pyclig

A Python Click wrapper with Command Line Interface Guidelines (CLIG) paradigms.

click owns parsing, help, shell completion & the test runner. pyclig adds the conventions on top: STDOUT for output & STDERR for messages, exit codes, error rendering, logging, output primitives, run flags & standard flags. Options, arguments & types are plain click.

CLIG is treated as a contract: a CLI conforms with no application code wherever pyclig can do it alone, gets one obvious interface for each rule that needs application input, and pyclig.testing.assert_conforms reports the rest by rule ID. docs/clig-coverage.md has the rule-by-rule status and docs/roadmap.md what each release adds. Pin pyclig~=0.6.0 while it is pre-1.0: a minor release may break, and ~=0.6 would admit 0.7.

Usage

import click
import pyclig

from dkais import exceptions


@pyclig.group(
    "dkais",
    short_help="DMA AIS data manager",
    examples=["dkais list --start 2026-10"],
    errors=(exceptions.Error,),
    epilog="For more information see https://github.com/andyrids/dkais.",
)
def cli() -> None:
    """Download DMA AIS archives and convert them to Parquet."""


@cli.command("list", examples=["dkais list --start 2026-10 --json"])
@pyclig.options.json_output()
@pyclig.options.plain()
@click.option("--start", metavar="DATE", help="The range start.")
@pyclig.pass_context
def list_(ctx: pyclig.Context, start: str | None, json_output: bool) -> int:
    """List bucket archives & local state."""
    if ctx.format == "human":
        ctx.out.print(f"from {start}")
    else:
        click.echo(f"start\t{start}")
    return pyclig.ExitCode.OK


if __name__ == "__main__":
    cli()

Point the entry point at the group: [project.scripts] dkais = "dkais.__main__:cli". Always name the group - the name is the program name, also under python -m, and sets the shell completion variable (_DKAIS_COMPLETE). A single-command tool uses pyclig.command() instead, and gets everything a command of a group does.

Commands take examples= too: their --help leads with them under EXAMPLES:. An example may show its output on the lines after the invocation, which print indented under it:

@cli.command(examples=["dkais count --start 2026-10\n31 archives"])

docs="https://..." on a group or a command adds a closing LEARN MORE: section to its full help. The root's also lists Issues, the bug tracker in the distribution's Project-URL metadata.

What every command gets

ctx.obj is left to the application.

Run flags

Run flags are about how a run behaves, never what a command does, so they are on every command and act as soon as they are parsed - -d applies even when a later option is invalid:

<PROG>_DEBUG=1 and <PROG>_NO_COLOR=1 do the same from the environment, read before parsing, so even -d's own parse-time records appear; a later -d or --no-color is a no-op. <PROG> is the package name upper-cased with _ for every other character (my-tool is MY_TOOL). A value is on when set, non-empty and not 0, false, no or off; bare NO_COLOR keeps Rich's rule, on whenever it is set. Neither is listed in --help. Run flags also read config files (quiet = true, see Configuration).

A command's own option with the same spelling wins: copy -d /tmp reaches the app, and --debug still works there. The root's --help lists the run flags under GLOBAL OPTIONS:; a subcommand's names them on one line.

An application declares its own with cls=pyclig.GlobalOption on the command whose subtree accepts it, usually the root group. The callback is a plain click callback. It runs only when the option was given - on the command line, through envvar= or from a config file - never for its default=, and may run once per level of a nested command line that sees it (each level reads the environment), so it must be idempotent:

def _trace(ctx: click.Context, param: click.Parameter, value: bool) -> None:
    ctx.meta["trace"] = True


@pyclig.group("dkais")
@click.option(
    "--trace", cls=pyclig.GlobalOption, callback=_trace, help="Trace the run."
)
def cli() -> None: ...

With type= (or is_flag=False) a GlobalOption takes a value, which the callback shares with the commands below through ctx.meta. A top-level key of a config file configures it (region = "eu"), which is how a setting shared by several commands is configured:

def _region(ctx: click.Context, param: click.Parameter, value: str) -> None:
    ctx.meta["app.region"] = value


@pyclig.group("dkais")
@click.option(
    "--region",
    cls=pyclig.GlobalOption,
    type=str,
    callback=_region,
    help="The region.",
)
def cli() -> None: ...

Configuration

Config is opt-in: a CLI without config= gets no new options and reads no files.

@pyclig.group(
    "dkais", config=pyclig.Config(project=".dkais.toml", profiles=True)
)
def cli() -> None: ...

The files are TOML, read into click's default_map before any other option, so click's own order gives the precedence, highest first:

1. The command line
2. The environment, via envvar=
3. ./.env, with dotenv=True
4. [profiles.NAME] - project, then user, then system file - with --profile NAME
5. The project file, nearest from the working directory upward
6. The user file, paths.config_dir(name)/config.toml
7. The system file, paths.system_config_dir(name)/config.toml
8. The application's default_map, then default=

Top-level keys configure the root command and a [sub] table its subcommand, as default_map does; they do not cascade, so a setting shared by several commands is a value GlobalOption (above) or is repeated per table. Tables merge recursively; scalars and lists are replaced. Run flags read config like any option (quiet = true), unknown keys are ignored, and -d logs the files it loaded.

--config FILE (or <PROG>_CONFIG) replaces every discovered file. It is read whole unless it is named like the project file: --config ../pyproject.toml selects table as discovery would. prog sub --config FILE applies the file to sub and below. prog --help, -v and completion read no file, and a file that cannot be read is reported only when a command runs, so a broken file never hides the help.

.env

With dotenv=True the root reads ./.env once, into the options that declare envvar= and found nothing in the environment. It never writes os.environ, so child processes see no change. The rules are deliberately small: KEY=value and export KEY=value lines; # comments and blank lines are skipped; an unquoted value loses a trailing # comment after a space; single or double quotes are removed; there are no escapes and no $VAR expansion. A key ending in a secret word (_TOKEN, _PASSWORD, _API_KEY...) is skipped, see Secrets. Any other line warns .env:LINE: not KEY=VALUE, ignored and the rest is still read.

In tests, invoke reads no config file or .env unless config_files=True, and clears <PROG>_DEBUG, _NO_COLOR, _CONFIG and _PROFILE, so a developer's own settings never leak in; --config and env={"<PROG>_CONFIG": ...} still work.

Paths

pyclig.paths returns the per-user and system directories, XDG everywhere but Windows, and creates none of them:

A relative XDG value is ignored ($XDG_CONFIG_DIRS gives its first absolute entry), and macOS follows XDG rather than ~/Library. The environment, platform and home are read on every call, so tests can patch them.

Secrets

A secret belongs neither on the command line nor in an environment variable (the kit reports ARG-16 and ENV-15). pyclig.options.secret("api-key") adds --api-key-file FILE and hands the command the file's text as api_key, one trailing newline removed; - reads a piped STDIN, so a secret store is reached by a pipe or process substitution:

pass show prog/token | prog login --token-file -
prog login --token-file <(pass show prog/token)

Errors & exit codes

Commands return an exit code (or None for 0). Anything else is mapped by the runner, with messages on STDERR:

pyclig.ExitCode names the codes: OK, FAILURE, USAGE, INTERNAL, INTERRUPT & TERMINATE. Messages are never parsed as Rich markup: user input such as [/b] or :warning: prints as typed.

The bug report hint links the distribution's Issues project URL on a line of its own, never wrapped. On github.com & gitlab.com the link opens a new issue, pre-filled with:

• the program & its version, the pyclig version;
• the Python version & implementation, the operating system & its release;
• the exception's type.

It never contains the exception's message, the command line arguments, a path or the environment, which may hold secrets. The user pastes the output themselves.

hint & exit_code are read by attribute, so the application's exceptions need no pyclig import - a class attribute on the exception type is enough:

class ToolNotFoundError(Error):
    hint = "Run the hook in the environment pkgdx is installed in, e.g. via `uv run`"

A command may catch Ctrl-C or its own errors to print a tailored message first - a returned code passes through untouched. Wrap third-party failures (e.g. network errors) in an errors= type, so users see a message instead of a crash. errors= goes on the root; add_command rejects it on a subcommand, where the runner would ignore it.

Signals & clean-up

Each run installs its own SIGINT, SIGTERM & (on Windows) SIGBREAK handlers and restores the previous ones when it ends. The notice prints as soon as the signal arrives, not after the stack unwinds, and the run exits 130 (interrupted) or 143 (terminated). On Windows os.kill(pid, SIGTERM) is TerminateProcess: the process ends at once with exit 15 and no handler runs - send CTRL_BREAK_EVENT to a process group instead, which exits 130.

Wrap clean-up that must run after an interrupt in ctx.cleanup, so it is bounded:

try:
    work(pool)
finally:
    with ctx.cleanup(timeout=10):
        pool.shutdown()

The timeout counts from the interrupt; timeout=None allows any time, though a second signal still skips. A skip or timeout flushes the streams and exits even while the block is stuck in a blocking call. Off the main thread, or with no interrupt, the block simply runs.

A blocking call, such as a socket read on Windows, holds Ctrl-C back until it returns. Wrap blocking network I/O in the same block: the notice then prints at once, and the run exits when the call returns or the timeout passes, whichever is first:

with ctx.cleanup(timeout=2):
    data = connection.recv(4096)

Output

ctx.out (STDOUT) & ctx.err (STDERR) are Rich consoles created for each run, with markup, emoji codes & highlighting off. Each console tells the truth about its own stream: is_terminal is the real TTY check, so FORCE_COLOR means colour only - a piped console keeps its colour but stays one line per message, never wraps and never animates. NO_COLOR, <PROG>_NO_COLOR, TERM=dumb & --no-color drop colour; TTY_COMPATIBLE=1 makes a pipe a terminal.

ctx.format says what a command producing records should print: human on a terminal, otherwise plain; the per-command flags --plain (pyclig.options.plain()) & --json (pyclig.options.json_output()) override it, --json winning in any order. Only commands that produce records take the flags, so help never advertises a flag a command ignores. ctx.emit(records, table=..., caption=...) prints the records in that format:

A plain cell never breaks its row: None is empty, bools are true/false, tabs & newlines become one space. Without records JSON prints [] and the others only the caption. table is a factory, so cells can depend on other fields and rich.table stays off the fast paths.

with ctx.pager(): pages what the block prints on ctx.out, in human format only - plain & json are for pipes. It goes through click.echo_via_pager, so PAGER is honoured and less gets LESS=-FIRX when LESS is unset. On Windows paging is opt-in: with PAGER blank the output prints directly, because click's fallback there is more, which garbles UTF-8 and cannot scroll; PAGER=less (Git for Windows ships one) pages in colour. Output is held until the block ends, and click.echo inside it bypasses the page.

Messages & feedback take the console first, then keyword-only arguments:

-q drops info, row (unless essential=True), summary (unless failed=True) & animation; notice & hint stay. status & progress animate only on an interactive console. A grouped failure is one notice(items=...) then a hint with the remedy. Control characters, including terminal escapes, are dropped from messages, and a str is never parsed as markup; a rich.text.Text with a style prints styled, as given. That holds for the description column & str columns of progress too, so prefer them to a TextColumn of your own, which renders as you built it.

Standard flags

pyclig.options holds the flags with standard names, each taking a help= override; the value flags also take default= and envvar=:

options.NoneOr is a type for an option whose value can be switched off: --limit none hands the command None, any other value goes to the inner type, and help shows --limit INTEGER|none. The word is matched in any case, from the command line, an envvar= or a config file:

@click.option("--limit", type=pyclig.options.NoneOr(int), default=50)

Prompts & confirmation

A prompt needs a terminal on STDIN to answer it and on STDERR to see it, and --no-input refuses every prompt; ctx.can_prompt is that test. pyclig.confirm() asks before a large or destructive action, and --force skips the question:

@cli.command()
@pyclig.options.force()
@pyclig.pass_context
def purge(ctx: pyclig.Context, force: bool) -> None:
    """Delete every local archive."""
    pyclig.confirm(ctx, message="Delete every local archive?", force=force)

A script cannot answer the prompt, so the run fails rather than read an answer from a pipe, and with 2>log it fails rather than hide the question in the log. Call confirm before starting a Progress or Status. A click option with prompt= follows the same rule: it prompts on STDERR, and when the run cannot prompt it takes its default, or fails naming the option when it has none - so --no-input never lies.

A severe action - one that cannot be undone, such as dropping a database - asks for the resource's name instead. pyclig.confirm_name() pairs with options.confirm(), and --force never skips it:

@cli.command()
@pyclig.options.confirm()
@pyclig.pass_context
def drop(ctx: pyclig.Context, confirmed: str | None) -> None:
    """Drop the production database."""
    pyclig.confirm_name(
        ctx, message="Drop prod?", name="prod", confirmed=confirmed
    )

Logging

Log through the package logger - logging.getLogger(__name__) in any module of the package that defines the group. Libraries add only a NullHandler, in the package __init__.py:

logging.getLogger(__name__).addHandler(logging.NullHandler())

During a run, warnings & errors print as warning: / error: notices on ctx.err. With -d, DEBUG records, times & call sites are added - laid out by a RichHandler when STDERR is a terminal. Otherwise each record is one line that never wraps, with any traceback after it in plain text:

[2026-10-31T09:14:03+0000] DEBUG    [convert] converting 'nope'  types.py:47

The logger's handlers, level & propagation are restored when the run ends.

When the root logger has no handlers as the run starts, it gets one too: third-party WARNING records and warnings.warn print as warning: notices - the message only, with file:line added under -d. The root's level is never changed, and an application that configures the root itself (e.g. logging.basicConfig) is left alone. Both are restored when the run ends.

<PROG>_LOG_FILE=FILE appends every record of the run, at DEBUG and in the plain -d format, to FILE (UTF-8), without changing what STDERR shows. A crash (exit 70) writes its traceback there too and adds the hint debug log: FILE. If the file cannot be opened, warning: cannot write the debug log to FILE: reason prints and the run goes on.

ctx.logs_on_failure() holds the info lines of a block and prints them only if it fails:

with ctx.logs_on_failure():
    for step in steps:
        logger.info("ran %s", step)
        run(step)

Warnings & errors still print at once. Raise to fail: when an exception leaves the block, the held lines print before it propagates, so error: is still last; on success, Abort, Exit or Ctrl-C they are dropped. -q hides the replayed lines, -d prints every record live, and the log file gets every record either way. Keep the block outside progress & status, so the replay is not drawn over.

Testing

from pyclig.testing import assert_conforms, invoke


def test_list_rows_when_piped() -> None:
    """Assert `list` prints tab-separated rows when STDOUT is a pipe."""
    result = invoke(cli, ["list"], out_tty=False, err_tty=True)
    assert result.exit_code == 0
    assert "\t" in result.stdout


def test_cli_conforms() -> None:
    """Assert the CLI follows every guideline pyclig can check."""
    assert_conforms(cli, ignore={"ARG-16": "`--token` also reads a file"})

invoke runs any pyclig.Command or Group and returns click's Result, with stdout & stderr apart. out_tty, err_tty, color & width set up each console; env sets (or, with None, unsets) variables. Colour & terminal variables from the developer's shell are cleared. Never patch Rich private attributes. in_tty=True & err_tty=True let pyclig.confirm & prompt= options prompt and read their answer from input; click echoes that answer to STDERR after the prompt (... [y/N]: y).

assert_conforms walks the CLI and reports deviations by the rule IDs of docs/clig-coverage.md. It never runs a command body unless given samples=. check returns the Findings instead of asserting; ignore needs a reason per rule.

samples= are command lines the kit runs for real, so name ones that are safe in a test:

assert_conforms(cli, samples=[["list", "--json"], ["show", "--value", "x"]])

A sample that exits non-zero or prints invalid --json is BAS-4, an error: or warning: line on STDOUT is BAS-5, and a traceback is ERR-2.

Parser-less entry points

pyclig.run() gives a pre-commit hook shim, which leaves sys.argv to the tool it wraps, the same exit codes, error rendering & logging. It enters a pyclig.Context, so the function may be a pyclig.pass_context function using ctx.err:

def ruff_format() -> NoReturn:
    pyclig.run(
        lambda: core_ruff_format(sys.argv[1:]),
        errors=(exceptions.Error,),
    )

Migrating from argparse

• Each subcommand becomes a @cli.command(); shared options become one decorator stacking click.options. Commands are listed in the order they are added.
• click shows a command's docstring as its help: put \f after the help paragraph, so the Args: & Returns: sections stay out of --help. Pass help= for a description that is not an imperative docstring line.
• Custom argument types become click.ParamType[T] subclasses calling self.fail().
• parser.error() becomes raise click.UsageError(...) (or ctx.fail()).
• A type= function checking a minimum becomes click.IntRange(min=1).
• An environment variable default becomes envvar=, with show_envvar=True so a missing value names it.
• Module-level consoles become ctx.out / ctx.err, passed to helpers as arguments.
• argparse accepted unambiguous option prefixes (--st for --start); click does not.