Caveman + GSD integration — has anyone explored this?

[Hypn0sis]

Hi there, great projects from both sides!
I've been using GSD for structured workflow management and was wondering if anyone has explored integrating Caveman into the GSD pipeline, specifically to compress the reasoning and chat layer while keeping GSD's natural language artifacts (PLAN.md, REQUIREMENTS.md, SUMMARY.md etc.) intact.
The ideal setup would be Caveman active for internal reasoning and agent-to-agent communication, with natural language reserved only for the final GSD artifacts that humans actually read.
The main challenge I see is that GSD spawns subagents with fresh context windows, so a Caveman instruction given at session start doesn't persist into those subagent prompts. Has anyone solved this via agent_skills config injection or another approach?
Would love to hear if this has been tested, and whether there are any known conflicts or gotchas.
Thanks!