LLM/AI usage auditability/reproducability/impact

We should have some kind of auditable oversight for code written with (agentic) AI, and how we are working with it (e.g. what models, what provider, are there options for local/open source models etc.). We could show which code was AI-written, what the AI was asked to do, who reviewed it, and roughly what it cost in tokens/emissions/energy.

There are CO2 emission measurement plugins for Claude Code (for example at https://github.com/gwittebolle/claude-carbon ), which give rough order of magnitude estimations for the CO2 emission per session. Codex doesn't have good equivalents as far as I could find, though https://github.com/thiga-co/tokstat with the --impact flag may be an option (that also covers Claude Code).