← All changes | includes/agents/runner.php +113 -12 1.0.1 → 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
		@@ -177,8 +189,9 @@
177	189	return $rate;
178	190	}
179	191
180	192	$instructions = openstation_agent_get_instructions( $user->ID );
	193	+ $instructions = openstation_agent_apply_vibes( $instructions, (int) $user->ID );
181	194	$abilities = openstation_agent_get_abilities( $user->ID );
182	195
183	196	list( $tool_defs, $slug_by_name ) = openstation_agent_runner_build_tools( $abilities );
184	197
		@@ -503,10 +516,15 @@
503	516	'text' => (string) $message,
504	517	);
505	518	$tool_trace = array();
506	519
	520	+ $turns_used = 0;
	521	+ $last_failure = '';
	522	+ $repeated_failures = 0;
	523	+
507	524	for ( $turn = 1; $turn <= OPENSTATION_AGENT_RUNNER_MAX_TURNS; $turn++ ) {
508		- $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 );
509	527	if ( is_wp_error( $generated ) && openstation_agent_generate_error_is_transient( $generated ) ) {
510	528	// One bounded retry for provider-side hiccups (a failed
511	529	// models-list fetch, a gateway timeout, a borderline
512	530	// refusal). A manual "try again" was already the working
		@@ -608,15 +626,27 @@
608	626	$history[] = array(
609	627	'type' => 'tool_results',
610	628	'results' => $results,
611	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	+ }
612	641	}
613	642
614		- // Cap reached with the model still asking for tools. Force one
615		- // last TOOL-LESS generate over the transcript so far: with nothing
616		- // to call, the model can only produce a final answer from what it
617		- // already gathered. A best-effort summary beats discarding the
618		- // 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
619	649	// re-searching before it answers).
620	650	$generated = openstation_agent_runner_generate( $agent_user_id, $history, array(), $instructions );
621	651	if ( is_wp_error( $generated ) && openstation_agent_generate_error_is_transient( $generated ) ) {
622	652	$generated = openstation_agent_runner_generate( $agent_user_id, $history, array(), $instructions );
		@@ -628,9 +658,9 @@
628	658	return array(
629	659	'text' => $answer['text'],
630	660	'callToActions' => $answer['callToActions'],
631	661	'toolCalls' => $tool_trace,
632		- 'turns' => OPENSTATION_AGENT_RUNNER_MAX_TURNS + 1,
	662	+ 'turns' => $turns_used + 1,
633	663	);
634	664	}
635	665
636	666	return new WP_Error(
		@@ -635,16 +665,45 @@
635	665
636	666	return new WP_Error(
637	667	'openstation_agent_runner_max_turns',
638	668	sprintf(
639		- /* translators: %d is the max-turn cap. */
	669	+ /* translators: %d is the number of turns the run made. */
640	670	__( 'Agent stopped after %d turns without a final answer.', 'desktop-mode' ),
641		- OPENSTATION_AGENT_RUNNER_MAX_TURNS
	671	+ $turns_used
642	672	)
643	673	);
644	674	}
645	675
646	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	+/**
647	706	* JSON Schema every agent's FINAL answer is constrained to (via the
648	707	* AI Client's structured output, `as_json_response()`): the markdown
649	708	* answer in `text`, plus optional `call_to_actions` the chat renders
650	709	* as buttons when the agent needs the user's confirmation instead of
		@@ -895,8 +954,19 @@
895	954	'detail' => $error->get_error_message(),
896	955	)
897	956	);
898	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	+ }
899	969	if ( 'openstation_ai_empty_answer' === $error->get_error_code() ) {
900	970	$data = $error->get_error_data();
901	971	return new WP_Error(
902	972	'openstation_agent_empty_answer',
		@@ -967,9 +1037,10 @@
967	1037	// confirmations arrive as renderable buttons, not typed-reply
968	1038	// requests. Tool-call turns are unaffected — the model either
969	1039	// calls a function or emits the JSON answer.
970	1040	openstation_agent_answer_schema(),
971		- (string) $instructions . "\n\n" . openstation_agent_answer_prompt_appendix()
	1041	+ (string) $instructions . "\n\n" . openstation_agent_answer_prompt_appendix(),
	1042	+ array( 'source' => 'agents/runner' )
972	1043	);
973	1044	}
974	1045	);
975	1046	}
		@@ -974,8 +1045,35 @@
974	1045	);
975	1046	}
976	1047
977	1048	/**
	1049	+ * Append an agent's voice line to its instructions.
	1050	+ *
	1051	+ * **After the instructions, never before.** The two can disagree — a
	1052	+ * voice that says "blunt" against a workflow that says "always explain
	1053	+ * your reasoning" — and when they do, the workflow should win. Later
	1054	+ * text is the one the model weights more heavily, so position is the
	1055	+ * whole mechanism here.
	1056	+ *
	1057	+ * The line is stored through `openstation_agent_sanitize_vibes()`,
	1058	+ * which strips line breaks. That matters more than it looks: the
	1059	+ * composed prompt marks operator turns, and a multi-line voice line
	1060	+ * could otherwise fake a turn boundary. `agentsSecurity.php` pins it.
	1061	+ *
	1062	+ * @param string $instructions The agent's system prompt.
	1063	+ * @param int $user_id Agent user id.
	1064	+ * @return string
	1065	+ */
	1066	+function openstation_agent_apply_vibes( $instructions, $user_id ) {
	1067	+ $vibes = openstation_agent_get_vibes( $user_id );
	1068	+ if ( '' === $vibes ) {
	1069	+ return $instructions;
	1070	+ }
	1071	+ $line = 'Voice: ' . $vibes;
	1072	+ return '' === $instructions ? $line : $instructions . "\n\n" . $line;
	1073	+}
	1074	+
	1075	+/**
978	1076	* Run a callback with the WordPress HTTP timeout raised for the
979	1077	* provider request it makes.
980	1078	*
981	1079	* Scoped to the generation call rather than the whole run: tool
		@@ -1214,10 +1312,13 @@
1214	1312	return $ability->execute( $args );
1215	1313	}
1216	1314
1217	1315	/**
1218		- * Append one invocation to the agent's persistent log. Most-recent
1219		- * 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.
1220	1321	*
1221	1322	* @param int $agent_user_id Agent user id.
1222	1323	* @param string $message Submitted message.
1223	1324	* @param array $result `{ text, toolCalls, turns }`.