Tarski/ docsJoin the waitlist

Feature status

A precise availability matrix for current, preview, and planned Tarski platform surfaces.

This page separates shipped contracts from roadmap direction. “Specified” means a machine-checked public contract exists; “available” means the named release exposes it. Do not design a production dependency around “planned”.

Current and next-preview contracts

Surface Availability Minimum version
Hosted installer and preview channel Available installer metadata 0.5.9-preview.7+
Underscore entity-name query normalization Next preview tarski 0.5.9-preview.8
Cloud automation credentials and shadow endpoint smoke Next preview tarski 0.5.9-preview.8
Reactive declared-query subscription SDK Next SDK release; local/hosted conformance-gated @tarski/client 0.8.0
Write-policy payload and delegated-actor bindings Next preview tarski 0.5.9-preview.8
Fixture write-policy attempts with expected denials Next preview tarski 0.5.9-preview.8
One-shot timer simulator and canonical fire envelope Next preview tarski 0.5.9-preview.8
Runtime schedule and blob client operations Next preview/SDK release tarski 0.5.9-preview.8, @tarski/client 0.8.0
Platform-builder delegated tokens and lineage provisioning Preview; capability-discovered tarski 0.5.9-preview.8, @tarski/client 0.8.0
TypeScript app-contract codegen Next preview tarski 0.5.9-preview.8
Fixture world export/import/compare Next preview tarski 0.5.9-preview.8

Until those artifacts are published, source or documentation on main is not an installable production contract. The release pages and registry are authoritative.

Planned agent-context surfaces

The following requests are accepted product direction but are not V1 runtime capabilities:

  • a first-class, batched judge harness that can amortize model calls across proposals;
  • backend-owned ranked/windowed ambient context slices such as recent-N messages;
  • dynamic cross-lineage projection import discovery for account-level “Home” views.

Current patterns remain explicit:

  • model a judge as a declared reviewer agent whose LLM output returns as observation evidence, and batch only in your application/provider contract;
  • derive bounded ranking with finite relations/aggregates or pass a deliberately bounded context projection;
  • declare cross-lineage projection imports statically and pin every source identity, or compose independently authorized query results in the client.

No runtime should advertise these planned items through capability discovery. A client that asks for one must receive a typed unsupported state rather than a silent fallback.