Every AI coding tool feels like magic on a fresh prototype and a liability on a large codebase. The reason is simple: most of them read your repo as flat text, not structure.

Comments

[nomos_ide]

So they forget your architecture, hallucinate file paths, and quietly break something three files away during a refactor. The industry calls it "context rot."
I spent the last few months building Nomos to attack that specific problem.
Instead of stuffing files into a context window, Nomos maps your codebase as a live dependency graph first — so the agent understands how your system actually connects before it touches anything. You hand it a GitHub issue or a plain-English spec, and it plans, writes, and updates code across the whole project, then fixes its own build errors along the way.
What's working so far:

Multi-file refactors that don't silently break distant code
Structural context instead of "dump the whole repo in the prompt"
Flat pricing — £7.99/month, no token metering or surprise bills

Where it still falls short: very large monorepos need smarter graph pruning, and some non-deterministic edge cases still slip through. Both are in active development, and I'd rather hear where it breaks than where it shines.
Built by 2505 Labs. Try it: nomos.wtf