Two years ago I lost most of a Friday to a single number. A retry loop in a payment path, set to seven. Not five, not ten. Seven. No comment, no test that explained it, and the person who wrote it had left the company eighteen months earlier. I decided it was arbitrary, changed it to three, shipped it, and found out that the provider rate-limits at six. The answer existed the whole time. It was in a pull request comment from two years earlier, one sentence into a thread that was mostly about something else: "they rate-limit at six, so seven is the floor". Nobody threw that away on purpose. It just stopped being findable the moment the branch merged. That is what Pruve is. Connect a repository, read-only, and it indexes the history that already exists: commits, pull request discussions, review threads, linked issues. Then you click a line and it tells you why that line is the way it is, with every claim linked to the thing it came from. You can open the receipts and disagree with them. It got more urgent rather than less. When an agent writes the code, the reasoning exists for the length of a session and is discarded at commit. The code ships, the argument does not. There is a one-line hook that keeps those sessions and links them to the commits they produced, so why-blame on AI-written lines returns the actual conversation. That part is beta. The rule the whole thing is built on: receipts, or silence. If there is no recorded reasoning, Pruve tells you what it searched, how much of it, and stops. It will not write you a confident paragraph it inferred from the code, because a wrong answer about why your code is the way it is costs more than no answer. You act on it. One thing I did not expect to matter as much as it does: it is exposed over MCP, so your coding agent can ask the same questions before it edits. It reads why the odd-looking constant is there, and leaves it alone. What it does badly today, so you find out from me rather than from it: repos with a thin pull request culture get thin answers, because there is genuinely less to find. The browser extension breaks when GitHub changes its DOM, which is why the web app stays fully capable. And the agent session hook is beta and only covers Claude Code and Codex properly. Free forever for public repositories, no card. GitHub and GitLab both, from day one, because GitLab teams never get day-one anything. Visit https://pruve.dev/launch gets you two months of Team free on top of the trial if you have private repositories. I would rather have hard questions than upvotes. Point it at a repository you know well and tell me where it is wrong.

Forum Threads

p/pruve-2

•

2mo ago

What's the weirdest load-bearing line in your codebase?

Every codebase has one. An odd retry count, a sleep that fixed a bug nobody can reproduce any more, a constant with no comment that survives every refactor because changing it is a gamble and leaving it is free.

Mine is maxRetries = 7. No comment, two years old. The answer turned out to be in a review thread from 2024: the provider rate-limits at six.

What's yours, and did you ever find out why?

View all

Pruve.dev

Ask any line of code why it exists, and get receipts

Ask any line of code why it exists, and get receipts

Forum Threads

What's the weirdest load-bearing line in your codebase?

Forum Threads

What's the weirdest load-bearing line in your codebase?