Branch G — Agent-to-Agent Workflow

The point of this pipeline: agents exchange via shared artifacts (URL + KB), not via the user.

① Cowork + User  → ý tưởng + HTML skeleton  → Drive draft + note to KB
② GPT            → normalize skeleton → KB spec.md ; review architecture
③ Claude Code    → deploy static preview to VPS ; update preview.meta.json + review-log (preview_url)
④ Claude Design  → short brief + preview URL → visual polish → design delta to review-log
⑤ Claude Code    → validate data/DOT contract (live PG survey, reuse-first) + mock JSON
⑥ GPT            → compare outputs vs spec → decision note to KB
⑦ User           → open preview URL → approve / change request
⑧ Claude Code    → ONLY after approval → implement Nuxt render shell (Đ28)

Roles (who does what with which artifact)

Agent	Receives	Produces	Channel
Cowork	direction from User	visual skeleton / prototype	Drive + KB note
Claude Code	KB spec, data contract, DOT map, constraints	deployed preview, preview_url, review-log updates, (later) Nuxt shell	VPS static + KB
Claude Design	short brief + preview URL	improved visual prototype / design delta	review-log
GPT	whole spec + preview	architecture/UX decision notes, blockers	KB review-log
User	preview URL	approval / change request	review-log

Key principle (from GPT direction)

• Design agents get a short visual brief + a preview URL — not the long architecture pack.
• Code agents get KB specs, data contracts, DOT maps, constraints — not vague screenshots.

Artifact triad (the shared workspace)

1. Spec knowledge/dev/ui/<project>/<surface>/spec.md — intent, interaction, contract, DOT, constraints, links, approval.
2. Preview …/ui-preview/<project>/<surface>/<version>/ — static HTML/CSS/JS + mock JSON + meta.
3. Review …/review-log.md — reviewer, decision, change requests, approval status, next agent task.