁拥有 `job.list({ sessionId })`——每轮聚合过的生命周期提交之后发一次整集帧,从不按追加发——其客户端安装 `ctx.jobs`(`watchRows`、`observe`、一份快照)。`dsh-api-session-controller` 只承载队列与投影;`ui-jobs` 在挂载期间关注其会话的名册,行在 job 进行中或留有保留输出时可展开。
`attachController` 与相对拥有者的 `servesOwner` 门保持原样;没有挂载控制器的生产者是否可以提供 `run_in_background`,因为这个缺口早于本次变更,暂不处理。
## 考虑过的替代方案
- **保留三个监听器并共享一个拥有者过滤。** 它保住了堆出来的名字,却保不住评审要的性质——一个地方就能学会注册表宣布什么——而且 `onJobDone` 的精确拥有者投递正是迫使注册表在每个投影里保留 `Agent` 的原因。
- **把 `reported`留在注册表。** 只有工具知道哪些投递到达了模型(等待的工具结果、kill 的确认);注册表只能近似,而这个近似正是评审列出的重复播报与漏播报 bug 的来源。
- **把名册留在会话控制流上。** 省了一条 Remote 流,却把会话控制器耦合到 job 注册表,还镜像了一个环的字节数如今直接回答的 `record` 标志(`output.total`)。
- **也把 `log` 块交给模型。** 生产者叙述是给观察者的;模型已经通过 `progress` 与终态 `detail` 收到同样的事实。
## 后果
模型可见文本保持不变:状态行、工具描述与 schema、播报措辞逐字节相同,唯一新增的模型可见事实是 `killed` 状态行现在携带的 kill 原因。需要流式输出的生产者选择拉取源或 `append`,从不格式化消耗式增量。Web 客户端把每个进行中的 job 渲染为可展开,因此从不写输出的 job 显示为空的运行中面板而非静态行。`JobKindMap` 的合并发生在宿主侧,浏览器程序把 `JobView.kind` 视为它认识的 kind 集合,其余当作不透明字符串——名册对它也只做这些。
合并 master 之后的评审跟进收紧了其中四条边。job 的第一个事件永远是 `registered`:通告先于泵,而泵的首轮 drain 同步执行且可能通告输出。已结算的环在模型的首次终态读取之前保留模型游标尚未消费的全部字节,因此在首次 `job_output` 之前结束的 job 不会丢失运行期上限保留的任何内容。`dsh-tool-jobs` 为每次等待调用保存一条带 `live` 标记的认领,调用方 abort 时在 abort 事件里同步清掉该标记,因此注册表的拒绝尚未传回工具之前宣布的结算仍会投递通知;超时的等待只丢弃自己的认领,并发的另一个等待仍覆盖该结算。浏览器把先于锚点发生的观察失败记录为带错误的视图,面板显示提示而不是留白。`@deepseek-ai/dsh-jobs/invariant` 伴生插件校验每个 job 的事件协议(先 `registered`、恰一次结算、最后 `removed`)以及每个通告的投影与注册表自身读取的一致性,这正是[伴生插件规则](../simplification/2026-08-28-omit-unneeded-invariant-companions.zh.md)要求的交叉观察;先前只针对单个视图的字段校验已删除。拉取源保留的 spill 文件是 job 元数据(`JobView.output.spillPaths`,每次泵读取都会上报,读取不再点名时即撤回),不是逐块元数据;模型读取对生产者侧缺口与环淘汰都渲染 shell 工具的丢失输出提示并列出这些文件,即使缺口块本身已被淘汰——本次变更曾引入的 `before this read` 变体已删除。观察生成过程若在流中看到其 job 被移除(拥有者 teardown),以被移除 job 的终态投影收尾,而不是一次失败的读取。被拒绝的后台 spawn 不产生任何进程输出,因此 shell provider 现在把 `spawn failed: …` 提示作为 observed stderr 的整条流提供(消耗式读取仍只折入一次),且 `ShellProcess.observed` 及其两个读取器改为必填:这类失败之后模型的首次 `job_output` 读到的正是 master 的消耗式读取所渲染的内容。
@@ -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/2026-09-05-workspace-files-service.md
2026-09-05-workspace-files-service.md: 40157508a83abc61245845a1dbe4f1749e175563
2026-09-05-workspace-files-service.zh.md: 51de0ad5b644f27e30557ea6b11ba62b8ef9a715
2026-09-05-workspace-files-service.md: 65bebf18175d9a94452c92cd36ada3c5ef267ce5
2026-09-05-workspace-files-service.zh.md: b4ca420042dc702f9bc1eaf5bb09f2667a19322d
@@ -25,15 +25,13 @@ Two constraints frame the service. File reads through `ctx.fs` use the Session's
`api/remotes` and both root aggregates reference the matching Host/Client leaf. The package exports `.`, `./client`, `./types`, `./typert`, and `./remote`, with one `workspace-files` web-app row supplying both faces. The Client plugin injects `['resources', 'remote', 'remote.workspaceFiles']`; the resource model takes result types directly from the protocol package, and the text preview owns the Sidebar parameter declaration, so the Client compilation graph has no reverse dependency on Remote assembly or Sidebar UI.
### The `workspaceFiles` Remote namespace
### The `workspaceFiles` file operations
Every Host method takes `WorkspaceFileScope` first. The Gateway resolves it from the wire Session identity by reading the live Session header or, for a cold Session, `SessionPersistence.stat`; it never activates an Agent, reads the event body, or falls back to a parent Session. The scope carries the selected Session id and its `cwd`, with the sandbox policy's deployment root used only when that header has no `cwd`. A Client passes its Session id and never names a root. The seven signatures, as `src/index.ts` declares them:
Every Host method takes `WorkspaceFileScope` first. The `workspaceFileScope` Gateway lookup resolves Session identity. It reads the live Session header or, for a cold Session, `SessionPersistence.stat`; it never activates an Agent, reads the event body, or falls back to a parent Session. The scope carries the selected Session id and its `cwd`, with the sandbox policy's deployment root used only when that header has no `cwd`. A Client passes its Session id and never names a root. The five signatures, as `src/index.ts` declares them:
```ts ignore-check
@Remote async read(workspaceFileScope: WorkspaceFileScope, path: string, range: WorkspaceFileRange, signal: AbortSignal): Promise<WorkspaceFileText>
@Remote async readBytes(workspaceFileScope: WorkspaceFileScope, path: string, range: WorkspaceByteRange, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readAll(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readRelated(workspaceFileScope: WorkspaceFileScope, path: string, relativePath: string, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readBytes(workspaceFileScope: WorkspaceFileScope, path: string, options: WorkspaceByteReadOptions, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async stat(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceFileStat>
@Remote async list(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceDirectoryListing>
@Remote({ mode: 'stream' }) changes(workspaceFileScope: WorkspaceFileScope, signal: AbortSignal): AsyncIterable<WorkspaceFileWatchFrame>
@@ -41,20 +39,20 @@ Every Host method takes `WorkspaceFileScope` first. The Gateway resolves it from
- **`stat`** returns `WorkspaceFileStat { absolutePath, version, bytes? }`: the file's identity, its opaque freshness token, and its size when the backend reports one. It accepts a regular file only.
- **`read`** returns one window of lines, `WorkspaceFileText = WorkspaceFileStat & { offset, text, lines, eof }`; `lines` counts the page's lines, so a page holding one empty line (`text: ''`, `lines: 1`) and a page past the end (`lines: 0`) read differently. `range.offset` is the 1-based first line and defaults to 1; `range.limit` is the largest number of lines and defaults to `maxLines`, which it may not exceed. Lines end at `\n` and a final `\n` terminates the last line rather than opening an empty one; `text` joins the page's lines with `\n` and carries no terminator; `eof` is true when the page includes the last line, and an offset past the end returns an empty page with `eof` true. The pager walks `streamText`, counts the lines before the window without keeping them, admits each in-window segment against `maxBytes` before buffering it, and returns at the first character past the window, so a file of any size costs one page of memory. The `version` and `bytes` on a page are the stat's, taken before the stream.
- **`readBytes`** returns one window of raw bytes, `WorkspaceFileBytes = WorkspaceFileStat & { offset, data, eof }`. `range.offset` is the 0-based first byte and defaults to 0; `range.length` is the largest byte count and defaults to `maxBytes`, which it may not exceed. `data` is base64, shorter than `length` where the file ends and empty at or past it; `eof` is true when the window includes the last byte. Nothing is decoded and nothing is refused as binary. `read` pages by lines and never by bytes; a byte window is `readBytes`.
- **`readAll` and `readRelated`** return complete `WorkspaceFileBytes` under `maxFileBytes`. `readRelated` resolves a relative filesystem path from the base file's directory; the Host applies the same regular-file checks and backend read authority to both files. [Document Preview](2026-09-08-document-preview-operations.md) owns their loading and address semantics.
- **`readBytes`** returns native bytes with file metadata, offset, and EOF. The required options object selects a complete read under `maxFileBytes` when `range` is absent, or a bounded window under `maxBytes` when present. `range: {}` uses offset zero and the configured window cap; windows can read files above the complete-file cap. The [binary transfer decision](2026-09-17-workspace-file-binary-transfer.md) owns the Remote response encoding.
- **`baseFile`** selects target resolution independently of `range`: the target `path` must be relative and resolves from the base file's directory. The Host applies regular-file checks to both files and preserves backend read authority. [Document Preview](2026-09-08-document-preview-operations.md) owns content loading and address semantics.
- **`list`** returns `WorkspaceDirectoryListing { path, entries, truncated }`: the listed directory as a workspace path relative to the root (empty for the root), its direct children in the backend's stable name order as `{ name, type, size? }`, and whether `maxEntries` cut the list. `type` is `file`, `directory`, or `other`; a symlink child reports the type of what it points to and a dangling one is `other`, while opening such a child still fails the link gate below. Dotfiles are listed; nothing is filtered.
- **`changes`** yields `WorkspaceFileWatchFrame`: `{ kind: 'ready' }` after the observation queue is registered and the workspace root resolves, followed by `{ kind: 'change', change }`. The `WorkspaceFileChange` payload is `{ absolutePath, version }` for a present file or `{ absolutePath, absent: true }` for one observed gone. Its source is `fs/observed` inside the workspace root, never an OS watcher. Observations after the first pull are queued, including during root resolution; cancellation or plugin disposal ends the generation.
### Paths on the wire
Two path vocabularies leave the service, and each method uses exactly one. `read`, `readBytes`, `readAll`, `readRelated`, `stat`, and `changes` name a file by `absolutePath`: its absolute path in the filesystem's execution world, symlinks resolved (`ctx.fs.processPath(target)`), so the Client provider matches a change frame to an open address by absolute path: the Client sends the address's path unchanged to the Host and binds the follower only to a successful `stat.absolutePath`, without reading a Session summary's cwd. `list` speaks workspace paths — the same syntax its `path` argument accepts, absolute or relative to the root — because its consumer is a tree rooted there. The field is called `absolutePath` and not `url` because it is not a resource address; the address grammar belongs to `dsh-util-workspace-path` and is described with the resource model. Input paths to `read`, `readBytes`, `readAll`, `readRelated`, `stat`, and `list` are absolute or relative to the session's workspace root, never to the backend's own cwd.
Two path vocabularies leave the service, and each method uses exactly one. `read`, `readBytes`, `stat`, and `changes` name a file by `absolutePath`: its absolute path in the filesystem's execution world, symlinks resolved (`ctx.fs.processPath(target)`), so the Client provider matches a change frame to an open address by absolute path: the Client sends the address's path unchanged to the Host and binds the follower only to a successful `stat.absolutePath`, without reading a Session summary's cwd. `list` speaks workspace paths — the same syntax its `path` argument accepts, absolute or relative to the root — because its consumer is a tree rooted there. The field is called `absolutePath` and not `url` because it is not a resource address; the address grammar belongs to `dsh-util-workspace-path` and is described with the resource model. Input paths to `read`, `readBytes`, `stat`, and `list` are absolute or relative to the session's workspace root, never to the backend's own cwd; `readBytes` with `baseFile` resolves the target from that file's directory.
`version` is an opaque string a consumer compares for equality and never parses: the local backend derives it from device, inode, size, and nanosecond mtime and ctime, so a rewrite that leaves the content identical still changes it. `offset` means a line on `read` and a byte on `readBytes`; the two units never mix, and `eof` on either means the window reached the file's end.
### File checks and workspace containment
`read`, `readBytes`, `readAll`, `readRelated`, and `stat` share regular-file checks and then rely on the filesystem backend's read authority. `list` shares path inspection but also checks workspace containment, while `changes` filters observations to the workspace root. The service applies the following checks:
`read`, `readBytes`, and `stat` share regular-file checks and then rely on the filesystem backend's read authority. `list` shares path inspection but also checks workspace containment, while `changes` filters observations to the workspace root. The service applies the following checks:
1. **The path itself.** `lstat` inspects the path before anything follows it: a missing path is `not-found`, and a symlink — wherever it points, including back inside the workspace — is `not-regular-file` (kind `symlink`) for the file methods and `not-directory` for `list`. An empty path is a `gateway/bad-request`.
2. **Workspace containment for `list`.** The directory resolves to a target and `ctx.fs.contains(root, target)` decides, where `root` is the `WorkspaceFileScope.workspaceRoot` resolved from the selected Session header. A `..` traversal or an absolute directory outside the root is `outside-workspace`. `changes` applies the same backend containment predicate to observed targets.
@@ -73,7 +71,7 @@ Each failure is one `RemoteError` code with typed details, declared beside the t
| `workspace-file/outside-workspace` | a `list` target is not inside the workspace root | `{ path }` |
| `workspace-file/too-large` | a page's text or byte window exceeds `maxBytes`, or a complete read exceeds `maxFileBytes` | `{ path, limit }` |
| `workspace-file/not-text` | invalid UTF-8 up to the page's end, or a NUL byte in the sample or the page (`read` only) | `{ path }` |
| `workspace-file/not-regular-file` | `read`, `readBytes`, `readAll`, `readRelated`, or `stat` on something that is not a regular file | `{ path, kind: 'directory' \| 'symlink' \| 'other' }` |
| `workspace-file/not-regular-file` | `read`, `readBytes`, or `stat` on something that is not a regular file | `{ path, kind: 'directory' \| 'symlink' \| 'other' }` |
| `workspace-file/not-directory` | `list` on something that is not a directory | `{ path, kind: 'file' \| 'symlink' \| 'other' }` |
| `workspace-file/unsupported-address` | Client-minted: a resource address this provider cannot serve | `{ address }` |
| `workspace-file/unknown-workspace` | Client-minted: an `absolute` address, which carries no Session | `{ address }` |
@@ -109,7 +107,7 @@ The Client export registers one `ResourceProvider<'file'>` into `ctx.resources`
### Related notes
The [resource model](2026-09-05-client-resource-model.md) owns `ctx.resources`, `useResource`, the `dsh-resource://<type>/…` address grammar, and the reasoning for one resource per address; the [text preview and file tree](../feature/2026-09-05-sidebar-text-preview-and-file-tree.md) are the shipped consumers of `read`, `list`, and the `file` provider; the [right Sidebar docking infrastructure](../feature/2026-09-04-right-sidebar-docking-infrastructure.md) is the surface they open into; [workspace file links](../feature/2026-07-31-web-workspace-file-links.md) is where serving files over HTTP was rejected. Anyone extending this system reaches the same seven methods through `remote.workspaceFiles` and the same `file` resource through `useResource<'file'>`; the wire types are published as `@deepseek-ai/dsh-api-workspace-files/types`. The [workspace file read authority](2026-09-09-workspace-file-read-authority.md) owns Host read access and the HTML security trade-off; [Document Preview](2026-09-08-document-preview-operations.md) owns content loading and per-tab freshness.
The [resource model](2026-09-05-client-resource-model.md) owns `ctx.resources`, `useResource`, the `dsh-resource://<type>/…` address grammar, and the reasoning for one resource per address; the [text preview and file tree](../feature/2026-09-05-sidebar-text-preview-and-file-tree.md) are the shipped consumers of `read`, `list`, and the `file` provider; the [right Sidebar docking infrastructure](../feature/2026-09-04-right-sidebar-docking-infrastructure.md) is the surface they open into; [workspace file links](../feature/2026-07-31-web-workspace-file-links.md) is where serving files over HTTP was rejected. Anyone extending this system reaches the same five methods through `remote.workspaceFiles` and the same `file` resource through `useResource<'file'>`; the wire types are published as `@deepseek-ai/dsh-api-workspace-files/types`. The [workspace file read authority](2026-09-09-workspace-file-read-authority.md) owns Host read access and the HTML security trade-off; [Document Preview](2026-09-08-document-preview-operations.md) owns content loading and per-tab freshness.
## Alternatives considered
@@ -117,7 +115,7 @@ The [resource model](2026-09-05-client-resource-model.md) owns `ctx.resources`,
**A dual-face package with reverse UI dependencies.** The split-package choice followed two project-reference cycles after `api/remotes` referenced the Client leaf: the resource model imported Remote assembly for result types, and the file provider imported Sidebar UI for its parameter map. TypeScript rejected these cycles with `TS6202`. [dual-face packaging](2026-09-07-workspace-files-dual-face-package.md) supersedes that split: result types come directly from the protocol package, and Sidebar parameter registration belongs to the text preview; both root aggregates retain explicit compiler entries.
**Serving workspace files over HTTP.** Already rejected by [workspace file links](../feature/2026-07-31-web-workspace-file-links.md) on origin grounds and not revisited: `read` and `readBytes` carry plain text and base64 over the authenticated Remote carrier, so no document is served, no URL is minted, and no origin question arises.
**Serving navigable workspace documents.** Rejected on origin grounds; the [binary transfer](2026-09-17-workspace-file-binary-transfer.md) returns multipart data while HTML remains isolated in a Blob iframe.
**Log-reachable or workspace-contained authorization for file reads.** The one precedent that sends file content over the wire, command attachments, authorizes only files that appear in the session log. That excludes typed paths, while workspace containment excludes readable files elsewhere on the Session backend. The [workspace file read authority](2026-09-09-workspace-file-read-authority.md) instead makes the backend's read decision authoritative and keeps containment only for workspace-shaped operations.
@@ -138,14 +136,14 @@ The [resource model](2026-09-05-client-resource-model.md) owns `ctx.resources`,
- Change frames report the agent's own operations only. A file edited by the user's editor, a shell, or a subprocess raises no frame; an agent merely reading a file that something else changed does raise one, because the read observes a new version.
- File-kind inspection precedes backend reads, and `list` reports kind before an outside position. A page's `version` may be one write behind its content, and a stalled `changes` consumer grows Host memory because a generation's queue is unbounded; each is a known trade-off recorded in the package README.
- The `file` resource pushes change, not content, so a preview learns a file moved on without a payload and reads the pages it wants; a failed open keeps following the address, so the agent creating the file brings the tab live without user action.
- `readBytes` has no shipped consumer yet; Document Preview uses `readAll` for complete-file formats.
- Office authorization probes use `readBytes` with a one-byte `range`; Document Preview uses the same method without `range` for complete-file formats.
## Testing
Host specs in `packages/api/workspace-files/tests` exercise header-only scope resolution for live and cold subagent Sessions, the deployment fallback, missing identities, and lookup disposal; the paged read (whole file, nested path, empty file, multi-byte UTF-8, the line window's edges, defaults and refused limits, carriage returns kept); the byte window (defaults, a middle window with more following, tail windows exact and short, past-end and empty files, NUL and invalid UTF-8 round-tripping through base64, version parity with `stat`, the cap as `too-large`, bad ranges, a window of a file far above the cap, and `eof` inferred without a size); `stat`; outside-workspace reads and backend refusals; `list` with containment, truncation, symlink children, and `not-directory`; and the `changes` stream driven by `fs/observed` and filtered by root. Client specs cover the provider's frames, the change feed, unsupported addresses, and registration and disposal. `fs/fs` and `fs-local` specs pin `readByteRange`; `dsh-util-workspace-path` specs pin the file-address grammar. The connection fixture serves `stat`, paged `read`, `list`, and an opt-in `changes` frame for the web e2e suite.
Host specs in `packages/api/workspace-files/tests` exercise header-only scope resolution for live and cold subagent Sessions, the deployment fallback, missing identities, and lookup disposal; the paged read (whole file, nested path, empty file, multi-byte UTF-8, the line window's edges, defaults and refused limits, carriage returns kept); the byte window (defaults, a middle window with more following, tail windows exact and short, past-end and empty files, NUL and invalid UTF-8 round-tripping through `Uint8Array`, version parity with `stat`, the cap as `too-large`, bad ranges, a window of a file far above the cap, and `eof` inferred without a size); `stat`; outside-workspace reads and backend refusals; `list` with containment, truncation, symlink children, and `not-directory`; and the `changes` stream driven by `fs/observed` and filtered by root. Client specs cover the provider's frames, the change feed, unsupported addresses, and registration and disposal. `fs/fs` and `fs-local` specs pin `readByteRange`; `dsh-util-workspace-path` specs pin the file-address grammar. The connection fixture serves `stat`, paged `read`, `list`, and an opt-in `changes` frame for the web e2e suite.
## Deferred
- A bound on a `changes` generation's queue.
- The shipped consumer of `readBytes` (image and binary previews) and any write, search, or media route; the service is read-only.
- Write and search operations; the service is read-only.
- Scopes other than `session` in the file address; the grammar leaves room, the provider serves one.
@@ -25,15 +25,13 @@ Web 客户端需要从一个未必在 Host 机器上的浏览器查看会话工
`api/remotes` 和两个根聚合分别引用匹配的 Host/Client 叶子。包导出 `.`、`./client`、`./types`、`./typert` 和 `./remote`,web-app 中单个 `workspace-files` 条目供应两面。Client 插件注入 `['resources', 'remote', 'remote.workspaceFiles']`;资源模型直接从协议包取结果类型,Sidebar 参数声明归文本预览,因此 Client 编译图不再反向依赖 Remote 装配或右栏 UI。
### `workspaceFiles` Remote 命名空间
### `workspaceFiles` 文件操作
每个 Host 方法首参都是 `WorkspaceFileScope`。Gateway 从线路上的 Session 身份解析它:优先读取 live Session header,cold Session 则只调用 `SessionPersistence.stat`;不会激活 Agent、读取事件正文或回退到父 Session。scope 携带所选 Session id 及其 `cwd`,仅当该 header 没有 `cwd` 时才使用沙箱策略的部署根。Client 传入 Session id,从不自行命名根。七个签名照 `src/index.ts` 的声明:
每个 Host 方法首参都是 `WorkspaceFileScope`。`workspaceFileScope` Gateway lookup 解析 Session 身份。它优先读取 live Session header,cold Session 则只调用 `SessionPersistence.stat`;不会激活 Agent、读取事件正文或回退到父 Session。scope 携带所选 Session id 及其 `cwd`,仅当该 header 没有 `cwd` 时才使用沙箱策略的部署根。Client 传入 Session id,从不自行命名根。五个签名照 `src/index.ts` 的声明:
```ts ignore-check
@Remote async read(workspaceFileScope: WorkspaceFileScope, path: string, range: WorkspaceFileRange, signal: AbortSignal): Promise<WorkspaceFileText>
@Remote async readBytes(workspaceFileScope: WorkspaceFileScope, path: string, range: WorkspaceByteRange, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readAll(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readRelated(workspaceFileScope: WorkspaceFileScope, path: string, relativePath: string, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async readBytes(workspaceFileScope: WorkspaceFileScope, path: string, options: WorkspaceByteReadOptions, signal: AbortSignal): Promise<WorkspaceFileBytes>
@Remote async stat(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceFileStat>
@Remote async list(workspaceFileScope: WorkspaceFileScope, path: string, signal: AbortSignal): Promise<WorkspaceDirectoryListing>
@Remote({ mode: 'stream' }) changes(workspaceFileScope: WorkspaceFileScope, signal: AbortSignal): AsyncIterable<WorkspaceFileWatchFrame>
@@ -41,20 +39,20 @@ Web 客户端需要从一个未必在 Host 机器上的浏览器查看会话工
- **`stat`** 返回 `WorkspaceFileStat { absolutePath, version, bytes? }`:文件身份、不透明的新鲜度令牌,以及后端报得出时的大小。它只接受普通文件。
- **`read`** 返回一个行窗口 `WorkspaceFileText = WorkspaceFileStat & { offset, text, lines, eof }`;`lines` 计页内行数,使只含一个空行的页(`text: ''`、`lines: 1`)与越过文件末尾的页(`lines: 0`)可区分。`range.offset` 是 1 起算的首行,缺省 1;`range.limit` 是最多行数,缺省 `maxLines` 且不得超过。行以 `\n` 结束,末尾的 `\n` 终止最后一行而不是开启一空行;`text` 以 `\n` 连接本页各行且不带终止符;页含最后一行时 `eof` 为 true,越过末尾的 offset 返回 `eof` 为 true 的空页。切页器沿 `streamText` 前进,数过窗口前的行而不保留,把每个窗内片段先按 `maxBytes` 核准再缓冲,并在越过窗口的第一个字符处返回,因此任意大小的文件只花一页内存。页上的 `version` 与 `bytes` 来自流之前的那次 stat。
- **`readBytes`** 返回一个原始字节窗口 `WorkspaceFileBytes = WorkspaceFileStat & { offset, data, eof }`。`range.offset` 是 0 起算的首字节,缺省 0;`range.length` 是最多字节数,缺省 `maxBytes` 且不得超过。`data` 为 base64,文件在窗内结束则短于 `length`,位于或越过末尾则为空;窗口含最后一个字节时 `eof` 为 true。不做任何解码,也不按二进制拒绝。`read` 按行分页、绝不按字节;字节窗口走 `readBytes`。
- **`readAll` 与 `readRelated`** 在 `maxFileBytes` 上限内返回完整的 `WorkspaceFileBytes`。`readRelated` 从基准文件所在目录解析相对文件系统路径;Host 对两个文件执行相同的普通文件检查和后端读取权限。[Document Preview](2026-09-08-document-preview-operations.zh.md) 负责其加载与地址语义。
- **`readBytes`** 返回原生字节、文件元数据、偏移量和 EOF。必填选项对象未传 `range` 时,在 `maxFileBytes` 上限内完整读取;传入时,在 `maxBytes` 上限内读取有界窗口。`range: {}` 使用零偏移和配置的窗口上限;窗口可读取超过完整文件上限的文件。[二进制传输决议](2026-09-17-workspace-file-binary-transfer.zh.md)负责 Remote 响应编码。
- **`baseFile`** 独立于 `range` 选择目标解析方式:目标 `path` 必须为相对路径,从基准文件所在目录解析。Host 对两个文件执行普通文件检查,并保留后端读取权限。[Document Preview](2026-09-08-document-preview-operations.zh.md)负责内容加载与地址语义。
- **`list`** 返回 `WorkspaceDirectoryListing { path, entries, truncated }`:被列目录相对根的工作区路径(根为空串)、其直接子项按后端的稳定名序以 `{ name, type, size? }` 给出,以及 `maxEntries` 是否截断了列表。`type` 为 `file`、`directory` 或 `other`;符号链接子项报告其指向目标的类型,悬空者为 `other`,而打开这样的子项仍会在下文的链接关被拒。dotfile 照常列出,不做任何过滤。
- **`changes`** 产出 `WorkspaceFileWatchFrame`:在观察队列注册且工作区根解析完成后先发 `{ kind: 'ready' }`,随后为 `{ kind: 'change', change }`。载荷 `WorkspaceFileChange` 对存在的文件为 `{ absolutePath, version }`,对消失的文件为 `{ absolutePath, absent: true }`。来源是工作区根内的 `fs/observed`,不监视操作系统。首次拉取后的观察都会排队,包括根解析期间的观察;取消或插件释放会结束该代流。
### 线路上的路径
离开服务的路径词汇有两套,每个方法只用其中一套。`read`、`readBytes`、`readAll`、`readRelated`、`stat` 与 `changes` 以 `absolutePath` 命名文件:它在文件系统执行环境中、符号链接已解析的绝对路径(`ctx.fs.processPath(target)`),因此 Client 提供者按绝对路径把变更帧匹配到已打开的地址:Client 把地址路径原样交给 Host,并只按成功的 `stat.absolutePath` 绑定跟随者,不读取会话摘要的 cwd。`list` 说工作区路径——与其 `path` 参数相同的语法,绝对或相对根——因为其消费方是一棵以根为起点的树。该字段叫 `absolutePath` 而不叫 `url`,因为它不是资源地址;地址语法归 `dsh-util-workspace-path` 所有,与资源模型一并描述。`read`、`readBytes`、`readAll`、`readRelated`、`stat` 与 `list` 的输入路径是绝对路径或相对会话工作区根的路径,从不相对后端自己的 cwd。
离开服务的路径词汇有两套,每个方法只用其中一套。`read`、`readBytes`、`stat` 与 `changes` 以 `absolutePath` 命名文件:它在文件系统执行环境中、符号链接已解析的绝对路径(`ctx.fs.processPath(target)`),因此 Client 提供者按绝对路径把变更帧匹配到已打开的地址:Client 把地址路径原样交给 Host,并只按成功的 `stat.absolutePath` 绑定跟随者,不读取会话摘要的 cwd。`list` 说工作区路径——与其 `path` 参数相同的语法,绝对或相对根——因为其消费方是一棵以根为起点的树。该字段叫 `absolutePath` 而不叫 `url`,因为它不是资源地址;地址语法归 `dsh-util-workspace-path` 所有,与资源模型一并描述。`read`、`readBytes`、`stat` 与 `list` 的输入路径是绝对路径或相对会话工作区根的路径,从不相对后端自己的 cwd;带 `baseFile` 的 `readBytes` 从基准文件所在目录解析目标。
`version` 是消费者只比较是否相等、从不解析的不透明字符串:本地后端由设备、inode、大小及纳秒级 mtime 与 ctime 导出,因此内容不变的重写也会改变它。`offset` 在 `read` 上指行、在 `readBytes` 上指字节;两套单位从不混用,二者的 `eof` 都表示窗口到达了文件末尾。
### 文件检查与工作区包含
`read`、`readBytes`、`readAll`、`readRelated` 与 `stat` 共享普通文件检查,之后依赖文件系统后端的读取权限。`list` 共享路径检查,但还会检查工作区包含关系;`changes` 则把观察过滤到工作区根内。服务执行以下检查:
`read`、`readBytes` 与 `stat` 共享普通文件检查,之后依赖文件系统后端的读取权限。`list` 共享路径检查,但还会检查工作区包含关系;`changes` 则把观察过滤到工作区根内。服务执行以下检查:
1. **路径本身。** `lstat` 在跟随任何东西之前检查路径:缺失路径为 `not-found`;符号链接——不论指向哪里,包括指回工作区内——对文件方法为 `not-regular-file`(kind 为 `symlink`),对 `list` 为 `not-directory`。空路径是 `gateway/bad-request`。
2. **`list` 的工作区包含。** 目录解析为目标,由 `ctx.fs.contains(root, target)` 判定,其中 `root` 是从所选 Session header 解析出的 `WorkspaceFileScope.workspaceRoot`。`..` 爬出或根外绝对目录为 `outside-workspace`。`changes` 对观察到的目标使用相同的后端包含判定。
@@ -73,7 +71,7 @@ Web 客户端需要从一个未必在 Host 机器上的浏览器查看会话工
| `workspace-file/outside-workspace` | `list` 的目标不在工作区根内 | `{ path }` |
| `workspace-file/too-large` | 一页文本或字节窗口超过 `maxBytes`,或全文读取超过 `maxFileBytes` | `{ path, limit }` |
| `workspace-file/not-text` | 到页末为止的非法 UTF-8,或样本或页内的 NUL 字节(仅 `read`) | `{ path }` |
| `workspace-file/not-regular-file` | 对非普通文件执行 `read`、`readBytes`、`readAll`、`readRelated` 或 `stat` | `{ path, kind: 'directory' \| 'symlink' \| 'other' }` |
| `workspace-file/not-regular-file` | 对非普通文件执行 `read`、`readBytes` 或 `stat` | `{ path, kind: 'directory' \| 'symlink' \| 'other' }` |
| `workspace-file/not-directory` | 对非目录执行 `list` | `{ path, kind: 'file' \| 'symlink' \| 'other' }` |
| `workspace-file/unsupported-address` | Client 铸出:本提供者无法服务的资源地址 | `{ address }` |
| `workspace-file/unknown-workspace` | Client 铸出:不携带 Session 的 `absolute` 地址 | `{ address }` |
@@ -109,7 +107,7 @@ Client 导出向 `ctx.resources` 注册一个 `ResourceProvider<'file'>`,存
### 相关记录
[资源模型](2026-09-05-client-resource-model.zh.md)拥有 `ctx.resources`、`useResource`、`dsh-resource://<type>/…` 地址语法以及"每个地址一份资源"的推理;[文本预览与文件树](../feature/2026-09-05-sidebar-text-preview-and-file-tree.zh.md)是 `read`、`list` 与 `file` 提供者随包交付的消费方;[右侧 Sidebar 停靠基础设施](../feature/2026-09-04-right-sidebar-docking-infrastructure.zh.md)是它们打开进去的界面;[工作区文件链接](../feature/2026-07-31-web-workspace-file-links.zh.md)是经 HTTP 供文件被否决之处。任何在这套体系上扩展的人都经 `remote.workspaceFiles` 触达同样的七个方法、经 `useResource<'file'>` 触达同样的 `file` 资源;线路类型以 `@deepseek-ai/dsh-api-workspace-files/types` 发布。[工作区文件读取权限](2026-09-09-workspace-file-read-authority.zh.md)负责 Host 读取权限与 HTML 安全取舍;[Document Preview](2026-09-08-document-preview-operations.zh.md)负责内容加载和逐 tab 新鲜度。
[资源模型](2026-09-05-client-resource-model.zh.md)拥有 `ctx.resources`、`useResource`、`dsh-resource://<type>/…` 地址语法以及"每个地址一份资源"的推理;[文本预览与文件树](../feature/2026-09-05-sidebar-text-preview-and-file-tree.zh.md)是 `read`、`list` 与 `file` 提供者随包交付的消费方;[右侧 Sidebar 停靠基础设施](../feature/2026-09-04-right-sidebar-docking-infrastructure.zh.md)是它们打开进去的界面;[工作区文件链接](../feature/2026-07-31-web-workspace-file-links.zh.md)是经 HTTP 供文件被否决之处。任何在这套体系上扩展的人都经 `remote.workspaceFiles` 触达同样的五个方法、经 `useResource<'file'>` 触达同样的 `file` 资源;线路类型以 `@deepseek-ai/dsh-api-workspace-files/types` 发布。[工作区文件读取权限](2026-09-09-workspace-file-read-authority.zh.md)负责 Host 读取权限与 HTML 安全取舍;[Document Preview](2026-09-08-document-preview-operations.zh.md)负责内容加载和逐 tab 新鲜度。
## Alternatives considered
@@ -117,7 +115,7 @@ Client 导出向 `ctx.resources` 注册一个 `ResourceProvider<'file'>`,存
**带有反向 UI 依赖的双面包。** 拆包选择源于 `api/remotes` 引用 Client 叶子后形成的两条工程引用环:资源模型为了结果类型引用 Remote 装配,文件提供者为了 Sidebar 参数表引用右栏 UI。TypeScript 以 `TS6202` 拒绝这些环。[双面包组织](2026-09-07-workspace-files-dual-face-package.zh.md)取代拆包选择:结果类型直接取自协议包,Sidebar 参数注册移至文本预览;保留两个根聚合中的显式编译入口。
**经 HTTP 供工作区文件。** 已被[工作区文件链接](../feature/2026-07-31-web-workspace-file-links.zh.md)以 origin 理由否决且未重议:`read` 与 `readBytes` 经认证的 Remote 载体传送纯文本与 base64,因此不供文档、不铸 URL,也不产生 origin 问题。
**提供可导航的工作区文档。** 因源隔离问题不采用;[二进制传输](2026-09-17-workspace-file-binary-transfer.zh.md)返回 multipart 数据,HTML 仍隔离在 Blob iframe 中。
**文件读取采用“日志可达”或工作区包含授权。** 唯一把文件内容送过线路的先例——命令附件——只授权出现在会话日志里的文件,这会排除手输路径;工作区包含则会排除 Session 后端其他位置的可读文件。[工作区文件读取权限](2026-09-09-workspace-file-read-authority.zh.md)改为以后端读取决策为准,只对工作区形态的操作保留包含限制。
@@ -138,14 +136,14 @@ Client 导出向 `ctx.resources` 注册一个 `ResourceProvider<'file'>`,存
- 变更帧只报告 agent 自己的操作。用户编辑器、shell 或子进程改动的文件不产生帧;agent 仅仅读取一个被别处改动的文件却会产生帧,因为读取观察到了新版本。
- 文件类型检查先于后端读取,`list` 也先报种类再报根外位置。页的 `version` 可能落后内容一次写入,停滞的 `changes` 消费者会让 Host 内存增长,因为一代流的队列无界;每一条都是包 README 记录在册的已知取舍。
- `file` 资源推送变更而非内容,因此预览不靠载荷就得知文件已更新并读取它想要的页;失败的打开继续跟随地址,因此 agent 创建该文件时 tab 无需用户动作即恢复正常。
- `readBytes` 尚无随包交付的消费方;Document Preview 对整文件格式使用 `readAll`。
- Office 授权探测使用带单字节 `range` 的 `readBytes`;Document Preview 对整文件格式使用同一方法并省略 `range`。
## Testing
`packages/api/workspace-files/tests` 中的 Host spec 覆盖 live 与 cold subagent Session 的 header-only scope 解析、部署 fallback、缺失身份与 lookup 释放;分页读取(整文件、嵌套路径、空文件、多字节 UTF-8、行窗口边界、缺省与拒绝的 limit、保留回车);字节窗口(缺省、中段与尾窗、越界与空文件、base64 往返、版本、上限、坏范围以及无大小时的 `eof`);`stat`;工作区外读取及后端拒绝;`list` 的包含、截断、符号链接与 `not-directory`;以及由 `fs/observed` 驱动并按根过滤的 `changes`。Client spec 覆盖提供者帧、变更流、不支持地址及注册与释放。`fs/fs` 与 `fs-local` spec 钉住 `readByteRange`;`dsh-util-workspace-path` spec 钉住文件地址语法。connection fixture 为 web e2e 套件提供 `stat`、分页 `read`、`list` 与一帧可选启用的 `changes`。
`packages/api/workspace-files/tests` 中的 Host spec 覆盖 live 与 cold subagent Session 的 header-only scope 解析、部署 fallback、缺失身份与 lookup 释放;分页读取(整文件、嵌套路径、空文件、多字节 UTF-8、行窗口边界、缺省与拒绝的 limit、保留回车);字节窗口(缺省、中段与尾窗、越界与空文件、`Uint8Array` 往返、版本、上限、坏范围以及无大小时的 `eof`);`stat`;工作区外读取及后端拒绝;`list` 的包含、截断、符号链接与 `not-directory`;以及由 `fs/observed` 驱动并按根过滤的 `changes`。Client spec 覆盖提供者帧、变更流、不支持地址及注册与释放。`fs/fs` 与 `fs-local` spec 钉住 `readByteRange`;`dsh-util-workspace-path` spec 钉住文件地址语法。connection fixture 为 web e2e 套件提供 `stat`、分页 `read`、`list` 与一帧可选启用的 `changes`。
## Deferred
- 给 `changes` 一代流的队列加上限。
- `readBytes` 随包交付的消费方(图片与二进制预览)以及任何写入、搜索或媒体路由;本服务只读。
- 写入和搜索操作;本服务只读。
- 文件地址中 `session` 之外的作用域;语法留有余地,提供者只服务一个。
@@ -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/2026-09-08-document-preview-operations.md
2026-09-08-document-preview-operations.md: be3a835a6296cb28d9a90c5c5e8b9bf7367598c2
2026-09-08-document-preview-operations.zh.md: efb87e16cc126001c1a6eb2bf938124d7e00044a
2026-09-08-document-preview-operations.md: 0b90c17ba94d0c7cd30afb0a6e6a033f8f12d0df
2026-09-08-document-preview-operations.zh.md: ec9c208c97dfe75afdb0915c655d9c1c7b4d42ff
@@ -16,9 +16,9 @@ Document Preview separates resource observation from content reads. The [resourc
Readable files use `dsh-resource://file/session/<sessionId>/<path>`. The path may be workspace-relative or absolute; an encoded absolute path retains its leading slash. `fileAddressFor` always emits this Session-address form. The provider and Preview RPC take the Session only from that address, never from the current selection, first holder, or owning tab. A Session-less `absolute` URI cannot be read; the provider reports `workspace-file/unknown-workspace`. Session authorization is a file-protocol rule, not an additional Resource identity.
[Document Preview](../../../../packages/client/ui-sidebar-documentpreview/README.md) owns format selection and loading policy. Metadata registers with `ctx.documentPreviews`; components register separately into the keyed `sidebar.right.tab.document` Slot. Extension registrations precede builtins, then longer suffixes and registration order decide. The toolbar lists supported alternatives and remembers a manual choice per tab. Plain text remains available for unknown extensions and text-compatible renderers. Registered binary suffixes determine text compatibility independently of loading mode; HTML and SVG remain outside those suffixes and retain source viewing. A single candidate renders no viewer control. Registered binary suffixes omit plain text; known binary suffixes without a renderer show an unsupported state without reading the file ([sidebar preview polish](../feature/2026-09-11-sidebar-document-preview-polish.md)). The child receives accumulated text or complete native bytes, the original resource address, and the standard `useResource` and `useTabInfo` hooks. Preview calls existing `read`, `readAll`, and `readRelated` through ordinary injection and decodes bytes in its own `rpc.ts`. Refresh remains per tab, with no resource reload, shared `changed` acknowledgement, extra resource wrapper, or content Session.
[Document Preview](../../../../packages/client/ui-sidebar-documentpreview/README.md) owns format selection and loading policy. Metadata registers with `ctx.documentPreviews`; components register separately into the keyed `sidebar.right.tab.document` Slot. Extension registrations precede builtins, then longer suffixes and registration order decide. The toolbar lists supported alternatives and remembers a manual choice per tab. Plain text remains available for unknown extensions and text-compatible renderers. Registered binary suffixes determine text compatibility independently of loading mode; HTML and SVG remain outside those suffixes and retain source viewing. A single candidate renders no viewer control. Registered binary suffixes omit plain text; known binary suffixes without a renderer show an unsupported state without reading the file ([sidebar preview polish](../feature/2026-09-11-sidebar-document-preview-polish.md)). The child receives accumulated text or complete native bytes, the original resource address, and the standard `useResource` and `useTabInfo` hooks. Preview calls `ctx.remote.workspaceFiles.read` for text and `ctx.remote.workspaceFiles.readBytes` for native bytes, with `baseFile` for related resources. Refresh remains per tab, with no resource reload, shared `changed` acknowledgement, extra resource wrapper, or content Session.
Markdown and code reuse the incremental primitives with cumulative paged text. HTML, PDF, and images read complete `Uint8Array<ArrayBuffer>` data; Host transport remains base64. Published buffers are borrowed read-only and never persist into layout or Session JSON. PDF.js runs in an owned Worker with version-matched bundled font and decoder data, and copies input before transfer to preserve Preview's retained buffer. With [Developer tools](../feature/2026-09-17-developer-tools-settings.md) enabled, HTML runs in a Blob iframe with `sandbox="allow-scripts"`, without same-origin, popup, form, download, or top-navigation privileges. The browser retains its normal external-network rules. Bounded static local JS/CSS reads stay in the parent; the opaque frame creates its own asset Blobs, because it cannot load parent-origin Blobs. PNG, JPEG, GIF, WebP, BMP, ICO, and SVG use image-specific Blob URLs in an `<img>` static-image context. An image wider than the pane scales down to its width at its aspect ratio; a smaller image keeps its intrinsic CSS-pixel dimensions centred by auto margins, and a taller image extends the shared scroller's vertical range ([sidebar preview polish](../feature/2026-09-11-sidebar-document-preview-polish.md)). The renderer provides no zoom or drag-to-pan. SVG markup never enters the application DOM or an iframe, so scripts remain inert and cannot reach the parent page. Replacing HTML or an image revokes its root Blob URL.
Markdown and code reuse the incremental primitives with cumulative paged text. HTML, PDF, and images read complete `Uint8Array<ArrayBuffer>` data; Host transport uses the [binary workspace file transfer](2026-09-17-workspace-file-binary-transfer.md). Published buffers are borrowed read-only and never persist into layout or Session JSON. PDF.js runs in an owned Worker with version-matched bundled font and decoder data, and copies input before transfer to preserve Preview's retained buffer. With [Developer tools](../feature/2026-09-17-developer-tools-settings.md) enabled, HTML runs in a Blob iframe with `sandbox="allow-scripts"`, without same-origin, popup, form, download, or top-navigation privileges. The browser retains its normal external-network rules. Bounded static local JS/CSS reads stay in the parent; the opaque frame creates its own asset Blobs, because it cannot load parent-origin Blobs. PNG, JPEG, GIF, WebP, BMP, ICO, and SVG use image-specific Blob URLs in an `<img>` static-image context. An image wider than the pane scales down to its width at its aspect ratio; a smaller image keeps its intrinsic CSS-pixel dimensions centred by auto margins, and a taller image extends the shared scroller's vertical range ([sidebar preview polish](../feature/2026-09-11-sidebar-document-preview-polish.md)). The renderer provides no zoom or drag-to-pan. SVG markup never enters the application DOM or an iframe, so scripts remain inert and cannot reach the parent page. Replacing HTML or an image revokes its root Blob URL.
PDF.js's official TextLayerBuilder owns selection boundaries and copy normalization over the width-fitted canvas, with shared page cleanup and a component-owned resize observer. Responsive sizing uses the CSS `scale` property independently of PDF.js's page rotation and translation transforms. Its end-of-content marker and stacking rules constrain selection in blank regions; line-break highlighting is suppressed. Per-page cancellation uses the builder's cleanup rather than aborting the first page's signal, because the official selection listeners are shared across pages.
@@ -28,7 +28,7 @@ PDF.js's official TextLayerBuilder owns selection boundaries and copy normalizat
**Methods attached to an Iterator or its values.** This conflates observation with commands and repeats capability identity in data frames. Frames carry data and failures; explicit Preview RPC callbacks perform reads.
**A core public-projection factory, or the same assembly inside `open`.** Separate stream values, operations bundles, and public interfaces add assembly without another current consumer that needs it. Preview's shared RPC adapter already keeps Session decoding and base64 out of renderers. Resource offers no provider-agnostic command interface or opening-bound command lifetime; adding either needs consumer evidence beyond file preview.
**A core public-projection factory, or the same assembly inside `open`.** Separate stream values, operations bundles, and public interfaces add assembly without another current consumer that needs it. Preview's shared RPC adapter already keeps Session addressing and byte transport out of renderers. Resource offers no provider-agnostic command interface or opening-bound command lifetime; adding either needs consumer evidence beyond file preview.
**UI Session as extra Resource identity, or authorization from the first holder or current selection.** A retained tab can belong to a different Session from the selected one, and the UI location does not identify the addressed file. Encoding the required Session in the file address preserves Host authorization while letting all readers of one address share observation.
@@ -16,9 +16,9 @@ Document Preview 将资源观察与内容读取分开。[资源模型](2026-09-0
可读取的文件使用 `dsh-resource://file/session/<sessionId>/<path>`。路径可以相对工作区,也可以是绝对路径;编码后的绝对路径保留前导斜杠。`fileAddressFor` 始终生成这种 Session 地址。提供方与 Preview RPC 只从该地址取 Session,不取当前选择、首个持有者或 tab 所属 Session。不带 Session 的 `absolute` URI 无法读取;提供方报告 `workspace-file/unknown-workspace`。Session 授权是文件协议规则,不是额外的 Resource 身份。
[Document Preview](../../../../packages/client/ui-sidebar-documentpreview/README.zh.md) 负责格式选择和加载策略。元数据通过 `ctx.documentPreviews` 注册;组件单独注册到 keyed `sidebar.right.tab.document` Slot。扩展注册优先于内置注册,其次比较后缀长度和注册顺序。工具栏列出受支持的候选,按 tab 记住手动选择。未知扩展名及文本兼容的渲染器仍可使用纯文本。文本兼容性由注册的二进制后缀决定,与加载方式无关;HTML 和 SVG 不属于二进制后缀,保留源码查看。只有一个候选时不显示查看器控件。注册声明为二进制的后缀不提供纯文本;已知二进制后缀没有注册渲染器时不发起读取,并显示不支持预览的空态([侧边栏预览打磨](../feature/2026-09-11-sidebar-document-preview-polish.zh.md))。子组件收到累积文本或完整原生字节、原始资源地址,以及标准 `useResource` 和 `useTabInfo` 钩子。Preview 经普通注入调用既有 `read`、`readAll` 与 `readRelated`,在自己的 `rpc.ts` 解码字节。刷新仍按 tab 独立进行,不引入资源 reload、共享 `changed` 确认、额外资源包装层或内容 Session。
[Document Preview](../../../../packages/client/ui-sidebar-documentpreview/README.zh.md) 负责格式选择和加载策略。元数据通过 `ctx.documentPreviews` 注册;组件单独注册到 keyed `sidebar.right.tab.document` Slot。扩展注册优先于内置注册,其次比较后缀长度和注册顺序。工具栏列出受支持的候选,按 tab 记住手动选择。未知扩展名及文本兼容的渲染器仍可使用纯文本。文本兼容性由注册的二进制后缀决定,与加载方式无关;HTML 和 SVG 不属于二进制后缀,保留源码查看。只有一个候选时不显示查看器控件。注册声明为二进制的后缀不提供纯文本;已知二进制后缀没有注册渲染器时不发起读取,并显示不支持预览的空态([侧边栏预览打磨](../feature/2026-09-11-sidebar-document-preview-polish.zh.md))。子组件收到累积文本或完整原生字节、原始资源地址,以及标准 `useResource` 和 `useTabInfo` 钩子。Preview 通过 `ctx.remote.workspaceFiles.read` 读取文本,通过 `ctx.remote.workspaceFiles.readBytes` 读取原生字节,关联资源使用 `baseFile`。刷新仍按 tab 独立进行,不引入资源 reload、共享 `changed` 确认、额外资源包装层或内容 Session。
Markdown 和代码通过累积的分页文本复用增量渲染原语。HTML、PDF 和图片读取完整 `Uint8Array<ArrayBuffer>` 数据;Host 传输保持 base64。发布后的缓冲区只读借用,绝不持久化进布局或 Session JSON。PDF.js 在自有 Worker 中运行,字体和解码数据以相同版本随包发布,转移输入前先复制,以保留 Preview 的缓冲区。开启[开发者工具](../feature/2026-09-17-developer-tools-settings.zh.md)时,HTML 在 Blob iframe 中运行,设置 `sandbox="allow-scripts"`,不授予同源、弹窗、表单、下载或顶层导航权限。浏览器保持正常的外部网络规则。有上限的静态本地 JS/CSS 读取由父页面负责;不透明源 iframe 创建自己的资源 Blob,因为它不能加载父源创建的 Blob。PNG、JPEG、GIF、WebP、BMP、ICO 和 SVG 使用图片专用 Blob URL,在 `<img>` 静态图片上下文中渲染。比面板宽的图片按纵横比缩小到面板宽度;较小的图片保留固有 CSS 像素尺寸并由 auto margin 居中,较高的图片扩展共享滚动区的纵向范围([侧边栏预览打磨](../feature/2026-09-11-sidebar-document-preview-polish.zh.md))。渲染器不提供缩放或拖拽平移。SVG 标记绝不进入应用 DOM 或 iframe,因此脚本保持不可执行,也无法访问父页面。替换 HTML 或图片时会撤销其根 Blob URL。
Markdown 和代码通过累积的分页文本复用增量渲染原语。HTML、PDF 和图片读取完整 `Uint8Array<ArrayBuffer>` 数据;Host 传输使用[工作区文件二进制传输](2026-09-17-workspace-file-binary-transfer.zh.md)。发布后的缓冲区只读借用,绝不持久化进布局或 Session JSON。PDF.js 在自有 Worker 中运行,字体和解码数据以相同版本随包发布,转移输入前先复制,以保留 Preview 的缓冲区。开启[开发者工具](../feature/2026-09-17-developer-tools-settings.zh.md)时,HTML 在 Blob iframe 中运行,设置 `sandbox="allow-scripts"`,不授予同源、弹窗、表单、下载或顶层导航权限。浏览器保持正常的外部网络规则。有上限的静态本地 JS/CSS 读取由父页面负责;不透明源 iframe 创建自己的资源 Blob,因为它不能加载父源创建的 Blob。PNG、JPEG、GIF、WebP、BMP、ICO 和 SVG 使用图片专用 Blob URL,在 `<img>` 静态图片上下文中渲染。比面板宽的图片按纵横比缩小到面板宽度;较小的图片保留固有 CSS 像素尺寸并由 auto margin 居中,较高的图片扩展共享滚动区的纵向范围([侧边栏预览打磨](../feature/2026-09-11-sidebar-document-preview-polish.zh.md))。渲染器不提供缩放或拖拽平移。SVG 标记绝不进入应用 DOM 或 iframe,因此脚本保持不可执行,也无法访问父页面。替换 HTML 或图片时会撤销其根 Blob URL。
PDF.js 官方 TextLayerBuilder 在适配宽度的 canvas 上负责选择边界和复制规范化,共享页面清理,并使用由组件拥有的 resize observer。响应式尺寸适配使用独立的 CSS `scale` 属性,与 PDF.js 的页面旋转和平移变换组合。其内容结束标记和堆叠规则限制空白区域中的选择;换行高亮被抑制。逐页取消使用 builder 的清理操作,而不 abort 第一页的信号,因为官方选择监听器跨页面共享。
@@ -28,7 +28,7 @@ PDF.js 官方 TextLayerBuilder 在适配宽度的 canvas 上负责选择边界
**把方法挂到 Iterator 或其值上。** 这会混淆观察与命令,并在数据帧中重复能力身份。帧携带数据和失败;显式 Preview RPC 回调负责读取。
**核心公开投影工厂,或在 `open` 内做同样的组装。** 分开的流值、operations 组合与公开接口增加了组装步骤,没有另一个当前消费方需要它。Preview 的共享 RPC 适配已让渲染器无需解码 Session 和 base64。Resource 不提供与提供方无关的命令接口,也不提供绑定于打开实例的命令生命周期;增加任一种都需要文件预览之外的消费方证据。
**核心公开投影工厂,或在 `open` 内做同样的组装。** 分开的流值、operations 组合与公开接口增加了组装步骤,没有另一个当前消费方需要它。Preview 的共享 RPC 适配已让渲染器无需处理 Session 寻址和字节传输。Resource 不提供与提供方无关的命令接口,也不提供绑定于打开实例的命令生命周期;增加任一种都需要文件预览之外的消费方证据。
**把 UI Session 作为额外 Resource 身份,或由首个持有者、当前选择决定授权。** 保留的 tab 可以属于不同于当前选择的 Session,UI 所在位置也不能标识地址指向的文件。将所需 Session 编入文件地址,既保留 Host 授权,也让同地址的所有读者共享观察。
@@ -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/2026-09-09-workspace-file-read-authority.md
2026-09-09-workspace-file-read-authority.md: d1a5b9d9634381cb74b17651c67f2b701402ff1c
2026-09-09-workspace-file-read-authority.zh.md: f8b0fae4325a90548c6b9dc33ac46b17248b04e7
2026-09-09-workspace-file-read-authority.md: 140816e8c48a6628133cf9208d5b48545e2e5eaf
2026-09-09-workspace-file-read-authority.zh.md: 75ffee32c40930c459018ab0fde42d5324371ba8
@@ -10,11 +10,11 @@ Workspace Files serves both file content and workspace navigation. Applying work
## Decision
`read`, `readBytes`, `readAll`, `readRelated`, and `stat` inherit the addressed Session filesystem backend's read authority. The workspace root is the base for relative input paths, not a read boundary; absolute paths and relative paths that leave the workspace are readable when the backend allows them. The service still requires regular files, refuses symlinks, and applies its text and byte caps.
`read`, `readBytes`, and `stat` inherit the addressed Session filesystem backend's read authority. The workspace root is the base for relative input paths, not a read boundary; absolute paths and relative paths that leave the workspace are readable when the backend allows them. The service still requires regular files, refuses symlinks, and applies its text and byte caps.
`list` and `changes` remain workspace-scoped because they expose workspace navigation and observation rather than a named file read. `list` rejects a directory outside the root, and `changes` filters observations through the backend's workspace-containment predicate.
`readRelated` resolves a relative path from the base file's directory. A `..` path may therefore read JavaScript or CSS outside the workspace when the Session backend permits it. With [Developer tools](../feature/2026-09-17-developer-tools-settings.md) enabled, Document Preview packages bounded, statically declared local scripts and stylesheets into an HTML Blob iframe with `sandbox="allow-scripts"`; the opaque origin blocks parent access, but the browser retains normal network access. This exposure is an intentional security trade-off for rendering static generated HTML.
`readBytes` with `baseFile` resolves a relative path from the base file's directory. A `..` path may therefore read JavaScript or CSS outside the workspace when the Session backend permits it. With [Developer tools](../feature/2026-09-17-developer-tools-settings.md) enabled, Document Preview packages bounded, statically declared local scripts and stylesheets into an HTML Blob iframe with `sandbox="allow-scripts"`; the opaque origin blocks parent access, but the browser retains normal network access. This exposure is an intentional security trade-off for rendering static generated HTML.
The [Workspace Files service](2026-09-05-workspace-files-service.md) owns paging, file checks, listing, and observation. [Document Preview](2026-09-08-document-preview-operations.md) owns which related files are packaged and the iframe sandbox.
@@ -10,11 +10,11 @@ Workspace Files 同时提供文件内容与工作区导航。对所有操作应
## Decision
`read`、`readBytes`、`readAll`、`readRelated` 与 `stat` 继承被寻址 Session 的文件系统后端读取权限。工作区根是输入相对路径的基准,而不是读取边界;只要后端允许,就可以读取绝对路径和离开工作区的相对路径。服务仍要求普通文件、拒绝符号链接,并应用文本和字节上限。
`read`、`readBytes` 与 `stat` 继承被寻址 Session 的文件系统后端读取权限。工作区根是输入相对路径的基准,而不是读取边界;只要后端允许,就可以读取绝对路径和离开工作区的相对路径。服务仍要求普通文件、拒绝符号链接,并应用文本和字节上限。
`list` 与 `changes` 仍限于工作区,因为它们暴露工作区导航和观察,而不是读取一个具名文件。`list` 拒绝根外目录,`changes` 通过后端的工作区包含判定过滤观察。
`readRelated` 从基准文件所在目录解析相对路径。因此,只要 Session 后端允许,`..` 路径就可以读取工作区外的 JavaScript 或 CSS。开启[开发者工具](../feature/2026-09-17-developer-tools-settings.zh.md)时,Document Preview 把有界、静态声明的本地脚本与样式表打包进带 `sandbox="allow-scripts"` 的 HTML Blob iframe;不透明源阻止访问父应用,但浏览器保留正常网络访问。这种暴露是为渲染静态生成 HTML 而有意接受的安全取舍。
带 `baseFile` 的 `readBytes` 从基准文件所在目录解析相对路径。因此,只要 Session 后端允许,`..` 路径就可以读取工作区外的 JavaScript 或 CSS。开启[开发者工具](../feature/2026-09-17-developer-tools-settings.zh.md)时,Document Preview 把有界、静态声明的本地脚本与样式表打包进带 `sandbox="allow-scripts"` 的 HTML Blob iframe;不透明源阻止访问父应用,但浏览器保留正常网络访问。这种暴露是为渲染静态生成 HTML 而有意接受的安全取舍。
[Workspace Files 服务](2026-09-05-workspace-files-service.zh.md)负责分页、文件检查、列举和观察。[Document Preview](2026-09-08-document-preview-operations.zh.md)负责选择要打包的关联文件及 iframe sandbox。
@@ -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/2026-09-11-node-office-kit.md
2026-09-11-node-office-kit.md: 6cab947d7cb8d4c2c416a83f0b765813c7f98ea9
2026-09-11-node-office-kit.zh.md: 8ed1a81922249cf0b9c1a74b06e5ce9d44696e0a
2026-09-11-node-office-kit.md: 132ac8b3dd709f9f16254a5569dec59f01d29fda
2026-09-11-node-office-kit.zh.md: c7bdab2322766c42cfee4092f158d7b6b530e875
@@ -14,7 +14,7 @@ The [office-to-pdf capability](../../../../packages/document/README.md) delegate
The [platform engine decision](2026-09-15-platform-office-engines.md) requires the kit’s declared native target engine, or WASM when no native target is declared. Missing or invalid required engines reject conversion. The shared [bounded provider](2026-09-15-bounded-office-conversion.md) owns admission, conversion reuse, and cancellation through scratch cleanup. Preview consumes that provider without registering another converter or requiring Office authoring skills.
The service, Remote methods, and Client registration accept DOC, DOCX, XLS, XLSX, PPT, and PPTX. The kit verifies bounded ZIP membership and content types for OOXML inputs and the OLE compound-file header for binary Office inputs before LibreOffice imports them. Renaming text to an Office suffix does not admit it. Binary formats return no missing-font diagnostics because the kit does not extract their font tables. It writes a fresh, exclusively created PDF in a caller-owned private directory. DSH reads and validates the complete output before removing scratch files. The [service’s Remote method](../../../../packages/document/office-to-pdf/README.md) authorizes source access through [Workspace Files](2026-09-09-workspace-file-read-authority.md), preserves the source path/version, and returns PDF bytes. Source read caps and generated PDF caps remain independent. One source path/version snapshot governs the access probe and deferred read, including size-failure rechecks, so conversion cannot publish bytes under another identity. Preview bytes do not enter Session storage or persistent caches.
The service and Remote methods accept DOC, DOCX, XLS, XLSX, PPT, and PPTX. The Client preview registers DOC, DOCX, PPT, and PPTX; XLS and XLSX use the unsupported-preview state and can open in the system default application without requesting conversion. The kit verifies bounded ZIP membership and content types for OOXML inputs and the OLE compound-file header for binary Office inputs before LibreOffice imports them. Renaming text to an Office suffix does not admit it. Binary formats return no missing-font diagnostics because the kit does not extract their font tables. It writes a fresh, exclusively created PDF in a caller-owned private directory. DSH reads and validates the complete output before removing scratch files. The [service’s Remote method](../../../../packages/document/office-to-pdf/README.md) authorizes source access through [Workspace Files](2026-09-09-workspace-file-read-authority.md), preserves the source path/version, and returns PDF bytes. Source read caps and generated PDF caps remain independent. One source path/version snapshot governs the access probe and deferred read, including size-failure rechecks, so conversion cannot publish bytes under another identity. Preview bytes do not enter Session storage or persistent caches.
A converter reuses font metadata obtained by its first conversion Worker; original font buffers and decoded glyph coverage remain conversion-local. Workers validate indexed files when reading them. Exact installed families precede configured alternatives, and complete family/style/weight/italic/width/pitch/language/code-point requests retain their distinct matches. WASM callbacks import original font files, including complete collections, into MEMFS. Native engines also retain their platform's font discovery. Neither path downloads or installs fonts; native OS-managed font memory is outside the explicit import budget. Recreating the converter refreshes its metadata after font changes.
@@ -24,7 +24,7 @@ DSH exports raster images at configurable resolution, defaulting to 192 DPI for
The [Office viewer](../../../../packages/client/ui-sidebar-documentpreview/README.md#office-preview) lives under Document Preview’s `client/office/` directory, alongside the loading lifecycle, PDF body, and reader types it uses. Keeping these components in one package removes an independent UI boot entry without creating cross-plugin runtime imports. Its bounded cache validates authorized source metadata, shares pending conversions between readers, cancels only when the final reader leaves, excludes failures, and clears on connection reset. Conversion starts when a user opens a preview. Missing declared font families accompany the PDF and are available from the document toolbar’s warning icon; font-table inventories and unrelated engine defaults are not warnings. The shared preview entry’s `office` cache settings use the existing page-global injection channel because the module boot graph carries package identities, not Loader configuration. Reloading the page adopts updated YAML values.
Ordinary file reads and Office responses share `documentFileBytes()`, which decodes into one typed byte buffer rather than materializing a JavaScript element array from the binary string. Element-array expansion can exhaust the browser heap for a PDF that the configured Host limits permit. A child-process regression checks byte equality within a fixed heap, while the built browser scenario verifies the transport and PDF Worker together. Cache byte limits do not bound transient transport or viewer memory.
Office responses use `documentFileBytes()`, which decodes into one typed byte buffer rather than materializing a JavaScript element array from the binary string. Element-array expansion can exhaust the browser heap for a PDF that the configured Host limits permit. A child-process regression checks byte equality within a fixed heap, while the built browser scenario verifies the transport and PDF Worker together. Cache byte limits do not bound transient transport or viewer memory.
The [kit ownership decision](2026-09-14-independent-libreoffice-kit.md) defines npm distribution and bundled offline conversion.
@@ -14,7 +14,7 @@ Status: implemented
[平台引擎决策](2026-09-15-platform-office-engines.zh.md)要求使用 kit 已声明的原生目标引擎,未声明原生目标时使用 WASM。缺失或无效的必需引擎会拒绝转换。共享的[有界提供方](2026-09-15-bounded-office-conversion.zh.md)负责准入、转换复用以及持续到临时文件清理完成的取消。预览消费该提供方,不注册另一个转换器,也不依赖 Office 创作 skills。
服务、Remote 方法和 Client 注册接受 DOC、DOCX、XLS、XLSX、PPT 和 PPTX。LibreOffice 导入前,kit 校验 OOXML 输入的有界 ZIP 成员和内容类型,以及二进制 Office 输入的 OLE 复合文件头。将文本改为 Office 后缀不能通过校验。kit 不提取二进制格式的字体表,因此这些格式不返回缺失字体诊断。kit 在调用方拥有的私有目录中独占创建新的 PDF。DSH 读取并校验完整输出后才删除临时文件。[服务的 Remote 方法](../../../../packages/document/office-to-pdf/README.zh.md)通过 [Workspace Files](2026-09-09-workspace-file-read-authority.zh.md)授权源文件访问,保留源路径和版本,并返回 PDF 字节。源文件读取上限与生成 PDF 上限相互独立。读取权限探测和延迟读取(包括超限失败后的复查)采用同一个源路径/版本快照,防止转换将字节发布到另一个源身份下。预览字节不会进入 Session 存储或持久缓存。
服务与 Remote 方法接受 DOC、DOCX、XLS、XLSX、PPT 和 PPTX。Client 预览注册 DOC、DOCX、PPT 和 PPTX;XLS 与 XLSX 使用不支持预览的空状态,可由系统默认应用打开,并且不请求转换。LibreOffice 导入前,kit 校验 OOXML 输入的有界 ZIP 成员和内容类型,以及二进制 Office 输入的 OLE 复合文件头。将文本改为 Office 后缀不能通过校验。kit 不提取二进制格式的字体表,因此这些格式不返回缺失字体诊断。kit 在调用方拥有的私有目录中独占创建新的 PDF。DSH 读取并校验完整输出后才删除临时文件。[服务的 Remote 方法](../../../../packages/document/office-to-pdf/README.zh.md)通过 [Workspace Files](2026-09-09-workspace-file-read-authority.zh.md)授权源文件访问,保留源路径和版本,并返回 PDF 字节。源文件读取上限与生成 PDF 上限相互独立。读取权限探测和延迟读取(包括超限失败后的复查)采用同一个源路径/版本快照,防止转换将字节发布到另一个源身份下。预览字节不会进入 Session 存储或持久缓存。
converter 复用首个转换 Worker 返回的字体元数据;原始字体缓冲区和解码后的字符覆盖范围仍只属于单次转换。Worker 读取字体时校验索引中的文件。已安装字体族的精确匹配优先于配置的替代字体,完整的字体族、样式、字重、斜体、宽度、字距、语言与码点请求保留各自的匹配结果。WASM 回调将包含完整字体集合的原始字体文件导入 MEMFS。原生引擎还保留各平台的字体发现能力。两条路径均不下载或安装字体;原生操作系统管理的字体内存不受显式导入预算约束。字体变化后,重新创建 converter 会刷新元数据。
@@ -24,7 +24,7 @@ DSH 按可配置分辨率导出栅格图片,默认 192 DPI,对应共享 PDF
[Office 查看器](../../../../packages/client/ui-sidebar-documentpreview/README.zh.md#office-preview)位于文档预览的 `client/office/` 目录,与其使用的加载生命周期、PDF 正文和读取器类型同属一个包。这些组件放在同一包中,既减少一个独立 UI 启动入口,也无需跨插件运行时导入。其有界缓存校验已授权的源元数据,在读取方之间共享待完成转换,仅在最后一个读取方离开时取消,不缓存失败,并在连接重置时清空。用户打开预览时才开始转换。缺失的已声明字体族随 PDF 返回,可通过文档工具栏的警告图标查看;字体表清单与无关的引擎默认字体不构成警告。共享预览入口的 `office` 缓存设置复用页面全局注入通道,因为模块启动图携带包标识而不传递 Loader 配置。重新加载页面后采用更新的 YAML 值。
普通文件读取与 Office 响应共享 `documentFileBytes()`,解码使用一个类型化字节缓冲区,不将二进制字符串物化为 JavaScript 元素数组。即使 PDF 符合 Host 配置的大小限制,元素数组展开也可能耗尽浏览器堆内存。子进程回归测试在固定堆容量内检查字节一致性,构建后的浏览器场景则一起验证传输和 PDF Worker。缓存字节限制不约束临时传输或查看器内存。
Office 响应使用 `documentFileBytes()`,解码使用一个类型化字节缓冲区,不将二进制字符串物化为 JavaScript 元素数组。即使 PDF 符合 Host 配置的大小限制,元素数组展开也可能耗尽浏览器堆内存。子进程回归测试在固定堆容量内检查字节一致性,构建后的浏览器场景则一起验证传输和 PDF Worker。缓存字节限制不约束临时传输或查看器内存。
[kit 归属决策](2026-09-14-independent-libreoffice-kit.zh.md)定义 npm 分发和随应用打包的离线转换。
@@ -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/2026-09-17-shared-office-runtime.md
2026-09-17-shared-office-runtime.md: 25b6ddccbb5b17b93bec2c0b024dfcb63e94abc3
2026-09-17-shared-office-runtime.zh.md: 99dffe8d0233e60480689ff983b6261cbb83bb7b
@@ -0,0 +1,37 @@
# Agent Note: Shared Office runtime for Desktop and SDK
Status: implemented
English | [中文](2026-09-17-shared-office-runtime.zh.md)
## Problem
SDK deployments need the same Office authoring libraries as Desktop while keeping interpreter payloads in read-only container image layers. A Desktop-owned query and builder require downstream carriers to duplicate dependency locks and path conventions.
## Decision
[`tool-workspace-dependencies`](../../../../packages/skill/tool-workspace-dependencies/README.md) owns manifest validation, interpreter paths, installation and the model query. It stays in the skill group beside its Office workflow consumer: the payload supplies those workflows, while the package exposes no workspace entity service. Desktop imports this package and retains copy-on-first-use installation. The packaged SDK profile enables the query and Office skills from its carrier default; `DSH_PRIMARY_RUNTIME` overrides that path, including an empty-string opt-out. It reads the payload in place. The common base and `sdk-minimal` do not enable Office implicitly.
The [shared build entry](../../../../scripts/primary-runtime/prepare.ts) owns the download lock, extraction and native smoke checks for Desktop targets and GNU/Linux x64/ARM64. Desktop supplies output paths and its release version, requires Node.js and pnpm, and applies its signing-specific subprocess environment. Other carriers choose their output directory and may build a Python-only payload. The existing `desktopVersion` manifest field retains its name and records the carrier release; `payloadDigest` distinguishes locked inputs and component selection.
The [Desktop primary-runtime decision](../feature/2026-09-14-desktop-primary-runtime.md) continues to own Desktop installation and platform signing. This decision replaces only its Desktop-only builder and query placement; both records remain active.
`runtime.json` stores Python, Node.js and pnpm versions at the top level and all Python distribution versions in `pythonPackages`. Build metadata does not single out numpy or pandas. The shared parser accepts legacy `components` files without rewriting them, retains their consistency checks, and returns only the canonical flat fields. Mixed formats are rejected. The build digest includes the assembly format so changed manifest bytes invalidate payload reuse.
The Python runtime wheel carries target-native CPython and Office libraries plus external skills in a sibling `<platform>-<arch>/` resource directory. Short platform directory names leave room for Python extension and dependency DLL paths after installation on Windows. Bundling supplies defaults; the skill registry still lets project, custom and user skills override same-name bundled entries, and profile patches can disable or replace Office skills without replacing Python. This retains configurable workflows while eliminating a separate environment installation step. Real filesystem resources let Python load native extensions and skills invoke their shared checker without startup extraction. The wheel grows by the compressed payload size; public-index size limits remain enforced by the release workflow.
## Alternatives considered
**Keep implementation in Desktop.** This leaves SDK and container builders depending on application packaging and signing code despite needing only interpreters, Office resources and a query.
**Enable Office in the common base.** A base deployment has no guaranteed bundled interpreter or resources. Explicit SDK carrier configuration makes that requirement observable without changing unrelated profiles.
**Always copy to the Harness home.** Copying duplicates image-layer content, requires a writable destination and cannot serve the intended immutable carrier deployment.
**One component field per Python library.** This duplicates the distribution map and requires schema changes for each library. A complete map records the locked package set independently of the interpreter and package-manager fields.
## Consequences
Successful tool preparation is shared by concurrent calls for one plugin lifetime. Failed preparation is retryable; disposal removes registration and waits for outstanding filesystem work. Copy mode verifies a staged replacement and restores the preceding installation after a failed publication. In-place mode performs no payload writes. Paths are absolute, and runtime metadata and declared entries are checked before results are returned.
SDK configuration changes require restart. Missing Office resources produce a startup warning and leave the optional skill provider inactive; runtime metadata and interpreter availability are validated on the first tool call. musl builds remain outside the lock. Cross-target assembly verifies archive hashes but requires execution on the target host; the Linux artifact CI lane owns native Linux checks. Unit coverage pins manifest, installation and lifecycle behavior; SDK profile tests observe optional tools, skill discovery and results, and the recorded workspace-dependencies scenario pins model-visible validation errors.
@@ -0,0 +1,37 @@
# Agent Note: Desktop 与 SDK 共享 Office 运行时
Status: implemented
[English](2026-09-17-shared-office-runtime.md) | 中文
## 问题
SDK 部署需要与 Desktop 相同的 Office 创作库,同时将解释器 payload 保存在只读容器镜像层中。由 Desktop 持有查询工具和构建器,会要求下游载体重复维护依赖锁与路径约定。
## 决策
[`tool-workspace-dependencies`](../../../../packages/skill/tool-workspace-dependencies/README.zh.md) 持有清单校验、解释器路径、安装逻辑和模型查询。它保留在 skill 组中,与 Office 工作流消费者相邻:payload 服务于这些工作流,本包不提供 workspace 实体服务。Desktop 引用本包并保留首次使用时复制安装。打包的 SDK profile 通过载体默认路径启用查询和 Office skills;`DSH_PRIMARY_RUNTIME` 覆盖该路径,空字符串表示禁用。它原位读取 payload。公共 base 与 `sdk-minimal` 不会隐式启用 Office。
[共享构建入口](../../../../scripts/primary-runtime/prepare.ts) 持有 Desktop 目标与 GNU/Linux x64/ARM64 的下载锁、解压和本机冒烟检查。Desktop 提供输出路径和发布版本,要求包含 Node.js 与 pnpm,并应用签名专用子进程环境。其他载体选择自己的输出目录,也可构建 Python-only payload。现有 `desktopVersion` 清单字段保留名字并记录载体发布版本;`payloadDigest` 区分锁定输入与组件选择。
[Desktop 第一方运行时决策](../feature/2026-09-14-desktop-primary-runtime.zh.md)继续持有 Desktop 安装与平台签名规则。本决策仅替代其中由 Desktop 独占构建器和查询实现的部分;两份记录均保持有效。
`runtime.json` 在顶层保存 Python、Node.js 和 pnpm 版本,并在 `pythonPackages` 中保存全部 Python 分发包版本。构建元数据不单独列出 numpy 或 pandas。共享解析器接受旧 `components` 文件而不改写它们,保留旧格式的一致性校验,并仅返回统一的扁平字段。混合格式会被拒绝。构建摘要包含组装格式,因此清单字节变化会使旧产物身份失效。
Python runtime wheel 将目标平台的 CPython、Office 库和外部 skills 放在可执行文件同级的 `<platform>-<arch>/` 资源目录中。较短的平台目录名为 Windows 安装后的 Python 扩展与依赖 DLL 路径保留长度空间。打包提供默认资源;skill 注册表仍允许项目、自定义目录和用户 skills 覆盖同名随包条目,profile patch 可在保留 Python 的同时禁用或替换 Office skills。这样既保留工作流配置能力,又省去单独安装环境的步骤。真实文件系统资源允许 Python 加载原生扩展,并让 skills 调用共用检查脚本,无需启动时解包。wheel 体积增加一个压缩 payload;发布流程仍执行公开索引的大小限制检查。
## 考虑过的替代方案
**将实现保留在 Desktop。** SDK 和容器构建器只需要解释器、Office 资源和查询,却仍会依赖应用打包及签名代码。
**在公共 base 中启用 Office。** base 部署不保证提供内置解释器或资源。显式 SDK 载体配置可以让该要求可见,并保持其他 profile 的行为。
**总是复制到 Harness home。** 复制会重复镜像层内容、要求目标可写,也无法满足预期的不可变载体部署。
**为每个 Python 库增加组件字段。** 这会重复分发包版本表,并使新增库要求修改 schema。完整版本表将锁定包集合与解释器、包管理器字段分开记录。
## 影响
同一插件生命周期内的并发调用共享成功的工具准备结果。准备失败后可重试;卸载移除工具注册并等待未完成的文件系统工作。复制模式验证暂存替换内容,并在发布失败后恢复原安装。原位模式不写入 payload。路径为绝对路径,返回结果前会检查运行时元数据和所声明的条目。
SDK 配置变更需要重启。缺少 Office 资源会产生启动警告,并使可选的 skill 提供方保持未激活;运行时元数据和解释器是否存在在首次工具调用时校验。musl 构建仍不在锁文件范围内。跨目标组装校验归档哈希,但需要在目标主机执行;Linux 产物 CI 负责本机 Linux 检查。单元覆盖固定清单、安装与生命周期行为;SDK profile 测试观察可选工具、skill 发现与结果,workspace-dependencies 录制场景固定模型可见的校验错误。
@@ -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/2026-09-17-workspace-file-binary-transfer.md
2026-09-17-workspace-file-binary-transfer.md: b419c8bd15f51246825ceddf2905ec055bbc70ae
2026-09-17-workspace-file-binary-transfer.zh.md: 9cf3d7074c3d536a5c8a39732231c6996d6f6ead
@@ -0,0 +1,31 @@
# Agent Note: Binary workspace file transfer
Status: implemented
English | [中文](2026-09-17-workspace-file-binary-transfer.zh.md)
## Problem
Document previews need native bytes. Encoding file contents as base64 adds about one third to the payload before compression and requires browser decoding. A separate file Fetch route also duplicates the Gateway's method dispatch, Session lookup, and error handling.
## Decision
The [Remote method architecture](2026-08-02-typert-remote-method-calls.md) owns method dispatch and lifetime; this note defines binary success encoding and Client validation. The [workspace file service](2026-09-05-workspace-files-service.md) exposes one binary `readBytes` Remote. Its required options object independently selects a byte `range` and a `baseFile` for relative target resolution. Without `range`, it reads a complete file. This avoids separate method names for orthogonal choices while preserving bounded `fs.readBytes` and `fs.readByteRange` reads.
Typert recursively recognizes `Uint8Array` in unary result types, including root values, optional fields, and containers. Generated result codecs provide `encode()` for subtrees whose types can contain bytes and `decode()` to validate the reconstructed value without iterating, copying, or freezing byte payloads; Client declarations narrow each byte buffer to `ArrayBuffer`. Gateway executes the generated encoder, keeps strict JSON values unchanged, and performs runtime byte detection for source-mode calls. It returns JSON-compatible metadata and result-relative byte attachments to Connection. Connection owns standard `FormData` parts and preserves the RPC envelope and correlation id in JSON metadata; it does not inspect business values or depend on Typert reflection. An attachment table associates byte parts with paths through the JSON result; field names remain business-owned. The Client restores `Uint8Array` views over `Blob.arrayBuffer()` and Gateway delegates validation to the generated decoder. Ordinary results and errors retain JSON responses. Gateway owns method dispatch, Session lookup, cancellation, result projection, and mount lifetime; feature code needs no multipart reader or dedicated route.
## Alternatives considered
- Separate `readAll`, range, and related-file methods mix independent path and extent choices. One options object permits their combination without adding another method.
- A workspace-only Fetch route keeps binary transport out of Remote but requires feature-specific dispatch, Session lookup, and error decoding.
- A whole-result binary marker or a reserved `data` field couples transport to one business result. Type-directed codecs and attachment paths support independently named and nested fields without adding result-wide categories.
- A custom metadata prefix needs owned framing and length validation. FormData delegates those mechanics to the platform, at the cost of Blob copies.
- Metadata HTTP headers constrain long Unicode paths. Multipart fields keep paths in the body; HTML stays in an isolated Blob iframe rather than a navigable Host response.
## Consequences
Byte ranges retain only the requested bytes under `maxBytes`, independently of `maxFileBytes`. Complete reads retain a bounded whole file; this is not a streaming preview or zero-copy optimization. On HTTP disconnect, the bridge aborts the request signal and drains materialized response chunks without writing to the socket: cancelling Node multipart bodies can race their producer and cause an unhandled `ERR_INVALID_STATE` rejection. Office authorization probes use the same Remote with a one-byte range; converted PDFs remain base64 in the separate Office Remote and use its bounded decoder. Plugin callers migrate to `readBytes` options and stop decoding its data as base64. Binary arguments, events, and stream items remain unsupported; recursive types require acyclic runtime values. No Session event or persistence format changes.
## Validation
Workspace tests cover complete and ranged reads, combined base-file resolution, limits, Session lookup, cancellation, and binary RPC round trips. Connection and Gateway tests cover nested and multiple attachments, framing, correlation, malformed multipart, and cancelled or withdrawn calls; Typert tests cover recursive result codecs, Client declarations, and rejected binary inputs or streams. The built browser scenario checks raw PNG bytes plus HTML, PDF, image, and text previews. Local benchmarks compare full and ranged reads using identical synthetic files, and distinguish transport/decoding from the complete read path.
@@ -0,0 +1,31 @@
# Agent Note: 工作区文件二进制传输
Status: implemented
[English](2026-09-17-workspace-file-binary-transfer.md) | 中文
## 问题
文档预览需要原生字节。将文件内容编码为 base64 会在压缩前使载荷增加约三分之一,并要求浏览器解码。独立的文件 Fetch 路由还会重复 Gateway 的方法分发、Session 查找和错误处理。
## 决策
[Remote 方法架构](2026-08-02-typert-remote-method-calls.zh.md)负责方法分发与生命周期;本记录规定二进制成功响应的编码及 Client 校验。[工作区文件服务](2026-09-05-workspace-files-service.zh.md)提供单一的二进制 `readBytes` Remote。其必填选项对象独立选择字节 `range` 和用于相对目标解析的 `baseFile`。未传 `range` 时读取完整文件。这避免为正交的选择设置不同方法名,同时保留有界的 `fs.readBytes` 和 `fs.readByteRange` 读取。
Typert 递归识别一元结果类型中的 `Uint8Array`,包括根值、可选字段和容器。生成的结果 codec 提供 `encode()` 处理类型可能包含字节的子树,并通过 `decode()` 校验还原后的值,不逐字节遍历、复制或冻结字节载荷;Client 声明将每个字节缓冲区限定为 `ArrayBuffer`。Gateway 执行生成的 encoder,保持 strict JSON 值不变,并为源码模式调用执行运行时字节识别;随后向 Connection 返回 JSON 兼容元数据及相对于结果的字节附件。Connection 负责标准 `FormData` 部分,并在 JSON 元数据中保留 RPC 信封和关联 id;它不检查业务值,也不依赖 Typert 反射。附件表将字节部分与 JSON 结果中的路径关联,字段名仍由业务拥有。Client 恢复基于 `Blob.arrayBuffer()` 的 `Uint8Array` 视图,Gateway 将校验交给生成的解码器。普通结果与错误继续使用 JSON 响应。Gateway 负责方法分发、Session 查找、取消、结果投影及挂载生命周期;功能代码无需 multipart 读取器或专用路由。
## 考虑过的替代方案
- 分立的 `readAll`、范围和关联文件方法混合了路径与读取范围两个独立选择。单一选项对象支持组合,无需再增加方法。
- 工作区专用 Fetch 路由让二进制传输独立于 Remote,但需要功能专属的方法分发、Session 查找与错误解码。
- 整体结果的二进制标记或保留的 `data` 字段会将传输与某一种业务结果耦合。按类型生成的 codec 和附件路径支持独立命名及嵌套字段,无需新增整体结果分类。
- 自定义元数据前缀需要自有的帧格式与长度校验。FormData 将这些机制交给平台,代价是 Blob 复制。
- 元数据 HTTP 头会限制较长的 Unicode 路径。Multipart 字段将路径保留在正文中;HTML 仍置于隔离的 Blob iframe,不作为可导航的 Host 响应。
## 影响
字节范围仅保留 `maxBytes` 内的请求字节,不受 `maxFileBytes` 约束。完整读取保留有上限的整个文件;这不是流式预览或零复制优化。HTTP 连接断开后,桥接器中止请求信号,并排空已准备的响应块,不再写入 socket:取消 Node multipart 响应体可能与其生产方竞争,导致未处理的 `ERR_INVALID_STATE` 拒绝。Office 授权探测使用同一 Remote 的单字节范围;转换后的 PDF 仍通过独立的 Office Remote 传输 base64,并使用其有界解码器。插件调用方迁移到 `readBytes` 选项,并停止将其 data 按 base64 解码。仍不支持二进制参数、事件和流条目;递归类型要求运行时值无环。不改变 Session 事件或持久化格式。
## 验证
Workspace 测试覆盖完整与分段读取、基准文件解析组合、上限、Session 查找、取消以及二进制 RPC 往返。Connection 和 Gateway 测试覆盖嵌套及多附件、帧格式、关联关系、畸形 multipart,以及取消或撤销的调用;Typert 测试覆盖递归结果 codec、Client 声明,以及对二进制输入和流的拒绝。构建后的浏览器场景验证原始 PNG 字节以及 HTML、PDF、图片和文本预览。本机 benchmark 使用相同合成文件比较完整与分段读取,并区分传输/解码和完整读取路径。
@@ -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/2026-09-18-volatile-config-references.md
2026-09-18-volatile-config-references.md: afaa6affdc3d18da182ec069126033cb55393b72
2026-09-18-volatile-config-references.zh.md: 3d8800cb92dec4da80b698cb8b3c073331ed266f
@@ -0,0 +1,33 @@
# Agent Note: Volatile configuration references
Status: implemented
English | [中文](2026-09-18-volatile-config-references.zh.md)
## Problem
Replacing an entire plugin to change a value also disposes its services and effects. Separate settings subscriptions require each consumer to implement another source of live configuration.
## Decision
Schemastery declares live fields with `.volatile()`, preserving node types and form metadata while returning stable references to immutable snapshots. Cosmokit owns the shared protocol. General comparison treats two volatile references as equal. Loader calculates raw config differences using schema-declared volatile paths and retains raw config for subsequent activation. When the raw diff reports no ordinary change but the raw config differs, Loader parses and validates the candidate through the fiber's `internal/config` hook, compares ordinary effective values, commits references when those values still match, and notifies the owning fiber; every update path, including profile reconciliation, HMR Include refreshes and direct `entry.update()` calls, reaches this step through `Entry.update`. Fiber does not own this update policy, and HMR supplies file watching.
Only volatile changes preserve the instance. Ordinary-field changes use the existing remount lifecycle and do not update the old references. An instance-local `loader/volatile-update` event reports paths after the complete candidate is validated and all references are committed. It uses ordinary `emit`; listener completion is not a configuration commit condition. References can be retained, while values may only be captured for one operation. Missing optional values retain a readable reference.
Config schemas own the field types, defaults, roles, and live-update declaration needed by schema-driven settings forms. Existing settings consumers remain unchanged; automatic form generation and their migration are separate work. The [configuration guide](../../../../docs/cordis-tutorial/05-config.md) owns plugin-author usage.
## Alternatives considered
**Restart for every configuration change.** This releases resources that remain valid when only operation-time parameters change.
**Add another settings-specific live-value API.** It duplicates runtime update semantics and makes plugins depend on settings to consume ordinary configuration.
**Wrap values only during mounting.** Direct schema parsing would return a different kind of value from the plugin's config, requiring separate consumer types.
## Consequences
Schemas reject independent volatile references inside dynamic containers; entire objects and arrays can be volatile values. Immutable snapshots accept plain data. Simplification unwraps references, and shared symbols support ESM/CJS copies. Group/Include retain dispatch while volatile-only target updates bypass `internal/update`. Direct `fiber.update()` retains its semantics, including custom handling, saving and `noSave`. Loader commits live volatile updates itself; no separate plugin is required.
Schema metadata keeps comparison independent of parsing: the diff does not execute expressions, validators or config hooks. Object nodes compare missing or null values using their declared defaults, even without volatile descendants; this avoids remounts for equivalent object defaults. Ordinary expression edits retain raw comparison even when they resolve to equal values. Loader compares parsed ordinary values, including URLs by normalized address. If an ordinary expression produces a different value, Loader applies the ordinary remount instead of committing; strict comparison treats class instances other than URL, Date and RegExp by identity, so an ordinary transform that builds a fresh instance per parse always remounts. Listener failures during the notification are logged and do not fail the update. Loader validates volatile candidates before committing; invalid edits remain in raw config but leave running references unchanged until a later activation.
The schema and lifecycle regression suite covers direct updates, invalid candidates, instance isolation, dependency replacement, real file watching, and save/reload behavior. The HMR module suite covers replacement after live config updates. These tests do not claim settings migration or automatic form rendering.
@@ -0,0 +1,33 @@
# Agent Note: Volatile 配置引用
Status: implemented
[English](2026-09-18-volatile-config-references.md) | 中文
## Problem
为了改变一个值而替换整个插件,也会销毁插件的服务和副作用。独立的 settings 订阅要求每个消费者额外实现一套动态配置来源。
## Decision
Schemastery 使用 `.volatile()` 声明动态字段,保留节点类型和表单元数据,解析结果是具有不可变快照的稳定引用。共享协议位于 Cosmokit。通用比较将两个 volatile 引用视为相等。Loader 根据 schema 声明的 volatile 路径计算原始配置差异,并保留原始配置供后续激活使用。原始 diff 未报告普通字段变化但原始配置有差异时,Loader 经 fiber 的 `internal/config` 钩子解析并校验候选配置,比较普通字段的有效值,在这些值仍匹配时提交引用并通知所属实例;所有更新路径,包括 profile 重载、HMR 的 Include 刷新和直接调用 `entry.update()`,都经 `Entry.update` 到达这一步。Fiber 不承担这项更新策略,文件监视由 HMR 提供。
仅 volatile 变化保留实例。普通字段变化沿用现有重新挂载生命周期,不更新旧引用。实例局部的 `loader/volatile-update` 事件在完整候选配置通过校验且所有引用提交后报告路径。事件使用普通 `emit`;监听器完成不是配置提交条件。可以保存引用,而值只能在单次操作中保存。缺省可选值仍具有可读取的引用。
Config schema 负责提供基于 schema 的 settings 表单需要的字段类型、默认值、角色和动态更新声明。现有 settings 消费者保持不变;自动表单生成及其迁移是独立工作。[配置指南](../../../../docs/cordis-tutorial/05-config.zh.md)负责插件作者的使用说明。
## Alternatives considered
**每次配置变化都重启。** 当只有操作时读取的参数变化时,这会释放仍然有效的资源。
**增加另一套 settings 专用动态值 API。** 这会重复运行时更新语义,并使插件为了读取普通配置而依赖 settings。
**仅在挂载时包装值。** 直接解析 schema 与插件配置将返回不同种类的值,需要独立的消费者类型。
## Consequences
Schema 拒绝动态容器内部的独立 volatile 引用;整个对象和数组可以作为一个 volatile 值。不可变快照接受普通数据。配置简化解包引用,共享 symbol 支持 ESM/CJS 副本互操作。Group/Include 保留分发,目标插件的仅 volatile 更新不进入 `internal/update`。直接调用 `fiber.update()` 保留原语义,包括自定义处理、保存及 `noSave`。动态 volatile 更新由 Loader 自己提交,不需要额外插件。
Schema 元数据使比较独立于解析:diff 不执行表达式、校验器或配置钩子。对象节点的缺省值或 null 按声明的默认值比较,即使没有 volatile 后代字段;等价的对象默认值不触发重新挂载。普通表达式即使求值结果相等,修改时仍按原始值比较。Loader 比较解析后的普通值,其中 URL 按规范化地址比较。普通表达式产生不同值时,Loader 改为按普通流程重新挂载而不提交;严格比较对 URL、Date、RegExp 之外的类实例按身份比较,因此每次解析都产生新实例的普通 transform 字段总是走重新挂载。通知阶段监听器抛错只记录日志,不使更新失败。Loader 在提交前校验 volatile 候选;非法编辑保留在原始配置中,但在后续激活之前不改变运行中的引用。
Schema 和生命周期回归测试覆盖直接更新、无效候选、实例隔离、依赖替换、真实文件监视及保存后重新加载。HMR 模块测试覆盖动态配置更新后的代码替换。这些测试不代表 settings 迁移或自动表单渲染已经完成。
@@ -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/2026-09-19-remote-duplex-stream.md
2026-09-19-remote-duplex-stream.md: 0f1f6a575f6871d37822c5829fe4531261b8e95b
2026-09-19-remote-duplex-stream.zh.md: d3982aa433aa3ce84b945bb5831cef8064472853
@@ -0,0 +1,597 @@
# Agent Note: Remote duplex stream: one kind of stream, an uplink channel, and an invocation context
Status: implemented
English | [中文](2026-09-19-remote-duplex-stream.zh.md)
## Problem
### Current state
`dsh-api-gateway` multiplexes every Typert Remote stream over one WebSocket at the fixed path `/api/remote.mux` (`packages/api/gateway/src/stream-protocol.ts`). A Host method decorated with `@Remote({ mode: 'stream' })` that returns an `Iterable` or `AsyncIterable` becomes one Host-to-Client logical stream; the optional final `signal: AbortSignal` is the only reserved parameter, never enters the wire arguments, and is appended by the Gateway after the decoded business parameters (`packages/api/gateway/src/index.ts`). The generated method the Client receives returns a bare `AsyncIterable`.
There are only five wire frames:
| Direction | Frame | Definition |
| --- | --- | --- |
| Client to Host | `{ type: 'open', streamId, endpoint, payload }` | `stream-protocol.ts` |
| Client to Host | `{ type: 'cancel', streamId }` | `stream-protocol.ts` |
| Host to Client | `{ type: 'item', streamId, value? }` | `stream-protocol.ts` |
| Host to Client | `{ type: 'error', streamId, error: { code, message, details } }` | `stream-protocol.ts` |
| Host to Client | `{ type: 'end', streamId }` | `stream-protocol.ts` |
The Client can send only open and cancel. The parser `parseRemoteStreamClientMessage` (`stream-protocol.ts`) throws on any other frame, and a Host connection that receives one closes the whole socket with 1008 (`stream-server.ts`). In short, the Gateway today supports only "a backend iterable sent to the frontend". Every feature that needs the frontend to send data to the backend continuously takes its own detour.
A Host method also has no notion of who initiated this call: `InvokeRemoteRequest = { namespace, method, args, signal }` (`packages/api/gateway/src/types.ts`) has no caller slot, `RemoteStreamOpener` has the signature `(endpoint, payload, signal)` (`stream-server.ts`), and the WebSocket checks the cookie once, at the handshake.
### Three unrelated uplink mechanisms
| Feature | Uplink path | Location | Cost |
| --- | --- | --- | --- |
| Web terminal keystrokes | One unary RPC `terminal.write(agent, id, attachmentId, data)` per xterm `onData`; the Client serializes them through a promise chain and caps its own byte budget | `packages/api/terminal-controller/src/client/model.ts`; Host `src/index.ts` | One HTTP round trip and one authentication per keystroke; two carriers correlated by hand through `attachmentId`; `inputFull` when the budget is exhausted |
| File upload | A `Blob` or `ReadableStream` goes to the native `POST /api/session/uploadFileBinary` outside the Remote layer, with `duplex: 'half'` | `packages/client/file-upload/src/client/runtime.ts`; route `src/index.ts` | Bypasses Typert descriptors; no type projection |
| Answers to approvals and questions | The Host sends waterfall frames down the `$events` stream, the Client answers through a separate unary RPC `$events/result`, correlated by `clientId` plus `eventId` | `packages/api/gateway/src/client/remote-events.ts`; Host `index.ts` | A second correlation key; `index.ts` records the race this creates |
These three mechanisms solve the same problem: the Client wants to write into a logical stream that is already open.
### Concrete costs
- **Aligning two carriers.** The terminal uses `attachmentId` to correlate the `follow` stream with the `write` RPC and to decide who owns input. Approvals use `clientId` to correlate the `$events` stream with the `$events/result` RPC.
- **No ordering guarantee.** A unary RPC travels over HTTP and is unordered relative to the WebSocket downlink; the terminal can preserve order only by serializing writes on the Client.
- **Separate authentication and lookup per uplink.** Every `write(agent, id, attachmentId, data)` resolves `agent`, looks up the terminal, and validates the attachment.
- **No EOF.** The Client cannot express "I am done writing"; a half-close needs yet another RPC.
- **Cancellation does not cover both halves.** Cancelling the downlink stream does not cancel queued uplink writes, and vice versa.
- **Backpressure implemented three times.** The downlink pulls the iterator at the pace of the socket write callback (`stream-server.ts`); the terminal uses a bounded follower that makes a slow consumer fail explicitly (`terminal-controller/src/stream.ts`); the uplink relies on the Client capping `maxInputBytes` itself.
- **The documentation is already stale.** `docs/api-gateway.md:160` still says Remote handles one request and one result, a sentence that lags the stream mode the README records.
### Relation to background jobs
The process-shaped duplex stream face of background jobs (archived as a draft PR on branch `worktree/job-stdio-stream-pair-archive`) models a job as a pair of mirrored stream faces; for the browser to attach to it, one Remote stream must carry both a downlink and an uplink. It is one motivation for this Note but not in its scope: whether jobs need a dedicated read-write stream is an architecture question for the jobs line itself. This Note covers only the transport layer.
### Goals
1. The transport layer has one kind of stream, duplex by nature: downlink `item` / `end` / `error`, uplink `item` / `end`, one `streamId`, one generation.
2. The uplink type and the downlink type are written in the same place: the method's return type.
3. A Host method takes its uplink from the invocation context; the invocation context also carries the caller (the Peer).
4. The Client receives a handle that can read, write, and close, not a bare `AsyncIterable`.
5. Upper-layer protocols (snapshot, journal, any business-defined stream) stay out of the transport layer.
## Decision
### In one sentence
Every Remote stream accepts a Client uplink. The method's return type `RemoteStream<Out, In = never>` declares the downlink item type and the uplink item type together; the generator derives two codecs from it. A Host method takes the uplink iterator through `this.ctx.invocation.uplink<In>()` and learns the caller through `this.ctx.invocation.peer`. The generated Client method returns a `RemoteStreamHandle<Out, In>`: `for await` reads the downlink, `send` / `end` write the uplink, `dispose` closes. Reopening after disconnection, cursor resumption, and baseline validation all belong to upper-layer protocols.
```
Client Host
───────────────────────────────── ───────────────────────────────────────────
const s = remote.job.attach(req, signal)
│ open { streamId, endpoint, payload } ───────▶ opener(endpoint, payload, uplink, peer, control)
│ │ prepareInvocation
│ s.send(v1) → item { streamId, v1 } ───────▶ inbox.push(v1) │ invocation = { request, service, peer, signal, uplink() }
│ s.send(v2) → item { streamId, v2 } ───────▶ inbox.push(v2) │ receiverContext.extend({ invocation }).get(service)
│ s.end() → end { streamId } ───────▶ inbox.end() │ method(...args, signal)
│ ◀─────── item { streamId, o1 } │ for await (v of this.ctx.invocation.uplink()) …
│ for await (o of s) … ◀─────── item { streamId, o2 } │ yield o …
│ s.dispose() → cancel { streamId } ───────▶ control.abort() │ signal aborts; uplink.return()
│ ◀─────── end | error
```
### Terms
| Term | Meaning |
| --- | --- |
| Logical stream | One call identified by one `streamId`, spanning `open` to the terminal frame |
| Generation | One physical lifetime of a logical stream. The transport layer knows exactly one generation; disconnection is failure, and reopening belongs to the upper layer |
| Downlink | The Host-to-Client `item` sequence, terminated by `end` or `error` |
| Uplink | The Client-to-Host `item` sequence, terminated by `end` (half-close) or ended together with `cancel` |
| Half-close | One direction has ended while the other remains open |
| Item | One business value in one direction; every item is validated by the codec independently |
| inbox | The bounded uplink queue the Host keeps per logical stream |
| Peer | The party the connection layer admitted, one `PeerScope`; this Host has exactly one Peer, the operator |
| Invocation context | `RemoteInvocation`: the request, receiving service, Peer, cancellation signal, and uplink entry of this call |
### Type signatures
```text
// @deepseek-ai/dsh-typert-protocol
/**
* One Remote stream. Host face: the method returns it, and at runtime it is AsyncIterable<Out>.
* On the Client face the generated method returns RemoteStreamHandle<Out, In>; each name has exactly one meaning.
* In is the item type the Client may send uplink; the default never means the method reads no uplink.
*/
export type RemoteStream<Out, In = never> = AsyncIterable<Out> & { readonly [STREAM_UPLINK]?: In }
// Host
@Remote({ mode: 'stream' })
async *attach(request: JobAttachRequest, signal: AbortSignal): RemoteStream<JobFollowFrame, JobInputFrame> {
const uplink = this.ctx.invocation.uplink<JobInputFrame>()
void (async () => { for await (const frame of uplink) … })()
yield …
}
// Generated Client signature: same parameters, returns the handle
attach(request: JobAttachRequest, signal?: AbortSignal): RemoteStreamHandle<JobFollowFrame, JobInputFrame>
// Client usage
const stream = remote.job.attach(req, signal)
for await (const frame of stream) …
stream.send(frame) // typed as JobInputFrame
stream.end()
stream.dispose()
```
`RemoteStream` is an alias rather than an interface, deliberately: if the transport layer later takes a Node stream form, only the alias definition and the adapters at both ends change; method signatures, descriptors, and upper-layer protocols stay untouched.
### Host face
```text
/** The context of this Remote call. A method reads it through this.ctx.invocation. */
export interface RemoteInvocation {
readonly request: {
readonly namespace: string
readonly method: string
readonly args: Readonly<Record<string, unknown>>
}
/** Cordis service key of the receiving service. */
readonly service: string
/** The Peer that initiated the call. In-process carriers and unadmitted calls are the operator. */
readonly peer: PeerScope
/** Carrier cancellation: Client cancel, socket closure, uplink failure. */
readonly signal: AbortSignal
/**
* The uplink items of this call. Can be taken once; the second call throws. With an uplink codec
* in the descriptor each item is decoded to In; without one, items are delivered as unknown with
* only a JSON-safety check.
* The iteration ends after the Client's end; when the method ends the downlink, the Gateway calls
* its return() and drops unconsumed items.
* The generic In is only the caller's type assertion; runtime decodes per the descriptor and does not cross-check.
*/
uplink<In = unknown>(): AsyncIterable<In>
}
declare module '@deepseek-ai/cordis' {
interface Context {
/** The Remote call this Context was derived for; undefined on a Context not derived from a Remote call. */
readonly invocation: RemoteInvocation | undefined
}
}
```
Rules:
1. Any `mode: 'stream'` method may call `this.ctx.invocation.uplink()`; a unary method may too, but its stream ends when the method returns, so uplink items are readable only while the method runs.
2. While `uplink()` has not been called, uplink items accumulate in the inbox until the limit is exceeded (see Backpressure) or the stream ends and they are dropped. This is deliberate: a method that does not read the uplink need not know the uplink exists.
3. When `In` defaults to `never`, the descriptor has no uplink codec; the Client handle's `send` has type `never`, so nothing can be sent at compile time. If frames still arrive at runtime (from a hand-built client), they are delivered as `unknown`.
4. `this.ctx.invocation` is `undefined` on a Context not derived from a Remote call; a method that reads it handles the optional as TypeScript requires, or asserts it on the method's first line. The accessor for that property is registered once when the `TypertRemoteService` base class is constructed, so any composition that installs a Remote service can read it.
#### How the invocation context reaches the method
The mechanism relies on two existing Cordis behaviors:
- `ctx.extend(meta)` is `Object.create(parent)` plus own properties (`vendor/cordis/src/context.ts:99-107`): no registration, no fiber, nothing to release, so deriving one per call costs nothing.
- `ctx.get(service)` returns the `createTraceable(ctx, service)` proxy (`vendor/cordis/src/reflect.ts:233-234`, `utils.ts:117-125`). Reading the `ctx` property returns the accessing Context directly (`utils.ts:175`); reading a method wraps it in `createShadowMethod`, whose `apply` replaces `this` with a shadow receiver whose `ctx` is `ctx.extend({ [symbols.shadow]: origin })` (`utils.ts:149-163`).
- The Gateway therefore writes `receiverContext.extend({ invocation }).get(descriptor.service)` to obtain a per-call view and calls `Reflect.apply(method, callReceiver, args)`; inside the method body, `this.ctx.invocation` reads this call through the prototype chain. The deferred execution of an async generator body is unaffected, because `this` was bound to the shadow object at call time.
### PeerScope
```text
/** The opaque identity of one Peer. */
export type PeerId = Branded<'PeerId'>
/** One Peer's session on this Host: opened and released by the admitting side; ctx owns connection-lifetime registrations. */
export interface PeerScope {
readonly id: PeerId
readonly ctx: Context
dispose(): Promise<void>
}
```
This Host has exactly one Peer: the operator. `@deepseek-ai/dsh-client-connection` creates it with `createScope(connectionCtx, peer)` when the service applies, the same mechanism an Agent uses to create its own scope: the Peer object is the ScopeKey, `peer.ctx` carries connection-lifetime effects, and `dispose()` runs with the Connection's release and quiets the fiber. It is exposed as `connection.operator`.
| Member | Semantics |
| --- | --- |
| `connection.operator: PeerScope` | The only Peer; lives and dies with the Connection |
| `connection.admit(request): PeerAdmission` | Runs the existing `requestRejection`: returns `{ rejection: 401 \| 403 }` when refused, otherwise `{ peer: operator }`. Future admitters plug in here |
The Gateway's `handleUpgrade(req, socket, head, peer)` binds an opener with a fixed Peer to that socket and registers socket closure in `peer.ctx.effect`; when the scope is already inactive, it closes the socket directly with 1001 instead. When `InvokeRemoteRequest.peer?` or the `peer` parameter of `RemoteStreamOpener` is absent, the answer is the operator, so existing in-process callers change nothing; in a composition without a Connection, the operator stand-in the Gateway builds for itself is likewise created through `createScope`, with the same `dispose()` contract. Looking up a Peer by id, associating a carrier with a second Peer, and Peer open and close events all wait until the first consumer that needs multiple Peers appears.
The Peer carries no access model: who the Peer is and what it may do are attached by business plugins on their own through `peer.ctx`.
### Client face
```text
/** The handle a generated method returns on the Client face. A different name from the Host's RemoteStream, exported from the same entry point. */
export interface RemoteStreamHandle<Out, In> extends AsyncIterable<Out> {
/** Send one uplink item. Throws on a value that is not lossless JSON, after termination, or after end. */
send(item: In): void
/** Half-close the uplink: sends the end frame. Idempotent. */
end(): void
/** Cancel the whole logical stream: sends the cancel frame (when no terminal frame has arrived); the downlink iterator ends quietly. */
dispose(): void
}
```
- The handle represents one generation. When the carrier is lost, `for await` fails with `RemoteStreamCarrierError` and the handle ends there; whether to reopen, and with which arguments, belongs to the upper-layer protocol.
- `send` is synchronous: it runs the `isRemoteJsonValue` check before enqueuing and throws synchronously on a value that is not lossless JSON. The browser's `WebSocket.send` has no write callback, so throttling sends would be meaningless; the Host-side inbox limit is the only backpressure point.
- When a terminal frame (`end` / `error`) arrives, the mux client stops the pump and closes the uplink queue immediately on the receive path, so later `send` / `end` throw without waiting for the consumer's next read.
- `dispose()` calls `return()` on the carrier iterator, so the `cancel` frame is sent even when nobody is reading the downlink; afterwards the handle's `next()` ends immediately and buffered items are dropped. A consumer that `break`s out of `for await` early is equivalent to `dispose()`.
- The `$stream` supervisor, `RemoteSnapshotStream`, and `RemoteJournalStream` are upper-layer tools outside the transport layer. The object their `open` factories receive is the handle itself, with unchanged `for await` behavior; exposing `send` to these upper-layer protocols is deferred.
### Wire frames
```text
export type RemoteStreamClientMessage =
| { readonly type: 'open'; readonly streamId: string; readonly endpoint: string; readonly payload: unknown }
| { readonly type: 'item'; readonly streamId: string; readonly value?: unknown } // new: uplink item
| { readonly type: 'end'; readonly streamId: string } // new: uplink half-close
| { readonly type: 'cancel'; readonly streamId: string }
/** Frames sent by the Host are unchanged. */
export type RemoteStreamServerMessage =
| { readonly type: 'item'; readonly streamId: string; readonly value?: unknown }
| { readonly type: 'error'; readonly streamId: string; readonly error: RemoteStreamFailure }
| { readonly type: 'end'; readonly streamId: string }
```
Parsing rules: `item` has exactly the keys `type` and `streamId` plus an optional `value`, and `value` must be a lossless JSON value (`isRemoteJsonValue`); `end` has exactly the keys `type` and `streamId`. Any other frame is a frame-format violation, and the socket closes with 1008. No binary frames are introduced.
### Logical stream state machine
#### Host side
```
open frame
(none) ───────────────────▶ opening ──── opener resolves ────▶ running
│ │
│ item frame: inbox.push │ item frame: inbox.push
│ end frame: inbox.end │ end frame: inbox.end (uplink half-close)
│ │ method yield: sends item frame
│ cancel frame / socket close ────▶│ cancel frame / socket close: control.abort
▼ ▼
aborted ◀───────────────────────── finished (sends end or error frame)
```
`item` frames that arrive during `opening` enter the inbox: `receive()` is synchronous and creates the `ActiveStream` and its inbox while handling the `open` frame, whereas the opener is asynchronous. Items the Client sends immediately after `open` are not lost.
#### Client side
```
call generated method ──▶ await socket ──▶ send open ──▶ handle usable
│ send(v): sends item; end(): sends end
│ downlink: inbox.next() → yield; end → ends; error → throws
│ downlink terminates first: send/end throw
│ dispose or caller signal abort: sends cancel (when no terminal frame has arrived)
```
### Half-close and termination
| Uplink | Downlink | Result |
| --- | --- | --- |
| Client `end` | Still open | The Host's `uplink()` iteration ends; the method keeps producing. This is the stdin EOF form |
| Still open | The Host method finishes and sends `end` | The Client handle terminates, and later `send` / `end` throw; the Host calls `return()` on the `uplink` iterator and drops unconsumed items |
| Still open | The Host method throws and sends `error` | Same as above; the Client downlink fails with `RemoteError` |
| Client `dispose` or signal abort | Any | Sends `cancel`; the Host runs `control.abort()`, `cancellableStream` closes `uplink` first and then calls `return()` on the method iterator; no terminal frame is sent |
| Socket closes | Any | The Host aborts every stream and awaits `done`; on the Client every stream fails with `RemoteStreamCarrierError` |
| `item` received after `end` | Any | The stream fails with a `gateway/protocol` error frame and aborts; the socket stays open |
| `item` / `end` with an unknown `streamId` | Any | Ignored, like `cancel`: after the Host ends a stream and deletes its id, uplink frames still in flight from the Client are normal and must not take down the other streams on the same socket |
### Ordering guarantees
Frames on one socket arrive in total order: a logical stream's `open`, `item`…, `end` reach the Host in send order, and uplink items enter `uplink()` in the order of `send`. There is no cross-direction ordering guarantee between uplink and downlink; a protocol that needs request-response pairing carries its own sequence numbers.
### Backpressure
WebSocket has no per-stream window like HTTP/2: one connection is one serial write chain, and a slow stream holds up the other streams on the same connection. This property of the downlink is unchanged. For the uplink:
**A bounded inbox plus whole-stream failure.** The Host meters each logical stream's inbox in UTF-8 bytes of its frames, with the limit set by the Gateway config `streamInboxBytes` (default 262144) and applied to every stream. On overflow the whole logical stream fails with `gateway/uplink-overflow`: the downlink breaks with it, an `error` frame goes to the Client, and the `uplink()` iterator ends with the same error. No data is discarded; a method that does not read its uplink while the Client keeps sending is a usage error, and the failure exposes it explicitly.
Credit frames wait until a consumer that needs sustained high-volume uplink appears; the inbox limit is the initial credit, and adding them is incremental.
### Security and validation
| Item | Rule |
| --- | --- |
| Authentication | Same as today: `connection.requestRejection(req)` at upgrade; descriptor and lookup-parameter validation at open. Uplink items belong to an already authenticated stream and are not authenticated again |
| Peer | Admitted at upgrade as one `PeerScope`; `invocation.peer` of every stream on that socket is that Peer |
| Item validation | With an uplink codec, every item is decoded strictly, and a failure fails the whole stream with the existing `gateway/input-invalid` (details `{ endpoint, field: 'uplink' }`); without a codec, only JSON-safety validation applies |
| Frame size | A single frame keeps the `maxPayload` of `ws`; the inbox limit bounds the accumulated total |
### In-process and worker carriers
| Carrier | Change |
| --- | --- |
| `ClientConnectionRpc.open` (`packages/client/connection/src/rpc.ts:245-250`) and `RpcStreamOpen` (`client/rpc.ts:21-25`) | Append an optional `uplink?: AsyncIterable<unknown>`; in process there are no frames and no inbox, the caller's iterator is the source of `uplink()` directly, and backpressure is the pace of the iterator itself |
| Worker tunnel (`packages/experimental/webworker-runtime`) | The page adds `stream-uplink-item` / `stream-uplink-end` frames; `serveStream` assembles them into an `uplink` handed to `seams.openStream`. Same-origin trusted boundary: no inbox limit, and items after `end` are dropped |
| `remote-mock` (`packages/test-support/remote-mock`) | `StreamHandle.uplink: AsyncIterable<unknown>`; `rpc.open` passes it through |
An in-process carrier with no `peer` means the operator.
### Relation to `$events`
`$events` is a stream on a reserved endpoint; with uplink frames, waterfall answers can become uplink items on the same `streamId`, `clientId` becomes redundant because `streamId` already identifies the generation, and `$events/result` disappears together with the race at `index.ts:527-529`. Out of scope for this Note; see Deferred.
## Type and wire definitions
### typert protocol (`packages/typert/protocol/src`)
```text
// types.ts
export type RemoteStream<Out, In = never> = AsyncIterable<Out> & { readonly [STREAM_UPLINK]?: In }
export type PeerId = Branded<'PeerId'>
export interface PeerScope { readonly id: PeerId; readonly ctx: Context; dispose(): Promise<void> }
export interface RemoteInvocation { … } // see Host face
declare module '@deepseek-ai/cordis' { interface Context { readonly invocation: RemoteInvocation | undefined } }
export interface InvocationDescriptor {
// existing fields unchanged; mode still has only 'stream'
readonly mode?: 'stream'
/** Uplink item codec, generated from In of the return type RemoteStream<Out, In>; absent when In is never. */
readonly uplink?: { readonly codec: TypertCodec }
readonly cancellation?: { readonly parameter: 'signal' }
readonly result: TypertCodec
}
```
The `Remote` decorator, `RemoteMethodOptions`, and `RemoteMethodMarker` recognize only `mode: 'stream'`.
### typert generator (`packages/typert/generator/src`)
- `model.ts`: `InvocationModel.uplink?: { boundary: RemoteBoundaryModel }`.
- `analyzer.ts` `remoteResultType`: for `mode: 'stream'`, the accepted return-type wrappers are `Iterable<Out>`, `AsyncIterable<Out>`, and `RemoteStream<Out, In?>`. `RemoteStream` is recognized the same way as the standard library's `AsyncIterable`: by symbol name plus declaring file (`types.ts` of `@deepseek-ai/dsh-typert-protocol`). The first type argument is the downlink item; when the second is present and is not `never`, an `uplink` boundary is generated under the key `${endpoint}:uplink`.
- `emitter.ts`: the descriptor literal emits `uplink: { codec }`; the generated Client signature returns `RemoteStreamHandle<Out, In>`.
- The parameter loop recognizes no parameter named `uplink`.
### Two names, one entry point
A Host method declares a stream with `RemoteStream<Out, In>`; the Client holds a `RemoteStreamHandle<Out, In>`. Both are exported from the main entry point of `@deepseek-ai/dsh-typert-protocol`, and the handle interface has no Host dependency. The generated Client contract writes the return type as `RemoteStreamHandle<Out, In>`. One name has one meaning: the `RemoteStream` that Client code obtains from the main entry point is always the declaration type and is never confused with the handle.
### Gateway Host (`packages/api/gateway/src`)
```text
// types.ts
export interface InvokeRemoteRequest {
readonly namespace: string
readonly method: string
readonly args: Readonly<Record<string, unknown>>
/** Uplink items; absent is equivalent to an iterator that ends immediately. */
readonly uplink?: AsyncIterable<unknown>
/** The Peer that initiated the call; absent means the operator. */
readonly peer?: PeerScope
readonly signal?: AbortSignal
}
export interface TypertGatewayWireStream {
open: (endpoint: string, payload: unknown, uplink: AsyncIterable<unknown>, peer: PeerScope | undefined, signal: AbortSignal) => Promise<AsyncIterable<unknown>>
failure: (error: unknown) => RemoteStreamFailure
}
// stream-server.ts
export type RemoteStreamOpener = (
endpoint: string,
payload: unknown,
uplink: AsyncIterable<unknown>,
peer: PeerScope,
control: AbortController,
) => Promise<AsyncIterable<unknown>>
// index.ts Config addition
/** Uplink inbox limit per logical stream, in UTF-8 bytes of frames (default 262144). */
readonly streamInboxBytes?: number
```
`prepareInvocation`:
```text
const args = await Promise.all(descriptor.parameters.map(parameter => this.resolveParameter(parameter, request.args, endpoint)))
const signal = request.signal ?? NEVER_ABORTED_SIGNAL
const invocation = new GatewayInvocation(
{ namespace: request.namespace, method: request.method, args: request.args },
descriptor.service,
request.peer ?? this.operatorPeer(),
signal,
new UplinkDecoder(request.uplink ?? EMPTY_ASYNC_ITERABLE, descriptor.uplink?.codec, endpoint, control),
)
const callReceiver = receiverContext.extend({ invocation }).get(descriptor.service) as object
if (descriptor.cancellation !== undefined) args.push(signal)
```
`GatewayInvocation.uplink()` returns the decoder on the first call and throws on the second. `UplinkDecoder` is a hand-written iterator: with a codec it runs `decode(codec, value, endpoint, 'uplink')` per item and, on failure, calls `control.abort(failure)` and throws; without a codec it applies only the `isRemoteJsonValue` check. The `finally` of `cancellableStream` closes the decoder first (`invocation.close()`) and then awaits the method iterator's `return()`, so a method blocked on `uplink.next()` can exit. The mode branches are the existing two, unary and stream. SRC fallback: the descriptor has no `uplink` codec, and `uplink()` delivers `unknown`.
### Gateway Host mux (`stream-server.ts`)
```text
interface ActiveStream { readonly control: AbortController; readonly inbox: UplinkInbox; done: Promise<void> }
/** Bounded uplink queue; iterated as the source of uplink(). */
class UplinkInbox implements AsyncIterable<unknown>, AsyncIterator<unknown> {
constructor(maxBytes: number, onOverflow: (error: RemoteError) => void)
push(value: unknown, frameBytes: number): void // push after ended fails the stream with gateway/protocol; overflow calls onOverflow
end(): void // idempotent
fail(error: unknown): void // ends the iterator with an error on cancel or carrier closure
next() / return() // hand-written: next races failure, return releases immediately
}
```
`receive(text)` dispatch:
```
open : exists → throw (socket 1008); otherwise create { control, inbox, done } and start the pump
item : missing → ignore; otherwise inbox.push(value, bytes)
end : missing → ignore; otherwise inbox.end()
cancel : missing → ignore; otherwise control.abort(new Error('Remote stream cancelled'))
```
`pump` passes the inbox and the socket's Peer to the opener and calls `inbox.fail(...)` in its `finally`. An abort whose reason is a `RemoteError` is sent by the pump as an `error` frame (a failure raised by the Gateway or the mux); an abort whose reason is a plain `Error` (cancel, socket closure) sends no terminal frame.
### Gateway Client (`packages/api/gateway/src/client`)
- `prepareInvocation` strips no uplink parameter; the positional parameters are the business parameters plus the optional `signal`.
- `invokeStream` returns a `RemoteStreamHandle`: it holds an uplink queue internally, `send` enqueues and a pump sends the `item` frames, `end` sends the `end` frame, and `dispose` aborts the generation signal. The downlink is still the iterator that `openRemoteStream` returns.
- `RemoteStreamMuxClient.open(endpoint, payload, signal, uplink?)`: after sending the `open` frame it starts the uplink pump, which sends one `item` per item and `end` when the source ends; when the downlink terminates or the signal aborts, it stops the pump and calls the source's `return()` (without awaiting, so a keyboard-style generator cannot hang it); when the source throws, it calls `inbox.fail(error)`, the downlink fails with that error, and `finally` sends `cancel` under the existing rules. The pump checks `this.socket === socket`, so an old generation never sends frames on a new socket.
- In-process carrier: `connection.rpc.open(channel, endpoint, payload, signal, uplink)`.
## Placement
| Package | Carries |
| --- | --- |
| `dsh-typert-protocol` (`types.ts`, `index.ts`) | `RemoteStream`, `RemoteStreamHandle`, `PeerId`, `PeerScope`, `RemoteInvocation`, the `ctx.invocation` declaration merge, `InvocationDescriptor.uplink`; the decorator recognizes only `mode: 'stream'`; `TypertRemoteService` registers the `invocation` accessor when constructed |
| `dsh-typert-registry` | Load-time validation that `uplink.codec` is a valid strict codec |
| `dsh-typert-generator` (`model.ts`, `analyzer.ts`, `emitter.ts`) | `remoteResultType` recognizes `RemoteStream<Out, In>` and generates the `uplink` boundary; the descriptor emits `uplink: { codec }`; the generated Client signature returns `RemoteStreamHandle<Out, In>`; the parameter loop accepts only business parameters and a final `signal` |
| `dsh-client-connection` (`operator-peer.ts`, `rpc.ts`, `rpc-host.ts`, `index.ts`) | The operator `PeerScope`; `connection.operator` and `admit`; the `/api` and upgrade routes obtain the Peer through `admit`; the optional `uplink` of `rpc.open` |
| `dsh-api-gateway` Host (`stream-protocol.ts`, `stream-server.ts`, `types.ts`, `index.ts`) | `item` / `end` frames; `UplinkInbox`; `Config.streamInboxBytes`; `GatewayInvocation` and `extend({ invocation })`; `UplinkDecoder`; `handleUpgrade(req, socket, head, peer)`; `operatorPeer()` |
| `dsh-api-gateway` Client (`client/index.ts`, `client/stream-client.ts`) | `invokeStream` returns the handle; the uplink pump is driven by the handle's queue |
| `dsh-webworker-runtime`, `dsh-remote-mock` | The worker tunnel's two uplink frames and the `serveStream` assembly; `StreamHandle.uplink`; the direct proxy returns the real handle |
| Documentation | The READMEs of gateway, the api group, connection, protocol, and generator; `docs/api-gateway`; the type-equiv blocks of `docs/subsystems/typert`; the config and Cordis catalogs |
## Key sequences
### Open, uplink, half-close, downlink end
```
Client caller mux client mux server gateway Host method
│ attach(req) │ │ │ │
│───────────────────▶│ open { id, ep, payload } ──────────────────▶│ create inbox │
│ s.send(v1) │ item { id, v1 } ───────────────────────────▶│ inbox.push(v1) │
│ │ │ │ prepareInvocation │
│ │ │ │ extend({ invocation }) │
│ │ │ │── method(req, signal) ─▶│
│ │ │ │ │ uplink() → v1
│ s.send(v2) │ item { id, v2 } ───────────────────────────▶│ inbox.push(v2) │ → v2
│ s.end() │ end { id } ────────────────────────────────▶│ inbox.end() │ iteration ends
│◀── yield o1 ───────│◀────────────────────── item { id, o1 } ─────│◀── yield o1 ────────────│
│◀── ends ───────────│◀────────────────────── end { id } ──────────│◀── method returns ──────│ gateway calls uplink.return()
```
### Client dispose
```
Client caller mux client mux server gateway Host method
│ s.dispose() │ │ │ │
│───────────────────▶│ stop pump; source return() (not awaited) │ │
│◀── iterator ends ──│ cancel { id } ─────────────────────────────▶│ control.abort() │
│ │ │ │ cancellableStream: invocation.close() closes uplink
│ │ │ │ iterator.return() ─────▶│ finally cleanup
│ │ │ no terminal frame │ │
```
### inbox overflow
```
Client caller mux client mux server gateway / Host method
│ s.send(vN) │ item { id, vN } ─────▶│ inbox.push: bytes > streamInboxBytes │
│ │ │ onOverflow → control.abort(RemoteError uplink-overflow)
│ │ │ │ uplink() ends with that error; method iterator return()
│◀── throw ──────────│◀──────────────────────│ error { id, uplink-overflow }: the pump sends an error frame because the reason is a RemoteError
```
## Boundary matrix
| Event | opening | running | Uplink ended | Downlink terminated |
| --- | --- | --- | --- | --- |
| `item` received | Buffered in the inbox | Into the inbox | The stream fails with `gateway/protocol` | Ignored |
| `end` received | inbox.end | inbox.end | Idempotent | Ignored |
| `cancel` received | abort | abort | abort | Ignored |
| inbox overflow | The stream fails with `uplink-overflow` | Same | Impossible | Impossible |
| Decode failure | Impossible | The stream fails with `input-invalid` | Same | Impossible |
| Host method finishes | Impossible | Sends `end`; `uplink.return()` | Sends `end` | Already sent |
| Host method throws | Sends `error` | Sends `error`; `uplink.return()` | Sends `error` | Already sent |
| Socket closes | abort, await done | Same | Same | Nothing |
| Client `send` | Queued, sent after open | Sent | Throws | Throws |
| Client `end` | Recorded, sent after open | Sent | Idempotent | Ignored |
| Client-side downlink ends first | Impossible | Stop the pump; source `return()`; no `end` sent | Nothing | Nothing |
## Error codes
| Code | When |
| --- | --- |
| `gateway/input-invalid` | Existing: a parameter codec fails to decode; new: an uplink item fails to decode |
| `gateway/uplink-overflow` | New: the inbox exceeds `streamInboxBytes` |
| `gateway/protocol` | New: an `item` after `end` |
| `gateway/cancelled` | Existing: `signal` aborted |
New codes are registered in `packages/api/gateway/src/remote-error-codes.ts` and in the README's error-code section.
## Minimal echo method
`packages/api/gateway/tests/echo-stream.host.spec.ts` runs the same method through the WebSocket carrier and through the in-process carrier:
```text
class EchoService extends TypertRemoteService {
constructor(ctx: Context) { super(ctx, 'echo', { namespace: 'echo' }) }
@Remote({ mode: 'stream' })
async *echo(prefix: string, signal: AbortSignal): RemoteStream<string, string> {
for await (const item of this.ctx.invocation!.uplink<string>()) {
signal.throwIfAborted()
yield `${prefix}${item}`
}
}
}
// Client
const stream = remote.echo.echo('> ')
stream.send('a'); stream.send('b'); stream.end()
const replies: string[] = []
for await (const reply of stream) replies.push(reply) // ['> a', '> b']
```
## Alternatives considered
**A separate `mode: 'duplex'` plus a reserved parameter `uplink: AsyncIterable<In>` alongside `signal`.** Special logic everywhere: the analyzer must recognize a second reserved parameter and its position rules, the generated signature gains a parameter, load time must validate `duplex ⟺ uplink`, and because the mux does not know the method mode, the Gateway must reject uplink frames on behalf of `'stream'` methods. Duplex is an inherent property of the transport-layer stream, not a mode of the method: the carrier already has `streamId`, `open`, and `cancel`; the uplink adds just two more frames, which any stream can receive and which are simply not sent when unused.
**Declaring the uplink type in a decorator option, `@Remote({ mode: 'stream', uplink: 'JobInputFrame' })`.** A string reference to a type name is not checked by the compiler and breaks on rename. The return-type alias puts both types in one compiler-visible place.
**Passing a schema at runtime: `uplink(schema)`.** Bypasses the generator, the Client loses type projection, and it splits from the "strict codec per item" system. The codec still comes from the generator; the generic in the method is only an assertion.
**Having the analyzer cross-check the hand-written generic of `uplink<In>()` against the return type.** A business-specific check whose mistakes the caller bears, consistent with the other hand-written assertions in the repository; not added.
**A transport-layer handle with reopening, cursors, and a `reopen` hook.** Reopening after disconnection belongs to the upper-layer protocol: a job resumes through an `opened` frame carrying `from` and a client that remembers `next` and calls again; snapshot and journal each have their own anchors and classification. The transport layer knows one generation.
**Putting the Client's `send` on the `AsyncIterable` the generated method returns.** Hanging methods on an `AsyncIterable` is unconventional; the handle type `RemoteStreamHandle` extends it explicitly.
**Copying the Access mechanism.** Access control is business semantics, served by `vouch` and the `remote/invoke` waterfall; this Note only needs "whose stream is this", and `PeerScope` suffices.
**Keeping unary RPC for the uplink.** One HTTP round trip, one authentication, and one lookup per uplink; two carriers with no ordering relation that only Client-side serialization can order; no EOF; cancellation that does not cover both halves; a hand-made correlation key per feature.
**A native HTTP request body stream.** A one-shot request body with no association to the downlink stream, bypassing Typert descriptors, and browser support for full-duplex fetch cannot be relied on.
**One dedicated WebSocket per duplex stream.** Loses the authentication, heartbeat, and multiplexing shared with the existing downlink streams, and every feature manages its own socket lifecycle.
**WebTransport or HTTP/2 bidirectional streams.** They have per-stream flow control, but support on the browser and Node sides, proxy traversal, and the Electron carrier are all immature; adding two frames to the existing mux is orders of magnitude cheaper than replacing the transport layer.
## Consequences
- **Bought**: one stream with two directions, so interactive scenarios such as terminal keystrokes, approval answers, and background-job stdin no longer each build their own "one stream plus one unary"; a Host method knows its caller; the uplink and downlink types are declared in one place, the return type; uplink items come from the browser and are validated strictly per item at the Host, while downlink items are Host-produced typed values that pass straight through; the generated output and wire frames of existing `AsyncIterable<Out>` methods and unary methods are completely unchanged.
- **`this.ctx.invocation` is `undefined` outside a Remote call**, so a method that reads it must handle the optional; a service method invoked only directly in process correctly reads `undefined`.
- **The inbox limit applies to every stream**: a method that does not read its uplink fails as a whole when it receives many uplink frames. This is a deliberate explicit failure, recorded in the README.
- **The operator is the only Peer**, and `admit` is the only admission point; the first consumer that needs a second Peer must add opening, association, and events in `dsh-client-connection`.
- **No cross-direction ordering guarantee between uplink and downlink**; a protocol that needs request-response pairing carries its own sequence numbers.
- **Upper-layer protocols do not yet expose `send`**: consumers of `$stream`, snapshot, and journal use only the downlink today; the object they receive is already the handle.
- **The worker tunnel behaves differently from the WebSocket carrier**: no inbox limit, and items after `end` are dropped, because page to worker is a same-origin trusted boundary.
## Testing
- `packages/api/gateway/tests/echo-stream.host.spec.ts`: the same echo method passes through the WebSocket carrier and through the in-process carrier.
- Gateway Host: exact-key validation in frame parsing; `UplinkInbox` buffering during opening, `item` after `end`, overflow, ignored unknown `streamId`, and cleanup on socket closure; `UplinkDecoder` delivery with and without a codec and abort on decode failure; `uplink()` taken only once; the `extend({ invocation })` view and the origin of `peer`; `cancellableStream` closing the uplink before `return()`.
- Gateway Client: the handle's `send` (JSON validation, queued before open, throws after termination), idempotent `end`, `dispose` driving the `cancel` frame and ending later reads; the pump's three exits; the pump stopping as soon as a terminal frame arrives.
- typert: the analyzer's handling of `RemoteStream<Out, In>`, `RemoteStream<Out>`, `never`, and the existing `AsyncIterable<Out>`; decorator options; registry load-time validation.
- connection: the three outcomes of `admit`, 401, 403, and the operator; the operator reaching handlers through the bridge and the shared handler; release with the Connection.
- remote-mock and the worker tunnel: uplink pass-through; the direct proxy returning the real handle.
- Existing stream and unary method tests pass without changed expectations.
## Deferred
- `$stream` / snapshot / journal expose the handle's `send` to upper-layer protocols.
- Web terminal: one `attach` stream replaces `follow` plus `write` plus `resize`, and `attachmentId` retires.
- Duplex `$events`: waterfall answers become uplink items on the same `streamId`, and `$events/result` and `clientId` retire.
- Credit frames: wait for a consumer that needs sustained high-volume uplink; the inbox limit is the initial credit.
- The `streamInboxBytes` default of 262144 is far more than keyboard and stdin need; if the first consumer is terminal paste, it may be aligned with the `maxInputBytes` of `terminal-controller`.
- A Host-initiated half-close frame (the Host stops reading the uplink but keeps writing the downlink): no consumer; not added.
- `uplink()` on unary methods: it can read items that arrive while the method runs, semantically valid but of doubtful use; allowed, noted in the README.
## Related
- [Remote event delivery](2026-08-10-remote-event-delivery.md): the unary `$events/result` answer path; the duplex `$events` item under Deferred partially supersedes it once it lands.
- [Session history and event transport](2026-08-18-session-history-and-event-transport.md): the `$stream()` supervisor and the journal / snapshot protocols above it, which this note leaves unchanged.
- [Web sidebar terminal](../feature/2026-09-09-web-sidebar-terminal.md): the unary terminal `write` and `attachmentId`; the `attach` stream under Deferred replaces it once it lands.
@@ -0,0 +1,597 @@
# Agent Note: Remote 双工流:一种流、上行通道与调用上下文
Status: implemented
[English](2026-09-19-remote-duplex-stream.md) | 中文
## 问题
### 现状
`dsh-api-gateway` 在一条 WebSocket 上复用全部 Typert Remote 流,路径固定为 `/api/remote.mux`(`packages/api/gateway/src/stream-protocol.ts`)。一个 Host 方法用 `@Remote({ mode: 'stream' })` 修饰并返回 `Iterable` 或 `AsyncIterable`,就成为一条 Host 到 Client 的逻辑流;可选的末位 `signal: AbortSignal` 是唯一的保留参数,不进线路参数,由网关在解码后的业务参数之后追加(`packages/api/gateway/src/index.ts`)。客户端拿到的生成方法返回一个裸 `AsyncIterable`。
线路帧只有五种:
| 方向 | 帧 | 定义 |
| --- | --- | --- |
| Client 到 Host | `{ type: 'open', streamId, endpoint, payload }` | `stream-protocol.ts` |
| Client 到 Host | `{ type: 'cancel', streamId }` | `stream-protocol.ts` |
| Host 到 Client | `{ type: 'item', streamId, value? }` | `stream-protocol.ts` |
| Host 到 Client | `{ type: 'error', streamId, error: { code, message, details } }` | `stream-protocol.ts` |
| Host 到 Client | `{ type: 'end', streamId }` | `stream-protocol.ts` |
客户端能发的只有打开与取消。解析器 `parseRemoteStreamClientMessage`(`stream-protocol.ts`)对其他任何帧抛错,Host 连接收到后以 1008 关闭整个 socket(`stream-server.ts`)。结论:网关今天只支持"后端 iterable 发到前端"。需要前端向后端持续发数据的功能,各自绕路。
Host 方法也没有"这次调用是谁发起的"这个概念:`InvokeRemoteRequest = { namespace, method, args, signal }`(`packages/api/gateway/src/types.ts`)没有调用者槽位,`RemoteStreamOpener` 的签名是 `(endpoint, payload, signal)`(`stream-server.ts`),WebSocket 只在握手时校一次 cookie。
### 三套各自为政的上行机制
| 功能 | 上行方式 | 位置 | 代价 |
| --- | --- | --- | --- |
| Web 终端按键 | 每次 xterm `onData` 一个 unary RPC `terminal.write(agent, id, attachmentId, data)`,客户端用 promise 链串行化并自限字节预算 | `packages/api/terminal-controller/src/client/model.ts`;Host `src/index.ts` | 一次按键一次 HTTP 往返与鉴权;两条载体靠 `attachmentId` 手工关联;满了报 `inputFull` |
| 文件上传 | `Blob` 或 `ReadableStream` 走 Remote 层之外的原生 `POST /api/session/uploadFileBinary`,`duplex: 'half'` | `packages/client/file-upload/src/client/runtime.ts`;路由 `src/index.ts` | 不经 Typert 描述符,没有类型投影 |
| 审批与提问的回答 | Host 在 `$events` 流上下发 waterfall 帧,Client 用一个独立的 unary RPC `$events/result` 回答,靠 `clientId` 加 `eventId` 关联 | `packages/api/gateway/src/client/remote-events.ts`;Host `index.ts` | 第二套关联键;`index.ts` 记录了由此产生的竞态 |
这三套机制解决的是同一个问题:一条已经打开的逻辑流,客户端想往里写。
### 具体代价
- **两条载体对齐**。终端要用 `attachmentId` 把 `follow` 流与 `write` RPC 关联起来,并用它决定谁拥有输入权。审批要用 `clientId` 把 `$events` 流与 `$events/result` RPC 关联起来。
- **没有顺序保证**。unary RPC 走 HTTP,与 WebSocket 下行互不排序;终端只能在客户端串行化写入来自保顺序。
- **每次上行单独鉴权与查找**。`write(agent, id, attachmentId, data)` 每次解析 `agent`、查终端、校验 attachment。
- **没有 EOF**。客户端无法表达"我这边写完了",半关闭要再加一个 RPC。
- **取消不覆盖两半**。取消下行流不会取消排队中的上行写;反之亦然。
- **背压各写一套**。下行按 socket 写回调节拍拉迭代器(`stream-server.ts`);终端用有界 follower 让慢消费者显式失败(`terminal-controller/src/stream.ts`);上行靠客户端自限 `maxInputBytes`。
- **文档已经过时**。`docs/api-gateway.md:160` 仍写着 Remote 只处理一请求一结果,这一句已经落后于 README 记录的 stream 模式。
### 与后台 job 的关系
后台 job 的进程形双工流面(已存档为草稿 PR,分支 `worktree/job-stdio-stream-pair-archive`)把 job 建成一对镜像的流面,浏览器要接它就需要一条既下行又上行的 Remote 流。它是本 Note 的动机之一,但不是本 Note 的范围:job 要不要一种专用的读写流,是 jobs 那条线自己的架构问题。本 Note 只覆盖传输层。
### 目标
1. 传输层只有一种流,天生双工:下行 `item` / `end` / `error`,上行 `item` / `end`,一个 `streamId`,一个代际。
2. 上行类型与下行类型写在同一个位置:方法的返回类型。
3. Host 方法通过调用上下文取上行;调用上下文同时携带调用者(Peer)。
4. 客户端拿到的不再是裸 `AsyncIterable`,而是一个能读、能写、能关的句柄。
5. 上层协议(snapshot、journal、任何业务自定义的流)不进传输层。
## 决定
### 一句话
所有 Remote 流都允许客户端上行。方法的返回类型 `RemoteStream<Out, In = never>` 同时声明下行项与上行项的类型;生成器从中生成两个 codec。Host 方法通过 `this.ctx.invocation.uplink<In>()` 取上行迭代器,通过 `this.ctx.invocation.peer` 知道调用者。客户端生成方法返回 `RemoteStreamHandle<Out, In>`:`for await` 读下行,`send` / `end` 写上行,`dispose` 关。断线重开、游标续传、基线校验都是上层协议的事。
```
Client Host
───────────────────────────────── ───────────────────────────────────────────
const s = remote.job.attach(req, signal)
│ open { streamId, endpoint, payload } ───────▶ opener(endpoint, payload, uplink, peer, control)
│ │ prepareInvocation
│ s.send(v1) → item { streamId, v1 } ───────▶ inbox.push(v1) │ invocation = { request, service, peer, signal, uplink() }
│ s.send(v2) → item { streamId, v2 } ───────▶ inbox.push(v2) │ receiverContext.extend({ invocation }).get(service)
│ s.end() → end { streamId } ───────▶ inbox.end() │ method(...args, signal)
│ ◀─────── item { streamId, o1 } │ for await (v of this.ctx.invocation.uplink()) …
│ for await (o of s) … ◀─────── item { streamId, o2 } │ yield o …
│ s.dispose() → cancel { streamId } ───────▶ control.abort() │ signal aborts; uplink.return()
│ ◀─────── end | error
```
### 术语
| 术语 | 含义 |
| --- | --- |
| 逻辑流 | 一个 `streamId` 标识的一次调用,跨越 `open` 到终止帧 |
| 代际 | 一条逻辑流的一次物理生命期。传输层只认一个代际;断线即失败,重开是上层的事 |
| 下行 | Host 到 Client 的 `item` 序列,以 `end` 或 `error` 终止 |
| 上行 | Client 到 Host 的 `item` 序列,以 `end` 终止(半关闭),或随 `cancel` 一起结束 |
| 半关闭 | 一个方向已结束而另一个方向仍开放 |
| 项 | 一个方向上的一个业务值;每项独立经 codec 校验 |
| inbox | Host 侧每逻辑流一个的有界上行队列 |
| Peer | 连接层被接纳的一方,一个 `PeerScope`;本 Host 只有操作者一个 Peer |
| 调用上下文 | `RemoteInvocation`:本次调用的请求、接收服务、Peer、取消信号与上行入口 |
### 类型签名
```text
// @deepseek-ai/dsh-typert-protocol
/**
* One Remote stream. Host face: the method returns it, and at runtime it is AsyncIterable<Out>.
* On the Client face the generated method returns RemoteStreamHandle<Out, In>; each name has exactly one meaning.
* In is the item type the Client may send uplink; the default never means the method reads no uplink.
*/
export type RemoteStream<Out, In = never> = AsyncIterable<Out> & { readonly [STREAM_UPLINK]?: In }
// Host
@Remote({ mode: 'stream' })
async *attach(request: JobAttachRequest, signal: AbortSignal): RemoteStream<JobFollowFrame, JobInputFrame> {
const uplink = this.ctx.invocation.uplink<JobInputFrame>()
void (async () => { for await (const frame of uplink) … })()
yield …
}
// Generated Client signature: same parameters, returns the handle
attach(request: JobAttachRequest, signal?: AbortSignal): RemoteStreamHandle<JobFollowFrame, JobInputFrame>
// Client usage
const stream = remote.job.attach(req, signal)
for await (const frame of stream) …
stream.send(frame) // typed as JobInputFrame
stream.end()
stream.dispose()
```
`RemoteStream` 是别名而不是接口,是刻意的:将来若把传输层做成 Node 流形态,只换别名的定义与两端的适配,方法签名、描述符、上层协议都不动。
### Host 面
```text
/** The context of this Remote call. A method reads it through this.ctx.invocation. */
export interface RemoteInvocation {
readonly request: {
readonly namespace: string
readonly method: string
readonly args: Readonly<Record<string, unknown>>
}
/** Cordis service key of the receiving service. */
readonly service: string
/** The Peer that initiated the call. In-process carriers and unadmitted calls are the operator. */
readonly peer: PeerScope
/** Carrier cancellation: Client cancel, socket closure, uplink failure. */
readonly signal: AbortSignal
/**
* The uplink items of this call. Can be taken once; the second call throws. With an uplink codec
* in the descriptor each item is decoded to In; without one, items are delivered as unknown with
* only a JSON-safety check.
* The iteration ends after the Client's end; when the method ends the downlink, the Gateway calls
* its return() and drops unconsumed items.
* The generic In is only the caller's type assertion; runtime decodes per the descriptor and does not cross-check.
*/
uplink<In = unknown>(): AsyncIterable<In>
}
declare module '@deepseek-ai/cordis' {
interface Context {
/** The Remote call this Context was derived for; undefined on a Context not derived from a Remote call. */
readonly invocation: RemoteInvocation | undefined
}
}
```
规则:
1. 任何 `mode: 'stream'` 方法都可以调用 `this.ctx.invocation.uplink()`;unary 方法也可以,但它的流在方法返回时结束,上行项只在方法运行期间可读。
2. `uplink()` 未被调用时,上行项在 inbox 里攒着,直到超限(见背压)或流结束被丢弃。这是刻意的:不读上行的方法不用知道上行存在。
3. `In` 缺省为 `never` 时描述符没有 uplink codec;客户端句柄的 `send` 类型为 `never`,编译期就不能发。运行时若仍有帧到达(手工构造的客户端),按 `unknown` 交付。
4. `this.ctx.invocation` 在非 Remote 调用的 Context 上是 `undefined`;方法内读它时按 TS 的可选处理,或在方法第一行断言。该属性的 accessor 由 `TypertRemoteService` 基类构造时注册一次,任何装有 Remote 服务的组合都能读它。
#### 调用上下文如何到达方法
机制依赖 Cordis 的两条既有行为:
- `ctx.extend(meta)` 是 `Object.create(parent)` 加自有属性(`vendor/cordis/src/context.ts:99-107`),无注册、无 fiber、不需释放,每次调用派生一个是零成本的。
- `ctx.get(service)` 返回 `createTraceable(ctx, service)` 代理(`vendor/cordis/src/reflect.ts:233-234`,`utils.ts:117-125`)。读 `ctx` 属性直接返回访问方的 Context(`utils.ts:175`);读方法时用 `createShadowMethod` 包一层,`apply` 时把 `this` 换成 shadow receiver,shadow 上的 `ctx` 是 `ctx.extend({ [symbols.shadow]: origin })`(`utils.ts:149-163`)。
- 网关因此写 `receiverContext.extend({ invocation }).get(descriptor.service)` 取一个调用视图,`Reflect.apply(method, callReceiver, args)`;方法体里 `this.ctx.invocation` 沿原型链读到本次调用。async generator 体延后执行不受影响,因为 `this` 在调用时已绑定为 shadow 对象。
### PeerScope
```text
/** The opaque identity of one Peer. */
export type PeerId = Branded<'PeerId'>
/** One Peer's session on this Host: opened and released by the admitting side; ctx owns connection-lifetime registrations. */
export interface PeerScope {
readonly id: PeerId
readonly ctx: Context
dispose(): Promise<void>
}
```
本 Host 只有一个 Peer:操作者。`@deepseek-ai/dsh-client-connection` 在服务 apply 时用 `createScope(connectionCtx, peer)` 建它,与 Agent 建自己的 scope 是同一机制:Peer 对象就是 ScopeKey,`peer.ctx` 承接连接期效果,`dispose()` 随 connection 释放让 fiber 静默。它以 `connection.operator` 暴露。
| 成员 | 语义 |
| --- | --- |
| `connection.operator: PeerScope` | 唯一的 Peer,随 connection 生灭 |
| `connection.admit(request): PeerAdmission` | 过既有的 `requestRejection`:被拒返回 `{ rejection: 401 \| 403 }`,否则返回 `{ peer: operator }`。将来的接纳方在这里接入 |
网关 `handleUpgrade(req, socket, head, peer)` 为该 socket 绑一个固定 Peer 的 opener,并在 `peer.ctx.effect` 里登记 socket 关闭;scope 已失效时改为直接以 1001 关闭 socket。`InvokeRemoteRequest.peer?` 与 `RemoteStreamOpener` 的 `peer` 形参缺席时按操作者作答,现有进程内调用方零改动;没有 connection 的组合里网关自建的操作者替身同样经 `createScope` 建立,`dispose()` 契约一致。按 id 查 Peer、把载体关联到第二个 Peer、Peer 的开关事件,都等第一个需要多 Peer 的消费者出现时再加。
Peer 不携带访问模型:Peer 是谁、能做什么,由业务插件通过 `peer.ctx` 自行附加。
### Client 面
```text
/** The handle a generated method returns on the Client face. A different name from the Host's RemoteStream, exported from the same entry point. */
export interface RemoteStreamHandle<Out, In> extends AsyncIterable<Out> {
/** Send one uplink item. Throws on a value that is not lossless JSON, after termination, or after end. */
send(item: In): void
/** Half-close the uplink: sends the end frame. Idempotent. */
end(): void
/** Cancel the whole logical stream: sends the cancel frame (when no terminal frame has arrived); the downlink iterator ends quietly. */
dispose(): void
}
```
- 句柄代表一个代际。载体丢失时 `for await` 以 `RemoteStreamCarrierError` 失败,句柄到此结束;要不要重开、带什么参数重开,是上层协议的事。
- `send` 是同步的:入队前做 `isRemoteJsonValue` 校验,非无损 JSON 值同步抛错。浏览器 `WebSocket.send` 没有写回调,发送节流没有意义;Host 侧的 inbox 上限是唯一的背压点。
- 收到终止帧(`end` / `error`)时,mux 客户端在收帧路径上立即停泵并关闭上行队列,之后 `send` / `end` 抛错,不等消费者下一次读。
- `dispose()` 调用载体迭代器的 `return()`,让 `cancel` 帧在无人读下行时也发出;之后句柄的 `next()` 直接结束,缓冲项丢弃。消费者提前 `break` 出 `for await` 等价于 `dispose()`。
- `$stream` 监督器、`RemoteSnapshotStream`、`RemoteJournalStream` 是上层工具,不在传输层内。它们的 `open` 工厂拿到的对象现在就是句柄,`for await` 行为不变;把 `send` 透出到这些上层协议是后续的事。
### 线路帧
```text
export type RemoteStreamClientMessage =
| { readonly type: 'open'; readonly streamId: string; readonly endpoint: string; readonly payload: unknown }
| { readonly type: 'item'; readonly streamId: string; readonly value?: unknown } // new: uplink item
| { readonly type: 'end'; readonly streamId: string } // new: uplink half-close
| { readonly type: 'cancel'; readonly streamId: string }
/** Frames sent by the Host are unchanged. */
export type RemoteStreamServerMessage =
| { readonly type: 'item'; readonly streamId: string; readonly value?: unknown }
| { readonly type: 'error'; readonly streamId: string; readonly error: RemoteStreamFailure }
| { readonly type: 'end'; readonly streamId: string }
```
解析规则:`item` 精确键 `type` `streamId` 加可选 `value`,`value` 须为无损 JSON 值(`isRemoteJsonValue`);`end` 精确键 `type` `streamId`。其余形状是帧格式违规,socket 以 1008 关闭。不引入二进制帧。
### 逻辑流状态机
#### Host 侧
```
open frame
(none) ───────────────────▶ opening ──── opener resolves ────▶ running
│ │
│ item frame: inbox.push │ item frame: inbox.push
│ end frame: inbox.end │ end frame: inbox.end (uplink half-close)
│ │ method yield: sends item frame
│ cancel frame / socket close ────▶│ cancel frame / socket close: control.abort
▼ ▼
aborted ◀───────────────────────── finished (sends end or error frame)
```
`opening` 期间到达的 `item` 帧进入 inbox:`receive()` 是同步的,`open` 帧处理时就创建 `ActiveStream` 与 inbox,而 opener 是异步的。客户端在 `open` 之后立刻发的项不会丢。
#### Client 侧
```
call generated method ──▶ await socket ──▶ send open ──▶ handle usable
│ send(v): sends item; end(): sends end
│ downlink: inbox.next() → yield; end → ends; error → throws
│ downlink terminates first: send/end throw
│ dispose or caller signal abort: sends cancel (when no terminal frame has arrived)
```
### 半关闭与结束语义
| 上行 | 下行 | 结果 |
| --- | --- | --- |
| 客户端 `end` | 仍开放 | Host 的 `uplink()` 迭代结束;方法继续产出。这是 stdin EOF 的形状 |
| 仍开放 | Host 方法结束,发 `end` | 客户端句柄终止,之后 `send` / `end` 抛错;Host 调用 `uplink` 迭代器的 `return()`,未消费项丢弃 |
| 仍开放 | Host 方法抛错,发 `error` | 同上;客户端下行以 `RemoteError` 失败 |
| 客户端 `dispose` 或 signal 中止 | 任意 | 发 `cancel`;Host `control.abort()`,`cancellableStream` 先关 `uplink`,再调方法迭代器的 `return()`;不发终止帧 |
| socket 关闭 | 任意 | Host 中止全部流并等待 `done`;客户端每条流以 `RemoteStreamCarrierError` 失败 |
| `end` 之后又收到 `item` | 任意 | 该流以 `gateway/protocol` 错误帧失败并中止;socket 不关闭 |
| 未知 `streamId` 的 `item` / `end` | 任意 | 忽略,与 `cancel` 相同:Host 结束流并删除 id 后,客户端在途的上行帧属正常现象,不能连累同一 socket 上的其他流 |
### 顺序保证
同一条 socket 上的帧全序到达:一条逻辑流的 `open`、`item`…、`end` 按发送顺序到达 Host;上行项进入 `uplink()` 的顺序就是 `send` 的顺序。上行与下行之间没有跨方向顺序保证;需要请求应答配对的协议自带序号。
### 背压
WebSocket 没有 HTTP/2 那样的每流窗口,一个连接一条串行写链,慢流会拖住同连接的其他流。下行的这一性质不变。上行:
**有界 inbox 加整流失败。** Host 每条逻辑流的 inbox 按帧的 UTF-8 字节计量,上限为网关配置 `streamInboxBytes`(默认 262144),对所有流生效。超限时整条逻辑流以 `gateway/uplink-overflow` 失败:下行一起断,向客户端发 `error` 帧,`uplink()` 迭代器以同一错误结束。数据不丢弃;一个方法不读上行而客户端持续发,是用法错误,由失败显式暴露。
credit 帧留待需要持续大流量上行的消费者出现时再加;inbox 上限就是初始授信,加法是增量。
### 安全与校验
| 项 | 规则 |
| --- | --- |
| 鉴权 | 与今天相同:升级时 `connection.requestRejection(req)`;打开时描述符与查找参数校验。上行项属于一条已鉴权的流,不重复鉴权 |
| Peer | 升级时接纳为一个 `PeerScope`,该 socket 上所有流的 `invocation.peer` 都是它 |
| 项校验 | 有 uplink codec 时逐项严格解码,失败即整条流以既有的 `gateway/input-invalid` 失败(details `{ endpoint, field: 'uplink' }`);无 codec 时只做 JSON 安全校验 |
| 帧大小 | 单帧沿用 `ws` 的 `maxPayload`;inbox 上限限制累计量 |
### 进程内与 worker 载体
| 载体 | 改为 |
| --- | --- |
| `ClientConnectionRpc.open`(`packages/client/connection/src/rpc.ts:245-250`)与 `RpcStreamOpen`(`client/rpc.ts:21-25`) | 追加可选 `uplink?: AsyncIterable<unknown>`;进程内没有帧也没有 inbox,调用方迭代器直接作为 `uplink()` 的源,背压就是迭代器本身的节拍 |
| worker 隧道(`packages/experimental/webworker-runtime`) | 页面新增 `stream-uplink-item` / `stream-uplink-end` 帧;`serveStream` 把它们组装成 `uplink` 交给 `seams.openStream`。同源可信边界:没有 inbox 上限,`end` 之后的项丢弃 |
| `remote-mock`(`packages/test-support/remote-mock`) | `StreamHandle.uplink: AsyncIterable<unknown>`;`rpc.open` 透传 |
进程内载体的 `peer` 缺席即操作者。
### 与 `$events` 的关系
`$events` 是一条保留端点的流;有了上行帧,waterfall 的回答可以成为同一 `streamId` 上的上行项,`clientId` 因 `streamId` 已标识代际而多余,`$events/result` 与 `index.ts:527-529` 的竞态一起消失。不在本 Note 范围,见「后续」。
## 类型与线路定义
### typert protocol(`packages/typert/protocol/src`)
```text
// types.ts
export type RemoteStream<Out, In = never> = AsyncIterable<Out> & { readonly [STREAM_UPLINK]?: In }
export type PeerId = Branded<'PeerId'>
export interface PeerScope { readonly id: PeerId; readonly ctx: Context; dispose(): Promise<void> }
export interface RemoteInvocation { … } // see Host face
declare module '@deepseek-ai/cordis' { interface Context { readonly invocation: RemoteInvocation | undefined } }
export interface InvocationDescriptor {
// existing fields unchanged; mode still has only 'stream'
readonly mode?: 'stream'
/** Uplink item codec, generated from In of the return type RemoteStream<Out, In>; absent when In is never. */
readonly uplink?: { readonly codec: TypertCodec }
readonly cancellation?: { readonly parameter: 'signal' }
readonly result: TypertCodec
}
```
`Remote` 修饰器、`RemoteMethodOptions`、`RemoteMethodMarker` 恢复为只认 `mode: 'stream'`。
### typert generator(`packages/typert/generator/src`)
- `model.ts`:`InvocationModel.uplink?: { boundary: RemoteBoundaryModel }`。
- `analyzer.ts` `remoteResultType`:对 `mode: 'stream'`,接受的返回类型包装器为 `Iterable<Out>`、`AsyncIterable<Out>`、`RemoteStream<Out, In?>`。识别 `RemoteStream` 的方式与识别标准库 `AsyncIterable` 相同:符号名加声明所在文件(`@deepseek-ai/dsh-typert-protocol` 的 `types.ts`)。第一个类型参数是下行项,第二个存在且不是 `never` 时生成 `uplink` boundary,键名 `${endpoint}:uplink`。
- `emitter.ts`:描述符字面量输出 `uplink: { codec }`;生成的 Client 签名返回 `RemoteStreamHandle<Out, In>`。
- 参数循环不再识别任何名为 `uplink` 的参数。
### 两个名字,一个入口
Host 方法用 `RemoteStream<Out, In>` 声明一条流;Client 持有的是 `RemoteStreamHandle<Out, In>`。两者都从 `@deepseek-ai/dsh-typert-protocol` 主入口导出,句柄接口没有任何 Host 依赖。生成的 Client 契约把返回类型写成 `RemoteStreamHandle<Out, In>`。一个名字只有一个含义:Client 代码从主入口拿到的 `RemoteStream` 永远是声明类型,不会与句柄混淆。
### 网关 Host(`packages/api/gateway/src`)
```text
// types.ts
export interface InvokeRemoteRequest {
readonly namespace: string
readonly method: string
readonly args: Readonly<Record<string, unknown>>
/** Uplink items; absent is equivalent to an iterator that ends immediately. */
readonly uplink?: AsyncIterable<unknown>
/** The Peer that initiated the call; absent means the operator. */
readonly peer?: PeerScope
readonly signal?: AbortSignal
}
export interface TypertGatewayWireStream {
open: (endpoint: string, payload: unknown, uplink: AsyncIterable<unknown>, peer: PeerScope | undefined, signal: AbortSignal) => Promise<AsyncIterable<unknown>>
failure: (error: unknown) => RemoteStreamFailure
}
// stream-server.ts
export type RemoteStreamOpener = (
endpoint: string,
payload: unknown,
uplink: AsyncIterable<unknown>,
peer: PeerScope,
control: AbortController,
) => Promise<AsyncIterable<unknown>>
// index.ts Config addition
/** Uplink inbox limit per logical stream, in UTF-8 bytes of frames (default 262144). */
readonly streamInboxBytes?: number
```
`prepareInvocation`:
```text
const args = await Promise.all(descriptor.parameters.map(parameter => this.resolveParameter(parameter, request.args, endpoint)))
const signal = request.signal ?? NEVER_ABORTED_SIGNAL
const invocation = new GatewayInvocation(
{ namespace: request.namespace, method: request.method, args: request.args },
descriptor.service,
request.peer ?? this.operatorPeer(),
signal,
new UplinkDecoder(request.uplink ?? EMPTY_ASYNC_ITERABLE, descriptor.uplink?.codec, endpoint, control),
)
const callReceiver = receiverContext.extend({ invocation }).get(descriptor.service) as object
if (descriptor.cancellation !== undefined) args.push(signal)
```
`GatewayInvocation.uplink()` 第一次调用返回解码器,第二次抛错。`UplinkDecoder` 是手写迭代器:有 codec 时逐项 `decode(codec, value, endpoint, 'uplink')`,失败即 `control.abort(failure)` 并抛出;无 codec 时只做 `isRemoteJsonValue` 校验。`cancellableStream` 的 `finally` 先关闭解码器(`invocation.close()`)再等待方法迭代器的 `return()`,阻塞在 `uplink.next()` 上的方法得以退出。unary 与 stream 的模式分支恢复为今天的两分支。SRC 回退:描述符无 `uplink` codec,`uplink()` 交付 `unknown`。
### 网关 Host mux(`stream-server.ts`)
```text
interface ActiveStream { readonly control: AbortController; readonly inbox: UplinkInbox; done: Promise<void> }
/** Bounded uplink queue; iterated as the source of uplink(). */
class UplinkInbox implements AsyncIterable<unknown>, AsyncIterator<unknown> {
constructor(maxBytes: number, onOverflow: (error: RemoteError) => void)
push(value: unknown, frameBytes: number): void // push after ended fails the stream with gateway/protocol; overflow calls onOverflow
end(): void // idempotent
fail(error: unknown): void // ends the iterator with an error on cancel or carrier closure
next() / return() // hand-written: next races failure, return releases immediately
}
```
`receive(text)` 分派:
```
open : exists → throw (socket 1008); otherwise create { control, inbox, done } and start the pump
item : missing → ignore; otherwise inbox.push(value, bytes)
end : missing → ignore; otherwise inbox.end()
cancel : missing → ignore; otherwise control.abort(new Error('Remote stream cancelled'))
```
`pump` 把 inbox 与该 socket 的 Peer 传给 opener;`finally` 里 `inbox.fail(...)`。以 `RemoteError` 为 reason 的中止由 pump 作为 `error` 帧发出(网关或 mux 发起的失败);以普通 `Error` 为 reason 的中止(cancel、socket 关闭)不发终止帧。
### 网关 Client(`packages/api/gateway/src/client`)
- `prepareInvocation` 不再剥离任何上行参数;位置参数仍是业务参数加可选 `signal`。
- `invokeStream` 返回 `RemoteStreamHandle`:内部持有一个上行队列,`send` 入队并由泵发 `item` 帧,`end` 发 `end` 帧,`dispose` 中止代际 signal。下行仍是 `openRemoteStream` 返回的迭代器。
- `RemoteStreamMuxClient.open(endpoint, payload, signal, uplink?)`:在发出 `open` 帧后启动上行泵,逐项发 `item`,源结束发 `end`;下行终止或 signal 中止时停泵并调用源的 `return()`(不等待,避免键盘类生成器卡死);源抛错时 `inbox.fail(error)`,下行以该错误失败,`finally` 按既有规则发 `cancel`。泵检查 `this.socket === socket`,旧代际不往新 socket 上发帧。
- 进程内载体:`connection.rpc.open(channel, endpoint, payload, signal, uplink)`。
## 落点
| 包 | 承载 |
| --- | --- |
| `dsh-typert-protocol`(`types.ts`、`index.ts`) | `RemoteStream`、`RemoteStreamHandle`、`PeerId`、`PeerScope`、`RemoteInvocation`、`ctx.invocation` 声明合并、`InvocationDescriptor.uplink`;修饰器只认 `mode: 'stream'`;`TypertRemoteService` 构造时注册 `invocation` accessor |
| `dsh-typert-registry` | 加载校验 `uplink.codec` 为有效的严格 codec |
| `dsh-typert-generator`(`model.ts`、`analyzer.ts`、`emitter.ts`) | `remoteResultType` 识别 `RemoteStream<Out, In>` 并生成 `uplink` boundary;描述符输出 `uplink: { codec }`;生成的 Client 签名返回 `RemoteStreamHandle<Out, In>`;参数循环只认业务参数与末位 `signal` |
| `dsh-client-connection`(`operator-peer.ts`、`rpc.ts`、`rpc-host.ts`、`index.ts`) | 操作者 `PeerScope`;`connection.operator` 与 `admit`;`/api` 与升级路由经 `admit` 取 Peer;`rpc.open` 的可选 `uplink` |
| `dsh-api-gateway` Host(`stream-protocol.ts`、`stream-server.ts`、`types.ts`、`index.ts`) | `item` / `end` 帧;`UplinkInbox`;`Config.streamInboxBytes`;`GatewayInvocation` 与 `extend({ invocation })`;`UplinkDecoder`;`handleUpgrade(req, socket, head, peer)`;`operatorPeer()` |
| `dsh-api-gateway` Client(`client/index.ts`、`client/stream-client.ts`) | `invokeStream` 返回句柄;上行泵由句柄的队列驱动 |
| `dsh-webworker-runtime`、`dsh-remote-mock` | worker 隧道的两种上行帧与 `serveStream` 组装;`StreamHandle.uplink`;直接代理返回真句柄 |
| 文档 | gateway、api 组、connection、protocol、generator 的 README;`docs/api-gateway`;`docs/subsystems/typert` 的 type-equiv 块;config 与 Cordis 目录 |
## 关键流程时序
### 打开、上行、半关闭、下行结束
```
Client caller mux client mux server gateway Host method
│ attach(req) │ │ │ │
│───────────────────▶│ open { id, ep, payload } ──────────────────▶│ create inbox │
│ s.send(v1) │ item { id, v1 } ───────────────────────────▶│ inbox.push(v1) │
│ │ │ │ prepareInvocation │
│ │ │ │ extend({ invocation }) │
│ │ │ │── method(req, signal) ─▶│
│ │ │ │ │ uplink() → v1
│ s.send(v2) │ item { id, v2 } ───────────────────────────▶│ inbox.push(v2) │ → v2
│ s.end() │ end { id } ────────────────────────────────▶│ inbox.end() │ iteration ends
│◀── yield o1 ───────│◀────────────────────── item { id, o1 } ─────│◀── yield o1 ────────────│
│◀── ends ───────────│◀────────────────────── end { id } ──────────│◀── method returns ──────│ gateway calls uplink.return()
```
### 客户端 dispose
```
Client caller mux client mux server gateway Host method
│ s.dispose() │ │ │ │
│───────────────────▶│ stop pump; source return() (not awaited) │ │
│◀── iterator ends ──│ cancel { id } ─────────────────────────────▶│ control.abort() │
│ │ │ │ cancellableStream: invocation.close() closes uplink
│ │ │ │ iterator.return() ─────▶│ finally cleanup
│ │ │ no terminal frame │ │
```
### inbox 超限
```
Client caller mux client mux server gateway / Host method
│ s.send(vN) │ item { id, vN } ─────▶│ inbox.push: bytes > streamInboxBytes │
│ │ │ onOverflow → control.abort(RemoteError uplink-overflow)
│ │ │ │ uplink() ends with that error; method iterator return()
│◀── throw ──────────│◀──────────────────────│ error { id, uplink-overflow }: the pump sends an error frame because the reason is a RemoteError
```
## 边界矩阵
| 事件 | opening | running | 上行已 end | 下行已终止 |
| --- | --- | --- | --- | --- |
| 收到 `item` | 入 inbox 缓冲 | 入 inbox | 流以 `gateway/protocol` 失败 | 忽略 |
| 收到 `end` | inbox.end | inbox.end | 幂等 | 忽略 |
| 收到 `cancel` | abort | abort | abort | 忽略 |
| inbox 超限 | 流以 `uplink-overflow` 失败 | 同 | 不可能 | 不可能 |
| 解码失败 | 不可能 | 流以 `input-invalid` 失败 | 同 | 不可能 |
| Host 方法结束 | 不可能 | 发 `end`;`uplink.return()` | 发 `end` | 已发 |
| Host 方法抛错 | 发 `error` | 发 `error`;`uplink.return()` | 发 `error` | 已发 |
| socket 关闭 | abort,等 done | 同 | 同 | 无事 |
| Client `send` | 入队,open 后发 | 发 | 抛错 | 抛错 |
| Client `end` | 记录,open 后发 | 发 | 幂等 | 忽略 |
| Client 侧下行先结束 | 不可能 | 停泵;源 `return()`;不发 `end` | 无事 | 无事 |
## 错误码
| 错误码 | 何时 |
| --- | --- |
| `gateway/input-invalid` | 既有:参数 codec 解码失败;新增:上行项解码失败 |
| `gateway/uplink-overflow` | 新增:inbox 超过 `streamInboxBytes` |
| `gateway/protocol` | 新增:`end` 之后的 `item` |
| `gateway/cancelled` | 既有:`signal` 中止 |
新码登记在 `packages/api/gateway/src/remote-error-codes.ts` 与 README 的错误码一节。
## 最小 echo 方法
`packages/api/gateway/tests/echo-stream.host.spec.ts` 让同一个方法分别穿过 WebSocket 载体与进程内载体:
```text
class EchoService extends TypertRemoteService {
constructor(ctx: Context) { super(ctx, 'echo', { namespace: 'echo' }) }
@Remote({ mode: 'stream' })
async *echo(prefix: string, signal: AbortSignal): RemoteStream<string, string> {
for await (const item of this.ctx.invocation!.uplink<string>()) {
signal.throwIfAborted()
yield `${prefix}${item}`
}
}
}
// Client
const stream = remote.echo.echo('> ')
stream.send('a'); stream.send('b'); stream.end()
const replies: string[] = []
for await (const reply of stream) replies.push(reply) // ['> a', '> b']
```
## 考虑过的替代方案
**独立的 `mode: 'duplex'` 加一个与 `signal` 同级的保留参数 `uplink: AsyncIterable<In>`。** 每一处都是特殊逻辑:analyzer 要认第二个保留参数及其位置规则,生成签名多一个参数,加载期要校验 `duplex ⟺ uplink`,mux 不知道方法模式所以网关要替 `'stream'` 方法拒绝上行帧。双工是传输层流的固有性质,不是方法的一种模式:载体本来就有 `streamId`、`open`、`cancel`,上行只是再加两种帧,任何流都能收,不用就不发。
**用装饰器选项声明上行类型,`@Remote({ mode: 'stream', uplink: 'JobInputFrame' })`。** 字符串引用类型名不受编译器约束,重命名即断。返回类型别名把两个类型写在同一个编译器可见的位置。
**运行时传 schema:`uplink(schema)`。** 不经生成器,客户端失去类型投影,与"每项严格 codec"的体系分裂。codec 的来源仍是生成器,方法里的泛型只是断言。
**analyzer 交叉校验 `uplink<In>()` 手填的泛型与返回类型一致。** 业务特化的校验,写错后果由调用方承担,与仓库其他处手填断言一致;不加。
**传输层句柄带重开、游标、`reopen` 钩子。** 断线重开是上层协议的事:job 的续传是 `opened` 帧带 `from`、客户端记住 `next` 再调一次;snapshot 与 journal 各有自己的锚点与分类。传输层只认一个代际。
**客户端 `send` 放在生成方法返回的 `AsyncIterable` 上。** `AsyncIterable` 上挂方法不常规;句柄类型 `RemoteStreamHandle` 显式扩展它。
**照搬 Access 机制。** 访问控制是业务语义,`vouch` 与 `remote/invoke` waterfall 为它服务;本 Note 只需要"这条流是谁的",`PeerScope` 就够。
**继续用 unary RPC 做上行。** 每次上行一次 HTTP 往返、一次鉴权、一次查找;两条载体没有顺序关系,只能靠客户端串行化;没有 EOF;取消不覆盖两半;每个功能各造一套关联键。
**原生 HTTP 请求体流。** 一次性请求体,与下行流没有关联,不经 Typert 描述符,浏览器对全双工 fetch 的支持不可依赖。
**每条双工流一个独立 WebSocket。** 失去与现有下行流共享的鉴权、心跳与多路复用,每个功能要自己管一条 socket 的生命周期。
**WebTransport 或 HTTP/2 双向流。** 有每流流控,但浏览器与 Node 两侧的支持、代理穿透与 Electron 载体都不成熟;在现有 mux 上加两种帧比换传输层便宜几个量级。
## 后果
- **买到的**:一条流两个方向,终端按键、审批回答、后台任务 stdin 这类交互式场景不再各造一套"一条流加一个 unary";Host 方法知道调用者;上行类型与下行类型在返回类型里一处声明,上行项来自浏览器,在 Host 逐项严格校验,下行项是 Host 自产的类型值,直接透传;既有 `AsyncIterable<Out>` 方法与 unary 方法的生成物与线路帧完全不变。
- **`this.ctx.invocation` 在非 Remote 调用下是 `undefined`**,方法读它要处理可选;一个只被进程内直接调用的服务方法读到 `undefined` 是正确行为。
- **inbox 上限对所有流生效**:一个不读上行的方法收到大量上行帧会整流失败。这是刻意的显式失败,README 记明。
- **Peer 只有操作者一个**,`admit` 是唯一的接纳点;第一个需要第二个 Peer 的消费者要在 `dsh-client-connection` 加上打开、关联与事件。
- **上行与下行之间没有跨方向顺序保证**;需要请求应答配对的协议自带序号。
- **上层协议尚未透出 `send`**:`$stream`、snapshot、journal 的消费者今天只用下行;它们拿到的对象已经是句柄。
- **worker 隧道与 WebSocket 载体行为不同**:没有 inbox 上限,`end` 之后的项丢弃,因为页面到 worker 是同源可信边界。
## 测试
- `packages/api/gateway/tests/echo-stream.host.spec.ts`:同一个 echo 方法分别穿过 WebSocket 载体与进程内载体。
- 网关 Host:帧解析的精确键校验;`UplinkInbox` 的 opening 缓冲、`end` 后 `item`、超限、未知 `streamId` 忽略、socket 关闭时的收尾;`UplinkDecoder` 有无 codec 两种交付与解码失败中止;`uplink()` 只能取一次;`extend({ invocation })` 视图与 `peer` 的来源;`cancellableStream` 先关 uplink 再 `return()`。
- 网关 Client:句柄的 `send`(JSON 校验、open 前入队、终止后抛错)、`end` 幂等、`dispose` 驱动 `cancel` 帧并结束后续读取;泵的三种退出;终止帧到达即停泵。
- typert:analyzer 对 `RemoteStream<Out, In>`、`RemoteStream<Out>`、`never` 与既有 `AsyncIterable<Out>` 的处理;修饰器选项;registry 加载校验。
- connection:`admit` 的 401、403 与操作者三种结果;操作者经桥与共享处理器到达;随 connection 释放。
- remote-mock、worker 隧道:上行透传;直接代理返回真句柄。
- 既有 stream 与 unary 方法的测试不改期望值即通过。
## 后续
- `$stream` / snapshot / journal 把句柄的 `send` 透出给上层协议。
- Web 终端:一条 `attach` 流取代 `follow` 加 `write` 加 `resize`,`attachmentId` 退场。
- 双工 `$events`:waterfall 的回答成为同一 `streamId` 上的上行项,`$events/result` 与 `clientId` 退场。
- credit 帧:等需要持续大流量上行的消费者出现;inbox 上限就是初始授信。
- `streamInboxBytes` 默认值 262144 对键盘与 stdin 远超需要;若第一个消费者是终端粘贴,可能与 `terminal-controller` 的 `maxInputBytes` 对齐。
- Host 发起的半关闭帧(Host 不再读上行但继续写下行):没有消费者,不加。
- unary 方法的 `uplink()`:能读到方法运行期间到达的项,语义成立但用处存疑;允许,README 注明。
## 相关
- [Remote 事件投递](2026-08-10-remote-event-delivery.zh.md):`$events/result` 的 unary 回答路径;「后续」里的双工 `$events` 落地后部分取代它。
- [会话历史与事件传输](2026-08-18-session-history-and-event-transport.zh.md):`$stream()` 监督器与其上的 journal / snapshot 协议,本 Note 不改它们。
- [Web 侧边栏终端](../feature/2026-09-09-web-sidebar-terminal.zh.md):终端按键的 unary `write` 与 `attachmentId`;「后续」里的 `attach` 流落地后取代它。
@@ -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/2026-09-20-sidebar-retained-tab-layout.md
2026-09-20-sidebar-retained-tab-layout.md: 9579a0ba47c56f1d533f066fddd37e7ddf443346
2026-09-20-sidebar-retained-tab-layout.zh.md: 365ad5682c8610cab5dc855984e58b4303bfae49
@@ -0,0 +1,82 @@
# Agent Note: Stable Sidebar tab mounting and Session retention
Status: implemented
English | [中文](2026-09-20-sidebar-retained-tab-layout.zh.md)
## Problem
Restoring Session-specific tab records and URLs does not preserve a live browsing context. Unmounting a body or changing its DOM ancestors during tab, pane or Session transitions disconnects its page. Electron 44's `WebViewElement.disconnectedCallback` detaches and resets its guest, so preserving a React key or caching an element reference cannot compensate for disconnected ancestors. Content size, clipping and overlap must also follow the actual parent layout.
## Decision
Sidebar owns live-body retention, independently of Browser navigation and Workspace storage. A retained page's DOM ancestors remain connected and unchanged during Session switches, tab selection, pane moves, collapse, floating and docking back. Browser Presentation attaches directly inside its content container and performs no placeholder measurement or overlay positioning.
### Layout ownership
- `DockLayout` renders the Sidebar's one or two horizontal panes with stable tab siblings. The existing recursive `DockSurface` and separate `FloatLayer` remain available for general split-tree rendering; the layout engine and serialized format are unchanged.
- Every tab has a stable Grid cell, frame, Header container and Body container. Explicit Grid columns determine docking placement; Flex assigns Header and Body height. Floating changes that same frame to `position: fixed` without changing its parent or replacing Body.
- A floating frame's zero-sized Grid cell establishes a stacking context at the float layer. CSS `order` expresses float depth within that layer; DOM order stays sorted by tab identity. Raising many floats cannot increment their layer past menus. Selection and strip order never reorder the content-node list.
- Common ancestors have no transform, clipping or enclosing stacking context. Docked cells own collapse transforms and visibility; each frame clips its content. Fullscreen changes the same root's width and docked layer. Existing window-drag region pulses remain, but no RAF copies content positions.
- Split and float gestures retain their existing measurements and operation state. These update user-selected split fractions and float rectangles, not a second Browser rectangle. Browser resize and content sizing use normal CSS layout.
### Session and tab lifetime
`SidebarRightTabDefinition.keepMounted` enables lazy retention for a type. Desktop Browser enables it; other types keep visibility-based body mounting. A retained body is first created on display, then survives hiding until its occurrence or provider lifetime ends. The policy is not persisted and does not select a storage partition.
`SidebarSessionView` acquires and releases one `SessionReference` using the independent `sidebarView` source, and owns its committed-mount count and retained-Body holds. `SidebarSessionViews` manages selection, retention policy and view indexes; it delegates lifetime operations to the view rather than mutating its counts or releasing its reference. `RightbarRoot` receives `SidebarSessionViewSnapshot` values and the existing injected callbacks, not resource-owner objects. It renders each through its own explicit `SessionProvider`, with a stable Session key; the renderer's general Session-remount semantics remain unchanged.
Tabs contribute Body-retention holds rather than independent Session references. The View's stable `retainTab` callback reaches its seat through owner props, so replacing Session injection bindings does not release and reacquire those holds. Those holds follow the existing occurrence signal, including a new signal after close-and-undo; the view does not create a second tab lifetime. A retired view stays in the collection's reference index until its own disposal removes it, so plugin shutdown can also reach views whose React mounts have not finished.
Only the selected Session reports frame occupancy or binds public Sidebar navigation. Background tab actions still address their own adopted store. `tab.visible` includes Session and global-panel visibility; floating does not make a background Session visible. A Session change or global panel hides its whole content tree, while Sidebar collapse hides only docked cells and leaves foreground floats visible. Hidden content is inaccessible to pointer and keyboard input, and hidden body focus is blurred.
A background Session with no initialized retained body leaves the view list. Its reference releases after its committed React root unmounts, avoiding scope disposal underneath a mounted subtree. Plugin disposal releases all references, including retired views awaiting unmount. The Sidebar plugin depends on `sessions` and `uiSession`, so Cordis unloads it when either dependency ends; `ui-session` owns binding invalidation. Views do not register disposal callbacks on the shared Session Context. Catalog absence alone is not treated as deletion. Unvisited Browser tabs and unrelated Sessions are not eagerly opened.
### Provider responsibilities
`BrowserFrame` remains navigation and observable page state. Presentation creates and attaches the carrier DOM, without pane, clipping or stacking knowledge. Electron's provider receives physical attach/detach callbacks: hiding retained content emits neither, while an actual unmount cancels pending attachment and releases the guest. A subsequent physical mount recreates it from its known address. Disposal waits for pending creation and guest releases; an acquisition completing after cancellation is released rather than attached.
The [Desktop Browser decision](../feature/2026-09-20-desktop-browser-webview.md) retains guest security, shared IPC declarations and Workspace partition ownership. Separate tabs and Sessions never share one page instance merely because they share a storage partition. Web remains disabled by default and retains its iframe carrier when explicitly enabled.
### Retention versus recovery
| Operation | Content handling |
|---|---|
| Switch tabs | Hide the old retained cell; display the target's existing body. |
| Split, reorder or move to another pane | Change logical placement and Grid assignment, without moving Body. |
| Float, dock back or raise | Change the same frame's CSS mode or cell paint order. |
| Collapse Sidebar | Hide docked cells; foreground floats remain visible. |
| Session A → B → A | Hide and reshow A's original tree instead of loading its saved URL. |
| Close, provider unload or actual generation termination | Release the body and guest; undo creates a new occurrence. |
| Application restart or window reload | Show the saved title and URL; load only after an explicit restore or address submission. |
| Guest crash | Retry through the provider; page memory and native history are not durable. |
[Layout and provider recovery](2026-09-14-sidebar-layout-provider-recovery.md) remains independent: persisted tab identity, placement and address support cold starts, not DOM serialization. A hidden page may continue scripts, media and networking.
## Alternatives considered
**Hide panes only.** Collapse works, but moving between panes or into another floating portal still changes content ancestors.
**Retain bodies only in the foreground Session.** Session-slot remounting disconnects the page and loses its live state on Session switches.
**Give each tab its own Session reference.** A view's tabs share one `SessionProvider`, whose reference must exist before their bodies render. Per-view ownership covers that parent and its children without an additional reference lifecycle for each tab.
**Cache and reparent DOM.** Ordinary reparenting invokes Electron disconnection. Atomic-move dependencies, private attachment APIs and lifecycle monkey-patches are not used.
**Position Browser content in a separate overlay.** Measurement or CSS anchors can align an overlay, but its content remains outside the parent Grid/Flex layout that must own its sizing and clipping.
**Use native popovers.** Top-layer elements cannot be covered by ordinary Portal content through `z-index`. Migrating menus and dialogs into another stacking system expands the work beyond Sidebar; the implementation preserves the existing Portal and layer model.
**Raise floats by increasing their root z-index.** An unbounded float count can cross menu and modal layers. Equal-layer Grid cells with CSS paint order retain both stacking and DOM identity.
**Save the URL and rebuild when hidden.** This cannot preserve forms, script state, scroll position or native history. It is recovery after loss, not live retention.
## Consequences
Layout correctness belongs to Sidebar/DockKit rather than Browser. The cost is stable content containers and explicit Session-reference ownership: background retention holds Session scopes and subscriptions as well as page memory. There is no implicit LRU or idle timeout; future reclamation requires an explicit suspension/recovery policy. Other tab types do not automatically acquire these costs.
Sidebar/DockKit own chrome, float stacking, focus and platform styles as well as retained content placement. The [docking infrastructure](../feature/2026-09-04-right-sidebar-docking-infrastructure.md) owns the layout engine, persistence and tab-type navigation; the [Desktop Browser decision](../feature/2026-09-20-desktop-browser-webview.md) owns guest navigation, storage and security.
## Verification
Focused tests cover retained bodies across Session-source registration and removal, persisted-layout validation, and existing Sidebar presentation. Host and Client TypeScript compilation passes. The user manually accepted a rebuilt Electron cold start; no GIF was recorded. Dedicated real-Electron evidence for guest identity, form and history retention, multi-float paint order, resizing, clipping, drop hints, focus and platform window controls remains separate from these unit tests.
@@ -0,0 +1,82 @@
# Agent Note: Sidebar 的稳定 Tab 挂载与会话保活
Status: implemented
[English](2026-09-20-sidebar-retained-tab-layout.md) | 中文
## Problem
恢复各会话的 Tab 记录和 URL, 并不能保留活的浏览上下文. 在 Tab、Pane 或会话切换期间卸载 Body 或改变其 DOM 祖先, 都会断开页面. Electron 44 的 `WebViewElement.disconnectedCallback` 会 detach 并重置 guest, 因此保持 React key 或缓存元素引用不能抵消祖先断开. 内容尺寸、裁剪与重叠也必须遵循真实父节点布局.
## Decision
Sidebar 负责活 Body 的保留,不依赖 Browser 导航和 Workspace 存储。保活页面的 DOM 祖先在切会话、选 Tab、跨 Pane 移动、收起、浮动和回停靠时保持连接与身份不变。Browser Presentation 直接挂载到内容容器内,不做占位元素测量或浮层定位。
### 布局所有权
- `DockLayout` 用稳定的 Tab 兄弟节点渲染 Sidebar 的一个或两个水平 Pane。现有递归 `DockSurface` 和独立 `FloatLayer` 仍可用于通用分割树渲染;布局引擎和序列化格式不变。
- 每个 Tab 都有稳定的 Grid cell、frame、Header 容器和 Body 容器。显式 Grid 列决定停靠归属,Flex 分配 Header 与 Body 的高度。浮动只把同一个 frame 改为 `position: fixed`,不改父节点或替换 Body。
- 浮动 frame 的零尺寸 Grid cell 在浮窗层建立层叠上下文。CSS `order` 表达该层内的浮窗深度,DOM 顺序始终按 Tab 身份排序。大量浮窗置顶不会把层级递增到菜单之上;选中项和 Tab 条排序均不重排内容节点列表。
- 共同祖先不设置 transform、裁剪或包住全部内容的层叠上下文。停靠 cell 负责收起 transform 和显隐,各 frame 裁剪自身内容。Fullscreen 修改同一个根的宽度和停靠层级。窗口拖动区域的既有脉冲仍保留,但没有 RAF 复制内容位置。
- 分栏与浮窗手势沿用原有测量和操作状态,更新用户选择的分栏比例与浮窗矩形,而非维护第二份 Browser 矩形。窗口缩放和内容尺寸变化由普通 CSS 布局处理。
### 会话与 Tab 生命周期
`SidebarRightTabDefinition.keepMounted` 为类型启用懒保活。Desktop Browser 启用它,其他类型继续按可见性挂载 Body。保活 Body 在首次显示时创建,随后隐藏不卸载,直到 occurrence 或提供方生命周期结束。该策略不持久化,也不选择存储分区。
`SidebarSessionView` 使用独立的 `sidebarView` 来源申请、释放一个 `SessionReference`,并管理自己的已提交挂载计数和 Body 保活持有。`SidebarSessionViews` 管理选中、保留策略和视图索引,将生命周期操作委托给视图,而不修改其计数或释放其引用。`RightbarRoot` 接收 `SidebarSessionViewSnapshot` 和已有注入回调,不接收资源所有者对象;它为各视图分别渲染显式 `SessionProvider`,保持会话 key 稳定,渲染器通用的会话重挂载语义不变。
Tab 提供 Body 保活持有,而不独立申请会话引用。View 的稳定 `retainTab` 回调通过 owner props 传给席位,因此替换 Session injection binding 不会释放再重新获取这些持有。这些持有跟随已有 occurrence signal,包括关闭后撤销恢复得到的新 signal;视图不创建第二套 Tab 生命周期。退出视图列表的视图仍留在集合的引用索引中,直到自行释放时移除索引,因此插件关闭时也能清理尚未结束 React 挂载的视图。
只有选中的会话向 frame 报告占位、绑定公共 Sidebar 导航。后台 Tab actions 仍指向自己的已接管存储。`tab.visible` 包含会话和全局面板的可见性,浮动不会让后台会话变成可见。切会话或打开全局面板隐藏整个内容树,收起 Sidebar 则只隐藏停靠 cell,前台浮窗仍可见。隐藏内容不可通过指针或键盘访问,隐藏 Body 内的焦点会被移除。
后台会话没有已初始化的保活 Body 时,从视图列表移除;等已提交的 React 根卸载后再释放引用,避免仍挂载的子树下方先销毁会话 scope。插件释放会清理全部引用,包括等待卸载的已移除视图。Sidebar 插件依赖 `sessions` 和 `uiSession`,任一依赖结束时由 Cordis 卸载它,绑定失效由 `ui-session` 负责。视图不向共享会话 Context 注册销毁回调。仅在目录列表中缺项不视为删除。未访问的 Browser Tab 和无关会话不会提前打开。
### 提供方职责
`BrowserFrame` 仍只负责导航与可观察页面状态。Presentation 创建、挂载载体 DOM,不了解 Pane、裁剪或层叠。Electron 提供方接收物理 attach/detach 回调:隐藏保活内容不触发这些回调,真正卸载则取消未完成的挂载并释放 guest;之后再次物理挂载时从已知地址重建。dispose(资源释放)会等待未完成的创建和 guest 释放,取消后才完成的申请只释放、不挂载。
[Desktop Browser 决策](../feature/2026-09-20-desktop-browser-webview.zh.md)继续负责 guest 安全、共享 IPC 声明和 Workspace 分区所有权。不同 Tab 和会话不能因为共享存储分区就共享一个页面实例。Web 默认仍关闭,显式启用时继续使用 iframe 载体。
### 保活与恢复
| 操作 | 内容处理 |
|---|---|
| 切换 Tab | 隐藏旧的保活 cell,显示目标已有 Body。 |
| 分栏、排序或移到另一个 Pane | 改逻辑归属和 Grid 分配,不搬移 Body。 |
| 浮动、回停靠或置顶 | 改同一个 frame 的 CSS 模式或 cell 绘制顺序。 |
| 收起 Sidebar | 隐藏停靠 cell,前台浮窗保持可见。 |
| 会话 A → B → A | 隐藏并重新显示 A 的原内容树,而非加载已保存 URL。 |
| 关闭、提供方卸载或真实 generation 结束 | 释放 Body 和 guest;撤销关闭创建新 occurrence。 |
| 应用重启或窗口刷新 | 展示保存的标题和 URL,显式恢复或提交地址后才加载。 |
| guest 崩溃 | 通过提供方重试;页面内存和原生 history 不持久化。 |
[布局与提供方恢复](2026-09-14-sidebar-layout-provider-recovery.zh.md)仍然独立:持久化的 Tab 身份、位置和地址支持冷启动,而不是序列化 DOM。隐藏页面可能继续执行脚本、播放媒体或联网。
## Alternatives considered
**只隐藏 Pane。** 可以覆盖收起,但跨 Pane 移动或进入独立浮动 Portal 仍会改变内容祖先。
**只保留前台会话的 Body.** 会话 slot 重挂载会断开页面, 导致切会话时丢失页面运行状态.
**每个 Tab 独立持有会话引用。** 同一视图的 Tab 共用一个 `SessionProvider`,它的引用必须先于这些 Body 渲染而存在。按视图持有即可覆盖父级及其子节点,无须再为每个 Tab 管理一份引用生命周期。
**缓存并搬移 DOM。** 普通 reparent 会触发 Electron 断开处理;实现不依赖原子搬移、私有挂载 API 或生命周期补丁。
**把 Browser 内容放在独立浮层中.** 测量或 CSS 锚点可以对齐浮层, 但内容仍处于负责其尺寸与裁剪的父级 Grid/Flex 布局之外.
**使用原生 popover。** 普通 Portal 内容无法通过 `z-index` 覆盖 top layer 元素。把菜单、对话框迁移到另一套层叠体系会扩大 Sidebar 之外的改动;实现保留既有 Portal 和层级模型。
**通过递增根 z-index 置顶浮窗。** 不设上限的浮窗数量可能越过菜单和模态层。同层 Grid cell 的 CSS 绘制顺序同时保留层叠与 DOM 身份。
**隐藏时保存 URL 再重建。** 无法保留表单、脚本状态、滚动位置和原生 history。这是丢失实例后的恢复,不是活页面保留。
## Consequences
布局正确性由 Sidebar/DockKit 负责,而不是 Browser。代价是稳定的内容容器与显式会话引用所有权:后台保活不仅持有页面内存,也持有会话 scope 和订阅。没有隐式 LRU 或空闲超时,未来回收需要明确的暂停与恢复策略;其他 Tab 类型不会自动承担这些成本。
Sidebar/DockKit 除了负责保活内容布局, 还负责控件、浮窗层级、焦点和平台样式. [停靠基础设施](../feature/2026-09-04-right-sidebar-docking-infrastructure.zh.md)负责布局引擎、持久化和 Tab 类型导航; [Desktop Browser 决策](../feature/2026-09-20-desktop-browser-webview.zh.md)负责 guest 导航、存储和安全.
## Verification
定向测试覆盖 Session 数据源注册与注销期间的正文保活、存档布局校验和既有 Sidebar 呈现。Host 与 Client TypeScript 编译通过。用户已手动验收重新构建后的 Electron 冷启动;未录制 GIF。guest 身份、表单与 history 保留、多浮窗绘制顺序、缩放、裁剪、放置提示、焦点和平台窗口控件仍需独立于这些单元测试的真实 Electron 证据。
@@ -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/2026-09-15-messages-v1-base-url.md
2026-09-15-messages-v1-base-url.md: 97f7294c28988e6f115963033cdac6794d70d6f3
2026-09-15-messages-v1-base-url.zh.md: ccb8c5930dbbfa6274770cc03087b4d8077fe27b
2026-09-15-messages-v1-base-url.md: d9014756eacd1752495bfecb52860634ece08ae1
2026-09-15-messages-v1-base-url.zh.md: 8f5803e4bdc68195164e187d1aa8cd537f71c5d6
@@ -12,7 +12,7 @@ The Messages transport appends the Anthropic-standard `/v1` namespace to a confi
The shared Messages API owner trims trailing slashes and treats only a final path segment exactly equal to `v1` as the existing API version. It preserves that root and appends `/messages` or `/files`; every other base receives `/v1/messages` or `/v1/files`. The same resolved root scopes cached file uploads. The official `https://api.deepseek.com/anthropic` base therefore resolves to `/anthropic/v1`, and an explicit `/anthropic/v1` base remains unchanged.
Chat Completions retains its independent URL behavior. The Messages rule does not infer support for `v1beta`, `v2`, `v4`, or other version-like suffixes; deployments that include those segments receive the standard `/v1` namespace beneath them.
The Messages rule does not infer support for `v1beta`, `v2`, `v4`, or other version-like suffixes; deployments that include those segments receive the standard `/v1` namespace beneath them.
## Alternatives considered
@@ -12,7 +12,7 @@ Messages 传输会在配置的基址后追加 Anthropic 标准 `/v1` 命名空
共享 Messages API 所有者移除末尾斜线,仅把严格等于 `v1` 的最后路径段视为已有 API 版本。它保留该根地址并追加 `/messages` 或 `/files`;其他基址均追加 `/v1/messages` 或 `/v1/files`。缓存文件上传也使用同一个解析后的根地址划分作用域。因此,官方 `https://api.deepseek.com/anthropic` 基址解析为 `/anthropic/v1`,显式 `/anthropic/v1` 基址保持不变。
Chat Completions 保留独立的 URL 行为。Messages 规则不会推断对 `v1beta`、`v2`、`v4` 或其他版本式后缀的支持;包含这些路径段的部署会在其下获得标准 `/v1` 命名空间。
Messages 规则不推断 `v1beta`、`v2`、`v4` 或其他类似版本后缀的支持;包含这些路径段的部署会在其下接收标准 `/v1` 命名空间。
## 考虑过的替代方案
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-09-16-messages-historical-tool-input.md
2026-09-16-messages-historical-tool-input.md: 201b0f21ad041395a4c4101d6919b2878e22c9ab
2026-09-16-messages-historical-tool-input.zh.md: bcc637982420124b401ef65c668b068394b6d915
2026-09-16-messages-historical-tool-input.md: 85e4047bb5274eb7c7adb458d564e1268d4ea81e
2026-09-16-messages-historical-tool-input.zh.md: 5a8c31dbee32094f124102ccac9ffb7b8b81b177
@@ -10,7 +10,7 @@ Chat Completions retains tool arguments as strings, including malformed JSON fro
## Decision
The [Messages serializer](../../../../packages/llm/llm-deepseek/src/protocols/messages/serialize.ts) follows the [pi-ai history conversion](../../../../packages/llm/llm-pi-ai/src/replay.ts): malformed JSON and non-object values become `{}` only in the outgoing historical tool input. Call ids, names, results, and original Session records remain intact. This applies with valid, absent, or unusable native replay metadata and does not execute the historical call again.
The [Messages serializer](../../../../packages/llm/llm-deepseek/src/serialize.ts) follows the [pi-ai history conversion](../../../../packages/llm/llm-pi-ai/src/replay.ts): malformed JSON and non-object values become `{}` only in the outgoing historical tool input. Call ids, names, results, and original Session records remain intact. This applies with valid, absent, or unusable native replay metadata and does not execute the historical call again.
This supersedes the historical argument rejection in the [Messages adapter decision](../feature/2026-09-07-deepseek-messages-adapter.md). New Messages responses still require valid object arguments before successful completion; output-limit truncation retains its existing pruning behavior. No Session event, persistence type, or protocol configuration changes.
@@ -10,7 +10,7 @@ Chat Completions 将工具参数保留为字符串,其中可能包含失败调
## 决策
[Messages 序列化器](../../../../packages/llm/llm-deepseek/src/protocols/messages/serialize.ts) 遵循 [pi-ai 历史转换](../../../../packages/llm/llm-pi-ai/src/replay.ts)的做法:只在发出的历史工具输入中,将非法 JSON 和非对象值替换为 `{}`。调用 ID、名称、结果和原始 Session 记录保持不变。原生回放元数据有效、缺失或不可用时均采用此规则,也不会重新执行历史调用。
[Messages 序列化器](../../../../packages/llm/llm-deepseek/src/serialize.ts) 遵循 [pi-ai 历史转换](../../../../packages/llm/llm-pi-ai/src/replay.ts)的做法:只在发出的历史工具输入中,将非法 JSON 和非对象值替换为 `{}`。调用 ID、名称、结果和原始 Session 记录保持不变。原生回放元数据有效、缺失或不可用时均采用此规则,也不会重新执行历史调用。
这取代了 [Messages 适配器决策](../feature/2026-09-07-deepseek-messages-adapter.zh.md)中的历史参数拒绝规则。新生成的 Messages 响应在成功完成前仍要求工具参数是有效对象;达到输出上限时仍按现有规则裁剪。不改变 Session 事件、持久化类型或协议配置。
@@ -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/2026-09-18-messages-input-history-compatibility.md
2026-09-18-messages-input-history-compatibility.md: 1775276deec6180f1fd186b11cb7a757a3328a54
2026-09-18-messages-input-history-compatibility.zh.md: 50367e33c29df0679a44aca1fe680d2bedc8c263
2026-09-18-messages-input-history-compatibility.md: dd974fb1ae883716002a1381f4c374dee792b3b3
2026-09-18-messages-input-history-compatibility.zh.md: 6e850328db63d5ae5cb508aa3529e9f2507182b0
@@ -10,7 +10,7 @@ Saved subagent settlement notices can contain child reasoning and tool calls in
## Decision
The [Messages serializer](../../../../packages/llm/llm-deepseek/src/protocols/messages/serialize.ts) omits `reasoning` and `tool-call` blocks from user and tool-result input. This matches Chat Completions and pi-ai for these two block types, without changing Session records or converting child reasoning into parent user text. The [provider README](../../../../packages/llm/llm-deepseek/README.md#model-experience) owns the input rules, including empty user messages, empty tool results, and rejection of other unsupported blocks.
The [Messages serializer](../../../../packages/llm/llm-deepseek/src/serialize.ts) omits `reasoning` and `tool-call` blocks from user and tool-result input. This matches Chat Completions and pi-ai for these two block types, without changing Session records or converting child reasoning into parent user text. The [provider README](../../../../packages/llm/llm-deepseek/README.md#model-experience) owns the input rules, including empty user messages, empty tool results, and rejection of other unsupported blocks.
This partially supersedes the settlement decision's rejection of serializer tolerance and the [Messages adapter decision](../feature/2026-09-07-deepseek-messages-adapter.md)'s input rejection. Notice construction still projects nonempty child text for every parent provider. Canonical child output and ordinary assistant reasoning and tool calls remain available to their existing consumers. The omission is based on input role and block type, not the notice source or creation date, so it also applies to newly supplied user and tool-result content.
@@ -10,7 +10,7 @@ Status: implemented
## 决策
[Messages 序列化器](../../../../packages/llm/llm-deepseek/src/protocols/messages/serialize.ts)省略用户与工具结果输入中的 `reasoning` 和 `tool-call` 块。这两类块的处理与 Chat Completions、pi-ai 一致,不改变 Session 记录,也不把子级推理转换成父级用户文本。[提供方 README](../../../../packages/llm/llm-deepseek/README.zh.md#model-experience)统一说明输入规则,包括空用户消息、空工具结果,以及对其他不支持块的拒绝。
[Messages 序列化器](../../../../packages/llm/llm-deepseek/src/serialize.ts)省略用户与工具结果输入中的 `reasoning` 和 `tool-call` 块。这两类块的处理与 Chat Completions、pi-ai 一致,不改变 Session 记录,也不把子级推理转换成父级用户文本。[提供方 README](../../../../packages/llm/llm-deepseek/README.zh.md#model-experience)统一说明输入规则,包括空用户消息、空工具结果,以及对其他不支持块的拒绝。
这部分取代了结算决策对序列化器容错的拒绝,以及 [Messages 适配器决策](../feature/2026-09-07-deepseek-messages-adapter.zh.md)中的输入拒绝规则。通知构建仍为每种父级提供方投影子级的非空文本。子级的规范输出以及普通助手推理和工具调用仍可供现有消费者使用。省略依据是输入角色和块类型,而不是通知来源或创建日期,因此也适用于新提供的用户和工具结果内容。
@@ -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/2026-09-19-session-local-subagent-migration.md
2026-09-19-session-local-subagent-migration.md: 26911ea6e2cb31126d13e4c8c6488ff3b45563e6
2026-09-19-session-local-subagent-migration.zh.md: 673040a3dc73d3665e42013a4704ef2d5911b23d
@@ -0,0 +1,33 @@
# Agent Note: Complete parent catalogs without migrating child catalogs
Status: implemented
English | [中文](2026-09-19-session-local-subagent-migration.zh.md)
## Problem
Opening a historical parent requires discovery facts from its direct children. A child's corrupt body or invalid descriptor, or an unreadable header elsewhere in the root, could reject the parent's V4 preparation and hide otherwise readable history.
## Decision
Opening A discovers candidate B Sessions through headers, reads their own descriptors, and completes A's catalog. Historical child reads use the V0–V3 catalog; current child reads use native validation. These reads neither prepare B's own catalog nor publish B's successor. Opening B separately performs B's migration and catalog preparation. Frontend discovery retains its parent-catalog source. Unknown-mode rows remain browsable; `list_agents` continues to select only known continuable children.
JSONL omits unreadable or unsupported headers from discovery. A child-body or descriptor-field failure produces a warning with the child path and retains the child's header identity. If the parent lacks a complete entry, migration appends `subagent/catalog`, projected as `mode: 'unknown'` alongside healthy siblings. Existing complete entries remain authoritative. Cancellation and source-consistency failures still stop the operation. Inspected child revisions, including failed decodes, are rechecked before reuse and publication so repairs cannot silently reuse stale evidence. Released predecessors remain byte-identical.
This partially supersedes child-failure propagation in [released migrations](../architecture/2026-08-31-released-session-format-migrations.md) and [incomplete child evidence](2026-09-19-v3-incomplete-child-catalog-evidence.md). Their format conversion, descriptor cardinality, conflict validation, and immutable publication rules remain active.
## Alternatives considered
**Skip catalog completion and discover children in each consumer.** This introduces incomplete tool results and additional UI states despite the parent's readable child descriptors. The parent catalog already serves these consumers.
**Drop unreadable children from the catalog.** Published parents would lose their navigation entries. Header identity is sufficient to retain membership without inventing a mode or label.
**Recursively prepare child catalogs.** A only needs B's own descriptor. Preparing B's descendants adds unrelated work and failure dependencies.
**Refuse A when one child cannot be read.** The child's discovery metadata does not justify withholding the parent's healthy history. Warnings retain the child location without inventing mode or label fields.
## Consequences
Initial parent preparation still scans headers and reads direct-child bodies one at a time. It does not claim to eliminate this I/O. Published V4 parents use their catalog without rescanning children. Unknown entries are not automatically rewritten after publication. Opening a child retries its log and determines browsing capabilities from its actual descriptor, or reports the child-local error. Catalog payload v0 retains its two known modes; v1 also supports `unknown`. Readers support both versions; healthy creation and complete historical facts retain v0, while unknown migration facts use v1. The V4 Session header is unchanged, and older readers reject the new payload version.
Raw and compressed tests cover catalog completion, child-local errors, source changes and repairs, unchanged child generations, and deferred child-catalog preparation. The shipped Web migration snapshot checks the parent catalog with a corrupt sibling, then opens the child independently. Its additional historical child raises the corpus budget from ten to eleven roles; the guard rejects a twelfth role.
@@ -0,0 +1,33 @@
# Agent Note:补齐父目录而不迁移子目录
Status: implemented
[English](2026-09-19-session-local-subagent-migration.md) | 中文
## 问题
打开历史父会话需要直属子会话的发现事实。子日志正文损坏、descriptor 无效,或根目录其他位置的 header 不可读,都可能拒绝父会话的 V4 准备,使原本可读的历史无法打开。
## 决策
打开 A 时通过 header 发现候选 B Session,读取其自身 descriptor,并补齐 A 的 catalog。历史子日志使用 V0–V3 catalog 读取,当前子日志使用原生校验。这些读取不准备 B 自己的 catalog,也不发布 B 的后继。单独打开 B 时才执行 B 的迁移与目录准备。前端发现继续读取父 catalog。模式未知的条目仍可浏览;`list_agents` 继续只选择已知可继续的子会话。
JSONL 在发现时跳过不可读或不支持的 header。子正文或 descriptor 字段失败会产生带子路径的警告,并保留该子的 header 身份。父目录没有完整条目时,迁移追加 `subagent/catalog`,与健康兄弟项一同投影为目录条目,其模式为 `mode: 'unknown'`。已有完整条目仍具有权威性。取消与来源一致性失败仍中止操作。复用和发布前会重新检查已检查子日志的修订,包括解码失败的日志,使修复不能静默复用旧证据。已发布前驱保持逐字节不变。
本决策部分取代[已发布迁移](../architecture/2026-08-31-released-session-format-migrations.zh.md)和[不完整子证据](2026-09-19-v3-incomplete-child-catalog-evidence.zh.md)中的子错误传播规则。其格式转换、descriptor 数量、冲突校验和不可变发布规则仍有效。
## 考虑过的替代方案
**跳过目录补齐,让每个消费者发现子项。** 父会话明明有可读的子 descriptor,却会因此产生不完整工具结果和额外 UI 状态。父 catalog 已经服务这些消费者。
**从目录删除不可读子会话。** 已发布父会话会丢失相应导航入口。Header 身份足以保留成员关系,无须编造模式或标签。
**递归准备子目录。** A 只需要 B 自身的 descriptor。准备 B 的后代会增加无关工作与失败依赖。
**一个子会话不可读就拒绝 A。** 子会话的发现元数据不足以成为阻断健康父历史的理由。警告保留子位置,不虚构 mode 或 label 字段。
## 后果
首次父准备仍扫描 header 并逐个读取直属子正文,不宣称消除此 I/O。已发布 V4 父会话使用其 catalog,不重新扫描子日志。发布后不自动重写未知条目。打开子会话会重试其日志,根据实际 descriptor 确定浏览能力,或报告该子会话的错误。Catalog 载荷 v0 保留两种已知模式,v1 另支持 `unknown`。读取器支持两版;正常创建及完整历史事实继续使用 v0,未知迁移事实使用 v1。V4 Session header 不变,旧读取器拒绝新载荷版本。
原始和压缩测试覆盖目录补齐、子会话错误隔离、来源变化与修复、子代际不变和延后的子目录准备。随产品 Web 迁移快照在损坏兄弟存在时检查父目录,再独立打开子会话。额外的历史子角色使语料预算由十个增加到十一个;守卫拒绝第十二个角色。
@@ -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/2026-09-19-v3-incomplete-child-catalog-evidence.md
2026-09-19-v3-incomplete-child-catalog-evidence.md: 22ca91502425257ed30aa4962f5a38497b8a8a92
2026-09-19-v3-incomplete-child-catalog-evidence.zh.md: c11e6de2c27b9892a72c28f327b6d06834d84038
2026-09-19-v3-incomplete-child-catalog-evidence.md: 400a84f376a7babc0a95492ca33a3bcd0761cd79
2026-09-19-v3-incomplete-child-catalog-evidence.zh.md: 535951630bc600647724d50107e74e83ab626739
@@ -10,9 +10,9 @@ A valid V3 child log can have no own descriptor, an unsupported descriptor versi
## Decision
The V3→V4 migration appends a missing parent catalog entry only when exactly one supported own child descriptor supplies its discovery fields. Other descriptor counts and unsupported versions contribute no new catalog fact. Existing parent entries and child events remain intact. This replaces only the missing-evidence refusal in the [adjacent migration decision](../architecture/2026-08-31-released-session-format-migrations.md).
The V3→V4 migration appends a complete missing parent catalog entry only when exactly one supported own child descriptor supplies its discovery fields. Other descriptor counts and unsupported versions retain unknown-mode membership through [session-local catalog preparation](2026-09-19-session-local-subagent-migration.md). Existing parent entries and child events remain intact. This replaces only the missing-evidence refusal in the [adjacent migration decision](../architecture/2026-08-31-released-session-format-migrations.md).
Known child creation times must still match existing parent entries. Mode/label and descriptor fields are checked only for exactly one supported own descriptor. Invalid fields in that descriptor, corrupt logs, unreadable membership, unsupported selected generations, and changed source revisions retain their existing failures.
The JSONL failure-propagation rules below are partially superseded by [session-local catalog preparation](2026-09-19-session-local-subagent-migration.md); the converter still validates supplied facts and conflicts. Known child creation times must still match existing parent entries. Mode/label and descriptor fields are checked only for exactly one supported own descriptor. JSONL isolates invalid descriptor fields and corrupt child bodies while retaining known header identity; unreadable or unsupported headers are excluded from discovery. Changed source revisions still refuse publication.
## Alternatives considered
@@ -24,6 +24,6 @@ Known child creation times must still match existing parent entries. Mode/label
## Consequences
A historical child without a usable descriptor remains readable by id but may be absent from its parent's direct-child catalog. Existing catalog entries remain visible. Once a V4 successor is published, opening it does not rescan historical children; automatic later catalog repair is outside this migration. Preparation still rechecks child revisions before publication, so newly available evidence cannot silently bypass source consistency checks.
A historical child with a readable header but no usable descriptor retains unknown-mode membership in its parent's direct-child catalog. Existing catalog entries remain visible. Once a V4 successor is published, opening it does not rescan historical children; automatic later mode backfill is outside this migration. Preparation still rechecks child revisions before publication, so newly available evidence cannot silently bypass source consistency checks.
Unit, JSONL read/write, and Preview packing tests cover unavailable evidence, unchanged source bytes, retained entries, identity conflicts, and evidence changes before publication.
@@ -10,9 +10,9 @@ Status: implemented
## Decision
V3→V4 迁移只在恰好一个受支持的自身子 descriptor 提供发现字段时,追加缺失的父目录项。其他 descriptor 数量及不受支持的版本不贡献新的目录事实。已有父目录项和子事件保持完整。这只替代[相邻迁移决策](../architecture/2026-08-31-released-session-format-migrations.zh.md)中的证据缺失拒绝规则。
V3→V4 迁移仅在恰好一个受支持的自身子 descriptor 提供发现字段时,追加完整的缺失父目录项。其他 descriptor 数量和不支持的版本通过[逐会话目录准备](2026-09-19-session-local-subagent-migration.zh.md)保留模式未知的成员关系。已有父目录项和子事件保持不变。这取代[相邻迁移决策](../architecture/2026-08-31-released-session-format-migrations.zh.md)中对缺失证据的拒绝。
已知子创建时间仍必须与已有父目录项一致。只有恰好一个受支持的自身 descriptor 时,才校验其字段并比较 mode/label。该 descriptor 字段无效、日志损坏、成员关系不可读、所选代际不受支持以及源修订变化仍保留原有失败行为。
JSONL 错误传播规则由[逐会话目录准备](2026-09-19-session-local-subagent-migration.zh.md)部分取代;转换器仍校验已提供事实及冲突。已知子创建时间必须匹配已有父目录项。只有恰好一个受支持的自身 descriptor 时才检查 mode/label 和 descriptor 字段。JSONL 隔离无效 descriptor 字段和损坏子正文,并保留已知 header 身份;不可读或不支持的 header 不参与发现。来源修订变化仍拒绝发布。
## Alternatives considered
@@ -24,6 +24,6 @@ V3→V4 迁移只在恰好一个受支持的自身子 descriptor 提供发现字
## Consequences
没有可用 descriptor 的历史子 Session 仍可按 id 读取,但可能不出现在其父级的直属子目录中。已有目录项仍可见。V4 后继发布后,打开它不会重新扫描历史子日志;后续自动修复目录不属于本迁移。准备过程仍会在发布前复查子修订,因此新出现的证据无法悄悄绕过源一致性检查。
历史子会话具有可读 header 但没有可用 descriptor 时,在父直属子目录中保留未知模式成员关系。已有目录项仍可见。V4 后继发布后,打开它不重新扫描历史子日志;自动补齐后续模式不属于本迁移。准备过程仍在发布前重新检查子修订,因此新出现的证据不能静默绕过来源一致性检查。
单元、JSONL 读写和 Preview 打包测试覆盖不可用证据、源字节不变、条目保留、身份冲突以及发布前的证据变化。
@@ -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/2026-09-20-session-open-read-coordination.md
2026-09-20-session-open-read-coordination.md: a4907df1d7dd38e4c532c38fb6ff10362a4a3898
2026-09-20-session-open-read-coordination.zh.md: 759a45d8c52a62ea7a0b30ceed4003f8804a42c3
@@ -0,0 +1,88 @@
# Agent Note: Session opening cache identity and read ordering
Status: implemented
English | [中文](2026-09-20-session-open-read-coordination.zh.md)
## Problem
Opening a historical Session can make `session.follow`, composer catalog prewarming, and RPCs with Agent parameters access the same cold log concurrently. `skills/list` requests a complete observation itself; `commands/list`, `goals/get`, `fileReferences/list`, and `sessionReferenceResolver/candidates` can initiate resume through the Agent lookup during parameter resolution. Auxiliary requests that do not wait for the history opening snapshot compete with the main conversation for reading, parsing, and projection computation instead of reusing completed results.
Reuse of completed observations has a separate problem: Cordis can return different proxy objects for accesses to the same service. Comparing the object references returned by `ctx.get('sessionPersistence')` can mistake one underlying service for different instances. Removing instance identity checks entirely is also unsafe because persistence revisions are comparable only within the corresponding service instance and Session identity.
Subagent navigation combines two needs: displaying a conversation at a known address, and discovering its parent or descendant catalogs. Refreshing parent projections before navigating through an existing parent-child address adds prerequisite work and can reject an address that the Host could validate and open when the Client catalog cache is absent.
## Decision
### Identify the cache producer with a service-owned Symbol
`SessionPersistence.identity` is a `Symbol('sessionPersistence')` created once per service instance. Proxy access preserves that value, while a replacement service instance receives a different value. The prepared observation cache looks up the Session id and compares both `persistence.identity` and `stat().revision`; the same instance and revision can reuse a completed preparation.
This identity exists only within the process and does not enter Session headers, events, or persisted files. Existing cache capacity, leases, live precedence, and revision invalidation remain unchanged. This fix does not merge Promises for unfinished cold reads; stable identity prevents false invalidation after completion, not duplicate concurrent work across all callers.
### Auxiliary RPCs wait for the current conversation's initial open to succeed
The Client uses the existing `sessions.using()` at the actual RPC call site, rather than only delaying prewarm hooks. Menu candidates, explicit reference queries, and later catalog refreshes therefore use the same waiting entry point. Each operation first checks that its target already has a Client binding, so background prewarming does not reopen a closed Session.
| Client consumer | RPC after waiting | Temporary reference source |
| --- | --- | --- |
| Skill catalog | `skills/list` | `skillCatalog` |
| Command catalog | `commands/list` | `commandCatalog` |
| Goal activation state | `goals/get` | `goalActivation` |
| `@` file and Session candidates | `fileReferences/list`, `sessionReferenceResolver/candidates` | Shared `referenceCandidates` |
`sessions.using()` waits for `reference.ready`, then for the operation's returned Promise, and finally releases its reference. `ready` means that the initial `Session.open()` attempt has settled, not that it succeeded; these consumers also require `openState === 'open'` and send no auxiliary RPC after a failed open. The Goal reader additionally checks that its captured binding has not been replaced and logs rejected reads without clearing existing display state.
Skills retain the shared catalog request's own cancellation signal; `@` queries use the current candidate request's signal for the history wait and both RPCs. Unquoted `@` queries still fetch files and Sessions concurrently after waiting, while quoted paths still query files only. Commands and Goals gain no new cancellation protocol. These references retain Client data and follow streams, grant no new Host Agent authority, and do not prohibit normal Agent resume after the opening snapshot.
### Conversations use follow; catalog branches retain explicit reads
The main conversation's opening `follow` snapshot already carries that Session's complete projection baseline. Later changes to shared projection values notify subscribers, so switching Sessions or opening a menu does not require another read of the same baseline.
- Main-view navigation does not separately refresh the selected Session's projections; restoring a saved subagent address does not prewarm its parent.
- The Sidebar retains the target child directly from its complete parent-child address, without making a parent-catalog read a prerequisite for opening the conversation.
- The header dropdown's `changeOpen` only changes presentation state and does not refresh `rootSessionId`. Child expansion and failed-read retries retain explicit refreshes under the UI callback name `refreshProjection`; the underlying `sessions.refreshProjections` API is not renamed.
- Team navigation uses the existing Lead and roster member ids to open `{ parentSessionId, childSessionId, mode: 'continuable' }` directly. It neither refreshes the parent catalog nor requires the target to be present in the Client catalog; member-role and main-view ownership checks, plus the UI restrictions on provisioning and failed members, remain unchanged.
A known address does not waive authorization checks. When opening child history, the Host still checks the target's own header parent and origin, plus the mode and descriptor ownership in its own `subagent` projection. Later continuation remains subject to the existing direct-parent authorization rules.
## Implementation scope
- Session formats, migrations, the Host observation's `all | none` policy, and the response fields of `session.projections` are unchanged.
- There is no global cold-read singleflight, and Agent lookups are not universally replaced with read-only lookups.
- `SessionManager.handleConnected()` still refreshes previously requested catalogs in bulk; this decision does not guarantee one cold read on reconnect.
- Tool-side `listDescendants` can still read child Sessions on cache misses, and expanding a child node in the Client can still cold-read that node. Ordinary main-session opening is not equivalent to every descendant-enumeration scenario.
- `@` Session candidates still enumerate headers and obtain names from live projections or projection caches, falling back to ids on misses. Existing ranking, the default limit of 50, and direct-subagent grouping remain unchanged; discovery does not recursively traverse `subagentCatalog`.
## Alternatives considered
**Remove persistence-instance comparison.** Rejected. A matching Session id and revision cannot establish that two persistence services provide the same data; replacing a service must still invalidate old preparations.
**Unwrap Cordis proxies in callers or cache service objects per Context.** Rejected. The cache needs the producer's identity, not the identity of an object returned by a particular service access. A per-instance Symbol expresses that requirement without propagating proxy implementation details.
**Add shared singleflight for every cold read.** Outside this implementation. It could merge more concurrent calls, but would require separate rules for independent waiter cancellation, leases, and task ownership during service replacement. It also would not replace the ordering requirement that auxiliary RPCs must not resume an Agent ahead of the opening snapshot.
**Delay only warm hooks, or merely await the extra projection request.** Rejected. Explicit menu queries could still bypass warm hooks, and serializing a redundant projection request does not remove its work. The wait belongs at the actual auxiliary RPC entry point; reads already supplied by follow are removed.
**Keep Team's parent-catalog presence check.** Rejected. The Team roster already provides member identity, and the existing Host history entry point can validate an explicit address. Requiring a Client catalog cache hit for navigation mistakes data that has not loaded for a child that does not exist.
## Consequences
An ordinary main Session completes its own history and projection read before auxiliary operations reuse the completed observation or attached state. The Client's direct-child list consumes the main Session's catalog without reading each child's log merely to display its row. Catalog and conversation opening are separate operations, so navigating to a known child no longer adds parent-projection prewarming.
Auxiliary catalogs become available after the conversation's opening snapshot, and operations retain temporary Client references until they finish. This wait covers only the current Client generation's initial open, not readiness after every underlying follow replacement. Reconnect and exceptional retry scenarios still require their own verification.
A missing header root catalog remains a specific coverage gap: if it belongs to an unopened parent and no other source supplies the catalog, opening the dropdown alone does not fetch it, and the current UI may continue to display a loading notice. This implementation adds neither a new parent-catalog loading entry point nor a shared unknown-state interaction protocol.
## Verification and known gaps
During local navigation to a long historical main Session, the user observed repeated `readColdSessionLog` calls converge to one using temporary timers and call stacks, and verified direct teammate navigation. This result describes that operation's cold-read helper invocation count, not every underlying file I/O, every reconnect scenario, or a single unit of end-to-end cost. Private Session content and identities from diagnosis are not test fixtures.
Focused TypeScript builds and bundles were run for affected packages during development; the complete change has no gate, behavior-test, or assembled Web snapshot verification result. Existing tests have only partial callback renames and still contain old fixture names, root-menu-opening refresh expectations, Sidebar parent-refresh prerequisites, and Team rejection of missing catalogs. These are known verification gaps, not passing checks.
## Relationship to existing records
- [Session observations and projection-owned Client state](../architecture/2026-08-25-session-observations-and-projection-owned-client-state.md) retains ownership of read cuts and shared values. This record adds cache-producer identity and opening-snapshot consumption order; the older claim of sharing unfinished cold loads does not match the current reader and is not evidence that this change implements singleflight.
- [Client Session references, reference sources, and UI state](../architecture/2026-09-15-client-session-references.md) retains ownership of `retain`, `ready`, `using`, and explicit-address semantics. This record limits temporary references to four classes of auxiliary reads as independent asynchronous operations, rather than adding references to every UI action.
- [Web subagent catalogs consume shared projections](../simplification/2026-09-08-web-subagent-catalog-projections.md) retains ownership of the generic projection store, explicit catalog reads, and control updates. This implementation applies its rule that opening a conversation uses the follow baseline, without replacing the catalog mechanism.
- [Web subagent catalogs and human continuation](../feature/2026-07-27-web-subagent-conversations.md) retains ownership of presentation and continuation authorization. This record changes only prerequisite navigation reads and root-menu refresh timing, not the child's own Host validation; the older description of interactive loading for missing root catalogs must be reconciled with the coverage gap recorded here.
@@ -0,0 +1,88 @@
# Agent Note: 会话首帧读取的缓存身份与调用时序
Status: implemented
[English](2026-09-20-session-open-read-coordination.md) | 中文
## 问题
打开历史会话时,`session.follow`、输入框目录预热和依赖 Agent 参数的 RPC 可能同时访问同一份冷日志。`skills/list` 自行请求完整 observation;`commands/list`、`goals/get`、`fileReferences/list` 与 `sessionReferenceResolver/candidates` 则可以在参数解析阶段经 Agent lookup 发起恢复。辅助请求不等待历史首帧,会与主对话争用读取、解析和 projection 计算,而不是复用已经完成的结果。
完成后的 observation 复用还有一个独立问题:Cordis 对同一服务的访问可以返回不同代理对象。以 `ctx.get('sessionPersistence')` 返回对象的引用相等性判断服务身份,会把同一底层服务误判成另一个实例。另一方面,彻底去掉实例身份检查也不安全,因为持久化 revision 只在对应服务实例和 Session 身份内可比较。
子会话导航混合了两种需求:展示一个已知地址的对话,以及发现该会话的父目录或后代目录。已有父子地址的导航再次刷新 parent projection,既增加前置工作,也会在目录缓存缺席时拒绝一个本可由 Host 校验并打开的地址。
## 决策
### 用服务自身的 Symbol 标识缓存生产者
`SessionPersistence.identity` 是每个服务实例创建一次的 `Symbol('sessionPersistence')`。代理访问保留这个值,服务被替换后新实例获得不同值。Prepared observation cache 按 Session id 查找,并同时比较 `persistence.identity` 与 `stat().revision`;相同实例、相同 revision 可以复用已完成的 preparation。
该身份只存在于进程内,不进入 Session header、事件或持久化文件。现有缓存容量、lease、live 优先级和 revision 失效规则不变。这个修复不增加尚未完成的冷读取之间的 Promise 合并;稳定身份解决的是完成后的误失效,不是所有调用方的并发去重。
### 辅助 RPC 等待当前对话的首次打开成功
Client 在实际 RPC 发送位置使用既有 `sessions.using()`,不只延迟预热钩子。菜单候选、显式引用查询和后续目录刷新因而经过相同的等待入口。每个操作先确认目标已经有 Client binding,不因后台预热重新打开一个已经关闭的会话。
| Client 消费方 | 等待后的 RPC | 临时引用来源 |
| --- | --- | --- |
| Skill 目录 | `skills/list` | `skillCatalog` |
| Command 目录 | `commands/list` | `commandCatalog` |
| Goal 激活状态 | `goals/get` | `goalActivation` |
| `@` 文件与会话候选 | `fileReferences/list`、`sessionReferenceResolver/candidates` | 共用 `referenceCandidates` |
`sessions.using()` 先等待 `reference.ready`,再等待操作返回的 Promise,最后释放自己的引用。`ready` 表示首次 `Session.open()` 尝试已结算,不保证成功;这些消费方还检查 `openState === 'open'`,失败时不发送辅助 RPC。Goal 读方还确认捕获的 binding 未被替代,并记录读取拒绝而不清除已有展示状态。
Skill 保留共享目录请求自己的取消信号;`@` 查询将当前候选请求的信号用于历史等待和两个 RPC。未加引号的 `@` 在等待后仍并行查询文件和会话,加引号的路径仍只查询文件。Command 与 Goal 不新增取消协议。这里的引用持有 Client 数据和跟随流,不赋予 Host Agent 新权限,也不禁止首帧之后发生正常 Agent 恢复。
### 对话使用 follow,目录分支保留显式读取
主对话的 `follow` 首帧已经携带该 Session 的完整 projection baseline。后续共享 projection 值变化会通知订阅者,不需要在切换会话或打开菜单时再次读取相同 baseline。
- 主视图切换不再额外刷新所选 Session 的 projection;恢复保存的子会话地址不再预先刷新其 parent。
- 侧栏从完整父子地址直接保留目标子会话,不把父目录读取作为打开对话的前提。
- Header 下拉菜单的 `changeOpen` 只改变展示状态,不刷新 `rootSessionId`。展开子节点和失败重试仍调用显式刷新,UI 回调命名为 `refreshProjection`;底层 `sessions.refreshProjections` API 不改名。
- Team 点击使用已有 lead 与 roster 成员 id,直接打开 `{ parentSessionId, childSessionId, mode: 'continuable' }`。它不刷新 parent catalog,也不要求目标预先出现在 Client catalog 中;成员角色、主视图持有检查及 UI 对 provisioning、failed 成员的限制保持不变。
已知地址不等于免除授权校验。Host 在打开子会话历史时,仍检查目标自身 header 的 parent、origin,以及自身 `subagent` projection 的 mode 和描述符归属。后续 continuation 仍受直接 parent 的既有授权规则约束。
## 实现边界
- 没有修改 Session 格式、迁移、Host observation 的 `all | none` 策略或 `session.projections` 返回字段。
- 没有实现全局 cold-read singleflight,也没有把所有 Agent lookup 改成只读 lookup。
- 没有删除 `SessionManager.handleConnected()` 对此前请求目录的批量刷新;本决策不保证重连只有一次冷读。
- 工具侧 `listDescendants` 的缓存缺失仍可能读取子会话;前端展开一个子节点也仍可能冷读该节点。普通主会话首开不等于所有后代枚举场景。
- `@` 会话候选仍枚举 header,并从 live projection 或 projection cache 取名称;缓存缺失回退到 id。原有排序、默认 50 条上限和直接 subagent 分组规则不变,没有改成沿 `subagentCatalog` 递归发现。
## 考虑过的替代方案
**取消 persistence 实例比较。** 不采用。同一个 Session id 和 revision 不能证明两个持久化服务提供同一份数据;服务替换仍必须使旧 preparation 失效。
**在调用方解包 Cordis 代理或按 Context 缓存服务对象。** 不采用。缓存需要的是生产者身份,而不是某次服务访问返回对象的身份;每实例 Symbol 足够表达这个要求,不需要传播代理实现细节。
**为所有冷读增加统一 singleflight。** 未纳入本次实现。它可以合并更多并发调用,但需要另外定义独立等待方取消、lease 和服务替换期间的任务归属;它也不能代替“辅助 RPC 不抢在首帧前恢复 Agent”的时序要求。
**只延迟 warm,或只给额外 projection 请求补 await。** 不采用。显式菜单查询仍能绕过 warm;串行执行一个重复 projection 请求也没有删除重复工作。等待放在实际辅助 RPC 入口,已有 follow 能提供的读取直接取消。
**保留 Team 的父目录存在性检查。** 不采用。Team roster 已给出成员身份,显式地址可以由现有 Host 历史入口验证;把 Client catalog 缓存命中作为导航前提,会把数据尚未加载误当成子会话不存在。
## 后果
普通主会话先完成自己的历史和 projection 读取,辅助操作再复用已完成的 observation 或已挂载状态。前端直接子代列表消费主会话的 catalog,不为显示每一行读取对应子会话正文。目录与对话的打开入口分开,导航到一个已知子会话不再附带 parent projection 预热。
代价是辅助目录首次可用时间晚于对话首帧;操作还会持有临时 Client 引用直到结束。该等待只对应当前 Client generation 的首次打开,不是每次底层 follow 重建的就绪屏障。所有重连和异常重试场景仍需各自验证。
Header 根目录缺失仍是明确的覆盖缺口:若它属于未打开的 parent,且没有其他来源提供 catalog,仅打开下拉菜单不会补读;现有 UI 仍可能显示加载提示。本次没有增加新的父目录加载入口,也没有实现统一的 unknown 交互协议。
## 验证与已知缺口
用户在本地长历史主会话切换流程中,通过临时 `readColdSessionLog` 计时和调用栈观察到重复调用收敛为一次,并验证了 teammate 直接导航。这个结果只描述该操作的冷读辅助函数调用次数,不代表所有底层文件 I/O、全部重连场景或端到端性能都只有一次成本。诊断期间的私有 Session 内容和身份不作为测试 fixture。
开发期对相关包执行过定向 TypeScript 构建与打包;尚无这批完整变更的门禁、行为测试或组装 Web 快照验证结果。现有测试只完成部分回调改名,仍有旧 fixture 名称、打开根菜单即刷新、侧栏先刷新 parent,以及 Team 拒绝缺失 catalog 的预期需要调整。这些是已知验证缺口,不是已经通过的检查。
## 与既有记录的关系
- [Session observation 与 projection 所有的客户端状态](../architecture/2026-08-25-session-observations-and-projection-owned-client-state.zh.md)继续拥有读取切面和共享值的职责划分。本记录补充缓存生产者身份与首帧消费顺序;旧记录中“包括未完成冷加载的共享”描述与当前 reader 实现不一致,不能作为本次 singleflight 已实现的依据。
- [Client Session 引用、引用来源与 UI 状态](../architecture/2026-09-15-client-session-references.zh.md)继续拥有 `retain`、`ready`、`using` 与显式地址语义。本记录限定四类辅助读取作为独立异步操作持有临时引用,不把额外引用扩展到所有 UI 动作。
- [Web 子代理目录消费共享 projection](../simplification/2026-09-08-web-subagent-catalog-projections.zh.md)继续拥有通用 projection store、显式目录读取和 control 更新。本次实现落实“打开会话使用 follow baseline”的规则,不替换整个目录机制。
- [Web subagent 目录与用户继续交互](../feature/2026-07-27-web-subagent-conversations.zh.md)继续拥有展示和 continuation 授权。本记录只调整导航前置读取及根菜单刷新时机,不取消子会话自身的 Host 校验;旧记录对缺失根目录交互加载的描述需与这里的覆盖缺口一起核对。
@@ -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/2026-09-20-windows-embedded-mandatory-update.md
2026-09-20-windows-embedded-mandatory-update.md: 99577a4d85adc3fe13ea0ae262d5aa1af6128237
2026-09-20-windows-embedded-mandatory-update.zh.md: abe4a9577d8e11d8de2089dc53ef07433e10bbdb
@@ -0,0 +1,25 @@
# Agent Note: Windows embedded mandatory update
Status: implemented
English | [中文](2026-09-20-windows-embedded-mandatory-update.zh.md)
## Problem
A second native Windows window does not resize atomically with its parent, producing a detached overlay during maximization (#4566). Native modality also disables the parent controls that users need to exit.
## Decision
Windows uses one native window and a shell-origin frame below the caption. This partially supersedes the presentation decision in the [mandatory-update client](../feature/2026-09-11-desktop-mandatory-update-client.md); that note retains policy and installation ownership. macOS retains its native overlay.
The isolated application preload transfers a private MessageChannel endpoint directly to the shell frame. Only messages on the paired private port authorize update IPC; synthetic window events cannot invoke actions. The main process validates the current application main frame and rebinds state delivery when the main window changes. Destroyed windows receive neither state nor focus operations.
## Alternatives considered
**Synchronize a second native window.** Parent movement, maximization and restoration expose independently scheduled geometry updates; removing the second window removes that synchronization problem.
**Treat source and origin on window events as authentication.** Product scripts can construct those event fields. A private port keeps action authority outside the shared product document.
## Consequences
Caption menus and Web controls remain reachable. The DOM overlay blocks ordinary background input but cannot prevent product scripts from hiding or imitating presentation; it is not a tamper-proof security surface. Main-process policy and installation authorization remain separate from that display. Regression coverage requires forged-event rejection, real channel delivery, caption keyboard operation, and destroyed/replacement-window lifecycle cases.
@@ -0,0 +1,25 @@
# Agent Note: Windows 内嵌强更弹窗
Status: implemented
[English](2026-09-20-windows-embedded-mandatory-update.md) | 中文
## 问题
Windows 的第二个原生窗口无法与父窗口原子地缩放,最大化时会出现脱离的蒙层(#4566)。原生模态还会禁用用户退出应用所需的父窗口控件。
## 决策
Windows 使用一个原生窗口和顶栏下方的壳源站 frame。此决策部分取代[强更客户端](../feature/2026-09-11-desktop-mandatory-update-client.zh.md)的展示决策;该说明继续负责策略与安装归属。macOS 保留原生蒙层。
隔离的应用预加载将私有 MessageChannel 端点直接传给壳 frame。只有配对私有端口上的消息才能授权更新 IPC;合成的 window 事件不能调用操作。主进程校验当前应用主 frame,并在主窗口变化时重新绑定状态投递。已销毁窗口不再接收状态或焦点操作。
## 考虑过的替代方案
**同步第二个原生窗口。** 父窗口移动、最大化与还原会暴露独立调度的几何更新;去掉第二个窗口即可消除这一同步问题。
**把 window 事件的 source 和 origin 当作身份认证。** 产品脚本可以构造这些事件字段。私有端口将操作权限保留在共享产品文档之外。
## 后果
顶栏菜单和 Web 控件仍可触达。DOM 蒙层阻止普通背景输入,但不能阻止产品脚本隐藏或仿造展示;它不是防篡改的安全界面。主进程策略和安装授权与展示相互独立。回归覆盖要求包括拒绝伪造事件、真实通道投递、顶栏键盘操作,以及窗口销毁和替换的生命周期用例。
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-08-web-background-job-display.md
2026-08-08-web-background-job-display.md: 4b8396c1c0504ddf8494a03c22b9af094c4eed17
2026-08-08-web-background-job-display.zh.md: 03a43b87c47d0ef4d9c08d32facd91ceb76dda38
2026-08-08-web-background-job-display.md: 049785e9cd5dea2d92d9fc5a14dfeddac9b64469
2026-08-08-web-background-job-display.zh.md: 1dd7ee9215fa034a0a7abbe75bd5f6eb6907e9ee
@@ -2,6 +2,10 @@
Status: implemented
Superseded: the roster this note put on the session control stream (`jobsBySession`, `onJobsChanged`) now streams from the job controller's `job.list` into `ctx.jobs` — see [the jobs seam consolidation](../architecture/2026-09-03-jobs-seam-consolidation.md). The display decisions below (one roster, sections, durations, no kill control) still hold.
Update: the output phase this note deferred now ships as the per-job observation record on `ctx.jobs` — see [jobs absorb the record](../architecture/2026-09-01-jobs-absorb-activity-record.md).
English | [中文](2026-08-08-web-background-job-display.zh.md)
## Problem
@@ -64,7 +68,7 @@ The listener is owner-granular rather than task-granular. The only consumer push
`onJobDone` is not a subset of this. It delivers the terminal record with the exact owner `Agent` under first-wins semantics that `dsh-tool-jobs` couples to `reported`; `onJobsChanged` is pure observation with no delivery meaning and marks nothing reported. Listener throws are contained and never awaited, matching `onJobDone`, and each registration is an effect on the calling fiber.
Service disposal deliberately announces nothing. Every `onJobsChanged` registration is an effect on the registry's own fiber, so the listeners are already gone by the time teardown clears the store; an observer learns the registry left through its own disposal, not through a final empty set.
Service disposal announces each removal: after the registry has cancelled and awaited its jobs, it drops every record and emits one `removed` event per job, so a subscriber on a longer-lived fiber sees the roster empty out instead of keeping a stale set (a subscriber on the registry's own fiber is already gone by then).
### The Session Controller carrier
@@ -87,7 +91,7 @@ Two replacement points keep it honest. Each control-stream generation clears the
### The header action
[`@deepseek-ai/dsh-client-ui-jobs`](../../../../packages/client/ui-jobs/README.md) registers one entry in `conversation.session.header.actions`, ordered after the subagent catalog. Its own README owns the presentation contract; the decisions worth recording here are that the control does not render at all until the session has a task, that the live badge is omitted at zero so a history-only session keeps a quiet entry point, and that settled rows stay visible because a failed task's `detail` is the only place its failure is legible.
`@deepseek-ai/dsh-client-ui-jobs` registers one entry in `conversation.session.header.actions`, ordered between the preset label and the subagent catalog (`order: 20` against the catalog's 30). Its own README owns the presentation contract; the decisions worth recording here are that the control does not render at all until the session has a task, that the live badge is omitted at zero so a history-only session keeps a quiet entry point, and that settled rows stay visible because a failed task's `detail` is the only place its failure is legible.
A running one-shot background subagent therefore appears both there and in the subagent catalog. The two answer different questions — the catalog navigates into the child's transcript, this list is the only handle a cancellation can ever attach to — and suppressing `kind: 'subagent'` here would leave the cancellation phase with no entry point for exactly those tasks.
@@ -117,7 +121,7 @@ A running one-shot background subagent therefore appears both there and in the s
The [web e2e scenario](../../../../apps/web/tests/background-job-list.e2e.ts) is the end-to-end proof and runs keyless: a real `run_in_background` bash call registers with `ctx.jobs`, the header count and row appear with no user interaction, and killing the task through the registry flips the open list to its producer detail. It asserts the whole delivery path rather than any single layer.
Below it, [`jobs-local`](../../../../packages/jobs/jobs-local/tests/jobs.spec.ts) pins the change feed at all four commit points, its containment of a throwing observer, and its removal on both explicit disposal and fiber teardown; [`control-jobs`](../../../../packages/api/session-controller/tests/control-jobs.host.spec.ts) pins the complete baseline, three change pushes, dropped internal fields, unowned fan-out, no-resume guarantee, registry-absent composition, and the prohibition on consuming model output; and the client suites pin baseline replacement, the last-wins fold, the absent-key representation, removal cleanup, and the component's ordering, duration, and dismissal behavior.
Below it, [`jobs-local`](../../../../packages/jobs/jobs-local/tests/jobs.spec.ts) pins the change feed at all four commit points, its containment of a throwing observer, and its removal on both explicit disposal and fiber teardown; [`rows`](../../../../packages/api/job-controller/tests/rows.host.spec.ts) pins the roster stream that replaced the control-stream fan-out: the complete visible set on open, a refresh after each lifecycle commit and none per append, owner removal, and a clean abort; and the client suites pin baseline replacement, the last-wins fold, the absent-key representation, removal cleanup, and the component's ordering, duration, and dismissal behavior.
## Consequences
@@ -2,6 +2,10 @@
Status: implemented
Superseded:本文放在会话控制流上的名册(`jobsBySession`、`onJobsChanged`)现在从 job 控制器的 `job.list` 流入 `ctx.jobs`——见 [jobs seam 收敛](../architecture/2026-09-03-jobs-seam-consolidation.zh.md)。下文的展示决策(单一名册、分组、时长、没有 kill 控件)仍然成立。
Update:本文推迟的输出阶段现已作为 `ctx.jobs` 上的逐 job 观测 record 交付——见 [jobs 吸收 record](../architecture/2026-09-01-jobs-absorb-activity-record.zh.md)。
[English](2026-08-08-web-background-job-display.md) | 中文
## 问题
@@ -64,7 +68,7 @@ abstract onJobsChanged(listener: JobsChangedListener): () => void
`onJobDone` 不是它的子集。后者按 first-wins 语义投递终态记录和确切的 owner `Agent`,`dsh-tool-jobs` 把这套语义与 `reported` 绑在一起;`onJobsChanged` 是纯观察,不含任何投递含义,也不把任何东西标为已上报。监听器抛错被包住且从不 await,与 `onJobDone` 一致,每次注册都是调用方 fiber 上的 effect。
服务销毁刻意什么都不通告。每个 `onJobsChanged` 注册都是注册表自身 fiber 上的 effect,等到 teardown 清空 store 时监听器早已消失;观察者通过自己的销毁而不是一份最终空集来得知注册表离开了。
服务销毁逐条通告移除:注册表取消并等完自己的 job 之后,丢弃每条记录并对每个 job 发出一个 `removed` 事件,因此注册在更长命 fiber 上的订阅者看到名册逐条清空而不是留着陈旧集合(注册在注册表自身 fiber 上的订阅者此时早已消失)。
### Session Controller 载体
@@ -87,7 +91,7 @@ abstract onJobsChanged(listener: JobsChangedListener): () => void
### header 入口
[`@deepseek-ai/dsh-client-ui-jobs`](../../../../packages/client/ui-jobs/README.zh.md) 在 `conversation.session.header.actions` 注册一个条目,排在 subagent 目录之后。呈现契约归它自己的 README;值得记在这里的决策是:会话没有任务时控件根本不渲染;活跃角标为零时省略,让只剩历史的会话保留一个安静的入口;终态行保持可见,因为失败任务的 `detail` 是其失败唯一可读之处。
`@deepseek-ai/dsh-client-ui-jobs` 在 `conversation.session.header.actions` 注册一个条目,排在 preset 标签与 subagent 目录之间(`order: 20`,目录为 30)。呈现契约归它自己的 README;值得记在这里的决策是:会话没有任务时控件根本不渲染;活跃角标为零时省略,让只剩历史的会话保留一个安静的入口;终态行保持可见,因为失败任务的 `detail` 是其失败唯一可读之处。
因此一个运行中的一次性后台 subagent 会同时出现在那里和 subagent 目录里。两者回答不同的问题——目录负责进入子会话的 transcript,而这个列表是中断能力唯一可能附着的句柄——在这里屏蔽 `kind: 'subagent'` 会让中断那一期恰好对这批任务没有入口。
@@ -117,7 +121,7 @@ abstract onJobsChanged(listener: JobsChangedListener): () => void
[web e2e 场景](../../../../apps/web/tests/background-job-list.e2e.ts)是端到端的证据,且无需密钥:一次真实的 `run_in_background` bash 调用注册进 `ctx.jobs`,header 的计数与行在没有任何用户操作的情况下出现,通过注册表杀掉该任务后打开着的列表翻到生产者给出的 detail。它断言的是整条投递链路,而不是其中某一层。
在它之下,[`jobs-local`](../../../../packages/jobs/jobs-local/tests/jobs.spec.ts) 钉住变更订阅的全部四个提交点、对抛错观察者的包容,以及显式销毁与 fiber 拆除两条路径上的注销;[`control-jobs`](../../../../packages/api/session-controller/tests/control-jobs.host.spec.ts) 钉住完整 baseline、三次变更推送、被丢弃的内部字段、无主扇出、不 resume 的保证、没有注册表的组合,以及不得消费模型输出;客户端各套件钉住 baseline 替换、last-wins 折叠、缺失键表示、移除清理,以及组件的排序、时长与关闭行为。
在它之下,[`jobs-local`](../../../../packages/jobs/jobs-local/tests/jobs.spec.ts) 钉住变更订阅的全部四个提交点、对抛错观察者的包容,以及显式销毁与 fiber 拆除两条路径上的注销;[`rows`](../../../../packages/api/job-controller/tests/rows.host.spec.ts) 钉住取代控制流扇出的名册流:打开时的完整可见集、每次生命周期提交后刷新一次且追加不刷新、拥有者移除,以及干净的中止;客户端各套件钉住 baseline 替换、last-wins 折叠、缺失键表示、移除清理,以及组件的排序、时长与关闭行为。
## 影响
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-11-background-job-completion-wakes-an-idle-owner.md
2026-08-11-background-job-completion-wakes-an-idle-owner.md: cb724d95f85900b3c55fdebfdc335e9da6439d29
2026-08-11-background-job-completion-wakes-an-idle-owner.zh.md: 1107fa2c8fdd9d3e4e45031c54fb87f5aac8ea18
2026-08-11-background-job-completion-wakes-an-idle-owner.md: b8a10a2150ead02036cb97506a25c3a51202adcf
2026-08-11-background-job-completion-wakes-an-idle-owner.zh.md: adc0770cd00adeb3b13c95dcc386464493dcd943
@@ -2,6 +2,8 @@
Status: implemented
Update: the `reported` bit and `onJobDone` this note relies on left the registry with the [jobs seam consolidation](../architecture/2026-09-03-jobs-seam-consolidation.md): `dsh-tool-jobs` now keeps the delivery ledger (a wait or an accepted `job_kill` claims the task) and skips teardown settlements by their `settled` cause, so the wake-or-inject decision below stands while the mechanism it names is the tool's ledger, not a registry flag.
English | [中文](2026-08-11-background-job-completion-wakes-an-idle-owner.zh.md)
## Problem
@@ -2,6 +2,8 @@
Status: implemented
Update:本文依赖的 `reported` 位与 `onJobDone` 已随 [jobs seam 收敛](../architecture/2026-09-03-jobs-seam-consolidation.zh.md)离开注册表:`dsh-tool-jobs` 现在持有投递台账(等待或被接受的 `job_kill` 认领任务),并按 `settled` 事件的 cause 跳过 teardown 结算,因此下文唤醒或注入的决策仍然成立,只是它提到的机制是工具的台账而非注册表标志。
[English](2026-08-11-background-job-completion-wakes-an-idle-owner.md) | 中文
## 问题
@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-20-unified-image-request-pipeline.md
2026-08-20-unified-image-request-pipeline.md: 7431a474b55a9dae82b1c9cd36b770b31fe39602
2026-08-20-unified-image-request-pipeline.zh.md: 247a57b8017f513c552204d951ef6931f7f9d045
2026-08-20-unified-image-request-pipeline.md: b79eeb7d908d664362851c28bda64dab163aac89
2026-08-20-unified-image-request-pipeline.zh.md: 444234715256c48167b0b45dbd8843a9c80079f7
@@ -34,7 +34,7 @@ Every retained request image is preceded by its display name or complete attachm
### DeepSeek Files lifecycle
The direct `deepseek-official` adapter normally uploads every retained request version through its selected protocol’s Files API and sends file-id references. [Messages](2026-09-07-deepseek-messages-adapter.md) and Chat Completions share the lifecycle while retaining their own endpoint and wire formats. A [bounded inline fallback](../../archived/bug-fix/2026-08-21-deepseek-files-inline-fallback.md) sends the same deterministic request versions when file resolution fails. The default catalog advertises `deepseek-v4-flash-vision-exp` as image-capable. Uploaded ids are indexed by Files endpoint and API-key scope plus `variantId`. Uploads request seven days by default. Chat records the returned `expires_at`; Messages metadata omits expiry, so its local reuse deadline uses the original upload time plus the configured lifetime without promising remote deletion. A mapping with no more than one hour of reuse remaining is replaced without a preceding retrieve call. The index never stores the API key.
The direct `deepseek-official` adapter uploads every retained request version through the [Messages Files API](2026-09-07-deepseek-messages-adapter.md) and sends file-id references. A [bounded inline fallback](../../archived/bug-fix/2026-08-21-deepseek-files-inline-fallback.md) sends the same deterministic request versions when file resolution fails. Uploaded ids are indexed by Files endpoint and API-key scope plus `variantId`. Uploads request seven days by default. Messages metadata omits expiry, so local reuse uses the original upload time plus the configured lifetime without promising remote deletion. A mapping with no more than one hour of reuse remaining is replaced without a preceding retrieve call. The index never stores the API key.
An upload is indexed only after the response returns a valid native file object with a matching byte count and the client establishes the protocol’s reuse deadline. A missing or inconsistent response leaves no local mapping, so a later request uploads again. Concurrent upload resolution for one scoped `variantId` shares one provider operation; one waiter cannot cancel another, and the upload stops when every waiter has cancelled. A malformed upload index is an empty cache and is replaced on the next successful upload; filesystem I/O failures remain errors. If the model endpoint reports expired, deleted, missing, or invalid ids and names one or more ids used by the request, only those mappings are removed. A stale-file response without a specific id removes every mapping used by that model request. The affected request bytes are uploaded again and the model request is retried once. A second stale rejection clears the mappings identified by its response and returns the error without a third model request. One upload quota error first lists the configured number of oldest harness-owned `dsh-` files, then deletes that collected set and retries once; deleting after pagination keeps provider cursors valid. Public file operations expose list, retrieve, delete, one-variant release, and namespace-wide release. Every Files request carries the shared Harness `User-Agent`. The client enforces a 128MiB upload limit, a 32MiB request-image limit, 10,000-file and 25GiB quotas, and one-hour to 30-day expiry range.
@@ -34,7 +34,7 @@ Status: implemented
### DeepSeek Files 生命周期
直接 `deepseek-official` 适配器通常通过所选协议的 Files API 上传每张保留的请求版本,并发送 file-id 引用。[Messages](2026-09-07-deepseek-messages-adapter.zh.md) 与 Chat Completions 共享生命周期,各自保留端点和协议格式。文件解析失败时,[有界内联回退](../../archived/bug-fix/2026-08-21-deepseek-files-inline-fallback.md)会发送相同的确定性请求版本。默认 catalog 把 `deepseek-v4-flash-vision-exp` 公布为支持图片。上传 ID 按 Files 端点和 API key 作用域以及 `variantId` 写入索引。上传默认请求 7 天有效期。Chat 记录返回的 `expires_at`;Messages 元数据不含过期时间,因此本地复用期限使用原始上传时间加配置的生存期,但不保证远端文件删除。本地映射剩余复用时间不超过一小时时会直接替换,不会先查询远端文件。索引绝不存储 API key。
直连 `deepseek-official` 适配器通过 [Messages Files API](2026-09-07-deepseek-messages-adapter.zh.md) 上传每个保留的请求版本,并发送 file-id 引用。文件解析失败时,[有界内联回退](../../archived/bug-fix/2026-08-21-deepseek-files-inline-fallback.md)发送相同的确定性请求版本。上传 id 按 Files 端点、API-key 作用域及 `variantId` 建立索引。上传默认请求七天有效期。Messages 元数据不含过期时间,因此本地复用使用原始上传时间加配置的有效期,不承诺远端删除。剩余复用时间不超过一小时的映射会直接替换,不先调用 retrieve。索引从不存储 API 密钥。
只有上传响应返回有效的原生文件对象、字节数匹配,且客户端确定了该协议的复用期限时,上传结果才会写入索引。缺失或不一致的响应不会留下本地映射,后续请求会重新上传。同一作用域和 `variantId` 的并发解析共享一次提供方上传;单个等待方无法取消其他等待方,全部等待方取消时才会停止上传。格式损坏的上传索引按空缓存处理,并在下一次成功上传时替换;文件系统 I/O 失败仍是错误。如果模型端点报告 ID 已过期、删除、缺失或无效,并指出本次请求使用的一个或多个 ID,适配器只删除这些映射。如果响应只说明文件状态失效而没有指出具体 ID,适配器会删除该次模型请求使用的全部映射。受影响的请求字节会重新上传,模型请求只重试一次。第二次仍报告文件失效时,适配器会按响应清理映射并返回错误,不会发起第三次模型请求。一次上传配额错误会先列出配置数量的最旧 `dsh-` 文件,再删除收集到的文件并重试一次;分页完成后才删除,避免游标失效。公开文件操作提供列表、查询、删除、单个变体释放和整个作用域释放。每个 Files 请求都携带 Harness 的共享 `User-Agent`。客户端执行 Files 单次上传 128MiB、请求单图 32MiB、10,000 个文件、25GiB,以及一小时到 30 天有效期限制。
@@ -0,0 +1,6 @@
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/feature/2026-08-24-activity-observation-seam.md
2026-08-24-activity-observation-seam.md: 21fe08fd21ad95eec7a83f6f29b9b94514358bb9
2026-08-24-activity-observation-seam.zh.md: ffb6766b3e5d226d324ea2894a6e3f758cf1932a
@@ -0,0 +1,53 @@
# Agent Note: The activity observation seam (`ctx.activities`) and live output streaming to the Web client
Status: implemented
Superseded: the standalone seam described here was folded into `ctx.jobs` as the per-job observation record — see [jobs absorb the record](../architecture/2026-09-01-jobs-absorb-activity-record.md); the durable-vs-live analysis below still holds.
English | [中文](2026-08-24-activity-observation-seam.zh.md)
## Problem
Every incremental reader in the shell, terminal, and jobs layers is single-consumer and consuming: `ShellProcess.readOutput()`, `TerminalSendOperation.readOutput()`, and `JobRegistry.read()` all destroy the delta they return, because the owning model is the intended reader. The [web background-job display](2026-08-08-web-background-job-display.md) shipped the roster on those terms — its carrier is test-pinned to never call `ctx.jobs.read()`, since a browser read would silently steal bytes the model's `job_output` would never see — and explicitly deferred the output phase to "a separate non-consuming observation API". So a human watching the Web client saw that a background build was running but never a line of its output, and the `workflow/log` / `workflow/phase` events scripts are told to narrate with had no consumer at all.
The substrate below bash already had the right shape: `SubprocessOutputReader.readFrom(fromByte)` is offset-based and explicitly non-consuming, and `bash-local` narrows it to a single cursor pair at the shell seam.
## Decision
A new capability seam, `packages/activity/`, is the non-consuming observation plane. An **activity** is one observable unit of long-running work: an append-only bounded output stream plus live status. The seam is pure observation — it starts nothing, cancels nothing, and is invisible to the model. Model-visible facts stay exactly where they were (tool results, `ctx.jobs` cursors), so the "model-visible ⟺ logged" invariant is untouched and a composition without the seam loses only live observation.
- **`@deepseek-ai/dsh-activity` (Service Definition)** — abstract `ActivityRegistry` on `ctx.activities`: `open(spec) → ActivityHandle { append, updateDetail, end }`, fenced `get`/`list`, non-consuming `read(id, from)` over absolute UTF-8 byte offsets, owner-relative `onActivitiesChanged` (roster) and `onOutput` (advancement signal, id only). `ActivityCorrelation { callId?, jobId? }` links a row to its tool call and job. `pumpActivityOutput` is the shared pull-substrate pump.
- **`@deepseek-ai/dsh-activity-local` (Service Provider)** — in-memory chunk ring per activity. Offsets stay absolute across eviction (`outputEarliest` marks the oldest retained byte; a read below it is `lossy`, never an error); a single over-cap chunk keeps its UTF-8-safe tail with `gapBefore`. Config `retainBytes` (256 KiB) live, trimmed to `settledRetainBytes` (16 KiB) at settlement. Records outlive producer fibers; owner disposal force-ends and removes; `end` is first-wins and post-end writes log-and-drop so a producer's trailing flush cannot break its own teardown.
- **`@deepseek-ai/dsh-api-activity-controller`** — the wire, in the workspace-controller shape: `activity.control` streams one roster baseline then whole-bucket replacement frames per owner session (the jobs-frame self-healing semantics), and `activity.observe({ activityId, from? })` streams one `opened` anchor, coalesced `output` frames (`flushMs` window, `maxFrameBytes` soft budget), then the terminal `status` on the same stream before closing — so settlement can never race a still-open output channel. Reconnects resume from the last frame's `next`. The client half installs `ctx.activityFeed` (roster mirror plus reference-counted observation streams with a bounded render tail).
- **`@deepseek-ai/dsh-client-ui-activity`** — the session-header task list ([unified with the job rows](../../archived/feature/2026-08-25-unified-task-list.md)); expanding a row opens its observation stream into a `TerminalBlock` panel, collapsing closes it, so output only flows while someone is watching. `TerminalBlock` itself gained live rendering: a running block with supplied `output` shows the text under the running state instead of the historical prompt-only frame.
Producers mirror best-effort behind `ctx.get('activities')`, never a declared inject, and every observation failure is logged and swallowed — the seam is strictly optional and can never break the work it observes:
- `ShellProcess` gained optional `observed` — the substrate's non-consuming offset readers re-exposed (bash-local and pwsh-local return `handle.collected`; the e2b subprocess provider satisfies the same contract). `dsh-tool-bash` / `dsh-tool-pwsh` open a correlated activity after `jobs.start()` commits and pump `observed` at `activityPollMs` (default 150 ms).
- `dsh-tool-workflow` mirrors each recorded top-level run as a push-mode `workflow` activity, giving the previously unconsumed `workflow/phase` / `workflow/log` events their consumer.
### Why not durable session events, the control stream, or the jobs seam
Peer products converge on the same split (Codex marks every exec-output delta "Transient, non-durable" in its rollout policy; Claude Code persists a file path, never deltas; Kimi's envelope has a first-class `volatile: true`; OpenCode's V2 corrected V1's per-chunk durable part updates to "stream fragments are live-only"). In this repository the durable route is also concretely bad: chunk-run packing is hardcoded to the three `assistant/chunk` shapes so a new output event gets zero compression, each event is one uncoalesced WebSocket frame, and reconnect gap-repair re-reads the window — and the job-display note already rejected it ("spill exists precisely so oversized tool output stays out of the log"). The session control stream's whole-snapshot frames are self-healing for rosters but quadratic for streams. Extending the jobs seam would cover only jobs — not workflow narration, not a future foreground stream — and would re-litigate the settled "producers own their buffers" decision; the activity plane instead recovers the substrate's existing offsets.
### Two cursor regimes, kept separate
The model keeps its consuming cursors (`job_output` semantics unchanged, zero prompt or tool changes anywhere in this work); observers get absolute offsets any number of readers can hold. This mirrors the strongest peer pattern (Codex: destructive drain for the model, seq-cursor reads for infrastructure) and is what makes the wire path provably non-stealing.
## Alternatives considered
**Durable `activity/output` session events with a conversation node.** Replay for free, but the log-growth math above, plus retention pressure on every follower, killed it; the transcript's durable story remains the producers' tool results.
**Extending `SessionControlFrame` with output.** One fewer stream, but whole-snapshot semantics resend the accumulated buffer per chunk — O(n²) on exactly the data that grows.
**A signal frame plus RPC pull.** The subagent-catalog shape; rejected for the split-authority apparatus the job-display note documents, and it fails worst at settlement (running row, dead stream).
**Observation inside the jobs seam.** One integration point, but only for jobs; foreground runs, terminals, and workflow narration would each need another mechanism, and the registry would grow a second buffer beside the producers' own.
## Testing
Unit suites pin the registry (offsets, eviction, UTF-8-safe tail trims, lossy reads, first-wins settlement, owner cleanup, scoped delivery, listener containment, HMR disposal, Loader composition with row config), the pump (cadence, per-source cursors, lossy `gapBefore`, settlement drain, timer hygiene), the controller (baseline/bucket frames, anchor/output/status ordering, resume, frame splitting, fenced owned reads, abort), the client model and reference-counted streams, and both producer taps (real subprocess output for bash including the never-steals-the-model's-cursor assertion; scripted readers for pwsh; event-driven mirror for workflow, including the throwing-registry containment the optionality promise requires). The keyless web e2e drives a real `run_in_background` bash through the full composition and pins ARIA goldens for the streaming panel while the command still runs and for the settled panel after a registry kill.
## Consequences
The Web client now shows live output for background bash/pwsh and workflow narration with zero model-facing change, and any plugin can join with three calls (`open`/`append`/`end`) against an optional service. Bought at the cost of one more seam family and a per-producer tap; the initial two-header-list duplication was folded into [one task list](../../archived/feature/2026-08-25-unified-task-list.md). Deferred, recorded on the package READMEs: foreground streaming (needs a shell-seam run/observe variant), terminal `pty-send` observation (needs a non-consuming face on terminal buffers), a human kill control (blocked on the jobs `reported` contract), SDK/ACP consumers, and durable replay after restart.
@@ -0,0 +1,53 @@
# Agent Note:activity 观察 seam(`ctx.activities`)与 Web 客户端实时输出流
状态:implemented
Superseded:本文描述的独立 seam 已折入 `ctx.jobs`,成为逐 job 的观测 record——见 [jobs 吸收 record](../architecture/2026-09-01-jobs-absorb-activity-record.zh.md);下文关于持久化与实时通道的分析仍然成立。
[English](2026-08-24-activity-observation-seam.md) | 中文
## 问题
shell、terminal 与 jobs 各层的增量读取器全都是单读者且消费型的:`ShellProcess.readOutput()`、`TerminalSendOperation.readOutput()` 与 `JobRegistry.read()` 返回即销毁增量,因为预期读者就是持有模型。[Web 后台任务显示](2026-08-08-web-background-job-display.zh.md)在这一前提下交付了 roster——其载体被测试钉死绝不调用 `ctx.jobs.read()`,否则浏览器读取会无声偷走模型 `job_output` 永远看不到的字节——并把输出阶段明确推迟给"一个独立的非消费观察 API"。于是人在 Web 客户端只能看到后台构建"在跑",看不到一行输出;脚本被告知用来叙述进度的 `workflow/log` / `workflow/phase` 事件更是全无消费者。
bash 之下的底座其实早有正确形态:`SubprocessOutputReader.readFrom(fromByte)` 是按 offset、明确非消费的,`bash-local` 在 shell seam 把它窄化成了单游标对。
## 决策
新建能力 seam `packages/activity/`,作为非消费观察面。**activity** 是一个可观察的长时工作单元:一条 append-only 的有界输出流加实时状态。该 seam 是纯观察——不启动、不取消任何工作,对模型不可见。模型可见事实原地不动(工具结果、`ctx.jobs` 游标),"模型可见 ⟺ 已记录"不变量不受影响;不装该 seam 的组合只失去实时观察。
- **`@deepseek-ai/dsh-activity`(Service Definition)**——`ctx.activities` 上的抽象 `ActivityRegistry`:`open(spec) → ActivityHandle { append, updateDetail, end }`、有栅栏的 `get`/`list`、按绝对 UTF-8 字节 offset 的非消费 `read(id, from)`、owner 相对投递的 `onActivitiesChanged`(roster)与 `onOutput`(推进信号,只带 id)。`ActivityCorrelation { callId?, jobId? }` 把行关联到工具调用与 job。`pumpActivityOutput` 是 pull 型底座的共享泵。
- **`@deepseek-ai/dsh-activity-local`(Service Provider)**——每个 activity 一个内存 chunk 环。offset 跨淘汰稳定(`outputEarliest` 指明最老保留字节;低于它的读取是 `lossy` 而非错误);单个超上限 chunk 保留 UTF-8 安全尾部并带 `gapBefore`。配置 `retainBytes`(256 KiB)存活,结算时裁到 `settledRetainBytes`(16 KiB)。记录存续期长于生产者 fiber;owner 销毁强制终结并移除;`end` first-wins,end 后写入记日志后丢弃,生产者的收尾 flush 砸不掉自己的 teardown。
- **`@deepseek-ai/dsh-api-activity-controller`**——wire 层,取 workspace-controller 的形态:`activity.control` 流出一份 roster baseline 加按 owner 会话的整桶替换帧(jobs 帧的自愈语义);`activity.observe({ activityId, from? })` 流出一帧 `opened` 锚点、合并的 `output` 帧(`flushMs` 窗口、`maxFrameBytes` 软预算)、再在同一条流上发终态 `status` 后关闭——结算永远不会与仍开着的输出通道竞态。重连以上一帧的 `next` 续传。客户端半区安装 `ctx.activityFeed`(roster 镜像 + 按引用计数的观察流与有界渲染尾巴)。
- **`@deepseek-ai/dsh-client-ui-activity`**——会话头部任务列表([与 job 行合并](../../archived/feature/2026-08-25-unified-task-list.md));展开某行即把其观察流打开进 `TerminalBlock` 面板,收起即关——只有有人在看时输出才流动。`TerminalBlock` 本身获得实时渲染能力:running 且提供了 `output` 的块在 running 状态下显示文本,而非历史上的仅提示行画面。
生产者一律经 `ctx.get('activities')` 尽力镜像,绝不声明 inject,一切观察失败只记日志并吞掉——该 seam 严格可选,永远不能破坏它所观察的工作:
- `ShellProcess` 新增可选 `observed`——底座的非消费 offset 读取器复出(bash-local 与 pwsh-local 返回 `handle.collected`;e2b subprocess provider 满足同一契约)。`dsh-tool-bash` / `dsh-tool-pwsh` 在 `jobs.start()` 提交后打开带关联的 activity,并以 `activityPollMs`(默认 150 ms)泵取 `observed`。
- `dsh-tool-workflow` 把每个被记录的顶层运行镜像为 push 形态的 `workflow` activity,让此前无人消费的 `workflow/phase` / `workflow/log` 事件有了消费者。
### 为何不用持久会话事件、control 流或 jobs seam
同行产品收敛于同一切分(Codex 在 rollout policy 里把所有 exec 输出 delta 标为 "Transient, non-durable";Claude Code 持久化文件路径而非 delta;Kimi 的信封有一等的 `volatile: true`;OpenCode V2 把 V1 逐 chunk 持久化 part 更新纠正为 "stream fragments are live-only")。在本仓库持久路线还有具体的坏处:chunk-run 打包只认三种 `assistant/chunk` 形态,新输出事件零压缩;每事件一条未合并的 WebSocket 帧;重连缺口修复要重读窗口——且 job 显示 note 已否决过它("spill 的存在就是为了让超大工具输出不进日志")。session control 流的整值帧对 roster 自愈、对流则是二次方开销。扩 jobs seam 只覆盖 jobs——workflow 叙述与未来的前台流都覆盖不到——还要重翻已定案的"生产者拥有自己的缓冲";activity 面转而复用底座既有的 offset。
### 两套游标制,各归其位
模型保留消费游标(`job_output` 语义不变,本次工作零 prompt/工具改动);观察者拿任意多读者可共持的绝对 offset。这正是最强的同行范式(Codex:模型用破坏性 drain,基础设施用 seq 游标读取),也是 wire 路径可证明"不偷字节"的原因。
## 曾考虑的替代方案
**持久 `activity/output` 会话事件 + conversation node。** 回放免费,但上述日志膨胀账加上对每个 follower 的保留压力否掉了它;transcript 侧的持久故事仍是生产者的工具结果。
**给 `SessionControlFrame` 加输出。** 少一条流,但整值语义会按 chunk 重发累计缓冲——恰好在会增长的数据上 O(n²)。
**信号帧 + RPC 拉取。** subagent catalog 的形态;因 job 显示 note 记录的权威割裂机制而否决,且它在结算时刻失效得最难看(行还在跑、流已死)。
**在 jobs seam 内做观察。** 一个集成点,但只对 jobs;前台运行、terminal 与 workflow 叙述各需另一套机制,注册表还得在生产者自己的缓冲旁再养一份。
## 测试
单测钉住注册表(offset、淘汰、UTF-8 安全裁尾、lossy 读取、first-wins 结算、owner 清理、scope 投递、监听容错、HMR 可逆、带行配置的 Loader 组合)、泵(节奏、每源游标、lossy `gapBefore`、结算排干、定时器卫生)、controller(baseline/桶帧、锚点/输出/状态排序、续传、帧切分、有栅栏的有主读取、abort)、客户端模型与引用计数流,以及两类生产者接入(bash 用真实子进程输出,含"绝不偷模型游标"断言;pwsh 用脚本化读取器;workflow 用事件驱动镜像,含可选性承诺要求的 throwing-registry 容错)。无钥 web e2e 经完整组合驱动真实 `run_in_background` bash,钉住命令仍在运行时流式面板与注册表 kill 后已结算面板的 ARIA golden。
## 后果
Web 客户端在零模型面改动下获得后台 bash/pwsh 实时输出与 workflow 叙述;任何插件对着一个可选服务用三个调用(`open`/`append`/`end`)即可接入。代价是多一个 seam 家族与每生产者一个接入点;最初的双头部列表重复已折并为[单一任务列表](../../archived/feature/2026-08-25-unified-task-list.md)。暂缓项记录在各包 README:前台流式(需 shell seam 的 run/observe 变体)、terminal `pty-send` 观察(需给 terminal 缓冲加非消费面)、人工 kill 控件(被 jobs `reported` 契约阻塞)、SDK/ACP 消费端、重启后的持久回放。
@@ -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/2026-09-04-right-sidebar-docking-infrastructure.md
2026-09-04-right-sidebar-docking-infrastructure.md: 3ebfe4b3a95ab5820682098be2a4c43f172d1a02
2026-09-04-right-sidebar-docking-infrastructure.zh.md: dc6516324a368dd86b3e01afb4051a4b20697363
2026-09-04-right-sidebar-docking-infrastructure.md: c5bfb75730f2037826723bc7fd0aa2bb01005d1e
2026-09-04-right-sidebar-docking-infrastructure.zh.md: 4959032b53f84654205a1973aee781acb356f999
@@ -16,6 +16,8 @@ The right column is a per-session docking surface — split panes, tabs, floatin
### Package topology
The [stable Sidebar mounting decision](../architecture/2026-09-20-sidebar-retained-tab-layout.md) owns one retained tab-content tree for docking, floating and Session switches. The engine, layout state and gesture ownership in this note remain applicable.
| Package | Kind | Owns |
|---|---|---|
| `packages/client/ui-dockkit` | static-linked library, zero DSH dependencies | the layout engine and the React components that render and drive it; consumers compile its sources, and it keeps exactly one stylesheet because a consumer de-duplicates injected sheets by file name |
@@ -33,7 +35,7 @@ Components render a snapshot and report settled intents, one per gesture: a drag
[Responsive Sidebar and tab information](../architecture/2026-09-07-sidebar-responsive-tab-info.md) supersedes this note's no-concession layout, overlay presentation and product pane limit. `ui-layout` still owns three-column geometry and pixel width preferences; the Sidebar occupant reports presentation through `ctx.layout.openRightbar(track, fullscreen)` and `closeRightbar()`, without the frame injecting the Sidebar package. Exact width rules belong to [ui-layout](../../../../packages/client/ui-layout/README.md).
The right Sidebar uses one mounted content tree in normal and fullscreen modes; hiding preserves tab state, and fullscreen covers the viewport while retaining underlying column reservation. Floats still use viewport coordinates through a portal and remain open when the Sidebar closes. The product limits docking to two horizontal panes and a 20–80% divider; the generic engine keeps its own defaults.
The right Sidebar uses one mounted content tree in normal and fullscreen modes; hiding preserves retained tab state, and fullscreen covers the viewport while retaining underlying column reservation. `DockLayout` renders stable Grid cells with fixed-position floating frames in the same tree; foreground floats remain open when the Sidebar closes. The product limits docking to two horizontal panes and a 20–80% divider; the generic engine keeps its own defaults.
### State
@@ -86,7 +88,7 @@ The surface renders tabs whose bodies it does not know: each tab carries a `kind
- The Detail panel and its duplicate card presentation are gone (a net removal of roughly 1,400 lines); cards are read in place, and `inspect` opens the trajectory view.
- The frame has no centre floor: a viewport narrower than the two edge columns squeezes the conversation toward zero instead of closing a column.
- The kit is compiled by its consumers, so a kit change requires a shell rebuild and a page reload; there is no HMR for it.
- The panel, the float host, and the portalled tab menu use hard-coded z-index values; the client still has no z-index token layer.
- Docked cells, floating cells and the portalled tab menu use fixed stacking levels; floating cells use CSS paint order within their level. The client still has no global z-index token layer.
## Testing
@@ -16,6 +16,8 @@ Agent 产出的文件是最尖锐的案例。产出文件 chip 或 `read` 行的
### 包拓扑
[Sidebar 稳定挂载决策](../architecture/2026-09-20-sidebar-retained-tab-layout.zh.md)负责停靠、浮动与会话切换共用的一棵保活 Tab 内容树。本文的引擎、布局状态与手势所有权仍然适用。
| 包 | 形态 | 所有物 |
|---|---|---|
| `packages/client/ui-dockkit` | 静态链接库,零 DSH 依赖 | 布局引擎与渲染/驱动它的 React 组件;消费方编译其源码,且它只保留一张样式表,因为消费方按文件名去重注入的样式表 |
@@ -33,7 +35,7 @@ Agent 产出的文件是最尖锐的案例。产出文件 chip 或 `read` 行的
[响应式 Sidebar 与标签信息](../architecture/2026-09-07-sidebar-responsive-tab-info.zh.md)取代本记录中的无让步布局、覆盖模式与产品窗格上限。`ui-layout` 仍拥有三列几何与像素宽度偏好,Sidebar 占位项通过 `ctx.layout.openRightbar(track, fullscreen)` 和 `closeRightbar()` 报告呈现方式,框架不注入 Sidebar 包。具体宽度规则见 [ui-layout](../../../../packages/client/ui-layout/README.zh.md)。
右栏在普通与全屏模式下使用同一棵已挂载内容树;隐藏保留标签状态,全屏覆盖视口并保留底层列占位。浮窗仍经 portal 使用视口坐标,不随右栏关闭。产品限制为两个水平窗格与 20–80% 分割比例,通用引擎保留自己的默认值。
右栏在普通与全屏模式下使用同一棵已挂载内容树;隐藏保留已保活的标签状态,全屏覆盖视口并保留底层列占位。`DockLayout` 在同一棵树中渲染稳定 Grid cell 与 fixed 定位的浮动 frame,前台浮窗不随右栏关闭。产品限制为两个水平窗格与 20–80% 分割比例,通用引擎保留自己的默认值。
### 状态
@@ -86,7 +88,7 @@ Agent 产出的文件是最尖锐的案例。产出文件 chip 或 `read` 行的
- Detail 面板及其重复的卡片展示消失(净删约 1,400 行);卡片就地阅读,`inspect` 打开 trajectory 视图。
- 框架没有中列下限:视口窄于两侧列之和时会话区被挤向零,而不是关掉某一列。
- 库由消费方编译,改库须重建壳并刷新页面;它没有 HMR。
- 面板、浮层宿主与 portal 出去的 tab 菜单使用硬编码 z-index;客户端仍没有 z-index token 层。
- 停靠 cell、浮动 cell 和 portal 菜单使用固定层级,浮动 cell 在所属层内通过 CSS 绘制顺序排序;客户端仍没有全局 z-index token 层。
## Testing
@@ -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/2026-09-05-sidebar-text-preview-and-file-tree.md
2026-09-05-sidebar-text-preview-and-file-tree.md: 8ddb1a336307011d8f6ce9bbe927f00642a83e60
2026-09-05-sidebar-text-preview-and-file-tree.zh.md: 294a2b456008837882f7f82e55e8e9e435254554
2026-09-05-sidebar-text-preview-and-file-tree.md: e978043545b7f172f2b67053fe44b5936ea9b529
2026-09-05-sidebar-text-preview-and-file-tree.zh.md: 088d4a58c973843ac5085cfaa1ca9980387bcc8f
@@ -34,7 +34,7 @@ The [Document Preview decision](../architecture/2026-09-08-document-preview-oper
A tab uses `dsh-resource://file/session/<sessionId>/<path>`, where the path may be relative or absolute ([Workspace Files](../architecture/2026-09-05-workspace-files-service.md) owns the grammar and the `fileAddressFor` / `parseFileAddress` helpers). `hostFileOf` accepts only this Session scope and takes both Session and path from the address; a Session-less `absolute` address is not claimed. A malformed claimed address throws as a programming error.
Metadata and content come from different places. `useResource<'file'>(tab.contentId)`, the global standard hook from the [client resource model](../architecture/2026-09-05-client-resource-model.md), yields `WorkspaceFileStat`; the body compares that observed version with the version of its loaded content. The Preview face reads text through `remote.workspaceFiles.read` and complete bytes through `readAll`. A later text page from a newer version restarts from page one, and a retired request cannot write after reload or tab disposal. The [Document Preview decision](../architecture/2026-09-08-document-preview-operations.md) owns renderer-specific loading.
Metadata and content come from different places. `useResource<'file'>(tab.contentId)`, the global standard hook from the [client resource model](../architecture/2026-09-05-client-resource-model.md), yields `WorkspaceFileStat`; the body compares that observed version with the version of its loaded content. The Preview face reads text through `remote.workspaceFiles.read` and complete bytes through `readBytes` with `{}` options. A later text page from a newer version restarts from page one, and a retired request cannot write after reload or tab disposal. The [Document Preview decision](../architecture/2026-09-08-document-preview-operations.md) owns renderer-specific loading.
The store is Slot-standard: one exclusive instance per session, bucketed by tab id, holding `{ version, pages, eof, loading, failure, scrollTop, wrap, revision }`. Bucketing by tab, not by file, is deliberate — two tabs of one file scroll independently. The face (`loadPage`, `reloadPages`) is the only asynchronous half: it marks a read in flight, awaits the Remote result, and writes a page or a failure through the store's actions, writing nothing if the owner's `signal` has fired. The `signal` also ends the bucket: the face arms one abort listener per tab at the tab's first read, and that listener forgets the bucket — not the body, which mounts and unmounts as tabs switch; a tab that never read has no bucket and no listener, and a record can end while its body is unmounted behind another tab. Scroll offset, wrap, and the navigation already answered therefore outlive the body: a tab comes back where the reader left it rather than re-reading or jumping again. Nothing persists across a page reload.
@@ -34,7 +34,7 @@ Sidebar 随包交付三个 tab 类型:**引导页**(`ui-sidebar-right`)、
tab 使用 `dsh-resource://file/session/<sessionId>/<path>`,其中路径可以是相对路径或绝对路径([Workspace Files](../architecture/2026-09-05-workspace-files-service.zh.md)负责该语法与 `fileAddressFor` / `parseFileAddress` 辅助函数)。`hostFileOf` 只接受这种 Session scope,并从地址取得 Session 与路径;不认领不带 Session 的 `absolute` 地址。被认领的地址若格式错误,则作为程序错误抛出。
元数据与内容来自不同的地方。`useResource<'file'>(tab.contentId)`——[client 资源模型](../architecture/2026-09-05-client-resource-model.zh.md)提供的全局标准 hook——产生 `WorkspaceFileStat`;正文把其观察版本与已加载内容版本比较。Preview face 通过 `remote.workspaceFiles.read` 读取文本,通过 `readAll` 读取完整字节。后续文本页若来自更新版本,则从第一页重新开始;被重载或 tab 销毁淘汰的请求不能再写入。[Document Preview 决议](../architecture/2026-09-08-document-preview-operations.zh.md)负责各渲染器的加载方式。
元数据与内容来自不同的地方。`useResource<'file'>(tab.contentId)`——[client 资源模型](../architecture/2026-09-05-client-resource-model.zh.md)提供的全局标准 hook——产生 `WorkspaceFileStat`;正文把其观察版本与已加载内容版本比较。Preview face 通过 `remote.workspaceFiles.read` 读取文本,通过带 `{}` 选项的 `readBytes` 读取完整字节。后续文本页若来自更新版本,则从第一页重新开始;被重载或 tab 销毁淘汰的请求不能再写入。[Document Preview 决议](../architecture/2026-09-08-document-preview-operations.zh.md)负责各渲染器的加载方式。
store 是 Slot 标准件:每会话一个独占实例,按 tab id 分桶,持有 `{ version, pages, eof, loading, failure, scrollTop, wrap, revision }`。按 tab 而非按文件分桶是有意的——同一文件的两个 tab 各自滚动。face(`loadPage`、`reloadPages`)是唯一的异步半边:它标记读取进行中,等待 Remote 结果,再经 store 的 action 写入一页或一次失败;若 owner 的 `signal` 已触发则什么也不写。`signal` 同时终结这个桶:face 在 tab 首次读取时挂一个 abort 监听器,由它忘掉桶——不是体,体随 tab 切换反复挂载卸载;从未读过的 tab 没有桶也没有监听器,而 tab 记录可能在其体被另一 tab 挡住而卸载时结束。因此滚动位置、换行与已答过的导航都活得比体久:tab 回来时停在读者离开的地方,而不是重读或再跳一次。刷新页面后什么都不保留。
Loaded 100 of 1448 files, more files were not shown because too many files have changed in this diff. Show more