Skip to content

Limited local check-policy handling

51 help pages

A change is being prepared for a narrow set of GitHub workflows. It is not yet available. Until it is released and enabled, the current security-update stop and maintenance-update warning behaviour described above remains in place.

For a supported project once enabled:

  • Confirmed advisory PHPUnit check: a failure can allow an otherwise-valid pull request to be opened, with a failed-test warning and the available comparison with the unchanged base. The check result stays failed; it is not recorded as passed.
  • Confirmed required PHPUnit check: a failure stops both security and maintenance pull-request delivery, even if the check also fails on the base. An earlier confirmed requirement is not waived because a later policy reading expires.
  • Unknown policy: missing, stale, changed, unreadable or unsupported policy does not make a failure advisory. Security updates retain the local test stop; maintenance updates retain their existing warning behaviour.

The policy is checked again after the baseline comparison and before delivery. Initial support covers only literal JSON-formatted workflow files for pull requests, with independent Ubuntu jobs and the supported checkout, Composer-install and PHPUnit steps. Ordinary YAML, reusable or dynamic workflows, matrices, wrapper scripts, unknown actions and ambiguous check identities remain unknown. Do not weaken your tests or change your pipeline just to make the policy recognisable.

This concerns preparing a pull request only. It does not change readiness requirements, give a required-check acknowledgement, authorise a merge or deployment, or remove their separate safeguards. Failed local tests still stop Autopilot. Review warnings and any items needing a person before deciding what to deploy.

When tests fail on a maintenance update (current behaviour). While limited local check-policy handling is unavailable, if PHPUnit fails or times out on a maintenance job, the pull request is still opened so you can see the update. Its description starts with a Tests failed note, which also says whether the tests fail on your production branch too. The job summary ends with "; tests failed", and the job page and the site's Activity show a row "The project's tests (phpunit) failed with this update" with a link to the pull request. Sentinel emails "Tests failed on the maintenance update for " to the person who added the site, the organisation's owner and every organisation administrator who is an active member. Review the failures before merging. A security update still stops when the update adds a failing test, with no pull request.

Tests that already fail without the update. Sentinel runs your PHPUnit tests on the update and, when they fail, again on the base branch without the update. That run has no database or other services, so tests that need them fail there too. When the same test cases fail on the base branch with the same failure messages, the same or more tests ran, PHPUnit printed its summary, and the update adds no new failure, the failure is pre-existing. Then no "Tests failed" email is sent, a security update is not stopped and opens its pull request. Autopilot still holds it: its terms say it never merges when tests failed in Sentinel's own update job. The pull request description starts with Tests already failing, the job page shows Tests already fail without the update, and the job summary ends "; tests already fail without the update". Anything Sentinel cannot prove is pre-existing keeps the rules above: a new failing test, a failure for another reason, a time-out, a base branch run that could not finish or be compared, or a report from an older runner. The email is sent, a security update stops, and Autopilot holds. New failing tests are named (a test that uses a data set is shown only as "(a data set)"). The pull request lists the tests that fail on both runs (up to 20). When your site's production branch is not the pull request's base branch, the base branch run says nothing about production, so the failure is not treated as pre-existing. These tests are a different matter from a failing check on the production branch, which holds deployment (see Checks failing on the production branch).

A run that cannot be done within its scope ends as Failed and the job summary gives the reason, for example when the fix exists only in a later minor release (run a maintenance update first), when the installed release already includes the fix, when an existing patch no longer applies, or when the tests of a security update fail. No pull request is opened then. The job page keeps the composer.lock changes, the pull request text and, when the agent ran, its transcript.

  • Organisation administrators and site managers can run or stop an update. Other members see the Runs table and a note that only an administrator or site manager can run one.
  • One job runs at a time for a site, so the buttons are hidden while a job is queued or running. A job can also wait while your organisation's allowance of running jobs is full.
  • A queued job shows Position N in the queue, estimated start in about M minutes in the Run update panel (should start shortly when the estimate is under a minute), and Queued, position N, estimated start in about M minutes in the status card at the top of the job page. On the job page, the same Waiting m:ss timer shows how long the job has been waiting (see Patching jobs and activity). Any member who can see the site's Patching tab or open the job page sees it. The estimate comes from the jobs ahead, the free runner slots and how long jobs took over the last 7 days (15 minutes when there is no history), so it is approximate. Other organisations' jobs count towards your position but are never shown. No estimate is shown while the job is held (the site already has a running job, or your organisation's allowance of running jobs is full) or while it is being cancelled; the panel says why it waits instead. A job also waits when another job is running on the same AI key and could still use up the rest of the key's monthly spend cap; the job list says "Position N in the queue; waits for another job on the same AI key", the job overview says "Waiting for another job on the same AI key to finish, so the two together stay under the key's monthly spend cap". If a runner had already picked the job up, the Activity also shows a Waiting for another job on this AI key to finish row with the month's spend and the cap. It starts with its agent when that job finishes and keeps its place. Jobs you start yourself (Run update, Rebase with Sentinel, Run security update and Run maintenance update on the Updates tab) go before scheduled runs, scheduled security before scheduled maintenance, and a scheduled run that has waited 2 hours moves up. When the platform queue is very busy, a scheduled maintenance run waits for a quieter moment and, if it still has not started after 24 hours, is skipped until the next maintenance slot. Scheduled security runs and runs you start are still queued. A job moves through Queued, Running, Succeeded, Failed or Stopped.
  • Stop job ends a job you no longer want. It replaces the old Cancel button. It is on the Run update panel while a job is queued or running, on the job page, and on the Summary tab of the job drawer on Jobs. Select it and a dialog asks Stop this job? with Keep it and Stop job. For a queued job it says the job has not started, so nothing on your site changes. For a running job it says the build server stops the job within about a minute and removes everything it started, nothing is deployed, and a pull request the job already opened stays open (Autopilot will not merge it; close it on your Git host if you do not want it). Both end with "You can run the update again at any time." Once you ask, the button goes away and the job ends as Stopped. The button shows only for people who may run updates and only while the job is queued or running and has not been asked to stop. Viewers and other members do not see it, and Sentinel refuses the request if they try another way. If the job finished first, Sentinel says so and nothing changes.
  • A job that never starts. If no build server takes a queued job for 60 minutes, Sentinel ends it for you. The 60 minutes do not count while the job is held by its own site, your organisation's allowance or another job on the same AI key, while Sentinel has paused all jobs, or while another job is running. The job shows as Failed with the summary "Stopped: no build server could start this job within 60 minutes, so it was ended. Nothing on the site changed. Run it again when you are ready." Organisation administrators get a message and an email titled "Patching job for (site) could not start", under the email category Patching needs attention. Run the update again when you are ready.
  • If you stopped a running job and its build server then stops responding, the job still ends as Stopped, never as a crash, and Sentinel does not retry it by itself.
  • A run is refused if no repository is chosen, if the CI tests rule is not met (the same notice as for deploying; see Automated tests in CI), or if a job is already active for the site. The message says which.
  • If a job runs out of memory it ends as Failed with the summary "The job ran out of memory; retried once on its own if possible". Sentinel then queues one automatic retry for the same site, which runs alone on the runner and appears in the Runs table as its own job. If the retry also runs out of memory it ends as Failed with "The job ran out of memory again on its own; not retried", and nothing more is queued.
  • Pressing Run security update or Run maintenance update opens the new job's page straight away. While a job is active the panel shows Open the job page (see Patching jobs and activity); the steps also appear in the site's Activity section on its Patching tab. The job works on a branch named like release/security-20261021-3f9a1c, using your branch prefixes. Pull requests opened before the prefixes existed keep their sentinel/ branch names. If no pull request is opened, Sentinel deletes that branch afterwards.

Runs table. Below the buttons, Runs lists the site's jobs with these columns: Kind, Requested, Status, Outcome and a Job page link to the job page. Outcome shows the status line, "Pull request #N" when the run opened one, and the reason it ended for a failed run only. The table shows the last five runs, newest first; select Show all to see the rest and Show fewer to go back.

The table refreshes every ten seconds while a job is active.

Runtime updates

The Runtime section on the site's Patching tab, next to the Drupal updates configuration, keeps PHP, Nginx and the database up to date. It is a separate, opt-in track: runtime updates change only version files such as .platform.app.yaml, docker-compose.yml or a Dockerfile, and never composer.json or composer.lock (except that a PHP release line change also sets config.platform.php and refreshes composer.lock). They open their own pull request ("Runtime update: ..."), and a runtime problem never blocks a security fix in a package.

The hosting platform chosen for the site's production environment decides what is listed:

  • Quant Cloud: PHP and Nginx. There is no database component; Quant Cloud runs the database as a managed service, and the section says so.
  • Platform.sh / Upsun: PHP and the database, release line changes only. The platform applies patch releases itself, so there is no Autopilot switch, and Nginx is not listed because the platform runs it.
  • Lagoon (a hidden preset): PHP, Nginx and the database. On a GovCMS PaaS site the images are tagged by GovCMS release (for example 10.5.5) and held in the variable GOVCMS_IMAGE_VERSION. A pinned release moves to the newest release of the same GovCMS minor (for example 10.5.5 to 10.5.6), and every place the variable is written changes together, so every GovCMS image built from it (CLI, PHP and Nginx) moves to the same release. The badge shows no PHP version, and the pull request reads "Runtime update: PHP (GovCMS 10.5.6)". Turn on both PHP and Nginx: they share the version, so it changes only when both are on and agree on the new release, and the pull request says which other GovCMS images (such as the CLI image) move with them. The database is managed by GovCMS and is not supported.
  • No hosting platform chosen: the section says Not available yet and links to Choose production environment. Sentinel does not read your runtime files until a component is turned on.

Every component is off by default. Each row has a checkbox, a badge from the last run (Pinned, Pinned to 8.3, Pinned by digest, Not pinned or Not read yet), and the current version and file. Sentinel never writes a floating tag such as latest: every update is to a pinned version. A component that is Not pinned (for example an image tagged latest) moves without a pull request, so a maintenance run pins it to the newest pinned version the tag stands for (uselagoon/php-8.3-fpm:latest to its newest build, php:8.3-fpm to the newest 8.3 release; on GovCMS, 11.x-latest to the newest GovCMS 11 release). It stays Not pinned only when no such version exists or the tag says too little to choose one safely (php:latest, php:8). Security runs leave it alone. The row says Pin this version so we can keep it up to date and shows an example for your platform.

  • Autopilot switches appear per component only when the site's mode is Autopilot. Autopilot never changes a release line.
  • Save runtime settings is for organisation administrators and site managers. Other members see the section read-only.
  • Needs a person. When a run finds a release line nearing end of life, the row says Needs a person: move PHP from 8.3 to 8.4. An administrator or site manager can press Start a PHP 8.4 maintenance run. It is an ordinary maintenance job that makes the line change in its own runtime pull request. On GovCMS a PHP line change is always Needs a person, because GovCMS chooses the PHP line; the value in .env is the one used when docker-compose.yml uses the variable; otherwise a repository that writes GOVCMS_IMAGE_VERSION with different values in different files is held for a person until the files agree, and one that gives it no value at all is Not supported. A Lagoon project or environment variable named GOVCMS_IMAGE_VERSION overrides the repository, and Sentinel does not see it. A database line change cannot be downgraded, and a run is refused where the platform has no backup step before it.
  • Pull request. A component that was updated links to its Runtime pull request. It also appears in the Pull requests panel, where it deploys after the Drupal pull request from the same run (see Runtime pull requests below).

A later run does not open a second runtime pull request. It updates the one that is already open: Sentinel rebuilds it from the base branch and gives it a new title and description. If nothing needs to change, it leaves the pull request alone. It also leaves it alone while it is being deployed.

On the job page, a Runtime group follows the Drupal phases. It shows one badge per component (Up to date, Updated, Not pinned, Needs a person or Not supported), the runtime pull request link, and a link from each Needs a person item to the site's Runtime section.

If Patching is not enabled for your organisation, the Runtime section and its routes are not available.

Pull requests

Sentinel's pull requests (merge requests on GitLab) appear in two places, refreshed every minute, and everyone in the organisation sees them:

  • Waiting on you, in the Runs section of the site's Patching tab, lists the open ones only, in deploy order. It is hidden while nothing is open.
  • Pull requests from this run, on each job page, lists the pull requests that job opened, with a runtime pull request that deploys after one of them. Below them, Deployment lists this job's deployments with their live status. While the job is queued or running and has no pull request, it says "No pull request yet. It appears here when the job opens one."; once the job has ended with none, it says "This run opened no pull request."

Past runs, below Waiting on you, shows each run's pull request numbers and whether each is open, merged or closed. Each open card ends with either a Deploy to production button or the reason it cannot be deployed yet (see Deploying a pull request).

Each card shows:

  • A status pill. For an open pull request it is one status: Ready to deploy when the deploy gate is open, Deploying while this pull request's own deployment runs, or Waiting for another deployment while a different pull request's deployment holds the site. Otherwise it shows the pull request's own status. Below it, in plain words, whether the pull request can be merged and why not, for example Needs 1 approval from alice, Branch is behind main; rebase it or Check phpunit failed. The sentence comes from the Git host's own state. It is not shown while the pill says one of the three statuses above.
  • The checks, each with its outcome, whether it is required or optional, its duration and a link. While checks run, the card says waiting for CI since and when it started.
  • Review requirements: approvals given out of approvals required, who was asked to review and who can approve. An unresolved conversation count links to the host.
  • Labels, and buttons that open the pull request (or merge request) on GitHub (or GitLab) and its checks.
  • The comments mirrored from the host, read only. Select Show N comments to expand them; Sentinel's own are marked ours. Reply on the host, not in Sentinel.
  • A freshness line: updated by webhooks from the host, polled every minute, or Polling slowed to every 5 minutes when the host's rate budget is low. After the pull request is merged or closed, Sentinel stops polling it.

Organisation administrators and site managers see Rebase with Sentinel when the branch is behind its base or has conflicts. It queues a short job that updates the branch. Asking again within five minutes reuses the same request. Members see everything but not the button. On a runtime pull request, the rebase rebuilds only the runtime change.

When checks fail. Where Sentinel's CI fixing is turned on (it is being rolled out and is off until then), Sentinel reads why a check failed on its own pull request and acts:

  • A network or runner hiccup: it re-runs the failed checks once. The card says Sentinel re-ran the failed checks once.
  • A failure the update caused: it pushes a fix to the same branch, up to two rounds. The card says Sentinel is fixing the failing checks (round N). It never changes your CI configuration and never weakens or skips a test.
  • A check that also fails on your base branch, a broken pipeline, or no fix after the last round: it stops and comments on the pull request with the reason. The card starts with Needs a person: and the reason.

When the checks pass after a fix, the card says Checks pass after N fix rounds by Sentinel. Once Sentinel has stopped, pushes you make to the branch get no more rounds or comments. While Autopilot is paused for the site, Sentinel does not fix checks there. Organisation administrators and site managers can lower the rounds to one or turn them off for a site through the API (PUT /api/v1/sites/{site_uuid}/patching/ci-loop with {"ci_fix_rounds": 0}); other members cannot.

Runtime pull requests. A runtime update (PHP, Nginx or database versions) has its own pull request and its own card. Each card has a badge: a violet Runtime badge on runtime pull requests and a blue Drupal badge on Drupal ones.

When the same run also opened a Drupal pull request, the two cards show as numbered steps: 1 Drupal, then 2 Runtime. The runtime card says Linked to #N · deploys after it. If this one fails, #N is not affected. While it is held, it also shows Waiting for #N. The card lists the changes in a table (component, file, from, to) and shows a Deploy consent line. That line names each change and the pull request it deploys after. A run with no Drupal changes gives a runtime pull request that is not linked and deploys on its own.

If a pull request is merged on the host rather than through Sentinel, the card warns you. Sentinel does not deploy it for you in that case, so confirm the production deployment yourself. A pull request you deploy with the button below is merged by Sentinel and is not reported as merged outside it.

Bell notifications, which also follow your notification rules under the Patching category: CI failed on pull request #N for the site (a warning, also emailed; it clears when the checks pass), Review requested on pull request #N, Pull request #N is ready to deploy (clears when it is no longer ready) and Pull request #N was merged outside Sentinel (a warning, also emailed). Security scan alerts are described under Deep scan. If Patching is switched off for your organisation, these messages leave your bell and Messages, and alerts still waiting to be sent by email, Slack or webhook are not sent. The messages come back if Patching is switched on again.

Where an advisory has a Sentinel pull request, the site's Findings tab says Sentinel opened pull request #N (or merged) and that the finding closes when the deployment is observed, not on merge. On the Releases timeline, Linked fixes can say Sentinel pull request #N was merged before this deployment was observed.