Context
After #158 (ccusage speed-tier + Claude day-window parity), Codex 7d dollars match ccusage within ~0.2% at both --speed auto and standard.
Codex 30d still lands ~8–9% under ccusage on the same machine (tokens and dollars scale together). Not the old 2× priority bug.
Likely causes
- Older
archived_sessions coverage / filename date gates
- Fork / subagent parent-history replay: Ceiling suppresses replayed parent totals; ccusage may count some of that
Preference
Keep anti-double-count behavior unless we prove ccusage is strictly more correct on real rollouts. Prefer per-day file audit over blindly matching ccusage totals.
Repro
npx ccusage@latest codex daily --json --speed standard --since <30d> --until <today>
codexbar cost -p codex -d 30 --codex-speed standard --json
Compare non-cache input, cache read, output, and total USD.
Acceptance
Either close the gap with a documented methodology change, or document intentional divergence (and when to use 7d for A/B).
Context
After #158 (ccusage speed-tier + Claude day-window parity), Codex 7d dollars match ccusage within ~0.2% at both
--speed autoandstandard.Codex 30d still lands ~8–9% under ccusage on the same machine (tokens and dollars scale together). Not the old 2× priority bug.
Likely causes
archived_sessionscoverage / filename date gatesPreference
Keep anti-double-count behavior unless we prove ccusage is strictly more correct on real rollouts. Prefer per-day file audit over blindly matching ccusage totals.
Repro
Compare non-cache input, cache read, output, and total USD.
Acceptance
Either close the gap with a documented methodology change, or document intentional divergence (and when to use 7d for A/B).