Changelog
What changed, and what it changed about detection.
Every release, with its accuracy figures tied to the corpus they came from and the limits still in place when it shipped. Install with npx myrqen@latest.
A detection figure is always quoted with the corpus it came from. The development corpus was tuned against and reads 1.000, which says nothing about unseen code. The benchmark page explains which corpora mean what, and why one number alone misleads however accurate it is.
0.3.02026-10-14Current
Reports that open with a verdict, a fix loop that branches, commits and retests, Python for FastAPI and Flask, data flow that crosses modules where a call resolves, and dependency advisories reported only where your code reaches them.
Detection
- As measured on 2026-10-14: on the newest blind JavaScript/TypeScript corpus (holdout-7, CWE-matched) recall is 0.250 and 4 of 32 deliberately-correct cases drew a finding. Against a gate of 0.90, the static pass alone.
- As measured on 2026-10-14: on holdout-3, after the fixes it prompted in 0.2.0, recall is 0.500 and precision 1.000. That corpus is burned, so this is not a generalisation figure.
- As measured on 2026-10-14: Python (FastAPI and Flask), scored once on a blind corpus (holdout-python-2) before this release's last round of Python fixes: recall 0.774, precision 0.923. Below the 0.95 precision bar; a new blind Python measurement follows in 0.3.x.
- JavaScript and TypeScript data flow now crosses modules where a call resolves, up to four calls deep, with zod/yup/joi/valibot schemas and framework-decorated parameters understood. On the corpora this moved one case: a holdout-2 detection now credited to the right route.
- Twelve new rule shapes: DOM XSS, client and Location-header open redirects, HTML responses reflecting request data, whole-body mass assignment, JWT misconfiguration, credentialed any-origin CORS, unthrottled credential routes, prototype-pollution paths, interpreter-argument command injection, unsafe deserialization libraries, and NoSQL aggregation injection.
- Dependency advisories from an offline OSV/GitHub snapshot (915 advisories across 128 watched packages), resolved from your lockfile, and raised as a finding only where your code calls an affected export.
Reports
- Every report opens with a verdict, the next step, and up to three findings to fix first, each with a paste-ready prompt for your agent.
- One source of truth for every count, checked across HTML, Markdown, JSON and SARIF. The old HTML stat cards could show 7 while the summary said 14.
- CWE, OWASP Top 10 (2021 and 2025) and ASVS references on every finding, a compliance section, and SARIF taxonomies.
- `.myrqen/baseline.json`: findings are marked new or already known by a stable identity that survives moved lines.
CLI
- `myrqen fix <id>` refuses a dirty tree, branches `myrqen/fix-<id>`, applies a mechanical fix or hands your agent a plan, and `myrqen retest <id>` re-runs only that finding's check and commits when it passes. Nothing is pushed.
- A live progress view, `myrqen report`, `myrqen diff`, `myrqen watch`, and a first run that detects how to start your app and remembers it in `.myrqen/config.json`.
- A run with no agent says plainly that it is a static pass and which agent on this machine can prove the findings.
Dashboard
- Finding triage that carries over between scans of the same project: accepted risk, false positive and fixed, with assignees and comments on team workspaces.
- A plan and billing page, and checkout for paid plans.
- A rebuilt interface, the report verdict and compliance in the browser, and real 404s for unknown or forbidden reports and share links.
Website and docs
- A redesigned myrqen.cc and docs.myrqen.cc.
Not in this release
- Java, Ruby and Go are not parsed. Python routes are modelled for FastAPI and Flask only; Django, aiohttp and Tornado are parsed but not followed from their routes.
- Recall is below its 0.90 gate on every held-out corpus. A clean report is not evidence that an application is secure.
- No destructive testing, at any effort. There is no opt-in flag.
- Windows support was fixed from CI logs; the first full Windows CI run is part of this release's verification.
0.2.02026-08-21Superseded
The first release worth installing: published to npm with a README and a licence, thirteen agent integrations, and a scanner that no longer reports anything against correct code.
Detection
- As measured on 2026-08-21: precision reached 1.000 and the false-positive rate on deliberately-safe code reached 0.000 on every corpus, the blind one included — reached, because on the blind run itself they were 0.833 and 0.059 and both failed their gates. Three of the four scanner fixes in this release were precision fixes.
- As measured on 2026-08-21: recall 0.294 on the one corpus nobody had tuned against — the only clean generalisation measurement this project has — and 0.441 on that same corpus after the fixes it prompted, which burned it. Against a gate of 0.90, and both figures are the static pass alone.
- `relative()`-based containment now counts as a guard; a producing helper still counts through `.toString()` and a ternary; a CORS origin predicate anchored at a label boundary gates the reflection; and a fluent query-builder chain. Drizzle, Kysely, is recognised as a record lookup.
- Entry-point adapters added for thirteen architectures: Astro, Bun, Deno, desktop CLI, editor extension, Nest, Nitro, realtime, serverless, SvelteKit, TanStack/Solid, and Cloudflare Workers.
- Ten regression tests, each asserting that the guarded form of a pattern is silent and that its unguarded sibling is still found.
Reliability
- A failing API no longer kills a scan. `resolveBinding` called two endpoints unguarded, so a linked device plus an unreachable API exited 1 and wrote no report, even though scanning needs no network. Both now degrade to the local outcome an unlinked device gets, and name the reason.
- `--scope` on `unlink`, so removing a project integration cannot reach a global one.
Packaging
- A single self-contained bundle with no runtime dependencies, a README and a licence in the tarball, and no sourcemap. Nine files, 1.33 MB packed.
- The packed tarball is installed into an empty project and taken through a full link and unlink round trip under npm, pnpm, and Bun before every publish.
The hosted service
- Deployed and verified end to end against production: signup, browser-approved device link, a scan, sync, and finish.
- An unauthenticated device approval was refused 401. A non-member reading a report got 404 rather than 403, so they do not learn it exists. A principal-scoped share showed an outsider exactly 2 of 14 findings with nothing withheld leaking. Both planted canary secrets appeared in neither the cloud copy nor the local report.
Privacy
- `myrqen config telemetry off` turns product analytics off persistently. The setting was already honoured, but the only ways to change it were an environment variable on every command or editing a JSON file by hand.
Not in this release
- Paid plans cannot be bought, checkout is not configured on the hosted deployment.
- Python, Java, Ruby, and Go are not parsed at all.
- No destructive testing, at any effort. There is no opt-in flag.
- No container image has been built and verified.
0.2.0-rc.12026-08-21Superseded
The release candidate for 0.2.0, published to the `next` dist-tag so a plain `npm install myrqen` would not pick it up.
Detection
- Functionally identical to 0.2.0.
Documentation
- Its README stated that the hosted service was not deployed. That was true when it was written and false a few hours later, which is why the site now has a check that compares its own prose to the product.
0.1.02026-08-18Deprecated
Published from a tree that predates the packaging work and the entry-point adapters. Deprecated on npm. Do not install it.
Detection
- Worse than 0.2.0 on every corpus, because it predates the entry-point adapters: without an adapter the engine cannot see a route, and without a route it cannot follow anything to it.
Why it was deprecated
- It shipped no README and no licence, declared no licence or repository in its manifest, and included its own sourcemap.
Versions are recorded here and in CHANGELOG.md, and CI checks the two against each other. The published tarballs and their integrity hashes are on npm.