[WIP] [Python/Beam] Add F# quotation support

[dbrattli]

Summary

FYI: This is just an investigation of what it would take to add Quotation support to Fable [WIP]

• Adds F# code quotation support (<@ ... @>) to the Python and Beam targets
• Quotation ASTs can be constructed, pattern matched (via Microsoft.FSharp.Quotations.Patterns), and evaluated (via LeafExpressionConverter.EvaluateQuotation)
• Shared QuotationEmitter.fs transforms quoted Fable.Expr into runtime library calls at compile time
• Deliberately left out JS to avoid increasing bundle size (FYI: @MangelMaxime)

Supported quotation nodes

Value, Lambda, Let, IfThenElse, Application, Call, Sequential, NewTuple, Operation (all arithmetic/comparison/logical), TypeCast, Get (TupleIndex, UnionField, UnionTag, FieldGet), Set (ValueSet, FieldSet), NewUnion, NewRecord, NewOption, NewList, DecisionTree (simple cases).

Files changed

Area	Files
AST	Fable.fs — adds Quote of quotedExpr: Expr * isTyped: bool (3 lines)
Frontend	FSharp2Fable.fs, FableTransforms.fs, Transforms.Util.fs — capture and propagate Quote
Shared engine	QuotationEmitter.fs (new) — transforms quoted Fable.Expr → runtime lib calls
Python	Fable2Python.Transforms.fs, Python/Replacements.fs, fable_quotation.py (new)
Beam	Fable2Beam.fs, Beam/Replacements.fs, fable_quotation.erl
Other targets	Fable2Babel.fs, Fable2Dart.fs, Fable2Rust.fs, Fable2Php.fs — error stubs for Quote
Tests	TestQuotation.fs (Python, 24 tests), QuotationTests.fs (Beam, 18 tests)
Changelogs	Fable.Cli/CHANGELOG.md, Fable.Compiler/CHANGELOG.md

Design note: AST approach vs original PR #1839

The original quotation PR (#1839) added a complex ExprData record to the AST:

// PR #1839
| Quote of typed: bool * data: ExprData * range: SourceLocation option

type ExprData = {
    typ: Type; variables: VarData[]; values: ValueData[]
    literals: Expr[]; types: Type[]
    members: array<FSharpEntity * Type * MemberInfo * Type[]>
    data: byte[]
}

This PR takes a minimal AST approach instead:

// This PR
| Quote of quotedExpr: Expr * isTyped: bool

The quoted Expr is kept as-is in the Fable AST. All decomposition into runtime calls happens in the transform layer (QuotationEmitter.fs), not the AST. This means:

• + Only 3 lines added to Fable.AST (no new types, no new Type variants)
• + No serialization/deserialization layer needed
• + QuotationEmitter can evolve without breaking the AST contract
• − Quote _ -> Any is less precise than Expr(Some typ) / Expr None

If a more precise type is desired, a | Quotation of gen: Type option could be added to the Type union — but that would be a separate AST change worth discussing.

Test plan

• All 2103 Python tests pass (./build.sh test python)
• All 2287 Beam tests pass (./build.sh test beam)
• .NET tests pass for both Python and Beam test projects
• Verify no regressions in other targets

🤖 Generated with Claude Code

[github-actions]

Python Type Checking Results (Pyright)

Metric	Value
Total errors	18
Files with errors	4
Excluded files	4
New errors	✅ No

Excluded files with errors (4 files)

These files have known type errors and are excluded from CI. Remove from pyrightconfig.ci.json as errors are fixed.

File	Errors	Status
temp/tests/Python/test_applicative.py	12	Excluded
temp/tests/Python/test_hash_set.py	3	Excluded
temp/tests/Python/test_nested_and_recursive_pattern.py	2	Excluded
temp/tests/Python/fable_modules/thoth_json_python/encode.py	1	Excluded

[dbrattli]

Question: Query expression support and FCS inlining

@ncave @MangelMaxime

While implementing quotation support, we investigated adding query { } expression support. We found that FCS fully inlines QueryBuilder methods before Fable sees the AST.

What happens:

1. query { for x in [1;2;3] do select x } gets desugared by FCS into QueryBuilder.Run(QueryBuilder.For(source, fun x -> QueryBuilder.Select(...)))
2. FCS then inlines these methods, which internally call LeafExpressionConverter.EvaluateQuotation
3. By the time Fable's FSharp2Fable processes the AST, the QueryBuilder calls are gone — the Replaced active pattern in FSharp2Fable.Util.fs is never reached
4. The inlined code becomes a Raise (since EvaluateQuotation can't be resolved at compile time)

We added replacement handlers for Microsoft.FSharp.Linq.QueryBuilder (Select, Where, For, Yield, etc. → Seq module operations), but they're unreachable because the calls are already inlined away.

Question: Does this analysis make sense? Are we right that FCS inlines the QueryBuilder before Fable can intercept it? If so, is there a way in the FCS fork to prevent this inlining, or to add a hook that lets Fable intercept QueryBuilder calls before they're resolved? This would enable mapping query { } to Seq operations on targets like Python and Beam.

[ncave]

@dbrattli

The FCS fork that Fable uses is intentionally a very thin layer/facade on top of the actual FCS, with no customization of the F# AST. That helps to shield us from the FCS internal churn and helps with migrating to new FCS versions, but it is true that it is sometimes hard to accurately reconstruct F# features from the F# AST, as exposed by FCS.

For example, in Fable we go through a lot of trouble to try to reconstruct the original F# match expressions, which get compiled into F# AST decision trees and targets by FCS, and that reconstruction back can never be perfect. I don't know if such reconstruction from F# AST is possible for quotations, does the F# AST retain enough fidelity for Fable to reconstruct the original form. Good question to ask in the dotnet/fsharp repo, perhaps.

That said, if you have a proposal for modifying the Fable FCS fork, we technically can make changes, assuming it is not a drastic modification of the F# AST that would impede future maintenance when adopting new FCS versions.

[MangelMaxime]

Deliberately left out JS to avoid increasing bundle size

In theory, if we don't use a code from Fable it gets removed by bundlers when generating the bundle.

This is for example, the case for things related to decimal which are quite heavy.

So this is probably fine, to add support for JavaScript too.

[dbrattli]

@ncave Thanks for the feedback. I will drop this for now. Probably easier to just make our own query builder outside the compiler if query syntax is needed.

[MangelMaxime]

When looking at the code from both Replacement.fs files, I am wondering if it would be worth sharing the code.

It seems like both code are identical match.