dsh-persona-select · 人设管理与人设注入
DSH 插件：把人设正文注入系统提示词，把历史对话注入上下文。设置页的人设管理负责保存与编辑，对话模式选择行下方可以随时切换人设。
Bilingual README. 中文 · English
啥子鱼思考
中文
这个插件做什么
它负责两件事：把一段文字送进系统提示词，把一段对话送进上下文。为了说清楚，本文与插件设置界面统一使用两个叫法：
| 叫法 | 指的是 |
|---|
| 人设正文 | 注入系统提示词的那段内容 |
| 历史对话 | 作为上下文注入的一段"已经发生过"的对话 |
两者按人设成对保存：在设置页把人设正文和历史对话写好、存成一条预设「人设」，之后在对话里选中这条人设，两者按各自的规则进入这一次请求。
注入的规则
人设正文进入系统提示词的人设段，注入位置可自行编辑：「段序权重」决定这一段排在系统提示词的哪个位置。
- 数值：按权重大小排序，数字越小越靠前（默认
-1，排在预设里那句 "You are a coding agent…" 之前）。
- 追随：点「追随」后，人设段被放到所有系统提示词段之后。
- 权重只影响排序，永远不会混进人设正文；每条人设各自保存自己的权重，编辑框跟着当前选中的那条人设走。
- 必须填整数（
-1000000000 ~ 1000000000）：小数、指数、带单位都会被拒绝保存。
历史对话在新对话的第一步作为 user 侧上下文注入一次，并按仿照 DSH 自身注入的格式包裹：
- 整段包进一对
<system-reminder>，开口之后是 CONTEXT_ROAD:，结尾是 CONTINUE_CONTEXT.；正文里出现的收尾标签会被转义，保证壳子不会被提前闭合。
- 往期对话的 user 侧内容包在
<user_message> 里；模型侧的思考包在 <thinking> 里、说的话包在 <assistant_message> 里；双方非对话内容包在 <file> 之类的标记里；压缩摘要保持 <compacted-summary>…</compacted-summary> 原样。
已经注入过的对话不会重复注入（那会把整段对话的缓存读取推倒重来）。所以历史对话只在每个新对话的第一步生效；如果想让一个已经聊起来的对话也带上历史，只能把历史写进人设正文——但效果不好，不推荐。
人设管理
- 新建、编辑、保存、删除人设，复制为新人设（内置人设只读，要改就复制一份副本再改）。
- 每一条人设包含：名称、人设正文、段序权重、历史对话。
- 设置分开保存：每个对话保存自己的选择与开关，切换人设、改动勾选项都不会影响新对话和进行中的对话。设置页有「当前 / 全局」两个页签，分别对应"这一个对话"和"新对话默认跟随的设置"。
- 保存前会先校验：正文为空、名称为空、权重不是整数都会拒绝保存并说明原因；正文超过建议上限会提示但完整保留。
- 未保存的改动在切换人设时会先询问，避免默默丢掉。
屏蔽与「纯聊天」
四个开关各自撤掉一类模型本来能看到的系统侧输入：
| 开关 | 撤掉什么 |
|---|
| 只留人设 | 系统提示词里除人设段以外的所有段：harness 身份、部署人设、工具指引、计划模式策略、harness 源码说明。人设为空时，系统提示词就是空的。 |
| 屏蔽 runtime 上下文 | 两类：① 沙箱策略、审批策略、委派规则等以 user 角色快照下发的动态上下文；② 包在 <system-reminder> 里的 runtime 快照与 skill 目录。 |
| 屏蔽 schema 工具 | 工具定义（它走独立的协议字段，不是系统提示词段，通常是最大的一笔 token）。打开后模型没有可调用的工具。 |
| 屏蔽其他注入内容 | AGENTS.md / CLAUDE.md 之类的指令基线与逐步时间读数——它们同样是以 user 角色消息追加进来的，只能按来源过滤。 |
四个开关全开、且没有选中人设，就是纯聊天：系统侧注入归零，模型只看到你发的内容。对话界面的选择框里有「一键纯聊天 / 恢复系统输入」按钮直接切换这两个状态。
工具保活通道（「可复用工具」）：屏蔽 schema 工具的保险丝。DSH 的工具账本要求每个对话有一份非空的工具声明基线，基线一旦被清空，回来的工具都会被打上 deferred 标记并被服务方以 INVALID_REQUEST 拒绝。勾上它，插件只在两个时刻保留完整工具列表各一轮——这个对话的第一次请求，以及疑似声明序列重启（例如压缩）时——先让基线建立起来，之后无论怎么开关屏蔽，工具都还能回来。
以下为关闭所有系统测注入后，你能够真实达到的token消耗
中文tok
导入与导出
导出（设置页编辑器里那条人设的正文 + 段序权重 + 历史对话一起走）：
- 导出 JSON：纯 JSON，不压缩、不丢信息。
- 导出图片：由你自己挑一张 PNG 当封面，同样的数据打进图片里一起作为一张角色卡下载。图片本身的数据一个字节都不改，包括AI生图或捏人游戏角色卡(导出为png格式的游戏，例如恋活)，任何看图软件照常显示它。
- 单次导出上限 32 MB；选了非 PNG、或 PNG 结构不完整（读不出 IHDR）会被直接拒绝并说明原因。
- 导出取的是编辑器里那条人设已保存的版本：有未保存改动会先询问，新建但还没保存的人设不能导出。
导入：
- 读本插件导出的
.json，或「导出图片」生成的那张 .png，作为新人设追加进列表——绝不覆盖已有条目。
- 直接导入 DSH 自己
/export 出来的 Session 日志：接受 /export 产生的 .zip，也接受裸的 session.vN.jsonl / session.jsonl，以及 .jsonl.zstd / .zst。导入时两个独立开关：
- 导入思考过程：决定模型侧的
<thinking> 跟不跟着进来。
- 展开压缩：默认只取压缩摘要
<compacted-summary> 与最后一次摘要之后的对话；勾上后反过来——取除摘要之外的全部内容，把压缩展开回原文。历史从未压缩过时，这个勾选项不改变任何东西。
对话界面的选择框
对话模式选择行下方那个「人设」展开：
- 切换当前对话的人设，或选「不注入」。
- 一键纯聊天 / 恢复系统输入。
- 层徽标告诉你这条设置是本对话的还是全局的。
- 状态点：
- 绿点：工具基线可用，一切正常。
- 黄点：本对话选定人设的历史对话与当初注入的那份对不上了；本对话的人设正文没进系统提示词（例如对话用了声明
complete: true 的「极简模式」）；本对话已经聊过，所以这条人设的历史对话不再注入；或「可复用工具」没勾。
- 红点：本对话选的人设已被删除（人设段现在是空的），或这个对话的首次请求没带任何工具（工具基线为空，本对话内工具回不来）。
- 设置页的「当前」页会按来源列出这个对话里被过滤掉的注入块：runtime 快照、skill 目录、工作区指令、时间读数。
数据文件
<DSH_HOME>/persona-select/（DSH_HOME 默认 ~/.dsh）：
| 文件 | 内容 |
|---|
personas.json | 全局设置、人设列表、每个对话的选择与开关 |
personas.json.v2-backup | 旧版升级时的备份，可删 |
history/<编号>.txt | 各条人设的历史对话正文 |
- 名称最长 40 字；人设正文建议不超过 20000 字（超过会提示，但完整保留、不截断）；最多 200 条人设；最多保留 500 条对话记录。
- 会话记录会被自动清理：已经不存在、也不在活动列表里的对话记录会被回收（设置页也可主动触发），避免文件无限膨胀。
界面
设置页与对话界面均跟随宿主语言，内置中文 / English 两套文案。
内置人设声明
插件中“鲸鱼娘”的形象描述来自@ZipZipPipe，遵守CC BY-NC-SA 4.0协议再创作并开源，请尊重相关形象创造者。
English
What this plugin does
Two jobs: put a block of text into the system prompt, and put a stretch of conversation into the context. To keep the wording precise, this README uses the same two names as the plugin UI:
| Term | Meaning |
|---|
| Persona text | the content injected into the system prompt |
| Conversation history | a stretch of "already happened" dialogue injected as context |
The two are stored together as one persona: write the text and the history in the settings page, save them as a persona preset, then pick that persona in a conversation and both enter the request by their own rules.
How injection works
Persona text becomes the persona section of the system prompt, and its position is yours to edit: the "Section order weight" decides where that section sits.
- A number: sections sort by weight, lower comes first (default
-1, ahead of the preset line "You are a coding agent…").
- Follow: click it and the persona section is appended after every system-prompt section.
- The weight only affects ordering and never enters the persona text. Each persona keeps its own weight, and the field follows whichever persona you are editing.
- It must be an integer (
-1000000000 to 1000000000): decimals, exponents and units are refused on save.
Conversation history is injected once, as user-side context, at the first step of a new conversation, wrapped in a shell that imitates how DSH itself injects context:
- the whole block sits inside one
<system-reminder> pair, opened by CONTEXT_ROAD: and closed by CONTINUE_CONTEXT.; any closing tag occurring in the body is escaped so the shell can never be closed early.
- inside it, past user-side content goes inside
<user_message>; the model side's thinking goes inside <thinking>, its spoken content inside <assistant_message>; non-conversational content from either side (tool calls, multimodal parts) goes inside markers such as <file>; compaction summaries stay wrapped in <compacted-summary>…</compacted-summary>, verbatim.
A conversation that has already been injected is never injected again — doing so would tear down the whole conversation's cached input(KV cache). History therefore only takes effect on the first step of each new conversation. If an already-running conversation must carry history across a persona switch, you can write that history into the persona text itself — the result is poor, so it is not recommended.
Managing personas
- Create, edit, save and delete personas, and fork as new (built-in personas are read-only; fork a copy to edit one).
- Each persona carries: name, persona text, section order weight, conversation history.
- Settings are saved per conversation: each conversation keeps its own selection and switches, so switching personas or toggling a switch never affects new conversations or conversations in progress. The settings page has a Current / Global tab pair — "this conversation" versus "what new conversations follow".
- Saving validates first: empty text, empty name or a non-integer weight is refused with a reason; text over the recommended limit is reported but kept in full.
- Unsaved edits trigger a confirmation before switching personas, so nothing is silently dropped.
Suppression and pure chat
Four switches, each removing one class of system-side input the model would otherwise see:
| Switch | What it removes |
|---|
| Persona only | Every system-prompt section except the persona: harness identity, deployment persona, tool guidance, plan-mode policy, harness-source notes. With no persona selected the system prompt is empty. |
| Suppress runtime context | Two families: (1) sandbox policy, approval policy and delegation rules delivered as user-role snapshots; (2) the runtime snapshot and skill catalog framed as <system-reminder> blocks. |
| Suppress tool schemas | Tool definitions (a separate wire field, not a prompt section, and usually the biggest token item). With this on the model has nothing to call. |
| Suppress other injections | The AGENTS.md / CLAUDE.md instruction baseline and per-step time readings — also user-role messages, so they can only be filtered by provenance. |
All four on with no persona selected is pure chat: zero system-side injection, the model sees only what you send. The picker in the conversation view has one-click Make it pure chat / Restore system input buttons.
Tool-keeping channel ("Reusable tools"): the safety valve for "Suppress tool schemas". DSH's tool ledger requires a non-empty tool-declaration baseline for the session; once that baseline is emptied, tools that come back are all marked deferred and the provider rejects the request with INVALID_REQUEST. When checked, the plugin keeps the full tool list for exactly one turn at two moments — the session's first request, and a suspected declaration-series restart (a compaction, for example) — so a non-empty baseline gets written. Afterwards tools can always come back no matter how you toggle suppression.
This is what token consumption really looks like after every system-side injection is disabled!
英文tok
Import and export
Export (the editor's persona text + section order + history together):
- Export JSON: plain, uncompressed, lossless.
- Export image: pick a PNG yourself as the cover and the same data is packed into the image and downloaded with it. The image is not modified by a single byte and every normal viewer still displays it.
- 32 MB per export. A non-PNG cover, or a structurally incomplete PNG (no readable IHDR), is refused outright with the reason.
- Export takes the persona's saved version: unsaved edits trigger a confirmation first, and a new persona that has never been saved cannot be exported.
Import:
- Read a
.json this plugin exported, or the .png from "Export image", and append it as new personas — existing entries are never overwritten.
- Import a DSH
/export session log directly: accepts the .zip produced by /export, or a bare session.vN.jsonl / session.jsonl, or a .jsonl.zstd / .zst. Two independent options:
- Include thinking: whether the model side's
<thinking> comes along.
- Expand compaction: by default the box gets the
<compacted-summary> blocks plus the dialogue after the last one; ticked it flips — everything except those summaries, with the compaction expanded back into the original text. With no summary in the imported text, ticking changes nothing.
The picker in a conversation
The "Persona" dropdown under the mode selector:
- Switch this conversation's persona, or choose Off.
- One-click pure chat / restore system input.
- A scope badge tells you whether the setting is for this chat or global.
- Status dots:
- Green: the tool baseline is usable, all good.
- Yellow: this session's persona history no longer matches what was injected; the persona text never made it into the system prompt (e.g. the conversation uses a "minimal" preset declaring
complete: true); the conversation has already been chatting, so its history will not be injected; or "Reusable tools" is unchecked.
- Red: the persona selected for this session has been deleted (the persona section is empty), or this session's first request carried no tools at all (empty tool baseline — tools cannot return within this session).
- The settings page's Current tab lists the injection blocks filtered out in this conversation, by kind: runtime snapshot, skill catalog, workspace instructions, time readings.
Data files
<DSH_HOME>/persona-select/ (DSH_HOME defaults to ~/.dsh):
| File | Contents |
|---|
personas.json | global settings, the persona list, and each conversation's selection and switches |
personas.json.v2-backup | migration backup, safe to delete |
history/<id>.txt | the conversation history body of each persona |
- Names up to 40 characters; persona text recommended under 20000 characters (over that it is reported but kept in full, never truncated); at most 200 personas; at most 500 session records.
- Session records are garbage-collected: records for conversations that no longer exist and are not in the active list are reclaimed (also triggerable from the settings page), so the file cannot grow without bound.
Interface
Settings page and conversation picker follow the host language, with Simplified Chinese and English built in.
Built-in persona notice
The “DeepSeek chan” likeness used by the built-in persona comes from @ZipZipPipe; it is re-created and open-sourced under the CC BY-NC-SA 4.0 licence. Please respect the creator of the likeness.
许可 / License
MIT
