Your Codex architect now demands a second opinion before shipping

What it does

Sol Advisor is a Codex plugin that turns a single GPT-5.6 chat into a governed software delivery pipeline. The primary Sol session owns requirements, architecture, and verification, while delegated Terra or Luna sessions handle implementation. Nothing is accepted until the architect inspects the diff and explicitly authorizes the next step.

The interesting bit

The project treats prompt engineering like org-chart design: it enforces strict role boundaries and mandates a context-independent “fresh Sol review” so the architect cannot blindly rubber-stamp its own earlier plans. It is essentially governance-as-code for agent-generated software.

Key highlights

• Splits work between an architect role (Sol / High) and implementer roles (Terra / High or Luna / Max)
• Native lane uses custom-agent TOML roles; Luna lane relies on explicit user-authorized Codex app tasks
• Primary task monitors threads, reads diffs, and must explicitly authorize pull requests
• Includes shell scripts to verify runtime routing metadata and enforce byte-exact companion agent installation
• Refuses to proceed if role, model, or effort metadata is missing, stale, or inconsistent

Caveats

• Locked to specific OpenAI model tiers (gpt-5.6 Sol, Terra, and Luna) that may not be broadly available
• Native mode requires custom-agent support, jq, and a fresh Codex task after every agent install or update
• Luna lane requires explicit per-request user authorization and stops entirely if any required app task tool is unavailable

Verdict

Worth a look if you are already deep in OpenAI Codex and want bureaucratic rigor around agent-generated code. Skip it if you lack GPT-5.6 access or prefer your AI assistant to simply write the file and move on.