conformance

type Capabilities struct {
	Name              string
	ExpectedModels    []string
	SupportsStreaming bool
	SupportsThinking  bool
	SupportsToolCalls bool
	MaxTokensSlack    int

	ChatContains      string
	StreamContains    string
	ReasoningContains string

	// ChatMaxTokens overrides the default chat/stream max-tokens budget
	// (8 tokens). Thinking-mode providers like luce burn output budget on
	// reasoning_content before the visible content; without headroom they
	// return empty content and the chat assertions fail. 0 = use default.
	ChatMaxTokens int
	// StreamMaxTokensCheck overrides the per-test cap for the
	// "streaming max_tokens honored" subtest. The check still asserts the
	// returned word count is bounded by this value + MaxTokensSlack. 0 =
	// use the existing default of 3.
	StreamMaxTokensCheck int
	// ScenarioTimeout overrides the per-subtest wall-clock budget
	// (default 5 seconds). Local thinking-capable providers can need
	// substantially more — set generously for those. 0 = use default.
	ScenarioTimeout time.Duration
}

type Factory func(t *testing.T) Subject

type Subject struct {
	Provider agent.Provider

	HealthCheck func(context.Context) error
	ListModels  func(context.Context) ([]string, error)
}