eval

type Assertion struct {
	Name   string
	Pass   bool
	Detail string
}

type Case struct {
	Name   string
	Prompt string
	Until  string
	Expect Expect
	// Judge is the optional LLM-judge rubric (evals/<case>/judge.md, design
	// 009 §3.2). Empty when the case has no judge.md: assertions alone grade
	// the case. When present, the judge runs AFTER assertions and both must
	// pass — the judge is a signal, structural assertions stay the hard gate.
	Judge string
}

type Expect struct {
	// FilesChanged: every changed file must match at least one of these globs
	// (scope containment — catches out-of-scope edits). Empty ⇒ not checked.
	FilesChanged []string `yaml:"files_changed"`
	// CommandsAllowed: if non-empty, every executed command must match one.
	CommandsAllowed []string `yaml:"commands_allowed"`
	// CommandsForbidden: no executed command may match any of these.
	CommandsForbidden []string `yaml:"commands_forbidden"`
	// OutputMatches: every regex must match somewhere in the output.
	OutputMatches []string `yaml:"output_matches"`
	// MaxTokens: tokens_used must be ≤ this (0 ⇒ not checked).
	MaxTokens int `yaml:"max_tokens"`
}

type Result struct {
	FilesChanged []string `json:"files_changed"`
	Commands     []string `json:"commands"`
	Output       string   `json:"output"`
	TokensUsed   int      `json:"tokens_used"`
}