MCP servers, over streamable HTTP

A server is a row with a URL; its tools are discovered by a button and
cached, then offered beside the built-in ones. Written by hand rather than
taken from the reference SDK, because that SDK's transport does its own
connecting -- and the one thing that must not be bypassed is check_url on
every hop. Owning the transport is the point; the framing beside it is the
small part.

Sessions are per call: initialize, initialized, the call, a best-effort
DELETE. Caching one wants an owner, a TTL, eviction, a lock and a shutdown
hook, and the server may expire it under all of that anyway -- ToolContext
is a session-free snapshot precisely so nothing in a tool holds live state.

A server's names and descriptions reach the model as instructions and are
bounded before they do; what it returns is escaped preformatted text, never
markdown. Tools are namespaced per server, so two servers exposing "search"
do not collide and neither shadows a built-in.

Also: a round's calls now run together under a semaphore, results indexed
so each tool turn stays paired with its call, and generation.status names
what is running -- a remote tool is latency-bound, and a silent pause is
what a hang looks like.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Jaroslav Beneš
2026-08-01 16:44:29 +02:00
parent bc84fec21d
commit ecb52e9978
17 changed files with 2270 additions and 28 deletions
+506
View File
@@ -0,0 +1,506 @@
"""The MCP client: framing, the session lifecycle, and what comes back.
The fake server below answers the way a real one does -- JSON-RPC over POST,
either as one JSON object or as an event stream, with a session id it expects
echoed. Everything is driven through `mock_http`, so no test touches a network.
"""
from __future__ import annotations
import json
import httpx
import pytest
from sqlalchemy import select
from lembas.db.models import McpServer
from lembas.services.mcp import client, protocol, registry
from lembas.services.mcp.protocol import McpError
@pytest.fixture(autouse=True)
def dns(monkeypatch):
"""`mcp.test` does not exist and check_url resolves for real."""
import ipaddress
import socket
def resolve(host, *_args, **_kwargs):
try:
ipaddress.ip_address(host)
except ValueError:
return [(2, 1, 6, "", ("93.184.216.34", 80))]
return [(2, 1, 6, "", (host, 80))]
monkeypatch.setattr(socket, "getaddrinfo", resolve)
def _spec(**overrides) -> client.McpSpec:
return client.McpSpec(
**{"slug": "github", "name": "GitHub", "url": "https://mcp.test/rpc", **overrides}
)
class FakeServer:
"""One server's worth of behaviour, as an httpx handler."""
def __init__(self, *, tools=None, stream=False, session_id="s-1", pages=None):
self.tools = tools if tools is not None else [
{"name": "search", "description": "Search things.", "inputSchema": {"type": "object"}}
]
self.stream = stream
self.session_id = session_id
self.pages = pages
self.requests: list[dict] = []
self.headers: list[httpx.Headers] = []
self.result_for = {}
def _frame(self, message: dict) -> httpx.Response:
if self.stream:
body = "".join(f"data: {line}\n" for line in json.dumps(message).splitlines()) + "\n"
return httpx.Response(
200, text=body, headers={"content-type": "text/event-stream"}
)
return httpx.Response(200, json=message)
def __call__(self, request: httpx.Request) -> httpx.Response:
self.headers.append(request.headers)
if request.method == "DELETE":
return httpx.Response(405)
message = json.loads(request.content)
self.requests.append(message)
method, request_id = message.get("method"), message.get("id")
if method == "notifications/initialized":
return httpx.Response(202)
if method == "initialize":
headers = {"mcp-session-id": self.session_id} if self.session_id else {}
response = self._frame(
{
"jsonrpc": "2.0",
"id": request_id,
"result": {
"protocolVersion": "2025-06-18",
"capabilities": {"tools": {}},
"serverInfo": {"name": "fake", "version": "1"},
},
}
)
return httpx.Response(
200, content=response.content, headers={**dict(response.headers), **headers}
)
if method == "tools/list":
if self.pages is not None:
cursor = (message.get("params") or {}).get("cursor", "")
index = int(cursor or 0)
page = self.pages[index]
result = {"tools": page}
if index + 1 < len(self.pages):
result["nextCursor"] = str(index + 1)
return self._frame({"jsonrpc": "2.0", "id": request_id, "result": result})
return self._frame(
{"jsonrpc": "2.0", "id": request_id, "result": {"tools": self.tools}}
)
if method == "tools/call":
name = (message.get("params") or {}).get("name")
result = self.result_for.get(
name, {"content": [{"type": "text", "text": "It worked."}]}
)
return self._frame({"jsonrpc": "2.0", "id": request_id, "result": result})
return self._frame(
{"jsonrpc": "2.0", "id": request_id, "error": {"code": -32601, "message": "Unknown"}}
)
# --- The lifecycle -----------------------------------------------------------
async def test_initialize_is_followed_by_the_initialized_notification(mock_http):
server = FakeServer()
mock_http(server)
tools, info, version = await client.list_tools(_spec())
methods = [request.get("method") for request in server.requests]
assert methods[:3] == ["initialize", "notifications/initialized", "tools/list"]
assert info == {"name": "fake", "version": "1"}
assert version == "2025-06-18"
assert [tool["name"] for tool in tools] == ["search"]
async def test_the_session_id_is_echoed_on_every_later_request(mock_http):
server = FakeServer(session_id="abc-123")
mock_http(server)
await client.list_tools(_spec())
assert "mcp-session-id" not in server.headers[0]
assert all(headers["mcp-session-id"] == "abc-123" for headers in server.headers[1:])
assert all("mcp-protocol-version" in headers for headers in server.headers[1:])
async def test_a_server_that_issues_no_session_is_fine(mock_http):
mock_http(FakeServer(session_id=""))
tools, _info, _version = await client.list_tools(_spec())
assert tools
async def test_an_expired_session_is_reinitialised_once(mock_http):
"""A 404 on a request carrying a session id means the server forgot us."""
server = FakeServer()
calls = {"list": 0}
inner = server.__call__
def handler(request: httpx.Request) -> httpx.Response:
if request.method == "POST":
message = json.loads(request.content)
if message.get("method") == "tools/list":
calls["list"] += 1
if calls["list"] == 1:
return httpx.Response(404, text="no such session")
return inner(request)
mock_http(handler)
tools, _info, _version = await client.list_tools(_spec())
assert tools, "the retry after re-initialising must succeed"
assert [r.get("method") for r in server.requests].count("initialize") == 2
async def test_a_json_response_and_an_event_stream_parse_the_same(mock_http):
plain = FakeServer(stream=False)
mock_http(plain)
from_json, _i, _v = await client.list_tools(_spec())
streamed = FakeServer(stream=True)
mock_http(streamed)
from_stream, _i, _v = await client.list_tools(_spec())
assert from_json == from_stream
async def test_tools_list_follows_the_cursor(mock_http):
server = FakeServer(
pages=[
[{"name": "one", "inputSchema": {"type": "object"}}],
[{"name": "two", "inputSchema": {"type": "object"}}],
]
)
mock_http(server)
tools, _i, _v = await client.list_tools(_spec())
assert [tool["name"] for tool in tools] == ["one", "two"]
async def test_a_jsonrpc_error_becomes_a_readable_message(mock_http):
def handler(request: httpx.Request) -> httpx.Response:
message = json.loads(request.content)
if message.get("method") == "notifications/initialized":
return httpx.Response(202)
return httpx.Response(
200,
json={
"jsonrpc": "2.0",
"id": message.get("id"),
"error": {"code": -32602, "message": "invalid params"},
},
)
mock_http(handler)
with pytest.raises(McpError, match="invalid params"):
await client.list_tools(_spec())
async def test_an_http_error_is_reported_with_its_status(mock_http):
mock_http(lambda _r: httpx.Response(502, text="bad gateway"))
with pytest.raises(McpError, match="502"):
await client.list_tools(_spec())
# --- Reaching the network ----------------------------------------------------
async def test_the_url_is_checked_before_the_first_request(mock_http):
server = FakeServer()
mock_http(server)
from lembas.services.fetch import FetchError
with pytest.raises(FetchError, match="private or local"):
await client.list_tools(_spec(url="http://127.0.0.1:9000/rpc"))
assert server.requests == [], "nothing may be sent before the address is checked"
async def test_a_302_is_refused_and_a_307_is_followed_and_checked(mock_http):
"""301, 302 and 303 turn a POST into a GET, which means nothing to a
JSON-RPC endpoint."""
inner = FakeServer()
def redirect_302(request: httpx.Request) -> httpx.Response:
return httpx.Response(302, headers={"location": "https://elsewhere.test/rpc"})
mock_http(redirect_302)
with pytest.raises(McpError, match="turn the request into a GET"):
await client.list_tools(_spec())
def redirect_307(request: httpx.Request) -> httpx.Response:
if request.url.host == "mcp.test":
return httpx.Response(307, headers={"location": "https://elsewhere.test/rpc"})
return inner(request)
mock_http(redirect_307)
tools, _i, _v = await client.list_tools(_spec())
assert tools
async def test_a_redirect_to_a_private_address_is_refused(mock_http, monkeypatch):
def resolve(host, *_args, **_kwargs):
if host == "inside.test":
return [(2, 1, 6, "", ("10.0.0.5", 80))]
return [(2, 1, 6, "", ("93.184.216.34", 80))]
monkeypatch.setattr("socket.getaddrinfo", resolve)
mock_http(lambda _r: httpx.Response(307, headers={"location": "https://inside.test/rpc"}))
from lembas.services.fetch import FetchError
with pytest.raises(FetchError, match="private or local"):
await client.list_tools(_spec())
async def test_the_secret_reaches_the_server(mock_http):
server = FakeServer()
mock_http(server)
await client.list_tools(_spec(secret="tok", secret_placement="bearer"))
assert server.headers[0]["authorization"] == "Bearer tok"
# --- Reading a tool result ---------------------------------------------------
async def test_text_blocks_are_joined(mock_http):
server = FakeServer()
server.result_for["search"] = {
"content": [{"type": "text", "text": "one"}, {"type": "text", "text": "two"}]
}
mock_http(server)
text, failed = await client.call_tool(_spec(), "search", {})
assert text == "one\n\ntwo"
assert failed is False
async def test_an_image_block_is_described_rather_than_forwarded(mock_http):
"""A tool turn is a string, images only reach models marked as having
vision, and base64 in a tool result fills a window with nothing."""
server = FakeServer()
server.result_for["search"] = {
"content": [{"type": "image", "mimeType": "image/png", "data": "A" * 400}]
}
mock_http(server)
text, _failed = await client.call_tool(_spec(), "search", {})
assert "A" * 40 not in text
assert "image/png" in text
assert "not shown to the model" in text
async def test_a_resource_block_with_text_keeps_it(mock_http):
server = FakeServer()
server.result_for["search"] = {
"content": [
{"type": "resource", "resource": {"uri": "file:///a.txt", "text": "contents"}}
]
}
mock_http(server)
text, _failed = await client.call_tool(_spec(), "search", {})
assert "file:///a.txt" in text and "contents" in text
async def test_a_resource_block_with_no_text_names_its_type(mock_http):
server = FakeServer()
server.result_for["search"] = {
"content": [
{"type": "resource", "resource": {"uri": "file:///a.bin", "mimeType": "application/x"}}
]
}
mock_http(server)
text, _failed = await client.call_tool(_spec(), "search", {})
assert "application/x" in text
async def test_an_unknown_block_type_degrades_to_a_note(mock_http):
server = FakeServer()
server.result_for["search"] = {"content": [{"type": "hologram"}]}
mock_http(server)
text, _failed = await client.call_tool(_spec(), "search", {})
assert "hologram" in text
async def test_structured_content_is_used_when_there_is_no_text(mock_http):
server = FakeServer()
server.result_for["search"] = {"content": [], "structuredContent": {"count": 3}}
mock_http(server)
text, _failed = await client.call_tool(_spec(), "search", {})
assert json.loads(text) == {"count": 3}
async def test_is_error_is_reported_as_a_failure(mock_http):
server = FakeServer()
server.result_for["search"] = {
"content": [{"type": "text", "text": "no such repo"}],
"isError": True,
}
mock_http(server)
_text, failed = await client.call_tool(_spec(), "search", {})
assert failed is True
# --- Bounding what a server can say ------------------------------------------
def test_a_giant_description_is_truncated():
cleaned = protocol.clean_tool({"name": "x", "description": "d" * 9000})
assert len(cleaned["description"]) == protocol.MAX_DESCRIPTION
def test_a_schema_that_is_not_an_object_is_replaced():
cleaned = protocol.clean_tool({"name": "x", "inputSchema": "nope"})
assert cleaned["schema"] == {"type": "object", "properties": {}}
def test_a_giant_schema_is_replaced():
huge = {"type": "object", "properties": {f"p{i}": {"type": "string"} for i in range(1000)}}
cleaned = protocol.clean_tool({"name": "x", "inputSchema": huge})
assert cleaned["schema"] == {"type": "object", "properties": {}}
def test_an_entry_with_no_name_is_dropped():
assert protocol.clean_tool({"description": "orphan"}) is None
assert protocol.clean_tool("not a dict") is None
# --- Namespacing -------------------------------------------------------------
def test_tool_names_are_namespaced_and_within_the_charset():
taken: set[str] = set()
assert registry.offer_name("github", "search", taken=taken) == "github_search"
assert registry.offer_name("github", "Create Issue!", taken=taken) == "github_create_issue"
assert all(registry.FUNCTION_NAME.match(name) for name in taken)
def test_a_name_colliding_with_a_builtin_is_renamed():
from lembas.services import tools as tools_service
taken = set(tools_service.REGISTRY)
name = registry.offer_name("", "web_search", taken=taken)
assert name != "web_search"
assert name not in tools_service.REGISTRY
def test_a_truncation_collision_is_disambiguated():
"""Two long names can collide once cut to 64 characters where the full ones
would not have."""
taken: set[str] = set()
first = registry.offer_name("srv", "a" * 200, taken=taken)
second = registry.offer_name("srv", "a" * 201, taken=taken)
assert first != second
assert len(first) <= 64 and len(second) <= 64
# --- Rows become tools -------------------------------------------------------
def _server(db, **overrides) -> McpServer:
row = McpServer(
**{
"slug": "github",
"name": "GitHub",
"url": "https://mcp.test/rpc",
"tools_json": [
{
"name": "search",
"offer_name": "github_search",
"description": "Search things.",
"schema": {"type": "object", "properties": {}},
}
],
**overrides,
}
)
db.add(row)
db.commit()
return row
def test_a_cached_tool_becomes_a_definition(db, user_id):
from lembas.db.models import User
_server(db)
defs = registry.tool_defs(db, db.get(User, user_id))
assert [tool.name for tool in defs] == ["github_search"]
assert defs[0].family == "mcp:github"
def test_a_tool_switched_off_is_not_offered(db, user_id):
from lembas.db.models import User
_server(db, tool_overrides_json={"search": False})
assert registry.tool_defs(db, db.get(User, user_id)) == []
def test_a_tool_absent_from_the_overrides_is_on(db, user_id):
"""Absent means on, the rule the model capability flags follow."""
from lembas.db.models import User
_server(db, tool_overrides_json={"something_else": False})
assert len(registry.tool_defs(db, db.get(User, user_id))) == 1
async def test_refresh_caches_the_tools_and_clears_the_last_error(db, mock_http):
mock_http(FakeServer())
server = _server(db, tools_json=[], last_error="it was broken")
count, error = await registry.refresh(db, server)
assert (count, error) == (1, "")
db.refresh(server)
assert server.tools_json[0]["offer_name"] == "github_search"
assert server.last_error == ""
assert server.protocol_version == "2025-06-18"
assert server.server_info["name"] == "fake"
async def test_a_failed_refresh_records_it_and_keeps_the_cached_tools(db, mock_http):
mock_http(lambda _r: httpx.Response(500, text="down"))
server = _server(db)
count, error = await registry.refresh(db, server)
assert count == 0 and "500" in error
db.refresh(server)
assert server.last_error
assert server.last_checked_at is not None
assert len(server.tools_json) == 1, "a bad refresh must not empty the list"
async def test_a_refresh_keeps_a_choice_about_a_tool_that_survives(db, mock_http):
mock_http(FakeServer(tools=[{"name": "search"}, {"name": "gone"}]))
server = _server(db, tool_overrides_json={"search": False, "vanished": False})
await registry.refresh(db, server)
db.refresh(server)
assert server.tool_overrides_json == {"search": False}
async def test_a_failing_tool_call_reaches_the_model_as_words(db, user_id, mock_http):
from lembas.db.models import User
mock_http(lambda _r: httpx.Response(500, text="down"))
_server(db)
tool = registry.tool_defs(db, db.get(User, user_id))[0]
outcome = await tool.run(None, {})
assert outcome.event["status"] == "error"
assert "GitHub" in outcome.content
def test_two_servers_exposing_the_same_tool_do_not_collide(db, user_id):
from lembas.db.models import User
_server(db)
_server(db, slug="gitlab", name="GitLab", tools_json=[{"name": "search", "schema": {}}])
names = [tool.name for tool in registry.tool_defs(db, db.get(User, user_id))]
assert len(names) == len(set(names)) == 2
def test_a_disabled_server_offers_nothing(db, user_id):
from lembas.db.models import User
_server(db, enabled=False)
assert registry.tool_defs(db, db.get(User, user_id)) == []
assert db.scalar(select(McpServer)) is not None