Skip to content

ROADMAP.md

Latest commit

 

History

History
76 lines (56 loc) · 3.21 KB

ROADMAP.md

File metadata and controls

76 lines (56 loc) · 3.21 KB

Roadmap

This file tracks the near-term release path for tywrap. It is the maintainer view of what is shipping now, what is next, and what is intentionally deferred.

The detailed technical appendix for the scientific data plane lives in docs/codec-roadmap.md.

Recently Shipped

v0.4.0: development hot reload and contract cleanup

v0.4.0 established the real development reload story for Tywrap:

  • tywrap/dev is the public development entrypoint
  • Node watch sessions regenerate wrappers and swap the active bridge
  • generated wrappers keep using the latest runtime bridge without re-imports
  • structured generation failures keep the last known good output and bridge live
  • placeholder config fields for reload behavior were removed in favor of the explicit dev API

The immediate 0.4.1 patch line is release hygiene, not a new roadmap theme. It exists to keep the release and CI baseline stable after the v0.4.0 cut.

Now (0.4.x)

Node 25 support, analyzer removal, and dev-contract stabilization

The 0.4.x line is for tightening what v0.4.0 introduced before the next minor release expands scope again.

  • remove the TypeScript-side analyzer and the required tree-sitter install path so npm install tywrap no longer depends on a native parser build for the supported Node story, including Node 25
  • stabilize tywrap/dev examples and hot-reload coverage, especially around failure and recovery behavior
  • define and document the tywrap to tywrap-ir compatibility contract so users get explicit expectations and explicit mismatch behavior
  • clean up deprecated runtime surface in docs and examples so new usage follows the shipped public API rather than compatibility shims

This release line should treat tywrap_ir as the single supported analysis path. The TypeScript analyzer is legacy internal code and should be fully removed rather than carried forward as a parallel implementation.

Next (0.5.0)

Chunked transport and the scientific data plane

v0.5.0 should be the release that makes large scientific payloads reliable and first-class.

  • add a versioned artifact or chunked transport path so large payloads no longer depend on single-line JSONL
  • make Arrow registration easier across the runtime story so the common scientific path does not require avoidable manual setup
  • expand scientific codec validation and performance gates so release quality is backed by real transport and payload-scale coverage
  • harden SciPy, Torch, and Sklearn envelope behavior so supported cases are explicit and unsupported cases fail clearly
  • document transport capability expectations so Node, Pyodide, and HTTP support boundaries stay easy to understand

See docs/codec-roadmap.md for the deeper technical plan behind this release theme.

Later

These items are intentionally not part of 0.4.x or 0.5.0:

  • GPU-native transport such as DLPack or Arrow CUDA
  • HTTP server lifecycle management owned by Tywrap
  • app-level HMR beyond Tywrap wrapper regeneration and bridge reload
  • unsafe default model-serialization paths such as implicit pickle or joblib
  • broader runtime surface removals that should wait for a later major-version contract pass