> ## Documentation Index > Fetch the complete documentation index at: https://docs.opper.ai/llms.txt > Use this file to discover all available pages before exploring further. # Replace the draft > Saves draft changes without deploying. Omitted description, graph and layout are preserved; an empty description clears it. Supply the complete graph to replace it. Structural errors reject the save; inspect meta.issues and data.draft.validation for remaining issues before deployment. Use simulate_dynamic_route without version_number to test the saved draft; deploy_dynamic_route performs full deployment validation. Requires the `dynamic_routes:write` scope. Graph format: an object with an `elements` array. Each element has a unique `id`, a `type`, optional `properties`, and `outputs`, an object mapping output names to `{"elementId":"target-node-id"}`. Include `id: START, type: start` and `id: END, type: end`; `input` and `output` are not node types. A minimal deployable graph (replace the model with an available catalog ID permitted by your policy): ```json {"elements":[{"id":"START","type":"start","outputs":{"next":{"elementId":"primary"}}},{"id":"primary","type":"model","properties":{"model":"cerebras/gpt-oss-120b"},"outputs":{"success":{"elementId":"END"},"fallback":{"elementId":"END"}}},{"id":"END","type":"end","outputs":{}}]} ``` Supported node types and properties: - `start`: the START node; `outputs.next` selects the first node. - `end`: the END node; `outputs` is `{}`. - `model`: `properties.model` is a catalog model ID. `outputs.success` targets END or an observe node; `outputs.fallback` targets another model, pool, or END. - `pool`: `properties.models` contains 2-10 `{model,metrics}` candidates; `properties.expression` is a numeric CEL expression such as `model.cost`, sorted ascending. Uses the same success/fallback outputs as model. - `conditional`: `properties.cases` contains `{id,expression}` entries; each expression is a CEL predicate over `metadata`, `auth`, and `input`. Connect an output for each case ID and a `default` output. The first true case wins. - `percentage`: `properties.buckets` contains at least two `{id,weight}` entries with positive integer weights totaling 100. Connect an output for each bucket ID. - `classifier`: `properties.cases` contains 2-10 `{id,label,instructions}` entries with unique labels, plus optional `context` and `model` properties. Connect an output per case ID; no default output. Classification failures fail the route. - `observe`: `properties.criteria` defines an asynchronous evaluation, with optional `score_type` (`score` or `binary`), `model`, `threshold`, and `sample_rate` (both greater than 0 and at most 1). Place after model/pool success or another observe node; `outputs.next` leads to another observe node or END. Graphs must be acyclic. Every node except END accepts at most one incoming edge; use separate model nodes on separate branches. Deployment requires each path from START to reach a model or pool before END. ## OpenAPI ````yaml put /management/v1/dynamic-routes/{name}/draft openapi: 3.1.0 info: description: Schema-driven generative API that orchestrates LLM-powered workflows. title: Task API version: 3.0.0 servers: - description: Production url: https://api.opper.ai - description: Local development url: http://localhost:8080 security: - BearerAuth: [] tags: - description: Schema-driven function management and execution name: Functions - description: OpenAI-compatible chat completions name: Chat - description: OpenAI Responses API compatible endpoint name: Responses - description: Google-compatible interactions endpoint name: Interactions - description: Model registry and capabilities name: Models - description: Synchronous image generation name: Images - description: Text-to-speech and speech-to-text name: Audio - description: Asynchronous video generation name: Videos - description: Reusable file storage for media inputs and generated outputs name: Files - description: Async generation status and downloads name: Artifacts - description: OpenAI-compatible embeddings name: Embeddings - description: Recorded HTTP request/response generations name: Generations - description: System health and status name: System - description: Roundtable endpoint — fan out a query to multiple LLMs and combine results name: Roundtable - description: Web search, fetch, and other utility tools name: Tools - description: Caller identity, credits, and usage name: Account - description: >- Programmatic project and API-key management. Authenticates with an `op-mak-…` management token; available on the control_plane and enterprise plans. name: Management paths: /management/v1/dynamic-routes/{name}/draft: put: tags: - Management summary: Replace the draft description: >- Saves draft changes without deploying. Omitted description, graph and layout are preserved; an empty description clears it. Supply the complete graph to replace it. Structural errors reject the save; inspect meta.issues and data.draft.validation for remaining issues before deployment. Use simulate_dynamic_route without version_number to test the saved draft; deploy_dynamic_route performs full deployment validation. Requires the `dynamic_routes:write` scope. Graph format: an object with an `elements` array. Each element has a unique `id`, a `type`, optional `properties`, and `outputs`, an object mapping output names to `{"elementId":"target-node-id"}`. Include `id: START, type: start` and `id: END, type: end`; `input` and `output` are not node types. A minimal deployable graph (replace the model with an available catalog ID permitted by your policy): ```json {"elements":[{"id":"START","type":"start","outputs":{"next":{"elementId":"primary"}}},{"id":"primary","type":"model","properties":{"model":"cerebras/gpt-oss-120b"},"outputs":{"success":{"elementId":"END"},"fallback":{"elementId":"END"}}},{"id":"END","type":"end","outputs":{}}]} ``` Supported node types and properties: - `start`: the START node; `outputs.next` selects the first node. - `end`: the END node; `outputs` is `{}`. - `model`: `properties.model` is a catalog model ID. `outputs.success` targets END or an observe node; `outputs.fallback` targets another model, pool, or END. - `pool`: `properties.models` contains 2-10 `{model,metrics}` candidates; `properties.expression` is a numeric CEL expression such as `model.cost`, sorted ascending. Uses the same success/fallback outputs as model. - `conditional`: `properties.cases` contains `{id,expression}` entries; each expression is a CEL predicate over `metadata`, `auth`, and `input`. Connect an output for each case ID and a `default` output. The first true case wins. - `percentage`: `properties.buckets` contains at least two `{id,weight}` entries with positive integer weights totaling 100. Connect an output for each bucket ID. - `classifier`: `properties.cases` contains 2-10 `{id,label,instructions}` entries with unique labels, plus optional `context` and `model` properties. Connect an output per case ID; no default output. Classification failures fail the route. - `observe`: `properties.criteria` defines an asynchronous evaluation, with optional `score_type` (`score` or `binary`), `model`, `threshold`, and `sample_rate` (both greater than 0 and at most 1). Place after model/pool success or another observe node; `outputs.next` leads to another observe node or END. Graphs must be acyclic. Every node except END accepts at most one incoming edge; use separate model nodes on separate branches. Deployment requires each path from START to reach a model or pool before END. operationId: updateManagementDynamicRouteDraft parameters: - description: Route name in: path name: name required: true schema: type: string requestBody: content: application/json: schema: $ref: '#/components/schemas/DynamicRouteDraftRequest' required: true responses: '200': content: application/json: schema: properties: data: $ref: '#/components/schemas/DynamicRouteDetail' meta: properties: issues: items: $ref: '#/components/schemas/DynamicRouteValidationIssue' type: array type: object type: object description: Successful response '400': content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' description: Bad request '401': description: Unauthorized - missing or invalid API key '403': content: application/json: schema: properties: error: type: string required_scope: description: Present on the scope refusal. type: string type: object description: >- The token lacks the scope this endpoint requires (`required_scope` is returned). Presenting a runtime `op-…` key instead of a management `op-mak-…` token also lands here. '404': content: application/json: schema: properties: error: type: string type: object description: >- No such resource in the calling organization. A resource belonging to another organization is indistinguishable from one that does not exist. '500': content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' description: Internal server error security: - ManagementBearer: [] components: schemas: DynamicRouteDraftRequest: properties: description: description: New description. Omit to leave it unchanged. type: string graph: description: >- Routing graph object with elements. Each node has id, type, outputs mapping output names to {elementId: targetID}, and optional properties. Include START/start and END/end. Supported types: start, end, model, pool, conditional, percentage, classifier, observe. Omit graph on create to seed a default; copy its returned draft.graph to customize. See tool description for node properties and topology. examples: - elements: - id: START outputs: next: elementId: model_default type: start - id: model_default outputs: fallback: elementId: END success: elementId: END properties: label: Default model model: cerebras/gpt-oss-120b type: model - id: END outputs: {} type: end properties: elements: items: properties: id: type: string outputs: additionalProperties: properties: elementId: type: string required: - elementId type: object type: object properties: type: object type: type: string required: - id - type - outputs type: object type: array required: - elements type: object layout: description: Editor canvas positions keyed by element id. Opaque to the gateway. type: object type: object DynamicRouteDetail: properties: active_version: properties: created_at: format: date-time type: string note: type: string uuid: format: uuid type: string version_number: type: integer required: - uuid - version_number - note - created_at type: object draft: properties: graph: description: The draft graph (DynamicRouteGraph). type: object layout: description: Editor canvas positions keyed by element id. type: object revision: type: integer validation: description: >- Validation issues on the draft (DynamicRouteValidationIssue); empty when it is deployable. type: array required: - graph - layout - validation - revision type: object latest_version_number: type: integer route: properties: created_at: format: date-time type: string deleted_at: format: date-time type: string description: type: string draft_graph: description: The draft graph (DynamicRouteGraph). type: object draft_layout: description: Editor canvas positions keyed by element id. type: object draft_revision: type: integer draft_validation: description: >- Validation issues on the draft (DynamicRouteValidationIssue); empty when it is deployable. type: array invocation_model: type: string name: type: string updated_at: format: date-time type: string uuid: format: uuid type: string required: - uuid - name - invocation_model - description - draft_revision - created_at - updated_at type: object version_count: type: integer required: - route - draft - version_count type: object DynamicRouteValidationIssue: properties: code: type: string field: type: string message: type: string node_id: type: string severity: type: string required: - code - message - severity type: object ErrorResponse: properties: error: properties: code: type: string details: description: Any value message: type: string required: - code - message type: object meta: type: object required: - error type: object securitySchemes: BearerAuth: bearerFormat: API Key description: API key authentication. Pass your API key as a Bearer token. scheme: bearer type: http ManagementBearer: bearerFormat: Management API Key description: >- Management API authentication. Pass an `op-mak-…` management token as a Bearer token. Runtime `op-…` API keys are rejected with 403 — they belong on the data-plane endpoints. Mint a management key from the platform UI under Settings → API keys. scheme: bearer type: http ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.