Codex totals are inflated when duplicate token_count events repeat unchanged total_token_usage #876

Summary

I can reproduce inflated Codex totals with the current released CLI (@ccusage/codex 18.0.8).

On my local logs, the overcount comes from summing info.last_token_usage even when consecutive token_count events carry the same cumulative info.total_token_usage.

I checked this against my own session JSONL files and local Codex SQLite state, and PR #824 appears to be the correct fix for this specific duplicate-total pattern.

Version

• @ccusage/codex: 18.0.8

Current released output

Running:

npx -y @ccusage/[email protected] monthly -s 2026-03-01 -u 2026-03-05 -j

reported:

• totalTokens = 8,321,045,582

Local audit of the same closed window

I audited the matching JSONL files under:

• ~/.codex/sessions/2026/03/01
• ~/.codex/sessions/2026/03/02
• ~/.codex/sessions/2026/03/03
• ~/.codex/sessions/2026/03/04
• ~/.codex/sessions/2026/03/05

and compared them with the corresponding thread totals in ~/.codex/state_5.sqlite.

Across that closed window on my machine:

• Sessions: 88
• token_count events: 66,386
• Duplicate events with unchanged total_token_usage: 42,869
• Current released logic (sum(last_token_usage.total_tokens)): 8,299,097,962
• PR fix(codex): ignore duplicate token_count totals #824 logic (ignore last_token_usage when total_token_usage is unchanged): 2,899,127,807
• SQLite threads.tokens_used: 2,899,127,807

So for this dataset, the PR #824 logic matches the local SQLite totals exactly, while the current released logic overcounts by about 2.86x.

Single-session reproduction

For this session file:

• rollout-2026-03-04T00-00-40-019cb46e-1a1c-79f2-bc67-6551633c9e03.jsonl

I measured:

• token_count events: 8,545
• Duplicate events with unchanged totals: 5,703
• Current released logic: 1,231,236,178
• PR fix(codex): ignore duplicate token_count totals #824 logic: 411,914,227
• sum(diff(total_token_usage.total_tokens)): 411,914,227
• SQLite threads.tokens_used: 411,914,227

So the single-session case also lines up exactly with PR #824.

Why I think PR #824 is the right fix

This is the exact pattern in my logs:

1. A token_count event is emitted.
2. The next token_count event repeats the same last_token_usage.
3. total_token_usage does not advance.
4. The current loader still adds last_token_usage again.

PR #824 adds the missing guard for that case and makes my local totals line up with both:

• the cumulative total_token_usage deltas in the JSONL logs
• the persisted threads.tokens_used values in state_5.sqlite

Related PR

• fix(codex): ignore duplicate token_count totals #824