Problem
In live use, the displayed spend/cost appeared to go down during or after sub-agent/fanout activity. Even if token accounting is approximate, the user-visible total should not decrease unless there is an explicit reset or session switch.
This may be related to overlapping parent/sub-agent usage accounting, prompt-cache adjustments, or replacing provisional estimates with lower final values, but the UI needs to make the source of truth clear.
Acceptance
- Session spend display is monotonic within a session unless a clearly labelled reset/session change occurs.
- Parent and sub-agent usage events are deduplicated without subtracting already-displayed totals in a confusing way.
- If provisional estimates are replaced by final provider usage, the UI communicates that or tracks separate estimated/final fields.
- Cache-hit/cached-token accounting does not make the visible total appear to move backwards.
- Add focused tests for out-of-order usage events, duplicate sub-agent events, and final-usage reconciliation.
Notes
Related to the broader telemetry/audit work in #204, but this issue tracks the concrete user-visible regression: spend should not visibly decrease during a live run.
Problem
In live use, the displayed spend/cost appeared to go down during or after sub-agent/fanout activity. Even if token accounting is approximate, the user-visible total should not decrease unless there is an explicit reset or session switch.
This may be related to overlapping parent/sub-agent usage accounting, prompt-cache adjustments, or replacing provisional estimates with lower final values, but the UI needs to make the source of truth clear.
Acceptance
Notes
Related to the broader telemetry/audit work in #204, but this issue tracks the concrete user-visible regression: spend should not visibly decrease during a live run.