termbench

type AgentInfo struct {
	Name      string `json:"name"`
	Version   string `json:"version"`
	ModelName string `json:"model_name,omitempty"`
}

type CaptureOptions struct {
	SessionID string
	Agent     AgentInfo
	TaskID    string
	StartedAt time.Time // used for relative timestamps; defaults to now()
}

type ExecutionPlan struct {
	// Task is the source task (kept for downstream reporting).
	Task *Task

	// Request is the ServiceExecuteRequest the caller should hand to
	// fizeau.New(...).Execute. WorkDir is left to the caller because a
	// real Harbor run mounts the task workspace at a container path
	// (/app), while a dry-run from cmd/bench may point at a temp dir.
	Request fizeau.ServiceExecuteRequest

	// Timeout matches the task's MaxAgentTimeoutSec budget, suitable for
	// a context.WithTimeout wrapping Execute.
	Timeout time.Duration
}

type GraderResult struct {
	// TaskID identifies which task produced this result.
	TaskID string

	// Reward is the canonical pass/fail signal: 1 = passed, 0 = failed.
	// Mirrors Harbor's /logs/verifier/reward.txt contract.
	Reward int

	// CTRFPath, if set, points at the verifier's pytest CTRF JSON. Reporters
	// can drill down for per-test detail.
	CTRFPath string

	// RewardPath is the absolute path the reward was read from.
	RewardPath string

	// Notes carries human-readable detail (e.g. "missing reward.txt").
	Notes string
}

type PlanOptions struct {
	// Harness is the agent harness label (e.g. "fiz", "claude-code",
	// "codex"). Passed through verbatim into ServiceExecuteRequest.
	Harness string

	// Model is the provider model ID (e.g. "openrouter/qwen/qwen3.6-plus").
	Model string

	// Provider, if non-empty, pins the named provider (e.g. "vidar",
	// "bragi"). Useful when the same model id is ambiguous across
	// providers, or when a pin-only catalog entry needs an explicit
	// provider hop.
	Provider string

	// WorkDir is the directory the agent operates in. For a real Harbor
	// trial this is /app inside the container; for a Go-side dry-run it
	// can be a tempdir seeded with the task workspace.
	WorkDir string

	// Permissions, if non-empty, overrides the default "safe" preset. The
	// TerminalBench tasks routinely require shell + edit access so the
	// caller may want "trusted" here.
	Permissions string

	// Seed enables deterministic sampling (matches cmd/bench parity runs).
	// Zero means "leave unset".
	Seed int64

	// Temperature, if non-nil, overrides the request temperature. The
	// default-zero behavior is "leave unset" so the agent's own bench
	// path can pin temperature to 0 separately.
	Temperature *float32
}

type Task struct {
	// ID is the task directory name (terminal-bench convention).
	ID string

	// Path is the absolute path to the task directory.
	Path string

	// Instruction is the natural-language prompt the agent receives. Resolved
	// from the descriptions[].instruction matching the chosen difficulty
	// key, falling back to the "base" description.
	Instruction string

	// Difficulty is one of "easy", "medium", "hard" (or empty if not declared).
	Difficulty string

	// Tags is the categorization labels from task.yaml, copied verbatim.
	Tags []string

	// MaxAgentTimeoutSec is the per-task agent wall-clock budget Harbor
	// enforces. The adapter uses this when building ServiceExecuteRequest
	// so the Go-side timeout matches what the grader will tolerate.
	MaxAgentTimeoutSec int

	// MaxTestTimeoutSec is the verifier's pytest timeout. We surface it so
	// reporters can record it; the adapter does not run tests itself.
	MaxTestTimeoutSec int

	// AuthorEmail comes straight from task.yaml; useful for provenance
	// tracking in result artifacts.
	AuthorEmail string
}

type Trajectory struct {
	SchemaVersion string           `json:"schema_version"`
	SessionID     string           `json:"session_id"`
	TaskID        string           `json:"task_id"`
	Agent         AgentInfo        `json:"agent"`
	Steps         []TrajectoryStep `json:"steps"`
	FinalMetrics  TrajectoryStat   `json:"final_metrics"`
	FinalStatus   string           `json:"final_status,omitempty"`
	ExitCode      int              `json:"exit_code"`
	DurationMS    int64            `json:"duration_ms"`
	Error         string           `json:"error,omitempty"`
}

type TrajectoryStat struct {
	InputTokens  int     `json:"input_tokens"`
	OutputTokens int     `json:"output_tokens"`
	Cost         float64 `json:"cost"`
}

type TrajectoryStep struct {
	StepID    int             `json:"step_id"`
	Timestamp string          `json:"timestamp"`
	Source    string          `json:"source"` // user|agent|system|tool
	Message   string          `json:"message,omitempty"`
	ToolCalls []TrajectoryTC  `json:"tool_calls,omitempty"`
	Metrics   *TrajectoryStat `json:"metrics,omitempty"`
}

type TrajectoryTC struct {
	Name   string          `json:"name"`
	Input  json.RawMessage `json:"input,omitempty"`
	Output string          `json:"output,omitempty"`
	Error  string          `json:"error,omitempty"`
}