Why LLM debugging fails on fragmented repository context

After spending a lot of time debugging large repositories with ChatGPT/Claude, I kept noticing the same failure pattern:

the model was not necessarily "bad at coding" — it was operating on a broken map of the repository.

Typical workflow looks something like:

• search manually across many files


• paste fragmented snippets


• omit surrounding implementation paths


• lose imports/dependencies/state flow


• ask the model to reason across missing context

At that point the model starts interpolating architecture that may not actually exist.

A lot of “hallucinated code” is really just missing repository state.

That was the motivation behind building grab.

grab is a small terminal-native workflow tool for iterative repository context extraction.

The workflow becomes:

search → extract → accumulate → recurse

Instead of repeatedly starting over with disconnected snippets, repository context is accumulated incrementally across extraction passes.

Example:

grab --tree

grab auth

grab --functions server.py

grab 500 635 auth.cs

Function indexing ended up being especially useful because it exposes exact extraction coordinates:

file:start-end

That allows the model to recursively request additional implementation context explicitly instead of guessing hidden code paths.

The accumulated context is continuously copied into the active clipboard/tmux buffer, which makes iterative debugging surprisingly fast in practice.

The tool is intentionally simple right now:

• ripgrep


• sed


• terminal workflows


• clipboard/tmux accumulation

But I think the underlying problem is larger than simple snippet retrieval.

Once repository context becomes fragmented, the model loses architectural continuity and starts constructing plausible-but-invented relationships between components.

I’ve started thinking about this more as:

• repository-state acquisition


• incremental context accumulation


• deterministic extraction


• semantic repository cognition

rather than “prompt engineering.”

There’s a short demo in the README showing the workflow in practice.

Repo:

https://github.com/johnsellin93/grab