gemini-cli - 💡(How to fix) Fix Agent Failure: Persistent Tool Execution Inaccuracy and Data Extraction Reliability

Bug report 053126.pdf

Description: The Gemini agent exhibits a systemic degradation in tool-calling reliability and data extraction. Simple, routine tasks consistently fail on the first attempt, requiring multiple corrective prompts to achieve accurate output. This lack of reliability significantly impacts workflow efficiency. Observed Issues: Tool Execution Inaccuracy (Calendar Integration): Behavior: The agent fails to populate key parameters (such as location) in tool-call JSON payloads, even when those details are explicitly provided in the prompt. Failure Pattern: The agent required three distinct prompt cycles to correctly include an address in a calendar event location field. Impact: Prevents effective scheduling and automated task management. Data Extraction Failure: Behavior: The agent consistently fails to extract simple, static information (such as phone numbers) from provided context or external sources like Google Maps. Failure Pattern: Multiple attempts (3-4 cycles) to retrieve a basic contact number failed, despite the data being clearly available in the source context. Impact: Compromises the agent's utility as a research and data retrieval tool. Expected Behavior: The agent should accurately map provided parameters to the correct tool schema on the first invocation. The agent should reliably extract and format data from provided sources without requiring repetitive prompting. When systemic reliability issues occur, the agent should support "Verify Before Finalize" protocols to allow user intervention before execution. System Context: Environment: Gemini CLI. Context: Recurring issue across multiple tool types (Calendar, Search/Extraction). Recommendation: Prioritize a review of the reasoning/tool-calling layer to address the decoupling of user intent from schema-compliant token generation.