v12.2.0 - Still Freeze

Fixed

• Synced model profiles keep their full guidance. An imported tuning profile's guidance text was cut at 600 characters on import, which silently truncated every profile in the shared library mid-sentence and dropped the tail of the instructions. The bound is now generous enough for a real multi-rule hint, and a profile that still exceeds it is logged instead of being trimmed in silence.
• Images keep your chosen chat model. When a dedicated Vision Provider read an attached image, the answer was silently handed to a different model than the one you picked (a local vision model, or a small cloud fallback), so the reply looked like it came from the wrong place and, on paid providers, billed the wrong model. The picture now goes to the vision model only when a raw image is actually sent; once the Vision Provider has described it, your chosen chat model answers. Vision-capability detection now reads what the local model actually reports instead of guessing from its name, a Vision URL that already points at the chat endpoint is no longer mangled into a broken address, an explicitly chosen local vision model is attempted instead of pre-blocked, and the image-related tips and error messages now point at the Vision Provider settings. Profiles can also carry a vision-capability override so a new model is recognised without an app update. A picture and its description are now saved with the conversation, so the very next question ("what did you see?") is answered from what was actually read instead of "I do not remember", and a failed image read no longer vanishes from the history. The saved description shows under the image as a small collapsed eye accordion (like the reasoning view), never as raw text in the bubble.
• The ChatGPT (Codex) subscription is selectable everywhere. A signed-in ChatGPT account never appeared in the model pickers (new-chat landing, in-session switcher, picker search, agent and Codework provider dropdowns) because those lists only recognised API-key credentials; the subscription sign-in now counts as configured and the Codex models show up like any other provider's. Flow already offered it.
• Every surface reads image capability from one place. The Telegram channel kept its own third list of which models can see (with the multimodal gemma3/gemma4 families wrongly marked text-only), and four code paths carried hand-copied checks for what counts as a local vision endpoint. All surfaces (chat, Telegram, WhatsApp, the desktop buddy, the orchestrator) now share the same detection and the same local-endpoint rule, with the pasted URL normalized first, and the buddy no longer requires a vision model id where every other surface accepts the built-in default.
• The ChatGPT (Codex) subscription is on one tested path. The plain-chat and tool-using code for a signed-in ChatGPT account had drifted into two hand-copied copies, which is how the subscription kept breaking one release at a time; both now share a single request builder and stream parser covered by tests, so a fix can no longer land in one and miss the other. The current model list gained GPT-5.5 (an up-to-date id was falling through to the manual entry field), and a short-lived access token is now refreshed automatically before a turn instead of leaving a still "signed in" account failing every message until a manual sign-out and sign-in.