agents

type Agent interface {
	Version() (string, error)
	Run(cwd string, llm LLMDefinition, session string, instructions string, opts RunOptions) RunResults
}

type Definition struct {
	Name         string   `yaml:"name"`
	Version      string   `yaml:"version"`
	SupportsLLMs []string `yaml:"supports-llms"`
}

type LLMCostDefinition struct {
	InputCost           *float64 `yaml:"input-cost"`
	CachedInputCost     *float64 `yaml:"cached-input-cost"`
	CacheWriteInputCost *float64 `yaml:"cache-write-input-cost"`
	OutputCost          *float64 `yaml:"output-cost"`
}

type LLMDefinition struct {
	Name           string             `yaml:"name"`
	Model          string             `yaml:"model"`
	ReasoningLevel string             `yaml:"reasoning-level"`
	PerAgent       map[string]string  `yaml:"per-agent"`
	Costs          *LLMCostDefinition `yaml:"costs"`
}

type Registry struct {
	Agents map[string]Definition
	LLMs   map[string]LLMDefinition
}

type RunContext struct {
	ScenarioName string
	ScenarioPath string
	ModelName    string
	LLM          *LLMDefinition
	Agent        Definition
	Instructions string
	Session      string
	Options      RunOptions
	Printer      *output.Printer
}

type RunOptions struct {
	// Package is an optional Go package path (relative to the workspace root)
	// that an agent may use to scope work (ex: "internal/cli").
	Package string
}

type RunOutcome struct {
	Progress *types.RunProgress
}

type RunResults struct {
	// Transcript is the full transcript
	Transcript string

	InputTokens            int     // number of uncached input tokens. This is essentially context used
	CachedInputTokens      int     // number of cached input tokens. As a convo is renetered with tool call results, the old parts accumulate here.
	WriteCachedInputTokens int     // number of tokens spent writing content to the cache
	OutputTokens           int     // number of reasoning/output tokens
	Cost                   float64 // total cost for the run (if available)

	// ScaleDuration, when >0, scales the wall-clock DurationSeconds recorded in RunProgress.
	// A value of 0 means "unscaled" (use the measured elapsed time).
	// The purpose of this is to, for example, adjust ChatGPT Pro's Priority Processing back to "apples-to-apples" times.
	// A better solution might be to just not use ChatGPT Pro's auth, but that would cost more money/time to run.
	ScaleDuration float64

	// If an agent supports it, this is the session ID (or resume ID). We can pass this ID to future Run calls to continue.
	Session string

	// Error is any error produced by trying to run the agent.
	Err error
}