`/plan` from a plugin collides with native plan mode: unrunnable for ~15s at session start, early input lands as plain chat text

Description:
In the desktop app, a plugin that provides a skill named plan hits a problem none of the plugin's other skills have: for roughly the first 15 seconds of a new session, typing /plan and hitting enter does not complete it to its prefixed form (/pluginname:plan) and does not run it - the text lands in the conversation as plain chat. The model then reacts to the bare text - in our case attempting to invoke the skill itself, which fails with a red error because the skill ships with model invocation disabled. The user's first experience of the session is an error plus a request to retype the command.

The same plugin's other skills (/setup, /next, /rescan, /done) don't show this - typed early they behave normally. Only /plan does, which suggests the input is being captured by Claude Code's native plan mode (or its keyword handling) before the plugin's command registration wins out. Waiting ~15 seconds and typing /plan again works reliably.

Expected: a plugin-provided /plan command resolves consistently - either the plugin command wins once installed, or the collision is surfaced to the user - rather than early input falling through as chat text.

Environment: Claude Code desktop app on Windows 11; plugin installed via a local marketplace.