Known issues
This page lists known problems in Context Control Center (CCC) that are not fixed yet, with a workaround where one exists. An entry is removed when a release fixes the problem.
The Content editor traps keyboard focus
Who is affected: people who use only the keyboard to edit context items.
The Content field on the context item form uses the Markdown editor from the AI module (MDXEditor). In rich text mode, Tab inserts a tab character and Shift+Tab stays in the editor, so Tab does not move focus out of the editor. This fails success criterion 2.1.2 (No Keyboard Trap) of the Web Content Accessibility Guidelines 2.2.
Workaround:
- Press Escape, then Tab, to move focus out of the editor, to the next control.
- Press Escape, then Shift+Tab, to move focus to the editor toolbar.
- In Source mode, Tab moves focus out of the editor.
Fix: planned in the AI module, in project/ai#3586814.
Affected AI CKEditor releases prevent the service container from building
Who is affected: within the Drupal versions CCC supports, sites on Drupal 10.5.x or 11.2.x that enable AI CKEditor 1.4.3–1.4.4 or 1.5.0-beta1 through 1.5.0-beta5. The CCC consumer for CKEditor first shipped in AI CKEditor 1.5.0-beta2.
The affected releases register Drupal\ai_ckeditor\Hook\AiCKEditorHooks
with autowire: true. Its constructor asks for
KeyValueExpirableFactoryInterface, whose service alias first appeared in
Drupal 10.6.0 and 11.3.0. On the affected Drupal series, container builds
fail and drush cr reports "Cannot autowire service ... AiCKEditorHooks".
Workaround: upgrade to Drupal 10.6 or later on the 10.x branch, or to Drupal 11.3 or later on the 11.x branch. Use a release of AI CKEditor that addresses SA-CONTRIB-2026-187, which affects AI CKEditor 1.4.2.
Fix: tracked in AI CKEditor in project/ai_ckeditor#3615859.
The Load tool does not apply the Roles scope
Who is affected: sites that use the Roles scope to tailor context
delivery and let agents use the Load tool
(ai_context:load_ai_context_items).
Ordinary selection for pushed context, Get Relevant and List applies the Roles filter. Always include overrides that filter. Load can return a directly requested item regardless of Roles if its translation is published and the current user has view access. So a user who may use an agent with the Load tool can read an item whose roles they do not have, if they know or guess its ID.
Workaround: limit the Load tool to agents whose users should receive the context those agents can load. Roles is not an access boundary: to keep an item out of AI tool responses completely, unpublish it or keep it out of CCC.
Follow-up: under discussion for 1.1.0 in #3586707.
An automator call without context can send the context of an earlier call
Who is affected: sites that push context to an AI Automator whose base field has more than one value, with a provider that sends the system role of the provider instance, for example an OpenAI-based provider.
AI Automators send all calls for one field through one provider instance,
and send no system prompt of their own. AI core copies a system prompt back
to the provider instance only when the prompt is not empty. So when one call
pushes context and a later call on the same field does not push, the later
call still sends the context block of the earlier call. The debug log shows
no_items or push_disabled for the later call, and the Usage tab does not
record it. For example, this can occur in Relevant mode when no item matches
the value of the later call. See
Consumer configuration.
Workaround: make every call on the field push context, for example with a published Global context item, which is selected in every selection mode, or turn off Push context automatically for the automator.
Fix: AI Automators must reset the system role for each call. It is reported in the AI module, in project/ai#3586900.