@NabhaniMehdi Spot the one line that deletes the wrong rows is exactly it. The harder part is when five agents touched the code and you have to work out which change came from which one.
@fraskey007 Fair framing, “proving intent without re-exposing credentials” is the hard part. We’ll write it up properly once it’s ready to be read. Thanks for pushing on this.
@MuraliMurugan0@Dhanush_Nehru The call tells you what happened. The reason tells you why, and that’s what you need to fix it. I’ll confirm exactly what we capture before I answer, so I don’t guess. What would you want in the log: the prompt, the model’s stated reason, or both?
“One leaked token can now act, not just read” is the right way to frame it. Our starting point is that the agent shouldn’t hold the real secret at all, so there’s less to leak. How mid-task access works exactly is something I’d rather write up properly at release than guess at here.
@fraskey007 I’d bet on secret hygiene. Isolation is a known problem with known fixes. For context, the cleaner answer is to keep real secrets out of the agent’s reach in the first place, instead of hoping nothing leaks out of it. That’s the problem we’re building Midsummer around.
@alexcloudstar@tiennguyendev Tests catch whether it breaks. They don’t tell you what changed or who made the change, and that’s usually where the fear comes from when agents wrote the code.
@tiennguyendev The gamble comes from not being able to see exactly what a fix changed, or who made the change. Small changes you can trace back make a fix a lot less scary, even if you didn’t write the code yourself.
@muso_am The middle fills up with review. Once agents run in parallel, the bottleneck stops being the wait and becomes knowing what each one changed, and whether to trust it.