RightNow-AI/openfang v0.2.0

on GitHub

Critical Bug Fixes

Fixed: UTF-8 Panics (#74, #73, #71)

• Added safe truncate_str() helper — never splits multibyte characters
• Fixed 20 crash sites across kernel, memory, API, runtime, channels, and CLI
• Chinese text, emoji, and em-dash no longer crash the daemon

Fixed: Shell_exec Infinite Loop (#69)

• Agent no longer retries denied tool calls endlessly burning API credits
• Approval denial now injects guidance message so the LLM stops retrying
• Works in both streaming and non-streaming paths

Fixed: Stale Groq Model (#60)

• Removed decommissioned deepseek-r1-distill-llama-70b (was returning HTTP 400)

Fixed: Volatile System Prompt Cache Misses (#82)

• Moved canonical context from system prompt to first user message
• System prompt is now stable across turns, enabling provider prompt caching

Fixed: Misleading Empty Response (#61)

• Now tracks whether tools actually executed
• Shows accurate message when model returns empty (no more false "tools executed" claims)

UI Improvements

• TUI chat now shows per-message cost alongside token counts ($0.0012)
• HTTP /api/agents/{id}/message response now includes cost_usd field
• All 16 TUI screens use safe UTF-8 truncation