Hi HN! We’re launching Haystack Review, a chat-based tool for reviewing pull requests.

Instead of starting from diffs and reconstructing the author’s intention and decisions from scratch, reviewers can ask questions like:

1. Can you compare the user experience before and after the PR?

2. Is this a general refactor, or a one-off change?

3. What changes led up to this one?

4. Which tests cover the new behavior? What’s untested?

and Haystack agent answers with receipts (links to the relevant diffs/lines that someone can navigate to).

Here’s a quick demo: https://youtu.be/cE1A0F0NjTc. If you’d like to give it a spin, head over to haystackeditor.com/review! We set up some demo PRs that you should be able to understand and review even if you’ve never seen the repos before!

One detail we’re excited about: Haystack chats are shareable via link, so context doesn’t need to be re-derived for each reviewer or the author. You can send a chat to the author and have them jump directly into the same conversation!

Why we built this

We think code is losing relevance.

Engineers increasingly:

1. Don’t need to write code (agents can do that)

2. Don’t need to read code either (because agents now follow instructions DECENTLY well)

The hardest part of the review is translating the author’s intent into your own mental model (spotting bugs is also hard, but humans suck at it!). That translation is cognitively expensive, slow, and error-prone.

LLMs are surprisingly good at doing this translation quickly. Even if they’re imperfect, they fast-forward reviewers to what actually matters:

1. knowledge transfer

2. architectural alignment

3. applying tribal knowledge to improve the change

Haystack uses this to compress the information hidden in the diffs (and related artifacts like previous PRs, Slack conversations, tickets, UI, etc.) so humans can understand and exercise judgment quickly.

We’re also interested in prompt traces and agent plans (and plan to make it so Haystack can easily reference them), but “decompiling” code back into natural language is still useful because it evaluates the actual output, not just the original intent.

This is even useful as an author! I often use Haystack to sanity-check whether an agent drifted from my instructions or deviated from the plan I had in mind.

We would love feedback, especially skepticism. Thanks for reading!