Skip to content

Self-improving AI Employees for everyone

fix(employees): keep the team plan on deployed listeners (#148)

trohitg0bfeacc4184c
diff --git a/docs-internal/node-logic-flows/ai_tools/taskManager.md b/docs-internal/node-logic-flows/ai_tools/taskManager.md
index 94973f92..2c316ee3 100644
--- a/docs-internal/node-logic-flows/ai_tools/taskManager.md
+++ b/docs-internal/node-logic-flows/ai_tools/taskManager.md
@@ -126,10 +126,12 @@ recipient. The durable job retains the original request, parameters, workspace
generation, and delivery identity across delegation, review, and retries.
The saved graph keeps node data to UI fields, so the plan is saved on the
-employee row, and Start and Apply copy it onto the lead in the generation's
-snapshot (`team_runtime.admit_team_plan`). A run holds only the nodes its
-trigger reaches, so the runtime reads the plan, and a job takes its graph,
-from that whole snapshot (`team_runtime.admitted_graph`).
+employee row. Start and Apply copy it onto the lead in the generation's
+snapshot, and every deploy (Start and the restart re-arm) copies it onto the
+lead in the deployed listeners, which a deploy sanitizes again
+(`team_runtime.admit_team_plan`). A run holds only the nodes its trigger
+reaches, so the runtime reads the plan, and a job takes its graph, from that
+whole snapshot (`team_runtime.admitted_graph`).
The acknowledgement confirms assignment only. Specialists submit work through
Task Manager, and the lead reviews and accepts it. The runtime delivers the
diff --git a/server/services/deployment/handlers.py b/server/services/deployment/handlers.py
index faeb8604..2a51d0e4 100644
--- a/server/services/deployment/handlers.py
+++ b/server/services/deployment/handlers.py
@@ -188,6 +188,7 @@ async def handle_deploy_workflow(data: Dict[str, Any], websocket: WebSocket) ->
load_node_parameters,
persist_parameter_aliases,
)
+ from services.employees.team_runtime import admit_team_plan
from services.workflow_migrations import normalize_workflow_graph
from services.workflow_sanitizer import sanitize_workflow_graph
@@ -205,7 +206,9 @@ async def handle_deploy_workflow(data: Dict[str, Any], websocket: WebSocket) ->
parameters_by_id,
)
safe_graph = sanitize_workflow_graph(normalization.graph_data())
- nodes = list(safe_graph["nodes"])
+ # A hired team's plan is saved beside the graph, not in it. Start and the
+ # restart re-arm deploy through here, so their listeners carry it too.
+ nodes = await admit_team_plan(container.database(), str(workflow_id or ""), list(safe_graph["nodes"]))
edges = list(safe_graph["edges"])
normalized_parameters = normalization.node_parameters
await import_legacy_context_receipts(
diff --git a/server/services/employees/email_input.py b/server/services/employees/email_input.py
index a67b0dca..fb5704fc 100644
--- a/server/services/employees/email_input.py
+++ b/server/services/employees/email_input.py
@@ -101,8 +101,8 @@ async def project_managed_gmail_parameters(database: Any, node_id: str, paramete
node_type = node.get("type") or context.get("node_type")
if node_type == "employeeJob" and parameters.get("operation") == "intake":
field = "mission"
- elif ((node.get("data") or context.get("node_data") or {}).get("employee_recipe_version") == 2
- and ((node.get("data") or context.get("node_data") or {}).get("employee_team_plan") or {}).get("lead_node_id") == node_id):
+ elif (context.get("employee_team_plan") or {}).get("lead_node_id") == node_id:
+ # Prepare resolved the plan from the generation's admitted graph.
field = "prompt"
else:
return parameters, context, False
diff --git a/server/tests/services/employees/test_email_input.py b/server/tests/services/employees/test_email_input.py
index 20bdc931..e44ceaed 100644
--- a/server/tests/services/employees/test_email_input.py
+++ b/server/tests/services/employees/test_email_input.py
@@ -194,7 +194,9 @@ def intake_context(body, *, version=1):
return params, {"workflow_id": "wf", "node_type": node["type"], "nodes": [
{"id": "gmail", "type": "googleGmailReceive", "data": {"label": "Gmail"}}, node],
"edges": [{"source": "gmail", "target": node["id"]}],
- "outputs": {"gmail": {"subject": "Actual email", "body": body}}, "node_id": node["id"]}
+ "outputs": {"gmail": {"subject": "Actual email", "body": body}}, "node_id": node["id"],
+ # Prepare puts the plan it resolved from the admitted graph on the context.
+ **({"employee_team_plan": plan} if version == 2 else {})}
@pytest.mark.parametrize("version", [1, 2])
diff --git a/server/tests/services/employees/test_intake_scope.py b/server/tests/services/employees/test_intake_scope.py
index af4a2966..d56ad9f0 100644
--- a/server/tests/services/employees/test_intake_scope.py
+++ b/server/tests/services/employees/test_intake_scope.py
@@ -60,7 +60,9 @@ async def admitted(database, monkeypatch, version=1, *, tail="TAIL", request="or
monkeypatch.setattr(database, "get_workflow", AsyncMock(return_value=SimpleNamespace(data={"nodes": nodes, "edges": edges})))
context = {"workflow_id": "wf", "node_id": node_id, "node_type": "employeeJob" if version == 1 else "ai_employee",
"execution_id": request, "session_id": "wf", "nodes": nodes, "edges": edges,
- "outputs": {"gmail": email}, "parameter_snapshot": parameters, "user_id": "owner"}
+ "outputs": {"gmail": email}, "parameter_snapshot": parameters, "user_id": "owner",
+ # Prepare puts the plan it resolved from the admitted graph on the context.
+ **({"employee_team_plan": plan} if version == 2 else {})}
resolver = ParameterResolver(database, AsyncMock(return_value=None))
resolved, context, handled = await project_managed_gmail_parameters(database, node_id, parameters[node_id], context, resolver.resolve)
assert handled
diff --git a/server/tests/services/employees/test_team_plan_admission.py b/server/tests/services/employees/test_team_plan_admission.py
index 7dade009..78b1acfe 100644
--- a/server/tests/services/employees/test_team_plan_admission.py
+++ b/server/tests/services/employees/test_team_plan_admission.py
@@ -15,14 +15,13 @@ import pytest
import nodes # noqa: F401
from models.employees import Employee
-from services.deployment.control import WorkflowControlService
from services.employees.apps import get_app
from services.employees.builder import BuildInputs, build_employee_graph
from services.employees.hire_request import HireEmployeeRequest
from services.employees.jobs import create_job
from services.employees.llm import LLMChoice
from services.employees.prompt import OwnerProfile
-from services.employees.team_runtime import admit_team_plan, employee_runtime_plan
+from services.employees.team_runtime import employee_runtime_plan
from services.plugin import NodeContext
from services.temporal.trigger_listener_workflow import _build_run_graph
from services.workflow_migrations import normalize_workflow_graph
@@ -43,7 +42,8 @@ def inbox_assistant():
@pytest.fixture
async def started(real_database, monkeypatch):
- """A hire saved the way commit_hire saves it, then admitted the way Start admits it."""
+ """A hire saved the way commit_hire saves it, then started through the real Start and deploy handlers."""
+ from services.deployment import handlers
built = inbox_assistant()
normalized = normalize_workflow_graph("7", built.nodes, built.edges, built.parameters)
saved = sanitize_workflow_graph({**normalized.graph_data(), "owner_id": "owner"})
@@ -52,29 +52,48 @@ async def started(real_database, monkeypatch):
team_plan=built.team_plan, node_roles=built.node_roles))
await session.commit()
monkeypatch.setattr(real_database, "get_workflow", AsyncMock(return_value=SimpleNamespace(data=saved, slug="Leo_1")))
- control, _ = await WorkflowControlService(real_database).begin_generation(
- workflow_id="7", nodes=await admit_team_plan(real_database, "7", saved["nodes"]), edges=saved["edges"],
- session_id="7", idempotency_key="start", graph_version=normalized.graph_version, owner_id="owner",
- parameters=normalized.node_parameters)
- return SimpleNamespace(database=real_database, saved=saved, control=control, roles=built.node_roles, plan=built.team_plan)
+ # Temporal and the status broadcasts are the only parts not run for real.
+ workflow_service = SimpleNamespace(
+ is_workflow_deployed=lambda _workflow_id: False,
+ deploy_workflow=AsyncMock(return_value={"success": True, "triggers_setup": [], "deployment_id": "deployment"}))
+ broadcaster = SimpleNamespace(
+ lock_workflow=AsyncMock(return_value=True), unlock_workflow=AsyncMock(), update_node_status=AsyncMock(),
+ update_workflow_status=AsyncMock(), update_deployment_status=AsyncMock())
+ monkeypatch.setattr("core.container.container", SimpleNamespace(database=lambda: real_database,
+ workflow_service=lambda: workflow_service))
+ monkeypatch.setattr("services.status_broadcaster.get_status_broadcaster", lambda: broadcaster)
+ monkeypatch.setattr("services.workflow_validator.validate_workflow", AsyncMock(return_value={"errors": []}))
+ monkeypatch.setattr(handlers, "_deployment_tasks", {})
+ monkeypatch.setattr(handlers, "_broadcast_control", AsyncMock(return_value={}))
+ monkeypatch.setattr(handlers, "_start_controller", AsyncMock(return_value=None))
+ monkeypatch.setattr(handlers, "_query_controller_state", AsyncMock(return_value=None))
+ result = await handlers.handle_start_workflow(
+ {"workflow_id": "7", "nodes": saved["nodes"], "edges": saved["edges"], "user_id": "owner",
+ "parameters_by_id": normalized.node_parameters, "expected_revision": 0, "idempotency_key": "start"}, None)
+ assert result["success"] is True, result
+ deployed = workflow_service.deploy_workflow.await_args.kwargs
+ return SimpleNamespace(database=real_database, saved=saved, control=await real_database.get_latest_workflow_control("7"),
+ listeners={"nodes": deployed["nodes"], "edges": deployed["edges"]},
+ roles=built.node_roles, plan=built.team_plan)
def run(started, trigger, node_id):
- """The context one agent gets in a run its trigger started."""
- snapshot = started.control.graph_snapshot
+ """The context one agent gets in a run its trigger's deployed listener started."""
nodes, edges = _build_run_graph(trigger_node_id=trigger, trigger_output={"subject": "Hello"},
- nodes=snapshot["nodes"], edges=snapshot["edges"])
+ nodes=started.listeners["nodes"], edges=started.listeners["edges"])
return {"workflow_id": "7", "generation": started.control.generation, "data_scope_id": started.control.data_scope_id,
"execution_id": f"run-{node_id}", "session_id": "7", "user_id": "owner", "node_id": node_id,
"nodes": nodes, "edges": edges, "outputs": {trigger: {"subject": "Hello"}},
- "parameter_snapshot": snapshot["parameters"]}
+ "parameter_snapshot": started.control.graph_snapshot["parameters"]}
async def test_start_admits_the_plan_the_saved_graph_cannot_hold(started):
lead = started.roles["agent"]
assert set(next(node for node in started.saved["nodes"] if node["id"] == lead)["data"]) == {"label"}
- admitted = next(node for node in started.control.graph_snapshot["nodes"] if node["id"] == lead)["data"]
- assert admitted["employee_recipe_version"] == 2 and admitted["employee_team_plan"] == started.plan
+ # The generation's snapshot and the listeners deployed from it both carry it.
+ for graph in (started.control.graph_snapshot, started.listeners):
+ admitted = next(node for node in graph["nodes"] if node["id"] == lead)["data"]
+ assert admitted["employee_recipe_version"] == 2 and admitted["employee_team_plan"] == started.plan
async def test_a_trigger_run_resolves_the_plan_and_its_job_holds_the_delivery_nodes(started):
diff --git a/server/tests/services/test_workflow_control_handlers.py b/server/tests/services/test_workflow_control_handlers.py
index e4392189..ec1203b8 100644
--- a/server/tests/services/test_workflow_control_handlers.py
+++ b/server/tests/services/test_workflow_control_handlers.py
@@ -1454,3 +1454,45 @@ async def test_start_admits_the_employee_team_plan_on_its_lead(monkeypatch):
admitted = {node["id"]: node["data"] for node in service.begin_generation.await_args.kwargs["nodes"]}
assert admitted["wf:ai_employee:1"] == {"label": "Lead", "employee_recipe_version": 2, "employee_team_plan": plan}
assert admitted["wf:aiAgent:1"] == {"label": "Talk"}
+
+
+async def test_deploy_admits_the_employee_team_plan_on_its_lead(monkeypatch):
+ """Start and the restart re-arm deploy through here, which sanitizes the graph again (#148)."""
+ plan = {"version": 2, "lead_node_id": "wf:ai_employee:1", "talk_node_id": "wf:aiAgent:1"}
+ workflow_service = SimpleNamespace(
+ is_workflow_deployed=lambda _workflow_id: False,
+ deploy_workflow=AsyncMock(return_value={"success": True, "triggers_setup": [], "deployment_id": "deployment"}),
+ )
+ broadcaster = SimpleNamespace(
+ lock_workflow=AsyncMock(return_value=True), unlock_workflow=AsyncMock(), update_node_status=AsyncMock(),
+ update_workflow_status=AsyncMock(), update_deployment_status=AsyncMock(),
+ )
+ monkeypatch.setattr("core.container.container", SimpleNamespace(workflow_service=lambda: workflow_service,
+ database=lambda: SimpleNamespace()))
+ monkeypatch.setattr("services.status_broadcaster.get_status_broadcaster", lambda: broadcaster)
+ monkeypatch.setattr("services.workflow_validator.validate_workflow", AsyncMock(return_value={"errors": []}))
+ monkeypatch.setattr("services.employees.store.get_by_workflow", AsyncMock(return_value=SimpleNamespace(team_plan=plan)))
+ monkeypatch.setattr(handlers, "_deployment_tasks", {})
+
+ result = await handlers.handle_deploy_workflow(
+ {
+ "workflow_id": "wf",
+ "nodes": [
+ # A plan the caller sends is dropped; the employee's is deployed.
+ {"id": "wf:ai_employee:1", "type": "ai_employee",
+ "data": {"label": "Lead", "employee_recipe_version": 2, "employee_team_plan": {"version": 2, "lead_node_id": "forged"}}},
+ {"id": "wf:aiAgent:1", "type": "aiAgent", "data": {"label": "Talk"}},
+ ],
+ "edges": [],
+ "parameters_by_id": {},
+ "generation": 1,
+ },
+ None,
+ )
+ await handlers._deployment_tasks["wf"]
+
+ assert result["success"] is True
+ deployed = {node["id"]: node["data"] for node in workflow_service.deploy_workflow.await_args.kwargs["nodes"]}
+ assert deployed["wf:ai_employee:1"] == {"label": "Lead", "employee_recipe_version": 2, "employee_team_plan": plan}
+ assert deployed["wf:aiAgent:1"] == {"label": "Talk"}