Recipe · OpenAI Codex CLI

Use Sverklo repo memory with Codex CLI.

Start by proving Sverklo can find useful repo evidence. Then let sverklo init wire the Codex CLI MCP server config and verify the setup with doctor --agent codex.

1. Run proof without setup writes

npm exec --yes --package=sverklo@latest -- sverklo prove --no-write --guided --markdown

This should print central files, one selected symbol, callers, and a paste-ready prompt. It does not write project files, MCP config, or agent instructions. First run may cache the local ONNX model and index under ~/.sverklo.

2. Install and preview

npm install -g sverklo
sverklo init --dry-run

The dry run previews setup targets. For Codex CLI, Sverklo looks for ~/.codex/ and previews the MCP server block that will be merged into ~/.codex/config.toml.

3. Apply and verify

sverklo init
sverklo doctor --agent codex

Re-running sverklo init from a different project rewires the per-machine Codex entry to that project. doctor --agent codex gives a focused prompt for checking that Codex can call Sverklo before you ask it to edit.

Starter Codex prompt:

Use sverklo impact on the symbol I am about to change, then tell me the caller/test blast radius before editing.

When to stop

If the proof receipt picks nonsense, do not continue setup. Share the receipt or the failure mode in the public proof thread. That is more valuable than a silent install.