QuotaTrace › sample review
Sample QuotaTrace Review
This is what a $19 review looks like. It was written from the sample report, which comes from synthetic logs: a Pro user, one long gpt-6.1-sol session and one shorter gpt-6-astra session on 2026-10-10.
1. Summary
- Most of your token use is self-inflicted and fixable: one 60,000-token command output was re-sent 109 times (≈6.5M of your 16.7M tokens), and 40 calls were polling steps after
sleep(28.6% of tokens). - One quota jump is not explained by your tokens: +9% of the weekly limit at 15:20 UTC after only 123k tokens. At your own average rate for that window, it should have cost about 0.1%. That one is worth a support ticket.
- Your cache hit rate (92.5% of input) is healthy, so cache misses are not the problem.
2. Quota vs tokens
Your weekly window went from 12% to 27% over 16.7M tokens: 0.90 points per 1M tokens, or 12 points per 1M uncached input + output. Excluding the 15:20 jump, the window cost about 6 points for 16.6M tokens, ≈0.36 points per 1M. That is the baseline to compare future weeks against. If next week runs at 2–3× that on the same model and similar work, collect the report again: that would be a metering change, not you.
3. Where your tokens went, ranked
| cause | tokens | share |
|---|---|---|
| Re-sending one 60k-token command output (session 50e5ff40, call 11) | 6.54M | 39% |
Polling: calls right after sleep, each at full context | 4.77M | 29% |
| Other context replay in the long session (100k+ context calls) | ≈1–2M | ≈10% |
| Actual new work (uncached input and output) | 1.25M | 7.5% |
4. Fixes, by expected saving
- Cap command output (≈35–40% saving on sessions like this). The 60k-token output was a full test log. Add to your
AGENTS.md: "Never print full logs. Run tests with a quiet reporter, and pipe long output throughtail -n 80; grep for failures instead of reading whole files." Once a big output is in context, start a new thread rather than continuing: it is re-sent on every call until compaction. - Stop sleep-and-check loops (≈25% saving). Each check re-sends ~120k tokens of context to learn "still running". Ask for one blocking command with a timeout (
timeout 900 npm test), or run long jobs in a separate terminal and paste only the result. - Split long sessions by task. The gpt-6.1-sol session reached 167k context with no compaction. A fresh thread per task, with a 5-line handoff note, keeps every call smaller.
- Keep gpt-6-astra for the hard parts. The jump below happened on Astra; until it's explained, compare its points per 1M tokens with Sol's on your own next report.
5. Support ticket draft (for the 15:20 jump)
Subject: Weekly Codex usage jumped 9 points after ~123k tokens (2026-10-10 15:20 UTC) Plan: ChatGPT Pro. Client: Codex CLI 0.162.1. Model: gpt-6-astra (effort xhigh). On 2026-10-10 my weekly usage (rate_limits.primary.used_percent, window 10080 min) went from 18% to 27% at 15:20 UTC. My local rollout logs show 123,300 tokens in the calls between those two readings. Over the same weekly window, the other 16.6M tokens cost about 6 points in total (≈0.36 points per 1M tokens), so this step is roughly 200× my average rate. I can provide the full per-call table (timestamps, input/cached/output tokens and the used_percent after each call). Could you check how this step was metered, and restore the usage if it was charged in error?
Get your own review, $19 · delivered within 6 hours