@linxin666/dsh-client-ui-task-board 0.1.16 → 0.1.18

package/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   node scripts/verify-docs.mjs --write <dir>
-README.md: bbbadebabdee529b7a0ea831519cf909a4716844
-README.zh.md: 64b6539d2a4505e3824657e62dcfb4dfab88704e
+README.md: ed814f18cf9668624c70d26ef8190d378aaf62f1
+README.zh.md: 54aa8959f94f196196e70107a3a18846f4a45dc5

package/README.md

@@ -11,11 +11,12 @@ A hot-pluggable DeepSeek Harness (DSH) client GUI plugin: it adds a **task board
 ## Features
 
 - **Sidebar entry**: injects a "任务看板" (task board) entry row inside the sidebar column (`[data-pane="sidebar"]` on older shells, `[class*="sidebarCol"]` on the DSH 0.1.0-rc.6 AppFrame layout) below the new-session button (wide rail shows icon + text, collapsed rail shows a bare icon, adapting to DSH skin tokens).
-- **Multi-column board**: five columns — 待规划 (to plan) / 待办 (to do) / 进行中 (in progress) / 已完成 (done) / 已失败 (failed); cards show title, description, status, update time, and execution count; the top supports search filter, new task, and back to chat.
+- **Multi-column board**: five columns — 待规划 (to plan) / 待办 (to do) / 进行中 (in progress) / 已完成 (done) / 已失败 (failed); cards show title, description, status, update time, and execution count; the top supports search filter, an archive view (done/failed tasks can be archived off the board and restored anytime, keeping execution history and session transcripts traceable), new task, and back to chat.
 - **Task details**: click a card to open details (title/description/execution prompt/execution log) — it does **not** execute on a single click; the details offer "执行 / 重新执行" (Run / Re-run), "删除" (Delete, with confirm), "查看会话" (View session, jumps to the execution transcript), and a manual move to 待规划/待办.
 - **Real execution**: on "执行" (Run), the plugin connects a workspace session through the client runtime (`workspaces.connectWorkspace`, reusing a blank session or letting the host create one), names the session after the task title, and drives a real agent via `session.prompt([{ type: 'text', text }], 'queue')`; it then subscribes to that session's snapshot and, once the round really finishes, sets the card to 已完成/已失败 and records the execution result. The execution session appears in the session list and can be opened to view the real transcript.
+- **Per-task execution targets**: a task can pin where and how it runs — **workspace** (the execution session lands in that workspace), **mode** (the agent preset the session is composed from, switched through `agentPresets.select` while the session is still blank), and **permission** (a sandbox preset applied through the `/permission <id>` slash command: read-only / workspace-write / danger-full-access). Blank pins fall back to the runtime defaults (recent workspace / deployment preset / session default). A pin that cannot be applied fails the run **before** the prompt is sent, so a task never silently runs under settings it did not ask for.
 - **Status write-back**: card status (进行中 → 完成/失败) is driven by the real session state; after a page refresh/restart, leftover running tasks are auto-reconciled against the current session state (reconcile).
-- **Scheduled tasks**: the details panel can schedule a task — an enable switch + a 5-field cron expression (分 时 日 月 周, supporting `*` / `*/n` / `a-b` / comma lists) + common presets (daily 09:00, every hour, every 10 minutes, Mondays 09:00); enabling computes and persists the "下次运行时间" (next run time), and the card shows a scheduled marker; at the due time it automatically takes the same real-execution path (as manual run), and the execution session remains linkable.
+- **Scheduled tasks**: the new-task dialog and the details panel can schedule a task — an enable switch + a 5-field cron expression (分 时 日 月 周, supporting `*` / `*/n` / `a-b` / comma lists) + common presets (daily 09:00, every hour, every 10 minutes, Mondays 09:00); enabling computes and persists the "下次运行时间" (next run time), and the card shows a scheduled marker; at the due time it automatically takes the same real-execution path (as manual run), and the execution session remains linkable.
 - **System-prompt injection**: the host half (`src/index.ts`) registers a `plugin:task-board` section (order 200) via `SystemPrompt.section`, declaring this plugin's existence, capabilities, and limits to every agent — it is injected when the plugin is in the composition (after mount + DSH restart) and disappears when removed (after unmount + restart), so an agent needs no external docs to know how to work with this board.
 
 ## Directory structure
@@ -45,8 +46,9 @@ scripts/dsh-task-board.js                          # one-click mount/unmount/sta
 - **The middle column cannot be replaced through a slot**: the `conversation` slot is single and already taken by ui-conversation. The board view mounts on the center column (`[data-pane="conversation"]` on older shells, `[class*="centerCol"]` on the DSH 0.1.0-rc.6 AppFrame layout) as a tail child node (outside React's ownership), toggled via the `<html data-dsh-taskboard-active>` attribute, keeping the chat subtree below mounted and stateful.
 - **Persistence uses browser localStorage**: client plugins run in the browser and DSH has no browser-writable file channel (matching skin-center's research on `cordis.patch.yml`); localStorage is also how DSH's own client snapshot store (`createSnapshotStore` persist) persists.
 - **Execution rides the client runtime**: `ctx.sessions.list` subscribes to session state (`running` / `byId`), `ctx.workspaces.connectWorkspace()` creates/reuses a session, `session.prompt()` drives a real agent, and `ctx.sessions.open()` jumps to the transcript.
+- **Execution targets ride the same runtime faces**: the workspace pin passes the task's id to `workspaces.connectWorkspace()` (validated against the workspace list first, so a stale pin fails locally); the mode pin recomposes the blank execution session via `api.agentPresets.select` — only legal before the first turn, so it runs before `session.prompt`, and `sessions.noteAgentPreset` keeps the list label current; the permission pin admits a `/permission <id>` slash command through `session.command` — the same mechanism the shell's own permission picker uses. A rejected admission or a line no command claims fails the run before the prompt.
 - **Background settlement relies on list reconciliation**: an unopened session has no chat-snapshot window (cold), so settlement keys off the session list — every list change reconciles running tasks; result judgment takes, in order, "missing from list → cancelled / still running → wait / chat snapshot visible → by lastAgentError / tail of raw history → a turn-error node proves failure / otherwise success", and reconciliation is idempotent.
-- **Scheduled tasks run in the browser scheduler**: the plugin is pure-client (no server channel), so "run at the due time" is done by the in-tab scheduler — a tick every minute, with an immediate catch-up tick when the page returns from the background; before firing it first moves "next run" forward to the next cron match so the same tick never fires twice; it does not fire early in page load (before the session-list baseline is ready), avoiding mis-execution. Limitation: the tab must stay open (schedules missed while closed are "miss = skip", and only already-deferred due tasks are caught up on the next open); a task that is 进行中 (in progress) skips the current due time and waits for the next cron match.
+- **Scheduled tasks run in the browser scheduler**: the plugin is pure-client (no server channel), so "run at the due time" is done by the in-tab scheduler — a tick every minute, with an immediate catch-up tick when the page returns from the background; before firing it first moves "next run" forward to the next cron match so the same tick never fires twice; it does not fire early in page load (before the session-list baseline is ready), avoiding mis-execution. Every tick first re-reads the persisted ledger, so a task deleted in another tab can never be fired from a stale in-memory copy. Limitation: the tab must stay open (schedules missed while closed are "miss = skip", and only already-deferred due tasks are caught up on the next open); a task that is 进行中 (in progress) skips the current due time and waits for the next cron match.
 - **Same-origin tabs share one ledger**: additions, edits, and deletions in any tab propagate to the others through storage events (`LocalStorageTaskStore.subscribeExternal`), so a task deleted in one tab can never keep firing from another tab's stale in-memory copy — nor be written back by that copy's later persistence (schedule roll-forward, execution settlement).
 
 ## Install
@@ -115,10 +117,11 @@ The rows registered in the profile manifest:
 
 1. `npm run build` → `node scripts/dsh-task-board.js mount` → refresh `http://127.0.0.1:3080`.
 2. A "任务看板" (task board) entry row appears below "新会话" in the sidebar; click it → the middle column switches to the five-column board.
-3. "+ 新建任务" (New task) with title/description/Prompt → the card appears in 待办 (to do).
-4. Click the card → details show content and Prompt; click "执行" (Run) → the card becomes 进行中 (in progress) (a session named after the task title appears in the session list); after the agent finishes the card lands in 已完成 (done) or 已失败 (failed), the detail execution log has a result and time, and "查看会话" (View session) jumps to the real transcript.
-5. Scheduled task: details → tick "定时运行" (Scheduled run) to enable, pick the preset "每 10 分钟" (every 10 minutes, cron `*/10 * * * *`); a scheduled marker appears on the card; wait for the next whole 10-minute mark, watch the card automatically enter 进行中 (in progress) and eventually complete, with "上次触发" (last trigger) showing a time and a new execution-log row (the session is linkable).
-6. Refresh the page / restart DSH → tasks remain; unmount the plugin → the GUI restores to its original state.
+3. "+ 新建任务" (New task) with title/description/Prompt → the card appears in 待办 (to do). The dialog also offers 工作区/模式/权限 (workspace / mode / permission) pins — leave them blank for runtime defaults.
+4. Pin a workspace/mode/permission on a task (in the dialog or the task detail) → run it → the execution session appears under the pinned workspace, its list row shows the pinned preset, and the session's permission selector shows the pinned permission.
+5. Click the card → details show content and Prompt; click "执行" (Run) → the card becomes 进行中 (in progress) (a session named after the task title appears in the session list); after the agent finishes the card lands in 已完成 (done) or 已失败 (failed), the detail execution log has a result and time, and "查看会话" (View session) jumps to the real transcript.
+6. Scheduled task: details → tick "定时运行" (Scheduled run) to enable, pick the preset "每 10 分钟" (every 10 minutes, cron `*/10 * * * *`); a scheduled marker appears on the card; wait for the next whole 10-minute mark, watch the card automatically enter 进行中 (in progress) and eventually complete, with "上次触发" (last trigger) showing a time and a new execution-log row (the session is linkable).
+7. Refresh the page / restart DSH → tasks remain; unmount the plugin → the GUI restores to its original state.
 
 ## Acceptance checklist
 
@@ -128,5 +131,6 @@ The rows registered in the profile manifest:
 - Execution really starts a session (its transcript is visible in the session list); card status follows the real execution progress; the details can jump to the execution session
 - Delete has a confirm step, and the local store is synced-removed after deletion
 - Scheduled tasks: cron config/preset/validation, next-run time, auto real execution at the due time, status write-back, scheduled card marker, scheduling resumes after refresh (browser-side scheduling, the tab must stay open)
+- Per-task execution targets: workspace/mode/permission pins persist across refresh, drive the execution session, and an un-appliable pin (missing workspace, locked preset, unknown permission command) fails the run with the reason visible in the execution log
 - One-click mount/unmount; after unmount the GUI restores and other managed segments are unaffected
 - README + automated tests covering storage read/write, state transitions, execution trigger, cron parsing, and the scheduler

package/README.zh.md

@@ -11,11 +11,12 @@
 ## 功能
 
 - **侧边栏入口**：侧边栏列（旧版 `[data-pane="sidebar"]`，DSH 0.1.0-rc.6 AppFrame 布局为 `[class*="sidebarCol"]`）内、新会话按钮下方注入「任务看板」入口行（宽栏显示图标+文字，折叠 rail 显示纯图标，随 DSH 皮肤 token 自适应）。
-- **多列看板**：待规划 / 待办 / 进行中 / 已完成 / 已失败 五列；卡片显示标题、描述、状态、更新时间、执行次数；顶部支持搜索过滤、新建任务、返回对话。
+- **多列看板**：待规划 / 待办 / 进行中 / 已完成 / 已失败 五列；卡片显示标题、描述、状态、更新时间、执行次数；顶部支持搜索过滤、归档视图（已完成/已失败任务可归档，归档后移出看板，可随时恢复，执行记录与会话 transcript 保留可追溯）、新建任务、返回对话。
 - **任务详情**：点卡片打开详情（标题/描述/执行 Prompt/执行记录），**不会**一点就执行；详情内提供「执行 / 重新执行」「删除（带确认）」「查看会话（跳转到执行 transcript）」以及手动移到待规划/待办。
 - **真实执行**：点「执行」后，插件通过客户端 runtime 连接工作区会话（`workspaces.connectWorkspace`，空白会话复用或 host 新建），把任务标题设为会话名，以任务 Prompt 调用 `session.prompt([{ type: 'text', text }], 'queue')` 驱动真实 agent；随后订阅该会话快照，轮次真实结束后把卡片置为 已完成/已失败 并记录执行结果。执行会话会出现在会话列表，可点进对话查看真实 transcript。
+- **任务级执行目标**：任务可钉住「在哪跑、怎么跑」——**工作区**（执行会话落在指定工作区）、**模式**（会话以指定 agent 预设组合，趁会话仍为空白时走 `agentPresets.select` 切换）、**权限**（经 `/permission <id>` 斜杠命令应用的沙箱预设：read-only / workspace-write / danger-full-access）。留空即回退运行时默认（最近工作区 / 部署预设 / 会话默认）。钉不住的目标会在发 Prompt **之前**失败，任务绝不会悄悄按没要求过的设置运行。
 - **状态回写**：卡片状态（进行中 → 完成/失败）由真实会话状态驱动；刷新页面/重启后，遗留的 running 任务会按会话现状自动对账（reconcile）。
-- **定时任务**：详情面板可为任务配置定时执行——启用开关 + 5 段 cron 表达式（分 时 日 月 周，支持 `*` / `*/n` / `a-b` / 逗号列表）+ 常用预设（每天 09:00、每小时、每 10 分钟、每周一 09:00）；启用即计算并持久化「下次运行时间」，卡片显示定时标识；到点自动走真实执行链路（同手动执行），执行会话照常可跳转。
+- **定时任务**：新建任务与详情面板均可配置定时执行——启用开关 + 5 段 cron 表达式（分 时 日 月 周，支持 `*` / `*/n` / `a-b` / 逗号列表）+ 常用预设（每天 09:00、每小时、每 10 分钟、每周一 09:00）；启用即计算并持久化「下次运行时间」，卡片显示定时标识；到点自动走真实执行链路（同手动执行），执行会话照常可跳转。
 - **系统提示词注入**：host 半边（`src/index.ts`）通过 `SystemPrompt.section` 注册 `plugin:task-board` 段（order 200），向每个 agent 声明本插件存在、能力与限制——插件在组合中（mount 后重启 DSH）即注入，移出组合（unmount 后重启）即消失，agent 无需任何外部文档就能知道如何与本看板协作。
 
 ## 目录结构
@@ -45,8 +46,9 @@ scripts/dsh-task-board.js                          # 一键挂载/卸载/状态
 - **中间列无法通过槽位替换**：`conversation` 槽位是 single 且已被 ui-conversation 占用。看板视图以 DOM 方式挂在中间列（旧版 `[data-pane="conversation"]`，DSH 0.1.0-rc.6 AppFrame 布局为 `[class*="centerCol"]`；挂载选择器两者都保留）内（React 不管的尾部子节点），通过 `<html data-dsh-taskboard-active>` 属性切换显隐，底下的对话子树保持挂载有状态。
 - **持久化用浏览器 localStorage**：客户端插件跑在浏览器里，DSH 没有浏览器可写的文件通道（与 skin-center 对 `cordis.patch.yml` 的调研结论一致）；localStorage 也是 DSH 客户端自身快照存储（`createSnapshotStore` persist）的持久化方式。
 - **执行走客户端 runtime**：`ctx.sessions.list` 订阅会话状态（`running` / `byId`），`ctx.workspaces.connectWorkspace()` 创建/复用会话，`session.prompt()` 真实驱动 agent，`ctx.sessions.open()` 跳转 transcript。
+- **执行目标走同样的 runtime 脸**：工作区钉子把任务指定的 id 传给 `workspaces.connectWorkspace()`（先对照工作区列表校验，失效的钉子就地失败）；模式钉子通过 `api.agentPresets.select` 重组仍为空白状态的执行会话——只有首轮之前合法，所以排在 `session.prompt` 之前执行，`sessions.noteAgentPreset` 让会话列表标签即时更新；权限钉子经 `session.command` 提交 `/permission <id>` 斜杠命令——与壳自带权限选择器同一机制。提交被拒或没有命令认领该行时，在发 Prompt 前失败。
 - **后台结算靠列表对账**：未打开的会话没有对话快照窗口（cold），所以执行结算以会话列表为准——每次列表变化都对账 running 任务；结果判定依次取「列表缺失→已取消 / 仍在跑→等待 / 对话快照可见→按 lastAgentError / 原始历史尾部→turn-error 节点证明失败 / 否则按成功」，对账幂等。
-- **定时任务在浏览器端调度**：插件是纯客户端（无服务端通道），所以「到点执行」由标签页内的调度器完成——每分钟 tick 一次，页面从后台恢复可见时立即补 tick；到点触发前先把「下次运行」顺延到下一个 cron 匹配点再执行，同一 tick 不会重复触发；页面加载早期（会话列表基线未就绪）不触发，避免误执行。限制：需要标签页保持打开（关闭期间错过的调度按「错过即跳过」处理，下次打开时只补跑已顺延的到期任务）；任务处于「进行中」时到点跳过本次，等下一个 cron 匹配点。
+- **定时任务在浏览器端调度**：插件是纯客户端（无服务端通道），所以「到点执行」由标签页内的调度器完成——每分钟 tick 一次，页面从后台恢复可见时立即补 tick；到点触发前先把「下次运行」顺延到下一个 cron 匹配点再执行，同一 tick 不会重复触发；页面加载早期（会话列表基线未就绪）不触发，避免误执行；每次 tick 先重读持久化账本，在别的标签页删除的任务绝不会从陈旧内存副本被触发。限制：需要标签页保持打开（关闭期间错过的调度按「错过即跳过」处理，下次打开时只补跑已顺延的到期任务）；任务处于「进行中」时到点跳过本次，等下一个 cron 匹配点。
 - **多标签页同源共享同一份台账**：任一标签页的增删改通过 storage 事件同步到其他标签页（`LocalStorageTaskStore.subscribeExternal`），删除的任务不会在其他标签页的内存副本里残留触发，也不会被其他标签页的后续持久化写回复活。
 
 ## 安装
@@ -115,10 +117,11 @@ profile 清单中注册的行：
 
 1. `npm run build` → `node scripts/dsh-task-board.js mount` → 刷新 `http://127.0.0.1:3080`。
 2. 侧边栏「新会话」下方出现「任务看板」入口行；点击 → 中间列切换为五列看板。
-3. 「+ 新建任务」填标题/描述/Prompt → 卡片出现在「待办」。
-4. 点卡片 → 详情可见内容与 Prompt；点「执行」→ 卡片变「进行中」（会话列表出现以任务标题命名的会话）；agent 跑完后卡片落「已完成」或「已失败」，详情执行记录有结果与时间，可「查看会话」跳转到真实 transcript。
-5. 定时任务：详情 →「定时运行」勾选启用，选预设「每 10 分钟」（cron `*/10 * * * *`），卡片出现定时标识；等待下一个整 10 分钟点，观察卡片自动进入「进行中」并最终完成，详情「上次触发」出现时间、执行记录新增一条（会话可跳转）。
-6. 刷新页面/重启 DSH → 任务仍在；卸载插件 → GUI 恢复原状。
+3. 「+ 新建任务」填标题/描述/Prompt → 卡片出现在「待办」；对话框同时提供 工作区/模式/权限 三项钉子，留空即运行时默认。
+4. 在对话框或详情里给任务钉上工作区/模式/权限 → 执行 → 执行会话落在指定工作区下，列表行显示指定预设，会话的权限选择器显示指定权限。
+5. 点卡片 → 详情可见内容与 Prompt；点「执行」→ 卡片变「进行中」（会话列表出现以任务标题命名的会话）；agent 跑完后卡片落「已完成」或「已失败」，详情执行记录有结果与时间，可「查看会话」跳转到真实 transcript。
+6. 定时任务：详情 →「定时运行」勾选启用，选预设「每 10 分钟」（cron `*/10 * * * *`），卡片出现定时标识；等待下一个整 10 分钟点，观察卡片自动进入「进行中」并最终完成，详情「上次触发」出现时间、执行记录新增一条（会话可跳转）。
+7. 刷新页面/重启 DSH → 任务仍在；卸载插件 → GUI 恢复原状。
 
 ## 验收对照
 
@@ -128,5 +131,6 @@ profile 清单中注册的行：
 - 执行真实启动会话（会话列表可见 transcript）；卡片状态随真实执行进度变化；详情可跳转到执行会话
 - 删除有确认环节，删除后本地存储同步移除
 - 定时任务：cron 配置/预设/校验、下次运行时间、到点自动真实执行、状态回写、定时卡片标识、刷新后调度恢复（浏览器端调度，标签页需保持打开）
+- 任务级执行目标：工作区/模式/权限 钉子刷新后仍在，驱动执行会话；钉不住的钉子（工作区缺失、预设被锁、权限命令无人认领）在发 Prompt 前失败，执行记录可见原因
 - 一键挂载/卸载；卸载后 GUI 恢复原状，其它 managed 段不受影响
 - README + 覆盖存储读写/状态流转/执行触发/cron 解析/调度器的自动化测试