Minglink/dsh-workflows

What This Is

dsh-workflows is a collection of ready-to-import .wflite.zip workflow packages for the dsh-jailbreak-workflow plugin, which runs inside DeepSeek Harness. Instead of building a workflow from scratch, users download a package and import it into the plugin.

The first published workflow is 破甲制作 (jailbreak authoring), a 15-step pipeline for systematic jailbreak research against a target LLM. The README says it takes a target model name and API key, then produces three outputs: a jailbreak prompt file, a usage manual (frame matrix plus reflection prompt library), and a research report. The pipeline covers automated probe testing, refusal-pattern analysis, customized jailbreak prompt generation, API validation with iteration, and final output. A bundled jailbreak-research skill contains test scripts and prompt templates (11).

This is security research tooling aimed at people studying or red-teaming model refusal behavior, not a general-purpose LLM application.

Getting Started

1. Download the workflow zip from the workflow list: 破甲制作.wflite.zip, saving it locally.
2. Open DeepSeek Harness, then go to the plugin's 工作流中心 (workflow center).
3. Choose 导入 (Import), select the downloaded zip, and confirm where resources are placed.
4. The full canvas should appear after import. Fill in the execution prompts it asks for (target model, API key, etc.) and click 执行 (Execute) (13).

The README notes the plugin itself must be installed first; install instructions live in the plugin repository. A screenshot of the full canvas is available in assets/canvas.webp.

Constraints

• Requires DeepSeek Harness 0.2.x as the host application (3).
• Workflows are consumed through the dsh-jailbreak-workflow plugin; the zip packages are not standalone programs.
• The workflow needs a target model name and API key, so runs consume API access and may incur provider costs.
• The README does not describe the scripts inside the zip, so the internal behavior and safety of the workflow steps should be reviewed before running them with real credentials.
• Licensed under MIT (16).

HotRepo's take

The repository entered HotRepo's new-project list on 2026-11-04 for the Python language board, reporting 10 stars gained over the roughly two-hour window from 2026-11-03 22:45 to 2026-11-04 00:45 UTC, reaching 112 stars and 58 forks. That is a short observation window, so it supports only the fact of early activity, not a trend. The narrow scope (one published workflow plus a companion plugin) and the sensitive jailbreak-research purpose are the more useful signals for judging fit than the star count.