hibira-approvals

Human approvals for AI agents over Hibira realtime channels, for Python. When a LangGraph node needs a person's OK, the call waits in a live inbox; reviewers approve, edit or reject it, and the graph resumes with their answer.

It speaks the same protocol as the npm package @hibira-io/approvals, so its React and Vue inbox components work against a Python backend unchanged.

Install

pip install "hibira-approvals[pusher,fastapi]"

Use with LangGraph

import pusher
from langgraph.types import Command, interrupt

from hibira_approvals import Approvals
from hibira_approvals.langgraph import approval, approval_calls, resume_value

client = pusher.Pusher(app_id=..., key=..., secret=..., host="ws.hibira.io", ssl=True)


def refund_node(state):
    answer = interrupt(approval("refund_order", {"order": state["order"]}))
    if not answer["approved"]:
        return {"log": [f"skipped: {answer['reason']}"]}
    return {"log": [refund(answer["input"])]}  # the edited input when the reviewer changed it


async def resume(request):
    config = {"configurable": {"thread_id": request["metadata"]["thread_id"]}}
    await graph.ainvoke(Command(resume=resume_value(request)), config)


approvals = Approvals(client, scope="support", resume=resume)

paused = await graph.ainvoke({"order": 1042}, config)
calls, title = approval_calls(paused)  # parallel interrupts become one request
if calls:
    await approvals.request(calls, agent="Support agent", title=title, metadata={"thread_id": "t1"})

approvals.wait(id) polls until a request is decided, for workers that keep the graph waiting in place.

Routes for the inbox components

from fastapi import FastAPI
from hibira_approvals.fastapi import approvals_router

app = FastAPI()
app.include_router(
    approvals_router(approvals, get_user=current_user, can_review=lambda user, request: user["role"] == "admin"),
    prefix="/api/approvals",
)

get_user(request) returns {"id", "name", "avatar"} or None. The router also serves POST /auth for the Pusher channel authorization.

Requests live in memory by default; pass store= with async create, get, list and an atomic update(id, fn) over your database.

License

MIT