What the agent does
The agent starts by reading what the repository already checks on a pull request, because a dependency review should fill the gap rather than duplicate a job. In a repository like actions/typescript-action there are five workflows under .github/workflows: ci.yml runs npm ci, npm run format:check, npm run lint and npm run ci-test on pull_request and push to main; check-dist.yml rebuilds dist/ with npm run bundle and fails on a stale bundle; linter.yml runs super-linter/super-linter/slim@v8; codeql-analysis.yml scans the source; and licensed.yml runs licensee/licensed against the cached metadata in .licenses/npm, but only on workflow_dispatch, because its pull_request and push triggers are commented out. None of them looks at what a pull request adds to package-lock.json, so a Dependabot bump that pulls in a package with a published advisory, or a contributor adding a GPL-licensed dependency, passes every check.
It then reads the policy the repository already states so the new check agrees with it. .licensed.yml allows apache-2.0, bsd-2-clause, bsd-3-clause, isc, mit, cc0-1.0 and other, and ignores @rollup/rollup-linux-x64-gnu. .github/dependabot.yml opens weekly grouped pull requests for npm and github-actions, which is where most dependency changes will arrive. package.json declares one runtime dependency, @actions/core, and a longer list of development dependencies for Jest, Rollup, ESLint and Prettier; since rollup.config.ts bundles everything into dist/index.js, the agent notes that a development dependency never ships to users of the action, which decides the fail-on-scopes setting below.
The agent writes one new file, .github/workflows/dependency-review.yml, triggered on pull_request like the other checks, with permissions: contents: read to match ci.yml, a checkout step on actions/checkout@v6 as the other workflows pin it, and actions/dependency-review-action@v5. It sets fail-on-severity: moderate so the check blocks on moderate, high and critical advisories and lets low ones through with a warning, fail-on-scopes: runtime so a vulnerable test runner does not block a release while an advisory in @actions/core does, and allow-licenses to the same SPDX list .licensed.yml allows, minus other, which is not an SPDX identifier, so the two license checks cannot disagree. It leaves comment-summary-in-pr at its default, never, which is why pull-requests: write is not needed; a FAQ below covers turning the comment on. The agent does not touch the five existing workflows or the Dependabot configuration.
Before pushing, the agent runs actionlint over the new file, using the repository's own actionlint.yml, which ignores the node24 runner-name warning, and runs the super-linter's YAML rules locally so linter.yml will not fail on indentation. It pushes the branch and opens the pull request, which is itself a pull_request event, so the new workflow runs on its own pull request alongside the existing checks. Because that pull request changes no dependency, the review finds nothing to flag and passes; the agent quotes the job summary in the description and says plainly that the first real test is the next Dependabot pull request, which changes package-lock.json.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Add a dependency review check to this repository's pull requests. Read every workflow under .github/workflows, .github/dependabot.yml, .licensed.yml and package.json first. Write .github/workflows/dependency-review.yml on pull_request with contents: read permissions and actions/dependency-review-action, failing on moderate or higher severity in runtime scope and allowing only the SPDX licenses the repository's existing license policy allows. Do not change the existing workflows. Validate the file with actionlint, open a pull request, and in its description show the new check's result on the pull request and explain what each setting does.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
What comes back is a pull request on its own branch that adds exactly one file, .github/workflows/dependency-review.yml, around twenty lines long. Nothing under src/, dist/, __tests__/ or .licenses/ changes, so check-dist.yml and the CI job pass without rebuilding anything.
The description lists the five existing workflows and what each already covers, then the new workflow's settings with a sentence each: the severity threshold, the runtime-only scope and the reasoning from rollup.config.ts, and the license list copied from .licensed.yml with other dropped and why. It includes the actionlint output and the new check's status on the pull request itself. The transcript at GET /o/{org}/sessions/{id}/messages holds every file read and every command run.
A reviewer checks three things before merging: that the license list matches the one in .licensed.yml, that fail-on-scopes matches how the project ships (an action that bundled nothing would want development too), and that the repository has the dependency graph enabled under its security settings, which the action needs to diff the two revisions. After merging, the check appears on the next Dependabot pull request, which is where it earns its place.
Run it on every terminal hand-off
A maintainer who just read an advisory for a transitive package starts this from the terminal with this body and the repository in resources, then goes back to what they were doing. The run reads the real workflows and the real license policy, so the new check fits this repository's conventions, pinned action versions included, instead of a copied template.
Send a request_id, a UUID, in the body so a retried start returns the run it already began. When the turn ends, GET /o/{org}/sessions/{id} names the pull request. The same body sent once per repository you maintain adds the same policy everywhere, each run in its own sandbox with its own pull request, and each reading that repository's own license file.
The use case this job follows, with its automation and trigger: Hand off a long task.
Why run it on Rush?
The workflow itself is short; the judgment is in the settings, and getting them wrong is silent. A fail-on-severity of low blocks Dependabot on advisories nobody will act on, and a license list typed from memory drifts from the one the project already enforces, so the two checks disagree the first time a dual-licensed package arrives. The agent reads the policy the repository already states and derives the settings from it, writing the reason beside each one.
Running it on Rush puts the work in its own sandbox with the repository's Node version from .node-version, keeps the actionlint output and every file read in the transcript, and ends with the check visible on the pull request that adds it. Several repositories can get the same treatment at once, each from its own license file.
Frequently asked questions
Will it change my existing workflows?
No. The agent reads ci.yml, check-dist.yml, linter.yml, codeql-analysis.yml and licensed.yml to learn the conventions and to avoid duplicating a check, but the pull request adds one new file and edits none of them. Leaving licensed.yml on workflow_dispatch is a separate decision the description points out.
Why runtime scope only?
Because of how the action ships. rollup.config.ts bundles the source and its runtime dependencies into dist/index.js, which is what uses: runs; Jest, ESLint and Prettier never reach a user. An advisory in a development dependency is still reported in the job summary, it just does not fail the check. If your project ships development dependencies, say so in the prompt and the agent sets fail-on-scopes: runtime, development.
Can it comment the summary on the pull request?
Yes. Set comment-summary-in-pr: on-failure or always and add pull-requests: write to the workflow's permissions. The agent leaves it at never by default so the new workflow keeps the same contents: read footprint as ci.yml; the job summary on the Actions tab has the same information.
Does the check prove anything on the pull request that adds it?
Only that the workflow is valid and runs: with no change to package-lock.json, there is nothing to flag, and the description says so rather than claiming a catch. The first meaningful result comes from the next pull request that changes a dependency, such as the weekly Dependabot group.
What does the repository need on GitHub's side?
The dependency graph, which the action uses to diff the base and head revisions, and which GitHub enables by default on public repositories. On a private repository, turn it on under the repository's security settings before merging, or the job fails with an API error rather than a review.
More GitHub Actions jobs
More jobs like this: Hand off a long task
- Add Dependency Verification to a Gradle Build (Gradle)
- Add a GitHub Pages Deploy Workflow to an Eleventy Site (Eleventy)
- Deploy Firebase Functions from GitHub Actions (Firebase Functions)