Codebraid – live code in Pandoc Markdown

Codebraid is a Python program that enables executable code in Pandoc Markdown documents. Using Codebraid can be as simple as adding a class to your code blocks' attributes, and then running codebraid rather than pandoc to convert your document from Markdown to another format. codebraid supports almost all of pandoc's options and passes them to pandoc internally. See Codebraid Preview for VS Code for editor support. See the Codebraid website for additional examples and documentation.

Codebraid provides two options for executing code. It includes a built-in code execution system that currently supports Python 3.7+, Julia, Rust, R, Bash, JavaScript, GAP and SageMath. Code can also be executed using Jupyter kernels, with support for rich output like plots.

Development: https://github.com/gpoore/codebraid

Citing Codebraid: "Codebraid: Live Code in Pandoc Markdown", Geoffrey M. Poore, Proceedings of the 18th Python in Science Conference, 2019, 54-61.

View example HTML output, or see the Markdown source or raw HTML (the Python and Rust examples demonstrate more advanced features at the end):

Python example [Pandoc Markdown source] [raw HTML]
Jupyter example [Pandoc Markdown source] [raw HTML]
Rust example [Pandoc Markdown source] [raw HTML]
Julia example [Pandoc Markdown source] [raw HTML]
R example [Pandoc Markdown source] [raw HTML]
Bash example [Pandoc Markdown source] [raw HTML]
JavaScript example [Pandoc Markdown source] [raw HTML]
GAP example [Pandoc Markdown source] [raw HTML]

Simple example

Markdown source test.md:

```{.python .cb-run}
var = 'Hello from Python!'
var += ' $2^8 = {}$'.format(2**8)
```

```{.python .cb-run}
print(var)
```

Run codebraid (to save the output, add something like -o test_out.md, and add --overwrite if it already exists):

codebraid pandoc --from markdown --to markdown test.md

Output:

Hello from Python! $2^8 = 256$

As this example illustrates, variables persist between code blocks; by default, code is executed within a single session. Code output is also cached by default so that code is only re-executed when modified.

Features

Comparison with Jupyter, knitr, and Pweave

	Codebraid	Jupyter Notebook	knitr	Pweave
multiple programming languages per document	✓	✓*	✓†	✓*
multiple independent sessions per language	✓
inline code execution within paragraphs	✓		✓	✓
no out-of-order code execution	✓		✓‡	✓
no markdown preprocessor or custom syntax	✓	✓
minimal diffs for easy version control	✓		✓	✓
insert code output anywhere in a document	✓		✓
can divide code into incomplete snippets	✓		✓	✓
support for literate programming	✓		✓
compatible with any text editor	✓		✓	✓

* One primary language from the Jupyter kernel. The IPython kernel supports additional languages via %%script magics. There is no continuity between %%script cells, because each cell is executed in a separate process. Some magics, such as those provided by PyJulia and rpy2, provide more advanced capabilities.
† knitr only provides continuity between code chunks for R, and more recently Python and Julia. Code chunks in other languages are executed individually in separate processes.
‡ Out-of-order execution is possible with R Markdown notebooks.

The table above summarizes Codebraid features in comparison with Jupyter notebooks (without extensions), knitr (R Markdown), and Pweave, emphasizing Codebraid's unique features. Here are some additional points to consider:

Jupyter notebooks — Notebooks have a dedicated, browser-based graphical user interface. Jupyter kernels typically allow the code in a cell to be executed without re-executing any preceding code, providing superior interactivity. Codebraid has advantages for projects that are more focused on creating a document than on exploratory programming.

knitr — R Markdown documents have a dedicated user interface in R Studio. knitr provides superior support for R, as well as significant Python and Julia support that includes R integration. Codebraid offers continuity between code chunks for all supported languages, as well as multiple independent sessions per language. It also provides unique options for displaying code and its output.

More about key features

Easy debugging — By default, stderr is shown automatically in the document whenever there is an error, right next to the code that caused it. It is also possible to monitor code output in real time during execution via --live-output.

Simple language support — Codebraid supports Jupyter kernels. It also has a built-in system for executing code. Adding support for a new language with this system can take only a few minutes. Just create a config file that tells Codebraid which program to run, which file extension to use, and how to write to stdout and stderr. See languages/ for examples.

No preprocessor — Unlike many approaches to making code in Markdown executable, Codebraid is not a preprocessor. Rather, Codebraid acts on the abstract syntax tree (AST) that Pandoc generates when parsing a document. Preprocessors often fail to disable commented-out code blocks because the preprocessor doesn't recognize Markdown comments. Preprocessors can also fail due to the finer points of Markdown parsing. None of this is an issue for Codebraid, because Pandoc does the Markdown parsing.

No custom syntax — Codebraid introduces no additional Markdown syntax. Making a code block or inline code executable uses Pandoc's existing syntax for defining code attributes.

Installation and requirements

Installation: pip3 install codebraid or pip install codebraid

Manual installation: python3 setup.py install or python setup.py install

Requirements:

Pandoc 2.4+ (2.17.1.1+ recommended for commonmark_x).
Python 3.7+ with setuptools, and bespon 0.6 (bespon installation is typically managed by pip/setup.py)
For Jupyter support, jupyter_client and language kernels
For YAML metadata support, ruamel.yaml (can be ruamel_yaml for Anaconda installations)

Converting a document

Simply run codebraid pandoc <normal pandoc options>. Codebraid currently supports Pandoc Markdown (--from markdown) and CommonMark with Pandoc extensions (--from commonmark_x) as input formats.

Note that --overwrite is required to overwrite existing files. If you are using a defaults file, --from, --to, and --output must be given explicitly and cannot be inherited from the defaults file. If you are using a defaults file and converting to a standalone Pandoc Markdown document, --standalone should be given explicitly rather than being inherited from the defaults file.

codebraid should typically be run in the same directory as the document, so that the default working directory for code is the document directory.

If you are converting from Pandoc Markdown to Pandoc Markdown with --standalone (basically using codebraid to preprocess Markdown documents), note that the following YAML metadata fields and command-line options are ignored in that situation:

header-includes and --include-in-header
include-before and --include-before-body
include-after and --include-after-body
toc/table-of-contents and --toc/--table-of-contents

This is typically what you want. Usually, "include" and a table of contents are desired in a final output format like HTML or PDF, not in a Pandoc Markdown file. In the rare cases where "includes" and a table of contents are needed in Markdown documents, this can be accomplished by piping the output of codebraid through pandoc.

Additional non-Pandoc command-line options

--live-output — Show code output (stdout and stderr) live in the terminal during code execution. For Jupyter kernels, also show errors and a summary of rich output. Output still appears in the document as normal.

Individual sessions can override this by setting live_output=false in the document.
--no-execute — Disables code execution. Only use available cached output.
--only-code-output={format} — Write code output in JSON Lines format to stdout as soon as it is available, and do not create a document.

This is intended for use with Codebraid Preview, so that document previews can be updated during code execution. Currently, the only supported format is codebraid_preview. One JSON data object followed by a newline is written to stdout for each code chunk. In some cases, the data for a chunk will be resent later if the data relevant for a chunk changes (for example, if code execution fails after the first chunk runs, but in such a way that an error message needs to be attached to the first chunk). Data for a chunk is sent as soon as it is available from code processing, from cache, or from code execution (as soon as the chunk completes, typically before the session completes). Additional JSON data may be sent to provide tracking of code execution progress or information such as metadata. The JSON data provided for format codebraid_preview may change between minor versions.

Caching

By default, code output is cached, and code is only re-executed when it is modified. The default cache location is a _codebraid directory in the working directory (directory where codebraid is run, typically the document directory). This can be modified using --cache-dir. Multiple documents can share a single cache location. A cache directory can be synced between different operating systems (such as Windows and Linux) while retaining full functionality so long as documents are in equivalent locations under the user's home directory (as resolved by os.path.expanduser()).

When multiple documents share the same cache location, each document will automatically clean up its own unused, outdated files. However, if a document is deleted or renamed, it may leave behind unused files in the cache, so it may be worth manually deleting and regenerating the cache in those circumstances. Future cache enhancements should be able to detect all unused files, making this unnecessary.

If you are working with external data that changes, you should run codebraid with --no-cache or delete the cache as necessary to prevent the cache from becoming out of sync with your data. Future releases will allow external dependencies to be specified so that caching will work correctly in these situations.

YAML metadata

Some document-wide settings can be given in the Markdown YAML metadata. Codebraid settings must be under either a codebraid or codebraid_ key in the metadata. Pandoc will ignore codebraid_ so it will not be available to filters; this distinction should not typically be important.

To use Jupyter kernels automatically for all sessions, simply set jupyter: true. For example,

---
codebraid:
  jupyter: true
---

It is also possible to set a default kernel and/or default timeout. For example,

---
codebraid:
  jupyter:
    kernel: python3
    timeout: 120
---

A Jupyter kernel and/or timeout can still be set in the first code chunk for a given session, and will override the document-wide default.

It is also possible to set live_output: <bool> in the metadata. Additional metadata settings will be added in future releases.

Code options

Commands (Classes)

Code is made executable by adding a Codebraid class to its Pandoc attributes. For example, `code`{.python} becomes `code`{.python .cb-run}.

When code is executed, the output will depend on whether the built-in code execution system or a Jupyter kernel is used.

When code is executed with the built-in system, the output is equivalent to collecting all code for each session of each language, saving it to a file, and then executing it (with an added compile step for some languages). For example, running Python code is equivalent to saving it to file.py and then running python file.py, while running R code is equivalent to saving it to file.R and then running Rscript file.R. Code is not executed as it would be in an interactive session (like running python or R at the command prompt). As a result, some output that would be present in an interactive session is absent. For example, in interactive sessions for some languages, simply entering a variable returns a string representation without explicit printing, and plotting opens a separate image window or displays an image inline. Such output is absent in Codebraid unless it is also produced when code is executed as a script rather than in an interactive session. The .cb-expr command is provided for when an inline string representation of a variable is desired.

An option for interactive-style code execution with the built-in system is planned for a future release. In the meantime, many interactive-style features are available between the .cb-expr command and Jupyter kernels.

When code is executed with a Jupyter kernel, the default output will be equivalent to executing it in a Jupyter notebook. Rich output such plots, images, and LaTeX math will be displayed automatically by default. This can be customized by using the show and hide options.

All classes for making code executable are listed below. These all have the form .cb-<command>. Classes with the form .cb.<command> (period rather than hyphen) are supported for Pandoc Markdown (--from markdown), but not for commonmark_x since it has a more restricted class syntax. The forms shown below (.cb-<command>) should be preferred for compatibility across Markdown variants supported by Pandoc.

.cb-code — Insert code verbatim, but do not run it. This is primarily useful when combined with other features like naming and then copying code chunks.
.cb-expr — Evaluate an expression and interpret the result as Markdown. Only works with inline code. This is not currently compatible with Jupyter kernels.
.cb-nb — Execute code in notebook mode. For inline code, this is equivalent to .cb-expr with verbatim output unless a Jupyter kernel is used, in which case rich output like plots or LaTeX will be displayed. For code blocks, this inserts the code verbatim, followed by any printed output (stdout) verbatim. If stderr exists, it is also inserted verbatim. When a Jupyter kernel is used, rich output like plots or LaTeX is also displayed.
.cb-paste — Insert code and/or output copied from one or more named code chunks. The copy keyword is used to specify chunks to be copied. This does not execute any code. Unless show is specified, display options are inherited from the first copied code chunk.

If content is copied from multiple code chunks that are executed, all code chunks must be in the same session and must be in sequential order without any omitted chunks. This ensures that what is displayed is always consistent with what was executed.

If content is copied from another cb-paste code chunk, only a single code chunk can be copied. This reduces the indirection that is possible when displaying the output of code that has been executed. This restriction may be removed in the future.
.cb-run — Run code and interpret any printed content (stdout) as Markdown. Also insert stderr verbatim if it exists. When a Jupyter kernel is used, rich output like plots or LaTeX is also displayed.

Keyword arguments

Pandoc code attribute syntax allows keyword arguments of the form key=value, with spaces (not commas) separating subsequent keys. value can be unquoted if it contains only letters and some symbols; otherwise, double quotation marks "value" are required. For example,

{.python key1=value1 key2=value2}

Codebraid adds support for additional keyword arguments. In some cases, multiple keywords can be used for the same option. This is primarily for Pandoc compatibility.

First chunk settings

These are only permitted for the first code chunk in a session (or the first chunk for a language, if a session is not specified and thus the default session is in use).

executable={string} — Executable to use for running or compiling code, instead of the default. This only applies to Codebraid's built-in code execution system.
executable_opts={string} — Command-line options passed to executable. This only applies to Codebraid's built-in code execution system.
args={string} — Command-line arguments passed to code during execution. For example, this could be used to add values to sys.argv for Python. This only applies to Codebraid's built-in code execution system.
jupyter_kernel={string} — Jupyter kernel to use for executing code instead of Codebraid's built-in code execution system. Multiple Jupyter kernels can be used within a single document, and multiple sessions are possible per kernel. Except when otherwise specified, Jupyter kernels should be usable just like the built-in code execution system.
jupyter_timeout={int} — Jupyter kernel timeout per code chunk in seconds. The default is 60.
live_output={true, false} — Show code output (stdout and stderr) live in the terminal during code execution. For Jupyter kernels, also show errors and a summary of rich output. Output still appears in the document as normal. Showing output can also be enabled via the command-line option --live-output.

When live_output=false is set for a session, this setting takes precedence over the command-line option --live-output, and output will not be shown for that session.

All output is written to stderr, so stdout only contains the document when --output is not specified. Output is interspersed with delimiters marking the start of each session and the start of each code chunk. The delimiters for the start of each code chunk include source names and line numbers.

With Codebraid's built-in code execution system, the output for a code chunk may be delayed until all code in the chunk has finished executing, unless code output is line buffered or code manually flushes stdout and stderr. For example, with Python you may want to use print functions like print("text", flush=True). Another option is to use Python in line-buffered mode by setting executable_opts="-u" in the first code chunk of a session.

With Jupyter kernels, the output for a code chunk will be delayed until all code in the chunk has finished executing.

Execution

complete={true, false} — By default, code chunks must contain complete units of code (function definitions, loops, expressions, and so forth). With complete=false, this is not required. Any stdout from code chunks with complete=false is accumulated until the next code chunk with complete=true (the default value), or until the end of the session, whichever comes first.

Setting complete is incompatible with outside_main=true, since the complete status of code chunks with outside_main=true is inferred automatically.
outside_main={true, false} — This allows code chunks to overwrite the Codebraid template code when code is executed with Codebraid's built-in code execution system. It is primarily useful for languages like Rust, in which code is inserted by default into a main() template. In that case, if a session starts with one or more code chunks with outside_main=true, these are used instead of the beginning of the main() template. Similarly, if a session ends with one or more code chunks with outside_main=true, these are used instead of the end of the main() template. If there are any code chunks in between that lack outside_main (that is, default outside_main=false), then these will have their stdout collected on a per-chunk basis like normal. Having code chunks that lack outside_main is not required; if there are none, the total accumulated stdout for a session belongs to the last code chunk in the session.

outside_main=true is incompatible with explicitly setting complete. The complete status of code chunks with outside_main=true is inferred automatically.
session={identifier-style string} — By default, all code for a given language is executed in a single, shared session so that data and variables persist between code chunks. This option allows code to be separated into multiple independent sessions. Session names must be Python-style identifiers.

Display

first_number/startFrom/start-from/start_from={integer or next} — Specify the first line number for code when line numbers are displayed. next means continue from the last code in the current session.
hide={markup, copied_markup, code, stdout, stderr, expr, rich_output, all} — Hide some or all of the elements that are displayed by default. Elements can be combined. For example, hide=stdout+stderr. Note that expr only applies to .cb-expr or .cb-nb with inline code using Codebraid's built-in code execution system, since only these evaluate an expression. rich_output is currently only relevant for Jupyter kernels.
hide_markup_keys={key(s)} — Hide the specified code chunk attribute key(s) in the Markdown source displayed via markup or copied_markup. Multiple keys can be specified via hide_markup_keys=key1+key2.

hide_markup_keys only applies to the code chunk in which it is used, to determined the markup for that code chunk. Thus, it only affects copied_markup indirectly.
line_numbers/numberLines/number-lines/number_lines={true, false} — Number code lines in code blocks.
show={markup, copied_markup, code, stdout, stderr, expr, rich_output, none} — Override the elements that are displayed by default. expr only applies to .cb-expr and to .cb-nb with inline code using Codebraid's built-in code execution system, since only these evaluate an expression. Elements can be combined. For example, show=code+stdout.

Each element except rich_output can optionally specify a format from raw, verbatim, or verbatim_or_empty. For example, show=code:verbatim+stdout:raw.
- raw means interpreted as Markdown.
- verbatim produces inline code or a code block, depending on context. Nothing is produced if there is no content (for example, nothing in stdout.)
- verbatim_or_empty produces inline code containing a single non-breaking space or a code block containing a single empty line in the event that there is no content. It is useful when a placeholder is desired, or a visual confirmation that there is indeed no output.
For rich_output, the format is specified as one or more abbreviations for the mime types of the output to be displayed. For example, rich_output:plain will display text/plain output if it exists, and otherwise nothing. rich_output:png|plain will display a PNG image if it exists, or otherwise will fall back to plain text if available. The following formats are currently supported:
- latex (corresponds to text/latex)
- html (text/html)
- markdown (text/markdown)
- plain (text/plain)
- png (image/png)
- jpg and jpeg (image/jpeg)
- svg (image/svg+xml)
- pdf (application/pdf)
For rich_output formats with a text/* mime type (latex, html, markdown, plain), it is possible to specify whether they are displayed raw, verbatim, or verbatim_or_empty. For example, show=rich_output:latex:raw and show=rich_output:latex:verbatim. raw treats latex and html as raw content with those formats embedded within Markdown. raw treats markdown and plain as Markdown. When a display style is not specified, all rich_output formats with a text/* mime type are displayed raw by default, except for plain which is displayed verbatim.

markup displays the Markdown source for the inline code or code block. Because the Markdown source is not available in the Pandoc AST but rather must be recreated from it, the Markdown source displayed with markup may use a different number of backticks, quote attribute values slightly differently, or contain other insignificant differences from the original document.

copied_markup displays the Markdown source for code chunks copied via copy.

expr defaults to raw if a format is not specified. rich_output defaults to latex|markdown|png|jpg|svg|plain. All others default to verbatim.
example={bool} — Insert a code block containing the Markdown source of the code chunk, followed by the rest of the output as normal. This is only valid for inline code if the code is in a paragraph by itself. This option is currently not compatible with --only-code-output and Codebraid Preview. This option is intended primarily for documentation about Codebraid.

Copying

copy={chunk name(s)} — Copy one or more named code chunks. When copy is used with a command like .cb-run that executes code, only the code is copied, and it is executed as if it had been entered directly. When copy is used with .cb-code, only the code is copied and nothing is executed. When copy is used with .cb-paste, both code and output are copied, and nothing is executed. Multiple code chunks may be copied; for example, copy=name1+name2. In that case, the code from all chunks is concatenated, as is any output that is copied. Because copy brings in code from other code chunks, the actual content of a code block or inline code using copy is discarded. As a result, this must be empty, or a space or underscore can be used as a placeholder.
name={identifier-style string} — Name a code chunk so that it can later be copied by name. Names must be Python-style identifiers.

Including external files

include_file={path} — Include the specified file. A leading ~/ or ~<user>/ is expanded to the user's home directory under all operating systems, including under Windows with both slashes and backslashes.

When include_file is used with a command like .cb-run that executes code, the file is included and executed as part of the current session just as if the file contents had been entered directly. When include_file is used with .cb-code, the file is included and displayed just as if it had been entered directly. Because include_file brings in code from another file, the actual content of a code block or inline code using include_file is discarded. As a result, this must be empty, or a space or underscore can be used as a placeholder.
include_encoding={encoding} — Encoding for included file. The default encoding is UTF-8.
include_lines={lines/line ranges} — Include the specified lines or line ranges. For example, 1-3,5,7-9,11-. Line numbers are one-indexed. Line ranges are inclusive, so 1-3 is 1 up to and including 3. If a range ends with a hyphen, like 11-, then everything is included from the line through the end of the file.

Cannot be combined with other include options that specify what is to be included.
include_regex={regex} — Include the first segment of the file that matches the provided regular expression.

Keep in mind that Pandoc's key-value attributes evaluate backslash escapes in values whether or not the values are quoted with double quotation marks, so two levels of backslash-escaping are always necessary (one for Pandoc's strings, one for the regex itself; there are no raw strings). Regular expressions use multiline mode, so ^/$ match the start/end of a line, and \A/\Z can be used to match the start/end of the file. Regular expressions use dotall mode, so . matches anything including the newline \n; use [^\n] when this is not desired.

Cannot be combined with other include options that specify what is to be included.
include_start_string={string} — Include everything from the first occurrence of this string onward.

Can only be combined with other include options that specify the end of what is to be included.
include_start_regex={regex} — Include everything from the first match of this regex onward.

Can only be combined with other include options that specify the end of what is to be included. See include_regex for notes on regex usage.
include_after_string={string} — Include everything after the first occurrence of this string onward.

Can only be combined with other include options that specify the end of what is to be included.
include_after_regex={regex} — Include everything after the first match of this regex onward.

Can only be combined with other include options that specify the end of what is to be included. See include_regex for notes on regex usage.
include_before_string={string} — Include everything before the first occurrence of this string.

Can only be combined with other include options that specify the start of what is to be included. If the start is specified, then the first occurrence after this point is used, rather than the first occurrence in the overall file.
include_before_regex={regex} — Include everything before the first match of this regex.

Can only be combined with other include options that specify the start of what is to be included. If the start is specified, then the first match after this point is used, rather than the first match in the overall file. See include_regex for notes on regex usage.
include_end_string={string} — Include everything through the first occurrence of this string.

Can only be combined with other include options that specify the start of what is to be included. If the start is specified, then the first occurrence after this point is used, rather than the first occurrence in the overall file.
include_end_regex={regex} — Include everything through the first match of this regex.

Can only be combined with other include options that specify the start of what is to be included. If the start is specified, then the first match after this point is used, rather than the first match in the overall file. See include_regex for notes on regex usage.