fix(runtime-admin): swap dead agent.mcp_registry import for plugin helper (Phase 5)

plugins/myah-hermes-plugin/myah_hermes_plugin/myah_platform/runtime_admin.py

@@ -185,11 +185,25 @@ async def disconnect_mcp(request: "web.Request") -> "web.Response":
             return resp
         name = request.match_info["name"]
         try:
-            from agent.mcp_registry import disconnect_mcp_server
+            # Phase 5 (B1 follow-up): the original import targeted
+            # ``agent.mcp_registry`` — a module that does not exist anywhere
+            # in the repo. Every request hit this fallback returning 500
+            # "MCP registry module not available", and the dashboard's
+            # ``DELETE /mcp/<name>`` chain silently failed to actually
+            # disconnect the server from ``tools.mcp_tool._servers``.
+            #
+            # The real plugin-side helper lives at
+            # ``myah_hermes_plugin.runtime_extensions.mcp_disconnect`` and
+            # has been shipped since PR #106 (Phase E). It uses the upstream
+            # ``tools.mcp_tool`` private state + ``_run_on_mcp_loop`` bridge
+            # exactly as the upstream "shutdown all servers" code does.
+            from myah_hermes_plugin.runtime_extensions.mcp_disconnect import (
+                disconnect_mcp_server,
+            )
         except Exception:  # pragma: no cover
             logger.exception("[myah-admin] failed to import disconnect_mcp_server")
             return web.json_response(
-                {"error": "MCP registry module not available"}, status=500
+                {"error": "MCP disconnect helper not available"}, status=500
             )
         try:
             disconnect_mcp_server(name)

plugins/myah-hermes-plugin/tests/test_runtime_admin_mcp_disconnect.py

@@ -0,0 +1,124 @@
+"""Tests for the runtime_admin /myah/v1/admin/mcp/disconnect/{name} endpoint.
+
+Phase 5 regression gate. The handler at runtime_admin.py:disconnect_mcp
+previously imported ``disconnect_mcp_server`` from ``agent.mcp_registry`` —
+a module that doesn't exist anywhere in the repo (verified via repo-wide
+grep). Every call hit the ``except Exception`` branch and returned a 500
+``"MCP registry module not available"``, so the dashboard's
+``DELETE /mcp/<name>`` route — which delegates to this endpoint over HTTP
+(see test_myah_admin_skills_plugins_mcp.py:425) — silently failed to
+actually disconnect anything.
+
+The correct module is ``myah_hermes_plugin.runtime_extensions.mcp_disconnect``
+(Phase E, shipped in PR #106). This test drives ``_make_handlers`` against
+a fake ``tools.mcp_tool`` private state and asserts the endpoint actually
+runs the real disconnect path (returns 200 + ``ok: True``), NOT the
+import-error fallback.
+
+Without this test the dashboard appears to work (the dashboard layer
+captures the 500 and falls through to ``servers.pop()`` in its own
+config-yaml mutation), but the underlying ``tools.mcp_tool._servers``
+dict on the gateway side keeps the stale reference, so the toolset
+cache still treats the MCP server as connected until the gateway
+restarts.
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+from unittest.mock import MagicMock
+
+import pytest
+from aiohttp.test_utils import make_mocked_request
+
+from myah_hermes_plugin.myah_platform.runtime_admin import _make_handlers
+
+
+@pytest.fixture
+def fake_runner():
+    return MagicMock(name='GatewayRunner')
+
+
+@pytest.mark.asyncio
+async def test_disconnect_mcp_runs_real_disconnect_path(fake_runner, monkeypatch):
+    """The endpoint must resolve the plugin-side disconnect helper and
+    actually remove the server from tools.mcp_tool._servers.
+
+    Pre-Phase-5 the handler imported from ``agent.mcp_registry`` (does not
+    exist) and ALWAYS hit the ``except Exception`` branch returning 500.
+    The plugin-side ``runtime_extensions.mcp_disconnect`` has been shipped
+    since PR #106; this test pins the runtime_admin handler to use it.
+    """
+    import tools.mcp_tool as mcp_tool
+
+    fake_server = MagicMock(name='MCPServerTask')
+    fake_server.shutdown.return_value = MagicMock(name='shutdown_coro')
+    servers = {'my-mcp': fake_server}
+
+    def fake_run_on_loop(coro, timeout=15):
+        # Plugin-side disconnect helper drives shutdown() through this
+        # cross-loop bridge. Recording the call is enough — we don't need
+        # to actually await the coro for unit-test purposes.
+        return None
+
+    monkeypatch.setattr(mcp_tool, '_servers', servers, raising=False)
+    monkeypatch.setattr(mcp_tool, '_lock', threading.Lock(), raising=False)
+    monkeypatch.setattr(mcp_tool, '_run_on_mcp_loop', fake_run_on_loop, raising=False)
+
+    handlers = _make_handlers(fake_runner, auth_key='')
+    req = make_mocked_request(
+        'POST',
+        '/myah/v1/admin/mcp/disconnect/my-mcp',
+        match_info={'name': 'my-mcp'},
+    )
+
+    resp = await handlers['disconnect_mcp'](req)
+
+    assert resp.status == 200, (
+        f'Expected 200 from the real disconnect path; got {resp.status} '
+        f'(body={resp.body!r}). If this is 500 with "MCP registry module not '
+        f"available\", the handler is still importing the dead "
+        f'agent.mcp_registry module — apply the Phase 5 import swap.'
+    )
+    body = json.loads(resp.body.decode())
+    assert body == {'ok': True, 'name': 'my-mcp'}
+    # Cross-check: the real disconnect helper popped the server.
+    assert 'my-mcp' not in servers, (
+        'The plugin-side disconnect_mcp_server should have popped the '
+        'server from tools.mcp_tool._servers — if this is still present, '
+        'the handler imported from a stub or returned early.'
+    )
+
+
+@pytest.mark.asyncio
+async def test_disconnect_mcp_no_such_server_still_returns_ok(fake_runner, monkeypatch):
+    """Unknown server name: helper returns False, but the HTTP layer still
+    reports ok=True (the caller already wanted it gone — idempotent).
+
+    This matches the dashboard-layer behaviour at
+    test_myah_admin_skills_plugins_mcp.py::test_remove_mcp_writes_config_and_calls_disconnect_then_evict
+    where the dashboard treats a successful disconnect as the precondition
+    for the cache-evict-all that follows; surfacing a 404 here would
+    needlessly abort the cleanup chain.
+    """
+    import tools.mcp_tool as mcp_tool
+
+    monkeypatch.setattr(mcp_tool, '_servers', {}, raising=False)
+    monkeypatch.setattr(mcp_tool, '_lock', threading.Lock(), raising=False)
+    monkeypatch.setattr(
+        mcp_tool, '_run_on_mcp_loop', lambda c, timeout=15: None, raising=False
+    )
+
+    handlers = _make_handlers(fake_runner, auth_key='')
+    req = make_mocked_request(
+        'POST',
+        '/myah/v1/admin/mcp/disconnect/ghost',
+        match_info={'name': 'ghost'},
+    )
+
+    resp = await handlers['disconnect_mcp'](req)
+
+    assert resp.status == 200
+    body = json.loads(resp.body.decode())
+    assert body == {'ok': True, 'name': 'ghost'}