wtf is soluna-workflow?

soluna-workflow is a configuration for GPT-5.6 Codex, a coding assistant, that sets up rules for how an AI coding agent should split work between two named roles, called Sol and Luna, to control cost while keeping code quality high. Rather than a program to install and run, this repository is a set of documentation and configuration files that tell Codex how to behave on a project. The core idea is that not every task needs the most expensive, deepest level of AI reasoning. Sol handles planning and reviewing work, while Luna handles the actual implementation. Both come in different effort levels, from a lighter setting for small mechanical edits up to a maximum setting reserved for cases where a worker gets stuck. The workflow starts most planning and coordination at a medium effort level and only escalates to a higher, more expensive level when there is concrete evidence the problem is genuinely difficult, such as conflicting requirements or a risky data migration. This escalation must be recorded rather than happening automatically just because a task looks big. The README also describes how the workflow manages the information given to the AI agents. A small file called AGENTS.md holds only rules that apply to nearly every task, such as routing policy and validation expectations. Bigger, project-specific details live in separate files under a docs and plans folder, so each worker only reads what it specifically needs rather than the whole repository and its history. Implementation tasks are handed to workers as bounded packets listing the exact files and constraints involved. A plan file acts as a contract with ordered steps, acceptance criteria, and clear status values like not started, in progress, blocked, or complete, and only the coordinating Sol role is allowed to update that shared plan file, so workers cannot create conflicting versions of the project state. A final review step, run at a high reasoning effort and restricted to read only access, checks the combined result of any tasks for integration problems across the pieces different workers touched, before the workflow considers the work finished.