Skip to content

v0.7.0 — UI planning and agent capability routing

Latest

Choose a tag to compare

@Hanjo92 Hanjo92 released this 09 Sep 08:18
d8be853

Changes

  • Plan new UI through sequential design/structure decisions, preserve prior answers, and reuse suitable project images even without an asset index.
  • Offer or perform image generation only when the generation executor has read an applicable skill and has a usable execution path. Only the designated coordinator may relay that verified owner's question; a skill-less implementation child cannot ask or forward it.
  • Qualify each delegated session for actual image input, generation, artifact access, and Unity operations. Give text-only workers approved structured handoffs and keep visual analysis/QA with qualified reviewers.
  • Add first-class UI Toolkit routing and reusable UXML/USS workflows, alongside layout snapshots, neutral mockup plans, an agent runbook, and review/assumption gates.
  • Refresh README, platform adapters, examples, and backlog. Keep changelog validation working when Unreleased notes move into a dated release.

주요 변경

  • UI 목적·동작·시안·구조를 순차적으로 확정하고, 시안에 맞는 프로젝트 이미지를 우선 재사용합니다.
  • 이미지 생성 담당자에게 스킬과 실행 경로가 있을 때만 생성 여부를 질문하고 생성합니다. 검증된 담당자의 질문은 지정된 총괄만 전달할 수 있습니다.
  • 하위 에이전트의 비전·도구·파일·Unity 접근을 세션별로 확인해 역할을 나눕니다. 코드 검사·촬영 성공과 시각 검증을 구분합니다.
  • UI Toolkit, layout snapshot, 중립 계획 문서, 실행 runbook과 검수 기준을 이번 버전에 함께 포함합니다.

Validation and limits

  • All 12 repository shell checks, shell syntax, skill validation, four YAML parses, 145 release document links, evidence JSON/observations, probe and tested-source hashes, and git diff --check passed.
  • The changelog check passed for dated and Unreleased notes and rejected missing required history in an isolated fixture.
  • Live Orca validation: real Codex and OpenCode sessions both interpreted the diagnostic image. The initial OpenCode relay contradiction and successful same-session follow-up after the correction are both retained.
  • Actual image generation, Unity import/assignment/runtime, Paseo/Claude, and a vision-disabled runtime were not exercised. The shell and stored-evidence checks do not establish those execution results.
  • 실제 검증은 Codex·OpenCode의 이미지 입력 및 제한된 역할 판단을 다룹니다. 실제 이미지 생성·Unity 실행과 나머지 런타임은 미검증이며, 재검증은 같은 OpenCode 세션에서 수행했습니다.

Included pull requests

#72, #73, #74, #75, #77, #81, #82. Issues #79 and #80 are resolved.

Changelog · Full comparison