Configure and run a first task
Prepare an API key for the intended account. The only config field is size, accepting small, default, or large, defaulting to small. Do not reuse another provider’s medium value, image, or CPU fields. Actual resources and quota depend on the account. These local management examples read a private key. See self-hosting for hosted initialization. Retain the connection ID rather than rerunning creation.Execution and file limits
Connector command deadlines are 1–590 seconds for foreground and 1–3600 for background, defaulting to 60. These are command limits, not total Run duration. A background pid still requires Runtime/application health checks. Truncated output fails instead of becoming an incomplete success. A single file write accepts at most 5 MiB and checks reported bytes. Reads bound base64 content to 7 MiB and verify decoded length; encoded length is not raw file size. Higher API/artifact layers can impose separate constraints. Verify small-file integrity first and design large deliverables within supported limits.Private preview
Access task services through Hexagon private preview. The connector requests public=false hosting, extracts the provider token, and uses a server-side_port_auth Cookie. It does not hand the token-bearing URL to browsers.
The application must listen on the intended port. Diagnose process, provider ingress, Cookie, and platform authorization separately. Do not expose a public port to bypass checks. Preview is development access, not production website deployment.
Snapshots, TTL, and recovery
Stop retains a recoverable snapshot. The next submit to the same Session resumes and refreshes runtime TTL. Managed files/conversation can continue, but process memory, connections, and service liveness are not guaranteed. Provider archived maps to paused; ready/idle/running map to running. These mappings do not establish precise billing state. Runtime TTL is different from final deletion: stopped compute may retain snapshots. The control plane owns durable deadlines and cleanup; stopping a local service does not prove remote release. Historical trial accounts had concurrency/TTL restrictions. Verify current account limits instead of turning historical counts into universal promises. Follow idle and wake: first-turn file write, second-turn read/conversation continuation, and separate service checks. A successful single turn does not prove restoration.Verify account environment isolation
For a dedicated test account, configure a harmless account-level marker namedHEXAGON_BOAT_ACCOUNT_PROBE; do not use a real secret or put this marker in Agent environment. In the first Run, ask the Harness to execute a presence check that prints only present or absent, never environment values. The expected result is absent. Repeat after confirmed stop/resume in the same Session. If it is present, stop the exercise and inspect account settings and the create/resume noEnv requests before running sensitive workloads. This checks one named marker; it is not an audit of all guest credentials. The connector sets noEnv: true on both requests, but source inspection alone does not prove provider enforcement. This exercise has not been newly executed for this page.
Confirm asynchronous deletion
Final deletion has a separate operation receipt. The platform saves its ID, polls it, confirms targetId matches the original Sandbox, waits for status=completed, and then confirms resource absence. Blocked operations and observation timeouts are not successful cleanup. If the deletion receipt is lost, a missing Sandbox alone does not prove completed purge. Retain unknown state and original IDs, following reconciliation. Creating another resource or deleting a different object does not resolve uncertainty. Stop, snapshots, and accepted deletion do not establish completed Session cleanup. For example, if the saved operation is still pending while the Sandbox lookup returns 404, keep cleanup unresolved and continue observing that operation. If it reports completed for a different target, treat this as a target mismatch. Only completed for the original target plus confirmed absence establishes completion. Keep the Session ID, Sandbox ID, platform operation ID, and provider deletion receipt in private recovery records; they identify different objects. Creation uses the original operation ID as Idempotency-Key. Exact names and labels support paginated reconciliation. Similar names or incomplete lists do not establish ownership.
This page is implementation-reviewed without new cloud execution. Historical journeys are in the expanded matrix, with recovery scope in the serverless ledger. Completion requires actual execution, both SDKs, required recovery, and confirmed final deletion.
Source: Boat connector. Next: reconciliation, capacity and cost, verified combinations.