open-design/QUICKSTART.md
lakatos 9858f496fd
feat(agents): add GitHub Copilot CLI as a code-agent option (#28)
* feat(agents): add GitHub Copilot CLI as a code-agent option

Wire `copilot -p "..." --allow-all-tools` into AGENT_DEFS so the daemon
can spawn it alongside the other CLIs. `--allow-all-tools` is required
in non-interactive mode (same tradeoff as Claude's `-p` and Codex's
`exec`). Add a matching GitHub-dark icon with the Copilot two-eye mark.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(copilot-stream): treat result.success as completed when exitCode missing

Strict `obj.exitCode === 0 ? 'completed' : 'error'` mis-flags turns
where Copilot emits a `result` event with `success: true` but no numeric
exitCode. Switch to `obj.success === true || obj.exitCode === 0`. The
asymmetry argument that applies to other defensive ideas in this file
(raw fallback, stringifyResult cap) doesn't apply here — Claude's
`result` event uses a string `stop_reason`, not exitCode, so this is a
Copilot-specific schema-variance fix rather than a one-sided hardening.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(agent-adapters): list Copilot CLI auth paths in §5.7

Adds a one-line note: detection assumes Copilot is already
authenticated, via either `copilot login` (subcommand, OAuth device
flow) or the interactive `/login` slash command inside `copilot` with
no args. Surfaces the assumption so anyone hitting an undetected /
unversioned Copilot install can trace it back.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 16:54:47 +08:00

8.7 KiB
Raw Blame History

Quickstart

Run the full product locally.

One-shot (dev mode)

nvm use                # uses Node 22 from .nvmrc
corepack enable
pnpm install
pnpm dev:all           # starts daemon (:7456) + Vite (:5173) together
open http://localhost:5173

On first load, the app detects your installed code-agent CLI (Claude Code / Codex / Gemini / OpenCode / Cursor Agent / Qwen), picks it automatically, and defaults to web-prototype skill + Neutral Modern design system. Type a prompt and hit Send. The agent streams into the left pane; the <artifact> tag is parsed out and the HTML renders live on the right. When it finishes, click Save to disk to persist the artifact under ./.od/artifacts/<timestamp>-<slug>/index.html.

The Design system dropdown ships with 71 built-in systems — 2 hand-authored starters (Neutral Modern, Warm Editorial) and 69 product systems imported from awesome-design-md, grouped by category (AI & LLM, Developer Tools, Productivity, Backend, Design Tools, Fintech, E-Commerce, Media, Automotive). Pick one to skin every prototype in that brand's aesthetic.

The Skill dropdown groups by mode (Prototype / Deck / Template / Design system) and shows the default skill per mode with a · default suffix. Bundled skills:

  • Prototypeweb-prototype (generic), saas-landing, dashboard, pricing-page, docs-page, blog-post, mobile-app.
  • Deck / PPTsimple-deck (single-file horizontal swipe) and magazine-web-ppt (the guizang-ppt bundle from op7418/guizang-ppt-skill — default for deck mode, ships its own assets/template + 4 references). Skills with side files get an automatic "Skill root (absolute)" preamble so the agent can resolve assets/template.html and references/*.md against the real on-disk path instead of its CWD.

Pair a skill with a design system and a single prompt produces a layout-appropriate prototype or deck in the chosen visual language.

Other scripts

pnpm daemon            # just the daemon (no web UI build)
pnpm dev               # just Vite (fails /api calls unless daemon is up)
pnpm build             # production build of the frontend → dist/
pnpm start             # build + daemon serving dist/ (single-process prod mode)
pnpm typecheck         # tsc -b --noEmit

Use Node 2022. The repo pins pnpm via packageManager; Node 24 is not supported because better-sqlite3 may lack matching prebuilt binaries and fall back to native compilation.

For the daemon-only production mode, the daemon serves the built SPA itself at http://localhost:7456, so no proxy involved.

Two execution modes

Mode Picker value How a request flows
Local CLI (default when daemon detects an agent) "Local CLI" Frontend → daemon /api/chatspawn(<agent>, ...) → stdout → SSE → artifact parser → preview
Anthropic API (fallback / no CLI) "Anthropic API · BYOK" Frontend → @anthropic-ai/sdk direct (dangerouslyAllowBrowser) → artifact parser → preview

Both modes feed the same <artifact> parser and the same sandboxed iframe. The only thing that differs is the transport and the system-prompt delivery (local CLIs have no separate system channel, so the composed prompt is folded into the user message).

Prompt composition

For every send, the app builds a system prompt from three layers and sends it to the provider:

BASE_SYSTEM_PROMPT   (output contract: wrap in <artifact>, no code fences)
   + active design system body  (DESIGN.md — palette/type/layout)
   + active skill body          (SKILL.md — workflow and output rules)

Swap the skill or the design system in the top bar and the next send uses the new stack. Bodies are cached in-memory per session so this is a single daemon fetch per pick.

File map

open-design/
├── daemon/                    # Node/Express — spawns local agents + serves APIs
│   ├── cli.js                 # `od` bin entry (also used by npm scripts)
│   ├── server.js              # /api/agents /api/skills /api/design-systems /api/chat /api/upload /api/artifacts/save
│   ├── agents.js              # PATH scanner for claude/codex/gemini/opencode/cursor-agent/qwen/copilot
│   ├── skills.js              # SKILL.md loader (frontmatter parser)
│   ├── design-systems.js      # DESIGN.md loader
│   └── frontmatter.js         # tiny YAML-subset parser (no deps)
├── skills/                    # SKILL.md — drops in from any Claude Code skill repo
│   ├── web-prototype/         # generic single-screen prototype (default for prototype mode)
│   ├── saas-landing/          # marketing page (hero / features / pricing / CTA)
│   ├── dashboard/             # admin / analytics dashboard
│   ├── pricing-page/          # standalone pricing + comparison
│   ├── docs-page/             # 3-column documentation layout
│   ├── blog-post/             # editorial long-form
│   ├── mobile-app/            # phone-frame single screen
│   ├── simple-deck/           # minimal horizontal-swipe deck
│   └── guizang-ppt/           # magazine-web-ppt — bundled deck/PPT default
│       ├── SKILL.md
│       ├── assets/template.html
│       └── references/{themes,layouts,components,checklist}.md
├── design-systems/            # DESIGN.md — 9-section schema (awesome-claude-design)
│   ├── default/               # Neutral Modern (starter)
│   ├── warm-editorial/        # Warm Editorial (starter)
│   ├── README.md              # catalog overview
│   └── …69 product systems    # claude · cohere · linear-app · vercel · stripe · airbnb …
├── scripts/sync-design-systems.mjs   # re-import from upstream getdesign tarball
├── src/                       # Vite + React SPA
│   ├── App.tsx                # orchestrates mode / skill / DS pickers + send
│   ├── providers/
│   │   ├── anthropic.ts       # SDK stream (BYOK path)
│   │   ├── daemon.ts          # fetch-SSE against /api/chat (local-CLI path)
│   │   └── registry.ts        # /api/agents /api/skills /api/design-systems fetchers
│   ├── prompts/system.ts      # composeSystemPrompt(base, skill, DS)
│   ├── artifacts/parser.ts    # streaming <artifact> parser
│   ├── runtime/srcdoc.ts      # sandbox wrapper for iframe srcDoc
│   ├── components/            # ChatPane, PreviewPane, AgentPicker, SkillPicker, DesignSystemPicker, SettingsDialog
│   └── state/config.ts        # localStorage persistence
├── docs/                      # product vision + spec
├── .od/                       # runtime data (gitignored, auto-created)
│   ├── app.sqlite              #   projects / conversations / messages / tabs
│   ├── artifacts/              #   one-off "Save to disk" renders
│   └── projects/<id>/          #   per-project working dir + agent cwd
└── vite.config.ts             # /api proxy to :7456

Troubleshooting

  • "no agents found on PATH" — install one of: claude, codex, gemini, opencode, cursor-agent, qwen, copilot. Or switch to "Anthropic API · BYOK" in the top bar and paste a key in Settings.
  • daemon 500 on /api/chat — check the daemon terminal for the stderr tail; usually the CLI rejected its args. Different CLIs take different argv shapes; see daemon/agents.js buildArgs if you need to tweak.
  • artifact never renders — the model produced text without wrapping in <artifact>. Confirm the system prompt is going through (check daemon log) and consider switching to a more capable model or a stricter skill.

Mapping back to the vision

This Quickstart is the runnable seed of the spec in docs/. The spec describes where this grows (see docs/roadmap.md). Highlights:

  • docs/architecture.md proposes Next.js; we picked Vite for a simpler dev loop. The daemon contract is identical, so migrating is a port, not a rewrite.
  • docs/skills-protocol.md describes the full od: frontmatter (typed inputs, sliders, capability gating). This MVP reads name / description / triggers / od.mode / od.design_system.requires only — extend daemon/skills.js to add the rest.
  • docs/agent-adapters.md foresees richer dispatch (capability detection, streaming tool-calls). Our daemon/agents.js is a minimal dispatcher — enough to prove the wiring.
  • docs/modes.md lists four modes: prototype / deck / template / design-system. We ship skills for the first two; the picker already filters by mode.