StreamedChatMessageIterator buffer corrupts HTML when consuming streamed responses server-side (relative URLs split mid-attribute)

[Tracker]
Update Summary: Buffer flushes mid-attribute on relative URLs, corrupting HTML in server-side consumers
Short Description: StreamedChatMessageIterator buffer corrupts HTML when consuming streamed responses server-side (relative URLs split mid-attribute)
Check-in Date: 03/18/2026
Metadata is used by the AI Tracker. Docs and additional fields here.
[/Tracker]

Problem/Motivation

When using the AI module's Chat operation type to process HTML-rich content server-side (e.g. via ai_integration_eca or any server-side batch consumer), the output HTML is consistently corrupted. Attributes are split mid-value, tags are broken, table structures are destroyed, and src/href URLs are truncated.

The root cause is in StreamedChatMessageIterator, where the buffer flushes when $maxBufferSize (100 characters) is reached or on every newline. This causes the buffer to flush mid-attribute when processing HTML content with relative URLs like src="/sites/default/files/...".

Note: as of Drupal 11.3, Fibers are supported and any AI provider call runs in streaming mode — including server-side consumers like ai_integration_eca. There is no way to avoid this code path.

Steps to reproduce (required for bugs, not feature requests)

1. Install Drupal 11.3+ with the AI module and ai_integration_eca
2. Configure an ECA Chat action that processes a node body field containing images with relative URLs (src="/sites/default/files/..."), HTML tables, and links
3. Trigger the ECA model
4. Observe the output — HTML attributes are split, tags are broken, table structure is destroyed

Example of corruption observed

Debugging with Xdebug inside flushInternal(), two consecutive flushes were observed:

Flush 1 (buffer limit reached mid-attribute):
src="/sites/default"

Flush 2 (rest of the URL):
/files/inline-images/image.png" width="2816" height="1536">

Resulting in broken HTML with the rest of the URL orphaned outside the tag.

Proposed resolution

1. Fix the regex in shouldFlush() to protect relative URLs starting with /:

// Before — only protects absolute URLs
if (preg_match('/"(?:http|\/\/:)[^"]*$/', $this->buffer)...

// After — also protects relative URLs
if (preg_match('/"(?:http|\/)[^"]*$/', $this->buffer)...

2. Consider increasing $maxBufferSize or making it configurable to reduce the risk of mid-attribute flushes on long attribute values.

Remaining tasks

• Fix the regex in shouldFlush() to protect relative URLs
• Add kernel test that verifies HTML with relative URLs is not corrupted when consuming the iterator server-side

User interface changes

None.

API changes

None.

Data model changes

None.

Release notes snippet

Fixed HTML corruption in StreamedChatMessageIterator when processing content with relative URLs server-side (e.g. via ECA or batch processing).