Inspiration

My research measures software supply-chain signals: npm provenance attestations, Sigstore log entries, npm dist-tags. The same pattern kept showing up. Each signal is real, each certifies something narrow, and the package page never says what it leaves out. Provenance names the source commit, but among the attested packages in the top 2,000, 76.7% of attestations record a branch rather than a tag, and 62.6% of the builds referenced at least one GitHub Action by a ref that can move later. Among the top 5,000 packages, 89.7% of tags that sit on a different major from latest point at an older release, so npm install pkg@next often installs older code than npm install pkg, and npm prints nothing. Developers decide in seconds on that page. I wanted the page to tell them.

What it does

Install Check is a Chrome extension for npmjs.com. Under the package title it adds a short strip that counts the signals needing a look and names the most important one. Open it and a side panel shows one card per signal: provenance, workflow inputs, Sigstore signature, dist-tags and install scripts. Each card says what the signal certifies for this exact version, what it does not, which facts were read live, and the measured context from my studies with its source.

At the top, Gemini Nano writes a three-line explanation on your device. For react@beta on 16 Nov 2026 the explanation said, in short: no build record for this version; the beta tag installs a release 883 days older than latest; pin an exact version instead. Follow-up questions are answered the same way, on device when Gemini Nano is available, and grounded only in the collected facts.

When a package runs code at install time, a deep check is one click away. The extension downloads this version and the previous one, unpacks both in the browser, and sends the install-time code plus the file diff to Gemini in the cloud. The verdict (risk level, behaviours with file names, network, process and environment flags, what changed) is cached in Firestore, so the cloud reads each version once for every user. A single "On-device only" switch keeps everything on the machine.

The value is a few seconds of the right doubt at the right moment: before npm install runs, a developer sees which green signal covers what, which tag is stale, and whether the package runs code at install time, with the private part done on their own machine.

How I built it

  • Facts, no model: the extension reads the npm registry, the npm attestation bundle (SLSA statement, Fulcio certificate, Rekor entry), the workflow file at the attested commit on GitHub, and checks for .github/workflows/actions.lock at that commit. Rules in findings.js set each card's level. Every number in the interface comes from one file of measured results with a cited source; the model never invents one.
  • Hybrid inference with Firebase AI Logic: getGenerativeModel in PREFER_ON_DEVICE mode for the summary and questions, ONLY_ON_DEVICE when the user turns the cloud off, ONLY_IN_CLOUD for the deep check and for inputs above an on-device limit. Structured JSON on both paths (responseConstraint for Gemini Nano, responseSchema in the cloud). Each answer shows inferenceSource, latency and the reason for the route.
  • App Check: enforced on AI Logic and Firestore. An MV3 extension cannot load reCAPTCHA, so development uses a debug token and the code ships a CustomProvider for a token-minting endpoint in production.
  • Firestore: shared deep-check verdicts per package version, with security rules that validate the schema and block deletes.
  • Remote Config: model names, the on-device input limit, the stale-tag threshold and the deep-check budget, so I can switch the cloud model without shipping a build.

Challenges

  • The cloud model I chose first for the deep check (gemini-3.8-flash) took 99 seconds on a 63,000-character input; gemini-3.5-flash-lite took about 2. Because the model name lives in Remote Config, I switched the deep check to the light model in the console, with no new build. A timeout still retries on the light model if a slower one is configured.
  • npm's version order and time order disagree: React's canary tag has a lower version than latest but was published after it. The tag analysis compares both before calling a tag stale.
  • The registry document for very large packages runs to tens of megabytes; the extension streams it and stops at a limit that Remote Config can change.

Accomplishments

The hybrid path runs on real hardware: in Chrome 157 with the built-in model, the summary on npmjs.com/package/react is labelled "Gemini Nano ยท on device", and nothing about which packages you browse goes to a cloud model unless you ask for the deep check. Each claim on screen is traceable: per-version facts link to the attested commit, the build run and the Rekor entry, and every context number names its study.

Built during the submission period

All of it, from 16 Nov 2026, with AI coding assistance: the extension, the fact collectors for the registry, attestation bundle, workflow and tarball, the findings rules, the hybrid routing, the Firebase setup and the tests. No earlier code was reused. The measured numbers come from my earlier studies and are used as data.

What I learned

Hybrid inference is a routing problem first. The useful question was not "which model is better" but "which input is small enough and private enough to stay on the device".

What's next

Next I want the workflow card to follow composite actions and reusable workflows, so it covers the full build. After that: deploy the App Check token endpoint, publish to the Chrome Web Store, and bring the same cards to PyPI attestations.

Built with

Firebase AI Logic, Gemini Nano (Chrome built-in AI), Gemini API, Firebase App Check, Cloud Firestore, Firebase Remote Config, Chrome Extensions Manifest V3, JavaScript, esbuild.

Built With

  • chrome
  • cloud-firestore
  • esbuild
  • firebase-ai-logic
  • firebase-app-check
  • firebase-remote-config
  • gemini
  • gemini-nano
  • javascript
  • manifest-v3
Share this project:

Updates

Submission history