Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import { createHash } from 'node:crypto'
import { type DbClient, pool } from '../db.js'
import {
type MediumIdentityInput,
type VerifiedMediumAccount,
findVerifiedMediumAccount,
normalizeMediumIdentity,
} from './workflowApprovalMediumIdentityService.js'
export type TelegramProviderChatType = 'private' | 'group' | 'supergroup' | 'channel'
export type ProviderTargetBindingInput = {
hostRef?: string | null
communicationChannelNamespace?: string | null
communicationChannelName?: string | null
// Figure D: 8-byte hash of communication_channel_ref, carried inside the
// provider action value by the reader. Present when the decision arrived via
// the reader path (DM private); absent on the channel-reader path (Figure C),
// which carries ns/name directly.
communicationChannelAlias?: string | null
providerBotId?: string | null
providerBotUsername?: string | null
}
// Must match CHANNEL_ALIAS_LEN in workflowApprovalNotificationDeliveryWorker.ts
// (16 hex / 64-bit). Kept local (not shared) to avoid a cross-module dependency
// for a single int.
const CHANNEL_ALIAS_LEN = 16
export type OperationalMediumIdentityInput = MediumIdentityInput & {
providerChannelType?: string | null
providerTarget?: ProviderTargetBindingInput | null
}
export function normalizeTelegramProviderChannelType(
value: unknown
): TelegramProviderChatType | null {
const normalized = typeof value === 'string' ? value.trim().toLowerCase() : ''
if (
normalized === 'private' ||
normalized === 'group' ||
normalized === 'supergroup' ||
normalized === 'channel'
) {
return normalized
}
return null
}
// Figure D: resolve a provider action channelAlias (8-byte sha256 of the ref)
// back to the full "namespace/name" ref by scanning the user's active accounts
// (typically 1–2 per user). The alias is a public hash carried in a
// provider-signed action payload, so a plain equality check suffices.
async function resolveChannelRefByAlias(
db: DbClient,
identity: ReturnType<typeof normalizeMediumIdentity>,
alias: string
): Promise<string | null> {
const result = await db.query(
`SELECT communication_channel_ref AS "communicationChannelRef"
FROM workflow_approval_medium_accounts
WHERE medium = $1
AND provider_user_id = $2
AND COALESCE(provider_workspace_id, '') = COALESCE($3, '')
AND ($4::text IS NULL OR provider_channel_id = $4)
AND communication_channel_ref IS NOT NULL
AND disabled_at IS NULL`,
[
identity.medium,
identity.providerUserId,
identity.providerWorkspaceId,
identity.providerChannelId,
]
)
// candidate is lowercase sha256 hex; the alias arrives lowercased from the
// reader, but normalize defensively so a mixed-case alias never causes a
// false medium_identity_not_verified rejection (lockout of a legit user).
const normalizedAlias = alias.trim().toLowerCase()
for (const row of result.rows as Array<{ communicationChannelRef: string }>) {
const candidate = createHash('sha256')
.update(row.communicationChannelRef)
.digest('hex')
.slice(0, CHANNEL_ALIAS_LEN)
if (candidate === normalizedAlias) return row.communicationChannelRef
}
return null
}
// Channel-binding mode for the operational provider lookup.
// - 'strict' (default): bind to the decision's CommunicationChannel
// (cross-bot block). A null/unresolved ref matches NO row.
// - 'identity-only': resolve by provider identity alone, no channel-ref
// filter. Used by the workflow-access resolve path, which has no channel
// context (mcp-host sends no providerTarget).
// The mode is an EXPLICIT caller intent — never inferred from "channelRef is
// null". Inferring it would let a failed alias resolution on the approval path
// silently fall through to an unbound match, reopening the cross-bot hole.
export type OperationalMediumChannelBinding = 'strict' | 'identity-only'
export async function findVerifiedOperationalMediumAccount(
identityInput: OperationalMediumIdentityInput,
db: DbClient = pool,
options: { channelBinding?: OperationalMediumChannelBinding } = {}
): Promise<VerifiedMediumAccount | null> {
const identity = normalizeMediumIdentity(identityInput)
if (
identity.medium !== 'telegram' &&
identity.medium !== 'slack' &&
identity.medium !== 'teams'
) {
return findVerifiedMediumAccount(identity, db)
}
if (identity.medium === 'telegram' && identity.providerWorkspaceId) {
return null
}
const enforceChannelBinding = options.channelBinding !== 'identity-only'
const providerChannelType = normalizeTelegramProviderChannelType(
identityInput.providerChannelType
)
const privateProviderChannelId =
providerChannelType === 'private' && identity.providerChannelId
? identity.providerChannelId
: null
// identity-only mode (workflow-access /resolve): match by provider identity,
// no channel-ref binding. Skip alias resolution entirely. This restores the
// pre-PR#547 behaviour for the path that has no channel context; the strict
// cross-bot binding below still governs the approval-decision path.
if (!enforceChannelBinding) {
const identityOnlyProviderChannelId =
identity.medium === 'telegram' ? privateProviderChannelId : identity.providerChannelId
const result = await db.query(
`SELECT id,
user_id AS "userId",
medium,
provider_user_id AS "providerUserId",
provider_workspace_id AS "providerWorkspaceId",
provider_channel_id AS "providerChannelId",
communication_channel_ref AS "communicationChannelRef",
display_name AS "displayName",
disabled_at AS "disabledAt"
FROM workflow_approval_medium_accounts
WHERE medium = $1
AND provider_user_id = $2
AND COALESCE(provider_workspace_id, '') = COALESCE($3, '')
AND ($4::text IS NULL OR provider_channel_id = $4)
AND disabled_at IS NULL
LIMIT 1`,
[
identity.medium,
identity.providerUserId,
identity.providerWorkspaceId,
identityOnlyProviderChannelId,
]
)
return (result.rows[0] as VerifiedMediumAccount | undefined) ?? null
}
// Figure D cross-bot fix (HIGH): provider user ids are not enough when the
// same user is verified through different CommunicationChannels. Bind the
// lookup to the CommunicationChannel of the decision context. The filter is
// STRICT (owner decision).
//
// Two provenance paths for the channel ref:
// - Figure C (channel-reader): providerTarget carries ns/name directly.
// - Figure D (reader + provider action): providerTarget carries a signed
// channelAlias (8-byte sha256 of the ref); resolve it before filtering.
//
// When no ref resolves, `communication_channel_ref = NULL` matches no row →
// the decision is rejected. Post-migration every active approval account has
// a non-null channel ref, so the real flows are unaffected.
const directChannelRef =
identityInput.providerTarget?.communicationChannelNamespace &&
identityInput.providerTarget?.communicationChannelName
? `${identityInput.providerTarget.communicationChannelNamespace}/${identityInput.providerTarget.communicationChannelName}`
: null
const channelAlias = identityInput.providerTarget?.communicationChannelAlias?.trim() || null
const channelRef = directChannelRef
? directChannelRef
: channelAlias
? await resolveChannelRefByAlias(db, identity, channelAlias)
: null
const result = await db.query(
`SELECT id,
user_id AS "userId",
medium,
provider_user_id AS "providerUserId",
provider_workspace_id AS "providerWorkspaceId",
provider_channel_id AS "providerChannelId",
communication_channel_ref AS "communicationChannelRef",
display_name AS "displayName",
disabled_at AS "disabledAt"
FROM workflow_approval_medium_accounts
WHERE medium = $1
AND provider_user_id = $2
AND COALESCE(provider_workspace_id, '') = COALESCE($3, '')
AND ($4::text IS NULL OR provider_channel_id = $4)
AND communication_channel_ref = $5
AND disabled_at IS NULL
LIMIT 1`,
[
identity.medium,
identity.providerUserId,
identity.providerWorkspaceId,
identity.medium === 'telegram' ? privateProviderChannelId : identity.providerChannelId,
channelRef,
]
)
return (result.rows[0] as VerifiedMediumAccount | undefined) ?? null
}