wingman

A Visual Studio Code extension with ChatGPT (3.5 and 4), LLaMa, or Claude integration with highly extensible and completely customizable prompting templates. These are your prompts, with a few built-in defaults to get you started.

Wingman is and always will be free and open source. If you're enjoying it and find value in it, please consider leaving a review on the VSCode Marketplace to help give it exposure.

The extension is an attempt to deliver a modular tool that supports a broad spectrum of use cases. If you have a use case that isn't supported and you think it might be a good fit for Wingman, please let me know by opening an issue.

A quick note on using LLaMa-based models as a completion source:

Support for LLaMa-based models assumes that you're using an endpoint that mimics the OpenAI API. LocalAI is a good example of a tool that does this. Set wingman.openai.apiBaseUrl to your API URL (e.g. http://localhost:1234/v1) and wingman.openai.model to your desired model (e.g. ggml-gpt4all-j). Make sure your command's provider is set to openai (it is by default), and you're all set.

• Demonstration
  • Comment-driven completion
  • Add documentation to what's selected
  • Refactor: decompose and modularize
  • Refactor: make it more functional
  • Refactor: input-driven modifications
  • Analysis: ask questions about your code
  • Translations
  • Custom prompts
• Quickstart
• Features
  • String interpolations
  • Command
    • Example commands
• Context
  • Text selection
  • Project text
    • Configuring context inclusion and exclusion
• FAQ

⚠️ Please note that this extension is currently under active development and its feature set is slowly evolving. As this happens, config property names may also be changed, and the way that commands are defined may also slightly change. I'll do my best to minimize this and preserve backward compatibility as much as possible.

Demonstration

The majority of the prompts below are built-in and can be used out-of-the-box. You can disable all of these by setting wingman.showBuiltinCommands to false in your settings.

It's really easy to create your own, and you are encouraged to do so.

Comment-driven completion

Add documentation to what's selected

Refactor: decompose and modularize

Refactor: make it more functional

Refactor: input-driven modifications

Select some code, click "Modify", and when prompted, enter the desired modifications. Wingman will replace the selected code with the newly generated code.

Analysis: ask questions about your code

Translations

Decide what language, framework, etc you want to translate the selected code into. In the below example, "python, using flask" is typed into the input field when prompted.

Try these:

• "typescript, using react and redux"
• "javascript, using vue and pinia"
• "go, using gin"
• "go, using martini and gorm"

Custom prompts

This is a custom prompt called "Reword README text". I use it when I'm drawing blanks. Like right now.

Quickstart

1. Install the extension.
2. Create an OpenAI account and get an API key (if you're using ChatGPT for generation).
3. Open the command palette (CTRL + SHIFT + P or CMD + SHIFT + P) and run Wingman: Set API key. Select openai as the provider and paste in your API key.

4. Open the bottom panel (CTRL + J or CMD + J) and select the Wingman tab.
5. In the Wingman panel, expand the "Analysis" category, highlight a block of code and click "Analyze for bugs".
6. Explore all of the other builtin commands.
7. Create your own commands that compliment your workflow in settings under wingman.userCommands. See the Command section for more details.

Features

• User-defined commands - Easily create your own commands with custom prompt templates.

• Language-specific elaboration - Use VSCode's language identifier to define language-specific elaboration for your prompts. Add {{language_instructions}} to your templates, and then define the instructions for each language in your command's languageInstructions property. Here's an example:

  {
    "label": "Write documentation",
    "userMessageTemplate": "I have the following {{language}} code:\n```{{filetype}}\n{{text_selection}}\n```\n\nWrite really good documentation using best practices for the given language. Attention paid to documenting parameters, return types, any exceptions or errors. Don't change the code. {{language_instructions}} IMPORTANT: Only return the code inside of a code fence and nothing else.",
    "languageInstructions": {
      "cpp": "Use doxygen style comments for functions.",
      "java": "Use javadoc style comments for functions.",
      "typescript": "Use TSDoc style comments for functions.",
      "javascript": "Use JSDoc style comments for functions."
    },
    "callbackType": "replace",
    "provider": "openai",
    "category": "Documentation"
  }
  

• Continue the conversation - Conversation state is preserved by default until a new command is issued. This gives you the opportunity to ask follow-up questions or fine-tune a reply.

• Elaboration/additional context - If your command includes {{input}} in its template, you will be prompted for elaboration on the command, and {{input}} will be replaced with your input. This is useful for commands that require additional input, e.g. the included "Modify" or "Translate" commands. You can specify additional text for the prompt using the syntax: {{input:Text for the prompt}}. For example, the "Translate" command:

• Configurable API url - Set either openai.apiBaseUrl or anthropic.apiBaseUrl to an atlernative base URL. This is particularly useful if you're using something like https://github.com/go-skynet/LocalAI, i.e. you want your wingman to be driven by a local LLaMa model.
• Configurable model - Set either openai.model or anthropic.model to your desired model. These are string fields to allow for more flexibility, e.g. if you're using LocalAI and want to use a custom model like ggml-gpt4all-j.
• String interpolation - Use {{language}}, {{input}}, {{text_selection}}, {{project_text}}, {{filetype}}, and {{language_instructions}} in your templates (userMessageTemplate and systemMessageTemplate supported) to automatically fill in values.

String interpolations

The table below lists some of the interpolations that you may have noticed in the example above. Wingman automatically replaces these interpolations for you, and the values they are replaced with are outlined in the table below. These interpolations can be utilized in both userMessageTemplate and systemMessageTemplate.

Command

You can create your own commands by adding them to your settings under wingman.userCommands. Your commands need to implement the Command interface. Here's what the default, base command looks like:

{
  // Required. Human readable label that appears in the UI.
  label: "Default",
  // Required. Your prompt, including any interpolations you need.
  userMessageTemplate: "",

  // Optional.
  systemMessageTemplate: "You are a {{language}} coding assistant.",
  // Optional.
  numberOfChoices: 1,
  // Optional.
  model: "gpt-3.5-turbo",
  // Optional.
  temperature: 0.3,
  // Optional.
  maxTokens: 4096,
  // Optional.
  // This determins what to do with the code that the model generates.
  // Possible values:
  // - "none": Do nothing - just show the reply in the conversation window.
  // - "buffer": Put the entire reply in a new window.
  // - "replace": Replace the originally selected text with the first code block of the reply.
  // - "beforeSelected": Places the text from the first code block of the reply before the originally selected text.
  // - "afterSelected": Places the text from the first code block of the reply after the originally selected text.
  callbackType: "none",
  // Optional.
  // Used to replace {{language_instructions}} interpolation.
  // e.g. If the language ID of the active editor is "typescript",
  // then {{language_instructions}} is replaced with languageInstructions.typescript
  languageInstructions: {
    javascript: "Use modern JavaScript syntax.",
    typescript: "Use modern TypeScript syntax.",
    cpp: "Use modern C++ features.",
    html: "Use modern HTML syntax.",
    csharp: "Use modern C# syntax.",
  },
  // Optional. Decides what category in the UI that this should appear under.
  category: "Misc",
  // Optional.
  // Possible values: "openai", "anthropic"
  provider: "openai",
};

When you create your own command, you can override any of these properties. The only required properties are command, label, and userMessageTemplate.

Example commands

Here are some more example commands.

// settings.json

"wingman.userCommands": [
  {
    "label": "Meaningful variable names",
    "userMessageTemplate":
      "I have the following {{language}} code:\n```{{filetype}}\n{{text_selection}}\n```\n\nGive the variables more meaningful names based on their usage. Return the refactored code inside of a code block and nothing else.",
    "callbackType": "replace",
    "category": "Refactoring",
    "provider": "anthropic",
  },
  {
    "label": "Decompose, modularize",
    "userMessageTemplate":
      "I have the following {{language}} code:\n```{{filetype}}\n{{text_selection}}\n```\n\nDecompose it by splitting functions, reducing responsibilities and enhancing modularity. Return the refactored code inside of a code block and nothing else.",
    "callbackType": "replace",
    "category": "Refactoring",
    "provider": "openai",
  },
  {
    "label": "Comment",
    "userMessageTemplate":
      "I have the following {{language}} code:\n```{{filetype}}\n{{text_selection}}\n```\n\nWrite really good comments using best practices for the given language. Attention paid to documenting parameters, return types, any exceptions or errors. Don't change the code. Return only the comment inside of a code block and nothing else.",
    "callbackType": "beforeSelected",
    "category": "Comments",
    "provider": "anthropic",
  }
]

Context

Context is included in a userMessageTemplate by using either {{text_selection}} or {{project_text}}.

Text selection

{{text_selection}} is a string interpolation that includes the currently selected text.

If you selected const x = 1; and your prompt was written as:

I have the following code:\n\n{{text_selection}}\n\nWhat's wrong with it?

Your final prompt would be:

I have the following code:\n\nconst x = 1;\n\nWhat's wrong with it?

Project text

{{project_text}} is a string interpolation that includes the contents of all files in the current vscode workspace root. Each file is stringified and formatted as:

// <filename>\n```<filetype>\n<file contents>\n```

If you had a file named index.js with the contents:

import { add } from "./common/utils.js";

And a file named common/utils.js with the contents:

export const add = (a, b) => a + b;

And your command prompt was:

Suggest improvements to my {{language}} project:\n\n{{project_text}}

Your final prompt would be:

Suggest improvements to my javascript project:

// file: src/index.js
\`\`\`javascript     <-- only escaped for the purposes of this README
import { add } from "./utils.js";
\`\`\`

// file: src/common/utils.js
\`\`\`javascript
export const add = (a, b) => a + b;
\`\`\`

This type of context inclusion can very quickly get out of hand, and you likely don't want to send tens of thousands of lines of code to whatever model you're using (at least not yet in 2023). For this reason, you are highly encouraged to smartly configure the options below with values that make sense for the project you are working on.

Sensible defaults are provided for each of these configuration options, but these may not be applicable to all.

In most cases, the only option you'll be changing is wingman.context.include.permittedFileExtensions, as the other two likely won't change project-to-project.

If you wanted to ask your project a question only about the CSS styling of your app, then you might want to set wingman.context.include.permittedFileExtensions to ["css"] before running your projectQuestion command. To set it back to its default value, just remove the entry from your settings.json entirely.

Configuring context inclusion and exclusion

If a .wmignore file is discovered, it's treated the same as a .gitignore file.

FAQ

Q: How do I remove or change my API key? A: Run the Wingman: Set API key command again, select the relevant provider. Paste in your new key, or leave it blank to remove it. Press enter.