Hacker News new | past | comments | ask | show | jobs | submit
Interesting idea, but wouldn't it use A LOT of usage?

I mean, it takes X usage to review a PR normally (in one go) ... to check every chunk in the entire diff, one by one, seems like it'd be significantly more?

Howdy! I'm the author of gavel.

For background, I use a $100/mo ChatGPT Pro plan and a $200 Ultra plan for Claude. I almost never use Fable: most work is done by Opus 5.5 high (not xhigh).

I'm in a power blackout right now and only have access to my laptop, but this is what Claude could see:

<claude>

Expect gavel to add about 2–3% to an agent session’s token cost, around $0.65 per session at Opus 5.5 API prices, and to take about 15k tokens of context at its peak (median; the worst session reached 42k).

In those 39 sessions the agent ran 162 launch or review create commands, about 4 per session, which works out to roughly $0.16 per review. That per-review figure is approximate, because the count includes retried launches.

</claude>

YMMV, especially because I'm using it in conjunction with my own "second-opinion" skill that does cross-model analysis, review, and remediation before it gets to me in gavel (so that usage pattern almost certainly impacts token consumption). More about that here: https://photostructure.com/coding/claude-code-review/#when-o...