| 1 |
<?php |
| 2 |
/** |
| 3 |
* OpenStation — AI Copilot: WordPress AI Client adapter. |
| 4 |
* |
| 5 |
* Thin wrappers around `wp_ai_client_prompt()` that the agentic search loop |
| 6 |
* and the comment-scoring job use to generate. Credentials are injected by |
| 7 |
* Core from the configured Connector — nothing here ever handles an API key. |
| 8 |
* |
| 9 |
* The search loop advertises its tools — built-in WordPress Abilities (see |
| 10 |
* abilities.php) plus client command tools — as function declarations, and |
| 11 |
* dispatches ability calls through `wp_get_ability()->execute()`. |
| 12 |
* |
| 13 |
* All SDK classes referenced here ship with WordPress 7.0+. The `use` |
| 14 |
* statements are compile-time aliases only; every call site is reached solely |
| 15 |
* through {@see openstation_ai_is_available()}, so this file is inert (and |
| 16 |
* never resolves the classes) on older WordPress. |
| 17 |
* |
| 18 |
* @package OpenStation |
| 19 |
*/ |
| 20 |
|
| 21 |
use WordPress\AiClient\Messages\DTO\Message; |
| 22 |
use WordPress\AiClient\Messages\DTO\MessagePart; |
| 23 |
use WordPress\AiClient\Messages\DTO\UserMessage; |
| 24 |
use WordPress\AiClient\Providers\Models\Contracts\ModelInterface; |
| 25 |
use WordPress\AiClient\Providers\Models\DTO\ModelConfig; |
| 26 |
use WordPress\AiClient\Tools\DTO\FunctionCall; |
| 27 |
use WordPress\AiClient\Tools\DTO\FunctionDeclaration; |
| 28 |
use WordPress\AiClient\Tools\DTO\FunctionResponse; |
| 29 |
|
| 30 |
defined( 'ABSPATH' ) || exit; |
| 31 |
|
| 32 |
/** |
| 33 |
* Builds SDK function declarations from the loop's tool definitions. |
| 34 |
* |
| 35 |
* Each definition is the neutral tool shape the registry already produces: |
| 36 |
* `{ type: 'function', name, description, parameters (JSON Schema) }`. |
| 37 |
* |
| 38 |
* @param array $tool_defs List of tool definitions. |
| 39 |
* @return FunctionDeclaration[] |
| 40 |
*/ |
| 41 |
function openstation_ai_build_function_declarations( array $tool_defs ) { |
| 42 |
$declarations = array(); |
| 43 |
foreach ( $tool_defs as $def ) { |
| 44 |
if ( ! is_array( $def ) ) { |
| 45 |
continue; |
| 46 |
} |
| 47 |
$name = isset( $def['name'] ) ? (string) $def['name'] : ''; |
| 48 |
if ( '' === $name ) { |
| 49 |
continue; |
| 50 |
} |
| 51 |
$description = isset( $def['description'] ) ? (string) $def['description'] : ''; |
| 52 |
$parameters = isset( $def['parameters'] ) && is_array( $def['parameters'] ) ? $def['parameters'] : null; |
| 53 |
|
| 54 |
$declarations[] = new FunctionDeclaration( $name, $description, $parameters ); |
| 55 |
} |
| 56 |
return $declarations; |
| 57 |
} |
| 58 |
|
| 59 |
/** |
| 60 |
* Wraps a user query as a text message for the conversation history. |
| 61 |
* |
| 62 |
* @param string $text |
| 63 |
* @return UserMessage |
| 64 |
*/ |
| 65 |
function openstation_ai_user_text_message( $text ) { |
| 66 |
return new UserMessage( array( new MessagePart( (string) $text ) ) ); |
| 67 |
} |
| 68 |
|
| 69 |
/** |
| 70 |
* Wraps tool results as a user message of function-response parts. |
| 71 |
* |
| 72 |
* @param array $tool_outputs List of `{ call_id, name, response }` entries. |
| 73 |
* @return UserMessage |
| 74 |
*/ |
| 75 |
function openstation_ai_tool_result_message( array $tool_outputs ) { |
| 76 |
$parts = array(); |
| 77 |
foreach ( $tool_outputs as $output ) { |
| 78 |
$parts[] = new MessagePart( |
| 79 |
new FunctionResponse( |
| 80 |
isset( $output['call_id'] ) && '' !== $output['call_id'] ? (string) $output['call_id'] : null, |
| 81 |
isset( $output['name'] ) && '' !== $output['name'] ? (string) $output['name'] : null, |
| 82 |
isset( $output['response'] ) ? $output['response'] : null |
| 83 |
) |
| 84 |
); |
| 85 |
} |
| 86 |
return new UserMessage( $parts ); |
| 87 |
} |
| 88 |
|
| 89 |
/** |
| 90 |
* Strips thought-channel parts from a message before it re-enters history. |
| 91 |
* |
| 92 |
* Providers cannot reliably round-trip reasoning blocks: the Anthropic |
| 93 |
* provider drops the cryptographic `signature` when parsing a `thinking` |
| 94 |
* block, and the API rejects any replayed thinking block without one |
| 95 |
* (`thinking.signature: Field required`). Thought parts carry no information |
| 96 |
* the next turn needs — the model re-reasons from the visible conversation — |
| 97 |
* so the agentic loop replays assistant turns without them. |
| 98 |
* |
| 99 |
* If every part is a thought (no text, no function call), the message is |
| 100 |
* returned unchanged rather than emptied; the loop never replays such a |
| 101 |
* turn anyway. |
| 102 |
* |
| 103 |
* @param Message $message Assistant message as returned by the AI Client. |
| 104 |
* @return Message Message safe to append to the conversation history. |
| 105 |
*/ |
| 106 |
function openstation_ai_strip_thought_parts( Message $message ) { |
| 107 |
$kept = array(); |
| 108 |
$stripped = false; |
| 109 |
foreach ( $message->getParts() as $part ) { |
| 110 |
if ( $part->getChannel()->isThought() ) { |
| 111 |
$stripped = true; |
| 112 |
continue; |
| 113 |
} |
| 114 |
$kept[] = $part; |
| 115 |
} |
| 116 |
|
| 117 |
if ( ! $stripped || empty( $kept ) ) { |
| 118 |
return $message; |
| 119 |
} |
| 120 |
|
| 121 |
return new Message( $message->getRole(), $kept ); |
| 122 |
} |
| 123 |
|
| 124 |
/** |
| 125 |
* Output-token ceiling for every generation turn the site's |
| 126 |
* `openstation_ai_model_config` filter leaves uncapped. |
| 127 |
* |
| 128 |
* The three default providers disagree about what "no ceiling" means. |
| 129 |
* The OpenAI and Google providers send none, so the model's own maximum |
| 130 |
* applies; the Anthropic provider must send one and falls back to a |
| 131 |
* hard-coded 4096. That is enough for a chat answer and nowhere near |
| 132 |
* enough for a tool call carrying a whole post: the model runs out of |
| 133 |
* room inside the call's JSON, the API returns only the argument pairs |
| 134 |
* that were complete before the cut, and the ability rejects the call |
| 135 |
* for its missing `content`. The Localizer failed seven translations of |
| 136 |
* one post in a row exactly that way, each attempt cut at the same place. |
| 137 |
* |
| 138 |
* 16384 is the largest value every current-generation model of the three |
| 139 |
* providers accepts (OpenAI's gpt-4o family caps output at exactly that). |
| 140 |
* A site that needs more, or pins an older model with a smaller limit, |
| 141 |
* sets `max_tokens` in the filter: the filter's value always wins. |
| 142 |
*/ |
| 143 |
const OPENSTATION_AI_DEFAULT_MAX_TOKENS = 16384; |
| 144 |
|
| 145 |
/** |
| 146 |
* Whether the provider stopped because the reply hit the output-token |
| 147 |
* ceiling. |
| 148 |
* |
| 149 |
* Anthropic's `max_tokens` and Google's `MAX_TOKENS` stop reasons both |
| 150 |
* map to the SDK's LENGTH finish reason. The OpenAI provider throws |
| 151 |
* instead and never builds a result; {@see openstation_ai_client_generate()} |
| 152 |
* maps that path from the WP_Error Core turns the exception into. |
| 153 |
* |
| 154 |
* @param mixed $result GenerativeAiResult. |
| 155 |
* @return bool |
| 156 |
*/ |
| 157 |
function openstation_ai_result_is_truncated( $result ) { |
| 158 |
try { |
| 159 |
foreach ( $result->getCandidates() as $candidate ) { |
| 160 |
if ( $candidate->getFinishReason()->isLength() ) { |
| 161 |
return true; |
| 162 |
} |
| 163 |
} |
| 164 |
} catch ( \Throwable $e ) { |
| 165 |
return false; |
| 166 |
} |
| 167 |
return false; |
| 168 |
} |
| 169 |
|
| 170 |
/** |
| 171 |
* Builds the error for a turn the output-token ceiling cut short. |
| 172 |
* |
| 173 |
* A truncated reply is never usable. A JSON answer no longer parses, |
| 174 |
* and a function call arrives with only the argument pairs that were |
| 175 |
* complete before the cut, so the ability rejects it for a missing |
| 176 |
* required field, and the model, reading its own truncated call back |
| 177 |
* from history, sends the same call again to the same end. Failing the |
| 178 |
* turn here turns a run of silent retries into one error that names |
| 179 |
* the cause. |
| 180 |
* |
| 181 |
* @param string $detail Underlying provider detail, preserved for logs. |
| 182 |
* @param array|null $usage Token usage of the truncated turn, if known. |
| 183 |
* @return WP_Error |
| 184 |
*/ |
| 185 |
function openstation_ai_output_truncated_error( $detail, $usage = null ) { |
| 186 |
return new WP_Error( |
| 187 |
'openstation_ai_output_truncated', |
| 188 |
__( 'The AI provider cut the reply short at the output-token limit.', 'desktop-mode' ), |
| 189 |
array( |
| 190 |
'status' => 502, |
| 191 |
'detail' => (string) $detail, |
| 192 |
'completion_tokens' => is_array( $usage ) && isset( $usage['completion'] ) ? (int) $usage['completion'] : null, |
| 193 |
) |
| 194 |
); |
| 195 |
} |
| 196 |
|
| 197 |
/** |
| 198 |
* Builds the error for a final turn that produced no answer text. |
| 199 |
* |
| 200 |
* Observed live with the Anthropic provider under agent runs: a hard task |
| 201 |
* spends the entire `max_tokens` budget inside a thinking block |
| 202 |
* (`stop_reason: "max_tokens"`, a single text-less thought part), so the |
| 203 |
* turn carries neither function calls nor extractable text. Callers that |
| 204 |
* can meaningfully degrade instead (the command follow-up turn) match on |
| 205 |
* this code and keep their own fallback. |
| 206 |
* |
| 207 |
* @param string $detail Underlying extraction failure, preserved for logs. |
| 208 |
* @return WP_Error |
| 209 |
*/ |
| 210 |
function openstation_ai_empty_answer_error( $detail ) { |
| 211 |
return new WP_Error( |
| 212 |
'openstation_ai_empty_answer', |
| 213 |
__( 'The AI provider returned no answer text.', 'desktop-mode' ), |
| 214 |
array( |
| 215 |
'status' => 502, |
| 216 |
'detail' => (string) $detail, |
| 217 |
) |
| 218 |
); |
| 219 |
} |
| 220 |
|
| 221 |
/** |
| 222 |
* Applies the site's model config to a prompt builder. |
| 223 |
* |
| 224 |
* @param mixed $builder WP_AI_Client_Prompt_Builder. |
| 225 |
* @param array $context Partial filter context; missing keys are defaulted. |
| 226 |
* @return mixed |
| 227 |
*/ |
| 228 |
function openstation_ai_apply_model_config( $builder, array $context ) { |
| 229 |
$context = array_merge( |
| 230 |
array( |
| 231 |
'user_id' => 0, |
| 232 |
'request_id' => '', |
| 233 |
'source' => '', |
| 234 |
'has_tools' => false, |
| 235 |
'has_schema' => false, |
| 236 |
), |
| 237 |
$context |
| 238 |
); |
| 239 |
|
| 240 |
/** |
| 241 |
* Filters the model config for one AI turn. |
| 242 |
* |
| 243 |
* Defaults to empty; the only value OpenStation fills in afterwards is |
| 244 |
* `max_tokens` ({@see OPENSTATION_AI_DEFAULT_MAX_TOKENS}), and only |
| 245 |
* when the filter left it unset. Recipe: `docs/examples/ai-model-config.md`. |
| 246 |
* |
| 247 |
* @param array $config { model?: string|ModelInterface, max_tokens?: int, temperature?: float, custom_options?: array<string, mixed> }. |
| 248 |
* @param array $context { user_id, request_id, source, has_tools, has_schema }. |
| 249 |
*/ |
| 250 |
$config = apply_filters( 'openstation_ai_model_config', array(), $context ); |
| 251 |
if ( ! is_array( $config ) ) { |
| 252 |
$config = array(); |
| 253 |
} |
| 254 |
|
| 255 |
$model_config = new ModelConfig(); |
| 256 |
|
| 257 |
$max_tokens = OPENSTATION_AI_DEFAULT_MAX_TOKENS; |
| 258 |
if ( isset( $config['max_tokens'] ) && is_numeric( $config['max_tokens'] ) && (int) $config['max_tokens'] > 0 ) { |
| 259 |
$max_tokens = (int) $config['max_tokens']; |
| 260 |
} |
| 261 |
$model_config->setMaxTokens( $max_tokens ); |
| 262 |
|
| 263 |
// Unlike max_tokens, 0.0 is a legitimate temperature (deterministic). The |
| 264 |
// 2.0 ceiling is the range the SDK's own schema declares. |
| 265 |
if ( isset( $config['temperature'] ) && is_numeric( $config['temperature'] ) |
| 266 |
&& (float) $config['temperature'] >= 0.0 && (float) $config['temperature'] <= 2.0 ) { |
| 267 |
$model_config->setTemperature( (float) $config['temperature'] ); |
| 268 |
} |
| 269 |
|
| 270 |
$custom_options = array(); |
| 271 |
if ( isset( $config['custom_options'] ) && is_array( $config['custom_options'] ) ) { |
| 272 |
foreach ( $config['custom_options'] as $key => $value ) { |
| 273 |
// A list would reach the provider as parameters named `0`, `1`, …. |
| 274 |
if ( is_string( $key ) && '' !== $key ) { |
| 275 |
$custom_options[ $key ] = $value; |
| 276 |
} |
| 277 |
} |
| 278 |
} |
| 279 |
|
| 280 |
if ( ! empty( $custom_options ) ) { |
| 281 |
$model_config->setCustomOptions( $custom_options ); |
| 282 |
} |
| 283 |
|
| 284 |
$builder = $builder->using_model_config( $model_config ); |
| 285 |
|
| 286 |
// After the config: `using_model()` merges the model's own defaults under |
| 287 |
// whatever the builder already carries, so ours has to land first. |
| 288 |
$model = isset( $config['model'] ) ? $config['model'] : null; |
| 289 |
if ( $model instanceof ModelInterface ) { |
| 290 |
$builder = $builder->using_model( $model ); |
| 291 |
} elseif ( is_string( $model ) && '' !== trim( $model ) ) { |
| 292 |
// `using_model()` needs a ModelInterface, so a bare model id goes |
| 293 |
// through `using_model_preference()`, which throws on anything that |
| 294 |
// isn't a non-empty string. |
| 295 |
$builder = $builder->using_model_preference( trim( $model ) ); |
| 296 |
} |
| 297 |
|
| 298 |
return $builder; |
| 299 |
} |
| 300 |
|
| 301 |
/** |
| 302 |
* Runs one generation turn through the AI Client. |
| 303 |
* |
| 304 |
* Rebuilds the prompt from the full ordered message list each turn (the |
| 305 |
* builder's `with_history()` prepends, so it can't append turns in a loop), |
| 306 |
* advertises the tools as function declarations, and constrains the final |
| 307 |
* answer to `$answer_schema` when given. Returns the assistant turn normalized |
| 308 |
* to the shape the loop consumes; `message` has thought-channel parts stripped |
| 309 |
* ({@see openstation_ai_strip_thought_parts()}) so it is safe to replay. |
| 310 |
* |
| 311 |
* @param int $user_id Requesting user id. |
| 312 |
* @param array $messages Ordered conversation as SDK Message objects. |
| 313 |
* @param array $tool_defs Tool definitions to advertise. |
| 314 |
* @param array|null $answer_schema JSON Schema for the final answer, or null. |
| 315 |
* @param string $instructions System instruction. |
| 316 |
* @param array $context Optional. `{ source?: string, request_id?: string }` |
| 317 |
* for the model-config filter. |
| 318 |
* @return array{ text: ?string, function_calls: array, message: mixed, usage: ?array, model: ?array }|WP_Error |
| 319 |
*/ |
| 320 |
function openstation_ai_client_generate( $user_id, array $messages, array $tool_defs, $answer_schema, $instructions, array $context = array() ) { |
| 321 |
$builder = wp_ai_client_prompt( $messages ); |
| 322 |
|
| 323 |
if ( is_string( $instructions ) && '' !== $instructions ) { |
| 324 |
$builder = $builder->using_system_instruction( $instructions ); |
| 325 |
} |
| 326 |
|
| 327 |
// Provider + model selection is delegated to the Core AI Client |
| 328 |
// (Connector-backed) unless the model-config filter says otherwise. |
| 329 |
|
| 330 |
$declarations = openstation_ai_build_function_declarations( $tool_defs ); |
| 331 |
if ( ! empty( $declarations ) ) { |
| 332 |
$builder = $builder->using_function_declarations( ...$declarations ); |
| 333 |
} |
| 334 |
|
| 335 |
if ( is_array( $answer_schema ) ) { |
| 336 |
// Strict structured output: providers reject an object subschema that |
| 337 |
// doesn't set `additionalProperties: false`, and one such node 400s the |
| 338 |
// whole turn. Normalize here so no schema author has to know that. |
| 339 |
$builder = $builder->as_json_response( openstation_ai_normalize_response_schema( $answer_schema ) ); |
| 340 |
} |
| 341 |
|
| 342 |
$builder = openstation_ai_apply_model_config( |
| 343 |
$builder, |
| 344 |
array_merge( |
| 345 |
$context, |
| 346 |
array( |
| 347 |
'user_id' => (int) $user_id, |
| 348 |
'has_tools' => ! empty( $declarations ), |
| 349 |
'has_schema' => is_array( $answer_schema ), |
| 350 |
) |
| 351 |
) |
| 352 |
); |
| 353 |
|
| 354 |
$result = $builder->generate_result(); |
| 355 |
if ( is_wp_error( $result ) ) { |
| 356 |
// The OpenAI provider reports the output ceiling as an exception |
| 357 |
// rather than a finish reason; Core maps it to this code. |
| 358 |
if ( 'prompt_token_limit_reached' === $result->get_error_code() ) { |
| 359 |
return openstation_ai_output_truncated_error( $result->get_error_message() ); |
| 360 |
} |
| 361 |
return $result; |
| 362 |
} |
| 363 |
|
| 364 |
$message = $result->toMessage(); |
| 365 |
$function_calls = array(); |
| 366 |
$has_text = false; |
| 367 |
foreach ( $message->getParts() as $part ) { |
| 368 |
if ( $part->getType()->isText() && ! $part->getChannel()->isThought() ) { |
| 369 |
$has_text = true; |
| 370 |
} |
| 371 |
if ( ! $part->getType()->isFunctionCall() ) { |
| 372 |
continue; |
| 373 |
} |
| 374 |
$call = $part->getFunctionCall(); |
| 375 |
if ( ! $call instanceof FunctionCall ) { |
| 376 |
continue; |
| 377 |
} |
| 378 |
$args = $call->getArgs(); |
| 379 |
$function_calls[] = array( |
| 380 |
'name' => (string) $call->getName(), |
| 381 |
'call_id' => (string) $call->getId(), |
| 382 |
'arguments' => wp_json_encode( is_array( $args ) ? $args : array() ), |
| 383 |
); |
| 384 |
} |
| 385 |
|
| 386 |
// Whatever the ceiling cut off is partial, and partial is unusable: |
| 387 |
// a function call missing its longest argument, or a JSON answer |
| 388 |
// missing its closing half. A budget spent entirely on reasoning |
| 389 |
// leaves nothing written at all; that case keeps its own error below. |
| 390 |
if ( ( ! empty( $function_calls ) || $has_text ) && openstation_ai_result_is_truncated( $result ) ) { |
| 391 |
return openstation_ai_output_truncated_error( |
| 392 |
'The provider stopped at the output-token ceiling (finish reason: length).', |
| 393 |
openstation_ai_result_token_usage( $result ) |
| 394 |
); |
| 395 |
} |
| 396 |
|
| 397 |
$text = null; |
| 398 |
if ( empty( $function_calls ) ) { |
| 399 |
// A turn with no function calls IS the final answer, so failing to |
| 400 |
// extract its text is a failed generation, not a valid empty one. |
| 401 |
// Swallowing it here used to surface as a "successful" run with an |
| 402 |
// empty answer, invisible to the retry and error paths alike. |
| 403 |
try { |
| 404 |
$text = $result->toText(); |
| 405 |
} catch ( \Throwable $e ) { |
| 406 |
return openstation_ai_empty_answer_error( $e->getMessage() ); |
| 407 |
} |
| 408 |
if ( ! is_string( $text ) || '' === trim( $text ) ) { |
| 409 |
return openstation_ai_empty_answer_error( 'The provider response contains no text part.' ); |
| 410 |
} |
| 411 |
} |
| 412 |
|
| 413 |
return array( |
| 414 |
'text' => $text, |
| 415 |
'function_calls' => $function_calls, |
| 416 |
'message' => openstation_ai_strip_thought_parts( $message ), |
| 417 |
'usage' => openstation_ai_result_token_usage( $result ), |
| 418 |
'model' => openstation_ai_result_model_metadata( $result ), |
| 419 |
); |
| 420 |
} |
| 421 |
|
| 422 |
/** |
| 423 |
* Extracts normalized token usage from a generation result. |
| 424 |
* |
| 425 |
* @param mixed $result GenerativeAiResult. |
| 426 |
* @return array{ prompt: int, completion: int, total: int }|null |
| 427 |
*/ |
| 428 |
function openstation_ai_result_token_usage( $result ) { |
| 429 |
try { |
| 430 |
$usage = $result->getTokenUsage(); |
| 431 |
return array( |
| 432 |
'prompt' => (int) $usage->getPromptTokens(), |
| 433 |
'completion' => (int) $usage->getCompletionTokens(), |
| 434 |
'total' => (int) $usage->getTotalTokens(), |
| 435 |
); |
| 436 |
} catch ( \Throwable $e ) { |
| 437 |
return null; |
| 438 |
} |
| 439 |
} |
| 440 |
|
| 441 |
/** |
| 442 |
* Extracts the resolved model's id + name from a generation result. |
| 443 |
* |
| 444 |
* @param mixed $result GenerativeAiResult. |
| 445 |
* @return array{ id: string, name: string }|null |
| 446 |
*/ |
| 447 |
function openstation_ai_result_model_metadata( $result ) { |
| 448 |
try { |
| 449 |
$model = $result->getModelMetadata(); |
| 450 |
return array( |
| 451 |
'id' => (string) $model->getId(), |
| 452 |
'name' => (string) $model->getName(), |
| 453 |
); |
| 454 |
} catch ( \Throwable $e ) { |
| 455 |
return null; |
| 456 |
} |
| 457 |
} |
| 458 |
|
