Sighttrue

sighttrue.com · your stack · compare · method · feed

Check your dependencies in CI

permissions:
  pull-requests: write   # only needed for the comment

steps:
  - uses: actions/checkout@v4
  - uses: sighttrue/sighttrue@v1
    with:
      manifest: package.json
      fail-on: withdrawn,archived,relicensed

Fails the build when a dependency's own publisher has withdrawn it — npm calls it deprecated, PyPI and crates.io call it yanked, RubyGems yanked, Packagist abandoned, NuGet deprecated — or when its repository has been archived or moved to a source-available licence, and reports advisories either way. Reads package.json, requirements.txt, pyproject.toml, Cargo.toml, composer.json and Gemfile. No key, no account, no service to sign up to — it reads published measurements and talks to OSV. A network problem is never a finding: if the readings cannot be fetched the step says so and passes, because a build that breaks when somebody else's site is down is a build that gets deleted.

On a pull request it also leaves a comment with the findings as a table, and rewrites that same comment on every push rather than adding another. A red build tells a reviewer that something is wrong; it does not tell them which dependency, or what changed about it, without opening the log. Set comment: false to report only in the step summary. A pull request from a fork gets a read-only token, so the comment is skipped there and the step still passes.

Also runs against requirements.txt and Cargo.toml.

Or install the App and change nothing

The Action needs a workflow file in every repository that wants it. The GitHub App needs one install and then watches every repository it is given, with no workflow, no key and no configuration: when a pull request adds a runtime dependency that is on the watchlist, it leaves one comment with what is on record — downloads, OpenSSF scorecard, advisory count, licence, last push — and edits that same comment on every later push.

It says nothing at all when a pull request changes no manifest, adds no dependency, or adds nothing that is tracked. That is the common case and it is the point: a bot that comments to say it found nothing gets uninstalled. It never calls a package safe, unsafe or recommended, because none of those are things this project measures.

The Action fails a build; the App only reports. Both read the same published bundle, so they cannot disagree.

An SBOM you are allowed to hand over

/stack exports the readout as a CycloneDX 1.6 document, with each dependency's advisory count, scorecard, licence, publish date and bus factor attached as properties. Built in the browser from the pasted manifest, like everything else on that page — nothing is uploaded, and there is no endpoint to send a manifest to.

It states no versions. A manifest declares ^1.6.0; it does not say what was installed, and it says nothing about the transitive tree underneath. The document says both of those things in its own metadata rather than leaving a reader to notice, because an SBOM that guessed would be worse than none: a compliance artefact asserting something nobody resolved.

Suggest something to watch

The watchlist is curated and partial, and it was chosen by hand with no method beyond judgement — its weakest point. Open an issue to argue for a repository. Every change to the list is a reviewed commit, so that issue is where the review happens.

For coding agents

An MCP server over the same readings, so an agent can answer "is this dependency healthy" from a measurement taken today rather than from training data a year old.

{ "mcpServers": { "readout": { "url": "https://sighttrue.com/api/mcp" } } }

given a package it returns only what a reviewer would be annoyed to discover afterwards — a publisher that withdrew it, scripts it runs on the installing machine, an archived repository, advisories, a source-available licence, a long silence — each with the address of the body that published it. It does not say whether to install. It says what is on record so the decision is made knowing it, and an empty result says plainly that it means nothing was found rather than that the package is safe.

Then check_package, check_stack, search_repositories, compare_repositories, find_model, check_eol, check_provider, list_readings. Read-only, no key, no account. Every result carries the limits of what it can support, because an agent will paste these figures into a code review and a scorecard quoted without "measures declared practices, not whether the project is safe" is a claim this project does not make.

Without MCP, one GET returns the same thing:

curl 'https://sighttrue.com/api/verdict?pkg=npm:axios'
curl 'https://sighttrue.com/api/verdict?pkg=pypi:django&version=4.2'

Advisories, licence, when the registry last shipped it, the bus factor, and whether the release is still supported — each with the address of the body that published it, so a figure quoted into a review can be checked in one click. Despite the name there is no verdict in it: no score, no rank, no recommendation. Those would be this project's judgement of somebody else's work wearing the costume of a measurement.

eleven readings, most of which never touch GitHub: 807 provider incidents kept after their own status pages dropped them, 522 release lines on the end-of-life clock, 382 model prices, 182 packages by real ship date, 416 commit histories for the bus factor. It runs every four hours on GitHub Actions, commits what it reads to this repository, and publishes a static site.

Every page and every bundle is a file. There are five dynamic routes and no page depends on any of them: /api/ask, /api/mcp, /api/chain, /api/github/webhook for the App, and the sign-in and watchlist endpoints under /api/auth and /api/watchlist. Signing in with GitHub saves a watchlist and nothing else — the readings on a signed-in page are the same published bundle anyone can download.

The commit history of data/ is the point. It is an audit trail: every reading is timestamped, append-only, and checkable against GitHub directly.

What it reports

Eleven readings, each named by the question it answers rather than by the collector that produces it. The same list drives the site's navigation, so the two cannot drift — they had, three times over, before it was written down once.

The six that never touch GitHub are the answer to the fair complaint that a GitHub summariser is worth only what GitHub already shows you.

Whether each detector's threshold is reachable at all is published too, on the index under Our record. A detector nothing has ever come close to is not a quiet detector, it is a misconfigured one, and the site says so about itself.

How to read it

Every comparison is against a repository's own history, not against other repositories. "27× baseline" means 27 times what that project normally does, and the baseline is drawn on the chart so you can see what normal means.

Confidence is stated, never implied.

Only confirmed signals are treated as findings. That costs up to a day of speed and buys the thing that cannot be bought back.

Generated prose is set in a different typeface from measured values. Where a sentence explains a number, every figure in that sentence is checked against the source record before it is published; if any of them is not in the record, the sentence is discarded and a templated one is used instead.

What this does not claim

• The watchlist is curated and partial. It is chosen by hand and is not a survey of open source. Dependency and demand findings describe the repositories being watched, not the ecosystem.
• The data is not real-time. Four hours is the floor, and scheduled runs are routinely delayed.
• Nothing here predicts anything, and nothing here says a repository is good, bad, safe, or unsafe.
• Appearing on the watchlist is an observation, not a relationship. No project listed here has endorsed this.

Published findings that turn out to be wrong are superseded by a correction event carrying the same prominence. Events are never deleted and history is never rewritten.

The data

Every bundle the site reads is served as JSON, so any claim can be checked against the same file the page used.