ActionLens reads untrusted input: a workflow pasted from the internet, possibly crafted to make a
parser work too hard or to smuggle a payload. It is not a sandbox for GitHub Actions — it never runs a
run: script, never starts a container, never resolves hashFiles() against your files, and never
evaluates expressions as code. A workflow containing curl … | sh is analysed, not obeyed
(ADR 0001).
Nothing leaves the browser. There is no backend, no telemetry, no analytics, no CDN font, script or stylesheet, and no GitHub API call (ADR 0003). Sharing and autosave are not implemented; if they are built, they will need this ADR reopened first (see its "Deferred" table). At present workflow text and event payloads live only in memory.
The local-first promise is enforced rather than asserted: LocalFirstGuardTests is a test in
ActionLens.Core.Tests (not a build step) that fails when the UI sources, the Node scripts or a
stylesheet construct a network client or reference an asset on another host. It matches code, not prose,
and both halves of it have their own self-test so a scan that silently matches nothing cannot pass.
The browser-level assertion is a Gate 2 Playwright case, still outstanding; the source guard alone
cannot prove that no request left the page.
Exceeding a bound produces a positioned diagnostic, never an exception:
| Bound | Value | Diagnostic |
|---|---|---|
| Workflow size | 1 MB | YAML006 |
| Event payload size | 1 MB | YAML006 (Scenario stage) |
| Anchor alias references | 128 | YAML004 |
| Expanded YAML nodes (aliases followed) | 200,000, max depth 64 | YAML009 |
| Nesting depth (token pass) | 64 | YAML005 |
| YAML tokens | 200,000 | YAML006 |
| Documents per file | 1 | YAML002 |
| Expression body length | 21,000 | EXPR002 |
| Expression parse depth | 50 | EXPR002 |
| Matrix combinations | 256 (truncated) | MTX001 |
| Matrix combinations materialised | 4,096 (refused above) | MTX001 |
The alias-reference and depth limits are checked on the raw event stream before an object graph is
built. A separate bounded walk of the loaded graph then counts every node visit with aliases followed —
counting distinct node objects would miss a bomb entirely, because aliases share one object — and rejects
the file with YAML009 before any recursive model walker sees it. The matrix product is likewise computed
arithmetically and refused before a single combination object is allocated; the 256 ceiling is applied
after exclude/include to the combinations that remain. Duplicate mapping keys are collected in the
token pass, so all of them are reported rather than only the first one a loader trips over.
Length and depth limits on expressions are not published by GitHub; they are taken from GitHub's own reference implementations, which agree, because a workflow GitHub would refuse must not be analysed as if it were fine.
ActionLens does not handle secrets, but you may paste them: a workflow's secrets.MY_TOKEN is a name,
and names are not sensitive. The risk is your own pasted text — event payloads and workflow bodies stay
in your browser's memory only; there is no persistence yet. Closing the tab clears them. Do not paste a
real secret value into the event payload editor to see what a condition would do; it is not needed,
because secrets.* is unknown by design and unavailable in if: regardless (rule AL004).
- Static analysis is not a sandbox. The engine is ordinary managed code: a crafted input that causes an unhandled exception or excessive memory is a bug to report, and the bounds above are the mitigation. One such bug was found and fixed during development — a resource-embedding glob that silently matched nothing and crashed the app on boot. That class is only visible in a browser, which is why a UI change is not considered verified until it has been loaded in one (see w4-verification.md).
- A wrong answer is a security problem here. A tool that reports "this job will not run" when GitHub
would run it teaches people to trust the wrong thing. Findings that cannot be established are therefore
reported as
unknownwith a reason, and the product promise is thatunknownis never collapsed into a definite result. - Third-party code is vendored as a package, not as source.
Atlas.Blazoris pinned to the published0.4.0(ADR 0004); the project does not track that library's development branch. GitHub's expression test data is vendored under its MIT licence with attribution intests/golden/languageservices/.
Report vulnerabilities privately. The public repository does not exist yet (Gate 2 must pass first), so
until it opens, contact the maintainers directly rather than filing a public issue. Once the repository
exists, use its security advisories (HestiaLab/ActionLens → Security). Do not open a public issue for a
reportable problem. Expect acknowledgement within five working days; fixes ship in a patch release with a
CHANGELOG.md entry that describes the impact without disclosing the exploit path.
Anything that would make the tool mislead rather than crash — a silent unknown turned into a definite
answer, a finding pointing at the wrong line, a network request the guard failed to catch — counts as a
security-relevant defect here, because misleading the reader is the failure mode this product exists to
avoid.