> [!note] > > Active MR: !41 ## Summary AI Translate post-processes the model's response before writing it to the field, and two steps of that post-processing corrupt valid content. This is not the model ignoring the prompt: the shipped prompt already instructs it to preserve HTML, and both defects reproduce with a stub provider that returns a fixed string, with no model involved. 1. **Entity decoding.** `EntityTranslationOrchestrator::translateTextMetadataItem()` runs `html_entity_decode()` on every translated column, unconditionally. On a formatted field this turns `&amp;` inside an `href` into a bare `&`, drops `&nbsp;`, and turns an intentionally escaped `&lt;strong&gt;` into a live tag. 2. **Fence and quote stripping.** `TextTranslator::translateContent()` cleans the response with `trim($cleaned, '```')` and `trim($cleaned, '"')`. `trim()`'s second argument is a character mask, not a literal suffix, so it removes _any_ backtick or double quote at either end of the value: `He said "stop"` is stored as `He said "stop`. Both affect 1.3.x, 1.4.x and 2.0.x — the three branches carry the same code here. ### Scope Two further causes were reported in this thread and are tracked elsewhere, so that closing this issue does not silently close them: - HTML tags split and re-closed mid-sentence on longer values. That is the AI module's streaming filter, not this module: project/ai#3586558. - A model returning prose with no markup at all. Prompt and model behavior rather than a code defect: follow-up issue TBD. ## Steps to reproduce 1. Install `ai_translate` and `content_translation`, add a second language, and set a chat provider as the default for the `translate_text` operation. 2. Create a content type with a body field using the Full HTML text format. 3. Create a node with the title `He said "stop"` and the body `<p>Fish &amp; chips cost 10&nbsp;euros.</p>`. 4. Open the node's Translate tab and translate it into the second language. 5. Inspect the stored values: the translation's edit form with the source view, or a direct database query. ## Expected result The stored translation matches what the model returned: entities still encoded, and the title's trailing quote character intact. ## Actual result - Body: `&amp;` is stored as a bare `&`, and `&nbsp;` as a raw U+00A0. - Title: the trailing `"` is gone. - An intentionally escaped `&lt;strong&gt;` anywhere in the body becomes a live `<strong>` tag. ## Proposed resolution See MR !41. 1. Decode entities only when the value came back genuinely escaped - it contains `&lt;` and no real `<` - while continuing to decode plain-text columns, which are escaped at display time and therefore do still need it. 2. Replace the character-mask `trim()` calls with an anchored match that unwraps a code fence only when the fence wraps the whole response, and unwraps quotes only when they wrap the whole value. ## Remaining tasks - [ ] Fix the plain-text regression in !41 - [ ] Widen the fence match so a leading newline or an unclosed fence is still handled - [ ] Re-test, including a plain-text field and a real provider - [ ] Port to 2.0.x and 1.3.x ## Environment - Drupal version: 10.4 / 11 - Module version: 1.3.x, 1.4.x, 2.0.x - PHP version: 8.1+ - Provider: any — both defects reproduce with a stub provider - Last known working version: not applicable, present since the module was split out of AI ## AI Usage - [x] AI Assisted Issue — This _edited_ issue description was generated with AI assistance, but was reviewed and refined by the creator. <details> <summary> Original report by @cbccharlie (11 June 2026), preserved verbatim </summary> ## Problem/Motivation When using AI Translate to translate content that contains formatted text fields (`text_with_summary`, `text_long`) with HTML markup (e.g. `<p>`, `<strong>`, `<ul>`, `<a href="...">`, etc.), the translated output does not preserve the original HTML structure. The AI model either strips, alters, or partially reconstructs the HTML tags, resulting in broken or invalid markup stored in the translated field value. This affects any content type that uses fields with a text format such as "Full HTML" or "Basic HTML" via CKEditor. ## Steps to reproduce 1. Install and configure the `ai_translate` submodule (Drupal AI). 2. Create a content type with a `text_with_summary` or `text_long` field using a text format that allows HTML (e.g. "Full HTML"). 3. Create a node with rich HTML content in the body field — e.g. paragraphs, bold text, lists, or links. 4. Navigate to the node's translation tab and trigger an AI translation to any target language. 5. Inspect the translated field value (edit form or database). ## Actual behavior The HTML markup in the translated field is corrupted: tags may be stripped, improperly nested, partially reproduced as plain text, or the entire value may be returned as plain text without any HTML structure. </details>