my-name-is-jev (nij)

Every name an agent introduces is checked at the moment it is chosen, inside the agent loop:

• files and directories
• variables and constants
• functions and methods
• classes, types and interfaces
• structured strings

Two kinds of check:

• Shape is decided in code: casing, position, pattern.
• Meaning is asked of Jev, TypeSafe AI's decision model: does this function name read as a command, are these words ordered from most stable to most variable. Jev answers typed yes/no and score questions with calibrated probabilities.

Coding agents now produce a large share of the names in a codebase.
Even the strongest models (at time of writing) still name things poorly:

• a function called like a noun (what does r2Bucket() do?!)
• multi-words-names in an order that doesn't sort well
• a name that reads too many ways They choose names 1 token at a time without knowing a team's conventions.
  A style guide does not reach them. A rule that runs before the tool call lands, and reports back in the agent's own context, does.

Two principles:

• No word lists. Whether a word is a verb, an extension or a domain entity is a question about meaning. Every such question goes to Jev.
• We diagnose, the model decides. A finding names the problem and the convention. It never proposes the replacement name.

The library ships primitives for writing rules and, separately, presets built from them. It runs without the presets.

Install: see docs/install.md.

Status: active development. Interfaces change without notice.