| @@ -1,357 +1,960 @@ | ||
| 1 | -<?php | |
| 2 | -/** | |
| 3 | - * MxChat Model Catalog — single source of truth for chat + embedding models. | |
| 4 | - * | |
| 5 | - * Plan: plan-mxchat-20260527-d14e89. Consolidates model definitions that | |
| 6 | - * previously lived in three places (Settings dropdown in | |
| 7 | - * class-mxchat-admin.php, autosave allowlist in admin/class-ajax-handler.php, | |
| 8 | - * modal picker catalog in js/mxchat-admin.js, plus a fourth in | |
| 9 | - * admin-onboarding-page.php). Adding a new chat/embedding model now means | |
| 10 | - * editing THIS file only. | |
| 11 | - * | |
| 12 | - * Schema: | |
| 13 | - * chat_models() / embedding_models() return: | |
| 14 | - * [ providerSlug => [ | |
| 15 | - * 'label' => Display name shown in pickers, | |
| 16 | - * 'key_option' => sub-key under get_option('mxchat_options'), | |
| 17 | - * 'requires_key_to_load_models' => bool — OpenRouter is true (key needed | |
| 18 | - * before model list can be fetched). Surfaces consult | |
| 19 | - * this to disable / gate the provider in pickers that | |
| 20 | - * can't run the inline key-then-fetch flow. | |
| 21 | - * 'models' => [ model_id => [ 'label' => …, 'description' => … ] ] | |
| 22 | - * ] | |
| 23 | - * ] | |
| 24 | - * | |
| 25 | - * @package MxChat | |
| 26 | - * @since 3.2.7 | |
| 27 | - */ | |
| 28 | - | |
| 29 | -if (!defined('ABSPATH')) { | |
| 30 | - exit; | |
| 31 | -} | |
| 32 | - | |
| 33 | -class MxChat_Model_Catalog { | |
| 34 | - | |
| 35 | - /** | |
| 36 | - * Chat models grouped by provider. Order within a provider matches the | |
| 37 | - * order shown in the Settings modal picker — keep the recommended / | |
| 38 | - * flagship model at top. | |
| 39 | - */ | |
| 40 | - public static function chat_models() { | |
| 41 | - return array( | |
| 42 | - 'openrouter' => array( | |
| 43 | - 'label' => __('OpenRouter', 'mxchat'), | |
| 44 | - 'key_option' => 'openrouter_api_key', | |
| 45 | - 'requires_key_to_load_models' => true, | |
| 46 | - 'disable_in_onboarding' => true, | |
| 47 | - 'models' => array( | |
| 48 | - 'openrouter' => array( | |
| 49 | - 'label' => 'OpenRouter', | |
| 50 | - 'description' => __('Access 100+ models from multiple providers (add API key to browse)', 'mxchat'), | |
| 51 | - ), | |
| 52 | - ), | |
| 53 | - ), | |
| 54 | - 'gemini' => array( | |
| 55 | - 'label' => __('Google Gemini', 'mxchat'), | |
| 56 | - 'key_option' => 'gemini_api_key', | |
| 57 | - 'requires_key_to_load_models' => false, | |
| 58 | - 'models' => array( | |
| 59 | - 'gemini-3.5-flash' => array('label' => 'Gemini 3.5 Flash', 'description' => __('Stable — newest Flash generation, recommended default', 'mxchat')), | |
| 60 | - 'gemini-3.1-pro-preview' => array('label' => 'Gemini 3.1 Pro', 'description' => __('Preview — most intelligent, multimodal & agentic', 'mxchat')), | |
| 61 | - 'gemini-3-flash-preview' => array('label' => 'Gemini 3 Flash', 'description' => __('Preview — balanced speed and scale', 'mxchat')), | |
| 62 | - 'gemini-3.1-flash-lite' => array('label' => 'Gemini 3.1 Flash-Lite', 'description' => __('Stable — cost-efficient, high throughput', 'mxchat')), | |
| 63 | - 'gemini-3.1-flash-lite-preview' => array('label' => 'Gemini 3.1 Flash-Lite (Preview)', 'description' => __('Preview — latest cost-efficient model', 'mxchat')), | |
| 64 | - 'gemini-2.5-pro' => array('label' => 'Gemini 2.5 Pro', 'description' => __('Stable — advanced thinking, code, math & long context', 'mxchat')), | |
| 65 | - 'gemini-2.5-flash' => array('label' => 'Gemini 2.5 Flash', 'description' => __('Stable — strong price-performance with thinking', 'mxchat')), | |
| 66 | - 'gemini-2.5-flash-lite' => array('label' => 'Gemini 2.5 Flash-Lite', 'description' => __('Stable — ultra fast and low cost', 'mxchat')), | |
| 67 | - ), | |
| 68 | - ), | |
| 69 | - 'openai' => array( | |
| 70 | - 'label' => __('OpenAI', 'mxchat'), | |
| 71 | - 'key_option' => 'api_key', | |
| 72 | - 'requires_key_to_load_models' => false, | |
| 73 | - 'models' => array( | |
| 74 | - 'gpt-5.5' => array('label' => 'GPT-5.5', 'description' => __('Latest Flagship — newest OpenAI reasoning and coding model', 'mxchat')), | |
| 75 | - 'gpt-5.4' => array('label' => 'GPT-5.4', 'description' => __('Flagship reasoning and coding model', 'mxchat')), | |
| 76 | - 'gpt-5.4-mini' => array('label' => 'GPT-5.4 Mini', 'description' => __('Fast and affordable with 400K context', 'mxchat')), | |
| 77 | - 'gpt-5.4-nano' => array('label' => 'GPT-5.4 Nano', 'description' => __('Fastest and cheapest for lightweight tasks', 'mxchat')), | |
| 78 | - 'gpt-5.3-chat-latest' => array('label' => 'GPT-5.3 Chat', 'description' => __('Optimized for natural conversations with reduced hallucinations', 'mxchat')), | |
| 79 | - 'gpt-5.2' => array('label' => 'GPT-5.2', 'description' => __('Best general-purpose & agentic model with fast responses', 'mxchat')), | |
| 80 | - 'gpt-5.1-chat-latest' => array('label' => 'GPT-5.1 Chat Latest', 'description' => __('Recommended for most use cases', 'mxchat')), | |
| 81 | - 'gpt-5.1-2025-11-13' => array('label' => 'GPT-5.1', 'description' => __('Flagship for coding & agentic tasks with low reasoning (400K context)', 'mxchat')), | |
| 82 | - 'gpt-5' => array('label' => 'GPT-5', 'description' => __('Flagship for coding, reasoning, and agentic tasks across domains', 'mxchat')), | |
| 83 | - 'gpt-5-mini' => array('label' => 'GPT-5 Mini', 'description' => __('Fast and lightweight', 'mxchat')), | |
| 84 | - 'gpt-5-nano' => array('label' => 'GPT-5 Nano', 'description' => __('Fastest and cheapest; ideal for summarization and classification', 'mxchat')), | |
| 85 | - ), | |
| 86 | - ), | |
| 87 | - 'claude' => array( | |
| 88 | - 'label' => __('Anthropic Claude', 'mxchat'), | |
| 89 | - 'key_option' => 'claude_api_key', | |
| 90 | - 'requires_key_to_load_models' => false, | |
| 91 | - 'models' => array( | |
| 92 | - 'claude-fable-5' => array('label' => 'Claude Fable 5', 'description' => __('Latest Flagship — newest and most capable Anthropic model', 'mxchat')), | |
| 93 | - 'claude-opus-4-8' => array('label' => 'Claude Opus 4.8', 'description' => __('Previous flagship — most capable Opus-tier model', 'mxchat')), | |
| 94 | - 'claude-opus-4-7' => array('label' => 'Claude Opus 4.7', 'description' => __('Previous Anthropic flagship model', 'mxchat')), | |
| 95 | - 'claude-opus-4-6' => array('label' => 'Claude Opus 4.6', 'description' => __('Most capable Claude model - recommended', 'mxchat')), | |
| 96 | - 'claude-sonnet-4-6' => array('label' => 'Claude Sonnet 4.6', 'description' => __('Latest Sonnet - excellent balance of speed and capability', 'mxchat')), | |
| 97 | - 'claude-opus-4-5' => array('label' => 'Claude Opus 4.5', 'description' => __('Highly capable for complex tasks', 'mxchat')), | |
| 98 | - 'claude-sonnet-4-5-20250929' => array('label' => 'Claude Sonnet 4.5', 'description' => __('Best for complex agents and coding', 'mxchat')), | |
| 99 | - 'claude-opus-4-1-20250805' => array('label' => 'Claude Opus 4.1', 'description' => __('Exceptional for specialized complex tasks', 'mxchat')), | |
| 100 | - 'claude-haiku-4-5-20251001' => array('label' => 'Claude Haiku 4.5', 'description' => __('Fastest and most intelligent Haiku', 'mxchat')), | |
| 101 | - ), | |
| 102 | - ), | |
| 103 | - 'xai' => array( | |
| 104 | - 'label' => __('xAI Grok', 'mxchat'), | |
| 105 | - 'key_option' => 'xai_api_key', | |
| 106 | - 'requires_key_to_load_models' => false, | |
| 107 | - 'models' => array( | |
| 108 | - 'grok-4-1-fast-reasoning' => array('label' => 'Grok 4.1 Fast (Reasoning)', 'description' => __('2M context window and reasoning', 'mxchat')), | |
| 109 | - 'grok-4-1-fast-non-reasoning' => array('label' => 'Grok 4.1 Fast (Non-Reasoning)', 'description' => __('2M context window and faster responses', 'mxchat')), | |
| 110 | - 'grok-4-0709' => array('label' => 'Grok 4', 'description' => __('Latest flagship model - unparalleled performance in natural language, math and reasoning', 'mxchat')), | |
| 111 | - 'grok-3-beta' => array('label' => 'Grok-3', 'description' => __('Powerful model with 131K context', 'mxchat')), | |
| 112 | - 'grok-3-fast-beta' => array('label' => 'Grok-3 Fast', 'description' => __('High performance with faster responses', 'mxchat')), | |
| 113 | - 'grok-3-mini-beta' => array('label' => 'Grok-3 Mini', 'description' => __('Affordable model with good performance', 'mxchat')), | |
| 114 | - 'grok-3-mini-fast-beta' => array('label' => 'Grok-3 Mini Fast', 'description' => __('Quick and cost-effective', 'mxchat')), | |
| 115 | - ), | |
| 116 | - ), | |
| 117 | - 'deepseek' => array( | |
| 118 | - 'label' => __('DeepSeek', 'mxchat'), | |
| 119 | - 'key_option' => 'deepseek_api_key', | |
| 120 | - 'requires_key_to_load_models' => false, | |
| 121 | - 'models' => array( | |
| 122 | - 'deepseek-chat' => array('label' => 'DeepSeek-V3', 'description' => __('Advanced AI assistant', 'mxchat')), | |
| 123 | - ), | |
| 124 | - ), | |
| 125 | - 'custom' => array( | |
| 126 | - 'label' => __('Custom (OpenAI-compatible)', 'mxchat'), | |
| 127 | - 'key_option' => 'custom_provider_api_key', | |
| 128 | - 'requires_key_to_load_models' => false, | |
| 129 | - 'disable_in_onboarding' => true, | |
| 130 | - 'models' => array( | |
| 131 | - 'custom-provider' => array( | |
| 132 | - 'label' => 'Custom Provider', | |
| 133 | - 'description' => __('OpenAI-compatible local LLM (Ollama, LM Studio, vLLM, llama.cpp, Azure OpenAI) — configure Base URL, key, and model in API Keys tab', 'mxchat'), | |
| 134 | - ), | |
| 135 | - ), | |
| 136 | - ), | |
| 137 | - ); | |
| 138 | - } | |
| 139 | - | |
| 140 | - /** | |
| 141 | - * Embedding models grouped by provider. Order within a provider matches | |
| 142 | - * the order shown in the Settings dropdown. | |
| 143 | - */ | |
| 144 | - public static function embedding_models() { | |
| 145 | - return array( | |
| 146 | - 'openai' => array( | |
| 147 | - 'label' => __('OpenAI', 'mxchat'), | |
| 148 | - 'key_option' => 'api_key', | |
| 149 | - 'models' => array( | |
| 150 | - 'text-embedding-ada-002' => array('label' => 'Ada 2', 'description' => __('1536 dim, recommended', 'mxchat')), | |
| 151 | - 'text-embedding-3-small' => array('label' => 'TE3 Small', 'description' => __('1536 dim, efficient', 'mxchat')), | |
| 152 | - 'text-embedding-3-large' => array('label' => 'TE3 Large', 'description' => __('3072 dim, powerful', 'mxchat')), | |
| 153 | - ), | |
| 154 | - ), | |
| 155 | - 'voyage' => array( | |
| 156 | - 'label' => __('Voyage AI', 'mxchat'), | |
| 157 | - 'key_option' => 'voyage_api_key', | |
| 158 | - 'models' => array( | |
| 159 | - 'voyage-3-large' => array('label' => 'Voyage-3 Large', 'description' => __('2048 dim, most capable', 'mxchat')), | |
| 160 | - ), | |
| 161 | - ), | |
| 162 | - 'gemini' => array( | |
| 163 | - 'label' => __('Google Gemini', 'mxchat'), | |
| 164 | - 'key_option' => 'gemini_api_key', | |
| 165 | - 'models' => array( | |
| 166 | - 'gemini-embedding-001' => array('label' => 'Gemini Embedding', 'description' => __('1536 dim', 'mxchat')), | |
| 167 | - ), | |
| 168 | - ), | |
| 169 | - ); | |
| 170 | - } | |
| 171 | - | |
| 172 | - /** | |
| 173 | - * Returns the display label for the given chat-or-embedding provider slug. | |
| 174 | - * Chat providers checked first; embedding-only providers fall through. | |
| 175 | - */ | |
| 176 | - public static function provider_label($slug) { | |
| 177 | - $chat = self::chat_models(); | |
| 178 | - if (isset($chat[$slug])) return $chat[$slug]['label']; | |
| 179 | - $emb = self::embedding_models(); | |
| 180 | - if (isset($emb[$slug])) return $emb[$slug]['label']; | |
| 181 | - return $slug; | |
| 182 | - } | |
| 183 | - | |
| 184 | - /** | |
| 185 | - * Returns the mxchat_options sub-key that holds the API key for the given | |
| 186 | - * provider slug. Chat providers checked first; embedding-only providers | |
| 187 | - * (voyage) fall through. | |
| 188 | - */ | |
| 189 | - public static function key_option_for_provider($slug) { | |
| 190 | - $chat = self::chat_models(); | |
| 191 | - if (isset($chat[$slug])) return $chat[$slug]['key_option']; | |
| 192 | - $emb = self::embedding_models(); | |
| 193 | - if (isset($emb[$slug])) return $emb[$slug]['key_option']; | |
| 194 | - return ''; | |
| 195 | - } | |
| 196 | - | |
| 197 | - /** | |
| 198 | - * Flat allowlist of every chat model id across all providers. Used by the | |
| 199 | - * sanitize() flow in class-mxchat-admin.php and the AJAX autosave validator | |
| 200 | - * in admin/class-ajax-handler.php so a new model id only needs editing in | |
| 201 | - * the catalog above. | |
| 202 | - */ | |
| 203 | - public static function chat_model_ids() { | |
| 204 | - $ids = array(); | |
| 205 | - foreach (self::chat_models() as $provider) { | |
| 206 | - foreach ($provider['models'] as $id => $_unused) { | |
| 207 | - $ids[] = $id; | |
| 208 | - } | |
| 209 | - } | |
| 210 | - return $ids; | |
| 211 | - } | |
| 212 | - | |
| 213 | - /** | |
| 214 | - * Flat allowlist of every embedding model id across all providers. | |
| 215 | - */ | |
| 216 | - public static function embedding_model_ids() { | |
| 217 | - $ids = array(); | |
| 218 | - foreach (self::embedding_models() as $provider) { | |
| 219 | - foreach ($provider['models'] as $id => $_unused) { | |
| 220 | - $ids[] = $id; | |
| 221 | - } | |
| 222 | - } | |
| 223 | - return $ids; | |
| 224 | - } | |
| 225 | - | |
| 226 | - /** | |
| 227 | - * Settings-dropdown-shaped chat model map keyed by group label, which is | |
| 228 | - * what mxchat_model_callback() in class-mxchat-admin.php renders as | |
| 229 | - * <optgroup>. Each model maps to its label-with-description string. | |
| 230 | - */ | |
| 231 | - public static function settings_dropdown_groups() { | |
| 232 | - $out = array(); | |
| 233 | - foreach (self::chat_models() as $provider_slug => $provider) { | |
| 234 | - $out[$provider['label']] = array(); | |
| 235 | - foreach ($provider['models'] as $id => $entry) { | |
| 236 | - $out[$provider['label']][$id] = $entry['label']; | |
| 237 | - } | |
| 238 | - } | |
| 239 | - return $out; | |
| 240 | - } | |
| 241 | - | |
| 242 | - /** | |
| 243 | - * Shape used by js/mxchat-admin.js's modal picker grid: provider slug => | |
| 244 | - * array of { value, label, description }. Mirrors the original inline | |
| 245 | - * `models = {…}` literal in js/mxchat-admin.js so the modal can read from | |
| 246 | - * a localized PHP source without changing its render path. | |
| 247 | - */ | |
| 248 | - public static function js_picker_shape() { | |
| 249 | - $out = array(); | |
| 250 | - foreach (self::chat_models() as $provider_slug => $provider) { | |
| 251 | - $out[$provider_slug] = array(); | |
| 252 | - foreach ($provider['models'] as $id => $entry) { | |
| 253 | - $out[$provider_slug][] = array( | |
| 254 | - 'value' => $id, | |
| 255 | - 'label' => $entry['label'], | |
| 256 | - 'description' => $entry['description'], | |
| 257 | - ); | |
| 258 | - } | |
| 259 | - } | |
| 260 | - return $out; | |
| 261 | - } | |
| 262 | - | |
| 263 | - /** | |
| 264 | - * Shape used by the Onboarding wizard's JS catalog (per-provider | |
| 265 | - * `{ label, chatModels, embeddingModels, hasKey, requiresKeyToLoadModels }`). | |
| 266 | - * `hasKey` is filled in by the caller after consulting mxchat_options. | |
| 267 | - */ | |
| 268 | - public static function onboarding_js_catalog() { | |
| 269 | - $chat = self::chat_models(); | |
| 270 | - $emb = self::embedding_models(); | |
| 271 | - $out = array(); | |
| 272 | - | |
| 273 | - $all_slugs = array_unique(array_merge(array_keys($chat), array_keys($emb))); | |
| 274 | - foreach ($all_slugs as $slug) { | |
| 275 | - $chat_models = array(); | |
| 276 | - if (isset($chat[$slug])) { | |
| 277 | - foreach ($chat[$slug]['models'] as $id => $entry) { | |
| 278 | - $chat_models[$id] = $entry['label']; | |
| 279 | - } | |
| 280 | - } | |
| 281 | - $embed_models = array(); | |
| 282 | - if (isset($emb[$slug])) { | |
| 283 | - foreach ($emb[$slug]['models'] as $id => $entry) { | |
| 284 | - $embed_models[$id] = $entry['label']; | |
| 285 | - } | |
| 286 | - } | |
| 287 | - $label = self::provider_label($slug); | |
| 288 | - $requires_key = isset($chat[$slug]) ? !empty($chat[$slug]['requires_key_to_load_models']) : false; | |
| 289 | - $disable_onb = isset($chat[$slug]) ? !empty($chat[$slug]['disable_in_onboarding']) : false; | |
| 290 | - | |
| 291 | - $out[$slug] = array( | |
| 292 | - 'label' => $label, | |
| 293 | - 'chatModels' => $chat_models, | |
| 294 | - 'embeddingModels' => $embed_models, | |
| 295 | - 'requiresKeyToLoadModels' => $requires_key, | |
| 296 | - 'disableInOnboarding' => ($disable_onb || $requires_key), | |
| 297 | - 'keyOption' => self::key_option_for_provider($slug), | |
| 298 | - ); | |
| 299 | - } | |
| 300 | - return $out; | |
| 301 | - } | |
| 302 | - | |
| 303 | - /** | |
| 304 | - * Whether the given chat model can do model-driven function calling (tool use). | |
| 305 | - * | |
| 306 | - * Used by the native function-calling loop (plan-mxchat-20260617-a41dee) to | |
| 307 | - * decide whether to offer tools to the model and to gate the admin toggle. | |
| 308 | - * | |
| 309 | - * Design: every modern flagship across the catalog's providers supports tool | |
| 310 | - * calling, so the DEFAULT is "capable" for any model whose id matches a known | |
| 311 | - * provider family prefix (or the OpenRouter meta). This deliberately avoids a | |
| 312 | - * per-model allowlist that would go stale the moment the catalog gains a model | |
| 313 | - * (the same drift trap documented in CLAUDE.md's model-registry checklist). | |
| 314 | - * A model can be force-excluded via the $no_tools list, and the whole verdict | |
| 315 | - * is overridable through the `mxchat_model_supports_tools` filter. | |
| 316 | - * | |
| 317 | - * @param string $model_id Chat model id (e.g. 'gpt-5.5', 'claude-opus-4-8') or | |
| 318 | - * the 'openrouter' meta selector. | |
| 319 | - * @return bool | |
| 320 | - */ | |
| 321 | - public static function supports_tools($model_id) { | |
| 322 | - $model_id = (string) $model_id; | |
| 323 | - | |
| 324 | - // The OpenRouter meta resolves to a per-account sub-model at request time; | |
| 325 | - // most routable models support tools, so default it capable (the loop | |
| 326 | - // no-ops gracefully if the chosen sub-model rejects a tools array). | |
| 327 | - if ($model_id === 'openrouter') { | |
| 328 | - return (bool) apply_filters('mxchat_model_supports_tools', true, $model_id); | |
| 329 | - } | |
| 330 | - | |
| 331 | - // Known tool-capable provider family prefixes across the catalog. | |
| 332 | - $capable_prefixes = array( | |
| 333 | - 'gpt-', 'o1-', 'o3-', 'o4-', // OpenAI | |
| 334 | - 'claude-', // Anthropic | |
| 335 | - 'gemini-', // Google | |
| 336 | - 'grok-', 'xai-', // xAI | |
| 337 | - 'deepseek-', // DeepSeek (OpenAI-compatible tools) | |
| 338 | - 'custom-', // OpenAI-compatible custom endpoints | |
| 339 | - ); | |
| 340 | - | |
| 341 | - // Explicit exclusions for any catalog model known NOT to support tools. | |
| 342 | - // Empty today — kept as the drift-safe escape hatch. | |
| 343 | - $no_tools = array(); | |
| 344 | - | |
| 345 | - $supported = false; | |
| 346 | - if (!in_array($model_id, $no_tools, true)) { | |
| 347 | - foreach ($capable_prefixes as $prefix) { | |
| 348 | - if (strpos($model_id, $prefix) === 0) { | |
| 349 | - $supported = true; | |
| 350 | - break; | |
| 351 | - } | |
| 352 | - } | |
| 353 | - } | |
| 354 | - | |
| 355 | - return (bool) apply_filters('mxchat_model_supports_tools', $supported, $model_id); | |
| 356 | - } | |
| 357 | -} | |
| 1 | +<?php | |
| 2 | +/** | |
| 3 | + * MxChat Model Catalog — single source of truth for chat + embedding models. | |
| 4 | + * | |
| 5 | + * Plan: plan-mxchat-20260527-d14e89. Consolidates model definitions that | |
| 6 | + * previously lived in three places (Settings dropdown in | |
| 7 | + * class-mxchat-admin.php, autosave allowlist in admin/class-ajax-handler.php, | |
| 8 | + * modal picker catalog in js/mxchat-admin.js, plus a fourth in | |
| 9 | + * admin-onboarding-page.php). Adding a new chat/embedding model now means | |
| 10 | + * editing THIS file only. | |
| 11 | + * | |
| 12 | + * Schema: | |
| 13 | + * chat_models() / embedding_models() return: | |
| 14 | + * [ providerSlug => [ | |
| 15 | + * 'label' => Display name shown in pickers, | |
| 16 | + * 'key_option' => sub-key under get_option('mxchat_options'), | |
| 17 | + * 'requires_key_to_load_models' => bool — OpenRouter is true (key needed | |
| 18 | + * before model list can be fetched). Surfaces consult | |
| 19 | + * this to disable / gate the provider in pickers that | |
| 20 | + * can't run the inline key-then-fetch flow. | |
| 21 | + * 'models' => [ model_id => [ 'label' => …, 'description' => … ] ] | |
| 22 | + * ] | |
| 23 | + * ] | |
| 24 | + * | |
| 25 | + * @package MxChat | |
| 26 | + * @since 3.2.7 | |
| 27 | + */ | |
| 28 | + | |
| 29 | +if (!defined('ABSPATH')) { | |
| 30 | + exit; | |
| 31 | +} | |
| 32 | + | |
| 33 | +class MxChat_Model_Catalog { | |
| 34 | + | |
| 35 | + /** | |
| 36 | + * Chat models grouped by provider. Order within a provider matches the | |
| 37 | + * order shown in the Settings modal picker — keep the recommended / | |
| 38 | + * flagship model at top. | |
| 39 | + */ | |
| 40 | + public static function chat_models() { | |
| 41 | + return array( | |
| 42 | + 'openrouter' => array( | |
| 43 | + 'label' => __('OpenRouter', 'mxchat'), | |
| 44 | + 'key_option' => 'openrouter_api_key', | |
| 45 | + 'requires_key_to_load_models' => true, | |
| 46 | + 'disable_in_onboarding' => true, | |
| 47 | + 'models' => array( | |
| 48 | + 'openrouter' => array( | |
| 49 | + 'label' => 'OpenRouter', | |
| 50 | + 'description' => __('Access 100+ models from multiple providers (add API key to browse)', 'mxchat'), | |
| 51 | + ), | |
| 52 | + ), | |
| 53 | + ), | |
| 54 | + 'gemini' => array( | |
| 55 | + 'label' => __('Google Gemini', 'mxchat'), | |
| 56 | + 'key_option' => 'gemini_api_key', | |
| 57 | + 'requires_key_to_load_models' => false, | |
| 58 | + 'models' => array( | |
| 59 | + 'gemini-3.8-flash' => array('label' => 'Gemini 3.8 Flash', 'description' => __('Stable — newest Flash generation, recommended default', 'mxchat')), | |
| 60 | + 'gemini-3.7-flash' => array('label' => 'Gemini 3.7 Flash', 'description' => __('Stable — efficiency-first Flash', 'mxchat')), | |
| 61 | + 'gemini-3.6-flash' => array('label' => 'Gemini 3.6 Flash', 'description' => __('Stable — previous Flash generation', 'mxchat')), | |
| 62 | + 'gemini-3.5-flash' => array('label' => 'Gemini 3.5 Flash', 'description' => __('Stable — older Flash, priced above 3.8 Flash on every token', 'mxchat')), | |
| 63 | + 'gemini-3.1-pro-preview' => array('label' => 'Gemini 3.1 Pro', 'description' => __('Preview — most intelligent, multimodal & agentic', 'mxchat')), | |
| 64 | + 'gemini-3-flash-preview' => array('label' => 'Gemini 3 Flash', 'description' => __('Preview — balanced speed and scale', 'mxchat')), | |
| 65 | + 'gemini-3.5-flash-lite' => array('label' => 'Gemini 3.5 Flash-Lite', 'description' => __('Stable — cheapest Gemini, no thinking', 'mxchat')), | |
| 66 | + 'gemini-3.1-flash-lite' => array('label' => 'Gemini 3.1 Flash-Lite', 'description' => __('Stable — cost-efficient, high throughput', 'mxchat')), | |
| 67 | + 'gemini-3.1-flash-lite-preview' => array('label' => 'Gemini 3.1 Flash-Lite (Preview)', 'description' => __('Preview — latest cost-efficient model', 'mxchat')), | |
| 68 | + 'gemini-2.5-pro' => array('label' => 'Gemini 2.5 Pro', 'description' => __('No longer offered by Google to new API users — pick Gemini 3.1 Pro', 'mxchat')), | |
| 69 | + 'gemini-2.5-flash' => array('label' => 'Gemini 2.5 Flash', 'description' => __('No longer offered by Google to new API users — pick Gemini 3.8 Flash', 'mxchat')), | |
| 70 | + 'gemini-2.5-flash-lite' => array('label' => 'Gemini 2.5 Flash-Lite', 'description' => __('No longer offered by Google to new API users — pick Gemini 3.5 Flash-Lite', 'mxchat')), | |
| 71 | + ), | |
| 72 | + ), | |
| 73 | + 'openai' => array( | |
| 74 | + 'label' => __('OpenAI', 'mxchat'), | |
| 75 | + 'key_option' => 'api_key', | |
| 76 | + 'requires_key_to_load_models' => false, | |
| 77 | + 'models' => array( | |
| 78 | + // OpenAI retired gpt-5.1-chat-latest and gpt-5.3-chat-latest on | |
| 79 | + // 2026-08-10 (replacement gpt-5.6-sol per the deprecations page); | |
| 80 | + // stored settings migrate via mxchat_migrate_deprecated_models() | |
| 81 | + // and the integrator carries a read-time rescue (plan e46b8f). | |
| 82 | + // | |
| 83 | + // 'deprecated_on' (Y-m-d) + 'replacement' are optional per-model | |
| 84 | + // fields: pickers label a future date as "retires <date>" and hide | |
| 85 | + // the entry once the date passes (a saved value keeps rendering via | |
| 86 | + // settings_dropdown_groups($current_model)). | |
| 87 | + // GPT-6 (announced 2026-09-22; plan 45c22d). Probed live on | |
| 88 | + // 2026-09-27: all three need max_completion_tokens, reject any | |
| 89 | + // temperature but 1 and reject reasoning_effort 'minimal'; | |
| 90 | + // Astra also rejects 'none'. Every "modern OpenAI" gate goes | |
| 91 | + // through is_modern_openai() so these ids are not sent legacy | |
| 92 | + // parameters (which 400 on the first message). | |
| 93 | + 'gpt-6-astra' => array('label' => 'GPT-6 Astra', 'description' => __('Recommended — newest OpenAI flagship; strongest reasoning, coding and agentic work', 'mxchat')), | |
| 94 | + 'gpt-6-sol' => array('label' => 'GPT-6 Sol', 'description' => __('Strong reasoning on demanding tasks at mid-tier cost', 'mxchat')), | |
| 95 | + 'gpt-6-luna' => array('label' => 'GPT-6 Luna', 'description' => __('Fastest and cheapest GPT-6 for high-volume, repeatable work', 'mxchat')), | |
| 96 | + 'gpt-5.6-sol' => array('label' => 'GPT-5.6 Sol', 'description' => __('Previous OpenAI flagship for reasoning, coding and chat', 'mxchat')), | |
| 97 | + 'gpt-5.6-terra' => array('label' => 'GPT-5.6 Terra', 'description' => __('Balanced GPT-5.6 — strong capability at lower cost', 'mxchat')), | |
| 98 | + 'gpt-5.6-luna' => array('label' => 'GPT-5.6 Luna', 'description' => __('Fastest and cheapest GPT-5.6 for lightweight tasks', 'mxchat')), | |
| 99 | + 'gpt-5.5' => array('label' => 'GPT-5.5', 'description' => __('Previous flagship — powerful reasoning and coding model', 'mxchat')), | |
| 100 | + 'gpt-5.4' => array('label' => 'GPT-5.4', 'description' => __('Flagship reasoning and coding model', 'mxchat')), | |
| 101 | + 'gpt-5.4-mini' => array('label' => 'GPT-5.4 Mini', 'description' => __('Fast and affordable with 400K context', 'mxchat')), | |
| 102 | + 'gpt-5.4-nano' => array('label' => 'GPT-5.4 Nano', 'description' => __('Fastest and cheapest for lightweight tasks', 'mxchat')), | |
| 103 | + 'gpt-5.2' => array('label' => 'GPT-5.2', 'description' => __('Best general-purpose & agentic model with fast responses', 'mxchat')), | |
| 104 | + 'gpt-5.1-2025-11-13' => array('label' => 'GPT-5.1', 'description' => __('Flagship for coding & agentic tasks with low reasoning (400K context)', 'mxchat')), | |
| 105 | + 'gpt-5' => array('label' => 'GPT-5', 'description' => __('Flagship for coding, reasoning, and agentic tasks across domains', 'mxchat'), 'deprecated_on' => '2026-12-11', 'replacement' => 'gpt-5.6-sol'), | |
| 106 | + 'gpt-5-mini' => array('label' => 'GPT-5 Mini', 'description' => __('Fast and lightweight', 'mxchat'), 'deprecated_on' => '2026-12-11', 'replacement' => 'gpt-5.6-terra'), | |
| 107 | + 'gpt-5-nano' => array('label' => 'GPT-5 Nano', 'description' => __('Fastest and cheapest; ideal for summarization and classification', 'mxchat'), 'deprecated_on' => '2026-12-11', 'replacement' => 'gpt-5.6-luna'), | |
| 108 | + ), | |
| 109 | + ), | |
| 110 | + 'claude' => array( | |
| 111 | + 'label' => __('Anthropic Claude', 'mxchat'), | |
| 112 | + 'key_option' => 'claude_api_key', | |
| 113 | + 'requires_key_to_load_models' => false, | |
| 114 | + 'models' => array( | |
| 115 | + 'claude-fable-5' => array('label' => 'Claude Fable 5', 'description' => __('Latest Flagship — newest and most capable Anthropic model', 'mxchat')), | |
| 116 | + 'claude-opus-5' => array('label' => 'Claude Opus 5', 'description' => __('Latest Opus — best for complex agentic and coding work', 'mxchat')), | |
| 117 | + 'claude-opus-4-8' => array('label' => 'Claude Opus 4.8', 'description' => __('Previous Opus generation', 'mxchat')), | |
| 118 | + 'claude-opus-4-7' => array('label' => 'Claude Opus 4.7', 'description' => __('Previous Anthropic flagship model', 'mxchat')), | |
| 119 | + 'claude-opus-4-6' => array('label' => 'Claude Opus 4.6', 'description' => __('Most capable Claude model - recommended', 'mxchat')), | |
| 120 | + 'claude-sonnet-5' => array('label' => 'Claude Sonnet 5', 'description' => __('Newest Sonnet - excellent balance of speed and capability', 'mxchat')), | |
| 121 | + 'claude-sonnet-4-6' => array('label' => 'Claude Sonnet 4.6', 'description' => __('Previous Sonnet - excellent balance of speed and capability', 'mxchat')), | |
| 122 | + 'claude-opus-4-5' => array('label' => 'Claude Opus 4.5', 'description' => __('Highly capable for complex tasks', 'mxchat')), | |
| 123 | + 'claude-sonnet-4-5-20250929' => array('label' => 'Claude Sonnet 4.5', 'description' => __('Best for complex agents and coding', 'mxchat')), | |
| 124 | + 'claude-haiku-4-5-20251001' => array('label' => 'Claude Haiku 4.5', 'description' => __('Fastest and most intelligent Haiku', 'mxchat')), | |
| 125 | + ), | |
| 126 | + ), | |
| 127 | + 'xai' => array( | |
| 128 | + 'label' => __('xAI Grok', 'mxchat'), | |
| 129 | + 'key_option' => 'xai_api_key', | |
| 130 | + 'requires_key_to_load_models' => false, | |
| 131 | + // Refreshed 2026-08-14 (plan-mxchat-20260814-e1aa4a) against TWO | |
| 132 | + // primary sources that agree: a live pull of xAI's /v1/models with | |
| 133 | + // a real key, and docs.x.ai/docs/models. Context windows below are | |
| 134 | + // that listing's own context_length; image-input support for every | |
| 135 | + // current-generation id was confirmed by a 32x32 probe returning a | |
| 136 | + // correct answer (all HTTP 200), not inferred from docs. | |
| 137 | + // | |
| 138 | + // The legacy block below is INTENTIONALLY still here. xAI has | |
| 139 | + // published no retirement dates for these ids and they still answer | |
| 140 | + // HTTP 200 (grok-4-0709 re-probed 2026-08-14), so 'deprecated_on' | |
| 141 | + // would be an invented date and deleting them would strand every | |
| 142 | + // install saved on one. Current generation simply sorts above them. | |
| 143 | + 'models' => array( | |
| 144 | + 'grok-4.6' => array('label' => 'Grok 4.6', 'description' => __('Newest xAI flagship — 500K context, accepts image input', 'mxchat')), | |
| 145 | + 'grok-4.5' => array('label' => 'Grok 4.5', 'description' => __('Previous flagship — 500K context, accepts image input', 'mxchat')), | |
| 146 | + 'grok-4.3' => array('label' => 'Grok 4.3', 'description' => __('Largest context in the Grok lineup — 1M tokens, accepts image input', 'mxchat')), | |
| 147 | + 'grok-4.20-0309-reasoning' => array('label' => 'Grok 4.20 (Reasoning)', 'description' => __('1M context with reasoning — dated release, pinned behaviour', 'mxchat')), | |
| 148 | + 'grok-4.20-0309-non-reasoning'=> array('label' => 'Grok 4.20 (Non-Reasoning)', 'description' => __('1M context without reasoning — faster replies at the same price', 'mxchat')), | |
| 149 | + // --- Previous generations. Still served by xAI, no announced | |
| 150 | + // retirement date, kept so saved selections keep working. --- | |
| 151 | + 'grok-4-1-fast-reasoning' => array('label' => 'Grok 4.1 Fast (Reasoning)', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 152 | + 'grok-4-1-fast-non-reasoning' => array('label' => 'Grok 4.1 Fast (Non-Reasoning)', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 153 | + 'grok-4-0709' => array('label' => 'Grok 4', 'description' => __('Previous generation — no longer listed by xAI, still served', 'mxchat')), | |
| 154 | + 'grok-3-beta' => array('label' => 'Grok-3', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 155 | + 'grok-3-fast-beta' => array('label' => 'Grok-3 Fast', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 156 | + 'grok-3-mini-beta' => array('label' => 'Grok-3 Mini', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 157 | + 'grok-3-mini-fast-beta' => array('label' => 'Grok-3 Mini Fast', 'description' => __('Older generation — no longer listed by xAI, still served', 'mxchat')), | |
| 158 | + ), | |
| 159 | + ), | |
| 160 | + 'deepseek' => array( | |
| 161 | + 'label' => __('DeepSeek', 'mxchat'), | |
| 162 | + 'key_option' => 'deepseek_api_key', | |
| 163 | + 'requires_key_to_load_models' => false, | |
| 164 | + 'models' => array( | |
| 165 | + // DeepSeek retired deepseek-chat / deepseek-reasoner on 2026-07-24; | |
| 166 | + // stored settings migrate via mxchat_migrate_deprecated_models(). | |
| 167 | + 'deepseek-v4-flash' => array('label' => 'DeepSeek V4 Flash', 'description' => __('Fast and cost-effective', 'mxchat')), | |
| 168 | + 'deepseek-v4-pro' => array('label' => 'DeepSeek V4 Pro', 'description' => __('Most capable DeepSeek model', 'mxchat')), | |
| 169 | + ), | |
| 170 | + ), | |
| 171 | + 'custom' => array( | |
| 172 | + 'label' => __('Custom (OpenAI-compatible)', 'mxchat'), | |
| 173 | + 'key_option' => 'custom_provider_api_key', | |
| 174 | + 'requires_key_to_load_models' => false, | |
| 175 | + 'disable_in_onboarding' => true, | |
| 176 | + 'models' => array( | |
| 177 | + 'custom-provider' => array( | |
| 178 | + 'label' => 'Custom Provider', | |
| 179 | + 'description' => __('OpenAI-compatible local LLM (Ollama, LM Studio, vLLM, llama.cpp, Azure OpenAI) — configure Base URL, key, and model in API Keys tab', 'mxchat'), | |
| 180 | + ), | |
| 181 | + ), | |
| 182 | + ), | |
| 183 | + ); | |
| 184 | + } | |
| 185 | + | |
| 186 | + /** | |
| 187 | + * Embedding models grouped by provider. Order within a provider matches | |
| 188 | + * the order shown in the Settings dropdown. | |
| 189 | + */ | |
| 190 | + public static function embedding_models() { | |
| 191 | + return array( | |
| 192 | + 'openai' => array( | |
| 193 | + 'label' => __('OpenAI', 'mxchat'), | |
| 194 | + 'key_option' => 'api_key', | |
| 195 | + 'models' => array( | |
| 196 | + 'text-embedding-ada-002' => array('label' => 'Ada 2', 'description' => __('1536 dim, recommended', 'mxchat')), | |
| 197 | + 'text-embedding-3-small' => array('label' => 'TE3 Small', 'description' => __('1536 dim, efficient', 'mxchat')), | |
| 198 | + 'text-embedding-3-large' => array('label' => 'TE3 Large', 'description' => __('3072 dim, powerful', 'mxchat')), | |
| 199 | + ), | |
| 200 | + ), | |
| 201 | + 'voyage' => array( | |
| 202 | + 'label' => __('Voyage AI', 'mxchat'), | |
| 203 | + 'key_option' => 'voyage_api_key', | |
| 204 | + 'models' => array( | |
| 205 | + 'voyage-3-large' => array('label' => 'Voyage-3 Large', 'description' => __('2048 dim, most capable', 'mxchat')), | |
| 206 | + ), | |
| 207 | + ), | |
| 208 | + 'gemini' => array( | |
| 209 | + 'label' => __('Google Gemini', 'mxchat'), | |
| 210 | + 'key_option' => 'gemini_api_key', | |
| 211 | + 'models' => array( | |
| 212 | + 'gemini-embedding-001' => array('label' => 'Gemini Embedding', 'description' => __('1536 dim', 'mxchat')), | |
| 213 | + ), | |
| 214 | + ), | |
| 215 | + ); | |
| 216 | + } | |
| 217 | + | |
| 218 | + /** | |
| 219 | + * Canonical retired-model map: every chat model id a provider has retired | |
| 220 | + * (or hard-deprecated) that stored settings may still carry, mapped to its | |
| 221 | + * current replacement (plan-mxchat-20260822-202df5). | |
| 222 | + * | |
| 223 | + * THE ONLY PLACE THIS MAPPING LIVES. mxchat_migrate_deprecated_models() | |
| 224 | + * consumes it for the main chat + content models, and add-ons that store | |
| 225 | + * their own model ids (mxchat-multi-bot's per-bot overrides) consume it for | |
| 226 | + * theirs — a second hardcoded list in any consumer is how a stranded-model | |
| 227 | + * bug recurs, so route every new retirement through this map. | |
| 228 | + * | |
| 229 | + * Claude ids map BY TIER so a rescued site lands on the current generation, | |
| 230 | + * not an older intermediate (plan dc91bd). Never map onto a model that is | |
| 231 | + * itself on a deprecation list (the e46b8f rule). | |
| 232 | + * | |
| 233 | + * Entry shape: retired_id => array( | |
| 234 | + * 'to' => replacement model id (must be current, never itself retired), | |
| 235 | + * 'label' => replacement display name for user-facing notices, | |
| 236 | + * 'reason' => notice clause explaining the retirement; deliberately NOT | |
| 237 | + * translated — migration messages are stored in an option and | |
| 238 | + * rendered later, matching the existing migration notices. | |
| 239 | + * ) | |
| 240 | + */ | |
| 241 | + public static function retired_model_map() { | |
| 242 | + return array( | |
| 243 | + // Anthropic Claude — Opus tier. | |
| 244 | + 'claude-3-opus-20240229' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Jan 5, 2026 | |
| 245 | + 'claude-opus-4-20250514' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 246 | + 'claude-opus-4-1-20250805' => array('to' => 'claude-opus-5', 'label' => 'Claude Opus 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Aug 5, 2026 (first-party API) | |
| 247 | + // Anthropic Claude — Sonnet tier. | |
| 248 | + 'claude-3-5-sonnet-20240620' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Oct 28, 2025 | |
| 249 | + 'claude-3-5-sonnet-20241022' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Oct 28, 2025 | |
| 250 | + 'claude-3-7-sonnet-20250219' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Retired Feb 19, 2026 | |
| 251 | + 'claude-3-sonnet-20240229' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Legacy | |
| 252 | + 'claude-sonnet-4-20250514' => array('to' => 'claude-sonnet-5', 'label' => 'Claude Sonnet 5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 253 | + // Anthropic Claude — Haiku tier. | |
| 254 | + 'claude-3-haiku-20240307' => array('to' => 'claude-haiku-4-5-20251001', 'label' => 'Claude Haiku 4.5', 'reason' => 'because Anthropic retired the older Claude model'), // Legacy | |
| 255 | + 'claude-3-5-haiku-20241022' => array('to' => 'claude-haiku-4-5-20251001', 'label' => 'Claude Haiku 4.5', 'reason' => 'because Anthropic retired the older Claude model'), // Deprecated | |
| 256 | + // OpenAI — GPT-4 series + GPT-3.5 Turbo (deprecated 2026-02-17). | |
| 257 | + 'gpt-4o' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 258 | + 'gpt-4.1-2025-04-14' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 259 | + 'gpt-4-turbo' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 260 | + 'gpt-4' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 261 | + 'gpt-3.5-turbo' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'due to OpenAI deprecating older models'), | |
| 262 | + // OpenAI — the gpt-5.x-chat-latest aliases retired 2026-08-10 (e46b8f). | |
| 263 | + 'gpt-5.1-chat-latest' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'because OpenAI retires the GPT-5.x Chat Latest models on August 10, 2026'), | |
| 264 | + 'gpt-5.3-chat-latest' => array('to' => 'gpt-5.6-sol', 'label' => 'GPT-5.6 Sol', 'reason' => 'because OpenAI retires the GPT-5.x Chat Latest models on August 10, 2026'), | |
| 265 | + // OpenAI — mini tier. | |
| 266 | + 'gpt-4o-mini' => array('to' => 'gpt-5-mini', 'label' => 'GPT-5 Mini', 'reason' => 'due to OpenAI deprecating GPT-4 series models'), | |
| 267 | + 'gpt-4.1-mini' => array('to' => 'gpt-5-mini', 'label' => 'GPT-5 Mini', 'reason' => 'due to OpenAI deprecating GPT-4 series models'), | |
| 268 | + // DeepSeek — removed at the vendor 2026-07-24 (hard cutoff, every request 400s). | |
| 269 | + 'deepseek-chat' => array('to' => 'deepseek-v4-flash', 'label' => 'DeepSeek V4 Flash', 'reason' => 'because DeepSeek retired its older API models on July 24, 2026'), | |
| 270 | + 'deepseek-reasoner' => array('to' => 'deepseek-v4-flash', 'label' => 'DeepSeek V4 Flash', 'reason' => 'because DeepSeek retired its older API models on July 24, 2026'), | |
| 271 | + ); | |
| 272 | + } | |
| 273 | + | |
| 274 | + /** | |
| 275 | + * The current replacement for a retired chat model id, or the id unchanged | |
| 276 | + * when it is not in retired_model_map(). Static, no I/O, no option read. | |
| 277 | + * Every request-time remap goes through here (plan 8c4eb4), so a newly | |
| 278 | + * retired id needs one map row and nothing else. | |
| 279 | + */ | |
| 280 | + public static function resolve_retired($model_id) { | |
| 281 | + $map = self::retired_model_map(); | |
| 282 | + if (is_string($model_id) && isset($map[$model_id]['to']) && $map[$model_id]['to'] !== '') { | |
| 283 | + return $map[$model_id]['to']; | |
| 284 | + } | |
| 285 | + return $model_id; | |
| 286 | + } | |
| 287 | + | |
| 288 | + /** | |
| 289 | + * Returns the display label for the given chat-or-embedding provider slug. | |
| 290 | + * Chat providers checked first; embedding-only providers fall through. | |
| 291 | + */ | |
| 292 | + public static function provider_label($slug) { | |
| 293 | + $chat = self::chat_models(); | |
| 294 | + if (isset($chat[$slug])) return $chat[$slug]['label']; | |
| 295 | + $emb = self::embedding_models(); | |
| 296 | + if (isset($emb[$slug])) return $emb[$slug]['label']; | |
| 297 | + return $slug; | |
| 298 | + } | |
| 299 | + | |
| 300 | + /** | |
| 301 | + * Returns the mxchat_options sub-key that holds the API key for the given | |
| 302 | + * provider slug. Chat providers checked first; embedding-only providers | |
| 303 | + * (voyage) fall through. | |
| 304 | + */ | |
| 305 | + public static function key_option_for_provider($slug) { | |
| 306 | + $chat = self::chat_models(); | |
| 307 | + if (isset($chat[$slug])) return $chat[$slug]['key_option']; | |
| 308 | + $emb = self::embedding_models(); | |
| 309 | + if (isset($emb[$slug])) return $emb[$slug]['key_option']; | |
| 310 | + return ''; | |
| 311 | + } | |
| 312 | + | |
| 313 | + /** | |
| 314 | + * The provider slug that owns a chat model id, or '' when the catalog | |
| 315 | + * doesn't list it (plan b65e8d). | |
| 316 | + * | |
| 317 | + * NOTE for callers doing retirement work: a model the provider has already | |
| 318 | + * retired is frequently ABSENT from this catalog while still saved on an | |
| 319 | + * install — that is precisely the state the liveness check exists to catch. | |
| 320 | + * Treat '' as "catalog doesn't know", not "not a real model", and fall back | |
| 321 | + * to a family-prefix guess if you need one. | |
| 322 | + */ | |
| 323 | + public static function provider_for_chat_model($model_id) { | |
| 324 | + $model_id = (string) $model_id; | |
| 325 | + foreach (self::chat_models() as $provider_slug => $provider) { | |
| 326 | + if (isset($provider['models'][$model_id])) { | |
| 327 | + return $provider_slug; | |
| 328 | + } | |
| 329 | + } | |
| 330 | + return ''; | |
| 331 | + } | |
| 332 | + | |
| 333 | + /** | |
| 334 | + * URL + auth headers for a provider's public model-listing endpoint | |
| 335 | + * (plan b65e8d — the daily liveness check). | |
| 336 | + * | |
| 337 | + * Only providers that actually publish a listing are mapped. OpenRouter is | |
| 338 | + * deliberately excluded: its 'openrouter' catalog entry is a meta-selector | |
| 339 | + * that resolves to a per-account sub-model at request time, so "is the | |
| 340 | + * configured id listed" is not a meaningful question. DeepSeek and the | |
| 341 | + * custom OpenAI-compatible provider are excluded too — a self-hosted or | |
| 342 | + * proxied endpoint's /models is not authoritative about retirement. | |
| 343 | + * | |
| 344 | + * @param string $provider Provider slug. | |
| 345 | + * @param string $api_key The install's key for that provider. | |
| 346 | + * @return array|null array('url' => string, 'headers' => array) or null. | |
| 347 | + */ | |
| 348 | + public static function models_listing_request($provider, $api_key) { | |
| 349 | + $api_key = (string) $api_key; | |
| 350 | + if ($api_key === '') { | |
| 351 | + return null; | |
| 352 | + } | |
| 353 | + | |
| 354 | + switch ($provider) { | |
| 355 | + case 'openai': | |
| 356 | + return array( | |
| 357 | + 'url' => 'https://api.openai.com/v1/models', | |
| 358 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key), | |
| 359 | + ); | |
| 360 | + case 'claude': | |
| 361 | + return array( | |
| 362 | + 'url' => 'https://api.anthropic.com/v1/models?limit=1000', | |
| 363 | + 'headers' => array( | |
| 364 | + 'x-api-key' => $api_key, | |
| 365 | + 'anthropic-version' => '2023-06-01', | |
| 366 | + ), | |
| 367 | + ); | |
| 368 | + case 'gemini': | |
| 369 | + return array( | |
| 370 | + 'url' => 'https://generativelanguage.googleapis.com/v1beta/models?pageSize=1000', | |
| 371 | + 'headers' => array('x-goog-api-key' => $api_key), | |
| 372 | + ); | |
| 373 | + case 'xai': | |
| 374 | + return array( | |
| 375 | + 'url' => 'https://api.x.ai/v1/models', | |
| 376 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key), | |
| 377 | + ); | |
| 378 | + } | |
| 379 | + | |
| 380 | + return null; | |
| 381 | + } | |
| 382 | + | |
| 383 | + /** | |
| 384 | + * A minimal 1-token chat request used ONLY to confirm whether a model id | |
| 385 | + * still exists (plan b65e8d). Never used for content — the response is | |
| 386 | + * thrown away; only the HTTP status and error code are read. | |
| 387 | + * | |
| 388 | + * WHY THIS EXISTS: a provider's model LISTING is not authoritative about | |
| 389 | + * retirement, in either direction (all probed live 2026-08-13): | |
| 390 | + * - OpenAI KEEPS retired ids listed — gpt-5.3-chat-latest is in | |
| 391 | + * /v1/models and 404s "model_deprecated" on use. | |
| 392 | + * - Anthropic DROPS working aliases — claude-opus-4-5 is absent from | |
| 393 | + * /v1/models and answers HTTP 200 perfectly. | |
| 394 | + * - xAI's listing is partial — grok-4-0709 is absent and answers 200. | |
| 395 | + * So "absent from the listing" is a suspicion, not a verdict. This probe is | |
| 396 | + * the verdict, and it is the only thing allowed to raise a warning. | |
| 397 | + * | |
| 398 | + * @return array|null array('url','headers','body') or null when unsupported. | |
| 399 | + */ | |
| 400 | + public static function inference_probe_request($provider, $api_key, $model) { | |
| 401 | + $api_key = (string) $api_key; | |
| 402 | + $model = (string) $model; | |
| 403 | + if ($api_key === '' || $model === '') { | |
| 404 | + return null; | |
| 405 | + } | |
| 406 | + | |
| 407 | + $msg = array(array('role' => 'user', 'content' => 'hi')); | |
| 408 | + | |
| 409 | + switch ($provider) { | |
| 410 | + case 'openai': | |
| 411 | + return array( | |
| 412 | + 'url' => 'https://api.openai.com/v1/chat/completions', | |
| 413 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key, 'Content-Type' => 'application/json'), | |
| 414 | + 'body' => array('model' => $model, 'messages' => $msg, 'max_completion_tokens' => 1), | |
| 415 | + ); | |
| 416 | + case 'xai': | |
| 417 | + return array( | |
| 418 | + 'url' => 'https://api.x.ai/v1/chat/completions', | |
| 419 | + 'headers' => array('Authorization' => 'Bearer ' . $api_key, 'Content-Type' => 'application/json'), | |
| 420 | + 'body' => array('model' => $model, 'messages' => $msg, 'max_tokens' => 1), | |
| 421 | + ); | |
| 422 | + case 'claude': | |
| 423 | + return array( | |
| 424 | + 'url' => 'https://api.anthropic.com/v1/messages', | |
| 425 | + 'headers' => array( | |
| 426 | + 'x-api-key' => $api_key, | |
| 427 | + 'anthropic-version' => '2023-06-01', | |
| 428 | + 'Content-Type' => 'application/json', | |
| 429 | + ), | |
| 430 | + 'body' => array('model' => $model, 'max_tokens' => 1, 'messages' => $msg), | |
| 431 | + ); | |
| 432 | + case 'gemini': | |
| 433 | + return array( | |
| 434 | + 'url' => 'https://generativelanguage.googleapis.com/v1beta/models/' | |
| 435 | + . rawurlencode($model) . ':generateContent', | |
| 436 | + 'headers' => array('x-goog-api-key' => $api_key, 'Content-Type' => 'application/json'), | |
| 437 | + 'body' => array( | |
| 438 | + 'contents' => array(array('parts' => array(array('text' => 'hi')))), | |
| 439 | + 'generationConfig' => array('maxOutputTokens' => 1), | |
| 440 | + ), | |
| 441 | + ); | |
| 442 | + } | |
| 443 | + | |
| 444 | + return null; | |
| 445 | + } | |
| 446 | + | |
| 447 | + /** | |
| 448 | + * Flat allowlist of every chat model id across all providers. Used by the | |
| 449 | + * sanitize() flow in class-mxchat-admin.php and the AJAX autosave validator | |
| 450 | + * in admin/class-ajax-handler.php so a new model id only needs editing in | |
| 451 | + * the catalog above. | |
| 452 | + */ | |
| 453 | + public static function chat_model_ids() { | |
| 454 | + $ids = array(); | |
| 455 | + foreach (self::chat_models() as $provider) { | |
| 456 | + foreach ($provider['models'] as $id => $_unused) { | |
| 457 | + $ids[] = $id; | |
| 458 | + } | |
| 459 | + } | |
| 460 | + return $ids; | |
| 461 | + } | |
| 462 | + | |
| 463 | + /** | |
| 464 | + * Flat allowlist of every embedding model id across all providers. | |
| 465 | + */ | |
| 466 | + public static function embedding_model_ids() { | |
| 467 | + $ids = array(); | |
| 468 | + foreach (self::embedding_models() as $provider) { | |
| 469 | + foreach ($provider['models'] as $id => $_unused) { | |
| 470 | + $ids[] = $id; | |
| 471 | + } | |
| 472 | + } | |
| 473 | + return $ids; | |
| 474 | + } | |
| 475 | + | |
| 476 | + /** | |
| 477 | + * Deprecation-aware picker label for a catalog entry (plan e46b8f). | |
| 478 | + * | |
| 479 | + * @param array $entry Model entry from chat_models(). | |
| 480 | + * @return string|null Label to show (with a "retires <date>" suffix when a | |
| 481 | + * future 'deprecated_on' is set), or null when the | |
| 482 | + * retirement date has passed and the entry must be | |
| 483 | + * hidden from NEW selections. | |
| 484 | + */ | |
| 485 | + private static function picker_label($entry) { | |
| 486 | + if (empty($entry['deprecated_on'])) { | |
| 487 | + return $entry['label']; | |
| 488 | + } | |
| 489 | + $ts = strtotime($entry['deprecated_on'] . ' 00:00:00 UTC'); | |
| 490 | + if ($ts !== false && time() >= $ts) { | |
| 491 | + return null; // retired — hide from new selections | |
| 492 | + } | |
| 493 | + return sprintf( | |
| 494 | + /* translators: 1: model label, 2: retirement date */ | |
| 495 | + __('%1$s — retires %2$s', 'mxchat'), | |
| 496 | + $entry['label'], | |
| 497 | + date_i18n(get_option('date_format', 'F j, Y'), $ts) | |
| 498 | + ); | |
| 499 | + } | |
| 500 | + | |
| 501 | + /** | |
| 502 | + * Settings-dropdown-shaped chat model map keyed by group label, which is | |
| 503 | + * what mxchat_model_callback() in class-mxchat-admin.php renders as | |
| 504 | + * <optgroup>. Each model maps to its label-with-description string. | |
| 505 | + * | |
| 506 | + * Retired models (past 'deprecated_on') are hidden — EXCEPT the install's | |
| 507 | + * currently saved model, which is re-injected with a "retired" marker so | |
| 508 | + * the settings screen never renders a blank select (plan e46b8f). | |
| 509 | + * | |
| 510 | + * @param string $current_model Optionally, the install's saved model id. | |
| 511 | + * @param array|null $provider_slugs Optional allowlist of provider slugs | |
| 512 | + * (catalog keys, e.g. 'openai', 'claude'). | |
| 513 | + * null = every provider. Lets a consumer | |
| 514 | + * that can only dispatch to some providers | |
| 515 | + * (the content generator — plan ccb751) | |
| 516 | + * derive its picker from the same catalog. | |
| 517 | + */ | |
| 518 | + public static function settings_dropdown_groups($current_model = '', $provider_slugs = null) { | |
| 519 | + $out = array(); | |
| 520 | + foreach (self::chat_models() as $provider_slug => $provider) { | |
| 521 | + if (is_array($provider_slugs) && !in_array($provider_slug, $provider_slugs, true)) { | |
| 522 | + continue; | |
| 523 | + } | |
| 524 | + $out[$provider['label']] = array(); | |
| 525 | + foreach ($provider['models'] as $id => $entry) { | |
| 526 | + $label = self::picker_label($entry); | |
| 527 | + if ($label === null && $id !== $current_model) { | |
| 528 | + continue; | |
| 529 | + } | |
| 530 | + if ($label === null) { | |
| 531 | + /* translators: %s: model label */ | |
| 532 | + $label = sprintf(__('%s — retired by provider, choose a replacement', 'mxchat'), $entry['label']); | |
| 533 | + } | |
| 534 | + $out[$provider['label']][$id] = $label; | |
| 535 | + } | |
| 536 | + } | |
| 537 | + return $out; | |
| 538 | + } | |
| 539 | + | |
| 540 | + /** | |
| 541 | + * Shape used by js/mxchat-admin.js's modal picker grid: provider slug => | |
| 542 | + * array of { value, label, description }. Mirrors the original inline | |
| 543 | + * `models = {…}` literal in js/mxchat-admin.js so the modal can read from | |
| 544 | + * a localized PHP source without changing its render path. | |
| 545 | + */ | |
| 546 | + public static function js_picker_shape() { | |
| 547 | + $out = array(); | |
| 548 | + foreach (self::chat_models() as $provider_slug => $provider) { | |
| 549 | + $out[$provider_slug] = array(); | |
| 550 | + foreach ($provider['models'] as $id => $entry) { | |
| 551 | + $label = self::picker_label($entry); | |
| 552 | + if ($label === null) { | |
| 553 | + continue; // retired — never offered as a new selection | |
| 554 | + } | |
| 555 | + $out[$provider_slug][] = array( | |
| 556 | + 'value' => $id, | |
| 557 | + 'label' => $label, | |
| 558 | + 'description' => $entry['description'], | |
| 559 | + ); | |
| 560 | + } | |
| 561 | + } | |
| 562 | + return $out; | |
| 563 | + } | |
| 564 | + | |
| 565 | + /** | |
| 566 | + * Shape used by the Onboarding wizard's JS catalog (per-provider | |
| 567 | + * `{ label, chatModels, embeddingModels, hasKey, requiresKeyToLoadModels }`). | |
| 568 | + * `hasKey` is filled in by the caller after consulting mxchat_options. | |
| 569 | + */ | |
| 570 | + public static function onboarding_js_catalog() { | |
| 571 | + $chat = self::chat_models(); | |
| 572 | + $emb = self::embedding_models(); | |
| 573 | + $out = array(); | |
| 574 | + | |
| 575 | + $all_slugs = array_unique(array_merge(array_keys($chat), array_keys($emb))); | |
| 576 | + foreach ($all_slugs as $slug) { | |
| 577 | + $chat_models = array(); | |
| 578 | + if (isset($chat[$slug])) { | |
| 579 | + foreach ($chat[$slug]['models'] as $id => $entry) { | |
| 580 | + $label = self::picker_label($entry); | |
| 581 | + if ($label === null) { | |
| 582 | + continue; // retired — not offered during onboarding | |
| 583 | + } | |
| 584 | + $chat_models[$id] = $label; | |
| 585 | + } | |
| 586 | + } | |
| 587 | + $embed_models = array(); | |
| 588 | + if (isset($emb[$slug])) { | |
| 589 | + foreach ($emb[$slug]['models'] as $id => $entry) { | |
| 590 | + $embed_models[$id] = $entry['label']; | |
| 591 | + } | |
| 592 | + } | |
| 593 | + $label = self::provider_label($slug); | |
| 594 | + $requires_key = isset($chat[$slug]) ? !empty($chat[$slug]['requires_key_to_load_models']) : false; | |
| 595 | + $disable_onb = isset($chat[$slug]) ? !empty($chat[$slug]['disable_in_onboarding']) : false; | |
| 596 | + | |
| 597 | + $out[$slug] = array( | |
| 598 | + 'label' => $label, | |
| 599 | + 'chatModels' => $chat_models, | |
| 600 | + 'embeddingModels' => $embed_models, | |
| 601 | + 'requiresKeyToLoadModels' => $requires_key, | |
| 602 | + 'disableInOnboarding' => ($disable_onb || $requires_key), | |
| 603 | + 'keyOption' => self::key_option_for_provider($slug), | |
| 604 | + ); | |
| 605 | + } | |
| 606 | + return $out; | |
| 607 | + } | |
| 608 | + | |
| 609 | + /** | |
| 610 | + * Whether the given chat model can do model-driven function calling (tool use). | |
| 611 | + * | |
| 612 | + * Used by the native function-calling loop (plan-mxchat-20260617-a41dee) to | |
| 613 | + * decide whether to offer tools to the model and to gate the admin toggle. | |
| 614 | + * | |
| 615 | + * Design: every modern flagship across the catalog's providers supports tool | |
| 616 | + * calling, so the DEFAULT is "capable" for any model whose id matches a known | |
| 617 | + * provider family prefix (or the OpenRouter meta). This deliberately avoids a | |
| 618 | + * per-model allowlist that would go stale the moment the catalog gains a model | |
| 619 | + * (the same drift trap documented in CLAUDE.md's model-registry checklist). | |
| 620 | + * A model can be force-excluded via the $no_tools list, and the whole verdict | |
| 621 | + * is overridable through the `mxchat_model_supports_tools` filter. | |
| 622 | + * | |
| 623 | + * @param string $model_id Chat model id (e.g. 'gpt-5.5', 'claude-opus-4-8') or | |
| 624 | + * the 'openrouter' meta selector. | |
| 625 | + * @return bool | |
| 626 | + */ | |
| 627 | + public static function supports_tools($model_id) { | |
| 628 | + $model_id = (string) $model_id; | |
| 629 | + | |
| 630 | + // The OpenRouter meta resolves to a per-account sub-model at request time; | |
| 631 | + // most routable models support tools, so default it capable (the loop | |
| 632 | + // no-ops gracefully if the chosen sub-model rejects a tools array). | |
| 633 | + if ($model_id === 'openrouter') { | |
| 634 | + return (bool) apply_filters('mxchat_model_supports_tools', true, $model_id); | |
| 635 | + } | |
| 636 | + | |
| 637 | + // Known tool-capable provider family prefixes across the catalog. | |
| 638 | + $capable_prefixes = array( | |
| 639 | + 'gpt-', 'o1-', 'o3-', 'o4-', // OpenAI | |
| 640 | + 'claude-', // Anthropic | |
| 641 | + 'gemini-', // Google | |
| 642 | + 'grok-', 'xai-', // xAI | |
| 643 | + 'deepseek-', // DeepSeek (OpenAI-compatible tools) | |
| 644 | + 'custom-', // OpenAI-compatible custom endpoints | |
| 645 | + ); | |
| 646 | + | |
| 647 | + // Explicit exclusions for any catalog model known NOT to support tools. | |
| 648 | + // Empty today — kept as the drift-safe escape hatch. | |
| 649 | + $no_tools = array(); | |
| 650 | + | |
| 651 | + $supported = false; | |
| 652 | + if (!in_array($model_id, $no_tools, true)) { | |
| 653 | + foreach ($capable_prefixes as $prefix) { | |
| 654 | + if (strpos($model_id, $prefix) === 0) { | |
| 655 | + $supported = true; | |
| 656 | + break; | |
| 657 | + } | |
| 658 | + } | |
| 659 | + } | |
| 660 | + | |
| 661 | + return (bool) apply_filters('mxchat_model_supports_tools', $supported, $model_id); | |
| 662 | + } | |
| 663 | + | |
| 664 | + /** | |
| 665 | + * The catalog's CURRENT recommended model for image analysis, per provider. | |
| 666 | + * | |
| 667 | + * plan-mxchat-20260717-eb371e: add-ons that need "a vision-capable model" | |
| 668 | + * (mxchat-vision's analysis paths, mxchat-admin-chat's in-admin image | |
| 669 | + * analysis) previously froze a model id at their own call sites — and those | |
| 670 | + * pins silently aged out (mxchat-vision was still sending gpt-4o). This is | |
| 671 | + * the ONE place that opinion lives now; add-ons ask, with their old pin | |
| 672 | + * demoted to a fallback for when mxchat-basic is absent or too old. | |
| 673 | + * | |
| 674 | + * Choices are deliberate: | |
| 675 | + * - openai: gpt-5.6-sol — the current flagship, vision-capable; OpenAI's | |
| 676 | + * stated replacement for the gpt-5.x-chat-latest aliases retired | |
| 677 | + * 2026-08-10 (which this map previously pinned; plan e46b8f). | |
| 678 | + * - claude: the current Haiku — image analysis wants fast + cheap; Haiku | |
| 679 | + * is vision-capable and the catalog's newest Haiku generation. | |
| 680 | + * - xai: grok-4.6, the current flagship. Promoted 2026-08-14 (plan | |
| 681 | + * e1aa4a) from grok-4-0709, which was two generations behind and no | |
| 682 | + * longer appears in xAI's /v1/models listing. Image input is CONFIRMED, | |
| 683 | + * not assumed: a 32x32 fixture probe on 2026-08-14 returned the correct | |
| 684 | + * colour on grok-4.6 / 4.5 / 4.3 and both 4.20 variants (all HTTP 200). | |
| 685 | + * grok-4.3 is the cheaper vision option (roughly half the image-token | |
| 686 | + * price, 1M context) if cost ever outweighs capability here. | |
| 687 | + * - gemini: current stable Flash — vision-capable, no per-image surcharge. | |
| 688 | + * | |
| 689 | + * NOT covered: image-GENERATION models (Gemini Imagen / Flash Image etc.). | |
| 690 | + * Those are a different capability the catalog does not model; add-ons that | |
| 691 | + * pin one (mxchat-vision's editing endpoints) are correctly pinned. | |
| 692 | + * | |
| 693 | + * @param string $provider 'openai' | 'claude' | 'xai' | 'gemini'. | |
| 694 | + * @return string Model id, or '' for an unknown provider (caller keeps its | |
| 695 | + * fallback). Overridable via mxchat_default_vision_model. | |
| 696 | + */ | |
| 697 | + public static function default_vision_model($provider) { | |
| 698 | + $map = array( | |
| 699 | + 'openai' => 'gpt-5.6-sol', | |
| 700 | + 'claude' => 'claude-haiku-4-5-20251001', | |
| 701 | + 'xai' => 'grok-4.6', | |
| 702 | + 'gemini' => 'gemini-3.8-flash', | |
| 703 | + ); | |
| 704 | + $model = isset($map[$provider]) ? $map[$provider] : ''; | |
| 705 | + return (string) apply_filters('mxchat_default_vision_model', $model, $provider); | |
| 706 | + } | |
| 707 | + | |
| 708 | + /** | |
| 709 | + * Per-model, per-surface reasoning_effort value. | |
| 710 | + * | |
| 711 | + * plan-mxchat-20260714-dcb71c: the single source of truth for the | |
| 712 | + * reasoning_effort ladder that was previously hand-maintained in FIVE | |
| 713 | + * parallel places (integrator streaming/non-streaming chat, integrator | |
| 714 | + * web-search Responses, content-generator, mxchat-advanced-content | |
| 715 | + * class-ai-client + class-content-calendar). Adding a model to | |
| 716 | + * chat_models() and updating THIS method now propagates the routing to | |
| 717 | + * every surface — no more drift-on-every-model-add. | |
| 718 | + * | |
| 719 | + * The mapping is DELIBERATELY per-context because the surfaces genuinely | |
| 720 | + * differ today (they are NOT a single clamped ladder): | |
| 721 | + * - 'chat' — /v1/chat/completions main chat. gpt-5.5 and gpt-5.4 use | |
| 722 | + * 'none'; the *-chat-latest / 5.4-mini / 5.4-nano / | |
| 723 | + * gpt-5.2 ids send NO reasoning_effort; the original | |
| 724 | + * gpt-5 / gpt-5-mini / gpt-5-nano send 'minimal' (their | |
| 725 | + * generation requires it — 25b972 probe). Unknown future | |
| 726 | + * gpt-5* ids send NOTHING (allowlist fall-through). | |
| 727 | + * - 'content' — content generation (content-generator, AC ai-client, | |
| 728 | + * AC content-calendar). gpt-5.5 / gpt-5.4 use 'low'; | |
| 729 | + * gpt-5.4-mini / gpt-5.4-nano use 'none' (they never | |
| 730 | + * supported 'minimal' — 25b972); gpt-5.3-chat-latest and | |
| 731 | + * the original gpt-5 / -mini / -nano use 'minimal'; only | |
| 732 | + * gpt-5.2 and gpt-5.1-chat-latest send nothing. This is | |
| 733 | + * an ALLOWLIST (5621dd safe-default): a future gpt-5* id | |
| 734 | + * not listed gets NO param rather than a guessed value | |
| 735 | + * that might 400. | |
| 736 | + * - 'websearch' — /v1/responses web-search. Only the gpt-5.6 trio, 5.5, | |
| 737 | + * 5.4, and 5.1-2025-11-13 send 'low'; every other gpt-5 id | |
| 738 | + * (including gpt-5 / mini / nano / chat-latest) sends | |
| 739 | + * nothing. | |
| 740 | + * | |
| 741 | + * @param string $model Chat model id. | |
| 742 | + * @param string $context 'chat' | 'content' | 'websearch'. Defaults to 'chat'. | |
| 743 | + * @return string|null 'none'|'low'|'minimal' to send, or null to omit the | |
| 744 | + * reasoning_effort / reasoning.effort param entirely. | |
| 745 | + */ | |
| 746 | + public static function reasoning_effort_for($model, $context = 'chat') { | |
| 747 | + $model = (string) $model; | |
| 748 | + | |
| 749 | + // plan a9774b: every Gemini 3.x output price is "including thinking | |
| 750 | + // tokens" and the model thinks at its default level unless told | |
| 751 | + // otherwise — measured on this site's own support prompt, 71% of | |
| 752 | + // billed output on 3.8 Flash was hidden thinking, with indistinguishable | |
| 753 | + // replies at 'low' (0 hidden tokens). So the chat surface sends 'low': | |
| 754 | + // as reasoning_effort on the OpenAI-compatible endpoint, and as | |
| 755 | + // generationConfig.thinkingConfig.thinkingLevel on the native one (see | |
| 756 | + // gemini_thinking_config()). Flash-Lite does not think and returns 400 | |
| 757 | + // on any reasoning parameter — it must send nothing. | |
| 758 | + if (strpos($model, 'gemini-3') === 0) { | |
| 759 | + if (strpos($model, 'flash-lite') !== false) { | |
| 760 | + return null; | |
| 761 | + } | |
| 762 | + return ($context === 'chat') ? 'low' : null; | |
| 763 | + } | |
| 764 | + | |
| 765 | + // Only the modern OpenAI family (gpt-5*, gpt-6*) carries | |
| 766 | + // reasoning_effort. Everything else omits. plan 45c22d widened this | |
| 767 | + // from a literal gpt-5 prefix: a gpt-6 id used to fall through all | |
| 768 | + // three maps below and be sent legacy parameters. | |
| 769 | + if (!self::is_modern_openai($model)) { | |
| 770 | + return null; | |
| 771 | + } | |
| 772 | + | |
| 773 | + switch ($context) { | |
| 774 | + case 'websearch': | |
| 775 | + // integrator web-search (Responses API). Explicit allowlist, no | |
| 776 | + // "else" — anything not listed omits reasoning entirely. | |
| 777 | + $map = array( | |
| 778 | + 'gpt-6-astra' => 'low', // plan 45c22d, probed 2026-09-27 on /v1/responses | |
| 779 | + 'gpt-6-sol' => 'low', | |
| 780 | + 'gpt-6-luna' => 'low', | |
| 781 | + 'gpt-5.6-sol' => 'low', | |
| 782 | + 'gpt-5.6-terra' => 'low', | |
| 783 | + 'gpt-5.6-luna' => 'low', | |
| 784 | + 'gpt-5.5' => 'low', | |
| 785 | + 'gpt-5.4' => 'low', | |
| 786 | + 'gpt-5.1-2025-11-13' => 'low', | |
| 787 | + ); | |
| 788 | + return isset($map[$model]) ? $map[$model] : null; | |
| 789 | + | |
| 790 | + case 'content': | |
| 791 | + // content-generator + AC class-ai-client + AC content-calendar. | |
| 792 | + // ALLOWLIST (5621dd): a gpt-5* id absent here sends nothing. | |
| 793 | + // | |
| 794 | + // plan-25b972 (probed live 2026-08-13): supported reasoning_effort | |
| 795 | + // VALUES are per-model. gpt-5.4-mini/nano never supported | |
| 796 | + // 'minimal' (their floor is 'none' — the old 'minimal' entries | |
| 797 | + // 400'd on every call); the original gpt-5/-mini/-nano still | |
| 798 | + // REQUIRE 'minimal' and reject 'none'. Do not "harmonize" these | |
| 799 | + // — the two generations genuinely differ. | |
| 800 | + $map = array( | |
| 801 | + // plan 45c22d (probed 2026-09-27): gpt-6-astra accepts only | |
| 802 | + // low|medium|high; gpt-6-sol/-luna accept none|low|medium|high. | |
| 803 | + // 'low' for all three, as for their 5.6 siblings. | |
| 804 | + 'gpt-6-astra' => 'low', | |
| 805 | + 'gpt-6-sol' => 'low', | |
| 806 | + 'gpt-6-luna' => 'low', | |
| 807 | + 'gpt-5.6-sol' => 'low', | |
| 808 | + 'gpt-5.6-terra' => 'low', | |
| 809 | + 'gpt-5.6-luna' => 'low', | |
| 810 | + 'gpt-5.5' => 'low', | |
| 811 | + 'gpt-5.4' => 'low', | |
| 812 | + 'gpt-5.1-2025-11-13' => 'low', | |
| 813 | + 'gpt-5.4-mini' => 'none', | |
| 814 | + 'gpt-5.4-nano' => 'none', | |
| 815 | + 'gpt-5.3-chat-latest' => 'minimal', | |
| 816 | + 'gpt-5' => 'minimal', | |
| 817 | + 'gpt-5-mini' => 'minimal', | |
| 818 | + 'gpt-5-nano' => 'minimal', | |
| 819 | + ); | |
| 820 | + return isset($map[$model]) ? $map[$model] : null; | |
| 821 | + | |
| 822 | + case 'chat': | |
| 823 | + default: | |
| 824 | + // integrator streaming + non-streaming main chat. | |
| 825 | + // gpt-5.1/5.3-chat-latest stay listed here (and in 'content') | |
| 826 | + // even though the catalog no longer offers them: bot-level / | |
| 827 | + // add-on-saved ids that miss the e46b8f migration must keep | |
| 828 | + // routing correctly until every surface is swept. | |
| 829 | + $no_reasoning = array('gpt-5.2', 'gpt-5.1-chat-latest', 'gpt-5.3-chat-latest', 'gpt-5.4-mini', 'gpt-5.4-nano'); | |
| 830 | + if (in_array($model, $no_reasoning, true)) { | |
| 831 | + return null; | |
| 832 | + } | |
| 833 | + if ($model === 'gpt-5.1-2025-11-13') { | |
| 834 | + return 'low'; | |
| 835 | + } | |
| 836 | + if ($model === 'gpt-5.5') { | |
| 837 | + return 'none'; | |
| 838 | + } | |
| 839 | + if ($model === 'gpt-5.4') { | |
| 840 | + return 'none'; | |
| 841 | + } | |
| 842 | + // plan 45c22d (probed 2026-09-27): gpt-6-astra accepts only | |
| 843 | + // low|medium|high; -sol/-luna accept none|low|medium|high. | |
| 844 | + // 'low' for all three, as for their 5.6 siblings. | |
| 845 | + if (in_array($model, array('gpt-6-astra', 'gpt-6-sol', 'gpt-6-luna'), true)) { | |
| 846 | + return 'low'; | |
| 847 | + } | |
| 848 | + if (in_array($model, array('gpt-5.6-sol', 'gpt-5.6-terra', 'gpt-5.6-luna'), true)) { | |
| 849 | + return 'low'; | |
| 850 | + } | |
| 851 | + // plan-25b972: the original gpt-5 generation REQUIRES 'minimal' | |
| 852 | + // (probed 2026-08-13: 'none' is rejected for these three). | |
| 853 | + if (in_array($model, array('gpt-5', 'gpt-5-mini', 'gpt-5-nano'), true)) { | |
| 854 | + return 'minimal'; | |
| 855 | + } | |
| 856 | + // plan-25b972: allowlist fall-through (5621dd philosophy). An | |
| 857 | + // unknown future gpt-5* id sends NO param — a guessed value is | |
| 858 | + // exactly how the 5.4-mini/nano 400s shipped. | |
| 859 | + return null; | |
| 860 | + } | |
| 861 | + } | |
| 862 | + | |
| 863 | + /** | |
| 864 | + * Whether a NON-DEFAULT temperature value may be sent to this model. | |
| 865 | + * | |
| 866 | + * plan-mxchat-20260714-dcb71c: consolidates the scattered temperature | |
| 867 | + * omissions. Two families reject a custom temperature: | |
| 868 | + * - gpt-5* — accept only the default (1); any other value 400s. Callers | |
| 869 | + * that send temperature=1 (the main chat path) need not consult this — | |
| 870 | + * 1 is the default and is accepted. Callers that send a NON-default | |
| 871 | + * value (content-generator 0.7, AC 0.7) gate on this and omit for gpt-5. | |
| 872 | + * - The Anthropic flagships that removed the temperature param entirely | |
| 873 | + * (claude-opus-4-7 / 4-8, claude-fable-5, claude-sonnet-5) — 400 if sent | |
| 874 | + * at all. This is the list the integrator's mxchat_claude_omits_temperature() | |
| 875 | + * and content-generator's call_claude() have hand-maintained. | |
| 876 | + * | |
| 877 | + * @param string $model Chat model id. | |
| 878 | + * @return bool true if a custom temperature may be sent. | |
| 879 | + */ | |
| 880 | + /** | |
| 881 | + * generationConfig.thinkingConfig for the native Gemini endpoint, or null | |
| 882 | + * to send none (plan a9774b). Derived from reasoning_effort_for() so the | |
| 883 | + * two Gemini paths — native generateContent and the OpenAI-compatible | |
| 884 | + * endpoint — always agree on the thinking level. Overridable via | |
| 885 | + * mxchat_gemini_thinking_level (return '' to send none). | |
| 886 | + * | |
| 887 | + * @param string $model Chat model id. | |
| 888 | + * @return array|null e.g. array('thinkingLevel' => 'low'). | |
| 889 | + */ | |
| 890 | + public static function gemini_thinking_config($model) { | |
| 891 | + $level = self::reasoning_effort_for($model, 'chat'); | |
| 892 | + if (strpos((string) $model, 'gemini-') !== 0) { | |
| 893 | + return null; | |
| 894 | + } | |
| 895 | + $level = apply_filters('mxchat_gemini_thinking_level', $level, $model); | |
| 896 | + if (!is_string($level) || $level === '' || !in_array($level, array('minimal', 'low', 'medium', 'high'), true)) { | |
| 897 | + return null; | |
| 898 | + } | |
| 899 | + return array('thinkingLevel' => $level); | |
| 900 | + } | |
| 901 | + | |
| 902 | + public static function supports_temperature($model) { | |
| 903 | + $model = (string) $model; | |
| 904 | + | |
| 905 | + // gpt-5 / gpt-6 families accept only the default temperature (1). | |
| 906 | + if (self::is_modern_openai($model)) { | |
| 907 | + return false; | |
| 908 | + } | |
| 909 | + | |
| 910 | + // Anthropic flagships that reject the temperature param outright. | |
| 911 | + // TRAP: exact-match list — every NEW Anthropic flagship (4.7 and newer) | |
| 912 | + // must be added here or it 400s on every chat message while looking | |
| 913 | + // fine in the picker. A claude-opus-* prefix rule would wrongly catch | |
| 914 | + // claude-opus-4-6 and older, which still accept temperature. | |
| 915 | + $claude_no_temp = array('claude-opus-5', 'claude-opus-4-7', 'claude-opus-4-8', 'claude-fable-5', 'claude-sonnet-5'); | |
| 916 | + if (in_array($model, $claude_no_temp, true)) { | |
| 917 | + return false; | |
| 918 | + } | |
| 919 | + | |
| 920 | + return true; | |
| 921 | + } | |
| 922 | + | |
| 923 | + /** | |
| 924 | + * The completion-token parameter key an OpenAI /v1/chat/completions call | |
| 925 | + * must use for this model. gpt-5* require 'max_completion_tokens'; the | |
| 926 | + * legacy 'max_tokens' 400s. All other OpenAI-compatible providers | |
| 927 | + * (xAI, DeepSeek, custom) use 'max_tokens'. | |
| 928 | + * | |
| 929 | + * plan-mxchat-20260714-dcb71c: centralizes the `$is_gpt5 ? | |
| 930 | + * 'max_completion_tokens' : 'max_tokens'` switch duplicated in | |
| 931 | + * content-generator and AC class-ai-client. | |
| 932 | + * | |
| 933 | + * @param string $model Chat model id. | |
| 934 | + * @return string 'max_completion_tokens' | 'max_tokens' | |
| 935 | + */ | |
| 936 | + public static function openai_token_param($model) { | |
| 937 | + return self::is_modern_openai($model) ? 'max_completion_tokens' : 'max_tokens'; | |
| 938 | + } | |
| 939 | + | |
| 940 | + /** | |
| 941 | + * plan 45c22d: ONE predicate for "a modern OpenAI chat model" — the | |
| 942 | + * generations that take max_completion_tokens, only the default | |
| 943 | + * temperature, and reasoning_effort. Every gate that used to test the | |
| 944 | + * literal prefix gpt-5 routes through here, so the next generation is a | |
| 945 | + * one-line change instead of nine edits. Before this, gpt-6-* ids fell | |
| 946 | + * through all nine and were sent legacy parameters: a 400 on the first | |
| 947 | + * message while looking fine in the picker. | |
| 948 | + * | |
| 949 | + * The frozen partial-upgrade fallbacks in the integrator, the content | |
| 950 | + * generator and the editor assistant carry the same two-prefix test | |
| 951 | + * inline (they exist for the window where this class is not loaded). | |
| 952 | + * | |
| 953 | + * @param string $model Chat model id. | |
| 954 | + * @return bool | |
| 955 | + */ | |
| 956 | + public static function is_modern_openai($model) { | |
| 957 | + $model = (string) $model; | |
| 958 | + return strpos($model, 'gpt-5') === 0 || strpos($model, 'gpt-6') === 0; | |
| 959 | + } | |
| 960 | +} | |
