Count every pass on the pull request#
An agent-glovebox pull request changed what happens when the program cannot save a command's execution record. On November 15, 2026, its automated reviewer read the proposed change. The checkout lacked code from the parent branch, so the reviewer assessed part of the behavior from the diff. On November 18 the reviewer returned after the author added a changelog entry. This time it read only that entry. Both reviews used Claude Opus 5.5. First review · Changelog review
| What the run covered | Model cost recorded by the workflow |
|---|---|
| Initial review of the proposed change | $0.656338 |
| Follow-up on the changelog | $0.132747 |
| Both recorded passes | $0.789085 |
The workflow displays the amounts rounded to cents and includes the higher-precision amounts in the review records. Counting only the latest run would put this pull request down as a 13-cent review. Counting both gives 79 cents of reported model usage. The public records do not include a provider invoice, verification runtime, or the cost of writing the patch. The second pass also covered less code, so its lower cost says little about what a second full review would have cost.
Reconcile a review made with your own provider key#
With bring your own key (BYOK), the model requests go to your provider account. Start with one pull request and the commits you asked the reviewer to check. List every run for those commits, including retries and follow-ups that produced no findings. A run can make several model requests.
For each run, match the provider request IDs to your usage export where both systems expose them. If you have only totals for a shared key and time window, mark the amount as an estimate. A quiet interval on a dedicated project or key is easier to reconcile. Keep the account's actual currency and apply discounts or credits according to how its invoice reports them.
If you have token counts but no cost field, calculate each category at the rate for the model and service tier used that day. For example, Anthropic prices input, cache writes, cache reads and output separately. Apply each rate to the matching tokens, divide by the rate's unit, then sum the results. A single price times “total tokens” loses the distinction between a cache hit and generated output. Anthropic's model pricing
Add verification on its own line#
A reviewer may also start a machine to build the change or run a test. Record the metered time or credits under verification and identify any plan coverage. On Proofline, BYOK model usage goes on the provider's bill. Verification VM time, when used, is a separate Proofline credit charge. Eligible coverage pays first. A blank verification field on a public review gives you no basis for entering zero. How Proofline bills verification
For your own PR, report the provider amount in currency, the verification usage in its original unit, and the coverage applied. Combine them into one cash figure only when you know what the credits cost your team. Keep the number of passes beside that figure so the next person can tell whether it bought one review, a second opinion, or a recheck after a fix.