#569 Claude Code plugin still injects using-agent-skills on every SessionStart, which is the double routing #557 tells users to avoid

#557 (merged 12 Sep) updated docs/getting-started.md to say:

If your host already discovers and activates skills from their descriptions, do not also paste using-agent-skills into an always-on system prompt or rules file. That creates two routers for the same task. Install the individual skills and let the host activate them on demand instead.

Claude Code is such a host. But the Claude Code plugin in this repo still does exactly that:

hooks/hooks.json registers a SessionStart command hook that runs hooks/session-start.sh.
hooks/session-start.sh reads skills/using-agent-skills/SKILL.md in full and returns it as additionalContext, with the comment "Injects the using-agent-skills meta-skill into every new session".

Per Claude Code's plugin reference, hooks/hooks.json is active whenever the plugin is enabled, so every Claude Code session with this plugin gets the meta-skill router on top of Claude Code's own skill routing.

So the manual and Codex setup paths now follow the merged guidance, and the Claude Code plugin path does not. Since #557 also closed #423 (the pack being slow on GPT-5.6), and the mechanism named there was procedural fragmentation costing turns, the plugin path may still carry the cost the fix was meant to remove.

Two possible shapes for a fix, whichever you prefer:

Remove the SessionStart hook from the Claude Code plugin, since Claude Code routes natively, and keep it for hosts that do not.
Make the hook host-aware: skip injection when it detects a natively routing host, inject otherwise.

Happy to open a PR for either. Scope here is only the mismatch between the merged guidance and the plugin's runtime behaviour; nothing about whether the meta-skill itself should change.