Scope: the ten most recently modified persisted Codex JSONL sessions at audit
time. Parsed with agent-token-ledger v3.2.0, which prefers cumulative
payload.info.total_token_usage over the last-request counter.
| Signal | Observed |
|---|---|
| Sessions | 10 |
| Provider total tokens | 654,720,502 |
| Provider input tokens | 652,769,135 |
| Cached input subset | 633,036,800 (96.98%) |
| Provider output tokens | 1,951,367 |
| Compactions | 32 |
| Captured tool-output bytes | 67,360,659 (64.24 MiB) |
| Shell execution calls | 3,512 |
| Shell calls with an RTK mention | 217 (6.18%, heuristic) |
| Checkpoint-required sessions | 6 |
| Warning sessions | 1 |
| Sessions with missing worker usage | 1 |
Six session totals exceeded 25M tokens or two compactions. One further session
was above the 10M warning threshold. The active implementation task crossed
25M during verification, and the installed Stop guard returned
checkpoint_required without automatically continuing or blocking STOP.
- This is provider-reported cumulative usage, not a character proxy.
- 96.98% cached input is not 96.98% billing or quota savings; cached tokens are still provider input and may have different price/quota treatment.
rtk gainreports 97.3% reduction inside commands that RTK actually wrapped. That denominator differs from all provider input and cannot prove 97.3% or 99% session savings.- Only 6.18% of observed shell-call payloads contained an RTK mention, and the classifier does not know which remaining commands were eligible. Treat it as a routing signal, not missed-savings proof.
- The valid three-task provider A/B measured 19.67% aggregate reduction after replacing automatic full-skill reads with a compact policy. One task still regressed 32.16%; 99%+ end-to-end savings is not established.
for file in ~/.codex/sessions/**/*.jsonl(.om[1,10]); do
agent-token-ledger --usage "parent=$file" --provider codex --format json
doneThe exact ten-session set is time-dependent because active sessions continue to append. Re-run before using these totals as a current status report.