Added#
The truncation suspects reach every rendering. --markdown-out, the MCP's
profile_usage and the web Bill tab all name which workloads pay for answers
cut off at max_tokens, with the same denominator — calls that recorded a
stop reason — and the same silence when one label is the whole log. The
markdown and the MCP also carry the ceiling the finished answers needed.
Truncation, with suspects. The report could say a bill paid for answers
cut off at max_tokens and not which workload was paying. It now names them,
ranked by wasted output, with the rate stated over calls that recorded a
stop reason — never over every call, because a workload logging the field
half the time is not a workload whose other half completed, and both numbers
print so the denominator is visible. A 40% rate is a max_tokens setting
that is wrong; 1% is a long tail, and the two call for opposite responses.
Beside them, the ceiling the finished answers actually needed — 95% fit
within N output tokens — measured on these calls and promised for nothing.
The drill-down reaches the browser: click a workload, see it alone. The
web Bill tab's per-label table becomes clickable — the CLI's --label,
without retyping the command. The banner carries the awkward half out loud:
every share below is a share of that workload's bill, not of the log, which
is otherwise read as "chat is 100% of our spend". Both logs of a comparison
are filtered the same way, as on the CLI, and a drill-down inside a
drill-down is not offered — it would filter an already-filtered report and
quietly produce an empty one. Verified in a real browser: $7.00 → $5.00 on
clicking chat, the sibling workload gone from the table, the whole log
restored on clearing, and zero network requests throughout.