config

type AlibabaClientConfig struct {
	// APIKey is the API key for the Alibaba provider.
	APIKey string `yaml:"api-key" validate:"required"`
	// Endpoint specifies the network endpoint URL for the API.
	Endpoint string `yaml:"endpoint" validate:"omitempty,url"`
}

type AlibabaModelParams struct {
	// TextResponseFormat indicates whether to use plain-text response format
	// for compatibility with models that do not support JSON mode (e.g., when
	// thinking is enabled on certain Qwen models).
	TextResponseFormat bool `yaml:"text-response-format" validate:"omitempty"`

	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Notes: Higher values (e.g. 0.8) make outputs more random; lower values
	// (e.g. 0.2) make outputs more focused and deterministic.
	// Notes: Use either `Temperature` or `TopP`, not both, for sampling control.
	// Valid range: 0.0 — 2.0. Default: 1.0.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=2"`

	// TopP controls diversity via nucleus sampling (probability mass cutoff).
	// Notes: Use either `Temperature` or `TopP`, not both, for sampling control.
	// Valid range: 0.0 — 1.0. Default varies by model.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Notes: Positive values encourage the model to introduce new topics.
	// Valid range: [-2.0, 2.0]. Default: 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in text so far.
	// Notes: Positive values encourage model to use less frequent tokens.
	// Valid range: [-2.0, 2.0]. Default: 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`

	// MaxTokens controls the maximum number of tokens available to the model for generating a response.
	MaxTokens *int32 `yaml:"max-tokens" validate:"omitempty,min=0"`

	// Seed makes text generation more deterministic. If specified, the system will
	// attempt to return the same result for the same inputs with the same seed value and parameters.
	Seed *uint32 `yaml:"seed" validate:"omitempty"`

	// DisableLegacyJsonMode toggles a compatibility behavior for certain models.
	// In the legacy mode (default), a standard response format instruction is included
	// in the prompt to guide the model to respond in a structured JSON format.
	// This is necessary for models that do not fully support schema-based structured JSON output.
	DisableLegacyJsonMode *bool `yaml:"disable-legacy-json-mode" validate:"omitempty"`
}

type AnthropicClientConfig struct {
	// APIKey is the API key for the Anthropic generative models provider.
	APIKey string `yaml:"api-key" validate:"required"`
	// RequestTimeout specifies the timeout for API requests.
	RequestTimeout *time.Duration `yaml:"request-timeout" validate:"omitempty"`
}

type AnthropicModelParams struct {
	// MaxTokens controls the maximum number of tokens available to the model for generating a response.
	// This includes the thinking budget for reasoning models.
	MaxTokens *int64 `yaml:"max-tokens" validate:"omitempty,min=0"`

	// ThinkingBudgetTokens specifies the number of tokens the model can use for its internal reasoning process.
	// It must be at least 1024 and less than `MaxTokens`.
	// If set, this enables enhanced reasoning capabilities for the model.
	ThinkingBudgetTokens *int64 `yaml:"thinking-budget-tokens" validate:"omitempty,min=1024,ltfield=MaxTokens"`

	// Temperature controls the randomness or "creativity" of responses.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// The default value is 1.0.
	// It is generally recommended to alter this or `TopP` but not both.
	Temperature *float64 `yaml:"temperature" validate:"omitempty,min=0,max=1"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// You usually only need to use `Temperature`.
	TopP *float64 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// TopK limits response tokens to top K options for each token position.
	// Higher values allow more diverse outputs by considering more token options.
	// You usually only need to use `Temperature`.
	TopK *int64 `yaml:"top-k" validate:"omitempty,min=0"`
}

type AppConfig struct {
	// LogFile specifies path to the log file.
	LogFile string `yaml:"log-file" validate:"omitempty,filepath"`

	// OutputDir specifies directory where results will be saved.
	OutputDir string `yaml:"output-dir" validate:"required"`

	// OutputBaseName specifies base filename for result files.
	OutputBaseName string `yaml:"output-basename" validate:"omitempty,filepath"`

	// TaskSource specifies path to the task definitions file.
	TaskSource string `yaml:"task-source" validate:"required,filepath"`

	// Providers lists configurations for AI providers whose models will be used
	// to execute tasks during the trial run.
	Providers []ProviderConfig `yaml:"providers" validate:"required,dive"`

	// Judges lists LLM configurations for semantic evaluation of open-ended task responses.
	Judges []JudgeConfig `yaml:"judges" validate:"omitempty,unique=Name,dive"`

	// Tools lists common tool configurations available to tasks.
	Tools []ToolConfig `yaml:"tools" validate:"omitempty,unique=Name,dive"`
}

type ClientConfig interface{}

type Config struct {
	// Config contains application-wide settings.
	Config AppConfig `yaml:"config" validate:"required"`
}

type DeepseekClientConfig struct {
	// APIKey is the API key for the DeepSeek generative models provider.
	APIKey string `yaml:"api-key" validate:"required"`
	// RequestTimeout specifies the timeout for API requests.
	RequestTimeout *time.Duration `yaml:"request-timeout" validate:"omitempty"`
}

type DeepseekModelParams struct {
	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Values range from 0.0 to 2.0, with lower values making the output more focused.
	// The default value is 1.0.
	// Recommended values by use case:
	// - 0.0: Coding / Math (best for precise, deterministic outputs)
	// - 1.0: Data Cleaning / Data Analysis
	// - 1.3: General Conversation / Translation
	// - 1.5: Creative Writing / Poetry (more varied and creative outputs)
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=2"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// You usually only need to use `Temperature`.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use new tokens,
	// increasing the model's likelihood to talk about new topics.
	// The default value is 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use less frequent tokens,
	// decreasing the model's likelihood to repeat the same line verbatim.
	// The default value is 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`
}

type GoogleAIClientConfig struct {
	// APIKey is the API key for the Google AI generative models provider.
	APIKey string `yaml:"api-key" validate:"required"`
}

type GoogleAIModelParams struct {
	// TextResponseFormat indicates whether to use plain-text response format
	// for compatibility with models that do not support JSON.
	// This setting applies to all tasks, including those with and without tools enabled.
	TextResponseFormat bool `yaml:"text-response-format" validate:"omitempty"`

	// TextResponseFormatWithTools forces plain-text response format when tools are enabled.
	// If true, forces plain-text mode when tools are used (required for pre-Gemini 3 models).
	// If false or unset, uses JSON schema mode with tools (Gemini 3+ default behavior).
	// This setting only applies to tasks with tools enabled.
	TextResponseFormatWithTools bool `yaml:"text-response-format-with-tools" validate:"omitempty"`

	// ThinkingLevel controls the maximum depth of the model's internal reasoning process.
	// Valid values: "low", "high". Gemini 3 Pro defaults to "high" if not specified.
	// - "low": Minimizes latency and cost, best for simple instruction following
	// - "high": Maximizes reasoning depth, the model may take longer but output is more carefully reasoned
	ThinkingLevel *string `yaml:"thinking-level" validate:"omitempty,oneof=low high"`

	// MediaResolution controls the maximum number of tokens allocated per input image or video frame.
	// Valid values: "low", "medium", "high". Higher resolutions improve fine text reading and small detail
	// identification but increase token usage and latency.
	// - "low": 280 tokens for images, 70 tokens per video frame
	// - "medium": 560 tokens for images, 70 tokens per video frame (same as low for video)
	// - "high": 1120 tokens for images, 280 tokens per video frame
	// If unspecified, the model uses optimal defaults based on media type.
	MediaResolution *string `yaml:"media-resolution" validate:"omitempty,oneof=low medium high"`

	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Values range from 0.0 to 2.0, with lower values making the output more focused and deterministic.
	// The default value is typically around 1.0.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=2"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// The default value is typically around 1.0.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// TopK limits response tokens to top K options for each token position.
	// Higher values allow more diverse outputs by considering more token options.
	TopK *int32 `yaml:"top-k" validate:"omitempty,min=0"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Positive values discourage the use of tokens that have already been used in the response,
	// increasing the vocabulary. Negative values encourage the use of tokens that have already been used.
	// This penalty is binary on/off and not dependent on the number of times the token is used.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Positive values discourage the use of tokens that have already been used, proportional to
	// the number of times the token has been used. Negative values encourage the model to reuse tokens.
	// This differs from PresencePenalty as it scales with frequency.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty"`

	// Seed is used for deterministic generation. When set to a specific value, the model
	// makes a best effort to provide the same response for repeated requests.
	// If not set, a randomly generated seed is used.
	Seed *int32 `yaml:"seed" validate:"omitempty"`
}

type JudgeConfig struct {
	// Name is the unique identifier for this judge configuration.
	Name string `yaml:"name" validate:"required"`

	// Provider encapsulates the provider configuration for the judge.
	Provider ProviderConfig `yaml:"provider" validate:"required"`
}

type JudgePrompt struct {
	// Template is the template string for the judge prompt.
	Template *string `yaml:"template" validate:"omitempty"`

	// VerdictFormat specifies how the judge should format its evaluation response.
	VerdictFormat *ResponseFormat `yaml:"verdict-format" validate:"omitempty"`

	// PassingVerdicts is the set of verdicts that count as a pass.
	PassingVerdicts *utils.ValueSet `yaml:"passing-verdicts" validate:"omitempty"`
	// contains filtered or unexported fields
}

type JudgeSelector struct {
	// Enabled determines whether judge evaluation is enabled.
	Enabled *bool `yaml:"enabled" validate:"omitempty"`

	// Name specifies the name of the judge configuration to use.
	Name *string `yaml:"name" validate:"omitempty"`

	// Variant specifies the run variant name from the judge's provider configuration.
	Variant *string `yaml:"variant" validate:"omitempty"`

	// Prompt specifies the judge prompt configuration.
	Prompt JudgePrompt `yaml:"prompt" validate:"omitempty"`
}

type LegacyJsonMode int

type MistralAIClientConfig struct {
	// APIKey is the API key for the Mistral AI generative models provider.
	APIKey string `yaml:"api-key" validate:"required"`
}

type MistralAIModelParams struct {
	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Values range from 0.0 to 1.5, with lower values making the output more focused and deterministic.
	// The default value varies depending on the model.
	// It is generally recommended to alter this or `TopP` but not both.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=1.5"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// The default value is 1.0.
	// It is generally recommended to alter this or `Temperature` but not both.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// MaxTokens controls the maximum number of tokens to generate in the completion.
	// The token count of the prompt plus max_tokens cannot exceed the model's context length.
	MaxTokens *int32 `yaml:"max-tokens" validate:"omitempty,min=0"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use new tokens,
	// increasing the model's likelihood to talk about new topics.
	// The default value is 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use less frequent tokens,
	// decreasing the model's likelihood to repeat the same line verbatim.
	// The default value is 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`

	// RandomSeed provides the seed to use for random sampling.
	// If set, requests will generate deterministic results.
	RandomSeed *int32 `yaml:"random-seed" validate:"omitempty"`

	// PromptMode sets the prompt mode for the request.
	// When set to "reasoning", a system prompt will be used to instruct the model to reason if supported.
	PromptMode *string `yaml:"prompt-mode" validate:"omitempty,oneof=reasoning"`

	// SafePrompt controls whether to inject a safety prompt before all conversations.
	SafePrompt *bool `yaml:"safe-prompt" validate:"omitempty"`
}

type ModelParams interface{}

type MoonshotAIClientConfig struct {
	// APIKey is the API key for the Moonshot AI provider.
	APIKey string `yaml:"api-key" validate:"required"`
	// Endpoint specifies the network endpoint URL for the API.
	Endpoint string `yaml:"endpoint" validate:"omitempty,url"`
}

type MoonshotAIModelParams struct {
	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Values range from 0.0 to 1.0, with lower values making the output more focused and deterministic.
	// The default value is 0.0.
	// Moonshot AI recommends 0.6 for kimi-k2 models and 1.0 for kimi-k2-thinking models.
	// It is generally recommended to alter this or `TopP` but not both.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=1"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// The default value is 1.0.
	// It is generally recommended to alter this or `Temperature` but not both.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// MaxTokens controls the maximum number of tokens available to the model for generating a response.
	MaxTokens *int32 `yaml:"max-tokens" validate:"omitempty,min=0"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use new tokens,
	// increasing the model's likelihood to talk about new topics.
	// The default value is 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use less frequent tokens,
	// decreasing the model's likelihood to repeat the same line verbatim.
	// The default value is 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`
}

type OpenAIClientConfig struct {
	// APIKey is the API key for the OpenAI provider.
	APIKey string `yaml:"api-key" validate:"required"`
}

type OpenAIModelParams struct {
	// ReasoningEffort controls effort level on reasoning for reasoning models.
	// Valid values are: "none", "minimal", "low", "medium", "high", "xhigh".
	ReasoningEffort *string `yaml:"reasoning-effort" validate:"omitempty,oneof=none minimal low medium high xhigh"`

	// Verbosity determines how many output tokens are generated.
	// Valid values are: "low", "medium", "high".
	// Note: May not be supported by legacy models.
	Verbosity *string `yaml:"verbosity" validate:"omitempty,oneof=low medium high"`

	// TextResponseFormat indicates whether to use plain-text response format
	// for compatibility with models that do not support JSON.
	TextResponseFormat bool `yaml:"text-response-format" validate:"omitempty"`

	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Values range from 0.0 to 2.0, with lower values making the output more focused and deterministic.
	// The default value is 1.0.
	// It is generally recommended to alter this or `TopP` but not both.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=2"`

	// TopP controls diversity via nucleus sampling.
	// Values range from 0.0 to 1.0, with lower values making the output more focused.
	// The default value is 1.0.
	// It is generally recommended to alter this or `Temperature` but not both.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use new tokens,
	// increasing the model's likelihood to talk about new topics.
	// The default value is 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Values range from -2.0 to 2.0, with positive values encouraging the model to use less frequent tokens,
	// decreasing the model's likelihood to repeat the same line verbatim.
	// The default value is 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`

	// MaxCompletionTokens controls the maximum number of tokens available to the model for generating a response,
	// including visible output tokens and reasoning tokens.
	MaxCompletionTokens *int32 `yaml:"max-completion-tokens" validate:"omitempty,min=1"`

	// MaxTokens controls the maximum number of tokens available to the model for generating a response.
	// This field is for internal use only and not exposed in YAML configuration.
	//
	// Deprecated: Use `MaxCompletionTokens` instead for user configuration.
	MaxTokens *int32 `yaml:"-"`

	// Seed makes text generation more deterministic. If specified, the system will
	// attempt to return the same result for the same inputs with the same seed value and parameters.
	// This field is for internal use only and not exposed in YAML configuration.
	Seed *int64 `yaml:"-"`

	// LegacyJsonMode specifies a compatibility mode for JSON response formatting.
	// When set to LegacyJsonSchema, adds format instruction to prompt while keeping json_schema response format.
	// When set to LegacyJsonObject, adds format instruction to prompt and uses json_object response format.
	// When nil, uses default behavior (no format instruction, json_schema response format).
	// This field is for internal use only and not exposed in YAML configuration.
	LegacyJsonMode *LegacyJsonMode `yaml:"-"`
}

type ProviderConfig struct {
	// Name specifies unique identifier of the provider.
	Name string `yaml:"name" validate:"required,oneof=openai google anthropic deepseek mistralai xai alibaba moonshotai"`

	// ClientConfig holds provider-specific client settings.
	ClientConfig ClientConfig `yaml:"client-config" validate:"required"`

	// Runs lists run configurations for this provider.
	Runs []RunConfig `yaml:"runs" validate:"required,unique=Name,dive"`

	// Disabled indicates if all runs should be disabled by default.
	Disabled bool `yaml:"disabled" validate:"omitempty"`

	// RetryPolicy specifies default retry behavior for all runs in this provider.
	RetryPolicy RetryPolicy `yaml:"retry-policy" validate:"omitempty"`
}

type ResponseFormat struct {
	// contains filtered or unexported fields
}

type RetryPolicy struct {
	// MaxRetryAttempts specifies the maximum number of retry attempts.
	// Value of 0 means no retry attempts will be made.
	MaxRetryAttempts uint `yaml:"max-retry-attempts" validate:"omitempty,min=0"`

	// InitialDelaySeconds specifies the initial delay in seconds before the first retry attempt.
	InitialDelaySeconds int `yaml:"initial-delay-seconds" validate:"omitempty,gt=0"`
}

type RunConfig struct {
	// Name is a display-friendly identifier shown in results.
	Name string `yaml:"name" validate:"required"`

	// Model specifies target model's identifier.
	Model string `yaml:"model" validate:"required"`

	// MaxRequestsPerMinute limits the number of API requests per minute sent to this specific model.
	// Value of 0 means no rate limiting will be applied.
	MaxRequestsPerMinute int `yaml:"max-requests-per-minute" validate:"omitempty,numeric,min=0"`

	// Disabled indicates if this run configuration should be skipped.
	// If set, overrides the parent ProviderConfig.Disabled value.
	Disabled *bool `yaml:"disabled" validate:"omitempty"`

	// ModelParams holds any model-specific configuration parameters.
	ModelParams ModelParams `yaml:"model-parameters" validate:"omitempty"`

	// RetryPolicy specifies retry behavior on transient errors.
	// If set, overrides the parent ProviderConfig.RetryPolicy value.
	RetryPolicy *RetryPolicy `yaml:"retry-policy" validate:"omitempty"`
}

type SystemPrompt struct {
	// Template is the template string for the system prompt.
	// It can reference `{{.ResponseResultFormat}}` to include the task's response format.
	Template *string `yaml:"template" validate:"omitempty"`

	// EnableFor controls when system prompt should be sent to AI models.
	// - "all": system prompt is sent for all tasks
	// - "text": system prompt is sent only for tasks with plain text response format
	// - "none": system prompt is never sent
	// Defaults to "text" when not specified.
	EnableFor *SystemPromptEnabledFor `yaml:"enable-for" validate:"omitempty,oneof=all text none"`
}

type SystemPromptEnabledFor string

type Task struct {
	// Name is a display-friendly identifier shown in results.
	Name string `yaml:"name" validate:"required"`

	// Prompt that will be sent to the AI model.
	Prompt string `yaml:"prompt" validate:"required"`

	// ResponseResultFormat specifies how the AI should format the final answer to the prompt.
	// Can be either a plain text instruction or a JSON schema object.
	ResponseResultFormat ResponseFormat `yaml:"response-result-format" validate:"required"`

	// ExpectedResult is the set of accepted valid answers for the prompt.
	// For plain text format: contains string values that must follow the `ResponseResultFormat` instruction precisely.
	// For structured schema format: contains object values that must be valid according to the `ResponseResultFormat` schema.
	// Only one needs to match for the response to be considered correct.
	ExpectedResult utils.ValueSet `yaml:"expected-result" validate:"required"`

	// Disabled indicates whether this specific task should be skipped.
	// If set, overrides the global TaskConfig.Disabled value.
	Disabled *bool `yaml:"disabled" validate:"omitempty"`

	// ValidationRules are validation settings for this specific task.
	// If set, overrides the global TaskConfig.ValidationRules values.
	ValidationRules *ValidationRules `yaml:"validation-rules" validate:"omitempty"`

	// SystemPrompt is the system prompt configuration for this specific task.
	// If set, overrides the global TaskConfig.SystemPrompt values.
	SystemPrompt *SystemPrompt `yaml:"system-prompt" validate:"omitempty"`

	// Files is a list of files to be included with the prompt.
	// This is primarily used for images but can support other file types
	// depending on the provider's capabilities.
	Files []TaskFile `yaml:"files" validate:"omitempty,unique=Name,dive"`

	// ToolSelector is the tool selector configuration for this specific task.
	// If set, overrides the global TaskConfig.ToolSelector values.
	ToolSelector *ToolSelector `yaml:"tool-selector" validate:"omitempty"`
	// contains filtered or unexported fields
}

type TaskConfig struct {
	// Tasks is a list of tasks to be executed.
	Tasks []Task `yaml:"tasks" validate:"required,unique=Name,dive"`

	// Disabled indicates whether all tasks should be disabled by default.
	// Individual tasks can override this setting.
	Disabled bool `yaml:"disabled" validate:"omitempty"`

	// ValidationRules are default validation settings for all tasks.
	// Individual tasks can override these settings.
	ValidationRules ValidationRules `yaml:"validation-rules" validate:"omitempty"`

	// SystemPrompt is the default system prompt configuration for all tasks.
	// Individual tasks can override this configuration.
	SystemPrompt SystemPrompt `yaml:"system-prompt" validate:"omitempty"`

	// ToolSelector is the default tool selector configuration for all tasks.
	// Individual tasks can override this configuration.
	ToolSelector ToolSelector `yaml:"tool-selector" validate:"omitempty"`
}

type TaskFile struct {
	// Name is a unique identifier for the file, used to reference it in prompts.
	Name string `yaml:"name" validate:"required"`

	// URI is the path or URL to the file.
	URI URI `yaml:"uri" validate:"required"`

	// Type is the MIME type of the file.
	// If not provided, it will be inferred from the file extension or content.
	Type string `yaml:"type" validate:"omitempty"`
	// contains filtered or unexported fields
}

type Tasks struct {
	// TaskConfig contains all task definitions and settings.
	TaskConfig TaskConfig `yaml:"task-config" validate:"required"`
}

type ToolConfig struct {
	// Name is the unique identifier for the tool.
	Name string `yaml:"name" validate:"required"`
	// Image is the name of the Docker image to use for the tool.
	Image string `yaml:"image" validate:"required"`
	// Description describes what the tool does. For optimal LLM understanding and tool selection,
	// provide extremely detailed descriptions including:
	// - What the tool does and its primary purpose
	// - When it should be used (and when it shouldn't)
	// - What each parameter in the schema means and how it affects behavior
	// - Any important caveats, limitations, or side effects
	// - Examples of usage if helpful
	// Aim for 3-4 sentences per tool description. Be specific and avoid ambiguity
	// to help the LLM choose the correct tool and provide appropriate parameters.
	Description string `yaml:"description" validate:"required"`
	// Parameters is the JSON schema for the tool's input parameters. Follow these best practices
	// to improve LLM parameter generation accuracy:
	// - Use standard JSON Schema format with detailed "description" fields for each parameter
	// - Specify precise types (string, integer, boolean, array, object)
	// - Use "enum" arrays for parameters with fixed sets of allowed values
	// - Include examples and constraints in parameter descriptions (e.g., "The city name, e.g., 'San Francisco'")
	// - Clearly mark all required parameters in the "required" array
	// - Use "additionalProperties": false for objects to prevent unexpected parameters
	// - Provide comprehensive descriptions that explain parameter purpose and format
	Parameters map[string]interface{} `yaml:"parameters" validate:"required"`
	// ParameterFiles maps parameter field names to file paths where argument values should be written.
	// This allows passing large or complex data to tools via files instead of inline JSON.
	// The tool's command should read these files as needed.
	ParameterFiles map[string]string `yaml:"parameter-files,omitempty"`
	// AuxiliaryDir specifies the directory path where task files will be automatically available.
	// If set, all files attached to a task will be copied to this directory using each file's
	// `TaskFile.Name` exactly as provided.
	// This directory is ephemeral: files are reset between tool calls and do not persist
	// across multiple invocations.
	AuxiliaryDir string `yaml:"auxiliary-dir,omitempty"`
	// SharedDir specifies the directory path that persists across all tool calls within a single task.
	// If set, files created in this directory will be available for any subsequent tool calls but
	// will be removed when the task completes.
	SharedDir string `yaml:"shared-dir,omitempty"`
	// Command specifies the command to execute as a list of its components.
	Command []string `yaml:"command,omitempty"`
	// Env specifies additional environment variables to set.
	Env map[string]string `yaml:"env,omitempty"`
}

type ToolSelection struct {
	// Name of the tool to select.
	Name string `yaml:"name" validate:"required"`
	// Disabled determines whether this specific tool is disabled.
	// If nil, uses the value from the ToolSelector.
	Disabled *bool `yaml:"disabled" validate:"omitempty"`
	// MaxCalls is the maximum number of times this tool can be called per task.
	// If nil, there is no limit.
	MaxCalls *int `yaml:"max-calls" validate:"omitempty,min=1"`
	// Timeout is the timeout for a single tool invocation.
	// If nil, there is no timeout.
	Timeout *time.Duration `yaml:"timeout" validate:"omitempty"`
	// MaxMemoryMB is the maximum memory limit in MB available for this tool per invocation.
	// If nil, there is no memory limit.
	MaxMemoryMB *int `yaml:"max-memory-mb" validate:"omitempty,min=1"`
	// CpuPercent is the CPU limit as a percentage of total host CPU (0-100) per invocation.
	// If nil, there is no CPU limit.
	CpuPercent *int `yaml:"cpu-percent" validate:"omitempty,min=1,max=100"`
}

type ToolSelector struct {
	// Disabled determines whether tools are disabled for the task.
	// Individual tools can override this setting.
	Disabled *bool `yaml:"disabled" validate:"omitempty"`
	// Tools lists the tools to be available in the task execution.
	Tools []ToolSelection `yaml:"tools" validate:"omitempty,unique=Name,dive"`
}

type URI struct {
	// contains filtered or unexported fields
}

type ValidationRules struct {
	// CaseSensitive determines whether string comparison should be case-sensitive.
	CaseSensitive *bool `yaml:"case-sensitive" validate:"omitempty"`

	// IgnoreWhitespace determines whether all whitespace should be ignored during comparison.
	// When true, all whitespace characters (spaces, tabs, newlines) are removed before comparison.
	IgnoreWhitespace *bool `yaml:"ignore-whitespace" validate:"omitempty"`

	// TrimLines determines whether to trim leading and trailing whitespace of each line
	// before comparison.
	TrimLines *bool `yaml:"trim-lines" validate:"omitempty"`

	// Judge specifies the judge configuration to use for evaluation.
	// When enabled, an LLM will be used to evaluate the correctness of the response
	// instead of simple string matching.
	Judge JudgeSelector `yaml:"judge" validate:"omitempty"`
}

type XAIClientConfig struct {
	// APIKey is the API key for the xAI provider.
	APIKey string `yaml:"api-key" validate:"required"`
}

type XAIModelParams struct {
	// Temperature controls the randomness or "creativity" of the model's outputs.
	// Notes: Higher values (e.g. 0.8) make outputs more random; lower values
	// (e.g. 0.2) make outputs more focused and deterministic.
	// Valid range: 0.0 — 2.0. Default: 1.0.
	Temperature *float32 `yaml:"temperature" validate:"omitempty,min=0,max=2"`

	// TopP controls diversity via nucleus sampling (probability mass cutoff).
	// Notes: Use either Temperature or TopP, not both, for sampling control.
	// Valid range: 0.0 — 1.0. Default: 1.0.
	TopP *float32 `yaml:"top-p" validate:"omitempty,min=0,max=1"`

	// MaxCompletionTokens controls the maximum number of tokens to generate in the completion.
	MaxCompletionTokens *int32 `yaml:"max-completion-tokens" validate:"omitempty,min=0"`

	// PresencePenalty penalizes new tokens based on whether they appear in the text so far.
	// Notes: Positive values encourage the model to introduce new topics.
	// Valid range: -2.0 — 2.0. Default: 0.0.
	PresencePenalty *float32 `yaml:"presence-penalty" validate:"omitempty,min=-2,max=2"`

	// FrequencyPenalty penalizes new tokens based on their frequency in the text so far.
	// Notes: Positive values discourage repetition.
	// Valid range: -2.0 — 2.0. Default: 0.0.
	FrequencyPenalty *float32 `yaml:"frequency-penalty" validate:"omitempty,min=-2,max=2"`

	// ReasoningEffort constrains how much "reasoning" budget to spend for reasoning-capable models.
	// Notes: Not all reasoning models support this option.
	// Valid values: "low", "high".
	ReasoningEffort *string `yaml:"reasoning-effort" validate:"omitempty,oneof=low high"`

	// Seed requests deterministic sampling when possible.
	// No guaranteed determinism — xAI makes a best-effort to return
	// repeatable outputs for identical inputs when `seed` and other parameters are the same.
	Seed *int32 `yaml:"seed" validate:"omitempty"`
}