Track how many turns it takes to complete a coding task, not just how much it costs.
Example: implementing stripe auth - did it one-shot or take seven rounds of edit/test/fix cycles before it worked?
Useful metrics:
- average turns per task type before completion
- one-shot vs multi-attempt rate per category
- edit/test/fix cycle detection (edit -> bash test -> edit again = retry)
Track how many turns it takes to complete a coding task, not just how much it costs.
Example: implementing stripe auth - did it one-shot or take seven rounds of edit/test/fix cycles before it worked?
Useful metrics: