dsh-everything-wp

English | 中文

DSH (DeepSeek Harness) adapter for everything-wp — WordPress plugin development AI toolkit.

Status: Phase 3 Complete (v1.7.0-dsh.3)

Full adaptation of everything-wp to DSH: 5 agents + 18 slash commands, with mode gating and inline rules injection. End-to-end verified through a complete plan → todo × 3 → review → verify workflow that produced a working WordPress contact form plugin (5 PHP files, 554 lines).

Component	Count	Status
Slash commands	18	✅
Agents (system prompt sections)	5	✅
Rules (inlined)	5	✅
Command docs (for doc-inline injection)	14	✅

Architecture

Two command registration patterns:

Pattern	Use case	Implementation
agent-backed	Command has invokes_agent	System prompt section (mode-gated) + agent.steer() trigger
doc-inline	Command has no agent, only docs	Handler reads command doc and injects as steer message

Key design decisions:

• Mode gating: System prompt sections include conditional comments so the model only follows agent instructions when the corresponding slash command was the most recent input. Multi-command agents (e.g., code-quality shared by /wp-test-generate, /wp-test, /wp-verify) use OR-joined gate conditions.
• Inline injection: All rules/*.md content is embedded directly into system prompt sections as INLINE APPENDIX. The inlineRefs() function rewrites @everything-wp/ references and skill auto-detection tables to "already inlined" notices, preventing the model from searching for external files.
• Agent collaboration via filesystem: Agents coordinate through spec/ (planner writes plans) and src/ (task-executor writes code, code-reviewer reads diffs) — no explicit orchestration needed.

Installation

Prerequisites

• Node.js 22.17+
• DSH v0.1.x (npm install -g @deepseek-ai/dsh)
• pnpm (npm install -g pnpm)

Build

cd dsh-everything-wp
npm install
npm run build

Register with DSH

# Register as a local plugin
dsh plugin --profile web add ./dsh-everything-wp

# Start DSH Web UI
dsh web

# Open http://127.0.0.1:3080

Verify

In the DSH Web UI, type / to see all 18 /wp-* commands in the autocomplete dropdown.

Example workflow:

/wp-plan Add a contact form shortcode [simple_contact_form] with validation, DB storage, and email notification
/wp-todo spec/<feature>/01-data-layer.md
/wp-todo spec/<feature>/02-email-layer.md
/wp-todo spec/<feature>/03-form-layer.md
/wp-review
/wp-verify

Commands

Agent-backed (7 commands)

Command	Agent	Purpose
/wp-plan <feature>	wp-planner	Generate implementation plan (6-step checklist → spec/)
/wp-todo <spec-file>	wp-task-executor	Execute tasks from a spec file (TDD)
/wp-review	wp-code-reviewer	Review current diff or latest commit
/wp-submit-review	wp-submission-reviewer	WordPress.org submission compliance check (6 items)
/wp-test-generate	wp-code-quality	Generate PHPUnit tests
/wp-test	wp-code-quality	Run PHPUnit tests
/wp-verify	wp-code-quality	Full quality verification (PHPStan → PHPUnit → PHPCS)

Doc-inline (11 commands)

Command	Purpose
/wp-api-wrapper <class>	Generate AJAX API wrapper class
/wp-custom-table <table>	Generate custom DB table + CRUD
/wp-frontend-page <slug>	Generate frontend page template
/wp-init-plugin <slug>	Initialize new plugin scaffold
/wp-init-theme <slug>	Initialize new theme scaffold
/wp-list-table <class>	Generate WP_List_Table child class
/wp-make-block <name>	Generate Gutenberg block
/wp-option-page <slug>	Generate settings/options page
/wp-release <version>	Prepare release (changelog, version bump)
/wp-rest-api <resource>	Generate WP_REST_Controller subclass
/wp-ajax <action>	Generate AJAX handler

File Structure

dsh-everything-wp/
├── package.json              # DSH bundle declaration
├── cordis.patch.yml          # Cordis plugin tree patch
├── tsconfig.json
├── src/
│   └── index.ts              # Entry: registers 5 agents + 18 commands
├── dist/                     # tsc output (gitignored)
├── agents/                   # 5 agent prompts (frontmatter removed)
│   ├── planner.md
│   ├── task-executor.md
│   ├── code-reviewer.md
│   ├── submission-reviewer.md
│   └── code-quality.md
├── commands/                 # 14 command docs (frontmatter removed)
│   ├── plan.md, todo.md, review.md
│   ├── api-wrapper.md, rest-api.md, wp-ajax.md, ...
└── rules/                    # 5 rule files (inlined into system prompt)
    ├── acceptance-criteria.md
    ├── wp-essentials.md
    ├── performance-lighthouse.md
    ├── org-submission.md
    └── phpstan.md

How It Works

Mode Gating

Each agent's system prompt section is wrapped in conditional comments:

<!-- MODE GATE: PLANNING -->
<!-- Active only when user's most recent slash command was /wp-plan. Otherwise ignore this section. -->

For multi-command agents:

<!-- MODE GATE: CODE_QUALITY -->
<!-- Active only when user's most recent slash command was /wp-test-generate OR /wp-test OR /wp-verify. -->

Inline Injection

The buildPrompt() function assembles each agent's system prompt:

function buildPrompt(modeName, commandNames, agentPath, appendixPaths) {
  const agentBody = inlineRefs(loadMarkdown(agentPath));
  const appendices = appendixPaths
    .map((p) => `\n\n## INLINE APPENDIX: ${p}\n\n${loadMarkdown(p)}`)
    .join("");
  return [
    `<!-- MODE GATE: ${modeName} -->`,
    `<!-- IMPORTANT: Do NOT attempt to load external files in rules/, skills/, or commands/ directories. -->`,
    agentBody,
    appendices,
  ].join("\n");
}

The inlineRefs() function rewrites all external references:

• @everything-wp/rules/X.md → Refer to INLINE APPENDIX: X.md
• skills/<name>/SKILL.md → Refer to INLINE APPENDIX (already inlined)
• @skill-name → inline skill reference
• Skill auto-detection tables → "All relevant content has been inlined"

Agent Collaboration

planner → spec/<feature>/     (writes plan files)
                ↓
task-executor → src/           (reads spec, writes code)
                ↓
code-reviewer → review report  (reads src/ diff)

Agents coordinate through the filesystem. No plugin code orchestrates them — each agent only knows its own input and output.

DSH API Reference

Verified API calls (DSH v0.1.0-rc.6):

API	Purpose
ctx.inject(['commands'], cb)	Register slash commands (one per callback to avoid routing conflicts)
ctx.inject(['systemPrompt'], cb)	Register system prompt sections
ctx.systemPrompt.section({name, order, text})	Add a system prompt section
ctx.commands.register({name, description, input, handler})	Register a slash command
handler: ({agent, rawInput}) => ({kind, text})	Command handler signature (synchronous)
agent.steer(createUserMessage({...}))	Inject a user message to trigger agent response
createUserMessage({content, source})	From @deepseek-ai/dsh-llm

Troubleshooting

See the DSH Plugin Dev Troubleshoot SKILL for a comprehensive 17-section troubleshooting guide covering:

• Environment setup (Node.js, pnpm, native modules)
• dsh.bundle and cordis.patch.yml schema
• Cordis inject mechanism
• Slash command routing conflicts
• Inline injection and mode gating
• End-to-end verification methodology

License

GPL-2.0-or-later (inherited from everything-wp)

Credits

Adapted from everything-wp by Oberon Lai. Built for the DeepSeek Harness community.