2026-05-04 — New Providers Day

Big architecture day. Three new providers landed in clang, replacing the old broken gemini provider.

The trigger was simple — user tried to use Gemini on MacBook and got "prepay credits depleted". Turns out AI Studio keys (AQ.*) use a separate prepay bucket from Cloud billing. The $520 sitting in Cloud billing was useless with the old endpoint.

One curl test proved aiplatform.googleapis.com works with the same key. From there it snowballed into a proper multi-provider architecture:

• gemini_vertex — pay-per-use via Cloud billing
• google_gemini — free via AI Pro subscription (cloudcode-pa)
• google_claude — same free path, Claude models

The worktree agent (new-providers) did the heavy lifting — 4930 lines of new C code across 9 files. It took about 2 hours of thinking time with multiple nudges. The main issues were:

• Off-by-one in thinkingConfig string length (63 vs 64 bytes — classic)
• Wrong model IDs for google_gemini (used internal names instead of matching Python codebase)
• Missing prompt caching (had to be told explicitly)

Also discovered the macOS exe_dir symlink bug — _NSGetExecutablePath doesn't resolve symlinks, so new binaries in ~/.scorpiox/ weren't found when sx lives at /usr/local/bin/ via symlink. Quick realpath() fix.

Satisfying day — went from "Gemini is broken" to three working providers, health check agents, and proper testing infrastructure in one session.