example ai model config

Tune the AI model config

Every AI turn OpenStation generates (Copilot search, the command follow-up, the comment scorer, the Agents runner, the Drafts writing assistant) passes its model config through openstation_ai_model_config first. It defaults to empty apart from max_tokens: OpenStation pins neither provider nor model, and fills in only the output ceiling (OPENSTATION_AI_DEFAULT_MAX_TOKENS, 16384) when the filter leaves it unset, because the Anthropic provider otherwise falls back to 4096 and truncates any tool call carrying a whole post.

apply_filters( 'openstation_ai_model_config', array $config, array $context );

Key	Type	Notes
model	string|ModelInterface	Model id, or an SDK model instance. Anything else is ignored.
max_tokens	int	Output-token ceiling. Must be > 0. Defaults to 16384 when unset; a reply that hits the ceiling fails the turn as openstation_ai_output_truncated instead of arriving cut short.
temperature	float	Sampling randomness. 0.0 is valid. Anthropic's Claude Opus 4.7 and later and Claude Sonnet 5 reject any non-default value with a 400, so leave it unset for those models.
custom_options	array<string, mixed>	Provider-native parameters, forwarded verbatim. Also feeds model discovery.

Model and ceiling

add_filter(
	'openstation_ai_model_config',
	function ( $config, $context ) {
		$config['model']      = 'claude-sonnet-5';
		$config['max_tokens'] = 16000;

		return $config;
	},
	10,
	2
);

Reasoning effort (custom_options)

add_filter(
	'openstation_ai_model_config',
	function ( $config, $context ) {
		$config['max_tokens']     = 16000;
		$config['custom_options'] = array(
			'thinking'      => array( 'type' => 'adaptive' ),
			'output_config' => array( 'effort' => 'low' ),
		);

		return $config;
	},
	10,
	2
);

That pair is the Anthropic Claude 5 shape; older Anthropic models reject it and want thinking: { type: 'enabled', budget_tokens: N } instead. Check your provider's reference before copying — this is why OpenStation ships no default.

Thinking tokens count toward max_tokens. A ceiling sized for a reply without thinking can end the turn inside its reasoning, with no answer text, so leave headroom.

Spend effort only where it pays

$context['source'] names the calling path, so a long agent run and the background comment scorer need not share a setting:

add_filter(
	'openstation_ai_model_config',
	function ( $config, $context ) {
		// Short, schema-constrained classification gets the least effort;
		// a long agent run gets the most.
		$effort = 'low';
		if ( 'agents/runner' === $context['source'] ) {
			$effort = 'medium';
		}

		$config['max_tokens']     = 16000;
		$config['custom_options'] = array(
			'thinking'      => array( 'type' => 'adaptive' ),
			'output_config' => array( 'effort' => $effort ),
		);

		return $config;
	},
	10,
	2
);

source is one of ai-copilot/search, ai-copilot/followup, ai-copilot/comment-analysis, agents/runner, agents/draft, widgets/drafts-suggestions, mio/window. The rest of $context: user_id, request_id (the /ai/search correlation UUID, '' where the path mints none), has_tools, has_schema.