Parsnips

Generate, search, and cite SWHIDs for Python code fragments like classes, functions, and expressions.

Overview

Parsnips is a Python tool that analyzes source files to create a citable hierarchy of folders and metadata files representing nodes in a parsed abstract syntax tree (AST). AST nodes correspond to meaningful code fragments such as classes, functions, and expressions. The metadata files and folders can be stored in your repository. If you archive your repository on Software Heritage, you can then retrieve and cite the SWHIDs for these fragments.

Parsnips uses the built-in ast module and asttokens to create fine-grained, reproducible identifiers for fragments of Python code. It is designed to support persistent software identifiers (SWHIDs) for structural code elements such as:

• Functions
• Classes
• Assignments
• Expressions
• Other AST nodes

The resulting parse tree is exported into a fully deterministic folder structure, where each node:

• Is stored in a dedicated folder inside a hidden .parsnips directory.
• Contains the exact source text corresponding to the node.
• Includes structured metadata for precise archival referencing.

Extraction Protocol

When Parsnips processes a Python file or directory, it follows this deterministic protocol:

1. Unified Extraction Root:

• Regardless of whether you provide a file or directory as input, Parsnips always creates a .parsnips/ subdirectory inside the directory where the file or directory resides.
• This ensures a fully consistent and predictable output structure.

1. Recursive Directory Traversal (when processing directories):

• Parsnips walks the entire directory tree starting at the input path.
• It identifies all .py files to process.
• Symlinks are not followed.
• Any pre-existing .parsnips/ subdirectories are deleted and regenerated to ensure full reproducibility.

1. Per-File Processing:

• For each Python file, Parsnips parses the Abstract Syntax Tree (AST) using ast and asttokens.
• The content SWHID for the entire file is computed using Python's hashlib.blake2s() directly, fully following the Software Heritage content SWHID specification, without external library dependencies.

1. Node Processing:

• Each AST node is visited in traversal order.
• If a node lacks a line number, it inherits the nearest parent node’s line number.
• Each node is assigned:
  • Its AST type (e.g., FunctionDef, ClassDef, Assign).
  • A label (e.g., function name, class name, or a sanitized version of the assigned target for assignments).
• A deterministic folder is created for each node with the structure: L{lineno}C{col_offset}T{traversal_index}__{node_type}__{node_label}

1. Label Sanitization:

• To ensure filesystem-safe and portable folder names, labels are strictly sanitized:
  • Only characters A-Z, a-z, 0-9, _, and - are allowed.
  • All other characters (including spaces, parentheses, brackets, symbols, and punctuation) are replaced with underscores (_).

1. Outputs Generated:

For each node folder, Parsnips creates one file:

### node_metadata.json

• Contains structured metadata for the AST node including:
  • type: The AST node type (e.g. FunctionDef).
  • label: The unsanitized label of the node.
  • text: The exact source text fragment for the node.
  • lineno: The node’s declared line number (or null if absent).
  • effective_lineno: The inherited line number after normalization.
  • col_offset: The column offset of the node.
  • file_swhid: The SWHID of the full source file.

1. Directory Structure:

• The extraction output for each processed file is always written into: {parent_dir}/.parsnips/parsnips__{file_stem}__{file_ext}/

1. Folder Sorting:

• Folders are deterministically named using the inherited line number, column offset, and traversal index.
• Sorting is lexicographic but naturally places nodes with lower line numbers first.

Fragment-Level SWHIDs

• The content SWHID for each node fragment is not generated during extraction.
• Instead, the node_metadata.json file itself becomes the fragment boundary.
• The fragment SWHID is computed as the blake2s() hash over the serialized node_metadata.json content.
• This ensures the SWHID includes both full structural metadata and the exact source text, not just the raw source text alone.

Fragment Identifier Semantics

Parsnips fragment-level SWHIDs are based on the full serialized metadata of each AST node — including both the exact source code fragment and its structural metadata. This means:

• The identifier uniquely represents a specific code fragment in its structural context.
• Any change to the source code, the AST structure, or even to the serialization format would change the SWHID.
• The SWHID does not solely identify the source text — it identifies the full parsed fragment as extracted.

This design enables reproducible, semantically-rich identifiers that reflect both the code content and its structure in the parse tree.

Citing Fragment-level SWHIDs with Context Qualifiers

Parsnip creates files for each node in an abstract parse tree so that code fragments like can be cited using SWHIDs. You can cite code fragments by citing their corresponding Parsnip node_metadata.json files.

Each node_metadata.json file corresponds to a node in the abstract syntax tree (AST). Like any other file, it can be cited directly using its SWHID. Like other content SWHIDs, it should cited with context qualifiers that describe its origin, anchor, and path.

You can use the Parsnip CLI to search for code fragments in AST nodes and their corresponding node SWHIDs. You can also discover the SWHID of the source code file which the node was parsed.

For example:

• Release SWHID (e.g., the SWHID for an annotated tag): swh:1:rel:abcdef1234567890abcdef1234567890abcdef12
• Unqualified Parsnip AST Node SWHID (identifies the node_metadata.json file that represents the AST node that contains the code fragment): swh:1:cnt:fedcba0987654321fedcba0987654321fedcba09
• Qualified Parsnip AST Node SWHID (e.g., the Python App class in main.py): swh:1:cnt:fedcba0987654321fedcba0987654321fedcba09;anchor=swh:1:rel:abcdef1234567890abcdef1234567890abcdef12;path=/src/.parsnips/parsnips__main__py/LOCOT1__Module__node/L1C0T2__ClassDef__App/node_metadata.json

Note Parsnips search currently only returns the Unqualified AST Node SWHID. The node_metadata.json files do not contain anchor qualifier information or path information relative to the anchor. In the future, the CLI will take parameters that will allow the search to create the qualified AST Node SWHID. So currently, authors must manually contrstruct the qualifiers.

CLI Parameters

Parsnips provides multiple command-line options:

| Argument | Description | | ---------- | ------------------------------------------------------------------------------------------- | | path (positional argument) | Path to Python file or directory to process (default: current directory) | | -c or --clean | Delete all .parsnips folders recursively | | -s or --search | Search using a regular expression inside the text field of all node_metadata.json files | | -n or --normalize-search | Apply Unicode normalization (NFC) to both the search pattern and node text before regex matching. This only applies to search operations and does not affect extraction. Extraction always preserves exact byte content for archival integrity. | | -q or --quiet | Suppress console output | | -l or --logfile | Write logs to specified JSON file | | --strict | Fail immediately on first error or missing .parsnips folder |

Example Usage

Process a directory recursively:

bash parsnips path/to/my/project

• This will recursively process all .py files under path/to/my/project, generate .parsnips/ directories for each directory containing Python files, and populate these directories with metadata files for the extracted fragment-level outputs.

Process a single file:

bash parsnips path/to/file.py

• This will process only the single file file.py and output results to:

bash path/to/.parsnips/parsnips__file__py/

Example Input file example.py:

```python class MyClass: def foo(self, x, y=(1, 2)): return x * y[0]

result = MyClass().foo(10) ```

Example Output structure:

bash .parsnips/parsnips__example__py/ L1C0T1__ClassDef__MyClass/ L2C4T2__FunctionDef__foo/ L3C8T3__Return__node/ L5C0T4__Assign__result_MyClass__foo_10/

Search for the SWHID of a code fragment by using regular expressions for the source code fragment

You can use Parsnips to search for the SWHID of a code fragment by using regular expressions that match on the text of the code fragment.

Example Search Usage

Suppose you have already extracted a directory and want to search for any AST nodes containing "hello world".

Run:

bash parsnips path/to/my/project --search "hello world"

Example output:

json { ".parsnips/parsnips__example__py/L8C4T7__Return__node/node_metadata.json": { "node_swhid": "swh:1:cnt:1234567890abcdef1234567890abcdef12345678", "node_metadata": { "type": "Return", "label": "node", "text": "return \"hello world\"", "lineno": 8, "effective_lineno": 8, "col_offset": 4, "file_swhid": "swh:1:cnt:abcdef1234567890abcdef1234567890abcdef12" } }, ".parsnips/parsnips__example__py/L9C0T8__Constant__hello_world/node_metadata.json": { "node_swhid": "swh:1:cnt:fedcba0987654321fedcba0987654321fedcba09", "node_metadata": { "type": "Constant", "label": "hello world", "text": "\"hello world\"", "lineno": 9, "effective_lineno": 9, "col_offset": 0, "file_swhid": "swh:1:cnt:abcdef1234567890abcdef1234567890abcdef12" } } }

If run with --strict mode and .parsnips folders are missing, the command will fail instead of silently returning empty results.

Licensing

Parsnips is licensed under the Apache License 2.0.

SPDX Identifier

Apache-2.0

Citation

If you use Parsnips in your research or projects, please cite it using the metadata provided in CITATION.cff.

Author

• Will Riley — wanderingwill@gmail.com

Related Projects

• Software Heritage
• SWHIDs specification

Disclaimer

Parsnips is an experimental tool designed to support research and archival workflows. Use at your own risk. Feedback and contributions are welcome.