Post

I don't see any skills loading - are you just using vanilla Claude Code?

Go issue #68111 is a kind of run-of-the-mill HTTP test flake. Claude tracked down the race and identified that the fix had landed attached to a different issue. I closed it as a duplicate. research.swtch.com/agentlogs/go...

Go issue #81021 is a Unicode 17 NFC bug, and I touched that code last. Claude explained what I did wrong and identified two other independent bugs in the same code. (Oops!) research.swtch.com/agentlogs/go...

Go issue #77515 was many mysterious darwin-amd64 flakes. Claude pulled the logs and identified some were from a deterministic bug already fixed, while others were due to bad hardware like, on one machine, bit 40 repeatedly flipping incorrectly. research.swtch.com/agentlogs/go...

> If you are not asking them to debug your code and review your bugs, you are doing it wrong. I don't ask Claude or any other commercial LLM to do those things because I can't afford to. These tools need to be accessible on personal computers for them to be useful to me and people like me.

That's entirely fair. The good news is that the smaller, open models will inevitably catch up to "good enough for these bugs" and then you won't need the lab models to do this. The lab models may still be ahead, but it will be like phones: good enough is good enough, no need to upgrade.

I keep an eye on the open models and try them occasionally. On my hardware (RTX5070) they're good enough to ask, "does this function do what I think it does". But not good enough (or more likely my hardware isn't good enough) to ask, "find security vulnerabilities in this (large) project".