v0.3.4

Changes in 0.3.4

• Smart Intercept of Model 1024x1024 Boilerplate: When a user configures a custom defaultSize (e.g. 1920x1080), LLMs frequently emit size: 1024x1024 due to pre-training habits or multi-turn chat history copying. The runtime now detects whether 1024x1024 was explicitly requested in the prompt (e.g. 1024, 1:1, square, 正方形). If not, it automatically intercepts and applies the user's configured defaultSize.
• E2E Tool Echo: Added size to the tool's result schema and rendered text so that the actual resolution used is clearly reported in the conversation.
• Purged 1024 from Schema Description: Removed all mentions of 1024x1024 from parameter description to avoid prompting the model.