Codex Luna Agent Config

Codex Luna Agent Config

This skill configures an opt-in local role after a real Luna dispatch failure. It is not part of ordinary subagent routing and does not authorize delegation.

Required Trigger

Proceed only when both conditions hold:

1. A direct active-spawn invocation rejected gpt-5.6-luna; report the tool, attempted model/effort, and observed rejection before changing configuration.
2. The user explicitly approves installing or configuring the luna_max role.

If either condition is missing, stop. State that the role can be configured after approval; do not read or edit global Codex configuration just to prepare.

Configure The Isolated Role

1. Resolve the active user-level Codex configuration root. It is normally ~/.codex; use a different root only when the runtime or an explicit test fixture identifies one.

2. Discover role names before writing: inspect every [agents.*] declaration in config.toml, resolve each referenced config_file, and inspect every existing agents/*.toml file. A declared role normally uses its table key; a standalone file uses its required name. Preserve unrelated settings, roles, model defaults, and comments.

3. If exactly one source resolves to luna_max and it already pins Luna/max, report it as configured and do not rewrite it. This includes a compatible existing [agents.luna_max] declaration or a compatible standalone file.

4. Stop and show a conflict if multiple sources resolve to luna_max, the one resolved role has incompatible behavior, or the intended agents/luna-max.toml path already contains an unrecognized role. A single compatible role at another path is already configured, not a conflict. Do not replace another role or overwrite an ambiguous file without explicit user direction.

5. Otherwise create one standalone agents/luna-max.toml. Do not add an [agents.luna_max] declaration for a new role; this keeps config.toml unchanged and avoids two definitions for the same role:

   name = "luna_max"
   description = "Use for clear, bounded subagent tasks that should run with GPT-5.6 Luna at max reasoning effort."
   developer_instructions = "Use this role only for a clear, self-contained task. Stay within the requested scope, run the requested verification when possible, and report unresolved decisions instead of expanding the task."
   model = "gpt-5.6-luna"
   model_reasoning_effort = "max"
   

Do not set agents.default_subagent_model or agents.default_subagent_reasoning_effort. This role is selected per child and must not convert all subagents to Luna/max.

Verify And Use

1. Use a read-only Codex configuration load check, such as codex features list, and report its result.

2. Tell the user to start a new Codex task or restart the client so the active spawn schema can expose agent_type: "luna_max".

3. For a self-contained worker, select the role and omit raw model fields:

   // V1
   spawn_agent({ agent_type: "luna_max", fork_context: false, message: handoff })
   
   // V2
   spawn_agent({ task_name: name, agent_type: "luna_max", fork_turns: "none", message: handoff })
   

The role avoids the raw model parameter path that previously rejected Luna. It cannot grant model access: if role-based dispatch still fails, report the server/runtime availability error and do not claim that configuration bypassed it.

Never

• Do not load this skill from codex-subagent-strategy before the user says yes.
• Do not silently replace a rejected Luna worker with another model.
• Do not rewrite a conflicting luna_max role.
• Do not test-spawn a child unless the user separately asks to delegate one.