A client-set header decided whether a Copilot request spent premium quota

Start with the field. Copilot's own model calls carried `X-initiator: agent`, and by j0selit0's account those calls did not draw down premium quota ^[5]. He then set the same value on requests he sent himself. The model answered, but the quota did not move and the request never appeared in billing analytics ^[6]. dev.to describes this as one developer's own testing rather than an audit, and characterises the mechanism as a billing system keying off a header the client controls ^[7].

That last phrase is the engineering story. A server handling that request already knows the authenticated seat and the model it is about to invoke. It does not need to be told who initiated the call, because the caller is the least reliable party to ask. Asking the client to declare whether its own request is billable is a generous default.

Whether this reaches your invoice depends on your billing setup and on whether the behaviour holds. Your org has to be paying for premium requests rather than working inside an included allowance, and the endpoint has to still behave the way one traffic capture said it did, which is an observation about someone else's session and not a measurement of yours ^[7]. Both are checkable in an afternoon with a logging proxy in front of the editor, which is also the only place a count originates that did not come from the vendor.

The Cursor complaints in the same roundup are usually read as taste, and one of them is not. jmuguy's list includes the editor changing his model to the latest Grok without prompting ^[3]. That is a provenance problem: a diff produced last Tuesday was produced by whatever model the client had selected at that moment, and if the client can reselect silently, the commit no longer identifies its own generator. Your review record ends up unable to tell you which model to credit or blame.

palmotea, commenting on the money thread, put the incentive plainly: leadership cares about velocity and cost, and would replace software development with worse AI software development if the numbers favoured it ^[9]. Cost, in that sentence, means the number on the dashboard. Two of the four grievances in this batch are about whether that number is a measurement.

dev.to is careful about what these are: each quote was located at its comment permalink and reproduced verbatim with username, platform and date, and framed as one practitioner's experience rather than a verdict ^[10]. The publication's own framing is that this batch came from the people using these tools most ^[2]. Of the specific claims, exactly one is a test rather than an impression, and reproducing it costs a proxy and an afternoon.