[saksham2001]

Refreshes the AI coach model pickers to the current lineups (checked against vendor docs on 2026-10-14) and adds per-model reasoning options.

Models

• OpenAI: gpt-6-luna, gpt-6.1-sol (new default), gpt-6-astra.
• Gemini: gemini-3.8-flash (default), gemini-3.5-flash-lite, gemini-4-argon.
  • gemini-2.0-flash is shut down.
  • New API keys can no longer use the 2.5 models.
• OpenRouter: Claude Sonnet/Opus/Haiku 5.5, gpt-6.1-sol, gpt-6-luna, gemini-4-argon, gemini-3.8-flash, deepseek-v4.1-flash.
• MiniMax: removed M2.5/M2.7. They close to new users Oct 19 and retire Nov 18.
• Pricing table: updated for every new preset. Old entries are kept so stored and Custom models still show a cost.

API changes

• Gemini 3+: the client now echoes thoughtSignature back on model parts. Without it, every tool-result turn fails with HTTP 400.
• Reasoning picker: now shown for OpenAI, OpenRouter and Gemini, listing only the levels the selected model accepts. The set now spans none/minimal/low/medium/high/xhigh/max.
  • A stored level the model doesn't accept falls back to the model default. For example, GPT-6.1 Sol rejects none with a 400.
  • On Gemini the level is sent as thinkingConfig.thinkingLevel.

Existing users keep their saved model, since the stored model is a free string.

Tests: full suite passes (1151), including new tests for thought-signature echo, thinking level, effort clamping and preset pricing.