We run an Electron desktop cockpit that watches and steers a fleet of Claude Code agent sessions spread across several subscription accounts at once. One of its jobs is to show how much of each account's quota has been consumed. The type that holds that number is called usedPercentageAtObservation, and its documentation describes the value as a lower bound on consumption. This note covers what that name and that documentation commit to, and what they leave out.

Why is the quota field called usedPercentageAtObservation?

The suffix AtObservation pins the value to a moment. A reading is the used percentage as of when it was observed. It is not a statement about the present.

The name itself does not carry the second property, the lower bound. That property lives in the documentation. Read together, the identifier says when the number was true, and the documentation says how far the number can be trusted. In general, putting the time qualifier into the identifier means any code that touches the value has to spell out that qualifier, whether or not its author reads the documentation.

Why is a quota reading only a lower bound on consumption?

The tool is not the only thing spending the quota. Browser use of claude.ai draws on the same subscription quota. So does any session the tool did not launch. Neither emits a statusline payload, so the tool has no view of that consumption.

Unseen consumption can only push real usage higher than what the tool recorded. It can never push it lower. Real consumption is therefore at least the reading and possibly more. That is why the reading is documented as a lower bound, and not as a measurement of total usage.

Does the wire format have a utilization or usedPercentage field?

No. The wire format has no field named utilization and no field named usedPercentage. This note records that as a fact about the format and makes no claim about how it came about.

It matters for one plain reason. In general, a field called utilization or usedPercentage reads like the account's complete, current usage, and this system's readings are neither complete nor current. With no such field on the wire, no consumer can pick up a value under either of those names.

How does the system turn a reading into a claim about now?

Storing a reading and making a claim about the present are separate steps. A separate function turns a reading into a claim about now, and that function returns one of 4 labelled bases. Every present-tense answer therefore comes with a label saying what kind of basis it rests on. It does not arrive as a bare number.

In general, this split lets the stored value stay a fixed fact about the past. Any interpretation of that fact is a distinct, labelled output, and nobody has to edit the reading to produce it.

What time windows and figures go with this design?

These are the figures recorded with this part of the system:

The rollover count shows how often the shorter window turns over relative to the longer one. In a week, the five-hour window rolls over 33 times. A percentage observed in one five-hour window can describe a different window from the one that is open when someone reads it. In general, that is the gap the AtObservation suffix and the separate claim-about-now step are there to keep visible.

What should teams building quota or usage dashboards take from this?

None of this removes the uncertainty. Browser use of claude.ai and sessions the tool did not launch still consume the same subscription quota without emitting a statusline payload. The type name and the documentation put that uncertainty on record so it does not disappear from the number.