dashboard

This skill never installs anything. It starts one Node process from the plugin directory, binds a random port on 127.0.0.1, and opens a browser at it. No packages, no services, no changes to the user's shell.

disable-model-invocation: true is deliberate: this is a command a person types when they want to look at something. Nothing in a conversation should decide, on its own, to open a web page.

Start it

node "${CLAUDE_PLUGIN_ROOT}/bin/dashboard.mjs"

That prints a URL and tries to open it. The output is the whole report — pass it on verbatim rather than paraphrasing it, because the URL carries the token and a retyped one will not work.

Exit	Meaning	What to tell the user
0	Running. The URL is in the output.	Give them the URL. If a browser did not open, they can paste it.
1	It did not come up inside the launch window.	Re-run it with --foreground, which keeps the server in this process and prints why it failed instead of detaching.

Add --no-open when there is no browser to open — over SSH or in a container — and --json when the result is going to be parsed rather than read.

What it shows

Three tabs, joining two sources that have never been joined before:

• Memory — every lesson the instance holds, across every run by default, filterable instantly and searchable properly. Each lesson shows its scope, and there is a filter for the ones visible outside the run that wrote them, which is the question nothing else in the plugin answers. One-click Worked / Did not work sends an outcome; deletion requires typing the lesson id. Activity mode shows the raw feed, which is fetched compactly and carries no scope — the scope and project filters switch off there and say so.
• Turns — one row per prompt: which rung recall used, how many memories it injected, what they cost in tokens, and how many were repeats rendered as a pointer. This is the local half, read from runs/<run_id>/turns/.
• Analytics — the same numbers as a trend, plus spool depth, ingest counts and breaker state.