← All changes | includes/ai/class-claude-client.php +167 -7 1.28.0 → 2.9.0 View file →

		@@ -11,21 +11,28 @@
11	11	declare(strict_types=1);
12	12
13	13	namespace ThinkRank\AI;
14	14
	15	+use ThinkRank\AI\Traits\Request_Timeout;
	16	+
15	17	// Prevent direct access
16	18	if (!defined('ABSPATH')) {
17	19	exit;
18	20	}
19	21
	22	+require_once __DIR__ . '/traits/trait-request-timeout.php';
	23	+
20	24	/**
21	25	* Claude Client Class
22		- *
	26	+ *
23	27	* Single Responsibility: Handle Claude API communication
24		- *
	28	+ *
25	29	* @since 1.0.0
26	30	*/
27	31	class Claude_Client {
	32	+
	33	+ use Request_Timeout;
	34	+
28	35
29	36	/**
30	37	* Claude API base URL
31	38	*/
		@@ -160,17 +167,22 @@
160	167
161	168	/**
162	169	* Whether the given model rejects sampling params (temperature/top_p/top_k).
163	170	*
164		- * Anthropic removed these on Opus 4.7+, Sonnet 5, and Fable 5 — including any
165		- * date-suffixed or "-latest" alias of them — so they must be omitted from the
166		- * request body or the API returns a 400.
	171	+ * Anthropic removed these on Opus 4.7+, Opus 5, Sonnet 5, and Fable 5 —
	172	+ * including any date-suffixed or "-latest" alias of them — so they must be
	173	+ * omitted from the request body or the API returns a 400.
167	174	*
	175	+ * Every generate_* method below sends a temperature, so a model missing from
	176	+ * this list fails on its first real call rather than at save time. `claude-opus-5`
	177	+ * was absent while being offered in the UI, which made the flagship model
	178	+ * unusable (#572).
	179	+ *
168	180	* @param string $model Model ID
169	181	* @return bool
170	182	*/
171	183	private function model_rejects_sampling_params(string $model): bool {
172		- foreach (['claude-opus-4-7', 'claude-opus-4-8', 'claude-sonnet-5', 'claude-fable-5', 'claude-mythos-5'] as $prefix) {
	184	+ foreach (['claude-opus-4-7', 'claude-opus-4-8', 'claude-opus-5', 'claude-sonnet-5', 'claude-fable-5', 'claude-mythos-5'] as $prefix) {
173	185	if (strpos($model, $prefix) === 0) {
174	186	return true;
175	187	}
176	188	}
		@@ -231,8 +243,145 @@
231	243	return $this->model;
232	244	}
233	245
234	246	/**
	247	+ * Output-token ceiling per model family, longest prefix wins.
	248	+ *
	249	+ * Matched by prefix so a dated snapshot (`claude-haiku-4-5-20251001`) and a
	250	+ * point release (`claude-fable-5-1`) resolve to their family. Order matters
	251	+ * only in that lookup walks longest-first, which is what keeps
	252	+ * `claude-fable-5-1` from matching `claude-fable-5`.
	253	+ *
	254	+ * @since 2.7.0
	255	+ * @var array<string, int>
	256	+ */
	257	+ private const MODEL_OUTPUT_LIMITS = [
	258	+ // 128K output.
	259	+ 'claude-fable-5-1' => 128000,
	260	+ 'claude-fable-5' => 128000,
	261	+ 'claude-mythos-5-1' => 128000,
	262	+ 'claude-mythos-5' => 128000,
	263	+ 'claude-opus-5' => 128000,
	264	+ 'claude-opus-4-8' => 128000,
	265	+ 'claude-opus-4-7' => 128000,
	266	+ 'claude-opus-4-6' => 128000,
	267	+ 'claude-sonnet-5' => 128000,
	268	+ 'claude-sonnet-4-6' => 128000,
	269	+ // 64K output.
	270	+ 'claude-haiku-4-5' => 64000,
	271	+ ];
	272	+
	273	+ /**
	274	+ * Upper bound per use case, applied after the percentage.
	275	+ *
	276	+ * Two reasons these exist rather than letting the percentage run against a
	277	+ * 128K ceiling.
	278	+ *
	279	+ * Requests here are a single blocking HTTP call with a 120s timeout and no
	280	+ * streaming, so 0.9 x 128000 would risk running past the timeout instead of
	281	+ * returning — trading a truncation failure for a timeout failure. 16000
	282	+ * leaves room for the brief's JSON plus reasoning tokens while staying
	283	+ * answerable; raise it only alongside streaming.
	284	+ *
	285	+ * And correcting the ceiling would otherwise inflate every other use case
	286	+ * as a side effect — seo_metadata would jump from ~1,229 tokens to ~19,200
	287	+ * purely because this bug was fixed. Metadata generation already works, so
	288	+ * it keeps its cost profile (#665).
	289	+ *
	290	+ * @since 2.7.0
	291	+ * @var array<string, int>
	292	+ */
	293	+ private const USE_CASE_TOKEN_CAPS = [
	294	+ 'content_brief' => 16000,
	295	+ 'llms_txt' => 16000,
	296	+ 'analysis' => 8000,
	297	+ 'seo_metadata' => 4000,
	298	+ 'optimization' => 4000,
	299	+ 'default' => 4000,
	300	+ ];
	301	+
	302	+ /**
	303	+ * Ceiling for a model this table does not know.
	304	+ *
	305	+ * The previous behaviour for every model, kept for older and unrecognised
	306	+ * ones: 8192 is accepted without an extended-output beta header, so it is
	307	+ * the safe answer when we cannot identify the family.
	308	+ *
	309	+ * @since 2.7.0
	310	+ * @var int
	311	+ */
	312	+ private const FALLBACK_OUTPUT_LIMIT = 8192;
	313	+
	314	+ /**
	315	+ * Maximum completion (output) tokens accepted for a single Claude request.
	316	+ *
	317	+ * This returned a flat 8192 for every model and ignored $model entirely, so
	318	+ * Content Brief was capped at a fraction of the available budget and
	319	+ * truncated before its structured JSON completed — on every Claude model,
	320	+ * every time. Current models also emit reasoning tokens from the same
	321	+ * output budget, which is why it failed so reliably rather than
	322	+ * intermittently (#665).
	323	+ *
	324	+ * @param string $model Model ID.
	325	+ * @return int Maximum output tokens.
	326	+ */
	327	+ private function get_max_completion_tokens(string $model): int {
	328	+ $model = strtolower(trim($model));
	329	+
	330	+ if ('' === $model) {
	331	+ return self::FALLBACK_OUTPUT_LIMIT;
	332	+ }
	333	+
	334	+ $limits = self::MODEL_OUTPUT_LIMITS;
	335	+
	336	+ // Longest prefix first, so a point release never matches the shorter
	337	+ // family id that is a prefix of it.
	338	+ uksort(
	339	+ $limits,
	340	+ static function (string $a, string $b): int {
	341	+ return strlen($b) <=> strlen($a);
	342	+ }
	343	+ );
	344	+
	345	+ foreach ($limits as $prefix => $limit) {
	346	+ if (0 === strpos($model, $prefix)) {
	347	+ return $limit;
	348	+ }
	349	+ }
	350	+
	351	+ return self::FALLBACK_OUTPUT_LIMIT;
	352	+ }
	353	+
	354	+ /**
	355	+ * Recommended output-token budget for a given use case.
	356	+ *
	357	+ * Mirrors the other clients so the Content Brief generator no longer falls
	358	+ * back to a hardcoded, model-blind budget for Claude (issue #287). Each
	359	+ * value is a fraction of the model's completion ceiling.
	360	+ *
	361	+ * @param string $use_case e.g. 'content_brief', 'seo_metadata', 'analysis'.
	362	+ * @return int Recommended max output tokens.
	363	+ */
	364	+ public function get_recommended_tokens(string $use_case): int {
	365	+ $max_tokens = $this->get_max_completion_tokens($this->model);
	366	+
	367	+ $recommendations = [
	368	+ 'content_brief' => 0.9, // Comprehensive brief incl. a full article body.
	369	+ 'seo_metadata' => 0.15,
	370	+ 'analysis' => 0.25,
	371	+ 'llms_txt' => 0.5,
	372	+ 'optimization' => 0.15,
	373	+ ];
	374	+ $percentage = $recommendations[$use_case] ?? 0.15;
	375	+
	376	+ $budget = (int) ($max_tokens * $percentage);
	377	+
	378	+ $cap = self::USE_CASE_TOKEN_CAPS[$use_case] ?? self::USE_CASE_TOKEN_CAPS['default'];
	379	+
	380	+ return max(1, min($budget, $cap));
	381	+ }
	382	+
	383	+ /**
235	384	* Test API connection
236	385	*
237	386	* @return bool True if connection successful
238	387	*/
		@@ -254,8 +403,14 @@
254	403	* @return array Response data
255	404	* @throws \Exception If request fails
256	405	*/
257	406	private function make_request(string $endpoint, array $body = []): array {
	407	+ // The user's daily ceiling and kill switch are enforced here, at the
	408	+ // one place every outbound Claude call passes through, so no feature
	409	+ // path can bypass them by forgetting to ask first (#448).
	410	+ Spend_Guard::guard();
	411	+ Spend_Guard::record();
	412	+
258	413	$url = self::API_BASE_URL . '/' . ltrim($endpoint, '/');
259	414
260	415	$args = [
261	416	'timeout' => $this->timeout,
		@@ -322,9 +477,14 @@
322	477
323	478	$is_transient = false;
324	479	$retry_after = 0;
325	480	if (is_wp_error($response)) {
326		- $is_transient = true;
	481	+ // A client-side timeout means the work genuinely needs longer
	482	+ // than the budget we allowed; re-running the identical prompt,
	483	+ // model and budget just times out again and multiplies the
	484	+ // wait (issue #288). Do not retry a timeout. Other WP_Error
	485	+ // results — DNS, connection refused, TLS — stay retryable.
	486	+ $is_transient = !$this->is_timeout_error($response);
327	487	} else {
328	488	$status = wp_remote_retrieve_response_code($response);
329	489	if (429 === $status || $status >= 500) {
330	490	$is_transient = true;