leopold-watch

name	leopold-watch
version	0.1.0
description	Launch a local live web dashboard to watch the current project's Leopold run — status, cost meters (context/subagents/forks), event feed, decisions, and a Stop button. Local-only (127.0.0.1), zero dependencies.
allowed-tools	["Bash"]
triggers	["leopold watch","watch the run","open the dashboard"]

/leopold-watch

Start the local live dashboard for this project's Leopold run and give the user the URL. It only reads .leopold/ (state, plan, decisions, events) and serves on loopback; nothing leaves the machine. The one action it offers is a Stop button, which touches .leopold/STOP — the same kill switch as /leopold-stop.

Run this (idempotent — if it is already up, just report the URL):

WATCH="$HOME/.claude/leopold/scripts/leopold-watch.py"; [ -f "$WATCH" ] || WATCH="scripts/leopold-watch.py"
PORT="${LEOPOLD_WATCH_PORT:-4179}"
if curl -sf "http://127.0.0.1:$PORT/api/state" >/dev/null 2>&1; then
  echo "Dashboard already running -> http://127.0.0.1:$PORT"
else
  nohup python3 "$WATCH" --project "$(pwd)" --port "$PORT" >/tmp/leopold-watch.log 2>&1 &
  sleep 1
  if curl -sf "http://127.0.0.1:$PORT/api/state" >/dev/null 2>&1; then
    echo "Dashboard started -> http://127.0.0.1:$PORT  (pid $!)"
  else
    echo "Failed to start; last log lines:"; tail -3 /tmp/leopold-watch.log
  fi
fi

Then tell the user, briefly:

• Open http://127.0.0.1:4179 (or the port shown) in a browser to watch the run live: run status, cost meters (context MB, subagents, forks, iterations, failures vs their budgets), the live event feed, the decisions log, and a Stop button.
• It shows real data only while a run is active (/leopold-run); otherwise it says "no active run" and updates the moment one starts.
• To stop the dashboard itself: pkill -f leopold-watch.py. It also runs via make watch or, from the npm package, npx leopold-driver watch.
• The Canvas tab renders the run as a live DAG (phases, agents, forks, adversarial verify, plan items as nodes; dependency edges). Click a node to inspect it (model, tokens, per-node cost, prompt/result preview, decisions). From a node you can steer the run — redirect / inject / kill / re-run — which the /leopold-run loop applies at the next turn boundary (a workflow node's steer becomes a directive for the next resume). Git stays locked throughout: steering never commits.

Do not do anything else. Start it, report the URL, stop.