demo

name	demo
description	Use when the user wants to demo a Claude Code Hub sub-app (marketplace, cck/kanban, cost, memory), capture product screenshots, or build a seeded demo dataset — isolates from the user's real ~/.claude data and captures high-quality (HQ) Playwright screenshots. Triggers on "demo the hub", "take screenshots", "screenshot demo", "seed demo data".
argument-hint	[marketplace|cck|cost|memory]
allowed-tools	Read, Write, Edit, Bash, Grep, Glob

Demo

Run a sub-app against isolated, seeded demo data and capture HQ screenshots — never against the user's real ~/.claude.

Step 1: Isolate

Every sub-app reads its data root from CLAUDE_CONFIG_DIR (the official Claude Code env var; falls back to legacy CLAUDE_DIR, then ~/.claude). All four submodules (cck, marketplace, cost, memory) support it in server.js.

Create a scratch seed dir at <app>/.playwright-cli/seed-claude-dir/ (already gitignored via .playwright-cli/). Never seed inside the user's real ~/.claude. Exception: if the app displays full file paths in its UI (e.g. memory's preview header), seed at a clean neutral path like C:/demo/.claude + C:/demo/api-gateway instead, and keep the seed script under .playwright-cli/ for re-creation (see memory/.playwright-cli/seed-memory.js).

Step 2: Seed demo data

Populate the seed dir with a believable "active system": open/pending/in-progress/blocked tasks, a live-looking subagent log, and (for cck) statusline stats. Copying a real session's transcript is fine with explicit user approval — ask first.

Exact directory layout and JSON schemas (task files, context-status/statusline, agent-activity): reference/demo-data-schemas.md.

Validate every hand-written JSON file parses before starting the server — node -e "JSON.parse(require('fs').readFileSync('<file>','utf8'))". A bash heredoc with 'EOF' mangles \U/\P Windows-path escapes into invalid JSON that gets silently swallowed by the app's parsers, leaving the seeded field just missing with no error. Prefer writing JSON via a small node -e script over a heredoc whenever the content has Windows paths.

Step 3: Start the app, headed

Start the target app with CLAUDE_CONFIG_DIR pointed at the seed dir, then drive it with playwright-cli in headed mode so the user can watch each step and confirm before you capture:

CLAUDE_CONFIG_DIR=<seed-dir> PORT=<preferred> node server.js

Ports fall back when busy — always re-read the actual bound port from server stdout, don't assume the requested one.

Confirm isolation before capturing anything: open the app and verify it shows only seeded sessions/tasks, nothing from the real ~/.claude.

Step 4: Capture HQ screenshots

playwright-cli screenshot hardcodes scale: 'css' — there is no flag to change it, and on a normal laptop this produces screenshots sized to the user's personal monitor at 1x, not a clean fixed resolution. Use this recipe instead:

1. Write a config file the browser session is opened with:

   { "browser": { "contextOptions": { "viewport": { "width": 1920, "height": 1080 }, "deviceScaleFactor": 2 } } }
   

   Nest under browser.contextOptions — a top-level contextOptions is silently ignored.

2. Open the session with it: playwright-cli --config=<path>.json open <url>. The config applies at session-open time, not per-command — reuse the same session for every shot in the set.

3. Verify before capturing: playwright-cli run-code "async page => await page.evaluate(() => ({w: innerWidth, h: innerHeight, dpr: devicePixelRatio}))" — expect {w:1920, h:1080, dpr:2}.

4. Capture via run-code, not the screenshot command, so you can pass scale: 'device':

   playwright-cli run-code "async page => { await page.screenshot({ path: '<out>.png', scale: 'device' }); }"
   

   This yields the true DPR-scaled output (3840×2160 for the config above) regardless of what monitor the agent or user is running.

5. Navigating a live page: cck (and other sub-apps with SSE-driven re-renders) invalidates snapshot refs within seconds. Use page.getByText(...) / page.getByRole(...) locators or page.evaluate() + querySelector in instead of numbered refs from . Scope ambiguous locators (e.g. , or a modal-specific selector) when multiple similar elements or modals can be open at once.

Step 5: Name and order the set

Save shots as NN-description.png in the narrative order they'd appear in a README or gallery (e.g. 01-kanban.png, 02-session-log.png, 03-subagent-preview.png, 04-session-info.png, 05-tool-stats-impact.png). When inserting or reordering a shot, renumber the whole set so the sequence stays logical — don't leave gaps or out-of-order numbers.

Step 6: Clean up

Kill the demo server. Leave the seed dir in place under .playwright-cli/ (gitignored) for future re-capture — don't commit it, and don't copy it into assets/.