Is this useful for coding work?

[cdarrigo]

Very cool project! I lean on Claude Code pretty hard for real coding work — not chat, actual edits. Exact diffs, line numbers, file paths, even line-ending normalization that breaks if a single character is off.

Given that Headroom compresses what the model sees, and that's lossy on purpose. For Q&A or RAG, that makes sense, but when the model's about to rewrite a real source file, could it be working off a view that's almost
right, but maybe missing something subtle that I don't catch until it's three commits downstream.

The benchmarks you've put out are all Q&A and math and function-calling ( SQUAD, GSM8K, BFCL, etc ) Have you measured anything on the coding side? I couldn't find it anywhere.

Finally, the README says don't bother if you're on a single provider with native compaction. That's literally me — Claude Code, Anthropic only, compaction already on. So am I outside your target audience?

[chopratejas]

These are very pertinent questions. Headroom, by design, does not want to compress things that are containing information. We disable any form of compression on local reads, so your question around coding lines and files is preserved. We do not compress any read calls. We do not compress any globs. These are calls that Claude Code makes in general. We inherently disable any other forms of compression that can impact any information.

Secondly, I think you mentioned the benchmarks. It's very hard to benchmark coding agents because we do not even compress them. In some sense, if you are just looking at source files and data, it would be very hard for us to show a benchmark that can compare it inherently, because of the non-deterministic nature of LLMs. But evals and benchmarks are something that we are doubling down on.

If you are a single provider with native compaction, that is exactly the way Headroom started. Like you, I use only Claude Code and Andromeda only with compaction already on. That is how Headroom started, and it is continuing to expand from that front. Give it a try. Try to use it with your Claude Code because I use it daily with my Claude Code and I have a setup exactly like yours.

[cdarrigo]

what about code that arrives via Bash output? git diff, git show, cat file output — that's the content an editing agent acts on. Its a command output, not a "read." If that gets compressed, bad things could happen.

[chopratejas]

We use tools like RTK, LeanCTX and SerenaMCP to compress these -
No loss of information.

absolutely agree with you - which is why we look at incoming data and choose a strategy based on that :)