Comment by lawrencechen

Comment by lawrencechen 4 days ago

Yeah this is honestly pretty expensive to run today.

> I’m not sure an LLM can really capture project-specific context yet from a single PR diff.

We had an even more expensive approach that cloned the repo into a VM and prompted codex to explore the codebase and run code before returning the heatmap data structure. Decided against it for now due to latency and cost, but I think we'll revisit it to help the LLM get project context.

Distillation should help a bit with cost, but I haven't experimented enough to have a definitive answer. Excited to play around with it though!

> which parts of the code change most often or correlate with past bugs

I can think of a way to do the correlation that would require LLMs. Maybe I'm missing a simpler approach? But agree that conditioning on past bugs would be great

CuriouslyC 4 days ago

Gemini is better than GPT5 variants for large context. Also, agents tend to be bad at gathering an optimal context set. The best approach is to intelligently select from the codebase to generate a "covering set" of everything touched in the PR, make a bundle, and fire it off at Gemini as a one shot. Because of caching, you can even fire off multiple queries to Gemini instructing it to evaluate the PR from different perspectives for cheap.

Reply View 1 reply

lawrencechen 4 days ago

Yeah, adding a context gathering step is a good idea. Our original approach used codex cli in a VM, so context gathering was pretty comprehensive. We switched to a more naive approach due to latency, but having a step using a smaller model (like SWE-grep) could be a nice tradeoff.

Reply View | 0 replies

kburman 4 days ago

For the correlation idea, you might take a look at how Sentry does it, they rely mostly on stack traces, error messages, and pattern matching to map issues back to code areas. It’s cheap, scalable, and doesn’t need an LLM in the loop, which could be a good baseline before layering anything heavier on top.

As for interactive reviews, one workflow I’ve found surprisingly useful is letting Claude Code simulate a conversation between two developers pair-programming through the PR. It’s not perfect, but in practice the dialogue and clarifying questions it generates often give me more insight than a single shot LLM summary. You might find it an interesting pattern to experiment with once you revisit the more context-aware approaches.

Reply View 0 replies