A Python typesetter for the reader

---

🪻 About

Prose formats Python source to be legible at a glance. It aligns equals signs and colons vertically across consecutive lines, places one entry per line in dictionaries and lists, alphabetizes methods and fields within their groups, applies a singleton rule for colon padding, and treats code like prose rather than minified text.

Note

Prose is still pre-1.0. 0.2.0 lifts the catalog to eighteen rules (fourteen auto-fix, four lint-only) and lands the surrounding surface (structured output formats, suppression directives, rule subsetting, an exit-code matrix), with additional rules and configuration knobs planned for later releases.

---

🗞️ Philosophy

Code is read far more often than it is written. A reader's eye moves down a page and across adjacent lines looking for parallels, patterns, and shape. When every = sits at a different column and every collection is compressed onto one line, that shape disappears, forcing the eye to slow down. Prose restores it, with aligned columns letting the eye skim, one-per-line collections making each entry a unit, and alphabetized groupings giving every reader the same landmarks.

The trade-offs minimalist formatters were built to avoid (wider diffs, more vertical scrolling, occasional re-alignment churn) no longer dominate the equation. Agentic assistants do much of the typing, and every modern code host offers whitespace-ignoring diffs. What remains is the daily experience of reading code.

---

🪄 Install & Usage

uv tool install prose-formatter

The binary exposes format, check, and completions:

prose format path/                          # rewrite files in place
prose check path/                           # exit non-zero on violations
prose format --diff path/                   # show the diff without writing
prose check --stdin < file.py               # read from stdin

Subset the active rules with --select and --ignore:

prose check --select align-equals path/         # run a single rule
prose check --ignore strip-trailing-commas path/  # subtract one
prose check --select align-equals,align-colons path/  # comma list

Pick a structured output format for editors, pre-commit, and CI:

prose check --output-format json path/      # newline-delimited JSON
prose check --output-format github path/    # GitHub Actions annotations
prose check --output-format sarif path/     # SARIF for Code Scanning
prose check --color always path/            # force color, NO_COLOR honored
prose completions zsh                       # shell completion script

Source ranges opt out of rule application through the # prose: directive namespace:

Directive	Effect
# prose: off / # prose: on	Suppress every rule across the span the pair opens
# prose: skip	Suppress every format rule on the directive's line
# prose: skip[<rule>]	Suppress one named format rule on the directive's line
# prose: ignore	Suppress every lint rule on the directive's line
# prose: ignore[<rule>]	Suppress one named lint rule on the directive's line

# prose: off
keep_this_block_exactly_as_written = (1,2,3)
# prose: on

inside(1, 2, 3,)  # prose: skip[strip-trailing-commas]
value: Union[int, str] = compute()  # prose: ignore[legacy-union-syntax]

A # prose: off at the top of a file (with no matching # prose: on) opts the whole file out of every rule. The # fmt: directives (# fmt: off, # fmt: on, # fmt: skip) and the # yapf: directives (# yapf: disable, # yapf: enable) are honored as Black-compatibility aliases that map onto their # prose: counterparts.

---

🦉 Exit Codes

The binary resolves every run into one of five exit codes, which CI gates and pre-commit hooks compile against:

Code	Meaning
0	Clean: no diagnostics, no rewrites pending
1	Format would change: at least one auto-fix diagnostic
2	Lint violation: at least one lint-only diagnostic
3	Parse error: input could not be parsed as Python
4	Config error: pyproject.toml, --select / --ignore, or argument validation

When two outcomes apply to the same run, the higher number wins. prose --help prints the same table beneath the option list. In format mode, code 1 is suppressed when the rewrite succeeds because the changes were applied rather than left pending. Codes 2, 3, and 4 apply identically across both subcommands.

---

🪶 Rules

Auto-fix rules:

Rule	Coverage
align-colons	Collection literals, Pydantic / dataclass fields, function signatures, and docstring Args: sections
align-equals	Consecutive assignments at the same indentation
align-imports	The import keyword in from ... import ... groups and as in import ... as ... groups
alphabetize	Classes, methods (grouped dunders → properties → privates → publics), enum members, Pydantic fields (required then optional), function parameters, keyword arguments, and from imports
bare-import-allowlist	Bare import X outside a configurable allowlist (default numpy, pandas)
blank-lines	Module-level def and class carry 2 blank lines before them, methods inside a class body carry 1, a module-level statement after if __name__ == "__main__": carries 1, and adjacent module-level bare import and from imports carry 1 between them
collection-layout	Expands dict, list, and set literals to one entry per line, even when they fit inline
docstring-wrap	Wraps docstring description prose to docstring-line-length and structured Args: / Returns: / Raises: sections to the budget chosen by docstring-structured-policy
match-case-align	Single-expression case bodies
multi-line-docstrings	Multi-line docstring placement, with opener and closer each on their own line
no-single-line-docstrings	Single-line triple-quoted docstrings, expanded into the canonical multi-line shape
singleton-rule	Skips colon padding when only one item exists in the aligned group
strip-trailing-commas	Multi-line collections and signatures
unused-future-annotations	Removes from __future__ import annotations when removal is provably safe for the file

Lint-only rules:

Rule	Coverage
legacy-union-syntax	Union[X, Y] and Optional[X] when target-version is 3.10 or higher, with the PEP 604 X | Y / X | None shapes as the recommendation
loose-constants	Module-level SCREAMING_CASE assignments that would read better as an Enum, a model field, or a function-local
no-step-narration	Own-line numbered-step comments (# 1. ..., # Step 2: ...) that signal extractable functions
single-use-variables	Bindings assigned and read exactly once, where inlining the right-hand side carries the same meaning

Example

Before:

from pydantic import BaseModel
from datetime import date
from collections.abc import Iterable
from decimal import Decimal

DEFAULT_LIMIT = 50
RETRY_INTERVAL = 30
PAGE_SIZE = 25

class Posting(BaseModel, extra="forbid"):
    title: str
    company: str
    skills: Iterable[str] | None = None
    location: str | None = None
    date_posted: date
    salary_max: Decimal | None = None

    def render(self, separator: str, include_location: bool, include_date: bool, max_width: int = 80) -> str: ...

config = {"threshold": 0.7, "metric": "euclidean", "linkage": "ward", "n_clusters": None, "random_state": 42,}

After:

from collections.abc import Iterable
from datetime        import date
from decimal         import Decimal
from pydantic        import BaseModel

DEFAULT_LIMIT  = 50
RETRY_INTERVAL = 30
PAGE_SIZE      = 25

class Posting(BaseModel, extra="forbid"):
    company     : str
    date_posted : date
    title       : str
    location    : str | None           = None
    salary_max  : Decimal | None       = None
    skills      : Iterable[str] | None = None

    def render(
        self,
        include_date     : bool,
        include_location : bool,
        separator        : str,
        max_width        : int = 80
    ) -> str: ...

config = {
    "linkage"      : "ward",
    "metric"       : "euclidean",
    "n_clusters"   : None,
    "random_state" : 42,
    "threshold"    : 0.7
}

---

⚖️ Configuration

Prose loads the nearest [tool.prose] section found by walking upward from the working directory. With no configuration, every rule runs at its default. To tune a rule, write its sub-table:

[tool.prose]
code-line-length = 88

[tool.prose.rules.align-equals]
enabled = false                # disable a rule outright

[tool.prose.rules.align-colons]
max-shift        = 12          # cap padding width
max-shift-policy = "drop"      # how to handle overrun groups

[tool.prose.rules.collection-layout]
max-atomics-per-line = 3       # keep short tuples on one line

Per-rule knobs:

Key	Type	Where	Meaning
enabled	bool	every rule sub-table	Toggle the rule on or off, defaulting to true
max-shift	positive int	alignment rules	Ceiling on per-line padding, defaulting to 8
max-shift-policy	"split" | "drop" | "skip"	alignment rules	How to handle a group whose widest member exceeds max-shift. split partitions the group, drop excludes the widest members from the padding calculation, skip leaves the whole group unaligned, defaulting to "split"
max-atomics-per-line	positive int	collection-layout	Keep short collections on one line when each entry is an atomic literal and the run fits the cap, defaulting to 8
allow	list of module names	bare-import-allowlist	Modules whose bare-import form is preserved, defaulting to ["numpy", "pandas"]
allow	list of names	loose-constants	Module-level names exempted from the lint, defaulting to []
allow-pattern	regex	single-use-variables	Binding names exempted from the lint, defaulting to "^_"
code-line-length	positive int	top-level [tool.prose]	Honored by line-length-aware rules, defaulting to 88
docstring-line-length	positive int	top-level [tool.prose]	Description-prose budget for docstring-wrap, defaulting to 76
docstring-structured-policy	"code-line-length" | "docstring-line-length"	top-level [tool.prose]	Source budget for structured docstring sections, defaulting to "code-line-length"
target-version	"3.X" version string	top-level [tool.prose]	Python runtime the project ships to. Consumed by version-gated rules, defaulting to unset

Docstrings carry two readings inside one triple-quoted region. Description prose between the opening """ and the first section heading reads as paragraphs, where 76 characters is the comfortable line for sustained reading. Structured sections (Args:, Returns:, Raises:) read as code-shaped tables and reuse code-line-length (88 by default) to match surrounding indentation, though docstring-structured-policy switches them to docstring-line-length if a project prefers a single narrower budget across the whole docstring. The docstring-wrap rule consumes both budgets.

target-version names the Python runtime a project ships to, taking the bare major.minor form ("3.13", "3.14") used by mypy's python_version setting. Rules whose safety depends on the runtime read this field directly, treating an unset value as the cue to skip every version-dependent arm rather than assume a default. legacy-union-syntax is the first such consumer, staying quiet on every project that has not opted into a target.

Alignment rules are align-colons, align-equals, align-imports, and match-case-align. Toggle-only rules carry only the enabled knob. They are alphabetize, blank-lines, docstring-wrap, legacy-union-syntax, multi-line-docstrings, no-single-line-docstrings, no-step-narration, singleton-rule, strip-trailing-commas, and unused-future-annotations. The remaining rules (bare-import-allowlist, collection-layout, loose-constants, single-use-variables) carry rule-specific knobs documented in the table above.

Per-invocation overrides via --select and --ignore (see Install & Usage above) take precedence over the configured-enabled set.

---

🗺️ Composition

Prose runs as the second pass in a two-stage pipeline. The first pass owns tokens (line wrapping, quote normalization, indentation, blank-line discipline) and the second pass owns layout (alignment, alphabetization, the singleton rule, one-entry-per-line collections, trailing-comma stripping). Ruff is the canonical first pass, in that ruff format matches the token-level scope and its lint config shares the pyproject.toml root with [tool.prose].

ruff format && prose format

Important

Running Prose first is incorrect, because Prose's alignment math depends on already-settled line breaks and an upstream re-wrap will undo per-line layout decisions, forcing a third pass.

Ruff Configuration

Code	Conflict	Reason
COM812	Lint re-adds trailing commas	strip-trailing-commas removes them in multi-line collections and signatures
E203	Lint flags whitespace before :	align-colons produces it in dict literals, dataclass fields, function signatures, and docstring Args: blocks
E221	Lint flags multiple spaces before =	align-equals produces it across consecutive assignments at the same indentation
E272	Lint flags multiple spaces before import / as	align-imports produces it across from ... import ... and import ... as ... groups
E501	Lint flags lines past line-length	A long member in an alignment group pads shorter lines rightward, occasionally past the configured limit
skip-magic-trailing-comma	Formatter re-expands collections by trailing-comma presence	prose format controls collection layout independently of comma signaling

Other Tools

Black formats, Flake8 lints, and isort sorts, so each pairs with Prose at a different layer:

Tool	Pairing
Black	Run Black with --skip-magic-trailing-comma, then Prose second. Black collapses collections that collection-layout re-expands and preserves trailing commas that strip-trailing-commas removes
Flake8	Add extend-ignore = E203, E221, E272 to .flake8 or setup.cfg (and C812 if flake8-commas is installed). Flake8 inherits the same pycodestyle codes Ruff inherits
isort	Run isort first, Prose second, with no configuration adjustment. Prose alphabetizes within isort's groups and aligns the import keyword that isort leaves un-aligned

---

🪡 Integrations

Wire Prose into anything that runs on save, on commit, or in CI.

GitHub Actions

The minimal check-on-CI shape:

- run: uv tool install prose-formatter
- run: prose check .

For inline annotations on the PR diff, use the github output format. Prose emits workflow commands that GitHub renders as native check-run annotations:

- run: uv tool install prose-formatter
- run: prose check --output-format github .

For findings that persist across runs and surface in Code Scanning, emit SARIF and upload it:

- run: uv tool install prose-formatter
- run: prose check --output-format sarif . > prose.sarif
- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: prose.sarif

CI gates compile against the exit-code matrix above. A non-zero status without continue-on-error fails the step.

Editor

Any editor that supports run-on-save (VSCode's runOnSave, Vim's autocmd BufWritePost, JetBrains File Watchers) can shell out to prose format <file>. For editors that consume structured diagnostics, prose check --output-format json --stdin emits one Ruff-shaped record per line.

Pre-Commit

Add a local hook to your .pre-commit-config.yaml:

- repo: local
  hooks:
    - id: prose
      name: prose
      entry: prose format
      language: system
      types: [python]

Swap entry: prose format for entry: prose check for the check-only variant. The hook surfaces the same exit codes the CLI uses, so a format hook never fails on rewrites it applies and a check hook fails the commit when changes are pending.

Shell Completions

prose completions zsh > "${fpath[1]}/_prose"
prose completions bash > /etc/bash_completion.d/prose
prose completions fish > ~/.config/fish/completions/prose.fish

Both elvish and powershell are supported targets for the completions subcommand.

---

🗜️ Development

One-Time Setup

Prose uses mise to manage every toolchain and CLI through a single mise.toml. Install mise, wire it into your shell, then mise install provisions the rest.

Install Mise

curl https://mise.run | sh

The installer drops the binary into ~/.local/bin/mise.

Wire Mise into Zsh

Three init files cover the three load contexts (every shell, login shells, interactive shells) so mise-managed tools resolve correctly whether you are inside an interactive terminal, a login shell, or a non-interactive subprocess:

# ~/.zshenv  (sourced for every shell, including non-interactive)
export PATH="$HOME/.local/bin:$PATH"

# ~/.zprofile  (sourced for login shells, before .zshrc)
eval "$(mise activate zsh --shims)"

# ~/.zshrc  (sourced for interactive shells)
eval "$(mise activate zsh)"

.zshenv puts mise itself on PATH so the later eval lines can find it. .zprofile's --shims activation makes mise-managed binaries resolvable in non-interactive contexts (scripts, editors, GUI launches). .zshrc's full activation gives interactive shells the per-directory tool resolution and task discovery.

Clone and Provision

git clone https://github.com/Jybbs/prose.git
cd prose
mise install

mise install provisions the base toolchain:

Tool	Purpose
maturin	Rust → Python wheel builder
python (3.14)	Python interpreter for wheel builds
rust (stable)	Rust toolchain via rustup
uv	Python package and venv manager

Additional task-scoped tools (cargo-insta, cargo-machete, cargo-llvm-cov, codecov-cli) install on demand when their owning task first runs, so no second provisioning step is required.

Tasks

Tasks live in mise.toml. GitHub Actions drives most of them across the CI workflows that gate every push and pull request, so what passes locally with mise ci mirrors what runs upstream. The same tasks are available locally for reproducing a failed CI step, accepting snapshot updates after a fixture change, or running an individual check during iteration. List them with mise tasks.

Command	What it does
mise audit	Detect unused dependencies via cargo machete
mise build	Compile in debug mode
mise check	Verify Rust source matches rustfmt without rewriting
mise ci	Full local sweep: audit, check, lint, test, wheel
mise coverage	Generate an LCOV coverage report at target/lcov.info
mise format	Format Rust source with rustfmt
mise lint	Run clippy with all warnings as errors
mise readme	Rewrite README.md in place to absolute URLs for PyPI
mise review	Interactively review pending snapshot diffs
mise test	Run all tests including insta snapshots
mise upload	Generate the coverage report and upload it to Codecov
mise wheel	Build the maturin wheel into the active virtualenv

A common iteration flow on a rule or fixture is to edit, run mise test, accept any snapshot diffs with mise review, then run mise ci before pushing.