Codex Workflow: Using GPT-5.6 Sol as Foreman and Luna Max as Bounded Worker

Key points

• A community workflow splits Codex into two roles: GPT-5.6 Sol on the main thread acts as foreman (decomposition, architecture, review); GPT-5.6 Luna Max runs as a subagent doing bounded implementation, bug fixes, tests, and refactors.
• This is not an OpenAI product announcement. It combines existing capabilities. Official Codex documentation confirms support for parallel subagents and custom agent configurations.
• Agent config files live in ~/.codex/agents/ or a project's .codex/agents/. Each must define name, description, and developer_instructions, and may set model and model_reasoning_effort.
• OpenAI's July 30, 2026 pricing update lowered the GPT-5.6 Luna API price to $0.20 per million input tokens and $1.20 per million output tokens, while Sol pricing stayed unchanged.
• Official Codex coding examples follow the same structure: Sol handles uncertainty, defines the plan; Luna executes the defined changes, writes and runs tests, and reports results.
• Community repo use-luna-subagents packages the pattern as auditable config: show diffs before installation, pin the Luna model, inherit parent sandbox and permissions, forbid silent model swaps, and require parent-agent verification of results.
• AYi's claim of "doubling output on the same subscription" is anecdotal, not a controlled experiment.
• Official docs flag three often-overlooked costs: subagents increase token consumption; sandbox and permissions inherit from the parent; overlapping parallel writes can cause conflicts.
• Recommended guardrails: only delegate rollback-safe, well-bounded, independently testable tasks; Luna should not directly modify shared config or critical interfaces; the main thread must review diffs, run tests, then merge; keep complex design, cross-module migration, and security judgement with Sol.
• The broader signal: AI coding cost is shifting from "one conversation" to "a tier of agents plus a verification pipeline." Model routing and harness design, not raw model choice, are becoming the productivity lever.

Sources and evidence

• https://x.com/AYi_AInotes/status/2083867265179537565
• https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6
• https://developers.openai.com/codex/subagents
• https://github.com/aitransformationdirector/use-luna-subagents