Port from google-gemini/gemini-cli#25062. Adds a concise system prompt
block that tells agents to stop and reconsider their approach after 3
failed attempts at fixing the same issue, instead of continuing to
apply small variations of a failing fix.
The guidance is injected for ALL models when tools are loaded (not just
enforcement-target models), since fix-loops affect every model.
3-step process:
1. Stop and re-read the original task description
2. List current assumptions and identify wrong ones
3. Propose a fundamentally different approach
Includes tests for the constant content and system prompt integration.