The reviewer reads this file first. It says what a failure looks like in this repo.
- Start with the count:
2 important, 1 nit, 0 pre-existing. - Every finding names
file:lineand the input that breaks it. No evidence, no finding. - At most 3 nits. Say nothing about style the formatter already fixes.
- A route, tool input or env read that skips its zod schema.
- Text shown to a customer that names a model, a token count or a stack trace.
- A worksheet that passes
checkBandbut overflows its page, or a band limit moved without the band table. - A bug fix with no test that fails without it.
- A permission, hook or CI workflow change that widens what a run may do.
supplied/,test-results/,docs/recordings/and lockfiles.- Existing problems the diff does not touch: label them pre-existing and stop.
Round one reads the whole diff. Later rounds read only what changed since the last round and never repeat a finding.