Chat-to-Board

Overview

Chat-to-Board is an AI-powered chat panel that lets you describe features in natural language and get actionable tasks added directly to your Kanban board. It also supports natural language board management — "mark X as done", "move Y to review".

Why

Typing tasks one-by-one is slow. Most people think in paragraphs, not ticket fields. Chat-to-Board bridges that gap — describe what you're building, and the AI breaks it into tasks with priorities.

How It Works

Frontend

• Floating chat bubble (bottom-right) opens a slide-up panel
• Mobile: full-width panel | Desktop: 420px fixed-width panel
• Sends conversation history + project context to /api/ai/chat
• AI responses can include:
  • Task suggestions — rendered as cards with "Add" / "Add All" buttons
  • Task actions — status updates or deletions executed against Supabase

Backend

The /api/ai/chat endpoint:

1. Authenticates the request (requires Supabase session)
2. Fetches BYOK settings from profiles
3. Falls back to server GROQ_API_KEY if no BYOK
4. Sends a structured system prompt forcing JSON output:

{
  "message": "Conversational response",
  "tasks": [{ "title", "description", "priority", "status" }],
  "actions": [{ "type": "update_status|delete_task", "taskTitle", "status?" }]
}

Supported Actions

Task matching is fuzzy — the AI extracts the closest task title from your phrasing.

Architecture

User types message
  → chat-to-board.tsx (client)
    → POST /api/ai/chat
      → reads BYOK from profiles table
      → calls AI provider (Groq/OpenAI/Anthropic/Gemini)
      → returns { message, tasks[], actions[] }
    → renders response + task cards
    → "Add" inserts task into Supabase tasks table
    → "Add All" bulk-inserts all suggested tasks
    → actions execute status updates via Supabase client
  → onTasksAdded() refreshes kanban board

Data Persistence

Tasks persist — when you click "Add" or "Add All", tasks are inserted into the Supabase tasks table immediately.

Chat conversations do not persist yet. The message history lives in React state. Opening the panel starts a fresh conversation; closing it discards the history. Added tasks remain on the board.

Planned: Persistent Chat History

• chat_sessions table (id, project_id, user_id, title, created_at)
• chat_messages table (id, session_id, role, content, tasks_json, actions_json, created_at)
• RLS policies scoped to the owning user
• Sidebar in chat panel to switch between past conversations

Design Decisions

• JSON-only system prompt: AI must return valid JSON (no markdown fences) for predictable parsing
• Context-aware: Current board tasks are sent to the AI so it can reference existing work
• No streaming: Full response returned at once — task cards need the complete response to render
• Conversation memory: Full chat history sent each request (within the panel session)

Key Files