GEL is a GitHub Action and a command line tool that runs 24 checks (guards) on every pull request, to catch what AI coding agents and tired people ship by mistake: a leaked key, an unpinned action, a workflow with an open token, a PR that says "all tests pass" while a check is red, an edited migration, a breaking API change.
It is plain Node with no dependencies, no model and no network. It starts in report mode, so nothing blocks until you promote a guard. This repo scans in about 0.1 s.
Add it as a step in a workflow you already have:
- uses: actions/checkout@v4 # pin it to a SHA; GEL will tell you
with: { fetch-depth: 0 }
- uses: flisk-sh/gel@v1.12.0Or run it on a folder, with no GitHub involved:
git clone https://github.com/flisk-sh/gel && cd gel
node bin/gel.mjs run --root /path/to/your/repoIt prints a table with one row per guard (pass or findings, and its level), then the first findings of each guard that failed. Every guard is in report mode until you say otherwise, so the exit code stays 0.
Pin to a tag or a commit if you want a fixed version. @v1 is a moving tag that follows main, so it picks up new guards without a change on your side.
As a whole job, .github/workflows/gel.yml. Pass gel_ref as well, so the guards run from the same version as the workflow.
name: GEL
on: [pull_request, push]
permissions: { contents: read }
jobs:
gel:
permissions: { contents: read, checks: read, pull-requests: read }
uses: flisk-sh/gel/.github/workflows/gel.yml@v1.12.0
with: { gel_ref: v1.12.0 }Across an org, one pull request per repo (needs gh signed in):
node scripts/rollout.mjs --org your-org # dry run
node scripts/rollout.mjs --org your-org --applyBench on 30 public repos (11,350 files) and 231 real pull requests:
| With GEL | Their own CI today | |
|---|---|---|
| Leaked keys caught (5 formats, planted) | 150 / 150 | 1 of 30 repos scans for keys |
| Unpinned actions caught (planted) | 90 / 90 | 3 of 30 repos check pins |
Open workflow token caught (planted: no permissions, write-all, fork head under pull_request_target) |
90 / 90 | 3 of 30 repos audit workflow permissions |
| Broken repo links caught (planted) | 90 / 90 | 1 of 30 repos checks links |
| PRs that said green while a check was red | 35 of 231 flagged | 0 of 30 repos check this |
| Findings on clean repos | secrets 0; workflow-permissions 45 in 18 of 30 repos, 5 of 5 read by hand real; dead links 16 of 21 not real |
Planted defects are ones the bench inserted itself, so "caught" measures recall on known cases, not on the wild. Method, limits and what GEL misses: bench/README.md. Raw numbers: bench/results.json.
| Guard | Fails when |
|---|---|
secret-scan |
an AWS, GitHub, Anthropic, OpenAI, Stripe live or Slack key, or a private key block, is in a file |
pinned-actions |
a workflow uses another org's action by tag or branch instead of a commit SHA |
workflow-permissions |
a job has no permissions, uses write-all, or checks out a fork's code under pull_request_target |
workflow-timeouts |
a workflow job that runs steps has no timeout-minutes, so a hang runs for GitHub's 360-minute default |
frozen-install |
a workflow installs without a frozen form (npm ci, --frozen-lockfile), or package.json has no lockfile |
public-env-secrets |
a secret-named variable has a browser-bundled prefix (VITE_, NEXT_PUBLIC_, and others) |
dead-links |
a markdown or HTML link points to a file that is not in the repo |
done-without-green |
a PR says "all checks pass" while a check on its last commit is red |
migrations-append-only |
a PR edits, renames or deletes a migration already on the base branch |
api-contract |
a PR changes an OpenAPI spec so an existing client breaks |
tests-with-change |
a PR changes product code and no test file |
stale-base |
a PR branch was cut before the base changed its CI or scoring files |
release-integrity |
a release has a non-SemVer tag, no CHANGELOG entry, a red check or assets without checksums |
repo-health |
the repo has no LICENSE, SECURITY.md, CONTRIBUTING.md or CODEOWNERS |
dependabot-config |
an ecosystem in use is not covered by Dependabot or Renovate |
adr-required, status-declared, north-star, product-brief |
opt-in: you declare a rule or a number in docs/gel.json, and GEL checks it |
five-second-test, input-in-viewport, agent-readiness, ice-score, click-map |
a landing page, a phone input, agent-discovery files, an ICE table or a measured click map does not hold up |
queue-merge-group-parity, queue-cost-budget, queue-main-duplicates, queue-required-satisfiable |
a merge queue would stall: a required check with no merge_group trigger, a shared concurrency group, a matrix over budget, a duplicate run on main, a ruleset rule nothing can satisfy. gel queue-ready scores all twelve items: docs/queue.md |
gel grade grades click-map, ice-score and product-brief as bad, passed, good, high or super high. e2e/README.md gives the thresholds and shows which grades have a real example.
The exact rules, options and edge cases for all 28 guards are in docs/guards.md. migrations-append-only, api-contract, tests-with-change, adr-required and stale-base compare a PR with its base, so a push scan passes them.
Optional. GEL runs on any repo without a config file, with every guard in report mode. To tune a guard or promote it to block, add docs/gel.json:
{
"mode": "report",
"guards": { "secret-scan": "block", "dead-links": "report" },
"trusted_owners": ["your-org"],
"ignore": ["vendor/**"]
}reportwrites findings and never fails the job.blockfails it. Promote a guard once its run records show it was right.trusted_ownersmay use their actions by tag. Your own org is always trusted.- A top-level
"mode": "report"keeps every guard advisory, whatever the per-guard level says.
Each run appends one JSON line to gel-record.jsonl: repo, commit, lane, each guard's result, its first 10 findings and the time. That record is how you decide to promote a guard.
GEL picks a risk lane from the changed files, never from a label: A docs and tests, B product code, C workflows, docs/gel.json, auth, payment, secrets, migrations, Dockerfiles and deploy config, where a person approves.
The lane also sets how much CI runs. gel lane --base <sha> prints the plan and writes it to the job outputs, and the shared Flisk CI (flisk-sh/.github) follows it:
| Lane | Package scripts | Python | CodeQL |
|---|---|---|---|
| A docs, tests, copy | lint, typecheck (and test when a test file changed) | pytest only when a test changed | skipped |
| B product code | lint, typecheck, test, build | pytest | yes |
| C sensitive paths | B plus test:integration, test:e2e, integration, e2e | pytest | yes |
A push to the default branch, the weekly run and a manual run have no diff to judge, so they always run lane C. "minLane": "B" (or "C") in docs/gel.json sets a floor for a repo; a change can raise the lane, never lower it.
Each guard has guards/<name>/cases.json, labelled good and bad. node bin/gel.mjs selftest must catch every bad case and pass every good one, or GEL's own CI fails. Every false alarm the bench found became a case first, then a fix.
- It does not replace a code scanner, dependency review or a test suite. It checks the repo and PR facts those tools do not look at.
- It reads files and PR metadata. It does not run your code, and it cannot judge whether a change is correct.
- It never promotes a guard to
blockon its own. A person editsdocs/gel.json, and that edit is lane C. - The five-second and phone-input guards read a committed browser measurement, because layout cannot be read from source.
docs/loops.md covers the parts that run around the guards: a scheduled north star measurement, per-guard precision and promote/demote advice, nightly fix PRs, the factory metrics, the Cilytics org view, and the release gate and deploy health workflows in flisk-sh/.github.
Version v1.12.0, built and run by flisk-sh on its own repos. It is not published to npm (package.json has "private": true); use it as an Action or from a clone. Releases: GitHub releases.
MIT. See LICENSE.