v1.6.1 - The luna line at the start is Codex's own

Why is there a luna line at the start?

When a Codex conversation starts, Codex sends a couple of requests of its own to gpt-5.6-luna, the cheap and fast model, for small chores such as naming the conversation. The live monitor lists them like any other request, as gpt-5.6-luna → gpt-5.6-luna, and several people read that line as a reroute.

It isn't one: Codex asked for luna, and luna answered. Your own messages are still answered by the model you picked. Only a line that pairs your model with another model (for example gpt-6-astra → gpt-5.6-luna) is a reroute.

• The live monitor guide now says so, in Korean and English, in both the web window and the fallback window.
• Both READMEs explain it in the live monitor section, and the READMEs have been rewritten to be shorter and friendlier, with an FAQ (account safety, cost, what is kept, network use).

Existing installs of the exe update themselves on the next start. 178 tests.

v1.6.0 - A softer window, friendlier words

A new window

The window has been rebuilt as one soft sheet. It is now drawn by a web view (pywebview), while every check still runs in Python.

• One big verdict card with the cat detective. It tells you in a sentence which model you called and which one answered; plan and usage figures stay in Details and in the copied report.
• Pill-shaped controls, an airy list of responses and a single teal accent.
• Help, the live monitor guide and About are laid out as cards, in Korean and English.
• The fonts are built in (Fredoka, NanumSquareRound for Korean, JetBrains Mono), so the window loads nothing from the network. Their licenses are in docs/FONT-LICENSES.txt.
• The window uses the Microsoft Edge WebView2 runtime, which comes with Windows 11 and current Windows 10. Without it, the previous window opens instead; --tk opens that one on purpose.

Fixes

• "Only server errors" with openai/gpt-6-astra: a router-style model name is now checked as gpt-6-astra, and a note says so. The server refuses the prefixed name, and the report's row for a refused model now says UNSUPPORTED instead of ERROR.
• Live monitor: an answer that is still streaming shows "waiting" instead of "the server sent errors".
• The live monitor's consent dialog is no longer skipped by the check dialog's "Don't ask again".

New models

• GPT-6 Sol and Luna (released on 2026-09-22) are checked like any other model. Tested with Codex CLI 0.155.1: gpt-6-sol was served as gpt-6-sol, in a check and in the live monitor.
• The effort list follows the chosen model (up to max for the GPT-6 models). ultra is never offered, because it spawns sub-agents and costs far more than a probe should.
• Long names such as gpt-daybreak-blue-latest keep the report's columns aligned.

Existing installs of the exe update themselves on the next start. 177 tests.

v1.5.0 - New pastel look with a mascot

A new look

The window has been redrawn from scratch in a pastel "retro window" style:

• Every panel is a little window of its own: thick rounded outline, a coloured title bar with three dots, a soft drop shadow and a small sticker (flower, leaf, sparkle, heart).
• Sticker-style buttons and tabs with an offset shadow that press down when you click them.
• A mascot: a lavender cat detective (made with Codex's image generation) greets you in the header and reacts in the result banner: happy when the model you chose answered, shocked when the request was routed to another model, sleepy while waiting. The exe and window icon use the mascot too.
• The Windows 11 title bar is tinted to match.
• The first-visit guide on the Live monitor tab now has the mascot and one idea per line.

Nothing changed in what the app measures or reports. Existing installs of the exe update themselves on the next start.

Also in this release

• Assets are embedded in the program (no image files next to the exe). tools/make_assets.py rebuilds them from the source pictures.
• 133 tests.