Introduction

In the DeepSeek Harness (DSH) development context, making models follow rigorous engineering methodologies (such as test-first development, fault isolation, and code review) usually requires complex prompt assembly. dsh-praxis aims to solve this problem by codifying senior engineers’ workflows into a standard skills library and integrating it into DSH as a plugin, enabling models to automatically enter structured engineering processes.

Plugin Positioning

dsh-praxis is an engineering methodology skills library plugin, authored by JohnXu22786. It addresses the lack of structured engineering workflows in models by codifying design conversations, plan writing, test-first development, and system debugging into a set of automatically triggerable skills.

Core Capabilities

The plugin includes 14 built-in skills, covering all stages of software development. These skills follow the Agent Skills standard and are compatible with DSH’s local skills discovery mechanism.

  1. method-compass: At task startup, understand the skills library and choose the first skill.
  2. design-conversation: When requirements are ambiguous or the design is undefined, first conduct a conversation to confirm details.
  3. implementation-blueprint: Once the design is confirmed, write an executable step-by-step plan.
  4. blueprint-execution: Execute according to the plan, set checkpoints for verification, and report deviations.
  5. test-first-cycle: When writing code, test by following a red-green-refactor cycle.
  6. fault-isolation: When encountering anomalies, identify root causes based on evidence rather than guessing.
  7. completion-proof: Before claiming completion, provide observable evidence.
  8. task-splitting: Split parallel tasks, organized around contracts, dispatch, and integration.
  9. delegated-build: Delegate the plan to subagents with two-round review gates.
  10. review-preflight: Before requesting code review, perform self-checks and self-tests.
  11. feedback-assimilation: After receiving feedback, digest it, fix the issues, and address each item.
  12. lane-isolation: During parallel development, ensure branches have independent workspaces.
  13. branch-conclusion: The full closing process before a branch is merged into the main branch.
  14. skill-authoring: Create, modify, or validate new skills.

Installation and Enablement

After installation, no manual configuration is required; the model automatically loads skills when it encounters trigger scenarios. The plugin also supports zero-code integration.

Standard Plugin Installation

dsh plugin --profile demo add github:JohnXu22786/skill-framework

Zero-code Integration

The skills/ directory in the plugin repository is itself a standard Agent Skills collection and can be directly copied into DSH’s local skills root directory:

cp -r skills/* ~/.dsh/skills/

This path is loaded by DSH’s built-in dsh-skill-filesystem discovery mechanism, so this plugin does not need to be installed.

Configuration and Usage

Configuration Options

The plugin supports specifying the skills directory through a configuration file. In cordis.patch.yml, you can configure:
- providerName: The skill provider name, defaulting to praxis-bundled.
- skillsDir: The skills root directory, defaulting to skills under the package root; array format is supported.

Usage Examples

In a conversation, you can directly specify a combination of skills. For example:

First produce an implementation plan, then execute it.
During debugging, prioritize the fault isolation workflow and provide evidence before completion.

Skills trigger one another via dependency declarations; for example, fault-isolation depends on method-compass and blueprint-execution.

Environment Requirements

  1. DeepSeek Harness dsh version must be ≥ 0.1.0-rc.6 (the standard configuration already mounts dsh-skill and dsh-tool-skill).
  2. Node.js version must be ≥ 22.
  3. Installation of dsh-skill and dsh-tool-skill is required.

Conclusion

dsh-praxis automates senior engineers’ workflows through a standardized skills library. It is suitable for DSH scenarios requiring rigorous engineering standards and can significantly improve the model’s compliance in code generation, debugging, and collaboration. For more details, see the plugin directory or GitHub repository.