Hello world, it's Wednesday, June 24, 2026, and here's what happened This Week in PHP Internals.

This week's episode is supported by Ballast — our new gauge for whether the software you ship actually lasts. Velocity tells you how fast you ship; Ballast tells you whether it stayed shipped — one durability score, read straight from your git history. No survey, no black box. Early access opens at artisan.build/ballast.

Right — Gina P. Banyard opened the annual Deprecations for PHP 8.6 RFC. It's the yearly omnibus: a year's worth of small removals, compiled by a dozen hands, each one voted on its own, and it has to be frozen by July 13. This year the sharpest exchange wasn't about any one removal — it was about evidence. Juliette Reinders Folmer pressed the point that the proposals ship without an impact analysis: "[N]one of the proposals contain a proper impact analysis," she wrote, calling it irresponsible to ask voters to decide partly blind — and she offered to build one, PHP_CodeSniffer plus PHPCompatibility, covering most of the list.

Tim Düsterhus pushed back — said the claim wasn't accurate, and reminded the list that deprecations aren't a breaking change. Juliette felt that talked past her, and answered plainly: "Tim, please don't try to gaslight me." It's worth sitting with, not laughing past: underneath the heat is a fair ask — give voters the data to decide. And by midweek it wasn't a two-person spat — Rowan Tommins backed the substance. A deprecation, he argued, is really a proposal to remove something later, so the impact of that removal is fair to weigh; he put it bluntly: "To knowingly omit information that might weaken your case would be dishonest."

So what's actually proposed for removal? A real grab-bag, each with its own argument. The _() shorthand for gettext() — and Rowan Tommins flagged that one as genuinely costly, since _() is the canonical gettext spelling and big codebases lean on it thousands of times. Reserving the keywords let, in, out, and inout. The CURLOPT_PROGRESSFUNCTION constant — where Ayesh Karunaratne argued don't deprecate it at all, just quietly remap it onto CURLOPT_XFERINFOFUNCTION and spare a few hundred maintainers the churn. Plus spl_object_hash() and metaphone().

One fresh proposal rode in on that same thread. Muhammed Arshid KV wants to deprecate SplFileObject's CSV methods — fgetcsv, fputcsv and the rest — in favour of a dedicated csv extension. But Ignace Nyamagana Butera, who maintains the league/csv library, hit the brakes: "deprecating the current API without providing at least the start of a replacement API seems premature [...] Anything else would be counterproductive at the moment." Gina's standing offer to anyone with an idea: send me the text and I'll fold it in — but I'm not writing it for you.

The single busiest thread of the week, though, was brand new. Tim Düsterhus and Derick Rethans proposed a Time\Duration class — a first-class, immutable stopwatch value, deliberately small, aimed at 8.6 so it can smooth the new Polling API's timeouts before the feature freeze. And it's the opening move of a modern date-and-time API. Fifty-some replies followed, almost all the productive kind of argument: should a negative duration carry a sign flag or just be signed seconds? What on earth do you call dividing one duration by another? Marc Bennewitz pushed hard on the representation; Paweł Kraśnicki showed up from the brick/date-time world and got a warm "welcome and thank you for contributing."

When the design drifted toward a static Duration::sum(), Larry Garfield drew a firm line: "I am 100% against a static method for adding durations. That's entirely pointless." He wants $dur1->add($dur2) — like a civilized person. By Tuesday Tim had already renamed the ISO parser and added an ->absolute() method, so this one is genuinely moving.

Rob Landers opened a pre-RFC for Primary Constructors, lifted out of his old Records proposal. Picture the whole constructor promoted up onto the class line itself. The catch in Rob's version: a class with one can't also declare a __construct, so real init logic has to run through property hooks — and Nick Sdot caught the snag immediately, that hooks-only means the class can't be readonly. Larry's in, keep-it-simple, with a side of Kotlin. Rowan Tommins pushed back: "I don't think we should add extra syntax to the language just to change people's habits."

Function autoloading came back for a fifth try — Paul Jones's mark 5 of a feature internals keeps almost-shipping and never quite does. It drew the exact objection that's sunk every prior run: lean on it, forget to fully-qualify a name, and your code works by accident because something else loaded that function first — then breaks silently after a refactor. Tim Düsterhus put it flatly: "Agreed on this being a non-starter for the proposal for me." Paul's answer is a declare(strict_namespace=1) switch to kill the global fallback, plus a patient reply to Bob Weinand on why the engine simply can't retry name resolution cheaply.

To the ballots. The headline features still can't find a two-thirds. Seifeddine Gmati's Bound-Erased Generics is, as of this recording, sitting at 7 in favour, 19 against, 10 abstaining — and that's the primary; on the syntax sub-vote the C#-style in/out spelling is killing it, 24 to 2. Nicolas Grekas's __exists() magic method is underwater too, 2 to 12 with 7 abstentions. And Daniel Scherzer read the room on ReflectionAttribute::getCurrent() — stuck at 4-7-8 — and simply pulled the vote to rework it.

The focused changes, meanwhile, are clearing without a single No. Tim Düsterhus's deprecation of returning values from constructors and destructors: 37 to nothing as of recording. Sjoerd Langkemper's cap on php://filter chains — last week's local-file-inclusion-to-RCE fix — running 30 to nothing, 2 abstaining. Weilin Du's Locale display-keyword additions are in — the vote closed and the RFC's accepted, 18 to nothing. And Jordi Kroon's plan to lift third-party extension docs out of the manual is cruising at 21 to 1 — though the room is split on where they should live.

Tough week to be a revival. Jorg Sowa brought back case-sensitive PHP, took a full week of fire — Rob Landers didn't hedge: "I honestly can't think of anything good that this RFC would bring" — and on Tuesday retired it himself, writing: "I'm sorry, but I'm going to stop work on this RFC. I see very little chance it passing." James Titcumb formally walked away from the deprecate-PEAR RFC, realising it isn't really the project's call to make, and explained: "[D]eprecating PEAR is not really within the remit of the wider PHP community [...] it's like making an RFC to deprecate Packagist." Quieter movers: Literal Scalar Types reached v0.2 with strict matching by default, and the .phpc pure-code idea mutated into an alternate opening tag, <?psf, that could even carry a new_scope directive.

On the governance bench: Khaled Alam's "write to objects held in constants" RFC cleared its cooldown and is headed to a vote June 29. And Ben Ramsey reshaped the Working Groups proposal — moving the real policy text out into a pull request so the RFC itself is just charter-and-link — still trying to give the project a little structure without crowning a BDFL.

One human note to end on. A few weeks back a newcomer, Sepehr Mahmoudi, floated a get_favicon() function and got the classic "what problem does this actually solve" welcome. This week he withdrew it himself — better suited to userland, he decided — pivoted to a str_mask() helper, heard Rowan explain that real masking needs graphemes and ICU rather than raw bytes, and pivoted again to grapheme_mask() for the intl extension, working C prototype in hand. And as of this morning he's got the keys to write it up — Ilija Tovilo granted his RFC karma and signed off with a simple "Good luck!" That's the list at its best: someone turning up, listening, and getting better in the open.

Quick hits. Sjoerd Langkemper is everywhere this week — beyond the filter-chain vote, he wants base_convert to stop silently shredding entropy when you build random tokens with it, and he's adding a ValueError for invalid characters to hexdec, bindec and octdec. Osama Aldemeery floated deprecating return inside a finally block — the one silent abrupt-exit PHP still allows — backed by an impact analysis showing it hits just 12 sites across 9 packages, 3 of them actual latent bugs; it's folding into the omnibus. Pratik Bhujel's terminal extension reached v0.5.0. Gianfrancesco Aurecchia floated an Openssl\Dtls class to bring DTLS — the UDP side of TLS, the piece WebRTC needs — into the openssl extension, proof-of-concept already green. And 8.4.23 and 8.5.8 both cut release candidates.

That's the week: a deprecations RFC that became an argument about evidence and respect, a stopwatch class that swallowed the mailing list, two revivals laid to rest, and the headline votes still stalled while the small ones glide through. Links to every thread are below. We're Artisan Build. See you next week.