Hand Codex reading to Luna and keep Sol for the main session

#What changed

Builder Hanako (@hanakoxbt) shared a Codex cost tip: keep GPT-6.1 Sol on the main session for writing code, hand explorer and researcher reads to Luna, and spawn Astra only before a large plan, when the same error repeats, and before calling a long task done. The post includes a paste-ready prompt to rebuild ~/.codex/agents and config.toml around that tree, with diffs shown before any edits.

#Who is affected

Codex users on GPT-6.1 Sol who burn high-end tokens on repo reading and docs lookup that a cheaper Luna agent could handle.

#What to do now

Review your Codex agent tree: put explorer and researcher on Luna, keep the main writer on Sol, and add an Astra architect that only reviews plans, repeated errors, and finished work. Ask Codex to draft the TOML diffs first and approve before applying.

#What is not confirmed

This is one builder's workflow tip on X, not an OpenAI pricing or product change. Exact Luna versus Sol cost savings depend on your plan and how often subagents run.