> :robot: _This was written by an AI agent on behalf of @lussoluca._ ## Problem `TraceableBigPipe` overrides `sendPlaceholders()` and `sendNoJsPlaceholders()` by copying the core method bodies and inserting one line each. Those copies have drifted from `Drupal\big_pipe\Render\BigPipe`, and every divergence is a behavior regression that only appears when webprofiler is installed. The `ajax_page_state` drift is tracked separately in project/webprofiler#3579662. The remaining ones, all comparing `src/Render/TraceableBigPipe.php` against `core/modules/big_pipe/src/Render/BigPipe.php` on Drupal 11.4.5: - **Placeholder errors are fatal.** Lines 59, 94, 175 and 213 use `E_USER_ERROR`; core uses `E_USER_WARNING` at lines 393, 429 and 635. `E_USER_ERROR` halts PHP, so the `continue` on the next line is unreachable and one failing placeholder aborts the whole response mid-stream instead of leaving the others intact. The file also declares `strict_types=1` while core's `BigPipe` does not, so passing `$e` to `trigger_error()`, which takes `string $message`, raises an uncaught `TypeError` rather than coercing through `__toString()`. The error path therefore fails before it can report anything. - **`EnforcedResponseException` is not handled.** Core catches it at lines 585 to 628 and emits a `RedirectCommand`, validating through `SecuredRedirectResponse` and `LocalRedirectResponse` and refusing external targets. `TraceableBigPipe` has only the generic `catch (\Exception)` at line 88, and since `EnforcedResponseException` extends `\Exception` it lands there and hits the point above. A form using GET as `#method` built inside a lazy builder no longer redirects. - **Fibers are gone.** Core wraps each placeholder in a `\Fiber` at line 501 and suspends with `FiberResumeType::Immediate` so slow placeholders interleave. `TraceableBigPipe` renders them sequentially, so total time becomes the sum of the individual renders and a BigPipe response nested inside an outer fiber never yields to it. For a profiler this is self-defeating: the timings it reports are not the timings production has. - **Messages raised inside placeholders are lost.** Core calls `$this->messenger->deleteAll()` at lines 544 to 551 and re-adds each message as a `MessageCommand`. `TraceableBigPipe` never touches the messenger, so a message set while rendering a placeholder is not shown on that page and surfaces on the next request instead. - **Cacheability debug output is missing.** Core accumulates `CacheableMetadata` per placeholder at lines 494 and 530 to 532 and emits the `big_pipe cache tags` and `big_pipe cache contexts` comments at lines 642 to 645. Neither exists in `TraceableBigPipe`, so those comments disappear when `http.response.debug_cacheability_headers` is on. ## Proposed resolution Both overrides exist for one reason: to set `X-Drupal-BigPipe-Placeholder` on the embedded response before `filterEmbeddedResponse()` dispatches `KernelEvents::RESPONSE`, which is where `ProfilerListener::onKernelResponse()` line 135 and `RequestDataCollector::collect()` line 60 read it. Core already exposes two `protected` seams called once per placeholder, so the tracing can wrap them instead of reimplementing their callers: - **`renderPlaceholder()`** receives the placeholder ID. Call the parent, then mark the returned elements with the ID. - **`filterEmbeddedResponse()`** receives the response immediately before the dispatch. Read the mark, remove it, set the header, then call the parent. The mark travels on the response rather than in a property because core interleaves fibers: with `\Fiber` suspension the most recently rendered placeholder is not the one the current loop iteration is about, so a "current placeholder" field would attribute profiles to the wrong placeholder. Carrying it on the response is immune to that. It has to be removed before the dispatch because `HtmlResponseAttachmentsProcessor` throws a `LogicException` on any unknown `#attached` key at lines 88 to 105. The mark reaches both paths already: the JS path calls `$ajax_response->setAttachments($elements['#attached'])` at `BigPipe.php:542`, and the no-JS path goes through `HtmlResponse::setContent()`, which calls `setAttachments($content['#attached'])` at `HtmlResponse.php:65`. The private copy of `splitHtmlOnPlaceholders()` goes away too. It is `private static` in core at `BigPipe.php:777` and only existed to serve the copied `sendNoJsPlaceholders()`. The two embedded responses that are not placeholder replacements stay untouched, because neither is ever marked: the bottom JavaScript chunk at `BigPipe.php:317` and the redirect built in the `EnforcedResponseException` branch at `BigPipe.php:615`. This takes the class from 271 lines to 71 and removes the drift by construction rather than by correction, since none of the core logic is duplicated any more. ## Verification On Drupal 11.4.5, webprofiler 11.2.1, PHP 8.4, with `big_pipe` enabled and an uncacheable JavaScript-driven block on the page: | Check | Result | | --- | --- | | `data-once` on the block element | `ai-deepchat ai-chatbot` | | `typeof Drupal.bigPipe` | `object` | | `Object.keys(Drupal.behaviors).length` | 23 | | `script[src*=jquery]` elements | 1 | | Toolbar blocks rendered | 11 | | Console errors | 0 | | Child profiles with placeholder attribution | 6, one per placeholder | | No-JS path with `big_pipe_nojs=1` | HTTP 200, no `LogicException`, no leftover placeholder markers | The child profile attribution matches a baseline captured before the change, including the placeholder for the chatbot block. ## Note on sequencing This supersedes the two lines fixed in project/webprofiler#3579662 by deleting the methods that contain them. That fix is worth landing on its own first, since it is small enough to backport, and this change is best reviewed separately.