v26.95.0 — gemini-consult opt-in advisor (4-CLI graceful)

Added

• gemini-consult — opt-in first-party advisor asset (--with gemini-consult / wizard, dev-tools). Consults Gemini (via Antigravity agy CLI) for natural Korean phrasing + multi-persona second-opinion review. Catalog now 61 assets.
  • Bundled like dev-method skills (templates/skills/gemini-consult/, incl. scripts/gemini-ask.sh wrapper), but opt-in (not default install).
  • 4-CLI graceful: wrapper ships to Claude scope; Codex/OpenCode/Antigravity get a SKILL.md that degrades to a direct agy call — no broken wrapper reference (no-false-ship).
  • Improvements over the user-scope original: project-scope path, generalized agy prereq, macOS timeout fallback.

Internal

• New INTERNAL_BUNDLED_SKILL_IDS superset (dev-method + opt-in) drives bundling/render/gen-compat; DEV_METHOD_SKILL_IDS keeps meaning the 7 has-dev-track methodology skills.

CI: publish ✓ / ci ✓ / install-matrix ✓. npm @uzysjung/agent-harness@26.95.0.

v26.94.0 — model-orchestration role split revision

model-orchestration 정책 개정 (사용자 2026-07-07)

역할 재편 (자산 수 60 불변):

• Fable (orchestrator) 직접: 서비스 방향성 수립·논의, 스펙문서 리뷰(multi-persona-review), 기능 개선, 성능·보안 문제발굴
• Opus @ xhigh+: 기획·스펙·계획 문서 작성·관리, 핵심 구현, V&V
• Sonnet @ high+: 반복 구현, E2E 테스트, 리서치 스윕

구조적 해소 2건:

• 문서 author≠reviewer: 작성이 Fable→Opus 이동 → Opus 작성, orchestrator 가 multi-persona-review 로 검수
• V&V 자기검증 충돌: Opus 핵심구현+검증을 instance 분리로 해소 — 구현 Opus 와 별개 fresh Opus 인스턴스가 검증(앵커링 방지)

유지: effort floor(opus≥xhigh, sonnet≥high)·강제 3경로·quota 핸드오프·위임 스펙·결정적변환 금지·parallel-write 규칙.

검증: CI green(60). 행동 A/B eval 은 백그라운드 에이전트 전달 이슈로 미회수 → 스킬 전문 통독 정합성 확인.

PR #194

v26.93.0 — model-orchestration: Orchestration & Model Policy as 7th dev-method skill

model-orchestration 추가 (dev-method 7번째)

사용자 확정 Orchestration & Model Policy를 스킬화해 모든 개발 트랙 기본설치(internal·official·has-dev-track·workflow). 자산 59→60.

• 역할분담: 오케스트레이터(top-tier)=기획·설계·문서 직접 / opus @ xhigh+=V&V·복잡 / sonnet @ high+=단순 코딩 / 결정적 변환=모델 위임 금지 / 소진 시 compaction-handoff→opus @ max 대행
• 리서치 검증(Sonnet 5 에이전트 2기): effort floor=공식 가이드 일치, Agent tool 호출별 effort 지정 불가 → 강제 3경로(agents frontmatter effort: / Workflow opts / 세션 상속) 수록, 위임 스펙 4요소, fresh-context+교차모델 V&V, 15× 토큰 가드레일
• 행동 A/B 실증: baseline 대비 4축(기획위임·effort floor·보안표면 라우팅·quota 핸드오프) 전부 정책 준수 유도
• 부수 fix: gen-compatibility dev-method CLI scope 하드코딩 → DEV_METHOD_SKILL_IDS derive (silent drift 실검출·차단)

검증: CI green(60) · Docker 4-CLI 배치 PASS · 회귀 시나리오 PASS. 실 Codex/OpenCode/Antigravity native 로드는 기존 templates 자산 공통 한계로 미검증.

PR #193

v26.92.0 — frontend-design (Anthropic official) has-dev-track default asset

frontend-design 추가 (Anthropic 공식)

frontend-design (Anthropic 공식 · claude-plugins-official · 984.5K installs)을 모든 개발 트랙 기본설치(has-dev-track) 자산으로 추가.

• impeccable의 official 보완재 — frontend-design=distinctive UI 코드 생성, impeccable=일관성 리뷰 (생성↔리뷰 짝)
• dev 트랙 자동설치(--with 불요) → claude plugin install frontend-design@claude-plugins-official
• tier=official(Anthropic 저자), category=frontend

자산 58→59 (official 11→12). 검증: CI green(track-matrix/interactive 기대값 갱신) · Docker(mock) --track tooling 자동설치 → pluginId 확인 · gh API upstream(Anthropic) 정합. 실 claude plugin install(GitHub resolve)은 mock claude 라 미검증(기존 plugin 자산 동일 한계).

PR #192

v26.91.0 — add coreyhaines31/marketingskills (35k★) opt-in asset

marketingskills (coreyhaines31) 추가

coreyhaines31/marketingskills (⭐35,249 · MIT · plugin v2.5.1, 45 스킬: CRO/copywriting/SEO/AI-SEO/ads/growth)를 opt-in business 자산으로 추가.

• 기존 marketing-skills(alirezarezvani 16k★)와 동명이물 → 병존(id 구분: marketingskills ≠ marketing-skills)
• 설치: --with marketingskills / wizard business 카테고리 토글 → claude plugin install marketing-skills@marketingskills
• "SEO 는 일반 개발에도 활용" → dev 트랙에서 번들 토글로 SEO 7종 포함 설치
• SEO 단일 분리는 폐기: 스킬 상호참조(seo-audit↔schema↔ai-seo, product-marketing 우선)로 부분 추출 시 정합 깨짐

자산 57→58. 검증: CI green / Docker(mock) --with marketingskills → pluginId 설치 확인 / gh API upstream 정합. 실 claude plugin install(GitHub resolve)은 mock claude 라 미검증(기존 plugin 자산 11종 동일 한계).

PR #191

v26.90.0 — bump pinned vetted assets to latest stable

Bump the version-pinned vetted assets to their latest stable releases (the pin is kept — "vetted" means an exact, vouched-for version installs):

• bmad-method 6.8.0 → 6.9.0
• vercel 54.11.1 → 54.17.3
• supabase 2.105.0 → 2.108.0
• agent-browser 0.27.2 → 0.31.0

openspec / netlify-cli / get-shit-done-cc were already latest. COMPATIBILITY.md regenerated from SSOT (57 assets).

v26.89.0 — remove the uzys 6-Gate workflow (BREAKING)

BREAKING: the uzys 6-Gate workflow is removed entirely.

A multi-persona audit found the opt-in gating only covered the /uzys:* commands — the hooks, agents, install output, and the OpenCode transform leaked the gate to users who never selected it. Rather than complete the gating across four CLIs, the feature is retired.

Removed: uzys-harness asset · /uzys:spec…ship (+auto) commands · gate-check.sh + agentshield-gate.sh hooks · gate-status.json · withUzysHarness + manifest gating · uzys emission in the Codex/OpenCode/Antigravity transforms · the global opt-in flags --with-codex-skills / --with-codex-prompts / --with-antigravity-global. Asset catalog 58 → 57.

Preserved (not 6-Gate): dev-method skills (6), tauri-desktop, the reviewer/plan-checker agents, the general hooks, gates-taxonomy, --with-codex-trust.

See ADR-023. Supersedes ADR-011/017/018.

v26.88.1 — dev-method skills, Visual & Media, honest 4-CLI curation

agent-harness curates vetted AI-coding skills, plugins, rules, and hooks by your tech stack and installs only what your project actually needs — across Claude Code (first-class), Codex, OpenCode, and Antigravity. Project-scoped by default; no global pollution unless you ask.

npx -y @uzysjung/agent-harness install

Highlights since v26.71

• dev-method skills (v26.87) — 6 first-party workflow skills (multi-persona-review, gap-analysis-e2e, ultracode-service-audit, asis-tobe-decision, compaction-handoff, northstar-roadmap), routed natively to Codex/Antigravity and as command fallbacks to OpenCode.
• Visual & Media stacks (v26.85–86) — code-first presentation / diagram / motion / video creation (Slidev, reveal.js, GSAP, Remotion, …).
• Trust-tier curation — every catalog asset is ≥1000★ + actively maintained + Docker install-verified (49/58 green today), with monthly star-drift monitoring. Honest by design: vetted ≠ a line-by-line security audit; npm/npx assets are version-pinned, plugin/skill assets resolve to upstream HEAD.
• 4-CLI, honestly labeled — Claude Code is first-class; Codex / OpenCode / Antigravity get the skills + rules layer. The compatibility matrix states exactly what reaches each CLI.

This release (v26.88.1)

Documentation honesty pass: refreshed catalog star counts (measured 2026-06-22 via GitHub API), corrected the vetted count to 49/58, and synced the roadmap SSOT.

Full history: CHANGELOG.md · Compatibility matrix: COMPATIBILITY.md

MIT.

v26.88.0 — persona multi-audit: trigger honesty + roadmap sync

5-persona Sonnet panel audited the dev-method skills + roadmap → P0/P1 fixes:

• Review-skill trigger disambiguation (multi-persona-review / ultracode-service-audit / gap-analysis-e2e) so a casual "review this" doesn't fire the heavyweight 87-agent audit.
• Honest 4-CLI labels — file-level (manifest copy) verified vs. real-CLI native load unverified, stated explicitly.
• Non-Claude plugin-reach notice at install time (plugin-kind assets are Claude-only).
• Roadmap SSOT sync + ship-checklist drift gate.

3 candidate findings were rejected via verification (arxiv citation real, Nielsen URL valid, backup already implemented) — no-false-ship: a panel generates candidates, you verify before committing. ci green, 716 tests.

v26.87.0 — dev-method skills: 6 first-party workflow skills across 4 CLIs

Six first-party workflow skills, bundled in-repo and installed as core on dev tracks (opt-out via --without <id>):

• Dev Tools: multi-persona-review, gap-analysis-e2e, ultracode-service-audit
• Workflow: asis-tobe-decision, compaction-handoff, northstar-roadmap

Routed to 4 CLIs: Claude (.claude/skills/) + Codex/Antigravity native skill (.agents/skills/<id>/) + OpenCode command fallback (.opencode/commands/<id>.md).

Honest caveat: real-CLI native recognition is unverified (Docker-isolated tests use mocks; host hooks block real-CLI runs). Catalog 52→58.