Start a run
Select a project and blueprint, provide the required inputs, and submit a task once.
A run is one execution of your task. Start one from Sessions → Dispatch run, from the CLI, or through a configured chat integration.
Dashboard and CLI dispatch require an owner or admin role. Chat uses the authorization configured for its channel; it is a separate access path.
Choose the execution settings
Select the project first. Devboxes checks which blueprints and models are available for that project and organization. Review the base branch, blueprint version, required inputs, and provider account before submitting.
The Implement GitHub Issue blueprint is designed around a GitHub issue. Other blueprints can describe different workflows. Use the instructions and inputs of the selected blueprint rather than assuming every run takes the same inputs or produces a pull request.
Describe the work
State the observed problem or requested change, the intended result, and the checks that will establish completion. Include constraints that would otherwise be easy to miss.
Fix the settings page so an empty display name cannot be saved.
Show a validation message beside the field and preserve the current value.
Add a regression test using the existing form test pattern.
Keep the public API unchanged and report the checks you ran.Avoid broad instructions such as “improve the app” for a first run. Supply missing context rather than requiring the agent to guess a product decision. Keep secrets in configuration, not in the task text.
Understand the interaction mode
Where offered, ask-when-unclear tells the agent to request missing decisions. Fully autonomous tells it to proceed without relying on that clarification step. Neither mode replaces your responsibility to choose appropriate repository and credential access.
After submission
Open the returned session and follow the run there. Queued means it was accepted, not that a container is already running. Image preparation, runner availability, and model access can delay the start.
Submit once. If the interface reports an uncertain result, check the session or run list before retrying. For related follow-up work, continue the session instead of creating an unrelated task.