Minglink/dsh-jailbreak-workflow

What it does

dsh-jailbreak-workflow is a plugin for DeepSeek Harness that lets you design a multi-step process on a canvas, compile it into a single plan prompt, and hand that prompt to the current session's model to execute. Steps can be connected with arbitrary conditions, covering sequential, parallel, merge, branch, loop, and rollback flows. The plugin ships no execution engine: scheduling and judgment are delegated to the model (workflow design and compilation).

It targets people who want repeatable, inspectable model workflows without building orchestration infrastructure. A companion repository, dsh-workflows, hosts ready-made .wflite.zip workflow packages that can be imported, including a "破甲制作" (jailbreak-production) research pipeline (supported workflows).

Quick start

Prerequisites: Node.js 22 or newer (Node.js badge) and DeepSeek Harness 0.2.0 or 0.2.1, including preview builds such as 0.2.1-rc.2.

For the web version:

dsh plugin --profile web add @xiaoso/dsh-workflow-lite
dsh web

The plugin, @xiaoso/dsh-workflow-lite, is installed through the dsh CLI (web install steps). On the desktop app, open 插件 in the sidebar, choose 添加插件 → 安装第三方插件, enter @xiaoso/dsh-workflow-lite, install, and restart DeepSeek Harness (desktop install steps).

After installation a 工作流 tab appears at the top of the session page. Either ask the model to create a workflow in one sentence (for example, write code, review it, and loop back to fix up to three rounds), or drag steps from the left palette onto the canvas, fill in prompts, and connect them. Click 执行 to run, or ask the model to execute it (usage). Enabling 记录运行状态 in workflow settings records each run and updates the canvas as it progresses (tracking).

Ready-made packages from dsh-workflows are imported through the 工作流中心 (workflow center), for example the 破甲制作 workflow package.

Features and constraints

The main capabilities include canvas editing with branches and loops, attaching files, folders, URLs, or Skills to steps, pre-execution questions to collect inputs, model-driven editing of workflows, per-step run status, version management, and import/export as .wflite.zip with bundled files and Skills (feature table). A workflow version can also be exported as a standalone skill (SKILL.md plus per-step task descriptions) usable without the plugin.

Paths support {workflow} for the workflow's own directory and {instance} for the current run's output directory (path variables).

The project is explicitly experimental: features and data formats may change incompatibly between versions (experimental note). Workflow memory is especially immature—concurrent edits to the same memory are not locked and last write wins, and reinstalling an exported skill overwrites accumulated memory. The author reports no stable speedup in early repeated-run trials and says more real workflows are needed to evaluate it (workflow memory). The license is marked NOASSERTION, so reuse terms are not established by the repository metadata.

HotRepo's take

On the HotRepo new-project list for 2026-11-03, the project recorded 15 new stars and 9 new forks over a roughly one-hour window, moving from 104 to 119 stars; 60 forks in total is notable for a plugin that depends on a specific harness version. No GitHub daily rank is recorded, and the sample is a single short snapshot, so it does not support a claim about sustained growth. Technically, the design of using the current session's model as the scheduler keeps the plugin thin and the workflows portable, but it also means execution quality depends on the model and the plan prompt rather than on deterministic orchestration.