← All changes | includes/agents/runner.php +83 -11 1.1.4 → 1.1.12 View file →

		@@ -67,8 +67,20 @@
67	67	*/
68	68	const OPENSTATION_AGENT_RUNNER_MAX_TURNS = 8;
69	69
70	70	/**
	71	+ * Consecutive turns in which every tool call failed with the same
	72	+ * errors before the loop gives up on tools and asks for a final answer.
	73	+ *
	74	+ * A model reads a tool error and usually fixes its next call. One that
	75	+ * sends the same failing call a third time is not going to fix it on
	76	+ * the eighth: the Localizer once spent seven turns re-sending a
	77	+ * `create_post` call the ability rejected identically every time. Three
	78	+ * is one honest correction attempt plus proof it did not help.
	79	+ */
	80	+const OPENSTATION_AGENT_RUNNER_STUCK_TURNS = 3;
	81	+
	82	+/**
71	83	* Seconds to allow one provider generation request, replacing the
72	84	* WordPress HTTP default of 5.
73	85	*
74	86	* The AI Client's HTTP adapter issues provider calls through
		@@ -504,10 +516,15 @@
504	516	'text' => (string) $message,
505	517	);
506	518	$tool_trace = array();
507	519
	520	+ $turns_used = 0;
	521	+ $last_failure = '';
	522	+ $repeated_failures = 0;
	523	+
508	524	for ( $turn = 1; $turn <= OPENSTATION_AGENT_RUNNER_MAX_TURNS; $turn++ ) {
509		- $generated = openstation_agent_runner_generate( $agent_user_id, $history, $tool_defs, $instructions );
	525	+ $turns_used = $turn;
	526	+ $generated = openstation_agent_runner_generate( $agent_user_id, $history, $tool_defs, $instructions );
510	527	if ( is_wp_error( $generated ) && openstation_agent_generate_error_is_transient( $generated ) ) {
511	528	// One bounded retry for provider-side hiccups (a failed
512	529	// models-list fetch, a gateway timeout, a borderline
513	530	// refusal). A manual "try again" was already the working
		@@ -609,15 +626,27 @@
609	626	$history[] = array(
610	627	'type' => 'tool_results',
611	628	'results' => $results,
612	629	);
	630	+
	631	+ $failure = openstation_agent_runner_failure_signature( $results );
	632	+ if ( '' !== $failure && $failure === $last_failure ) {
	633	+ ++$repeated_failures;
	634	+ } else {
	635	+ $repeated_failures = '' === $failure ? 0 : 1;
	636	+ }
	637	+ $last_failure = $failure;
	638	+ if ( $repeated_failures >= OPENSTATION_AGENT_RUNNER_STUCK_TURNS ) {
	639	+ break;
	640	+ }
613	641	}
614	642
615		- // Cap reached with the model still asking for tools. Force one
616		- // last TOOL-LESS generate over the transcript so far: with nothing
617		- // to call, the model can only produce a final answer from what it
618		- // already gathered. A best-effort summary beats discarding the
619		- // whole run (observed on Anthropic: a model happily spends the cap
	643	+ // Cap reached (or the model stuck re-sending the same failing
	644	+ // call) with it still asking for tools. Force one last TOOL-LESS
	645	+ // generate over the transcript so far: with nothing to call, the
	646	+ // model can only produce a final answer from what it already
	647	+ // gathered. A best-effort summary beats discarding the whole run
	648	+ // (observed on Anthropic: a model happily spends the cap
620	649	// re-searching before it answers).
621	650	$generated = openstation_agent_runner_generate( $agent_user_id, $history, array(), $instructions );
622	651	if ( is_wp_error( $generated ) && openstation_agent_generate_error_is_transient( $generated ) ) {
623	652	$generated = openstation_agent_runner_generate( $agent_user_id, $history, array(), $instructions );
		@@ -629,9 +658,9 @@
629	658	return array(
630	659	'text' => $answer['text'],
631	660	'callToActions' => $answer['callToActions'],
632	661	'toolCalls' => $tool_trace,
633		- 'turns' => OPENSTATION_AGENT_RUNNER_MAX_TURNS + 1,
	662	+ 'turns' => $turns_used + 1,
634	663	);
635	664	}
636	665
637	666	return new WP_Error(
		@@ -636,16 +665,45 @@
636	665
637	666	return new WP_Error(
638	667	'openstation_agent_runner_max_turns',
639	668	sprintf(
640		- /* translators: %d is the max-turn cap. */
	669	+ /* translators: %d is the number of turns the run made. */
641	670	__( 'Agent stopped after %d turns without a final answer.', 'desktop-mode' ),
642		- OPENSTATION_AGENT_RUNNER_MAX_TURNS
	671	+ $turns_used
643	672	)
644	673	);
645	674	}
646	675
647	676	/**
	677	+ * Fingerprint of a turn in which every tool call failed: the sorted
	678	+ * tool names with their error messages. Empty when any call succeeded,
	679	+ * so a turn that got something done never counts as stuck.
	680	+ *
	681	+ * Arguments are deliberately left out. The failure that motivated this
	682	+ * had the model vary a title between attempts while the ability
	683	+ * rejected each one for the same missing field, and that IS the same
	684	+ * failure.
	685	+ *
	686	+ * @param array $results Tool results of one turn (`{ name, response }` rows).
	687	+ * @return string
	688	+ */
	689	+function openstation_agent_runner_failure_signature( array $results ) {
	690	+ if ( empty( $results ) ) {
	691	+ return '';
	692	+ }
	693	+ $failures = array();
	694	+ foreach ( $results as $row ) {
	695	+ $response = isset( $row['response'] ) ? $row['response'] : null;
	696	+ if ( ! is_array( $response ) || ! isset( $response['error'] ) ) {
	697	+ return '';
	698	+ }
	699	+ $failures[] = ( isset( $row['name'] ) ? (string) $row['name'] : '' ) . "\0" . (string) $response['error'];
	700	+ }
	701	+ sort( $failures );
	702	+ return implode( "\n", $failures );
	703	+}
	704	+
	705	+/**
648	706	* JSON Schema every agent's FINAL answer is constrained to (via the
649	707	* AI Client's structured output, `as_json_response()`): the markdown
650	708	* answer in `text`, plus optional `call_to_actions` the chat renders
651	709	* as buttons when the agent needs the user's confirmation instead of
		@@ -896,8 +954,19 @@
896	954	'detail' => $error->get_error_message(),
897	955	)
898	956	);
899	957	}
	958	+ if ( 'openstation_ai_output_truncated' === $error->get_error_code() ) {
	959	+ $data = $error->get_error_data();
	960	+ return new WP_Error(
	961	+ 'openstation_agent_output_truncated',
	962	+ __( 'The reply ran past the output-token limit before it finished, so it was discarded rather than acted on incomplete. Ask for something shorter, or raise max_tokens with the openstation_ai_model_config filter.', 'desktop-mode' ),
	963	+ array(
	964	+ 'status' => 502,
	965	+ 'detail' => is_array( $data ) && isset( $data['detail'] ) ? (string) $data['detail'] : '',
	966	+ )
	967	+ );
	968	+ }
900	969	if ( 'openstation_ai_empty_answer' === $error->get_error_code() ) {
901	970	$data = $error->get_error_data();
902	971	return new WP_Error(
903	972	'openstation_agent_empty_answer',
		@@ -1243,10 +1312,13 @@
1243	1312	return $ability->execute( $args );
1244	1313	}
1245	1314
1246	1315	/**
1247		- * Append one invocation to the agent's persistent log. Most-recent
1248		- * entries surface in the chat window's history strip.
	1316	+ * Append one invocation to the agent's persistent log: an audit trail
	1317	+ * of who ran the agent and what came back, capped at
	1318	+ * OPENSTATION_AGENT_RUNNER_LOG_CAP entries and readable from PHP with
	1319	+ * {@see openstation_agent_runner_get_log()}. No UI shows it; the chat
	1320	+ * window's history is the human's saved conversations, not this log.
1249	1321	*
1250	1322	* @param int $agent_user_id Agent user id.
1251	1323	* @param string $message Submitted message.
1252	1324	* @param array $result `{ text, toolCalls, turns }`.