wehub-resource-sync
74 lines
2.8 KiB
Python
74 lines
2.8 KiB
Python
"""UI journey: a second turn typed into the composer sees turn-1 context.
|
|
|
|
The smoke test proves one prompt round-trips; this proves the part
|
|
users actually depend on: history threading. Turn 1 establishes a
|
|
random token, turn 2 asks for it back, and the token can only appear
|
|
in turn 2's bubble if UI -> server -> runner -> LLM replayed the
|
|
conversation.
|
|
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import uuid
|
|
|
|
from playwright.sync_api import Page, expect
|
|
|
|
from tests.e2e_ui.conftest import configure_mock_llm
|
|
|
|
_COMPOSER = "Ask the agent anything…"
|
|
_ASSISTANT = '[data-testid="message-bubble"][data-role="assistant"]'
|
|
_WORKING = '[data-testid="working-indicator"]'
|
|
|
|
# The recall prompt's stable substring — present only in turn 2, so the mock
|
|
# can content-route turn 2 to the token-echo response.
|
|
_RECALL_PROMPT = "What token did I ask you to remember? Reply with the token only."
|
|
|
|
|
|
def _send(page: Page, text: str) -> None:
|
|
"""Type *text* into the composer and click Send."""
|
|
composer = page.get_by_placeholder(_COMPOSER)
|
|
expect(composer).to_be_visible()
|
|
composer.fill(text)
|
|
page.get_by_role("button", name="Send", exact=True).click()
|
|
|
|
|
|
def test_multi_turn_recall_through_ui(
|
|
page: Page,
|
|
seeded_session: tuple[str, str],
|
|
mock_llm_server_url: str,
|
|
) -> None:
|
|
base_url, session_id = seeded_session
|
|
token = f"ui-recall-{uuid.uuid4().hex[:8]}"
|
|
|
|
# The mock routes by message content: turn 1 (the only message carrying
|
|
# the token) replies "stored"; turn 2 (the recall prompt) replies with the
|
|
# token. Both match strings are unique to their turn, so each queue fires
|
|
# exactly once and the recall assertion is deterministic.
|
|
configure_mock_llm(mock_llm_server_url, [{"text": "stored"}], key="recall-store", match=token)
|
|
configure_mock_llm(
|
|
mock_llm_server_url, [{"text": token}], key="recall-echo", match=_RECALL_PROMPT
|
|
)
|
|
|
|
page.goto(f"{base_url}/c/{session_id}")
|
|
|
|
_send(
|
|
page,
|
|
f"Remember this token exactly, you will be asked to repeat it "
|
|
f"verbatim later: {token}. Reply with just the word 'stored'.",
|
|
)
|
|
# Turn 1 terminal: an assistant bubble rendered AND the working
|
|
# shimmer is gone (sending turn 2 mid-stream would test steering,
|
|
# not history replay).
|
|
expect(page.locator(_ASSISTANT).first).to_be_visible(timeout=60_000)
|
|
expect(page.locator(_WORKING)).to_have_count(0, timeout=60_000)
|
|
|
|
_send(page, _RECALL_PROMPT)
|
|
# Two user bubbles = turn 2 actually left the composer.
|
|
expect(page.locator('[data-testid="message-bubble"][data-role="user"]')).to_have_count(
|
|
2, timeout=15_000
|
|
)
|
|
# The literal token in an assistant bubble is only producible from
|
|
# turn-1 history; a fresh-context reply cannot contain it.
|
|
expect(page.locator(_ASSISTANT, has_text=token).first).to_be_visible(timeout=60_000)
|