Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
JavaScript agentagentsai-agentdeepagentsdesignhyperframeslanggraphmanaged-agentsmotion-graphicspaperslackvideo
"""Real graph: a post-job review saves proposals for the skill, a later thread reads them, and a
rating moves the counts (SPEC §7.7)."""
from __future__ import annotations
import json
import re
from collections.abc import Iterator, Sequence
from pathlib import Path
import pytest
from langchain_core.messages import AIMessage, BaseMessage, ToolMessage
from design_agent.harness import inbox as inbox_module
from design_agent.harness.learning import ObservationStore
from design_agent.harness.tasks import REGISTRY
from design_agent.testing.scripted_model import (
RespondingChatModel,
ScriptedChatModel,
reply,
tool_call,
)
from tests.contract.test_background_contracts import build
NOTE = "The person cuts any intro longer than two seconds; open on the product."
@pytest.fixture(autouse=True)
def fast(monkeypatch: pytest.MonkeyPatch) -> Iterator[None]:
monkeypatch.setattr(inbox_module, "COALESCE_SECONDS", 0.05)
yield
REGISTRY.clear()
def review(_messages: Sequence[BaseMessage]) -> AIMessage:
proposal = {
"skill": "launch-film",
"observations": [{"text": NOTE, "evidence": ["v2 review: 'skip the intro'"]}],
}
return reply("Read the job.\n" + json.dumps(proposal))
async def test_a_job_review_is_saved_read_by_the_next_thread_and_rated(tmp_path: Path) -> None:
read: list[str] = []
def after_read(messages: Sequence[BaseMessage]) -> AIMessage:
read.append(str(messages[-1].content))
observation_id = re.search(r"\b(o-[0-9a-f]+)\b", read[0])
assert observation_id, read[0]
return tool_call(
"rate_observation",
{"skill": "launch-film", "id": observation_id.group(1), "helpful": True},
)
lead = ScriptedChatModel(
steps=[
lambda _m: tool_call(
"learn_from_job",
{
"skill": "launch-film",
"summary": "Made a 30 s launch film; the person cut the intro at v2.",
},
),
lambda _m: reply("Reviewing the job for next time."),
lambda _m: reply("[quiet]"),
lambda _m: tool_call("read_observations", {"skill": "launch-film"}),
after_read,
lambda m: reply("Rated." if isinstance(m[-1], ToolMessage) else "?"),
]
)
runtime, posted = build(tmp_path, lead, low=RespondingChatModel(respond=review))
await runtime.ask("Wrap up the launch film job.", thread_id="job-1")
assert await runtime.settle(timeout_seconds=30, poll_seconds=0.05)
store = runtime.graph.store
saved = await ObservationStore(store).list("launch-film")
assert [o.text for o in saved] == [NOTE] and saved[0].status == "proposed"
added = await runtime.ask("Start a new launch film.", thread_id="job-2")
assert NOTE in read[0], "a later thread reads what the earlier job showed"
assert added[-1].content == "Rated."
(rated,) = await ObservationStore(store).list("launch-film")
assert rated.helpful == 1 and rated.harmful == 0
assert posted == [], "the review's end was [quiet]"
async def test_a_configured_learning_model_runs_the_review(tmp_path: Path) -> None:
from design_agent.config import Settings
from design_agent.runtime.local import build_local_runtime
from design_agent.testing.fakes import folder_binding
from tests.contract.test_background_contracts import ROOT
lead = ScriptedChatModel(
steps=[
lambda _m: tool_call(
"learn_from_job", {"skill": "launch-film", "summary": "A film; the intro was cut."}
),
lambda _m: reply("Reviewing."),
lambda _m: reply("[quiet]"),
]
)
reviewer = RespondingChatModel(respond=review)
settings = Settings(
_env_file=None, workspace_root=tmp_path / "ws", learning_model="anthropic:claude-haiku-5-5"
) # type: ignore[call-arg]
runtime = build_local_runtime(
settings,
ROOT,
model=lead,
role_models={"learning": reviewer},
binding=folder_binding(tmp_path / "sb", ROOT),
replies=lambda *_: None,
)
await runtime.ask("Wrap up.", thread_id="job-3")
assert await runtime.settle(timeout_seconds=30, poll_seconds=0.05)
assert reviewer.call_count >= 1, "the review ran on the learning model"
assert [o.text for o in await ObservationStore(runtime.graph.store).list("launch-film")] == [
NOTE
]