tree 7aa467c4e5782a82677ff03c1098172c0c873b7c
parent 022ec60f69932cbb358d7561977d69341bdfe612
author Jacek Centkowski <geminica.programs@gmail.com> 1791360357 +0200
committer Jacek Centkowski <geminica.programs@gmail.com> 1791459594 +0200

Offer only available Gemini chat models, newest first

The Gemini provider listed every model returned by ListModels on the
v1 API in API order, so the chat dropdown started with outdated or
unsuitable entries:

* gemini-2.5-pro is still listed, but generateContent fails with
  "This model models/gemini-2.5-pro is no longer available to new
  users".
* Image models such as gemini-2.5-flash-image were offered for a
  text-only code review.
* Preview models and the gemini-*-latest aliases, including the
  replacement Google suggests for gemini-2.5-pro, exist only on v1beta
  and were never offered.

Move all Gemini API calls to v1beta: not only the models listing but
also generateContent, which serves the actual reviews, since the
preview models and -latest aliases fail with HTTP 404 on v1. Request
up to 1000 models per page, as the v1beta catalog is larger than the
default page of 50. Drop the image, tts, transcribe, live,
customtools, robotics, computer-use, omni and nano-banana variants.

ListModels carries no lifecycle information, so probe every remaining
model with countTokens, which fails with HTTP 404 for retired models
like generateContent does, without consuming generation quota. Any
other failure keeps the model listed. The provider plugin caches the
resulting list for a day, so the probes run at most once per day.

Sort the -latest aliases first, then by version, newest first; within
a version pro comes before flash and flash-lite, and stable models
before previews.

Bug: https://github.com/GerritForge/ai-review-agent-provider/issues/8
Change-Id: Ibc3cdc71de7efc4b9f5cee02cacc2dcfef961fe6