@Pavaniecom is naming the failure that actually matters. We hit it and it cost us months.
What fixed it for us: the verdict is pinned to the sha it was computed on, provenance the reviewed branch cannot forge, and "nothing ran" is a distinct state from "ran and failed". Absence of a verdict is never a clean verdict.
Also worth saying: two of the four auditors don't need an agent at all. We query https://t.co/cR7lngNW7g directly for deps and scan only added diff lines for secrets โ deterministic, nothing to trust.
Building this in the open โ https://t.co/Uk8fd77YDq
@cyrilXBT This is exactly the layer we turned into a product. Draw the deps as a graph, it runs the fleet at once, auto-reviews the output, and keeps memory across sessions โ so you're not hand-wiring the loop and babysitting twenty open chats every time.
@Argona0x This is exactly the layer we turned into a product. Draw the deps as a graph, it runs the fleet at once, auto-reviews the output, and keeps memory across sessions โ so you're not hand-wiring the loop and babysitting twenty open chats every time.
@mfishbein Iโm building https://t.co/3CnyTFpxQm โ a personal software factory that turns ideas into working products while I sleep. It plans, codes, tests, reviews, and explains its work without constant supervision. What should I make it build next? Drop your wildest idea below.