Current beta scope and limitations

Stripe

The implemented surface has real-API drift evidence. It covers the documented customer, payment, refund, invoice and subscription workflows. The dated fidelity record states what was checked, and the scope reference lists omissions and refused parameters.

Examples outside this subset include Checkout, Connect, disputes, tax, coupons, trials and multi-item subscriptions. Unimplemented routes and recognized unsupported writes fail explicitly.

Zendesk

The native profile supports the declared directory, ticket, comment, audit, bounded Search, assignment and lifecycle operations. Local conformance is tested; live API qualification is pending. It is included as beta functionality with those limits, not as a fully verified copy of your tenant.

Macros, triggers and SLAs are not implemented. Custom fields/statuses, attachments, deletion, bulk jobs and tenant-specific automation are also outside the current profile. Import is from supported offline snapshots or reviewed Full JSON exports, not live Zendesk ingestion.

Local directory guards protect other staff accounts from agent name updates and reject duplicate organization names or external IDs according to the documented comparison rules. Agents may update end users and their own numeric user ID; administrators may update all supported roles. User email is accepted on creation and refused on update. These are local rules; exact provider permissions, identity behavior and uniqueness semantics remain unqualified. See the native reference for comparison and validation details.

The older Zendesk-shaped ticket desk is a separate demo-grade profile used by legacy billing samples. Its behavior and evidence do not establish native Zendesk parity.

Starter project boundaries

Customer Desk and Team Inbox use the demo-grade ticket desk, not native Zendesk workflows. Digital Shop selects existing prices and quantities for unpaid draft invoices; it has no purchase, cart editing, inventory or fulfillment. Reset to revise a proposal.

Outcome checks evaluate explicit records, changes and request attempts. Required reply facts do not judge general writing or factual accuracy. The examples are model-free teaching scripts; their passes do not establish agent performance. Model credentials, external calls and costs belong to the participant.

Review exported inputs and reports before sharing. Evidence omits connection details and configured secrets, but cannot identify every sensitive value in arbitrary input. Reports are unsigned and should be inspected and rerun. See the starter guide for supported checks and project recovery.

CLI setup directories

Create the destination project folder before native Zendesk or combined CLI setup. These modes require an existing directory; a missing destination currently produces a generic input error. The dashboard setup guide also starts in an existing project folder.

Combined workflows and agent connections

Supported combined worlds share a clock and state. Starting data must satisfy composition rules, and grading requires explicit billing-to-ticket bindings. The dashboard offers supported combinations and retains each connector's qualification.

Agents run as local commands with the documented input/output contract. Native Zendesk requires explicit endpoint configuration; supported Stripe SDKs can use the supplied preload. Hosted-agent endpoints and tools whose destination cannot be changed are outside this release.

Results, replay and local storage

Native standalone HTML and JSON include world-level findings for protected or unexplained changes as well as task outcomes. Keep the safe JSON beside the report. A passing task does not override a failed overall result. The published beta.1 and beta.2 HTML reports retain their documented limitation.

World replay is deterministic for the same starting data and request sequence. A model may choose a different sequence on another run. Optional model judgments are advisory; inspect authoritative state expectations and any ungraded tasks.

Scripted samples contain synthetic data and teach the workflow. They are not customer performance measurements. Local evidence stays until you delete it. Beta upgrades may require recreating disposable worlds from reviewed saved inputs; preserve your project before changing versions.

Beta.4 and beta.5 refuse schema-6 world files, including Stripe-only worlds from beta.3. There is no in-place migration. Export needed settled state with the old runtime, back up your original project, seeds and suites, stop the old host, then recreate worlds. Replay caches do not transfer. Optional connector profiles and browser commerce have further schema boundaries. Downgrading the executable does not restore old files. Follow the storage and upgrade guide.

The local runtime is designed for a trusted harness. It is not a security boundary for hostile code running on the same machine. Keep it on trusted infrastructure and follow the deployment reference before exposing it beyond loopback.

Platforms and optional model drafting

macOS and Linux with Node 24 or newer are the beta platforms. Windows is not verified. The first installation needs network access. Installed dashboard assets and scripted samples work offline; your own agent, source imports and selected model providers may need the network.

Optional dashboard drafting has not undergone live-provider acceptance for this beta. It remains available with that limitation. A configured provider, local request validation or a mocked test does not establish a live-verified integration.

Aggregate reporting

Optional aggregate reporting covers usage and test outcomes across the CLI, dashboard, starters, browser workflows and runtime API. New onboarding discloses a default-on choice with an opt-out. Existing users and unknown CLI/CI installations stay off until a choice is saved. Use worlds telemetry disable, Reporting settings, WORLDS_TELEMETRY=0 or DO_NOT_TRACK=1 to disable it. Tests, customer records, prompts, code and evidence remain local; reporting never changes test limits or verdicts. See the reporting policy.

Other connectors

Agentforce Runtime has a bounded local session, message, action and feedback profile with dedicated credentials and an intake workflow. Experimental SAP S/4HANA, Dynamics 365 Dataverse, Confluence and Freshservice profiles are included for their documented local subsets. These do not establish complete provider compatibility or general dashboard task-suite support. Live provider qualification remains pending. JSM includes seed/state/client foundations only; native JSM endpoints are not implemented.

Browser workflows include a local storefront, worker and MCP execution, deterministic interruption, state-based grading and evidence export. The storefront is a separate commerce profile; Digital Shop starter invoices do not establish storefront coverage. Other catalog entries continue to collect interest.