Codex diagnostics stopped leaking config values and credentials

Codex's codex doctor diagnostic command was echoing raw configuration values, credentials included, into the reports it generates. Anyone who pasted that diagnostic output into a bug thread for help was pasting their secrets too.

The fix, merged into openai/codex on September 21, replaces those raw error strings from config.load failures with typed metadata. When a ConfigLoadError is available, even one wrapped inside an I/O error, the report now shows file, line and column instead. When it isn't, the report falls back to just the I/O error kind or a generic failure message. The config contents themselves never make it into the output either way.

The PR's regression tests cover four scenarios: malformed user config, malformed project config, an unknown model provider, and invalid header config. The tests assert both JSON reporting modes omit the credentials built into those fixtures.

What the PR doesn't say is how long this behavior shipped before someone caught it. It also doesn't say whether any pasted diagnostic reports already leaked real credentials into public issue threads or support channels. If you've run codex doctor and shared the output anywhere before September 21, check it for real config values and rotate whatever it exposed.