Eaon 2026.2.0

Added

• MCP & Skill integration — connect any custom MCP server with a pasted
  token, and install a Skill Library from GitHub or your own local Claude
  Code skills. Invoke any enabled skill directly in chat with
  /skill-name.
• A rebuilt memory system — Eaon can learn from your conversations over
  time (with consent), remembering specific things you've shared instead
  of only static facts, and can learn from a file you pick or from a
  connected plugin's results.
• Live code streaming with syntax highlighting — watch code appear in
  color as the model writes it, and the chat no longer yanks you back to
  the bottom while you're scrolled up reading.
• A per-model CPU/GPU control for downloaded Hugging Face (llama.cpp)
  models — force CPU-only, force max GPU, or leave it on auto.

Fixed

• Chat mode not actually writing code to disk when asked.
• The sidebar re-sorting your chats out of order just from clicking
  between them.
• Several Hugging Face model download/run issues: a hard crash on
  Gemma-family models, misleading "try a different quantization" advice
  when a model's whole architecture isn't supported, and a truncated
  error log that hid the real reason a model failed to load.
• Deleting a downloaded model reporting success without actually freeing
  disk space.
• A confusing "wait for models to load from the Aqua API" message
  appearing even when the real issue had nothing to do with Aqua.
• pip's "externally-managed-environment" error when Eaon's agent tries
  to install a Python package.