Sandbox Lifecycle

Agent calls the Daytona SDK to create a sandbox for a single codegen task (e.g. Daytona().create(...)). The call returns a sandbox handle carrying an id. The agent must run code in it and tear it down afterward.

Persist the returned sandbox id to the agent's durable task record BEFORE doing any work inside the sandbox, so a crash between create and delete cannot orphan a running (billed) sandbox. Treat the create call as the resource-acquisition step of a try/finally whose finally deletes the sandbox. [REQ…

An agent loop creates one sandbox per task across thousands of tasks. Some tasks raise exceptions mid-run before the explicit delete call is reached.

Delete the sandbox in a finally / context-manager so it is reclaimed on both the success and the exception path. A leaked sandbox keeps consuming the org's concurrency quota and compute budget. Prefer the SDK's context-manager form if it exists; otherwise wrap create/delete yourself. [REQUIRES-VERI…

Agent resumes a previously stopped sandbox by id and immediately calls a code-exec method without starting it.

A stopped sandbox must be start()ed (and reach a running state) before code/fs/git operations will succeed. Wait for the running state rather than firing operations optimistically, and surface a clear error if start fails (e.g. quota, deleted). [REQUIRES-VERIFICATION] for the exact state machine an…