Pyrefly Type Coverage Skill

Prerequisites

• The file must live in a project with a pyrefly.toml.
• pyrefly, lintrunner, and the project's test runner must be on PATH. **If any

are missing, stop and ask whether a conda environment needs activating** — don't install or substitute (per repo CLAUDE.md).

Step 1: Remove file-level type-check suppressions

Delete any of these from the top of the file (pyrefly honors # mypy: ignore-errors for mypy compat, so that one must go too):

# pyre-ignore-all-errors
# pyre-ignore-all-errors[16,21,53,56]
# @lint-ignore-every PYRELINT
# mypy: ignore-errors

Step 2: Add a sub-config entry to pyrefly.toml

[[sub-config]]
matches = "path/to/directory/**"
[sub-config.errors]
implicit-import = false
implicit-any = true
bad-param-name-override = false
unannotated-return = true
unannotated-parameter = true

IMPORTANT: Setting any error key in [sub-config.errors] overrides only that key relative to the parent — but enabling unannotated-return / unannotated-parameter / implicit-any will resurface errors that were previously hidden file-wide. If you see unrelated errors (e.g., bad-param-name-override) flooding the output, mirror the parent config's setting for that key in the sub-config to silence them.

Step 3: Run pyrefly

pyrefly check <FILENAME>

Goal: resolve all unannotated-return, unannotated-parameter, and implicit-any errors by adding annotations — see Step 4's ladder. These three target categories are always resolvable; never suppress them with # pyrefly: ignore. The single exception is @compatibility(isbackwardcompatible=True) (Step 4).

Other categories (bad-argument-type, missing-attribute, …) are real type bugs. Handle them by where pyrefly reports them:

• Reported in another file (path != target): leave it. Don't widen scope. If

the error is now blocking the target, suppress at the report site with # pyrefly: ignore[<category>] # TODO.

• Reported in the target file but the message names a symbol defined elsewhere

(e.g., bad-return because an imported function's annotation is wrong): suppress locally with the same TODO comment. Don't invent a cast() that papers over the upstream gap.

• Reported in the target file, originates locally: fix it.

Use # pyrefly: ignore[...] only as a last resort, and only on non-target categories.

Step 4: Add annotations

Examine call sites when the right type isn't obvious from the function body.

Annotation conventions

• Use PEP 604 / PEP 585 syntax (int | None, list[str]) — assume Python >= 3.10.
• Prefer collections.abc over typing for ABCs (Callable, Sequence, Generator, ...).
• For generic helpers, import from typing when available on the project's minimum

Python version, and from typingextensions only when you need a newer feature (e.g., Self and override if supporting < 3.11/3.12, or PEP 696 default= for TypeVar / ParamSpec). Don't blanket-import from typingextensions.

• Always parameterize Callable (never bare Callable). Prefer

Callable[..., object]; reach for Callable[..., Any] only when a caller genuinely consumes the dynamic return — if the result is just passed through (or the callable isn't even invoked), object is stricter and equally correct. (See ParamSpec below for the signature-preserving wrapper case.)

• Give any module-local global you introduce a leading underscore —

TypeVar/ParamSpec (matching the string arg: T = TypeVar("T"), P = ParamSpec("P"), R = TypeVar("R")), TypeAliases, helper constants, and sentinels alike. This is the prevailing torch convention for non-public names (_P outnumbers P ~6:1 in the tree). Exceptions (leave un-underscored): a name imported by other modules, listed in all, or used as a runtime token (e.g. an annotation-string dispatch marker). Applies only to names you add — do not rename pre-existing globals; that's an unrelated refactor outside this skill's scope.

• A boolean predicate — is/has name, takes a broad type (often object),

returns bool — usually wants TypeGuard[X] (or TypeIs[X], which also narrows the negative branch). TypeGuard is in typing (>= 3.10, so import from there); TypeIs only entered typing in 3.13, so import it from typingextensions (>= 4.10) to stay 3.10-compatible. An issubclass-style helper taking klass: type[T] should return TypeGuard[type[_T]]. Prefer an explicit isinstance(x, type) guard over try/except TypeError around issubclass() — clearer, and it lets the checker narrow.

• When a return type is derived from a parameter — passthroughs/identity

functions, "return one of these args" helpers, decorators, registries keyed by type — reach for a TypeVar (or, for a callable arg whose signature flows through, Callable[P, R] with ParamSpec/TypeVar) rather than widening to object/Any. "Output type == some input type" is exactly what a TypeVar encodes; object in / object out discards it. Caveat: if the function transforms the value so the output type differs from the input (e.g. converts an array to an int), a single TypeVar is wrong — name the actual domain type instead.

• Class attributes assigned in init should get a class-level annotation so pyrefly can see them.
• Break import cycles with if TYPE_CHECKING: — annotation-only imports go inside the

guard, and use from future import annotations (or string forward refs) so runtime imports stay lazy: ``python from future import annotations from typing import TYPECHECKING if TYPECHECKING: from torch.fx import GraphModule def transform(gm: GraphModule) -> GraphModule: ... ``

• Never suppress the three target categories. unannotated-return,

unannotated-parameter, and implicit-any are always resolvable by adding an annotation; # pyrefly: ignore[<one of those>] is not an acceptable outcome. The single exception is the Backward compatibility carve-out below.

• Widen, don't bail. When the right type is hard to infer, walk down this

ladder rather than reaching for an ignore: 1. Most specific concrete type observable from call sites and return paths. 2. A union (X | Y), Sequence[X]-style abstract type, or a bound TypeVar for genuinely generic functions (identity-passthrough, container helpers). 3. object — strictest fallback that still type-checks. Forces callers to narrow before use, e.g., def serialize(value: object) -> str:. Visually similar to Any but stricter — pyrefly rejects value.foo() without an isinstance. 4. Any — last rung. Always preferred over a # pyrefly: ignore on a target category, but only after rungs 1–3 fail. Be able to articulate why each earlier rung doesn't fit (e.g., "union exceeds 8 types", "no observable common bound", "callers genuinely never narrow").

• Be especially wary of object/Any in return position — a function

usually knows more about what it produces than its callers do. A wide return is right only at a genuine boundary (it returns its input unchanged, or the value is handler/caller-defined); if the body builds a known shape, name it (a domain alias or union beats object).

• Read at least three call sites before deciding a parameter must be Any —

don't pattern-match "looks dynamic" on the first try.

• Narrow-scope # pyrefly: ignore[...] (on a non-target category) is reserved

for cases where pyrefly is actually wrong about a specific local error — dynamic metaprogramming, third-party stub gaps: ``python # pyrefly: ignore[attr-defined] result = getattr(obj, dynamic_name)() ``

• When an inline # pyrefly: ignore[...] would push a line past the length

limit, put it on the line immediately above the flagged line rather than reaching for # fmt: skip to keep it inline — pyrefly honors a previous-line ignore. (Exception: the backward-compat carve-out below, where it must sit on the def line.)

Backward compatibility (the one exception to never-suppress)

CRITICAL: Functions decorated with @compatibility(isbackwardcompatible=True) must NOT have their signatures changed. The backward-compat test (testfunctionback_compat) compares stringified inspect.signature against a golden file — adding annotations (even -> None) changes that string and the test fails. Use pyrefly ignore comments instead:

@compatibility(is_backward_compatible=True)
def my_function(  # pyrefly: ignore[unannotated-return]
    self,
    arg1,  # can't add type here either
):
    ...

The # pyrefly: ignore comment must be on the def line (where pyrefly reports the error), not on the closing ).

ParamSpec for signature-preserving wrappers (decorators, functools.wraps-style helpers). Use Callable[P, R] so the wrapped function's signature flows through to the caller — Callable[..., Any] loses it. Skip ParamSpec if the wrapper genuinely accepts arbitrary callables. Pair with Concatenate[X, P] when the wrapper prepends or appends args.

from collections.abc import Callable
from typing import ParamSpec, TypeVar

_P = ParamSpec("_P")
_R = TypeVar("_R")

def log_calls(fn: Callable[_P, _R]) -> Callable[_P, _R]:
    def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> _R:
        return fn(*args, **kwargs)
    return wrapper

Step 5: Iterate

Re-run pyrefly check. New annotations often surface bad-return errors where the function actually returns an incompatible type — fix those. Repeat until clean.

Tightening a shared helper (e.g. adding a TypeGuard or a precise return) can make pre-existing # pyrefly: ignore comments in its callers unused. Re-check and delete now-dead suppressions and any stale explanatory comments — don't leave them.

Step 6: Lint

Required before handing off — annotations frequently shift import order and line length:

lintrunner -a <files...>

Resolve anything lintrunner can't auto-fix manually.

Step 7: Test

Precedence when something fails: tests passing > pyrefly clean > annotation strictness. If a freshly-added annotation breaks a test, narrow it one rung in the discipline ladder (e.g., concrete → object, or remove an Any widening that broke a downstream isinstance check) before reverting the file.

1. Backward-compat check. Run iff

grep -l '@compatibility(isbackwardcompatible=True)' <target> returns the file — the decorator is the actual precondition for the golden file. The broader "imports torch.fx" heuristic catches half of torch/. ``bash python -m pytest test/test_fx.py::TestFXAPIBackwardCompatibility -x -v ``

1. Unit tests for the modified module. Search both ways before concluding

no coverage exists: ``bash # torch/foo/bar.py is usually covered by test/testfoo.py or test/testbar.py ls test/ | grep -i <module-name> # or by import grep -rl "from torch.foo.bar import\|import torch.foo.bar" test/ ``

If both come up empty, tell the user — don't silently skip. Type changes can introduce real runtime regressions (Optional[X] vs X, Sequence vs list when .append is called, etc.).

Notes

• Forward refs in class bodies without from future import annotations

still need string quoting: ``python class MyClass: def new(cls) -> "MyClass": ... ``

• Committing: don't commit unless the user explicitly asks (per repo

CLAUDE.md). Stop and surface the diff for review when the file is clean.