Enhance CLI interactivity for LLM/automation scenarios

[spboyer]

Problem Statement

LLMs (GitHub Copilot, Claude, etc.) and automation tools running azd commands like init, up, provision encounter interactive prompts that block execution. The current --no-prompt flag only uses default values or fails when no default exists—it does not allow callers to provide answers programmatically.

Current Limitations

When an LLM runs azd init or azd provision:

1. Prompts block execution - The CLI waits for user input that LLMs cannot provide through normal terminal interaction
2. --no-prompt is insufficient - It only accepts defaults or fails; there is no mechanism to supply specific answers
3. External prompt service is complex - The AZD_UI_PROMPT_ENDPOINT mechanism requires setting up an HTTP server, which is impractical for simple LLM automation

Affected Commands

• azd init - Template selection, environment naming, service configuration
• azd provision - Subscription selection, location selection, parameter inputs
• azd up - Combines init + provision prompts
• azd down - Confirmation prompts (partially addressed by --force)
• Any command with IaC parameter prompts

Proposed Solution

Implement complementary mechanisms for LLM-friendly interactivity:

1. JSON Prompt Mode (--json-prompts)

When enabled, prompts output structured JSON to stdout and read JSON responses from stdin, enabling real-time LLM interaction:

Prompt Output (stdout):

{
  "type": "prompt",
  "prompt": {
    "id": "subscription",
    "kind": "select",
    "message": "Select an Azure subscription",
    "choices": [
      {"value": "sub-123", "detail": "My Subscription 1"},
      {"value": "sub-456", "detail": "My Subscription 2"}
    ],
    "defaultValue": "sub-123"
  }
}

Response Input (stdin):

{"id": "subscription", "value": "sub-456"}

2. Upfront Answers Flag (--answers)

Accept pre-defined answers as JSON, allowing all inputs to be provided before execution:

# Inline JSON
azd init --answers '{"template":"todo-nodejs-mongo","environment":"dev"}'

# File reference  
azd provision --answers @./answers.json

3. Input Introspection (--describe-inputs) [Future]

List all possible prompts a command might ask before running:

azd init --describe-inputs --output json

Benefits

• LLM Compatibility: Enables GitHub Copilot, Claude, and other LLMs to run azd commands interactively
• CI/CD Flexibility: Provides more control than --no-prompt for complex automation scenarios
• Backwards Compatible: Additive changes that do not affect existing --no-prompt behavior
• Aligned with Existing Patterns: JSON format matches the existing external prompt API specification

Technical Notes

• Implementation builds on existing Console abstraction in pkg/input/console.go
• JSON schema compatible with AZD_UI_PROMPT_ENDPOINT external prompt API
• Prompt types supported: string, password, select, multiSelect, confirm, directory

Related

• External prompting documentation: cli/azd/docs/external-prompting.md
• MCP server support: cli/azd/internal/mcp/

[spboyer]

As a note here is an example of the Copilot CLI trying to do azd up on a project and just spins on "Read Shell" for 30+ minutes until user kills session. It was waiting for
Enter unique environment name: