Until now Conn2Flow talked to one AI provider at a time, in that provider's format. The ia-provedores library gives every provider the same input and output: the caller does not need to know anybody's format.
What it covers
- Four server types in the registry: Gemini, Claude, OpenAI and any service that speaks the OpenAI format, including a model running on your own machine.
- Text on all of them, with a system instruction and multi-message conversations; images on Gemini and OpenAI.
- Text results include input and output tokens, which feed the credits.
Key safety
The Gemini key used to travel in the request URL, which tends to end up in access and proxy logs. It now goes in a header for every provider, and error messages go through a filter that strips anything that looks like a key.
Extension points
Before each request, pedido.autorizar lets a module refuse the call; after the answer, pedido.concluido reports the usage and who asked. A third point, interface / pagina, lets a module add a feature to any panel screen. The panel's AI features are built on these three points.
Status
On the 3.0 line. Only Gemini was exercised with a real key; Claude, OpenAI and compatible services were checked by offline tests. Acceptance pending.
Source: Core · REQ-260 to REQ-262 · BATCH-269 to BATCH-271