Telemetry from ai_reviews
Pass-1 is the default diff review; pass-2 is an optional agentic re-review after pass-1 when enabled.
Final-pass input (code + security)
1.4M
Final-pass output (code + security)
658.5k
Pass-1 input (non-agentic)
972
Pass-1 output (non-agentic)
595.4k
Pass-2 tokens (agentic)
3.2M
Pass-1 share (of pass-1 + pass-2)
15.9%
Avg core IO tokens / PR
24.6k
Avg duration
1m 47s
Total duration (hours)
2.4h
Cache hit rate
71.3%
Est. cost
$56.70
Re-review rate
22%
Avg runs / PR
1.54
Activation rate
59%
Avg planner iterations
2.3
Avg context retrieved
36.6 KB
Planner token overhead
870.2k
Avg planner duration
13s
Activation rate
36%
Avg planner iterations
2.2
Avg context retrieved
30.2 KB
Planner token overhead
409.8k
Avg planner duration
12s
Avg cost to non-blocking: $0.73 across 9 resolved PR lifecycles.
Cache hit % is cache-read tokens divided by cache read plus cache-creation tokens, split by whether the review row is the first run on a PR or a later re-review run.
Cohort | Cache hit rate |
|---|---|
| First runsRun index 1 on each pull request | 71.4% |
| Re-runsRun index 2 and above on the same PR | 93.2% |
Re-runs show cache hits 21.8 percentage points above first runs in this filter.