Fixes and improvements

[Chenglong Wang (Chenglong-MS)]

This pull request introduces several improvements and refinements to the Data Formulator agent and context handling, focusing on clearer terminology, enhanced connector and skill management, and dependency updates. The most important changes are summarized below.

Agent and Context Terminology Improvements:

• All references to "primary table(s)" and "other available tables" have been replaced with "primary analysis inputs" and "other analysis inputs" throughout agent summaries, context, and system prompts for greater clarity and consistency. The [AVAILABLE TABLES] section is now [ANALYSIS INPUT TABLES]. [1] [2] [3] [4] [5] [6]

Connector Availability and Summary Enhancements:

• The connector summary block now displays all currently loadable connectors, including those not yet cached, and distinguishes between connected/disconnected sources and catalog availability. Disconnected sources are hidden from the agent, and error handling/logging is improved. [1] [2] [3]

Skill Preloading and System Prompt Improvements:

• The agent now preloads the "data-loading" skill when no analysis input tables are present, exposing its tools and actions immediately. System prompts include all preloaded skills with clear banners. Skill state is properly rehydrated from both loaded and preloaded markers in the trajectory. [1] [2] [3] [4] [5]

Dependency Updates:

• Several frontend dependencies have been updated for security and compatibility, including vite, dompurify, echarts, js-yaml, and new resolutions for postcss, esbuild, and tmp. [1] [2] [3]

Documentation Cleanup:

• The obsolete docs/desktop-portable.md file has been removed.
• The old model evaluation plan in loops/model-evaluation/plan.md has been deleted.