JEO - Integrated Agent Orchestration
Keyword: jeo | Platforms: Claude Code, Codex CLI, Gemini CLI, OpenCode
Workflow: Plan (ralph
- plannotator ) -> Execute (team or bmad ) -> Verify (agent-browser ) -> Cleanup (worktree-cleanup )
JEO is the release-oriented orchestration skill package shipped in this repository. It standardizes one path through planning, implementation, browser verification, and worktree cleanup across the supported AI coding tools.
agentation , annotate , and agentui are intentionally removed in v2.0.0 . UI review is handled with agent-browser snapshots and normal edit loops only.
- Execution Contract
Run the phases in order. Do not skip PLAN. Do not enter EXECUTE without an approved plan.
Unity3D 모드 (unity-mcp 감지)
unity-mcp 서버가 실행 중인 경우 (localhost:8080 연결 확인 시) Unity3D 검증 루프가 활성화됩니다.
STEP 0: Bootstrap
Create the working state directories:
mkdir -p .omc/state .omc/plans .omc/logs
If .omc/state/jeo-state.json does not exist, create it with:
{ "mode": "jeo", "phase": "plan", "task": "<detected task>", "plan_approved": false, "plan_gate_status": "pending", "team_available": null, "retry_count": 0, "last_error": null, "checkpoint": null, "created_at": "<ISO 8601>", "updated_at": "<ISO 8601>" }
User-facing activation message:
JEO activated. Phase: PLAN.
STEP 0.1: Recovery Rules
-
Update checkpoint whenever a phase starts.
-
Update last_error and increment retry_count before failing pre-flight.
-
Resume from .omc/state/jeo-state.json if a prior run already exists.
-
If retry_count >= 3 , stop and ask the user whether to continue.
- PLAN
PLAN is mandatory in every JEO run.
Required output
Write plan.md with:
-
goal
-
implementation steps
-
risks
-
completion criteria
Required plan gate
Run the blocking plannotator loop:
bash scripts/plannotator-plan-loop.sh plan.md /tmp/plannotator_feedback.txt 3
Rules:
-
Never use & .
-
If plannotator is missing, auto-run bash scripts/ensure-plannotator.sh first.
-
Proceed only when the result contains approved=true .
-
If the plan is rejected, update plan.md and run the loop again.
-
If the loop exits with 32 , use manual conversation approval and do not execute until the user explicitly approves.
When approved, update the state file:
{ "phase": "execute", "plan_approved": true, "plan_gate_status": "approved" }
- EXECUTE
JEO supports two execution paths.
Claude Code
Use team execution:
/omc:team 3:executor "<task>"
JEO expects CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 .
Codex, Gemini CLI, OpenCode
Use BMAD as the fallback orchestration path:
/workflow-init /workflow-status
Execution rule:
-
keep the implementation aligned to the approved plan
-
update phase to execute
-
return to PLAN only if the plan is invalidated by new information
- VERIFY
Use browser verification when the task includes UI or browser behavior.
Primary command:
agent-browser snapshot http://localhost:3000
Useful follow-ups:
agent-browser screenshot http://localhost:3000 -o verify.png playwriter --help
Verification rule:
-
confirm the changed behavior, not only page load
-
capture evidence before cleanup when the task is browser-facing
-
keep the state file in phase = "verify" until verification is complete
STEP 3: VERIFY — Unity3D 검증 루프
Unity3D 모드가 활성화된 경우 브라우저 검증 대신 아래 루프를 실행합니다.
① unity-mcp: run_tests → pass/fail 집계 ② unity-mcp: read_console → Error/Exception 패턴 탐지 ③ unity-mcp: editor_state → 씬 로드 상태 확인 ④ unity-mcp: find_gameobjects → 필수 오브젝트 확인
검증 결과:
-
모두 통과 → CLEANUP
-
실패 → Fix 루프 (code-refactoring 또는 unity-mcp: validate_script → 재검증, 최대 3회)
-
3회 초과 → 사용자 확인 요청
jeo-state.json 업데이트:
{ "unity_verify": { "tests_passed": true, "console_errors": 0, "retry_count": 0 } }
- CLEANUP
After verification:
bash scripts/worktree-cleanup.sh || git worktree prune
Then set:
{ "phase": "done", "checkpoint": "cleanup" }
Cleanup rule:
-
warn before cleanup if there are uncommitted changes
-
clean extra worktrees only
-
never delete unrelated user work
- Unity3D 워크플로우 예제
Workflow 1: 게임 기획 → 씬 프로토타이핑
jeo "씬 프로토타이핑: <게임명>" [PLAN] bmad-gds-brainstorm-game → bmad-gds-gdd [EXECUTE] unity-mcp: manage_scene → manage_gameobject → manage_probuilder [VERIFY] unity-mcp: run_tests → read_console → editor_state → Fix 루프 (max 3) [CLEANUP]
Workflow 2: 스프린트 스토리 → C# 구현
jeo "스토리 구현: <스토리명>" [PLAN] bmad-gds-sprint-planning → bmad-gds-create-story [EXECUTE] bmad-gds-dev-story → unity-mcp: create_script → validate_script → script_apply_edits [VERIFY] unity-mcp: run_tests → read_console → Fix 루프 → bmad-gds-code-review [CLEANUP]
Workflow 3: 에셋 파이프라인 자동화
jeo "에셋 파이프라인: <에셋 종류>" [PLAN] file-organization [EXECUTE] unity-mcp: manage_asset → manage_texture → manage_material → manage_prefabs → batch_execute [VERIFY] unity-mcp: read_console → run_tests → performance-optimization → Fix 루프 [CLEANUP]
Workflow 4: Unity UI/비주얼 개발
design-system (Unity3D Design Guide 탐색) → ui-component-patterns → unity-mcp: manage_ui → manage_animation → manage_vfx → [VERIFY] unity-mcp: run_tests → read_console
Workflow 5: 성능 최적화 & 디버깅
jeo "성능 최적화: <증상>" [PLAN] log-analysis (unity-mcp: read_console) → find_gameobjects → codebase-search [EXECUTE] performance-optimization → unity-mcp: manage_components → batch_execute [VERIFY] unity-mcp: run_tests → read_console → bmad-gds-performance-test → Fix 루프 [CLEANUP]
- Quick Start
Install dependencies and helpers
bash scripts/install.sh --all
Check current environment
bash scripts/check-status.sh
Configure each platform
bash scripts/setup-claude.sh bash scripts/setup-codex.sh bash scripts/setup-gemini.sh bash scripts/setup-opencode.sh
- Installed Components
Tool Purpose
omc
Claude Code team orchestration
omx
OpenCode orchestration
ohmg
Gemini multi-agent support
bmad
Fallback workflow orchestration
ralph
specification-first planning loop
plannotator
blocking plan review gate
agent-browser
browser verification
playwriter
optional browser automation helper
- Platform Notes
Claude Code
-
setup-claude.sh enables team mode and plannotator review hooks.
-
JEO does not fall back to single-agent execution in Claude Code.
Codex CLI
- setup-codex.sh writes developer_instructions , /prompts:jeo , and a PLAN_READY notify hook.
Gemini CLI
- setup-gemini.sh installs a plannotator-oriented AfterAgent helper and appends JEO guidance to GEMINI.md .
OpenCode
- setup-opencode.sh registers plugins and slash commands for plan, execute, verify, and cleanup.
- State File
Path:
.omc/state/jeo-state.json
Example:
{ "mode": "jeo", "phase": "verify", "task": "Ship the JEO release package", "plan_approved": true, "plan_gate_status": "approved", "team_available": false, "retry_count": 0, "last_error": null, "checkpoint": "verify", "created_at": "2026-03-09T00:00:00Z", "updated_at": "2026-03-09T00:00:00Z" }
- Troubleshooting
Problem Action
plannotator unavailable Run bash scripts/ensure-plannotator.sh or bash scripts/install.sh --with-plannotator
Claude team execution unavailable Run bash scripts/setup-claude.sh , restart Claude Code, confirm CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Codex prompt not found Re-run bash scripts/setup-codex.sh
Gemini hook missing Re-run bash scripts/setup-gemini.sh
Browser verification unavailable Install agent-browser with bash scripts/install.sh --with-browser
Worktrees remain after completion Run bash scripts/worktree-cleanup.sh and then git worktree prune
- Release Notes
v2.0.0
-
removed agentation integration from the skill contract
-
removed annotate and agentui workflow branches
-
simplified verification around agent-browser
-
aligned setup scripts and docs to the reduced release surface