env

type Env struct {
	AgentUC        *usecase.AgentUseCase
	JobRunner      *job.JobRunner
	Repo           interfaces.Repository
	Registry       *model.WorkspaceRegistry
	Entry          *model.WorkspaceEntry
	Slack          *fakeSlack
	Recorder       *toolsim.Recorder
	Trace          *agentarchive.MemoryTraceRepository
	SeededCases    []*model.Case
	MonitorChannel string
	Language       string
	// contains filtered or unexported fields
}

type Options struct {
	// LLM is the system-under-test agent's LLM client (required).
	LLM gollem.LLMClient
	// Models is which model each run generates through and what it may spend
	// (required). A scenario Job naming a model reaches it the same way a
	// deployed one does, so a scenario exercises the real resolution rather than
	// a harness-only default.
	Models agentkernel.ModelPolicy
	// Completer drives the simulated tools (required when any tool is sim).
	Completer evaltype.Completer
	// Live* are real tool clients, used for tools marked live=true. Nil unless
	// the corresponding live tool is requested.
	LiveSlackSearch slacktool.SearchService
	LiveNotion      notiontool.Client
	GitHub          *githubtool.Client
	// WebFetch holds the live-only webfetch HTTP settings; the eval LLM is
	// injected as the screening client when the tool is built.
	WebFetch *webfetch.ClientConfig
	// JiraTools carries the already-expanded Jira read tools (see
	// pkg/agent/tool/jira). Live-only, like GitHub.
	JiraTools []gollem.Tool
	// SlackToolLimits bounds what a Slack read tool's result may inject into
	// the model context, exactly as it does in serve. It applies to the
	// simulated Slack tools too: a scenario whose simulated result would be
	// truncated in production must be truncated here, or the harness measures
	// behaviour production never sees. The zero value disables both bounds.
	SlackToolLimits slacktool.Limits
}

type Post struct {
	Kind      string // "thread_reply" | "thread_message" | "message"
	ChannelID string
	ThreadTS  string
	Text      string
	TS        string
}