git4data.ai

The marketing site for Git4Data — a database-native workspace for proposed data changes from AI agents. Agents can work on an isolated table branch; teams can inspect row-level diffs and merge with an explicit conflict policy. Git4Data is implemented in MatrixOne.

The homepage film replays a measured 10-million-row catalog run with 20 independent Codex tasks. catalog-run.html and data/catalog-run.json document actual results, synthetic data, deliberately injected price violations, and execution limits. The public Playground is a separate 124-row live SQL tutorial with a branch per visitor.

Static frontend with a MySQL-backed Cloudflare Worker API. Deployment requires Node.js 22 or later.

index.html            landing page
playground.html       live SQL sandbox — one branch per visitor
benchmark.html        BranchBench results, jsonbench-style
worker/index.js       playground API and public skill download/statistics routes
scripts/seed-playground.sql   one-time dataset setup for the playground
data/branchbench.json every number shown on the benchmark page
assets/css/style.css  the whole design system
assets/js/dag.js      animated git-DAG canvas backdrop
assets/js/i18n.js     EN / 中文 overlay (English lives in the HTML)
assets/js/bench.js    renders the benchmark tables from the JSON
assets/js/main.js     nav, reveal-on-scroll, copy buttons, step switcher
assets/img/og.png     social card (regenerate: see below)
404.html              themed not-found page
wrangler.jsonc        Cloudflare Workers static-assets config
.assetsignore         files that stay out of the deployed bundle
_headers              CSP + cache policy

Local development

No build step. Serve the folder over HTTP — benchmark.html fetches data/branchbench.json, so file:// will not work.

python3 -m http.server 8080

Then open http://localhost:8080.

Deploying to Cloudflare

The repo is configured for Cloudflare Workers static assets, which is the flow the dashboard offers today (Workers & Pages → Create → Import a repository).

Field	Value
Project name	git4data-ai
Build command	(leave empty)
Deploy command	npx wrangler deploy

wrangler.jsonc does the rest:

• assets.directory: "./" — the repo root is the site; no build step, no output directory.
• html_handling: "auto-trailing-slash" — /benchmark serves benchmark.html, which is what the canonical tags and sitemap.xml assume.
• not_found_handling: "404-page" — unknown paths render 404.html.
• .assetsignore keeps README.md, LICENSE, wrangler.jsonc, package.json and scripts/ out of the deployed bundle.
• _headers sets the CSP and cache policy — Workers static assets honours it, same as Pages.

Deploy from your laptop instead:

npx wrangler deploy

After any analytics-related deploy, confirm acquisition fields are live (they disappeared once when an older Worker build was redeployed — see ACQUISITION_REDEPLOY.md):

curl -sS 'https://git4data.ai/api/analytics/report?days=7' \
  -H "Authorization: Bearer $ANALYTICS_READ_TOKEN" \
  | jq 'has("acquisitionEnabledAt") and has("rowsV2")'

Pointing git4data.ai at it

Nothing to click — wrangler.jsonc declares the custom domain:

"routes": [{ "pattern": "git4data.ai", "custom_domain": true }]

Every deploy binds both hostnames and creates their DNS records. Without this the Worker deploys fine but git4data.ai has no DNS record at all and the site is unreachable, which is easy to mistake for a build failure.

www serves the same assets rather than redirecting. The canonical tags point to the apex. For a true redirect, configure a Cloudflare Redirect Rule in the dashboard.

If you use Cloudflare Pages instead

Pages can serve the static frontend (framework None, build output /), but the live playground requires the Worker API and its secrets. Use the Workers configuration above for the complete site.

The playground

playground.html is a live SQL console. Each visitor gets their own branch of one demo table:

DATA BRANCH CREATE TABLE g4d_s_<id>.customers
  FROM g4d_demo.customers{snapshot='g4d_base'};

The public console accepts a bounded tutorial dialect compiled by worker/playground-sql.mjs. Visitor SQL is never forwarded directly: the server constructs each supported statement with a session-scoped database and snapshot names, and parameterizes country values. Supported operations are the eight guided steps, SELECT *, SELECT COUNT(*), and DESCRIBE on the two demo tables. Other SQL, including access to other sessions or the seed, is rejected.

Sessions expire after 20 minutes idle. New sessions reclaim up to five expired sessions; a scheduled job runs every ten minutes to reclaim up to 30, including their snapshots. Admission pauses when 30 session records already exist (this is a load guard, not an atomic concurrency limit). Queries have a five-second timeout, a 200-row result cap, and an 80-statement session budget.

Only /api/* reaches the Worker (run_worker_first in wrangler.jsonc); every other path is served straight from the asset store, so page views never become Worker invocations.

Pointing it at a database

1 — seed the instance, once. This creates g4d_demo.customers, the _sessions bookkeeping table, and the g4d_base snapshot every visitor branches from. The seed script replaces the sample table; run it only for initial setup on a dedicated demo instance:

mysql -h <host> -P 6001 -u <user> -p < scripts/seed-playground.sql

2 — configure the production Hyperdrive binding. The deployed Worker uses HYPERDRIVE for its database connection; origin credentials are stored in Cloudflare, never in this repository. Disable Hyperdrive query caching so snapshots, session bookkeeping and reads after writes are current. Use g4d_demo as the origin database. The Worker uses fully qualified names because Hyperdrive does not support changing the connection database with USE.

The current demo connection uses Hyperdrive's REQUIRED TLS mode. The instance presents a self-signed server certificate with CA:FALSE, which Cloudflare rejects as an uploaded CA. This configuration provides encryption but does not provide the explicit certificate/hostname verification of VERIFY_IDENTITY. For that mode, obtain a compatible CA-signed certificate chain from the instance operator, upload its CA, and update the Hyperdrive TLS configuration.

For a different Cloudflare account, create your own Hyperdrive and replace its ID in wrangler.jsonc. Direct MySQL TLS from Workers is not supported by this driver/runtime combination; use Hyperdrive.

For a Node.js backend or direct local development, the API also supports MO_HOST, MO_PORT, MO_USER, and MO_PASSWORD, plus MO_CA_CERT for a trusted PEM certificate. This path verifies certificates and hostnames by default; MO_TLS=off is only for local plain connections. Production prefers the Hyperdrive binding when both configurations are present.

3 — check it. /api/health reports exactly which step is failing:

curl https://git4data.ai/api/health

stage: "config" means a secret is missing, "connect" means the host is unreachable, "seed" means the dataset is not there, and "snapshot" means the seed ran but the snapshot did not.

Database permissions

Use a dedicated demo instance/account with no business data. The service needs to create and drop session databases and snapshots, branch from the seed snapshot, and run diff/merge. Snapshot permissions depend on the MatrixOne release and may require an administrative role. Keep that credential server-side in Cloudflare secrets; the public compiler is the application access boundary. Do not expose arbitrary SQL through the provisioning connection.

Local development

wrangler dev serves the site and the API together on port 3000. Without the secrets set, the playground renders its offline state — a notice with the Docker command instead of an error — which is also what visitors see if the instance goes down.

npm ci
npm run dev

For local Hyperdrive, set its CLOUDFLARE_HYPERDRIVE_LOCAL_CONNECTION_STRING_HYPERDRIVE environment variable as described in Cloudflare documentation. To use direct MO_* values in an ignored .dev.vars, remove the Hyperdrive binding from a separate local Wrangler configuration. Run npm test for the tutorial compiler checks, including cross-session access rejection.

Editing the benchmark data

data/branchbench.json is the single source of truth for the results page. One suite corresponds to one table in one paper:

{
  "id": "macro",
  "unit": "s",
  "lowerIsBetter": true,
  "source": "CIDR '27 Git4Data, Table 4",
  "modes": [{ "id": "cold", "name": "Cold run" }, { "id": "warm", "name": "Warm run" }],
  "cols": [{ "id": "git4data", "name": "Git4Data", "self": true }, { "id": "dolt", "name": "DoltDB" }],
  "rows": [{
    "id": "software_dev",
    "values": { "warm": { "git4data": 122.1, "dolt": 1925.6 } }
  }]
}

• Suites without modes put their values under the key "_".
• A value may be a number, or { "v": 0.314, "label": "314 KB" } when the printed unit differs.
• A missing value renders as —, never as zero.
• self: true marks the column that gets the accent colour.

Where the numbers come from

Source	Used for
Git4Data: Database-Native Version Control for AI Agents (CIDR '27) — MatrixOrigin & Purdue	Tables 1–5: clone, diff, merge, BranchBench macro runtimes, scaling
BranchBench: An Extensible Benchmark for Agentic Database Branching (arXiv:2604.17180)	Benchmark definition, third-party capability matrix

Every value is transcribed from a printed table. Where a paper only plots a figure, the metric is left out rather than estimated from the chart.

Regenerating the social card

python3 scripts/make-og.py

Requires Pillow. Writes assets/img/og.png at 1200×630.

Translations

English copy lives directly in the HTML. assets/js/i18n.js holds the Chinese overlay, keyed by the data-i18n attribute on each element. To add a string: put the English in the markup with a data-i18n="some.key", then add 'some.key': '中文' to the ZH object.

Licence

Site code MIT (see LICENSE). MatrixOne itself is Apache 2.0.

Cache busting

_headers marks /assets/* as immutable for a year, so every asset URL carries a ?v=N query. When you change a file under assets/, bump N in every HTML file (index.html, playground.html, benchmark.html, 404.html) and in the fetch() call inside assets/js/bench.js when the dataset changes — otherwise returning visitors keep the old copy. Bump it after you finish editing, not before: republishing different content under a version number a browser has already cached is the same as not bumping at all.

Catalog film

See scripts/catalog-proof/README.md for the measured experiment and outputs/catalog-proof/ for its retained task plans, query log and results. After a successful run, publish-results.py creates the public evidence and render-video.py renders the captioned 60-second replay. The renderer uses Pillow, FFmpeg and macOS system fonts. The film contains no audio and starts only on user action.

Downloadable Agent Skill and impact measurement

/agent-skill provides the bilingual installation guide for skills/matrixone-safe-data-changes/. The versioned ZIP is hosted as a GitHub Release asset so its download count survives Worker redeployments. /api/skill/download redirects to the fixed release asset; /api/skill/stats reads that asset's public download_count, cached for 15 minutes at each edge location. These endpoints run before any database configuration checks and do not depend on MatrixOne availability. No new Cloudflare storage, analytics subscription, or OAuth scope is required. API rate limits/outages reuse a timestamped last-good observation for up to 24 hours when available, or render an unavailable count; neither path invents a zero. GitHub can rate-limit shared Cloudflare egress IPs. Downloads continue via the release link. The apex and www hostnames share the same per-release cache key. The browser falls back to GitHub's public release API without cookies or a referrer if that cache is unavailable; this third-party metadata request is disclosed on the download page and allowed by CSP.

The current-version count includes repeated requests, bots and our verification downloads. It does not deduplicate people or prove installation or use. Source clones, mirrors and copied instructions are not counted. The installed package contains no telemetry. GitHub handles asset downloads under its own policies; we do not collect identity, database credentials, SQL or runtime events for metrics.

Publish a version

1. Update metadata.version in the skill and the version/tag/asset in data/agent-skill.json.
2. Run python3 scripts/package-agent-skill.py; it writes a deterministic ZIP and SHA256SUMS to outputs/skill-release/, and the hash/size to the public manifest. Update the download page's version, release/source links, size, checksum, and translations from that manifest when releasing a new version.
3. Commit the source and manifest, push, and create a release at that exact commit with both artifacts: gh release create skill-v0.1.0 outputs/skill-release/* --target <commit> --title "MatrixOne Agent Skill v0.1.0" --notes-file <notes-file>. Substitute the new version for future releases. Upload before deploying website links. Never replace an existing version's ZIP: create a new version to preserve integrity and historical counts.
4. Deploy the website. Read-only verification downloads count as downloads; report them as such.

Measure influence without tracking installations

Run node scripts/skill-metrics.mjs to obtain current and historical release-asset download counts, plus repository stars/forks. It reads only public GitHub metadata; optional GH_TOKEN raises API limits. Save periodic JSON snapshots outside the deployed assets to calculate weekly changes. This is an on-demand command; no background scheduler is installed.

Keep these signals separate:

Stage	Signal	Interpretation
Awareness	First-party page views and click counts	Events, not unique people or adoption; see Site analytics below
Distribution	Release ZIP downloads by version	Requests, not unique users or installations
Interest	Repository stars/forks	Covers the whole site repository, not only this skill
Activation	Voluntary GitHub experience reports	Self-reported successful branch → review → merge
Adoption	Consented case studies / repeat reports	Qualitative evidence of sustained use

The website invites voluntary public GitHub reports without asking for confidential data. Do not call downloads MAU, paid conversion, or production adoption. Add opt-in aggregate usage events only if users explicitly choose them in a future executable tool; a documentation skill should stay offline.

Tool roadmap

This release ships instructions, SQL guidance, permission/approval design and a review-report template. An operator can use an existing MySQL client and separate branch/merge credentials. A custom CLI/API/MCP becomes useful for unattended or multi-user execution: it must enforce workspace scope, immutable proposals, policy, approvals, target-version checks, resource limits and durable operation state. MCP is an interface option, not an access-control guarantee. Self-hosting supports either approach.

Blog

blog.html lists engineering articles. The first case study lives at blog/10-million-products-20-codex-agents.html (English) and blog/10-million-products-20-codex-agents-zh.html (Chinese). Both are static HTML with canonical URLs, reciprocal language links, article metadata, video, SQL excerpts, and links pinned to the completed experiment commit. The article distinguishes measured execution, policy-based approval, injected faults and the reconstructed video interface.

Edit the two article files together when changing facts. Their evidence comes from the retained catalog run, not from another database execution. Add new article URLs to the blog index and sitemap. Shared editorial styles live in assets/css/style.css; bump its cache version in all HTML pages, including nested blog pages, whenever it changes.

Site analytics

All public HTML pages (including both article languages and 404) load assets/js/analytics.js. The private /analytics dashboard does not track itself. scripts/instrument-pages.py assigns persistent control IDs and creates data/analytics-catalog.json, the server-side event allowlist. Run it after adding pages, links or buttons, or regenerating catalog-run.html; do not reuse an existing ID for a different action. Dynamic Playground steps and benchmark buttons are explicitly instrumented in their own scripts. New dynamic controls must be added to the catalog deliberately.

Metrics and limits

• page_view: once when a page becomes visible per document load, and again on BFCache restoration. Reloads count; hidden prerenders wait until visible. Internal anchor navigation does not add a view.
• click: trusted link/button activation, including keyboard activation. Every occurrence counts; dynamic benchmark button counts are grouped by control type. No arbitrary page/body-click heatmap.
• video_start / video_complete: first play and first ended event per video per document lifetime. The latter is an ended event, not proof that the viewer watched every second.
• Ratios are clicks per 100 views, not unique-user conversion rates. No users, sessions, IPs, cookies, fingerprints, referrers, query strings, SQL or form values are stored in analytics.
• Acquisition tracking: Records normalized UTM parameters (source, medium, campaign) and coarse referrer categories to measure promotion channel effectiveness. Only allowlisted values are stored; others collapse to 'other'. Campaign values are validated against a strict lowercase alphanumeric slug pattern (max 50 chars). Raw query strings, full referrer URLs, IPs, and user identifiers are never stored. CTA clicks may be attributed to landing source within the same browser session (sessionStorage only), cleared when the tab closes, with no cross-session or cross-device tracking. See worker/acquisition-tracking.mjs for allowlists and /privacy for full disclosure.
• DNT, GPC, local opt-out and ?analytics=off suppress collection. Automated/headless clients and recognizable bots are filtered where possible. Ad blockers, navigation loss, quotas, rate limits and network failures can reduce counts; forged or undetected bot events remain possible.
• No historical data is backfilled. Reports clamp the requested range to collection activation.

Events go to /api/analytics/events as small same-origin batches, with omitted credentials/referrer. The Worker validates the page/event/target against the catalog, limits body size and batch length, and uses Cloudflare's rate limiter (60 requests/minute per edge/IP key). IPs are transient rate-limit keys, not analytics dimensions or stored identifiers. The browser caps 200 events per document.

SiteAnalytics is a SQLite Durable Object; it atomically increments daily UTC counters, independent of MatrixOne. Counts survive deploys. It retains 180 UTC days and prunes older buckets daily through an alarm, plus reads/writes. Cloudflare recovery backups may have additional retention. The site-v1 object name, class name and site-analytics-v1 migration are stable identities: do not rename or remove them casually. This uses existing Cloudflare services and their plan quotas, not unlimited free storage; quotas can cause missing events. No new paid subscription is purchased by deployment.

Owner dashboard

Open https://git4data.ai/analytics and paste the private viewing key. ANALYTICS_READ_TOKEN is a Cloudflare secret. Its initial local copy is outputs/private/analytics-access-key.txt (mode 0600, Git-ignored, and excluded from deployed assets). It is never embedded in HTML or a URL. Anyone holding this read key can see aggregate statistics, so treat it as a secret and rotate on exposure. To rotate, securely generate a new key and upload via wrangler secret put ANALYTICS_READ_TOKEN; no source change is required. The dashboard keeps the key in memory and sends it only in the Authorization header to /api/analytics/report. Reports use Cache-Control: no-store.

The dashboard provides 1/7/30/90/180-day ranges, page filtering, daily totals, page totals, per-control click counts, acquisition source breakdowns (by UTM source/medium/campaign and referrer category), CTA click attribution by source, CSV export and optional one-minute refresh while visible. Login failures never return data. The public page shell is intentionally reachable but has no statistics embedded. Privacy disclosure and opt-out live at /privacy. Installed Skill telemetry remains absent; website clicks and GitHub release downloads remain separate measures.