Log resolved model/config when using .prompt.yml

[mingzaily]

Summary

When actions/ai-inference runs in structured .prompt.yml mode, the workflow log still shows the action input defaults in the with: block, for example:

• model: openai/gpt-4o
• system-prompt: You are a helpful assistant
• max-tokens: 200

However, the implementation later loads the .prompt.yml file and resolves the effective model and model parameters from promptConfig.

That means a run can actually use values from the prompt file while the visible log output strongly suggests the defaults are in effect.

Why this is confusing

From the repo:

• action.yml defines defaults for model, system-prompt, and max-tokens
• dist/index.js then does:
  • const modelName = promptConfig?.model || core.getInput('model')
  • const maxCompletionTokensInput = promptConfig?.modelParameters?.maxCompletionTokens ?? core.getInput('max-completion-tokens')
  • const maxTokensInput = promptConfig?.modelParameters?.maxTokens ?? core.getInput('max-tokens')
  • const temperature = promptConfig?.modelParameters?.temperature ?? ...

So behavior is correct, but observability is misleading.

Reproduction pattern

Use a .prompt.yml file with values such as:

model: openai/gpt-4.1
responseFormat: json_schema
modelParameters:
  temperature: 0.0
  maxCompletionTokens: 900

and call the action like this:

- uses: actions/ai-inference@v1
  with:
    prompt-file: ./.github/prompts/sample.prompt.yml
    input: |
      foo: bar

The log shows:

• Using prompt YAML file format
• but the earlier with: block still displays the input defaults (openai/gpt-4o, You are a helpful assistant, max-tokens: 200)

This makes it hard to tell which model/config was actually used.

Expected behavior

One of these would solve it:

1. Add an explicit log section after loading .prompt.yml, for example:
   • resolved model
   • resolved response format
   • resolved max completion tokens / max tokens
   • resolved temperature / topP
2. Or document clearly in the README that the with: block shows action input defaults and is not the final resolved config in .prompt.yml mode.
3. Ideally both.

Notes

I do not think the inference behavior is wrong here. This looks like a logging / diagnostics gap specific to .prompt.yml mode.