Operations and deployment
Manage the service, execution target, and persistent research records as separate responsibilities. First verify a local connection with the quickstart, then choose the operation you need.
Choose an operation
| Goal | Guide |
|---|---|
| Execute, wait, follow logs, stop, or delete Tasks | Task management |
| Combine installed Tasks into a synchronous flow | Composite Tasks |
| Browse and inspect workspace artifacts | Workspace files |
| Configure service access and tokens | Authentication |
| Host a service with or without Studio | Deployment |
| Use CLI direct access or Studio forwarding | Remote machines |
| Forward configured HTTP requests | HTTP proxy |
| Schedule recurring local Task commands | Scheduling |
| Send task notifications through DingTalk | Notifications |
| Diagnose failures, back up, and recover records | Troubleshooting and recovery |
Establish the execution environment
Each execution target has its own Python environment, plugins, data, and workspace. Direct CLI access uses the target service token; Studio uses the local token and forwards through its same-origin backend to a configured target. Use the same execution target for execution, waiting, logs, and artifact queries.
Prepare raw data, plugin dependencies, credentials and Agent environments separately. See workspace concepts and client configuration.
Preserve evidence while maintaining the service
Retain distinct Task identities for experiments. A fixed-name rerun replaces a finished directory; deletion removes records and artifacts. Explicit Task termination stops execution. The default remote cli Job continues after client disconnection; service shutdown stops its manual commands without resuming them later. Daemon-owned scheduled commands survive service restart. Read task lifecycle before choosing cleanup or recovery actions.