Repository navigation
Atalaia 1.2.0
Dependency coverage went from 14 of the 61 manifest and lock files people actually commit to 58, and the test suite stopped failing at random.
Full notes: CHANGELOG.md.
Lockfiles
Nothing was reading lockfiles. Every parser read a manifest, which declares a constraint: ^4.17.0 does not say whether the installed copy is 4.17.11, which is vulnerable to CVE-2021-23337, or 4.17.21, which is not.
Sixteen are read now — the npm family (package-lock.json schemas 1 and 2/3, pnpm-lock.yaml v5–v9, yarn.lock classic and Berry, bun.lock), Python's four, and composer.lock, Gemfile.lock, Cargo.lock, go.sum, packages.lock.json, paket.lock, gradle.lockfile, mix.lock, rebar.lock, pubspec.lock, Chart.lock, conan.lock, cabal.project.freeze, stack.yaml.lock, opam.locked, Cartfile.resolved and conda-lock.yml.
A lockfile now supersedes the manifest beside it. A repository holding both package.json and package-lock.json was storing lodash twice — once as ^4.17.0 and once as 4.17.21 — so every count doubled and the Dependencies tab listed the package twice with two different verdicts. Scoped by directory, so a monorepo keeps its per-app rows.
When a lockfile holds the same package twice, the lower version wins. rust-lang/cargo's own lockfile has 23 crates at two or three versions each. If 1.3.2 is vulnerable and 2.13.0 is patched, keeping the newer one turns a real exposure into a row that reads as clean.
New ecosystems
Dart and Flutter, Helm, Elixir and Erlang, C and C++ (Conan and vcpkg), Haskell, OCaml, Zig, Lua, Conda, Carthage, sbt and Ivy — with version lookups against pub.dev, hex.pm and Hackage.
Four ecosystems had been declared and never produced. The Helm one was visible: the console's Containers & CI only filter and the weekly digest both counted HELM, so they offered a category no row could ever match.
pyproject.toml
It was matched but barely read — one regular expression, and then every entry stored with a null version. Poetry below 2.0 writes no [project] table at all, so a Poetry project scanned as empty. Now: PEP 621, extras, PEP 735 groups as uv writes them, and the Poetry, PDM and Hatch tables, with PEP 503 name normalisation.
Sources and notifications
Debian security advisories, off by default. Read from the DSA/DLA lists rather than tracker/data/json — that endpoint is a status database, not a feed: 85.9 MB, 15.2 seconds, and no publication date anywhere, so every item would have been discarded.
A source that was never called no longer reports as empty. OpenCVE and CVE Details showed EMPTY with "Source responded but returned no vulnerabilities" — neither had responded. The status is NOT_CONFIGURED now, naming the setting.
NVD's 503 is a rate limit, not an outage. NVD_API_KEY is sent when set.
Discord, under Settings → Discord: an incoming webhook and an embed, truncated to Discord's limits because it rejects an over-long payload rather than trimming it.
The test suite
It was failing about one run in eight, on a different test each time. supertest starts a server for every single request, which across 1190 tests in eleven workers is thousands of bind-and-close cycles per run — and a port that has just been released still has the previous connection in TIME_WAIT. Each suite now binds one port and keeps it: 0 failures in 26 runs, against 3 in 8 before.
The migration advisory lock was keyed per database rather than per schema, so eleven workers queued twenty-two acquisitions on one key. The full suite went from 44–47 seconds to 21. That was the obvious suspect and it was not the cause.
Alerting
Old CVEs stopped being announced as new. CISA KEV serves its whole list on every fetch — 1670 items back to 2021 — and the feeds were substituting today's date for a missing one. Advisories are filtered by publication date, an undated one is discarded rather than dated to now, and a cycle sends at most 20 alerts with a pause between them.
Batch actions on the vulnerabilities table, model answers stored as answers (no preamble, no fences, no restated titles), and Ollama asked for the answer rather than the reasoning — think: false took a 12B model from 57.7 seconds to 6.2 for the same paragraph.
Verification
1229 tests, 67 suites
26 consecutive clean full runs