Rehearsal
Rehearsal simulates thousands of user journeys across your site to find where visitors drop off, before they ever do.
About the project
Rehearsal rehearses your website before visitors arrive. Give it a public URL and a visitor goal. It maps every reachable page into a decision graph, simulates thousands of journeys, replays the risky ones in real browsers, tests a fix on an isolated copy of the site, then writes spec.md - a ranked fix list plus a rebuild plan, each item citing Steve Krug’s “Don’t Make Me Think, Revisited”.
Demo target: https://ai.engineer/code/2026. On the live site, visitors hunting the conference schedule hit 407 links and 16 screens of scroll; 9 of 10 simulated journeys never found it. Rehearsal found that, verified it in real browsers, wrote the fix list, and shipped a rebuilt mockup.
Pipeline: discover → simulate → verify → analyze → compare → spec. The deployed instance runs all five integrations live (Gemini, Jev, Bright Data, Tiger Data, Sentry); each adapter degrades to a deterministic fixture so the pipeline still runs keyless in CI.
Built with
AI tools
Prize Categories Entered
- Submitted

Best Use of Gemini API
Rehearsal's analyze stage calls the Gemini API - gemini-4-hydrogen for fast classification and gemini-4-argon for deep analysis, with automatic fallback when the deep model is quota-limited. Gemini classifies each piece of evidence (verified failure, reproducible concern, hypothesis), writes Krug-cited usability findings, and drafts the prioritized spec. REHEARSAL_GEMINI_MAX_CALLS caps calls per run; a deterministic fixture replaces it offline so the pipeline still runs keyless in CI. Live-verified on a real key during the hackathon.
- Submitted

Best Use of DigitalOcean
The live demo runs on a DigitalOcean droplet provisioned entirely by terraform/digitalocean: a digitalocean_droplet with cloud-init that installs Node, clones the repo, runs npm ci, builds, and serves the app behind Caddy - plus a digitalocean_firewall. App secrets (GEMINI_API_KEY, JEV_API_KEY, BRIGHTDATA_*, SENTRY_DSN, DATABASE_URL) are mapped from Terraform variables into the service env. terraform fmt/validate run in CI on every PR. Deployed at rehearsal.alikhani.co.
- Submitted

Best Use of Tiger Data
TigerDataStore persists every experiment - the state graph (nodes/edges), simulated journey stats, browser run results, and findings - to Tiger Data's managed Postgres (PostgreSQL 18 + TimescaleDB) via TIGERDATA_URL. The dashboard reads experiments back from it; live-verified locally and on the deployed droplet. JsonFileStore is the offline fallback, selected automatically when no connection string is set.
- Submitted
Sentry Raffle
@sentry/nextjs instrumentation wraps every pipeline stage (discover, simulate, verify, analyze) in traced spans tagged with stage, experimentId, and targetHost, with a lazy SDK init gated on SENTRY_DSN. We also built POST /api/hooks/sentry - a webhook receiver that converts Sentry alert payloads into first-responder GitHub issues so a coding agent picks them up (setup in docs/telemetry.md).
- Submitted
Best Use of Bright Data
Two zones provisioned via the Bright Data Account Management API: rehearsal_unlocker (Web Unlocker - first in selectFetcher() for page fetches; it fetched keysight.com, which blocks local Playwright) and rehearsal_browser (Browser API over CDP - first in selectRunner(); loaded ai.engineer in ~6s and ran the live demo's browser verification stage). Every outbound URL passes an assertPublicUrl SSRF guard first, and both fall back cleanly to local fetch/Playwright when unconfigured.
- Submitted
Best Demo Video Using HyperFrames
The demo video at demo/video/ is a hand-built HyperFrames composition: 1920x1080, a paused GSAP root timeline registered on window.__timelines, data-start/data-duration scene clips, real recorded footage of the product running a live audit, Geist typography, and data-media-start/data-playback-rate time remaps - rendered with [email protected] to out/rehearsal-demo.mp4. Voiceover generated in ElevenLabs and mixed as an <audio> track.
- Event prizeSubmitted
Best Overall Hack
Rehearsal closes the loop most demos leave open: point it at a live public site and a visitor goal, and it maps every reachable page into a decision graph, simulates thousands of journeys with a decision model, replays the riskiest ones in real browsers, tests a fix on an isolated variant copy, and writes spec.md - ranked changes plus a rebuild plan, each citing Steve Krug's Don't Make Me Think. On ai.engineer/code/2026 it found that visitors hunting the schedule face 407 links and 16 screens of scroll, and 9 of 10 simulated journeys never found it - verified failure, not a guess. It then shipped a rebuilt mockup driven by its own spec.
- Event prizeSubmitted
Best Use of System One Models
Jev - a system-one decision model - scores each discovered state once: a choice distribution keyed by real actionIds (so the model can't pick an action that doesn't exist) plus an ambiguity score. Those distributions become transition weights for the seeded journey simulator, so 1,000 trajectories reuse one decision per deduped state - cost grows with unique page states, not journeys. The client (HttpJevClient, OpenRouter-compatible decisions API) falls back to a deterministic fixture scorer when no endpoint is configured.
- Event prizeSubmitted
Cleanest Use of a Software Factory
The repo was built by a software factory: 30+ PRs opened by fanned-out Devin subagents on non-overlapping file slices, reviewed by Devin Review, merged by an orchestrator session. The factory itself is code in the repo - factory/ holds the branch-protection ruleset, review policy, and automations; npm run factory:check diffs them against the live repo in CI. Lint rules encode review feedback as back-pressure (contracts-first, finding-cites-krug), check:skill-sync keeps the agents' testing skill in lockstep with the pipeline, a spec-delta golden diffs spec.md on every PR, hourly snapshot-health rebuilds catch a red main (it filed issue #16 on itself), and npm run check:facts verifies env/contract/doc drift in under a second.