See the work behind the numbers.

Follow the change.

Compare periods and open the activity behind a busy day. A rise in spend is a starting point for investigation, not evidence that the work was better or worse.

Know what the price means.

Cost carries its source. Subscription usage can be valued using public model prices; that equivalent value is not a subscription invoice. A model waiting for a dated price stays pending instead of appearing free.

Watch the current block.

See what has been spent in the active usage block and the projection based on its elapsed share. The projection describes a possible continuation of the current rate, not measured future usage.

Take the evidence further.

The usage and block reports are available through MCP. Ask your agent to explain the pattern using the available day and vendor data, and to name any dimensions the report cannot supply.

Input, cache and output.

Inspect token totals and their split into uncached input, cache read, cache write and output where recorded, alongside turns, replies and words written. For a Claude session, cache read is usually most of what the model read. These describe the interaction; none is a measure of the quality of the work.

Tools and failures.

See tool-call counts and failed calls where available. Use failures to decide which session deserves a closer look, without assuming every error means an unsuccessful outcome.

Timing and steering.

See the recorded start and end of the session, and steering activity when the vendor provides it. Missing fields remain absent, so an unavailable measurement does not turn into a zero.

A clear metadata boundary.

The current Sessions page does not display a session title, repository, model mix or outcome, and has no repository filter or transcript reader. Local transcript coaching is a separate path through your AI client.

Understand how time is counted.

Attended time combines overlapping segments within each vendor. The timesheet adds those vendor totals, so time spent with Claude Code and Codex together can be counted in both. The timeline withholds combined attended time on mixed-vendor days. These measurements describe recorded activity, not total working hours or performance.

A useful record of activity.

The timesheet lists days, people, attended hours and session counts. Attendance shows recorded activity across the current visible roster. Neither turns presence into a performance score.

1. Notice a repeated pattern Use the repeated-prompt signal to choose what to investigate. The comparison is within a person’s day; sending the same prompt on a different day does not count as a repeat here.
2. Ask what caused it A legitimate retry can look like a loop. Review the context before deciding whether to change a prompt, a tool or the workflow. The signal is a question, not a verdict.

Questions to ask your connected agent

› How did my spend change across this window?
› What are the cost and token totals for session <session ID>?
› What does the attended-time pattern show?
› Which measurements are still missing or unpriced?

Ask for the evidence, the limits, and one useful next step.

An analysis is only as complete as its inputs. Repository, branch, work-item and target links are not interchangeable with session totals; if a requested link or metric is unavailable, the answer must name that gap rather than infer a result.

Coverage before conclusions.

Check which machines and members contributed data and whether the reporting window is complete. A machine that has not reported is a collection gap, not a person doing no work.

Review work without ranking people.

Attended hours, token counts and prompt repeats are observations. They are not a league table. Personal reviews compare you with your own history, and a small window can be too noisy to support a trend.