anotherpanacea-eng

Most cross-host AI tooling treats MCP as the portability layer. That makes sense. MCP is the cleanest common denominator.

This post is about a different problem:

> How do you move a workflow system authored around one host’s command model into other hosts with different invocation, packaging, and runtime assumptions, without maintaining multiple divergent versions?

The system here started in a Claude-oriented environment. The first real port was Codex. Once that architecture was made explicit, Gemini/Antigravity became another generator target instead of a rewrite.

The rule that made this manageable was simple: