PHP digest — 26 October 2026
Rebuilt PHP Windows packages fix a Brotli vulnerability, a "Scope" RFC opens for discussion, the ext/csv author plans a vote for 2 November (Tim Düsterhus suggested waiting), and Rector 2.8 and Symfony AI 1.0.0-BETA1 are out. There were no new PHP source releases and no RFC status changes in the official index.
Windows builds 8.5.12-1, 8.4.27-1, 8.3.36-1 and 8.2.35-1 bundle Brotli 1.2.1
php.net announced updated Windows builds (also
posted to internals at about 15:49 UTC) that include Brotli 1.2.1,
which fixes CVE-2026-111497, an out-of-bounds write in Brotli decompression. According to php.net, the decoder can be
reached in PHP's Windows builds through the cURL extension when automatic HTTP response decompression is enabled. That
is off by default in cURL, but applications and HTTP client libraries can enable it. Windows users are advised to
update. The -1 suffix marks a package rebuild with updated dependencies; the PHP source and runtime versions are
unchanged. Brotli's v1.2.1 release (21 October) says the
out-of-bounds write "permits remote code execution". No CVSS score was found (the CVE record could not be fetched), and
the announcement does not discuss other platforms. See the release page.
"Scope" RFC opens for discussion
Frode Børli started the discussion of the Scope RFC
(version 0.9.1, proposed PHP 8.7) at about 16:04 UTC, although the page header still reads Draft. A final class Scope would identify a unit of work (an HTTP request, queue job or CLI command) in long-running servers:
(new Scope())->run($handler, $request), Scope::getCurrent(), Scope::getRootScope(), bind(), plus
Fiber::getScope() and Generator::getScope(). Fibers and generators keep the scope they were created in. The primary
vote (2/3) adds only this identity. Two independent secondary votes (50%+1) would make request superglobals including
$_SERVER, and the session, follow the scope (A), and the default locale and timezone (B). The karma request on
13 October had said $_SERVER would stay shared. The only listed BC break is a global Scope class name. The reference
implementation is an extension for PHP 8.2–8.5 (not checked here); a php-src PR is planned before the vote. See the
RFC page.
ext/csv: vote planned for 2 November, request to wait
The ext/csv RFC author wrote
at about 14:52 UTC that there had been no further feedback for about two weeks and that, without new concerns, they
intend to open the vote on 2 November. The page still reads version 1.2 (11 October), Under Discussion, with one
proposed 2/3 question on adding the extension to core. The message does not mention the directional vote between
improving ext/standard, ext/csv and an object-oriented API that the author had floated on 11 October. At about
22:06 UTC Tim Düsterhus replied that the quiet period included last
week's security release and suggested waiting with the vote. Without having read the RFC yet, he questioned whether CSV
warrants its own core extension and reserved top-level namespace, and asked to hear from Gina Peter Banyard, author of
the original girgias/csv extension. See the RFC page.
Rector 2.8 and 2.8.1
Rector 2.8.0 (12:16 UTC) reports a core engine about 26%
faster with about 8% less memory than 2.7.0 on the project's own benchmark, and runs its parallel workers with PHPStan
Turbo on PHP 8.3+. Rules now converge in a single pass (fixed-point iteration, idempotency enforced in tests). Many
niche rules that belong to no set are deprecated (they still run, with a warning), and the notes list the removal of
deprecated public API, including the withPhp53Sets() … withPhp74Sets() configuration methods.
2.8.1 (13:06 UTC) fixes a cache crash caused by PHPStan
Turbo. See Rector 2.8.
Symfony AI 1.0.0-BETA1
Symfony released Symfony AI 1.0.0-BETA1, the
first pre-release towards 1.0 (composer require symfony/ai-bundle:^1.0@beta). The post says the 1.0 feature set is
settled, but the backward compatibility promise starts only at 1.0.0 final, so the beta is the last window for
breaking changes. The official PHP MCP SDK (mcp/sdk) is said to be heading for 1.0 alongside it. An online hackathon
from 26 October to 22 November is meant to collect bug reports for both.
Symfony 8.2: Messenger inbox skips already-handled redeliveries
A post on 26 October describes an inbox transport
option in the Symfony 8.2 development branch (not released). It complements the Messenger outbox that the post says
was introduced earlier. PR #66957 was merged into the 8.2 branch.
Brokers deliver messages at least once. A worker can commit its handlers' changes and then crash before it acknowledges
the message, and the handlers then run again when the message is redelivered. With inbox: true (or
inbox: { connection: … }) on a transport, Symfony writes the transport name and message ID to a messenger_inbox
table inside the transaction that runs the handlers. When a redelivered message is already in the table, its handlers
are skipped and the message is acknowledged. If a handler fails, the transaction rolls back both the changes and the
inbox entry. A second worker that receives the same message waits for the first one's transaction to finish.
According to the post:
- it needs
identity_stamps: true, thedoctrine_transactionmiddleware andsymfony/doctrine-messenger, and the inbox must use the same database connection as the handlers.messenger:setup-transportsor a Doctrine migration creates the table; - retention defaults to 7 days and must be longer than the transport's possible redelivery delay;
- messages retried from a failure transport are recorded under the transport where they first failed. Messages handled by batch handlers are not recorded;
- messages dispatched with
DispatchAfterCurrentBusStampmust go to a transport with an outbox, and Messenger refuses to send them otherwise. This is because the inbox entry is committed before they are sent.
The post describes how the inbox behaves with the database transaction. It does not discuss handler side effects outside that transaction.
Also in discussion
The first reply to the array_partition() RFC
(132928) suggested a general array_group() with seeded keys
instead. The author kept the two-way function: seeded [true, false]
keys still become 1 and 0, so destructuring swaps the buckets. The RFC page (still 0.1) now records this argument.
Later, Tim Düsterhus said he favors the RFC and listed editorial
points (a rendering error, the BC wording, a full userland polyfill). See the RFC page.
Coverage
At 05:01, 11:01 and 17:01 UTC fetch --due collected all 32 due feed, release and index sources without failures; the
25 daily sources were not due. At 23:01 UTC it collected all 57 sources, including the daily ones, without failures. health --fail-stale reports 57 sources ok and none stale. The official RFC index grew
from 954 to 955 entries (Scope, draft). New records before 11:01 UTC: the Symfony Messenger inbox post (above); a reply
(132927) in the thread on first-time contributors' php-src PRs, in which Sjoerd
Langkemper says the biggest friction for them was getting a response to their PRs; and an edit to the
Uri\QueryParams RFC at 10:43 UTC, followed by another at 15:05 UTC. After
both edits the page still reads version 1.0, Under Discussion, with placeholder vote dates, so it is not a story (see
the RFC page). Not included: three replies (132929,
132930, 132936) in the idea thread on effect declarations and capability-gated
process execution (no RFC); a reply in the native async/await idea thread that has no text content in the list archive
(132935; covered on 23 October); a blog post on
Laravel queue deadlocks, a Laravel testing course and a SymfonyCon schedule post. The CVE record for CVE-2026-111497
could not be fetched (network proxy refused the MITRE CVE API). Two RFCs are voting. At about 23:05 UTC number-base
exceptions stood at 24–1–0 (closes 1 November 15:30 UTC) and Filter enums at 2–10–0 (closes 4 November), both unchanged.
Daily package and version observations showed no notable change against 25 October, and support dates were unchanged.