토큰 절약과 결과물 향상을 위한 Sol->Terra->Luna 복사 붙여넣기 설정 공유
My copy paste Sol->Terra->Luna setup to save tokens and get good outputs
핵심 요약
Codex에서 토큰 효율을 높이고 작업 품질을 개선하기 위한 3단계 에이전트 워크플로우 설정법을 공유합니다.
- 에이전트 워크플로우 — Sol(기획), Terra(작업 지시), Luna(실행)로 구성된 효율적인 구조를 구현함
- 토큰 최적화 — 복잡한 작업을 세분화하여 토큰 사용량을 줄이고 결과물의 정확도를 높임
- 설정 공유 — config.toml 및 각 에이전트별 설정 파일을 통해 즉시 적용 가능한 템플릿을 제공함
- 사용자 피드백 — 에이전트 과잉 설계 문제와 역할 분담에 대한 커뮤니티의 활발한 논의가 이루어짐
Hey everyone! I am not an expert by any means or a coder, but I love Codex as it has enabled me to turn a lot of my ideas into reality.
Since the Luna cost drop I have seen lots of people discussing strategically leveraging Luna as a worker for hyper defined tasks. I spent a bit trying to understand how this could work and went through 4+ iterations of an agent flow that was constantly over-engineering, requiring extreme baby sitting, or generally making the experience harder than just feeding codex a prompt.
I finally landed on this! Seems to work very well so far, I am using it on 3 apps and getting great results and finding token usage lower. Wanted to share hoping it might be helpful. The first part is instructions on changing config so that Luna workers can be implemented and the second part is a prompt that I give codex in a dedicated folder. I’ll then have that project install it in dedicated app folders I’m working on. Simplicity and efficiency was my guiding principles.
Hope it’s helpful and very open to feedback or improvements!
- CONFIG SETTINGS
From the repository root, create the following files. To make the agents available in every repository, use ~/.codex/config.toml and ~/.codex/agents/ instead.
FILE: .codex/config.toml
[agents]
enabled = true
max_concurrent_threads_per_session = 3
FILE: .codex/agents/sol.toml
name = "sol"
description = "AFOS planner, architect, approver, and final reviewer."
model = "gpt-5.6-sol"
model_reasoning_effort = "xhigh"
sandbox_mode = "workspace-write"
developer_instructions = """
Follow the AFOS skill. Choose one useful, small slice and approve Terra's work order. Review Luna's result only against the approved criteria. Send defects back as focused corrections on the same branch; leave improvements for later orders.
"""
FILE: .codex/agents/terra.toml
name = "terra"
description = "Creates small, concrete AFOS work orders."
model = "gpt-5.6-terra"
model_reasoning_effort = "high"
sandbox_mode = "read-only"
developer_instructions = """
Follow the AFOS skill. Convert Sol's approved slice into one independently reviewable work order with exact file operations, observable acceptance criteria, checks, safety boundaries, and non-goals. Use at most five criteria and three checks unless Sol approves more.
"""
