--- title: "Human-in-the-loop" description: "Pause workflows for human approval, input, or review." source: https://flowdrop.io/docs/editor/interrupts site: FlowDrop documentation --- # Human-in-the-loop FlowDrop's interrupt system enables workflows to pause execution and request user input before continuing. This is essential for approval workflows, data collection, decision points, and quality control. Every interrupt prompt below ships as a live, interactive component. Explore its states, props, and edge cases in [FlowDrop Storybook](https://flowdrop-demo.netlify.app/storybook/). ## Interrupt types ### Confirmation Simple yes/no prompt for binary decisions: ![Interrupt prompts rendered in chat showing confirmation buttons.](/images/screenshots/hitl-interrupt-confirm.webp) ```json { "interrupt_type": "confirmation", "message": "Do you approve sending this email to 150 recipients?", "confirm_label": "Yes, send email", "cancel_label": "No, cancel" } ``` **Response**: `boolean` Try the confirmed, declined, submitting, and error states. ### Choice Single or multiple selection from predefined options: ```json { "interrupt_type": "choice", "message": "Select the output format:", "options": [ { "value": "json", "label": "JSON", "description": "Structured data" }, { "value": "csv", "label": "CSV", "description": "Spreadsheet format" } ], "multiple": false } ``` **Response**: `string` (single) or `string[]` (multiple) Compare single-select and multi-select variants. ### Text input Free-form text entry: ```json { "interrupt_type": "text_input", "message": "Provide additional context:", "placeholder": "Enter your notes...", "multiline": true, "max_length": 1000 } ``` **Response**: `string` See single-line, multiline, and length-constrained inputs. ### Form Complex data entry using JSON Schema: ```json { "interrupt_type": "form", "message": "Complete the configuration:", "schema": { "type": "object", "properties": { "priority": { "type": "string", "enum": ["low", "medium", "high"] }, "notify": { "type": "boolean", "title": "Send notification" } } }, "default_values": { "priority": "medium", "notify": true } } ``` **Response**: `object` (matching schema structure) ### Review Review proposed field changes with per-field accept/reject decisions and visual diffs: ```json { "interrupt_type": "review", "message": "Review these proposed changes:", "changes": [ { "field": "title", "label": "Page Title", "original": "About Us", "proposed": "About Our Company" }, { "field": "body", "label": "Body Content", "original": "

Welcome to our site.

", "proposed": "

Welcome to our company website.

" } ] } ``` **Response**: `ReviewResolution` with per-field decisions and summary counts. Inspect per-field diffs, accept/reject controls, and many-change views. ## Architecture ```mermaid sequenceDiagram participant W as Workflow Execution participant B as Backend participant F as Frontend participant U as User W->>B: Pause & create interrupt B->>F: Send interrupt request F->>U: Render prompt U->>F: Submit response F->>B: Resolve interrupt B->>W: Resume workflow ``` ## Frontend integration The `ChatPanel` automatically detects and renders interrupts in messages. For manual integration: ```typescript import { getInstance } from '@flowdrop/flowdrop/editor'; import { InterruptService } from '@flowdrop/flowdrop/playground'; const fd = getInstance(); const interruptService = InterruptService.getInstance(); // Read pending interrupts reactively const pending = $derived(fd.interrupts.getPending()); // Resolve an interrupt async function resolveInterrupt(interruptId: string, value: unknown) { const result = fd.interrupts.startSubmit(interruptId, value); if (!result.valid) return; try { // Pass the instance's endpoint config first await interruptService.resolveInterrupt(fd.api.config, interruptId, value); fd.interrupts.submitSuccess(interruptId); } catch (error) { fd.interrupts.submitFailure(interruptId, String(error)); } } ``` ## Using prompt components directly ```svelte handleConfirm()} onCancel={() => handleCancel()} /> ``` ## Backend integration ### Message metadata format When a workflow requires input, the backend sends a message with interrupt metadata: ```json { "id": "msg-123", "role": "assistant", "content": "I need your approval to proceed.", "metadata": { "type": "interrupt_request", "interrupt_id": "int-456", "interrupt_type": "confirmation", "message": "Do you approve this action?", "confirm_label": "Approve", "cancel_label": "Reject" } } ``` ### API endpoints | Endpoint | Method | Purpose | | -------------------------------------- | ------ | ----------------------- | | `/interrupts/{id}` | GET | Get interrupt details | | `/interrupts/{id}` | POST | Resolve interrupt | | `/interrupts/{id}/cancel` | POST | Cancel interrupt | | `/playground/sessions/{id}/interrupts` | GET | List session interrupts | ## State management The interrupt store uses a state machine with these transitions: - **idle** — Awaiting user input - **submitting** — User response being sent - **resolved** — Successfully processed - **error** — Submission failed (can retry) Resolved interrupts remain visible but disabled, showing the user's selection. ## Best practices 1. **Clear messages** — Write actionable prompts ("Do you approve sending this email to 150 recipients?" not "Proceed?") 2. **Meaningful labels** — Use descriptive button labels ("Yes, send email" not "Yes") 3. **Default values** — Provide sensible defaults for form fields 4. **Cancel behavior** — Only set `allowCancel: false` for mandatory interrupts 5. **Error handling** — Always handle resolution failures gracefully