Skip to main content

Roadmap

These are concrete, RFC-track proposals for after the core spec stabilizes and the reference implementation matures further. They're ordered roughly by dependency — later items generally assume earlier ones have landed. None of this is committed or scheduled; it's the direction the design is headed.

  • Effect-tracked parallelism beyond with. An effect-tracking type-system extension (in the spirit of algebraic effects) that lets the compiler prove independence for a broader class of programs than today's syntactic sibling-reference restriction allows — without giving up the soundness guarantee. with would become sugar for a more general, effect-typed scheduling primitive.

  • Distributed execution engine. A distributed variant of the current single-process engine: sharding the scheduler across a worker fleet, replicating the trace engine for high-availability checkpoint writes, and defining exactly-once semantics for effect nodes across worker failover — building on the lessons of Temporal's distributed durable-execution model.

  • Formal verification of scheduling and replay. A machine-checked proof (e.g. in Lean or Coq) that the two-pass scheduler never reorders across a true dependency, and that replay is exactly equivalent to live execution given an identical trace — moving today's informal argument to a verified one, in the spirit of seL4/CompCert.

  • Cost-aware automatic provider routing. Extending today's static routing policies (cheapest, pinned(...)) into a learned/adaptive router that uses accumulated trace data to pick a provider per call based on observed quality-cost tradeoffs for that specific capability and input distribution.

  • Multi-language host bindings beyond FFI. Today, Python/JS/shell are an escape hatch called from Ulexite. A stable C ABI would let a host application in any language embed the Ulexite runtime and drive conversations from outside .ulx source entirely — useful for gradually introducing Ulexite into an existing large application.

  • A standardized capability conformance suite. A published, versioned test suite that any provider plugin author can run to certify which structured_output tier and which artifact types their plugin actually delivers under load — turning today's "depends on honest plugin authors" limitation into a certifiable, third-party-auditable property, the way TCK suites work for JVM languages.

  • A visual/notebook authoring surface. A structured-editing surface for authoring with-block dataflow and match branching visually, extending the trace viewer's timeline UI from a read-only debugging tool into an optional bidirectional authoring tool — aimed at prompt engineers and analysts who aren't primarily software engineers.

  • Judge ensembles and adversarial verification as language sugar. Promoting patterns like multiple independent judges voting or refutation-biased verifiers from a standard-library convention into dedicated grammar, once real-world usage shows which ensemble shapes recur often enough to deserve first-class syntax.

  • Formal semantics for cross-conversation supervision. Extending today's single-conversation supervise scope across nested conversations — a parent conversation declaring a supervision policy that governs how child-conversation failures propagate, generalizing Elixir's supervision-tree model to Ulexite's conversation-nesting structure.

For the full design rationale, see §25 of the spec.