accelerating typing efforts

[henrydingliu]

Copilot gave this suggestion:

✅ Recommended workflow (tailored to you)
Given:

high test coverage ✅
strong engineering discipline ✅
care about API/docs quality ✅

Step 1 — Bootstrap with runtime inference

monkeytype run -m pytest
monkeytype apply your_package

Step 2 — Immediately run a strict type checker

mypy --strict
# or
pyright --strict

This step:

forces consistency across call graph
reveals incorrect runtime-derived hints

Step 3 — Manually fix interfaces, not implementations
Focus your effort on:

public APIs
class boundaries
key data structures

Typical upgrades:

# from MonkeyType
def f(x: list[int]) -> list[int]

# to something meaningful
from collections.abc import Sequence

def f(x: Sequence[int]) -> list[int]

Step 4 — Use static inference to fill gaps
Run:

pytype .

This helps:

discover untested paths
suggest missing annotations

(pytype can even emit .pyi files you can merge back)

[genedan]

If it gives us any comfort, it wasn't until March of this year that Pandas became type complete. For this reason, I think it's ok if this gets completed after 1.0, as it certainly was for Pandas.

https://pyrefly.org/blog/pandas-type-completeness/

Interestingly, Ruff was part of the process. We've been mentioning linters more and more these days, so I think we'll need one soon.

[henrydingliu]

using monkeytype wouldn't be a bad idea. i think the community really came together for the api docstring work. if one person used monkeytype on the whole repo, but then commit each .py to a separate branch. we can then crowd source the reviewing and editing. i mean, maybe it won't be as perfect as you going through it yourself over the course of a couple of months, but we'd get there a lot faster.

[genedan]

I have a pretty good vision on what I'd like out of chainladder.core, other people can take over the rest of the package, especially the estimators. For estimators, I think there should be protocol similar to TriangleProtocol since that part of the package also uses a mixin pattern, but other than that let's see what other ideas people can come up with. I'll see what monkeytype does when I take a look at the dunders, which is next on my list.

For now we can take a sledgehammer approach to prioritize coverage, just making sure the number goes up when you run pyright --verifytypes chainladder on each PR. Once the coverage % goes up we can shift attention to the consistency errors.