Some engineering tasks require reasoning across the codebase: Which code is safe to delete? What queries are slowing performance? Introducing Code Scans: codebase-wide audits for any goal. Devin investigates, reports findings, and opens the PRs. Powered by Agentic MapReduce.

We asked Devin to speed up compilation in the @dioxuslabs repo. Devin analyzed thousands of files and applied changes to reduce the debug build time from 58.6s to 21.0s – a 64% reduction across 22 workspace crates tested.

What you can scan for: ▪️ Slow paths and N+1 queries ▪️ Untested flows and edge cases ▪️ Dead code and stale feature flags ▪️ Accessibility against your WCAG criteria ▪️ Your team's own coding standards

Scans work best with a clear goal. Devin helps you scope what to inspect, what to skip, and what counts as a finding before it starts. Type /scan in Devin to try it.

Read more about how Code Scans work:

“just throw agents at it” actually is a totally reasonable solution now

What's going on w. Devin?? I'm on the 200$ plan.

We need to check the kitchen because all this cooking, something has to be on fire right now😭😭

can it scan really massive codebases?

Please bring it to the CLI!

SWE-2 on two occasions now decided to not only wipe my last PR but my entire backup folder in another directory as it decided it was 'stale code.' Practice what you preach.

@imjaredz This looks interesting. Does it work across repositories or only mono repos?

why is devin eating 49gb of my disk space after running sessions?

letting an agent open prs to delete code is high trust

@dabit3 Alright @dabit3, you finally got me. Spinning up an account.

code scans as codebase wide audits is a neat framing, reasoning across whole repo to decide what to delete or which queries drag performance beats file by file linting

Codebase-wide context is the real unlock

Which code is safe to delete is the expensive one to get wrong. What does Devin check before proposing a deletion: static reachability only, or reflection and dynamic call paths too? And is there a measured false positive rate on that class of PR?

Most codebase audits start only after someone already knows where to look. Code Scans flips that model: define the outcome, shard the whole repo, verify every batch, then turn findings into PRs. That is a big contrast with the usual manual, needle-in-a-haystack approach, and Cognition's reported 64% Dioxus debug-build reduction is a strong first demo. Takeaway for developers: stop waiting to search. Build scans that cover the whole repo, verify continuously, and make fixes actionable.

@grok does this work on Devin Desktop or only cloud?

Code review automation feels most useful when it explains why a change matters, not just whether it passes. I would love to see how teams handle false positives at scale.

Great new feature! I've tried some skills for that but the results would vary a lot. It's nice to see a simple workflow for that baked in Devin!

@andrew_locke when in CLI?

Pretty awesome.

is it also a seperate limit like Review?

Safe to delete is the hard one. Reflection and config driven dispatch never show up as call sites, so the confident answer is the wrong one.

Devin isn't letting me work today; this error keeps popping up constantly.

The interesting implication is goal-scoped audits, not “AI read my repo.” If Devin scopes what counts as a finding before it runs, that’s closer to a real eng review than a dump of nits.

“Safe to delete” is the question that turns a five-line function into a three-day investigation.

hey guys ive been making a game can i get a free trial to test devin so, i can keep developing using swe 2? please 😶

For codebase-wide agent audits, what do you track after Devin opens the PRs — finding precision, false-positive rate, and regressions introduced by the fix itself?

@alexkaplan0 super cool

Can't wait to try this on my projects soon.

I track agent sessions and outcomes in DailyWork. For Code Scans, I'd want to trace a finding back to the workers that investigated it. Does the report expose where workers disagreed, or only the final conclusion?

Every day i get more reasons to try devin...

how is it any different than just telling devin "scan the entire codebase for X using subagents"?

'which code is safe to delete' has been the question for 20 years. someone finally shipped it as a button

👀🔥🚀