dsh-mindmark 0.0.0-stage → 0.4.0
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 +25 -0
- package/LICENSE +21 -0
- package/README.en.md +103 -0
- package/README.md +125 -2
- package/client.js +3239 -0
- package/cordis.patch.yml +14 -0
- package/docs/demo.gif +0 -0
- package/icon.svg +9 -0
- package/index.js +629 -0
- package/locale/en.json +6 -0
- package/locale/zh.json +6 -0
- package/package.json +55 -4
- package/src/inject.js +367 -0
- package/src/lessons.js +187 -0
- package/src/store.js +1028 -0
package/CHANGELOG.md
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
1
|
+
# 更新记录
|
|
2
|
+
|
|
3
|
+
## v0.4.0
|
|
4
|
+
|
|
5
|
+
新增两种标记,标记从五种变成七种;选区浮条分成两排。
|
|
6
|
+
|
|
7
|
+
- **依据**:选中一段,挂上参考材料,助手只按材料改这一段,并说明改动依据了材料的哪一句。
|
|
8
|
+
- 材料可以粘贴文本、贴链接,或选一个纯文本文件(txt、md、csv、json、html 等)。
|
|
9
|
+
- 2000 字以内的材料随标注整段发给助手;更长的只发开头摘录,助手需要时自己分段读全文。
|
|
10
|
+
- 材料单独存放,删标注时一起删。
|
|
11
|
+
- **移动**:像剪切粘贴一样挪动一段文字。
|
|
12
|
+
- 选中一段按 Alt+C 拿起,在同一条回答里点一下,出现闪烁竖线,这段会插到竖线之后。
|
|
13
|
+
- 拖选一段文字作为目标,则用被移动的那段替换它。
|
|
14
|
+
- 按 Alt+V 放下,Esc 取消;提示条里可以写一句为什么要移。
|
|
15
|
+
- **浮条分两排**:第一排点一下就生效(引用、高亮、下划线、删除、清除标记),第二排要补充内容(批注、改写、依据、移动)。
|
|
16
|
+
- 修复:精简模式发给助手的上下文只取回答正文,不再混进思考过程等页面文字。
|
|
17
|
+
|
|
18
|
+
## v0.3.0
|
|
19
|
+
|
|
20
|
+
- 沉淀为习惯:让助手把反复出现的标注归纳成一句话偏好,写进全局 `AGENTS.md`,新对话自动遵守。
|
|
21
|
+
- 演示 GIF、英文 README,发布到 npm。
|
|
22
|
+
|
|
23
|
+
## v0.2.0
|
|
24
|
+
|
|
25
|
+
- 第一版:高亮、下划线、删除、批注、改写五种标记;精简和完整两种注入模式;助手处理后回传确认。
|
package/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 BaiKe-Zhang
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
package/README.en.md
ADDED
|
@@ -0,0 +1,103 @@
|
|
|
1
|
+
# dsh-mindmark
|
|
2
|
+
|
|
3
|
+
**Show the model your mind: mark up its answer the way you would in Word.**
|
|
4
|
+
|
|
5
|
+
English | [中文](README.md)
|
|
6
|
+
|
|
7
|
+
A [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) plugin. Select any passage of an answer and **highlight, underline, strike, comment on, rewrite, back with a source or move** it. On the next turn the model receives your markup as context: not just its own answer, but the answer as you want it understood. It confirms back when it has handled every mark. After a few annotations, one click asks the model to distill them into **long-term preferences** in your global `AGENTS.md`, which every new conversation then follows.
|
|
8
|
+
|
|
9
|
+

|
|
10
|
+
|
|
11
|
+
In the demo the model drafts a short social-media post, then the answer gets, in order: a highlight on the analogy to keep, an underline on a mechanism that needs explaining, a strike on a filler clause, a comment asking for one more sentence, a rewrite turning a vague repo hint into the full address, the README attached as a source to list all seven marks, and a move that puts the new-features sentence first. The next turn applies every mark and confirms it; "Keep as habit" then distills seven lasting preferences.
|
|
12
|
+
|
|
13
|
+
## What it feels like
|
|
14
|
+
|
|
15
|
+
1. Select text in any answer. The Chat view's own selection bar appears in two rows:
|
|
16
|
+
- one click and done: **Quote · Highlight · Underline · Strike · Remove marks**;
|
|
17
|
+
- needs your input: **Comment · Rewrite · Source · Move**.
|
|
18
|
+
2. Comment and rewrite open a small input (confirm with the button or Ctrl+Enter).
|
|
19
|
+
3. **Source** attaches reference material to the passage so the model revises only that passage to match it: paste text, give a link, or pick a plain-text file (txt / md / csv / json / html…; for PDF or Word, copy the text). An optional instruction can go with it.
|
|
20
|
+
4. **Move** works like cut and paste: select a passage and press **Alt+C** (or click Move), then click in the same answer: a blinking orange caret appears and the passage goes right after it. Drag over some text instead and the passage replaces that text. Press **Alt+V** (or "Drop here" in the banner) to drop it; Esc cancels. The banner takes an optional reason. The source shows an orange dashed underline.
|
|
21
|
+
5. Send your next message. The model adjusts its answer to your marks, then calls the `annotations` tool to confirm they are handled.
|
|
22
|
+
6. The composer gains an **Annotation feed** switch with a pending-count badge:
|
|
23
|
+
- **Brief**: the next turn sends only the annotated passages with a little context;
|
|
24
|
+
- **Full**: the next turn sends the whole answer as a revision view with inline markers.
|
|
25
|
+
7. The marker icon in the session header opens the panel: every annotation with its state (pending / handled by the assistant / auto-settled), per-item delete, mark handled, send again, and **Keep as habit**, which asks the model to distill the marks into lasting preferences. Distilled preferences are listed there and can be removed.
|
|
26
|
+
|
|
27
|
+
## What each mark tells the model
|
|
28
|
+
|
|
29
|
+
| Mark | Meaning |
|
|
30
|
+
|---|---|
|
|
31
|
+
| Highlight | This is what I want: keep it and strengthen this direction |
|
|
32
|
+
| Underline | Not enough here: expand, clarify or justify |
|
|
33
|
+
| Strike | Drop this next time |
|
|
34
|
+
| Comment | Act on the note and respond to each one |
|
|
35
|
+
| Rewrite | Treat my wording as the correct version |
|
|
36
|
+
| Source | Revise only this passage, following the attached material; the material wins on conflict; cite the line used |
|
|
37
|
+
| Move | Move this passage right after the caret (or onto the selected text, replacing it), remove it from where it was, leave the rest alone |
|
|
38
|
+
|
|
39
|
+
## How the model sees and answers your marks
|
|
40
|
+
|
|
41
|
+
- **Pre-step injection.** At the start of every turn the Host renders the conversation's pending annotations into one user-role context message entered right after your prompt, through the official `agent/pre-step` hook. Like every other context plugin it lands in the session log, so you can inspect it in the Trajectory view.
|
|
42
|
+
- **Protocol section.** While annotations are pending, a short system-prompt section explains the marks and the confirmation rule.
|
|
43
|
+
- **Reference material.** Materials are stored separately under `$DSH_HOME/annotations/materials/`, never inside `state.json`, and are deleted with their mark. Up to 2,000 characters travel whole with the mark; longer ones send an excerpt and the length, and the model reads the rest with the `annotations` tool's `material` action (8,000 characters per call, continue with `offset`). Links are opened by the model's own web tools; if it cannot, it asks you to paste the text. One material holds up to 200,000 characters.
|
|
44
|
+
- **Confirmation signal.** The model calls the `annotations` tool (`action: resolve`, optional `ids`) when done. Confirmed marks stop being injected and show as "handled by the assistant"; unconfirmed ones come back next turn, up to `maxInjections` (default 3) before they settle automatically.
|
|
45
|
+
- **Long-term preferences.** "Keep as habit" prefills a distillation request. The model reviews the marks and calls `annotations` with `action: distill`, which writes one-sentence preferences into a managed block of `$DSH_HOME/AGENTS.md` (`<!-- dsh-mindmark:start -->` … `<!-- dsh-mindmark:end -->`; everything outside the block is preserved). The official workspace-instructions plugin feeds that file to every new session. `lessons` lists, `forget` removes; the panel can remove too.
|
|
46
|
+
|
|
47
|
+
## Install
|
|
48
|
+
|
|
49
|
+
```sh
|
|
50
|
+
dsh plugin --profile web add dsh-mindmark
|
|
51
|
+
```
|
|
52
|
+
|
|
53
|
+
Or in the Harness plugin page → **Add plugin** → `dsh-mindmark`; for local development use `link:<absolute path>`. A Client-side change needs a Harness restart.
|
|
54
|
+
|
|
55
|
+
Options in `cordis.patch.yml`:
|
|
56
|
+
|
|
57
|
+
| Field | Default | Meaning |
|
|
58
|
+
|---|---|---|
|
|
59
|
+
| `defaultMode` | `brief` | Injection view for new sessions (`brief` / `full`); the composer switch overrides it per session |
|
|
60
|
+
| `maxInjections` | `3` | How many turns an unconfirmed mark is re-sent before it settles |
|
|
61
|
+
|
|
62
|
+
## Layout
|
|
63
|
+
|
|
64
|
+
```
|
|
65
|
+
dsh-mindmark/
|
|
66
|
+
├── package.json # dsh.bundle.patch + dsh.client(platform: web)
|
|
67
|
+
├── cordis.patch.yml # one profile row (defaultMode / maxInjections)
|
|
68
|
+
├── index.js # Host: HTTP route, annotations tool, pre-step injection, protocol section
|
|
69
|
+
├── src/store.js # annotation persistence (state, injection counters, per-session mode, message capture, material files)
|
|
70
|
+
├── src/inject.js # injection text: brief view / full revision view
|
|
71
|
+
├── src/lessons.js # long-term preferences: the managed AGENTS.md block
|
|
72
|
+
├── client.js # Client: selection bar, mode switch, panel, CSS highlight painting
|
|
73
|
+
├── locale/{en,zh}.json # plugin-page name and description
|
|
74
|
+
└── test/ # node:test suites (store, HTTP, injection, listener, lessons)
|
|
75
|
+
```
|
|
76
|
+
|
|
77
|
+
## Design notes
|
|
78
|
+
|
|
79
|
+
- **Selection UX reuses the official seam.** The plugin registers into `conversation.chat.text-interaction` (taking over the built-in Quote surface while keeping the Quote action), so the Chat view positions the bar and keeps it attached while scrolling. The Source form lives in the plugin's own overlay, because the bar closes whenever the window loses focus and a native file dialog takes it.
|
|
80
|
+
- **No dependency on host DOM structure.** Each mark stores the selected text, surrounding context and the scroll container's CSS path, and is re-anchored by scoring every occurrence of the quote; message-level offsets and the full message text are captured from the official `data-text-interaction-source` element for the full revision view.
|
|
81
|
+
- **Painting uses the CSS Custom Highlight API** with pending/handled colours per mark; no nodes are inserted into message DOM.
|
|
82
|
+
- **Move uses Alt.** Shift would clash with typing, Fn is invisible to web pages, and Harness binds no Alt+C / Alt+V. The shortcuts are ignored while focus is in a text field.
|
|
83
|
+
- **HTTP channel:** `/plugins/dsh-mindmark/annotations` (GET and POST: list / add / remove / clear / resolve / reopen / set-mode / forget-lesson).
|
|
84
|
+
- Annotations live in `$DSH_HOME/annotations/state.json`, grouped by session, independent of the session log.
|
|
85
|
+
- No `@deepseek-ai/*` imports, Node built-ins only, so the bundle loads straight from a linked directory.
|
|
86
|
+
|
|
87
|
+
## Limits
|
|
88
|
+
|
|
89
|
+
- Rewrite tells the model what you changed; the model's own history is left untouched by design.
|
|
90
|
+
- Source asks for a one-passage edit, but the model rewrites the whole answer each time; keeping the other passages word-for-word is up to the instruction and cannot be enforced.
|
|
91
|
+
- Move works within one answer only; moving across messages is not supported yet.
|
|
92
|
+
- In full mode overlapping marks fall back to the brief list.
|
|
93
|
+
- Marks are painted only for messages currently rendered; virtualised messages re-anchor when scrolled back.
|
|
94
|
+
- Multi-agent sessions get no special handling: marks are stored per session.
|
|
95
|
+
- Message anchoring relies on the Chat view's `data-chat-node-key` attribute; if upstream renames it, the plugin degrades to text anchoring (affects full mode only).
|
|
96
|
+
|
|
97
|
+
## Changelog
|
|
98
|
+
|
|
99
|
+
See [CHANGELOG.md](CHANGELOG.md) (Chinese).
|
|
100
|
+
|
|
101
|
+
## License
|
|
102
|
+
|
|
103
|
+
MIT
|
package/README.md
CHANGED
|
@@ -1,3 +1,126 @@
|
|
|
1
|
-
#
|
|
1
|
+
# dsh-mindmark · 思维标注
|
|
2
2
|
|
|
3
|
-
|
|
3
|
+
**把你的想法直接标在回答上,让模型看见你怎么想。**
|
|
4
|
+
|
|
5
|
+
[English](README.en.md) | 中文
|
|
6
|
+
|
|
7
|
+
像在 Word 里一样,对助手的回答做**高亮 / 下划线 / 删除 / 批注 / 改写 / 依据 / 移动**。
|
|
8
|
+
你的标注会在下一轮作为上下文送给模型:它看到的不再只是自己的回答,而是「你修改后希望它理解的样子」;处理完后它会回传确认。
|
|
9
|
+
标注积累几条后,一键让模型把它们归纳成**长期偏好**,写进你的全局 `AGENTS.md`,新对话自动遵守。
|
|
10
|
+
|
|
11
|
+

|
|
12
|
+
|
|
13
|
+
演示里让模型写一篇小红书笔记,然后依次:高亮要保留的比喻、给没说清的机制画下划线、删掉一句废话、
|
|
14
|
+
批注要求补一句说明、把含糊的仓库提示改写成完整地址、挂上 README 作为依据要求把七种标记列全、
|
|
15
|
+
把新功能那句移到前面。下一轮模型逐条照做并确认,最后点「沉淀为习惯」,归纳出 7 条长期偏好。
|
|
16
|
+
|
|
17
|
+
## 用起来是什么样
|
|
18
|
+
|
|
19
|
+
1. 在任意一条回答里选中一段文字,Harness 自带的选区浮条会出现,分两排:
|
|
20
|
+
- 第一排点一下就生效:**引用 · 高亮 · 下划线 · 删除 · 清除标记**;
|
|
21
|
+
- 第二排要你补充内容:**批注 · 改写 · 依据 · 移动**。
|
|
22
|
+
2. 批注和改写会展开一个小输入框,写完按「确定」(或 Ctrl+Enter)。
|
|
23
|
+
3. **依据**:给这一段挂参考材料,让助手只按材料改这一段。材料可以粘贴文本、贴链接或选一个纯文本文件
|
|
24
|
+
(txt / md / csv / json / html 等;PDF、Word 请先复制文字)。可以顺手写一句要求。
|
|
25
|
+
4. **移动**像剪切粘贴:选中一段按 **Alt+C**(或点「移动」)拿起,在同一条回答里点一下,
|
|
26
|
+
出现一条闪烁的橙色竖线,这段会插到竖线之后;如果拖选一段文字作为目标,就用被移动的那段替换它。
|
|
27
|
+
按 **Alt+V**(或点底部提示条里的「放到这里」)放下,Esc 取消。提示条里可以写一句为什么要移。
|
|
28
|
+
原处显示橙色虚线。
|
|
29
|
+
5. 发下一条消息。模型会先按你的标注调整回答,再调用 `annotations` 工具确认「已处理」。
|
|
30
|
+
6. 输入框左侧有「标注注入」开关,可在**精简**和**完整**之间切换,并显示待处理数量:
|
|
31
|
+
- 精简:下一轮只发送被标注的段落和少量上下文;
|
|
32
|
+
- 完整:下一轮发送整条回答的修订视图(标记内嵌在原文里)。
|
|
33
|
+
7. 顶栏右上角的标注图标打开面板:看全部标注、状态(待处理 / 助手已处理 / 自动收起)、单条删除、
|
|
34
|
+
标为已处理、重新提交;底部「沉淀为习惯」让模型把标注归纳成长期偏好,面板里能看到并移除已沉淀的偏好。
|
|
35
|
+
|
|
36
|
+
## 七种标注对助手的含义
|
|
37
|
+
|
|
38
|
+
| 操作 | 含义 |
|
|
39
|
+
|---|---|
|
|
40
|
+
| 高亮 | 这是我要的:保留并强化这个方向 |
|
|
41
|
+
| 下划线 | 这里不够:展开、澄清或给出依据 |
|
|
42
|
+
| 删除 | 去掉:下一版不要再出现 |
|
|
43
|
+
| 批注 | 按批注内容处理,并逐条回应 |
|
|
44
|
+
| 改写 | 以用户改写后的文字为准 |
|
|
45
|
+
| 依据 | 只改这一段:以用户附的材料为准,与材料冲突时听材料的,并说明改动依据了材料的哪一句 |
|
|
46
|
+
| 移动 | 把这段挪到用户指定的位置(竖线之后;选了目标文字就替换它),原处不再保留,其余内容不动 |
|
|
47
|
+
|
|
48
|
+
## 助手怎么「看到」并回应标注
|
|
49
|
+
|
|
50
|
+
- **步骤前注入**:每一轮开始时,Host 把当前对话里所有待处理的标注渲染成一条用户上下文消息,
|
|
51
|
+
紧跟在你的消息之后进入模型(走官方 `agent/pre-step` 钩子,像其它上下文插件一样落进会话日志,可在「轨迹」里看到)。
|
|
52
|
+
- **协议段落**:对话有待处理标注时,系统提示词里多一小段,解释各种标注的含义和确认方式。
|
|
53
|
+
- **参考材料**:材料单独存在 `$DSH_HOME/annotations/materials/`,不混进 `state.json`,删标注时一起删。
|
|
54
|
+
2000 字以内的材料随标注整段注入;更长的只注入开头摘录和字数,模型需要时调用
|
|
55
|
+
`annotations` 工具的 `material` 动作分段读取(每次 8000 字,用 `offset` 续读)。
|
|
56
|
+
链接由模型用自己的联网工具打开,打不开时会请你改为粘贴。单份材料最多 20 万字。
|
|
57
|
+
- **回传信号**:助手处理完后调用 `annotations` 工具(`action: resolve`,可带 `ids`)确认;
|
|
58
|
+
已确认的标注不再注入,面板里显示「助手已处理」。没确认的会在下一轮再次出现,
|
|
59
|
+
连续 `maxInjections`(默认 3)轮未确认则自动收起,避免无限重发。
|
|
60
|
+
- **沉淀为长期偏好**:面板里的「沉淀为习惯」会把一条归纳请求填进输入框;模型回顾标注后调用
|
|
61
|
+
`annotations` 工具的 `distill` 动作,把每条一句的偏好写进 `$DSH_HOME/AGENTS.md` 的一个受管区块
|
|
62
|
+
(`<!-- dsh-mindmark:start -->` … `<!-- dsh-mindmark:end -->`,区块外内容原样保留)。
|
|
63
|
+
官方的工作区指令插件会把这个文件喂给每个新会话。`lessons` 列出、`forget` 移除;面板里也能逐条移除。
|
|
64
|
+
|
|
65
|
+
## 安装
|
|
66
|
+
|
|
67
|
+
```sh
|
|
68
|
+
dsh plugin --profile web add dsh-mindmark
|
|
69
|
+
```
|
|
70
|
+
|
|
71
|
+
或者在 Harness 的插件页 → **添加插件** → 填 `dsh-mindmark`;本地开发用 `link:<本目录绝对路径>`。
|
|
72
|
+
Client 半边改动后需要重启 Harness。
|
|
73
|
+
|
|
74
|
+
`cordis.patch.yml` 里可配:
|
|
75
|
+
|
|
76
|
+
| 字段 | 默认 | 说明 |
|
|
77
|
+
|---|---|---|
|
|
78
|
+
| `defaultMode` | `brief` | 新会话的注入模式(`brief` 精简 / `full` 完整),输入框开关按会话覆盖 |
|
|
79
|
+
| `maxInjections` | `3` | 一条未确认的标注最多重发几轮,超过自动收起 |
|
|
80
|
+
|
|
81
|
+
## 结构
|
|
82
|
+
|
|
83
|
+
```
|
|
84
|
+
dsh-mindmark/
|
|
85
|
+
├── package.json # dsh.bundle.patch + dsh.client(platform: web)
|
|
86
|
+
├── cordis.patch.yml # 往 profile 插件树插入一行(可配 defaultMode / maxInjections)
|
|
87
|
+
├── index.js # Host:HTTP 路由、annotations 工具、pre-step 注入、协议段落
|
|
88
|
+
├── src/store.js # 标注持久化(状态、注入计数、按会话的模式、消息全文捕获、参考材料文件)
|
|
89
|
+
├── src/inject.js # 注入文本:精简视图 / 完整修订视图
|
|
90
|
+
├── src/lessons.js # 长期偏好:AGENTS.md 受管区块的读写
|
|
91
|
+
├── client.js # Client:选区浮条、模式开关、面板、CSS 高亮绘制
|
|
92
|
+
├── locale/{en,zh}.json # 插件管理页里的名称与描述
|
|
93
|
+
└── test/ # node:test 单测(store + HTTP + 注入 + 监听器 + lessons)
|
|
94
|
+
```
|
|
95
|
+
|
|
96
|
+
## 实现要点
|
|
97
|
+
|
|
98
|
+
- **选区交互复用官方机制**:注册 `conversation.chat.text-interaction` 槽位(接管内置「引用」的面板并保留引用按钮),
|
|
99
|
+
浮条由 Chat 视图自己定位、随滚动移动,不再自己监听鼠标。需要填写的依据表单放在插件自己的浮层里,
|
|
100
|
+
因为浮条在窗口失去焦点时会关闭,而系统的选文件对话框正会抢走焦点。
|
|
101
|
+
- **不依赖宿主 DOM 结构**:标注记录选中原文 + 前后文 + 滚动容器 CSS 路径,恢复时按原文和前后文打分重新定位;
|
|
102
|
+
另外在消息正文(官方的 `data-text-interaction-source` 元素)里记录偏移并捕获全文,供完整模式渲染修订视图。
|
|
103
|
+
- **绘制用 CSS Custom Highlight API**,七种样式各有待处理/已处理两种颜色;不往消息 DOM 里插节点。
|
|
104
|
+
- **移动的快捷键选 Alt**:Shift 会和输入冲突,Fn 网页读不到,Harness 自己没有占用 Alt+C / Alt+V;
|
|
105
|
+
焦点在输入框里时快捷键不生效,不影响打字。
|
|
106
|
+
- **HTTP 通道**:`/plugins/dsh-mindmark/annotations`,GET 与 POST(list / add / remove / clear / resolve / reopen / set-mode / forget-lesson)。
|
|
107
|
+
- 标注存在 `$DSH_HOME/annotations/state.json`,按会话分组;与会话日志无关,删会话不会自动删标注。
|
|
108
|
+
- 不 import 任何 `@deepseek-ai/*` 包,只用 Node 内置模块,所以能从链接目录直接加载。
|
|
109
|
+
|
|
110
|
+
## 限制
|
|
111
|
+
|
|
112
|
+
- 改写只能让助手「知道你改成了什么」,助手看到的历史仍是它原来的回答;这是有意为之,不改会话日志。
|
|
113
|
+
- 「依据」要求只改那一段,但模型每次都会重写整条回答,其余段落一字不动只能靠指令约束,无法从技术上保证。
|
|
114
|
+
- 移动只支持同一条回答之内;跨消息移动暂不支持。
|
|
115
|
+
- 完整模式按消息全文内嵌标记,重叠的标注会退回到精简列出。
|
|
116
|
+
- 标注只对当前渲染在页面里的消息绘制;长对话里被虚拟化滚出屏幕的消息,滚回来时重新定位。
|
|
117
|
+
- 多智能体场景未做特别处理:标注按会话存储,子会话各自独立。
|
|
118
|
+
- 消息锚定依赖 Chat 视图的 `data-chat-node-key` 属性;上游改名时退化为纯文本锚定(只影响完整模式)。
|
|
119
|
+
|
|
120
|
+
## 更新记录
|
|
121
|
+
|
|
122
|
+
见 [CHANGELOG.md](CHANGELOG.md)。
|
|
123
|
+
|
|
124
|
+
## 许可
|
|
125
|
+
|
|
126
|
+
MIT
|