[FEATURE] [MER-4945] Add adaptive AI activation points

[manelli]

See: https://eliterate.atlassian.net/browse/MER-4945

[github-actions]

	Warnings
⚠️	PR is large (2629 LOC changed). Consider splitting.

Risk score: 22 → risk/high

Generated by 🚫 dangerJS against 7bcac6c

[github-actions]

AI Review — performance

Full revision JSON fetched on hot trigger path

file: lib/oli/conversation/triggers.ex
line: 303
Description: Every adaptive trigger resolution loads rev.content (full revision JSON) from the DB, then uses only partsLayout to locate one part/prompt. On a high-frequency trigger endpoint, transferring and decoding full content increases DB, CPU, and memory pressure.
Suggestion: Query only the needed JSON fragment (for example partsLayout and the specific fields under custom) instead of full content, or cache pre-extracted adaptive trigger metadata by revision_id in SectionResourceDepot/Cachex to avoid repeated DB reads.

Fallback resolver query adds expensive join path per request

file: lib/oli/conversation/triggers.ex
line: 306
Description: Cache misses call DeliveryResolver.section_resource_revisions(section_slug) and filter by resource_id, introducing a heavier multi-join query in the request path. Under load or low cache hit rate, this increases latency and DB contention.
Suggestion: Avoid resolver fallback in this path by requiring/priming section resource cache first, or use a direct indexed lookup keyed by section_id + resource_id to fetch revision_id/metadata without the broader resolver join.

Duplicate suppression logs at info can become hot-path I/O

file: lib/oli/conversation/triggers.ex
line: 176
Description: Suppressed duplicates are expected behavior during rapid UI retries; logging each duplicate at info can create high log volume and unnecessary I/O/CPU in bursty traffic.
Suggestion: Downgrade duplicate logs to debug (or sample/rate-limit them) and keep info for accepted or anomalous paths only.

[github-actions]

AI Review — elixir

Malformed authoring content can crash edit validation

file: lib/oli/authoring/editing/activity_editor.ex
line: 812
Description: content |> map_value(:authoring) |> map_value(:parts) assumes authoring is a map. If a client sends malformed content (for example, "authoring" => []), map_value/2 has no matching clause and raises FunctionClauseError, turning a validation failure into a server error.
Suggestion: Make map_value/2 total (add a catch-all clause) and defensively normalize containers before enumeration, e.g. treat non-map authoring as %{} and non-list parts as [].

partsLayout type is not validated before Enum.find/2

file: lib/oli/conversation/triggers.ex
line: 321
Description: parts_layout = map_value(content, "partsLayout") || [] does not guarantee a list. If stored content is malformed (for example, "partsLayout" => "bad"), Enum.find/2 later raises Protocol.UndefinedError, causing trigger resolution to fail with an exception instead of returning {:error, :invalid_trigger}.
Suggestion: Guard with is_list(parts_layout) (or coerce non-lists to []) before calling Enum.find/2, and return {:error, :invalid_trigger} for invalid shapes.

[github-actions]

AI Review — typescript

AI trigger prompt is never sent to the trigger API

file: assets/src/components/parts/aiTrigger.ts
line: 28
Description: The new AI-trigger flow validates that a prompt exists, but buildAdaptiveAiTriggerPayload does not include prompt in the payload. This means author-configured prompts (aiTriggerPrompt / prompt) are ignored at runtime, so backend trigger behavior can be incorrect.
Suggestion: Extend AdaptiveAiTriggerConfig and buildAdaptiveAiTriggerPayload to accept prompt, include prompt in the returned payload, and pass the configured prompt from AITrigger.tsx, janus-image/Image.tsx, and janus-navigation-button/NavigationButton.tsx when invoking.

Selector always allocates a new array and can cause avoidable rerenders

file: assets/src/apps/authoring/store/app/slice.ts
line: 361
Description: selectPartComponentTypes now calls .filter(...) unconditionally, returning a new array reference even when triggers are allowed and nothing is filtered. Components consuming this selector may rerender on unrelated app state changes due to referential instability.
Suggestion: Return state.partComponentTypes directly when state.allowTriggers is true, and only allocate a filtered array when allowTriggers is false.

[github-actions]

AI Review — security

No issues found

[github-actions]

AI Review — ui

Interactive image can become an unlabeled button

file: assets/src/components/parts/janus-image/Image.tsx
line: 177
Description: The image is now keyboard-focusable with role="button" when AI trigger is enabled, but no explicit accessible name is added for that button state. If alt is empty/decorative, screen readers can announce an unnamed control.
Suggestion: When aiTriggerAvailable is true, set an explicit accessible name (for example aria-label="Open DOT AI assistant" or a configurable label), and keep alt focused on image meaning rather than button labeling.

Global MutationObserver risks unnecessary UI churn

file: assets/src/components/parts/janus-ai-trigger/AITrigger.tsx
line: 90
Description: Observing document.body with childList: true and subtree: true can fire on most DOM updates, repeatedly calling setTriggerAvailable and causing avoidable re-renders/perf overhead in complex pages.
Suggestion: Replace with a narrower signal (for example a targeted observer on a stable container, or an explicit event when #ai_bot mounts/unmounts), and guard state updates so setTriggerAvailable runs only when availability actually changes.

[github-actions]

Preview deployed to: https://pr-6286.plasma.oli.cmu.edu