hermes-agent/tests/gateway/test_stream_edit_flood_retry_after.py
teknium1 7fdeb2094a test(streaming): trim salvaged #116385 suites to the two invariants per fix
Keep interim-skip/final-never-gated + send-waits-for-slot, and mid-word/one-long-token
for the continuation cut; the dropped cases restate the same two invariants.
2026-09-20 12:17:23 -07:00

65 lines
2.7 KiB
Python

"""A flood-refused streaming edit honours the server's ``retry_after`` (#116312).
Telegram answers a flood-refused ``editMessageText`` with the seconds to wait. The stream
consumer used to ignore it and only double its own 0.8s interval, so every one of the three
allowed strikes was spent — and three more refused requests earned — inside the first ~5s of
a penalty that is usually 9s or longer.
"""
import asyncio
from types import SimpleNamespace
from unittest.mock import AsyncMock, MagicMock
import pytest
from gateway.stream_consumer import GatewayStreamConsumer, StreamConsumerConfig
def _flood(wait: float):
return SimpleNamespace(success=False, error=f"flood_control:{wait}", retry_after=wait)
async def _stream(adapter, edit_interval: float, deltas: int, gap: float):
consumer = GatewayStreamConsumer(
adapter, "chat_1", StreamConsumerConfig(edit_interval=edit_interval, buffer_threshold=1, cursor=" ▉"))
consumer.on_delta("Hello")
task = asyncio.create_task(consumer.run())
await asyncio.sleep(gap)
for i in range(deltas):
consumer.on_delta(f" w{i}")
await asyncio.sleep(gap)
consumer.finish()
await task
return consumer
@pytest.mark.asyncio
async def test_flood_refused_edit_waits_the_servers_retry_after_before_the_next_interim_edit():
adapter = MagicMock()
adapter.MAX_MESSAGE_LENGTH = 4096
adapter.send = AsyncMock(return_value=SimpleNamespace(success=True, message_id="m1"))
adapter.edit_message = AsyncMock(return_value=_flood(9.0))
consumer = await _stream(adapter, edit_interval=0.02, deltas=6, gap=0.05)
# The first refused interim edit reads retry_after=9s; every later interim tick (all inside
# that window) is skipped instead of being sent into the same penalty.
# Interim previews carry the cursor; the finalize path's cursor-strip edit does not.
interim_edits = [c for c in adapter.edit_message.await_args_list if "▉" in c.kwargs["content"]]
assert len(interim_edits) == 1
assert consumer._current_edit_interval >= 9.0
@pytest.mark.asyncio
async def test_flood_without_retry_after_keeps_doubling_backoff():
adapter = MagicMock()
adapter.MAX_MESSAGE_LENGTH = 4096
adapter.send = AsyncMock(return_value=SimpleNamespace(success=True, message_id="m1"))
adapter.edit_message = AsyncMock(return_value=SimpleNamespace(success=False, error="flood_control:9"))
consumer = await _stream(adapter, edit_interval=0.02, deltas=6, gap=0.05)
interim_edits = [c for c in adapter.edit_message.await_args_list if "▉" in c.kwargs["content"]]
assert len(interim_edits) >= 2 # 0.04s, 0.08s, 0.16s: the legacy doubling ladder keeps firing
assert consumer._current_edit_interval <= 10.0
assert consumer._flood_strikes >= 2