Skip to content

docs(adr): ADR-0033 — AI-assisted metadata authoring#1528

Merged
xuyushun441-sys merged 2 commits into
mainfrom
docs/adr-0033-ai-authoring-flow
Jun 2, 2026
Merged

docs(adr): ADR-0033 — AI-assisted metadata authoring#1528
xuyushun441-sys merged 2 commits into
mainfrom
docs/adr-0033-ai-authoring-flow

Conversation

@xuyushun441-sys

Copy link
Copy Markdown
Contributor

Summary

Proposes the authoring loop for an AI-authored platform — how an AI generates/modifies any metadata type and how a human safely confirms it before it goes live. Status: Proposed (for discussion).

Decisions:

  1. One brain, many surfaces — AI authoring lives in the agent + global chat; designers contribute context + an in-context "Ask AI", never bespoke per-type generators.
  2. AI never publishes — every agent write lands in the ADR-0027 draft; the change is confirmed in the designer's review/diff; the human's Publish is the only path to live. The draft is the gate (retires the unenforced requiresConfirmation).
  3. Type-agnostic apply — one update_metadata(type,name,patch,mode:'draft') (+ create/describe) validated against the per-type Zod schema (ADR-0005), instead of N per-type write tools.
  4. Plan-first for vague intent — for "build me a project-management system," the agent emits a structured blueprint the human confirms before batch-drafting. No mass-generating unreviewed artifacts.
  5. Type-agnostic review in the shared host — generic draft↔published structural diff for all types; rich per-type diffs as enhancements; flow (graph) last.
  6. Open-core boundary — agent framework + generic tools + BYO-model (incl. local/offline) ship OSS; managed AI + governance (approval gate, env promotion, audit, publish RBAC) are EE.

Builds on ADR-0010 (NL→flow) / ADR-0011 (actions-as-tools) / ADR-0027 (authoring lifecycle) / ADR-0032 (validate-by-default) / ADR-0005 / ADR-0019.

Why now

The pieces exist but aren't a safe loop: the metadata tools publish immediately (no draft, requiresConfirmation unenforced); write tools cover only objects/fields; and nothing handles low-specificity "design me a system" input. This ADR specifies the target end-state (pre-launch, no deprecation path).

Note

Doc-only. The objectui side already shipped the client foundations referenced here (per-designer designers, review/diff, designer→chat context + "Ask AI"). The framework-side changes (draft-routed tools, generic update_metadata, blueprint) are the Phase A/C work this ADR motivates.

🤖 Generated with Claude Code

Proposes the authoring loop for an AI-authored platform: one agent/chat
brain across all metadata types (not per-designer generators); every AI
write lands in the ADR-0027 draft (never publishes) and is confirmed via
the designer's review/diff; a type-agnostic update_metadata tool validated
per-type (ADR-0005); plan-first blueprint for vague intent; and the
open-core boundary (OSS BYO-model authoring vs EE managed AI + governance).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@vercel

vercel Bot commented Jun 2, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
spec Ready Ready Preview, Comment Jun 2, 2026 7:56am

Request Review

@github-actions github-actions Bot added documentation Improvements or additions to documentation size/m labels Jun 2, 2026
@xuyushun441-sys xuyushun441-sys merged commit 7469ff2 into main Jun 2, 2026
10 of 11 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation size/m

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants