Pull request reviews
Replies and reactions
Push back on a finding, ask questions, and react — the reviewer rechecks and learns from it.
A review is the start of a conversation, not the end. You can dispute a finding, ask about the pull request, and react to comments; each of these changes what Audit Bench Ai does next.
Reply to a finding#
Reply in a finding’s thread, for example “This can’t be null here, the caller checks it on line 40” or “Fixed in the last commit”. Audit Bench Ai rechecks the code at the pull request’s latest commit and answers in the thread with one of:
| Answer | Meaning |
|---|---|
| ✅ Resolved | The fix is in. The thread is resolved. |
| 👍 Dismissed | You were right, it isn’t an issue here. The thread is resolved. |
| ↘️ Downgraded | It’s real but less serious than first stated; the new severity is given. |
| 🔁 Confirmed | It still applies, with the evidence from the code. |
| ❓ Under review | It needs a bit more detail from you. |
It judges your reply against the code, not against how confident it sounds: “we have tests for this” leads it to look for those tests. If you disagree with its answer, reply again. Each thread gets up to five replies.
Ask about the pull request#
Mention @auditbench anywhere in a pull request comment to ask a question about the change:
- “@auditbench what happens to existing sessions after this deploy?”
- “@auditbench is there anything here that needs a database migration?”
It answers in the conversation, using the changed files. If the answer is about one place in the code, it also opens a thread on that line and links to it. Mentioning @auditbench in a new inline comment starts a recheck of that spot.
@auditbench works on GitHub and GitLab. It’s the only command: there are no slash commands to remember.
React to findings#
| Reaction | Means |
|---|---|
| 👍 | Useful, keep raising this kind of thing |
| 👎 | Not useful here, stop raising this kind of thing |
| 👀 | Seen; you’ll deal with it (counted as acknowledged) |
GitHub doesn’t send reaction events, so reactions are read when the pull request closes. At that point each finding gets an outcome: accepted (👍, the thread was resolved, or its lines were changed), dismissed (👎) or ignored. Those outcomes feed the review quality metrics.
How it learns#
When the same kind of finding in a repository is pushed back on at least twice (a 👎, an agreed “not an issue”, or “won’t fix”) and that outweighs the positive signals, Audit Bench Ai creates a learned rule to stop raising it there.
Managing learned rules#
- Learned rules appear on the Rules page, marked as learned, next to the rules you wrote.
- Switch one off to start seeing that kind of finding again.
- If the evidence behind a learned rule is withdrawn, the rule is removed by itself.
Learning is per repository, so a convention your team follows in one project doesn’t silence the same finding everywhere.