Introduction

DeepSeek Harness (DSH) adopts a plugin-based architecture. oh-my-deepseek-harness is an OMX-style workflow skill set rewritten for DSH.

The plugin is ported from the skill design of oh-my-codex and has been comprehensively rewritten for DSH’s native capabilities (such as subagent, workflow, goal, and ralph). It provides developers with a set of structured agent role routing, team orchestration, and quality assurance tools.

Plugin Positioning

  • Core value: Introduces structured workflow skills into the DSH ecosystem.
  • Design source: Based on oh-my-codex (MIT), adapted for DeepSeek models.
  • Maintainer: hatsuyuki0103.

Installation and Enablement

Install the plugin using the official command:

dsh plugin --profile web add oh-my-deepseek-harness

After installation, you need to restart dsh web. After restarting, the skills included in this package will appear in the session skill catalog, and the model can load them by name through the skill tool.

Skill List

The skill workflow designs include the following functional categories:

Planning

  • deep-interview: Socratic deep interview. It converges into an executable specification through round-by-round questioning and ambiguity scoring.
  • plan: Strategic planning. It supports four modes: interview, direct, consensus, and review.
  • ralplan: Consensus planning. It reviews in the sequence Planner→Architect→Critic and includes RALPLAN-DR deliberation.
  • prometheus-strict: Strict planning. It achieves a strict logical closed loop through Metis clarification, Momus challenge, and Oracle self-check.

Execution

  • ralph: Persistent execution loop. It uses DSH’s native ralph tool, combining session discipline with independent architectural review.
  • visual-ralph: Visual Ralph. Based on reference images/baselines, it proceeds through user approval, implementation, read_image scoring (threshold 90), and pixel-level iteration.
  • autopilot: Strict autonomous delivery loop. It integrates the deep-interview, ralplan, ultragoal, code-review, and ultraqa processes.
  • team: Coordinates parallel teams. It uses the workflow tool to orchestrate shared task lists and verification channels.
  • ultrawork: Parallel execution engine. Acceptance criteria come first, with support for both direct and evidence-based channels.
  • ultragoal: Persistent multi-goal execution. It aggregates goals and manages the .omx/ultragoal ledger.
  • ultraqa: Adversarial dynamic e2e QA. It includes a matrix of 9 malicious scenario categories and a test-diagnose-fix loop.

Quality

  • code-review: Comprehensive code review. It uses two independent channels (code-reviewer + architect) + deterministic merge gates.
  • security-review: Security audit. It performs a seven-point health check covering secrets, dependencies, injection, authentication, configuration, and network.
  • analyze: Read-only in-depth analysis. It classifies and ranks explanations of evidence, inferences, and unknowns.
  • build-fix: Build repair. It covers reproduction, root cause analysis, minimal fixes, and full revalidation.
  • tdd: Test-first. It follows the Red-Green-Refactor rule.
  • ai-slop-cleaner: Anti-redundancy cleanup. It locks tests first, then cleans up smells one by one.
  • git-master: Git expert. It handles atomic commits, rebasing, and branch management (Conventional Commit).
  • design: DESIGN.md design-source workflow.
  • aliyun-media: Alibaba Cloud Token Plan image generation / video generation / video editing.

Operations and Bootstrapping

  • cancel: Stop the workflow. It ends goals, terminates background tasks, and finalizes states.
  • doctor: DSH environment diagnosis. It health-checks the harness, skills, plugins, and environment.
  • note: Session notes. It generates .omx/notepad.md.
  • skill-authoring: DSH skill authoring guide. It covers contracts, registration, packaging, and publishing.
  • ecomode: Token-saving discipline. It prioritizes self-execution, batches delegations, and closes out background work.
  • planner / architect / critic / analyst / momus / oracle / executor / verifier / code-reviewer / test-engineer: Accompanying role prompts for reuse by subagent.

Usage

Simply say the skill name in the conversation to trigger it; no special prefix is required (the skill catalog hints to the model).

Examples:
* Input: “deep interview”
* Input: “Interview me”
* Input: “Don’t assume”
* Input: “Use the deep-interview skill”

Dependencies and Permissions

  • Dependency: @deepseek-ai/cordis
  • Permissions: Requires filesystem:read and filesystem:write permissions.
  • Compatibility: DSH 0.1.0-rc.6

Notes

  • Visual features: visual-ralph depends on a vision model and the read_image tool.
  • Internal mechanisms: visual-verdict, frontend-ui-ux, and vision are internal mechanisms and are not independently ported.
  • Workspace: Retains the .omx/ workspace convention (context / interviews / specs / plans).
  • Open-source license: MIT. The skill workflow design is derived from oh-my-codex (MIT).

Conclusion

This plugin provides DSH with a full-process toolchain from planning and execution to quality assurance. It is suitable for development scenarios that require structured workflows and deep agent collaboration.