Coding agents are useful precisely because they can apply a large body of explicit knowledge quickly.
They are dangerous when that knowledge is generic Lua advice plus confident autocomplete.
Cod3x is intended to be agent-readable engineering context.
What an agent needs to know
A competent OpenMW-Lua agent must understand:
- script contexts;
- object versus record identity;
- interface and event ownership;
- storage/persistence boundaries;
- engine-backed userdata;
- OpenMW API availability;
- the difference between Lua cost and engine cost;
- project style and error policy;
- how to validate claims against source.
A model that knows Lua but not those constraints can produce code that looks professional while being architecturally wrong.
Repository instructions should be short and hard
content/cod3x/AGENTS.md is the condensed operational contract for work in Cod3x.
It should point into the field manual rather than duplicating every explanation.
The documentation explains why.
The agent file says do this.
Agents should investigate history
When an existing pattern looks unusual, do not immediately “simplify” it.
Search:
git log -- <file>;- commit messages around the code;
- related
FIX:/PERF:/REVERT:commits; - current OpenMW source/annotations;
- tests and benchmarks.
A weird branch may be obsolete. It may also be a scar covering a hole you are about to reopen.
Generated code is held to the same standard
Do not keep code because it was generated cheaply.
Do not reject code because a model wrote it.
Review the architecture, contracts, types, performance, and readability exactly as you would human code.
The whole point of Cod3x is to raise the floor of machine-assisted work, not to create a separate lower standard for it.