workspaceeval

type AgentDoc struct {
	ID           string           `yaml:"id"`
	StarterTasks []StarterTaskDoc `yaml:"starterTasks"`
	Profile      struct {
		Publish bool `yaml:"publish"`
	} `yaml:"profile"`
}

type EvalDoc struct {
	ID              string `yaml:"id"`
	Title           string `yaml:"title"`
	UserPrompt      string `yaml:"user_prompt"`
	EntryAgent      string `yaml:"entry_agent"`
	ExpectedRouting struct {
		Agent   string   `yaml:"agent"`
		Profile string   `yaml:"profile"`
		FanOut  []string `yaml:"fan_out"`
	} `yaml:"expected_routing"`
	ExpectedPreDelegationTools []EvalToolExpectation `yaml:"expected_pre_delegation_tools"`
	ExpectedOutput             struct {
		Template string `yaml:"template"`
	} `yaml:"expected_output"`
}

type EvalToolExpectation struct {
	Name string `yaml:"name"`
}

type Options struct {
	Workspace            string
	ContractTests        []string
	RequiredProfiles     []string
	Behavioral           bool
	BehavioralCases      string
	BehavioralTimeoutSec int
	BehavioralAPI        string
	BehavioralOOB        string
	BehavioralToken      string
	BehavioralAgentlyBin string
}

type PromptDoc struct {
	ID               string `yaml:"id"`
	EvidenceContract struct {
		Required   []string `yaml:"required"`
		Optional   []string `yaml:"optional"`
		Forbidden  []string `yaml:"forbidden"`
		Completion []string `yaml:"completion"`
	} `yaml:"evidenceContract"`
}

type StarterTaskDoc struct {
	ID              string   `yaml:"id"`
	Title           string   `yaml:"title"`
	CoverageEvalIDs []string `yaml:"coverageEvalIds"`
}