What the agent does
The agent reads the file the error names and the image the project ships, because the call is usually one old guard and the project's own PHP version explains why nobody saw it. In an app like thibaud-rohmer/PhotoShow, index.php line 74 reads if (!get_magic_quotes_gpc()){ and the three lines inside it pass $_POST, $_COOKIE and $_GET through protect_user_send_var, a function defined a few lines above that maps addslashes over arrays. docker/Dockerfile starts from alpine:3.10.1 and installs php7-fpm, php7-gd and the other php7-* packages, so the project's own container never ran PHP 8.
It reproduces the failure on the PHP that hosts run now. On PHP 8.2 php -r 'get_magic_quotes_gpc();' prints PHP Fatal error: Uncaught Error: Call to undefined function get_magic_quotes_gpc(), and function_exists('get_magic_quotes_gpc') is false. Serving the repository with php -S and requesting index.php shows how users see it: PhotoShow's own error handler catches the Error and renders <div class='exception'>Call to undefined function get_magic_quotes_gpc()</div> with an HTTP 200, so a monitor that checks status codes never notices.
The fix keeps the app's behaviour. Magic quotes were removed from PHP in 5.4, so on every PHP the project has run in years get_magic_quotes_gpc() returned false and the three protect_user_send_var calls always ran. The agent deletes the if and keeps the three calls unconditional, which is exactly what the old code did on PHP 5.4 through 7.4. php -l over the project's own files reports no syntax errors, and a search of the project's code outside the vendored src/ directory finds no each(), create_function(), ereg(), split() or mysql_* call to trip on next.
The vendored phpThumb copy under src/phpthumb also calls @get_magic_quotes_gpc(), on line 3922 of phpthumb.class.php, inside its debug dump, which runs only when config_disable_debug is turned off; the class ships with it set to true. The @ does not silence an Error on PHP 8, so the agent leaves the vendored file alone and records the line in the pull request as what would break if debug output were ever enabled. After the change, the served index.php answers 200 with <title>PhotoShow</title> and the gallery's normal page.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Our PHP app shows `Call to undefined function get_magic_quotes_gpc()` after the host moved to PHP 8. Find every call to it, read the code around each one and the PHP version the project's own Dockerfile installs, and serve the app on PHP 8 to show what the page returns now. Remove the magic-quotes guard in our own code so the input handling stays exactly what it was on PHP 5.4 and later, run php -l over our files, and search our code for other functions PHP 8 removed. Leave vendored libraries alone but list any call they make. Serve the app again and open a pull request quoting the response before and after.",
"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
A pull request that changes index.php only: the if (!get_magic_quotes_gpc()) line and its closing brace are gone, and the three protect_user_send_var assignments stay, now unconditional. The description quotes the fatal error from the PHP 8 command line, the HTML the app's error handler returned before, and the <title>PhotoShow</title> response after.
A reviewer checks that addslashes is still applied to the same three superglobals, that no vendored file under src/ changed, and that the Dockerfile still installs PHP 7, which the pull request notes but does not change.
The description lists the one vendored call in phpThumb's debug dump with its line and the setting that keeps it unreachable, so the team knows what breaks if they ever turn that debug mode on.
Run it on every PHP version change on a host
A host upgrade breaks every old app at once, so the useful unit is each PHP repository the org still runs. Your server sends this request once per repository, each with its own UUID as request_id so a retry returns the run it already started, and reads the results from the key's run.completed webhooks. A key starts at most max_concurrent runs at once and queues the rest.
A Grafana alert on 200 responses that contain the string exception can call a Rush webhook trigger directly, since a Grafana contact point sends an Authorization: Bearer header with the trigger's secret; the alert payload reaches the agent as data after the prompt.
The use case this job follows, with its automation and trigger: Fix a bug from an alert.
Why run it on Rush?
The app hides this failure behind a 200 page, so proof means serving it, not reading it. In a Rush sandbox the agent installs PHP 8.2, starts the built-in server against the checkout, requests the page before and after, and keeps both responses with the run, so the pull request shows what a user saw and what they see now.
The right fix is to remove a guard without changing input handling, and the argument for that rests on what the old code did on the PHP versions the project supported. The agent reads the Dockerfile for that history, searches for the other removed functions so the next fatal error is already known, and leaves vendored code untouched with the one risky line named.
Frequently asked questions
Why not replace the call with a stub that returns false?
Because the guard has done nothing since PHP 5.4 removed magic quotes: the function returned false, so the three addslashes calls always ran. Deleting the condition keeps that behaviour with less code, and a stub would keep a dead concept alive.
Why does the monitor show the site as up?
PhotoShow's error handler renders the exception inside the page and answers HTTP 200. The agent reproduces that with php -S and quotes the response, so a reviewer sees why status checks missed it.
Does the vendored phpThumb need patching?
Not for this failure. Its only call sits in a debug dump that runs when config_disable_debug is false, and the class ships with it true. The pull request records the line so turning debug on is a known risk.
What else from PHP 8 would break this app?
The agent searches the project's own code for each(), create_function(), ereg(), split() and mysql_* calls and runs php -l on every file. In PhotoShow's own code it finds none, and says so; if it did, each would be its own fix.