> ## Documentation Index > Fetch the complete documentation index at: https://docs.tetherai.ca/llms.txt > Use this file to discover all available pages before exploring further. # Conversational prompt assistant (streaming) > Conversational prompt assistant that streams assistant reply deltas and may emit a draft prompt update. Classifies the user turn into chat, clarify, or draft mode, then streams a series of newline-delimited JSON events (mode, status, assistant-delta, reasoning-delta, draft-ready, done, error). Response uses application/x-ndjson; treat each line as a discrete event. ## OpenAPI ````yaml /api-reference/openapi.yaml post /api/ai/prompt-assistant/stream openapi: 3.0.3 info: title: Tether Relay and Tether API version: 1.0.0 description: >- Route-derived OpenAPI contract generated from tether-relay backend sources. Regenerate with scripts/sync-openapi-from-backend.mjs. servers: - url: https://your-instance.example.com description: Production - url: http://localhost:2212 description: Local API service security: [] tags: - name: AI - name: Applications - name: Approval - name: Auth - name: Automations - name: Calls - name: Campaigns - name: Contact Analytics - name: Contact Metrics - name: Contact Notes - name: Contacts - name: Conversation Monitor - name: Conversations - name: Departments - name: Email - name: Escalation Dashboard - name: Event Logs - name: Google Sheets - name: Marketplace - name: Messages - name: Notification Sounds - name: Notifications - name: Org Snapshots - name: Organizations - name: Outreach - name: Outreach Metrics - name: Outreach Sources - name: Pipelines - name: Processes - name: Prompts - name: Reminders - name: Sms - name: Templates - name: Traces - name: Training - name: Upload - name: User Snapshots - name: Users - name: Vertical - name: Webchats - name: Webhook paths: /api/ai/prompt-assistant/stream: post: tags: - AI summary: Conversational prompt assistant (streaming) description: >- Conversational prompt assistant that streams assistant reply deltas and may emit a draft prompt update. Classifies the user turn into chat, clarify, or draft mode, then streams a series of newline-delimited JSON events (mode, status, assistant-delta, reasoning-delta, draft-ready, done, error). Response uses application/x-ndjson; treat each line as a discrete event. operationId: post_ai_prompt_assistant_stream requestBody: required: true content: application/json: schema: $ref: '#/components/schemas/AIPromptAssistantStreamRequest' responses: '200': description: >- NDJSON event stream of assistant reply deltas and optional draft prompt content: application/x-ndjson: schema: $ref: '#/components/schemas/AIPromptAssistantStreamEvent' text/event-stream: schema: $ref: '#/components/schemas/AIPromptAssistantStreamEvent' '400': description: instruction is required content: application/json: schema: $ref: '#/components/schemas/AIErrorResponse' '401': description: Unauthorized content: application/json: schema: $ref: '#/components/schemas/AIErrorResponse' '500': description: Failed to process prompt assistant request content: application/json: schema: $ref: '#/components/schemas/AIErrorResponse' security: - bearerAuth: [] components: schemas: AIPromptAssistantStreamRequest: type: object properties: instruction: type: string description: >- Latest user turn in the assistant conversation. Classified into chat/clarify/draft mode before any prompt edit is attempted. currentPrompt: type: string description: >- The prompt XML currently in the editor; used as read-only grounding for chat/clarify and as the basis for diff when drafting. conversationHistory: type: array description: >- Prior assistant turns in this editor session, replayed so the model can carry context. Empty or malformed entries are dropped server-side. items: type: object properties: role: type: string enum: - user - assistant content: type: string required: - role - content variableContext: type: object additionalProperties: true description: >- Available runtime variables and field scopes; the assistant only references `@variables` that appear here and never invents new ones. assistedMode: type: boolean description: >- When true, the assistant clarifies ambiguity instead of drafting directly, and any draft uses the stronger XML architecture system prompt. modelOverride: $ref: '#/components/schemas/AiModelOverride' description: >- Pin to a specific provider/model (e.g. `{ provider: 'anthropic', model: 'claude-sonnet-4-6' }`) overriding the org's default assignment. required: - instruction additionalProperties: true example: instruction: Tighten the tone — make it more concise and less salesy. currentPrompt: >- Leasing assistantConvert leads into tours. conversationHistory: - role: user content: It feels a bit pushy. - role: assistant content: Got it — what tone would you like instead? assistedMode: true modelOverride: provider: anthropic model: claude-sonnet-4-6 AIPromptAssistantStreamEvent: type: object description: >- One newline-delimited JSON event from the prompt-assistant stream. Possible types: mode, status, assistant-delta, reasoning-delta, draft-ready, ai-setup-required, error, done. properties: type: type: string enum: - mode - status - assistant-delta - reasoning-delta - draft-ready - ai-setup-required - error - done mode: type: string enum: - chat - clarify - draft label: type: string delta: type: string prompt: type: string code: type: string reason: type: string section: type: string provider: type: string message: type: string error: type: string required: - type additionalProperties: true example: type: assistant-delta delta: Sure — I can tighten that up for you. AIErrorResponse: type: object description: >- Error envelope used across AI endpoints. The `error` field is a stable machine-readable code on approval-related failures (`approval_owner_mismatch`, `approval_org_mismatch`, `approval_expired`, `approval_service_unavailable`) and a human-readable string elsewhere. `message` is set alongside on the approval paths so the UI can render both. properties: success: type: boolean enum: - false error: type: string message: type: string required: - error additionalProperties: true example: success: false error: approval_expired message: >- This approval request has expired or was already handled. Please send a new playground message to retry. AiModelOverride: type: object properties: provider: type: string enum: - anthropic - openai - google model: type: string required: - provider - model additionalProperties: false example: provider: anthropic model: claude-sonnet-4-6 securitySchemes: bearerAuth: type: http scheme: bearer bearerFormat: JWT ````