What is Sol Advisor?
A Codex-native workflow that splits delivery into roles so one session keeps the architecture and the acceptance decision, while a separate worker does the implementation. Here are the pieces and how they hand off.
One Sol session, two implementation lanes
Implementation is delegated; verification and acceptance stay in the primary Sol session. The diagram adapts to your theme.
Sol — the architect
The primary session. Owns requirements, architecture, the five-part implementation spec (native lane) or the complete task packet (Luna lane), verification, and the final acceptance.
Terra — the implementer
A separately installed custom-agent role pinned to GPT-5.6 Terra / High. Does the bounded implementation within the settled architecture. Used by the native lane by default.
Luna — the task worker
An opt-in, user-visible GPT-5.6 Luna / Max task created through Codex app-task tools. Only runs when you explicitly authorize the Luna lane in your request.
Sol reviewer — the gate
A fresh-context Sol reviews the diff in the native lane; in the Luna lane the primary Sol reviews directly. It returns ship, fix-first, or rethink — and the session cannot report done until it says ship.
Verifiable, not magical
No silent fallback
If the expected role, model, or reasoning effort can't be observed, the lane stops with an actionable error. It never quietly downgrades to a weaker model.
Routing evidence
Spawn metadata and (when needed) a read-only local inspector prove which role and model actually ran. Public and local evidence must agree.
Fresh review
The native reviewer runs with a fresh context, not the orchestrator's own — so acceptance isn't the architect rubber-stamping their own work.