MEDIA PREVIEW
Overview & result
This Codex CLI wrapper presents Jev with the current goal, a small set of candidate native reads and completion metadata before the coding model is invoked. Jev picks a bounded read or hands control back; Codex executes the selected call under its own permissions and handles coding and the final answer.
From the original post
Jev selects a native Codex tool call before the coding model is invoked. Codex executes that call through its own tool runtime, approvals and sandbox. The result stays in Codex's conversation. The coding model handles reasoning, edits and the final answer.
How Jev fits in the loop
- Extract at most eight bounded read offers for relative source paths explicitly named in the current user message.
- Jev chooses one offered native call or hands off to the coding model.
- Codex executes at most four selected reads through its own tool runtime before the model continues.
How to reproduce
- Use Node.js 22+, an authenticated Codex CLI and a TypeSafe key stored in a private environment file.
- Build the repository and launch a Codex CLI session through its documented router wrapper.
- Run the published live smoke check with an explicitly named relative source file and compare the trace with the normal Codex path.
Why this build matters
A concrete, narrow example of Jev taking a tool-choice decision before Codex reasoning while leaving execution permissions and code work with Codex.
Limitations
- The current adapter only offers bounded reads of relative paths named in the user message; arbitrary commands, test selection and browser use are not implemented.
- Unsupported requests and provider failures fall back to the ordinary coding-model request.
- Historical benchmarks in the repository refer to an earlier architecture and do not establish savings for the native tool selector.