The full plugin authoring contract (hooks, tools, OAuth, memory providers, overrides) lives as an opt-in skill so it is not injected into every system prompt:
.catalyst-code/skills/plugin-authoring/SKILL.md
(Also staged to ~/.catalyst-code/skills/plugin-authoring/SKILL.md on first run.)
Apply it with /skill:plugin-authoring, or have the agent read that SKILL.md when
authoring or debugging a plugin.
A plugin command can be prompt-backed (prompt_file + mode: "agent_turn")
instead of script-backed: the harness renders the template and submits it as a
normal agent turn (same path as a user send), so a single slash command like
/deep-research can drive a full multi-agent workflow with no script, no
interpreter, and no recompile. See the "Prompt-backed commands" section of the
plugin-authoring skill. The bundled deep-research plugin is the reference
implementation.