feat(mcp): per-user catalog scoping (Phase 7 — tools)

Light up production reachability of pool dispatch (RFC §3, invariant 8)
by widening the public catalog API to optionally take a ``user_id``:

- ``MCPClientManager.get_tools(user_id=None)`` returns the merged
  static + per-user pool view when ``user_id`` is supplied; the default
  preserves the legacy global-only contract.
- ``is_mcp_tool(name, *, user_id=None)`` extends the lookup to the
  per-user ``_user_tool_map``. Pool tools become reachable from
  ``ChatSession._prepare_tool`` only when the session-bound user_id
  flows through — flipping invariant 8 from "must hold" to "satisfied".
- Listener identity becomes ``(user_id, callback)``. Static-path
  changes fire ALL listeners (admin + every user); pool-entry
  changes fire only matching-user + admin (``None``) listeners.
  RFC §3.3.
- Pool sessions discover their tool list on first connect
  (``_connect_one_pool`` → ``await session.list_tools()``); the
  notification closure binds to ``(user_id, server_name)`` so
  push-driven ``list_changed`` updates target the correct user's
  catalog. R6 verified empirically: ``list_tools()`` 401 propagates
  through anyio TaskGroup unwinding, no hang — plain ``await`` is
  fine, no carrier-race shape needed for discovery.
- ``_evict_session`` drops ``entry.tools`` and rebuilds the user's
  index so an evicted-then-reconnected session doesn't carry
  stale catalog state.
- ``web_search.resolve_web_search_client`` refuses
  ``auth_type=oauth_user`` backends (per-node web search can't
  carry per-user tokens).

Resources / prompts pool dispatch deferred to Phase 7b — invariant 8
is satisfied by the tool path alone, and the resource/prompt path
needs sibling ``_dispatch_pool_resource_sync`` /
``_dispatch_pool_prompt_sync`` helpers each with their own
carrier-race plumbing (~400 LOC). Phase 7b will follow the patterns
established here.

CLI sessions default ``user_id=""`` and so cannot use oauth_user
MCP servers — documented limitation; users must use the web UI.

Round-1 review fixes (4-finder review applied, no push yet):
- bug-1: get_tools(user_id) was iterating _user_pool_entries from sync
  threads while the mcp-loop concurrently mutated it (RuntimeError:
  dictionary changed size during iteration). Now reads from a sibling
  _user_tools dict updated atomically by _rebuild_user_tool_map.
- bug-2: _close_pool_entry_if_idle (LRU/TTL eviction) skipped the
  catalog cleanup that _evict_session does — stale tools persisted
  in _user_tool_map and ChatSession's tool list never rebuilt. Now
  mirrors _evict_session.
- perf-1: _last_pool_notification_refresh debounce dict was never
  pruned in either eviction path. Now popped alongside the entry.
- perf-3: web_search resolver was issuing a sync SQL query per LLM
  turn to gate oauth_user backends. Now reads from the cached
  in-memory config.
- sec-1: bearer token could leak into exc_info-rendered tracebacks
  via Sentry/faulthandler. log.debug now uses structured fields,
  not exc_info.
- sec-2: tools-per-server response now capped at 1000 (defensive,
  mirrors _MAX_ERROR_LEN / _MAX_INSUFFICIENT_SCOPE_REPORTED).
- Test cleanup: dropped two listener fan-out tests duplicating
  test_mcp_client.py coverage; renamed test_pool_session_notification_handler
  to match its actual scope (_refresh_pool_server_tools); removed
  stale comments referencing /tmp/r6-spike*.py scratchpads and a
  misleading "copy-on-write" comment.

Round-2 pre-push review fixes (focused single-pass review applied):
- round2-1: bug-2's catalog-cleanup block in _close_pool_entry_if_idle
  had no integration test (exactly the failure mode flagged in
  feedback_tests_through_boundaries.md). Added
  test_close_pool_entry_if_idle_clears_catalog_and_fires_listener
  driving the LRU/TTL eviction path through real streamablehttp_client +
  MockTransport. Negative-test verified: reverting the
  _rebuild_user_tool_map / _notify_user_tool_listeners calls makes
  the new test fail.
- round2-3: documented the _oauth_user_server_names cache invariant
  in add_server_sync / remove_server_sync docstrings. Cache is
  reconcile_sync's sole owner — direct callers leave it stale, but
  _db_servers_to_config strips oauth_user rows so production paths
  are unaffected. Static→oauth_user transitions correctly leave the
  name in the cache because remove_server_sync drops the static
  connection, not the cache identity.
- round2-6: strengthened test_rebuild_user_tool_map_populates and
  test_rebuild_user_tool_map_drops_empty_user to assert on the
  _user_tools sibling cache (bug-1 fix). Without this, a future
  revert dropping the sibling write would still pass the unit
  tests because get_tools coverage lives in separate tests.

Round-3 full-stack review fixes (multi-stage review on the final
state caught what the layered apply passes missed):
- q-1 REGRESSION: pool tool-discovery used asyncio.wait_for around
  session.list_tools(), the exact pattern the f6a3b66 fix (and
  feedback_asyncio_timeout_vs_wait_for.md) put in place to avoid.
  Python 3.11's asyncio.wait_for wraps the inner coroutine in a
  fresh task → cross-task scope-exit when the SDK's anyio TaskGroup
  unwinds on a 401. Switched to `async with asyncio.timeout(...):`
  pattern used by _safe_close_stack.
- sec-2: TOCTOU in _connect_one_pool — entry.tools was published
  (via _rebuild_user_tool_map + listener fan-out) BEFORE entry.session
  was assigned. A sync-thread reader could observe a tool whose
  backing entry has session=None. Defence-in-depth — dispatch
  re-fetches its own token and lazy-reconnects on session=None — but
  reordering catches the race at the source. entry.session now
  publishes BEFORE catalog visibility.
- bug-1: _close_pool_entry_if_idle's _user_pool_locks.pop ran
  unconditionally after the try/finally, but the early-return
  branches (entry None on re-check, in_flight > 0 under lock) skip
  it via Python's return-through-finally semantics. The lock was
  never popped on those paths. Now gated behind an `evicted` flag
  set only on the success path; in_flight > 0 leaves the lock for
  the active dispatcher to reuse, entry-None races leave the lock
  for re-allocation by _ensure_pool_entry. Comment now describes
  the actual semantics, not the original promise.
- bug-2: softened the _rebuild_user_tool_map docstring's atomicity
  claim. The two-dict write is technically non-atomic across Python
  statements; in practice the window is sub-microsecond on the
  mcp-loop with no awaits between writes, and the listener fan-out
  fires AFTER both writes complete. Docstring now says "back-to-back
  on the mcp-loop" instead of "atomically alongside".
- q-3: dropped `hasattr(mcp_client, "server_auth_type")` defensive
  check in web_search.py. The method ships in this commit; the
  hasattr created a silent fallthrough that would let a future
  rename silently re-enable oauth_user backends.
- q-4: surfaced the CLI / empty-user_id limitation in a docstring
  comment at ChatSession.__init__'s self._user_id assignment. The
  note previously lived only inside is_mcp_tool's docstring — a
  future maintainer wiring CLI features against MCP pool servers
  wouldn't think to read is_mcp_tool to find the constraint.
- q-2 + q-5: deleted a tautological duplicate test in
  test_mcp_user_catalog.py whose docstring claimed to test
  ChatSession.close but never instantiated a ChatSession (the
  manager-level identity semantics are already covered by
  test_listener_identity_includes_user_id in the same file and by
  test_session_close_removes_listener_with_same_user_id in
  test_mcp_client.py which DOES drive a ChatSession). Reworded a
  misleading "fixture provides only 5s" comment to point at the
  actual `_run_on_loop(..., timeout=5)` site.
- q-6: the `self._user_id or None` collapse repeated at 8 sites
  across session.py. Cached once at __init__ as
  ``self._mcp_user_id`` (since ``_user_id`` is set once and never
  mutated); 8 call sites now read the cached value. The empty-
  string-is-CLI-sentinel invariant is documented at the assignment
  site, not re-asserted at each consumer.

Deferred to follow-up:
- sec-1: a hostile MCP server bound to user-A could craft a
  tool.name containing `__` to synthesize a prefixed-name collision
  in user-A's own catalog. Bounded impact: cross-tenant dispatch is
  prevented by the per-tenant token gate in _dispatch_pool, and
  user-B's get_tools(user_id="B") never includes user-A's pool
  entries. The fix needs policy decisions (reject vs. sanitize)
  and touches _mcp_to_openai which is shared between static and
  pool paths; better discussed in its own follow-up where the
  policy applies uniformly to static-path servers too. The threat
  model already requires user-A to have consented to a malicious
  server, who has many more dangerous vectors than tool-name
  shenanigans.

Test count delta: +31 tests (5435 → 5466, ``-m "not live"``; one
test deleted in round-3 apply per q-2):
- ``tests/test_mcp_client.py`` +20 (per-user catalog state, listener
  identity, session thread-through)
- ``tests/test_mcp_user_catalog.py`` +9 NEW (integration tests
  driving real ``streamablehttp_client`` + ``httpx.MockTransport`` per
  invariant 14: discovery on connect, user isolation, eviction +
  reconnect, LRU/TTL eviction (round2-1), R6 401-propagation
  regression, static byte-identical canonical regression; review
  passes dropped duplicate listener fan-out tests from earlier
  drafts whose coverage lived in test_mcp_client.py)
- ``tests/test_web_search.py`` +2 (oauth_user backend rejection +
  static backend acceptance regression; updated to use the new
  ``server_auth_type`` in-memory accessor)
This commit is contained in:
Patrick Buckley
2026-05-06 12:52:41 -07:00
parent 0fbf31e713
commit dad98c062d
7 changed files with 1702 additions and 54 deletions
+358
View File
@@ -397,6 +397,217 @@ class TestMCPClientManager:
mgr = MCPClientManager({})
mgr.shutdown() # should be a no-op
# -- Phase 7: per-user catalog scoping ---------------------------------
def test_is_mcp_tool_user_id_default_none_unchanged(self):
"""Sanity: default ``user_id=None`` answers static-only.
The legacy single-arg call still works, and unknown names still
return False — Phase 7 adds an optional keyword without
rewriting the static-path semantics.
"""
mgr = MCPClientManager({})
mgr._tool_map["mcp__static__list"] = ("static", "list")
# Legacy single-arg call still works.
assert mgr.is_mcp_tool("mcp__static__list") is True
assert mgr.is_mcp_tool("mcp__static__list", user_id=None) is True
assert mgr.is_mcp_tool("nonexistent") is False
assert mgr.is_mcp_tool("nonexistent", user_id=None) is False
def test_is_mcp_tool_user_keyed_pool_tool(self):
"""A name visible only via ``_user_tool_map`` resolves only for
the matching ``user_id``.
Verifies the new branch: ``_tool_map`` miss + ``user_id`` hit.
"""
mgr = MCPClientManager({})
mgr._user_tool_map["user-1"] = {
"mcp__pool-srv__do": ("pool-srv", "do"),
}
# Visible to user-1.
assert mgr.is_mcp_tool("mcp__pool-srv__do", user_id="user-1") is True
# Invisible to None caller (admin / web-search backend resolution).
assert mgr.is_mcp_tool("mcp__pool-srv__do", user_id=None) is False
# Invisible to a different user.
assert mgr.is_mcp_tool("mcp__pool-srv__do", user_id="user-2") is False
def test_is_mcp_tool_static_wins_for_any_user(self):
"""Static-path tools are visible regardless of ``user_id`` —
the merged view is ``static user-pool``."""
mgr = MCPClientManager({})
mgr._tool_map["mcp__static__list"] = ("static", "list")
# Even for an unknown user, a static tool is still reachable —
# static-path is process-global.
assert mgr.is_mcp_tool("mcp__static__list", user_id="user-1") is True
assert mgr.is_mcp_tool("mcp__static__list", user_id="anybody") is True
def test_get_tools_user_id_none_returns_static_only(self):
"""``get_tools(user_id=None)`` returns the global static catalog.
Pool tools are NEVER included in the default-arg view — that's
the legacy contract every pre-Phase-7 caller relies on.
"""
mgr = MCPClientManager({})
mgr._tools = [_fake_openai_tool("mcp__static__list")]
# Seed a pool entry that should NOT appear in the default view.
from turnstone.core.mcp_client import PoolEntryState
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
entry.tools = [_fake_openai_tool("mcp__pool-srv__do")]
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
tools = mgr.get_tools()
names = [t["function"]["name"] for t in tools]
assert names == ["mcp__static__list"]
def test_get_tools_user_id_merges_pool(self):
"""``get_tools(user_id='user-1')`` merges static + that user's pool tools.
Other users' pool entries MUST NOT leak into the result —
privacy / RBAC invariant.
"""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
mgr._tools = [_fake_openai_tool("mcp__static__list")]
e1 = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
e1.tools = [_fake_openai_tool("mcp__pool-srv__do")]
e2 = PoolEntryState(key=("user-2", "pool-srv"), open_lock=MagicMock())
e2.tools = [_fake_openai_tool("mcp__pool-srv__other")]
mgr._user_pool_entries[("user-1", "pool-srv")] = e1
mgr._user_pool_entries[("user-2", "pool-srv")] = e2
# Production invariant: ``_connect_one_pool`` /
# ``_refresh_pool_server_tools`` / ``_evict_session`` /
# ``_close_pool_entry_if_idle`` all call ``_rebuild_user_tool_map``
# immediately after mutating ``_user_pool_entries``. Tests that
# seed pool entries directly must mirror that invariant —
# ``get_tools(user_id=...)`` reads from the ``_user_tools``
# snapshot (built by ``_rebuild_user_tool_map``), never iterating
# ``_user_pool_entries`` directly.
mgr._rebuild_user_tool_map("user-1")
mgr._rebuild_user_tool_map("user-2")
u1_names = [t["function"]["name"] for t in mgr.get_tools(user_id="user-1")]
assert sorted(u1_names) == ["mcp__pool-srv__do", "mcp__static__list"]
u2_names = [t["function"]["name"] for t in mgr.get_tools(user_id="user-2")]
assert sorted(u2_names) == ["mcp__pool-srv__other", "mcp__static__list"]
# Default still global-only — unaffected by either user's entries.
default_names = [t["function"]["name"] for t in mgr.get_tools()]
assert default_names == ["mcp__static__list"]
def test_get_tools_user_id_returns_copies(self):
"""Mirror existing ``test_get_tools_returns_copy``: caller mutation
of the returned list MUST NOT affect the manager's catalog."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
mgr._tools = [_fake_openai_tool("mcp__static__a")]
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
entry.tools = [_fake_openai_tool("mcp__pool-srv__b")]
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
# See note in ``test_get_tools_user_id_merges_pool``.
mgr._rebuild_user_tool_map("user-1")
tools = mgr.get_tools(user_id="user-1")
assert len(tools) == 2
tools.clear()
# Re-fetch — original catalog unchanged.
assert len(mgr.get_tools(user_id="user-1")) == 2
def test_get_tools_user_with_none_tools_skipped(self):
"""A pool entry that hasn't completed discovery (``entry.tools is None``)
contributes no tools — the merged view skips it cleanly."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
mgr._tools = [_fake_openai_tool("mcp__static__a")]
# Brand-new pool entry, discovery not yet run.
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
assert entry.tools is None
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
# Rebuild observes ``entry.tools is None`` and skips this entry.
mgr._rebuild_user_tool_map("user-1")
names = [t["function"]["name"] for t in mgr.get_tools(user_id="user-1")]
assert names == ["mcp__static__a"]
def test_rebuild_user_tool_map_populates(self):
"""``_rebuild_user_tool_map`` materializes the per-user index from
pool entries owned by that user."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
entry.tools = [_fake_openai_tool("mcp__pool-srv__do")]
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
mgr._rebuild_user_tool_map("user-1")
assert mgr._user_tool_map["user-1"] == {"mcp__pool-srv__do": ("pool-srv", "do")}
# Sibling _user_tools cache (bug-1 fix) MUST be populated alongside
# the map — otherwise get_tools(user_id="user-1") would silently
# return the static-only view despite is_mcp_tool returning True.
assert mgr._user_tools["user-1"] == [_fake_openai_tool("mcp__pool-srv__do")]
def test_rebuild_user_tool_map_drops_empty_user(self):
"""Rebuilding for a user with no pool entries removes the key
rather than retaining an empty-dict sentinel."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
entry.tools = [_fake_openai_tool("mcp__pool-srv__do")]
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
mgr._rebuild_user_tool_map("user-1")
assert "user-1" in mgr._user_tool_map
assert "user-1" in mgr._user_tools
# Drop the entry, rebuild — user_id key should be removed from BOTH
# the map and the sibling tool list (bug-1 fix). A drop in only one
# would leave get_tools and is_mcp_tool out of sync.
mgr._user_pool_entries.clear()
mgr._rebuild_user_tool_map("user-1")
assert "user-1" not in mgr._user_tool_map
assert "user-1" not in mgr._user_tools
def test_rebuild_user_tool_map_isolates_users(self):
"""Rebuilding for ``user-1`` MUST NOT touch ``user-2``'s entry."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
e1 = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
e1.tools = [_fake_openai_tool("mcp__pool-srv__one")]
e2 = PoolEntryState(key=("user-2", "pool-srv"), open_lock=MagicMock())
e2.tools = [_fake_openai_tool("mcp__pool-srv__two")]
mgr._user_pool_entries[("user-1", "pool-srv")] = e1
mgr._user_pool_entries[("user-2", "pool-srv")] = e2
mgr._rebuild_user_tool_map("user-1")
mgr._rebuild_user_tool_map("user-2")
assert mgr._user_tool_map["user-1"] == {"mcp__pool-srv__one": ("pool-srv", "one")}
assert mgr._user_tool_map["user-2"] == {"mcp__pool-srv__two": ("pool-srv", "two")}
# Clear user-1's entry only; rebuild user-1; user-2 must remain.
mgr._user_pool_entries.pop(("user-1", "pool-srv"))
mgr._rebuild_user_tool_map("user-1")
assert "user-1" not in mgr._user_tool_map
assert mgr._user_tool_map["user-2"] == {"mcp__pool-srv__two": ("pool-srv", "two")}
def test_rebuild_user_tool_map_does_not_touch_static(self):
"""Invariant 1: per-user rebuild must NOT mutate ``_tool_map``."""
from turnstone.core.mcp_client import PoolEntryState
mgr = MCPClientManager({})
mgr._tool_map["mcp__static__list"] = ("static", "list")
entry = PoolEntryState(key=("user-1", "pool-srv"), open_lock=MagicMock())
entry.tools = [_fake_openai_tool("mcp__pool-srv__do")]
mgr._user_pool_entries[("user-1", "pool-srv")] = entry
before = dict(mgr._tool_map)
mgr._rebuild_user_tool_map("user-1")
assert mgr._tool_map == before
# ---------------------------------------------------------------------------
# Session integration (mock MCP client)
@@ -597,6 +808,84 @@ class TestSessionIntegration:
assert "MCP tool error" in output
assert "server crashed" in output
# -- Phase 7: per-user catalog scoping ---------------------------------
def test_session_passes_user_id_to_get_tools(self, tmp_db):
"""ChatSession threads its ``user_id`` into ``get_tools`` so the
merged static + pool view is scoped to the session's user."""
mock_mcp = MagicMock()
mock_mcp.get_tools.return_value = [_fake_openai_tool()]
self._make_session(mcp_client=mock_mcp, user_id="user-7")
mock_mcp.get_tools.assert_called_with(user_id="user-7")
def test_session_get_tools_empty_user_id_passes_none(self, tmp_db):
"""Sentinel ``user_id=""`` (CLI / service / unknown) collapses to
``user_id=None`` so the static-only view is returned."""
mock_mcp = MagicMock()
mock_mcp.get_tools.return_value = [_fake_openai_tool()]
self._make_session(mcp_client=mock_mcp, user_id="")
mock_mcp.get_tools.assert_called_with(user_id=None)
def test_session_passes_user_id_to_add_listener(self, tmp_db):
"""ChatSession registers its tool-change listener under its own
``user_id`` so pool-only changes for OTHER users do not fire it."""
mock_mcp = MagicMock()
mock_mcp.get_tools.return_value = [_fake_openai_tool()]
self._make_session(mcp_client=mock_mcp, user_id="user-7")
# ``add_listener`` was called with ``user_id="user-7"``.
listener_calls = mock_mcp.add_listener.call_args_list
assert listener_calls, "ChatSession did not register a tool listener"
first_call = listener_calls[0]
assert first_call.kwargs.get("user_id") == "user-7"
def test_session_close_removes_listener_with_same_user_id(self, tmp_db):
"""R4 critical: register and remove MUST agree on ``user_id`` —
the listener identity is ``(user_id, callback)``."""
mock_mcp = MagicMock()
mock_mcp.get_tools.return_value = [_fake_openai_tool()]
session = self._make_session(mcp_client=mock_mcp, user_id="user-7")
session.close()
# ``remove_listener`` must be called with the same ``user_id``.
remove_calls = mock_mcp.remove_listener.call_args_list
assert remove_calls, "ChatSession.close did not unregister a tool listener"
first_remove = remove_calls[0]
assert first_remove.kwargs.get("user_id") == "user-7"
# And the callback identity must match what was registered.
registered_cb = mock_mcp.add_listener.call_args_list[0].args[0]
removed_cb = first_remove.args[0]
assert registered_cb is removed_cb
def test_session_unknown_tool_lists_user_scoped_catalog(self, tmp_db):
"""The "Unknown tool" error message lists tools the session can
actually invoke — drawn from the merged user-scoped catalog,
not the manager's private static-only ``_tool_map``."""
mock_mcp = MagicMock()
# Pretend the user's merged view contains a static + pool entry.
mock_mcp.get_tools.return_value = [
_fake_openai_tool("mcp__static__list"),
_fake_openai_tool("mcp__pool-srv__do"),
]
mock_mcp.is_mcp_tool.return_value = False
session = self._make_session(mcp_client=mock_mcp, user_id="user-7")
# Reset the call counter so we observe only the _prepare_tool call.
mock_mcp.get_tools.reset_mock()
tc = {
"id": "call_unknown",
"function": {"name": "no_such_tool", "arguments": "{}"},
}
prepared = session._prepare_tool(tc)
assert "error" in prepared
# The error mentions both static and pool tools — proves we're
# consulting the merged catalog rather than ``_tool_map``.
assert "mcp__static__list" in prepared["error"]
assert "mcp__pool-srv__do" in prepared["error"]
# And the catalog request was scoped to this session's user.
assert any(
call.kwargs.get("user_id") == "user-7" for call in mock_mcp.get_tools.call_args_list
)
# ---------------------------------------------------------------------------
# Server name validation
@@ -794,6 +1083,75 @@ class TestListeners:
mgr.add_listener(lambda: 1 / 0) # will raise ZeroDivisionError
mgr._rebuild_tools() # should not raise
# -- Phase 7: user-keyed listener fan-out ------------------------------
def test_add_listener_records_user_id(self):
"""``add_listener`` stores ``(user_id, callback)`` tuples — the
listener identity carries the user_id."""
mgr = MCPClientManager({})
cb_admin = lambda: None # noqa: E731
cb_user = lambda: None # noqa: E731
mgr.add_listener(cb_admin) # default: user_id=None (admin)
mgr.add_listener(cb_user, user_id="user-1")
assert (None, cb_admin) in mgr._listeners
assert ("user-1", cb_user) in mgr._listeners
def test_remove_listener_requires_matching_user_id(self):
"""Removing with a different ``user_id`` must NOT remove the
original registration — listener identity is the pair."""
mgr = MCPClientManager({})
calls: list[int] = []
cb = lambda: calls.append(1) # noqa: E731
mgr.add_listener(cb, user_id="user-1")
# Try to remove with the wrong user_id — should be a no-op.
mgr.remove_listener(cb, user_id="user-2")
# The user-1 listener should still be live.
mgr._notify_user_tool_listeners("user-1")
assert calls == [1]
# Now remove with the right user_id.
mgr.remove_listener(cb, user_id="user-1")
mgr._notify_user_tool_listeners("user-1")
assert calls == [1] # not invoked again
def test_static_change_fires_all_listeners(self):
"""``_rebuild_tools`` (static-path change) fires ALL registered
listeners — admin + every user. RFC §3.3."""
mgr = MCPClientManager({})
admin_calls: list[int] = []
u1_calls: list[int] = []
u2_calls: list[int] = []
mgr.add_listener(lambda: admin_calls.append(1))
mgr.add_listener(lambda: u1_calls.append(1), user_id="user-1")
mgr.add_listener(lambda: u2_calls.append(1), user_id="user-2")
_seed_static_state(mgr, "a", tools=[_fake_openai_tool("mcp__a__x")])
mgr._rebuild_tools()
assert admin_calls == [1]
assert u1_calls == [1]
assert u2_calls == [1]
def test_user_tool_listeners_only_fire_for_matching_user(self):
"""``_notify_user_tool_listeners('user-1')`` fires admin (None)
and user-1 listeners; user-2's listener is silent."""
mgr = MCPClientManager({})
admin_calls: list[int] = []
u1_calls: list[int] = []
u2_calls: list[int] = []
mgr.add_listener(lambda: admin_calls.append(1))
mgr.add_listener(lambda: u1_calls.append(1), user_id="user-1")
mgr.add_listener(lambda: u2_calls.append(1), user_id="user-2")
mgr._notify_user_tool_listeners("user-1")
assert admin_calls == [1]
assert u1_calls == [1]
assert u2_calls == []
mgr._notify_user_tool_listeners("user-2")
assert admin_calls == [1, 1]
assert u1_calls == [1]
assert u2_calls == [1]
class TestServerNames:
def test_server_names_property(self):
+718
View File
@@ -0,0 +1,718 @@
"""Phase 7 integration tests — per-user catalog scoping.
These tests drive ``_connect_one_pool`` against a real
``streamablehttp_client`` with the SDK's response hook bound to an
``httpx.MockTransport`` (R6 verified the happy path closes cleanly and
a mid-discovery 401 propagates through anyio TaskGroup unwinding).
Direct method injection on ``MagicMock`` as the SOLE gate is
forbidden — invariant 14 of the OAuth-MCP RFC.
The test scaffolding mirrors ``tests/test_mcp_pool_auth_introspection.py``
(per-test ``running_loop_mgr`` fixture, ``_seed_oauth_server`` helper,
``_run_on_loop``) and adds a tool-discovery transport that returns
``initialize`` + ``tools/list`` responses programmatically.
"""
from __future__ import annotations
import asyncio
import contextlib
import threading
import time
from typing import Any
from unittest.mock import MagicMock
import httpx
import pytest
from turnstone.core.mcp_client import (
MCPClientManager,
PoolEntryState,
_AuthCapture,
_make_capturing_http_factory,
)
# ---------------------------------------------------------------------------
# Fixtures and helpers
# ---------------------------------------------------------------------------
@pytest.fixture
def running_loop_mgr() -> Any:
"""Background mcp-loop fixture. Mirrors test_mcp_pool_auth_introspection."""
cfg: dict[str, Any] = {}
mgr = MCPClientManager(cfg)
loop = asyncio.new_event_loop()
thread = threading.Thread(target=loop.run_forever, daemon=True, name="mcp-pool-test-loop")
thread.start()
mgr._loop = loop
try:
yield mgr, loop, thread
finally:
async def _drain(m: MCPClientManager) -> None:
task = m._user_pool_eviction_task
if task is not None:
task.cancel()
with contextlib.suppress(BaseException):
await task
m._user_pool_eviction_task = None
with contextlib.suppress(Exception):
asyncio.run_coroutine_threadsafe(_drain(mgr), loop).result(timeout=2)
loop.call_soon_threadsafe(loop.stop)
thread.join(timeout=2)
def _run_on_loop(loop: asyncio.AbstractEventLoop, coro: Any, timeout: float = 10) -> Any:
fut = asyncio.run_coroutine_threadsafe(coro, loop)
return fut.result(timeout=timeout)
def _build_mock_transport_factory(
mgr: MCPClientManager,
monkeypatch: pytest.MonkeyPatch,
handler: Any,
) -> None:
"""Wrap ``_make_capturing_http_factory`` so the resulting httpx
client uses an ``httpx.MockTransport(handler)``.
The production capturing factory is preserved (we still get the
response-hook plumbing); only the underlying transport is swapped.
Mirrors the spike v3 pattern that R6 used to verify discovery
behaviour against the real SDK.
"""
real_factory = _make_capturing_http_factory
def _wrapped_factory(
capture: _AuthCapture,
fired_event: asyncio.Event | None = None,
) -> Any:
inner = real_factory(capture, fired_event=fired_event)
def _factory(*args: Any, **kwargs: Any) -> httpx.AsyncClient:
client = inner(*args, **kwargs)
client._transport = httpx.MockTransport(handler)
return client
return _factory
monkeypatch.setattr(
"turnstone.core.mcp_client._make_capturing_http_factory",
_wrapped_factory,
)
def _patch_tcp_probe(mgr: MCPClientManager, monkeypatch: pytest.MonkeyPatch) -> None:
"""Skip the TCP probe — MockTransport never opens a real socket."""
async def _noop_probe(*_args: Any, **_kwargs: Any) -> None:
return None
monkeypatch.setattr(MCPClientManager, "_tcp_probe", _noop_probe)
def _make_jsonrpc_handler(
*,
init_response: dict[str, Any] | None = None,
list_tools_response: dict[str, Any] | None = None,
list_tools_status: int = 200,
list_tools_error_payload: dict[str, Any] | None = None,
list_tools_seq: list[dict[str, Any]] | None = None,
counter: list[int] | None = None,
record_bodies: list[str] | None = None,
) -> Any:
"""Build an httpx async handler for the ``streamable-http`` shape.
Returns a coroutine the MockTransport invokes per request.
"""
if init_response is None:
init_response = {
"jsonrpc": "2.0",
"id": 0,
"result": {
"protocolVersion": "2025-06-18",
"capabilities": {"tools": {"listChanged": False}},
"serverInfo": {"name": "fake", "version": "1"},
},
}
if list_tools_response is None:
list_tools_response = {
"jsonrpc": "2.0",
"id": 1,
"result": {"tools": []},
}
counter = counter if counter is not None else [0]
list_tools_seq = list_tools_seq or []
list_tools_index = [0]
async def _handler(req: httpx.Request) -> httpx.Response:
if req.method == "GET":
return httpx.Response(405)
if req.method == "DELETE":
return httpx.Response(200)
body = req.content.decode() if req.content else ""
if record_bodies is not None:
record_bodies.append(body)
if "notifications/initialized" in body:
return httpx.Response(202)
counter[0] += 1
if "method" in body and '"initialize"' in body:
return httpx.Response(
200,
headers={"content-type": "application/json", "mcp-session-id": "sess-1"},
json=init_response,
)
# ``tools/list`` — by request order: explicit override first, then
# the staged sequence, then fall back to ``list_tools_response``.
if list_tools_status != 200:
return httpx.Response(
list_tools_status,
headers={"www-authenticate": 'Bearer error="invalid_token"'},
json=list_tools_error_payload or {"error": "unauthorized"},
)
if list_tools_seq and list_tools_index[0] < len(list_tools_seq):
payload = list_tools_seq[list_tools_index[0]]
list_tools_index[0] += 1
return httpx.Response(
200,
headers={"content-type": "application/json"},
json=payload,
)
return httpx.Response(
200,
headers={"content-type": "application/json"},
json=list_tools_response,
)
return _handler
def _list_tools_payload(tools: list[dict[str, Any]], req_id: int = 1) -> dict[str, Any]:
"""Build a minimal ``tools/list`` JSON-RPC result payload."""
return {
"jsonrpc": "2.0",
"id": req_id,
"result": {"tools": tools},
}
def _tool_spec(name: str, description: str = "") -> dict[str, Any]:
return {
"name": name,
"description": description or f"do {name}",
"inputSchema": {"type": "object", "properties": {}},
}
def _connect_pool(
mgr: MCPClientManager,
loop: asyncio.AbstractEventLoop,
*,
user_id: str,
server_name: str,
url: str = "https://mcp.example.com/sse",
access_token: str = "access-aaa",
) -> PoolEntryState:
"""Drive ``_connect_one_pool`` once; returns the resulting entry."""
cfg: dict[str, Any] = {"type": "streamable-http", "url": url, "headers": {}}
capture = _AuthCapture()
fired = asyncio.Event()
async def _go() -> PoolEntryState:
return await mgr._connect_one_pool(
(user_id, server_name),
cfg,
access_token,
auth_capture=capture,
auth_fired_event=fired,
)
return _run_on_loop(loop, _go())
# ---------------------------------------------------------------------------
# Discovery on first connect — drives REAL streamablehttp_client + httpx
# MockTransport per invariant 14.
# ---------------------------------------------------------------------------
def test_discovery_on_pool_connect_via_real_streamable_http(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""``_connect_one_pool`` discovers the user's tool catalog after
``initialize()`` and stores it on the entry.
Drives through the real ``streamablehttp_client`` and the real
SDK ``ClientSession`` (only the underlying httpx transport is
swapped). Asserts on observable state rather than mock call count
so a refactor that produces the same end state still passes.
Verified by reverting the discovery block in ``_connect_one_pool``
(the ``await session.list_tools()`` + ``_rebuild_user_tool_map``
calls): this test fails because ``entry.tools`` stays ``None`` and
``_user_tool_map`` never gains the user_id key.
"""
mgr, loop, _ = running_loop_mgr
handler = _make_jsonrpc_handler(
list_tools_response=_list_tools_payload(
[_tool_spec("do_thing"), _tool_spec("other")],
),
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
entry = _connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
assert entry.session is not None
assert entry.tools is not None
tool_names = {t["function"]["name"] for t in entry.tools}
assert tool_names == {"mcp__pool-srv__do_thing", "mcp__pool-srv__other"}
# Per-user index reflects the discovery.
user_map = mgr._user_tool_map.get("user-1")
assert user_map is not None
assert "mcp__pool-srv__do_thing" in user_map
assert user_map["mcp__pool-srv__do_thing"] == ("pool-srv", "do_thing")
# is_mcp_tool with the user's id sees the discovered tool.
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing", user_id="user-1") is True
# Default (None) caller does NOT — pool is per-user.
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing") is False
def test_pool_tool_visibility_user_isolation(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""Two users connecting to the same server have independent
``_user_tool_map`` entries; one user never sees another's catalog."""
mgr, loop, _ = running_loop_mgr
# Both users see the same set, but the per-user maps must stay
# independent — privacy is a structural property even when the
# contents happen to match.
handler = _make_jsonrpc_handler(
list_tools_response=_list_tools_payload([_tool_spec("shared_tool")]),
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
_connect_pool(mgr, loop, user_id="user-2", server_name="pool-srv")
map_1 = mgr._user_tool_map.get("user-1")
map_2 = mgr._user_tool_map.get("user-2")
assert map_1 is not None and map_2 is not None
assert map_1 is not map_2
assert "mcp__pool-srv__shared_tool" in map_1
assert "mcp__pool-srv__shared_tool" in map_2
# Cross-user visibility is forbidden via is_mcp_tool: user-1's name
# presence implies nothing about user-3.
assert mgr.is_mcp_tool("mcp__pool-srv__shared_tool", user_id="user-3") is False
def test_eviction_drops_catalog_and_fires_listener(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""``_evict_session`` clears ``entry.tools``, rebuilds the user's
map (drops the now-empty entry), and fires user + admin listeners.
Verified by reverting the catalog-cleanup block in ``_evict_session``
(drop the ``evict.tools = None`` / ``_rebuild_user_tool_map`` /
``_notify_user_tool_listeners`` calls): the test fails because the
listener never fires AND ``is_mcp_tool`` keeps returning True for
the now-evicted tool.
"""
mgr, loop, _ = running_loop_mgr
handler = _make_jsonrpc_handler(
list_tools_response=_list_tools_payload([_tool_spec("do_thing")]),
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing", user_id="user-1") is True
# Register one user-keyed and one admin (None) listener; both
# MUST fire on this user's eviction.
user_calls = [0]
admin_calls = [0]
other_calls = [0]
def _user_cb() -> None:
user_calls[0] += 1
def _admin_cb() -> None:
admin_calls[0] += 1
def _other_cb() -> None:
other_calls[0] += 1
mgr.add_listener(_user_cb, user_id="user-1")
mgr.add_listener(_admin_cb) # admin / None
mgr.add_listener(_other_cb, user_id="user-2")
mgr._evict_session(("user-1", "pool-srv"))
# Catalog cleared.
entry = mgr._user_pool_entries[("user-1", "pool-srv")]
assert entry.session is None
assert entry.tools is None
# User map dropped (no remaining pool entries for this user).
assert "user-1" not in mgr._user_tool_map
# is_mcp_tool no longer surfaces the evicted name.
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing", user_id="user-1") is False
# Listener fan-out: matching user + admin fire, OTHER user does not.
assert user_calls[0] == 1, f"user-keyed listener fired {user_calls[0]} times; expected 1"
assert admin_calls[0] == 1, f"admin listener fired {admin_calls[0]} times; expected 1"
assert other_calls[0] == 0, (
f"unrelated user-2 listener fired {other_calls[0]} times; expected 0 — "
"RFC §3.3 privacy violation."
)
def test_close_pool_entry_if_idle_clears_catalog_and_fires_listener(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""LRU/TTL eviction (``_close_pool_entry_if_idle``) mirrors
``_evict_session``'s catalog-cleanup contract: drops the entry,
prunes the notification-debounce dict, rebuilds the user's tool
map, and fires user + admin listeners.
Phase 7 round-2 review hardening (round2-1): the bug-2 fix added
the catalog-cleanup block to this method but no integration test
drove it — exactly the failure mode flagged in
``feedback_tests_through_boundaries.md``. Negative test: drop the
``_rebuild_user_tool_map`` / ``_notify_user_tool_listeners`` calls
from ``_close_pool_entry_if_idle``'s post-pop block; this test
fails because ``is_mcp_tool`` keeps returning ``True`` for the
evicted name AND no listener fires.
"""
mgr, loop, _ = running_loop_mgr
handler = _make_jsonrpc_handler(
list_tools_response=_list_tools_payload([_tool_spec("do_thing")]),
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
key = ("user-1", "pool-srv")
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing", user_id="user-1") is True
# Sanity: in_flight must be 0 for the eviction path to proceed.
assert mgr._user_pool_entries[key].in_flight == 0
# Seed the debounce dict so the perf-1 prune is observable.
mgr._last_pool_notification_refresh[key] = 0.0
user_calls = [0]
admin_calls = [0]
other_calls = [0]
def _user_cb() -> None:
user_calls[0] += 1
def _admin_cb() -> None:
admin_calls[0] += 1
def _other_cb() -> None:
other_calls[0] += 1
mgr.add_listener(_user_cb, user_id="user-1")
mgr.add_listener(_admin_cb) # admin / None
mgr.add_listener(_other_cb, user_id="user-2")
# Drive the LRU/TTL eviction path directly. ``open_lock`` is
# uncontested (no concurrent dispatch) and ``in_flight`` is 0,
# so the close proceeds without retry.
_run_on_loop(loop, mgr._close_pool_entry_if_idle(key))
# Entry fully removed (LRU eviction pops the dict — unlike
# ``_evict_session`` which keeps the entry as a ``session=None``
# phantom for the next dispatch to re-connect).
assert key not in mgr._user_pool_entries
assert key not in mgr._user_pool_last_used
assert key not in mgr._user_pool_locks
# perf-1 prune: debounce dict no longer carries the key.
assert key not in mgr._last_pool_notification_refresh
# Catalog cleanup ran in BOTH dicts (bug-1 sibling + bug-2 cleanup).
assert "user-1" not in mgr._user_tool_map
assert "user-1" not in mgr._user_tools
assert mgr.is_mcp_tool("mcp__pool-srv__do_thing", user_id="user-1") is False
# Listener fan-out: matching user + admin fire, OTHER user does not.
assert user_calls[0] == 1, f"user-keyed listener fired {user_calls[0]} times; expected 1"
assert admin_calls[0] == 1, f"admin listener fired {admin_calls[0]} times; expected 1"
assert other_calls[0] == 0, (
f"unrelated user-2 listener fired {other_calls[0]} times; expected 0 — "
"RFC §3.3 privacy violation."
)
def test_reconnect_after_eviction_repopulates_catalog(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""After eviction, the next ``_connect_one_pool`` re-populates the
catalog from the SDK by construction — no extra glue required.
Verified by reverting the discovery block in ``_connect_one_pool``:
the reconnected entry's ``tools`` stays ``None`` and the user map
never re-emerges.
"""
mgr, loop, _ = running_loop_mgr
# Sequence: first connect sees [tool_a]; eviction clears; second
# connect (after backend rotates) sees [tool_b].
handler = _make_jsonrpc_handler(
list_tools_seq=[
_list_tools_payload([_tool_spec("tool_a")]),
_list_tools_payload([_tool_spec("tool_b")], req_id=1),
],
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
assert mgr.is_mcp_tool("mcp__pool-srv__tool_a", user_id="user-1") is True
mgr._evict_session(("user-1", "pool-srv"))
assert mgr.is_mcp_tool("mcp__pool-srv__tool_a", user_id="user-1") is False
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
# Reconnect picked up the rotated catalog.
assert mgr.is_mcp_tool("mcp__pool-srv__tool_b", user_id="user-1") is True
# Old name was correctly purged — not still hanging around.
assert mgr.is_mcp_tool("mcp__pool-srv__tool_a", user_id="user-1") is False
def test_refresh_pool_server_tools_isolates_user(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""``_refresh_pool_server_tools`` refreshes THIS user's catalog only,
leaving the static path and other users' catalogs untouched.
Drives ``_refresh_pool_server_tools`` directly (which is also what
the pool session's notification handler invokes) and asserts:
* The pool entry's ``tools`` field reflects the new catalog.
* ``_user_tool_map`` for the owning user is updated.
* ``_tool_map`` (static) is untouched — invariant 1.
* Other users' ``_user_tool_map`` entries are untouched.
Verified by reverting ``_refresh_pool_server_tools`` to call
``_rebuild_tools()`` (the static path) instead of
``_rebuild_user_tool_map(user_id)``: this test fails because
``_tool_map`` is mutated and the user map doesn't pick up the
rotated catalog.
"""
mgr, loop, _ = running_loop_mgr
handler = _make_jsonrpc_handler(
list_tools_seq=[
_list_tools_payload([_tool_spec("v1")]),
_list_tools_payload([_tool_spec("v2")], req_id=2),
],
)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
# Pre-seed a static-path entry to assert it stays untouched.
from turnstone.core.mcp_client import StaticServerState
sentinel_static = StaticServerState(name="static-srv", session=MagicMock())
sentinel_static.tools = [
{"type": "function", "function": {"name": "mcp__static-srv__static_one"}}
]
mgr._static_servers["static-srv"] = sentinel_static
mgr._tool_map["mcp__static-srv__static_one"] = ("static-srv", "static_one")
# Pre-seed an unrelated user's pool entry so we can assert it's
# untouched too.
async def _seed_other_user() -> None:
entry = await mgr._ensure_pool_entry(("user-2", "pool-srv"))
entry.session = MagicMock()
entry.tools = [{"type": "function", "function": {"name": "mcp__pool-srv__user2_only"}}]
mgr._rebuild_user_tool_map("user-2")
_run_on_loop(loop, _seed_other_user())
# Sanity: user-2's map is set.
assert "mcp__pool-srv__user2_only" in (mgr._user_tool_map.get("user-2") or {})
_connect_pool(mgr, loop, user_id="user-1", server_name="pool-srv")
assert mgr.is_mcp_tool("mcp__pool-srv__v1", user_id="user-1") is True
# Trigger a refresh (simulating the notification handler invoking it).
added_removed = _run_on_loop(loop, mgr._refresh_pool_server_tools(("user-1", "pool-srv")))
added, removed = added_removed
assert added == ["mcp__pool-srv__v2"]
assert removed == ["mcp__pool-srv__v1"]
# User-1 sees the new tool.
assert mgr.is_mcp_tool("mcp__pool-srv__v2", user_id="user-1") is True
assert mgr.is_mcp_tool("mcp__pool-srv__v1", user_id="user-1") is False
# Static path untouched — hard invariant 1.
assert mgr._tool_map == {"mcp__static-srv__static_one": ("static-srv", "static_one")}
assert mgr._static_servers["static-srv"].tools == [
{"type": "function", "function": {"name": "mcp__static-srv__static_one"}}
]
# User-2's pool catalog untouched.
assert mgr._user_tool_map["user-2"] == {
"mcp__pool-srv__user2_only": ("pool-srv", "user2_only"),
}
# ---------------------------------------------------------------------------
# Invariant 1 — canonical regression: static path byte-identical when
# oauth_user is disabled.
# ---------------------------------------------------------------------------
def test_static_path_byte_identical_with_oauth_user_disabled(
running_loop_mgr: Any,
) -> None:
"""In a static-only deployment (no oauth_user config rows), the
Phase 7 changes MUST leave every static-path catalog API byte-
identical to pre-Phase-7 behaviour. This is invariant 1.
Drives ``_connect_one`` (static path) and asserts:
* ``get_tools()`` (no user_id) returns the static catalog only.
* ``is_mcp_tool(name)`` (no user_id) works as before.
* ``_tool_map`` is the only catalog index touched.
* No ``_user_*`` state exists (no pool entries / no per-user maps).
Static-path connect uses ``StaticServerState`` instead of pool
entries, so this test seeds it directly rather than driving through
``_connect_one`` (which is exercised by the existing integration
suite). The point is to assert the catalog API contract stays
intact under Phase 7's signature widening.
Verified by widening the default of ``get_tools(user_id)`` to
``user_id=""`` (instead of ``None``) — the test fails because
``get_tools()`` now goes through the per-user merge branch and
starts paying for an empty-loop iteration that pre-Phase-7
callers never saw.
"""
mgr, _loop, _ = running_loop_mgr
# Seed the static-path catalog directly (no pool plumbing needed).
from turnstone.core.mcp_client import StaticServerState
sentinel = StaticServerState(name="static-srv", session=MagicMock())
static_tool = {
"type": "function",
"function": {
"name": "mcp__static-srv__list",
"description": "list",
"parameters": {"type": "object", "properties": {}},
},
}
sentinel.tools = [static_tool]
mgr._static_servers["static-srv"] = sentinel
mgr._tools = [static_tool]
mgr._tool_map["mcp__static-srv__list"] = ("static-srv", "list")
# Pre-Phase-7 caller pattern: no kwargs.
pre_phase7_tools = mgr.get_tools()
assert pre_phase7_tools == [static_tool]
assert mgr.is_mcp_tool("mcp__static-srv__list") is True
assert mgr.is_mcp_tool("nonexistent") is False
# Per-user state does NOT exist when no pool entries are present.
assert mgr._user_pool_entries == {}
assert mgr._user_tool_map == {}
# ---------------------------------------------------------------------------
# R6 verification (production form): a 401 mid-discovery propagates
# rather than hangs. This test pins the empirical conclusion so a future
# SDK upgrade that re-introduces the hang surfaces here.
# ---------------------------------------------------------------------------
def test_pool_connect_list_tools_401_propagates(
running_loop_mgr: Any, monkeypatch: pytest.MonkeyPatch
) -> None:
"""A 401 returned to the discovery POST surfaces as an exception
from ``_connect_one_pool`` within a bounded timeout (no hang).
The SDK propagates the 401 through anyio TaskGroup unwinding (raises
an ``ExceptionGroup`` from the surrounding ``streamablehttp_client``
context); the production reading is "discovery does not need the
carrier-race shape from ``_dispatch_pool_with_entry``."
If a future SDK bump silently re-introduces the hang, this test
will time out — ``_run_on_loop`` is invoked below with a 5s budget.
"""
mgr, loop, _ = running_loop_mgr
handler = _make_jsonrpc_handler(list_tools_status=401)
_build_mock_transport_factory(mgr, monkeypatch, handler)
_patch_tcp_probe(mgr, monkeypatch)
cfg: dict[str, Any] = {
"type": "streamable-http",
"url": "https://mcp.example.com/sse",
"headers": {},
}
capture = _AuthCapture()
fired = asyncio.Event()
t0 = time.monotonic()
# The 401 surfaces as ``ExceptionGroup`` (anyio TaskGroup unwinding
# wraps the underlying ``HTTPStatusError``); both are ``Exception``
# subclasses. The point of the assertion is "the call returned/raised
# within the timeout instead of hanging" — the type-discrimination
# tests cover the carrier classification path.
with pytest.raises(Exception): # noqa: B017 — generic catch documents the no-hang assertion
_run_on_loop(
loop,
mgr._connect_one_pool(
("user-1", "pool-srv"),
cfg,
"access-aaa",
auth_capture=capture,
auth_fired_event=fired,
),
timeout=5,
)
elapsed = time.monotonic() - t0
# Hard upper bound — the spike's happy path closed in ~0.05s, so a
# 401 propagation that takes longer than 5s is the hang regression.
assert elapsed < 5.0, f"_connect_one_pool 401 hang regression: {elapsed:.2f}s"
# Carrier captured the 401 — the response hook fired before the
# SDK propagated the failure, so the dispatcher would have the
# auth signal even though discovery aborted.
assert capture.status == 401, (
f"401 mid-discovery did NOT surface to carrier; capture.status={capture.status}"
)
# ---------------------------------------------------------------------------
# Listener fan-out — RFC §3.3 verification (Chunk 3)
# ---------------------------------------------------------------------------
def test_listener_identity_includes_user_id() -> None:
"""Listener identity is the ``(user_id, callback)`` tuple, not the
callback alone — removing for user A leaves user B's registration
intact even when the same callable was registered for both.
Negative-test reproducer: revert ``add_listener`` /
``remove_listener`` to keep storage as a flat list of callables —
this test fails because removing one registration accidentally
removes BOTH (callable identity is shared).
"""
mgr = MCPClientManager({})
calls = []
def _shared_cb() -> None:
calls.append(time.monotonic())
mgr.add_listener(_shared_cb, user_id="user-A")
mgr.add_listener(_shared_cb, user_id="user-B")
# Removing user-A's registration leaves user-B's intact.
mgr.remove_listener(_shared_cb, user_id="user-A")
# Static-path notify fires both registrations; with user-A removed,
# only user-B fires now.
mgr._notify_listeners()
assert len(calls) == 1, (
f"shared callback fired {len(calls)} times after removing user-A's "
"registration; expected 1 (user-B's registration intact)"
)
# Now remove user-B's; static-path notify fires nobody.
mgr.remove_listener(_shared_cb, user_id="user-B")
calls.clear()
mgr._notify_listeners()
assert calls == [], "callback fired after BOTH user_id-keyed registrations removed"
+5
View File
@@ -199,6 +199,11 @@ class TestLazyConnect:
fake_session = MagicMock()
fake_session.initialize = AsyncMock(return_value=None)
# Phase 7: ``_connect_one_pool`` discovers the user's tool
# catalog after ``initialize()`` returns. This stub returns a
# zero-tool result so the test can keep its narrow focus on
# the bearer-injection contract.
fake_session.list_tools = AsyncMock(return_value=MagicMock(tools=[]))
def _stream_factory(*, url: str, headers: dict[str, str]) -> _AsyncCM:
observed_kwargs["url"] = url
+47
View File
@@ -173,3 +173,50 @@ class TestResolveClient:
def test_unknown_backend_returns_none(self):
client = resolve_web_search_client("typo_backend", tavily_key="key")
assert client is None
def test_resolve_web_search_client_rejects_oauth_user_backend(self):
"""A web_search backend pointing at an ``auth_type=oauth_user``
MCP server MUST be rejected at boot — per-node web_search
cannot carry per-user tokens, so resolving the backend would
guarantee a 401-on-call instead of a clean disablement.
Phase 7 invariant 8 corollary: pool tools are user-scoped;
every entry point that lacks per-user identity (web_search
boot resolver, eval harness, CLI default) MUST refuse them
rather than silently produce a broken client.
Verified by reverting the ``server_auth_type(...) == 'oauth_user'``
guard in ``resolve_web_search_client``: the resolver returns
an ``MCPSearchClient`` whose ``call_tool_sync`` would surface
a 401 / consent_required structured error on every search.
"""
mcp = MagicMock()
mcp.is_mcp_tool.return_value = True # name resolves
mcp.server_auth_type.return_value = "oauth_user"
client = resolve_web_search_client(
"mcp:oauth-search:search", tavily_key=None, mcp_client=mcp
)
assert client is None, (
"oauth_user-backed web_search backend resolved to a non-None client; "
"boot-time guard missing or regressed."
)
# Per-turn callers must read from the in-memory cache, never
# the SQL helper — perf regression guard.
mcp.server_auth_type.assert_called_with("oauth-search")
assert not mcp._lookup_server_row.called, (
"resolver issued a SQL roundtrip via _lookup_server_row; "
"per-turn web_search backend resolution must use the "
"in-memory server_auth_type accessor."
)
def test_resolve_web_search_client_accepts_static_backend(self):
"""Static-path (``auth_type=none`` or ``static``) MCP backends
still resolve cleanly — the new guard ONLY rejects oauth_user.
"""
mcp = MagicMock()
mcp.is_mcp_tool.return_value = True
mcp.server_auth_type.return_value = None
client = resolve_web_search_client(
"mcp:static-search:search", tavily_key=None, mcp_client=mcp
)
assert isinstance(client, MCPSearchClient)
+502 -44
View File
@@ -119,6 +119,16 @@ def _validate_oauth_user_url(url: str) -> None:
# bloating either surface via a thousand-token ``scope=`` value.
_MAX_INSUFFICIENT_SCOPE_REPORTED = 32
# Defensive cap on the number of tools we accept from any single MCP
# server's ``tools/list`` response. Real servers expose at most a few
# dozen tools; a misconfigured or hostile upstream returning thousands
# would amplify both memory (one OpenAI tool dict per entry) and
# downstream BM25 reindex cost. Mirrors the ``_MAX_ERROR_LEN`` /
# ``_MAX_INSUFFICIENT_SCOPE_REPORTED`` defensive ceilings: we truncate
# rather than reject so partial visibility beats zero visibility, and
# emit a warning so operators can investigate.
_MAX_TOOLS_PER_SERVER = 1000
@dataclass
class _AuthCapture:
@@ -247,6 +257,25 @@ def _mcp_to_openai(server_name: str, tool: Any) -> dict[str, Any]:
}
def _cap_server_tools(server_name: str, tools: list[Any]) -> list[Any]:
"""Apply ``_MAX_TOOLS_PER_SERVER`` cap with operator-visible warning.
Identity for inputs at or below the cap (no copy); slice + warn on
overflow. Caller is responsible for converting the returned list to
the OpenAI shape via :func:`_mcp_to_openai`.
"""
if len(tools) <= _MAX_TOOLS_PER_SERVER:
return tools
log.warning(
"MCP server '%s' returned %d tools — truncating to %d "
"(_MAX_TOOLS_PER_SERVER cap). Misconfigured or hostile upstream?",
server_name,
len(tools),
_MAX_TOOLS_PER_SERVER,
)
return tools[:_MAX_TOOLS_PER_SERVER]
# ---------------------------------------------------------------------------
# Per-server state containers
# ---------------------------------------------------------------------------
@@ -366,8 +395,12 @@ class MCPClientManager:
self._last_error: dict[str, str] = {}
self._MAX_ERROR_LEN = 256
# Listener infrastructure (tool-change callbacks for ChatSession)
self._listeners: list[Callable[[], None]] = []
# Listener infrastructure (tool-change callbacks for ChatSession).
# Each entry is ``(user_id, callback)``: ``user_id=None`` is the
# admin / global listener (fires on every tool-change), a string
# ``user_id`` fires only on changes scoped to that user OR on
# global static-path changes. RFC §3.3.
self._listeners: list[tuple[str | None, Callable[[], None]]] = []
self._listeners_lock = threading.Lock()
# Merged resource catalog
@@ -408,6 +441,31 @@ class MCPClientManager:
# so each ``asyncio.Lock`` binds to the correct loop on first use.
self._user_pool_locks: dict[tuple[str, str], asyncio.Lock] = {}
# Per-(user, server) catalog state. Tools live on
# ``PoolEntryState.tools`` (the dataclass already carries the
# field — it's populated lazily by per-user discovery on first
# connect). ``_user_tool_map`` is the per-user prefixed-name
# index, mirroring static ``_tool_map``: outer key is ``user_id``,
# inner is ``prefixed_name → (server_name, original_name)``.
# Resource/prompt mirrors are deferred to Phase 7b — Phase 7
# only lights up the tool path needed for invariant 8.
self._user_tool_map: dict[str, dict[str, tuple[str, str]]] = {}
# Per-user merged tool list (one snapshot per user_id), updated
# atomically alongside ``_user_tool_map`` in
# ``_rebuild_user_tool_map``. ``get_tools(user_id=...)`` reads
# this via a single dict-get (atomic under GIL) so sync-thread
# callers (ChatSession) never iterate ``_user_pool_entries``
# concurrently with the mcp-loop's mutations of the same dict
# (insert in ``_ensure_pool_entry`` / pop in
# ``_close_pool_entry_if_idle`` / ``_evict_session``).
self._user_tools: dict[str, list[dict[str, Any]]] = {}
# Notification debounce for pool sessions, keyed ``(user_id, server)``.
# Mirrors ``_last_notification_refresh`` (static) but per-pool-key
# so a noisy server in one user's pool doesn't suppress a refresh
# in another user's pool of the same server.
self._last_pool_notification_refresh: dict[tuple[str, str], float] = {}
# Pool tuning (read at construction; falls back to defaults).
mcp_cfg = load_config("mcp")
self._user_pool_idle_ttl_s = float(mcp_cfg.get("user_session_idle_ttl_seconds", 600))
@@ -419,6 +477,15 @@ class MCPClientManager:
# never hit it.
self._app_state: Any = None
# In-memory cache of server names whose ``auth_type='oauth_user'``.
# ``_db_servers_to_config`` strips oauth_user rows on the way into
# ``_server_configs``, so neither that dict nor ``_static_servers``
# carries auth_type. This set is populated alongside ``reconcile_sync``
# / ``set_oauth_user_servers`` so per-turn callers (web_search
# backend resolution) can answer "is this server pool-backed?"
# without a SQL roundtrip.
self._oauth_user_server_names: set[str] = set()
# Idle-eviction task handle. Scheduled lazily on the mcp-loop the
# first time a pool entry is created (start() runs before pool
# rows exist, so deferring keeps the task count at zero in
@@ -866,9 +933,8 @@ class MCPClientManager:
# Discover tools
result = await session.list_tools()
server_tools: list[dict[str, Any]] = []
for tool in result.tools:
server_tools.append(_mcp_to_openai(name, tool))
capped = _cap_server_tools(name, result.tools)
server_tools: list[dict[str, Any]] = [_mcp_to_openai(name, tool) for tool in capped]
state.tools = server_tools
self._rebuild_tools()
@@ -975,9 +1041,12 @@ class MCPClientManager:
* Streamable-HTTP transport only — pool servers are remote.
* ``Authorization: Bearer {access_token}`` injected into headers
alongside any operator-supplied static headers.
* No catalog discovery (tools / resources / prompts) and no
notification handler — those land later once per-user catalog
state is in place.
* Tool catalog discovery runs after ``initialize()``; the
notification handler is bound to ``(user_id, server_name)`` so
push-driven ``tools/list_changed`` updates only refresh the
owning user's catalog (the static refresher must NEVER fire
from a pool session — it would clobber static-path state).
Resource / prompt discovery is deferred to Phase 7b.
When ``auth_capture`` is supplied, the underlying ``httpx``
client is built via a factory whose response hook records 401/403
@@ -1064,8 +1133,66 @@ class MCPClientManager:
await self._safe_teardown_on_connect_failure(key, stack)
raise
# Pool-scoped notification handler — bound to ``(user_id,
# server_name)`` via closure so a push-driven
# ``tools/list_changed`` only refreshes THIS user's catalog,
# never the static path's. Calling the static
# ``_refresh_server_tools(server_name)`` from a pool session
# would mutate ``_static_servers[server_name].tools`` and
# ``_tool_map`` — breaking invariant 1 (static-path
# byte-identical) and broadcasting one user's tool view to all
# other sessions. Phase 7 ToolListChangedNotification only;
# resource / prompt branches log + skip until Phase 7b adds
# ``_refresh_pool_server_resources`` / ``_refresh_pool_server_prompts``.
async def _on_pool_notification(
msg: Any, # RequestResponder | ServerNotification | Exception
) -> None:
if not isinstance(msg, mcp_types.ServerNotification):
return
root = msg.root
now = time.monotonic()
last = self._last_pool_notification_refresh.get(key, 0.0)
if now - last < self._NOTIFICATION_DEBOUNCE:
log.debug(
"Debouncing pool notification user=%s server=%s (%.1fs since last refresh)",
user_id,
server_name,
now - last,
)
return
try:
if isinstance(root, mcp_types.ToolListChangedNotification):
log.info(
"Received tools/list_changed from pool user=%s server=%s",
user_id,
server_name,
)
self._last_pool_notification_refresh[key] = now
await self._refresh_pool_server_tools(key)
elif isinstance(root, mcp_types.ResourceListChangedNotification):
log.debug(
"pool resources/list_changed (deferred to Phase 7b) user=%s server=%s",
user_id,
server_name,
)
elif isinstance(root, mcp_types.PromptListChangedNotification):
log.debug(
"pool prompts/list_changed (deferred to Phase 7b) user=%s server=%s",
user_id,
server_name,
)
except Exception:
log.warning(
"Pool refresh after notification failed user=%s server=%s",
user_id,
server_name,
exc_info=True,
)
try:
session = await stack.enter_async_context(ClientSession(read, write))
session = await stack.enter_async_context(
ClientSession(read, write, message_handler=_on_pool_notification) # type: ignore[arg-type]
)
except Exception:
await self._safe_teardown_on_connect_failure(key, stack)
raise
@@ -1089,9 +1216,70 @@ class MCPClientManager:
entry.stack = None
await self._safe_teardown_on_connect_failure(key, stack)
raise
# Discover this user's tool catalog. R6 verified: a 401 here
# propagates through anyio TaskGroup unwinding (raises an
# ``ExceptionGroup`` from the surrounding ``streamablehttp_client``
# context) — plain ``await`` under ``asyncio.timeout`` is
# sufficient. The carrier-race shape used by
# ``_dispatch_pool_with_entry`` defends a different scenario
# (reused-session 401 from inside a SECOND dispatch) that
# doesn't apply to first-connect discovery. Resource / prompt
# discovery deferred to Phase 7b.
#
# Why ``asyncio.timeout``, not ``asyncio.wait_for``: per
# ``feedback_asyncio_timeout_vs_wait_for.md`` and the f6a3b66
# fix, Python 3.11's ``asyncio.wait_for`` wraps the inner
# coroutine in a fresh task. When the SDK's ``streamablehttp_client``
# TaskGroup unwinds (e.g. on a 401), ``aclose`` on the surrounding
# anyio scope runs from a different task than entered it →
# ``RuntimeError("Attempted to exit cancel scope in a different
# task")``. ``asyncio.timeout`` runs the inner coroutine in the
# current task and is the safe shape for any await that may
# traverse anyio cleanup.
try:
async with asyncio.timeout(self._CONNECT_TIMEOUT):
tools_result = await session.list_tools()
except asyncio.CancelledError:
entry.stack = None
task = asyncio.current_task()
if task is not None and task.cancelling():
await self._safe_teardown_on_connect_failure(key, stack)
raise
await self._safe_teardown_on_connect_failure(key, stack)
raise TimeoutError(f"Pool discovery failed for '{server_name}'") from None
except TimeoutError:
entry.stack = None
await self._safe_teardown_on_connect_failure(key, stack)
raise TimeoutError(f"Pool discovery timed out after {self._CONNECT_TIMEOUT}s") from None
except Exception:
entry.stack = None
await self._safe_teardown_on_connect_failure(key, stack)
raise
capped_tools = _cap_server_tools(server_name, tools_result.tools)
entry.tools = [_mcp_to_openai(server_name, tool) for tool in capped_tools]
# Publish session readiness BEFORE catalog visibility.
# ``_rebuild_user_tool_map`` makes ``is_mcp_tool(name, user_id=U)``
# return True for the discovered names; if catalog visibility
# preceded ``entry.session`` assignment, a sync-thread reader
# racing with this coroutine could observe a tool whose backing
# entry has ``session=None``. Defence-in-depth — dispatch
# re-fetches its own token through ``get_user_access_token_classified``
# and lazy-reconnects on session=None — but ordering catches the
# race at the source rather than relying on the dispatch-time
# recovery path.
entry.session = session
entry.last_used = time.monotonic()
self._user_pool_last_used[key] = entry.last_used
# Loop-only mutation; sync-thread readers observe the new tool
# list atomically via the per-user dict-get on ``_user_tools``.
self._rebuild_user_tool_map(user_id)
# Wake user-keyed AND admin (None) listeners; per-user fan-out
# ensures another user's session never observes this user's
# tool change.
self._notify_user_tool_listeners(user_id)
return entry
# -- pool eviction --------------------------------------------------------
@@ -1181,6 +1369,7 @@ class MCPClientManager:
)
except (TimeoutError, asyncio.CancelledError):
return
evicted = False
try:
entry = self._user_pool_entries.get(key)
if entry is None:
@@ -1198,11 +1387,27 @@ class MCPClientManager:
await self._safe_close_stack(stack)
self._user_pool_entries.pop(key, None)
self._user_pool_last_used.pop(key, None)
# Mirror ``_evict_session``'s catalog cleanup: dropping the
# entry without rebuilding ``_user_tool_map`` would leave
# ``is_mcp_tool`` returning True for tools whose backing
# pool is gone, and ChatSession's ``_tools`` would never
# rebuild because no listener fires.
self._last_pool_notification_refresh.pop(key, None)
user_id, _server_name = key
self._rebuild_user_tool_map(user_id)
self._notify_user_tool_listeners(user_id)
evicted = True
finally:
lock.release()
# Drop the now-orphaned lock so the dict doesn't grow without
# bound across the process lifetime. A new key will reallocate.
self._user_pool_locks.pop(key, None)
# bound across the process lifetime — but ONLY on the success
# path. The early-return branches (entry already removed by a
# racing path, or in_flight > 0) leave the lock in place: an
# in-flight dispatcher needs the same lock object on its next
# acquire, and a key whose entry races a concurrent eviction
# will be re-allocated by ``_ensure_pool_entry`` next time.
if evicted:
self._user_pool_locks.pop(key, None)
# -- failure classification (pool dispatch) ------------------------------
@@ -1270,6 +1475,52 @@ class MCPClientManager:
self._tool_map = new_map
self._notify_listeners()
def _rebuild_user_tool_map(self, user_id: str) -> None:
"""Rebuild the per-user prefixed-name index from pool entries.
Mirrors :meth:`_rebuild_tools` for one user's pool entries:
scan ``_user_pool_entries`` for keys whose first element matches
``user_id``, materialize a fresh ``prefixed_name → (server, original)``
dict and a parallel tool-list, assign each to its dict in
``_user_tool_map`` and ``_user_tools`` respectively. Each
per-key write is individually atomic under the GIL; the two
writes happen back-to-back on the mcp-loop with no awaits
between them, so a sync-thread reader cannot interleave at the
Python statement level — but the cross-dict write is not a
single atomic operation. In practice the window is sub-microsecond
and the listener fan-out (which fires AFTER both writes
complete) is the trigger for any session-side rebuild that
might re-read both dicts.
``_tool_map`` (the static index) is NEVER mutated here, so
invariant 1 (static path byte-identical) is preserved.
Empty rebuilds drop the user_id key from BOTH dicts so an idle
user with no pool entries doesn't retain permanent empty-list
sentinels.
MUST run on the mcp-loop. The pool dict scan here cannot race
with sync-thread reads because sync threads never touch
``_user_pool_entries``; they read ``_user_tools`` instead.
"""
new_map: dict[str, tuple[str, str]] = {}
new_tools: list[dict[str, Any]] = []
for (uid, _server_name), entry in self._user_pool_entries.items():
if uid != user_id or entry.tools is None:
continue
for tool in entry.tools:
prefixed: str = tool["function"]["name"]
# Extract original name from the mcp__server__original pattern.
original = prefixed.split("__", 2)[2] if prefixed.count("__") >= 2 else prefixed
new_map[prefixed] = (_server_name, original)
new_tools.append(tool)
if new_map:
self._user_tool_map[user_id] = new_map
self._user_tools[user_id] = new_tools
else:
self._user_tool_map.pop(user_id, None)
self._user_tools.pop(user_id, None)
async def _refresh_server_tools(self, name: str) -> tuple[list[str], list[str]]:
"""Re-fetch tools for one server. Returns ``(added, removed)`` names."""
state = self._static_servers.get(name)
@@ -1283,7 +1534,8 @@ class MCPClientManager:
old_names = {t["function"]["name"] for t in state.tools}
result = await session.list_tools()
server_tools = [_mcp_to_openai(name, tool) for tool in result.tools]
capped = _cap_server_tools(name, result.tools)
server_tools = [_mcp_to_openai(name, tool) for tool in capped]
new_names = {t["function"]["name"] for t in server_tools}
state.tools = server_tools
@@ -1300,6 +1552,51 @@ class MCPClientManager:
)
return added, removed
async def _refresh_pool_server_tools(self, key: tuple[str, str]) -> tuple[list[str], list[str]]:
"""Re-fetch tools for one pool entry. Returns ``(added, removed)`` names.
Mirror of :meth:`_refresh_server_tools` for the pool path:
targets a single ``(user_id, server_name)`` entry, mutates ONLY
``entry.tools`` + the per-user index, fires the user-scoped
listener fan-out. Static-path catalogs are NEVER touched, so
invariant 1 (static path byte-identical) is preserved.
MUST run on the mcp-loop. Caller does not need to hold
``open_lock`` — ``_refresh_pool_server_tools`` is invoked from
the pool session's notification handler, which already runs in
the SDK's receive task on the loop.
"""
entry = self._user_pool_entries.get(key)
if entry is None or entry.session is None:
raise RuntimeError(f"Pool entry {key!r} is not connected")
# Snapshot session locally — a concurrent transport-error
# eviction can clear ``entry.session`` after our reads, so a
# post-await ``entry.session.<...>`` call would raise
# ``AttributeError``. Mirrors the static path's pattern.
session = entry.session
user_id, server_name = key
old_names = {t["function"]["name"] for t in (entry.tools or [])}
result = await session.list_tools()
capped = _cap_server_tools(server_name, result.tools)
server_tools = [_mcp_to_openai(server_name, tool) for tool in capped]
new_names = {t["function"]["name"] for t in server_tools}
entry.tools = server_tools
self._rebuild_user_tool_map(user_id)
# Fire user-keyed AND admin (None) listeners. Other users'
# listeners do NOT see this change — the pool catalog is private.
self._notify_user_tool_listeners(user_id)
added = sorted(new_names - old_names)
removed = sorted(old_names - new_names)
if added or removed:
log.info(
"Refreshed pool MCP server user=%s server=%s: +%d/-%d tool(s)",
user_id,
server_name,
len(added),
len(removed),
)
return added, removed
async def _refresh_server(self, name: str) -> tuple[list[str], list[str]]:
"""Re-fetch tools, resources, and prompts for one server.
@@ -1533,20 +1830,52 @@ class MCPClientManager:
# -- listener infrastructure ---------------------------------------------
def add_listener(self, callback: Callable[[], None]) -> None:
"""Register a callback invoked when the tool list changes."""
with self._listeners_lock:
self._listeners.append(callback)
def add_listener(self, callback: Callable[[], None], *, user_id: str | None = None) -> None:
"""Register a callback invoked when the tool list changes.
def remove_listener(self, callback: Callable[[], None]) -> None:
"""Unregister a tool-change callback."""
``user_id=None`` (default) registers a global / admin listener
that fires on every tool-change (static or pool). A string
``user_id`` scopes the listener: it fires on global static-path
changes AND on pool changes for that user only — never on
another user's pool change. RFC §3.3.
"""
with self._listeners_lock:
self._listeners.append((user_id, callback))
def remove_listener(self, callback: Callable[[], None], *, user_id: str | None = None) -> None:
"""Unregister a tool-change callback.
``user_id`` MUST match the value used at registration; the
``(user_id, callback)`` pair is the listener identity.
"""
with self._listeners_lock, contextlib.suppress(ValueError):
self._listeners.remove(callback)
self._listeners.remove((user_id, callback))
def _notify_listeners(self) -> None:
"""Invoke all registered tool-change listeners."""
"""Static-path tool change — fires ALL registered listeners.
The static catalog is a process-wide concern: a static-server
notification (or reconcile) updates ``_tool_map`` which every
user-scoped session relies on, so the fan-out is unconditional.
"""
with self._listeners_lock:
listeners = list(self._listeners)
for _uid, cb in listeners:
try:
cb()
except Exception:
log.warning("Tool-change listener raised", exc_info=True)
def _notify_user_tool_listeners(self, user_id: str) -> None:
"""Pool-entry tool change — fires only listeners that should see it.
A pool catalog change touches one user's view; firing every
registered listener would broadcast that user's tool set to
unrelated sessions. Scoped fan-out targets the matching
``user_id`` AND admin (``None``) listeners. RFC §3.3.
"""
with self._listeners_lock:
listeners = [cb for uid, cb in self._listeners if uid == user_id or uid is None]
for cb in listeners:
try:
cb()
@@ -1807,6 +2136,16 @@ class MCPClientManager:
"""Connect a new MCP server at runtime (blocks the calling thread).
Returns status dict with keys: connected, tools, resources, prompts, error.
Note on ``_oauth_user_server_names``: this method does NOT update
the oauth_user-name cache. ``_db_servers_to_config`` strips
``auth_type='oauth_user'`` rows before this method is reached, so
production callers (only :meth:`reconcile_sync`) never pass an
oauth_user cfg here — the cache is rebuilt wholesale by
``reconcile_sync`` at the top of every reconcile, which is the
canonical update point. Direct test callers passing an oauth_user
cfg would leave the cache stale; route through ``reconcile_sync``
instead.
"""
if "__" in name:
return {
@@ -1929,6 +2268,17 @@ class MCPClientManager:
with notification handlers and refresh tasks.
Returns True if the server was connected and successfully removed.
Note on ``_oauth_user_server_names``: this method does NOT discard
``name`` from the oauth_user-name cache. The cache is the source of
truth for "this server exists in the DB as oauth_user", not "this
server is connected on the static path" — those are separate
concerns. A static→oauth_user transition during reconcile_sync
keeps ``name`` in the cache (the new identity) AND calls this
method to drop the old static-path connection; discarding here
would silently make web_search resolve to a now-pool-only server.
The cache is updated only by :meth:`reconcile_sync`'s wholesale
rebuild at line 2340.
"""
existing = self._static_servers.get(name)
was_connected = existing is not None and existing.session is not None
@@ -2039,6 +2389,14 @@ class MCPClientManager:
log.warning("reconcile_sync: failed to read mcp_servers table", exc_info=True)
return {"added": [], "removed": [], "updated": []}
# Refresh the in-memory oauth_user name cache from the rows we
# just read — feeds :meth:`server_auth_type` so callers (e.g.
# ``web_search.resolve_web_search_client``) avoid a per-turn SQL
# roundtrip.
self._oauth_user_server_names = {
row["name"] for row in rows if row.get("auth_type") == "oauth_user"
}
desired = _db_servers_to_config(rows)
desired_names = set(desired)
@@ -2093,9 +2451,29 @@ class MCPClientManager:
# -- query methods -------------------------------------------------------
def get_tools(self) -> list[dict[str, Any]]:
"""Return MCP tools in OpenAI function-calling format."""
return [dict(t) for t in self._tools]
def get_tools(self, user_id: str | None = None) -> list[dict[str, Any]]:
"""Return MCP tools in OpenAI function-calling format.
``user_id=None`` returns the global static-path catalog only —
the legacy behaviour every pre-Phase-7 caller relies on. A
string ``user_id`` returns the merged view: static catalog
followed by that user's cached pool tools, read from
``_user_tools`` (a single dict-get, atomic under GIL). The
list snapshot is materialised on the mcp-loop by
:meth:`_rebuild_user_tool_map` and replaced atomically — sync
threads never iterate ``_user_pool_entries`` directly, so the
mcp-loop is free to insert/pop entries concurrently.
Callers that don't carry a session-bound user (boot-time
logging, web_search backend resolution) MUST use the default.
Returned dicts are shallow-copied; nested objects are shared
with the manager's catalog, mirroring the pre-Phase-7 contract.
"""
base = [dict(t) for t in self._tools]
if user_id is None:
return base
base.extend(dict(t) for t in self._user_tools.get(user_id, []))
return base
def get_resources(self) -> list[dict[str, Any]]:
"""Return discovered MCP resources (shallow-copied dicts)."""
@@ -2115,24 +2493,47 @@ class MCPClientManager:
"""Number of discovered prompts (no allocation)."""
return len(self._prompts)
def is_mcp_tool(self, func_name: str) -> bool:
def is_mcp_tool(self, func_name: str, *, user_id: str | None = None) -> bool:
"""Check whether *func_name* belongs to an MCP server.
Caveat: only static-path tools (``auth_type ∈ {none, static}``)
populate ``_tool_map``; pool tools become reachable via this
method only once per-user catalog scoping lands. Until then,
pool dispatch is reachable from production callers like
``ChatSession._exec_mcp_tool`` only when the LLM produces a
``mcp__{server}__{tool}`` name that bypasses ``is_mcp_tool``-
style gating, or via direct ``call_tool_sync`` with a known
prefixed name (the path the new pool tests use).
``user_id=None`` (default) asks "is this a static-path tool?"
the answer is process-global. Boot-time and per-node callers
(e.g. ``resolve_web_search_client``) MUST use the default
because they don't carry a session-bound user identity.
A string ``user_id`` extends the lookup to that user's pool
catalog: returns ``True`` if ``func_name`` is in either the
static map OR the user's per-user tool map. Pool tools become
reachable via this gate ONLY once a session-bound caller
threads its ``user_id`` through; CLI sessions default
``user_id=""`` and so cannot see pool tools — documented
limitation.
"""
return func_name in self._tool_map
if func_name in self._tool_map:
return True
if user_id is None:
return False
user_map = self._user_tool_map.get(user_id)
return user_map is not None and func_name in user_map
def is_mcp_prompt(self, name: str) -> bool:
"""Check whether *name* is a known MCP prompt."""
return name in self._prompt_map
def server_auth_type(self, server_name: str) -> str | None:
"""Return ``'oauth_user'`` for pool-backed servers, else ``None``.
In-memory accessor for the per-turn callers that need to
distinguish pool-backed servers from static-path ones without a
SQL roundtrip. ``None`` means "either static-path or unknown"
the boot-time / per-node web_search resolver only uses this as
a defence-in-depth gate, so a missing-cache miss is safe (the
outer ``is_mcp_tool`` check already proves the server is in
``_tool_map``, which by construction excludes oauth_user).
Populated by ``reconcile_sync`` and ``create_mcp_client``.
"""
return "oauth_user" if server_name in self._oauth_user_server_names else None
@property
def server_count(self) -> int:
return sum(1 for s in self._static_servers.values() if s.session is not None)
@@ -2605,11 +3006,28 @@ class MCPClientManager:
# ``streamablehttp_client`` TaskGroup — see
# :meth:`_dispatch_pool_sync` for the architectural
# rationale.
log.debug("mcp_pool.auth_401_initial", exc_info=exc)
#
# exc_info=False: the underlying httpx exception
# carries ``request.headers["authorization"]`` with
# the rejected bearer; standard tracebacks don't
# render locals but Sentry / faulthandler hooks
# capture frame state. Structured fields below
# provide the diagnostic signal without the secret.
log.debug(
"mcp_pool.auth_401_initial server=%s user=%s exc=%s",
server_name,
user_id,
type(exc).__name__,
)
raise _PoolDispatchRetryRequested from None
# retry_count == 1 — refreshed bearer also rejected;
# emit consent_required so the user/operator re-grants.
log.debug("mcp_pool.auth_401_retry_failed", exc_info=exc)
log.debug(
"mcp_pool.auth_401_retry_failed server=%s user=%s exc=%s",
server_name,
user_id,
type(exc).__name__,
)
return _structured_error(
code="mcp_consent_required",
server=server_name,
@@ -2617,7 +3035,12 @@ class MCPClientManager:
)
if classification == "auth_403":
self._evict_session(key)
log.debug("mcp_pool.auth_403", exc_info=exc)
log.debug(
"mcp_pool.auth_403 server=%s user=%s exc=%s",
server_name,
user_id,
type(exc).__name__,
)
return await self._handle_auth_403(
user_id=user_id,
server_name=server_name,
@@ -2627,7 +3050,15 @@ class MCPClientManager:
if classification == "transport":
self._cb_record_failure(server_name)
self._evict_session(key)
log.debug("mcp_pool.transport_failure", exc_info=exc)
# See auth_401 branch for why exc_info=False — pool
# dispatch errors all share the same request object
# whose headers carry the bearer.
log.debug(
"mcp_pool.transport_failure server=%s user=%s exc=%s",
server_name,
user_id,
type(exc).__name__,
)
raise
# protocol / other — don't trip the breaker.
raise
@@ -2636,18 +3067,40 @@ class MCPClientManager:
return result
def _evict_session(self, key: tuple[str, str]) -> None:
"""Drop the cached session on a pool entry. Stack/streams left for reconnect.
"""Drop the cached session AND catalog on a pool entry.
Auth/transport branches both call this — the next connect's
``_connect_one_pool`` tears down the stale stack lazily via the
stale-entry guard at the top of the method. Closing eagerly
from here is incorrect under cancellation: ``stack.aclose()``
must run inside the same anyio scope it was entered in, which
the next connect arranges.
Stack/streams left for reconnect. Auth/transport branches both
call this — the next connect's ``_connect_one_pool`` tears
down the stale stack lazily via the stale-entry guard at the
top of the method. Closing eagerly from here is incorrect
under cancellation: ``stack.aclose()`` must run inside the
same anyio scope it was entered in, which the next connect
arranges.
Catalog cleanup (Phase 7): clearing ``entry.tools`` here
ensures an evicted-then-not-yet-reconnected entry contributes
no stale entries to ``_user_tool_map``. Without this,
``is_mcp_tool(name, user_id=user_id)`` could return True for a
name whose backing pool is gone, then ``_resolve_pool_target``
would dispatch into a session-less entry and the next call
would surface as a generic transport error instead of a
clean reconnect path.
"""
evict = self._user_pool_entries.get(key)
if evict is not None:
evict.session = None
evict.tools = None
# Prune the debounce stamp in lockstep with the entry — the
# dict grows otherwise (slow leak) across (user, server)
# churn. Mirrors the cleanup in ``_close_pool_entry_if_idle``.
self._last_pool_notification_refresh.pop(key, None)
user_id, _server_name = key
self._rebuild_user_tool_map(user_id)
# Wake the user's session so its merged tool list shrinks
# back to static-only until the next connect populates the
# entry. Admin (None) listeners also fire — operator
# tooling tracking pool catalog state observes the drop.
self._notify_user_tool_listeners(user_id)
async def _handle_auth_403(
self,
@@ -3129,11 +3582,15 @@ def create_mcp_client(
"""
# Check DB first to know which servers are DB-managed
db_names: set[str] = set()
oauth_user_names: set[str] = set()
if storage is not None:
try:
rows = storage.list_mcp_servers(enabled_only=True)
if rows:
db_names = {r["name"] for r in rows}
# Cache oauth_user names so per-turn callers (web_search
# backend resolution) can answer auth_type without SQL.
oauth_user_names = {r["name"] for r in rows if r.get("auth_type") == "oauth_user"}
except Exception:
log.warning("Failed to load DB-managed MCP servers", exc_info=True)
@@ -3144,5 +3601,6 @@ def create_mcp_client(
mgr = MCPClientManager(servers)
# Mark DB-sourced servers so reconcile_sync won't remove config-file servers
mgr._db_managed = {name for name in servers if name in db_names}
mgr._oauth_user_server_names = oauth_user_names
mgr.start()
return mgr
+52 -10
View File
@@ -712,7 +712,24 @@ class ChatSession:
self.debug = False
self.auto_approve = False
self._node_id = node_id
# ``user_id`` is the authenticated principal's UUID for HTTP-borne
# sessions (server, console) and the empty string ``""`` for CLI
# / eval / unauthenticated callers. The empty string is a SENTINEL
# that collapses to ``None`` for MCPClientManager's optional
# ``user_id`` arguments (cached once below as ``_mcp_user_id``),
# which short-circuits per-user OAuth pool lookup → CLI / eval
# sessions cannot use ``auth_type=oauth_user`` MCP servers
# (no per-request user identity to attach a bearer to). End
# users running such servers must pre-link via the web UI;
# static-path MCP servers continue to work in all session
# contexts.
self._user_id = user_id
# ``_user_id`` is set once here and never reassigned, so we
# compute the MCP-API form (empty-string-to-None collapse) once
# rather than re-asserting the invariant at each call site.
# Listener identity, catalog merge, and dispatch all consume
# this through ``_mcp_user_id``.
self._mcp_user_id: str | None = user_id or None
self._username = username
self._client_type = client_type
self._config_store = config_store
@@ -841,13 +858,15 @@ class ChatSession:
self._task_tools = []
self._agent_tools = []
elif mcp_client:
mcp_tools = mcp_client.get_tools()
mcp_tools = mcp_client.get_tools(user_id=self._mcp_user_id)
self._tools = merge_mcp_tools(INTERACTIVE_TOOLS, mcp_tools)
self._task_tools = merge_mcp_tools(TASK_AGENT_TOOLS, mcp_tools)
self._agent_tools = merge_mcp_tools(AGENT_TOOLS, mcp_tools)
# Register for tool-change notifications from MCP servers
# Register for tool-change notifications from MCP servers.
# ``user_id`` is the listener identity component — pool-only
# changes for OTHER users must not fire this callback.
self._mcp_refresh_cb = self._on_mcp_tools_changed
mcp_client.add_listener(self._mcp_refresh_cb)
mcp_client.add_listener(self._mcp_refresh_cb, user_id=self._mcp_user_id)
# Register for resource-change notifications
self._mcp_resource_cb = self._on_mcp_resources_changed
mcp_client.add_resource_listener(self._mcp_resource_cb)
@@ -1203,7 +1222,11 @@ class ChatSession:
# is fixed at COORDINATOR_TOOLS. Ignore MCP server changes.
if self._kind == WorkstreamKind.COORDINATOR:
return
mcp_tools = self._mcp_client.get_tools()
# Phase 7: pass session-bound user_id so the merged tool list
# includes this user's pool catalog. The static path is included
# by ``get_tools`` regardless; ``user_id=None`` would silently
# drop pool tools that the LLM is allowed to call.
mcp_tools = self._mcp_client.get_tools(user_id=self._mcp_user_id)
self._tools = merge_mcp_tools(INTERACTIVE_TOOLS, mcp_tools)
self._task_tools = merge_mcp_tools(TASK_AGENT_TOOLS, mcp_tools)
self._agent_tools = merge_mcp_tools(AGENT_TOOLS, mcp_tools)
@@ -1369,7 +1392,11 @@ class ChatSession:
if self._judge_cancel_event is not None:
self._judge_cancel_event.set()
if self._mcp_client and self._mcp_refresh_cb:
self._mcp_client.remove_listener(self._mcp_refresh_cb)
# ``user_id`` MUST match the value used at registration —
# the listener identity is ``(user_id, callback)``, not
# callback alone. ``self._user_id`` is set once in
# ``__init__`` and never mutated, so identity is stable.
self._mcp_client.remove_listener(self._mcp_refresh_cb, user_id=self._mcp_user_id)
self._mcp_refresh_cb = None
if self._mcp_client and self._mcp_resource_cb:
self._mcp_client.remove_resource_listener(self._mcp_resource_cb)
@@ -4725,13 +4752,24 @@ class ChatSession:
}
preparer = preparers.get(func_name)
if not preparer:
# Check if this is an MCP tool
if self._mcp_client and self._mcp_client.is_mcp_tool(func_name):
# Check if this is an MCP tool. Phase 7: pass session-bound
# ``user_id`` so per-user pool tools become reachable here —
# without this kwarg the gate stays static-only and pool
# dispatch is structurally unreachable from
# ``ChatSession._prepare_tool`` (RFC §3, invariant 8).
if self._mcp_client and self._mcp_client.is_mcp_tool(
func_name, user_id=self._mcp_user_id
):
return self._prepare_mcp_tool(call_id, func_name, args)
self.ui.on_error(f"Model called unknown tool: {func_name!r}")
available = list(preparers)
if self._mcp_client:
available.extend(sorted(self._mcp_client._tool_map))
available.extend(
sorted(
t["function"]["name"]
for t in self._mcp_client.get_tools(user_id=self._mcp_user_id)
)
)
return {
"call_id": call_id,
"func_name": func_name,
@@ -7782,7 +7820,7 @@ class ChatSession:
output = self._mcp_client.call_tool_sync(
func_name,
args,
user_id=self._user_id or None,
user_id=self._mcp_user_id,
timeout=self.tool_timeout,
)
except TimeoutError:
@@ -10148,7 +10186,11 @@ class ChatSession:
elif arg and arg.split()[0] == "refresh":
self._handle_mcp_refresh(arg)
else:
tools = self._mcp_client.get_tools()
# Phase 7: pass session-bound user_id so the /mcp listing
# surfaces this user's pool tools alongside the static
# catalog. Resource / prompt query stays user_id-less
# (deferred to Phase 7b).
tools = self._mcp_client.get_tools(user_id=self._mcp_user_id)
resources = self._mcp_client.get_resources()
prompts = self._mcp_client.get_prompts()
mcp_lines = []
+20
View File
@@ -168,7 +168,27 @@ def resolve_web_search_client(
if len(parts) == 3 and mcp_client is not None:
_, server, tool = parts
prefixed = f"mcp__{server}__{tool}"
# Boot-time gate: ``is_mcp_tool`` without ``user_id`` returns
# True only for static-path catalogs. Pool-backed
# (``auth_type=oauth_user``) servers are NEVER reachable via
# the per-node web_search client because the boot-time
# resolver has no per-user identity to attach a bearer to —
# the resolved client would be shared across requests, but
# the bearer can't be (RFC §3, invariant 8 corollary).
if mcp_client.is_mcp_tool(prefixed):
# Defence-in-depth: even if a future change widens
# ``_tool_map`` to include oauth_user names by accident,
# refuse the backend explicitly. ``server_auth_type``
# is an in-memory accessor — this resolver is invoked
# per LLM turn, so a SQL hop here would amplify token
# cost on every chat round.
if mcp_client.server_auth_type(server) == "oauth_user":
log.warning(
"web_search_backend %r points at oauth_user MCP server; "
"per-node web search cannot use per-user tokens — disabling",
backend,
)
return None
return MCPSearchClient(mcp_client, prefixed, timeout=timeout)
return None