Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import * as k8s from '@kubernetes/client-node'
import { config } from './config.js'
import { llmAllowlistConfigMapWriteFailuresTotal } from './observability/metrics.js'
import {
type AllowedModelsConfigMapMaterializer,
LlmAllowedModelsConfigMapWriter,
} from './services/llmAllowedModelsConfigMap.js'
/**
* Re-materialize the `clerum-llm-allowed-models` ConfigMap from Postgres on
* boot (spec §3-R3.4 / V7 anti-drift). Runs after `initDb()` in main.ts.
*
* Wiring pattern mirrors `registryBootGuard.ts` — an isolated, testable module
* with a narrow dependency graph — but with a DELIBERATE difference: the
* registry guard is fail-fast and aborts the boot; this reconcile is
* **non-fatal**. V7 asks for fail-loud without aborting: if the write fails we
* log ERROR + increment the failure metric and let the process continue (the
* next CRUD mutation, or the next boot, reconciles). Bricking control-api on a
* transient K8s write would be worse than serving a slightly stale ConfigMap
* that Postgres remains the source of truth for.
*/
export async function reconcileAllowedModelsConfigMapOnBoot(
materializer?: AllowedModelsConfigMapMaterializer
): Promise<void> {
try {
const writer = materializer ?? buildDefaultWriter()
await writer.materialize()
console.log('[ControlAPI] llm allowed-models ConfigMap reconciled on boot')
} catch (err) {
llmAllowlistConfigMapWriteFailuresTotal.inc({ phase: 'boot' })
console.error(
'[ControlAPI] llm allowed-models ConfigMap boot reconcile failed (continuing):',
err instanceof Error ? err.message : String(err)
)
}
}
function buildDefaultWriter(): LlmAllowedModelsConfigMapWriter {
const kc = new k8s.KubeConfig()
kc.loadFromDefault()
const coreApi = kc.makeApiClient(k8s.CoreV1Api)
// The allowlist ConfigMap always lives with the interactive Host pods.
return new LlmAllowedModelsConfigMapWriter(coreApi, config.hostsNamespace)
}