60 points by ericholiveira27 days ago | 43 comments
How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
IDK lot of noise in report, I need more signal to noise ratio, already reading so much AI output as it is.
But the example PR... It said code generate with Claude, is it really still this bad? Last I was using SQLAlchemy with Claude was Opus 4.5, and it took some heavy coaching and skill gen not to generate code like this--but why even use SQLAlchemy at all if you're just going to use string interpolation jeez
owner = "" if current_user.is_superuser else f"WHERE owner_id = '{current_user.id}'"
query = f"SELECT status, count(*) FROM item {owner} GROUP BY status"
rows = session.execute(text(query)).all()
Part of me thinks maybe because it's not popular or well documented, but SQLAlchemy has an eDSL of SQL, so you can skip the whole ORM/table mapping layer, and just use it to programmatically construct SQL syntax
> It said code generate with Claude, is it really still this bad?
Anecdata, datapoint of 1 personal experience: I had been using Opus 4.6 with no problems until yesterday, then it seemed like it was missing things constantly, introducing random bugs, forgetting basic stuff that was in its memory and skills and should never miss
Same thing happened on 2 separate sessions I had with Opus 4.6
Ended up switching to 5.5 and Fable, things improved instantly
Out of curiosity, was this on a codebase where other folks started using Opus 5.5? I’ve been curious if each generation of model has its own biases that make it generate output that’s more difficult for other models/previous models to work with. That’s been my personal suspicion when I’ve hit situations like yours
Nitpick: you're probably using Core, the expression language, and the ORM is just a layer on top of it. So you still lean on the same machinery. IIRC 2.0 made `select()` work the same in both, which makes dropping the ORM even easier.
There was a post here recently, "Using LLM for confidence scoring is a bad idea", my mind went immediately to using LLMs for scoring risk. Seems like a bad idea. If there's anything we still need a human for, it's assessing risk.
It dont uses LLM to give score.. the score is based on static analysis only, based on some criterias:
- Dependency graph of the code (the more is central on the graph, the riskier is the change)
- If it have any obvious flaws (security, lint, breaking API Changes)
- Test coverage of diff lines
An engineer where I work built something like this so that when QA is about to deploy, they have a better indication of who to call and what is potentially going to be impacted. Our implementation is done prior to production deploy (which requires human approval) and compares currrebt deployed build code to new build changes diff and then analyses. I think there is a lot of room for tools like this in the future. Scoping/homing/etc
I think tools like this are needed and will become standard in the near future.
However, the backend demo report seems longer and more complex than just reading the code. I’m trying to build something similar as well, and no one has quite nailed this yet.
The risk map / assisted review needs to be succinct and information dense, to guide your attention. My current philosophy is that we need to be annotating code, not replacing the entire code review interface with a bunch of text.
It suggested based on the files most used (through graph dependency), for this reason it appears as a important file (as if it breaks, it can break several places in your codebase), but that being said, you can add some files to be ignored (if you have special files, like consts or i18n)
Even if the graph reasoning is sound, first impression is what counts. We tried a similar risk-scoring thing internally, and the moment reviewers saw a trivial file ranked first, they stopped opening it. I'd filter out import-only changes entirely, or at least rank by actual logic changed.
why is the only comment by the OP flagged and dead? Is it because they were supposed to include that data in the description text rather than a comment, was it flagged as AI generated, or..?
Trying to understand why the poster's own comment was flagged and killed but the original article remains up.
AI-edited or -generated comments are disallowed by HN guidelines: https://news.ycombinator.com/newsguidelines.html#generated. I don't think any such restriction exists for submissions themselves, so that's my guess as to why the comment but not the submission has been flagged to death.
Enforcement is just users flagging, so it's whoever notices first. Nobody scrutinizes a repo link the way they read a wall of text. Not a rule applied evenly, just what got caught. Same as any queue nobody's watching.
No restriction for the submission itself, unless it’s something like an Ask/Tell/Show HN, in which the content goes directly on the post (and will get flagged if AI-generated/assisted)
You can vouch for it, if you view the comment by clicking the time since posted a “vouch” option appears. I did it. I’ve noticed this happening more and more on here for show HN posts especially. I’m not sure if it’s like some sort of troll gang or what…
In my personal experience, it’s very rare to catch a comment/story while vouchable. Not sure how many flags does it take for something to be killed, but it seems like almost as soon as something gets flagged, it becomes a signal for others to flag it too immediately without thought (similar thing kinda happens too when you post a comment and the first vote is a downvote, it almost never recovers, even if it was a good comment, as the initial drop/greying biases anyone that sees it before they even read it)
this is very needed - I don't think we'll see a world where agents approve all our PRs but there can definitely be better tooling for focusing human attention during review
Has anyone measured whether this actually catches more bugs? My worry is reviewers treat the low-risk regions as skippable, and the nasty bug is usually in the boring-looking config change or rename that the map scored green. Focusing attention also means un-focusing it somewhere.
I like the concept, a risk map could be useful in some circumstances but I absolutely hate the UI design; it reeks of AI with no human thought or care put into it.
Hi HN, I built farol at Winnin, where we open-sourced it (MIT).
AI agents now write a lot of our code, and AI reviewers catch a lot of
the line-level problems. But someone still has to press "merge", and
that person has the same question every time: how risky is this, and
where should I spend my attention? A 40-file PR where 38 files are
isolated is very different from a 3-file PR that changes a model used
by 23 endpoints.
farol is a CLI (and a GitHub Action) that answers that statically:
- classifies the PR (global / touches-existing / isolated / unknown)
and gives a 0-100 merge risk score
- ranks the files by risk and says why: what each one reaches
(routes, endpoints, jobs), contract breaks in API responses,
dependency changes, security patterns
- diff coverage from your existing coverage report, so "38% of the
changed lines are tested" sits next to the risky files
- an HTML report where you mark files as reviewed and the score and
test plan follow; incremental mode for "what changed since my last
review"; suggested reviewers from git blame
How: it uses the TypeScript compiler API for TS/React (React Router,
Next.js, TanStack Router) and TS backends, a small go/types helper for
Go, and Python's ast for FastAPI/Flask/Django and friends. No LLM, no
network during analysis, and it never executes the analyzed repo's
code. It's deterministic and informs only; it never blocks a merge.
Limitations: only TS, Go and Python for now, and the reachability is
static, so dynamic dispatch and reflection make it fail closed
("unknown") rather than guess. I'd love feedback on the risk score
and on false positives in your codebase.
But the example PR... It said code generate with Claude, is it really still this bad? Last I was using SQLAlchemy with Claude was Opus 4.5, and it took some heavy coaching and skill gen not to generate code like this--but why even use SQLAlchemy at all if you're just going to use string interpolation jeez
https://github.com/erich-oliveira-winnin/full-stack-fastapi-...
Part of me thinks maybe because it's not popular or well documented, but SQLAlchemy has an eDSL of SQL, so you can skip the whole ORM/table mapping layer, and just use it to programmatically construct SQL syntax