anthropics
[Feature Request] Implement Thrashing Detection and Auto-Intervention
Bug Description Implement Thrashing Detection and Auto-Intervention
Priority: CRITICAL
Problem: Claude Code can get stuck in trial-and-error loops, trying variations of the same failing approach without stopping to research or ask for help. Example from real session: Hit "EAS CLI couldn't find credentials" error 10+ times, tried different environment variable combinations, created 5 different expect scripts, never stepped back to research the root cause.
Current Behavior:
Same error appears multiple times AI tries slight variations without hypothesis User watches 794 lines of transcript with minimal progress No self-awareness that approach isn't working Expected Behavior:
After same error appears 3 times, automatically pause Prompt: "I'm stuck in a loop with [error]. Let me research the root cause before trying more approaches." Force a research/analysis phase Ask user for guidance before continuing Impact:
Wastes hours of user time Creates frustration and loss of trust Makes complex debugging tasks nearly unusable Suggested Implementation:
Track error messages across tool calls in a session When same error signature appears 3+ times, trigger intervention Require AI to articulate hypothesis before next attempt Surface this to user: "I notice I'm repeating attempts. Should I step back and research?"
