pyjson5
A Python implementation of the JSON5 data format.
JSON5 extends the JSON data interchange format to make it slightly more usable as a configuration language:
- JavaScript-style comments (both single and multi-line) are legal.
- Object keys may be unquoted if they are legal ECMAScript identifiers
- Objects and arrays may end with trailing commas.
- Strings can be single-quoted, and multi-line string literals are allowed.
This project implements a reader and writer implementation for Python; where possible, it mirrors the standard Python JSON API package for ease of use.
There is one notable difference from the JSON api: the load() and
loads() methods support optionally checking for (and rejecting) duplicate
object keys; pass allow_duplicate_keys=False to do so (duplicates are
allowed by default).
This is an early release. It has been reasonably well-tested, but it is SLOW. It can be 1000-6000x slower than the C-optimized JSON module, and is 200x slower (or more) than the pure Python JSON module.
Please Note: This library only handles JSON5 documents, it does not allow you to read arbitrary JavaScript. For example, bare integers can be legal object keys in JavaScript, but they aren't in JSON5.
One more note: It's long past time to declare this a 1.0 stable
product and move on, so I am preparing to do so. If the latest version
proves to be stable (e.g., no bug reports for a month or so), I plan to
switch the directory layout to move json5 under a src directory
(i.e., use a src-layout to avoid accidentally importing an unpackaged
version of the code), bake that for a month or so as well, and then
release this as 1.0.
Known issues
- Did I mention that it is SLOW?
load()reads the entire remaining stream into memory, including when
consume_trailing=False. That option ignores trailing content during
parsing; it does not preserve unread stream data for a later load().
To read successive values, read the input into a string once and use
parse(..., consume_trailing=False), passing its returned position as
the next start. Positions refer to decoded characters, not byte offsets
or file seek positions. The parse() docstring includes an example.
- The implementation is based on recursive descent parsing and relies on the
RecursionError, but this doesn't seem to be a problem in
normal practice. If it is, please report it and I can raise the priority
of fixing this properly :).
- The implementation follows Python3's
jsonimplementation where
encoding method to dump() is
ignored, and unicode strings are always returned. There is a slight
difference to the strict argument to load()/loads() in order
to be JSON5-compliant.
- The
clskeyword argument thatjson.load()/json.loads()accepts
JSONDecoder is not and will not be
supported, because this implementation uses a completely different
approach to parsing strings and doesn't have anything like the
JSONDecoder class.
- The
cls keyword argument that json.dump()/json.dumps() accepts is also not supported, for consistency with json5.load(). The default
keyword is supported, though, and might be able to serve as a
workaround.
Contributing:
json5 has no runtime dependencies and it is supported on Python version
3.8 or later.
However, in order to develop, test, build and publish the package you need a
bunch of extra tools that are only supported on 3.11 or newer; on older
versions, you can run the test, but not do anything else.
When changing the code, do not add anything that only runs on 3.9 or later
versions, and do not rely on
from __future__ import annotations.
Use a 79-character line and prefer single quotes to double; otherwise follow
PEP-8 and the normal
ruff format coding style.
On Mac
uv and
use uv and the //run script to develop things. See ./run --help for
the various commands that are supported.
$ brew install uv
$ git clone https://github.com/dpranke/pyjson5
$ cd pyjson5
$ source $(./run devenv) # To activate a venv w/ all the needed dev tools.
On other platforms
Install
uv via whatever mechanism is appropriate.
Creating a venv
You do not have to create a venv to do anything, if you use the
run
script, but if you want to invoke tools like ruff directly, you
can do so as follows:
$ ./run devenv # which really just calls uv sync.
$ source .venv/bin/activate
Run tests and other checks
$ ./run presubmit
//run supports a number of other commands:
regen will regenerate the parser when the grammar is updated; in order to run this you need glop to be available as described above.
format will format all the code using ruff format for the Python code and pyproject-fmt for the pyproject.toml file; format --check will
report whether the code is formatted properly without actually changing
anything. It checks for the coding style described above.
checks checks the code for problems without actually executing the code. It runs:
* ruff check
* pylint
* mypy
tests runs all of the tests. Write tests using either doctest or unittest. Aim for 100% line coverage in tests (see next bullet).
coverage runs all of the tests using the coverage tool in order to collect code coverage. You can pass -b/--branch to collect branch
coverage. You should target 100% line coverage on any changes, as there
shouldn't be a reason you can't do this, and actually ensuring that you
execute every line can catch problems in Python that might otherwise be
caught earlier in statically-typed languages (though the checks go
a long way to accomplishing the same thing).
build builds the package to get it ready to publish
publish tries to publish the package. You must specify --dev or --prod to indicate which PyPI repository you want to publish to; explicitly having
to mention one or the other helps ensure that you don't try to publish
something by accident. You can run publish --check in order to ensure
that the package is built correctly without actually trying to publish it.
You can only actually publish the package if you have permission to do so,
of course.
presubmit just runs all of the above commands at once (though not regen
at the moment, because you might not have glop installed; this'll change
when glop is properly packaged and installed in the venv.
Other
run commands:
devenv creates a virtual environ using uv and echoes the command neededto activate it.
clean will remove any unversioned files from the repo.
help prints usage.
Releasing a new version:
- Update the version in
json5/version.py to $VERSION, which should be of the form X.Y.Z, where X, Y, and Z are numbers. Use
semantic versioning when increasing the numbers:
* X increases only when making backwards-incompatible or other very
significant changes. Try very hard to avoid this.
* Y increases when adding functionality in a backwards-compatible way.
* Z increases when you are fixing bugs in a backwards-compatible way.
In between releases, the version may be set to
X.Y.Z.dev0' to indicate
that this is not an officially released version.
./run presubmitonce more and make sure everything passes.
git commit -a -m "Bump the version to $VERSION"
git tag "v$VERSION"
git push origin
git push --tags origin
./run publish --prod(which will build the package if need be first)
Version History / Release Notes:
- Unreleased
frozendict; add
use_duck_typing=True to retain structural mapping and sequence
detection.
- v0.17.3 (2026-10-09)
urllib3 to 2.8.0 to pick up a
security fix.
- v0.17.2 (2026-10-09)
- v0.17.1 (2026-10-09)
- v0.17.0 (2026-10-09)
glop, rather than relying on it
being checked out alongside this repo. This should make the
package properly self-contained, and now we can enforce that
the grammar is up-to-date in presubmit.
* Rework the project dependencies. As before, the module still has
no runtime dependencies. However, we will now only support doing
most development on Python 3.11 or newer; older versions will only
support running the tests. This means the dev venv config will
only install the dev packages in that newer environment; this
should make Dependabot happier.
* Rework how the //run script works;
the previous pylint and mypy checks have been moved into
the checks command, which will now run all of the static
analyzers, not just ruff check. The script also now consistently
supports -f/--failfast across the commands, not just during tests.
* There are no code changes in this release other than bumping the
version numbers.
- v0.16.0 (2026-10-07)
RecursionError issue.
* Fix an issue w/ spec compliance where \u2028 and \u2029 were not
being allowed in string literals.
* "Fix" an issue w/ spec compliance in how control characters
(\x00 - 0x1f) were being handled in string literals; in JSON,
they are not allowed, and the Python library has the strict=False
option to allow them. In JSON5, they are allowed (except for \r
and \n). json5 also has the strict option, and was initially
supposed to match json here (I think), but due to a bug only \r and
\n were being disallowed. However, this behavior is actually correct
(they should always be allowed), and now strict=False will just
control whether \r and \n will be allowed. This isn't a behavior
change, but it might change what you expected the behavior to be.
I've updated the docs.
* Fix an issue where start wasn't being handled properly in the
load() call.
* Fix an issue where the error message was misformatted for illegal
NaN values.
* Fix #92 and
#115, and update
the type declarations for load() and loads() to reflect that either
bytes or strings might be passed in.
* Updated all the dev dependencies and fixed all of the issues raised
by newer versions of ruff lint and pylint.
- v0.15.0 (2026-06-19)
* Fix #113, where parsing
of signed hexadecimal literals such as -0x1f and +0xff was raising
a ValueError instead of being allowed.
* Fix #111, with a slight
docstring typo fix/improvement.
* upgraded package dependences to their "latest" versions (without changing
any direct dependencies and not pulling in anything newer than 30 days
old). This resolves some potential security issues in a dev environment
that Dependabot was warning about; no effect on the production code
since the production library has no dependencies.
- v0.14.0 (2026-03-27)
This is really just a dependency bump release.
* No (non-test) code changes.
* Upgraded to latest packages for dev dependencies
* Silenced a couple of warnings from the latest pylint where it
can't deal with both unreachable code and a bad return value.
- v0.13.0 (2026-01-01)
* No code changes.
* Add Python 3.14 to supported version, project config, dependencies
* Update dependencies to latest stuff < 2025-12-01
* This relaxes the versions specified in pyproject.toml to just
use 'newer than' rather than exact matches.
* Sets 'uv.tool.exclude-newer' in pyproject.toml to tell uv not
to look at packages published within the past 30 days; this will
hopefully help prevent dependencies on compromised projects.
* Switch to using dependency-groups for 'dev' group.
- v0.12.1 (2025-08-12)
* Fix #94, where objects
returned from a custom encoder were not being indented properly.
- v0.12.0 (2025-04-03)
* Roll back pyproject.toml change for licensing so that we can still
build the package on 3.8.
* Upgrade devenv package dependencies to latest versions; they now need
Python 3.9 or newer, though json5 itself still supports 3.8.
- v0.11.0 (2025-04-01)
* Add a couple examples to the documentation and run doctest over
them.
* Fix a typing issue in dump and dumps with the cls argument; turns
out mypy was right and I was wrong and I didn't realize it :).
* Introduce a new parse method that can be used to iterate through
a string, extracting multiple values.
* Add a new consume_trailing parameter to load/loads/parse
that specifies whether to keep parsing after a valid object is
reached. By default, this is True and the string must only contain
trailing whitespace. If set to False, parsing will stop when a
valid object is reached.
* Add a new start parameter to load/loads/parse to specify
the zero-based offset to start parsing the string or file from.
* GitHub issue #60.
Fix a bug where we were attempting to allow '--4' as a valid number.
- v0.10.0 (2024-11-25)
* GitHub issue #57.
Added a JSON5Encoder class that can be overridden to do custom
encoding of values. This class is vaguely similar to the JSONEncoder
class in the standard json library, except that it has an
encode() method that can be overridden to customize any
value, not just ones the standard encoder doesn't know how to handle.
It does also support a default() method that can be used to
encode things not normally encodable, like the JSONEncoder class.
It does not support an iterencode method. One could probably
be added in the future, although exactly how that would work and
interact with encode is a little unclear.
* Restructured the code to use the new encoder class; doing so actually
allowed me to delete a bunch of tediously duplicative code.
* Added a new quote_style argument to dump()/dumps() to control
how strings are encoded by default. For compatibility with older
versions of the json5 library and the standard json library, it
uses QuoteStyle.ALWAYS_DOUBLE which encodes all strings with double
quotes all the time. You can also configure it to use single quotes
all the time (ALWAYS_SINGLE), and to switch between single and double
when doing so eliminates a need to escape quotes (PREFER_SINGLE and
PREFER_DOUBLE). This also adds a --quote-style argument to
python -m json5.
* This release has a fair number of changes, but is intended to be
completely backwards-compatible. Code without changes should run exactly
as it did before.
- v0.9.28 (2024-11-11)
* Fix GitHub CI to install uv so ./run tests works properly.
* Mark Python3.13 as supported in package metadata.
* Update dev package dependencies (note that the latest versions
of coverage and pylint no longer work w/ Python3.8)
- v0.9.27 (2024-11-10)
* Fix typo in //README.md
- v0.9.26 (2024-11-10)
* GitHub issue #82
Add support for the strict parameter to load()/loads().
* Significantly rework the infra and the run script to be
contemporary.
- v0.9.25 (2024-04-12)
* GitHub issue #81
Explicitly specify the directory to use for the package in
pyproject.toml.
- v0.9.24 (2024-03-16)
* Update GitHub workflow config to remove unnecessary steps and
run on pull requests as well as commits.
* Added note about removing hypothesize in v0.9.23.
* No code changes.
- v0.9.23 (2024-03-16)
* Lots of cleanup:
* Removed old code needed for Python2 compatibility.
* Removed tests using hypothesize. This ran model-based checks
and didn't really add anything useful in terms of coverage to
the test suite, and it introduced dependencies and slowed down
the tests significantly. It was a good experiment but I think
we're better off without it.
* Got everything linting cleanly with pylint 3.1 and ruff check
using ruff 0.3.3 (Note that commit message in 00d73a3 says pylint
3.11, which is a typo).
* Code reformatted with ruff format
* Added missing tests to bring coverage up to 100%.
* Lots of minor code changes as the result of linting and coverage
testing, but no intentional functional differences.
- v0.9.22 (2024-03-06)
* Attempt to fix the GitHub CI configuration now that setup.py
is gone. Also, test on 3.12 instead of 3.11.
* No code changes.
- v0.9.21 (2024-03-06)
Moved the benchmarks/.json data files' license information
to //LICENSE to (hopefully) make the Google linter happy.
- v0.9.20 (2024-03-03)
* Added json5.__version__ in addition to json5.VERSION.
* More packaging modernization (no more setup.{cfg,py} files).
* Mark Python3.12 as supported in project.classifiers.
* Updated the //run script to use python3.
- v0.9.19 (2024-03-03)
* Replaced the benchmarking data files that came from chromium.org with
three files obtained from other datasets on GitHub. Since this repo
is vendored into the chromium/src repo it was occasionally confusing
people who thought the data was actually used for non-benchmarking
purposes and thus updating it for whatever reason.
* No code changes.
- v0.9.18 (2024-02-29)
* Add typing information to the module. This is kind of a big change,
but there should be no functional differences.
- v0.9.17 (2024-02-19)
* Move from setup.py to pyproject.toml.
* No code changes (other than the version increasing).
- v0.9.16 (2024-02-19)
* Drop Python2 from setup.py
* Add minimal packaging instructions to //README.md.
- v0.9.15 (2024-02-19)
* Merge in Pull request #66
to include the tests and sample file in a source distribution.
- v0.9.14 (2023-05-14)
* GitHub issue #63
Handle +Infinity as well as -Infinity and Infinity.
- v0.9.13 (2023-03-16)
* GitHub PR #64
Remove a field from one of the JSON benchmark files to
reduce confusion in Chromium.
* No code changes.
- v0.9.12 (2023-01-02)
* Fix GitHub Actions config file to no longer test against
Python 3.6 or 3.7. For now we will only test against an
"oldest" release (3.8 in this case) and a "current"
release (3.11 in this case).
- v0.9.11 (2023-01-02)
* GitHub issue #60
Fixed minor Python2 compatibility issue by referring to
float("inf") instead of math.inf.
- v0.9.10 (2022-08-18)
* GitHub issue #58
Updated the //README.md to be clear that parsing arbitrary JS
code may not work.
* Otherwise, no code changes.
- v0.9.9 (2022-08-01)
* GitHub issue #57
Fixed serialization for objects that subclass int or float:
Previously we would use the objects __str__ implementation, but
that might result in an illegal JSON5 value if the object had
customized __str__ to return something illegal. Instead,
we follow the lead of the JSON module and call int.__repr__
or float.__repr__ directly.
* While I was at it, I added tests for dumps(-inf) and dumps(nan)
when those were supposed to be disallowed by allow_nan=False.
- v0.9.8 (2022-05-08)
* GitHub issue #47
Fixed error reporting in some cases due to how parsing was handling
nested rules in the grammar - previously the reported location for
the error could be far away from the point where it actually happened.
- v0.9.7 (2022-05-06)
* GitHub issue #52
Fixed behavior of default fn in dump and dumps. Previously
we didn't require the function to return a string, and so we could
end up returning something that wasn't actually valid. This change
now matches the behavior in the json module. *Note: This is a
potentially breaking change.*
- v0.9.6 (2021-06-21)
* Bump development status classifier to 5 - Production/Stable, which
the library feels like it is at this point. If I do end up significantly
reworking things to speed it up and/or to add round-trip editing,
that'll likely be a 2.0. If this version has no reported issues,
I'll likely promote it to 1.0.
* Also bump the tested Python versions to 2.7, 3.8 and 3.9, though
earlier Python3 versions will likely continue to work as well.
* GitHub issue #46
Fix incorrect serialization of custom subtypes
* Make it possible to run the tests if hypothesis isn't installed.
- v0.9.5 (2020-05-26)
* Miscellaneous non-source cleanups in the repo, including setting
up GitHub Actions for a CI system. No changes to the library from
v0.9.4, other than updating the version.
- v0.9.4 (2020-03-26)
* GitHub pull #38
Fix from [email protected] for dumps() crashing when passed
an empty string as a key in an object.
- v0.9.3 (2020-03-17)
* GitHub pull #35
Fix from pastelmind@ for dump() not passing the right args to dumps().
* Fix from [email protected] to remove the tests directory from
the setup call, making the package a bit smaller.
- v0.9.2 (2020-03-02)
* GitHub pull #34
Fix from roosephu@ for a badly formatted nested list.
- v0.9.1 (2020-02-09)
* GitHub issue #33:
Fix stray trailing comma when dumping an object with an invalid key.
- v0.9.0 (2020-01-30)
* GitHub issue #29:
Fix an issue where objects keys that started with a reserved
word were incorrectly quoted.
* GitHub issue #30:
Fix an issue where dumps() incorrectly thought a data structure
was cyclic in some cases.
* GitHub issue #32:
Allow for non-string keys in dicts passed to dump()/dumps().
Add an allow_duplicate_keys=False to prevent possible
ill-formed JSON that might result.
- v0.8.5 (2019-07-04)
* GitHub issue #25:
Add LICENSE and README.md to the dist.
* GitHub issue #26:
Fix printing of empty arrays and objects with indentation, fix
misreporting of the position on parse failures in some cases.
- v0.8.4 (2019-06-11)
* Updated the version history, too.
- v0.8.3 (2019-06-11)
* Tweaked the README, bumped the version, forgot to update the version
history :).
- v0.8.2 (2019-06-11)
* Actually bump the version properly, to 0.8.2.
- v0.8.1 (2019-06-11)
* Fix bug in setup.py that messed up the description. Unfortunately,
I forgot to bump the version for this, so this also identifies as 0.8.0.
- v0.8.0 (2019-06-11)
* Add allow_duplicate_keys=True as a default argument to
json5.load()/json5.loads(). If you set the key to False, duplicate
keys in a single dict will be rejected. The default is set to True
for compatibility with json.load(), earlier versions of json5, and
because it's simply not clear if people would want duplicate checking
enabled by default.
- v0.7 (2019-03-31)
* Changes dump()/dumps() to not quote object keys by default if they are
legal identifiers. Passing quote_keys=True will turn that off
and always quote object keys.
* Changes dump()/dumps() to insert trailing commas after the last item
in an array or an object if the object is printed across multiple lines
(i.e., if indent is not None). Passing trailing_commas=False will
turn that off.
* The json5.tool command line tool now supports the --indent,
--[no-]quote-keys, and --[no-]trailing-commas flags to allow
for more control over the output, in addition to the existing
--as-json flag.
* The json5.tool command line tool no longer supports reading from
multiple files, you can now only read from a single file or
from standard input.
* The implementation no longer relies on the standard json module
for anything. The output should still match the json module (except
as noted above) and discrepancies should be reported as bugs.
- v0.6.2 (2019-03-08)
* Fix GitHub issue #23 and
pass through unrecognized escape sequences.
- v0.6.1 (2018-05-22)
* Cleaned up a couple minor nits in the package.
- v0.6.0 (2017-11-28)
* First implementation that attempted to implement 100% of the spec.
- v0.5.0 (2017-09-04)
* First implementation that supported the full set of kwargs that
the json` module supports.