Inline Reply Suggestions

When you click an inbound message in the communication feed, an inline suggestions panel appears directly below it. The panel shows three AI-generated reply cards ranked by confidence, with controls to generate, regenerate, and use suggestions.

How It Works

Click inbound message bubble
  → Toggle InlineSuggestionsPanel below the bubble
  → GET /api/text-agents/suggestions/{message_id}
  → If suggestions exist → render 3 cards with confidence badges
  → If none exist → show "Generate Suggestions" button
  → Click generate → POST /api/text-agents/generate-suggestions
  → Agentic AI loop produces 3 ranked replies
  → Click "Use This Reply" → populate compose bar with suggestion text

UI States

The panel cycles through five states depending on data availability and user actions:

State	Trigger	Display
Loading	Panel opens, fetching data	Three skeleton placeholder cards
Empty	No suggestions saved yet	Sparkles icon + “Generate Suggestions” button
Generating	Generate/regenerate clicked	Spinner + “Generating suggestions…” + skeletons
Loaded	Suggestions returned	Three cards in a responsive grid
Error	API call failed	Error message + “Try Again” button

Suggestion Cards

Each card displays three elements:

pub struct TextAgentSuggestionItem {
    pub body: String,        // The suggested reply text
    pub confidence: f64,     // Score between 0.0 and 1.0
}

Confidence badges use color-coded thresholds:

• ≥ 80% — green badge (emerald-500)
• 50–79% — blue badge (primary)
• < 50% — gray badge (muted)

The “Use This Reply” button on each card calls the on_use handler, which populates the compose bar with that suggestion’s body text.

Inbound Bubble Interactions

Inbound message bubbles gained interactive behaviors to support the suggestions panel:

• Hover — a subtle ring highlight (hover:ring-1 ring-primary/30) and sparkles icon appear
• Click — toggles the suggestions panel for that message
• Selected state — active ring (ring-1 ring-primary/40) persists while the panel is open
• Toggle — clicking the same bubble again closes the panel
• Switch — clicking a different inbound bubble closes the previous panel and opens a new one

Selection resets automatically when you navigate to a different contact.

Auto-Scroll

The communication feed now auto-scrolls to the most recent message when:

• The conversation first loads
• A new message arrives via WebSocket
• You switch to a different contact

The scroll uses a reactive use_effect hook tied to the message list signal, with a short timeout to ensure the DOM has rendered before scrolling.

Authorization

Both suggestion endpoints use contact-based access control instead of text agent management permissions. You need permission to view the contact — not permission to manage text agents.

The system calls check_contact_access with ContactInstanceAction::View, which checks:

1. The contact exists in the caller’s organization
2. The caller has either ContactInstancePermission::View (org-wide) or ContactInstancePermission::ViewAssigned (restricted to assigned contacts)

Super-admins and org owners bypass these checks automatically.

Role	Can see suggestions?
Super admin / org owner	Always
Org admin with contact view permission	Yes, for all contacts
Org member assigned to the contact	Yes, via ViewAssigned
Org member not assigned to the contact	No (403 Forbidden)

This means any team member who can view a contact’s conversation also sees AI suggestions for that contact’s messages — no separate text agent permission required.

API Endpoints

Get suggestions

GET /api/text-agents/suggestions/{message_id}
Body/Query: { contact_id: string }

Returns the most recent TextAgentSuggestion for the given message, or null if none exist. Requires ContactInstanceAction::View permission for the specified contact.

Status codes: 200 success, 400 invalid UUID, 403 no contact access, 500 server error.

Generate suggestions

POST /api/text-agents/generate-suggestions
Body: { contact_id: string, message_id: string }

Resolves the text agent assigned to the contact (via direct link or phone number fallback), runs the agentic suggestion loop, and persists the result to the text_agent_suggestion table. Requires ContactInstanceAction::View permission for the specified contact.

Status codes: 200 success, 400 invalid UUID, 403 no contact access, 404 message or contact not found, 500 server error.

Note

Generation uses the same agentic loop as inbound processing — up to 5 tool-calling steps with knowledge base access. The assigned text agent’s instructions and provider configuration apply.

Key Components

Component	File	Role
InlineSuggestionsPanel	text_agent/components/text_agent_suggestions_panel_component.rs	Panel with fetch, generate, and card rendering
SuggestionCard	Same file (private)	Single card with confidence badge and use button
SkeletonCards	Same file (private)	Three animated placeholder cards
MessageEntry	messaging/components/message_entry_component.rs	Bubble with click, hover, and selection states
CommunicationFeed	messaging/components/communication_feed_component.rs	Feed container managing selection state and auto-scroll

Database

Suggestions are persisted in the text_agent_suggestion table:

text_agent_suggestion (
    id              UUID PRIMARY KEY,
    organization_id UUID NOT NULL,
    contact_id      UUID NOT NULL,
    text_agent_id   UUID NOT NULL,
    message_id      UUID NOT NULL,
    suggestions     JSONB NOT NULL,  -- Vec<TextAgentSuggestionItem>
    auto_sent_body  TEXT,            -- Set if auto-reply was triggered
    created_at      TIMESTAMP NOT NULL
)

Each generation creates a new row. The GET endpoint returns the most recent row for a given message_id, ordered by created_at DESC.