jpml-py

Python package for JPML

A configuration language that borrows TOML's sections and JSON's nesting.

.jp files use [SECTION] headers at the top level and {...} / [...] structures inside them. Keys need no quotes, # starts a comment, trailing commas are fine, and a value is allowed to be empty.

[SERVER_ID]
config: {
  disabled_channels:,
  disabled_users: [9892, 82082, 8209]
}

[SERVER_ID_2]
prefix: "!"
modules: {
  moderation: true,
  fun: false
}

>>> import jpml
>>> jpml.load("data/servers.jp")
{'SERVER_ID': {'config': {'disabled_channels': None,
                          'disabled_users': [9892, 82082, 8209]}},
 'SERVER_ID_2': {'prefix': '!', 'modules': {'moderation': True, 'fun': False}}}

---

Why another format

JSON has no comments, demands quotes on every key, and rejects a trailing comma. TOML has comments and headers, but nesting anything non-trivial means either deeply dotted keys or a table per level.

.jp takes the half of each that suits configuration files people edit by hand:

• Sections for the top level. [SERVER_ID] reads better than another brace.
• JSON for everything below it. Nest objects and arrays as deep as you like.
• No ceremony. Unquoted keys, comments anywhere, trailing commas ignored.
• Empty values are legal. disabled_channels:, means the key exists and has no value yet — a real state in configs that JSON can only spell as null.

It is a small, fully specified format with a strict parser, precise error messages, and a deterministic writer, so files stay stable when a program rewrites them.

pip install jpml     # or: uv add jpml

No runtime dependencies. Python 3.14+.

---

The format

Sections

A [NAME] header opens a root key. Everything below it, until the next header, belongs to that section.

[SERVER_ID]
prefix: "!"

Headers may be dotted to nest, and quoted when a name contains a dot:

[guild.limits]        # -> {"guild": {"limits": {...}}}
["weird.name"]        # -> {"weird.name": {...}}

Key/value pairs written before the first header land at the document root:

version: 2

[SERVER_ID]
prefix: "!"

Entries

An entry is key: value. Keys need no quotes; a bare key may contain spaces but not brackets, commas or quotes — quote it if it needs those.

Entries are separated by a line break, a comma, or both. Trailing and repeated commas are accepted:

[SERVER_ID]
a: 1
b: {x: 1, y: 2,}
c: [1, 2, 3,]

Empty values

A key with nothing after the colon parses to None:

config: {
  disabled_channels:,      # -> None
  timeout:                 # -> None
}

Because of this, a value must start on the same line as its :. An opening { or [ goes on the colon's line; its contents may then wrap freely.

Values

Type	Examples
String	"hello", 'hello', hello world (unquoted)
Integer	42, -7, 1_000, 0xff, 0o755, 0b1010
Float	3.5, 1e3, inf, -inf, nan
Boolean	true, false (case-insensitive, so True works too)
Null	null, none, nil, or nothing at all
Object	{a: 1, b: 2}
Array	[1, 2, 3]

Unquoted values are read as a keyword first, then a number, then a plain string. Quote a value if it contains a #, a comma, a bracket, or leading/trailing whitespace you want to keep.

Strings honour the usual escapes — \n, \t, \\, \", \uXXXX, \U0001F600, plus \ at end of line to continue onto the next.

Comments

# runs to the end of the line and is allowed anywhere, including inside objects and arrays.

---

What you can do with it

Read and write files

import jpml

data = jpml.load("data/servers.jp")        # -> dict
jpml.dump(data, "data/servers.jp")         # formatted, atomic write

text = jpml.dumps(data)                    # -> str
data = jpml.loads(text)                    # -> dict

Writes are atomic by default: the file goes to a temporary neighbour and is renamed into place, so a crash or a concurrent reader never sees half a config.

Options worth knowing:

jpml.load("servers.jp", duplicate_keys="last")   # "error" (default), "first", "last"
jpml.dumps(data, indent=4, sort_keys=True)       # also: width, ensure_ascii
jpml.dumps(data, default=str)                    # convert datetimes and friends

Edit a config in place

JPConfig is a MutableMapping that remembers the file it came from.

from jpml import JPConfig

cfg = JPConfig.load("data/servers.jp", missing_ok=True)

cfg["SERVER_ID"]["prefix"]                              # plain dict access
cfg.get_path("SERVER_ID.config.disabled_users", [])     # never raises
cfg.set_path("SERVER_ID.config.disabled_users", [9892]) # creates missing sections
cfg.has_path("SERVER_ID.prefix")
cfg.section("NEW_SERVER", create=True)["prefix"] = "?"
cfg.merge({"SERVER_ID": {"modules": {"fun": True}}})    # deep merge
cfg.save()                                              # atomic, back to its own path
cfg.reload()                                            # discard in-memory changes
cfg.to_dict()                                           # deep copy as a plain dict

missing_ok=True gives an empty config bound to the path, which is what you want for a program that writes its config on first run. Formatting options given to the constructor are remembered by save():

cfg = JPConfig.load("data/servers.jp", indent=4, sort_keys=True)

Load a whole folder

config = jpml.load_dir("data")                 # {'servers': {...}, 'roles': {...}}
guilds = jpml.load_dir("data/guilds")          # {'1234567890': {...}, ...}
everything = jpml.load_dir("data", recursive=True)

Each file becomes one key, named after the file.

Find mistakes quickly

Every error derives from jpml.JPError. JPDecodeError (a ValueError) points at the exact character:

data/servers.jp:2:8: expected ':' after key 'prefix', found '"'
    prefix "!"
           ^

It carries .line, .col, .pos, .filename and .raw_message if you want to render the failure yourself. JPEncodeError (a TypeError) explains what could not be serialised — an unsupported type, a non-string key, a circular reference.

By default a repeated key is an error rather than a silent overwrite; pass duplicate_keys="first" or "last" if you would rather it not be.

Work from the shell

jpml check data/*.jp                      # validate; non-zero exit on failure
jpml fmt -w data/servers.jp               # reformat in place
jpml get data/servers.jp SERVER_ID.prefix # read one value
jpml to-json data/servers.jp -o out.json
jpml from-json out.json -o data/servers.jp

python -m jpml ... works identically, and - reads stdin.

---

Round trips

dumps is deterministic, so a file rewritten twice is byte-identical:

• every top-level mapping becomes a [SECTION], separated by a blank line;
• section entries sit one per line, with no separating commas;
• nested objects always expand across lines, {} being the only inline form;
• arrays stay inline while they fit inside width (default 88), then break one element per line;
• None is written as an empty value inside mappings (key:) and as null inside arrays, since an array element cannot be empty;
• insertion order is preserved unless sort_keys=True.

Two things do not survive a rewrite:

• Comments are dropped. Rewriting a hand-annotated file loses its notes.
• Root-level scalars move above the first section, because anything after a header would be read back as part of that section.

---

Organising your configs

Nothing is enforced, but this layout is what load_dir is built for:

your-project/
├─ data/
│  ├─ servers.jp            # one file per concern
│  ├─ roles.jp
│  ├─ servers.example.jp    # committed template, safe to publish
│  └─ guilds/               # optional: one file per entity
│     ├─ 1234567890.jp
│     └─ 9876543210.jp
└─ src/

A few habits that save pain later:

1. One file per concern. A parse error then takes out one feature, not everything.
2. Keep live data out of git, and commit a template instead:

   data/*.jp
   !data/*.example.jp

3. Use IDs as section names. [1234567890] parses to the string key "1234567890", and integer keys are stringified on write, so {1234567890: {...}} round-trips.
4. Write through JPConfig.save() rather than by hand, so an interrupted write cannot truncate a live config.
5. Validate in CI with jpml check data/*.jp.

---

Licence

MIT. Contributing, tests and release process: CONTRIBUTING.md.