integration test: exercise CLI skill via Claude Agent SDK

[zzstoatzz]

idea

add an integration test that spawns a claude-code instance with our plugin installed and verifies the CLI skill actually teaches the agent proper usage patterns.

flow

1. spawn claude-code instance via agent SDK
2. /plugin marketplace add prefecthq/prefect-mcp-server
3. /plugin install prefect@prefect
4. give it a mutation task (e.g., "cancel flow run X")
5. assert it:
   • uses --no-prompt as top-level flag
   • uses prefect api for JSON output when needed
   • doesn't fail on truncated IDs

questions

• how does the agent SDK handle plugin installation?
• can we mock the prefect API for deterministic testing?
• what's the right assertion model for "agent used skill correctly"?

refs

would need to understand claude agent SDK patterns - see claude-code-guide