[&] Stack E2E Dashboard

Single-command proof artifact for STACK_COMPLETION.md. Runs every test suite, hits every deployed endpoint, executes the 5 live smoke tests against production Graphonomous, screenshots every marketing site, and emits a self-contained dashboard.html.

Usage

cd /home/travis/ProjectAmp2/e2e-dashboard
./run.sh                 # full run (~15–30 min)
./run.sh --fast          # skeleton check: suites + deploys only, no smokes or screenshots
./run.sh --only suites   # only run test suites
./run.sh --no-live       # skip the 5 live smokes
./run.sh --open          # xdg-open dashboard.html on success

Output lands in out/<ISO-timestamp>/:

dashboard.html — self-contained, portable (embeds CSS, references images relatively)
results.json — structured results for programmatic consumers
logs/{suites,smokes,deploys}/<name>.{log,json} — full captured output
screenshots/marketing/*.png — full-page screenshots of every portfolio site

out/latest symlinks to the most recent run.

What it proves

Test suites (12) — graphonomous, PRISM, ampersand_core, specprompt, agentelic, fleetprompt, body-browser, body-os, delegatic, PULSE, WebHost.Systems, bendscript. Each card shows passed/total vs the claimedTests number from STACK_COMPLETION.
Live smoke tests (5) — fleetprompt-crystallization, prism-os011, body-browser-cross-machine, body-browser-agent-browser, body-os-computer-use. First three hit $GRAPHONOMOUS_LIVE_URL (defaults to https://graphonomous-mcp.fly.dev/mcp).
Fly.io deployments (10) — HTTP probes for sites, full MCP Streamable HTTP handshake (initialize → notifications/initialized → tools/list) for MCP endpoints. Cold-start apps get a warm-up GET first.
Marketing sites (11) — full-page screenshots via agent-browser against a local stdlib static server (port 7777).
Dark-factory 7-step loop — aggregated from the above; each step shows its evidence cards.

Environment variables

Variable	Default	Purpose
`GRAPHONOMOUS_LIVE_URL`	`https://graphonomous-mcp.fly.dev/mcp`	Target for the 3 Graphonomous-dependent live smokes

Dependencies

Node.js ≥ 18 — uses stdlib fetch, AbortController, node:http, child_process. No npm install required.
`agent-browser` in PATH — required for marketing screenshots (phase skipped gracefully if missing)
`fly` CLI — optional, enriches deployment cards when present

Exit codes

0 — rollup.overall == "pass"
1 — any failure, timeout, or down deployment
2 — runner crash (partial results.json still written)

Debugging a red card

Open out/latest/dashboard.html in a browser
Expand the failing card's <details> block
Copy the rerun command shown, paste into a terminal
Fix the underlying issue, re-run ./run.sh --only <phase> to confirm

For failed MCP deploy probes, each card includes a curl reproducer under its details.

Not in v1

Video capture
Lighthouse / SEO / a11y scores
Historical diff across runs
CI wiring (the exit code is CI-friendly but no GHA is committed)
Automatic retry of flaky tests

Architecture

Zero-runtime-dependency Node orchestrator. Five phases; 1–4 parallelize independently, 5 aggregates. Full spec lives in /home/travis/.claude/plans/cheeky-wondering-metcalfe.md.