9.x Series

IPython 9.5

Featuring improvements for numerous magics (%autoreload, %whos, %%script, %%notebook), a streaming performance regression fix, completer policy overrides improvements, and initial support for Python 3.14.

• #14938 Fix printing long strings in %whos magic command

• #14941 Fix performance of streaming long text

• #14943 Simplify overriding selective evaluation policy settings for modules

• #14955 Populate notebook metadata when exporting with %notebook magic

• #14960 Better handling in deduperreload for patching functions with freevars

• #14964 Fix traceback logic for non-SyntaxError exceptions in plain mode

• #14966 Do not warn repeatedly if policy overrides are not applicable

• #14967 Support Python 3.14.0rc2, test on CI

• #14969 Fix truncated output in %script magic

• #14970 Fix exceptions in %whos magic command

The %notebook magic now stores the language and kernel information in notebook metadata, allowing users to quickly open the exported notebook with syntax highlighting and an appropriate kernel.

The completer Completer.policy_overrides traitlet handling was improved. It no longer repeatedly warns on each completion after switching away to a policy that does not support previously specified overrides. Allow-listing attribute access on all objects in a given library is now possible. The specification now also accepts dotted strings (rather than requiring tuples to specify the path) which should make configuration easier and less error-prone.

c.Completer.policy_overrides = {
    "allowed_getattr_external": {
        "my_trusted_library"
    }
}

A number of recent regressions were fixed:

• %autoreload now again shows the correct module name in traceback

• standard output/error streaming of long text/logs is now as fast as in IPython 9.0

• in the %whos magic handling of long strings and class objects that implement __len__ was fixed.

IPython 9.4

Featuring %autoreload, %whos, %%script, %%time magic improvements, along with a fix for use of list comprehensions and generators in the interactive debugger (and ipdb).

• #14922 Improved reloading of decorated functions when using %autoreload

• #14872 Do not always import all variables with %autoreload 3

• #14906 Changed behaviour of %time magic to always interrupt execution on exception and always show execution time

• #14926 Support data frames, series, and objects with __len__ in the %whos magic

• #14933 List comprehensions and generators now work reliably in debugger on all supported Python versions

• #14931 Fix streaming multi-byte Unicode characters in the %script magic and its derivatives

The %time magic no longer swallows exceptions raised by the measured code, and always prints the time of execution. If you wish the execution to continue after measuring time to execute code that is meant to raise an exception, pass the new --no-raise-error flag. The --no-raise-error flag does not affect KeyboardInterrupt as this exception is used to signal intended interruption of execution flow.

Previously the debugger (ipdb) evaluation of list comprehensions and generators could fail with NameError due to generator implementation detail in CPython. This was recently fixed in Python 3.13. Because IPython is often used for interactive debugging, this release includes a backport of that fix, providing users who cannot yet update from Python 3.11 or 3.12 with a smoother debugging experience.

The %autoreload magic is now more reliable. The behaviour around decorators has been improved and %autoreload 3 no longer imports all symbols when reloading the module, however, the heuristic used to determine which symbols to reload can sometimes lead to addition of imports from non-evaluated code branches, see issue #14934.

IPython 9.3

This release includes improvements to the tab and LLM completer, along with typing improvements:

• #14911 Implement auto-import and evaluation policy overrides

• #14910 Eliminate startup delay when LLM completion provider is configured

• #14898 Fix attribute completion for expressions with comparison operators

• #14908 Fix typing of error_before_exec, enhance mypy coverage

Notably, the native completer can now suggest attribute completion on not-yet-imported modules. This is particularly useful when writing code which includes an import and the use of the imported module in the same line or in the same cell; the default implementation does not insert the imported module into the user namespace, for which an actual execution is required.

The auto-import of modules by completer is turned off and requires opting-in using a new Completer.policy_overrides traitlet. To enable auto-import on completion specify:

ipython --Completer.policy_overrides='{"allow_auto_import": True}' --Completer.use_jedi=False

This change aligns the capability of both jedi-powered and the native completer. The function used for auto-import can be configured using Completer.auto_import_method traitlet.

IPython 9.2

This is a small release with minor changes in the context passed to the LLM completion provider along few other bug fixes and documentation improvements:

• #14890 Fixed interruption of %%time and %%debug magics

• #14877 Removed spurious empty lines from prefix passed to LLM, and separated part after cursor into the suffix

• #14876 Fixed syntax warning in Python 3.14 (remove return from finally block)

• #14887 Documented the recommendation to use ipykernel.embed.embed_kernel() over ipython.embed.

IPython 9.1

This is a small release that introduces enhancements to %notebook and %%timeit magics, and a number of bug fixes related to colors/formatting, performance, and completion.

%notebook saves outputs

The %notebook magic can be used to create a Jupyter notebook from the commands executed in the current IPython session (since the interpreter startup).

Prior to IPython 9.1, the resulting notebook did not include the outputs, streams, or exceptions. IPython 9.1 completes the implementation of this magic allowing for an easier transition from an interactive IPython session to a Jupyter notebook.

To capture streams (stdio/stderr), IPython temporarily swaps the write method of the active stream class during code execution. This ensures compatibility with ipykernel which swaps the entire stream implementation and requires it to remain an instance of IOStream subclass. If this leads to undesired behaviour in any downstream applications, your feedback and suggestions would be greatly appreciated.

%%timeit -v argument

New -v argument allows users to save the timing result directly to a specified variable, e.g.

%%timeit -v timing_result
2**32

Completer improvements

The LLM-based completer will now receive the request number for each subsequent execution.

The tab completer used when jedi is turned off now correctly completes variables in lines where it previously was incorrectly attempting to complete attributes due to simplistic context detection based on the presence of a dot.

Thanks

A big thank you to everyone who contributed towards the 9.1 release, including new contributors: @Darshan808, @kwinkunks, @carschandler, returning contributors (shout out to @wjandrea!), and of course @Carreau whom I would like to thank for the guidance in the preparation of this release and stewardship of IPython over the years - Mike.

As usual, you can find the full list of PRs on GitHub under the 9.1 milestone.

IPython 9.0

Welcome to IPython 9.0. As with any version of IPython before this release, it should not be majorly different from the previous version, at least on the surface. We still hope you can upgrade as soon as possible and look forward to your feedback.

I take the opportunity of this new release to remind you that IPython is governed by the Jupyter code of conduct. And that even beyond so we strive to be an inclusive, accepting and progressive community, Here is a relevant extract from the COC.

We strive to be a community that welcomes and supports people of all backgrounds and identities. This includes, but is not limited to, members of any race, ethnicity, culture, national origin, color, immigration status, social and economic class, educational level, sex, sexual orientation, gender identity and expression, age, physical appearance, family status, technological or professional choices, academic discipline, religion, mental ability, and physical ability.

As a short overview of the changes in 9.0, we have over 100 PRs merged since 8.x, many of which are refactors, cleanups and simplifications.

• (optional) LLM integration in the CLI.

• Complete rewrite of color and theme handling, which now supports more colors and symbols.

• Move tests out of tree in the wheel with a massive reduction in file size.

• Tips at startup

• Removal of (almost) all deprecated functionalities and options.

• Stricter and more stable codebase.

Removal and deprecation

I am not going to list the removals and deprecations, but anything deprecated since before IPython 8.16 is gone, including many shim modules and indirect imports that would just re-expose IPykernel, qtconsole, etc.

A number of new deprecations have been added (run your test suites with -Werror), as those will be removed in the future.

Color and theme rewrite

IPython's color handling had grown many options through the years, and it was quite entrenched in the codebase, directly emitting ansi escape sequences deep in traceback printing and other places.

This made developing new color schemes difficult, and limited us to the 16 colors of the original ansi standard defined by your terminal.

Syntax highlighting was also inconsistent, and not all syntax elements were always using the same theme.

Using (style, token) pairs

Starting with 9.0, the color and theme handling has been rewritten, and internally all the printing is done by yielding pairs of Style and token objects (compatible with pygments and prompt_toolkit), then as much as possible, IPython formats these objects at the last moment, using the current theme.

256-bit colors and unicode symbols

This means that new themes can now use all of pygments's color names and functionalities, and you can define for each token style, the foreground, background, underline, bold, italic and likely a few other options.

In addition, themes now provide a number of symbols, that can be used when rendering traceback or debugger prompts. This let you customize the appearance a bit more. For example, instead of using dash and greater-than sign, The arrow pointing the current frame can actually use horizontal line and right arrow unicode symbol, for a more refined experience.

New themes using colors and symbols

All the existing themes (Linux, LightBG, Neutral and NoColor) should not see any changes, but I added two new pride themes, that show the use of 256bits colors and unicode symbols. I'm not a designer, so feel free to suggest updates and new themes to add.

Themes currently still require writing a bit of Python, but I hope to get contributions for IPython to be able to load them from text files, for easier redistribution.

Tips at startup

IPython now displays a few tips at startup (1 line), to help you discover new features. All those are in the codebase, and can be displayed randomly or based on date. You can disable it via a configuration option or the --no-tips flag.

Please contribute more tips by sending pull requests!

Out-of-tree tests

And more generally I have changed the folder structure and what is packaged in the wheel to reduce the file size. The wheel is down from 825kb to 590kb (-235kb) which is about a 28% reduction. This should help when you run IPython via Pyodide – when your browser needs to download it.

According to https://pypistats.org/packages/ipython, IPython is downloaded about 13 million times per week, so this should reduce PyPI bandwidth by about 2Tb each week, which is small compared to the total download, but still, trying to reduce resource usage is a worthy goal.

Integration with Jupyter-AI LLM

This feature allow IPython CLI to make use of Jupyter-AI provider to use LLM for suggestion, and completing the current text. Unlike many features of IPython this is disabled by default, and need several configuration options to be set to work:

• Choose a provider in jupyter-ai and set it as default one: c.TerminalInteractiveShell.llm_provider_class = <fully qualified path> You likely need to setup your provider with API key or other things.

• Choose and available shortcut (I'll take Ctrl-Q as an example) and bind to trigger llm_autosuggestion only while typing.

c.TerminalInteractiveShell.shortcuts = [
     {
         "new_keys": ["c-q"],
         "command": "IPython:auto_suggest.llm_autosuggestion",
         "new_filter": "navigable_suggestions & default_buffer_focused",
         "create": True,
     },
 ]

See llm_suggestions for more.

Thanks as well to the D. E. Shaw group for sponsoring this work.

For something completely different

Ruth Bader Ginsburg 1933-2020 was an American lawyer and jurist who served on the Supreme Court of the United States. Ginsburg spent much of her legal career as an advocate for gender equality, women's rights, abortion rights, and religious freedom.

Thanks

Thanks to everyone who helped with the 9.0 release and working toward 9.0.

As usual you can find the full list of PRs on GitHub under the 9.0 milestone.