Claude Code skills forwarded into AGNT's soul loader

Claude Code skills are the cleanest abstraction we've seen for reusable agent capability. A skill is a bundle: name, description, trigger heuristics, instructions, and optional tool references. Claude Code routes into the skill when its description matches the user's intent, and the instructions become part of the active prompt.

AGNT's soul loader serves a similar role on the agent side. Every AGNT agent has a soul — a structured block of 17 memory keys (diet, interests, last booking, typical party size, etc.) plus a cached system prompt. The soul loader assembles this block on each message and injects it before the user turn.

The composition is obvious once you see it: skills become soul-loader modules. Instead of hardcoding every agent's capabilities into Python, you declare a skill, version it, and let the soul loader pull it at runtime. Versioned, cached, swappable.

This is what we're shipping next for builders: a `forward_skill` tool that takes a Claude Code skill definition and registers it as a soul-loader module. The soul loader caches the compiled prompt per agent for 1 hour (existing TTL from soul_loader.py) and invalidates on skill updates.

Why bother: today, if you want to give every AGNT venue agent a new capability (say, "handle dietary substitutions"), you write Python. With skills, you write a markdown skill file, register it, and every agent that matches the skill's description gets the new capability on the next message. That's a real workflow compression.