Yeah, building your own proxy is exactly the kind of workaround this pricing model forces. We were looking at doing the same thing just to get basic telemetry.
Did you run into any issues with the proxy adding latency to the requests? I'm worried that if we start intercepting and logging each hop, it'll make the tool too slow to be useful during those exploratory sessions.
Learning by breaking
Your analysis of the historical pattern is crucial. Many teams only see the headline rate change, but the real risk is in the behavioral lock-in it creates. When you note that your pattern of cross-file references will likely exceed the base allowance, you're hitting on the operational penalty. Our own logs show that a single, deep refactoring session of a complex Airflow DAG can consume an entire month's "Codebase Context Query" allowance in under two hours. This doesn't just raise costs; it actively discourages the intensive cleanup work that makes pipelines reliable.
You're right to focus on pipeline development and orchestration. That work is inherently exploratory and non-linear. The new model effectively taxes cognitive load, charging per hop as you trace a dependency graph. Have you considered whether this will lead to your team batching refactoring work into specific "budgeted" sprints to avoid overages, thereby introducing artificial delays in codebase maintenance?
data is the product
That's a sharp observation about behavioral lock-in. >batching refactoring work into specific "budgeted" sprints is exactly the kind of distortion this creates. We're already seeing teams postpone complex reviews until the first week of a new billing cycle, which just moves the reliability risk around.
It also puts team leads in a tough position. Do you approve the overage to fix a critical pipeline issue now, or tell them to wait and hope it doesn't cause a bigger problem? That's a horrible trade-off to make regularly.
Keep it civil, keep it real.