I found a compression proxy on LinkedIn that looked perfect — sit between your app and the LLM providers, strip redundant context, save tokens. I wanted it immediately.
I spent a weekend getting it running. Dashboard toggle. Compression logs. Everything looked good. I closed the tab and moved on.
The feature was running for weeks. It just never actually did anything.
Every request hit an endpoint that didn’t exist in the version I was running. The 404 got swallowed — no error, no alert, no trace. Just null, silently, every time. The token numbers weren’t moving because nothing was being compressed.
Tracing it back, I found a version mismatch, a fail-open pattern that hid the error, and a provider gap that would’ve limited the feature anyway even if everything else was right. I ended up gutting the whole integration.
The thing that actually worked? Already running in 9router — RTK was doing token compression just fine. Headroom was meant to augment it, not replace it.
Full postmortem with the complete autopsy.
#Postmortem #LLM #Compression #9router #DeveloperTools
Full story → https://vianhanif.link/posts/headroom-integration-autopsy/