Saw a lot of discussion around Codex cache hits and had the same confusion, so I built a small tool for the info and analysis

Codex Token Watch v0.1.0: github.com/premk134/codex-token-watch Codex already records useful token and prompt-cache data locally, but that information isn’t easy to inspect per turn in the Desktop app. So I built Codex Token Watch , a lightweight terminal tool that turns the existing local data into readable reports. It shows: Input, cached, output, and reasoning tokens Model-call count and turn duration Model and reasoning-effort changes API-equivalent cost estimates Complete cache misses and the gap since the preceding model call Cache misses across one task or all recently active tasks You can inspect an entire task, only the latest few turns, or cache misses within a rolling period: codex-token-watch THREAD --last 5 codex-token-watch --zero-cache --since 6h It runs locally, reads Codex files in read-only mode Feedback and ideas are welcome.