GT AI Provider for OpenAI

A WordPress AI Client OpenAI provider with content-site enhancements. Fork of wordpress/ai-provider-for-openai, extended with:

• Latest-model-first defaults (GPT 5.4, GPT Image 2) with a UI to pick any other model from a live, cached dropdown
• A second OpenAI-compatible connector entry (Azure OpenAI, OpenRouter, LiteLLM, local LLM server) registered as a first-class entry on wp-admin/options-connectors.php
• 7 new WordPress Abilities content sites actually need — slop scanner, voice rewriter, internal-link suggester, FAQ-to-ACF-accordion generator, 16:9 featured-image generator, TTS, Rank Math meta sync
• Two one-click editor sidebar panels in the block editor
• Automatic rename of AI-generated image filenames so you never end up with ai-generated-image-1776823794.png again
• Cached 7-day model discovery, shared-transient HTTP layer

Requires WordPress 7.0+ (which ships the AI Client and Abilities API in core).

Installation

Manual

1. Download the latest release zip or clone this repo into wp-content/plugins/gt-ai-provider-for-openai/.
2. In wp-admin → Plugins, activate GT AI Provider for OpenAI.
3. Open wp-admin → Settings → Connectors and paste your OpenAI API key into the OpenAI connector.
4. (Optional) Open Settings → GT Custom OpenAI Instance to pick default models or wire a custom OpenAI-compatible endpoint.

Composer

composer require gauravtiwari/gt-ai-provider-for-openai:dev-main

What's inside

1. Smarter OpenAI provider defaults

• Prefers the latest GPT 5.x text model and the most recent gpt-image-2 snapshot when models are auto-discovered.
• Recognizes and sorts dated snapshots (gpt-image-2-2026-04-21) above their undated family (gpt-image-2).
• Expanded aspect-ratio metadata for gpt-image-* so the AI Client advertises 1:1, 3:2, 2:3, 7:4, 4:7, 16:9, 9:16, 2:1, 1:2 instead of three ratios.
• /v1/models response cached for 7 days via a single WordPress transient — cuts admin-page load time from 2-3s to near-zero.

2. Second OpenAI-compatible connector

Adds a second connector row on options-connectors.php called GT Custom OpenAI Instance. Use it to plug in any OpenAI-compatible endpoint:

The API key lives on the Connectors page, the base URL lives on Settings → GT Custom OpenAI Instance. Both can be overridden via wp-config.php constants or environment variables for infra-as-code setups.

3. Default-model picker

Settings → GT Custom OpenAI Instance adds three dropdowns populated from the live OpenAI model catalog (cached 7 days):

• Text generation model — default gpt-5.4. Used by gt/rewrite-for-voice and any code that reads apply_filters('gt_openai_default_text_model', '...').
• Image generation model — default gpt-image-2-2026-04-21. Used by gt/set-featured-image.
• Meta / summarization model — default gpt-4.1-mini. Used by gt/generate-faqs-accordion and any other short-form generation where cost matters.

A "Refresh now (clear 7-day cache)" button purges the transient and re-fetches.

4. New WordPress Abilities

All registered under the gt category. Invoke via /wp-json/wp-abilities/v1/abilities/{name}/run or programmatically via wp_get_ability('gt/...')->invoke($input).

5. Block editor sidebar panel

Under the GT AI Tools panel in the post sidebar:

• Generate Featured Image (16:9) — one-click invocation of gt/set-featured-image, sets the result as the featured image without reloading.
• Generate FAQs (8) — one-click invocation of gt/generate-faqs-accordion, inserts the accordion block at the end of the post.

Both buttons disable while in-flight, show spinners, and guard against running on unsaved posts.

6. Human-friendly AI-generated image filenames

The upstream ai/image-generation ability defaults its filename to ai-generated-image-<timestamp>. GT patches multiple points in the upload pipeline so any matching filename is rewritten to a title derived from (in order): attachment alt/excerpt → parent post's title → REST input.post_id → site name. Then a 6-char hex suffix keeps filenames unique.

Result: how-to-run-windows-apps-on-mac-a7k2df.png with post_title = "How To Run Windows Apps On Mac a7k2df" instead of ai-generated-image-1776823794.png.

Detection patterns (case-insensitive): ai-generated-image, ai-image, openai-image, dalle-*, gpt-image-*, img-<epoch>. Extend via standard WP filters if needed.

7. Content-site system instruction

A site-wide filter ai_client_default_system_instruction nudges any OpenAI call toward specific numbers over adjectives, named entities, no AI-slop phrases, contractions, 1-4 sentence paragraphs, American English.

Configuration

Constants (wp-config.php)

Filters

REST examples

Scan a post for AI-slop

curl -u "admin:app-password" \
  -X POST "https://example.com/wp-json/wp-abilities/v1/abilities/gt/stop-slop-scan/run" \
  -H "Content-Type: application/json" \
  -d '{"input":{"post_id":123}}'

Generate and set a featured image

curl -u "admin:app-password" \
  -X POST "https://example.com/wp-json/wp-abilities/v1/abilities/gt/set-featured-image/run" \
  -H "Content-Type: application/json" \
  -d '{"input":{"post_id":123}}'

Programmatic call from another plugin

$ability = wp_get_ability( 'gt/generate-faqs-accordion' );
$result  = $ability->invoke( array(
    'post_id' => 123,
    'count'   => 8,
) );

// $result['block'] contains the single-line acf/accordion block, ready to prepend/append.
// $result['faqs']  contains the raw { question, answer }[] array.

Architecture

gt-ai-provider-for-openai/
├── gt-ai-provider-for-openai.php       # Plugin manifest, constant defaults
├── src/
│   ├── autoload.php                    # PSR-4 for WordPress\OpenAiAiProvider\*
│   ├── Provider/
│   │   ├── OpenAiProvider.php          # Upstream base provider (unchanged signature)
│   │   └── GtCustomOpenAiProvider.php  # Extends base, overrides baseUrl() for custom instances
│   ├── Metadata/
│   │   └── OpenAiModelMetadataDirectory.php  # Patched: expanded aspect ratios, better sort
│   ├── Models/
│   │   ├── OpenAiTextGenerationModel.php
│   │   └── OpenAiImageGenerationModel.php
│   └── GT/
│       ├── Bootstrap.php               # Attachment rename, model cache, admin notice, default-model resolution
│       ├── Abilities.php               # 7 abilities + shared helpers
│       ├── Defaults.php                # Slop phrase list, stopwords, voice instruction
│       ├── CustomInstance.php          # Settings page + Custom connector registration
│       ├── ModelList.php               # 7-day transient cache for /v1/models
│       └── EditorSidebar.php           # Enqueues the block editor panel
└── assets/
    └── editor-sidebar.js               # Vanilla JS, no build step needed

Uninstall

Deactivating the plugin leaves the upstream OpenAI provider registered via WordPress core. To remove GT additions entirely, deactivate and delete the plugin in wp-admin → Plugins. The following options and transients are left in place for safety (delete manually if desired):

• connectors_ai_gt-custom-openai_api_key
• connectors_ai_gt_custom_openai_base_url
• gt_openai_default_text_model
• gt_openai_default_image_model
• gt_openai_default_meta_model
• gt_agent_readiness_version
• Transient gt_openai_model_list_v1

License

GPL-2.0-or-later. Fork of wordpress/ai-provider-for-openai (same license). See LICENSE.

Contributing

Issues and PRs welcome. See CHANGELOG.md for version history.