Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
/**
* Unpersisted model-selection intents (R2 "Option A").
*
* When a user changes the model while the agent host is SUSPENDED (`replicas=0`)
* the `POST /model` write cannot reach the runtime, so the choice cannot be
* persisted server-side. Rather than silently drop it, the selection is held
* here — keyed by `(agentRef, chatId)` like sessions — and PIGGYBACKED onto the
* next send, which wakes the host and applies the model to that task.
*
* A PRE-CHAT pick (no chatId yet) is keyed by agent alone; the send path
* migrates it into the chat-keyed slot once the first send creates the chat.
*
* This is a module-level singleton (not React state): it is written by the
* shared `hostModelSelectionStore` and read by the composer/send path, two
* consumers that never share a render tree. Keeping it out of React state avoids
* a context/provider just to shuttle one imperative value between them.
*/
const pendingModelByChat: Record<string, string> = {}
const preChatModelByAgent: Record<string, string> = {}
/** Composite key mirroring the session `chatKey` convention (`agentRef::chatId`). */
export function pendingModelKey(agentRef: string, chatId: string): string {
return `${agentRef}::${chatId}`
}
export function setPendingModelIntent(agentRef: string, chatId: string, model: string): void {
if (!agentRef || !chatId || !model) return
pendingModelByChat[pendingModelKey(agentRef, chatId)] = model
}
export function getPendingModelIntent(agentRef: string, chatId: string): string | undefined {
if (!agentRef || !chatId) return undefined
return pendingModelByChat[pendingModelKey(agentRef, chatId)]
}
export function clearPendingModelIntent(agentRef: string, chatId: string): void {
if (!agentRef || !chatId) return
delete pendingModelByChat[pendingModelKey(agentRef, chatId)]
}
export function setPreChatModelIntent(agentRef: string, model: string): void {
if (!agentRef || !model) return
preChatModelByAgent[agentRef] = model
}
export function getPreChatModelIntent(agentRef: string): string | undefined {
return agentRef ? preChatModelByAgent[agentRef] : undefined
}
export function clearPreChatModelIntent(agentRef: string): void {
if (!agentRef) return
delete preChatModelByAgent[agentRef]
}
/**
* Drops every intent. Called on logout/user/team scope changes: an unpersisted
* model choice must never carry across identity boundaries where the same
* agent/chat identifiers may reappear.
*/
export function resetHostModelIntentStore(): void {
for (const key of Object.keys(pendingModelByChat)) delete pendingModelByChat[key]
for (const key of Object.keys(preChatModelByAgent)) delete preChatModelByAgent[key]
}