Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import type { DbClient } from '../db.js'
import type { Logger } from '../observability/logger.js'
import { mcpServerUninstallTeardownFailuresTotal } from '../observability/metrics.js'
import type { DcrDeps } from './dcr.js'
import { bestEffortRfc7592Delete } from './dcrCleanup.js'
import { claimDeleteDynamicClientForResource } from './dynamicClientStore.js'
import { deleteOAuthGrantsForServer } from './store.js'
/**
* Shared OAuth teardown for a mcp-server uninstall, fenced by the uninstalled CR's
* `metadata.uid` (R3-H5). Replaces the previous name-only pair — `dcrCleanup`'s
* `cleanupDynamicClientForServer` and the loose `deleteOAuthGrantsForServer` call in
* the uninstall route — so a same-name REINSTALL's state is never torn down or revoked
* by the uninstall of a previous installation.
*
* Two race windows the uid fence + the single-statement claim close:
* - B: the reinstall wrote its row before the old teardown ran → the old name-only
* read+revoke would revoke the NEW client. Here the DELETE…RETURNING only revokes
* the client of the row it actually removed (this uid's, or a legacy row's).
* - C: the old read→await(7592)→delete gap let a reinstall slip a new row in between.
* Here there is no gap: the claim is one statement.
*
* Runs while the CR still exists (see mcpServerUninstall.ts) and never throws: a
* throw is caught, counted (metric, per stage) and logged (names-only), and reported
* as `'failed'` so the caller can keep the CR and answer "repair required" — the
* retry reuses the same uid.
*/
export type TeardownStageResult = 'done' | 'none' | 'failed'
export interface McpServerOAuthTeardownResult {
/** 'done' = a DCR row was removed, 'none' = none matched (clean), 'failed' = threw. */
dynamicClient: TeardownStageResult
/** 'done' = grant rows were purged, 'none' = none matched (clean), 'failed' = threw. */
grants: TeardownStageResult
}
export async function teardownMcpServerOAuthState(
db: DbClient,
encryptionKey: Buffer,
dcrDeps: DcrDeps,
input: { namespace: string; name: string; crUid: string },
logger: Logger
): Promise<McpServerOAuthTeardownResult> {
const { namespace, name, crUid } = input
const result: McpServerOAuthTeardownResult = { dynamicClient: 'none', grants: 'none' }
// DCR: single-statement claim+read. Deletes the row bound to this uid (or a legacy
// row) and returns its RFC 7592 handle; a pending row or a row bound to a different
// uid (a reinstall) is left intact. Only the actually-deleted row's client is revoked.
try {
const claimed = await claimDeleteDynamicClientForResource(
db,
encryptionKey,
{ ownerKind: 'mcpserver', serverNamespace: namespace, serverName: name },
crUid
)
if (claimed.deleted) {
result.dynamicClient = 'done'
if (claimed.handle) {
await bestEffortRfc7592Delete(
dcrDeps,
claimed.handle.registrationClientUri,
claimed.handle.registrationAccessToken
)
}
logger.info(
{ serverName: name, namespace, attemptedRemoteDelete: Boolean(claimed.handle) },
'Revoked remote OAuth dynamic client on uninstall'
)
}
} catch (err) {
result.dynamicClient = 'failed'
mcpServerUninstallTeardownFailuresTotal.inc({ stage: 'dynamic_client' })
logger.error({ serverName: name, namespace, err }, 'Dynamic client cleanup failed on uninstall')
}
// Grants: purge this installation's rows (fenced by cr_uid + legacy). A reinstall's
// grant (a different uid) survives.
try {
const purged = await deleteOAuthGrantsForServer(db, {
recipeNamespace: namespace,
recipeName: name,
crUid,
})
if (purged > 0) {
result.grants = 'done'
logger.info(
{ serverName: name, namespace, count: purged },
'Purged OAuth grants on uninstall'
)
}
} catch (err) {
result.grants = 'failed'
mcpServerUninstallTeardownFailuresTotal.inc({ stage: 'oauth_grants' })
logger.error({ serverName: name, namespace, err }, 'OAuth grants purge failed on uninstall')
}
return result
}