dsh-suggest-prompt

为 DeepSeek Harness 开发的「建议提示词」插件：每个 agent 回合完成后，通过一次有界的辅助 LLM 调用，在会话日志中写入一条建议的下一条提示词；Web 输入框在草稿为空时把它渲染成浅灰色幽灵文本，按 Tab（默认）即可采纳进草稿（与 Claude Code 一致）。

本仓库是这两个包的权威源码（source of record），由两个包构成：

特性

• 默认轻量：建议模型默认走 deepseek-official / deepseek-v4-flash；通过 provider / model 可换成任意路由（例如本地 OpenAI 兼容网关）。
• 只发最后一轮：默认只把最后一轮的用户输入与 AI 最终回答发给建议模型（maxRecentTurns 默认为 1），中间的工具调用 / 推理过程一律不发送。
• 有界调用：字节 / 令牌 / 超时上限、转录长度预算、建议可见字符上限，全部可配置。
• 安全：转录在发送前脱敏（密钥形状被掩蔽）；输出净化（控制序列、围栏、引号剥离、单行化）并做语义过滤（元文本、评价套话、助手口吻等被当作「无建议」丢弃）。
• 无建议是常态：模型回复为空或不合格时静默跳过，不报错、不写事件、不打扰。
• 免调用重显：删回空草稿会重新显示已持久化的建议，不再发新的模型请求。
• 快捷键可配：采纳快捷键通过 acceptKey 配置，默认 Tab。

安装

前置条件

• Node.js ^22.19 或 >=24、pnpm。
• 一个基于 deepseek-harness 的 dsh 部署（web profile）。Web 输入框需要具备「幽灵输入」能力（InputActions.setGhost）；请确认所用 harness 版本已内置该输入机能力。

通过 npm 安装（发布后）

注意：@studyzy/dsh-suggest-prompt 与 @studyzy/dsh-client-ui-suggest-prompt 的完整依赖链尚未全部发布到 npm（上游 @deepseek-ai/dsh-compact、@deepseek-ai/dsh-environment 等仍缺失）。等 registry 补齐后：

npm install @studyzy/dsh-suggest-prompt @studyzy/dsh-client-ui-suggest-prompt

然后在 profile 的 cordis.yml / 补丁层挂上两行（见下「配置」示例）。

从本仓库源码接入（当前方式）

把这两个包放进 harness 工作区（或通过 file: 依赖引用本仓库），并在 profile 补丁层插入两行。例如 ~/.dsh/profiles/web/cordis.patch.yml：

- id: suggest-prompt
  name: '@studyzy/dsh-suggest-prompt'
  config:
    maxInputBytes: 4096
    maxOutputTokens: 512
    timeoutMs: 60000
    maxRecentTurns: 1
    maxTranscriptChars: 12000
    maxSuggestionChars: 240
    provider: ccr
    model: ttswitch/deepseek-v4-flash-ioa
    acceptKey: Tab

- id: ui-suggest-prompt
  name: '@studyzy/dsh-client-ui-suggest-prompt'

provider / model 同时省略时，会继承当前主请求最近一次记录的路由，无需为建议单独配置模型。

配置

maxOutputTokens 提示：如果建议模型是推理模型（回答前会「思考」），思考会消耗输出预算；maxOutputTokens 偏小时，流会在输出建议文本之前就以 max-tokens 结束。给推理模型留足预算（例如 512）。

工作方式

• 宿主在 turn/end（reason=completed）时触发生成；按会话 + 回合去重，下一个完成回合会中止上一个在途生成。
• 建议写入会话日志的 suggest-prompt/suggested 事件，suggestPrompt 投影把它暴露给 Web 端。
• 幽灵文本只在满足以下条件时显示：建议对应最新完成回合、agent 空闲、草稿为空；键入即隐藏，删回空草稿重新显示。
• 按 acceptKey（默认 Tab）把建议填入草稿（可编辑后再发送）；焦点不在输入框或处于 IME 组合输入时不触发，Tab 也只在显示幽灵时才被拦截（否则保持默认焦点行为）。

模型体验

• 系统提示词：把模型限定为「以用户口吻预测下一条提示词」，禁止生成内容或元文本，给出具体正反例；回复语言跟随会话（最后一条用户消息含 CJK → 简体中文，否则 English）。
• 模型看到的输入：默认只有最后一轮的 [User Message] / [Assistant Response] 带标签块（已脱敏、受 maxTranscriptChars 约束）。
• 请求前记录：确切的框架化输入与系统提示在派发前写入 suggest-prompt/request 事件，满足「模型可见 ⟺ 日志可重建」。
• 成本：每个完成回合至多一次辅助请求，受 maxInputBytes / maxOutputTokens 约束；主 agent 请求不增加任何 token。

安全

• 转录脱敏：AWS AKIA…、OpenAI sk-…、GitHub ghp_/gho_/ghu_、Slack xox-…、JWT、Stripe rk_… 等密钥形状在发送前被掩蔽为占位标签。
• 输出净化：ANSI/OSC/CSI/DCS 序列、C0/C1 控制符、双向覆盖符、孤立代理项被剥离；引号与代码围栏被去除；压缩为单行并截断到 maxSuggestionChars。
• 语义过滤：元文本（"no suggestion"、"stay silent"）、错误回显、评价套话（"thanks"、"looks good"、谢谢、不错）、助手口吻（"Let me…"、"I'll…"、我来、我帮你）、多句 / 过长回复、孤立单词会被当作「无建议」丢弃，而不是显示。

已知限制

• 每个完成回合都会生成（与输入框是否已有内容无关），幽灵文本只在草稿为空时显示。
• 被中止（取代）的生成不会为较早回合留下建议。
• 空回复或被过滤的回复 = 该回合无建议：不写 suggest-prompt/suggested 事件，投影保持 null，也不记录警告。
• 投影保留最后一条建议：重新打开旧会话会显示其最终建议，且不发起新的模型调用。
• 建议模型的路由与预算由部署配置决定；推理模型需要更大的 maxOutputTokens。

开发

pnpm install
pnpm build      # host tsc + client tsdown bundle
pnpm test       # vitest
pnpm typecheck

安装说明：本仓库依赖已发布的 @deepseek-ai/* 包（deepseek-harness 工作区）。上游少量内部包（@deepseek-ai/dsh-compact、@deepseek-ai/dsh-environment）尚未出现在 npm registry，pnpm install 可能失败，直到 harness 的 registry 补齐。完整测试矩阵在 harness monorepo 内运行；本仓库是两个包的权威源码副本。

许可

MIT

dsh-suggest-prompt

Suggested-next-prompt plugin for the DeepSeek Harness. After every completed agent turn, a bounded auxiliary LLM call writes one suggested next prompt into the session log; the web composer renders it as ghost text in the empty input — press Tab (default) to adopt it into the draft (the Claude Code behavior).

This repository is the authoritative source of record for the two packages:

Features

• Lightweight by default: the suggestion model defaults to deepseek-official / deepseek-v4-flash; set provider / model to route anywhere (for example a local OpenAI-compatible gateway).
• Last turn only: by default only the last completed turn's user input and assistant final answer are sent to the suggestion model (maxRecentTurns defaults to 1); intermediate tool calls / reasoning are never included.
• Bounded: byte / token / timeout caps, a transcript budget, and a visible-character cap on the suggestion — all configurable.
• Safe: transcripts are secret-redacted before framing; output is sanitized (control sequences, fences, quotes stripped, single line) and semantically filtered (meta-text, evaluative filler, assistant-voice phrasing are dropped as "no suggestion").
• Silent no-suggestion: an empty or rejectable model reply is skipped quietly — no error, no event, no noise.
• Re-arm without a call: deleting back to an empty draft re-shows the persisted suggestion with no new model request.
• Configurable shortcut: the adopt shortcut is set via acceptKey, default Tab.

Install

Prerequisites

• Node.js ^22.19 or >=24, pnpm.
• A dsh deployment built from the DeepSeek Harness (web profile). The web composer must expose the ghost-input capability (InputActions.setGhost); verify your harness version ships that input-machine seam.

From npm (once published)

Note: the full dependency chain of @studyzy/dsh-suggest-prompt and @studyzy/dsh-client-ui-suggest-prompt is not fully on the npm registry yet (upstream @deepseek-ai/dsh-compact, @deepseek-ai/dsh-environment, etc. are still missing). Once the registry is complete:

npm install @studyzy/dsh-suggest-prompt @studyzy/dsh-client-ui-suggest-prompt

Then mount the two rows in your profile's cordis.yml / patch layer (see the config example below).

From this repository's source (current)

Put the two packages into the harness workspace (or reference this repo via a file: dependency), and insert the two rows in your profile patch layer. For example ~/.dsh/profiles/web/cordis.patch.yml:

- id: suggest-prompt
  name: '@studyzy/dsh-suggest-prompt'
  config:
    maxInputBytes: 4096
    maxOutputTokens: 512
    timeoutMs: 60000
    maxRecentTurns: 1
    maxTranscriptChars: 12000
    maxSuggestionChars: 240
    provider: ccr
    model: ttswitch/deepseek-v4-flash-ioa
    acceptKey: Tab

- id: ui-suggest-prompt
  name: '@studyzy/dsh-client-ui-suggest-prompt'

Omit both provider and model to inherit the route of the most recently logged main request — no need to configure a model just for suggestions.

Configuration

On maxOutputTokens: if the suggestion model reasons before answering, thinking consumes the output budget. A small maxOutputTokens ends the stream with max-tokens before any suggestion text is produced — leave a generous budget (e.g. 512) for reasoning models.

How it works

• The host triggers generation on turn/end (reason completed), deduplicated per session and turn; the next completed turn aborts the in-flight generation.
• The suggestion is appended to the session log as the suggest-prompt/suggested event, and the suggestPrompt projection exposes it to the web side.
• The ghost text shows only when the suggestion answers the latest completed turn, the agent is idle, and the draft is empty; typing hides it, deleting back to an empty draft re-shows it.
• Pressing acceptKey (default Tab) fills the draft (editable, not sent). It is ignored while focus is outside the composer or during IME composition; Tab is intercepted only while a ghost is displayed (otherwise it keeps its default focus behavior).

Model Experience

• System prompt: binds the model to predicting the user's next prompt in the user's own voice, forbids generating content or meta-text, and gives concrete examples and anti-examples; the reply language follows the conversation (简体中文 when the last user message contains CJK, otherwise English).
• What the model sees: by default only the last turn, framed as labelled [User Message] / [Assistant Response] blocks (redacted, bounded by maxTranscriptChars).
• Pre-dispatch logging: the exact framed input and system prompt are recorded in the suggest-prompt/request event before dispatch, satisfying the model-visible ⟺ logged invariant.
• Cost: at most one auxiliary request per completed turn, bounded by maxInputBytes / maxOutputTokens; the main agent request gains zero tokens.

Security

• Transcript redaction: AWS AKIA…, OpenAI sk-…, GitHub ghp_/gho_/ghu_, Slack xox-…, JWTs, and Stripe rk_… secret shapes are masked before the transcript reaches the model.
• Output sanitization: ANSI/OSC/CSI/DCS sequences, C0/C1 control characters, bidirectional overrides, and lone surrogates are stripped; quotes and code fences are removed; text is collapsed to one line and truncated to maxSuggestionChars.
• Semantic filtering: meta-text ("no suggestion", "stay silent"), error echo, evaluative filler ("thanks", "looks good"), assistant-voice phrasing ("Let me…", "I'll…"), multi-sentence or over-long replies, and stray single words are dropped as "no suggestion" instead of shown.

Known Limitations

• Generation runs after every completed turn regardless of whether the composer already holds text; the ghost is only displayed while the draft is empty.
• A superseded (aborted) generation leaves no suggestion for the older turn.
• An empty or filtered reply means "no suggestion" for that turn: no suggest-prompt/suggested event is written, the projection stays null, and no warning is logged.
• The projection persists the last suggestion, so reopening an old session shows its final suggestion without a new model call.
• The suggestion route and budget are deployment configuration; reasoning models need a larger maxOutputTokens.

Development

pnpm install
pnpm build      # host tsc + client tsdown bundle
pnpm test       # vitest
pnpm typecheck

Install caveat: this repo depends on the published @deepseek-ai/* packages (the DeepSeek Harness workspace). A small number of internal packages referenced by the published dsh-* releases are not yet on the npm registry (@deepseek-ai/dsh-compact, @deepseek-ai/dsh-environment), so pnpm install may fail until the harness registry is complete. The full test matrix runs inside the harness monorepo; this repo is the source-of-record copy for the two packages.

License

MIT