Find out which Python version added each feature your code uses.
Point sincewhen at a file and it will tell you what's in there and how long each piece of it has been in Python, back to the first public release in 1991.
sincewhen parses with the standard library's ast and ships no dependencies, so the whole tool runs in a browser tab.
Paste code in at pym.dev/since to see what it reports without installing anything.
The analysis runs on your own machine, in the tab, so nothing you paste is uploaded anywhere.
Installing with uv tool:
uv tool install sincewhenInstalling with pipx:
pipx install sincewhenYou can also install sincewhen globally with pip, but I usually recommend installing command-line tools in their own separate environment.
Give sincewhen a file to see every dated feature it uses:
# example.py
import tomllib
def load(path, /):
with open(path, "rb") as config:
return tomllib.load(config)$ sincewhen example.py
example.py:1 tomllib module 3.11 2022-10-24
example.py:4 positional-only parameters (/) 3.8 2019-10-14
example.py:5 with statement 2.5 2006-09-19
example.py:5 open() 0.9 1991-02-20
example.py:6 tomllib.load() 3.11 2022-10-24The last column is the day that release shipped, so a version number reads as an age. Every release the dataset can name has one, back to 1991-02-20.
Pass as many files as you like to see them all in one report.
Only the first use of each feature is shown by default, so a feature used a hundred times costs one line.
Pass --all to see every occurrence.
The dataset reaches back to Python 0.9.1, so a file will report things like str() and open() that have been there since 1991.
That is usually the point, but pass --since when you only want the recent arrivals:
$ sincewhen --since 3.0 example.py
example.py:1 tomllib module 3.11 2022-10-24
example.py:4 positional-only parameters (/) 3.8 2019-10-14
example.py:6 tomllib.load() 3.11 2022-10-24Read from standard input with -:
$ echo 'x: int = 1' | sincewhen -
<stdin>:1 variable annotation 3.6 2016-12-23
<stdin>:1 int() 0.9 1991-02-20Look a single feature up by name instead of analyzing code:
$ sincewhen --search walrus
walrus operator (:=) - Python 3.8 (released 2019-10-14)
PEP: https://peps.python.org/pep-0572/
Docs: https://docs.python.org/3/whatsnew/3.8.htmlA method of a builtin type answers to its own name, without the type in front of it:
$ sincewhen --search removeprefix
removeprefix() on str, bytes and bytearray - Python 3.9 (released 2020-10-05)
Docs: https://docs.python.org/3/whatsnew/3.9.htmlA module member with no entry of its own is answered from the member index, which covers every documented member of every stdlib module:
$ sincewhen --search platform.system
platform.system - Python 2.3 (released 2003-07-29)
$ sincewhen --search os.path.relpath
os.path.relpath - Python 2.6 (released 2008-10-02)A member nothing can date is bounded, exactly as an entry would be:
$ sincewhen --search os.path.join
os.path.join - Python 1.5 or earlier (released 1997-12-31)A bare name searches every module's member list, which is how a name usually gets typed:
$ sincewhen --search TimeoutError
No entry for 'TimeoutError'. Did you mean one of these?
multiprocessing.TimeoutError - Python 3.3 (released 2012-09-29)
asyncio.TimeoutError - Python 3.4 (released 2014-03-17)
concurrent.futures.TimeoutError - Python 3.5 (released 2015-09-13)Pass --json to either mode for machine-readable output.
A search that answers from the member index emits member records, which carry module where a feature record carries id and category.
Dating 1,847 features against seven independent sources turns up a lot of history that is easy to misremember, and nearly all of it is one search away:
$ sincewhen --search sorted
sorted() - Python 2.4 (released 2004-11-30)
Docs: https://docs.python.org/3/whatsnew/2.4.htmlHere are a few of my favorites.
The standard library is full of arcs: a series of tools for one problem, each solving it a bit better than the last.
Counting.
How do you count things in Python?
The answer kept improving for thirteen years: dict.get arrived in 1997 (Python 1.5), dict.setdefault in 2000 (2.0), dict.fromkeys in 2003 (2.3), collections.defaultdict in 2006 (2.5), and collections.Counter in 2010 (2.7).
Counting things in Python walks this same history as code, refactoring one loop forward through every release.
String formatting.
Python has gone through several waves of string interpolation.
Percent formatting was there at the beginning, in 1991.
Then string.Template in 2004 (2.4), str.format in 2008 (2.6), f-strings in 2016 (3.6), and t-strings in 2025 (3.14).
All five still work.
T-strings in Python lays them out side by side.
Records.
"A class that just holds some fields" sounds like a solved problem.
Tuples were the answer in 1991, but Python has added new answers since: collections.namedtuple in 2008 (2.6), types.SimpleNamespace in 2012 (3.3), typing.NamedTuple in 2015 (3.5), dataclasses in 2018 (3.7), and typing.TypedDict in 2019 (3.8).
Running a subprocess.
os.system is from 1997 (1.5).
subprocess.Popen arrived in 2004 (2.4) with the stated aim of replacing it, subprocess.check_output in 2010 (2.7), and subprocess.run in 2015 (3.5).
os.system is still there in Python 3.14.
Comprehensions came six years after map() and filter().
map(), filter(), and lambda are all from 1994 (1.0), so the whole functional trio was there together before any comprehension syntax existed.
List comprehensions arrived in 2000 (2.0), generator expressions in 2004 (2.4), and dict and set comprehensions in 2010 (2.7).
sorted() is nearly fourteen years younger than list.sort().
Sorting a list in place worked in 1991 (0.9); sorting anything into a new list arrived in 2004 (2.4).
reversed() is also from 2004 (2.4), against a list.reverse() from 1994 (1.0).
enumerate() took twelve years.
It arrived in 2003 (2.3), so for i in range(len(items)) was simply how you numbered things for a long time.
zip() beat it by three years, so from 2000 to 2003 you could pair two lists together but still could not number one.
Curly braces meant only dictionaries for sixteen years.
Dict literals date to 1994 (1.0).
Sets arrived as a module in 2003 (2.3), became the set() and frozenset() builtins in 2004 (2.4), and finally got literals, and set comprehensions with them, in 2010 (2.7).
String methods did not exist for Python's first nine years.
Until 2000 (1.6), splitting a string meant string.split(line) rather than line.split().
That release added twenty-nine string methods at once, and the string module kept growing anyway: string.ascii_lowercase is from 2001 (2.2) and string.Template from 2004 (2.4), both newer than the methods that took away the module's original job.
Today string has twelve public names left.
The documentation is a record kept by many people over thirty-five years, and a module's early history is genuinely hard to reconstruct after the fact.
An "Added in version" marker attaches to whatever signature sits above it, which is not always the thing it was written about: compile() is from 1994 (1.0), and the nearest marker in the current docs says 3.8 because it documents a flags value.
The Python 2.7 docs date bisect to 2.1, while Lib/bisect.py is already in the 1.0 tarball, from 1994.
And a PEP records what was planned rather than what shipped: class decorators are headed Python-Version: 3.0 by PEP 3129, and CPython's 2.6 grammar already has the rule.
That is why every version here is checked against seven independent sources rather than trusted to any one of them. Twenty entries in the dataset carry an evidence note recording that the documentation dates them to a different release than the one they shipped in.
>>> import sincewhen
>>> sincewhen.minimum_version("import tomllib")
Version(major=3, minor=11)
>>> [d.feature.name for d in sincewhen.detect("if (n := 1): pass")]
['walrus operator (:=)']
>>> sincewhen.lookup("tomllib")[0].added
Version(major=3, minor=11)
>>> sincewhen.minimum_version("x = 1") is None
True
>>> sincewhen.lookup_member("platform.system").since
'2.3'
>>> [answer.dotted for answer in sincewhen.find_members("system")]
['os.system', 'platform.system']minimum_version() returns None when nothing detected sets a floor, which means no known feature requires a particular release rather than that the code runs anywhere.
A bounded feature sets no floor either: "1.5 or earlier" is a limit on what the sources could read rather than a date, so import zlib on its own also gives None.
sincewhen parses code with the standard library's ast module, which can only understand syntax that the running interpreter understands.
A Python 3.9 interpreter cannot parse a match statement, so it could not report one either.
Requiring the newest Python is what lets sincewhen recognize the newest syntax.
The Python version you run sincewhen on has nothing to do with the versions it reports on, which reach back to Python 0.9.1.
- Some features are genuinely ambiguous in an AST.
a | bcould be a Python 3.9 dict merge, a Python 3.10 union type, or an integer bitwise-or that has worked since forever, and the AST alone cannot tell you which. Ambiguous features are left out rather than guessed at. - Detection is syntactic.
sincewhensees that you called something namedmath.isclose, not that you called the real one. Shadowed builtins are skipped, but a shadowed module attribute is not. - The dataset is curated and incomplete, at about 1,800 entries today. A feature that isn't in it won't be reported, so the minimum version is a lower bound on the true answer.
addedis the oldest release from which a feature has been available ever since, ignoring Python 3.0 and 3.1. Nobody shipped code on those two, so a gap there is not a gap anyone lived through:argparseshipped in 2.7 and again in 3.2, and is dated 2.7. A feature missing from 3.2 as well has a real gap and takes the later date.- Some features cannot be dated, only bounded, and those say so:
zlibreads as "1.5 or earlier" rather than as 1.5. Three entries are bounded today.resourceandzlibare already in Python 1.5, the oldest archived documentation build, so they are at least that old and may be older.os.pathreads "1.2 or earlier" for a different reason: Python 1.1 ships it, but the 1.1 interpreter built from that release's own tarball cannot import it, so that absence belongs to the build rather than to the release and cannot date anything. A bound is a limit on what the sources could read rather than a date, so a bounded feature is left out of the minimum version entirely. - Anything present in Python 0.9.1 is dated 0.9, the first public release, and reads like any other version. Python began before it was published, so a few of those are genuinely older, but there is no earlier release for them to have been added in: nothing has been in Python longer than Python has been public. Those entries are dated rather than bounded for that reason, and each one's evidence still records that it may predate the public record.
- Release dates come from python.org's downloads database back to 2.2, and from CPython's release tags before that. Python 0.9 and 1.6 are the two rows neither source reaches, 0.9 because CPython's history begins after it and 1.6 because it was cut by BeOpen and has no tag, so both are taken from Wikipedia's table of versions. The 0.9 date is that table's date for the whole 0.9 line, while the corpus reads the 0.9.1 tarball, which was cut within days of it and which no source dates on its own.
- Searching for a module member with no entry of its own falls back to the member index, which covers about 3,700 members across 242 modules. Its versions are derived by the same machinery that rechecks every entry in the dataset, so an answer from it is the answer an entry would carry; what it does not carry is the evidence, which is most of what an entry is for. It holds only names the newest Python still documents, so a member Python 3 removed is not in it; only names the sources agree about; and only names something corroborates, so a member whose sole evidence is a 3.x inventory diff is left out rather than dated by the age of its markup. A member the index has never heard of falls back to the module it lives in, as before.
- A method of a builtin type is dated for searching but is mostly not detected, because
value.removeprefix(...)says nothing about whatvalueis. Only a receiver whose type is certain reports one: a literal, as in"Mr. Smith".removeprefix("Mr. "), or the type's own name, as indict.fromkeys(keys).
This project uses uv and just.
Run just to see every available task.
$ just test # run the test suite
$ just check # format, lint, typecheck, and testNo setup step is needed: uv creates the virtual environment and installs dependencies on the first uv run.
If you would rather not install just, every task is a short uv command that you can run directly (check the justfile for the commands):
uv run pytestFeatures live in src/sincewhen/features.toml, one [[features]] table each.
Give the feature an id, a human-readable name, the version that added it, a category, exactly one matcher, and the evidence for the version:
[[features]]
id = "walrus"
name = "walrus operator (:=)"
added = "3.8"
category = "syntax"
pep = 572
nodes = ["NamedExpr"]
[features.evidence]
method = "pep"
pep = 572
python_version = "3.8"
checked = "2026-07-28"The matcher kinds are nodes (AST node class names), builtins, modules, attributes (dotted module.name paths), and methods (dotted type.method paths for the builtin types).
Node matchers can be narrowed with requires (a node attribute that must be truthy) or check (a predicate registered in detect.py).
Documentation links are generated from added and pep, so only set docs when you have a better link than the "What's New" page.
Nobody should be typing version numbers from memory. For anything in the standard library, let the archived documentation say what the version is:
$ just fetch-docs # one-time, ~2 GB into a gitignored .cache/
$ just whenadded math.lcm # what each source says, and whether they agree
$ just propose math.lcm math.isqrt # entries with evidence, ready to paste
$ just typemethods --compare # what the builtin types' method tables date
$ just verify-dataset # re-derive every claim in the datasetjust verify-dataset also runs in CI, so a pull request that edits a version without editing its evidence fails.
Evidence has eight method values, seven of which a machine can recheck:
| method | what it means |
|---|---|
objects.inv |
the symbol is absent from one release's Sphinx inventory and present in the next |
archive |
the same diff over the module lists and built-in function pages in the pre-Sphinx doc builds, back to the 0.9.1 LaTeX |
source |
the name is absent from one release's own C or Python implementation and present in the next, which reaches back further than any doc build. For a method of a builtin type this is the type's own method table, which is the only thing that can date dict.setdefault or str.split at all |
interpreter |
that release's own interpreter, built from its tarball, was asked whether the name resolves. Thirty-one of them, 0.9.1 to 3.14 |
annotation |
the documentation dates it itself, in an "Added in version" marker quoted in the entry |
grammar |
the token is absent from one release's grammar and present in the next, which is what shipped rather than what a PEP intended |
pep |
the feature's PEP carries a Python-Version header |
manual |
a human read the archives and wrote down what they found, and why the other seven do not settle it |
interpreter is the only method that reads Python rather than a description of it, and it outranks the rest across the whole timeline, 0.9.1 to 3.14.
It is what settles a name no text can account for: re.finditer is defined in Python 2.2's sre.py and left out of its __all__, so from sre import * never bound it and the answer is 2.3, not the 2.2 the docs claim.
On the 3.x line it is the only cross-check objects.inv has, and the inventory dates documentation: shutil.SpecialFileError is first indexed in 3.13 and resolves from 2.7 on.
Building the interpreters needs Docker and the better part of an hour, so the result is committed as scripts/interpreters.json and nothing downstream needs a compiler:
$ just build-pythons # build 0.9.1 through 3.14 (slow)
$ just build-pythons modern # or just one half of it
$ just probe-pythons # ask them all, and record it
$ just interpreters-vs-dataset # where they disagree with the datasetmanual is for the cases where the sources genuinely disagree, and every one of them is printed on every verify-dataset run so the override stays visible.
src/sincewhen/members.txt is the other file the package ships, and it is generated rather than curated.
It is the member index search falls back to, and it is derived from the same cache by the same rules, so it is rebuilt rather than edited:
$ just memberindex --write # regenerate the index from the cache
$ just memberindex --module platform # what the index has for one module
$ just memberindex --grep system # every module with a member of that namejust verify-dataset re-derives it and fails if the committed file is not what the sources produce, exactly as it does for every version claim.
A new entry, or a corrected version on an existing one, also gets a line under Unreleased in CHANGELOG.md.
Dataset changes are the ones that alter what sincewhen reports about code that did not change, so they are worth spelling out.
Move the Unreleased notes in CHANGELOG.md under a heading for the new version, then bump, tag, and push:
$ just bump patch
$ just releasejust release refuses to tag a version that the changelog has nothing to say about, so the notes have to be written before the release goes out rather than after.
Pushing a v* tag runs the release workflow, which publishes to PyPI with trusted publishing and creates a GitHub release whose notes are that version's changelog section.
sincewhen is distributed under the terms of the MIT license.