feat(azure): Support Responses API alongside Completions API (#1337)

[zjl4616]

Summary

Adds transparent dual-backend support to gptel-make-azure, enabling it to use the OpenAI Responses API for OpenAI models (GPT-4o, GPT-5, etc.) while falling back to the Completions API for non-OpenAI models — identical to how gptel-gh (GitHub Copilot) handles this.

Closes #1337.

Motivation

Azure OpenAI now serves OpenAI models that expect the Responses API, but also hosts non-OpenAI models (Llama, Mistral, etc.) that only support the Completions API. The current gptel-make-azure only supports Completions, causing errors with newer OpenAI models.

Implementation

Follows the exact pattern established by gptel-gh.el:

1. New struct gptel--azure — inherits from gptel-openai, adds a responses-backend field storing an embedded gptel-openai-responses instance.

2. 11 cl-defmethod dispatchers — each checks (gptel--model-capable-p 'responses-api gptel-model) and routes to either the responses backend or falls through to the parent completions method:

   • gptel-curl--parse-stream
   • gptel--parse-response
   • gptel--request-data
   • gptel--parse-schema
   • gptel--parse-tools
   • gptel--inject-tool-call
   • gptel--parse-tool-results
   • gptel--inject-prompt
   • gptel--parse-list
   • gptel--parse-buffer
   • gptel--inject-media

3. Modified gptel-make-azure — uses the new struct, creates an embedded responses backend, dynamically selects endpoint URL based on model capability.

Usage

Users mark models with responses-api capability:

(gptel-make-azure "Azure-OpenAI"
  :host "myresource.openai.azure.com"
  :endpoint "/openai/deployments/gpt-5/chat/completions?api-version=2025-04-01-preview"
  :key "..."
  :models '((gpt-5 :capabilities (responses-api))
            gpt-4o))

• gpt-5 → automatically uses Responses API (/openai/deployments/gpt-5/responses?api-version=...)
• gpt-4o → falls back to Completions API (existing behavior)

Backward Compatibility

✅ Fully backward compatible. Existing Azure configurations work unchanged. Models without responses-api capability continue using the Completions API exactly as before.

Changes

• gptel-openai.el: +124/-16 lines (single file change)
  • New gptel--azure struct definition
  • 11 dispatch methods
  • Refactored gptel-make-azure constructor

[karthink]

Thanks for the PR @zjl4616! Please move all the azure-specific code to gptel-openai-extras.el. (If we need even more azure-specific code in the future, we can move it to a dedicated gptel-azure.el feature.)