[Agent]: Add support for DeepSeek Harness (DSH)

[AnInteger]

Agent Name

DeepSeek Harness (DSH)

Official Website

https://github.com/deepseek-ai/deepseek-harness

Agent Type

CLI tool (command-line interface)

It also ships a browser-based GUI (dsh web); the CLI and GUI share the same skill-discovery machinery.

CLI Command (if applicable)

dsh

Installation Method

npm install -g @deepseek-ai/dsh

Command/Workflow Structure

DSH is skills-based. It discovers skills from (in precedence order):

• <projectRoot>/.dsh/skills (native project root, highest rank)
• <projectRoot>/.agents/skills (shared agents standard)
• user-level roots (~/.dsh/skills, ~/.agents/skills)

Skill format: one-level-deep directory bundles <name>/SKILL.md (or flat <name>.md) with YAML frontmatter requiring a kebab-case name and a description — the same agentskills.io layout Spec Kit already scaffolds for Codex, Zed, Antigravity, and Devin. Skills are invoked through the /-trigger input in the Web GUI (e.g. /speckit-specify build photo albums); the harness injects the rendered skill content into the turn. Project guidance in AGENTS.md at the repo root is loaded natively.

Non-interactive dispatch is available via the one-shot headless profile:

dsh --profile headless "/speckit-specify <feature description>"

Argument Passing Pattern

No file placeholder substitution. The text the user types after the /name token ships verbatim as the task; the skill body sees it in the triggering message. (Spec Kit's $ARGUMENTS placeholder is therefore left literal in the skill body, same as other skills-based agents like Zed — the core templates already phrase user-input handling accordingly.)

Popularity/Usage

Widely used (thousands+ of users) — 196k stars on GitHub.

Documentation Links

• Agent/skill discovery: @deepseek-ai/dsh-skill-filesystem (roots, ranks, bundle format, frontmatter contract)
• Skill catalog and /name invocation: @deepseek-ai/dsh-tool-skill
• Workspace instructions (AGENTS.md/CLAUDE.md loading): @deepseek-ai/dsh-agent-instructions
• Headless one-shot profile: @deepseek-ai/dsh-headless

Use Case

We develop with DSH as the primary coding agent and want to use Spec Kit's spec-driven workflow (/speckit-constitution → /speckit-specify → /speckit-plan → /speckit-tasks → /speckit-implement → /speckit-converge) directly inside DSH sessions. The integration is a natural fit because DSH already speaks the skills format Spec Kit generates for other agents.

Example Command File

specify init --integration dsh would scaffold .dsh/skills/speckit-specify/SKILL.md:

---
name: "speckit-specify"
description: "Create or update the feature specification from a natural language feature description."
compatibility: "Requires spec-kit project structure with .specify/ directory"
metadata:
  author: "github-spec-kit"
  source: "templates/commands/specify.md"
---

(command body — standard Spec Kit specify workflow)

Contribution

• I can help test the integration
• I can provide example command files
• I can help with documentation
• I can submit a pull request for the integration

Additional Context

I have a working implementation ready and will open a PR referencing this issue: a minimal SkillsIntegration subclass installing to .dsh/skills/ (isolated root, multi-install safe, so it coexists with other agents), slash-style invocations (/speckit-<command>), and CLI dispatch via dsh --profile headless. Verified end-to-end: specify init --integration dsh scaffolds 10 skills; the real DSH filesystem skill provider discovers all of them as user- and model-invocable; /speckit-constitution and /speckit-specify were executed by a real DSH agent and produced the expected artifacts.

AI disclosure: this issue and the accompanying integration were prepared with AI assistance (DeepSeek Harness, glm-5.3) under human supervision — the design decisions were reviewed and approved step by step, and all changes and test evidence were personally verified before submission.