ai_context 1.0.0-rc2

Changes since 1.0.0-rc1 (compare)

1.0.0-rc2

64 credited issues

The 1.0.0-rc2 release of Context Control Center (CCC) is the second release candidate. It gives QA a fixed build to test the changes made after rc1, before the 1.0.0 release. Most of this release comes from a security and bug review of the code before 1.0. CCC now delivers only published translations of context items, filters subcontext children by view access and hard-filter scope, and reads the Entity Types bundle only from entities that the user can view. The core "view any unpublished content" permission no longer shows unpublished context items, Diff pages require the revision permission, Duplicate always saves the copy as a draft, and context items and usage records are no longer available through JSON:API. Drupal tokens in context item text are no longer replaced on any path, so item text cannot send user data to the AI provider. When scope cleanup removes the last value of a hard-filter scope, the item is unpublished, and the uninstall of Dynamic Entity Reference is blocked so that Specific Entities values are not deleted. CCC now requires drupal/ai 1.4.11 or later on 1.4, or 1.5.2 or later, because older versions show tooltip text as HTML.

New sites get a default token budget of 20,000 instead of 1,200. Existing sites keep their saved value. The stored token count of each item now includes its rendered header, so the Tokens column matches what the budget uses, and the truncation warning explains what it means. Usage records and invocation results count only the items that the budget delivered, "Last used" changes each time an item is used, usage records keep a limited number of request paths, and CCC uses approximate token counts when the tiktoken vocabulary cannot load. Other fixes keep consumer overrides during any module uninstall, resolve duplicate consumer IDs the same way everywhere, set the moderation state when Scheduler publishes on save, and vary Site Sections results by request path. Site builders can add CCC to the admin menu, the item and consumer listings now show published items and pushed consumers by default and have Reset buttons, and the scope settings table and status badges have accessibility fixes. For developers, the new load_ai_context_items function call replaces load_ai_context_item_by_id, which is deprecated and will be removed in 2.0.0. The entity_context keys are documented as public API, and the docs are updated for 1.0.0 with a new known issues page. This release also credits the work for the DrupalCon Rotterdam Inside AI demo, which configured CCC for AI Answers, Canvas AI, CKEditor and automator consumers. Run the database updates after you upgrade.

Upgrade notes

Update the code

• CCC now requires drupal/ai 1.4.11 or later on 1.4, or 1.5.2 or later (#3586712). Update both modules together: composer update drupal/ai_context drupal/ai -W. To stay on AI 1.4, add --with drupal/ai:~1.4.11. If your site uses a drupal/ai dev branch, update it to the current branch head.
• If you use Document Loader: Context File Importer, require Document Loader 2.0.7 or later, which needs AI 1.5 or later: composer require 'drupal/document_loader:^2.0.7' (#3586650).
• Run drush updatedb, then drush cr. Update 10032 recounts the token count of every revision and translation of every context item (#3586677), so it can take some time on sites with many items.

Check permissions

• Core View any unpublished content and View the latest version no longer give access to unpublished context items. Give View any unpublished AI Context Item to the roles that must see context drafts, for example Content editor on Drupal CMS (#3586701).
• Duplicate now always creates a draft copy, and it needs the AI Context editorial workflow: Use Create new draft transition permission (#3586706).
• Diff pages for context items now need View all AI Context Item revisions (#3586703).
• The Entity Types scope reads the bundle only from content that the current user can view. Give editor roles permission to view the content that they edit. Background jobs need an account that can view the entity, or a bundle hint (#3586593).

Check settings and content

• Existing sites keep their Max tokens for injection value. New sites get 20,000. After update 10032, the token badges and budget totals show higher numbers, and a consumer that is near its budget can show Context budget exceeded. Raise the site or consumer budget if necessary (#3586648, #3586677).
• Scope cleanup now unpublishes an item when it removes the last value of a hard-filter scope. Sites that ran post-update 0006 on 1.0.0-beta5 or 1.0.0-rc1 should review published items that have no Roles, Languages, Site Sections, Entity Types, Taxonomy Terms or Specific Entities value (#3586603).
• To review copies that earlier versions published through Duplicate, look for published items that have the same label or the same created date as another item (#3586706).

Changes that can affect integrations and custom code

• Context items and usage records are no longer available through core JSON:API. There is no replacement for sites that read or synced context items this way, for example with Entity Share (#3586619).
• Drupal tokens in context item text are no longer replaced. For example, [user:mail] is sent as [user : mail] (#3586618).
• The Load tool has a new ID, ai_context:load_ai_context_items. The old ID, ai_context:load_ai_context_item_by_id, still works but is deprecated and is removed in 2.0.0. Change agents, default information tools, agent overrides and prompts that use the old ID by hand (#3586696).
• Code that caches selection results must use the cache metadata of the result. Results now vary by the new ai_context.request_path cache context and by the view access of the request entity (#3586687, #3586593).

Known issues

• On AI 1.4.11, context that CCC pushes on the generic path (automators and other consumers that use PreGenerateResponseEvent) does not reach OpenAI-based providers. The result status and the usage records still say pushed. The fix needs AI 1.4.12, which is not released yet. Use AI 1.5.2 or later to get pushed context on these consumers (#3586709).
• On AI 1.5.2 and later, an automator in base mode on a multi-value field adds one more context block for each field value, so the third request sends three blocks. The fix is planned for 1.0.0 (#3586709).
• A user who can translate context items, but cannot edit them, can change the shared settings of an item from the translation form: the scopes, Priority, Specific Entities, Same as parent, Parent context and Subcontext type. The change applies to all languages. Until the fix in 1.0.0, give translation permissions only to roles that can also edit context items (#3586721).

Contributors

People: kristen pol (64), pravesh_poonia (33), kieran.cott (14), dcgoodwin (13), jucs7 (10), ajabhinavjha30 (8), csakiistvan (7), cadence96 (6), kati.vera (6), afoster (5), ahmad-khalil-imagex (5), emma horrell (4), lbesenyei (4), shqshqnk (4), ultimike (4), tazo_vekua (3), tushar1 (3), joekrukosky (3), yautja_cetanu (2), cmoeser (2), robloach (2), dschmidtadv (2), a.dmitriiev (1), bbruno (1), davidbp (1), aivazashvilit (1), akhil babu (1), marcus_johansson (1), brunaemerich (1), pranavdoodleweb (1), scott falconer (1), scott_euser (1), aherczeg (1), jeroen dost (1)

Organizations: Salsa Digital (64), Itty Bitty Byte (64), OpenSense Labs (40), DrupalFit (40), Reading Room (14), CivicActions (13), esinergia. (10), 1xINTERNET (9), Mearra (6), SeeD EM (6), Foster Interactive Inc. (5), ImageX (5), The University of Edinburgh (4), Omedia (4), Pantheon (4), QED42 (4), Brainsum (4), DrupalEasy (4), Specbee (4), FreelyGive (3), Kalamuna (2), Pare & Co (2), Cool Fire Inc (2), Acquia (1), DoodleWeb (1), Zoocha (1), Soapbox (1)

New Features (5)

• Issue #3586426: Update CCC recipe for Inside Rotterdam demo: CKEditor / non-agent consumer configuration.
• Issue #3586527: Update CCC recipe for Inside Rotterdam demo: Canvas AI agent/consumer configuration.
• Issue #3586573: Add setting to add CCC to main menu.
• Issue #3586577: Update CCC recipe for Inside Rotterdam demo: automator/consumer (non-agent) configuration.
• Issue #3586696: Add the load_ai_context_items function call and deprecate load_ai_context_item_by_id.

Bug Fixes (27)

• Issue #3586516: Automator consumers disappear from listing when consumer type is unavailable.
• Issue #3586541: Make duplicate consumer_id rows resolve consistently and map validator errors to the right row.
• Issue #3586551: Keep consumer overrides during any module uninstall, including automators on automator chain fields.
• Issue #3586571: Scheduled publish on save can leave context item Draft in the listing.
• Issue #3586581: Loading documents for CCC in Rotterdam demo causes fatal error and fails.
• Issue #3586593: Entity Types scope loads the request entity without a view access check.
• Issue #3586603: Scope cleanup can leave an item published with an emptied hard-filter scope.
• Issue #3586605: Correct the Language scope fail-open text and the conditional parent cap label.
• Issue #3586616: Unpublished context item translations can be delivered to AI prompts.
• Issue #3586617: Filter subcontext children by view access and hard scope.
• Issue #3586618: Drupal tokens in context item text are replaced on the agent path only.
• Issue #3586620: Context items listing replaces its cache metadata instead of merging it.
• Issue #3586638: Scope settings table: Edit links do not say which scope they are for (WCAG 2.4.4).
• Issue #3586639: Status badge contrast (WCAG 1.4.3): Unavailable fails on row hover.
• Issue #3586640: API stability docs link to Post-update 0010 uses the wrong anchor.
• Issue #3586645: Usage records and the invocation result count items that the token budget left out.
• Issue #3586649: Uninstalling Dynamic Entity Reference deletes Specific Entities values and removes the item restriction.
• Issue #3586660: Warm the tiktoken vocab cache before the parallel phpunit runs in CI.
• Issue #3586661: Saving or rendering context items fails when the tiktoken vocab cannot be loaded.
• Issue #3586677: Getting a "AiContextSelector: Items cut off due to token limit: 4" warning and I don't understand why.
• Issue #3586678: "Last used" value of "Context usage" table acts more like "first used" when used on new content entities.
• Issue #3586687: Site Sections hard-filters on the supplied request path, but the selection's cache contexts carry only url.path and route.
• Issue #3586701: Core "view any unpublished content" permission shows unpublished context items.
• Issue #3586702: Usage records keep every request path with no limit on number or length.
• Issue #3586703: Diff pages show older context item revisions without the revision permission.
• Issue #3586706: Duplicate publishes the copy for users who may not use the publish transition.
• Issue #3586712: Require drupal/ai 1.4.11 or later on 1.4, or 1.5.2 or later.

Other Major Issues (1)

• Issue #3586700: Split the phpunit CI jobs by test type so that they finish within the runner time limit.

Additional Code Issues (10)

• Issue #3586570: Remove temporary Twig 3.29 development pin in CCC composer.
• Issue #3586586: Update context items listing status filter to published and consumers listing push filter to on and add reset buttons.
• Issue #3586612: Update the README and docs for the 1.0.0 release.
• Issue #3586619: Hide context items and usage records from JSON:API.
• Issue #3586642: Document the return value of getSelectedValueLabels().
• Issue #3586646: Document the entity_context keys as public API and how to pull context from background jobs.
• Issue #3586648: Raise the default token budget for new sites to 20,000.
• Issue #3586650: Correct the feature docs from the rc1 QA.
• Issue #3586664: [Accessibility] Document the Content editor keyboard trap as a known issue before 1.0.0.
• Issue #3586679: Document that full Automator consumer support needs AI core 1.5.0 or later.

Additional Non-Code Contributions (21)

• Plan: 7
• Task: 14

Release notes from merge requests

These notes come from the merge requests in this release. They describe what changed and what to do after you update.

Issue #3586541: Make duplicate consumer_id rows resolve consistently and map validator errors to the right row.

When ai_context.consumers has more than one row with the same consumer ID, the first row now takes effect everywhere: on the Consumers settings page, at runtime and in the consumer editor. Before this change, the listing showed the last row while runtime used the first row, and saving any consumer failed with a "Duplicate consumer ID" error. Now, saving any consumer removes the later rows and shows a warning that names the IDs, and Remove configuration deletes every row with that ID. The UI and config import cannot create a duplicate, so this applies only after a direct write to active config, for example from custom code or a database edit.

Issue #3586551: Keep consumer overrides during any module uninstall, including automators on automator chain fields.

Saved consumer settings now stay when a module uninstall deletes their agent or automator, and deleting one agent or automator outside an uninstall still removes its row. Before this change, an uninstall deleted some rows in ai_context.consumers without a message, for example when you uninstalled ai_automators with automators on automator chain fields, or a module that an agent or automator depends on. After an uninstall, the Consumers page lists the row as Off (type unavailable) or Off (stale) until you use Remove configuration, or until an agent or automator with the same ID exists again. Run drush cr after the update.

Issue #3586570: Remove temporary Twig 3.29 development pin in CCC composer.

None. This change affects only CCC's own development dependencies. composer.json no longer pins twig/twig to ~3.29.0 in require-dev, so CI now tests with Twig 3.30 or later. Sites see no change.

Issue #3586571: Scheduled publish on save can leave context item Draft in the listing.

When scheduling is on, saving a context item with a Publish on date that has already passed now sets its moderation state to Published. Before this change, the item was published, but its moderation state stayed Draft: the context items list showed Draft, while selection and the global count treated the item as published. Publishing through cron for a future date does not change. Run drush cr after the update; this change does not repair items that were saved this way before the update.

Issue #3586573: Add setting to add CCC to main menu.

You can now add a Context item to the top level of the admin menu with the new Add to main menu setting on the CCC General Settings page (/admin/config/ai/context/settings/general). The setting is off by default, and Menu item title changes the link text; run drush updatedb to add both settings to existing sites. The item is in the Navigation sidebar when Navigation is installed, and in the Toolbar admin menu otherwise; it opens the overview page when Show overview page is on, and the context items list when it is off. Scope value checkboxes, for example role names, now show HTML in their labels as plain text.

Issue #3586586: Update context items listing status filter to published and consumers listing push filter to on and add reset buttons.

The context items listing (/admin/config/ai/context/items) now opens with Status set to Published, and the consumers listing (/admin/config/ai/context/settings/consumers) opens with Automatic push set to On, instead of All. Published includes every workflow state that publishes an item, so a published workflow state, for example a custom Approved state, no longer has its own Status option. After you change the search or a filter, a new Reset button shows on both listings and clears the search and restores the default filters, but on the consumers listing it keeps a Type that a ?type= link selected. Run drush cr after the update so that the listings use the new JavaScript and CSS.

Issue #3586593: Entity Types scope loads the request entity without a view access check.

Entity Types scope matching now reads the bundle only from a request entity the current user can view. An entity the user cannot view is treated like an entity ID that does not exist, so items scoped to its bundle are not selected. This also affects:

• edit forms for editors who can edit an entity but cannot view it
• background runs on unpublished or restricted content, such as cron or queue workers running as Anonymous

Give editor roles permission to view the content they edit. Callers that need bundle matching in background runs can run as an account that can view the entity, or send a bundle hint with setEntityBundle() or entity_context metadata. The hint matches only when that account has Use Published AI Context in AI Features.

Selection results for a request entity now also carry the cache metadata of the view access check on that entity, also when access is denied. A result varies by user when the access check varies by user. Examples in core:

• The request entity is an unpublished node, and the current user is authenticated and has View own unpublished content.
• The request entity is an unpublished media item, and the current user has View own unpublished media.

In both cases this applies to the owner and to all other users with that permission. Two users with the same roles therefore do not share a cached result. On sites with a node access module, results for nodes also vary by user.node_grants:view. Users who can view published content through their permissions still share one result. Consumers that cache selection results must use the cache metadata of the result.

AiContextRequestInfoResolver (internal) now also takes the current_user service and has a new method, getContentEntityAccessResult(). After the update, run drush cr. There is no update hook.

Issue #3586603: Scope cleanup can leave an item published with an emptied hard-filter scope.

Scope cleanup now unpublishes a context item when removing a stale value would let the item reach more users, pages, or content. This happens when cleanup removes the last value of Roles, Languages, Site Sections, Entity Types, Taxonomy Terms or a custom stored-value scope, or the last Specific Entities reference. Context Tags only rank items. Removing a tag, including the last one, leaves the item published. Global items stay published.

Every published translation is unpublished; on moderated sites the state becomes Archived. Scheduled publishing of the item is cancelled, and a new revision says why. The user who started the cleanup sees one warning with the number of items, and Recent log messages has one warning for each item. A custom workflow needs an unpublished state that is a default revision state, like Archived. Without one, cleanup does not save the item. The stale value stays stored and is not shown on the item. The user sees a warning, and Recent log messages has an error.

After you delete roles, languages, terms, vocabularies, bundles or referenced content, change Site Section or Entity Types settings, or uninstall modules, check the context item list for newly unpublished or archived items.

Sites that already ran post-update 0006 on 1.0.0-beta5 or 1.0.0-rc1 do not run it again. On those sites, review published items that have no Roles, Languages, Site Sections, Entity Types, Taxonomy Terms or Specific Entities value.

Run drush updb, then drush cr, because the cleanup service has new constructor arguments and there is a new helper service.

Issue #3586605: Correct the Language scope fail-open text and the conditional parent cap label.

Corrected the Language scope help text, the conditional parent limit label, and the consumer help about which scopes leave a context item out (it now names Roles). Behavior did not change.

Translators: seven source strings changed: two Language scope strings, three consumer form help strings, and the title and description of the cap field.

Issue #3586612: Update the README and docs for the 1.0.0 release.

Documentation: The README and the docs no longer say that CCC is in beta. They now say that from 1.0.0, public APIs follow the deprecation policy on the API Stability page, and they link to the upgrade policy for upgrades from beta and release candidate versions. Three headings have new names, so links to the old anchors #beta-api-stability (docs home page), #beta-status-and-release-line and #skipping-intermediate-beta-releases (upgrade policy) no longer open those sections. The docs site now shows !!! note blocks, for example on the Usage Tracking page, as note boxes.

Issue #3586616: Unpublished context item translations can be delivered to AI prompts.

On multilingual sites, an unpublished translation of a context item (for example, Draft or Archived) is no longer delivered to AI prompts or tool output. The published default-language text is used instead, and the item is left out when that is not published either. This applies to every user. Users with Administer AI Context or permission to view unpublished items no longer receive unpublished translations in prompts. A published translation of a draft source is now delivered to non-admin users in that language. Context push no longer fails for non-admin users when a selected item has an unpublished translation in the request language.

Issue #3586617: Filter subcontext children by view access and hard scope.

Subcontext children are now checked before they are delivered with a parent, on both the selector path and the Load tool. Every child is checked on the translation that is delivered: that translation must be published, and the current user must be able to view it. A child with its own scope must also pass the hard filters (for example, Roles) for the current request. A child that fails these checks is no longer sent to the AI provider for the conditional decision. Children that use the parent's scope behave as before. Sites that rely on a custom-scope child reaching users outside its scope must change the child's scope.

Issue #3586618: Drupal tokens in context item text are replaced on the agent path only.

Drupal tokens in context item text are no longer replaced when context reaches AI. Before this change, AI Agents replaced tokens such as [user:mail] in item text with data of the user who ran the agent, and [user:uuid] stopped the agent run with an error. This happened in pushed context and in the output of the List, Load and Get Relevant tools when a site used them as default information tools. CCC now neutralizes Drupal tokens on every path, for example [user:mail] becomes [user : mail]: pushed agent and chat context, the three context tools for every caller, and the public selection API, including text that TEXT_RENDERED subscribers add. The token budget is checked on the neutralized text. Undeclared types that look like tokens, such as [current_user:mail] from ECA, and ECA data paths such as [current_user.mail.value] are neutralized too. Bracketed text that looks like a token, such as [note:intro] or items[i:j], also gets the spaces, and so do some IPv6 addresses, such as [fd12:3456::1], and dotted Markdown links whose first part is a token type, such as [node.js](https://nodejs.org). Other bracketed text, such as times, JSON, URLs and [2], is unchanged. Some forms that other modules replace are not covered, for example [current_user] with ECA, which becomes the ID of the user who runs the agent, and undeclared tokens whose name has whitespace, / or @ before its first colon. See "Authoring trust" in access-boundaries.md. Tokens in the Context prefix setting are still replaced on the agent path. Sites that used tokens in item text on purpose must write the values into the text instead. Items that document real token patterns, for example for Metatag or URL aliases, show them with a space on both sides of the colon. Custom code that needs the item text with its tokens as written must load the context item entity.

Issue #3586619: Hide context items and usage records from JSON:API.

Context items and usage records are no longer available through core JSON:API. Before this change, a site that enabled JSON:API gave every account with Use Published AI Context in AI Features a bulk HTTP read of all published context items with their raw fields, without the CCC delivery rules. Usage records are now an internal entity type, so core REST and other integrations that skip internal types no longer offer them either. Context items remain non-internal for Content Moderation and Workspaces. Sites that read or synced context items through JSON:API, for example with Entity Share, lose that access, and this release has no supported replacement. Run drush cr after the update. Until the caches are rebuilt, the resources stay available. If the site enabled a core REST resource for usage records, delete it before you rebuild the router, that is before you run drush cr or drush updatedb: run drush config:delete rest.resource.entity.ai_context_usage, and remove it from config sync. Otherwise the router rebuild fails, because the REST plugin for usage records no longer exists. Workspaces now ignores usage records, so usage is also recorded while a non-default workspace is active.

Issue #3586620: Context items listing replaces its cache metadata instead of merging it.

The context items listing (/admin/config/ai/context/items) now keeps its ai_context_item_list cache tag and its view access cache data, such as the user cache context for Draft items. Before this change, the listing replaced this cache data instead of merging it. Sites with default caching see no change, because Dynamic Page Cache does not cache admin routes. If custom code makes this listing cacheable (for example, by removing _admin_route from its route), new items now show without a cache clear, and a cached listing no longer shows a Draft item to a user who cannot view it, or hides a Draft item from its owner.

Issue #3586638: Scope settings table: Edit links do not say which scope they are for (WCAG 2.4.4).

On Scope Settings (/admin/config/ai/context/settings/scope), each Edit link now has an accessible name that includes its scope, for example "Edit Global", and the scope name is now the row header of its row. Before this change, screen readers announced every link only as "Edit", so users could not tell which scope each link opened. The visible link text is still Edit, so voice control users can still say "Edit". Custom tests or theme code that look for the scope name in a td cell must now look for a th cell with scope="row".

Issue #3586639: Status badge contrast (WCAG 1.4.3): Unavailable fails on row hover.

Status badges on CCC admin pages now keep the same contrast when the mouse is over their table row. Before this change, the color of a hovered row showed through the semi-transparent badge colors, and the Unavailable badge on Scope Settings fell below the 4.5:1 contrast that WCAG 2.2 Level AA requires. This change does not fix the red badges, for example Disabled and Unpublished, which have low contrast in light mode also without hover (#3586662). Run drush cr after the update so that the changed CSS loads.

Issue #3586640: API stability docs link to Post-update 0010 uses the wrong anchor.

Documentation: On the API Stability page, the "Configuration: Post-update 0010" link now opens the "Post-update 0010: tool consumer_id limits" section of the Configuration page. Before this change, the link used an incorrect anchor and opened the Configuration page at the top.

Issue #3586642: Document the return value of getSelectedValueLabels().

Documentation: The docblock of AiContextScopeInterface::getSelectedValueLabels() and the Scope API docs now say what the method returns. The result is a list of display labels, not a map keyed by value. The list can be shorter than the selected values, so custom code must not pair labels with values by position. getSelectedValueLabelsWithCacheability() and AiContextScopeBase::mapSelectedValueLabels() now declare @return array instead of @return string[], because a label can be translatable markup.

Issue #3586645: Usage records and the invocation result count items that the token budget left out.

Usage records and the generic-path invocation result (item_ids and item_count) now list only the context items that reached the prompt, so an item that the token budget left out is no longer recorded as used. An item that the budget only shortened still counts as used, and existing usage records do not change. For custom code, AiContextSelectionResult has the new methods getOmittedItemIds() and getDeliveredItemIds(), AiContextSelectionTextRenderedEvent has the new method getOmittedItemIds(), and AiContextSelectionResult::getSelectedItemIds() does not change (see "Beta upgrade: usage and item_ids count delivered items" in docs/developers/api-stability.md). If the budget leaves out every selected item but an AiContextSelectionEvents::TEXT_RENDERED subscriber added text, that text is still pushed, with status pushed and an empty item_ids list.

Issue #3586646: Document the entity_context keys as public API and how to pull context from background jobs.

Documentation: The developer docs now list the entity_context request metadata keys and the top-level path key as public API, with the rules that CCC applies to them (for example, CCC ignores an entity ID that is not numeric, such as a UUID). A new "Background jobs and queue workers" section in the consumer API docs tells how to pull context from a queue worker, a cron hook or a Drush command. It says that the account that runs the job decides what is selected, that entity context stays on the current request between pulls, that pull does not enforce the push gates, and that pull writes no usage record. The usage tracking docs now say that a usage record lists the selected items before the token budget is applied, so a record can list items that were shortened or not delivered.

Issue #3586648: Raise the default token budget for new sites to 20,000.

New sites now get a default token budget of 20,000 in Max tokens for injection. Before this change, the default and the code fallback AiContextSelection::DEFAULT_MAX_TOKENS were 1,200, which fits only one or two typical context items. Existing sites keep their saved value, because there is no update hook: after the update, check Max tokens for injection at /admin/config/ai/context/settings/general and raise it if your consumers need more context. A consumer can also set its own budget in Max tokens for injection under Advanced on its edit form, for example a higher budget for an agent with many context items.

Issue #3586649: Uninstalling Dynamic Entity Reference deletes Specific Entities values and removes the item restriction.

Uninstalling Dynamic Entity Reference is now blocked while a context item uses Specific Entities on its current revision or a pending draft. Old revisions get a revision log note that lists the entities they referenced, so the values can be entered again by hand.

Issue #3586650: Correct the feature docs from the rc1 QA.

Documentation: The feature docs now include the corrections from the rc1 QA in #3586589 for the overview page, context items, consumer configuration, usage tracking, revisions and scheduling. A new Test page and entity context section in the context delivery docs shows how to run an agent consumer with an entity or page path that you choose, with the AI Agents Debugger and its Spoof Tokens. The docs now specify Document Loader ^2.0.7, which fixes a fatal error during import and needs AI 1.5 or higher. CCC does not enforce this version, so sites that use Document Loader: Context File Importer should run composer require 'drupal/document_loader:^2.0.7'.

Issue #3586660: Warm the tiktoken vocab cache before the parallel phpunit runs in CI.

None. This change affects only CCC's own CI. The phpunit jobs now fill the tiktoken vocab cache one time before the tests run in parallel, which stops random Could not decode token failures. Sites see no change.

Issue #3586661: Saving or rendering context items fails when the tiktoken vocab cannot be loaded.

On sites where the tokenizer cannot download or read its vocabulary file, saving context items no longer fails. AI Context stores an approximate token count, trims context by bytes, logs one warning, and recalculates the stored counts on cron after the vocabulary loads. See "When the tokenizer vocabulary cannot be loaded" in docs/developers/configuration.md. The token estimator and the token recount services have new constructor arguments, so run drush cr after the update.

Issue #3586664: [Accessibility] Document the Content editor keyboard trap as a known issue before 1.0.0.

Documentation: A new Known issues page lists problems that are not fixed yet and their workarounds, and README.md and the docs home page link to it. The first entry is about the Content editor on the context item form: in rich text mode, Tab and Shift+Tab do not move keyboard focus out of the editor (WCAG 2.2 SC 2.1.2). To move focus out, press Escape and then Tab, or switch to Source mode. The fix belongs in the AI module (project/ai#3586814) and is not in a release yet.

Issue #3586677: Getting a "AiContextSelector: Items cut off due to token limit: 4" warning and I don't understand why.

The stored token count of each context item (AiContextItem::getTokenCount()) now includes the ID, label and purpose header and the indentation of its rendered block, not only the guidance text. Before this change, the badges in the Tokens column on /admin/config/ai/context/items and the consumer budget summary showed too low a figure, so items could be cut off while the summary said that they were within budget. Run drush updatedb, then drush cr. Update 10032 recounts every revision and translation. Selection and rendering do not change, but expect higher token badges, tooltip percentages and budget totals, and possibly a Context budget exceeded message for a consumer that is near its Max tokens for injection. The warning about cut-off items now gives the number of items that did not fit, the token budget, the consumer ID when the selection has a consumer, and the item IDs.

Issue #3586678: "Last used" value of "Context usage" table acts more like "first used" when used on new content entities.

• Last Used on the Context usage page (/admin/config/ai/context/usage) now changes every time a context item is used. Before, a repeat use through the same consumer on the same route stopped changing it after the second use, so it showed an early use instead of the last one.
• Usage cleanup on cron reads Last Used. A usage record that is still in use is no longer deleted by the maximum age rule.
• Run drush cr after the update. The ai_context.usage_tracker service has a new constructor argument.

Issue #3586679: Document that full Automator consumer support needs AI core 1.5.0 or later.

Documentation: The docs now say that the Automator consumer type works fully only with AI 1.5.0 or later, which sends the ai_automator:id:{id} request tag and the entity_context metadata. On AI 1.4.x, direct or cron runs on a field with more than one automator get no context, and the Entity Types, Taxonomy Terms and Specific Entities scopes do not match automator requests. The drupal/ai requirement stays ^1.4 || ^2. See docs/features/consumer_configuration.md.

Issue #3586687: Site Sections hard-filters on the supplied request path, but the selection's cache contexts carry only url.path and route.

Context selection results now vary by a new ai_context.request_path cache context, which holds the path that the Site Sections scope filters on. That path can come from an ai_context_path request attribute, a posted path parameter, a path key in a JSON body, or a same-site Referer on an asynchronous request, and not only from the URL. Before this change, two requests to one endpoint with different supplied paths could select different items but get the same cache keys, so custom code that cached a selection, or a render array built from one, could serve items from the wrong site section. Run drush cr after the update so that Drupal registers the new cache context, and make sure that custom code that caches a selection uses all of the selection's cache metadata.

Issue #3586696: Add the load_ai_context_items function call and deprecate load_ai_context_item_by_id.

The Load tool has a new plugin ID, ai_context:load_ai_context_items, with the tool name "Load AI Context Items". The old plugin ID, ai_context:load_ai_context_item_by_id, is now a deprecated alias with the same parameters, permission checks and output. The alias is removed in 2.0.0. The agent tool list shows it as "Load AI Context Item By ID (deprecated)". Run drush cr after the update, so that the new plugin is discovered.

For new agents, select Load AI Context Items. Existing agents that use the old ID keep working until 2.0.0. No database update changes them. To move an agent to the new ID, make these changes by hand, and then export the configuration:

• Agent tools: select Load AI Context Items, clear the old tool, and copy its tool settings.
• Default information tools on the agent form: change tool: ai_context:load_ai_context_item_by_id to tool: ai_context:load_ai_context_items.
• AI Agents agent overrides: change the old ID in the tools to add or remove, the tool settings and the tool usage limits.
• Agent prompts: this update does not change prompt text. Edit each prompt that names the old tool. Change ai_context:load_ai_context_item_by_id to ai_context:load_ai_context_items, and ai_context_load_ai_context_item_by_id to ai_context_load_ai_context_items.

Each run of the old tool triggers an E_USER_DEPRECATED notice. Listing the tools does not. Usage records show the old ID for agents that were not moved. Use them to find the agents to change. See "Upgrade from 1.0.0-rc1: Load tool plugin ID" in docs/developers/api-stability.md.

Issue #3586700: Split the phpunit CI jobs by test type so that they finish within the runner time limit.

None. This change affects only CCC's own CI. The phpunit job now runs as three jobs, split by test type: unit and kernel tests, functional tests, and functional JavaScript tests. The phpunit (previous major) and phpunit (max PHP version) jobs get the same split, and the test-only changes job stays one job. The GitLab runner stops every job at 30 minutes, and the full test suite in one job took longer than that on a busy runner. Sites see no change.

Issue #3586701: Core "view any unpublished content" permission shows unpublished context items.

Fixes a bug with a security impact. Core's View any unpublished content and View the latest version permissions no longer give access to unpublished context items. This applies to the item page, the CCC listing, the Latest version tab, and the revision pages. Before this fix, a role with View any unpublished content, for example the Content editor role on Drupal CMS, could read draft and archived context items without an AI Context permission. Sites that used these core permissions so that editors can see context drafts must give those roles View any unpublished AI Context Item. Custom access hooks can no longer allow view of unpublished context items. Run drush cr after the update.

Issue #3586702: Usage records keep every request path with no limit on number or length.

On sites with usage tracking on, a user with "Use Published AI Context in AI Features" who could use a chatbot could make usage records very large, so AI requests became slow or failed. Each usage record now keeps only the last 20 request paths, each 255 characters or less, and a path hint that is longer than 2,048 characters is ignored. A record that was saved before this update is cut the next time a path is added to it. An error in usage tracking during an agent run is now logged, and the agent run continues. If a consumer already fails because of large usage records, first turn off usage tracking at /admin/config/ai/context/settings/usage. Then clear the records at /admin/config/ai/context/settings/usage/clear. If the clear page also fails, empty the ai_context_usage table (for example, drush sql:query "TRUNCATE TABLE ai_context_usage") and rebuild the caches.

Issue #3586703: Diff pages show older context item revisions without the revision permission.

Diff pages for context items now need the View all AI Context Item revisions permission, as the revision history page does. Before, on sites with Diff 8.x-1.x, 2.0.0, or 2.1.0, any user with Use Published AI Context in AI Features could compare older published revisions. Review which roles have View all AI Context Item revisions. After the update, rebuild the caches (drush cr), so that the router uses the new requirement.

Issue #3586706: Duplicate publishes the copy for users who may not use the publish transition.

Duplicate now always creates a draft copy. This applies to every user, also to users who may use the publish transition, and also when Publish by default is on. Review the copy, and then publish it. Before this change, a user who could not use the publish transition could get a published copy without the publish transition. To duplicate an item, a user now needs the Create new draft transition (AI Context editorial workflow: Use Create new draft transition). A user without it gets a form error, also when the user may use the publish transition. The copy belongs to the user who made it, and each translation of the copy gets new created and changed dates. The copy does not keep the publishing schedule or the revision log message of the original. The description of Publish by default now says that Duplicate always creates a draft copy. On earlier versions, a copy has the same label and the same created date as its original. To find copies that were published this way, look for published items that have the same label or the same created date as another item, and review them. This check also finds copies that a user with the publish transition made. It does not find a copy whose original was deleted.

Issue #3586712: Require drupal/ai 1.4.11 or later on 1.4, or 1.5.2 or later.

CCC now requires drupal/ai 1.4.11 or later on the 1.4 branch, or 1.5.2 or later (~1.4.11 || ^1.5.2 || ^2). These drupal/ai versions show admin tooltips as text. With older drupal/ai versions, HTML in a long context item description or AI agent description runs as script when an administrator opens the CCC consumer pages. Update drupal/ai together with CCC, for example with composer update drupal/ai_context drupal/ai -W, and then run drush updatedb. To stay on AI 1.4, use composer update drupal/ai_context drupal/ai -W --with drupal/ai:~1.4.11. Composer does not check the commit of a dev branch. If your site uses a drupal/ai dev branch, for example 1.4.x-dev, 1.5.x-dev, 1.6.x-dev, 1.x-dev, 2.0.x-dev or 2.x-dev, update it to the current branch head. drupal/ai 2.x has no release yet, so every site on drupal/ai 2 uses a dev branch. With these drupal/ai versions, scope pill tooltips, the +N overflow tooltip and the Automator description tooltip show HTML entities such as & or '. #3586705 (milestone 1.0.1) fixes this.