Developer Fixes Unreliable Tool Calls in Local AI Assistant with Retry Logic

Developer AlainDevs published a pull request on August 21, 2026, addressing an intermittent bug in a local-first AI assistant where the model occasionally returned plain text instead of executing a required local-data tool call. The root cause was the use of tool_choice='auto', which allowed the model to bypass mandatory tool calls despite prompt-level instructions. The fix introduces a one-time retry mechanism that switches to tool_choice='required' only when no tool call is detected, and only before any tool has executed, preventing duplicate database writes. Response buffering during streaming ensures users never see an ungrounded answer before the recovery step completes. The solution was validated against 67 automated regression tests and a live local Gemma model, reinforcing that reliable AI agent systems require programmatic safeguards beyond prompt instructions alone.