dsh-quote-followup — native select-to-quote for DSH Web

[tr1v3r]

I built dsh-quote-followup, a Web-only plugin that adds native select-to-quote interaction to DeepSeek Harness.

What it does

1. Select any excerpt inside the conversation transcript.
2. Click the floating ❐ Quote action.
3. The excerpt appears in the composer as a compact conversation-reference chip.
4. Quote more excerpts or write a focused follow-up and send.

Each quote behaves as one atomic, removable chip. The composer shows only the excerpt, while the conversation bubble icon carries the quote meaning. Existing drafts remain intact, and multiple excerpts can be collected before sending.

Native DSH integration

The plugin reuses DSH's registered ReferenceChipNode, conversation icon, and business color. Its codec expands each chip into a model-readable Markdown blockquote only when the prompt is sent, keeping the composer compact without hiding context from the model.

The floating action and serialized quote frame follow the active DSH locale. The plugin falls back to plain text when native chip support is unavailable, and its browser-independent path is verified on Chromium and Firefox.

Install

Requires DSH >=0.1.2-rc.1.

dsh plugin --profile web add dsh-quote-followup

Repository: https://github.com/tr1v3r/dsh-quote-followup
npm: https://www.npmjs.com/package/dsh-quote-followup

Feedback and integration suggestions are welcome.

[argszero]

Nice — reusing the same registered ReferenceChipNode / icon / business color as the @file / @session references is exactly the right way to make a quote feel native instead of bolted-on. Two questions from the consumer side:

1. Attribution provenance — when the codec expands a chip into the Markdown blockquote at send time, does the model-visible text carry which message the excerpt came from (message id / session seq)? For follow-up tasks ("fix what you quoted") source traceability matters as much as the excerpt itself; if it's only the raw text, consider embedding a hidden citation marker or emitting the source ref alongside the blockquote.

2. Durability after compaction/rotation — if a chip ever stores a session-message reference rather than just text, the target can disappear after a compaction fold or a tail rotation. Text-only chips are immune; ref-bearing ones need a stale-ref fallback (degrade to plain text) rather than a dead chip.

Also worth stating in the README: this is a Web client-side extension — the composer/send path is client-owned, so nothing here needs a host-side plugin, which is a clean boundary to document for anyone auditing the wire format (the model only ever sees the expanded blockquote, never the chip).

[tr1v3r]

Thanks for the thoughtful review — both questions hit exactly the right design axes.

1. Attribution provenance — Today the model-visible text is > [Quote · conversation] plus the raw excerpt; the role is display-only metadata, so the serialized frame carries no message id / session seq yet. Source traceability for "fix what you quoted" is a fair ask. I've noted it as a follow-up: the clean shape would be a hidden citation marker emitted only in the serialized frame (keeping the chip visually slim). Whether it can be a stable message anchor depends on what the host eventually exposes to client plugins, so it's parked until that surface exists.

2. Durability after compaction/rotation — Already covered by design: chips are text-only. The stored ref is just {text, role, truncated} with no session-message references, so compaction folds and tail rotation cannot produce a dead chip.

3. Client-side boundary — Agreed and done: the README now has a Scope section stating that the composer/send path is client-owned, the model only ever sees the expanded blockquote (never the chip), and no host-side plugin surface is involved.

[tr1v3r]

v0.2.5 — 引用轮次溯源 / Turn provenance for serialized quotes

中文：选中内容来自 DSH 聊天行时，发送时的序列化引用现在会附带对话轮次编号，模型可以定位"第 3 轮引用的内容"，让"修复你引用的那个问题"这类追问有据可依；无轮次标记的场景保持原有框架。同批更新：README 明确声明这是 Web 客户端侧扩展（composer/发送路径由 client 拥有，模型只看到展开的 Markdown 引用块，从不接触 chip 本身，无需任何 host 侧接口）；chip 刻意只存文本（摘录正文 + 显示用角色提示 + 可用轮次编号），不含全局 session-message 引用，compaction 折叠与会话轮转后依然有效；重录干净会话的 demo。npm latest 已是 0.2.5，发布流程改为 CI tag 门控发布（npm provenance + 测试矩阵 + pack 断言）。要求 DSH >=0.1.2-rc.1，安装不变：

dsh plugin --profile web add dsh-quote-followup

English: When the selection comes from a DSH chat line, the serialized quote now carries the conversation turn number, so the model can locate "the content quoted from turn 3" — follow-ups like "fix what you quoted" get real source traceability; quotes without turn markers keep the existing frame. Also in this release: the README now states this is a Web client-side extension (the composer/send path is client-owned; the model only ever sees the expanded Markdown blockquote, never the chip — no host-side plugin surface); chips deliberately store text only (excerpt + display-only role hint + available turn number), so they survive compaction folds and session rotation; demo re-recorded in a clean session. npm latest is 0.2.5, now published via a CI tag-gated workflow (npm provenance + test matrix + pack assertion). Requires DSH >=0.1.2-rc.1; install unchanged:

dsh plugin --profile web add dsh-quote-followup