@codegraph Agents are stuck in the “asking for directions at the gas station” era of navigation It’s ya boi

@codegraph I am gonna give it a try, also I think it’s super thoughtful way to give one command to remove all of them. I can just use them pre git push to remove those automagically

Per-file summaries solve the discovery problem well. The harder one they don't touch is multi-hop reasoning. If I change parseConfig, what breaks three hops out? That's what symbol-graph approaches give you, which is why Aider's repomap uses PageRank rather than file-local prose. Have you benchmarked on refactors where the blast radius crosses five or six files?

@codegraph I wonder how we can better use files own metadata to store this. What attributes would need to be added. Delta Lake / iceberg have separate metadata logs, this is a similar idea except more granular which makes me think it needs to be OS level

@codegraph This is overly complicated. LSP already does all of this using tree-sitter. And I've stripped it all down into a very efficient CLI tool for agents. Check it out. github.com/ind-igo/cx

The quieter benefit at the end is the one that matters most. Everyone benchmarks on token cost, but the real value is consistency. Same task, same output, every run. Structural comprehension is where that leaks today, and once you've seen an agent wander a different path on the same problem twice it stops feeling like optimisation and starts feeling like a requirement.

@codegraph Well written!

@codegraph why do you think cursor/cc don’t do this natively? seems very straightforward

Grok Imagine now has dramatically improved lip sync and sharper audio quality on all image-to-video generations. Dialogue tracks the mouth. Sound matches the scene. Your videos look and sound the way you imagined them. Try it today in the Grok app.