I built an AI agent that found a critical bug a human reviewer would've missed published: true

The diff looked harmless

A teammate changes a function's return type. bool becomes dict.
Looks like a minor improvement — more detail in the response, why not?

Three files away, that one-line change quietly breaks payment validation.

This is the exact scenario I built Tripwire to catch.

What Tripwire does

Tripwire is a GitLab Duo agent + custom flow built on GitLab Orbit —
GitLab's knowledge graph that maps real relationships across your
codebase, not just file contents.

Assign Tripwire as a reviewer on any merge request, and it:

Traces the actual dependency chain of every changed function
Finds downstream callers that aren't even part of the diff
Checks for open security findings in those code paths
Posts a structured risk report directly on the MR
Suggests who should actually review it, based on code ownership

No chat window. No copy-pasting a prompt. Assign it like a teammate,
and it runs.

The bug it actually caught

I built a tiny test codebase with a real dependency chain:
notifications.py → payments.py → validation.py

Then I opened an MR that changed validate_card()'s return type from
bool to dict. The MR looked clean. One file changed. Reasonable-looking diff.

Tripwire flagged it CRITICAL and said: don't merge this.

Here's why. In Python, any non-empty dict is truthy — regardless
of what's inside it. So this code:

if validate_card(card_number):
    process_payment(...)

...now always evaluates to True. Even for an invalid card. Silently.
No exception. No test failure unless you specifically assert on the
return type. Just a payment that should have been rejected, going
through anyway.

It also caught a second, smaller issue I hadn't even been thinking
about: the new logic only accepts 16-digit cards. American Express
(15 digits) got silently dropped, with zero mention in the MR description.

Building on a beta platform means assumptions break constantly

Half of what I assumed going in turned out to be wrong:

There's no merge_request:opened trigger (yet) — flows trigger on mention, assignment, or being assigned as reviewer
Custom agents aren't defined in a repo file — they're created through the AI Catalog UI directly
Orbit's API isn't a set of clean REST routes per resource type — it's one generic graph query endpoint

Every one of these required throwing out an assumption and rebuilding
against what the platform actually does, not what seemed like the
obvious design.