Blog

4 min readrelease

Four alerts, three findings, zero regressions: the security round

How Unown cleared every open Dependabot advisory and CodeQL finding in one gated round — including the one Dependabot's own patch would have left broken.

  • #release
  • #security
  • #verification

v0.0.55 is a security release: both open Dependabot advisories closed, all three open CodeQL findings fixed, nothing regressed. Like the dependency round before it, the interesting part is not the diff — it is the discipline that lets you believe the diff. Here is the shape of it.

The intake came from the API, not from memory

The round opened by reading the repo's actual security state through the GitHub API: four open Dependabot alerts (follow-redirects at HIGH severity, minimatch at MEDIUM, three manifests), three open code-scanning findings (one js/insecure-randomness, two js/polynomial-redos), zero secret alerts, and seven open Dependabot PRs. That census is the contract — "address dependabot and code sense" means that list goes to zero, and every decline has to be argued, not shrugged.

The advisory that needed five surfaces, not one

minimatch's fix (GHSA-g3hv-75wg-m249) is a major bump: v11 swaps the matcher to a non-recursive engine that upstream explicitly will not backport. Dependabot's PR touched two files — the manifest and the shrinkwrap. But this repo's vendored tree carries five surfaces that pin that version, and leaving three of them behind would either keep the alert open or, worse, break the lock generators' consistency checks. So the fix regenerated all five through the repo's own tooling: the exact pin, npm-shrinkwrap.json, install-lock/package-lock.json, the workspace package-lock.json, and bun.lock — whose diff came out at four lines. npm audit --omit=dev now reads zero vulnerabilities in both trees that carry lockfiles, and the 288 tests covering minimatch's two consumers (model-resolver globs, package-manager path patterns) pass.

follow-redirects (GHSA-mrwp-f3q8-3g9p) was the quieter one: the bun lockfile had already drifted up to the fixed 1.16.2 on its own; only the npm lock lagged — a three-line fix once you know that npm install --package-lock-only alone will not move a range that is already satisfied, because npm conservatively preserves existing resolutions. The fix forces exactly one re-resolution.

The finding that was eleven bugs wearing two hats

CodeQL flagged two lines: .replace(/\/+$/, "") on a user-supplied base URL in setup.ts and providers-cli.ts. That regex is polynomial — a paste ending in N slashes makes the engine rescan the run from every offset, O(N²) for what should be one backward scan. But grepping the tree found the same idiom inline at eleven call sites, including one normalizing the Origin header of live HTTP requests. Fixing only the two flagged lines would have bought a clean dashboard until the next scan flagged the next copy. So the round extracted one linear, regex-free helper (charCodeAt backward scan — no engine, no backtracking, allocation-free when nothing trims), swept all eleven sites, and pinned the contract with tests — including a hostile 100 000-slash URL that must trim in a single linear pass, and byte-identity proofs for every shape the old regex produced, down to the Windows-path normalizers in the uninstall code.

The randomness finding got the same treatment class: the sigil demo's task ids drew from Math.random(). They now draw from crypto.getRandomValues — the CSPRNG — same id shape, and a pinning test that stubs Math.random deterministic and proves the ids still vary, so the fix cannot silently regress.

The decline that is also a receipt

One Dependabot PR was declined, on the record: @vitest/coverage-v8 4 → 5. The registry shows its peer contract pins vitest to exactly 5.0.5 — the isolated bump cannot satisfy the frozen vendored workspace it targets, and the paired vitest 5 migration is an upstream-sync decision, not a routine bump. Declining with the receipt is the difference between "ignored the robot" and "audited the robot." The same reasoning left the root lint chain's old minimatch alone: dev-only, majors-ignored by policy, clean bun audit --prod.

The actions pins got the audit treatment too — every new SHA verified against its tag through the API — and one improvement over the machine's own patch: Dependabot's codeql-action PR moves only the analyze step, which would leave init and analyze on different versions inside one workflow. Both moved together.

Receipts

Every claim above is checkable: the session log carries the full gate table — engine suite 5331/0, bridge 243/0, root 2067/0, 15/15 e2e, frozen installs, audits at both roots, and a real-user smoke that round-tripped a prompt through the live bridge to the model and back (UNOWN_OK, exit 0). The changelog stop links it from the release notes.

What ships next

The queue recorded in docs/SESSION.md — the S2 ledger's head first, then the injection-scanner classifier. Security rounds like this one are what make the feature rounds safe to ship fast.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts