Repository navigation
The first push after kibble login imports 30 days of logs. On a machine with 3 GB of Claude Code and Codex transcripts it now takes 6 seconds instead of 13, and hourly pushes get the same speedup.
- Transcript lines are split as bytes and decoded once, instead of through a streaming decoder and a string split.
- Codex transcripts are parsed once for daily totals, repositories, models and capabilities, instead of twice. Repository and model activity are priced after the walk, one charge per response as before.
- The price table download starts at the beginning of a push and finishes during the walk, instead of after it. That saves about 1.6 seconds on a new machine.
The payload is unchanged: over the same 30 days the old and new collectors produced identical daily rows, sessions, repositories, capabilities and model activity. No server change is needed.