Download the PHP package wp-media/fleet-bridge without Composer

fleet-bridge

Lets a WordPress plugin accept commands from Fleet, verified.

What Fleet is, and why a plugin needs this

Fleet is a WP Media application where an agency manages a plugin across every site on its licence from one dashboard. Today they open one wp-admin per site.

To read or change a setting, Fleet has to reach the site — and it deliberately holds no customer credential: no WordPress login, no application password, not even the licence key. An application password would be the easy route and the wrong one. It authenticates a user, so it grants whatever that user can do rather than "this plugin's allowlisted options", and Fleet would have to store one secret per site — tens of thousands of secrets whose leak is a total compromise of every one of them.

So Fleet signs a short-lived token per request, and this package checks it.

Two proofs, from two parties

Every request must carry both:

Neither substitutes for the other, and the reasons are worth understanding before changing anything here.

Fleet must not sign the consent. It is the party being granted the access. A grantee attesting its own grant proves nothing at all.

The consent must not be cached. An install that remembers "allowed" for a day keeps accepting commands for a day after the owner withdraws it — and a permission that takes a day to withdraw is not a permission, it is a delay. So it arrives with each command, minted seconds earlier, and revocation is immediate.

What is safe to cache for a long time is a key set location, because a rotation publishes two keys at once for exactly that reason. That is why the vendor tells an install where the keys are, not whether it is allowed.

What it deliberately cannot tell you

Which of Fleet's own customers asked. There is no per-site secret shared with Fleet, so there is nothing to check that against. Isolation between Fleet's customers is enforced in Fleet, before anything leaves it.

This package verifies that a command is Fleet's, that it names this site, and that the owner allows it. That is the whole of it, and the boundary is deliberate rather than an omission.

Install

PHP 7.4+, and libsodium — which has shipped with PHP since 7.2, so in practice there is nothing to install.

Using it

Type your route against Contract\Verifier, not against Bridge. Bridge is final — verification is security critical, and a subclass overriding half a check is a hole that looks like a customisation — which means a route holding the concrete class cannot be doubled, and the only way left to test the route's own refusals is to sign real tokens in your suite. That tests this package again, not your route.

The only thing you must write is a TrustStore, because where trust comes from is your business:

Empty means trust nobody, never trust anybody. An install that has not yet polled its vendor is expected to answer empty and refuse everything.

Design notes worth reading before you change it

The host's site identity is port free. Config takes a host, and both sides derive the subject from it. A port in one and not the other is a refusal that looks like a signature problem.

Only EdDSA, compared by equality. Not a blocklist — RS256 and none fail the same check, and a future algorithm name cannot slip past by not being on a list. Accepting RS256 would let anyone who can read the published public key sign their own commands.

Every published key is tried. Tokens carry no kid, because a rotation publishes two keys and neither side should have to be redeployed in step with the other. An empty key set refuses — "we could not fetch the keys" must never read as "no key objected".

A ver claim, refused when unknown. Everything else here freezes on the first plugin release — old installs live for years and nobody can patch them — so a newer signer talking to an older verifier has to be refused cleanly rather than read with the old meanings. Config::withAcceptedVersions() widens it for a transition; widening it permanently defeats the point.

The identifier is spent last. A token that fails a later check is not burned, so a wrong-scope request is retryable rather than a dead end.

WpdbNonceStore uses a single INSERT, not a transient. get_transient() then set_transient() is not atomic: two requests carrying one identifier can both read "unused" before either writes, and both are honoured — which is exactly the replay the store exists to prevent. The unique index on option_name is the only place that decision can be made atomically.

One fixed refusal. A route using this is reachable by anyone, so every cause returns Bridge::REFUSED. Reflecting the reason turns it into an oracle: "unknown issuer" against "bad signature" tells an attacker whether their guessed issuer is the one this site trusts.

Key sets are fetched over HTTPS only (localhost and .test excepted for development, by host name rather than by a flag — a flag is one wrong environment variable away from doing it in production). Every key in a fetched set is tried, so a substituted set would be enough to forge commands.

Layout

Tests

The suite is written around the refusals, and each one says what would happen without the check. A test named for what it allows proves the happy path; a test named for what it refuses proves the thing an attacker would try.

src/WordPress/ is not unit tested here — it is integration surface, covered by the host plugin's own suite where WordPress exists.

Code style

PSR-12, enforced by composer lint. Deliberately not the WordPress standard: this package is framework agnostic and only its adapters touch WordPress.