[GPT-6.1 Sol][Codex] on behalf of shivam.
Handing the issue-watch investigation for https://github.com/pingdotgg/t3code/issues/28714 over to this PR. The macOS baseline test was already running when you took over; it finished, and no further tests or implementation were started.
Reproduced on physical shiv, arm64, macOS 27.0.1 build 26A434, using the released 0.1.7-nightly.20261101.4035 package at 80120b1aa34a3125c05da393c8616c2076f5888d. The app had private HOME/state, no cache, zsh, launcher PATH /usr/bin:/bin:/usr/sbin:/sbin, and a shell profile that added a private CLI directory after 1.2 seconds. A version-only Claude fixture was missing on first launch and detected as 2.1.132 after quitting and relaunching with the resulting cache.
First launch: the provider probe at 15:54:24.957Z preceded shell completion at 15:54:25.618Z and publish at 15:54:25.662Z. The missing status persisted about 45 seconds after discovery. On relaunch, cached publish at 15:55:19.419Z preceded the successful provider check at 15:55:19.745Z. This supports the cache-miss preparation race described here.
First launch:
first-launch.mp4
Relaunch of the same affected build with its cache:
cache-relaunch.mp4
The test launched the package executable through Python subprocess.Popen with an explicit environment and --use-mock-keychain. HOME and CFFIXED_USER_HOME pointed to a private test home; T3CODE_HOME, ZDOTDIR, TMPDIR, and --user-data-dir also pointed inside the task folder. SHELL=/bin/zsh; the private .zprofile slept 1.2 seconds and then prepended the fixture directory to PATH. The fixture returned 2.1.132 (Claude Code) for --version and exited 64 for other commands. The same environment and state were used for both launches; no provider refresh button was clicked. Recordings begin after process startup and show navigation into Settings > Providers.
This is baseline evidence, not fixed-head verification. Your head 8dc3e6ff446dc1f1a714288091249ee86c91735b was not tested on macOS by this investigation. No real authentication or provider turn was tested; the fixture's account warning on relaunch is expected. Test apps/processes and private state were cleaned up. Your thread owns further testing and follow-up.