This is being rolled out. Until it is turned on for your organisation, Deploy waits for tests alone does not qualify a website for Autopilot, and the row described here does not appear. A valid attestation or a current required-check acknowledgement can qualify it separately. The rest of this section describes how it works once it is turned on for your organisation.
Some websites run their tests after a merge, on the production branch, and the deploy job waits for them. Sentinel shows this as Deploy waits for tests (see Automated tests in CI). A failing test stops the deploy, so it does not reach the live site. Autopilot normally needs tests that run on the pull request, before the merge. With this option, Autopilot may merge first and rely on the deploy waiting for the tests.
Where it is. Open the enrolled website's Patching tab. The Autopilot card has a Merge before the tests run row, including on Copilot websites. It shows only when the result is Deploy waits for tests, the deploy method is Sentinel merges, my pipeline deploys, and Deploy workflow or check matches the assessed workflow or deploy job (on GitLab, the deploy job). If the website is not enrolled yet, enrol it in Copilot first.
- Off. The row says "Off. This website's deploy waits for its tests, so Autopilot can merge before they run if you agree." This agreement does not yet allow Autopilot to rely on the deploy gate.
- On. The row says "On. The deploy waits for the tests. Sentinel may open a revert pull request if they fail." It also shows who agreed and when.
- The earlier agreement no longer applies. This note shows when the Git connection, repository, production branch, Deploy workflow or check setting or the wording changed since the agreement. Autopilot can no longer rely on this agreement; someone must agree again unless the website qualifies separately through its tests or required-check acknowledgement. If the deploy no longer waits for its tests, or its method or workflow no longer matches, Autopilot cannot rely on the agreement. The row goes away, unless an agreement is still stored: then it says so and keeps Turn off.
Who can change it. Only an organisation administrator. Site managers and members see the row, with "An organisation administrator can change this." and no buttons. Turning it on needs two-factor sign-in set up and a recent second-factor code.
If the row cannot load, select Try again. A load failure is not an Off state.
Turn it on.
- Select Review on the row.
- Read the wording and the explanation that a revert is not guaranteed, the Deploy jobs that wait, the Tests they wait for and the production branch. How this works opens the knowledge base article in a new tab.
- Tick the box and select Confirm for this website. Enter your two-factor code if asked.
If the Git connection, repository, production branch, deploy workflow or wording changed while the dialog was open, nothing is recorded: read the refreshed details and tick the box again. If they cannot be loaded, select Read details again. If the deploy no longer waits for its tests, restore a qualifying setup before trying again. Each change is recorded in the Patching activity. On a Copilot website, the agreement does not change its mode: select Switch to Autopilot in the Mode card under Configuration afterwards and complete its terms and other checks.
What changes when it is on.
- When Autopilot relies on this option, it may merge before tests run and does not wait for a check if none reports on the pull request. Checks that do report must pass under the website's existing policy; a current required-check acknowledgement can allow optional checks to fail.
- For a merge that relies on this deploy gate, Sentinel watches the run on the production branch for up to 90 minutes. If the website qualifies through a test attestation or required-check acknowledgement instead, this special test-failure and revert handling does not apply. Passing tests allow the deploy to run; the deploy itself must still succeed, and the site must report the expected versions.
- When Sentinel confirms that a gated test failed and the deploy jobs were skipped, it marks the deployment as failed, names the tests and pauses Autopilot. That run did not change the live site. Further pushes cannot deploy while their tests keep failing.
- Sentinel tries to open a revert pull request (a merge request on GitLab) only when the same tests passed on the preceding commit, its merge is still the latest commit on the production branch, and the agreement is still current. A person must merge the revert; Sentinel never merges it.
- No revert is opened when other commits landed, earlier tests failed or their result is unknown, the test run cannot be read in full, deploy jobs did not skip, the failure was elsewhere, or the agreement or feature is no longer available. An existing revert branch, a conflict or a Git host error can also prevent it. The deployment shows the available explanation, which may be general. Check the run, fix the branch and select Resume if Autopilot is paused.
- Cancelled, unreadable or inconclusive runs do not always fail the deployment or pause Autopilot. Sentinel can show Deploy workflow not confirmed and continue waiting for the site's report; it opens no revert for that outcome.
Turn it off. An organisation administrator selects Turn off. It takes effect at once: subsequent merge checks cannot rely on this agreement. Any separate test attestation or required-check acknowledgement still applies. No second-factor code is needed.
Turning this on does not change the other Autopilot limits, such as the Autopilot terms, required checks, a missing review, a connector that has stopped reporting, or another deployment running.
Automated tests in CI
Updates and deployments are offered only to a site whose CI/CD pipeline runs automated tests on every pull request and fails when those tests fail. A pipeline that only lints, or that marks failing tests as allowed failures, does not qualify. The minimum we expect for a Drupal site: composer validate; composer install succeeding against the production platform; PHP lint of the custom code; the project's PHPUnit or Behat suite where one exists; and, recommended, a smoke test of key pages on a review environment. Sentinel does not write or maintain your tests.
The Automated tests in CI item on the site's Patching tab shows what Sentinel found in the last 20 pull requests of your production branch:
- Adequate: a test suite (PHPUnit, Behat, Playwright, Cypress, Jest, Vitest, pytest or
composer validate) ran on the pull requests and a failing test made a pull request fail, or every test check is required on the branch. - Lint only: checks ran, but none of them was a test suite.
- Tests allowed to fail: a test suite ran, but a failure did not fail the pull request.
- No CI: no check or pipeline ran on the recent pull requests.
- Deploy waits for tests (GitHub and GitLab): tests do not block pull requests, but a workflow or pipeline runs them on every push to the production branch and its deploy jobs wait for them, confirmed on the last 20 push runs. The item names the workflow, deploy job and test jobs. It counts for running updates and for deploying only while the site's deployment is Sentinel merges, my pipeline deploys and its Deploy workflow or check names that workflow or its deploy job (on GitLab, the deploy job), and for Autopilot only after an organisation administrator has let it merge before the tests run, once that is turned on for your organisation (see Autopilot when your deploy waits for tests).
- Not assessed: Sentinel could not read the checks (for example the Git host connection lacks permission to read them), there were no pull requests to learn from, or the tests never failed in the sample so nothing shows that a failure would block. This is also what a site shows before its first assessment.
Until the item is Adequate or attested (or Deploy waits for tests for runs and deployments through your pipeline), deployments and Autopilot are refused (a website that waits for its tests can have an administrator's agreement for Autopilot, see Autopilot when your deploy waits for tests), and the refusal repeats this rule with what was detected. Once required checks are turned on for your organisation, a current acknowledgement also lifts this for the website (see Required checks). Administrators and site managers can select Assess again once a repository is chosen on an active Git host connection; until Git host support is released it answers that assessing is not available yet. Members see the result and no buttons. If an assessment has run and Sentinel still reports Not assessed, or Deploy waits for tests where the pull requests did run tests (for example your pull request tests run in a system Sentinel cannot see, or they never failed in the sample), an organisation administrator or site manager can select Attest that tests run and give a reason of at least 10 characters. The button is not offered before the first assessment, nor for Lint only, Tests allowed to fail or No CI. The attestation is recorded with the time and your reason and shown on this tab. A later assessment with any result other than Not assessed or Deploy waits for tests removes it, and changing the site's Git host connection, repository or production branch clears the assessment and the attestation, so the new pipeline is assessed afresh. Withdraw attestation removes it yourself. Sentinel does not write or run your tests: fix the pipeline in your repository, then assess again. The public guide "Preparing a Drupal repository for automated updates" will describe a suitable pipeline when it is published.
When you have tests but the item is not Adequate. The usual causes:
- Tests run only after a merge. The workflow runs on a push to the branch, not on pull requests. CI runs on pull requests can still pass, because any check counts there, such as a dependency scan. Add the pull request trigger for the production branch.
- Tests are not required. They ran and always passed, so nothing shows a failure would block. Make every test check, including each shard, a required check on the production branch (on GitLab, Pipelines must succeed). Use the names exactly as the pull request shows them.
- Only a summary job is required. A job such as
testthat waits for the real test jobs is not recognised as standing for them. Also require the test jobs themselves, using their names exactly as the pull request shows them. - A generic job name. A job called
CIcounts only when a step names the suite and Sentinel can read the steps. - Allowed to fail.
continue-on-errororallow_failureon a test job, or a pull request merged past a failed test, gives Tests allowed to fail. - Skipped on some pull requests. Path filters or conditions that skip the tests leave the item Not assessed.
- Old pull requests. The sample is the last 20 pull requests, so a fixed pipeline can stay Not assessed until 20 newer pull requests replace the old ones. You can attest in the meantime.
- Your deployment waits for the tests. On GitHub and GitLab this shows as Deploy waits for tests, which lets updates and deployments through your own pipeline run. It is not enough for Autopilot on its own, because Autopilot would merge before the tests run. Either run the tests on pull requests as well and make them required, or, once it is turned on for your organisation, have an organisation administrator let Autopilot merge before the tests run (see Autopilot when your deploy waits for tests). Do not attest for this setup: the attestation says tests run on every pull request.
Configuration. Administrators and site managers can change:
- Production branch (live site, security updates), Branch for maintenance updates, Security branch prefix and Maintenance branch prefix.
- The path to
composer.jsoninside the repository. - Update kinds: security, maintenance, or both (at least one).
- Data residency: the country where this site's builds run (see Data residency). A change applies from the next job.
The mode is not in the Edit form. Saving the form keeps the stored mode; the Mode card changes it (see below).
Choose Save. A branch name, prefix or path that is not valid is refused with a plain message. Security update pull requests go into the production branch and maintenance updates into the branch for maintenance updates; Security branch prefix (default release/) and Maintenance branch prefix (default maintenance/) set how the branches Sentinel creates begin. A prefix must end with / and must not clash with either target branch. The repository is shown as Not chosen yet until you choose one; you never type it.
Mode. The Mode card is the first card in Configuration and is always shown. It shows the Copilot or Autopilot pill and a sentence about what that mode does. Only the Mode card switches mode; the Edit form does not.
- On Copilot, for organisation administrators and site managers. A panel reads Autopilot is available for this site or Autopilot is not available yet. It says what Autopilot does (it prepares each update on your schedule, merges, deploys and checks the live site, and stops and asks a person at the first failed check, or when a required check fails once required checks are on) and lists what it needs. The first line names how the site meets the tests rule: Tests in CI that block a bad merge when Sentinel found them, or your attestation, your required checks acknowledgement or your deploy gate agreement. While none applies, it has a Set up CI tests link that opens Readiness at the CI tests panel, and the CI tests rule. The second line, One update deployed through Sentinel by a person, appears only while Sentinel asks for it before Autopilot, with a Deploy one from Runs link when that is missing: press Deploy on an update pull request there (the switch waits for it). It is left out when the site has more deployments than the list shows, because Sentinel checks the whole history when you switch. Where merging and deploying without a person is not turned on yet, the card says Autopilot runs stop at the pull request, as in Copilot. Switching changes only the mode: the schedules stay as saved. While the Configuration form is open, the switch waits until you save or cancel it. On a site set to I merge and deploy myself, the card says that Sentinel only opens pull requests in either mode. Switch to Autopilot stays disabled while the CI tests are missing. Once required checks are turned on for your organisation, Review required checks can unlock it (see Required checks).
- Switching. Select Switch to Autopilot. If the website's required checks need an acknowledgement, that panel opens first. Then the change is saved, and if the website has no current Autopilot terms, the Autopilot terms dialog opens. Anyone who can switch the mode (an organisation administrator or site manager) can accept them; the change is then saved again with the latest saved settings. Any other refusal is shown in red inside the card, for example "Deploy one update through Sentinel on this website yourself before turning on Autopilot", which applies where Autopilot also needs one deployment run by a person. A recent second-factor code is asked for as for any switch to Autopilot.
- On Autopilot. Administrators and site managers see Switch to Copilot.
- Other members see the mode and its sentence only, with no buttons.
- From the Enrolled card. While a website is on Copilot, administrators and site managers also see Switch to Autopilot · merges and deploys for you in the Enrolled card's header. It opens Configuration at the Mode card.
Repository. An administrator or site manager selects Choose repository (or Change), picks a Git connection, searches the list and picks a repository. The list is cached for five minutes; Refresh list fetches it again, at most once every 15 seconds for each Git connection (in between, the cached list is shown). Each repository can belong to one site, across every Sentinel organisation: a repository another site uses is greyed out with no Choose button (see Choosing a repository under Git hosts). Your site's own repository shows Current and Check again. Choosing a repository checks it and shows a summary: the default branch, whether the production branch exists, branch protection (required checks, required reviews, and whether the Sentinel App can bypass them), whether continuous integration was seen on recent pull requests, whether composer.json is present, and anything Sentinel could not read. If that choice fails, the same reason as on the enrolment Repository step is shown, with Try again and Choose another. Members see the repository but not the controls. My repository is not listed is also on this tab.
Integration requested. After you send a request for a repository host or hosting platform, the matching step on the checklist shows "Integration requested: ticket SNTL-n, typical turnaround one week" until Support closes the ticket. The ticket link opens for the member who sent the request; a colleague sees the reference and "Raised by another member of your organisation", because a ticket is visible to the person who raised it only. When the request is accepted it adds "Accepted: the plan is on the ticket".
Who deploys updates. The Deployment panel in the Configuration section of the Patching tab (the address ends &view=configuration#hosting) asks "How do updates reach production?" and has three choices. Each saves as soon as you pick it. The enrolment wizard's Deployment step offers the same choice.
- Sentinel merges, my pipeline deploys. Your CI, Quant Cloud GitHub Actions, a Platform.sh Git push, or a deploy script on your own server. No hosting connection is needed. Sentinel merges the pull request at the commit you reviewed, and your pipeline does the deploying. See Deploying through your own pipeline below.
- Sentinel merges and follows my hosting platform. A Platform.sh or Quant Cloud connection. It takes a backup before each deployment and waits for it to finish before merging. Connect a hosting platform and map the production environment (see below). After the pull request passes, you press Deploy (Copilot) or Autopilot merges and deploys.
- I merge and deploy myself. No hosting connection is needed. Every update ends at the pull request. You merge it and deploy it on your Git host and hosting the way you usually do. Pull request cards show a sentence instead of Deploy to production. Autopilot still runs on its schedule and opens pull requests, but never merges or deploys.
Organisation administrators and site managers can change the choice or the deploy workflow; other members see them but not the controls. Switching between choices keeps any environment you mapped before and any workflow you named.
Deploying through your own pipeline. Choose Sentinel merges, my pipeline deploys when a pipeline you run deploys the site after a merge. It can be a GitHub Actions deploy workflow, Quant Cloud's GitHub Actions, a Platform.sh Git push, or a deploy script on your own server. The panel shows an optional field, Deploy workflow or check, with a Save button.
- What to enter. The workflow file (for example
.github/workflows/deploy.yml), the workflow name or a check name. On GitLab, enter a job name. The field suggests the workflows it finds in the repository and says how many it found. Sentinel follows the run of that workflow on the merge commit. - When to leave it empty. Leave it empty when a push deploys with no workflow, as on Platform.sh. The site's connector then confirms the deployment on its own. On GitLab, an empty field follows the whole pipeline.
- Tip. End your deploy script with
drush sentinel_connector:deploy. The site then reports the new versions within minutes instead of at the next cron run. - No backup. Sentinel takes no backup for a pipeline deployment. The panel, the deployment terms dialog and the deployment card all say so ("No backup is taken: this site deploys through its own pipeline."). Take your own backup in your pipeline if you need one.
- Already running. Before it merges, Sentinel checks that the deploy workflow is not already running on the production branch. If it is, nothing is merged and the deployment fails with "A deployment is already running". Try again when the run has finished.
- No access to your server. A pipeline deployment needs no SSH key or other access to your server. Sentinel learns what happened only from your Git host (the GitHub App or GitLab.com) and from the connector's own signed report. If you connect a hosting platform, Sentinel stores that platform's API token, encrypted. A site that moved to a pipeline keeps its earlier connection until you remove it.
- Autopilot. Autopilot merges for a pipeline site just as it does for a site that follows a hosting platform.
After the merge, the deployment card shows three steps:
- Merged, with the merge commit and the note that no backup was taken.
- Deploy workflow, with a link to the run.
- Waiting for the site to report the new versions, which becomes Verified on site when the connector reports every expected version.
What the card can say:
- Failed. The deploy workflow failed. To roll back, revert the merge commit and let your pipeline deploy the revert, or restore your own backup.
- Deployed, awaiting site confirmation. The workflow succeeded. Sentinel now waits for the site's report, until about 24 hours after the deployment. The card shows that time, says to run
drush sentinel_connector:deploy --forceon the deployed Drupal environment (not in CI) to report sooner, and links Open this site's Connector log. - Deploy workflow not confirmed. The workflow did not start within 15 minutes, ran for more than 90 minutes (Sentinel cancels nothing), was cancelled or skipped, or GitHub could not be read. Sentinel still waits for the site's report, so the deployment can still end as Verified on site.
- Verified on site. The connector reported the expected versions.
- Deployed, not confirmed. No matching report arrived within 24 hours of the merge. Organisation administrators get a notice. A later report cannot confirm this deployment, but it still updates the site's versions in Sentinel. Check the deployment reached the site, then run
drush sentinel_connector:deploy --forceon the deployed Drupal environment. Check the site and its connector.
Production environment. Only an organisation administrator can map the environment. They select Map environment (or Change), chooses a hosting connection, selects List environments, picks one, ticks the box to confirm it is the production environment, and saves. For Quant Cloud the form also asks for the branch your GitHub Actions deploy workflow triggers on (it starts as the site's production branch) and, optionally, the workflow file. The environment's branch must match the site's Production branch. If it does not, nothing is saved and the message shows both branch names; fix the branch, then try again. A saved mapping completes the hosting step in the checklist. This applies when Sentinel deploys through your hosting platform. Everyone else, including site managers, sees the environment but not the controls. Changing the mapping needs a recent second-factor code if the administrator has set one up.
