safe-debug

Diagnose deep learning training failures by providing the error traceback to generate a conservative diagnostic package. The system analyzes symptoms to identify root causes, outputting structured files like DIAGNOSIS.md before any code changes are applied.

Yes, you can resolve CUDA memory errors without automatic code modification. The debugging workflow requires no code changes by default, enforcing a patch workflow with explicit approvals and savepoints to protect experimental integrity.

The safest way to address shape mismatches is using a policy-driven patch plan that requires explicit approval before applying minimal fixes. It generates a PATCH_PLAN.md and status.json to ensure structured, reproducible rollbacks.

Troubleshoot checkpoint loading problems by analyzing the error traceback through a conservative diagnostic workflow. The system identifies root causes and guides minimal fixes using policy-based decisions and structured savepoints.

Use a conservative debug workflow when reproducibility is critical and risky changes must be avoided. It escalates risky modifications, requires explicit patch approvals, and maintains savepoints for safe rollbacks during training or inference failures.