Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
/**
* Canonical agent resolution for workflow recipes.
*
* A recipe's mcp-host provider/model comes from spec.agent, or — for a genuine
* workflow — from the first step that declares a complete agent. This is the
* single source of truth shared by the reconciler (which needs the AgentSpec
* to build the mcp-host pod) and the Plugin Workload SDK validator (which only
* needs to know whether one is resolvable). Keeping both on this helper
* prevents the two from drifting apart.
*/
import { isRunnableLlmModelId } from '@clerum/llm-providers'
import type { WorkflowRecipeSpec } from '../types'
import type { AgentSpec } from './types'
/**
* Resolve the complete agent (provider + model) for a recipe, or undefined when
* neither spec.agent nor any step agent declares both fields. spec.agent is
* returned verbatim (preserving soulRef/secretRef); a step agent is narrowed to
* provider + model to match the prior reconciler behavior.
*/
export function resolveMcpHostAgent(spec: WorkflowRecipeSpec): AgentSpec | undefined {
// An explicitly declared top-level binding is authoritative. Do not silently
// replace an invalid operator declaration with a step agent: that can boot a
// different provider/model than the recipe author requested.
if (spec.agent !== undefined) {
if (spec.agent.provider && isRunnableLlmModelId(spec.agent.model)) {
return spec.agent as AgentSpec
}
return undefined
}
const stepAgent = (spec.steps ?? []).find(
step => step.agent?.provider && isRunnableLlmModelId(step.agent.model)
)?.agent
if (!stepAgent?.provider || !stepAgent.model) return undefined
return { provider: stepAgent.provider, model: stepAgent.model }
}
/** True when the recipe resolves to a complete agent (provider + model). */
export function hasResolvableAgent(spec: WorkflowRecipeSpec): boolean {
return resolveMcpHostAgent(spec) !== undefined
}
/**
* Agent used to build the eager SDK mcp-host pod. promptBridge requires a
* real resolvable agent; clientNotifications-only recipes deliberately return
* no agent so the notification runtime does not invent a provider/model.
*/
export function resolveEagerSdkMcpHostAgent(spec: WorkflowRecipeSpec): AgentSpec | undefined {
const resolved = resolveMcpHostAgent(spec)
if (resolved) return resolved
return undefined
}