kur-compact-trigger

Category / 分类： dsh-plugin · DeepSeek Harness 第三方插件

Per-session compaction for DeepSeek Harness. Official auto-compact is global only. This plugin lets each chat pick its own threshold from the header — without keeping the settings window open.

会话级压缩：官方自动压缩只能调全局。本插件在标题栏给每个会话单独设阈值，不必一直开着设置窗。

---

English

Why this exists

Official @deepseek-ai/dsh-compaction-basic auto is a single global policy (typically 80% of the model window). Every session shares it.

That is awkward once token prices go up (DeepSeek and others): a long research thread may need a large window, while a short task chat should compact early. Opening Settings for every switch is not how people actually work.

This plugin adds session-level control: a compact button on the chat header, optional overrides in sessions.<id>, and a global default only as fallback.

It does not replace the official summarizer. It decides when and which range to compact, then calls compaction.compactRegion().

Setting	Default	Meaning
thresholdTokens	200000	Compact when this session reaches this many tokens
retainTokens	32000	Keep roughly this many recent tokens verbatim
maxAttempts	3	Max compact loops in one turn
sessions.<id>	—	Per-session override; empty = follow global

The compact cut is moved to a tool-call / tool-result pairing boundary so a tool result is not split from its call.

Host plane, shared by every preset. Do not inject compaction (that service lives in the preset realm). At agent/pre-step it uses ctx.agentPresets.serviceFor(agent, 'compaction').

Screenshots

Header — change this session without leaving the chat:

Session override — leave fields blank to follow the global default:

Settings → Auto Compact — global fallback only; slider max follows the model context window:

Install

1. Copy this package to ~/.dsh/profiles/node_modules/kur-compact-trigger/ (or clone / npm pack into that path).
2. Merge examples/cordis.patch.yml into each profile’s cordis.patch.yml (desktop, web, feishu, …).
3. Add examples/settings.yaml under kur-compact-trigger: in ~/.dsh/settings.yaml.
4. Restart DSH Desktop / the profile process.

Tune the global default in Settings → Auto Compact. Tune one chat from the header Compact control.

Official auto vs this plugin

	Official compaction-basic auto	This plugin
Scope	Global (one ratio for all sessions)	Per session, plus a global default
Typical trigger	80% of the model context window	Absolute token threshold (default 200k)
UI	Settings only	Header + Settings

Both can run. If you want only session-level policy, turn official auto off and use this plugin.

License

MIT

---

中文

为什么做这个

官方 @deepseek-ai/dsh-compaction-basic 的 auto 只能控制全局（通常是模型窗口的 80%）。所有会话共用同一套。

模型（尤其 DeepSeek）涨价之后，长上下文很贵：有的会话需要大窗口慢慢做，有的短任务根本不该带着几十万 token 往下跑。每次去设置页改全局、改完再改回来，也不方便——设置窗不能一直开着。

所以才有会话级压缩：标题栏给当前对话单独设阈值；sessions.<id> 可覆盖；全局默认只当没覆盖时的兜底。

本插件不实现摘要模型，只决定何时、压哪一段，再调用官方 compaction.compactRegion()。

配置	默认	含义
thresholdTokens	200000	本会话总 token 达到此值时压缩
retainTokens	32000	尽量保留尾部最近约这么多原文 token
maxAttempts	3	同一轮最多连压几次
sessions.<id>	—	按会话覆盖；留空跟随全局

切分点会挪到工具调用 / 工具结果配对平衡处。挂在 host 平面；不要 inject compaction。在 agent/pre-step 里用 ctx.agentPresets.serviceFor(agent, 'compaction') 取当前 preset。

截图

标题栏入口：不离开对话、也不用开设置窗，改这一路的阈值：

本会话覆盖；留空则跟随全局：

设置 → 自动压缩：只是全局兜底；滑条上限跟随模型窗口：

安装

1. 把本包放到 ~/.dsh/profiles/node_modules/kur-compact-trigger/（或 clone / npm pack）。
2. 将 examples/cordis.patch.yml 合并进各 profile 的 cordis.patch.yml。
3. 在 ~/.dsh/settings.yaml 写入 examples/settings.yaml 的 kur-compact-trigger: 段。
4. 重启 DSH Desktop 或对应 profile。

全局默认在 设置 → 自动压缩；单个会话用标题栏 压缩阈值。

官方自动压缩 vs 本插件

	官方 compaction-basic auto	本插件
范围	只能全局（所有会话同一比例）	会话级 + 全局兜底
典型触发	模型窗口的 80%	绝对 token 阈值（默认 20 万）
界面	只有设置页	标题栏 + 设置页

两条线可以同时开。若只想按会话控成本，关掉官方 auto，只用本插件。

许可证

MIT