3억 토큰 사용, 주간 한도 50%만 소진 — 방법 공유함

서브 에이전트 써봐라. 방금 Codex한테 재사용 가능한 팀 하나 꾸려달라고 시켰다. 리더는 Sol, 코딩/리서치/디버깅은 Luna, 검토는 Terra 이렇게 구성함.

https://learn.chatgpt.com/docs/agent-configuration/subagents?surface=app

혹시 너네도 괜찮은 방법 있으면 좀 공유 좀 해줘라.

근데 지금 이 방식은 Codex가 돌아가는 시간이 좀 길어지네. 더 좋은 아이디어 있으면 공유 부탁한다.

필요한 사람 쓰라고 내가 쓴 프롬프트 남겨둔다. 그냥 Codex한테 보내면 Sol이 알아서 도와줄 거임:

Set up a reusable local Codex multi-agent team for me.

Use the current Sol model only as the main lead/orchestrator for planning, architecture, hard debugging, delegation, and final decisions.

Create these agents under ~/.codex/agents/:
- coder: gpt-5.6-luna — implements plans, edits code, runs tests
- researcher: gpt-5.6-luna — researches the repo and official docs, returns concise findings
- browser_debugger: gpt-5.6-luna — uses Chrome DevTools MCP to reproduce UI issues and collect screenshots, console, and network evidence; must not edit application code
- reviewer: gpt-5.6-terra — reviews actual diffs for correctness, regressions, security, performance, edge cases, and maintainability

For browser_debugger use:
[mcp_servers.chrome_devtools]
url = "http://localhost:3000/mcp"
startup_timeout_sec = 20

STRICT MODEL-ROUTING POLICY:
- Sol is exclusively reserved for the main/root agent.
- NEVER spawn, create, delegate to, or configure ANY subagent using Sol.
- This applies to named agents, generic agents, temporary agents, ad-hoc agents, background workers, reviewers, researchers, and dynamically created agents.
- A subagent must NEVER inherit the main/root Sol model.
- Prefer existing named agents first.
- All unspecified or dynamically spawned subagents must use gpt-5.6-luna by default.
- Use gpt-5.6-terra only when stronger reasoning is genuinely needed.
- If a task appears to require a Sol subagent, DO NOT spawn it. Ask me for explicit approval first.
- Sol approval is per-task only and must never be treated as permanent permission.
- Configure Codex's default subagent model to gpt-5.6-luna if supported.

SUBAGENT CONTEXT POLICY:
- Configure `fork_turns = "none"` for subagent execution if supported by my installed Codex version.
- Subagents should not automatically inherit/fork the parent conversation history.
- Pass only the task, plan, files, constraints, and context actually needed for that delegation.
- Apply this globally/default where supported, while preserving any existing unrelated configuration.
- If `fork_turns` must be configured at a different scope in my installed Codex version, inspect the supported config and place it correctly instead of inventing syntax.

Preferred routing:
Sol plans -> researcher/browser_debugger investigates when needed -> coder implements -> reviewer reviews -> coder fixes valid findings -> browser_debugger verifies UI fixes when relevant -> Sol makes the final decision.

Inspect my existing ~/.codex/config.toml, ~/.codex/agents/, and AGENTS.md first. Preserve existing settings and agents, do not overwrite unrelated configuration, and adapt to the Codex version actually installed instead of inventing unsupported keys.

Actually create/update the local files, add reusable AGENTS.md instructions for this routing, strict no-Sol-subagent policy, and fork_turns policy. Validate the TOML/config, verify the agents are recognized if possible, and report exactly what you changed.