부모가 판단에 초점을 맞추는 경량 코덱 멀티 시약 설정은 실제로 도움이 될 때 경계 구현을 위임하면서.

현재 설정은 Astra-only 보다는 오히려 부모 인식입니다:

• GPT-6 Astra Low - 계획, decomposition, delegation 및 리뷰에서 혜택이 큰 또는 진정한 에이전트 작업을 선호하는 부모
• Sol - 더 작거나 직진된 코딩 작업에 대한 유효한 부모가 될 수 없습니다.
• Spark - 사용할 수 있을 때 작은 deterministic 편집을 위한 선택적인 매우 빠른 마이크로 노동자
• Luna High - 기본 경계 구현 작업자
• Sol Medium - Astra가 부모와 Luna가 충분하지 않을 때 더 열심히 경계 작업자
• Astra Medium/High - Astra Low가 충분하지 않을 때 탁월한 부모 수준의 이유

이동적인 부모는 글로벌 문제를 소유합니다. Delegate는 컨텍스트, 시간 또는 비용을 절약 할 때만 작동을 경계했습니다.

이전 설정 보존

• v1-sol-luna - 원래 Sol + Luna 설정
• v2-sol-spark-luna-terra - 솔 중간 부모, 스파크/루나 노동자, 테라 높은 에스컬레이션
• v3-astra-low-parent - Astra Low parent, Spark/Luna workers and Sol Medium 에스컬레이션

Astra/Sol + Spark/Luna Codex Orchestrator

A lightweight Codex multi-agent setup that keeps the parent focused on judgement while delegating bounded implementation when it actually helps.

The current setup is parent-aware rather than Astra-only:

• GPT-6 Astra Low - preferred parent for larger or genuinely agentic work that benefits from planning, decomposition, delegation and review
• Sol - also a valid parent for smaller or straightforward coding work where delegation overhead may not be worth it
• Spark - optional ultra-fast micro-worker for tiny deterministic edits when available
• Luna High - default bounded implementation worker
• Sol Medium - harder bounded worker when Astra is the parent and Luna is not enough
• Astra Medium/High - exceptional parent-level reasoning when Astra Low is not enough

The active parent owns the global problem. Delegate bounded work only when it saves context, time or cost.

Preserved older setups

• v1-sol-luna - original Sol + Luna setup
• v2-sol-spark-luna-terra - Sol Medium parent, Spark/Luna workers, Terra High escalation
• v3-astra-low-parent - Astra Low parent, Spark/Luna workers and Sol Medium escalation

Why the instructions are leaner now

Newer models need less scaffolding than earlier coding agents. The current template deliberately keeps AGENTS.md focused on durable repository-wide behaviour instead of turning it into a long execution recipe.

The delegation skill also uses progressive disclosure: its root file is a small router, while the detailed work-package format lives in a supporting reference that is only needed when work is actually delegated.

Verification is proportional to the change rather than ritualistic, and persistence is explicit so the active parent continues through implementation, relevant checks and fixes when the requested outcome clearly requires end-to-end completion.

Parent-aware routing

When Astra Low is the parent

Astra Low
    |
    |-- tiny + deterministic + Spark available --> Spark
    |
    |-- normal bounded implementation -----------> Luna High
    |
    |-- harder bounded work ----------------------> Sol Medium
    |
    `-- exceptional parent-level problem --------> Astra Medium/High

When Sol is the parent

Sol
    |
    |-- small / straightforward work ------------> handle directly
    |
    |-- tiny delegated work ----------------------> Spark if available
    |
    |-- normal delegated implementation ----------> Luna High
    |
    `-- harder work ------------------------------> handle directly in Sol