DeepSeek Harness 插件

JessenReinhart/dsh-tokenslash

Star 数 ★ 10 分类 用量与计费 收录于 2026-10-08

基于 Jev System One 分流、动态工具裁剪与子代理路由的智能 Token 优化插件。

安装

# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)

dsh plugin --profile web add github:JessenReinhart/dsh-tokenslash

装任何插件都等于在你的机器上跑第三方代码,权限和你本人一样大——能读你的文件、用你的凭据、访问网络,工具审批管不到它。GitHub 来源的插件还会在安装时执行构建脚本——pnpm 默认拦截,所以安装可能停在 ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED 或 ERR_PNPM_IGNORED_BUILDS;dsh 会打印出需要添加的确切键名,把它加进该 profile 的 pnpm-workspace.yaml 的 allowBuilds 下,重跑一次即可装上。放行构建本身就是一次信任判断:请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。

README

该插件的 README 只有英文版本。

TokenSlash

Keep the context that matters. Drop the overhead that doesn't.

TokenSlash is a Cordis plugin for DeepSeek Harness (DSH) that makes agent context more deliberate. It uses lightweight, structured decisions to select task-relevant tools and prompt sections, keep tool output manageable, and help agents route work without permanently losing access to capabilities.

Your agent should spend its context solving the task, not rereading the descriptions of tools it will never call.

What it does

Capability What you get
Task-aware prompt pruning Selects relevant tool schemas and prompt sections for the current task.
Output and observation hygiene Compacts oversized tool results and can mask older observations to keep context focused.
Subagent routing Uses structured decisions to select model tiers and decide when a fork can run with less parent context.
Tool recovery Lets an agent inspect a pruned tool's schema or unlock that tool for the current turn.
Workflow guardrails Optional goal, turn-completion, and tool-safety hooks.
Visibility In-chat savings telemetry, prune statistics, history, and a settings UI.

TokenSlash is configurable rather than one-size-fits-all. Tool pruning supports normal, extreme, and off modes; tools can be pinned, and individual modules can be enabled or disabled.

Quick start

Install from GitHub using the DSH plugin manager:

dsh plugin add github:jessenreinhart/dsh-tokenslash

Then open the DSH settings for TokenSlash, choose a triage provider, and configure the model and module switches for your setup.

  • Default preset: 9Router Proxy
  • Other presets: OpenCode Zen, TypeSafe Jev, OpenRouter, Inception Labs (Mercury Decide), LocalJev, and a custom endpoint
  • Runtime: Node.js 20+

The provider is used for narrow, structured decisions. It is not the conversational model that writes code or answers general questions.

Built-in agent tools

TokenSlash adds focused utilities that help an agent use its existing tools more efficiently.

Tool Use it for
tokenslash_index Get a structural outline of a file: signatures, exports, symbols, and line ranges. Use it to locate the relevant part before requesting a larger read.
tokenslash_batch Run independent tool calls in parallel, or run calls sequentially when order matters.
tokenslash_peek Inspect the schema of a pruned tool, or unlock it so the agent can use it during the current turn.
tokenslash_diagnose Classify errors and test failures, suggest files to inspect, or rank candidate tests by relevance.
tokenslash_triage Ask the configured decision engine a narrow probability or multiple-choice question.

A note on triage

The triage engine is a micro-decision engine, not a chat model. Use it for decisions such as:

  • Is this task suitable for a cheaper subagent model?
  • Can this fork proceed without the parent's full conversation history?
  • Which tool group is relevant to this prompt?
  • Is this test failure likely to be flaky?

Do not use it to write code, inspect files, search the web, or answer open-ended questions. Keep those jobs with the main agent and its normal tools.

How pruning stays reversible

  1. Read the task. TokenSlash evaluates the current prompt with the configured decision provider.
  2. Keep the useful surface. Relevant tool groups and prompt sections stay available; irrelevant schemas can be pruned.
  3. Recover on demand. The agent can call tokenslash_peek to inspect or unlock a tool that turns out to be needed.
  4. Fail open. The default configuration favors preserving normal tool access when triage is unavailable or fails.

That last point matters: pruning should reduce clutter, not strand an agent halfway through a task.

Configuration

Most users can configure TokenSlash from the DSH settings UI. The main controls include:

Setting Purpose
Provider / model Select a supported decision endpoint and model.
Tool pruning mode normal for conservative pruning, extreme for more aggressive pruning, or off to disable tool pruning.
Pinned tools Keep named tools available regardless of the pruning decision.
Fail open Preserve tool access when the decision provider cannot be used.
Modules Toggle routing, fork decoupling, prompt pruning, output compaction, observation masking, prefix-cache alignment, and guardrails.

See docs/TRIAGE.md for provider and decision-format details, and docs/COMPATIBILITY.md for the tested DSH/Cordis host contracts.

Benchmarks and measurement

Token savings depend on the task, available tool set, DSH runtime, and decision provider. A pruning pass can add overhead when most tools are still relevant, so there is no fixed savings percentage promised by this project.

The repository includes the benchmark harnesses and raw result files so you can inspect the measurements and compare them with your own workload:

Development

Requires Node.js 20 or later.

npm install
npm test

The test command runs the project's Node.js test suite.

Project links

License

MIT © Jessen Reinhart

内容来自项目 README(GitHub)↗

链接

同类插件

查看整个分类 →

社区评论

评论公开保存在 GitHub Discussions。加载评论会连接 GitHub 和 Giscus;发表内容需要 GitHub 账号。