AI Inference in GitHub Actions

You can separate your instructions from your user prompt by using a system-prompt-file. This file contains the system instructions that guide the model's behavior. If both system-prompt (inline) and system-prompt-file are provided, the file takes precedence.

steps:
  - name: Run AI Inference with System Prompt File
    id: inference
    uses: actions/ai-inference@v1
    with:
      prompt: Hello
      system-prompt-file: ./.github/prompts/system.txt

When the prompt-file input points to a YAML file, the action uses a structured configuration. This allows you to define the model and template variables more robustly.

Variables used in the prompt can be injected using the input and file_input parameters. The action parses these and merges them into the PromptConfig used to build the messages sent to the provider.

The action produces the following outputs:

The action provides two ways to access the model's response:

1. response: The raw text of the model response.
2. response-file: The path to a temporary file containing the response text. This is useful for large responses that might exceed shell command length limits.

steps:
  - name: Run AI Inference
    id: inference
    uses: actions/ai-inference@v1
    with:
      prompt: Hello

  - name: Use Response File
    run: |
      echo "Response saved to: ${{ steps.inference.outputs.response-file }}"
      cat "${{ steps.inference.outputs.response-file }}"

The loadPromptFile function is the primary way to ingest a prompt configuration. It performs the following steps:

1. Reads the file from the provided filePath.
2. Parses the YAML content into a PromptConfig object.
3. Validates that the messages array exists and contains valid roles ('system', 'user', or 'assistant').
4. Performs template variable substitution on all message contents using the provided templateVariables.

Note: If no templateVariables are provided, it defaults to an empty object.

const config = loadPromptFile('./my-prompt.prompt.yml', { user: 'Bob' });
// config is now a fully parsed PromptConfig with variables replaced

The parseFileTemplateVariables function allows you to define template variables by mapping names to file paths in a YAML object. The function reads the content of each file and uses that content as the variable value.

Example YAML input:

context: ./docs/context.txt
user_data: ./data/user.json

Error Handling:

• Throws an error if the input is not a valid YAML object.
• Throws an error if a value is not a string (file path).
• Throws an error if a specified file does not exist.
• Throws an error if a file cannot be read.

const fileInput = "my_var: ./path/to/file.txt";
const variables = parseFileTemplateVariables(fileInput);
// variables: { my_var: '[contents of file.txt]' }

The copilotInference function accepts an optional spawner argument of type Spawner. This allows you to inject custom logic for how the CLI command is executed (e.g., for testing or using a different process execution library).

Type Definition: type Spawner = (cmd: string, args: string[]) => Promise<RunResult>

RunResult Shape:

interface RunResult {
  stdout: string
  stderr: string
  exitCode: number | null
}

type Spawner = (cmd: string, args: string[]) => Promise<{ stdout: string; stderr: string; exitCode: number | null }>;

// Example of a custom spawner
const mySpawner: Spawner = async (cmd, args) => {
  // custom execution logic
  return { stdout: 'mock response', stderr: '', exitCode: 0 };
};

await copilotInference(request, mySpawner);

You can inject dynamic data into your prompt files using the {{variable_name}} syntax. The replaceTemplateVariables function scans the text for these placeholders and replaces them with values provided in a TemplateVariables object.

If a placeholder is found in the text but is missing from the provided variables, a warning is issued and the original {{placeholder}} is left intact in the text.

const text = "Hello {{name}}, welcome to {{location}}!";
const variables = { name: "Alice", location: "Wonderland" };

const result = replaceTemplateVariables(text, variables);
// result: "Hello Alice, welcome to Wonderland!"