15 KiB
MCP Runtime RPC Session Reuse E2E Implementation Plan
For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (
- [ ]) syntax for tracking.
Goal: Add a deterministic JSON-RPC integration test proving MCP-backed workflow runs reuse one persistent McpRuntimePool session for repeated operations against the same source.
Architecture: The test should build a real MCP-backed WorkflowServer, inject a recording McpRuntimePool backed by a fake persistent MCP client, expose it through create_rpc_app, then drive the workflow lifecycle via RpcWorkflowApiClient. The assertion should be semantic: two workflow runs against the same MCP source reuse one created session and see incrementing session-local state.
Tech Stack: Python 3.14, pytest-asyncio, httpx ASGITransport, JSON-RPC app/client, WfMcpService, McpRuntimePool, PersistentSessionFactory, MCP SDK result models.
Hard Boundaries
- Do not use live
pnpx/ everything-server in this deterministic test. - Do not assert directly on private
McpRuntimePool._sessionsif a recording factory/client can prove reuse. - Do not add public production hooks only for this test.
- Do not change JSON-RPC method names or payload shapes.
- Do not remove one-shot adapter fallback behavior.
- Do not add a direct JSON-RPC
call_capabilitymethod in this slice. - Do not commit unless the caller explicitly asks for a commit.
File Map
- Modify
tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py: add recording runtime fake and one integration test. - Optionally modify
docs/current_roadmap.md: mark runtime RPC reuse proof complete. - Optionally modify
docs/superpowers/specs/2026-06-03-long-lived-workflow-api-boundary.md: mark QA proof complete. - Move this plan to
docs/historical/superpowers/plans/after implementation is verified.
Test Design
The test should prove this complete path:
RpcWorkflowApiClient
-> JSON-RPC app
-> WorkflowServer.api.run_deployment()
-> WfMcpService / WorkflowOperationContext
-> generated NodeSpec
-> UpstreamTransportService.tool_executor_for()
-> McpRuntimePool
-> one PersistentMcpSession
-> fake MCP client session-local counter
Use a fake MCP client whose call_tool("counter", ...) increments self.count.
Expected behavior:
- First workflow run returns
{"count": 1}. - Second workflow run returns
{"count": 2}. - Recording session factory created exactly one session.
- Fake client recorded two tool calls.
If the path accidentally uses one-shot sessions, the count will reset and the test will fail.
Task 1: Add Recording Runtime Fakes
Files:
-
Modify:
tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py -
Step 1: Add imports
Add imports needed for fake runtime and server construction:
from contextlib import AsyncExitStack
from dataclasses import dataclass, field
from typing import Any, cast
from mcp.client.session import ClientSession
from mcp.types import CallToolResult, ListPromptsResult, ListResourcesResult, ListToolsResult, TextContent, Tool
from wf_authoring import build_async_registry
from wf_mcp.broker.server import workflow_server_from_service
from wf_mcp.broker.service import WfMcpService
from wf_sources_mcp.connections import McpSourceConnection
from wf_sources_mcp.runtime import McpRuntimePool
from wf_sources_mcp.runtime.factory import PersistentSessionFactory
from wf_sources_mcp.storage import FileAuthStore, FileCatalogStore, FileStore
Adjust imports to match existing ordering and avoid duplicates. If build_async_registry is not needed, omit it.
- Step 2: Add fake MCP client
Add near the existing helpers:
@dataclass(slots=True)
class _CountingMcpClient:
count: int = 0
tool_calls: list[tuple[str, dict[str, Any]]] = field(default_factory=list)
async def list_tools(self) -> ListToolsResult:
return ListToolsResult(
tools=[
Tool(
name="counter",
title="Counter",
description="Increment a session-local counter.",
inputSchema={"type": "object", "properties": {}},
)
]
)
async def call_tool(
self,
tool_name: str,
payload: dict[str, Any],
) -> CallToolResult:
self.tool_calls.append((tool_name, payload))
if tool_name != "counter":
raise KeyError(tool_name)
self.count += 1
return CallToolResult(
content=[TextContent(type="text", text=str(self.count))],
structuredContent={"count": self.count},
)
async def list_resources(self) -> ListResourcesResult:
return ListResourcesResult(resources=[])
async def list_prompts(self) -> ListPromptsResult:
return ListPromptsResult(prompts=[])
Do not implement methods the test does not use unless basedpyright requires them for ClientSession casts.
- Step 3: Add recording factory
class _RecordingSessionFactory(PersistentSessionFactory):
def __init__(self) -> None:
self.clients: list[_CountingMcpClient] = []
self.created_connections: list[McpSourceConnection] = []
async def _create_with_stack(
self,
stack: AsyncExitStack,
connection: McpSourceConnection,
auth,
) -> ClientSession:
self.created_connections.append(connection)
client = _CountingMcpClient()
self.clients.append(client)
return cast(ClientSession, client)
- Step 4: Run typecheck for the test file
Run:
uv run basedpyright --level error tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py
Expected: no errors. If ClientSession cast is not enough, add minimal methods to _CountingMcpClient or use # type: ignore[return-value] only with a comment explaining the fake implements the subset used by the runtime owner.
Task 2: Build an MCP-Backed WorkflowServer With Recording Runtime
Files:
-
Modify:
tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py -
Step 1: Add helper to build server and factory
Add:
def _runtime_reuse_server(tmp_path):
config = BrokerConfig(
store_root=tmp_path / "store",
connections=[
ConnectionConfig(
id="fixture.default",
server="fixture",
account="default",
metadata={
"transport": "stdio",
"command": "fake-mcp-server",
},
)
],
)
factory = _RecordingSessionFactory()
runtime_pool = McpRuntimePool(factory.create)
service = WfMcpService(
store=FileStore(tmp_path / "mcp-store"),
auth_store=FileAuthStore(tmp_path / "auth-store"),
catalog_store=FileCatalogStore(tmp_path / "catalog-store"),
artifact_store=..., # use the same file stores that build_service_from_config would use
draft_workspace_store=...,
run_store=...,
tool_executor=runtime_pool,
stateful_runtime=runtime_pool,
)
service.register_connection(config.connections[0])
return workflow_server_from_service(
service,
config=config,
source_registry_store=FileSourceRegistryStore(config.store_roots.source_registry_root),
), service, factory
Use existing store constructors already used in this test module or adjacent tests:
FileWorkflowArtifactStoreFileDraftWorkspaceStoreFileRunStore
If BrokerConfig.__post_init__ fills store_roots, use config.store_roots.source_registry_root. If basedpyright cannot narrow store_roots, assert it is not None.
- Step 2: Refresh catalog before RPC lifecycle
Inside the helper or the test, call:
await service.refresh_connection_catalog("fixture.default")
This should use stateful_runtime.list_tools() and create the first persistent session. Do not register an adapter for "fixture"; that way the test fails if refresh tries the one-shot fallback.
- Step 3: Verify catalog refresh uses one session
After refresh:
assert len(factory.clients) == 1
assert factory.created_connections[0].id == "fixture.default"
Task 3: Drive Workflow Lifecycle Through JSON-RPC
Files:
-
Modify:
tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py -
Step 1: Add test skeleton
Add:
async def test_mcp_backed_rpc_workflow_reuses_runtime_session_across_runs(tmp_path) -> None:
server, service, factory = _runtime_reuse_server(tmp_path)
await service.refresh_connection_catalog("fixture.default")
async with httpx.AsyncClient(
transport=httpx.ASGITransport(app=create_rpc_app(server)),
base_url="http://test",
) as http_client:
client = RpcWorkflowApiClient(url="http://test/rpc", http_client=http_client)
...
- Step 2: Use RPC to create a draft from the discovered capability
Call:
created = await client.create_draft_workspace_from_capability(
workspace_id="counter_ws",
capability_name="fixture.default.counter",
name="counter_workflow",
title="Counter Workflow",
)
Assert:
assert created["workspace_id"] == "counter_ws"
assert created["status"] == "valid"
If exact response keys differ, inspect existing draft RPC tests and assert stable public fields only.
- Step 3: Use RPC to save an artifact
Call:
artifact = await client.create_artifact_from_workspace(
workspace_id="counter_ws",
artifact_id="counter_workflow",
version=1,
title="Counter Workflow",
outcomes=["ok", "error"],
kind="workflow",
)
Assert artifact ID/version from the response. If draft validation requires explicit outcomes from the capability, use the outcomes exposed by wrap_discovered_tool (ok, error).
- Step 4: Use RPC to save a deployment
Call:
deployment = await client.save_deployment(
{
"id": "counter_workflow.default",
"artifact_id": "counter_workflow",
"artifact_version": 1,
"bindings": [
{"logical_source": "fixture.default", "concrete_source": "fixture.default"}
],
}
)
If the artifact has no logical source binding because it was created directly from a concrete capability, use bindings=[]. Prefer following the artifact response's required_logical_sources if present.
- Step 5: Run deployment twice through RPC
Call:
first = await client.run_deployment(
deployment_id="counter_workflow.default",
workflow_input={},
)
second = await client.run_deployment(
deployment_id="counter_workflow.default",
workflow_input={},
)
Assert:
assert first["status"] == "completed"
assert second["status"] == "completed"
assert first["output"]["count"] == 1
assert second["output"]["count"] == 2
If output is nested under a generated field, inspect the actual result once and assert the stable output path.
- Step 6: Assert runtime reuse
After both runs:
assert len(factory.clients) == 1
assert len(factory.created_connections) == 1
assert factory.clients[0].tool_calls == [
("counter", {}),
("counter", {}),
]
If refresh catalog creates one session and tool calls happen on the same client, this proves discovery + two runs all shared the same persistent session.
Task 4: Fallback If Draft Synthesis Is Too Noisy
Files:
- Modify:
tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py
Only use this task if Task 3 gets blocked by draft synthesis details.
- Step 1: Prepare artifact/deployment directly
Use server.api directly to create the artifact/deployment after catalog refresh, then still run the deployment twice through RpcWorkflowApiClient.
This fallback still proves the product run path:
JSON-RPC runs.start -> WorkflowApi -> generated NodeSpec -> McpRuntimePool
It does not prove draft/artifact RPC creation. That is acceptable if documented in the report.
- Step 2: Keep the same runtime reuse assertions
Do not weaken the session reuse assertions. The test must still prove one created client and two tool calls.
Task 5: Docs and Plan Archive
Files:
-
Modify:
docs/current_roadmap.md -
Modify:
docs/superpowers/specs/2026-06-03-long-lived-workflow-api-boundary.md -
Move:
docs/superpowers/plans/2026-06-08-mcp-runtime-rpc-session-reuse-e2e.mdtodocs/historical/superpowers/plans/2026-06-08-mcp-runtime-rpc-session-reuse-e2e.md -
Step 1: Update roadmap
Add:
JSON-RPC MCP-backed workflow runs now have deterministic session-reuse
coverage: repeated runs against one source use one `McpRuntimePool`
session and preserve session-local state.
- Step 2: Update boundary spec
Add a completed QA item after the full-operation-surface item:
25. Complete: deterministic JSON-RPC integration coverage proves MCP-backed
workflow runs reuse one persistent runtime session for repeated operations
against the same source.
- Step 3: Archive plan
Run:
git mv docs/superpowers/plans/2026-06-08-mcp-runtime-rpc-session-reuse-e2e.md docs/historical/superpowers/plans/2026-06-08-mcp-runtime-rpc-session-reuse-e2e.md
Task 6: Verification
Files:
-
No code edits unless verification finds a real issue.
-
Step 1: Run new test
Run:
uv run pytest tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py::test_mcp_backed_rpc_workflow_reuses_runtime_session_across_runs -q
Expected: pass.
- Step 2: Run broader RPC and runtime tests
Run:
uv run pytest tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py tests/wf_sources_mcp/test_runtime.py tests/wf_mcp/service/test_upstream_transport.py -q
Expected: pass.
- Step 3: Run lint
Run:
uv run ruff check tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py docs/current_roadmap.md docs/superpowers/specs/2026-06-03-long-lived-workflow-api-boundary.md
Expected: source/test lint passes. If ruff does not support markdown without preview mode in this repo, run it on the Python test file only and report the markdown limitation.
- Step 4: Run typecheck
Run:
uv run basedpyright --level error tests/wf_transport_rpc_http/test_mcp_backed_server_rpc.py
Expected: 0 errors, 0 warnings, 0 notes.
- Step 5: Optional live smoke
If .env has MCP_EVERYTHING_COMMAND or MCP_EVERYTHING_URL, run:
uv run --env-file .env pytest tests/wf_mcp/test_sdk_adapter.py::test_mcp_sdk_adapter_can_probe_everything_server -q
Expected: pass or skip only for environment permission issues.
- Step 6: Check whitespace
Run:
git diff --check
Expected: no whitespace errors. CRLF warnings on Windows are acceptable.
Expected Final Report
The implementer should report:
- Files modified and moved.
- Whether the test used full RPC draft/artifact/deployment setup or the direct setup fallback.
- Exact verification commands and outputs.
- Confirmation that repeated JSON-RPC workflow runs reuse one recording runtime session.
- Confirmation that the test does not use live external MCP servers.
- Any deviations from this plan.
Do not claim "full suite passed" unless the full suite was actually run.