fix(mcp): follow tools/list pagination so later pages are not dropped

[Johnx-w]

Description

POST /mcp called ClientSession.list_tools() once. The MCP Python SDK returns a single page and leaves nextCursor to the caller, so servers that paginate tools/list silently dropped later tools from the connect payload and the composer MCP list.

This walks nextCursor (and the next_cursor alias), caps the walk at 100 pages, and stops if a cursor repeats.

Why

MCP pagination is part of the spec; list_tools() does not auto-paginate. The UI only sees the tools returned here.

Testing

• uv run pytest backend/tests/test_mcp.py::TestListAllMcpTools
• uv run pytest backend/tests/test_mcp.py::TestConnectMcpEndpoint::test_connect_follows_list_tools_pagination (needs frontend build)
• uv run scripts/lint.py backend/chainlit/mcp.py backend/chainlit/server.py backend/tests/test_mcp.py
• uv run scripts/format.py --check on the same files

---

Summary by cubic

Fixes POST /mcp so paginated MCP servers no longer silently drop later tools from the connect payload and composer MCP list, and ensures a failed or hung tools/list cannot evict a working session.

• Walks nextCursor (and the next_cursor alias) until the server omits it, capped at 100 pages and stopping on a repeated cursor.
• Bounds each page and the whole walk with the connect timeout, so a server that never finishes paginating cannot hang the connect handler.
• Lists tools before swapping in the replacement session, so a reconnect that fails during listing keeps the old session intact.
• Adds tests covering pagination, cursor aliases, repeated cursors, the page cap, timeouts, and session eviction.

^{Written for commit de819a6. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 3 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

2 issues found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="backend/chainlit/server.py">

<violation number="1" location="backend/chainlit/server.py:1825">
P2: When `/mcp` disconnects the same name while pagination is in progress, the later `swap_mcp_session` recreates the session after DELETE has completed. Coordinate pending connects with disconnects, or invalidate the in-flight connect before allowing it to store the new session.</violation>

<violation number="2" location="backend/chainlit/server.py:1830">
P1: When the request is cancelled during `list_all_mcp_tools`, `asyncio.CancelledError` bypasses this `except Exception`, so `stop_mcp_task` never runs. Catch cancellation, stop the task, then re-raise; the pre-swap listing otherwise leaves the transport unreachable from `mcp_sessions`.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P1: When the request is cancelled during list_all_mcp_tools, asyncio.CancelledError bypasses this except Exception, so stop_mcp_task never runs. Catch cancellation, stop the task, then re-raise; the pre-swap listing otherwise leaves the transport unreachable from mcp_sessions.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At backend/chainlit/server.py, line 1830:

<comment>When the request is cancelled during `list_all_mcp_tools`, `asyncio.CancelledError` bypasses this `except Exception`, so `stop_mcp_task` never runs. Catch cancellation, stop the task, then re-raise; the pre-swap listing otherwise leaves the transport unreachable from `mcp_sessions`.</comment>

<file context>
@@ -1821,11 +1821,41 @@ async def _mcp_session_runner() -> None:
+            page_timeout=connect_timeout,
+            total_timeout=connect_timeout,
+        )
+    except Exception as e:
+        # Same teardown as a rejecting on_mcp_connect: tools/list is part of
+        # making the connection usable, and it runs *before* swap so a
</file context>

Suggested change

	except Exception as e:
	except asyncio.CancelledError:
	await stop_mcp_task(task, stop_event, payload.name)
	raise
	except Exception as e:

[cubic-dev-ai]

P2: When /mcp disconnects the same name while pagination is in progress, the later swap_mcp_session recreates the session after DELETE has completed. Coordinate pending connects with disconnects, or invalidate the in-flight connect before allowing it to store the new session.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At backend/chainlit/server.py, line 1825:

<comment>When `/mcp` disconnects the same name while pagination is in progress, the later `swap_mcp_session` recreates the session after DELETE has completed. Coordinate pending connects with disconnects, or invalidate the in-flight connect before allowing it to store the new session.</comment>

<file context>
@@ -1821,11 +1821,41 @@ async def _mcp_session_runner() -> None:
             )
 
+    try:
+        tools = await list_all_mcp_tools(
+            mcp_client_session,
+            page_timeout=connect_timeout,
</file context>