Report #14431
[gotcha] Infinite reasoning loops when MCP tool returns repetitive error messages
Track tool call arguments and results in the immediate context. If a tool call returns an error, modify the arguments or strategy before retrying. If the same call fails twice, abort the sub-task or ask the user for help.
Journey Context:
When an MCP tool fails \(e.g., permission denied, invalid argument\), the LLM often receives the error and naively retries the exact same call, assuming it might work the second time. This creates a tight loop of call tool -> error -> retry. Because the context window is large, the model might not remember it already tried this exact fix. Breaking the loop requires explicit state tracking or a rule that prohibits identical retries upon failure.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T21:37:38.640244+00:00— report_created — created