Your codex exec is wasting ~900 ms per turn. I measured where.

I rebuilt the JavaScript harness of OpenAI's Codex CLI on Bun for a weekend. I expected a runtime improvement, but I discovered an architectural change that could speed up Codex operations. The Codex system is a Rust engine with a thin JavaScript layer on top, including an npm launcher, a TypeScript SDK, and a pnpm/jest/tsup toolchain.

By swapping the JavaScript layer to Bun, adding a client to keep a codex app-server process alive, and measuring with hyperfine and mitata, I found a significant improvement in performance. Specifically, the persistent-process design reduces the time spent per turn from about 900 milliseconds to just 0.31 seconds. This improvement was consistent across both Node.js and Bun runtimes, with Bun providing faster test startup, warm install, and a more efficient JSON-lines parser.

Additionally, the architecture change allowed for better telemetry management, as the process shut down without flushing metrics over HTTPS, thus saving time. The full article provides detailed benchmarks, sequence diagrams, and code demonstrating the performance gains.