Semi-automatic Snakemake workflow generation from unstructured Bash commands or Python Notebooks.

If you use Snakemaker in your work, please cite:

Marco Masera, Alessandro Leone, Johannes Köster, Ivan Molineris.
Snakemaker: Seamlessly transforming ad-hoc analyses into sustainable Snakemake workflows with generative AI. arXiv preprint arXiv:2505.02841, 2025.

@article{masera2025snakemaker,
  title   = {Snakemaker: Seamlessly transforming ad-hoc analyses into sustainable Snakemake workflows with generative AI},
  author  = {Marco Masera, Alessandro Leone, Johannes Köster, Ivan Molineris},
  journal = {arXiv preprint arXiv:2505.02841},
  year    = {2025},
  url     = {https://arxiv.org/abs/2505.02841}
}

• Install Snakemaker
• Setting up for first time
• Bash Command Support
• Notebook Support

The simplest way to install Snakemaker is through the VSCode Extension Marketplace. The official Snakemaker page can be found at this link or by typing snakemaker into VSCode extension tab.

Snakemaker can be built and installed from the source code. A requirement for local build is NodeJS.

• Clone the repository: git clone https://github.com/molinerisLab/snkmaker.git
• Set the desired version: open "package.json", modify field "version"
• Build: npx @vscode/vsce -- package. This command produces a *.vsix file.
• Install: from VSCode, open the "Extension" sidebar, click on the three dots on the top-right corner of the sidebar, click Install from VSIX.

In order to work with Snakemaker, it's necessary to connect it to a LLM provider. There are two ways to connect it with some LLM:

• Use Github Copilot via VSCode language API (recommended)
• Manually connect remote or local LLM via OpenAI-compatible APIs.

Github Copilot allows users to access most state-of-the-art models, and provides APIs that extensions as Snakemaker can use to directly connect to it. Setting up Copilot inside Snakemaker is straightforward:

In order to work with it, Copilot must be active inside your VSCode.

• Follow the official guidelines to install and enable Github Copilot.
• Ensure Copilot is working: try to open copilot chat and verify the models are available.

Once Github Copilot is installed and available, it should be visible from the Snakemaker model tabs:

• Open Snakemaker view in the VSCode left-side bar.
• Expand the models section of the Snakemaker view.
• A list of models labelles as (copilot) should be visible. The models can be activated with a double click. If the list of models is empty:: Snakemaker cannot communicate with VSCode language APIs. This sometimes happens even if Copilot is correctly installed. In order to fix it:
• Ensure Copilot is actually available - test it using the standard Copilot chat.
• Close and re-open VSCode, go back to the models section of Snakemaker.

Snakemaker also supports external models via OpenAI-compatible APIs, supported by most vendors (OpenAI, Gemini, Openrouter) and local deploy tools (Ollama). In order to connect Snakemaker to an external LLM:

• Open the Snakemaker view in the left-side Vscode panel.
• Locate the models section of Snakemaker view.
• Click "Add Model".
• Provide the model's URL.
• Provide the model's API key, or leave it empty if the LLM server requires no key (i.e. locally hosted models).
• Provide the model's name, as required by the OpenAi APIs.
• Provide the maximum tokens the model can produce in a single request.

Snakemaker provides a custom assistant inside the GitHub Copilot Chat. Tag @snakemaker to access the assistant and get help with the extension.

The Notebook feature has a separate assistant, which can be reached with the tag @snakemaker-notebook.

Turn on/off listening and recording of bash commands manually.

Run Manually add commands to Snakemaker history to manually add commands to the history.

Snakemaker distinguishes between important commands, which can contribute to the Snakemake workflow, and unimportant, one-timer commands.

Non-important commands will be shown in the Snakemaker panel in a dark-gray color, and by default they are not exported as rules. Importance of a command can be changed manually.

Snakemaker extracts details like input/output files and possible rule names. These can be edited manually.

Merge multiple commands into composite commands using drag-and-drop.

Snakemaker can also generate Make rules. The user can switch between Snakemake and Make rules generation by searching for "Rules output format" in the VSCode settings. Alternatively, ask @snakemaker in the chat to open the setting for you.

For the Snakemake rules, some additional options are offered in the settings:

• Log directive: whether to include the log directive in the rules. Log directive is recommended in the Snakemake best practices and is on by default.
• Prefer generic rules: when exporting multiple rules, generic rules with wildcards are preferred. This can be turned off in the settings.

Settings related to Snakemake rule generation are grouped under "Snakemake Best Practices" in the VSCode settings.

Generated Snakemake rules are checked for errors, and fed back to the model with the error message for correction. This feature makes rule generation more reliable, but can slow down the process and consume more tokens of the LLM API. It can be disabled in the settings: Snakemaker: Validate Snakemake Rules.

In order for automatic correction to work, a path to snakemake must be provided in the settings, or snakemake must be on the user's $PATH.

• Open the VSCode settings with Ctrl + ,
• Search for Snakemaker: Snakemake Absolute Path
• Provide the absolute path to the snakemake executable
• Leave empty to use the $PATH

The Snakemaker custom assistant integrated in the GitHub Copilot Chat can assist in many ways during the workflow, providing help and performing operations on behalf of the user. Tag @snakemaker in the Copilot Chat to access the assistant and get help with the extension.

The direct chat can be used for a variety of purposes:

• Retrieve information about the usage of Snakemaker, troubleshooting, understand the principles of the extension.
  • What can Snakemaker do for me?
  • Why aren't my commands being recorded?
  • How do I export my workspace?
• Flexible rule generation.
  • Can you generate the rule for my last command using wildcards instead of fixed filenames?
  • Set a ruleorder between my two commands that write to "output.txt".
• Queries related to your own history:
  • When did I create the file named "columns.csv"?
  • Do I have multiple commands generating the same file?
• Automatically re-organize your history:
  • Can you set all the commands writing to "output" as important?
  • Can you turn all the commands writing to "output" into a single composite command?

The workspace contains all the recorded commands and their details. By default, the workspace is preserved between VSCode sessions, an option that can be disabled in the settings ("Keep history between sessions").

Explicit import and export of the workspace to a JSON file can be done:

• Open the VSCode command palette with Ctrl + Shift + P
• Search for "Snakemaker: Load Workspace" or "Snakemaker: Save Workspace"

Open the notebook, select More actions (three dots), and click Process with Snakemaker.

The first step in the conversion process involves resolving the data dependencies between the cells of the notebook into discrete, file-based rules.

Each cell is characterized by three sets of variables:

• Read: variables that the cell (might) read from other cells.
• Write: variables that the cell (might) write to be used by other cells.
• Wildcards: variables that the cell (might) read that will be provided as Wildcards.

From these sets, the dependency graph is constructed by matching each Read variable of a cell with the closest Write of this variable in the previous cells. Wildcards on the other hand will be mathed to patterns in the filenames during the next step, where actual rules will be generated.

In the image below, for example, Cell 1 first reads the variable var1, which is received from some previous cell, and then modifies it. Cell 2 also reads var1, but it is the modified version produced by Cell 1.

Each Cell will become either a Snakemake rule or a script that will be imported by other scripts or rules. The user can split, merge or remove cells to better fit the workflow.

Dependencies can be added by selecting a variable in the code and clicking on the context buttons that appear.

Similarly, variables can be removed from the sets by clicking on the X buttons displayed next to them under the cell's code.

Under each cell, buttons are provided to set its state as a Rule or a Script, with an additional option Undecided, indicating cells for which a decision hasn't been taken yet.

Cells' states are constrained by the states of the cells they depend on. In particular:

• A Rule can read output files from other rules and can import scripts, so it can depend on both Rules and Scripts. This means, every cell can become a Rule.
• A Script can import other scripts, but can not directly read output files from another rule, so it can depend only on Scripts. This means, setting a cell as a Rule might have a cascading effect on other cells set as Scripts, forcing them to become Rules as well.

In order to continue to the next step, all dependencies must be resolved, and all cells must be set either as a Rule or a Script. Unresolved dependencies and undecided cells are shown at the top of the panel.

In the second step, Snakemaker uses the information provided in the first step to automatically generate the Snakemake workflow. This process involves code generation.

• A prefix code is generated for each cell. This code manages the import statements, the reading of command line arguments, and the reading of input files.
• A suffix code is generated for the Rules. This code manages the writing of output files.
• Snakemake rules are generated for each cell set as a Rule.

In this step the user can review the generated code, and make modifications to it.

The user can modify the generated code on the cells as in a regular editor. Changes are fed to the LLM, which will propagate them to the following cells, if necessary.

For example, if the user modifies the suffix of a cell to write the output file with a different format, the LLM will automatically:

• Update the Snakemake rule with the correct output filename.
• Update the Snakemake rule and prefix code of the cells that read this output file.

When the user is satisfied with the generated code, he can export the workflow. A directory where to save the Snakefile and the scripts will be asked.

The process described above can be performed manually using the GUI. However, a more efficient way is to use the Snakemaker-Notebook assistant, which can be reached by tagging @snakemaker-notebook in the Copilot Chat.

The assistant can answer prompts regarding the current state of the process, identify issues, provide suggestions to fix them or fix them directly, and perform batch operations from natural language, such as changing the output format of all the rules, or the output directory, set outputs as temp etc. The user is encouraged to engage with the chat agent to get help with the entire process.

This project is financed by the Essential Open Source Software for Science Program of the Chan Zuckerberg Initiative. https://chanzuckerberg.com/eoss/