README

Varbase AI Content Operations Base

An opt-in Drupal recipe that gives the editors of a Varbase 11 site a content operations assistant in the admin toolbar. Editors ask in their own words, and it audits, reports on and updates the site's content:

"Which pages have broken internal links or missing alt text?"
"Show me articles with no meta description and draft one for each based on the body."
"Which content types or taxonomy terms have no published content, and which tags are near-duplicates?"
"List everything published in the last 30 days by author and section, and flag anything still in draft for more than two weeks."
"Find all articles mentioning [old name] and replace it with [new name], then show me the diffs before saving."

It works with the editor's own permissions, shows every change first, and saves only what the editor approves, as a new revision with a log message. It never deletes, and it never touches users, permissions or the content model.

It is not the visitor chatbot. Varbase AI Chatbot Base answers visitors from published content and cannot change anything; this assistant is for signed-in editors only.

Experimental. Tool API and Tool Belt have no stable release (1.0.0-beta11 and 1.0.0-alpha6 when this was written). There is no custom PHP and no custom shell code in the recipe. Its no-key checks run inline in CI, so the repository holds no PHP file at all.

What it adds

Part	What it does
Modules	Installs `ai_assistant_api`, `ai_chatbot`, `tool`, `tool_ai_connector` (Tool API tools as AI function calls), `tool_belt`, `tool_belt_content` and `tool_belt_entity`.
Agent `varbase_ai_content_operations`	"Content operations assistant". Seven tools: list content, read image alt text (pinned to image media), look up by a field value, list content types and vocabularies, load one item and read its full field values (for a text field's format), and update one item as a new revision. No tool that creates, deletes, or changes fields or configuration.
Assistant `varbase_ai_content_operations`	Runs the agent for Varbase's editorial roles only: Content editor, Content admin, SEO admin, Site admin (and administrators). Low temperature, conversation kept per user.
Block `gin_varbase_ai_content_operations`	The chat in the Gin admin toolbar, shown to those editorial roles.
Permissions	`access deepchat api` for the four editorial roles.

How it stays safe

The editor's own permissions. Every tool checks access as the signed-in editor: listing skips what they cannot view, field by field, and a save is refused on anything they cannot edit.
Approval before any save. It shows the item, the field and the value before and after, asks "Shall I save these N changes?", and saves only what the editor approves in the next message. A "yes" for changes it has not shown yet saves nothing: it shows them and asks.
Text formats kept. Before saving a formatted field it reads the item's current format and sends it back unchanged; it never picks a format.
A revision for every save, logged as "Content operations assistant: ..., approved by ". Undo is the usual Revisions tab. Status, author, moderation state and dates are never changed.
Honest reports. It says what it checked and what it could not, and reports only what the tools returned.
Content is data. Instructions written inside content are ignored.

Works on every Varbase site template

Field names differ between content types, so the agent reads one item of each type before a task and uses the Varbase conventions: body text in field_content (or body), summary in field_description, meta description in field_seo_description, featured image in field_featured_image, and the type's own category field as its section. Those are the fields of Varbase Starter, Educare, Horizon Aid and The Rightup.

Requirements

Drupal core ^11.4, Varbase 11, with Varbase's editorial roles (content_editor, content_admin, seo_admin, site_admin) and the Gin admin theme
drupal/ai ^1.5 and drupal/ai_agents ^1.3
drupal/tool ^1.0@beta and drupal/tool_belt ^1.0@alpha
A default chat model that supports tools (Configuration > AI > AI settings), as Varbase AI Base sets up. The answers in this README came from Claude Opus.

Install and apply

# 1. Fetch the recipe.
ddev composer config --json extra.drupal-recipe-unpack.ignore '["drupal/varbase_ai_content_operations_base"]'
ddev composer require drupal/varbase_ai_content_operations_base:1.0.x-dev

# 2. Apply it.
ddev drush recipe ../recipes/varbase_ai_content_operations_base
ddev drush cr

# 3. Check it. Without arguments no AI key is needed. "ask <uid> <question>" asks the agent one real question as
#    that user.

Re-applying the recipe is safe. The ignore line keeps the recipe a locked package: without it core's recipe-unpack plugin moves its requirements into the project's composer.json and drops the recipe itself.

Then sign in as an editor, open any admin page and open Content operations in the toolbar.

Automated testing

GitLab CI on git.drupalcode.org (.gitlab-ci.yml) runs on every merge request and every push to 1.0.x, like the other Varbase AI recipes: composer validate and a lint of every YAML file.

The functional tests in tests/features (Playwright and Cucumber through @vardot/varbase-e2e) run against a site of your own: yarn install, then LAUNCH_URL=https://your-site.ddev.site yarn test.

Checked

On 11 October 2026, on a Varbase 11 development site with 21 content items written by three authors, the five example questions were asked as a Content admin, through Claude Opus with the real tools:

Question	Time	What it did
Broken links and alt text	70 to 90 s	Found both broken internal links, the image with no alt and the one with an empty alt, and a featured image whose alt text is a file hash.
Missing meta descriptions	37 s	Drafted five descriptions of 122 to 143 characters; after "yes", saved five revisions.
Content model hygiene	78 s	Found the content type with nothing published, the unused tags, and the AI / Artificial Intelligence, SEO / Search Engine Optimization and Open-Source / Open Source pairs.
Editorial report	51 s	Grouped ten items by author and section, totals adding up, and flagged the five items stalled for more than 14 days.
Rename a term	37 s, then 42 s	Showed 10 occurrences in 9 fields of 5 items, a draft included; after "yes", saved them with the text format kept.

Each save was checked afterwards: one new revision per item, by the editor, with the log message, and the status, moderation state, author and text format unchanged.

The content model question was also asked through the toolbar chat in a browser, as a Content admin: the answer came back in about 100 seconds as short lists, which is what the narrow chat panel shows well (it does not show tables).

Known issues

Saves can fail on sites with Varbase AI Taxonomy Tagging. That recipe creates the ai_automator_status field with no allowed values, and AI Automators then stores "finished" in it, so every validated save outside the edit form fails with "ai_automator_status.0: The value you selected is not a valid choice". The assistant reports the failure and changes nothing. The fix is in Varbase AI Taxonomy Tagging; until then, an administrator can give the field storage its four values (pending, processing, failed, finished).
AI Agents' list_content_entities ends the whole chat request when it is asked for a field an item does not have or the editor cannot view: it throws instead of returning a tool error, and the chat shows only "Sorry, I could not finish that". The recipe pins that tool to image media (name and field_media_image, for alt text) with tool_usage_limits; everything else goes through Tool Belt's entity_list, which skips missing fields.
Tool Belt's entity_update drops a text field's format when it is given only the value, and entity_list returns text fields without their format. The agent therefore reads the format with entity_load_by_id and entity_field_values before it saves a formatted field.
Raise the AI request timeout. A step of a tool-heavy answer can take more than the AI module's default 60 seconds ("cURL error 28"). Set Configuration > AI > AI settings > request timeout to 180, or drush config:set ai.settings request_timeout 180.
Content admins cannot list URL aliases or content types (those need "administer url aliases" and "administer content types"). The assistant reads each item's own path field instead, and says when it could not list the content types.
Answers take 20 to 90 seconds and several model calls: it reads the content in pages of 50. On a large site, ask for one content type or section at a time.
It suggests merging near-duplicate terms but does not merge them: that means retagging content and deleting a term, which it never does.

Related recipes

Recipe	What it does
Varbase AI Content Review Base	Scores one content item against the site's style guide, from its Review tab.
Varbase AI Chatbot Base	A visitor chatbot that answers from published content only.
Varbase AI Answers Base	Ask AI: cited answers over the site's vector index.

License

GPL-2.0-or-later. See LICENSE.

drupal / varbase_ai_content_operations_base

Package info

Security

Statistics