wtf is gpt-5-6-relay?

GPT-5.6 Relay is a Codex skill that routes a single task through three separate model-specific child threads instead of handling everything in one conversation. Each phase of the work gets handed to whichever model and reasoning effort setting fits it best, while keeping the full thread history visible and passing concrete handoffs between phases rather than hiding the process. The default flow starts with a coordinator model called Sol, which handles planning, architecture decisions, and hard diagnosis at a high reasoning effort. Work then typically moves to a model called Terra for implementation, tests, and bounded debugging, then back to Sol for review, and finally to a third model called Luna for reconnaissance, mechanical checks, and release execution. The exact path is not fixed: small mechanical tasks might use only Luna, while genuinely ambiguous architecture work can stay with Sol for longer. Deployment is always treated as its own separate phase handled by Luna, following a project's own deployment rules. Each child thread in the relay is scoped to one role, one deliverable, and one checkable gate, and a thread that depends on another's output will not start until that output actually exists. If something needs correcting, the fix happens in the same child thread rather than spinning up a new one, and only one thread is allowed to write to a project checkout at any given time. At the end, the skill reports back with the actual thread IDs, which models and effort levels were used, what artifacts were produced, and what checks and deployment evidence back up the result. To use it, someone copies the skill folder into a project that supports Codex project skills and invokes it directly with a task description. It depends on the host application exposing tools for creating and reading threads and sending messages with specific model and effort settings, if those tools are not available, the skill reports the route it would have taken instead of pretending the routing happened. It is released under the MIT license.