3.0.6

Horme 3.0.6

Agent loop overhaul

• Fixed terminal skill flag killing the model-initiated loop after one call. The model now always gets a chance to synthesize results and chain further skill calls for multi-part questions.
• Removed ghost assistant bubbles ([Called skills: ...]) from pure tool-call rounds.

New skill: Deep Research

• Multi-query research skill (deep_research) that runs 1-5 queries in parallel via Wikipedia + DuckDuckGo, fetches top page extracts, and compiles a dossier for the model to synthesize.

Timezone support

• date_calc skill now supports a timezone operation. Accepts IANA timezones or 60+ city names (e.g. "Shanghai", "NYC"). Uses Intl.DateTimeFormat -- no internet required.

History persistence fixes

• Tool result messages (tool_result role) now survive save/reload.
• All skill result bubbles persist across history navigation, not just the last one.
• Source pills (e.g. en.wikipedia.org) aggregate correctly on the final answer bubble in both live and history views.

UI

• Green glowing dot on "Reasoning process" while active, turns red when complete.
• Tighter spacing between reasoning/skill/timeline bubbles.
• Consistent assistant bubble width (min(92%, 620px)).

Default settings

• Agent mode now enabled by default.