dashboard

What it does

Serves the read-only browser dashboard on http://127.0.0.1:7777 without tying up a terminal. The server runs as a background task owned by the current Claude Code session — it dies when the session ends (with the session-end hook's reap as deterministic insurance), so there are no PID files and nothing to daemonize. If a dashboard is already listening, the skill reuses it instead of starting a second one.

When to use

• "dashboard", "open the dashboard", "show me the dashboard"
• "show me the graph / memory / tasks / sessions in the browser"
• "stop the dashboard", "kill the dashboard"
• /anton-core:dashboard [surface]

Surfaces: overview (default), memory, repos, tasks, sessions, improvements, news, health, graph.

How

Default port is 7777. Honour an explicit operator port ("on port 8080") everywhere $PORT appears below; never pick a different port silently.

Listener check (shared by every path below): lsof -ti tcp:$PORT -sTCP:LISTEN — the LISTEN filter is required; a bare port query also matches established browser sockets and returns multiple PIDs, which breaks ps -p/kill quoting — then confirm each PID's ps -o command= -p <pid> output matches the dashboard binary: grep -Eq '(^|/)anton-core[^/ ]* dashboard( |$)'. The anchored pattern covers every install layout (anton-core pinned/dev, anton-core-vX.Y.Z bootstrap) and never matches a foreign command line that merely contains similar words (e.g. encore dashboard). Iterate per PID.

1. Probe: curl -fsS -m 2 http://127.0.0.1:$PORT/healthz.

   • Answers ok and the listener check matches → already running (this or another session). Open http://127.0.0.1:$PORT/<surface path> (open on macOS; otherwise print the URL) and stop here.
   • Answers (any body, ok or not) but no listening PID matches the dashboard binary → a foreign app owns the port. Report it and suggest an explicit alternative port; never reuse it, never kill it.

2. Start (probe failed — nothing listening, or the listener doesn't serve /healthz): run with the Bash tool and run_in_background: true:

   "${CLAUDE_PLUGIN_ROOT}/scripts/core" dashboard --port $PORT <surface>
   

   Do NOT detach — no nohup, no disown, no setsid. Session-scoped lifetime is the design: the server binds 127.0.0.1 only, opens the browser itself, and is torn down with this session's background tasks.

3. Confirm: re-probe /healthz up to 5 times, 1s apart. On success report the URL. On exhaustion, read the background task's output and relay its typed error verbatim — a missing data dir surfaces a precondition error (set CORE_DATA_DIR). Exception for the busy-port error: run the listener check first — a PID matching the dashboard binary means an existing dashboard is wedged on the port (it holds the bind but failed the probe); report it as unresponsive and offer the stop gesture instead of relaying the "try --port" hint. If the output carries no typed error (e.g. a slow cold start), relay the raw output and say the server has not bound yet.

Stopping ("stop the dashboard"): run the listener check, then kill each matched PID — plain SIGTERM; the server drains in-flight requests for up to 3 seconds. Kill only matching PIDs; report any non-matching PID on the port as a foreign listener left untouched (refuse entirely when nothing matches). Then verify: re-probe /healthz after ~4s (drain allowance) — report stopped when the probe fails, or that the port is still bound when it answers. Works even when an earlier session started the server.

Output

• One line — Dashboard: http://127.0.0.1:$PORT/<path> — plus whether the server was started fresh or an existing one was reused.
• Lifecycle: the server lives until this session ends (background-task teardown, plus the session-end hook reap on port 7777) or until the stop gesture. Custom-port instances rely on session teardown alone.
• No DB writes — the dashboard is a read-only surface; the skill itself records nothing.