Skip to content

Run an update

51 help pages

Once a repository is chosen, the Runs section of the site's Patching tab starts with a Run an update panel (after the drift pause notice, when one shows). It has one tile per kind of update. The info button next to the heading (What does a run do?) opens a short explanation; select it again to close it. Run security update applies the security fixes due for this site on its production branch and opens a pull request if anything changes. Run maintenance update prepares dependency updates within the repository constraints and the scope you choose; when a maintenance schedule is set, its tile shows the next scheduled start. When there is update work, the run works in a disposable container on a Sentinel runner, stops after 45 minutes and opens a pull request if it produces changes. Nothing is deployed from this panel.

Choose what a run updates. Run security update and Run maintenance update (on the Patching tiles, the Updates tab and the header) open Choose what to update instead of starting at once. Only organisation administrators and site managers see these buttons. Each package shows the version the run would move it to.

  • On a maintenance update, Patch and minor releases are ticked. On a security update, patch-release fixes under Fix on the installed major are ticked and minor-release fixes are not. New major releases (on a security update, Fix only on a new major) are not ticked; each links to its release notes. A run takes a limited number of majors (3 by default; the dialog shows the limit), so the pull request stays reviewable.
  • On a maintenance update, Other Composer libraries is on by default. Turn it off to change only the ticked packages and what Composer must move with them.
  • When the repository check confirms standalone Drupal Core, selecting it also allows its installed production companion packages, such as drupal/core-recommended, to move within the selected release step. Protected or excluded packages stay held. Selecting Core does not also select development-only Core packages.
  • When a distribution includes Core, select the distribution, such as govcms/govcms. Core has no separate checkbox and appears under Not offered with Core is managed by and the distribution name. Sentinel uses the repository dependencies to identify ownership, including distributions from other vendors. A distribution update can move its required Core packages together, within the permitted scope. Core stays held if no newer usable distribution release resolves; Sentinel does not ask the AI to guess a version or bypass the distribution constraints.
  • Not offered lists what the run will not touch and why: Drupal core keeps its major, Distributions keep their major, Needs a newer Drupal core, Set by distribution, Excluded from Patching, and the Core ownership or distribution-release reasons below.
  • Run update with N components runs the choice. Run the whole update runs everything, as before. If Sentinel refuses the choice, the message shows in the dialog. If the choices cannot be loaded, the button asks you to try again; Sentinel never starts a whole update by itself in that case.
  • The job page shows What this run was asked to update. A pull request that takes a major has the label Contains a major upgrade. Autopilot never merges or deploys it: the run pauses and you deploy it yourself with Deploy.

No verified distribution update means saved data offers no newer same-major distribution release; it does not prove none exists. Core ownership needs a repository check means ownership is unverified, so standalone Core cannot be selected. A successful repository update refreshes this information. Run the whole update checks the repository and enforces distribution ownership, but may update other packages too. A security run without a repository check does not refresh ownership. When all rows are held, their reasons remain visible and Run update with 0 components is disabled.

A request to update a package does not guarantee a version change. The pull request lists Requested packages left unchanged, including their locked versions and any known hold. If Composer resolved no version change within the repository constraints and the permitted scope, a newer release may still exist. Compare this section with Lock diff to see what actually changed. When no locked versions changed, read the job's steps and output before deciding whether another run is needed; an unchanged result alone does not mean every package is up to date. Earlier jobs may not have this unchanged-package report.

No applicable advisories. A security run with no applicable advisories succeeds with No applicable advisories to apply. No changes were needed. When there is no opted-in runtime work, it finishes without starting a runner container or AI agent and opens no pull request. No action is needed for that run. This result describes the run’s advisory workload; it does not mean the whole site is secure, mark advisories as fixed or deploy anything. Enabled runtime updates are still checked and may produce their own pull request; the summary then names that runtime pull request. A runtime failure still fails the run. Rebasing an existing pull request still processes its existing work.

Composer audit. Where Composer audit is switched on for Sentinel, a security update you start yourself, and every maintenance update, also run composer audit on your repository. This catches vulnerable Composer packages that are not Drupal projects, for example Symfony, Twig or Guzzle, in production and development packages. Sentinel fixes them as part of the update. Scheduled and Autopilot security updates do not run the audit. They still end at once when Sentinel has no drupal.org finding. When neither source has anything due, the summary reads Drupal.org advisories and Composer audit found nothing due. No changes were needed. If the audit cannot run and there is nothing else to do, the job fails instead of reporting the site clean. A maintenance update fixes security advisories too, even beyond the site's maintenance scope (for example a minor release on a patch-only schedule, or a development package), and its pull request has a Security fixes table. A run that finds nothing still uses server time and is billed like any run. Advisories found only by Composer audit show on the job page and in the pull request. They do not yet appear on the site's vulnerabilities page or in the monthly report.

What a run does:

  • A security update moves each affected package to the patch release that fixes its advisory (for example 10.3.0 to 10.3.1). A whole maintenance update considers direct and indirect dependencies within their current major version, while a selected run follows the steps you chose. Packages can remain unchanged because of repository constraints or holds. A whole update never takes a new major version (unless the schedule's scope includes majors for contributed modules, themes and libraries; Drupal core and distributions never take one) and never moves to a development release or pre-release (such as 7.10.x-dev or 1.1.0-beta1), even when your project allows that stability.
  • It checks that only the expected packages changed, re-checks your composer.json patches and runs the project's PHPUnit tests (and PHP CodeSniffer, recorded only) where the repository has them configured.
  • It pushes one commit to a branch named like release/security-20261021-3f9a1c (or maintenance/maintenance-20261021-3f9a1c; the start comes from your branch prefixes) and opens a pull request (a merge request on GitLab) labelled sentinel plus security or maintenance.
  • The description lists the advisories and versions, a table of the composer.lock changes, the checks that ran, what to do before deploying (for example run database updates) and a link to the Sentinel job. Review it on your Git host as usual.
  • When Composer cannot resolve an ordinary package update by itself, the AI agent from the site's AI key is asked to adjust composer.json or your patches, within the same limits; the description then says what the agent changed. An unresolved distribution-managed Core update stops for a person instead of asking the AI to guess how to update the distribution. Without a connected provider, the job stops at that point instead.
  • Once the update resolves, the job reads the drupal.org release notes of every release each Drupal project (core, modules and themes) moved through, oldest first. The AI agent then applies what they ask for in one round, with no approval: a replacement for a removed module or file, changes needed in your custom code (web/modules/custom, themes/custom, profiles/custom), special update instructions, and changes to exported configuration under config/. Steps that must run on the site (database updates, Drush commands, configuration import) are not run; they are listed in the pull request for you to do. The checks then run again. The pull request has a Release notes reviewed section with an Outcome line (reviewed and applied by the agent, not reviewed by an agent with the reason, no release notes could be read, or the agent's review did not finish) and, for each package, the releases read (linked to drupal.org) and any that could not be read. The job page keeps the notes as release-notes.md. Without a connected AI provider the notes are only listed. If the agent's review does not finish, its changes are undone. The pull request still opens and lists the item under Needs a person. The same happens if the agent's changes would move, downgrade or remove a package version the update chose, or move another package. Adding a replacement package is fine, except on a security update, which never adds a package; a replacement for a removed module is listed for a person instead. Role and permission configuration (user.role.*) is never changed.
  • The job also scans your custom modules, themes and profiles for deprecated and removed Drupal API, against the updated packages. It uses the same tools as Upgrade Status (PHPStan with phpstan-drupal and the deprecation rules). They are used only for the check and are never added to your project's composer.lock. The AI agent fixes what the scan finds in the same round as the release notes, then the job scans again. The pull request has a Custom code scan section with how many uses were found, how many remain after the agent's changes, and up to 30 remaining ones with file and line. Anything that remains is listed under Needs a person. The job page keeps the findings as deprecations.md, and as deprecations-after.md when the agent changed code. Maintenance updates always scan when the project has custom code. Security updates scan only when a module or core moved through more than one release. If the tools cannot be added alongside your locked packages, the scan is skipped and the pull request says why. Without a connected AI provider the findings are only listed.

A job never waits for an answer. A command that asks a question fails at once and the job records in the Activity what it did instead. The job's authority is fixed: it opens a pull request and never merges, deploys or touches production. Instructions in your repository's own files (such as AGENTS.md or CLAUDE.md) are not followed; a line asking for more, for example "merge the PR when green", is recorded in the Activity as ignored.

When one package cannot be updated within the job's authority (the fix is only in a minor or major release, or the package is protected) or does not update cleanly with the rest, the job leaves it out, delivers the rest, and the pull request gets a Needs a person section listing what was left out, why, and a Recommendation. A security fix that also moves other packages by patch releases, including development-only packages such as PHPUnit or Symfony test tools, is delivered: the pull request has an Also changed, and why section that counts the production packages (deployed) and the development-only packages (never deployed) separately. A development-only package that would move by a minor or major release still stops that fix. A maintenance update does the same for a package it will not move (a new major version, a development release or pre-release, a protected package, or a change to development dependencies when your site does not include them): the package stays at its current version (a package offered a new major version still takes the releases of its current major) and is listed under Needs a person. If the AI agent stops making progress, Sentinel stops it and asks it to plan again once; if that fails, a security update delivers the packages that update cleanly, or stops with no pull request when none do, and a maintenance update stops with no pull request. If your project's tests fail on the base branch without the update too, and the update adds no new failure, the failure is pre-existing and the security update still opens its pull request (see When tests fail on a maintenance update). If the update adds a failing test, or Sentinel cannot prove the failure was already there, the job stops as blocked and opens no pull request; the Activity says so.

Check policy and existing failures

A required check must meet the project's effective merge rules. An advisory or allowed-to-fail check can report a failure without preventing that merge. The effective rules can depend on the branch, workflow and reporting app. A check's name, a comment saying it is optional, or a green overall pipeline does not establish those rules by itself.

Sentinel's checks while preparing an update are separate from checks reported on a pull request by your Git host. Where required-check handling is enabled for your organisation, the merge decision follows the supported repository rules and any acknowledgement required for the website. Confirmed optional failures are reported as warnings; required checks and the other deployment safeguards still apply. Unreadable or incomplete rules are not evidence that no checks are required.

Current limitation: a security update can stop before opening a pull request when PHPUnit fails in Sentinel's update job, even if your Git host allows it to fail. It does not stop when the same tests fail the same way on the base branch and the update adds no new failure. The required-check acknowledgement for merging does not remove that local test stop. A maintenance update can open a pull request with a Tests failed note. This is not a way to bypass required checks or deployment safeguards.

A failure reproduced on the unchanged base branch is evidence of an existing problem. It does not prove that every failure in the update is the same, that the update caused no regression, or that the check is optional. If the baseline could not run, the comparison is unavailable. Read the recorded results rather than treating an unavailable check as passed.

If a check you intentionally allow to fail stops a security update, select Flag for review on the job page and explain which check you expected to be advisory. Support can review the available evidence. Do not weaken a required check or repeatedly run the same job just to get past the stop.