dsh-chief 1.4.0 → 1.5.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +42 -0
- package/README.en.md +66 -50
- package/README.md +67 -50
- package/api-routes.js +165 -3
- package/archive-core.js +136 -8
- package/archive-dossiers.js +450 -20
- package/archive-precipitated.js +115 -36
- package/auto-continue.js +13 -5
- package/board-actions.js +59 -4
- package/board-schema.js +1 -1
- package/chief-approvals.js +176 -11
- package/chief-archives.js +157 -30
- package/chief-commands.js +565 -101
- package/chief-models.js +158 -28
- package/chief-review.js +211 -70
- package/chief-state.js +362 -32
- package/chief-view.js +17 -14
- package/client.js +3744 -2207
- package/digest.js +18 -9
- package/model-tools.js +15 -5
- package/package.json +9 -4
- package/runtime-hooks.js +115 -12
- package/subagent-runtime.js +257 -1
package/CHANGELOG.md
CHANGED
|
@@ -5,6 +5,48 @@ All notable changes to **`dsh-chief`** (**DSH 参谋长**) will be documented in
|
|
|
5
5
|
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
|
|
6
6
|
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
|
|
7
7
|
|
|
8
|
+
## [1.5.0] - 2026-11-20
|
|
9
|
+
|
|
10
|
+
### Added
|
|
11
|
+
|
|
12
|
+
#### 🗂️ 认知桌面、三层钩子渐变隔离与动态俯视纸质卷宗 UI (`archive-core.js` / `archive-dossiers.js` / `tabs/archives.js`)
|
|
13
|
+
- **工程实现**:
|
|
14
|
+
- **三层钩子平滑可见性模型**:本局档案柜未整理草稿对主对话不可见、参谋随时可读,经参谋整理写入 `hookKeywords`/`hookCue`/`hasHook: true` 后仅向主对话注入轻量 `🪝` 索引并新增 `chief_board({ action: "archive.pull", id })` 按需拉阅(点亮 `leadPulledSessions` 与 `📖 主代理已拉阅` 徽章);全局档案柜未织钩草稿对主对话和参谋双盲,织钩后仅向参谋开放 `globalHookIndex` 与 `pull_global` 拉阅(主对话仍 0 泄露);冷藏档案柜(`precipitated`)新增 `precipitatedFolders` 折叠抽屉及人类速览钩子,对主对话和参谋维持 0-byte 隔离并拦截参谋 `archiveOps` 访问。
|
|
15
|
+
- **认知桌面预算驱动与双通道注入**:`digest.js` 与 `board-actions.js` 彻底肃清 `sessionArchiveBriefs` 全文泄露,确立置顶 `desktop`(上限 `2,400 Tokens / 6 份`)为唯一全文注入源;超额自动按 `copper < silver < gold`/Pin 时间挤退入 `recentlyUnpinnedWatch` 观察队列;新增 `executeChiefArchiveInjectionOp` 双通道(`pin_to_desktop`/`unpin_from_desktop` 通道 A、`deliver_to_lead` 一次性执行信封通道 B,记入 `chiefDeliveredSessions`),严格服从 L0–L3 权限门控。
|
|
16
|
+
- **动态俯视纸质卡片 UI**:新增 `.dnb-desktop-zone`(含实时 Token 占用条与「🌙 交参谋梦境漫游归入抽屉」快捷按钮)、`.dnb-vault-banner` 谁在看总览横幅、全三层 `.dnb-drawer-box` 折叠抽屉及 `.dnb-paper-dossier` 动态纸质卷宗卡(折叠态仅呈现标题、关键词钩子与谁在看徽章,点击后平滑展开操作栏与分块记忆点)。
|
|
17
|
+
- **用户感知**:打开档案页顶部即见置顶「认知桌面」和 Token 额度条,三层档案像俯瞰书桌上的实体文件夹与纸页,未点开只看标题和关键词钩子,谁能看、谁在看一眼了然,主对话上下文不再被无关历史档案挤爆。
|
|
18
|
+
|
|
19
|
+
#### 🤖 正交运行模式、托管「🌙 梦境漫游」、5 条件自动休眠与一站式执行 (`chief-review.js` / `auto-continue.js` / `runtime-hooks.js`)
|
|
20
|
+
- **工程实现**:
|
|
21
|
+
- **三态正交总控 (`runMode: 'off' | 'patrol' | 'managed'`)**:统一替代旧版零散开关;`off` 后台 0 review LLM 消耗;`patrol` 常规回合 0 Token,仅在 `guard_wall_blocked`、`turn_error`、`stream_interrupted` 等异常事件唤醒巡检;`managed` 启用持续督导。
|
|
22
|
+
- **托管后台「🌙 梦境漫游」 (`runManagedDreamRoam`)**:托管态 L1+ 权限时自动分拣未织钩草稿与 `recentlyUnpinnedWatch` 入抽屉并生成钩子,同时点亮星空紫蓝毛玻璃遮罩 `data-tone="dream"`(`'🌙 参谋长正在梦境漫游 · 轻拢卷宗与织就钩子中...'`),严禁改写工作区代码或回推桌面。
|
|
23
|
+
- **5 条件自动休眠守护**:`checkAndStopIdleAutoPilot` 仅在(1)闲置达 `managedIdleTimeoutMs`(默认 15 分钟)且无运行中代理、(2)0 `doing` 任务、(3)0 待审批/提权、(4)0 未结 `inquiries`、(5)0 待复盘定时器且 `recentlyUnpinnedWatch` 清空时自动切回 `'off'`。
|
|
24
|
+
- **`Auto continue` 双模式唤醒与 L3 一站式建议收口**:`patrol` 与 `managed` 下断流事件过红线与物理墙检查后按 L0–L3 分级处置(L3 下发 `[EXECUTION_REQUEST #taskId]`);`L3 + 强介入` 时 `executeBatchSuggestions` 自动执行安全建议(跳过 `guard_wall`、`goal_approval` 及 `pinGlobal`)并产出 `isOneStopReport: true` 合并战报卡片 `'🏁 一站式战报回执'`。
|
|
25
|
+
- **用户感知**:关停即零消耗,巡检只在真出事或断线时介入,开启托管后参谋长会在空闲期悄悄「🌙 梦境漫游」归整卷宗并织好检索钩子,忙完 15 分钟所有待办归零后自动退回关停休眠。
|
|
26
|
+
|
|
27
|
+
#### 🧠 思考强度、自适应换挡与跨模型每角色 `🔒 锁挡` 保护 (`chief-models.js` / `subagent-runtime.js` / `tabs/chief.js`)
|
|
28
|
+
- **工程实现**:
|
|
29
|
+
- **6 档配置与模型能力对齐 (`thinkingEffort`)**:支持 `auto / off / low / medium / high / xhigh`,调用 `llm.resolveModelInfo` 自动降级或对非推理模型剥离参数;`decideAdaptiveChiefEffort` 在 `chat_summary` / `dream_roam` / `quick_status` 自动降为 `off`,常规复盘为 `medium`,`blank_explore` / `wall_blocked` / `craft_wall` / `precipitate` / `strong_intervention` 升为 `high`,`hazard_diagnosis` 升为 `xhigh`,并在聊天与回执卡落盘 `thinkingEffort` 与 `thinkingEffortMode` 角标元数据。
|
|
30
|
+
- **L2/L3 跨代理热调挡与 `🔒 锁挡` 一票否决**:通过 `effortOps`、看板 `task.dispatch({ reasoningEffort })` 与 `controlMainAgent({ op: 'set_effort' })` 静默热调 `@lead` 与子代理下一轮推理深度;新增 `lockedModelEfforts` 每角色锁挡,用户一键锁定后参谋 L2/L3 自动改挡被硬拦截并记入审计。
|
|
31
|
+
- **用户感知**:可在模型选择器与参谋回复右上角直接看清每条回答使用的推理强度,默认「🪄 自适应」自动把好钢用在复杂排障与筑墙上;也能直接在前端一键锁定主对话或任意子代理的思考挡位,防止参谋越权改动。
|
|
32
|
+
|
|
33
|
+
### Fixed
|
|
34
|
+
- **加固模型思考锁挡 (`lockedModelEfforts`) 权限隔离与 `toggleLockAgentEffort` 落地 (`chief-state.js` / `api-routes.js`)**:封堵非人类调用方通过 `updateChiefConfig` 直接篡改 `lockedModelEfforts` 或已锁挡 `chief` 思考强度的旁路(Fail-Closed 拦截并写入 `model_effort_lock` 审计日志);仅允许经宿主鉴权面板路由 (`/chief/config`) 或显式 `byUser: true` 人工操作修改。
|
|
35
|
+
- **拦截 `inquiry.reply` 重复重放 (`chief-state.js`)**:`answerChiefInquiry` 校验问询当前状态,对已 `answered` 结案的 `inquiryId` 幂等拒绝并写入 `inquiry_reply_replay_rejected` 审计记录,杜绝覆盖原答案及重复写入聊天回执、审计日志与任务账本。
|
|
36
|
+
- **修复 DSH 浅色主题下弹层仍为深色问题(Fixes [#3](https://github.com/gluztm/dsh-chief/issues/3), thanks `@iridiumgit`,`styles.js`)**:将 `--dnb-popover` 从固定深色的 `--dsw-alias-hovercard-bg` 切换为随主题变的 `--dsw-alias-bg-elevated`,并显式适配 `html[data-ds-theme-source="light"]` 与 `[data-ds-theme="light"]`(独立于系统 `prefers-color-scheme`,覆盖四级权限弹层、战术驾驶舱、更多菜单、流光毛玻璃与俯视角纸质卷宗,并防止系统浅色误染 DSH 深色主题)。
|
|
37
|
+
- **子代理会话自动隐藏参谋长入口与防会话污染(Fixes [#4](https://github.com/gluztm/dsh-chief/issues/4), thanks `@iridiumgit`,`store.js` / `api.js` / `components/Panel.js`)**:新增 `isSubagentSessionId` 复用宿主 `sessions.subagentAddress` 与快照元信息识别子代理会话,在子代理对话框内自动隐藏输入框参谋长按钮(`ComposerEntry`)及主视口右下角悬浮入口(`CornerTrigger`),并阻止 `adoptSession` 被子代理窗格覆盖主会话焦点。
|
|
38
|
+
- **自适应顶栏版本号、锁挡状态同步、强介入先锁目标即推与极低概率请示拆墙(`Panel.js` / `tasks.js` / `modals.js` / `chief.js` / `chief-commands.js` / `chief-approvals.js` / `api-routes.js`)**:
|
|
39
|
+
- 使用 `@container dnb-head` 在宽敞面板展示版本号、窄面板自动隐藏(未同步警示始终亮显);
|
|
40
|
+
- 修复子代理/主对话思考强度 `🔒 锁挡` 按钮因 `effortLocked` 缓存及别名键导致的无法切换/解锁问题,并区分清晰锁定态图标文案;
|
|
41
|
+
- 开启「主动强介入」必须先填写并点选锁定核心总目标(未锁点击则目标框红框闪烁提醒),确认开启后按当前权限自动立即唤醒/审批首轮推进,无需再发“继续”;
|
|
42
|
+
- 参谋长请示拆墙机制落地(`kind: 'wall_removal'`):仅当主/子代理真实撞墙无法绕过并上报、参谋独立复核确认无替代方案时方可提交审批单,任何 L0–L3 权限均不可自拆或由自动流水线跳过,必须人类点击批准才暂挂/拆除目标墙;
|
|
43
|
+
- 参谋摘要页新增「💡 咨询」按钮,坦诚给出建议或直接说明暂无建议/管不了的外部事项,不强行凑执行按钮,可执行项按标准权限与审批流程推进。
|
|
44
|
+
- **扇形思考强度仪表盘、固定高度防抽搐、弹窗实底防透与团队思考强度「自动 / 手动」简洁切换(`chief-state.js` / `chief-commands.js` / `subagent-runtime.js` / `api-routes.js` / `components/common.js` / `components/modals.js` / `tabs/chief.js` / `tabs/tasks.js` / `styles.js`)**:
|
|
45
|
+
- 将模型选择器与战术驾驶舱中的思考强度替换为 180° 扇形档位表(`EffortFanGauge`):支持拖拽/点击切换固定挡位或点按中心一键切回 `🪄 自动调节`,移除自动扫掠循环动画与底部自适应说明栏,保持原生气泡解释且彻底消除鼠标跨档悬停导致的弹窗高度抽搐;
|
|
46
|
+
- 取消针对单个主/子代理的思考档位下拉调控与 `锁挡` 按钮,改为在扇形表盘下方提供「团队思考强度(主/子代理):自动 / 手动」一键切换(`teamEffortMode`,默认 `auto` 由参谋长按当前权限范围自动调节,切换 `manual` 则由人类掌控且参谋主动调挡门禁与执行均严格拦截);
|
|
47
|
+
- 精简移除战术驾驶舱顶部 `拉起 / 复位` 把手,保留主代理与子代理卡片及调度标签前置 `agentLead` / `agentSub` 矢量图标,并为 `.dnb-tactical-pop` / `.dnb-perm-pop` / `.dnb-more-menu` / `.dnb-menu` 补充深/浅色实底背景修复深色模式下弹层半透明透底叠字问题。
|
|
48
|
+
- **支持鼠标单击毛玻璃遮罩即穿透恢复操作且保留外框流光提示 (`components/Panel.js` / `styles.js`)**:当后台「🌙 梦境漫游」或自动审查点亮毛玻璃舞台时,新增 `点击任意处收起遮罩继续操作 · 外框流光表示仍在运作` 提示;用户鼠标单击毛玻璃任意区域即可立即卸除模糊与点击拦截层恢复自由操作面板,同时四边 `.dnb-chief-fx-border` 彩色流光外框依然保持运行,直至本轮后台运作结束。
|
|
49
|
+
|
|
8
50
|
## [1.4.0] - 2026-11-19
|
|
9
51
|
|
|
10
52
|
### Added
|
package/README.en.md
CHANGED
|
@@ -1,11 +1,15 @@
|
|
|
1
1
|
<div align="center">
|
|
2
2
|
|
|
3
|
-
**Right-sidebar supervisor,
|
|
4
|
-
<sub>Saves paths and findings before `/compact` wipes context, tracks subagent DAGs in one view, and
|
|
3
|
+
**Right-sidebar supervisor, Cognitive Desktop, pre-execute Guard Walls & multi-agent gate for DeepSeek Harness (DSH)**
|
|
4
|
+
<sub>Saves paths and findings before `/compact` wipes context, tracks subagent DAGs in one view, and enforces Guard Walls + per-action approvals before file edits or commands run.</sub>
|
|
5
5
|
|
|
6
6
|
<img src="assets/chief-10s-compare.gif" alt="Side-by-side comparison: plain long session loses context on compaction vs. dsh-chief saves findings to cold vaults, guards actions, and carries structured handoffs across windows" width="800" />
|
|
7
7
|
|
|
8
|
-
<sub>Read-only by default · Per-action workspace approval · `dsh plugin --profile
|
|
8
|
+
<sub>Read-only by default · Per-action workspace approval · `dsh plugin --profile desktop add dsh-chief`</sub>
|
|
9
|
+
|
|
10
|
+
<br />
|
|
11
|
+
|
|
12
|
+
<img src="assets/v150-highlights.png" alt="dsh-chief v1.5.0 feature highlights: 180-degree fan effort gauge, click-through frosted Dream Roam keeping flowing neon border, Cognitive Desktop & stack-of-papers vaults, pre-execute Guard Walls" width="100%" />
|
|
9
13
|
|
|
10
14
|
<br />
|
|
11
15
|
|
|
@@ -17,7 +21,9 @@
|
|
|
17
21
|
### DeepSeek Harness Right-Sidebar Supervisor, Zero-Token Archive Vault & Safety Gate
|
|
18
22
|
|
|
19
23
|
[](https://dsh.deepseek.com)
|
|
24
|
+
[](https://github.com/gluztm/dsh-chief/releases)
|
|
20
25
|
[](https://www.npmjs.com/package/dsh-chief)
|
|
26
|
+
[](https://github.com/gluztm/dsh-chief/actions/workflows/ci.yml)
|
|
21
27
|
[](https://awesome-dsh-plugin.com/p/gluztm/dsh-chief/)
|
|
22
28
|
[](https://www.bilibili.com/video/BV1DqsV6QEGi)
|
|
23
29
|
[](https://github.com/deepseek-ai/dsh-superagent-prototype)
|
|
@@ -25,15 +31,26 @@
|
|
|
25
31
|
[](#)
|
|
26
32
|
[](LICENSE)
|
|
27
33
|
|
|
28
|
-
> **Current release `v1.
|
|
34
|
+
> **Current release `v1.5.0`**: Adds the **Cognitive Desktop (`2,400 Tokens / 6 items` sole prompt-injected vault)** + top-down stack-of-papers drawers with 3-tier hook isolation, **orthogonal runtime control (`⏸ Off | 🛡 Patrol | 🤖 Managed`)** with background `🌙 Dream Roam` (**click anywhere on the frosted overlay to dismiss the blur and regain panel control while the flowing neon border stays lit to signal background work**), 5-condition auto-sleep + `L3` One-Stop Report, **`180°` fan-shaped thinking effort gauge** + Team `Auto / Manual` toggle, plain-language Consult mode, mandatory Core Objective lock for Strong Intervention, blocker-gated wall removal requests, full DSH light-theme support, and subagent entry hiding on DSH `0.2.2-rc.5`.
|
|
29
35
|
> Shares the Router + Workers split from DeepSeek's **[SuperAgent (`dsh-superagent-prototype`)](https://github.com/deepseek-ai/dsh-superagent-prototype)** while keeping Chief state and prompts isolated from the main session.
|
|
30
36
|
|
|
31
|
-
[简体中文 (`README.md`)](README.md) · **English** · [Changelog (`CHANGELOG.md`)](CHANGELOG.md)
|
|
37
|
+
[简体中文 (`README.md`)](README.md) · **English** · [Changelog (`CHANGELOG.md`)](CHANGELOG.md) · [Architecture (`ARCHITECTURE.md`)](ARCHITECTURE.md)
|
|
38
|
+
|
|
39
|
+
[v1.5.0 Highlights](#what-changed-in-v150) · [Problems Fixed](#daily-friction-this-plugin-fixes) · [Inside the Sidebar](#inside-the-sidebar) · [Architecture](#modular-architecture-v150) · [Installation](#installation--host-compatibility) · [Security](#local-storage--security-guarantees) · [Contributors](#contributors)
|
|
32
40
|
|
|
33
41
|
</div>
|
|
34
42
|
|
|
35
43
|
---
|
|
36
44
|
|
|
45
|
+
## What Changed in `v1.5.0`
|
|
46
|
+
|
|
47
|
+
- **🗂️ Cognitive Desktop & Stack-of-Papers Hook Isolation**: Eliminates full-text archive leaks into the main prompt — only documents explicitly pinned to the **📌 Cognitive Desktop (`2,400 Tokens / 6 items` cap)** are injected per turn. Session / Global / Precipitated drawers use collapsible paper-dossier cards with lightweight `🪝` keyword hooks so agents pull dossiers on demand, while un-flipped Precipitated records remain `0-byte` double-blind to AI.
|
|
48
|
+
- **🤖 Orthogonal Runtime Modes & Click-Through `🌙 Dream Roam`**: Header switches cleanly between `⏸ Off` (0 background calls), `🛡 Patrol` (0 idle cost; wakes only on guard blocks, errors, or stream interrupts), and `🤖 Managed` (active oversight + idle Dream Roam organization). Clicking anywhere on the frosted Dream Roam overlay immediately dismisses the blur/blocker so you can keep using the sidebar while the animated outer neon border continues to indicate background work. Auto-sleeps back to `Off` after all 5 idle checks clear.
|
|
49
|
+
- **🧠 `180°` Fan Thinking Effort Gauge & Team Control**: Drag along the 180° rainbow arc (`Off / Low / Medium / High / Max`) or tap the center hub for `🪄 Auto` (lowers effort on summaries/roam, raises effort for diagnostics & wall crafting), paired with a clean `Team Thinking Effort: Auto / Manual` toggle and per-reply effort badges.
|
|
50
|
+
- **🛡️ Sharpened Guardrails & Theme Fixes**: Strong Intervention now enforces locking a Core Objective before activation; `💡 Consult` answers honestly without forcing synthetic action buttons; blocker-only `wall_removal` requests require human approval across every L0–L3 tier; solid popover contrast fixes dark-mode transparency overlap; and `SidebarChiefTab` hides cleanly inside subagent threads alongside full DSH light-theme styling.
|
|
51
|
+
|
|
52
|
+
---
|
|
53
|
+
|
|
37
54
|
## Daily Friction This Plugin Fixes
|
|
38
55
|
|
|
39
56
|
If you run multi-step coding tasks inside DeepSeek Harness, these six problems tend to come up quickly.
|
|
@@ -45,64 +62,61 @@ If you run multi-step coding tasks inside DeepSeek Harness, these six problems t
|
|
|
45
62
|
|
|
46
63
|
### 2. Agents stop halfway ("I edited the file, please test it")
|
|
47
64
|
- **What goes wrong**: Left unattended or overnight, an agent updates one file and marks the task done without running checks or builds.
|
|
48
|
-
- **How `dsh-chief` handles it**:
|
|
65
|
+
- **How `dsh-chief` handles it**: Requires locking a Core Objective before enabling Strong Intervention (`0 = ∞` turns with 3-strike identical-error circuit breakers), wakes cleanly across `🛡 Patrol` and `🤖 Managed` on stream interruptions, executes safe L3 steps into a merged `🏁 One-Stop Report`, and holds final completion until your `allowed` checks pass and you approve.
|
|
49
66
|
- Hit the red `⏹ 急停` button at any time to stop background drives and take back manual control.
|
|
50
67
|
|
|
51
68
|
### 3. Always-on memory plugins bloat prompts and burn tokens
|
|
52
69
|
- **What goes wrong**: Injecting kilobytes of old notes on every turn wastes context window and distracts the model from the current file.
|
|
53
|
-
- **How `dsh-chief` handles it**:
|
|
54
|
-
- Click `📌 摆上桌面` or `📤 取出` only when a specific note is needed, and put it back as soon as you finish that step.
|
|
70
|
+
- **How `dsh-chief` handles it**: Archives live in top-down stack-of-papers folders across three vaults (Session, System Global, Precipitated). **Only items pinned to the Cognitive Desktop (`2,400 Tokens / 6 items`) enter the prompt automatically; unpinned dossiers add 0 full-text tokens**, exposing only lightweight `🪝` keyword hooks for on-demand `archive.pull` retrieval while keeping Precipitated secrets double-blind.
|
|
55
71
|
|
|
56
72
|
### 4. Full autonomy feels unsafe, but chat-only mode leaves Chief idle (or bossy)
|
|
57
73
|
- **What goes wrong**: Clicking "allow everything" risks `git push --force`, destructive deletes, or reading `.env`, while high-tier supervisors often try to edit files themselves instead of letting your main agent work.
|
|
58
|
-
- **How `dsh-chief` handles it**:
|
|
74
|
+
- **How `dsh-chief` handles it**: Separates permission tiers (`L0 advisor` read-only default / `L1 assistant` / `L2 strategist` per-action approval / `L3 chief`) from header runtime modes (`⏸ Off / 🛡 Patrol / 🤖 Managed`), backed by a pre-execution firewall.
|
|
59
75
|
- **Every new session starts read-only**: Observes code, answers questions, and updates the board without touching your workspace.
|
|
60
|
-
- **Pre-execution guard walls (`tools/pre-execute`)**: Before the main agent, any subagent, or Chief writes a file or runs a command, calls pass through `~/.dsh/chief/guard-walls.json`. Walls
|
|
61
|
-
- **Lazy intervention & background check-ins**: Unless you explicitly ask Chief to edit files directly, it sends structured `[EXECUTION_REQUEST]` items to the main conversation instead of taking over or speaking as you. In a blank new chat, it
|
|
62
|
-
- **Auto-resume after approval & restrict-only policies**: `TaskLedger` picks up unfinished steps once you click Allow
|
|
76
|
+
- **Pre-execution guard walls (`tools/pre-execute`)**: Before the main agent, any subagent, or Chief writes a file or runs a command, calls pass through `~/.dsh/chief/guard-walls.json`. Walls switch between `🌐 Global`, `📁 Workspace`, and `⏸ Paused`, merge same-type rules, or get drafted from plain language for your approval. Models in any L0–L3 tier can never mute or remove locked walls — genuine dead-end blockers can only submit a human-approved removal request.
|
|
77
|
+
- **Lazy intervention & background check-ins**: Unless you explicitly ask Chief to edit files directly, it sends structured `[EXECUTION_REQUEST]` items to the main conversation instead of taking over or speaking as you. In a blank new chat, it guides `@lead` to inspect the repo and plan first, and asks clarifying `[INQUIRY]` questions quietly without pausing `@lead`.
|
|
78
|
+
- **Auto-resume after approval & restrict-only policies**: `TaskLedger` picks up unfinished steps once you click Allow; hard rules (compound shells, `node -e`/`python -c`, quote-backslash `.env` tricks, symlinks, NTFS ADS) hold in every tier with logs written to `~/.dsh/chief/chief-audit.jsonl`, alongside optional restrict-only JSON `policyPlugins`.
|
|
63
79
|
|
|
64
80
|
### 5. Dozens of subagents run out of sight and ignore verbal rules
|
|
65
81
|
- **What goes wrong**: Workers live across scattered tabs, and rules mentioned in chat scroll out of context after ten turns.
|
|
66
|
-
- **How `dsh-chief` handles it**: Shows the main coordinator and subagents in one tree
|
|
82
|
+
- **How `dsh-chief` handles it**: Shows the main coordinator and subagents in one pyramid tree with status controls, `@agent` direct steering pills along the bottom bar, plain-language step explanations, and one-click `分屏` live streams. Pin rules with `📌` in Boundaries so every new task inherits them.
|
|
67
83
|
|
|
68
84
|
### 6. Long threads slow down, but opening a new window resets everything
|
|
69
85
|
- **What goes wrong**: After 100 turns you want a clean window without re-explaining workspace paths and task progress, and you do not want a background plugin quietly calling paid APIs every turn.
|
|
70
|
-
- **How `dsh-chief` handles it**: Defaults to `⏸
|
|
86
|
+
- **How `dsh-chief` handles it**: Defaults to `⏸ Off` / `🛡 Patrol` (0 idle LLM calls), mirrors your active session's provider/model, and exposes a `180°` fan effort gauge with `🪄 Auto` adaptive depth. Click `解释` on any plan step for plain-language explanations, freeze handoff notes with `📌 定住防改`, and deploy a linked window carrying only the archive scope you pick.
|
|
71
87
|
|
|
72
88
|
---
|
|
73
89
|
|
|
74
90
|
## Inside the Sidebar
|
|
75
91
|
|
|
76
|
-
### 01 · Chief Cockpit (`参谋` Tab)
|
|
77
|
-
- Send tasks in plain language from the bottom input box
|
|
78
|
-
-
|
|
79
|
-
- Filter the feed by `全部 / 💬 问答 / 📋 回执`. User prompts and Chief receipts sit inside high-contrast grouped frames that collapse individually or all at once, multi-step suggestions merge into one top `▶ 执行建议` button, and background Chief work switches to the active tab with a frosted overlay and sweeping light-beam border.
|
|
92
|
+
### 01 · Chief Cockpit & Click-Through FX Stage (`参谋` Tab)
|
|
93
|
+
- Send tasks in plain language from the bottom input box, toggle `⚡ 强介入`, `✦ 智选压缩`, `🛡 红线哨兵`, `L0–L3` tiers, and `⏹ 急停`, or adjust `⏸ Off / 🛡 Patrol / 🤖 Managed` and the `180°` fan thinking effort gauge.
|
|
94
|
+
- When Dream Roam or auto-review raises the frosted working stage overlay, **a single click dismisses the blur so you keep full control while the glowing outer border continues to show Chief is active**.
|
|
80
95
|
|
|
81
96
|

|
|
82
97
|
|
|
83
98
|
---
|
|
84
99
|
|
|
85
|
-
### 02 · Coordinator & Subagent
|
|
86
|
-
- Displays the root coordinator and active subagent lanes
|
|
100
|
+
### 02 · Coordinator & Subagent Pyramid Tree (`任务` Tab)
|
|
101
|
+
- Displays the root coordinator and active subagent lanes with dedicated coordinator/worker icons, inline status toggles, bottom `@steer` pills, and `分屏` live stream panes.
|
|
87
102
|
|
|
88
103
|
<div align="center">
|
|
89
|
-
<img src="assets/screenshots/02-full-spread-pyramid-dag.png" alt="02 - Subagent DAG View" width="
|
|
104
|
+
<img src="assets/screenshots/02-full-spread-pyramid-dag.png" alt="02 - Subagent DAG View" width="80%" />
|
|
90
105
|
</div>
|
|
91
106
|
|
|
92
107
|
---
|
|
93
108
|
|
|
94
|
-
### 03 ·
|
|
95
|
-
-
|
|
96
|
-
- Dossiers stay out of prompts until you click `📌 摆上桌面`.
|
|
109
|
+
### 03 · Cognitive Desktop & Stack-of-Papers Vaults (`档案` Tab)
|
|
110
|
+
- Features the pinned **📌 Cognitive Desktop** (`2,400 Tokens / 6 items` bar — sole per-turn prompt injection source) above Session, System Global, and Precipitated drawers with top-down stack-of-papers cards and `🪝` keyword hooks.
|
|
97
111
|
|
|
98
112
|
<div align="center">
|
|
99
|
-
<img src="assets/screenshots/03-cold-archive-cabinet.png" alt="03 - Three-Tier Cold Vaults and
|
|
113
|
+
<img src="assets/screenshots/03-cold-archive-cabinet.png" alt="03 - Three-Tier Cold Vaults and Cognitive Desktop" width="90%" />
|
|
100
114
|
</div>
|
|
101
115
|
|
|
102
116
|
---
|
|
103
117
|
|
|
104
|
-
### 04 ·
|
|
105
|
-
- Click `🎯 设为目标` to
|
|
118
|
+
### 04 · Objective Lock & Step Q&A (`规划` Top)
|
|
119
|
+
- Click `🎯 设为目标` to lock your Core Objective before Strong Intervention, or click `解释` on any step to get a plain-language breakdown without interrupting `@lead`.
|
|
106
120
|
|
|
107
121
|
<div align="center">
|
|
108
122
|
<img src="assets/screenshots/04-dual-tier-goal-and-step-qa.png" alt="04 - Goal Supervision and Step Q&A" width="68%" />
|
|
@@ -111,8 +125,8 @@ If you run multi-step coding tasks inside DeepSeek Harness, these six problems t
|
|
|
111
125
|
---
|
|
112
126
|
|
|
113
127
|
### 05 · Cross-Window Peer Linking & Mailbox (`协同` Tab)
|
|
114
|
-
- Connect multiple open DSH windows to share live progress, with automatic cache and mailbox purge on detach.
|
|
115
|
-
- Screenshots pasted into the cockpit are saved under `<cwd>/.dsh-chief/uploads/` so models
|
|
128
|
+
- Connect multiple open DSH windows to share live progress and coordinate ports/contracts, with automatic cache and mailbox purge on detach.
|
|
129
|
+
- Screenshots pasted into the cockpit are saved under `<cwd>/.dsh-chief/uploads/` so models inspect real files instead of guessing.
|
|
116
130
|
|
|
117
131
|
<div align="center">
|
|
118
132
|
<img src="assets/screenshots/05-cross-session-collab-bus.png" alt="05 - Cross-Window Linking and Mailbox" width="68%" />
|
|
@@ -121,8 +135,8 @@ If you run multi-step coding tasks inside DeepSeek Harness, these six problems t
|
|
|
121
135
|
---
|
|
122
136
|
|
|
123
137
|
### 06 · Guard Walls, Pinned Redlines & Must-Do Checks (`边界` Tab)
|
|
124
|
-
- The top **🧱 Physical Guard Walls** section blocks dangerous commands or paths before execution: add rules manually
|
|
125
|
-
- Below that, `allowed
|
|
138
|
+
- The top **🧱 Physical Guard Walls** section blocks dangerous commands or paths before execution: add rules manually or let Chief draft them from plain text, switch `🌐 Global`, `📁 Workspace`, or `⏸ Paused`, and merge same-type walls.
|
|
139
|
+
- Below that, `allowed`, `forbidden`, and `notes` separate global `📌` rules from session-only ones. No L0–L3 model can delete locked walls without an approved blocker removal ticket.
|
|
126
140
|
|
|
127
141
|
<div align="center">
|
|
128
142
|
<img src="assets/screenshots/06-pinned-bounds-and-handoff.png" alt="06 - Pinned Global Redlines and Proposals" width="68%" />
|
|
@@ -131,7 +145,7 @@ If you run multi-step coding tasks inside DeepSeek Harness, these six problems t
|
|
|
131
145
|
---
|
|
132
146
|
|
|
133
147
|
### 07 · Session Handoff & Deployment (`规划` Bottom)
|
|
134
|
-
- Edit the Markdown handoff packet directly and
|
|
148
|
+
- Edit the Markdown handoff packet directly and freeze it via `📌 定住防改`.
|
|
135
149
|
- Choose which archive scope follows into a new session (`None`, `Spotlight Only`, `Entire Drawer`, or `Full Session Cabinet`), while Strong Intervention safely resets to off.
|
|
136
150
|
|
|
137
151
|
<div align="center">
|
|
@@ -146,32 +160,32 @@ If you run multi-step coding tasks inside DeepSeek Harness, these six problems t
|
|
|
146
160
|
> Keep these two constraints in mind when using Wishing Box tools, `⚡ Strong Intervention`, or `✦ 智选压缩`:
|
|
147
161
|
>
|
|
148
162
|
> 1. **Guardrails set the safety floor; your LLM sets the ceiling**
|
|
149
|
-
> The plugin guarantees 0-token vault isolation, permission gating, verified pre-compact saves, loop breakers, and `⏹ 急停`. Planning accuracy and code judgment still depend on the model assigned to Chief. Use a strong reasoning model for real engineering work.
|
|
163
|
+
> The plugin guarantees 0-token vault isolation, permission gating, verified pre-compact saves, loop breakers, guard walls, and `⏹ 急停`. Planning accuracy and code judgment still depend on the model assigned to Chief. Use a strong reasoning model for real engineering work.
|
|
150
164
|
> 2. **Avoid stacking heavy external prompt injections**
|
|
151
|
-
> Skip stacking hundreds of conflicting global rule lines or always-on memory plugins alongside `dsh-chief`.
|
|
165
|
+
> Skip stacking hundreds of conflicting global rule lines or always-on memory plugins alongside `dsh-chief`. Keep only essentials on the **Cognitive Desktop** and store reference notes in the hooked drawers instead.
|
|
152
166
|
|
|
153
167
|
---
|
|
154
168
|
|
|
155
|
-
## Modular Architecture (`v1.
|
|
169
|
+
## Modular Architecture (`v1.5.0`)
|
|
156
170
|
|
|
157
|
-
`v1.
|
|
171
|
+
`v1.5.0` spans **38 backend modules (including 9 isolated modules under `pkg/safety/`) and 16 frontend modules** with zero `export let` bindings and zero circular imports. See [ARCHITECTURE.md](ARCHITECTURE.md) for boundaries and module contracts.
|
|
158
172
|
|
|
159
173
|
```mermaid
|
|
160
174
|
flowchart TB
|
|
161
175
|
subgraph UI["DSH Right Sidebar (client.js <- 16 source modules)"]
|
|
162
|
-
T1["Chief Tab (
|
|
163
|
-
T2["Tasks Tab (Coordinator
|
|
176
|
+
T1["Chief Tab (180° Effort Gauge · Click-Through FX · Grouped Q&A)"]
|
|
177
|
+
T2["Tasks Tab (Coordinator Pyramid · Split Stream · @steer)"]
|
|
164
178
|
T3["Peers Tab (Cross-Window Link + Mailbox)"]
|
|
165
|
-
T4["Plan Tab (Objective Lock
|
|
166
|
-
T5["Archives Tab (
|
|
167
|
-
T6["Boundaries Tab (
|
|
179
|
+
T4["Plan Tab (Core Objective Lock · Step Explain · Pinned Handoff)"]
|
|
180
|
+
T5["Archives Tab (2400T Cognitive Desktop · Hooked Stack-of-Papers)"]
|
|
181
|
+
T6["Boundaries Tab (Guard Walls · Blocker Removal · 📌 Redlines)"]
|
|
168
182
|
end
|
|
169
183
|
|
|
170
184
|
subgraph Core["Local Backend & Safety Kernel (Zero External Deps · 38 modules)"]
|
|
171
185
|
Hot["Hot Boards: ~/.dsh/chief/board.json & boards/<sid>.json\nCross-process/worker_threads O_EXCL spin locks"]
|
|
172
186
|
Walls["Guard Walls: ~/.dsh/chief/guard-walls.json\nGlobal / Workspace / Paused"]
|
|
173
|
-
Cold["3-Tier Vaults:
|
|
174
|
-
Bus["
|
|
187
|
+
Cold["3-Tier Vaults: 2400T Cognitive Desktop + Hooked Drawers + 0-byte Blind Vault"]
|
|
188
|
+
Bus["Runtime & Safety: safety/*.js + Off/Patrol/Managed(5-Cond Sleep) + Audit"]
|
|
175
189
|
end
|
|
176
190
|
|
|
177
191
|
subgraph Agents["Host Execution Agents (tools/pre-execute gate)"]
|
|
@@ -185,7 +199,7 @@ flowchart TB
|
|
|
185
199
|
Hot -->|"Compact Hot State"| Subs
|
|
186
200
|
Walls -->|"Pre-Execution Block or Ask Prompt"| Lead
|
|
187
201
|
Walls -->|"Pre-Execution Block or Ask Prompt"| Subs
|
|
188
|
-
Cold -.->|"
|
|
202
|
+
Cold -.->|"Cognitive Desktop Full Text · Hooked Pull on Demand"| Lead
|
|
189
203
|
Bus <-->|"chief_board / chief_memory\nchief_boundaries / chief_handoff"| Agents
|
|
190
204
|
```
|
|
191
205
|
|
|
@@ -196,21 +210,21 @@ flowchart TB
|
|
|
196
210
|
Search **`dsh-chief`** inside [dshmarket](https://github.com/dsh-market/dsh-market) / DSH Plugin Market, or install via CLI (replace `desktop` with `web` for the web profile):
|
|
197
211
|
|
|
198
212
|
```bash
|
|
199
|
-
# Option 1: Install latest release (v1.
|
|
213
|
+
# Option 1: Install latest release (v1.5.0) from npm
|
|
200
214
|
dsh plugin --profile desktop add dsh-chief
|
|
201
215
|
|
|
202
216
|
# Option 2: Install directly from GitHub main branch
|
|
203
217
|
dsh plugin --profile desktop add github:gluztm/dsh-chief
|
|
204
218
|
|
|
205
|
-
# Option 3: Install prebuilt GitHub Release tarball (v1.
|
|
206
|
-
dsh plugin --profile desktop add https://github.com/gluztm/dsh-chief/releases/download/v1.
|
|
219
|
+
# Option 3: Install prebuilt GitHub Release tarball (v1.5.0)
|
|
220
|
+
dsh plugin --profile desktop add https://github.com/gluztm/dsh-chief/releases/download/v1.5.0/dsh-chief-1.5.0.tgz
|
|
207
221
|
```
|
|
208
222
|
|
|
209
223
|
### Host Version Compatibility
|
|
210
224
|
|
|
211
225
|
| DeepSeek Harness (DSH) Version | Panel / Board / Vaults / Pre-Execute Walls | Named Subagent Tree (`task.dispatch` / `team.*`) | Notes |
|
|
212
226
|
| :--- | :---: | :---: | :--- |
|
|
213
|
-
| **`>= 0.2.2-rc.5` (current desktop target)** | Full support | Full support | Matches `dsh-tool-fs`
|
|
227
|
+
| **`>= 0.2.2-rc.5` (current desktop target)** | Full support | Full support | Matches `dsh-tool-fs`, `str_replace_editor`, dynamic `reasoningEffort` switching, and native `{ kind: 'ask' }` dialogs |
|
|
214
228
|
| `0.2.2-rc.2 ~ 0.2.2-rc.4` | Full support | Full support | Provides native `subagents.resolveName`, `rootOf`, `namesOf`, and `tools/pre-execute` |
|
|
215
229
|
| `0.2.0 ~ 0.2.2-rc.1` | Panel works | Upgrade notice | Blocks unsafe `team.*` dispatch and prompts host upgrade |
|
|
216
230
|
|
|
@@ -220,7 +234,7 @@ dsh plugin --profile desktop add https://github.com/gluztm/dsh-chief/releases/do
|
|
|
220
234
|
|
|
221
235
|
| Tool Name | Purpose | Common Actions |
|
|
222
236
|
| :--- | :--- | :--- |
|
|
223
|
-
| **`chief_board`** | Task DAG updates, cold archive queries, peer & chief requests, inquiry replies, wall listing,
|
|
237
|
+
| **`chief_board`** | Task DAG updates, cold archive queries/pulls, peer & chief requests, inquiry replies, wall listing, effort tuning, compaction | `graph.plan`, `task.add`, `task.update`, `plan.set`, `cursor.set`, `mail.send`, `chief.request`, `inquiry.reply`, `wall.list`, `archive.save`, `archive.search`, `archive.pull`, `model.effort`, `bounds.propose`, `permission.request`, `context.compact` |
|
|
224
238
|
| **`chief_memory`** | Session-scoped hot items (`decisions`, `progress`, `verify`, `risks`) | `list`, `get`, `add`, `update`, `remove` |
|
|
225
239
|
| **`chief_boundaries`** | Global `📌` redlines, must-do checks (`allowed`), and notes | `get`, `set`, `add`, `pin`, `remove` |
|
|
226
240
|
| **`chief_handoff`** | Preview and render self-contained Markdown handoff packets | `preview` / session handoff |
|
|
@@ -241,7 +255,9 @@ dsh plugin --profile desktop add https://github.com/gluztm/dsh-chief/releases/do
|
|
|
241
255
|
| Contributor | Role |
|
|
242
256
|
| --- | --- |
|
|
243
257
|
| [gluztm](https://github.com/gluztm) | Creator, product design, and maintainer |
|
|
244
|
-
| [
|
|
258
|
+
| [@iridiumgit](https://github.com/iridiumgit) | Root-caused and contributed the DSH light-theme popover fix ([#3](https://github.com/gluztm/dsh-chief/issues/3)), subagent entry visibility diagnosis ([#4](https://github.com/gluztm/dsh-chief/issues/4)), and UX feedback ([#5](https://github.com/gluztm/dsh-chief/issues/5)) |
|
|
259
|
+
| [@edz017](https://github.com/edz017) | Day-one DSH `0.2.2-rc.1~rc.3` host compatibility audit, `engines.dsh` constraint diagnosis, and `preferNewPane` fix ([#1](https://github.com/gluztm/dsh-chief/issues/1)) |
|
|
260
|
+
| [Gemini](https://github.com/gemini-code-assist) | Feature development, `pkg/safety/` kernel, Cognitive Desktop, orthogonal runtime modes, fan gauge & CI gates |
|
|
245
261
|
| [Codex](https://github.com/codex) | Modular architecture design, concurrency stress tests, and red-team boundary audits |
|
|
246
262
|
| [deepseek-ai](https://github.com/deepseek-ai) | DeepSeek Harness host runtime and SuperAgent workflow alignment |
|
|
247
263
|
|