feat: support hierarchical permission wildcards

[memleakd]

This PR intends to finalize the hierarchical permissions work from #1253 and #1309, which had stalled for some time.

Shield already supports wildcard permissions, but the current behavior is limited when permissions are organized into deeper namespaces.

In larger, real-world applications, permissions often grow beyond two segments. The flat/single-level wildcard behavior makes those permission sets harder to express and maintain cleanly, often requiring custom workarounds.

This PR makes wildcard permissions work with deeper structures while keeping the existing $user->can() and $group->can() APIs unchanged.

A trailing wildcard now matches descendant permissions:

'forum.posts.*'

matches forum.posts.create and forum.posts.comments.delete, but not forum.posts.

Wildcards can also be used in the middle of a permission, so forum.*.create matches forum.posts.create and forum.comments.create, but not forum.posts.comments.create.

I tried to address the concerns and review feedback from the earlier attempts:

• wildcard matching is shared instead of duplicated;
• both group-level and direct user-level permissions use the same matcher;
• direct user wildcard permissions still need to be listed in Config\AuthGroups::$permissions before assignment;
• * must be a complete segment;
• * cannot be the first segment;
• standalone * does not grant everything;
• trailing wildcards do not grant the parent permission itself;
• tests cover the matcher itself and the public $user->can() / $group->can() paths.

I’d appreciate any feedback. Hopefully this can help move this long-standing and requested feature forward.

resolves#1253, resolves#1309,

[michalsn]

I would like to make wildcards simpler: trailing wildcards match children only, never the parent. So forum.posts.* would match forum.posts.create but NOT forum.posts.

The current "parent matching" feels like hidden privilege escalation. If someone sets up forum.posts.* thinking "all actions on posts", they probably don't expect it also grants the bare forum.posts scope - which could be used as a different kind of gate elsewhere in their app. Current behavior is non-intuitive to me.

[datamweb]

@michalsn I agree, and I think this is the direction we should standardize on as a team.

The overall goal of supporting deeper hierarchical wildcard permissions still seems valid, but I do not think trailing wildcards should also grant the parent scope itself.

Allowing scope.* to match scope introduces implicit privilege expansion. In an authorization system, that is too broad a default because the granted access becomes wider than the permission string naturally suggests. If forum.posts is used elsewhere as its own gate, then assigning forum.posts.* may unintentionally grant that too.

The clearer contract, in my view, is:

1- scope => exact match only
2- scope.* => descendants only
3- never the parent itself

That keeps permission grants explicit, avoids hidden coupling between namespace-like labels and actual permission checks, and gives us simpler semantics to document and test consistently across both group-level and direct user permissions.

So my preference would be to keep the feature, but adjust the trailing-wildcard semantics to follow this rule before merging.

[memleakd]

Thanks for your reviews @michalsn, @datamweb

Just pushed an update. Trailing wildcards now match descendants only, never the parent scope itself. Tests and docs were updated to cover the new semantics.

[michalsn]

Thanks, this looks good!

[datamweb]

<?phpdeclare(strict_types=1);
/** * This file is part of CodeIgniter Shield. * * (c) CodeIgniter Foundation <admin@codeigniter.com> * * For the full copyright and license information, please view * the LICENSE file that was distributed with this source code. */namespaceCodeIgniter\Shield\Authorization;
/** * Matches permission grants against requested permission names for Shield authorization internals. */finalclass PermissionMatcher
{
/** * @param list<string> $grants */publicstaticfunctionmatches(string$permission, array$grants): bool
{
if (! self::isValid($permission)) {
returnfalse;
}
// Cache the exploded permission to prevent redundant string parsing in loops$permissionSegments = explode('.', $permission);
$permissionSegCount = count($permissionSegments);
foreach ($grantsas$grant) {
// Exact match first// If it matches exactly, it's inherently valid (since $permission is valid)// This skips the isValid() overhead entirely for exact matchesif ($grant === $permission) {
returntrue;
}
if (! self::isValid($grant)) {
continue;
}
if (str_contains($grant, '*') && self::matchesWildcardGrant($grant, $permissionSegments, $permissionSegCount)) {
returntrue;
}
}
returnfalse;
}
/** * @param list<string> $permissionSegments */privatestaticfunctionmatchesWildcardGrant(string$grant, array$permissionSegments, int$permissionSegCount): bool
{
$grantSegments = explode('.', $grant);
$grantSegCount = count($grantSegments);
// Check trailing wildcard efficientlyif ($grantSegments[$grantSegCount - 1] === '*') {
array_pop($grantSegments);
$grantSegCount--; // Decrement count after pop// Trailing wildcard matches children only, never the parent itself.return$permissionSegCount > $grantSegCount
&& self::segmentsMatch($grantSegments, $permissionSegments, $grantSegCount);
}
if ($permissionSegCount !== $grantSegCount) {
returnfalse;
}
returnself::segmentsMatch($grantSegments, $permissionSegments, $grantSegCount);
}
/** * @param list<string> $grantSegments * @param list<string> $permissionSegments */privatestaticfunctionsegmentsMatch(array$grantSegments, array$permissionSegments, int$matchLimitCount): bool
{
// Using "for" with a strict limit instead of array_slice() overheadfor ($i = 0; $i < $matchLimitCount; $i++) {
if ($grantSegments[$i] !== '*' && $grantSegments[$i] !== $permissionSegments[$i]) {
returnfalse;
}
}
returntrue;
}
privatestaticfunctionisValid(string$permission): bool
{
// Catch invalid starts right away before allocating an array.if ($permission === '' || str_starts_with($permission, '*')) {
returnfalse;
}
$segments = explode('.', $permission);
foreach ($segmentsas$segment) {
if ($segment === '' || ($segment !== '*' && str_contains($segment, '*'))) {
returnfalse;
}
}
returntrue;
}
}

I was playing around with the code to see if we could squeeze out a bit more performance (mostly by avoiding array_slice overhead and optimizing the explode logic). I ran a benchmark test with 100,000 iterations, and the results were actually quite promising:

Running Benchmark (100000 iterations)...Original Code:Time: 488.31 msOptimized Code:Time: 395.79 msPerformance Improvement: 18.95%

@memleakd What are your thoughts on this?

[memleakd]

Thanks for digging into this and putting together the benchmark idea @datamweb.

I played with it a bit more locally and compared the implementations:

Click to see the optimized matcher class

<?phpdeclare(strict_types=1);
/** * This file is part of CodeIgniter Shield. * * (c) CodeIgniter Foundation <admin@codeigniter.com> * * For the full copyright and license information, please view * the LICENSE file that was distributed with this source code. */namespaceCodeIgniter\Shield\Authorization;
/** * Matches permission grants against requested permission names for Shield authorization internals. */finalclass PermissionMatcher
{
/** * @param list<string> $grants */publicstaticfunctionmatches(string$permission, array$grants): bool
{
$permissionSegments = self::splitIfValid($permission);
if ($permissionSegments === null) {
returnfalse;
}
$permissionSegmentCount = count($permissionSegments);
foreach ($grantsas$grant) {
if ($grant === $permission) {
returntrue;
}
$wildcardPosition = strpos($grant, '*');
if ($wildcardPosition === false) {
continue;
}
if (
$wildcardPosition > 0
&& $wildcardPosition === strlen($grant) - 1
&& $grant[$wildcardPosition - 1] === '.'
) {
$prefix = substr($grant, 0, $wildcardPosition - 1);
if (self::isLiteral($prefix) && str_starts_with($permission, $prefix . '.')) {
returntrue;
}
continue;
}
$grantSegments = self::splitIfValid($grant);
if ($grantSegments === null) {
continue;
}
if (self::matchesWildcardGrant($grantSegments, $permissionSegments, $permissionSegmentCount)) {
returntrue;
}
}
returnfalse;
}
/** * @param list<string> $grantSegments * @param list<string> $permissionSegments */privatestaticfunctionmatchesWildcardGrant(array$grantSegments, array$permissionSegments, int$permissionSegmentCount): bool
{
$grantSegmentCount = count($grantSegments);
if ($grantSegments[$grantSegmentCount - 1] === '*') {
$grantSegmentCount--;
return$permissionSegmentCount > $grantSegmentCount
&& self::segmentsMatch($grantSegments, $permissionSegments, $grantSegmentCount);
}
return$permissionSegmentCount === $grantSegmentCount
&& self::segmentsMatch($grantSegments, $permissionSegments, $grantSegmentCount);
}
/** * @param list<string> $grantSegments * @param list<string> $permissionSegments */privatestaticfunctionsegmentsMatch(array$grantSegments, array$permissionSegments, int$segmentCount): bool
{
for ($index = 0; $index < $segmentCount; $index++) {
if ($grantSegments[$index] !== '*' && $grantSegments[$index] !== $permissionSegments[$index]) {
returnfalse;
}
}
returntrue;
}
/** * @return list<string>|null */privatestaticfunctionsplitIfValid(string$permission): ?array
{
if ($permission === '' || str_starts_with($permission, '*')) {
returnnull;
}
$segments = explode('.', $permission);
foreach ($segmentsas$segment) {
if ($segment === '' || ($segment !== '*' && str_contains($segment, '*'))) {
returnnull;
}
}
return$segments;
}
privatestaticfunctionisLiteral(string$permission): bool
{
return$permission !== ''
&& ! str_contains($permission, '*')
&& ! str_contains($permission, '..')
&& ! str_starts_with($permission, '.')
&& ! str_ends_with($permission, '.');
}
}

• Original: the first implementation in this PR
• Your version: the optimization you suggested
• Current: the version I pushed after building on the same idea

I used 100,000 iterations, same as your benchmark:

XDEBUG_MODE=off php -d opcache.enable_cli=0 permission-matcher-compare.php 100000

Here is one representative run:

Case	Original	Your version	Current
Exact match first	34.14 ms	24.39 ms (-28.57%)	20.66 ms (-39.49%)
Exact match late	67.93 ms	58.55 ms (-13.80%)	25.39 ms (-62.62%)
Trailing wildcard	68.92 ms	61.57 ms (-10.67%)	38.70 ms (-43.84%)
Middle wildcard	58.60 ms	56.28 ms (-3.96%)	51.95 ms (-11.35%)
Many grants miss	139.78 ms	126.41 ms (-9.57%)	100.69 ms (-27.96%)
Many grants late wildcard	165.55 ms	141.61 ms (-14.46%)	99.29 ms (-40.02%)

I kept the direction you suggested, then adjusted it a bit further: the matcher now parses the requested permission once, avoids array_slice(), and has a small fast path for common literal trailing wildcards like admin.users.*.

I also tried simplifying the class more, but those versions were slower in the trailing wildcard path, so I kept this as the best balance I could find between readability and performance.

I have not committed this yet, so I’m happy to go either way. If the team thinks the performance gains are worth the slightly larger matcher, I can commit this version here.

I'd appreciate your thoughts on the tradeoff.

[michalsn]

Thanks, these numbers look good. The absolute savings are small in application terms, but parsing the requested permission once and avoiding repeated allocations are sensible improvements for code that may run frequently. I'm open to the optimized version as long as the fast path remains straightforward and covered by the existing edge-case tests. Please commit it so we can review it.

[memleakd]

Thanks @michalsn, that sounds good. I just pushed the optimized version. I kept the fast path limited to trailing wildcards, with the existing edge-case tests covering it.

[datamweb]

While the new implementation introduces a bit more complexity, I believe the significant performance and efficiency gains are absolutely worth the trade-off in this context.

[datamweb]

LGTM!

[datamweb]

@memleakd@michalsn Thank you both for your time and effort on this contribution.

[memleakd]

Thank you @michalsn@datamweb@bgeneto@christianberkman

[christianberkman]

@memleakd Thank you for finally making this happen.