nepkit

                      _     _  _
 _ __    ___   _ __  | | __(_)| |_
| '_ \  / _ \ | '_ \ | |/ /| || __|
| | | ||  __/ | |_) ||   < | || |_
|_| |_| \___| | .__/ |_|\_\|_| \__|
              |_|

Typed Bikram Sambat ↔ Gregorian date conversion for Python, as a library and a command-line tool.

Status: first release. The library and CLI both work and are tested; the API may still change before 1.0.

DEMO.md walks through every command, option, and failure mode with real captured output.

Why

Bikram Sambat is Nepal's official calendar, and converting to and from it is not arithmetic. Gregorian leap years follow a rule you can write down; BS month lengths do not. They vary between 29 and 32 days with no generating formula, are fixed by observation, and are published by Nepal's Panchanga authority. Every correct converter is therefore a lookup table plus one verified anchor date — which means the data matters more than the code, and most of the work in this repo went into the data.

The existing Python options are small, mostly unmaintained packages with undocumented year ranges, no type hints, and no statement of where their numbers came from or how far they can be trusted. nepkit aims to be one typed, tested converter that is explicit about all three.

Install

Requires Python 3.12+.

pip install nepkit       # library and CLI
uv add nepkit            # into a uv project
uv tool install nepkit   # the CLI on its own, isolated

Typer is the only direct dependency, plus pyreadline3 on Windows alone, which gives the interactive prompt the line editing CPython omits there. The calendar table is bundled, so nepkit never touches the network.

Windows (Git Bash)

uv tool install puts nepkit.exe in uv's executable directory, which Git Bash does not carry on PATH by default — the install reports success and the command is still not found. Add the directory, then confirm:

uv tool install nepkit
uv tool dir --bin        # C:\Users\you\.local\bin
uv tool update-shell     # add it to the Windows user PATH

uv tool update-shell writes to the HKCU\Environment registry key rather than to ~/.bashrc, so only a newly launched tab sees it, and Windows does not reliably notify running processes of the change (uv#17331). If a fresh tab still comes up empty, set it from bash instead — under Git Bash $HOME is /c/Users/you, so this resolves to the same directory uv tool dir --bin prints:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

Tab completion is the same one-liner as on Linux:

nepkit --install-completion

It appends a source line to ~/.bashrc pointing at ~/.bash_completions/nepkit.sh, spelled as a Windows path (C:\Users\...). That looks wrong inside a bash file, but MSYS2 resolves both spellings, so it works as written.

Library

from datetime import date

from nepkit import BSDate, ad_to_bs, bs_to_ad

bs_to_ad(BSDate(2081, 4, 15))  # date(2024, 7, 30)
ad_to_bs(date(2024, 7, 30))  # BSDate(year=2081, month=4, day=15)

BSDate validates on construction, so if you are holding one it is a real date in the supported range:

from nepkit import BSDate, days_in_month

days_in_month(2081, 4)  # 32
days_in_month(2081, 9)  # 29

BSDate(2081, 4, 33)  # raises InvalidDateError
BSDate(2081, 13, 1)  # raises InvalidDateError
BSDate(2095, 1, 1)  # raises DateOutOfRangeError

ad_to_bs takes a datetime.date, so a malformed Gregorian date is impossible by construction — Python's own constructor rejects it before nepkit is involved.

Errors

NepkitError
├── CalendarDataError   the bundled table is malformed; raised at import
└── DateError
    ├── InvalidDateError      not a real BS date (month 13, day 33, ...)
    └── DateOutOfRangeError   a real date, but outside the bundled range

The split between the last two is the one that earns its keep. BS 2095-03-12 is a perfectly real date that nepkit simply has no data for, and a caller can reasonably catch that and fall back or report the supported range. BS 2081-13-01 is not a date at all, and catching it is always a mistake. Catch DateError if you only need "the user gave me something I can't convert".

CLI

pip install and uv tool install put a nepkit command on your PATH. uv add installs into the project's .venv without activating it, so reach the command with uv run nepkit — or activate the venv first.

$ nepkit bs2ad 2081-04-15
2024-07-30 Tue

$ nepkit ad2bs 2024-07-30
2081-04-15 Tue

$ nepkit today
BS 2083-04-27 Wed
AD 2026-08-12 Wed

$ nepkit range
BS 2000-01-01 .. 2090-12-30  (years 2000-2090)
AD 1943-04-14 .. 2034-04-13

Direction is always explicit, and has to be: the BS and AD year numbers overlap from 2000 to 2034, so 2024 is a valid year in both calendars and nothing could reliably guess which one you meant.

Interactive

Run nepkit with no arguments in a terminal and it clears the screen and opens a session:

$ nepkit
                      _     _  _
 _ __    ___   _ __  | | __(_)| |_
| '_ \  / _ \ | '_ \ | |/ /| || __|
| | | ||  __/ | |_) ||   < | || |_
|_| |_| \___| | .__/ |_|\_\|_| \__|
              |_|
nepkit v0.2.0 - Bikram Sambat (BS) <-> Gregorian (AD) date conversion
Today  BS 2083-04-27   AD 2026-08-12 Wed

Type a command, 'help', 'clear', or 'quit'.  Up/Down recalls history.

nepkit> today
BS 2083-04-27 Wed
AD 2026-08-12 Wed

nepkit> bs2ad 2081-04-15
2024-07-30 Tue

nepkit> quit

It accepts exactly the commands above — the same table, not a parallel interface — so anything you can type at the shell works here unchanged. A bad line reports the error and returns you to the prompt rather than ending the session. clear (or cls) wipes the screen and redraws the banner. quit, exit, q, and Ctrl-D all leave; Ctrl-C abandons the current line only.

Those four are prompt-only words, not subcommands — nepkit clear at a shell stays a usage error rather than clearing your terminal.

Up and Down recall previous commands, and the usual readline editing keys work (Ctrl-A, Ctrl-E, Ctrl-R, and so on). History lasts for the session and is not written to disk. CPython ships no readline on Windows, so nepkit depends on pyreadline3 there to supply one — it needs a real console, which Windows Terminal provides. Should it fail to load, the prompt still works; it loses the editing keys and says so in its banner.

Only on a terminal. With stdin redirected — a pipeline, a script, CI — nepkit prints help and exits 2 exactly as before, so nothing ever blocks waiting for a prompt that isn't there.

Calendars

$ nepkit calbs 2081 4
        Shrawan 2081
    16 Jul - 16 Aug 2024
Sun Mon Tue Wed Thu Fri Sat
          1   2   3   4   5
  6   7   8   9  10  11  12
 13  14  15  16  17  18  19
 20  21  22  23  24  25  26
 27  28  29  30  31  32

$ nepkit calad 2024 7
         July 2024
Ashadh 17 - Shrawan 16, 2081
Sun Mon Tue Wed Thu Fri Sat
      1   2   3   4   5   6
  7   8   9  10  11  12  13
 14  15  16  17  18  19  20
 21  22  23  24  25  26  27
 28  29  30  31

Both default to the current month. A Gregorian month never lines up with a BS month, so the subtitle names both ends of the span rather than pretending a single corresponding month exists.

Grids are boxed and coloured on a terminal and plain when redirected, following the same convention as ls and git. Force it either way with --color always|never|auto.

On a terminal, today's date is picked out in bold bright cyan — the bright form of the box's own border colour, so the highlight belongs to the grid containing it rather than arriving from a different palette. The highlight exists only in the coloured path: piped output is byte-for-byte identical whether or not today falls in the month shown, so nothing parsing stdout breaks on the one day a month it would otherwise appear. --json reports it as a today field instead, which is null when today is elsewhere.

Scripting

Every command takes --json:

$ nepkit ad2bs 2008-05-28 --json
{"bs": "2065-02-15", "ad": "2008-05-28", "weekday": "Wed"}

stdout carries results, stderr carries errors, and neither ever carries both. stdout contains no ANSI escapes unless you ask for colour explicitly, so piping is always safe.

Exit codes come straight from the exception hierarchy, so a script can branch without parsing any text:

The 3/4 split is the one that matters when scripting: 4 is worth retrying against another source, 3 never is. Collapsing both into 1 would throw that distinction away at exactly the boundary where it is most useful.

Supported range

That is 91 years, 33,238 days. Anything outside it raises DateOutOfRangeError rather than extrapolating, because there is no rule to extrapolate with — dates beyond the table would have to be invented.

The bounds are computed from the bundled data, not written down separately, so extending the table moves them automatically.

Design decisions

One anchor, everything else derived. The whole library hangs on a single verified correspondence: BS 2000-01-01 = AD 1943-04-14. That is the only Gregorian fact in the package that cannot be computed, because month lengths alone cannot tell you where the calendar sits against the Gregorian one. Every other bound — the last BS date, both ends of the AD window — is derived from it plus the table. A second hardcoded date would be free to drift out of sync, and the failure would be silent and total.

Both directions collapse to a day count. A BS date becomes "days since the anchor", integer arithmetic happens there, and the result expands out the other side. datetime.date is already a correct expander for the Gregorian side, so bs_to_ad is one line; the real work is the inverse, which has no equivalent in the standard library.

Types instead of validation where possible. ad_to_bs accepts a datetime.date rather than three integers, which removes an entire error class from its contract at no cost.

Limitations

• Dates only. No time of day, no timezones, no Nepali-language month names or numeral formatting.
• The range is hard-bounded at BS 2000–2090 and will not extrapolate.
• Correctness rests on the data, and the tests cannot prove it. The test suite verifies self-consistency exhaustively — every one of the 33,238 days round-trips, and consecutive day counts produce consecutive dates. But both directions read the same table, so a wrong month length cancels out exactly and every property still passes. This was verified by deliberately corrupting the table: all properties passed while conversions were silently wrong. Only the sourcing described in src/nepkit/data/DATA.md stands behind the numbers themselves.

Data provenance

src/nepkit/data/DATA.md records where the calendar table came from: two independently maintained sources with different authors, languages, and conversion epochs, pinned at specific commits, diffed row by row over all 91 years with no disagreements, and cross-checked by walking each source forward from its own epoch to the anchor.

Development

This project uses uv.

git clone https://github.com/akakritagya/nepkit
cd nepkit
uv sync --group dev          # pytest, ruff, mypy, pre-commit
uv run pytest                # run tests
uv run ruff check .          # lint
uv run ruff format .         # format
uv run mypy                  # type check

uv run pre-commit install                         # lint/format/type-check on commit
uv run pre-commit install --hook-type commit-msg  # enforce Conventional Commits
uv run pre-commit install --hook-type pre-push    # full test suite before push

pre-commit runs lint, format, and type-check on every commit; blocks large files, private keys, and direct commits to main; runs the full pytest suite before push (not on every commit — too slow to survive contact with a growing suite); and enforces Conventional Commits on the message.

Releasing

Releases go to PyPI through Trusted Publishing — GitHub Actions proves the repository's identity over OIDC, so no API token exists to leak or rotate. Bump version in pyproject.toml, then:

git tag v0.2.0 && git push origin v0.2.0

The tag runs .github/workflows/publish.yml, which re-runs the full gate, refuses to continue if the tag disagrees with the packaged version, and only then uploads. Run the workflow manually to rehearse against TestPyPI first.

License

MIT — see LICENSE.