Skip to content
Menu

Proof#

This page says how Squire was proven on real Wayne State applications: what ran, what held, what the run caught in Squire itself, and what it does not prove yet. Next, see Doctor for the rules the run checked and Local parity for the setup the boots used.

A squire is judged on the field, not in the tent. So on 2026-11-05 Squire went out to every PHP application in the stack's shared proof corpus, the same set PROCTOR and Preflight are measured on.

Last run#

The table below records the most recent time the checks on this page were run against Squire itself, so a reader can see when the page was last verified and what the run found.

Check Date Commit Result Duration
composer ci 2026-11-07 88fd2ec pass: 404 tests, 2771 assertions, 0 phpstan errors, 0 advisories 8m 14s
Local parity rig (parity.sh up --fallback, shell twin under WSL 2 bash) 2026-11-07 88fd2ec pass: the release ran all eight steps in 6s and /up.php answered 200 with body ok through the proxy 27s to healthy (images prebuilt on this machine)
Image contract on the fallback (contract-image.sh, linux/amd64, Git Bash on Docker Desktop) 2026-11-10 c33f789 plus the --base-image change pass: 26 of 26, under the compose file's least privilege 4m 19s (base built from cache)
Image contract on Citadel's own image (contract-image.sh --base-image=citadel/php-apache:8.5r3, image id 950dcf67937a) 2026-11-10 the same pass: 23 checks passed, 3 gaps that only the image can close (below) 1m 20s
Local parity rig on Windows PowerShell 5.1.26100 (parity.ps1 up -Fallback, then up on citadel/php-apache:8.5r3, status, pin, reset) 2026-11-10 the same plus the 5.1 fix pass: /up.php answered 200 with body ok through the proxy on both bases; before the fix, up with no base image died on a raw NativeCommandError instead of its own message 49s fallback, 26s Citadel's image

To refresh this section, rerun composer ci and the local parity rig from Local parity, then rewrite the table. The 82-app corpus run and the four real-app boots were not rerun: they need the shared proof corpus and the internal app repositories, which are not available on this machine, so the parity row uses the synthetic fixture that squire init writes, the same way CI does.

What ran#

Squire 0.5.0 from main, plus the init notice described below, in php:8.4-cli with no network. The 82 applications are 64 raw PHP apps (33 with a composer.json, 31 without one), 17 Laravel apps from 5.8 to 13, and 1 Slim 2 app. Six libraries, one template and six test sites ran too and are counted on their own line.

Each app was a throwaway copy of its default branch, checked out with Linux line endings, with .env files, auth.json and config.files/ removed, and made into a fresh git repository so doctor could see what is tracked. Then, in order:

  1. squire doctor --format=json on the app as it was;
  2. squire init, letting Squire pick the target;
  3. squire init --check, which must find no drift in what init just wrote;
  4. squire init again, which must change nothing;
  5. squire doctor --format=json on the result;
  6. squire update --dry-run;
  7. a lint of every file init wrote: sh -n or bash -n for the shell scripts, php -l for the PHP, a JSON parse for .squire.json, and a search for carriage returns.

An app passes when every step does what it should and doctor ends with exit 0. It is flagged when Squire did its job and still has something to tell the app (a fail or an untracked file the app wrote itself). A tool error is Squire breaking: a crash, an exit 2, invalid JSON, drift in its own output, a second init that changes something, or a written file that fails its lint.

Results#

Units Pass Flagged Tool error
82 applications 76 6 0
13 libraries, template and test sites 10 3 0
Shape Apps Target init picked
Raw PHP, no composer.json 31 raw
Raw PHP with Composer 33 raw
Slim 2 1 raw
Laravel 5.8 to 13 17 laravel (from artisan)
  • init wrote 18 to 21 files per app, and every one of them passed its lint. A second init changed nothing anywhere.
  • 36 apps already had their own public/index.php and 2 their own public/.htaccess. init reported each as app-owned and left it alone, as documented.
  • The six flagged apps had their own build.sh or health.php. init skipped those files without --force, init --check named them as untracked, and in one of them doctor failed S1: its own build.sh runs composer require, which cannot work at image build time.
  • The median app took 3 seconds for all seven steps.

Four apps on local parity#

Doctor reads files; it never runs an app. To see whether the kit actually brings an app up, four apps of different shapes were booted on local parity with the fallback base image, HAProxy in front and MySQL 5.6.51 behind, with every Wayne State host mapped to 127.0.0.1 so no app could reach a real service.

App shape What happened
Raw PHP, no composer.json The first release stopped at step 1: composer.json not found. See below.
Raw PHP with Composer Healthy in 23 seconds. /up.php answered ok through the proxy. The app's own front page answered 500 without its configuration, which the proof does not supply.
Slim 2 The release stopped at step 3: the lock file installs one package from Subversion, and the image has no svn. The message named the step and the cause.
Laravel 13 Without a .env, step 1 refused, as documented. With one made from .env.example, as the Citadel dashboard would, the release ran all ten steps in 106 seconds, including the migrations on MySQL 5.6. The app's own .htaccess then redirects every plain HTTP request to HTTPS, so /up answered 307 and the health check said unhealthy. That is the trusted-proxy class of problem parity exists to show before Citadel does.

The release on Citadel's own image#

The boots above used the fallback base. On 2026-11-10 the image contract suite ran on Citadel's own PHP image instead: citadel/php-apache:8.5r3, loaded byte for byte by Parity's parity image get (image id 950dcf67937a), with the app container under the same least privilege the kit's compose file gives it. Citadel's entrypoint fetched Composer and ran the first release as citadel; a second release ran with the smoke test on; /up.php answered 200 from the host and health.php passed inside; opcache checked mounted code every 2 seconds and never checked baked code; the web had no exec and kept proc_open; the CLI kept everything; a Composer script shelled out; and the mounted secret's value was in no log, layer or image config.

Three findings belong to the image, not to the kit, and no app file can change them:

  • The Server header is Apache/2.4.69 (Debian): every answer names the Apache version and the OS (ServerTokens Prod fixes it).
  • Apache's own error pages carry the version footer (ServerSignature Off fixes it). TRACE already answers 405.
  • The entrypoint does not turn /run/secrets files into variables for Apache, so on Citadel ./.env stays the secret path, as Image hardening says.

What it caught in Squire#

A run that only finds praise was not looking hard enough. This one found these, and every finding is about Squire, not the apps:

  • Fixed: the release needs a lock file, and init never said so. 31 of the 82 apps have no composer.json and no composer.lock. init wrote a release into each, doctor passed it, and the first boot showed that release stopping at step 1. init now ends with a "Heads up" line naming the missing files and how to make them; see Troubleshooting.
  • Fixed: Yarn. Two apps keep a yarn.lock. The release installed JavaScript only from package-lock.json or pnpm-lock.yaml, so the frontend step stopped with "commit a lock file", which read oddly when one was committed. It now installs from yarn.lock too, through corepack: Yarn 1 with --frozen-lockfile, Yarn 2 and later with --immutable (see The release script). Both were run by hand inside citadel/php-apache:8.5r3 on 2026-11-10: corepack's Yarn 1.22.22 and Yarn 4.5.3 each installed from their lock with scripts off and then ran the build.
  • Open: fresh output that warns about itself. Right after init, doctor warns on every app that .env.squire is not ignored at the repository root (init only edits .gitignore in a folder below the root) and that the Citadel image in the compose file is not pinned by digest (S67). Both warnings are true and both have a documented fix, but a kit should not hand its owner homework on day one.
  • Windows checkouts. The first pass used clones made on Windows, where Git had turned the line endings of existing deploy scripts into CRLF. doctor failed S8 on 11 apps that are LF in their repositories, exactly the trap the rule describes. The numbers above are from the Linux checkout Citadel sees.

What this does not prove#

  • No app was deployed to Citadel and no Quick Update ran. The four app boots used the fallback base image, an approximation of Citadel's image; only the image contract's fixture app ran on Citadel's own image, and never behind Citadel's proxy.
  • No real configuration, database contents, Oracle, CAS or Entra were used, so an app's own pages are only shown to start, not to work.
  • doctor decides only the rules a machine can read from files; the rest of the best practices are not checked here.
  • The per-app record names the applications and stays in the stack's internal notes.

Run it on your app#

The same steps on your own app, from its root:

composer require waynestate/squire --dev
vendor/bin/squire doctor
vendor/bin/squire init
vendor/bin/squire init --check
vendor/bin/squire doctor

Then bring it up on local parity:

Linux / macOS

scripts/parity.sh up --fallback

Windows

scripts\parity.ps1 up -Fallback