chat

name

chat

description

This skill should be used when adding new chat functionality to the application or modifying existing chat behaviors. Use this skill when creating a new domain-specific chat interface (e.g., for Sessions, Specs, Pull Requests), adding new document types to existing chats, or debugging chat-related issues. The skill provides comprehensive patterns for both frontend Vue components and backend Python orchestrators.

Chat System Implementation

Overview

Add and modify AI-powered chat functionality in the application. The chat system uses a unified architecture with:

• Frontend: Universal ChatInterface.vue component for all chat types
• Backend: BaseChatOrchestrator abstract class providing 95% of orchestration logic
• Streaming: Real-time event broadcasting via Pushpin/WebSocket
• Tools: Extensible tool system for LLM capabilities

When to Use This Skill

Use this skill when:

• Creating a new chat for a different kind of domain object (e.g., Tasks, Reports, Deployments)
• Adding a new document type to an existing chat (e.g., adding "Roadmap" to Spec chat)
• Implementing inline AI editing for a new document type (text selection → AI-assisted rewrite/expand/fix)
• Modifying chat behavior or adding new tools
• Debugging chat streaming, tool execution, or event handling issues
• Understanding how chat works end-to-end

Core Architecture

The chat system has three main layers:

1. Frontend Layer (Vue.js)

• ChatInterface.vue: Universal component handling all chat types
• Domain Wrappers: Thin wrappers (SessionChat.vue, SpecFullChat.vue) for domain-specific configs
• Event System: WebSocket-based real-time updates via $lastWsEvents

2. Backend Layer (Python)

• BaseChatOrchestrator: Abstract base providing conversation management, tool execution, streaming
• Domain Orchestrators: Thin subclasses (PlanOrchestrator, SpecOrchestrator) configuring tools and prompts
• StreamableObject Interface: Domain objects implement get_stream_id(), get_event_log(), save()

3. Communication Layer

• Events: AIEvent, UserEvent, ToolCallEvent, ToolResultEvent, etc.
• Broadcasting: async broadcast_event_async() sends events to Pushpin/WebSocket
• Streaming: Server-Sent Events (SSE) for historical event replay

Quick Start: Adding a New Chat

Step 1: Implement StreamableObject Interface

Domain object must provide:

def get_stream_id() -> str         # Unique identifier (e.g., "tasks_abc123")
def get_event_log() -> list[Event] # Chat history
async def save(org) -> None        # Persist changes

Step 2: Create Domain Orchestrator

Extend BaseChatOrchestrator and implement two methods:

class MyDomainOrchestrator(BaseChatOrchestrator):
    async def get_tool_configuration(self, streamable_object, org, **kwargs):
        # Configure domain-specific tools

    async def create_domain_prompt(self, streamable_object, org, **kwargs):
        # Create initial context prompt

Step 3: Create Frontend Wrapper

Wrap ChatInterface.vue with domain-specific props:

<IssuesChatInterface
  domain-type="my_domain"
  :domain-id="myDomainId"
  @stop="handleStop"
  @session-updated="handleUpdate"
/>

Step 4: Register Domain Type with Unified Endpoints

Update THREE files for complete streaming integration:

A. Add to sleuth/apps/issues/utility/ninja_api.py

A1. Add to get_streamable_object() function (around line 650):

elif object_type == "my_domain":
    return await get_my_domain_manager().get_my_domain(object_id, org=org)

A2. Add to get_orchestrator_for_domain() function (around line 717) if you have a custom orchestrator:

"my_domain": MyDomainOrchestrator,

A3. Add stream auto-registration in stream_events_generic() (around line 260):

elif stream_id.startswith("mydomain_"):
    domain_id = stream_id[9:]  # Remove 'mydomain_' prefix
    my_domain = await get_my_domain_manager().get_my_domain(domain_id, org=org)
    if my_domain:
        await get_or_create_streaming_context(stream_id, my_domain, org)
        streamable_object = await get_streaming_context(stream_id, org)

B. CRITICAL: Add to sleuth/apps/issues/streaming/disk_context.py

Add your domain to get_streaming_context() function (around line 40):

elif stream_id.startswith("mydomain_"):
    domain_id = stream_id.replace("mydomain_", "")
    from sleuth.apps.issues.mydomain.manager import get_mydomain_manager

    return await get_mydomain_manager().get_mydomain(domain_id, org)

Why this is critical: Without this, you'll get 404 Not Found when connecting to the stream and chat will not work at all. This is the most commonly forgotten step!

Note: The unified endpoints /issues/{domain_type}/{domain_id}/chat/ and /issues/{domain_type}/{domain_id}/stop/ work automatically once your domain is registered!

Step 5: Register Streaming Tools (If Applicable)

CRITICAL: If your domain uses sub-orchestrator tools (tools that themselves run orchestrators), you MUST register them in ChatInterface.vue's isStreamingTool() function. Otherwise, users won't see streaming progress in the UI.

// frontend/components/issues/ChatInterface.vue
function isStreamingTool(toolName: string): boolean {
  return [
    // ... existing tools ...
    'your_sub_orchestrator_tool',  // Don't forget this!
  ].includes(toolName)
}

See references/frontend_architecture.md "Sub-Orchestrator Streaming" section for details.

Detailed Documentation

Refer to the bundled reference files for comprehensive technical details:

references/frontend_architecture.md

• ChatInterface.vue component API and patterns
• Wrapper component patterns (SessionChat, SpecFullChat)
• Event handling and state management
• Custom input slots and file uploads
• Auto-scroll behavior and user interaction patterns

references/backend_architecture.md

• BaseChatOrchestrator responsibilities and lifecycle
• Tool configuration and ToolConfiguration class
• Event system and broadcasting patterns
• Two-stage model strategy (research vs generation)
• Stop handling and error management
• Context collection and metadata tracking

references/streamable_object_interface.md

• Interface requirements and implementation patterns
• Stream ID conventions and naming
• Event log management
• Save/persistence patterns
• Examples from Session, Spec, PullRequest

references/code_examples.md

• Complete working examples for common scenarios:
  • Creating a simple chat for a new domain
  • Adding document generation to existing chat
  • Implementing sub-orchestrators with dual-streaming
  • Custom tool creation and streaming tools
  • Complex event handling patterns

references/common_pitfalls.md

• Frequent mistakes and how to avoid them
• Debugging tips and techniques
• Performance considerations
• Testing strategies

references/inline_editing.md

• Inline AI editing pattern for document section updates
• TipTapEditor text selection and AI action menu
• use-document-section-update composable pattern
• TextSelectionModal integration
• Extending UpdateSectionTool to new document types
• Document ID format conventions and routing
• Complete implementation guide with examples

Key Patterns to Follow

Frontend Patterns

1. Use ChatInterface.vue - Never create a new base chat component
2. Thin wrappers - Domain components should only pass props and handle events
3. domainType/domainId - Use consistent naming (e.g., "session"/"session_123")
4. Event emission - Emit domain-specific events for UI refresh triggers

Backend Patterns

1. Extend BaseChatOrchestrator - Never reimplement orchestration logic
2. OrchestratorConfig - Use configuration for behavior customization
3. Tool organization - Separate research and domain-specific tools
4. System prompts - Write clear, structured prompts with explicit workflows
5. Auto-stop via stop_reason - Orchestration stops automatically via Claude's API, not via tools

Communication Patterns

1. Event broadcasting - Always use broadcast_event_async()
2. Dual streaming - Support parent callbacks for sub-orchestrators
3. Event types - Use appropriate event types (AI, User, Tool, System)
4. Ephemeral events - Mark streaming content as ephemeral to avoid storage

Testing Your Implementation

1. Frontend: Test chat UI, message display, event handling, stop button
2. Backend: Test tool execution, event broadcasting, error handling
3. Integration: Test full conversation flow with real LLM
4. Streaming: Test real-time updates and reconnection behavior

Next Steps

After reviewing this documentation:

1. Read the relevant reference files for your use case
2. Look at code examples similar to your needs
3. Implement following the patterns documented
4. Test thoroughly using the testing checklist
5. Iterate based on behavior and user feedback