Preface¶
Complex workflows often stall not because there are not enough tools, but because executors cannot be swapped, state lacks a single source of truth, and boundaries lack assertions. For example, a task that has been running for hours may become impossible to recover progress from once it crashes; if you want to change how work is executed, it usually means modifying core code; adding a new team source is the same.
dsh-flow addresses these three issues at the kernel shape level: it defines rules/ as a pure core (no IO, no ctx), declares execution and sources via declarative seams, and uses kernel.js as the single composition root to decide which concrete implementations to mount at deployment time.
What Is This¶
DeepSeek Harness’s unified agent canvas: weaving requirement sessions, multi-agent teams, member tasks and messages, and dependency DAGs into a single interactive timeline.
- Owner: rootkiller6788
- Category: Workflow
- License: MIT
Core Features¶
This plugin adopts a microkernel architecture pattern, standing in dialectical contrast with the flat kernel of dsh-agent-teams.
- Pure core architecture: The
rules/directory contains no IO operations orctxcontext, keeping core logic pure. - Declarative seams: By defining service interfaces, providers, and consumers, implementations are dynamically mounted at runtime.
- Single composition root:
kernel.jsis the only composition root, responsible for deciding which implementation to mount at deployment time. - Toolset: Provides 13
flow_*tools. - Runtime dependencies: Zero runtime dependencies; no build step.
- Runtime injection: Automatically injects
@deepseek-ai/dsh-client-runtime, supporting the Web platform.
Installation and Enablement¶
In the currently verified facts, no explicit installation command is provided.
It is recommended to obtain the plugin source code and directory information through the following links:
* Directory page: https://www.skillhub.cn/plugins/rootkiller6788/dsh-flow
* GitHub: https://github.com/rootkiller6788/dsh-flow
Typical Usage¶
The core value of this plugin lies in making complex multi-agent collaboration processes visible. It integrates requirement timelines, multi-agent team hierarchies, member tasks and message records, and task dependency DAGs into a single interactive canvas. Developers can configure kernel.js to mount different executors or team source implementations.
Applicable Scenarios and Notes¶
- Environment requirements: Requires Node.js >= 22.19.0.
- Architecture characteristics: The core layer performs no IO operations; dependency injection is completed at runtime.
- Permission risks: The plugin runs with the permissions of the current DSH process; review the source code and license before installation.
- Ecosystem comparison: If you need a flattened architecture and have no requirement for custom seams, consider
dsh-agent-teams; if you need flexible kernel extension and composition,dsh-flowprovides a corresponding solution.
Conclusion¶
dsh-flow introduces declarative seams and the composition root pattern, providing another kernel implementation form for DeepSeek Harness. It maintains a capability surface consistent with dsh-agent-teams, while decoupling executors from the core in its internal structure. For scenarios requiring fine-grained control over workflow execution logic and state management, this plugin offers a practical technical path.