Snabb controlled beta reviewer guide
Support matrix dated 29 October 2026. The current release manifest identifies the deployed source and client version.
Bring a Node24 project, make and test a change with personal ChatGPT, save the workspace and review a private delivery.
Prisma and SQLite example
Use the Prisma + SQLite archive and its checksum and commands for a database-backed project. It includes matching database configuration, explicit client generation and a strict HTTP/database verification. Saved SQLite data belongs to the workspace; separate image previews require their own persistence design.
Start with your own account
- Open the console and register with your own email. Open the delivered verification message, set a password and sign in. Record the time and any failure. An operator-created session does not count as independent signup.
- Create an empty project. Choose Node24 and use the review sample archive. Verify it against its SHA-256 manifest. The sample has no package dependencies.
- Upload the archive through project source settings, or open a source-backed workspace. Review the selected folder and source identity before running
node --test. A terminal result receipt proves command completion. - To use the agent, open Organisation → AI providers and connect your own personal ChatGPT account with the device-code flow. Keep credentials out of chat and project variables.
For local inspection, extract the sample and run node --test with Node 24.22.1 or newer. Start it with node server.mjs; it listens on port 4000 and serves / and /health.
JavaScript and Python SDK downloads · Developer documentation
Complete one useful delivery
- Ask personal ChatGPT to add a
GET /versionroute returning{"version":"review-1"}, add a native test, run the full test suite and save the workspace. Inspect changes and command receipts. - Ask for a pipeline that tests and builds the saved source with the included Dockerfile. Create a project registry destination credential when publication requests it. Select secrets explicitly.
- Review the exact source, pipeline version, destination and limits before approving publication. A successful test step is separate from a verified, scanned image.
- Review the exact image and current price quote before approving an owner-private preview on port 4000. Open the preview through the console. Verify root, health and version responses.
- Inspect the actual compute receipt and storage usage. Stop the preview and save/stop the workspace when finished.
- Delete only the disposable review project through project settings after retaining your result. Access closes first; final erasure waits for existing upload authorizations to expire. Keep its deletion receipt.
A browser interruption can leave a submitted action unconfirmed. Recover its original project, source, run or pipeline identity. Do not submit a different action to guess whether the first one happened.
Supported scope and known limits
| Capability | Current beta boundary |
|---|---|
| Execution | Node24 is the reviewer path. Node22, Python, Ubuntu and OCI are active platform profiles with separate acceptance scopes. |
| Saved workspaces | Published saved files can be recovered. Uncheckpointed writes are not guaranteed after physical hardware loss. |
| RAM recovery | Original-process recovery is qualified for personal ChatGPT on Node24 with the current pm6 worker. It operates at supported safe boundaries. The measured recovery sample took about five minutes; progress and current limits matter. |
| Agent permissions | Workspace commands follow their selected permissions. Permitted commands with network and credentials can cause external effects. Platform publication and preview launch require separate owner review. |
| Private images | Docker Hub and GHCR import/run/cache/isolation have bounded acceptance. Dedicated read-only provider-token and provider-revocation qualification remains open. |
| Network | Named command policies and supported HTTP/HTTPS choices follow the selected destinations and permissions. Public outbound IPv4 TCP/UDP/SSH is available with an explicit broad network policy. Private, reserved and metadata addresses and unsolicited inbound traffic remain denied. New permissions apply through the supported settings handoff; existing admitted commands retain their original authority. |
| Capacity | The initial cohort is up to five reviewers, with at most two review workflows started together. Resource admission can queue or refuse work. A configured worker ceiling is not reserved reviewer capacity. |
| Availability | This preview uses one execution worker. Production failover, cross-host RAM continuity and an availability SLO are not offered. |
| Other features | Team preview extension, arbitrary CI conversion, verified test skipping, broad provider/runtime/device combinations and customer-published OS templates are incomplete or deferred. The second console is a demonstration with sample data. |
| Support | General product support and a response-time commitment are not offered for this controlled preview. Account data requests retain their manual request process. |
Understand the cost before starting
Current beta plans have zero recurring fees. New verified personal accounts receive one EUR 50 promotional opening credit; plan and account quotas have no cap. This does not create additional wallet credit: funded-wallet admission, explicit user budgets and physical worker capacity still apply. Automated checkout is disabled.
Ordinary compute uses wallet reservations and actual settlement with fractional-cent carry. First-time imports can charge measured preparation work, including a failed import. Cached execution still uses compute. Platform-startup refusals and never-started cancellations have their own zero-charge policy. Inspect the receipt rather than inferring a charge from a status.
Standard shared payer allowances are 30 decimal GB of workspace data and 50 decimal GB of object storage, with EUR 0.02/GB-month overage when charging is enabled. Retained objects and registry image content count; row sizes are not added to charge shared image layers twice. Effective account policy and complete metering observations govern actual charges.
Personal model usage and provider limits are separate from Snabb compute. Preview quotes use the effective account setting; measured resources can be recorded even when a beta quote is zero.
Keep an actionable failure record
- No verification email: record request time, provider and displayed error. Check your inbox/spam and use the normal resend control; do not request an operator login to count the journey as completed.
- Project quota: inspect your active projects. Never delete another project or raise a plan limit to hide a refusal.
- Model reconnect: use Organisation → AI providers and the device-code connection flow for the same Snabb owner account. Do not paste tokens.
- Queued work or low capacity: inspect the displayed state and limits; avoid duplicate submissions. Pause new reviewer starts while the operator checks capacity.
- Build or scan failure: inspect the failed step and exact source/image receipt. A successful command does not establish a published image or healthy preview.
- RAM or checkpoint limit: inspect progress, quota and the saved-state boundary. Retain the original recovery identity and explicit unknown result; avoid starting a second recovery to guess the first outcome.
- Result still being reconciled: a terminal run may briefly return
RESULT_NOT_READY. Wait and read its same result receipt again; do not start another run. Keep the original job ID if your local wait ends. - Interrupted response: preserve the original action ID and recover it. Stop separately if an admitted run should no longer continue.
Record your observation in the reviewer worksheet. Include identifiers and error codes, not passwords, provider tokens or secret values.