nominee-openai
v3.0.0
Published
Bridges OpenAI Agents' native approval into nominee's one-shot evidence: an approved call cannot change before the resumed token is minted.
Maintainers
Readme
nominee-openai
Decision-bound tools for the OpenAI Agents SDK.
import { Agent } from '@openai/agents'
import { Nominee, allow, ask } from 'nominee'
import { nomineeTool } from 'nominee-openai'
import { z } from 'zod'
const nominee = new Nominee({
policy: {
rules: [allow('github.issue.read'), ask('github.issue.close')],
fallback: 'deny',
},
})
const closeIssue = nomineeTool({
name: 'close_issue',
description: 'Close one GitHub issue',
parameters: z.object({ repo: z.string(), issue: z.number() }),
nominee,
action: 'github.issue.close',
user: ({ context }) => context.context.userId,
resource: ({ input }) => `repo:${input.repo}#${input.issue}`,
connection: 'github',
scopes: ['issues:write'],
execute: async ({ repo, issue }, { token }) => {
// The credential is fetched only after the exact call is authorized.
return closeGitHubIssue({ repo, issue, token })
},
})
const agent = new Agent({ name: 'support-agent', tools: [closeIssue] })Nominee ask rules are evaluated by the SDK's needsApproval hook. OpenAI
pauses the run, and on resume the adapter verifies the approved tool-call id
from RunContext before recording approval evidence and executing. Denials
remain exceptions and never call the underlying tool.
What happens on ask
This adapter maps Nominee ask (and needsApproval) into the OpenAI Agents SDK's native, resumable approval flow: the run pauses for a human decision, and on resume the adapter verifies the approved tool-call id from RunContext and seals it as framework approval evidence on the action before execute runs. A denial remains an exception and never calls the tool. If execute runs without verifiable framework approval (e.g. a direct call that bypassed the SDK's pause), the portable path applies instead: the ask is settled inline when possible, or the call throws ActionPendingError with a durable actionId — persist that id and the original input, then resume with resolveActionApproval() → resumeAction() → executeCapability(). Full walkthrough: Approvals that outlive the request.
Observe before enforcing
Use the same tool with new Nominee({ mode: 'observe' }) to inventory the
callbacks that actually run before writing a policy. The adapter's
needsApproval hook returns false in this mode—including for an explicitly
configured approval—while the original ask/deny verdict remains on the action
and receipts. Observation reports do not retain raw string/boolean values or
user IDs; numeric aggregates may be sensitive. Observe mode is not a security
control and cannot be combined with production: true.
