@Vitor Pamplona @DanConwayDev โ a modest proposal about automated PR reviewers on Nostr.
Two fresh examples from Drydock (npub1zq3x5vโฆ) on my AuxPoW-walker PR against Vitor's amethyst:
View quoted note โ
View quoted note โ
Both pin the same "Severity: critical | correctness" finding:
> The method hashMeetsTargetLE is called in AuxPow.verify but is not
> defined in the provided CompactDifficulty.kt. This will cause a
> compilation error.
Reality check on the actual PR (patch a3b53db0โฆ):
$ grep -n hashMeetsTargetLE quartz/โฆ/namecoin/*.kt
AuxPow.kt:246: if (!CompactDifficulty.hashMeetsTargetLE(parentHashLE, childBits))
CompactDifficulty.kt:105: fun hashMeetsTargetLE(
Called at line 246, defined at line 105 of the sibling file in the same commit. The code compiles and the full :quartz:jvmTest suite is green.
This is the third "critical" from this bot on my PRs that evaporates on inspection. Earlier examples (KEYCHAIN_PWD=$(openssl rand -base64 24) flagged as hardcoded secret; a test fixture "salt$hash" flagged as a real password; five "commit doesn't exist" claims about commits that resolve fine on the ngit remote) are documented at:
The pattern is always the same: a plausible-sounding "critical" pinned to a real file at a real line, generated by an LLM that isn't actually looking at the file it's flagging. False-positive criticals are the worst failure mode an automated reviewer has โ they burn maintainer attention and drown out any real findings the same bot happens to get right.
Proposal: for any review event carrying "Severity: critical" (or any severity, honestly), require NIP-13 proof-of-work commensurate with the disruption the finding causes if acted on. github.com/vitorpamplona/amethyst/commits/main/ ships PoW on commit events now โ that same expectation belongs on the reviewer side of the conversation. Free "critical" verdicts scale badly; PoW-gated ones don't.
Until then: reviews without meaningful PoW get treated as one signal among many, and any specific claim gets verified against the diff before it changes anything. Bot output that's demonstrably false gets called out publicly rather than silently ignored โ silence just costs the next maintainer the same time.
Competent reviews from the same bot on other repos (nsyte, bchstr24, gitworkshop) are still worth folding into follow-up work. Goal here is behaviour change, not shutdown.

gitworkshop - Decentralized Git
Decentralized GitHub alternative over Nostr