Beacon is a Rust implementation of the language server protocol & a Hindley-Milner type system for Python, inspired by languages like F# & OCaml and the work of Astral.

Goals

Deliver an incremental, performant, pragmatic Hindley-Milner (HM) type inference and checking engine for Python that integrates with modern editor tooling via the Language Server Protocol (LSP). The system supports Python’s dynamic features thoughtfully, interoperates with typing hints, and scales to multi-file projects.

HM for Python

HM type systems provide principled inference (no annotations required), compositional reasoning, strong guarantees, & fast unification-based algorithms (Algorithm W family).

Challenges

• Pervasive dynamism (monkey-patching, __getattr__, metaclasses, duck typing, runtime reflection)
• Nominal & structural patterns mixed
• Subtyping-ish expectations (None, unions, protocols)
• First-class classes & functions
• Decorators
• Generators
• Async
• Pattern matching (PEP 634)

Features

”Smart” Completions

The CompletionProvider looks at the latest document snapshot (text, AST, and symbol tables) to suggest in-scope names, imports, and attribute members the moment you type or hit the "." trigger.

Confident Feedback

Beacon’s static analyzer combines Hindley-Milner style inference with control/data-flow passes. It flags type mismatches, use-before-def bugs, unreachable code, and unused bindings, then caches the results per file so hovers, squiggles, and inlay hints stay consistent without extra latency.

Shared Insight

Both systems reuse the same analyzer output, so upcoming features like type-filtered completions or richer hover details piggyback on the existing substitution data instead of re-walking files.

Design

HM core + pragmatic extensions, with a gradual boundary to accommodate Python idioms and annotations:

• HM for expressions and local bindings
• Controlled subtyping-like features via unions/optionals and protocols/structural constraints
• Annotation-aware: treat PEP 484/PEP 695 types as constraints and hints
• Soundness modes: “strict”, “balanced”, “loose” (affecting treatment of Any, unknown attributes, dynamic imports)