Automatic Headroom compression without maintaining a QM core fork #1091

I'd like Headroom to run automatically in the background for QM sessions, rather than requiring the agent to call headroom_compress. Ideally users can turn it on/off per session across the supported harnesses while keeping their existing model-provider login.

Headroom's MCP docs distinguish on-demand tools from its automatic proxy mode. Our local attempt needed changes to QM's turn lifecycle and harness routing. I want to keep updating QM easily, so I've stopped building that as a downstream core patch and would prefer an upstream-supported extension point.

We already have a local prototype with working automatic lossless compression and local routing checks. It is not deployed, and full end-to-end validation across the harnesses is unfinished. I can open a PR with the prototype if you'd like.

Is there an existing supported way to plug this into model traffic, or would a small generic per-session transport extension make sense? The goal is to reuse QM's authentication, security screening, history and compaction—not build parallel versions of them. Headroom itself would need no per-user account; existing provider credentials should stay protected, and compression should not expose data across users/sessions.

Happy to follow your ADR proposal workflow if that's the preferred next step.