FSharpened

How it works

A layered satellite architecture around the F# toolchain.

Wrap the platform

FSharpened.Platform hides every OS difference (file watching, IPC, paths, sandboxing, terminals) behind interfaces and a capability manifest. Nothing else may use #if on a platform.

Share one state substrate

FSharpened.Statepoints stores content-addressed snapshots with fork, rollback and merge. Hot reload sessions, AI edits and training runs are all statepoint consumers.

Ground everything semantically

FSharpened.Semantic exposes symbol resolution over the F# compiler services and MCP, so tools and agents work from facts, not text guesses.

Compose the best tools

Product pillars connect Fantomas, analyzers, debuggers and editors. We write glue and contracts, not replacements.

Add ML behind contracts

ML pillars reach frameworks only through ML.Portability, pick a backend per device through ML.Hardware, and always end in a deterministic fallback chain.

Prove it on every RID

One local-first pipeline (fsh_ci.py) runs identically in hooks, on your machine and in the six-RID GitHub matrix, including NativeAOT.

The fallback chain

Any model-backed feature resolves through five deterministic layers, so the tool keeps working when a model is missing, slow or untrusted:

  1. Specialised model
  2. Sibling model
  3. Symbolic rules
  4. Warm checkpoint
  5. Structured escalation to a human

Invariants, not vibes

Fourteen constitutional invariants are written down, mirrored in the registry and checked mechanically. They are what lets agents and humans change the project safely for years.

Read the invariants