Summing the token usage reported in a transcript over-counts it, because the harness repeats one response's usage on every content-block line it writes. A single assistant response is written as several lines, each carrying the same message id and the same usage object, so a naive sum multiplies that response's cost by its block count. Measured on one sub-agent transcript here the naive total was 26.7 million against 15.0 million de-duplicated by message id, a factor of 1.77; measured across ten transcripts the same comparison gave 132,336 against 25,046, a factor of 5.28. The factor varies with how many blocks a response happens to have, so it cannot be corrected after the fact by a constant. Any telemetry derived from these transcripts must state the rule that one usage counts once per message id, and must be tested against a fixture containing a multi-block response with a known total. Without that rule a telemetry file is not merely imprecise, it is confidently wrong by a factor that changes between sessions, which is worse than absent for the comparison across runs that such a file exists to support.

Summing the token usage reported in a transcript over-counts it, because the harness repeats one response's usage on every content-block line it writes. A single assistant response is written as several lines, each carrying the same message id and the same usage object, so a naive sum multiplies that response's cost by its block count. Measured on one sub-agent transcript here the naive total was 26.7 million against 15.0 million de-duplicated by message id, a factor of 1.77; measured across ten transcripts the same comparison gave 132,336 against 25,046, a factor of 5.28. The factor varies with how many blocks a response happens to have, so it cannot be corrected after the fact by a constant. Any telemetry derived from these transcripts must state the rule that one usage counts once per message id, and must be tested against a fixture containing a multi-block response with a known total. Without that rule a telemetry file is not merely imprecise, it is confidently wrong by a factor that changes between sessions, which is worse than absent for the comparison across runs that such a file exists to support.

Grounds

• pursued: we expect one usage per response id to be the whole of the correction, because the harness repeats an identical usage object on every line of one response and varies nothing else; it is shown wrong if a response ever reports different usage across its own lines, or if some lines carry usage with no response id to group them under, which the completeness block counts