Metrics
Cline reports cumulative session totals rather than per-step deltas, so the last
record in the stream that carries usage wins. That is the terminal run_result
record for a completed session.
Token classes
Section titled “Token classes”The raw JSON keys map onto the normalized classes as follows:
| Normalized class | Cline JSON key |
|---|---|
| Uncached input | inputTokens − cacheReadTokens (plus cacheWriteTokens) |
| Cached input | cacheReadTokens |
| Output | outputTokens |
| Reasoning | not reported, recorded as null |
Cline’s inputTokens is cache-inclusive: it already contains the
cacheReadTokens restated alongside it, so the cache reads are subtracted to
recover the uncached input. Cache-creation tokens (cacheWriteTokens) are
billed as input and are folded into the uncached input class.
Cline reports no reasoning-token count, so the reasoning class is null (not
determinable) rather than 0, even for a model that reasons. Those tokens are
still accounted within outputTokens, so a Cline run’s token total stays
meaningful and the run participates in token comparisons. Only the reasoning
breakdown is unavailable; the event stream still
surfaces reasoning content when Cline reports it as a distinct block.
Cline reports no run cost of its own, so the comparable cost is computed from the list price curated on the model’s catalog entry, applied to the recorded token classes. See Cost.