Skip to content

roadmap.md

Latest commit

 

History

History
96 lines (84 loc) · 5.51 KB

roadmap.md

File metadata and controls

96 lines (84 loc) · 5.51 KB
Raw

RSScript roadmap

This roadmap implements the product boundary in product.md. The language specification and tests remain authoritative for existing behavior.

Completed product-boundary milestones

  1. Removed active policy/capability examples and obsolete product descriptions.
  2. Replaced catalog-size assertions with uniqueness, completeness, signature, determinism, and orphan-entry checks.
  3. Split neutral package analysis from review, provider, native, and build metadata, with checked-in schemas.
  4. Enforced dependency direction using Cargo metadata.
  5. Established syntax, structural semantics, Typed HIR, executable IR, provider ABI, host-neutral runtime defaults, and concrete leaf providers.
  6. Added bounded rsscript.bytecode.v1, structural verification, verified-only VM construction, a stable embedding façade, build/inspect commands, and the provider-replacement demonstration.
  7. Defined the strict execution-report contract, added a reusable Provider Conformance Kit for all official Providers, and added property/fuzz coverage for ownership, retention, resource handles, Artifact bytes, bindings, and report consumers.
  8. Established a versioned Core SLO/reporting gate and three-platform CI evidence with checksums and provenance.
  9. Split the reference VM into rsscript-vm, established the owned CFG rsscript-mir model and rsscript-codegen-vm boundary, and made checked-HIR lowering a one-way compiler projection. Cargo metadata tests now reject any VM dependency on compiler, syntax, semantics, or lowering internals.

Current priority: generation oracle and machine feedback

The compiler-purity migration is complete. Its checked baseline is retained as historical evidence, not as the active roadmap. The active work is to define a sound, Agent-authored generation oracle around the existing Core path. It must make feedback usable by machines without creating a parallel parser, checker, compiler, or trust boundary.

  1. Maintain the parser-owned v1 generation query: syntax supplies prefix and terminal facts, semantics supplies and composes scoped/type/effect facts, and compiler callers reuse that contract without owning a second oracle. See ADR 0232.
  2. Expand parser terminal coverage without weakening the explicit completeness flags or the rule that may_stop requires complete syntax and valid semantics. Keep rss check --json and rss fix --json as the diagnostic and repair contracts.
  3. Strengthen the generated machine context and interface identity beyond the current versioned language card, grammar hash, Core interface sources, Core policy, and per-session interface revision. It excludes host secrets, deployment policy, and Provider implementation state.
  4. Grow the offline evaluation corpus beyond its initial fixtures and populate all four generation modes with caller-supplied model samples. Evaluations may demonstrate progress, but no incomplete query result is a successful compile or an execution authorization.
  5. Keep the archived research archived. Rust AOT, REIR, and self-hosting live on the archive/experiments-2026-09 branch; they are not workspace members and do not return as SDK, VM, or CLI dependencies.
  6. Preserve the current repository and workspace shape: do not split the repository, delete backends, or reorder the Cargo workspace as part of this work. Keep rsscript.bytecode.v1, host deployment boundaries, and the capture/build/verify/inspect/default-isolated-run golden path intact.

Parallel priority: native JIT as a supported VM tier

The Cranelift JIT is a required part of the product, not an experiment awaiting a retention verdict. CPU-bound embedded workloads need native execution, and the interpreter alone does not meet that need. The JIT remains an opt-in VM feature that consumes the same verified bytecode and cannot be selected by source or Artifact, and the interpreter remains the semantic oracle for every native path. Within those invariants, JIT work is active feature work:

  1. Close the accounting gap: native execution must report the same deterministic step, allocation, cancellation, and deadline facts as the interpreter, so bounded and isolated execution can use it.
  2. Grow native coverage of the language so fewer regions fall back to the interpreter, with differential parity gates for every new region kind.
  3. Keep the controlled workload scorecard as the evidence for each optimization; measured regressions are removed, but the engine itself is not a removal candidate.
  4. Promote the engine to Core once accounting parity, supported-platform CI, and the threat model align, following the maturity rules in product.md.

Frozen scope

Until the priorities above are complete, do not expand language syntax, public intrinsics, speculative JIT tiers without scorecard evidence, the C backend, full self-host bootstrap, package publishing, native plugin surface, or language-level policy. Correctness, security boundary, maintenance, and measured-regression fixes remain allowed. Syntax sugar that desugars in the parser to an existing AST node and is justified by a measured generation failure is also permitted; new semantics are not.

Rust AOT, REIR, and self-hosting are archived on the archive/experiments-2026-09 branch and receive no changes. The Cranelift JIT is Experimental in maturity but is an active product surface; its promotion follows the criteria in product.md.