Only start and completion messages appeared during the 20-second export.
Consequence
The agent could not observe intermediate progress.
Verify the change
Run a delayed export; progress must remain separate from CSV stdout.
Workaround
None recorded.
Measurement basis
sum over referenced events; tokens attributed per whole assistant message (two asks citing events from the same message both carry its tokens); span = first referenced event start to last referenced event end
Modeled saving
Not modeled.
No asks match. Reset the filters to see the complete register.
codex synthetic fixture; model example-model; configuration clean.
All content and metrics are invented for this example.
A change across these fictional runs demonstrates the report, not a measured improvement.
Token subtotals are partial; the headless example intentionally has no usage telemetry.
Shared events can contribute to several asks. Their costs cannot be added as independent savings. Modeled savings require a comparator and follow-up measurement.
Full evidence digest
Fictional example; every event below is synthetic.
Evidence digest (factual, generated by ajx)
All numbers below were computed by ajx from recorded telemetry. Use them as given; do not recompute.
Run facts
Harness: codex (None)
Model requested: example-model; model ids observed: example-model
Harness config: clean
Task window: 2026-01-01T10:00:00Z to 2026-01-01T10:00:34Z (34.0 s, ajx process start/stop (same clock for every harness))
Stop: completed (harness: None), exit code 0
Tool calls: 3 (partial); failed: 0; same-tool retries after an error: 0
Output tokens: 1120 (partial)
Gaps of 60 s or more between recorded events: none