dsh-engineering-workflow

An engineering workflow layer for DeepSeek Harness (dsh). One install adds the 工程工作流 (Engineering Workflow) agent preset — a disciplined-engineer mode with five hard-gated phases — plus six workflow skills that carry the methodology.

The workflow methodology is adapted from obra/superpowers (MIT): brainstorming, writing-plans, TDD, subagent-driven development, and verification-before-completion, reworked for dsh's native tools (plan mode + exit_plan_mode, background subagent/subagent_fork, workflow orchestration, goals, and the preset/skill system).

The five phases

Phase	Skill	Gate	dsh mechanism
① Requirements clarification	workflow-requirements	Intent approved before any code	ask_user_question, one question at a time
② Plan approval	workflow-planning	Plan approved via exit_plan_mode	plan mode, todo_write after approval
③ TDD implementation	workflow-tdd	Failing test before production code	pwsh/bash test runs
④ Parallel subagent execution	workflow-subagents	Per-task review + ledger	background subagent, send_message, list_agents
⑤ Verified finishing	workflow-verification	Fresh evidence before claims	full suite + branch-finish menu

The master skill engineering-workflow routes every non-trivial task to the right phase and enforces the discipline rules (rationalization red flags included).

Install

dsh plugin --profile <name> add github:82c86b8z86-stack/dsh-engineering-workflow

(Or npm install the package into your profile and add dsh-engineering-workflow to dsh.profile.bundles.)

Restart dsh once so the host plugin mounts. On startup it syncs the preset into ~/.dsh/.agent-presets/engineering-workflow; the preset then appears in the new-session preset picker as 工程工作流. The sync is idempotent — upgrading the plugin updates the preset and its skills automatically.

Manual / development fallback without a restart:

node scripts/sync-presets.mjs

dsh re-discovers presets on every roster read, so the synced preset is selectable immediately.

How it works

dsh-engineering-workflow (bundle)
├── cordis.patch.yml        inserts one host plugin row
└── lib/index.js            host plugin: syncs presets/ → ~/.dsh/.agent-presets,
│                           announces the workflow via a system-prompt section
└── presets/engineering-workflow/
    ├── agent.cordis.yml    full toolset composition (adapted from the shipped
    │                       cordis preset, MIT): shell, filesystem, jobs, goals,
    │                       plan mode, compaction, delegation (subagent/subagent_fork/
    │                       workflow/ralph), ask-user, todo, web, skills
    ├── preset.yml          roster metadata (name / description / order)
    ├── skills/             6 workflow skills (one SKILL.md per directory)
    └── NOTICE              attribution

The preset wires its skills through @deepseek-ai/dsh-skill-filesystem with customSkillDirs rooted at the preset's own directory — the same pattern the shipped cordis preset uses, so the skill catalog travels with the preset wherever it is installed.

Development

pnpm install
pnpm test          # preset-sync unit tests
pnpm run validate  # structural validation of the bundled preset
pnpm run sync      # sync the preset into ~/.dsh/.agent-presets

License

MIT. The preset composition is adapted from the DeepSeek Harness built-in cordis preset (MIT); the workflow methodology is adapted from obra/superpowers (MIT); the preset-sync host-plugin pattern follows @linxin666/dsh-liangshen (Apache-2.0). See presets/engineering-workflow/NOTICE.