Proposal: Support Unpacked `TypeVarTuple` and `tuple` in `Concatenate`

先從連結的 typing 規範中關於有效使用位置的部分開始，然後將提議的語法和語意與此處描述的 PEP 646 和 PEP 612 行為進行比較。完成的標準是選擇並記錄一條用於拆分 ParamSpec P 的解包前綴的規則，其中包括未錨定和無界的情況。

Abstract

PEP 612 introduced ParamSpec and Concatenate to prepend fixed positional parameters to a callable's signature. PEP 646 introduced TypeVarTuple for variadic positional typing in Callable[[*Ts], R]. Because the two PEPs were developed independently, the typing specification does not allow unpacked types (*Ts or *tuple[...]) inside Concatenate.

This proposal extends Concatenate to accept unpack_expressions in its prefix, enabling Callable[Concatenate[*Ts, P], R].

---

Motivation

Higher-order abstractions like partial application helpers and execution wrappers need to capture an arbitrary number of leading positional arguments while preserving the remaining signature (keyword-only params, defaults, **kwargs) via ParamSpec. Today, this requires repetitive overload ladders:

from typing import Any, Callable, Concatenate, overload

class Wrapper[**P, R]:
    @overload
    def __init__(self, fn: Callable[P, R]) -> None: ...

    @overload
    def __init__[G1](
        self,
        fn: Callable[Concatenate[G1, P], R],
        __g1: G1,
        /,
    ) -> None: ...

    @overload
    def __init__[G1, G2](
        self,
        fn: Callable[Concatenate[G1, G2, P], R],
        __g1: G1,
        __g2: G2,
        /,
    ) -> None: ...

    # Must repeat up to arbitrary maximum arity...
    def __init__(self, fn: Callable[..., R], *args: Any) -> None:
        self._fn = fn
        self._args = args

    def __call__(self, *args: P.args, **kwargs: P.kwargs) -> R:
        return self._fn(*self._args, *args, **kwargs)

With this proposal, the entire ladder collapses to a single generic signature:

from __future__ import annotations
from typing import Callable, Concatenate

class Wrapper[**P, R, *Ts]:
    def __init__(self, fn: Callable[Concatenate[*Ts, P], R], *args: *Ts) -> None:
        self._fn = fn
        self._args = args

    def __call__(self, *args: P.args, **kwargs: P.kwargs) -> R:
        return self._fn(*self._args, *args, **kwargs)

def f(a: str, b: int, *, flag: bool = False, x: float) -> bool: ...

# Ts = () -> P = (a: str, b: int, *, flag: bool = ..., x: float)
w0 = Wrapper(f)
r0 = w0("hello", 42, x=3.14, flag=True)  # type: bool

# Ts = (str,) -> P = (b: int, *, flag: bool = ..., x: float)
w1 = Wrapper(f, "hello")
r1 = w1(42, x=3.14)  # type: bool

# Ts = (str, int) -> P = (*, flag: bool = ..., x: float)
w2 = Wrapper(f, "hello", 42)
r2 = w2(x=3.14)  # type: bool

---

Specification

Grammar

Update the Concatenate grammar in the typing specification from:

concatenate ::= "Concatenate" "[" type_expression ("," type_expression)* "," parameter_specification_variable "]"

to:

concatenate_prefix_item ::= type_expression | unpack_expression
concatenate ::= "Concatenate" "[" concatenate_prefix_item ("," concatenate_prefix_item)* "," parameter_specification_variable "]"

Semantics

Expansion follows existing PEP 646 semantics: when *Ts is bound to tuple[T1, T2, ..., Tn], Concatenate[*Ts, P] is equivalent to Concatenate[T1, T2, ..., Tn, P]. When *Ts is bound to tuple[()], Concatenate[*Ts, P] simplifies to P. Individual type expressions and unpack expressions may be freely combined in the prefix (e.g. Concatenate[LeadingArg, *Ts, P]).

---

Open Question: Splitting Boundary

The core design question is: how does a type checker determine the split between the prefix and ParamSpec P?

When the prefix length is statically known, splitting is unambiguous. This covers concrete bounded tuples (Concatenate[*tuple[int, str], P] — always length 2) and value-anchored TypeVarTuples where a companion *args: *Ts pins the length at the call site (the Wrapper example above). These are the primary use cases.

Ambiguity arises when the prefix length is not statically determined:

Case A — Unanchored *Ts (no companion *args: *Ts):

class TaskRunner[**P, R, *Ts]:
    def __init__(self, fn: Callable[Concatenate[*Ts, P], R]) -> None: ...

def compute(user_id: int, query: str, *, timeout: float = 5.0) -> bool: ...

# How many positional params should *Ts capture vs. leave in P?
task = TaskRunner(compute)

Case B — Unbounded tuple (*tuple[T, ...]):

def strip_leading_ints[**P, R](
    fn: Callable[Concatenate[*tuple[int, ...], P], R]
) -> Callable[P, R]: ...

def example(x: int, y: int, z: int, *, flag: bool = False) -> None: ...

# *tuple[int, ...] could match 0, 1, 2, or 3 leading int parameters.
wrapped = strip_leading_ints(example)

Options:

• Option 1 — Greedy prefix: The prefix consumes all matching positional-capable parameters. In Case A, *Ts = (int, str) and P = (*, timeout: float = 5.0). In Case B, all 3 ints are consumed, leaving P = (*, flag: bool = False).

• Option 2 — Restrict to fixed-length prefixes initially: Require the prefix length to be statically determined (concrete bounded tuples, companion *args: *Ts, or explicit specialization). Reject unanchored/unbounded prefixes as ambiguous and defer them to a future extension.