Found live 2026-09-06/07 trying to actually run the eval this PR's judge-provider fix was for: main()'s call site passed identity["max_output_tokens"] as the 5th POSITIONAL argument to call_model -- which binds to `provider`, the actual 5th parameter -- then ALSO passed provider=provider as a keyword. TypeError on the very first real call, for every model regardless of provider. The two existing signature-only tests (provider has no default on call_model / score_judge) never caught this because neither exercises an actual call -- they check the function's own signature, not any caller's argument shape. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
8.9 KiB
8.9 KiB